Welcome to METEO 410

Welcome to METEO 410
Looking for the lesson content? Registered METEO 410 students can access and navigate through the lessons in the "Lessons" menu (you might need to log in with your PSU user ID and password).

Quick Facts about METEO 410

METEO 410 is the capstone course in Penn State's online program that confers a Certificate of Achievement in Weather Forecasting (opens in a new window). It is offered every Fall (August - December) and Spring (January - May) semester.

Course Prerequisite(s): METEO 101, METEO 241, METEO 361

Course Overview

METEO 410 is designed specifically for adult students seeking a Certificate of Achievement in Weather Forecasting. The course will review and expand upon major topics and themes in its prerequisite courses, and emphasize the application of conceptual models and advanced forecasting strategies in making deterministic weather forecasts. Student participation in the national collegiate weather forecasting competition (WxChallenge (opens in a new window)) is required and will be the foundation for various types of course assessment.

Photograph of Elliot Abrams presenting a television forecast in the 1970s.

Elliot Abrams presents a weather forecast on Penn State's Weather World television program in the 1970s. A lot has changed since the days of hand-drawn maps and low-resolution satellite images!
Credit: Penn State Department of Meteorology

Why learn about advanced topics in weather forecasting?

In recent years, weather forecasts have become ubiquitous. The days when weather forecasts could only be found in the morning newspaper, on the radio or on television are long gone. Today, local weather forecasts are easily found "on demand" via your favorite weather website or weather app on your smartphone or tablet. With weather forecasts so easily accessible, why bother learning about forecasting?

First, all weather forecasts are fallible. Yes, your favorite weather app is wrong sometimes! No person or computer can create perfect forecasts consistently. Second, publicly available forecasts from various media outlets, websites, and apps, don't meet everyone's specific needs. Even as computer-generated forecasts continue to improve, skilled weather forecasters can still outperform the models, and a need still exists for humans to understand weather forecasting and effectively communicate key aspects of the forecast to the public (or clients).

What will you learn in this course?

Your exploration of advanced topics in weather forecasting will begin by learning about the basics of weather forecasting and the attributes of a good forecaster. By the end of the course, you'll be making your own deterministic weather forecasts and critically analyzing forecast mistakes. Along the way, you'll also learn about the many factors that forecasters have to consider when making decisions, and see how these factors impact the forecast through numerous case studies. While "learning by doing" is a major component of METEO 410 (through making your own forecasts), you'll also learn about a number of forecast tools and advanced strategies, as the course outline below demonstrates:

Lesson 1: Forecasting Fundamentals (introduction to various types of forecasts and verification metrics, the philosophy and attributes of a good forecaster, observation of surface meteorological variables, decoding METARS, important aspects of local climatology)

Lesson 2: Ensemble Forecasting (model jumpiness and lagged-average forecasts, generation of ensemble forecasts, commonly used ensemble systems, statistics needed for ensemble interpretation, interpreting ensemble forecast products)

Lesson 3: Statistical Models (utility of statistical models, Model Output Statistics (MOS) tables, MOS temperature, wind, and precipitation predictors, MOS performance near climatological extremes, LAMP guidance, MOS reliance on dynamic models, the National Blend of Models (NBM), other statistically-post processed forecast products)

Lesson 4: Forecasting Temperature--Controllers and Strategies: (local climate factors, impacts of various cloud types, wind, and water on temperature forecasts, utilizing statistical temperature guidance in the context of the synoptic-scale pattern, recognition of problematic synoptic patterns, the Delta Method, the Analogous Thickness Method)

Lesson 5: Forecasting Wind--Controllers and Strategies (local wind climatology data, large-scale forces and the impact of latitude on wind speed, the impact of vertical mixing and mesoscale circulations (sea- and lake-breezes and mountain-valley circulations) on wind forecasts, the isallobaric wind)

Lesson 6: Forecasting Precipitation--Controllers and Strategies (sources of synoptic-scale upward motion and incorporating the synoptic-scale pattern into your forecast, quantifying vertical motion, model QPF and its weaknesses, consensus forecasting, understanding precipitation climatology)

Lesson 7: Introduction to Medium- and Long-Range Forecasting (selecting a forecast format, medium- and long-range computer guidance and ensembles, teleconnections, the El Niño Southern Oscillation, the Arctic and North Atlantic Oscillations, the Madden-Julian Oscillation, the Pacific-North American Pattern)

How does this course work?

Much like the other courses in the Certificate of Achievement in Weather Forecasting program, all course materials are presented online. The course lessons include many animations and interactive tools to provide a tactile, visual component to your learning. METEO 410 also has a highly interactive component focused on creating and discussing weather forecasts for the WxChallenge forecasting competition. Your instructor will assess your progress through lab exercises, forecast analysis projects, and your forecasting performance in WxChallenge. You should expect to log into the course on several days each week in order to participate fully and meet forecast and discussion deadlines during the portion of the course focused on WxChallenge. In general, the course requires a total of 8 to 10 hours per week studying the lesson material and/or creating and discussing weather forecasts.

atb3

Lesson 1. Forecasting Fundamentals

Lesson 1. Forecasting Fundamentals

Motivate...

In your coursework so far, you have focused on learning how the atmosphere works and how to analyze synoptic and mesoscale weather. But, in this course, it's time to put that knowledge to work and make some forecasts! I believe that "learning by doing" represents the highest form of learning, and in this course, you'll be "getting your hands dirty" as you apply many of the concepts from past courses to real weather situations in order to make actual forecasts.

As you've learned, weather forecasters today have a wide array of computer model guidance to work with. Therefore, being able to correctly interpret this guidance is a critical part of forecasting. Sounds easy enough, right? Anyone can read the models with a bit of training, but being a really good forecaster is much more difficult. It requires solid conceptual models of how the atmosphere works and years of experience, because models are far from perfect. Even as computing power increases and model performance improves, models produce highly-detailed, but still erroneous simulations of future weather. Having sound conceptual models and years of experience keeps us from being completely at the mercy of model predictions and gives us a chance to create better forecasts than a computer model can.

30-hour forecast of simulated radar reflectivity

A 30-hour computer model forecast for simulated radar reflectivity and 10-meter winds valid at 18Z on August 10, 2020 (from the 12Z run on August 9). Was this model prediction exactly right? Spoiler alert: no.

As a quick example, take the 30-hour computer model forecast valid at 18Z on August 10, 2020 (above), and focus on Iowa. Now compare it to the 18Z radar image (opens in a new window) for verification. The handful of isolated thunderstorms predicted to be over western Iowa at 18Z actually turned out to be a wicked derecho racing across the Midwest, which spawned more than 30 reports of wind gusts of 75 miles per hour or more and a few wind gusts in excess of 100 miles per hour (SPC filtered storm reports (opens in a new window)). Two weeks later, thousands of Iowans were still without power! That's a pretty significant model forecast bust only a day in advance! As a forecaster, if you live solely by the models, you'll die by the models.

How much experience makes for a good forecaster? Well, no specific amount of time makes a forecaster an "expert," but practicing your forecasting skills over time will make you a better forecaster. If we apply the "10,000-hour rule" that Malcom Gladwell talks about in his 2008 book, Outliers: The Story of Success (opens in a new window), to really become good at forecasting, you need 10,000 hours of practice. Think about that: If you made weather forecasts as a full-time job, eight hours a day, five days a week, 50 weeks a year (only two weeks of vacation!), it would take five full years to get to 10,000 hours. While you won't approach your "10,000 hours" this semester, this course will hopefully help you establish good forecasting habits and set you on the path of lifelong learning about forecasting.

All the practice in the world won't really help you become a better forecaster, though, if you don't have sound conceptual models of how the atmosphere works. This is where your knowledge from the previous three courses comes into play, especially the basics of synoptic and mesoscale weather. These concepts will play a pivotal role in your experience in this course. If you are feeling a bit rusty, don't get nervous. Throughout the lessons, I'll include some review material, as well as try to make targeted suggestions for review reading from previous courses.

Before we get into more advanced controllers of forecast variables and the strategies used in forecasting, we have to answer some basic questions first. What are we forecasting? How are forecasts verified? How do prudent forecasters go about making a forecast?

We'll look at these issues in this lesson. Let's get started!

sas405

What is forecasting?

What is forecasting?

Prioritize...

When you've completed this page, you should be able to distinguish deterministic forecasts from probabilistic forecasts and point forecasts from areal forecasts. You should also be able to answer the three key questions near the top of the page as they apply to the WxChallenge forecasting competition (particularly relating to forecast deadlines and beginning / ending times for forecast periods).

Read...

What is forecasting? This might seem like a silly question, but it's an important one! Of course, to forecast means to make a prediction about the future state of something. So, in weather forecasting, we're trying to predict the future state of the atmosphere (duh!).

But, what exactly are we trying to forecast? That question could have a variety of answers. It depends on the atmospheric variable(s) that you are trying to forecast and what exactly you want to know about those variable(s). Indeed, before making any weather forecast, we have to start with a few important questions:

  1. What do I want to know (what variables am I forecasting)?
  2. What time period am I interested in?
  3. What geographic region am I interested in?

Answering these questions helps you to define the objective(s) of your forecast. These questions can have many answers, and thus, many different types of weather forecasts exist. But, before we address these questions, we have to cover a basic distinction that you're already aware of -- the difference between deterministic and probabilistic forecasts.

  • A deterministic forecast is one in which forecasters provide only a single solution. For example, "tonight's low will be 31 degrees Fahrenheit," or "0.46 inches of rain will fall tomorrow."
  • A probabilistic forecast is one in which forecasters convey uncertainties by expressing forecasts as probabilities of various outcomes. For example, "the probability that tonight's low will be below 32 degrees Fahrenheit is 40 percent," or "the probability of receiving at least 0.25 inches of rain tomorrow is 60 percent."

I'm sure you have encountered both classifications of forecasts before. The daily high and low temperature forecasts produced by the National Weather Service, most private forecasting companies, and your local TV weathercaster are usually deterministic. Probabilistic forecasts, on the other hand, are more commonly produced for precipitation (for example, "tomorrow's chance of rain is 80 percent"), but the Storm Prediction Center also uses a probabilistic approach for their convective outlooks (see the tornado, damaging wind, and hail forecast graphics in this example from May 20, 2019 (opens in a new window)).

Hale Stone's deterministic forecasts

A forecast for nearly three feet of snow on day 6? Good luck with that!
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

In many cases, probabilistic forecasting is a more realistic approach because it allows forecasters to acknowledge and express the fact that the forecast contains uncertainty (and just about all weather forecasts do contain some degree of uncertainty). But, despite their shortcomings, deterministic formats still rule the day in the form of the five, seven, or even 10-day "tombstone forecasts" that you see on TV (see the example on the right) or on many weather apps.

Some private forecasting firms produce similar deterministic forecasts out a few months into the future, which is a completely ridiculous format for that length of time. Such deterministic forecasts imply that the forecaster has the exact same confidence in the first day of the forecast period as he or she does in the 50th day, for example (which isn't true), and the accuracy of the exact forecast values that far into the future is laughable. If you're interested in reading more about the differences between deterministic and probabilistic forecasts, check out the links in the Explore Further section below.

Forecasting Issues of Space and Time

Now that we've discussed probabilistic and deterministic forecasts, we can deal with those key questions that I mentioned at the top of the page. Think about it: The format of your forecast and the variables that you forecast will differ based on what you want to know (as well as when and where). If you're forecasting for a sailing race between Chicago, Illinois and Mackinac Island, Michigan (opens in a new window), you need to predict weather conditions (especially wind and threats of thunderstorms) along your race path for more than a day. On the other hand, if you just want to know whether the afternoon will be hot enough to take the kids to the pool, you're interested in the afternoon weather conditions (particularly temperature, sky conditions, wind, and chance of precipitation) at a single specific location for a time period of a few hours.

To account for the variety of variables, geographic areas, and time periods that forecasters may be interested in, forecasts take on a variety of formats. The most common are point forecasts (or zone forecasts) and areal forecasts.

  • Point forecasts, as you might have guessed, are forecasts valid at a given point in space (or a very small area), such as your backyard, neighborhood, or zip code. When you look at daily forecasts (which often include high and low temperature, probability of precipitation, amount of precipitation, and wind speed) from the National Weather Service or most private forecasting companies, you're looking at point forecasts.

    Another twist on a point forecast is a "zone forecast."  The forecast format is similar to that of a point forecast, but instead of the forecast being valid for a single point (or very small area), it's valid for a "zone"--an area with similar topographical and climatological characteristics (usually the size of a county or part of a county). In the modern era of automation in forecasting, point forecasts are the norm because computers are able to interpolate gridded computer model output to many specific points. Thus, zone forecasts are being utilized less and less.

  • Areal forecasts are typically issued for individual forecast parameters over a much larger area (portions of states, or entire regions of a country, for example). Because of the size of the forecast area, it's common to only view one forecast variable at a time (an areal forecast for precipitation, for example).

You're already quite familiar with areal forecasts, too, I'm sure. The example below is a Day 1 quantitative precipitation forecast (QPF) map from the Weather Prediction Center.

Day 1 QPF map from the Weather Prediction Center

The 24-hour QPF map valid from 12Z Wednesday, October 28, 2020 through 12Z Thursday, October 29, 2020 indicated several swaths of very heavy rain in the eastern half of the United States. Such forecasts are examples of areal forecasts for a single variable (amount of precipitation in this case).
Credit: Weather Prediction Center

During the 24-hour period covered by this forecast, a large area of the central and eastern U.S. was predicted to receive at least 0.01 inches of rain (light green shading). Successive colors (darker greens, blues, purples, and reds) indicate areas of heavier rain, with a few areas predicted to receive at least 3.00 inches of rain during the 24 hour period (associated in part with Hurricane Zeta (opens in a new window)). Although areal QPF maps are quite common, other regularly published areal forecasts include snowfall accumulation maps, as well as daily Convective Outlooks (opens in a new window) from the Storm Prediction Center (SPC).

Of course, in addition to the point or area that we're forecasting for, we also need to define what time span we're interested in. We can forecast for the very near future (perhaps a few hours), or produce short (1-2 days), medium (3-7 days), or long-range forecasts (you may find varying definitions of these forecast ranges, but these will give you the basic idea). Obviously, as we move from the short range, to the medium range, to the long range, the types of forecasts that we can reasonably make change, because the uncertainties become greater. For example, check out the format of the medium and long range temperature and precipitation forecasts produced by the Climate Prediction Center (opens in a new window). They merely identify areas (probabilistically) where temperatures and precipitation will either be above, below, or near long-term averages for the given time period.

So far, I've just scratched the surface in describing various types of forecasts because I want you to realize that forecasting can be a very broad field! Different forecasting scenarios will have different objectives, so answering the three questions near the top of the page is important any time you begin to make a forecast. Check out the WxChallenge Application section below to better understand our forecasting focus in this course.

WxChallenge Application

In this course, we're primarily going to focus on short-range, deterministic point forecasts because they are the most commonly consumed weather forecasts (and that's the type of forecast you'll be creating for WxChallenge (opens in a new window)). Let's look at the three key questions I posed above on this page as they relate to WxChallenge forecasts:

  1. What do we want to know? Maximum and minimum temperature, maximum sustained wind speed, and total liquid precipitation
  2. What time period are we interested in? The following 06Z - 06Z time period (a total of 24 hours)
  3. What geographic region are we interested in? An observation site (usually at a city's airport) in or near the WxChallenge forecast city

Once the competition begins, you'll have four WxChallenge forecasts per week, due by 00Z Tuesday, Wednesday, Thursday, and Friday (Monday - Thursday evenings in the U.S.), each covering the following 06Z - 06Z time period. Check out the chart below to visualize the weekly schedule.

Chart detailing the weekly forecast deadlines and forecast periods for WxChallenge

The weekly template that shows the four deadlines and corresponding forecast periods during each week of WxChallenge.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

For each forecast, you'll be forecasting the maximum and minimum temperatures, maximum sustained wind speed, and total liquid precipitation for a specific observation site (usually at a city's airport). The quality of your forecasts will be assessed according to the rules of the contest (opens in a new window). But, since we've already discovered that many different types of forecasts exist, there must be more than one way to assess the quality of forecasts. That's what we'll look at in the next section.

Explore Further...

If you want to delve deeper into the differences between deterministic and probabilistic forecasting, you may find the links below to be of interest:

sas405

Assessing Forecast Accuracy

Assessing Forecast Accuracy

Prioritize...

Upon completion of this page, you should be able to perform simple calculations for absolute error, mean absolute error, and Brier Scores, and interpret them to assess the quality of a particular forecast.

Read...

Darts on a dartboard

Contrary to popular belief, throwing darts has nothing to do with forecasting or measuring forecast accuracy.

You've probably heard the old adage, "weather forecasters are the only people who can be wrong most of the time and still get paid." Hilarious, right? Most folks just don't realize the complex science that is involved in making good forecasts. Throwing darts won't cut it! I don't think most people even realize weather forecasters have methods of verifying and assessing the accuracy of their forecasts (or even if the methods exist, they assume that nobody cares enough to use them). Well, methods for verifying and assessing the accuracy of forecasts do exist--many types, in fact. Here, we're just going to cover a few of the most commonly used ones. If you're a bit "math averse," don't worry too much. We'll limit the discussion to simple arithmetic and other basics.

Mean Absolute Error

Perhaps the simplest way to assess the accuracy of a deterministic point forecast is to calculate the absolute error. The absolute error is merely the absolute value of the difference between the observed and the forecast value (absolute value just means that we're ignoring the sign of the difference). So, if you forecast a high temperature of 65 degrees Fahrenheit, and the observed high temperature ends up being 68 degrees Fahrenheit, then the absolute error would be |65ºF - 68ºF| = |-3ºF| = 3ºF. Or, if you forecast 0.82 inches of rain to fall in a 24-hour period, and 0.50 inches actually falls, the absolute error would be |0.82 inches - 0.50 inches| = 0.32 inches. Simple enough, right? Then, if you take the absolute errors over a number of forecast periods and average them, you get mean absolute error (which you may see abbreviated as MAE). The mean absolute error is useful for telling us the average difference between our forecasts and the observed values that occur. As you probably realize, more accurate forecasts yield smaller absolute errors (and smaller mean absolute errors over the long haul).

You might be asking, "why do we take the absolute value of our forecast errors to eliminate the sign?" If we didn't, then we would get misleading results when we calculate the average of our errors. Take the example from above: You forecast a high temperature of 65 degrees Fahrenheit, and the observed high temperature is 68 degrees Fahrenheit. The error would be 65ºF - 68ºF = -3ºF. The next day, you forecast a high of 65 degrees Fahrenheit again, but the observed high is 62 degrees Fahrenheit. Your forecast error would be 65ºF - 62ºF = +3ºF. If we averaged your two errors we would get an average error of zero degrees Fahrenheit over the two day period. That's not an accurate reflection of your error because the positive and negative errors canceled each other out. But, your mean absolute error over the two-day period would be 3 degrees Fahrenheit, which is more telling. Calculating mean error (including the signs) over time doesn't tell us much about the size of the difference between forecasts and observations, but it can tell us about whether or not forecasts exhibit any bias (as in, whether forecasts tend to be lower or higher than the observed values over time).

Check your understanding of absolute error and mean absolute error with the practice questions below:

Brier Scores

How would we assess the accuracy of a probabilistic forecast? Mean absolute error doesn't really help us. So, I introduce to you the Brier Score, a common way of assessing the accuracy of probabilistic forecasts. Calculating the Brier Score (BS) for a single forecast event is pretty simple, using this formula:

BS = (p - o)2

In this formula, p is the forecast probability of an event occurring, and o is the occurrence of the event ("0" if the event does not occur, or "1" if the event does occur). The Brier Score for a single forecast event, then, is just the difference between the forecast probability and "0" or "1" depending on whether the event occurs. Let's think about a simple example. Say you forecast a 70 percent chance of measurable rain at your hometown on a given day, and indeed, measurable rain ends up falling. The Brier Score calculation would be:

BS = (0.7 - 1)2 = (-0.3)2 = 0.09

So, your Brier Score for the forecast would be 0.09, which is pretty good! Now, if you had made the same forecast (a 70 percent chance of measurable rain), and no rain fell at all, then your Brier Score would be:

BS = (0.7 - 0)2 = (0.7)2 = 0.49

This time, your forecast wasn't as good because you predicted a 70 percent probability of something that ended up not happening. From these two examples, it's clear that lower Brier Scores are more desirable. A perfect Brier Score for a single event would be "0" (from a forecast of 0 percent or 100 percent that ends up being correct), while the worst possible Brier Score for a single event would be "1" (from a forecast of 0 percent or 100 percent that ended up being incorrect...a complete bust). If we want to track Brier Scores for a particular forecast variable (say, precipitation) over a long period of time, we can merely average them. Or, if we wanted to combine Brier Scores for multiple forecast events on a given day (say, the probability of measurable rain, the probability of the high temperature exceeding 80 degrees Fahrenheit, and the probability of a thunderstorm occurring), we can add up the individual Brier Score for each event to compute a Total Brier Score.

Check your understanding of Brier Scores (and their interpretation) with the practice questions below:

Threat Scores

Mean absolute error and Brier Scores help us assess the accuracy of deterministic and probabilistic forecasts at a single point, but what about areal forecasts? That's where threat scores come into play. Forecasters at the Weather Prediction Center (WPC), for example, use threat scores to assess their areal forecasts for heavy precipitation. For a given event, forecasters compute threat scores by comparing the area where they predicted heavy precipitation with the area that ultimately received heavy rain or heavy snow.

You can think of a threat score as the ratio of the area where the forecast was accurate to the area where the forecast didn't verify correctly. In the figure below, the Forecast area (F) is the region for which WPC forecasted heavy precipitation (shaded in red). The observed area (OB) indicates the region where heavy precipitation fell and is shaded in green. The hatched area, C, represents the region where the forecast for heavy precipitation was correct. By definition, the threat score (T) is calculated using the equation: T = C / (F + OB - C)

Schematic showing the components of a threat score calculation

A schematic defining the areas used to compute a threat score (T). The hatched area, C, represents the region where the forecast for heavy precipitation verified. F, the red shaded area, is the region for which WPC forecasters predicted heavy precipitation. OB, the green shaded area, indicates the region where heavy precipitation fell.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

Unlike with Brier Scores, higher threat scores indicate better forecasts. A threat score for a perfect forecast is "1", while a completely busted forecast gets a big fat "0" (you can read more about computing threat scores (opens in a new window), if you're interested).

The figure below shows the trend in annual threat scores based on predictions for one inch of rain (or a liquid-equivalent of one inch) during each year from 1960 through 2020. To put this graph in proper context, I point out that WPC bases all QPF verifications on the 12Z-to-12Z period. As for the graph itself, please note that Day 1 represents the forecast period from 12 to 36 hours (WPC calculates forecast hours based on the 00Z model runs). Day 2 forecasts cover 36 to 60 hours, while Day 3 spans from 60 to 84 hours.

Graph showing WPC's annual threat scores for 1 inch of precipitation from 1960-2020

The annual threat scores for predicting one inch of rain or liquid equivalent from 1960 to 2020. Threat scores are improving, but lots of room for improvement still exists.
Credit: Weather Prediction Center

These threat scores indicate that WPC typically attains a threat score right around 0.28 for Day-2 forecasts of one inch of rain (or liquid equivalent). Such a score translates to WPC getting only about 45 percent of the predicted area of one inch of rain or liquid equivalent correct 36 to 60 hours in advance. The Day-3 threat scores are a bit lower yet. Keep in mind that the forecasters at WPC are very good!

The take-home message here is that threat scores in the field are improving (note the increasing trends on the graph), but our ability to accurately forecast regions of heavy rain or snow 48-hours or more from an event is not very good. Kind of paints extended forecasts for blizzards (opens in a new window) and daily deterministic point forecasts weeks into the future in a sobering light, doesn't it?

Forecast Tip

I highly recommend keeping a "forecast journal" when making forecasts. Keep notes about key forecast issues, model predictions, the reasoning behind your forecast, the final verification, and what your forecast errors were. These notes can really help you learn from your mistakes, and help build your mental log of forecasting experience!

Before we move on, check out how what you learned in this section applies to forecasts in the WxChallenge competition in the WxChallenge Application section below.

WxChallenge Application

As I mentioned previously, in this course, our focus is going to be short-range, deterministic point forecasting for WxChallenge. So, you won't need to calculate threat scores in this course (the exact calculations can be quite daunting, and I promised we'd keep to basic math). I just wanted to give you a taste for the various types of forecast assessment schemes that exist. For our purposes, realize that the scoring formulas in WxChallenge (opens in a new window) are based on daily absolute errors for high and low temperature, maximum sustained wind speed, and amount of precipitation (although they assign an artificial point scheme to the absolute errors for wind and precipitation).

Now that we've covered the basics of forecasting and accuracy assessment, let's think more about how prudent, responsible forecasters go about making forecasts. Applying the basic approach described in the next section will help you create quality forecasts more consistently given any forecast objective, including those in WxChallenge.

sas405

A Wise Forecasting Philosophy

A Wise Forecasting Philosophy

Prioritize...

From this page, you should be able to explain why "swinging for the fences" is a risky forecasting approach that won't lead to success over the long haul. Also, you should be able to identify the key steps to making a good forecast, and be able to execute a "probabilistic" internal dialogue with yourself when making deterministic forecasts.

Read...

How do good forecasters go about making a forecast? What general approach do they use? As we've already discussed, anybody can interpret model guidance with a little bit of training, but the goal here is to put you on the path to becoming a good weather forecaster (a rare breed). To get us started in discussing the forecasting philosophy advocated in this course, I'm going to use an analogy that involves baseball (although softball would work, too). So, if you're not into baseball, forgive me, but I think the analogy is still effective at getting the point across even for those who aren't into baseball or softball.

A batter awaiting a pitch.

What does baseball have to do with weather forecasting? As it turns out, the prudent approach of successful "singles" hitters can serve as a basis for a successful forecasting strategy.
Credit: Steve Seman

When it comes to forecasting, you can take one of two approaches, which I associate with a batter's approach in baseball (or softball). I'll call one approach the "Dave Kingman (opens in a new window) approach" and the other the "Tony Gwynn (opens in a new window) approach." The first paragraphs of their Wikipedia biographies highlight their contrasts in hitting style. To boil down the differences, here are a few key stats from their careers:

  • Dave Kingman (played from 1971-1986): 16 seasons, 442 home runs, 1,816 strike outs
  • Tony Gwynn (played from 1982-2001): 20 seasons, 135 home runs, 434 strike outs

What should you note from those stats? Well, for starters, both players had fairly long, but different careers. Dave Kingman hit a lot of home runs, but he also struck out a lot. In fact, when he retired, he had the fourth most strike outs in Major League Baseball history. On the other hand, Tony Gwynn didn't hit many home runs, but he rarely struck out. Tony Gwynn was the ideal "singles hitter," he had a very high batting average (top-20 all-time, in fact), so he was very effective at getting hits and getting on base (not making outs).

When it comes to forecasting, the "Tony Gwynn approach" is more successful over the long haul. When forecasting, I try to hit a lot of "singles," and I'm very careful about taking large risks to avoid strike outs (major forecast busts). When you take the "Dave Kingman approach" and always try to swing for the fences to hit home runs in weather forecasting, you might hit the occasional home run (make risky forecast that turns out great), but more often you're going to strike out (like Dave Kingman) and end up with a bad forecast.

The name of the game in forecasting is to minimize your error as often as possible, which is akin to consistently hitting "singles" in baseball. Such an approach will allow you to minimize or eliminate huge forecast busts. Now, you might be thinking, "but I'd rather hit home runs -- home runs are better!" I assure you, however, that even great weather forecasters have a hard time hitting home runs consistently. Also, keep this in mind: Tony Gwynn, the ultimate singles-hitter, was inducted into the baseball hall of fame on the first ballot. Dave Kingman, the monster home-run hitter, hardly got hall-of-fame consideration, despite the fact that when he retired, 400 home runs was almost a guarantee for hall-of-fame induction.

What does applying this philosophy to a weather forecast look like? Allow me to show an example: METEO 410 students had to grapple with a rather challenging forecast for Atlantic City, New Jersey a number of years ago, when a powerful low-pressure system was slated to move up the Atlantic Coast and dump heavy rain on Atlantic City. In fact, one model predicted nearly three and a half inches for the day! You can see the basic set-up in the GFS four-panel forecast prog below.

GFS forecast prog valid at 00Z on March 17, 2007.

The GFS forecast valid at 00Z on March 17, 2007 predicted that a strong area of low pressure would bring heavy rain to Atlantic City, New Jersey. This model run was initialized at 12Z on March 14.
Credit: Penn State University / Phil Lutzak

But, the daily record for precipitation at Atlantic City was only 1.14 inches, meaning that many models were forecasting double or even triple the daily-record rainfall. That was a red flag that predicting such huge precipitation amounts could be risky. In the final analysis, the models were too high (surprise, surprise) and students that swung for the fences along with the models had huge forecast errors (they struck out big time). Atlantic City did set a daily rainfall record, but only 1.24 inches actually fell.

How could a forecaster have gone beyond mere "model reading" and come up with a prudent forecast? The key is to assess the overall weather pattern to look for clues that could make or break a forecast. For example, you may recall from previous studies that forecasters look for strong low-level jet streams at 850 mb to import deep moisture in synoptic-scale heavy rain events. A skilled forecaster in this case may have noticed that the strongest portion of the low-level jet stream was likely to be offshore, as suggested by this GFS forecast from the day before the event (opens in a new window) (note the 850-mb prog, bottom middle), which meant that the strongest deep moisture convergence (opens in a new window) might be east of Atlantic City. With only a peripheral encounter with the best moisture convergence in the cards, doubling or tripling the daily rainfall record was pretty unlikely, and a forecaster could have gone with a more conservative forecast. This case is a microcosm of the notion that swinging for the fences and/or forecasting exactly what the models predict will happen (or what you want to happen) in an extreme situation is a risky approach that often doesn't work well.

The attributes of a good forecaster ...

The Atlantic City case is pretty scary, isn't it? Models (and many forecasters) predicted double or triple the amount of rain that actually fell. The models can be your worst enemy if you don't use them in an insightful (and cautious) way. Of course, insight comes with experience, and to gain worthwhile forecasting experience, you need to develop specific attributes. What attributes to good forecasters possess? The list below shouldn't surprise you after all of your time in the certificate program. To become a good forecaster, you must:

  1. Possess a working knowledge of the behavior of the atmosphere.
  2. Possess a working knowledge of forecasting principles and techniques.
  3. Gain enough experience to know which principles to apply to any given situation.
  4. Develop the ability to interpret statistical and numerical model guidance.
  5. Acquire knowledge of how a general forecast must be modified to account for local effects at a specific site.

Hopefully, all the certificate courses have given you a solid foundation for attribute #1. Don't get too comfortable, though. Number 1 is a "work in progress," even for experienced forecasters! Weather forecasting is all about life-long learning.

This course provides a solid foundation for you to learn and apply a working knowledge of forecasting principles and techniques (#2), and the process of making forecasts in this course will be critical to building your body of forecasting experience (#3). This, too, will be a "work in progress." The lessons in this course will also enhance your ability to interpret statistical and numerical model guidance (#4). Finally, #5 is crucial: Know your local climatology! We'll talk more about the types of climatological information that good forecasters need to know later in the lesson.

Radar mosaic of the eastern U.S.

Observations, satellite imagery, and radar imagery, like this radar image of the eastern U.S. from 0930Z on August 27, 2020 should be an important starting point for your forecasting routine.
Credit: WSI / Penn State University

Good forecasters also develop a forecasting routine that allows them to apply the knowledge and skills described above. How exactly do good forecasters go about making a forecast? Ultimately, different forecasters will approach the forecasting process a bit differently based on personal preferences for data types and their forecasting objective in a particular situation. Still, for making a short-range forecast, good forecasters tend to follow the same basic recipe, which I've outlined below for you:

Forecast Tip

When you create your own short-range forecasts...

  1. Start with the "big picture." Look at local observations and also observations "upstream" (regions from which relevant systems will approach the local forecast area). Is the air that's moving into your area notably warmer / cooler or moister / drier?
  2. Look at surface and upper-air analyses, satellite and radar imagery -- really, any tool that helps form a clearer "big picture" of the present state of the atmosphere.
  3. Peruse the model runs to get an overall sense of the forecasting issues (shortwave troughs, fronts, jet streaks, other sources of vertical motion, etc.). Compare models. Look for consensus or a lack of consensus (that is, the degree of uncertainty in the forecast). Look for trends in the models (so you need to look at previous model runs). Also look at ensemble solutions for the overall synoptic-scale pattern, trying to identify where the models are having problems (or where there is a relatively high confidence in the forecast). In general, try to get your arms around the key issues of the forecast. The big picture is the boss!
  4. Now it's time to get down and dirty with details: statistical guidance (like the National Blend of Models), ensemble solutions for specific forecast parameters, forecast skew-T's, and other advanced techniques that we'll learn about later. Your big picture analysis should tell you which techniques will be appropriate, so this stage of the forecasting process might differ from day to day.

You will learn that statistical guidance like the National Blend of Models (NBM) and some ensemble mean forecasts tend to be good over the long haul. But, they may or may not be good for tomorrow's specific forecast. Of course, you should always consider these tools, but don't automatically accept them or reject them out of hand. Indeed, you must have a clear "big picture" in mind (be one with the atmosphere) before you make any decisions about tomorrow's forecast. In other words, you must put model guidance into a proper meteorological context. Developing this kind of an approach takes practice, of course, and you'll get plenty of it this semester.

While sorting out the details of the forecast, you have to keep your wits about you. The amount of data can be overwhelming, and sometimes there's little consensus among the models. Especially in cases with great model uncertainty, I think it's wise to establish an internal dialogue with yourself. For example, if the range in ensemble precipitation forecasts for tomorrow is from 0.10 inches to 0.75 inches, ask yourself questions like "Is it more likely that something closer to 0.10 inches or 0.75 inches will fall tomorrow? Why?" Then back up your answers to your internal questions with sound meteorological reasoning (perhaps based on various lifting mechanisms available and the overall moisture characteristics of the air mass, for this example). Even though we'll focus on deterministic forecasts this semester, it's still important to think probabilistically. This will help you hit more "singles" and avoid swinging for the fences, trying to hit a home run instead (but often striking out with a big forecast bust).

I hope you noticed in the forecasting process outlined above that good forecasters start with observations and the atmospheric "big picture." On that note, we're going to look at a case study to give you an example of how forecasters can use the big picture to diagnose the expected weather in a particular region. Keep reading!

sas405

Case Study: Blocking Highs and Closed Lows

Case Study: Blocking Highs and Closed Lows

Prioritize...

After completing this section, you should be able to identify blocking highs, discern between cut-off lows and closed lows, and identify areas that may be persistently wet (or dry) based on the presence of a highly meridional pattern.

Case Study...

In this course, I often preach that becoming a consistently good weather forecaster requires a focus on "the big picture" (the overall synoptic pattern). But, what types of things do forecasters look for in assessing the big picture? The specifics, of course, vary day by day, but I'm going to show an example to introduce you to the basics. A "big-picture" diagnosis starts with a survey of the upper-air pattern, and there's no better place to start than at 500 mb. Check out the short video below, when forecasters could identify an upper-level blocking pattern, and take note of the weather consequences that forecasters can infer from such patterns.

Text on screen: Here’s an example of a blocking pattern that set up over Canada and U.S. Northern Border states as shown on this daily composite of 500-mb heights. There were two blocking highs — one located over Canada's Northwest Territories and the other simultaneously anchored over Newfoundland in eastern Canada. The result of such a pattern is that the flow across Canada and the U.S. northern border states is not very "zonal" (west-east). Indeed, it's highly "meridional" (north-south). That means the natural west-to-east progression of mid-latitude weather systems across Canada and the U.S. northern border states gets interrupted by the two blocking highs and weather systems tend to move slowly.

So, what formally makes these highs blocking highs? Well, a 500-mb closed high or ridge qualifies as a blocking high if several criteria are met: First, the basic westerly current must split into two branches (like the split in the flow of water around a rock in a fishing stream), and each branch must transport appreciable mass, as seen here. Second, the double-current system must extend over several tens of degrees longitude, so the split must span a substantial west-east distance. Third, a sharp transition from zonal flow upstream to meridional flow downstream must be observed across the split in the westerly current. And finally, the pattern must persist with recognizable continuity for at least five days, although some stricter criteria would require ten days.

These criteria were easily met in this case, so this was indeed a blocking pattern. I should point out that two closed lows formed on either side of each blocking high, which compounded the stagnancy of this blocking pattern. So, what are the weather consequences of such patterns? Well, near closed (or cut-off) lows sometimes we get protracted rains and flooding because of their slow movement. In contrast, the weather underneath the blocking high tends to be sunny and dry. Long-lived blocks can actually create drought conditions under the high.

The stagnancy of this blocking pattern paved the way for protracted, recurrent rains across New England for nearly a week, as these multi-sensor rainfalls over a 7-day period show. A wide area of New England received at least 5 inches of rain, with some areas getting more than 10 inches. Needless to say, flooding was rampant.

The flooding rains in this stagnant pattern were aided by a tropical connection, as suggested by this visible satellite image, indicating a long plume of clouds extending from low latitudes all the way up into New England. The sprawling closed low over the Great Lakes region had helped draw moist Atlantic Air northward in concert with a low-level jet stream, which we can see on this daily composite of 850-mb vector winds. The end result was plenty of moisture convergence in New England, feeding the heavy rain.

Credit: Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

Note that in the video, I referred to the 500-mb lows in this case as "closed lows" instead of "cut-off lows." While some forecasters might use the terms interchangeably, there's actually a big difference between a "closed low" and a "cut-off low." With a true "cut-off" low, the low's circulation drops out of the prevailing westerly 500-mb flow (the low is "cut off" from the prevailing westerly flow). On the other hand, there are traveling 500-mb lows embedded in the prevailing westerly flow that display a "closed" circulation. But, these 500-mb lows are not cut off. They're just plain old closed lows. Generally speaking, closed lows tend to continue their propagation with the prevailing westerly flow, albeit typically rather slowly.

Schematic showing the difference between a closed low and a cut-off low.

(Left) Any upper-level low with at least one closed height contour (or isobar) encircling its center is a closed low. (Right) A cut-off low is a closed low that has detached from the main westerly current of air and has taken a position equatorward of this current.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

So, all cut-off lows are closed lows, but not all closed lows are actually cut-off lows. However, both closed and cut-off lows can be prone to slow movement, and their slow movement can cause prolonged rainy periods, especially on their eastern flanks. In this case, the stagnancy of this blocking pattern over North America paved the way for protracted, recurrent rains across New England for nearly a week. And, after a week of recurrent rains, flooding was significant.

Analysis of 500-mb heights from May 12, 2006

The rampaging Little River in Lebanon, Maine, during the spate of rain over New England around the middle of May, 2006.
Credit: The Boston Globe

Recognizing blocking patterns as part of a "big-picture" analysis can help you get an initial basic feel for what type of weather to expect in various parts of the pattern. This understanding may help you in preparing both short-range and longer-range forecasts. Another note of interest is that models tend to break down blocking patterns too quickly, so that's something to keep in mind should you encounter one this semester (and beyond). Finally, I should note that the subtropical high pressure systems that you learned about in your previous studies do not qualify as blocking highs. They typically lie too far south (near latitude 30 degrees) to cause a significant split in the 500-mb westerly current.

While we didn't look at every aspect of the "big picture" here, did you notice just how much information we gathered by carefully diagnosing the 500-mb pattern and recognizing its consequences? Simply by recognizing the blocking highs and closed lows, forecasters knew that weather systems would be slow moving, resulting in some persistently wet and dry areas. When you start making your own forecasts, you should always start with an analysis of the "big picture" pattern, including current observations. Speaking of observations, we need to look more deeply into the observations of critical forecast variables (temperature, wind, and precipitation). Read on.

sas405

It All Starts With Observations

It All Starts With Observations

Prioritize...

From this page, you should become familiar with the standard instruments for measuring temperature, wind speed, and precipitation. You should also be able to identify the standard height at which "surface" wind observations are taken and discern the difference between "sustained" wind speed and wind gusts as measured at ASOS sites.

Read...

Since good forecasters start their forecasting process with observations, we need to look more closely at how weather observations are taken--especially observations of temperature, wind speed and precipitation. While measuring these weather variables might seem straightforward, as you're about to see, measurements aren't taken in the same way at all stations. Furthermore, for decades, weather observations were taken solely by humans, but the number of completely automated observations taken with no human intervention continues to increase. Even today, however, students at Penn State still take daily observations for the campus weather station at University Park. Check out the observation card (below) from January 25, 1985. For the record, the icon under the heading of general observations is the universal symbol for blowing snow.

An observation card from the Penn State weather station.

The observation card at the Penn State Weather Station on January 25, 1985. The weather observer on this date was none other than original certificate program author, Lee Grenci.
Credit: Penn State Department of Meteorology and Atmospheric Science

Why do I show this antique observation card? I want you to note the variety of different observations it includes. It's not just high and low temperature, wind speed, and total precipitation! Even something as seemingly simple as "temperature" has observations for maximum, minimum, and set. What's the "set" temperature?  Let's find out, and explore some other nuances of measuring temperature.

The Highs and Lows of Temperature

Students at Penn State aren't alone in taking these "old fashioned" observations. Many cooperative observers (opens in a new window) and some smaller airports rely on the same type of liquid-in-glass max / min thermometers that have been used for decades (see image on the right). While modern digital and electronic equipment has largely replaced less-sophisticated tools and methods, these "low-tech" thermometers are critical for recording daily high and low temperatures at locations where observations are not electronically monitored throughout the day.

Maximum / minimum temperature thermometers

A minimum thermometer (top) and a maximum thermometer (bottom) in the Cotton Region Shelter at the National Weather Service Office in San Francisco.
Credit: NOAA

In locations where observations aren't constantly monitored electronically, maximum and minimum thermometers (yes, there are two!) are typically housed in Cotton Region Shelters (opens in a new window), or "CRS." Cotton Region Shelters are typically painted white to reflect sunlight because too much absorption of direct sunlight could dramatically raise the temperature of the thermometer above the true air temperature. CRS are also mounted about five feet above the surface, away from the drastic temperature changes that occur very near the ground. Their louvers allow ventilation so that the air in contact with the thermometer has essentially the same temperature as the outside air (check out the interior of the CRS (opens in a new window)).

But, the two thermometers in the CRS are not ordinary thermometers. The most important part of the maximum thermometer is the constriction in the bore (opens in a new window). Think of the constriction as a trap door. As the air temperature increases, the mercury expands and forges through the constriction. But, as the air temperature eventually starts to decrease, the mercury remains trapped beyond the constriction so that the weather observer can determine the high temperature later on. Like a clinical thermometer used to measure a patient's temperature, the observer resets the maximum thermometer to the current temperature (called the "set temperature") by shaking or spinning (centrifuging) the mercury back through the constriction.

The minimum thermometer typically contains alcohol (which has a lower freezing point than mercury). Within the bore of this thermometer, there is a barbell-shaped marker (opens in a new window) called an index. The barbell on the right of the index registers the minimum temperature. How? As air temperature decreases, the alcohol contracts, and when its retreating meniscus (opens in a new window) encounters the right side of the barbell, it drags the index along with it. When temperatures increase, however, the expanding alcohol draws the meniscus away from the index and alcohol flows up the bore around the index (the index doesn't move; it stays in a position that corresponds to the minimum temperature). The weather observer can simply read-off the 24-hour minimum temperature (the right barbell) at the observation time and reset the minimum thermometer to the current temperature by tilting the thermometer downward.

Instead of using max / min thermometers inside a CRS, increasing numbers of cooperative observers are using the Maximum Minimum Temperature System (opens in a new window), or MMTS. Essentially, an MMTS is an electrical thermometer quartered in what I can only describe as a beehive (a circular, louvered shelter). More specifically, the MMTS is a special type of sensor called a thermistor which electronically measures temperature based on the electrical resistivity of a circuit. The MMTS transmits the information from the temperature sensor to a digital display, which can be easily read by the observer.

Meanwhile, at most primary airports, the temperature sensor is an integral component of an Automated Surface Observing System (opens in a new window) (ASOS). Smaller airports may instead have an Automated Weather Observing System (opens in a new window) (AWOS), which is basically a less sophisticated version of an ASOS. ASOS and AWOS sites commonly measure temperature with a type of thermistor called a platinum-wire resistance sensor (PRT). A hygrothermometer (opens in a new window) houses the PRT sensor and also a chilled mirror that measures dew-point temperature.

An ASOS station.

An Automated Surface Observing System (ASOS) at the airport in Caribou, Maine.

Getting a Handle on the Wind

Of course, ASOS stations measure more than temperature. The ASOS sensor that measures wind speed is an ultrasonic anemometer (opens in a new window), which is mounted atop a tower that is typically 10 meters high (some are a bit lower, depending on local restrictions). How can ultrasonic sound waves measure the wind? Well, the transmission of sound waves through the air varies with wind speed. If the wind speed is blowing in the same direction as the sound waves are traveling, the sound waves travel faster (or slower, if the wind is blowing in the opposite direction). With three equally spaced ultrasonic transducers (a transducer is a device that can convert energy from one form to another), the sensor measures how long it takes for ultrasonic waves to travel between transducers. A total of six measurements are taken (one in each direction along each path between transducers) and from those measurements, a computer calculates wind speed and direction.

The two common wind speeds that get reported are sustained wind speeds and gusts. What does "sustained" really mean? At ASOS stations, It's a two-minute average speed, but the two-minute averages are computed from three-second averages of one-second wind speeds, so measuring a sustained wind speed is not straightforward. The smoothing of data through averaging provides a fairly reliable measurement, though. Gusts, on the other hand, are quick "bursts" of faster winds, technically defined as three-second averages that are faster than those immediately before or after.

The ASOS standards for wind measurement, however, are not universal. Reported sustained winds in tropical cyclones over the ocean, for example, are often measured or estimated one-minute averages. But, some agencies, like the Japan Meteorological Agency, use 10-minute averages as a standard for sustained wind speeds in tropical cyclones. In addition to using different averaging periods than ASOS sites, other observing networks don't always take observations at a standard height of around 10 meters. Wind observations from home weather stations, for example, are often measured much lower to the ground and therefore suffer from increased friction and various sheltering issues because of surrounding trees and buildings inherent in many neighborhood settings. The height of a wind observation really matters (in terms of frictional effects and surrounding obstructions), so beware of observation height when interpreting or comparing wind observations. Wind observations that aren't taken near a standard height of 10 meters can be notably different!

Reining in Precipitation

Several types of rain gauges are commonly used to measure precipitation. At the Penn State Weather Station, we use a standard rain gauge to measure precipitation (see image below). For starters, the circle of triangular metal slats helps to buffer the gauge from the wind, lessening its effect on the amount of precipitation falling into the gauge (allowing precipitation, particularly snow, to fall more vertically into the gauge). The outer shell of a standard rain gauge is a metallic cylinder eight inches in diameter. Inside the outer shell, there is a clear, plastic inner cylinder (opens in a new window) just over 2.5 inches in diameter into which rain funnels. Given that the inner cylinder's cross-sectional area is one-tenth of the outer cylinder's area, rain funneling into the inner cylinder will rise to a height ten times the actual rainfall - horizontal "squeezing" yields vertical "stretching." Thus, a meager 0.01 inches of rain balloons to a depth of 0.10 inches on a measuring stick, making the otherwise unwieldy task of obtaining 0.01" accuracy much easier.

A standard rain gauge

The standard rain gauge on the University Park campus at Penn State.
Credit: Penn State University

When the forecast calls for snow, weather observers remove the inner cylinder and the snow accumulates in the hollow outer cylinder. They then determine the liquid equivalent by taking the gauge indoors, melting the snow, and then pouring the liquid into the small beaker to accurately measure its depth. Observers similarly obtain the liquid equivalent of unmelted hail and ice pellets (sleet).

The weighing gauge (opens in a new window) is another instrument that measures precipitation. Rain collected by the gauge funnels into a hole, below which is a catch bucket. As water accumulates in the bucket, its weight gets converted to a liquid-equivalent depth. Routinely, a mechanically driven pen scrolls across a chart attached to a rotating drum and traces out a record of this depth as a function of time (the drum rotates once every 24 hours). Yet another instrument for measuring rainfall is the tipping-bucket rain gauge (opens in a new window). For this gauge, collected rainwater gets funneled into a two-compartment bucket; 0.01 inches of rain will fill one compartment and overbalance the bucket so that it tips, emptying the water into a reservoir and moving the second compartment into place beneath the funnel (see a close-up view of the top of a tipping-bucket (opens in a new window) rain gauge). As the bucket tips, it actuates an electric circuit that records rainfall.

Observing stations that report hourly precipitation are likely equipped with a Fischer Porter automatic rain gauge (opens in a new window) The gauge is approximately five feet tall and has a diameter of two feet. Much like a weighing gauge, precipitation collecting in a catch bucket presses down on a scale. Every 15 minutes, holes are punched on a "ticker tape" (as a function of the weight of the bucket). In this way, the punched tape keeps a running tally of precipitation since the catch bucket was last emptied. Check out the "guts" of a Fisher Porter and a close-up of the punched ticker tape (opens in a new window).

Naturally, different instruments and standards for taking weather observations can lead to some issues with consistency and quality. If you're interested in reading about a few of these issues, along with a more technical description of ASOS instruments, check out the Explore Further section below. In the meantime, now that we've examined the ways that temperature, precipitation, and wind are measured, we need to familiarize ourselves with the published format of all of these observations. Put on your decoder rings: It's time to learn about decoding METAR observations!

Explore Further...

If you're interested in reading up on some issues arising from the use of different instruments, and / or more technical information about some instruments presented on this page, you may want to check out the following links:

sas405

METARS

METARS

Prioritize...

When you've completed this page, you should be able to decode a METAR observation. In particular, focus on the station ID, time and date, wind observations, present weather, sky coverage, hourly temperature / dew point observations, the T-group, the 1-group, 2-group, and 6-group.

Read...

Throughout the certificate program, you've surely come across METAR observations at least a few times. But, you never actually had to know how to decode them. You might be wondering, "Why do we need to know how to decode METARS? We can get already decoded versions lots of places online!" That's certainly true, but by looking at raw METARS, you can glean more information than from the standard decoded versions (many decoded versions don't include the complete observation). Plus, if you become well-versed decoding raw METARS, you'll find that you can actually gather information from the coded versions much more quickly since the data is presented in a denser format. So, by knowing how to decode raw METARS, you can more easily and quickly see how your forecasts are doing, and gather data as you prepare your forecast for the next day!

Note that this page is quite long, but I don't intend for you to read every single part of it in detail. Focus on the sample METAR observations below and the parts emphasized in the Prioritize statement (they're in red as you go down the page). Also, focus on the Key Skill and Quiz Yourself sections, because they'll outline important things you'll need to know and give you some practice. Consider the rest of the page as a reference guide for times when you need to look up the meaning of a specific code or remark.

Let's get right to it. Here are a couple of METAR observations from Concord, New Hampshire on May 13, 2006.

  • METAR KCON 131151Z AUTO 09009KT 1 3/4SM +RA BR OVC010 09/07 A3005 RMK AO2 CIG 007V013 SLP177 P0015 60056 70066 T00890072 10094 20089 53018
  • METAR KCON 131751Z AUTO 00000KT 2SM +RA BR OVC013 08/06 A3012 RMK AO2 SLP202 P0017 60115 T00780061 10089 20072 53004

What do all of these letters and numbers mean? Let's break it down and decode the following METAR:

METAR KCON 131151Z AUTO 09009KT 1 3/4SM +RA BR OVC010 09/07 A3005 RMK AO2 CIG 007V013
SLP177 O0015 60056 7066 T00890072 10094 20089 53018
  • METAR - METAR is a French acronym that, loosely translated, means "routine aviation weather observation." But sometimes you'll see "SPECI", which translates to a special (unscheduled) report.
  • KCON is the four-character ICAO (International Civil Aviation Organization (opens in a new window)) identifier for Concord, New Hampshire. You can use the station list (opens in a new window) at the National center for Atmospheric Research to help you decipher any identifier.
  • 131151Z is the observation date and time. The observation was taken on the 13th (May, 2006) at 1151Z (the month and the year are not part of the report, so it's important to know whether you're looking at a current or past observation).
  • AUTO indicates a fully automated report with no human intervention. If an observer takes or augments observations, this tag does not appear. Sometimes you might see COR, which indicates a corrected observation.
  • 09009KT indicates wind direction and speed. Here, the wind blew from 90 degrees (an easterly wind) at 9 knots. What happens if winds are gusty? Let's look at a METAR from Mount Washington in New Hampshire (see photograph below) at the same time as the first METAR from Concord:

    KMWN 131147Z 13043G58KT 1/16SM FZRA PL FZFG VV001 M01/M01 RMK PLB40 VRY LGT GICG 60074 70148 931000 10017 21013

    Winds were blowing from 130 degrees sustained at 43 knots and gusting to 58 knots. That's just a ho-hum "breeze" compared to the world-record setting 231 miles an hour (opens in a new window) clocked at the summit on April 12, 1934! Yes, Mount Washington is a windy place indeed.

    Photograph of the summit of Mount Washington

    Mount Washington, New Hampshire, in December, 2005. If you look closely, you can see the Mount Washington Observatory.
    Credit: Mount Washington Observatory

    In stark contrast to windy Mount Washington, a METAR entry of 00000KT represents a calm wind. When the wind is light (a speed of six knots or less) and it varies in direction with time, the data encoded on a METAR might look like VRB004KT (variable direction blowing at four knots). If the wind speed is greater than six knots and the wind direction varies, the data encoded on a METAR might look like "32014KT 290V350". Translation: the wind direction was 320 degrees and the wind speed was 14 knots, but the direction varied from 290 to 350 degrees. Such a varying wind direction might occur in the immediate wake of a cold front. Variable wind directions are always encoded in the clockwise direction (just for the record).

  • 1 3/4SM translates to a horizontal visibility of one and three-fourths statute miles (opens in a new window). Visibilities below one fourth of a mile appear as M1/4SM in METARS from automated stations.
  • +RA BR is the present weather in this case. "+RA" represents heavy rain, while "BR" is the METAR code for mist. You should become familiar with the other codes for precipitation and restrictions to visibility.
The various codes for reporting present weather on METARS.

QUALIFIER

INTENSITY OR
PROXIMITY

1

QUALIFIER

DESCRIPTOR

2
WEATHER PHENOMENA

PRECIPITATION

3
WEATHER PHENOMENA

OBSCURATION

4
WEATHER PHENOMENA

OTHER

5
- Light

Moderate2

+ Heavy

VC In the Vicinity3
MI Shallow
PR Partial
BC Patches
DR Low Drifting
BL Blowing
SH Shower(s)
TS Thunderstorm
FZ Freezing
DZ Drizzle
RA Rain
SN Snow
SG Snow Grains
IC Ice Crystals
PE Ice Pellets
GR Hail
GS Small Hail and/or Snow Pellets
UP Unknown Precipitation
BR Mist
FG Fog
FU Smoke
VA Volcanic Ash
DU Widespread Dust
SA Sand
HZ Haze
PY Spray
PO Well- Developed
Dust/Sand Whirls
SQ Squalls
FC Funnel Cloud Tornado Waterspout4
SS Sandstorm
SS Duststorm
1. The weather groups shall be constructed by considering columns 1 to 5 in the table above in sequence, i.e. intensity, followed by description, followed by weather phenomena, e.g. heavy rain shower(s) is coded as +SHRA
2. To denote moderate intensity no entry or symbol is used.
3. See paragraph 8.4.1.a.(2), 8.5, and 8.5.1 for vicinity definitions.
4. Tornados and waterspouts shall be coded as +FC.

On with the rest of our METAR translation:

  • OVC010 represents the current sky condition, which, at this time, was overcast at 1000 feet (the three-digit code corresponds to the ceiling (or cloud base) in hundreds of feet). In general, please note that METARS can list data about more than one layer of clouds. Moreover, when the sky is obscured, METARS should include the vertical visibility in hundreds of feet. For example, VV004 corresponds to an obscured sky with a vertical visibility of 400 feet.
  • A3005 is the altimeter (opens in a new window) setting - in this case, 30.05 inches of mercury.
  • 09/07 represent the temperature and the dew point reported to the nearest degree Celsius (more precise data sometimes appear near the end of METARS - more in a bit). In this observation, the temperature was 9 degrees Celsius and the dew point was 7 degrees Celsius. A letter "M" appearing before either number indicates a value below 0 degrees Celsius.
  • RMK stands for "Remarks." There are a multitude of possible remarks, but I'll only go over the ones in this particular METAR. In this case, A02 indicates that the automated station has a precipitation sensor (A01 means that the automated station does not have a precipitation sensor).
  • CIG 007V013. When the ceiling (as measured by a ceilometer, which uses a laser or other light source to determine the height of a cloud base) is less than 3000 feet and variable, this group typically appears in METARS. In this case, the ceiling was variable between 700 and 1300 feet.
  • SLP177 indicates the sea-level pressure in millibars, using the same convention as on a standard station model (1017.7 mb, in this case).
  • P0015 is the hourly liquid precipitation (in hundredths of an inch). In this case, 0.15 inches of rain fell in the hour ending at 12Z.
  • 60056 represents the three- or six-hour total liquid precipitation (in hundredths of an inch). In this case, 0.56 inches of rain fell in the six-hour period ending at 12Z. For the record, six-hour totals appear at 00Z, 06z, 12Z and 18Z. Three-hour totals appear at 03Z, 09Z, 15Z and 21Z. 60000 translates to a trace of liquid precipitation during the three- or six-hour period.
  • 70066 indicates the total 24-hour liquid precipitation ending at 12Z (in hundredths of an inch). In this case, 0.66 inches fell at Concord from 12Z on May 12 to 12Z on May 13.
  • T00890072, the T-group, indicates the hourly temperature and dew point to the nearest tenth of a degree Celsius. The "0" after the "T" indicates that the temperature is greater than or equal to 0 degrees Celsius (a "1" will follow the "T" when the temperature is lower than 0 degrees Celsius). After the three digits for temperature, a "0" indicates that the dew point is greater than or equal to 0 degrees Celsius (a "1" indicates that the dew point is less than 0 degrees Celsius). In this case, the 12Z temperature at Concord was 8.9 degrees Celsius (48 degrees Fahrenheit) and the dew point was 7.2 degrees Celsius (45 degrees Fahrenheit).
  • 10094, the 1-group, indicates the highest temperature, in tenths of a degree Celsius, during the previous six-hour period. If the digit following the "1" is a "0", then the temperature is higher than 0 degrees Celsius (a "1" following the first "1" indicates that the temperature is less than 0 degrees Celsius). So the highest temperature at Concord between 06Z and 12Z on May 13, 2006, was 9.4 degrees Celsius (49 degrees Fahrenheit). Note that the "1" group is only reported at 00Z, 06Z, 12Z and 18Z.
  • 20089, the 2-group, indicates the lowest temperature during the previous six-hour period. If the digit following the "2" is a "0", then the temperature is higher than 0 degrees Celsius (a "1" following the "2" indicates that the temperature is less than 0 degrees Celsius). So the lowest temperature at Concord between 06Z and 12Z on May 13, 2006, was 8.9 degrees Celsius (48 degrees Fahrenheit). Like the "1" group, the "2" group is only reported at 00Z, 06Z, 12Z and 18Z.
  • 53018 indicates the pressure tendency (the "5 group"). The digit following the "5", which can vary from 0 to 8, describes the behavior of the pressure over the past three hours (for guidance, consult the table below). The last three digits represent the amount of pressure change in tenths of a millibar. Thus, the pressure at Concord increased 1.8 mb in the three-hour period ending at 12Z on May 13, 2006.
Descriptions of the behavior of pressure over the past three hours and the corresponding METAR code (third column).
Primary RequirementDescriptionCode Figure
Atmospheric
pressure now
higher than 3
hours ago.
Increasing, then decreasing.0
Increasing, then steady, or increasing then increasing more slowly.1
Increasing steadily or unsteadily.2
Decreasing or steady, then increasing; or increasing then increasing more rapidly.3
Atmospheric
pressure now
same as 3 hours
ago.
Increasing, then decreasing.0
Steady4
Decreasing then increasing.5
Atmospheric
pressure now
lower than 3
hours ago.
Decreasing, then increasing.5
Decreasing, then steady, or decreasing then decreasing more slowly.6
Decreasing steadily or unsteadily.7
Steady or increasing, then decreasing; or decreasing then decreasing more rapidly.8

I realize that translating one METAR hardly qualifies as an entire lesson, but at least you now know the general guidelines and where to find information in case you run across a METAR that gives you pause. I encourage you to expand your aptitude for decoding METARS - they hold a lot of information! If you want to completely master METAR code, you can study Chapter 12 of the Federal Meteorological Handbook (opens in a new window) No. 1, which is basically the METAR "bible" (section 12.7 in particular covers the wide variety of "remarks" that can appear). If you have aspirations for becoming an official weather observer, you may be interested in checking out the entire handbook (it's very long, but there is a table of contents to help you find specific topics).

Key Data Resources

You can access METAR observations taken at ASOS and AWOS sites at this interactive FAA Web site (opens in a new window) or the NCAR surface observation site (opens in a new window) (which also includes plots of regional station models).

What are the key take-away points from this section? The Key Skill box below highlights some key data groups with which you must become proficient in this course, and the Quiz Yourself section will give you a chance to get a little practice, too. Once you get the hang of it and memorize the general format and some common codes, it's a snap!

All of these observations we've been talking about, over long periods of time, constitute a station's climatology, and good forecasters investigate the climatology of the location they're forecasting for. We'll take a look at common climate data and what components are needed for a good understanding of a city's climatology next.

Key Skill...

This semester, you'll need to work with some METAR data groups repeatedly, so you must be proficient in translating them and applying them to forecasts and verifications. Forgetting about them can easily lead to an incorrect forecast verification, or perhaps making your next day's forecast with the wrong impression of today's observations! When interpreting METAR data, keep the following important points in mind:

Observation times: As you may remember from your previous studies, a certain hour's observation is typically issued several minutes before the hour. So, our KCON observation example on this page came from 1151Z, but that's actually the official 12Z observation.

Temperature observations: Many METARS have two temperature (and dew point) observations -- one that's rounded to the nearest degree Celsius, and a more precise version to the nearest tenth of a degree Celsius in the T-group. Make sure you use the T-group for conversions to Fahrenheit (if it's available). Using the less precise values rounded to the nearest degree Celsius can lead to incorrect Fahrenheit conversions!

Verifying high and low temperatures: We can't just use T-groups to determine the highest and lowest temperatures over a given period of time because they only tell us the precise temperature right at the observation time. They don't tell us anything about what happened in between observations, so you must look at the 1-groups and 2-groups at specific synoptic times in order to determine six-hour maximum and minimum temperatures, respectively. For example, if you wanted to determine a station's maximum and minimum temperatures over a 24-hour period from 06Z to 06Z, you would need to evaluate the 1- and 2-groups at 12Z, 18Z, 00Z, and 06Z to cover the entire 24 hour period.

Verifying total liquid precipitation: The "P" groups give us total liquid precipitation that has fallen in the past hour, but summing the "P" groups to get a liquid precipitation total for a longer period of time can be fraught with error--especially when "P" groups get reported with special observations in-between the hourlies. Thus, it's easiest to look at the 6-groups at specific synoptic times, and add them together in order to determine the total liquid precipitation over a period of time. For example, if you wanted to determine a station's total precipitation over a 24-hour period from 06Z to 06Z, you would need to evaluate the 6-groups at 12Z, 18Z, 00Z, and 06Z to cover the entire 24 hour period (each would give a six-hour total). You would ignore the 6-groups at 09Z, 15Z, 21Z, and 03Z because those three-hour totals would be redundant in this scenario.

Verifying maximum sustained wind: Hourly METARS only show us sustained winds at the time of the observation, but of course sustained wind speeds will vary (perhaps greatly) over the course of each hour. So, the maximum sustained wind speed will not likely occur exactly at the time of the hourly observations (we often can't verify maximum sustained wind speed from hourly METARS). For major stations, the National Weather Service tracks sustained wind speeds continuously and publishes the maximum sustained wind speed for each calendar day in its Daily Climate Summary.

Quiz Yourself...

Quickly gathering information from coded METARS is a useful skill for forecasters. To give you some practice in gathering the types of information you'll regularly need to gather in this course, here's 24-hours of METARS from Cheyenne, Wyoming (opens in a new window), spanning 06Z December 30 through 06Z December 31, 2014. Older observations are toward the bottom, while more recent ones are toward the top.

Question #1

During this 06Z - 06Z period, what was the maximum temperature observed at KCYS? What data group confirms it?

Answer: -18.9 degrees Celsius (-2 degrees Fahrenheit) as evidenced by the 1-group in the 12Z observation (11189). The 1-groups give us six-hour maximum temperatures at the synoptic times (12Z, 18Z, 00Z, and 06Z). So, we can quickly tell the maximum temperature during this period by merely looking at the 1-groups at the synoptic times and determining which one shows the highest temperature. There's no need to look at the temperature reported in every single observation.


Question #2

During this 06Z - 06Z period, what was the minimum temperature observed at KCYS? What data group confirms it?

Answer:  -30 degrees Celsius (-22 degrees Fahrenheit), from the 2-group in the 06Z observation on the 31st (21300). The 2-groups give us six-hour minimum temperatures at the synoptic times (12Z, 18Z, 00Z, and 06Z). So, we can quickly tell the minimum temperature during this period by merely looking at the 2-groups at the synoptic times and determining which one shows the lowest temperature. In fact, trying to answer this question by looking only at each observation (and ignoring the 2-groups) would have led you to the wrong answer since the lowest temperature occurred in between observation times.


Question #3

The temperatures reported in the 22Z and 23Z observations were both "M21". Were the temperatures actually the same at each time?

Answer: No. We can get more precise values using the T-groups. At 22Z, the temperature was -20.6 degrees Celsius (from the T-group: T12061256), which converts to -5 degrees Fahrenheit. At 23Z, the temperature had dropped to -21.1 degrees Celsius (from the T-group: T12111261), which converts to -6 degrees Fahrenheit.


Question #4

How much liquid precipitation fell during this 06Z - 06Z period?

Answer:  0.03", which we can determine by summing the totals in the 6-groups (which give us six-hour total precipitation at 12Z, 18Z, 00Z, and 06Z). There were two 6-groups at the synoptic times, "60002" at 12Z and "60001" at 18Z. Note that you'll also get three-hour precip totals at 09Z, 15Z, 21Z, and 03Z sometimes, but those values are included in the six-hour totals at the synoptic times (so we didn't need to consider them). Using the six-hour totals from the proper 6-groups is much faster than summing hourly precipitation from P-groups (which can get pretty difficult when there's lots of precipitation and lots of special observations coming in).

Explore Further...

While METARS are typically published every hour, the observations are being recorded continuously, and you can find some sites on the Web that provide more frequent METAR data. For example, here are translated 5-minute METAR observations (opens in a new window) for Los Angeles International Airport (if you want to see the actual encoded 5-minute METARS, just click on "Raw Observations"). By the way, if you want to view a different station, just change the "KLAX" in the URL to your desired station ID.

However, I urge caution when interpreting translated 5-minute METARS. The raw 5-minute METARS typically lack a T-group, which means decoders get the temperature in degrees Fahrenheit from rounded values in Celsius in the raw 5-minute METAR. That sometimes results in incorrect Fahrenheit conversions, so the temperatures you see in degrees Fahrenheit can sometimes be a degree off from the actual observation. Be careful with 5-minute METARS!

sas405

Climatology, Part I: Temperature and Precipitation Stats

Climatology, Part I: Temperature and Precipitation Stats

Prioritize...

When you've completed this page, you should be able to identify a station's period of record, discern between a city's ThreadEx data and climatology data from an individual station, differentiate between daily raw averages and normals, and retrieve and interpret daily normals and records for temperature and precipitation.

Read...

You can gain all the sophisticated knowledge of advanced forecasting techniques in the world, but if you fail at the fundamentals, you're not going to be a very good forecaster over the long haul. That's why the fundamentals in this lesson, while perhaps seeming very basic, are critically important to quality forecasting.

We've focused on observations so far because adding ground truth to your forecasting procedure through observations is vital, especially over short time periods (say, 12 hours or less). Current observations can still offer clues even for longer-range forecasts, but as we go out in time, specific current observations may be less helpful. Meanwhile, the importance of various types of numerical weather prediction (NWP) models and statistically post-processed forecasting products grows. Forecasters, however, must still view the guidance with well-grounded conceptual models of the atmosphere in mind. We'll be looking more at NWP and statistically post-processed forecasting products in upcoming lessons, but before we move on, there's one more critical "fundamental" we must cover--climatology.

Studying climatological statistics and local and regional topography around a given location can provide important insights about the historical bounds of weather variables and how the weather "behaves" in various weather patterns. This background knowledge gives forecasters essential context for what's typical and atypical at their forecast location. In order to assess a location's climatology, let's look at some important variables and issues that forecasters must be aware of.

Know Your Location and Period of Record!

This might seem obvious, but forecasters need to know the specific location where their forecasts will be verified and how long weather observations have been taken there (the "period of record") in order to understand a station's climatology. Imagine you're given the task of forecasting for Pittsburgh, Pennsylvania. If you're verifying your forecast at a single point (as is the case for many forecasters), your first question should be, "Where in Pittsburgh?" Are you forecasting for a location in downtown Pittsburgh? In a suburb? At the airport?

Map showing the location of KPIT compared to the old observation site in Pittsburgh

Pittsburgh International Airport (the current location of the official weather observations for Pittsburgh) is about 10 miles west and 200 feet higher than the old observation site downtown.
Credit: Applied Climate Information System

If you're forecasting for the downtown Pittsburgh area (where the city's official observations were taken from 1875 until 1952), its location is about 200 feet lower in elevation and about 10 miles east of Pittsburgh International Airport--KPIT (where the city's official observations have been taken since 1952). Recalling from previous studies that higher elevation locations tend to be a bit cooler than their lower elevation counterparts (all else being equal), it stands to reason that the specific climatology of the two locations might differ, at least slightly.

Since the location of official observations has changed over time in many large cities, scientists at NOAA have created what's called ”ThreadEx“ data (opens in a new window), which is a "threaded" historical record that stitches together observations from multiple historical observation sites into one "continuous" thread. For cities where a ThreadEx data set exists, it's considered the city's official climate record. For Pittsburgh, the period of record goes back to 1875, even though the period of record from KPIT at the airport doesn't go back nearly that far.

ThreadEx data is a great starting point to understand a city's climatology, but using it by itself can be problematic. Are the observations taken at the old observation site in downtown Pittsburgh (200 feet lower in elevation and more in the midst of the urban heat island) representative of conditions at the airport? Not really. For example, if we examine the history of the hottest days on record in Pittsburgh (opens in a new window) (highs of 100 degrees Fahrenheit or higher), it's easy to see that most of them occurred prior to 1952, when the official observation site switched to KPIT. In other words, historically, 100-degree days are more common downtown than they are at KPIT. In fact, two dates exist when the downtown observation site reached 100 degrees Fahrenheit and weather records were also being taken at KPIT. The highs at KPIT on those days were 99 degrees and 96 degrees, providing further evidence that 100 degrees is a harder threshold to reach at the higher elevation of the airport.

Forecast Tip

If you want to really dig into a city's climatology, just using the official climate record (the ThreadEx data) may not be sufficient. If climatological data exists for the exact location where your forecasts are being verified, the data for that site may be more useful, even if its period of record is shorter (as long as it's not too short). More information for retrieving a city's climate data is coming later in this section.

What Exactly is "Normal" Anyway?

When studying a location's climatology, a common place to start is with "normal" temperatures and precipitation. But, what does "normal" really mean? In your previous studies, you learned that normal temperatures are essentially 30-year averages, which get updated once every decade. That's basically true, but there's a little more to it than that. As it turns out, raw averages over a 30-year period can be a bit noisy, where one or two very warm / cold / wet days in the sample can skew the average for that date a bit. So, to eliminate the noise and paint a clearer picture of day-to-day and seasonal trends, the raw averages undergo some statistical smoothing to create the official climate normals for a given city. To see the difference between raw averages and official 1991-2020 normal high temperatures for State College, Pennsylvania, check out the graph below.

Graph of 1991-2020 raw averages and official normal high temperatures at State College, PA.

Raw 30-year averages (blue line) can be "noisy" from day to day, so official climate normals (orange line) are statistically smoothed to create a clearer picture of day-to-day and seasonal variability.
Credit: Steve Seman @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window) (Data from ACIS)

The blue line on the graph, which represents the raw average of daily high temperatures from 1991 to 2020 shows a number of "bumps" that are just noise and have no meteorological significance (a few are highlighted on this annotated version of the graph (opens in a new window)). So, to avoid being misled by the noise in day-to-day raw averages, forecasters typically prefer using the official (smoothed) normals (orange line). With precipitation, the "noise problem" with raw averages gets even worse (check out the corresponding graph for daily raw average and official normal precipitation at State College (opens in a new window)). The abrupt spikes and lulls that occur on a day-to-day basis are noise, so official (smoothed) normals are calculated to better identify seasonal trends. Note that there is a slight uptick in daily normal precipitation in the warm season (daily normals are generally greater than 0.10 inches), but it's hard to tell from the raw averages.

Forecast Tip

Official climate normals (which are commonly communicated to the public) are typically more useful for understanding average day-to-day weather behavior at a particular location than 30-year raw averages. Official normals are statistically smoothed values and typically get reported as whole numbers for temperature and as inches of liquid (to the nearest hundredth) for precipitation, which is consistent with measurement practices in both cases in the U.S.

What Records are Important?

While normals tell forecasters the (approximate) 30-year average conditions at a location, it's also helpful to know the extreme (record) values of temperature and precipitation. You may be most accustomed to "record high" and "record low" temperatures since they're what most commonly get reported, but they don't tell the entire story of temperature extremes.

For example, if forecasters wanted to know what constituted an extremely hot day at a particular location during a particular time of year, they would look up the record maximum temperature (or, "record high") for the date (and perhaps some surrounding days, too). But, the record maximum temperature only tells us the historical upper bound for maximum temperatures on a given date. The historical lower bound for maximum temperatures is important, too, because it lets forecasters know what constitutes an extremely cool (or cold) "daytime" in most cases. Knowing the record maximum temperature and recorded lowest maximum temperature for any given date tells the forecaster what the historical range of maximum temperatures is for the date.

Graph of 1991-2020 raw averages and official normal high temperatures at State College, PA.

Daily record maximum temperatures (red dots) and record lowest maximum temperatures (blue dots) tell forecasters what the historical range of maximum temperatures is for any given date.
Credit: Steve Seman @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window) (Data from ACIS)

Knowing the range of historical maximum temperatures at a given time of year can be of great use to forecasters. For example, the graph above shows record daily maximum (red dots) and daily record lowest maximum temperatures (blue dots) at State College, Pennsylvania for each day during the year. Very quickly, we can see that an extremely hot day during the hottest time of year (mid-July) in State College would have high temperatures in the upper 90s to low 100s. On the flip side, a very cool mid-July day would have high temperatures in the 60s! In mid-winter, the warmest days in January historically have had high temperatures in the 60s mainly, while the coldest days have had high temperatures in the single digits, or (in a few cases) even below zero.

The same idea applies to daily minimum temperatures. The daily record minimum (or "record low") tells forecasters the lower bound for minimum temperature. But, to get a complete picture of daily minimum temperatures, we need the upper bound, too -- the record highest minimum temperature, which tells us what constitutes an extremely mild or warm night (in most cases) at a given time of year. Precipitation, on the other hand, is a bit of a different story. Yes, daily record maximum precipitation values tell us the greatest amount of liquid precipitation that has fallen on any particular day, but the daily record minimum precipitation value would presumably be zero everywhere on Earth (it's pretty safe to assume that it's been totally dry at least once on every calendar day, everywhere).

Key Data Resource

Where can you find daily temperature and precipitation normals and records for a particular city? On the NWS Climate Page (opens in a new window), click on the NWS forecast office on the map that includes your desired location, and then select NOWData to access the data you need. On the list of NOWData stations, note that "[CityName] Area" represents the city's ThreadEx data. If you want data from a specific site (such as a major airport), you'll need to find the specific location on the list. Select, "Daily / Monthly Normals" for normal daily high and low temperatures and precipitation and the official normals will be displayed in several tables beneath a graph. For record temperatures and precipitation, select "Calendar Day Summaries" and the appropriate variables (for example, the variable "Min Temp" and summary type "Daily Maximum" will give you the record highest minimum temperatures for each day).

The bottom line: For maximum temperatures, the relevant extremes that mark the historical upper and lower bounds are record maximum temperatures and record lowest maximum temperatures. For minimum temperatures, the relevant extremes are record minimum temperatures and record highest minimum temperatures.

In the next section, we'll focus on how forecasters can analyze local terrain and historical wind observations to further assess a location's climatology. But, before you move on, check out the Quiz Yourself tool, so that you can test your ability to retrieve basic climatological statistics for a given city. The Explore Further section has some additional links that may be of interest to you, as well.

Quiz Yourself...

Imagine that you are going to be forecasting for Pittsburgh, Pennsylvania, with forecasts verified at Pittsburgh International Airport (KPIT). To retrieve some basic climatological stats, use the NWS Climate Page (opens in a new window), along with the guidance above, to answer the questions in the quiz tool below. Can you answer all five questions correctly? If not, make sure that you're searching for the correct variable and using the correct data set.

Explore Further...

If you want to dig a bit deeper into a particular location's temperature and precipitation climatology, you may be interested in the following links:

  • Threaded Station Extremes (opens in a new window): The home of ThreadEx data. You can investigate the official station history for many large U.S. cities., as well as gather a limited set of daily records from each available city's ThreadEx data set.
  • xmACIS (opens in a new window): This site contains a complete set of temperature and precipitation records for thousands of U.S. stations. The data is from the Applied Climate Information System (ACIS), which is the same database that runs the NWS Climate Page (linked above), but this site has a larger set of reports and graphs that can be generated (and therefore, is a bit less user friendly). Various interfaces into the ACIS database exist online in addition to xmACIS and the NWS Climate Page (SC-ACIS (opens in a new window) is another option with similar functionality), and if you're want to dissect a location's climatology, these sites have tons of data to explore.
sas405

Climatology, Part II: Winds and Terrain

Climatology, Part II: Winds and Terrain

Prioritize...

When you've completed this page, you should be able to retrieve monthly summary wind statistics from Local Climatology Data documents, interpret key tendencies from wind roses (most / least common wind directions and directions most / least likely to produce fast winds), and analyze important regional terrain features (elevation changes and bodies of water) that may impact a location's climatology, as well as their impacts based on local wind direction.

Read...

The temperature and precipitation statistics that we covered in the last section are a big piece of a particular location's climatological puzzle, but they don't paint the whole picture. A location's wind climatology also provides important clues about typical / atypical weather, and perhaps more importantly, when coupled with knowledge of the surrounding terrain like bodies of water and elevation changes, can help forecasters understand local influences that winds can have on other weather variables (like temperatures and precipitation).

First off, however, it's important to gauge what constitutes "windy" conditions at your forecast location. Perhaps you've never dug into wind data before, but your experience has probably told you that some locations are typically windier than others. Unfortunately, climatological wind data can be somewhat hard to come by because of a relative lack of properly sited observation stations (many privately owned and home weather observing equipment is not set up to measure wind speed at a standard height of 10 meters). Furthermore, publicly available wind data from ASOS sites is generally limited to hourly observations and some summary statistics for stations at larger airports (anything more typically requires some data mining).

Still, we can get a sense for what constitutes a windy day at select forecast locations by viewing summary statistics compiled by the National Centers for Environmental Information. For larger airports in the U.S., they produce "Local Climatological Data" documents, which contain a wealth of statistics, including mean and record wind speeds (sustained and gusts) for each month. For example, compare the historical wind data at Cold Bay, Alaska (opens in a new window) and Seattle, Washington (opens in a new window). Cold Bay is clearly a windier place, with monthly mean wind speeds that are roughly twice those in Seattle. Furthermore, the monthly record sustained wind speeds in Cold Bay are significantly higher, as are the record gusts (multiple months have had gusts of over 80 miles per hour in Cold Bay). The data for monthly record sustained wind speeds and gusts can be particularly helpful for gauging what constitutes unusually fast winds at a particular location during various times of the year.

Key Data Resource

Local Climatology Data documents from NCDC/NCEI (opens in a new window) contain a wealth of climatological statistics and other useful information. In addition to the monthly summaries of wind data described above, they also contain the exact location (latitude / longitude coordinates) and elevation of major airport stations. They also include a brief, but informative climatology discussion toward the end of the document that can really help forecasters get their local bearings.

A very common visualization for studying a location's wind climatology is a wind rose, which is a plot that displays historical wind direction and speed information. You encountered wind roses during your previous studies, and if you need a refresher on how to interpret them, I recommend this quick discussion on the basics (opens in a new window) (keep in mind, however, that display conventions can vary from website to website). Ultimately, wind roses are a great way to quickly assess from which directions winds most and least commonly blow, and assess which directions are more (or less) prone to particularly speedy winds.

Image slider (hold your mouse over the left or right side of the image and click the arrow to change the image) showing wind roses from January and July in Miami, Florida.
Credit: Credit: Natural Resources Conservation Service

Recurrent patterns in winds may become apparent at a particular location because of regional terrain (winds being channeled through mountains or valleys, for example), persistent mesoscale circulations (like sea breezes or mountain-valley circulations), or persistent storm tracks, and these patterns may change throughout the year. For example, check out the wind roses for Miami, Florida (above) from January and July.

Note that the Miami wind roses for January and July look very different. The most frequent wind direction (marked by the longest spoke) is from the north-northwest in January (blowing from that direction nearly 14 percent of the time). Winds from the east and southeast are also fairly common in January, but not quite as common as north-northwest winds. In July, however, east-southeast winds are the most common (occurring nearly 19 percent of the time), and winds from the north-northwest are much less frequent. It's not even close! In fact, if we sum the percentages of winds from the east, east-southeast, and southeast (the three longest spokes), winds from those directions occur nearly 50 percent of the time combined!

Any ideas why winds from those three directions are so common in July? A big part of it is the regular occurrence of sea breezes as the Florida peninsula warms up during the day and a mesoscale sea-breeze circulation develops, resulting in onshore flow at Miami (opens in a new window). During the summer, this onshore flow ushers in slightly cooler, moist air into south Florida. The reliable sea breeze is a big reason why temperatures have only hit 100 degrees in Miami once since records began in 1895 (per the ThreadEx data). Daily record highs in Miami during the summer are typically in the middle and upper 90s, which is comparable to or even lower than many locations that are much farther north in the U.S. So, wind direction, and the trajectory of the air as it approaches your forecast location are important considerations!

Given the forecasting impacts that air trajectories can have, getting the lay of the land around your forecast site and the surrounding region is critical. Large bodies of water and elevation changes are particularly influential. To get a feel for the lay of the land around your forecast location, it's always a good idea to study a detailed topographic map of the local area, in order to judge the local elevation changes and distance from your forecast site. I'll use Los Angeles International Airport (KLAX) as an example. As you can see from the Google map below, the terrain around Los Angeles is complicated! Many terrain features influence the weather at KLAX!

Topographic map of the Los Angeles area.

Many important topographic features can influence the weather at Los Angeles International Airport. Full-sized image (opens in a new window).
Credit: Google Maps

For starters, the Pacific Ocean is just west of the airport, and the coastline wraps around such that the ocean also lies about 15 miles to the south. Therefore, any time low-level winds blow onshore (having a large westerly or southerly component), they'll bring a substantial maritime influence, which during certain times of year can lead to fog or low cloud development. But, given the substantially higher elevations that are located around the Los Angeles Basin (KLAX sits at an elevation of 97 feet), considerable downsloping occurs when low-level winds blow from the northwest all the way around to the southeast.

Such downsloping brings a warming and drying influence (from compressional warming), with the greatest downsloping occurring on northeast and east-northeast flow since that's where the highest elevations are locally. Indeed, within 30 to 40 miles northeast of KLAX, multiple elevations in excess of 7,000 feet can be found! That's a lot of downsloping! Even downsloping from smaller ranges of hills and mountains (like the Hollywood Hills about 12 miles north of the site, or the Palos Verdes Hills about 12 miles south) is considerable. If air parcels descended the Palos Verdes Hills dry adiabatically (warming at a rate of 5.5 degrees Fahrenheit per 1,000 feet), that's worth almost 8 degrees Fahrenheit of compressional warming between the hills and the airport (assuming a starting elevation of 1,500 feet)! So, it pays to study the terrain surrounding your forecast location even for fairly subtle elevation changes. It's worth considering the impacts for even elevation changes of a few hundred feet or more within several miles.

The farther you go from your forecast site, it takes a more significant topographic feature (body of water or elevation change) to make a significant contribution to a location's climatology. In general, it's wise to assess the terrain within several hundred miles for major bodies of water or elevation changes. And, sometimes, the elevation changes don't take the form of well-defined, obvious hills or mountains. To get an idea of how a forecaster might assess the larger-scale terrain around a forecast location, check out the short video below (4:16), which contains a few examples.

Each forecast location can have its own unique topographic influences that are an integral part of its weather and climate, and it’s important to look at both local factors, which may be within only a few miles or tens of miles, and larger-scale regional factors. In this video, I’ll show some examples of the types of larger-scale regional factors that forecasters need to consider. Some of the factors and their influences are fairly obvious, while others might be more surprising. To start, imagine that you’re forecasting for Omaha, Nebraska. Nebraska is located in the Central Plains of the United States, so it’s easy to think that the terrain in the state is flat, but it’s not.

A closer look at a topographic map of Nebraska shows Omaha over here in extreme eastern Nebraska. The official observation site at the airport sits at an elevation of 982 feet, but elevation steadily increases toward western Nebraska. In fact, the town of Scottsbluff, Nebraska sits at an elevation of nearly 4,000 feet! So, there’s an elevation difference of about 3,000 feet between these two cities across the supposedly flat plains, and there are even some higher elevations to the south of Scottsbluff. The elevation changes are very gradual – the distance between Omaha and Scottsbluff is about 400 miles, but they’re elevation changes are still significant. Any wind with a large westerly component will downslope across Nebraska toward Omaha. Of course, farther to the west, just off the map is the Rocky Mountains, so forecasters have to think about downsloping anytime low-level wind flow has a large westerly component across Nebraska toward Omaha. Downsloping causes warming and drying, so if there’s no cold advection occurring, such wind flow can have a considerable warming influence. Even if cold advection is occurring, downslope warming can counteract some or all of it, depending on the speed of the wind and the strength of the advection.

Now, imagine you’re forecasting for Philadelphia, Pennsylvania. Philadelphia International Airport sits at an elevation of 10 feet above Mean Sea Level. But, there are a number of hills west and northwest of the city, so elevations of 300-500 feet are common even within 20 miles of the airport. Back in Central and Northern Pennsylvania, numerous ridge tops in the Appalachians are between 1,000 and 2,000 feet and we’re still less than 200 miles from the airport. The Poconos are less than 100 miles north of Philadelphia, with some elevations exceeding 2,000 feet. So, winds with a notable northerly or westerly component will downslope toward Philadelphia, bringing a warming and drying influence. Off the map to the east, the Atlantic Ocean sits a little more than 50 miles to the east, and Delaware Bay sits less than 50 miles to the south. So, winds from the south or with notable easterly components will bring a maritime influence, which may moderate temperatures and bring moisture.

Finally, let’s look at Portland, Oregon. The airport lies at a low elevation of 19 feet above mean sea level. About 65 miles to the west is the Pacific Ocean, but in between the ocean and Portland is the coastal range, with numerous elevations between 2,000 and 3,000 feet within 40 miles of Portland. So, winds with large westerly components bring a moderating, maritime influence from the Pacific, but they also downslope considerably between the coastal range and Portland. The drying influence makes the airport one of the drier locations in the area, on average. About 30 miles to the east, we get up into the Cascades, with numerous elevations in excess of 4,000 feet, and Mount Hood in excess of 11,000 feet. So, winds with a significant easterly component will downslope significantly from the Cascades toward Portland, also bringing a warming, drying influence.

As I mentioned, each location can have its own unique topographic influences, so forecasters must take the time to get to know the lay of the land, and think through what impacts changes in elevation and bodies of water have on local weather and climate.

Credit: @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

From the video, take note that quantitatively estimating elevation changes and the distance of a particular feature from your forecast location is important, as it gives context for how significant a particular feature may be. Performing these types of analyses also helps forecasters understand the influences that various wind directions may have on the weather at their forecast location in terms upsloping (which brings cooling and potential cloud / precipitation development) or downsloping (warming and drying), or a potential maritime influence.

Ultimately, each location can have its own unique topographic influences that are an integral part of its weather and climate, and it's up to the forecaster to analyze the location's surroundings and determine the relevant features and impacts. I hope the sampling of examples gives you an idea of the types of features and influences that forecasters need to look for, and with your fundamental knowledge from your previous studies, you should be able to perform basic analyses like these on your own. If you want a little practice in making these kinds of assessments for yourself, check out the Quiz Yourself section below for a few more examples. In the next section, I'll show you a Case Study detailing how a forecaster might compile relevant climatology information for a particular city.

Key Data Resources

Quiz Yourself...

Question #1

Imagine that you are forecasting for KMAF in Midland, Texas (elevation 2,862 feet). Given these local-scale (opens in a new window) and state-level (opens in a new window) terrain maps (here's the elevation key (opens in a new window)), describe the important topographical features (including reference elevations and distances) and the impacts of wind direction on local weather based on the terrain. Feel free to explore the area around Midland International Air and Space Port in Google Maps (opens in a new window), if you wish.

Answer: Midland is located about 500 miles northwest of the Gulf of Mexico on the Plains of Texas. Elevations near the city don't change very much, but there is a gradual decrease in elevation toward the northeast, east, and southeast, such that elevations fall below 2,000 feet less than 100 miles from the site and continue decreasing from there. Toward the southwest, west, and northwest, elevations gradually increase toward the mountains of West Texas and New Mexico, where elevations exceed 5,000 feet less than 200 miles from the site.

Flow from the southeast would import moisture from the Gulf of Mexico, which combined with gradual upsloping (a cooling mechanism) can sometimes lead to low clouds, fog, and/or drizzle. Some upsloping occurs on any wind with a substantial easterly component. Flow with a large westerly component downslopes from higher terrain in Mexico, West Texas, and New Mexico, which favors compressional warming and drying.

Note that the exact elevation references and distances that you chose may differ from the ones mentioned here, but a good description needed to reference the Gulf of Mexico and the elevation changes to the east and west (along with upsloping and downsloping impacts).


Question #2

Imagine that you are forecasting for KALB in Albany, New York (elevation 280 feet). Given these local-scale (opens in a new window) and state-level (opens in a new window) terrain maps (here's the elevation key (opens in a new window)), describe the important topographical features (including reference elevations and distances) and the impacts of wind direction on local weather based on the terrain. Feel free to explore the area around Albany International Airport in Google Maps (opens in a new window), if you wish.

Answer: KALB is located in the Hudson River Valley, with rolling hills generally ranging from 200-500 feet near the station (lower elevations are near the river, about 6 miles east). The Helderberg escarpment rises to about 1800 feet 11 miles to the west, while about 12 miles to the east, terrain rises up near 2000 feet. Even higher elevations (3000-4000 ft+) lie about 40+ miles northwest and northeast of the site in the Adirondacks and Green Mountains, respectively. The Atlantic Ocean lies about 120 miles south and east of KALB, while Lake Ontario is about 100 miles northwest.

Winds from most directions downslope into Albany, which has a warming and drying influence, with the greatest downsloping occurring on northwest, southwest, and east/east-northeast winds,. This largely inhibits lake-effect precip (and to a lesser extent, clouds) on northwesterly flow. Long-duration southerly – easterly flow brings maritime, moderating influences from the Atlantic, but they are mitigated somewhat (primarily on easterly flow) by downsloping.

A good description needed to reference Albany's valley location, rolling hills nearby, and the higher elevations surrounding it (along with distance and elevation references), as well as the major bodies of water in the region (Atlantic Ocean and Lake Ontario) and potential downsloping and maritime influences.


Question #3

Imagine that you are forecasting for KRDU in the Raleigh-Durham Metropolitan Area in North Carolina (elevation 416 feet). Given these local-scale (opens in a new window) and state-level (opens in a new window) terrain maps (here's the elevation key (opens in a new window)), describe the important topographical features and the impacts of wind direction on local weather based on the terrain. Feel free to explore the area around Raleigh-Durham International Airport in Google Maps (opens in a new window), if you wish.

Answer: Terrain in the Raleigh-Durham metro area is generally rolling with elevations within 100-200 feet of the station elevation of KRDU. Higher elevations start about 10-20 miles west of the area on the Piedmont Plateau (generally 500-1000 feet). Higher elevations in the Appalachian Mountains are located about 150 miles to the west, up to 3000-6000 ft+. The Atlantic Ocean lies about 150 miles east and south of Raleigh, with lower elevations on the coastal plain in the eastern part of North Carolina.

Northwest to west-southwest winds will downslope from the Appalachians and Piedmont Plateau toward KRDU (a warming, drying influence). Winds from the south and those with a large easterly component will have a maritime influence from the Atlantic, with slight upsloping (a slight cooling influence, which combined with moisture from the Atlantic could lead to low cloudiness and precipitation depending on the large-scale weather situation).

A good description needed to reference rolling hills near KRDU, and the higher elevations to the west (along with distance and elevation references), as well as the Atlantic Ocean and resulting downsloping, upsloping, and maritime influences on specific wind directions.

sas405

Case Study: A Climatology Example

Case Study: A Climatology Example

Prioritize...

When you've completed this page, you should be able to create a climatology similar to the one below.

Case Study...

I recommend that you use the example climatology below as a reference guide. It serves as a template that you can follow when you collect climatological information for forecast cities this semester. The most important parts of the example are the annotations (gray text), which provide additional explanations and descriptions about how each piece of data was collected (much of which came from the Local Climatological Data Document (LCD) for Los Angeles International Airport (KLAX) (opens in a new window), a KLAX wind rose, and the NWS Climate Page for Los Angeles (opens in a new window)).

Climatology of Los Angeles International Airport (KLAX)

Time Period

This climatology focuses on February 9-12 and 16-19 (These were the forecast dates for Los Angeles in the WxChallenge forecasting competition in 2021. General time periods are listed on the WxChallenge schedule page (opens in a new window).)

Topography and Geography

Local Time Zone (Conversion from UTC): Pacific Standard Time (UTC-8) (note that if daylight saving time was in effect, the conversion would be to Pacific Daylight Time (UTC-7)).

Station Elevation: 97 feet (Taken from page 2 of the LCD for KLAX (see the annotation for station elevation in the LCD (opens in a new window)).

Station Location: KLAX is located about 10 miles southwest of downtown Los Angeles, just inland from the Pacific Coast, as shown on this Google map (opens in a new window). (Make sure to describe the location of the station with respect to the center of the nearest major city, which can help to assess urban heat island effects).

Important Topographical Features: The Pacific Ocean is located just west and about 15 miles south of KLAX. The southern California coastal ranges line the inland side of the coastal plain that surrounds the airport. About 12 miles south of the airport, the Palos Verdes Hills rise to about 1,500 feet, while the Hollywood and Santa Monica Hills rise to near 2,000 feet about 12 miles north of the station. Other various hills rise to several hundred feet between the airport and downtown LA. A bit farther away, elevations rise in excess of 3,000 feet about 40 miles to the east-southeast and up to 6000-7000 feet + in Angeles National Forest, about 20 miles north and northeast. These high elevations extend off to the northwest, as well. (Generally, bodies of water and regional elevation differences are the important features to note. Make sure to include some reference distances from the station for significant bodies of water, mountain ranges, and even hills (along with representative elevations).

Topographic map of California

A relief map of California. KLAX is located in the Los Angeles Basin along the Pacific Coast.
Credit: John Hopkins University

Winds

Wind Rose: The February wind rose is below. (Note that the color-coded key to quantify wind speeds at the bottom of the wind rose can be a bit difficult to read, so here is a zoomed-in version (opens in a new window). Winds on these particular wind roses are expressed in meters per second.)

February wind rose for KLAX.

The February wind rose at Los Angeles International Airport
Credit: Natural Resources Conservation Service

Frequency (percentage) of the single most common wind direction: Winds from the west-southwest occur roughly 16.5 percent of the time during February (estimate based on the length of the longest spoke).

Directions that are most and least common: Clearly, winds from the WSW, W, and E are the most common (largest spokes). Winds from the SSW, S, SSE, and WNW-NNW are the least common (smallest spokes).

Direction(s) most likely to produce the fastest winds: W, WSW (largest area of green (and some cyan) shading).

Direction(s) least likely to produce the fastest winds: NE, ENE, SSW. Winds from these directions almost never blow faster than 16.5 knots (smallest blue shading along the spokes and minimal, if any, green areas).

Impacts of wind direction on local weather: Onshore wind directions (large westerly components or southerly components) will bring moist, moderating flow, that can lead to fog or low cloud development. Considerable downsloping occurs on winds from the northwest all the way around clockwise to the southeast, as air flows down surrounding hills and mountains into the LA basin. Such flow would have a warming, drying contribution. The most substantial downsloping comes on northeast and east-northeast flow. (Make sure to note the implications of the wind interacting with the topographic features you mentioned earlier in the climatology. Impacts on precipitation and temperature caused by wind flow interacting with bodies of water or elevation changes are particularly important.)

Maximum observed two-minute wind speed for the month (or months): 38 mph = 33 knots (converted from mph to knots from the "Normals, Means, and Extremes" table on page 3 of the LCD—see the annotation for monthly maximum wind speed (opens in a new window)). (Make sure that you don't use the table containing data only from the most recent year).

Temperatures

All data taken from the NWS local climate page for Los Angeles by clicking on "NOWData." I chose data for "Los Angeles Intl, CA" so that I could get data specifically for KLAX (I did not choose the "LA Airport Area" or "LA Downtown Area" options since those are the ThreadEx datasets that may incorporate multiple sites). For normals, I selected "Daily / Monthly Normals" and "Daily" for data type, then retrieved the relevant values from the tables (not the graph). For records, I used "Calendar Day Summaries" and selected my variables accordingly. For daily records, make sure you're using the entire period of record at the station (you may need to enter "por" into the year range if it's not there by default).

Each WxChallenge city consists of two forecasting weeks (4 days each) and forecasts are submitted Monday - Thursday, so the days we're actually forecasting for are Tuesday - Friday. In this section, you want to list the data for each day we're forecasting for (Tuesday - Friday each week). This climatology was prepared in February, 2021 and our forecast days were February 9-12 and 16-19, so the data below come from those dates.

Normal and Extreme Daily Temperatures
DateNormal Maximum (ºF)Normal Minimum (ºF)Record Maximum (ºF)Record Minimum (ºF)Record Lowest Maximum (ºF)Record Highest Minimum (ºF)
Tuesday 2/9665088385160
Wednesday 2/10665085374960
Thursday 2/11665088365558
Friday 2/12665088345361
Tuesday 2/16665088365759
Wednesday 2/17665087375458
Thursday 2/18655082395460
Friday 2/19655087385458
RANGE65-66All days are 5082-8834-3949-5758-61

Precipitation

All data taken from the NWS local climate page for Los Angeles. Data can be retrieved by similar methods used to retrieve daily temperature data.

Normal and Extreme Daily Precipitation
DateNormal (inches of liquid)Record Maximum (inches of liquid)
Tuesday 2/90.101.66
Wednesday 2/100.112.50
Thursday 2/110.101.98
Friday 2/120.112.15
Tuesday 2/160.121.82
Wednesday 2/170.121.88
Thursday 2/180.111.29
Friday 2/190.123.49
RANGE0.10-0.121.29-3.49

With a complete climatology at your disposal, you should keep these statistics in the back of your mind (refer to your climatology as often as needed) as you make your daily forecasts. In addition to helping you get to "know" your forecast location and common weather influences, this data will help you see when weather patterns depart dramatically from "normal" and near the extremes. During these times, you'll want to employ some strategies that we'll learn about later in the course.

sas405

Lesson 2. Ensemble Forecasting

Lesson 2. Ensemble Forecasting

Motivate...

Today, weather forecasters have a vast array of tools available to them when making weather forecasts -- far more than forecasters in prior decades! While studying observations is a critical first step in making any short-term weather forecast, forecasters use numerical weather prediction models (NWP models) as guidance when making their forecasts, especially for forecasts beyond a few hours into the future. NWP models are very powerful tools, but as you've learned in your previous studies, they produce flawed simulations of future states of the atmosphere. I won't rehash the basics of computer models and their flaws in this lesson, but if you're in need of a refresher, I recommend revisiting Lesson 10 (opens in a new window) in Meteo 101 (opens in a new window) (login required).

With many models to choose from, rookie forecasters often choose the "model of the day" -- the model they think is going to perform better for some reason. Too often, however, these choices are grounded in poor reasoning based on past model performance or reputation (or in some cases complete guesses), and when model forecasts are vastly different, choosing the "model of the day" can be very risky and hazardous to the accuracy of your forecast!

For example, check out the four deterministic forecasts for total precipitation over a five-day period, from 12Z on August 26, 2021, through 12Z on August 31, in the animation below. These forecasts came from the ECMWF, Canadian (GDPS), GFS, and UKMET computer models. They all show a swath of heavy rain predicted from the landfall of Hurricane Ida, but disagreed on the details (location of heaviest rain, etc.).

Deterministic forecasts for total precipitation over a five-day period from 12Z on August 26, 2021 through August 31.

Five-day total predicted precipitation from the ECMWF, GDPS (Canadian), GFS, and UKMET computer models ending at 12Z on August 31, 2021. All models showed a swath of very heavy rain from Hurricane Ida, but disagreed on its placement.
Credit: Pivotal Weather

Differences like the ones shown in the forecasts above aren't uncommon, but they can have major consequences for the weather at any single location. Imagine you had to forecast for Alexandria, Louisiana (annotated on this version of the animation (opens in a new window)). Choosing any one of these models would be extremely risky. The ECMWF predicted more than six inches of rain at Alexandria during the period, while the GDPS and GFS both predicted less than one inch! The UKMET was in the middle at 3.9 inches. Choosing a model of the day could result in forecast messaging for potentially serious flooding in Alexandria (if the ECMWF or UKMET was reasonably close), or a forecast of hardly any impact from rain at all if the GDPS or GFS was closer to reality!

This forecast was definitely uncertain! So, what tools do we have to deal with uncertainty in weather forecasts? From your previous studies, you're already familiar with a primary tool that forecasters use to assess uncertainty: Ensemble forecasts, which are a set of different forecasts all valid at the same time. Taken together, the deterministic forecasts above could be considered a crude ensemble, showing a range of possibilities that forecasters could use to gauge the uncertainty in the forecast. This forecast was clearly sensitive to the exact path that Hurricane Ida would take, which was also uncertain at the time, as shown by the 12Z run of the ECMWF ensemble on August 26 (opens in a new window) (you may remember ensemble track forecasts like this one from when you studied tropical forecasting).

Gauging the uncertainty in a forecast is critical and is a huge challenge for forecasters! Utilizing ensemble forecasts can help, but ensemble forecasting is not perfect (as you will learn). Still, using ensemble forecasts appropriately will help you think probabilistically even while creating deterministic forecasts. When combined with sound meteorological analysis of the weather pattern, such an approach will, on average, make you a better forecaster, and is part of a wise forecasting philosophy.

In this lesson, we'll delve into the details of how ensemble forecasts are created and how best to interpret them (which will require some discussion of statistics), and we'll talk about commonly used ensemble forecast products. But, before we get into those details, I think it's worth checking out a Case Study -- a high-stakes forecast where initial deterministic forecasts didn't turn out very well. Let's go!

sas405

Case Study: When Deterministic Forecasts Go Wrong

Case Study: When Deterministic Forecasts Go Wrong

Prioritize...

From this page, you should focus on the insights that ensemble forecasts can provide, in contrast to choosing a "model of the day" from the deterministic operational model runs. Also note the significant difference in "weather" that occurs at a given location when a storm track forecast goes awry.

Case Study...

In September of 2018, forecasters along the East Coast were fixated as Hurricane Florence (opens in a new window) made its trans-Atlantic journey toward the United States. While the National Hurricane Center's forecasts for Florence's path were excellent leading up to landfall near Wrightsville Beach, North Carolina early on the morning of September 14 (the center of NHC's forecast cone (9/09) (opens in a new window) was nearly over Wrightsville Beach five days in advance), deterministic forecasts for rainfall after Florence made landfall left a lot to be desired.

This forecasting failure was a classic case of forecasters creating a deterministic forecast that didn't effectively communicate the true uncertainty of the forecast situation. Roughly three days before Florence made landfall in North Carolina, forecasters started to beat the drum for apocalyptic rainfall inland in the Mid-Atlantic. On the morning of September 11, the Weather Prediction Center issued a rainfall forecast (opens in a new window) calling for a swath of double-digit rainfall totals from the North Carolina coast northward into Virginia through the morning of the 18th. The forecast showed 15 to 20 inches near Richmond, Virginia, which at the time would have crushed the local record for rainfall over a seven-day period (opens in a new window), likely resulting in catastrophic flooding.

This forecast was recreated and broadcast by some television stations (opens in a new window), and was passed around social media without further explanation or context by news reporters (example (opens in a new window)). Not surprisingly, such a dire rainfall forecast garnered the attention of residents of eastern Virginia (not to mention municipalities, emergency managers, and businesses hoping to maintain continuous operations). But, if you remember what you learned about threat scores in Lesson 1 (recall WPC's statistics for 1 inch QPF (opens in a new window)), you realize how risky such forecasts were three or more days in advance (in 2018 and today). Of course, a three- to seven-day forecast related to impacts from a tropical cyclone has considerable uncertainty, and the forecast cones from NHC reflect this notion. The National Hurricane Center issued the five-day forecast shown below at 5 A.M. EDT on Tuesday, September 11. The black dots show NHC's official forecast positions for Florence's center of circulation, while the cone represents the area within which the eventual path of Florence's center could fall (the width of the cone is based on historical forecast errors (opens in a new window)). The expansion of the cone with time and the different shading convention for days four and five indicate the growing uncertainty in the forecast.

NHC forecast cone for Hurricane Florence, issued at 5 A.M. EDT on September 11, 2018.

The National Hurricane Center forecast cone for Florence, issued at 5 A.M. on September 11, 2018 showed that in five days, the center of Florence could likely be anywhere from southern South Carolina to northern Virginia or West Virginia.
Credit: National Hurricane Center

The cone of uncertainty was pretty darn wide after Florence made landfall, wouldn't you agree? The large uncertainty in Florence's path after landfall was inconsistent with dire, deterministic forecasts for cataclysmic rainfall in specific inland locations. Such deterministic forecasts failed to convey the fact that while historically heavy rain was likely somewhere inland, pinpointing the location of such heavy rain three days before landfall even occurred was practically impossible because the location of heavy rainfall would greatly depend on Florence's path after landfall (which was predicted to be slow and meandering). Focusing in on a deterministic forecast from a "model of the day," like the GFS run at 12Z on September 11 (opens in a new window), which generally supported the idea of extremely heavy rain in eastern Virginia, was a risky forecast strategy.

While some deterministic model out put like the GFS was in line with WPC's rainfall forecast, other reasonable possibilities existed, and ensemble forecasts were critical to assessing those other possibilities. To see how ensembles signaled that WPC's deterministic rainfall forecast was far from a lock, check out the short video below.

Three days before Hurricane Florence made landfall near Wrightsville Beach, North Carolina, WPC issued this 7-day forecast for total rainfall, for the 7-day period ending at 8 AM Eastern Time on September 18, 2018. They predicted more than 20 inches of rain near the predicted point of landfall in North Carolina, with a swath of extremely heavy rain northward into eastern Virginia. Areas around Richmond were predicted to receive in excess of 15 inches of rain, which would have crushed the all-time record for 7-day rainfall in Richmond at the time.

Florence was predicted to slow to a crawl near landfall as upper-level steering currents over the storm weakened, as suggested by these GEFS forecasts of 500-mb heights, valid at 12Z on September 14. The footprint of Florence is evident in the height field near the North Carolina Coast in all members, but the lack of nearby height contours outside the storm’s circulation in any of the members showed a strong consensus for weak middle tropospheric winds, which would have led to slow, possibly erratic motion. Such a slow moving tropical cyclone made extremely heavy rainfall likely, but the exact locations receiving extreme rainfall would be determined by the storm’s exact path, which was not set in stone when WPC issued their deterministic precipitation forecast.

At the time that WPC issued their forecast for record-setting rainfall in eastern Virginia, the most recent GEFS track forecasts for Florence showed a consistent message from many members of slow movement near and after landfall, as indicated by the forecast position dots all bunched together. However, many members showed a northward drift. Such paths very well could have led to heavy rain in eastern Virginia capable of producing catastrophic flooding.

The 12Z run of the ECMWF ensembles on September 12 strengthened the message from the run a day earlier – they strongly suggested that Florence would drift west-southwestward toward South Carolina after landfall, a scenario that would spare eastern Virginia from apocalyptic rainfall totals.

The GEFS ensemble tracks from 18Z on September 12 fell in line with the ECMWF ensemble, boosting confidence that eastern Virginia would likely be spared the worst of the rainfall, and instead parts of South Carolina needed to prepare for extreme rainfall and major flooding.

WPC followed these messages, and began shifting their forecast on the 12th. By the afternoon of September 13, their forecast showed the heaviest rainfall near landfall and into northeastern South Carolina. But, by the time this forecast was issued, Florence was less than a day from landfall.

So, what happened? The heaviest rain did fall from eastern North Carolina, where some locations received in excess of 30 inches of rainfall, into northeastern South Carolina. Areas where the original WPC forecast predicted up to a few inches of rain received more than 10 to 15 inches. Flooding was rampant. Meanwhile, the area of eastern Virginia that had originally been identified in the swath of heaviest rain – where upwards of 15 inches or more was predicted to fall, received no more than a few inches. And, many areas received less than one inch.

Credit: Penn State

WPC's deterministic forecast for record-setting rain in eastern Virginia was predicated on Florence taking a path more in line with the GFS or many of the GEFS members, but the ECMWF ensembles suggested that such a path was not a lock. Indeed, the forecast for 10 to 15 inches of rain (or more) in eastern Virginia failed miserably, as did the forecast for only a few inches of rain in northeastern South Carolina (where 10 inches or more fell in several areas). Forecasters who consulted ensembles knew that WPC's prediction was not a lock, but unfortunately, many members of the public did not know that.

Tell them what you know, not what you don't know ...

So, how could forecasters have accounted for the uncertainty in the forecast situation? For starters, there's a wise old saying in the weather forecasting community: Tell them what you know, not what you don't know. Did forecasters on September 11 know that 15 to 20 inches of rain would fall near Richmond, Virginia from Florence? No! But, at that point they knew that if the center of Florence meandered northward after landfall, eastern Virginia would likely be inundated with heavy rain. On the other hand, if Florence took a more west or west-southwestward path, heavy rain likely wouldn't materialize over eastern Virginia, and would instead occur toward the southern flank of NHC's forecast cone (9/11) (opens in a new window). Such messaging would have alerted the public and other stakeholders of the possibilities for heavy rain, but also given some attention to the fact that the forecast was still uncertain.

A day later (on September 12), the picture was becoming clearer that the southern part of the NHC forecast cone was the more likely target for flooding rains, and forecasters that worked hard to make informative probabilistic forecasts could now express more confidence. Knowing that multiple realistic possibilities existed through the use of ensembles was the first step in communicating an informative probabilistic forecast.

As it was, deterministic forecasts for record rainfall in eastern Virginia (opens in a new window) viewed by the public with little or no additional messaging or context sounded false alarms for folks in eastern Virginia and perhaps brought a false sense of security for folks in northeastern South Carolina (where major flooding occurred). I hope that this case gives you a clear message that using ensemble forecasts can help you better communicate forecasts and, if nothing else, can help you keep probabilities in mind even when you have to make deterministic forecasts (allowing you to hedge your bets toward more likely scenarios). But, how do ensembles really work? Let's take the first step in really understanding the process and look at something called "lagged-average forecasting" on the next page.

sas405

Model Jumpiness and Lagged Average Forecasts

Model Jumpiness and Lagged Average Forecasts

Prioritize...

When you've completed this page, you should be able to use a "lagged average forecast" technique as a crude ensemble to help deal with "model jumpiness."

Read...

In a perfect world, all computer model initializations would be dead-on accurate, and computer models would precisely simulate all atmospheric processes (like the deep, moist convection in the image below). If you hold out hope for such utopia, keep on dreaming, folks. First, we will never be able to exactly measure all meteorological parameters at all relevant atmospheric levels; there will always be space between observing sites, and we can never, ever expect to know the precise conditions at every point within the domain of each model. Second, we can never expect to model all aspects of the atmosphere perfectly. Recall that numerical weather prediction models use imperfect parameterizations (oversimplifications) to handle complex processes that occur on very small scales.

A thunderstorm.

Thunderstorm clouds near Duck, North Carolina in June, 2016.
Credit: Steve Seman

That's all bad news because model errors on small scales grow faster in time, on average, than errors that occur on larger scales. Forecast errors at small scales, propagating relatively quickly to larger scales, is the underpinning of chaos theory. You may recall from your previous studies that the premise of chaos theory helps to explain the growing uncertainty of computer guidance with increasing time.

"Model Jumpiness"

Of all the red flags associated with computer guidance, none should be more alarming than "model jumpiness." I unofficially define "model jumpiness" as a sequence of runs with varying solutions. For example, the 12Z runs of the GFS (opens in a new window) and ECMWF (opens in a new window) models on November 15, 2021 both showed a major mid-latitude cyclone impacting the Mid-Atlantic and Northeast U.S. on the morning of November 23, during the typically busy Thanksgiving holiday travel period. The potential for heavy rain, snow, and gusty winds led to a flurry of stories in popular media about travel headaches leading up to the holiday (USA Today via Yahoo! (opens in a new window); full story (opens in a new window)).

But, just one day later on November 16, the GFS and ECMWF forecasts valid at the same time (12Z on November 23) looked very different. Both models (as shown in the slider below) showed a weaker low-pressure system centered much farther northeast compared to the runs the day before.

Image slider (click the arrow to change the image) Showing the 168-hour forecasts from the GFS and ECMWF models initialized at 12Z on November 16, 2021 (valid at 12Z on November 23).

Credit: Tropical Tidbits

Over the next several days, the trend toward a weaker storm that was farther offshore continued, as shown by this loop of ECMWF forecasts all valid at 12Z on November 23 (opens in a new window). The first frame in the animation is the "oldest" forecast (from 12Z on November 15), and the last frame is the most recent forecast (a 24-hour forecast from the 12Z run on November 22). Changes in the model's depiction of the 500-mb trough (from very deep and negatively tilted in the 192-hour forecast (opens in a new window) to less deep and slightly positively tilted in the 24-hour forecast (opens in a new window)) led to a weaker storm much farther east, having minimal impacts in the Eastern U.S.

The big jump in model solutions between the 12Z run on November 15 and the 12Z run the next day was a huge red flag that model solutions were not very reliable, but the hype train had already left the station. You'll sometimes hear forecasters say, "the trend is your friend" with model output, and that's certainly true here.By November 20, the story of a messy pre-Thanksgiving travel period had changed dramatically (see the headline from The New York Times (opens in a new window)). Indeed, weaker areas of low pressure did form far offshore, leaving merely breezy, chilly conditions in the East with some lake-effect snow in the classic snow belts of the Great Lakes (check out the 12Z surface analysis from November 23 (opens in a new window)).

Lagged Average Forecasts

The adage that "the trend is your friend" is based on the idea that model jumpiness tends to settle down as the event approaches. This idea is also the basis for a technique that uses a Lagged Average Forecast (LAF), which is basically the simplest approach to ensemble forecasting because in the most general sense, any set of forecasts all valid at the same time is considered an ensemble forecast.

In general terms, a lagged average forecast is simply the mean forecast of a series of successive model runs, all valid at the same time. So, for example, the mean of the 24-hour forecast from today's 12Z run, the 36-hour forecast from last evening's 00Z run, and the 48-hour forecast from yesterday's 12Z run shows the "lagged average forecast" for any field since all of those forecast progs (from different model runs) are valid at the same time.

The image below shows the lagged average forecast for MSLP valid at 00Z on December 31, 2000. The LAF was computed using the 60-hour forecast (generated by the 12Z run on December 28), the 48-hour forecast (from the 00Z run on December 29), the 36-hour forecast (generated by the 12Z run on December 29), and the 24-hour forecast (from the 00Z run on December 30). In case you're wondering, the shaded areas on the LAF panel represent the spread ("dispersion") of the member forecasts from the average.

A series of weather forecast maps with isobars over the eastern U.S., concluding with a composite map.

Successive GFS forecasts for MSLP from the 12Z run on December 28, 2000 (60 hr forecast), 00Z on December 29 (48 hr forecast), 12Z on December 29 (36 hr forecast), and 00Z on December 30 (24 hr forecast). All forecasts are valid at 00Z on December 31. The Lagged Average Forecast appears on the bottom. The shaded areas represent the dispersion of the members from the average.

Each of the top maps is labeled as a forecast for 24, 36, 48, and 60 hours, respectively, and shows contour lines of pressure values (in millibars) with numbers like 1008, 1004, 1000, 996, and 992. The location of the contours shifts slightly in each forecast as a low-pressure system approaches from the west over several days. Below these maps is a larger composite map labeled "Derived Lagged Average Forecast (LAF)" which combines elements from the forecast maps above. This main map provides an averaged view of the forecast with pronounced shading to indicate the intensity of the pressure system. Additionally, it highlights a concentration of isobars around the 992 millibar center. Green arrows connect the top maps to this bottom map. There is additional red text "Derived Lagged Average Forecast (LAF)" to the right of the bottom map.

Credit: NOAA

To determine the size of the dispersion, we compute the average of the forecasts and then gauge how much each member varies from the mean. It's tantamount to assuming that the average is the "verification" and then taking the difference between each member and the mean and statistically coming up with a value to measure the impact of all members on the LAF. It should come as no surprise that there's less confidence in a lagged average forecast when the dispersion is relatively large.

Now, let's leave the realm of statistics and talk some meteorology. Note that the 60-hour forecast had an intense, sub 988-mb low off the Delmarva Peninsula. In subsequent runs (after 12Z on December 28), the model trend weakened the low and moved it north-northeastward. As a forecaster on Long Island, going with the lagged average forecast was a prudent way to go, given the low's observed position at 00Z on December 31 (opens in a new window). Clearly, the lagged average forecast did a pretty good job.

On a daily basis, however, most forecasters don't have access to forecast products that formally compute lagged average forecasts and the corresponding spread from the means. Rather, we look for trends in the models, making a qualitative lagged average forecast. A seasoned forecaster looking at the model trends in this case may have predicted the low to be slightly north and east of the 24-hour forecast position. Remember: The trend is your friend.

Key Data Resources

Both Tropical Tidbits (opens in a new window) and Pivotal Weather (opens in a new window) give you the ability to create forecast trend animations, which is a great way to qualitatively assess trends and make a lagged average forecast.

Ultimately, when applying the LAF approach, you're using looking for model trends and using the idea that "the trend is your friend." But, there are a few caveats you should always keep in mind when you decide to embrace an LAF approach:

Tips whenapplying the LAF approach

  • Use three (or more) successive model runs instead of two runs (a trend from two successive runs might simply turn out to be model jumpiness).
  • If you can wait, the trends that come into greater focus from runs 24 to 12 hours in advance of an event are the most valuable for making a sound lagged average forecast (do not "hang your hat" on any trends that appear in runs more than a few days in advance of the event).
  • Before you use the LAF technique, make sure the model has a decent handle on the current weather pattern (make sure that the model's initialization doesn't differ greatly from reality).
  • Do not use the LAF technique in a situation that highly depends on parameterizations (for example, a trend characterized by deep, moist convection).
  • The LAF technique should be generally reserved for dynamic models, not statistically post-processed guidance like MOS or the National Blend of Models (which we'll cover more in depth later). There are just too many other statistical (non-meteorological) factors in these forecasts that can sometimes mask meteorological trends (or create trends for non-meteorological reasons).

While lagged average forecasts qualify as a crude form of ensemble forecasting, they do not account for the uncertainty associated with the initialization of a model or models. The technique of true ensemble forecasts deals with that. Let's explore!

sas405

Ensemble Forecasting: Part I

Ensemble Forecasting: Part I

Prioritize...

When you've completed this page, you should be able to discuss the differences between a "true" ensemble forecast and a lagged average forecast, as well as how true ensemble forecast members are generated.

Read...

As you just learned, some model jumpiness falls into the category of model trends, which as a lagged average forecast can serve as a basis for making a sound forecast. But, some trends, especially those that might seem promising more than a few days in advance, can lead you astray. Even trends established at 24 and 12 hours are occasionally fallible. So, really, the trend is usually your friend, but not always.

For example, check out this loop of three successive model runs of the GFS (opens in a new window), all valid at 00Z on September 4, 2019. The first frame is the "oldest" forecast (126 hours), while the last frame is the most recent (114 hours). At roughly a five-day forecast lead time, the model trend for the position of Hurricane Dorian was toward the west, with the storm located near the west coast of Florida in the most recent forecast run.

A forecaster following the model trends at this time and remembering that "the trend is your friend" could have easily concluded that this westward trend had merit. But, the trend wasn't a great friend in this case. In fact, Hurricane Dorian never even made landfall in Florida (Dorian's track (opens in a new window)). It approached from the east, stalled off the coast (over Grand Bahama Island), and then turned northward. The westward trend shown around five days in advance was misleading (as can often be the case with trends several days in advance).

In principle, a reliable lagged average forecast converges to the correct forecast with each successive run that gets closer to the event (the time the forecast is valid). In the schematic below, which serves as a conceptual model for an ideal lagged average forecast, successive runs of a computer model converge on the correct solution. To get your bearings, please note that an "X" marks each model initialization of a particular forecast parameter (such as wind speeds, precipitation, etc.), while the dark lines depict the evolution of that forecast parameter ending in a single deterministic forecast for each model run (large black dot). The green bar at the valid forecast time is the climatological spread of the forecast variable (the long-term average range in values). The orange bar represents the set of appropriate probabilistic forecasts based on model uncertainty (from imperfect physics, parameterizations, etc.).

Schematic showing how an ideal lagged average forecast converges to a solution.

The ideal lagged average forecast converges to a solution with each successive run leading up to the time the forecast is valid.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

But, the model trend, as you just saw in the Hurricane Dorian example, may not "hit the bullseye." In a worst-case scenario, the trend of a single model may, in fact, converge on a "solution" that's far removed from what eventually happens. So, what's a forecaster to do--look for trends in other models, too?

There's no doubt that when all the models converge on the same solution, forecasters can probably feel fairly confident in the forecast (like in this idealized schematic of precipitation forecasts (opens in a new window)from three different models). But, how much confidence would you have in the forecast if these three models showed the trends below? 

An example showing three different models trending toward different solutions for a hypothetical QPF prediction.

An example of three different models trending toward a different solution for a hypothetical precipitation forecast. What's a forecaster to do?
Credit: Adapted from NOAA

Yikes! Is it time to flip a coin? Well, some might take that approach and just pick a "model of the day," but we have better ways.

True ensembles to the rescue!

What's the better answer? True ensembles to the rescue! They give forecasters a way to identify and better deal with signals of uncertainty from computer guidance. True ensemble forecasts include several different solutions that attempt to get a handle on the inherent uncertainty in any given forecast. What does that mean exactly?  Perhaps a metaphor will be useful.

Two bulls-eye targets, one with a single off-target arrow, the other hit by many arrows.

Which would you rather have: One chance to hit the bullseye (left target), or several tries (right target)?
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

Imagine you're competing in an archery contest. Would you rather shoot one arrow at a target (left target) or increase your chances of hitting the bullseye by shooting a quiver full of arrows (right target)?

Think of choosing a "model of the day" (or even a lagged average forecast from a single computer model) as an archery contest that allows you only one shot at the target. In the same light, think of true ensembles as an archery contest that allows you to shoot a quiver full of arrows at the same target. The bottom line is that for a pending forecast fraught with uncertainty, utilizing true ensembles will give you a better chance of hitting at least something.

To get a more "meteorological" idea of what I mean, check out the schematic below. It shows the spread in the solutions of two different ensemble forecasts (one is a medium-range forecast; the other is a shorter-range forecast). In turn, this spread defines the model uncertainty for the given forecast (note that the uncertainty decreases as forecast time decreases). In most cases, the bullseye lies somewhere within that spread. Although we can never really hope to hit the bullseye dead center, we can at least use ensemble forecasts to make informed probabilistic forecasts that better convey the uncertainty of the situation. In archery terms, we have a better chance of hitting at least something.

Schematic showing how true ensembles create a spread of solutions, which usually contains the bull's eye.

By looking at true ensemble forecasts, you increase your chances of "hitting at least part of the target." To get your bearings, the green bar represents the climatological spread of the forecast parameter and the lines drawn to the right of the initialization values are the various forecasts for the parameter by the ensemble members. The orange bar represents the model uncertainty for a 48-hour forecast.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

So, ensembles help us to identify the signals of uncertainty associated with the forecast (even lagged average forecasts help to convey these uncertainties). Since you'll deal with some level of uncertainty daily when making forecasts, it behooves us to dig a bit deeper into the theory of ensemble forecasting.

Tweaking initializations: Generating a quiver of arrows

What allows us to shoot more than one "arrow" at the target? As it turns out, the initial conditions fed into the model lie at the heart of the problem of model uncertainties (imperfect physics and parameterizations don't help either). That's because model initializations rely heavily on the previous run's forecast. Granted, these initializing forecasts are tempered by real observations, but in the grand scheme of things, they fall short of the mark because observing stations, particularly upper-air sites, are simply too far apart over land and too infrequent. To make matters worse, in-situ upper-air observations are virtually non-existent over the vast expanse of oceans (this deficiency obviously has more impact on global models).

Needless to say, initializations are fraught with error. For example, model initializations can be quite suspect with regard to the intensity and evolution of a compact but vigorous short-wave 500-mb trough over the Pacific Ocean, and these can inject huge uncertainties into a pending forecast period. The National Centers for Environmental Prediction have taken some innovative approaches to help remedy the problem presented by large data voids, such as using targeted aircraft reconnaissance to try to observe important features over the Pacific. If you're interested, you read more about such adaptive observations (opens in a new window). However, regardless of these targeted efforts, initializations will never be perfect.

In some cases, initializations more accurately capture the essence of the current pattern, like a slowly-evolving blocking pattern (opens in a new window). In general, forecasters have greater confidence making predictions in such patterns. Still, even such highly meridional patterns, there are always relatively large uncertainties in the forecast in some regions of the country. And, not surprisingly, those uncertainties grow with time. As an example, check out this loop of GFS ensemble forecasts of the 528- and 570-dekameter 500-mb height contours (opens in a new window). By the way, this is a Medium-Range Ensemble Forecast (MREF). Each colored contour corresponds to an ensemble member with slightly different initial conditions. You can see little spread in the forecasts early on (great confidence), but the spread then increases with time (growing uncertainty). You can see why these types of forecasts are appropriately called spaghetti plots.

How do we decide to tweak the initial conditions to get an ensemble forecast like this, and how do we decide whether these "tweaks" will have any real impact on the forecast? We'll see in the next section! Read on.

sas405

Ensemble Forecasting: Part II

Ensemble Forecasting: Part II

Prioritize...

Upon completion of this page, you should be able to discuss the basic structure of ensemble systems and be able to competently interpret the ensemble mean forecast. You should also be able to interpret the spread of ensemble members in order to determine whether the ensemble mean might be a reasonable (useful) forecast.

Read...

In recent decades, numerous ensemble forecast systems have been developed. But, before we get into the details of commonly-used ensemble systems, let's start with a short discussion about how ensemble members are generated. For starters, a deterministic model (such as the GFS or ECMWF) is run at the highest model resolution possible, with the best estimate of the initial conditions in the atmosphere and the best-performing model physics (on average).

But, even with the highest resolution possible, the best estimate of initial conditions, and the best-performing model physics (on average), the forecasts from deterministic models are clearly imperfect (in part because even the best estimate of initial conditions and the best-performing model physics are both still imperfect). Running ensembles with many members (a couple dozen or more, typically) allows forecasters to explore "what if" scenarios based on small variations in the initial conditions and / or variations in the model physics package. But, to run that many model simulations with available computing resources, ensemble members are typically run at a lower resolution than the deterministic model. This lower resolution sacrifices some detail in the forecasts, but the ensemble members still provide a useful estimate of forecast uncertainty.

How are ensemble members generated?

Think of an ensemble forecast as an experiment (like this one during which David Babb blew bubbles over a hot plate (opens in a new window) to simulate convection). In this experiment, Lee Grenci performed a "control experiment" off camera, by blowing bubbles relatively far away from the hot plate and observing that they sank to the floor, so that we could scientifically compare the results of the real experiment to a "control run." The "control run" that you did not see was "unperturbed" (there was no hot plate to disturb the natural chain of events). And so it is with the ensemble control forecasts, which provide a reference "base line" so that we can gauge how far each model's subsequently "perturbed" forecasts (ensemble members) deviate from the "unperturbed" control run. The control run is much like the deterministic version of a model, but is run at a lower resolution like all the other ensemble members.

But, now we have to go about the process of generating individual ensemble members. Our initial goal is to find the most rapidly growing modes by perturbing the initial state of the control run. You can think of "modes" this way: The perturbations of the initial state of a control run are essentially "potential error patterns" that evolve with increasing forecast time. These "potential error patterns" are synonymous with "modes." Depending on the current organization of the atmosphere, certain modes (perturbations / error patterns) will grow much faster than others. Logically, the fastest growing modes are the ones that will primarily govern the overall uncertainty of the forecast. The search for these fastest growing modes is the initial step in the process of generating an ensemble forecast. Once found, these fastest growing modes then become "optimal perturbations" that breed the various ensemble members.

As an example, the initial perturbations for an ensemble run might be the differences between the latest control analysis and the 12-hour forecast from the control run initialized 12 hours earlier. For a visual, consider the idealized 500-mb pattern corresponding to the initial state of a control run below. The perturbations are used to determine the regions where rapidly growing error patterns occur. In more metaphorical terms, you can think of perturbing the initial state of the control run as "tickling" the virtual atmosphere to see if you get any reaction. When you tickle in sensitive places and the virtual atmosphere reacts heartily, you know you have your finger on a fast growing mode.

An idealized 500-mb pattern corresponding to the initial state of a control run.

An idealized 500-mb pattern corresponding to the initial state of a control run. Any rapidly growing error patterns in the shaded regions would create great uncertainty in the forecast.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

I hope we all agree that rapidly growing error patterns in heights to the east of a 500-mb trough would create great uncertainty in the forecast, so let's "tickle" these regions to see how fast potential error patterns grow. To test the sensitivity of the initial conditions to potential error patterns, suppose we tickle the virtual atmosphere by introducing a perturbation of +10 meters and +20 meters within the purple zones above (consistent with the magnitude of errors that routinely occur when initializing 500-mb height fields). Such perturbations would then represent the initial conditions of two ensemble members, "p1" and "p2". In similar fashion, we introduce perturbations of -10 meters and -20 meters ("n1" and "n2") that would serve as initial conditions for another two members. From this process, we would have five ensemble members - the control run, p1, p2, n1 and n2. I emphasize that this is an oversimplification of how the actual ensemble cycle works, but it gives you a basic idea of how ensemble members are generated.

In addition to tweaking initial conditions, another way that modelers add variety to ensemble systems is through including versions of the same model with different parameterization schemes. Recall that no parameterization scheme is perfect (parameterizations are oversimplifications of highly-complex physical processes), so it makes sense that ensemble systems can include versions of models with different schemes. For example, check out the comparison of two models below, one of which was run with the Betts-Miller-Janjic (BMJ) convective parameterization and physics packages, and the same model run with the Kain-Fritsch (KF) convective parameterization and physics packages. The details of these parameterizations aren't important for our purposes, but the key message is that different parameterization schemes can help identify uncertainties in the forecast (note that the forecasts for precipitation differed markedly in some areas).

The 36-hour ETA forecasts for 24-hour total precipitation ending at 12Z on March 17, 2000, using the Betts-Miller-Janjic (BMJ) convective parameterization (left) and the Kain-Fritsch convective parameterization (right).

The 36-hour ETA forecasts for 24-hour total precipitation ending at 12Z on March 17, 2000 (the model was initialized at 00Z on March 16) using the Betts-Miller-Janjic (BMJ) convective parameterization (left) and the Kain-Fritsch convective parameterization (right). Details of these parameterizations are beyond the scope of the certificate program. Although there are agreements, there are dramatic differences, supporting the notion that such schemes are fraught with error.
Credit: NCEP

The Bottom Line

So, what does all this mean? Ultimately, ensembles give you (typically) dozens of forecasts, each of which represents an arrow in your quiver (if you recall the metaphor we used previously), which provides many chances to hit the bullseye. But, how do we gather relevant messages from all those forecasts? For starters, we look at the ensemble mean (the average of the ensemble forecasts) and the spread. For example, take a look at the image below, showing the five-day track forecasts from the ECMWF ensemble for Tropical Storm Henri, initialized at 06Z on August 20, 2021. Each line represents the forecast of an ensemble member, and the thick black line represents the ensemble mean forecast.

ECMWF ensemble forecasts for Hurricane Henri initialized at 06Z on August 20, 2021.

Five-day track forecasts from the ECMWF ensemble for Tropical Storm Henri, initialized at 06Z on August 20, 2021.
Credit: weathernerds.org

In certain situations, the ensemble mean has a higher statistical chance of striking the target. In the example above, for instance, the individual members are spread fairly evenly around the mean. Even though there's substantial spread (predictions ranging from landfall in southern New Jersey to no landfall at all), the fairly even spread around the mean and large cluster of members near the mean suggests that the mean might be a reasonable forecast in this case. And, indeed, it was! Henri's center tracked just east of Long Island before curving westward as it made landfall (Henri's Track (opens in a new window)). Was the ensemble mean perfect? Not quite. But, it was still a very good track forecast in this case!

But, please don't think that the mean ensemble forecast is infallible. To the contrary, it can fail miserably at times. Think about a situation in which ensemble members cluster themselves into two distinct "camps." For example, check out the track forecasts for Hurricane Joaquin from the GFS ensemble (opens in a new window), initialized at 00Z on October 1, 2015. Each oval highlights a cluster of ensemble members. One cluster of members turned the storm westward toward the U.S., suggesting a landfall in the Carolinas or Mid-Atlantic. Another cluster of members suggested Joaquin would turn northeastward and go out to sea.

While the ensemble mean isn't drawn explicitly on the graphic, it's not hard to imagine that it would be somewhere in between those two camps, with very few members near it! In fact, the official NHC forecast cone (opens in a new window) issued later on the morning of October 1 was centered on a path right in between those two camps (near the ensemble mean), headed for southern New England -- a solution which virtually no GFS ensemble members predicted. As it turned out, the "out to sea" solutions were on the right track, and Joaquin never came anywhere close to the U.S. East Coast (Joaquin's Track (opens in a new window)). In this case, the mean ensemble forecast was an unlikely scenario (it was unlikely to hit the target).

Lesson learned: Always consider the mean ensemble forecast, but before you accept it, have a sound meteorological reason for doing so, and make sure to assess the overall qualitative spread of the individual ensemble members. That assessment will give you a better feel for the uncertainty of the forecast and give you insight about how useful the ensemble mean might be.

For some forecast variables, like precipitation, over the long haul the ensemble mean does very well, but it may not do very well tomorrow. So, you better have some sound meteorological reasoning before you accept or reject the mean ensemble forecast in any given situation. For other forecast variables (like surface temperatures or winds) ensembles can help us gauge uncertainty, but we have even better forecast products for those variables (which involve statistical post processing -- we'll cover these in the next lesson). Still, when assessing the big picture, I add that, over the long haul, the ensemble mean forecast has a much better track record than choosing the "model of the day" and going with its solution.

Now that you know the story on how ensembles work and what their advantages are, let's turn the page and explore some commonly used ensemble systems and data resources.

sas405

Ensemble Systems

Ensemble Systems

Prioritize...

Upon completion of this page, you should be able to identify and access commonly used ensemble systems, and describe their general characteristics (global versus regional domain, convection-allowing, etc.).

Read...

Now that we've covered the basic utility of ensemble forecasts and how they're created, I want to quickly summarize the basics of ensemble systems most commonly used by forecasters so that you can start accessing and using ensemble forecasts. Note that some ensemble systems commonly go by multiple names or acronyms, so there's a bit of "alphabet soup" in this section, but I want you to be familiar with the different acronyms since you may see these variations on various model websites and apps.

Global Ensembles

The major modeling centers around the world have been running global ensembles since the 1990s, though they have become considerably more robust since then. The National Centers for Environmental Prediction (NCEP) in the U.S., for example, began running a "Medium-Range Ensemble Forecast" (MREF) system in 1992, which had just three members -- a control run and two perturbations, producing forecasts out to 12 days. That's not much of an ensemble by today's standards! Technology and computing power have greatly improved over the years, allowing for ensembles with many more members (usually a few dozen or more). Modeling centers are also increasingly adding ensembles based on AI versions of global models into the mix. When using various websites to access model data, you'll likely encounter these major ensemble systems (among others):

Global Ensemble Forecasting System (GEFS): The National Centers for Environmental Prediction's (NCEP's) flagship global ensemble, based on the GFS model. It's most often just referred to as the GEFS, but you may also hear it referred to as "the GFS ensemble." It has grown to have more than 30 members and is run more than two weeks into the future. The AI-GFS model also has an ensemble (AI-GEFS), and NCEP also runs a "hybrid" GFS-based ensemble consisting of both the GEFS and AI-GEFS members. This "H-GEFS" ensemble has more than 60 members, and verification statistics indicate that its forecasts are more accurate, on average, than either the traditional, "physics-based" GEFS or the AI-GEFS alone.

AI-GEFS and H-GEFS Forecasts

Forecasts for the probabilities of 2-m temperatures greater than 70º from the AI--GEFS (slide 1) and H-GEFS (slide 2). Slide pagination controls located after next slide button.

Now showing slide 1 of 2

  •  

    AI-GEFS forecast for probabilities of 2-m Temps greater than 70ºF
    The AI-GEFS 90-hour forecast, valid at 18Z on November 4, 2026, for probabilities of 2-m temperatures greater than 70ºF showed rather high probabilities as far north as Pennsylvania and parts of New England, suggesting the likelihood of unusual warmth for early November. Advance the image slider to see the corresponding H-GEFS forecast for comparison.
    Credit: NOAA-GSL

     

Global Ensemble Prediction System (GEPS): This ensemble system is run by the Canadian Meteorological Centre (CMC), based on their global model, the GDPS -- Global Deterministic Prediction System (also known as "the CMC"). You may also hear the GEPS referred to as the CMCE (CMC Ensemble) or even "the Canadian ensemble." The GEPS has more than 20 members and is run more than two weeks into the future.

North American Ensemble Forecast System (NAEFS): A joint venture between the national meteorological services of the United States, Canada, and Mexico, this system combines the GEFS and the GEPS to create an ensemble with more than 50 members. Much like with the hybrid approach of the H-GEFS, verification statistics have shown that the NAEFS forecasts are of higher quality (on average) than the guidance from either the GEFS or GEPS alone. In addition, the ensemble mean of the NAEFS, is, on average, superior to any single deterministic model solution for forecast lead times greater than three days. It's important to note the "on average" caveat in these previous two statements. While the NAEFS provides superior results most of the time, it may not do so on any single day.

ECMWF Ensemble Prediction System (ECMWF EPS): Sometimes called "the Euro ensemble," this system is run by the European Centre for Medium-Range Weather Forecasts (ECMWF). The ECMWF EPS can trace its roots back to 1992, just like the GEFS, but has grown to include more than 50 members, so it's a powerful ensemble system. The AI-based version of the ECMWF model also has its own ensemble, the EPS-AIFS (Ensemble Prediction System-AI Integrated Forecast System), or sometimes just referred to as the "ECMWF AI Ensemble" or AIFS-Ensemble." Verification statistics indicate slightly higher forecast accuracy (on average) from the AI ensemble than the traditional physics-based ensemble at longer forecast lead times.

Regional Ensembles

While global models and ensemble systems can obviously provide forecast guidance for anywhere on the planet, some ensemble systems only cover limited areas. In the U.S., the flagship ensemble system is the Rapid-Refresh Ensemble Forecast System (REFS). This ensemble is based on the Rapid-Refresh Forecast System convection-allowing model, and it is run at a high-enough resolution that the ensemble members can simulate convective structures and processes (convection does not have to be parameterized). You may recall learning about such "convection-allowing models" in your previous studies.

REFS forecast showing high probabilities of at least 1" of rain in the next 6 hours over parts of Florida.
The 6-hour REFS forecast valid at 06Z on November 2, 2026 shows high probabilities (in excess of 80 percent in some areas) of at least 1 inch of rain in the 6-hour period ending at 06Z.
Credit: Storm Prediction Center

Such convection-allowing models need to have a grid spacing of no more than four or five kilometers in order to realistically simulate convection, and running such high-resolution models as part of an ensemble requires a lot of computing power. So, the tradeoff is that the REFS has fewer members than most other ensemble systems, and the forecasts are only run a few days into the future. The REFS can be particularly useful when mesoscale factors will be important players in the forecast (convective precipitation situations, small-scale heavy snow bands, etc.).

Other ensembles do exist, but the ones outlined on this page are the most commonly available on public (free) websites. Now that you have a little background on various ensembles, where can you access their forecast products? I encourage you to take some time to explore the sites below. They all offer unique ensemble-based products:

Key Data Resources

Combined, these data sources provide a wealth of ensemble forecast products that you can use to better understand and express the uncertainty in any given forecast, which is an important part of thinking probabilistically about future weather. If you want to dig deeper to learn more about the ensemble systems presented on this page (I kept the discussion pretty general), you might enjoy the links below in the Explore Further section. Otherwise, in order to get the most out of ensembles, we have to discuss some commonly-used forecast products, and in order to do so, we need to run through a quick "boot camp" in statistics so that you can competently interpret ensemble forecast products.

Explore Further...

If you're interested in learning more about the "guts" of various ensemble systems, you might enjoy the following links. Some of the resources here are on the technical side, but may be of interest to folks who are into modeling or the evolution of these systems:

sas405

Statistics Bootcamp: Part I

Statistics Bootcamp: Part I

Prioritize...

When you've finished this page, you should be able to interpret and discuss statistical metrics useful for interpreting ensemble forecasts, such a the mean, range (spread), standard deviation, and standardized (normalized) anomaly.

Read...

In order to get the most out of ensemble guidance, you need to know some statistics. Some of the statistical measures we'll cover on this page you're already familiar with from your previous studies in the course (or earlier courses in the program). Others you may have studied in a high-school or college science or math course, but even if they're new to you, keep in mind that you won't be required to actually calculate these statistics on your own. Instead, I'm going to emphasize their practical meaning and how you can apply them to interpreting ensemble forecasts. So, don't worry too much about the specific calculations. Focus on how to use these metrics. Let's get started!

Mean and Range (Spread)

You're already familiar with the ensemble mean forecast. The mean is simply the average of all ensemble member forecasts. As you recently learned, the ensemble mean for some variables (like precipitation or the track of a tropical cyclone) tends to perform better than any single deterministic model over the long haul (although that might not be true every single day).

In order to determine whether the ensemble mean is a reasonable forecast on any given day, forecasters need to do a thorough analysis of the big picture weather situation, as well as assess the size and nature of the spread (or range) of solutions. The spread of solutions is just the size of the difference between all members. So, for example, in this ensemble forecast for the track of Tropical Storm Henri (opens in a new window), the overall spread as the storm was predicted to move northward toward southern New England increased in time and was defined by the western most track and the eastern most track (all forecasts fall within that range). Or, in a precipitation case, if the wettest ensemble member predicted one inch of rain at a city and the driest member predicted 0.25 inches, the spread would be from 0.25 inches to one inch. All of the ensemble members fall somewhere within the spread of solutions, and forecasts with more uncertainty tend to have larger ensemble spreads.

Standard Deviation

One common way to measure the size of the spread is with the standard deviation, which is commonly represented with the Greek letter sigma (σ ). If you're interested, check out this video from Khan Academy (12:33) about how to calculate the standard deviation (opens in a new window) (along with a brief discussion of how it connects to the mean and range). In a nutshell, the standard deviation of a set of ensemble forecasts gives us an idea of how far individual member forecasts are from the mean forecast, on average. Therefore, a larger standard deviation (a larger difference between member forecasts and the mean, on average), would represent a larger spread and greater forecast uncertainty typically. On the flip side, a smaller standard deviation signals a smaller spread of forecasts, and less uncertainty typically.

If you're looking at short-term ensemble forecasts for temperature, it's pretty common to see a standard deviation of 3 degrees Fahrenheit or less (that would indicate a "typical" amount of uncertainty). Larger standard deviations for a short-term forecast would indicate that the spread (and therefore the uncertainty) would be greater than typical. Of course, the standard deviations of ensemble forecasts tend to grow as forecast lead time increases (uncertainty grows in time).

Another way that standard deviations come into play in forecasting is through the use of standardized (or "normalized") anomalies, typically based off a deterministic model run or an ensemble mean, like in the image below.

GEFS mean forecast of 500-mb heights and normalized anomalies, valid at 18Z on December 14, 2021.

The GEFS mean forecast for 500-mb heights and normalized (standardized) anomalies, valid at 12Z on December 14, 2021 (from the 12Z run on December 6).
Credit: Tropical Tidbits

The above forecast graphic plots two things -- the GEFS mean forecast for 500-mb heights (black contours) and "normalized anomalies" (shaded) valid at 12Z on December 14, 2021 (from the 12Z run on December 6). The GEFS mean clearly predicted a large ridge in 500-mb heights in the eastern U.S. at this time. But, what are "normalized anomalies?" Well, from your previous studies, you should know that an anomaly is just the difference between an observation (or in this case a forecast value) and the long-term climatology. But, what does "normalized" mean?

If we looked at a corresponding forecast of GEFS mean 500-mb heights and anomalies (not normalized anomalies), it looks a little different. Note that the anomalies are expressed in dekameters (since they represent the difference between the long-term average 500-mb heights and the ensemble mean forecast). The large ridge in the eastern U.S. is apparent from the GEFS mean forecast, and the largest positive anomalies (nearly 30 dekameters) are located across parts of Michigan. So, the largest predicted positive departure from climatology is over Michigan. Is that where the predicted 500-mb ridge is supposed to be most climatologically significant? Not exactly.

GEFS mean forecast of 500-mb heights and anomalies, valid at 18Z on December 14, 2021.

The GEFS mean forecast for 500-mb heights and raw anomalies, valid at 12Z on December 14, 2021 (from the 12Z run on December 6).
Credit: Tropical Tidbits

As it turns out, the typical range of 500-mb heights at a given time of year differs from place to place. As you may recall, for example, over the tropics, 500-mb heights (and many other variables) tend to have much less variation in space and time. Meanwhile, 500-mb heights over higher latitudes vary much more considerably at any given time of year. So, 500-mb heights that are 30-dekameters higher than the long-term average might be very significant in one part of the world, but less so in another.

In order to make "apples to apples" comparisons, the anomalies are often "standardized" (or "normalized"), by dividing the anomalies by the climatological standard deviation. The result of this calculation is a number of standard deviations from the long-term average in a given region, which allows forecasters to see where the anomalies are most climatologically significant. By normalizing the anomalies, we can see that the most climatologically significant anomalies were predicted to be over Missouri and surrounding states, not over Michigan (where the largest raw differences between the mean forecast and climatology were). Check out the side-by-side comparison (opens in a new window) of these GEFS forecasts showing raw vs. normalized anomalies.

How do we interpret normalized anomalies? I'll skip the statistical background, but for many weather variables (like 500-mb heights) most historical observations are clustered within one standard deviation of the climatological mean, and observations beyond one standard deviation from the mean become increasingly rare. Specifically:

  • Observations are within one standard deviation of the climatological mean about 68 percent of the time (they fall outside of one standard deviation about 32 percent of the time)
  • Observations are within two standard deviations of the climatological mean about 95 percent of the time (they only stray outside of two standard deviations from the mean about 5 percent of the time)
  • Observations are within three standard deviations of the climatological mean about 99.7 percent of the time (they only stray outside of three standard deviations from the mean about 0.3 percent of the time)

So, when you see normalized anomalies in ensemble forecasts that are two or three (or more) standard deviations from the long-term climatology, that's an extreme forecast. The largest normalized anomalies in the GEFS forecast above (2 - 2.5 standard deviations above normal), are pretty extreme. Where the normalized anomalies are largest here, the ensemble mean forecast for 500-mb heights is essentially in the top few percent of climatological values in mid-December. Since we know that 500-mb ridges tend to go hand in hand with warmer-than-normal air masses, this GEFS forecast sent a pretty strong message for unusual December warmth in the Midwest (and that general message verified quite well -- check out the actual temperature anomalies for the date (opens in a new window)).

One final point: One limitation of model-based anomalies and normalized anomalies is that the "climatology" used as their basis is a computer-model derived climatology, which can vary from model to model, and generally covers a few to several decades. So, if you see a rather extreme forecast, it's extreme in the context of the model's climatology, and may not always be extreme in the context of a location's total period of climatological records (which sometimes cover longer periods of time). Still, these tools can help forecasters identify signals for possible extreme weather.

By the way, if you want to dig a bit deeper into the statistics underpinning the guidelines for interpreting standardized (normalized) anomalies, check out the Explore Further section below. Otherwise, you may want to test your knowledge of a few key concepts on this page in the Quiz Yourself section below before moving on to Part II of Statistics Bootcamp in the next section.

Quiz Yourself...

Explore Further...

In this section, I provided some guidelines to help you interpret standardized (normalized) anomalies, which can be summarized by the empirical rule (or the 68-95-99.7 rule) (opens in a new window). The implication of this rule is that weather observations that are more than 2-3 standard deviations from the long-term average are quite rare, so when a forecaster sees normalized anomalies of this magnitude in ensemble mean forecasts, it's a sign that pretty extreme weather could be on the way. But, in order to make these interpretations, we had to make some assumptions.

For many atmospheric variables, if we plot a large sample of historical observations, the distribution of the observations can be reasonably approximated with a normal distribution, which is sometimes commonly called the "bell curve." The normal distribution is symmetrical, with the mean right in the middle of the distribution, and most of the values are clustered somewhat near the mean (the peak of the "bell").

A normal distribution plot with standard deviation bands labeled.

A normal distribution (a "bell curve") is symmetrical, an has the mean right in the middle of the distribution (marked by "0" along the bottom axis of this graph).

The graph above has the mean marked by "0" on the axis along the bottom, and each colored "band" marks one standard deviation from the mean. With a normal distribution, statistically, 68.2 percent of values lie within one standard deviation of the mean (34.1 percent are in each of the bands within one standard deviation on either side of the mean). Values more than one standard deviation from the mean become increasingly rare, with 27.2 percent of values located between one and two standard deviations from the mean (13.6 percent between one and two standard deviations above, and the same amount below). Therefore, 95.4 percent of values lie within two standard deviations of the mean (68.2 percent + 27.2 percent), and so on. Very few values are more than three standard deviations from the mean.

So, our guidelines for interpreting normalized anomalies of ensemble mean forecasts were based on the assumption that the climatology of the forecast variable during a given time of year roughly looks like a normal distribution. This assumption works well for a number of weather variables (like temperatures or heights at constant pressure levels), but doesn't work very well for variables like precipitation, which have a mean that isn't close to the middle of the distribution.

sas405

Statistics Bootcamp: Part II

Statistics Bootcamp: Part II

Prioritize...

When you've completed this page, you should be able to interpret ensemble probabilities (including the difference between historically-calibrated and uncalibrated probabilities) and percentiles, and define the median.

Read...

We'll round out our "Statistics Bootcamp" with a couple of statistics very commonly used in ensemble forecast products. You may be familiar with these statistics already, but folks sometimes confuse them when making interpretations, so as with the statistics we've covered already, make sure you focus on how to use and interpret them properly.

Probabilities (Percentages)

You're already familiar with probabilities, which are typically expressed as the percent chance that some event will (or won't) occur. You may be most familiar with probabilities in weather forecasting via the probability of precipitation, which is the percent chance that any point in a forecast area will receive measurable (at least 0.01 inches) of liquid-equivalent precipitation over a specified period of time. However, probabilities can be applied to many types of weather forecasts, such as a probabilistic forecast that temperature will exceed a certain threshold. Determining a true probability in a given forecast situation requires knowing how many instances of the event have occurred in a large sample of cases where atmospheric conditions were similar to what's being predicted.

In other words, a forecast with a true probability is historically calibrated. The historical "knowledge" of how often the event occurred when similar atmospheric conditions occurred helps generate the probabilities. A hallmark of a well-calibrated probabilistic forecast is that the percent chance accurately reflects how often the event actually occurs. So, for example, measurable precipitation will actually occur in 30 percent of the cases when the forecast indicates a 30 percent chance of measurable precipitation. What does this have to do with ensemble forecasts? Well, with many member forecasts available, ensembles are great tools for generating probabilistic forecast products, like the AI-GEFS forecast below showing the ensemble probability of temperatures exceeding 70 degrees Fahrenheit at 18Z on November 4.

AI-GEFS forecast for probabiliies of 2-m temperatures in excess of 70ºF.
The AI-GEFS forecast for probabilities of temperatures exceeding 70 degrees Fahrenheit showed high probabilities (80 percent or higher) as far north as Pennsylvania and New York at 18Z on November 4.
Credit: NOAA-GSL

The yellow-shaded areas forecast above indicate at least an 80 percent probability of temperatures exceeding 70 degrees Fahrenheit at 18Z on November 4 (which would be remarkably warm in areas of Pennsylvania and New York in early November). But, how was this probability calculated? Most ensemble probabilities simply represent the fraction of ensemble members that show a particular event happening (or threshold being met). So, if an ensemble has 10 members, and 8 of them predicted temperatures above 70 degrees Fahrenheit at this time, the ensemble probability would be 8 out of 10, or 80 percent.

In other words, this probability is uncalibrated. It doesn't tell us whether there's an 80 percent chance based on what's happened in similar weather patterns in the past. Because the ensemble probability is uncalibrated, a forecast of an 80 percent chance doesn't necessarily mean the event will actually occur in 80 percent of the cases. It just tells us where at least 80 percent of the ensemble members predicted temperatures above 70 degrees Fahrenheit.

Ultimately, ensemble probabilities are still a very useful way to view information (they tell you the percentage of ensemble members that predict a certain outcome), but they're not the same as historically-calibrated probabilities. Uncalibrated ensemble probabilities aren't based on historical information and assume that the actual solution will fall somewhere in the spread of ensemble forecasts, which isn't always true. In other words, if an ensemble forecast shows a 100 percent chance of at least one inch of rain falling in a given time period (all ensemble members predict at least one inch), but less than one inch actually falls, it's because the actual solution fell outside of the ensemble range. Because an ensemble spread doesn't truly represent all possibilities, this scenario can happen occasionally (although hopefully not often), causing the ensemble probability to misrepresent the true probability.

You may be able to find ensemble-based products that have undergone an extra layer of statistical analysis based on past events for a true, calibrated probability, but those products are usually labeled as "calibrated." Otherwise, you should assume that ensemble probabilities are uncalibrated and just represent the percentage of ensemble members that predict a particular outcome.

Percentiles

In the context of ensembles, a percentile is essentially a "marker" to help us identify where a certain value falls within the range of forecast solutions (on a scale from 0 to 100). For example, the 50th percentile represents the median--the middle value from the range of ensemble solutions. In other words, half of the ensemble members lie on one side of the median solution and half lie on the other side. A percentile does NOT directly represent a probability or percent chance! The 50th percentile solution does not necessarily have a 50 percent chance of occurring.

Since percentiles give us a quick way to mark where a value falls within the range, they provide a useful tool for interpreting ensemble forecasts. For example, the 90th percentile forecast is near the high end of the spread of ensemble forecasts (about 90 percent of the forecasts were lower). On the flip side, the 10th percentile solution marks a forecast higher than only about 10 percent of the solutions. In other words, the 10th percentile solution is toward the low end of the spread of ensemble forecasts. The 10th percentile solution, however, is not the same as saying that the forecast has a 10 percent chance of occurring, for example. The closer the percentile is to the 50th (the median), the farther it is from the extreme ends of the spread.

NAEFS mean forecast for 500-mb heights and climatological percentiles, valid at 12Z on December 14, 2021.

The 168-hour NAEFS mean forecast for 500-mb heights and climatological percentiles, valid at 12Z on December 14, 2021 showed that the mean forecast for 500-mb heights was above the climatological maximum (red shadings) in a swath from northern Mexico to Iowa.
Credit: Storm Prediction Center

Percentiles can also be used to compare ensemble forecasts to climatology so that forecasters can identify anomalous weather patterns (much like we saw with normalized anomalies). The image above shows the 168-hour NAEFS mean forecast for 500-mb heights (black contours) and climatological percentiles (shaded areas show percentiles less than 10 or greater than 90), valid at 12Z on December 14, 2021. How can we interpret this graphic? The large shaded area in the eastern U.S. represents a region where the ensemble mean forecast for 500-mb heights was near (or above) the high end of the range in the climatological period used for comparison (areas greater than the climatological 90th percentile are shaded). The large red shaded area from Mexico to Iowa represents an area where the ensemble mean forecast for 500-mb heights actually exceeded the maximum in the climatological period used for comparison!

This forecast map sends a very similar message as the forecast for normalized anomalies of GEFS mean 500-mb heights (opens in a new window) valid at the same time: The anomalously high 500-mb heights in much of the Midwest were predicted to be climatologically extreme. The two forecasts are slightly different since they came from different ensemble systems (GEFS and NAEFS), which were run at different times, but combined, they certainly sent a consistent message about the prospects for above normal warmth in the eastern U.S.

The bottom line is that viewing ensemble forecasts with percentiles allows forecasters to quickly assess where a certain forecast value lies within the overall range of ensemble solutions (or allows forecasters to see where the ensemble mean forecast falls within the range of climatology). If you're having trouble with the general concept of percentiles, this short video from Khan Academy (3:37) may help (opens in a new window). Otherwise, I recommend testing your basic knowledge of the concepts on this page in the Quiz Yourself section below before moving onto the next section, where we'll examine some common ensemble forecasting products (which will allow you to apply the concepts from Statistics Bootcamp).

Quiz Yourself...

sas405

Ensemble Forecasting Products: Part I

Ensemble Forecasting Products: Part I

Prioritize...

When you're finished with this page, you should be able to correctly interpret common ensemble forecasting products, namely spaghetti plots, maps showing the ensemble mean forecast, ensemble max forecast, percentages and percentiles.

Read...

Now that you have a little background in statistics, we're going to apply that knowledge to interpret some common ensemble forecasting products. The products described in this lesson don't constitute an exhaustive list, but they'll help you go beyond the ensemble mean to assess a range of possibilities and help you think probabilistically, even when you have to create a deterministic forecast. Forecast decisions that take into account for the likelihood of various scenarios will tend to perform better over the long haul, and can help prevent forecasters from wildly swinging for the fences (predicting an extreme, unlikely outcome without any meteorological backing). With that in mind, let's get to the products!

Spaghetti Plots

You've already encountered spaghetti plots a number of times throughout your studies (even as far back as METEO 101), but I want to offer a few reminders. Spaghetti plots of ensemble forecasts can help a forecaster quickly gauge the spread in ensemble forecasts. Hurricane track spaghetti plots (like this ensemble forecast for the track of Tropical Storm Henri (opens in a new window), initialized at 06Z on August 20, 2021) are just one example.

Spaghetti plots are also commonly used to gauge the forecast spread in other variables like 500-mb heights. Typically, the ensemble forecasts for one (or a few) of the contours are plotted. For example, check out the Northern Hemispheric spaghetti plot for 500-mb heights from the GEFS ensemble (opens in a new window), initialized at 00Z on December 28, 2021. The only contours that are plotted are the 5460 meter and 5700 meter contour in this particular case. It's easy to see the high-degree of agreement (small spread) in short-term forecasts, but the "spaghetti" becomes increasingly tangled (spread increases) such that discerning a clear message from the pattern over much of the Northern Hemisphere becomes increasingly difficult beyond about 240 hours (10 days) in this case.

Ensemble Mean and Maximum Forecast Maps

Perhaps the most common ensemble forecast map you'll encounter shows the ensemble mean forecast, which is often useful for evaluating the overall weather pattern. But, I want to reiterate a word of caution about interpreting ensemble mean forecasts. For example, check out the 240-hour GEFS mean 500-mb forecast on the left below, valid at 00Z on September 28, 2019. The ensemble mean 500-mb pattern over the United States looks fairly zonal, with some low-amplitude ridges and troughs. Now, look at the corresponding spaghetti plot on the right. The spaghetti is very tangled, meaning that there's a huge spread in ensemble forecasts! Many individual members have a much higher amplitude pattern (larger ridges and troughs), but because of the vast disagreement on location and magnitude, the signals get "washed out" when looking at the mean.

(Left) Mean GFS ensemble 500-mb forecast. (Right) Corresponding spaghetti plot. See the image caption for additional details.
(Left)The 240-hour GEFS mean forecast for 500-mb heights, valid at 00Z on September 28, 2019 shows a mostly zonal pattern over the United States. (Right) The corresponding GEFS 500-mb spaghetti plot (for the 5700-meter and 5880-meter contours) shows a much different story, with great uncertainty and a number of much higher amplitude possibilities.
Credit: National Centers for Environmental Prediction

Because the forecast uncertainty was large here (large ensemble spread) the "washed out" pattern of 500-mb heights depicted by the ensemble mean wasn't a particularly helpful forecast depiction. But, the only way for forecasters to make this assessment is by looking at the mean and the spread.

In addition to ensemble mean forecasts for 500-mb heights, forecasters frequently assess the ensemble mean forecasts for any number of other variables, like mean sea-level pressure or precipitation. For example, forecasters expected a series of low-pressure systems to bring a swath of snow from the Upper-Midwest into the Northeast on December 28, 2021 (check out the WPC surface forecast valid at 00Z on December 29 (opens in a new window)).

How much snow might fall in this area? Well, since ensemble mean forecasts tend to outperform single deterministic model forecasts, the ensemble mean might be a good place to start. The high-resolution ensemble mean forecast for 24-hour snowfall ending at 12Z on December 29 (opens in a new window) suggested a wide area of 1 to 4 inches of snow, with higher amounts in north-central Minnesota and along the northwestern shore of Lake Superior (in excess of 6 inches). But, the ensemble mean alone gives us an incomplete picture of the possibilities. We need information about the spread.

If forecasters wanted to assess a "high-end" scenario, they could examine the corresponding "ensemble max" snowfall forecast map (below), which shows the highest snowfall prediction in each area from all the ensemble members.

HREF ensemble max 24-hour snowfall forecast ending at 12Z on December 29, 2021.
The high-resolution ensemble max snowfall forecast for the 24 hours ending at 12Z on December 29, 2021.
Credit: Storm Prediction Center

The "snowiest" ensemble members predicted more areas of at least 6 inches, and a small band along the northwest shore of Lake Superior with more than a foot (small area of yellow shading). It's important to keep in mind that such ensemble max forecast graphics represent the high end of the ensemble spread (in other words, all other solutions had less snow in each area). Because no ensemble system truly represents all possibilities, it is possible that more snow than what was shown on the ensemble max forecast map could fall (although forecasters hope this doesn't happen very often). Still, forecasters who locked in on the ensemble mean forecast would have missed these higher-end possibilities.

Occasionally you may be able to find ensemble minimum forecasts, as well, to help visualize a "low-end" scenario, but even ensemble maximum, minimum, and mean forecasts don't give a complete picture of the spread of ensemble members, and they don't give us much indication about how many ensemble members predict certain scenarios to occur.

Ensemble Probabilities and Percentiles

Using forecast maps showing ensemble probabilities and percentiles can give us a more complete picture of the ensemble spread and can help forecasters think more probabilistically about the forecast. For example, in the snowfall scenario above, a helpful forecast question might be "how likely is it that at least X inches of snow will fall?" Ensemble probability maps can help with this question (keeping in mind the caveats that most ensemble probabilities are uncalibrated).

You may encounter ensemble probability products online from a number of ensemble systems. But, for precipitation (both liquid and snow), I want to draw your attention to some "superensemble" products available from NOAA's Weather Prediction Center (WPC). By "superensemble," mean an ensemble created from different models and ensemble systems (kind of "an ensemble of ensembles"). You can read more about the WPC probabilistic winter precipitation forecast products (opens in a new window), if you're interested. For the snowfall scenario above, check out the WPC "superensemble" probabilities of at least 1-inch of snow and 4-inches of snow below (click the arrow to toggle between the two images).

Image slider (click the arrow to change the image) showing the WPC "superensemble" probability of at least 1 inch and at least 4 inches of snow during the 24-hour period ending at 12Z on December 29, 2021.
Credit: Weather Prediction Center

When viewing the ensemble forecasts in this fashion, some useful messages jump out. First, the vast majority of ensemble members (at least 80 percent) predicted at least an inch of snow across a large area of the Upper Midwest, with another area of high ensemble probability across northern Pennsylvania and southern New York. Also of note, they suggested that at least 1 inch of snow across northern Ohio was far from a sure thing, with ensemble probabilities generally less than 50 percent except for in the far northwestern corner of the state. Recall that the high-resolution ensemble mean forecast was for a widespread 1 to 4 inches across northern Ohio, but the probabilities from the WPC superensemble should have given forecasters pause.

So, ensemble probability maps give forecasters a quick way to see the percentage of ensemble members that predict a certain threshold (like "at least 1-inch of rain," "at least 4-inches of snow," "temperatures of at least 90 degrees," etc.). That's a useful framing for thinking probabilistically about the forecast. Meanwhile, if forecasters want a quick way to assess "low end," "middle ground," and "high end" scenarios based on the ensemble guidance, maps of ensemble percentiles can help. For example, check out the 10th percentile, 50th percentile (median), and 90th percentile snow forecasts from the WPC "superensemble" covering the 24-hours ending at 12Z on December 29 (below).

 

Image slider (click the arrow to change the image) showing the WPC "superensemble" 10th percentile, 50th percentile (median), and 90th percentile snow accumulation forecasts for the 24-hour period ending at 12Z on December 29, 2021.
Credit: Weather Prediction Center

The 10th percentile snow accumulation forecast (a "low end" scenario for the storm), showed a maximum snowfall of 2 to 4 inches confined to northern Minnesota and southern Ontario. The median snow forecast (50th percentile) showed a more widespread snowfall of an inch or more, with 4 to 6 inches in areas of northern Minnesota (dark green). Finally, a "high end" scenario as represented by the 90th percentile solution showed a larger area of 4 to 6 inches, with a small area of 6 to 8 inches (blue) in northern Minnesota and southern Ontario.

Taken together, these three graphics give forecasters a better look at the overall ensemble spread; roughly 80 percent of the ensemble members fall within the range of these forecasts. But, none of the ensemble forecast products in this section (mean, max, probabilities, or percentiles) ever have all the right answers. To see how these forecast products verified in this case, check out the short video (4:07) below (you'll be able to see the images better if you view in "Full Screen" mode).

Here’s the national snowfall analysis for the 24-hour period ending at 12Z on December 29. We can use this analysis to assess the forecast messages of a number of ensemble forecasting products. As we’ll see no single ensemble forecasting product had all the right answers, which is pretty common. Even the best performing deterministic or ensemble model products perform better in some areas than others during any given storm.

The high-resolution ensemble mean had indicated a wide area of 1 to 4 inches of snow from the Upper Midwest extending down into northern Ohio and into Pennsylvania and New York, with higher amounts in north-central Minnesota and along the northwestern shore of Lake Superior – in excess of 6 inches in the green shading. But, it underestimated the magnitude and area of the heaviest snow in much of northern Minnesota, as well as the area of heavy snow in northern lower Michigan. Meanwhile, across northern Ohio, the ensemble mean was generally too high, as most of the area had less than 1 inch of snow.

The high-resolution ensemble max forecast had some merit, as it suggested some higher totals of 8-12 inches plus in parts of northern Minnesota, although it didn’t have the location or areal extent quite right. It also did pretty well with the area of 4 to 6 inches of snow in western Wisconsin. Forecasters who locked in on the ensemble mean forecast would have missed out on these higher-end possibilities. But, even the ensemble max didn’t capture the localized area of up to 8 inches of snow in Michigan, and it was way too high across northern Ohio – it had a swath of 4 to 8 inches of snow where largely less than an inch fell.

How about ensemble probabilities? Were their forecast messages useful? Well, certainly the low probabilities of getting at least an inch of snow across northern Ohio should have given forecasters pause, and indeed, much of the area did get less than an inch. Most of the area with a very high ensemble probability of getting at least an inch across Minnesota and Wisconsin – with ensemble probabilities of at least 80 percent, did get at least an inch of snow.

Turning to the ensemble probability of at least 4 inches of snow, much of the area with the greatest ensemble probabilities (greater than 50 percent) did get at least 4 inches of snow. But several areas where ensemble probabilities were quite low – between 0 and 10 percent, did get 4 inches or more. These areas of Wisconsin and Michigan really stick out.

Finally, onto ensemble percentiles. The 10th percentile solution was notably too low in many areas, which we might expect in many cases because most ensemble forecasts were higher. Although the idea of getting less than an inch was right across northern Ohio and into Pennsylvania.

The median, or 50th percentile solution was still too low across areas of the Upper Midwest that got the most snow, with its highest totals around 4-6 inches, but was closer to reality away from the areas of heaviest snow, like in northern Wisconsin, where the median solution was 1 to 2 inches.

Finally, the 90th percentile solution did a better job at pointing out the possibility of exceeding 6 inches in northern Minnesota, although it was still generally too low in that region – 6 inch totals were much more widespread. It also didn’t reflect the magnitude of the localized heavy snow in Michigan. But, on the other hand, it was too high across northern Ohio, where it had a widespread 1 to 4 inches and less than an inch fell in most areas.

So, each of these products may have contained a useful message in a particular area, but they each had flaws, some more significant than others. None of the products did a particularly good job of getting most mesoscale details correct, which is common. No single product, even from a powerful superensemble will have all the right answers. But, they do help forecasters quickly assess various possibilities. However, wise forecasters never look at model data – even ensemble data – in a vacuum, and they couple their probabilistic thinking about the forecast with an assessment of the big picture factors that will affect the forecast.

Credit: Penn State

Before moving on, I should mention that this was a snow forecasting case, which brings extra complications to the forecasting situation. Recall that most model snow maps are based off a "first guess" 10-to-1 snow-to-liquid ratio, which is sometimes a poor estimate in a given area for a particular storm (you may want to review caveats about using model snow maps from METEO 101 (opens in a new window) (login required)). The WPC probabilistic snow forecasting products do incorporate "dynamic" ratios that attempt to account for atmospheric conditions, but of course, the snow-to-liquid ratio estimates are still imperfect.

While using ensembles to think probabilistically about a forecast can help a forecaster make wise forecast decisions and limit error, these products shouldn't be used in a vacuum. It's still up to human forecasters to perform a solid analysis of the "big picture" -- the synoptic and mesoscale features affecting the forecast situation, in order to make forecast decisions (and in a snow forecasting case, that includes factors that can impact the snow-to-liquid ratio).

I've included some helpful links for accessing ensemble products (like spaghetti plots and WPC's probabilistic rain and winter precipitation products) in the Explore Further section below. Now it's time to dial in on ensemble products focused on individual locations. Read on.

Explore Further...

Key Data Resources

 

sas405

Ensemble Forecasting Products: Part II

Ensemble Forecasting Products: Part II

Prioritize...

After completing this section, you should be able to properly interpret and box-and-whisker plots and ensemble plume diagrams.

Read...

The ensemble forecast products we covered in the previous section were focused on maps that allowed us to view forecast information over large areas. But, what about if we want to dial in on ensemble forecast data at a specific town or city? From large-scale maps, we might have to do a lot of interpolating or estimating to glean values for a specific location. But, I want to introduce two common, very useful products for viewing ensemble forecasts at specific cities -- box-and-whisker plots and plume diagrams.

Box-and-Whisker Plots

Box-and-whisker plots (also sometimes just called "box plots") are convenient tools for summarizing the spread of a particular set of data. For example, check out the box-and-whisker plot below, which shows the range of 0 - 1 km storm-relative helicity observed for three storm types (tornadic supercells, non-tornadic supercells, and ordinary convection) in a sample collected as part of a research study.

Box-and-whisker plot showing the range of storm-relative helicity values by storm type in a research study.

Box-and-whisker diagram relating 0 - 1 km storm-relative helicity to ordinary thunderstorms, non-tornadic supercells, and tornadic supercells. Clearly, tornadic supercells tended to have greater 0 - 1 km storm-relative helicity, but each storm type had a considerable range.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

This plot tells us that tornadic supercells tend to form in environments with greater 0 - 1 km storm-relative helicity, but there's considerable overlap in the ranges observed for each type of storm. How do we interpret the boxes and whiskers specifically? Well, each box marks the "interquartile range," meaning that if we break the data set into four parts ("quartiles"), the interquartile range is just the middle two quartiles. Therefore, the top of the box marks the 75th percentile of the data set and the bottom of the box marks the 25 percentile. Since half the observations lie between the 75th and 25th percentiles, half of the observations lie within the box. The median (50th percentile) is marked by a red line in this case. So, for the tornadic supercell storms in this sample, half of them had 0 - 1 km storm-relative helicity values between roughly 30 m2s-2 (25th percentile) and 180 m2s-2 (75th percentile), and the median storm-relative helicity for tornadic supercells was 89.2 m2s-2.

The whiskers in this case represent the maximum and minimum observed values. So, the maximum observed 0 - 1 km storm-relative helicity in the sample of tornadic supercells was 240 m2s-2, while the minimum was just 7.7 m2s-2. That's quite a range, and its practical implication is that tornadoes can occasionally form with low values of 0 - 1 km storm-relative helicity.

However, using the maximum and minimum values for the ends of the whiskers is not a universal practice. Using the 90th percentile and 10th percentile or the 95th and 5th percentile is also pretty common as markers for the high end and low end of a range (while omitting the actual extremes). Some folks also choose to plot the mean of the sample either in addition to, or instead of, the median. But, the bottom line is that box-and-whisker plots summarize the range of data by focusing on five data points -- a marker for the high end (often the maximum, 95th, or 90th percentile), the 75th percentile, the median (or mean), the 25th percentile, and a marker for the low end (often the 10th percentile, 5th percentile, or minimum).

How does this translate into an ensemble forecasting product? Check out the box-and-whisker plot for 2-meter temperatures from the ECMWF EPS for State College, Pennsylvania, initialized at 12Z on January 19, 2022 (below). This run of the EPS was in the lead up to an Arctic outbreak in central Pennsylvania. Note that the coldest ensemble members were predicting temperatures on Saturday, January 22 to dip near -29 degrees Celsius (-20 degrees Fahrenheit), but there was a very large forecast spread as highlighted on this annotated version of the box-and-whisker forecast (opens in a new window) (the top of the range was around -13 degrees Celsius, or about 9 degrees Fahrenheit).

ECMWF EPS box-and-whisker meteogram showing forecasts for  2-m temperature in State College, Pennsylvania.

ECMWF EPS box-and-whisker meteogram showing forecasts for 2-meter temperatures at State College, Pennsylvania, from the 12Z run on January 19, 2022. Note that the plotting convention for the boxes and whiskers is a bit different than the storm-relative helicity box-and-whisker plot shown above.
Credit: ECMWF

You should notice that the plotting convention on this box-and-whisker meteogram is different than the storm-relative helicity plot shown above. The "large" box represents the interquartile range with the black horizontal line representing the median. The smaller boxes above and below the larger ones encompass the forecasts between the 10th percentile and 25th percentile (bottom box) and 75h percentile and 90th percentile (top box). The whiskers extend out to the extremes. This plot also contains the deterministic model forecast (blue line) as well as the ensemble control run (red dashed line).

These types of plots can help you get a quick feel for the size of the range of ensemble forecasts for a variable at a given location, as well as where the median (or sometimes, the mean) fits within the overall range. If you'd like to learn a bit more about interpreting box-and-whisker plots in general, this video from Kahn Academy (7:44) (opens in a new window) may help. Otherwise, you may want to practice your interpretations in the Quiz Yourself section below.

Plume Diagrams

For tracking individual ensemble member forecasts over time and seeing the details of the ensemble spread, plume diagrams are a great tool. For example, check out the GEFS plumes for total precipitation at Pittsburgh, Pennsylvania (KPIT) from the 12Z run on January 14, 2022, a few days before a major storm was to arrive in the Mid-Atlantic and Northeast.

GEFS plumes for total liquid precipitation at Pittsburgh, PA, from the 12Z run on January 14, 2022

GEFS plumes for total liquid precipitation (QPF) from the 12Z run on January 14, 2022. Each line represents an ensemble member forecast, while the thick black line represents the ensemble mean, and the blue line marks the operational GFS forecast. The ensemble spread was very large, due mostly to uncertainty around how much precipitation would fall late on January 16th into the 17th.
Credit: NCEP EMC

In a nutshell, plume diagrams display a time series of forecasts for a specific variable at a specific point. In the GEFS plume diagram above, time runs from left to right along the bottom (the model's initialization time is on the left here and the end of the forecast period -- 12Z on January 22 -- is on the right). Each line on the diagram represents a running total for each member's precipitation forecast, and in this case, the ensemble mean is marked by the thick black line. The blue line represents the operational GFS forecast. This diagram shows us that there was substantial disagreement in how much precipitation would fall late on January 16th into the 17th, with a fair number of members predicting nearly an inch of liquid precipitation or more. On the other hand, a few GEFS members predicted very little precipitation (0.1 inches or less) during this time frame. The operational GFS predicted a little more than 1.2 inches of liquid precipitation by 18Z on the 17th, which is more than all but four GEFS members.

For another visual perspective on this GEFS forecast, we could look at the corresponding plumes for 3-hour precipitation at Pittsburgh (opens in a new window). Note that this product isn't a running total, but is a time-series of 3-hour precipitation forecasts, which makes it easier to isolate times when precipitation might fall (although discerning totals over a long period of time is much harder than using the total QPF plumes shown above). The large spike in many member forecasts from about 18Z on January 16th through about 12Z on the 17th signaled the time frame for the most significant precipitation. Please note that the amount of 3-hour precipitation at a given time indicates the forecast for precipitation in the three hours ending at that time (so, a member's forecast at 12Z covers the 09Z-12Z total, for example).

The real value of plumes is that they allow us to quickly get a handle on the mean and the details of the spread of ensemble forecasts at a single point. They also are great tools for discerning timing of meteorological events, overall trends, and times when forecasts may become more or less uncertain. For example, check out the GEFS plumes for 2-meter temperatures at Pittsburgh (opens in a new window) from the 12Z run on January 14. It's easy to pick out diurnal temperature trends (daily increases and decreases), and clearly, the spread in the members grew much larger as forecast lead time increased. There was considerable disagreement among the GEFS members on surface temperatures toward the end of the forecast period (at times the range was nearly 30 degrees Fahrenheit).

If you're interested in learning more about plume diagrams (including more about their advantages and disadvantages), I recommend this document about plume diagrams (opens in a new window). Otherwise, I'll leave you with the following tips:

Tips for using plume diagrams

When using plumes, keep the following things in mind:

  • Assessing the range (spread) in the forecasts of individual members can help you gauge uncertainty, and possibly determine whether the ensemble mean might be useful (or not useful) in a particular case.
  • Look for trends in forecast variables in time. This can be particularly helpful in assessing the timing of particular events, like when the high or low temperature might occur, when a front might pass, or when heavy precipitation might start. Even identifying that there's great uncertainty in the timing of an event is helpful information!

That wraps up our lesson on ensemble forecasts. Before you move on, test your ability to interpret box-and-whisker plots and plume diagrams in the Quiz Yourself section below, and be sure to check out the data resources in the Explore Further box. They're great free resources for ensemble box-and-whisker plots and plume diagrams. There's no doubt that ensembles are extremely powerful forecast tools, but we have another extremely powerful set of forecast tools yet to cover -- statistically post-processed models. We'll cover those in the next lesson! Onward!

Quiz Yourself...

Answer the following questions to test your ability to interpret box-and-whisker plots and plume diagrams. If you need a larger version of the plume diagrams used in Questions 1 and 2, here are the full-sized versions -- Question 1 (opens in a new window); Question 2 (opens in a new window)

Explore Further...

Key Data Resources

  • ECMWF Ensemble Forecasts (opens in a new window): Lots of ensemble-based products here (feel free to explore!), but if you search for "ENS Meteograms," you'll get box-and-whisker forecast meteograms like the one shown above.
  • NAEFS Ensemble Meteograms (opens in a new window): Box-and-whisker forecast meteograms from the NAEFS.
  • GEFS Plumes (opens in a new window): You can generate plume diagrams for a variety of forecast fields (including 2-m temperatures, dew points, winds, QPF, etc.) from the GEFS for any of the stations on the map.
  • REFS Plumes (opens in a new window): You can generate plumes for a number of forecast fields (including 2-m temperatures, dew points, winds, QPF, etc.) by typing the station ID in the "site" box on the left. The link defaults to Olympia, Washington (KOLM), so you just need to change the site to generate plumes for your location of interest.
sas405

Lesson 3. Statistical Models

Lesson 3. Statistical Models

Motivate...

Creating an accurate weather forecast can be a difficult task, even with today's constantly improving computer technology. Our atmospheric models are able to make very detailed predictions on spatial and temporal scales, which were impossible just a few decades ago. As for the role that models play, think about what it would be like to create a weather forecast with absolutely no computer model guidance whatsoever. I think even most professional meteorologists today would be pretty lost! As much as we like to talk about how imperfect the models are (and they are imperfect...), we rely on them a great deal, and they are the main reason behind increases in forecast accuracy at longer lead times.

Still, perfect weather forecasts are elusive. Why? The primary reason is that observations which serve as a reality check for model initializations tend to be coarse and imperfect (yes, there are inherent instrument errors). We simply can't know and measure all of the relevant processes going on in the atmosphere everywhere at any given moment. Moreover, we cannot accurately model nonlinear weather phenomena (opens in a new window) and related physical processes, particularly on the mesoscale and microscale. Common nonlinear phenomena include turbulent motions that arise in concert with convection and the mechanical eddies that form when the wind blows over rough ground. If you're rusty on how computer models work and what their main flaws are, I strongly recommend that you look back at Lesson 10 in METEO 101 (opens in a new window) (login required).

These examples don't even account for other factors that affect the movement of air, like planes zipping through the air (opens in a new window) or butterflies flapping their wings. The atmosphere sounds like a pretty chaotic place, doesn't it? Indeed, if you've heard of chaos theory, it's related to our discussion of nonlinear processes that we can't model accurately. If you're interested in reading a bit more about chaos theory (opens in a new window), the beginning of the Wikipedia entry will give you a taste.

Photograph of a monarch butterfly

Are the flaps of this butterfly's wings responsible for your most recent rainy day? No, but the atmosphere is still full of nonlinear processes that we can't regularly measure and model.

In popular culture, the phrase "butterfly effect" is often used as a metaphor for chaos theory (here's a mildly amusing cartoon linking the butterfly effect to weather (opens in a new window)). The metaphor suggests that a seemingly random flapping of a butterfly's wings in Brazil (or wherever) could eventually affect weather on a larger scale. In reality, this idea is pretty silly. Any flap of a butterfly's wings in Brazil would be dampened out locally by other larger-scale forces, and thus there's no way that a single butterfly could alter the weather a month later in some distant part of the world.

But, despite the chaotic nature of the atmosphere, today's computer models usually produce helpful forecasts of the overall weather pattern at least several days into the future. Even when a computer model predicts the overall weather pattern fairly well (say, the pattern of 500-mb heights or 850-mb temperatures), it still might struggle more with the details of the weather we experience at the surface. Indeed, a computer model's predictions for key surface weather variables--the variables that most folks care about like temperatures, dew points, winds, and precipitation--tend to be less reliable than its predictions for the overall weather pattern in the upper atmosphere.

Fortunately, meteorologists have developed tools that, on average, produce better forecasts for surface weather variables than raw computer model output can. You've already learned a bit about these tools, like Model Output Statistics (MOS) and the National Blend of Models (NBM) in your previous studies. We'll do a little reviewing in this lesson, but if you're rusty on the basics of these tools, I suggest reviewing Lesson 11 in METEO 101 (opens in a new window) (login required). These "statistically post-processed" types of model guidance are incredibly useful for forecasters, so they warrant a closer look. In this lesson, we'll look "under the hood" of MOS and the NBM so that you better understand how they work, what their strengths are, and what their limitations are. Let's get started!

sas405

Statistical models: Why and How?

Statistical models: Why and How?

Prioritize...

By the end of this page, you should be able to discuss why forecasters use statistically post-processed models, and explain the difference between "dynamic" and AI models versus "statistically post-processed models" like MOS and the NBM. You should also be able to generally describe how MOS and NBM forecasts are produced, the accuracy of their 24-hour temperature forecasts, as well as an overall philosophy for using statistically post-processed forecasts (whether it's wise to frequently make large adjustments).

Read...

A Little Review

Before we get into the dirty details of how statistical models work, it's important to remember a few things. First, why do we even need them in the first place? Recall that the GFS, ECMWF, and other major computer models are dynamic models. By "dynamic", I mean that these models solve mathematical equations which describe atmospheric behavior in order to predict how the atmosphere will evolve with time (these equations include variables such as temperature, moisture, and wind). Major modeling centers also run artificial intelligence (AI) versions of these models, which are trained on decades of historical weather patterns to "learn" how the atmosphere evolves in time in order to make their predictions.

The problem with dynamic models is that, on average, these models don't predict important near-surface forecast variables very well (like temperatures, dew points, winds, etc.) because the physical processes near the earth's surface (such as radiation fluxes, turbulence, etc.) are just too complicated to mathematically model very accurately. Instead, meteorologists create oversimplifications (called parameterizations) to try to simulate these processes. Not surprisingly, oversimplification breeds forecast errors in the dynamic models. Predicting detailed surface weather is also not a strength of global AI models, but for a different reason. Their training data (from reanalysis data sets) are imperfect and not sufficiently high resolution to capture all the "real" details of surface weather, which reduces forecast accuracy. So, forecasters are faced with the conundrum that the variables the general public cares about the most are the ones that dynamic and AI models predict with the least accuracy.

Dynamic and AI model forecasts for surface temperatures winds, dew points, etc. are easily available online, and in many cases come with very colorful, detailed-looking maps like this ECMWF forecast of 2-meter temperatures (opens in a new window). But, the raw forecasts from dynamic models and AI models for these variables are, on average, less accurate than those from statistically-post processed models (I'll just call them "statistical models" for brevity), which apply statistical techniques to dynamic or AI model forecasts in order to produce improved forecasts. Keep that in mind next time you see dynamic or AI model forecast maps for variables like surface temperature or wind featured on websites or social media!

The primary statistical models in use today essentially use statistics to "post-process" dynamic or AI model guidance in order to improve upon the original model forecasts. For decades, Model Output Statistics (MOS) was the dominant statistical model applied to improve raw dynamic model forecasts for surface temperatures, surface dew points, surface wind speeds and directions, horizontal visibility, cloud coverage, cloud ceilings, probability of precipitation, and other variables.

At its core, MOS uses statistical equations that correlate variables predicted by the dynamic models to observed weather variables in order to correct for biases in the dynamic model forecasts. So, MOS is a form of bias correction (a process that adjusts model forecasts based on their past errors in order to create an improved forecast). But, other more direct approaches to bias correction exist, and with increases in computing power in the past few decades, they can be applied to large ensembles, as is the case with NCEP's flagship statistically post-processed forecast product -- the National Blend of Models (NBM).

MOS / NBM graphic

MOS and the National Blend of Models are the primary statistical models produced by the National Centers for Environmental Prediction
Credit: Meteorological Development Lab

Ultimately, statistical guidance from MOS and the NBM is, on average, more accurate than forecasts for surface variables from raw dynamic or AI models. And, over the long haul, MOS and NBM forecasts are usually pretty good. How good? Well, in a sample of 24-hour temperature forecasts collected by the course author over a three-year period as part of a national forecasting competition, MOS's predictions were within 5 degrees Fahrenheit about 88 percent of the time, and within 10 degrees Fahrenheit 99 percent of the time. The NBM's performance was similar. It was within 5 degrees Fahrenheit 92 percent of the time and within 10 degrees Fahrenheit 98 percent of the time. MOS has a very long track record, and has been able to approach this level of accuracy on short-term forecasts for a few decades. The NBM's track record is shorter (it became operational in 2016), but it, too, provides reliable guidance for short-term weather forecasts.

In other words, MOS and the NBM are usually in the right ballpark, and forecasters don't need to make large adjustments to their forecasts very often (subtle tweaks are more common). But, MOS and NBM guidance is certainly not perfect, and rare occasions do come along when their forecasts have large errors. Therefore, it's up to the human forecaster to recognize those situations and create better forecasts when appropriate. How do we decide when to adjust this reliable guidance and by how much? For starters, we have to learn a bit more about how they produce their forecasts.

A Short Lesson in Statistics

It's important to remember that statistical models are not the same as dynamic models, which are based on a system of primitive equations (equations that include the equations of motion (opens in a new window) based on Newton's Second Law of Motion, the Hydrostatic Equation, the Ideal Gas Law, the First Law of Thermodynamics, the Conservation of Mass, and the Conservation of Mass of Water). The primitive equations, along with the equation that governs surface pressure tendency, can be solved to predict future states of the atmosphere. AI models don't solve the primitive equations, but instead they "learn" atmospheric behavior from being trained on decades of reanalysis data. I know I'm repeating myself here, but make sure you understand the differences in the output from dynamic and AI models and the output from MOS and the NBM. 

MOS, on the other hand, relies on statistical equations that are developed by correlating raw dynamic model output to observed weather conditions using the technique of linear regression (opens in a new window). The resulting statistical equations are then used to predict a select group of forecast variables. To get an idea of what's involved in creating a MOS forecast, check out this short video (2:06) showing a simplified explanation of how MOS can generate daily high temperature forecasts (opens in a new window) (video transcript (opens in a new window)) from dynamic model forecasts of 1000-850-mb thickness during the month of January at Columbus, Ohio.

Note in the video how the MOS equation takes a variable that is well predicted by the dynamic model (1000-850-mb thickness, in this example) and uses it to predict a variable that is not well predicted by the dynamic model (surface high temperature). Of course, there's more than just one important predictor of the maximum temperature (the predictand) at each selected site. As an extreme example, there can be a whopping 15 model predictors for MOS visibility. Thus, as the number of predictors increases, linear regression equations for a given predictand look increasingly more complicated:

y = a 0 + a 1 x 1 + a 2 x 2 + a 3 x 3 + … + a n x n 

where:
y is the predictand (dependent variable)
a0, a1, a2, etc. represent constants
x0, x1, x2, etc. represent predictors (independent variables)

Feeding raw output from dynamic models into these linear regression equations generates MOS at a number of selected cities across the United States (check out this map of the sites for GFS MOS (opens in a new window), for example). As you may recall, each city has its own set of linear regression equations, although the linear regression equations for PoP (Probability of Precipitation) at single sites are "pooled" to create regional equations for PoP (hence, the color-coding on the map you just saw). All in all, accounting for all the sites at which MOS is generated, there are on order of 10 million unique regression equations! For the record, MOS has a long history and was developed using dynamic models only, though it could be created for use with AI models as well (NCEP currently only produces GFS MOS operationally). 

While the NBM is still a statistical model, which uses dynamic and AI model forecasts as its input, it doesn't work quite like MOS. Instead, the NBM takes a more direct approach to bias correction. Consider this simple example of bias correction: Imagine that the GFS dynamic model, on average, has been 3 degrees Fahrenheit too low on its high temperature forecasts over the past month (the model's recent average bias would be 3 degrees Fahrenheit too low). A "bias-corrected" forecast for tomorrow's high temperature would be to add 3 degrees Fahrenheit to the GFS dynamic model's prediction, in order to correct for this recent bias. Pretty straightforward, right?

A simple bias-corrected forecast from any single model is likely to perform better, on average, than raw dynamic model forecasts, but would still typically be inferior to a more sophisticated MOS approach. So, to improve the forecast further, the NBM takes the simple concept of bias correction and applies it to a very large ensemble of dynamic models, AI models, and MOS forecasts (yes, some MOS forecasts are actually included in the NBM) to create a bias-corrected "superensemble." The bias-corrected members that comprise the NBM are then blended together to create a consensus forecast, but the consensus isn't just a simple average. In the NBM, the accuracy of each bias-corrected ensemble member over the past 60 days is then used to determine its weighting in creating a final consensus forecast. Those members that have performed better in the last 60 days are weighted more heavily, while those members with less accurate forecasts over the last 60 days are weighted less in the final consensus forecast. Through this blending technique, the final consensus forecast is calculated and becomes the NBM's deterministic forecast. In addition, the NBM provides a range of bias-corrected ensemble solutions, so it's a very powerful tool indeed!

The Bottom Line

Photograph of a slot machine

Using MOS and the NBM effectively has nothing to do with blind luck. Keep big gambles out of your weather forecasts!
Credit: Slot Machine / Jeff Kubina / CC BY-SA 2.0 (opens in a new window)

The statistical techniques used to create MOS and NBM forecasts tend to remove model biases, but in different ways. MOS does so through linear regression. If a dynamic model, for example, has a persistent cold bias in its 2-meter temperature forecasts over a long period of time, MOS will statistically account for that and make a better forecast. The NBM, on the other hand, more or less accounts for recent model biases (and uses the power of a large ensemble). So, even without any fancy physics equations, MOS and the NBM can out-forecast dynamic and AI models by using statistics to reduce their biases!

That's not to say that MOS and the NBM don't make errors (sometimes big ones). These post-processed forecasting products are reliant on dynamic and/or AI models for their starting data, so if those models are "out to lunch" completely, then MOS and the NBM are starting with "garbage," motivating a popular expression in the forecasting community, "Garbage in, garbage out." The large ensemble approach of the NBM, however, is less susceptible to a complete set of "garbage" inputs because it has many members, whereas MOS relies on a single deterministic model.

I don't mean to absolve MOS and the NBM from any blame whenever forecasts go awry. Indeed, in MOS's case the lines of "best fit" are not perfect fits, so the statistical equations that comprise MOS no doubt have shortcomings. And there are indeed weather patterns for which MOS falls from favor. In terms of the NBM, a weighted average of a large bias-corrected ensemble won't always be a good forecast (take a convective precipitation forecast situation, or a situation where models disagree on the placement of a very strong front, for example). As is the case when we discussed ensembles, the ensemble mean isn't a great forecast in all situations, and so it is with the deterministic NBM forecast. With these thoughts in mind, I suggest approaching your use of MOS and the NBM with the following mindset:

Getting Philosophical About Statistical Models

  • Realize that beating MOS and the NBM every single day is impossible. Some days, MOS and/or the NBM (or a MOS/NBM consensus) provides a great forecast, so on those days you want to stick with MOS and/or the NBM because they can't be beat. Dramatically attempting to beat them every day is, by and large, a losing strategy.
  • Often, MOS and/or the NBM is a pretty good forecast, and only a minor tweak is the best approach (just be sure you have sound meteorological reasoning based on the big picture situation).
  • Recall that MOS does struggle in specific weather patterns. The NBM also has particular weather patterns where its forecasts are more likely to struggle. When you recognize one of those weather patterns, that's your time to pounce and take full advantage of their weaknesses. But, these situations don't come around very often, so it's important to be selective. Keeping in mind the statistics about accuracy from the earlier on the page, you better have a really good reason to dramatically depart from MOS or the NBM (we'll talk specifically about some of the "good reasons" later in the course). Remember, swinging for the fences is a risky strategy, and you'll strike out much more often than you'll hit a home run.

Statistical guidance is an essential part of a forecaster's toolkit, but to most effectively use these tools, we have to dig deeper. We'll start with a closer look at MOS. Read on.

sas405

MOS Tables

MOS Tables

Prioritize...

When you've completed this page, you should be able to decode a MOS table and extract relevant forecast information correctly (namely for maximum and minimum temperature, wind direction and speed, and precipitation) regardless of forecast context.

Read...

In your previous studies, you learned how to decode most parts of a MOS table, but you mainly used them to look at forecasts of high and low temperatures. Each MOS table, however, contains forecasts for many other variables, too. If you're rusty on decoding MOS, I highly recommend that you refer to this GFS MOS guide (opens in a new window), which take you through decoding an entire table. In this course, you'll need to become more familiar with the rest of the MOS table because some of the "other" forecast variables can provide extra clues that can help you assess MOS's behavior.

Of course, it's important to understand what each line of the MOS table actually tells you, so that you can interpret the data correctly for your forecast context. Indeed, before interpreting a MOS table (or any tabular forecast data), you must answer one of the fundamental forecast questions posed previously: What time period am I interested in? This is especially important for when the weather may not follow typical diurnal patterns (cases when, say, high temperature may not occur during the daytime and/or low temperatures may not occur overnight). To highlight some issues that sometimes crop up in interpreting MOS tables, I'll use the GFS MOS forecast for Hobby Airport in Houston, Texas (KHOU) initialized at 18Z on February 2, 2022 (below) as an example.

GFS MOS table for KHOU from the 12Z run on February 2, 2022

The GFS MOS tabular output for KHOU (Hobby Airport in Houston, Texas) from the 18Z run on February 2, 2022.
Credit: National Centers for Environmental Prediction.

MOS Table Temperature Issues

I'm not going to go through a complete translation of the table (that's what the review links above are for), but I want to point out some interpretation issues with tabular data. You may recall that the "N/X" line shows the minimum and maximum temperatures during the local nighttime and daytime periods, respectively. So, the nighttime minimum and daytime maximum temperatures for February 3 listed at 12Z February 3 and 00Z February 4, respectively are both 44 degrees Fahrenheit (opens in a new window). How can the nighttime low and the daytime high be the same? Does that mean MOS thinks temperatures will remain constant on February 3?

No! The maximum and minimum temperatures on the N/X line do not cover a full calendar day (or any full 24-hour period). They cover local nighttime (for the minimum temperature) and local daytime (for the maximum temperature). If you were forecasting the maximum and minimum temperature for the calendar day of February 3 at KHOU, which runs from 06Z February 3 to 06Z February 4 (Central Standard Time is UTC - 6 hours), MOS's forecast wouldn't be 44 degrees for either the high or the low! In order to better understand what's going on, take a look at the "TMP" line, which shows the predicted temperatures at three-hour intervals. These forecasts show temperatures declining through much of the period (opens in a new window), starting at 63 degrees Fahrenheit at 06Z on February 3 (local midnight Central Standard Time) and dropping into the low 30s before 06Z on February 4. As best we can tell from the temperature forecasts at three-hour intervals, for the calendar day of February 3, the GFS MOS maximum temperature forecast was actually 63 degrees Fahrenheit (opens in a new window) and the minimum temperature forecast was 32 degrees Fahrenheit (opens in a new window) (as best we can tell from the three-hourly forecasts).

In this case, we found the 24-hour maximum and minimum temperatures on the TMP line because the maximum temperature actually occurred overnight (early on February 3) instead of in the mid-to-late afternoon (as typically happens), while the end of the calendar day was colder than the time around dawn (when the low temperature typically occurs). It's always up to the forecaster to have situational awareness of the big-picture weather situation, and in this case, a strong cold front passing through Houston around 06Z on February 3 (opens in a new window) (and cold advection in its wake) was the culprit behind the decline in temperatures throughout the day. Forecasters who weren't in-touch with the big picture could have easily been fooled by MOS's N/X forecasts. In certain weather situations, especially in the cold season (involving passages of fronts, major changes in sky cover or wind, etc.), daily temperature swings don't always follow the classic diurnal model, so it's important that you take the time to check both the N/X and TMP forecasts.

MOS Table Wind Issues

Note that the MOS table above also includes hourly forecasts (at three-hour intervals) for wind direction and sustained wind speed (the "WDR" and "WSP" lines, respectively). But, these forecasts are merely snapshots of predicted wind direction and speed. What if we wanted to precisely time a wind shift or predict the maximum sustained wind speeds during a particular time period? We can't directly gather that information from a MOS table. For a wind shift, we can only identify a three-hour window when it's predicted to occur. For sustained wind speed, if we keep with the February 3 example above, the highest MOS wind speed listed is 22 knots at 03Z on February 4 (opens in a new window) (which is late in the evening on February 3 in Houston). But, that's just the predicted wind speed right at 03Z. MOS really doesn't predict the maximum sustained wind speed for a given period of time. So, if we're interested in maximum sustained wind speeds during a time period, we have to take MOS's forecasts and apply some additional strategies, which we'll cover later.

MOS Table Precipitation Issues

MOS tables include predictions for probability of precipitation over six and twelve-hour periods ("P06" and "P12," respectively), as well as QPF for six and twelve-hour periods ("Q06" and "Q12," respectively). Refer again to the GFS MOS table for KHOU from the 18Z run on February 2, 2022 below (precipitation forecasts are highlighted in yellow):

GFS MOS table for KHOU from the 18Z run on February 2, 2022.

The GFS MOS tabular output for KHOU from the 18Z run on February 2, 2022. The MOS forecasts for probability of precipitation and QPF are highlighted in yellow.
Credit: National Centers for Environmental Prediction.

It's critically important that you realize that the numbers listed in these rows refer to the 6- or 12-hour periods ending at the time the number is listed. In other words, if we wanted to know which 6-hour period had the highest probability of precipitation, it's the 06Z to 12Z period on February 3 (a probability of 92 percent (opens in a new window), listed at 12Z). To determine how much precipitation MOS predicts during that time period, we go to the Q06 line and see a "3", which converts to 0.25 inches to 0.49 inches of liquid during the 6-hour period ending at 12Z.

If you need to determine MOS's precipitation forecast for a longer period of time, you need to select the appropriate ranges covering shorter time periods and sum them. For example, what if we wanted to know the GFS MOS precipitation forecast for 06Z to 18Z on February 3? While that's a 12-hour period, we can't use the Q12 line because its periods run from 00Z to 12Z and 12Z to 00Z. Those forecasts don't align with our time period of interest, so we need to sum the 6-hour ranges covering the 06Z to 12Z period and the 12Z to 18Z period by adding the bottom ends of each range and the top ends of each range, yielding a total range of 0.26 inches to 0.58 inches, as shown in the table below.

MOS Total Liquid Precipitation Forecasts
GFS MOS QPF
Time PeriodQ06 CategoryLiquid Precipitation Range
06Z-12Z30.25 inches to 0.49 inches
12Z-18Z10.01 inches to 0.09 inches
06Z-18Z Total—0.26 inches to 0.58 inches

If you wanted to know MOS's forecast for a longer time period (say, 24 hours), you might be able to use the Q12 ranges if your forecast period aligns perfectly with the Q12 periods (00Z to 12Z and 12Z to 00Z). If not, I recommend simply summing the appropriate Q06 ranges that cover your 24-hour period of interest. For example, if we wanted the GFS MOS total QPF prediction for 06Z February 3 to 06Z February 4 (essentially the calendar day of February 3 in Houston), we have to sum the appropriate 6-hour ranges (opens in a new window) as in the table below:

MOS Total Liquid Precipitation Forecasts
GFS MOS QPF
Time PeriodQ06 CategoryLiquid Precipitation Range
06Z-12Z30.25 inches to 0.49 inches
12Z-18Z10.01 inches to 0.09 inches
18Z-00Z10.01 inches to 0.09 inches
00Z-6Z10.01 inches to 0.09 inches
06Z-06Z Total—0.28 inches to 0.76 inches

Note that in the examples above, we didn't just add the Q06 Category numbers (3+1+1+1=6 for our 24-hour example) and then convert to a range. Doing so can lead you to an erroneously high QPF total. If you're wondering why we didn't include the "2" at 06Z on February 3 (the very beginning of our period of interest), remember that these are 6-hour QPF categories for the preceding six hours. So, that ”2“ refers to the six-hours from 00Z to 06Z (opens in a new window) -- before our period of interest started.

Other Variables

MOS predictions for other variables, such as sky cover ("CLD"), wind direction, ceiling height category ("CIG"), and visibility category ("VIS") can also provide clues that can help with our high and low temperature, wind speed, and precipitation forecasts if you apply some sound meteorology and have a good handle on the big picture. So, I strongly recommend taking some time to get familiar with the format of these predictions on MOS tables so that you can decode them quickly.

Key Data Resources

MOS Tables

My goal for this page was to refresh your memory on how to interpret MOS tables, as well as point out some nuances related to interpreting them. Trust me: These nuances are important! Your forecast could get off to a pretty bad start if you don't read MOS properly! But, to this point, we really haven't dug deeper into MOS and focused on how it really works. We'll cover that in the next several sections. But, first, I suggest you try your hand at the Quiz Yourself questions below to get a little practice interpreting MOS tables.

Quiz Yourself...

When interpreting MOS forecasts, it's critical that you can extract the correct forecast information for your forecast context. To give you some practice, here's the GFS MOS forecast for Bismarck, North Dakota from the 18Z run on April 23, 2022. To answer the questions below, imagine you're forecasting for the period from 06Z on April 24 through 06Z on April 25.

GFS MOS table for KBIS from the 12Z run on April 23, 2022.

The GFS MOS tabular output for KBIS (Bismarck, North Dakota) from the 18Z run on April 27, 2022.
Credit: National Centers for Environmental Prediction.

sas405

MOS Temperature Forecasts

MOS Temperature Forecasts

Prioritize...

Upon completion of this page, you should be able to discuss, in general terms, the most commonly used predictors for MOS temperature forecasts, including different ways that MOS tends to account for clouds in its temperature forecasts. You should also be able to discuss how MOS simulates climatology, and why we have to be extra careful with MOS temperature forecasts around April 1 and October 1.

Read...

So far, we've been focusing on how MOS generally works and interacts with the dynamic models. But, in order to really get to know MOS and understand how it creates typically reliable (albeit imperfect) forecasts, we need to look a bit closer at MOS predictions for specific variables. Let's start with temperature.

Photograph of a bank vault

Are the MOS predictors and regression equations safe and sound behind this vault door? Probably not, but the public doesn't get to see the station-specific MOS equations or predictors.

In order to get a better idea of how MOS arrives at its temperature forecasts, we have to examine the predictors from the dynamic models that MOS uses in its regression equations. But, this is not a simple topic! For starters, the predictors of maximum and minimum temperatures vary from station to station. Also, the predictors and the regression equations for each station aren't public, so we're left to make generalizations, and in some cases, speculate.

How are the temperature predictors determined at each station? The screening process is really complex! In a nutshell, many potential predictors (perhaps 50 or 100) are offered up as possibilities, and they are screened automatically by a statistical software package. The software "weeds out" the predictors that offer no predictive information or are redundant to others, leaving only the valuable predictors (usually only 8 to 12 out of the original pool). For example, 850-mb temperature is tried as a predictor in every regression equation at every station. But, it actually survives the screening process rather infrequently. The details of this screening process (technically, a "stepwise regression scheme") are well beyond the scope of this course, but if you're curious, I've included a description of the screening process in the Explore Further section below.

Typical MOS Temperature Predictors

So, what types of forecast variables from the dynamic models make it through the screening process and are accepted as valid MOS temperature predictors? Let's take a look at the major ones (remembering, of course, that we're broad-brushing here since we don't have access to the specific predictors and equations for each station):

  • Dynamic model temperature forecasts: Specifically, boundary-layer temperatures and dew points (either two-meter temperatures and dew points or those at some other representative pressure level like 1000 mb). In the West (especially in the Rockies), the 850-mb temperature is likely used a bit more often than it is in the East because of the higher surface elevations.
  • Recent temperature and dew point observations (short-term forecasts only): Recent observations for temperature and dew point are only used for short-term MOS forecasts. These observational predictors are important for the next 12 hours, but they aren't usually selected for forecast times greater than 24 hours.
  • Other thermal predictors: At some cities, thermal variables like 1000-850 mb thickness pass screening as MOS temperature predictors, but such predictors tend to carry less weight in the final regression equation.
  • Attempts to account for clouds: Predicted sky coverage is not directly used in MOS temperature forecasts. Instead, MOS uses dynamic model forecast variables that are related to potential cloud development, such as mean relative humidity between 1000 and 500 mb (or other atmospheric layers), which is frequently used in GFS MOS.
  • Wind-based predictors: At some stations, the east-west and north-south components of the wind are sometimes selected as MOS temperature predictors.
  • Simulated climatology (longer-term forecasts only): Climatology is also a MOS-temperature predictor for longer-term forecasts (beyond 24 hours in most cases), but the daily average high and low temperatures are never used as temperature predictors. Instead, the sine and cosine (trigonometric functions (opens in a new window), see below) of the day of the year serve as potential predictors that simulate climatology. There is a separate set of equations for the warm season and cool season, too, but this simulated scheme is imperfect because climatology doesn't exactly track with the arithmetic sum of the sine and the cosine of the day of year. For really short-range forecasts (6 to 12 hours), climatology does not make it into the final regression equations (observations trump climatology in the short term), but becomes increasingly more important for longer-term forecasts.

Graph of sine and cosine curves

Sine and cosine waves are far too smooth to perfectly mimic climatology.
Credit: Wikimedia Commons

Curiously, hourly observations of clouds, visibility, winds, etc., are not used as predictors, and observations from nearby stations are also not selected as MOS predictors. In case you're wondering, the regression equations for the daytime maximum and nighttime minimum temperatures are developed separately and have different predictors and/or coefficients. Once temperature predictors are chosen, they don't change at a given station with two exceptions:

  • The two days each year when regression equations switch between warm season and cool season. The equations switch from cool season to warm season on April 1, and they switch from warm season to cool season on October 1.
  • If the the most recent weather observations are missing at a station on any given day, then MDL will switch to back-up (secondary) regression equations created for short-term forecasts at the station. These back-up (secondary) regression equations only use output from dynamic models as predictors since there aren't any observational predictors available.

Points to remember...

  • As a general rule, the #1 MOS temperature predictor is boundary-layer temperature (two-meter temperature or a temperature representative of the boundary layer, say, at 1000 mb, for example).
  • "Climatology" is a MOS predictor for longer term forecasts (more than 24 hours), but is simulated with sine and cosine functions, which imperfectly mimic climatology.
  • MOS temperature equations switch from cold to warm season on April 1, and warm to cold season on October 1. Of course, the weather does not magically undergo any artificial switch on those dates, so keep an eye on MOS performance around the transition dates.

MOS Temperatures at Synoptic Times

The three-hourly MOS-temperature forecasts use the same two equations that generate the maximum and minimum temperatures, but the coefficients will vary for each of the three-hourly forecasts (this is how we get to the astounding total of 10 million MOS equations, roughly). Specifically, the regression equation that generates the 06Z, 09Z, 12Z, and 15Z MOS temperatures use the same predictors as the equation for the minimum temperature forecast, while the 18Z, 21Z, 00Z, and 03Z temperature forecasts use the same predictors as the equation for the maximum temperature forecast.

By using the same predictors as those used in the maximum and minimum regression equations, the chance of "inconsistencies" goes way down (but it's not completely eliminated). To eliminate inconsistencies, there is an extra layer of post-processing that occurs before MOS data gets disseminated. If one of the three-hourly temperatures happens to be greater than the maximum temperature predicted for the daytime, for example, the maximum MOS temperature is set to the highest three-hourly temperature forecast during that period (similar logic applies to MOS's minimum temperature). What about a final check for dew points? If the three-hourly temperature forecast happens to be lower than the corresponding three-hourly dew-point forecast, then post-processing uses the average of the two for the forecasts of both the three-hourly temperature and dew point.

Now that we've covered how MOS makes its temperature predictions, let's move on to MOS's wind and precipitation forecasts.

Explore Further...

The screening process (called a "stepwise regression scheme") used to determine relevant MOS predictors is pretty complicated, but this highly simplified take will give you an idea of how it works (if you're interested).

The first step in any regression procedure is to create a "data matrix." Any data matrix has one column for each potential predictor (variable) and one row for each "case." Now, suppose we had three years of dynamic-model historical forecasts, and we wanted to predict high temperature using a statistical approach. First, we'd have to create a bunch of different statistical models (regression equations):

  • one model to forecast the 12- to 24-hour high temperature based on the 00Z dynamic model run
  • one model to forecast the 36- to 48-hour high temperature based on the 00Z dynamic model run
  • one model to forecast the 24- to 36-hour high temperature based on the 12Z dynamic model run
  • one model to forecast the 48- to 60-hour high temperature based on the 12Z dynamic model run

Each model would require its own separate data matrix; therefore, each model would have its own separate regression equation. In other words, there would be one regression equation for each data matrix. Realistically, we would develop eight equations for maximum temperature: one set of four equations for the warm season, and another set of four equations for the cool season. To make life simpler, let's suppose we simply want to develop a set of four equations that apply to the entire year.

Let's build the first statistical model listed above -- the 12-24h high temperature based on the 00Z dynamic model run. If we had three years of history, the data matrix would look like a much larger version of the one I created below. In a nutshell, each row represents one data point (a "case", in other words). Given that we have three years of dynamic-model data, the data matrix would have 1095 rows of data (365 x 3), and, of course, each column would represent a predictor (variable) from the dynamic model, except for Column A, which lists the initialization date. The second column (Column B), which contains the observed high temperature that occurred 12 to 24 hours after initialization date/time, is also not a predictor, but it's obviously used to create the regression equation for the predictand (high temperature).

Example of a simple data matrix created in Microsoft Excel

A simple data matrix created in Microsoft Excel, containing columns for potential predictors and rows for each 00Z model run ("case").
Credit: Lee Grenci

With Columns A and B accounted for, Column C contains one of the predictors (variables), say the dynamic model 18-hour forecast for two-meter temperature. Column D lists the next possible predictor (variable), say the 18-hour dynamic model forecast for 850-mb temperature. Column E contains the next possible predictor (variable), and so on and so forth.

With the data matrix completed, MDL feeds it into the screening regression program (part of a statistics software package). There are a number of ways to screen (backwards, forwards, all possible models, maximum correlation, maximum significance, etc.). Also, a number of different types of regression exist (linear, additive, logistic, etc.). However, in all cases, the actual "screening" is done automatically by the software. A human being only has to enter the specific screening method and the relevant software settings.

The gist of each screening method is that the software will find a subset of relevant predictors that show some capability for predicting the predictand (MOS high temperature, etc.). Moreover, the software will discard all the other predictors that are either redundant to "valuable" predictors already selected or otherwise offer no additional predictive information. You might start out with 50 to 100 potential predictors, but end up with only eight or ten predictors in the final model.

What would happen if you skipped the screening process and tried to fit a regression model with 50 to 100 predictors? Such an approach is called overfitting. If you use too many predictors, then you'll probably be confronted with spurious statistical relationships between variables that make absolutely no physical sense. To use a baseball analogy of statistical overfitting, a statistical relationship like, "Pedro Martinez is undefeated pitching in the World Series on Sundays when there is a full moon" might be true, but it's irrelevant for analysis or prediction.

The specific screening method used by MDL is a forward stepwise method. It goes through all predictors individually and correlates each variable one by one with the predictand (hence the word, stepwise). The variable with the highest correlation is the first predictor selected in the model. Then it goes through the remaining predictors and finds the variable that makes the best two-predictor model (assuming the previously selected #1 predictor is included in the model). The procedure repeats itself until the next best selected predictor doesn't add any more correlation to the combined model. Usually, the procedure halts after selecting approximately 8 to 12 predictors.

Then we'd repeat the process for the last three statistical models. Keep in mind that these details are outside the scope of our course, but I hope this description gives interested folks an idea of how complex and tedious the process of developing MOS equations is!

sas405

MOS Wind and Precipitation Forecasts

MOS Wind and Precipitation Forecasts

Prioritize...

When you've completed this page, you should be able to execute a simple adjustment to MOS's forecasts in order to predict a maximum sustained wind over a period of time. You should also be able to discuss common MOS predictors for Probability of Precipitation (PoP), scenarios in which MOS precipitation forecasts may struggle, and reconcile how MOS can predict measurable precipitation in situations where its PoP is less than 100 percent.

Read...

MOS Wind Forecasts

As you've learned, MOS makes predictions for daytime maximum and nighttime minimum temperatures, but that's not the case with its wind forecasts. MOS only forecasts wind speeds at synoptic times. Indeed, the corresponding regression equations were created from the two-minute average wind-speed observations at 00Z, 03Z, 06Z, 09Z, 12Z, 15Z, 18Z, and 21Z in the original developmental data. Thus, they only predict a single wind speed at each synoptic time.

Of course, winds can vary substantially within a three-hour time period, so MOS's three-hourly "snapshots" often don't capture the true maximum sustained wind speed. In order to estimate a maximum sustained wind speed, forecasters will often add a few knots to the highest MOS wind speed forecast at the synoptic times as a "first guess." In reality, wind speed forecasting is a bit more complicated than that (we'll explore more later), but it's a good starting point if you're interested in predicting the maximum sustained wind speed over a period of time. For now, it's important to understand a bit more about how MOS wind speed forecasts are created.

Inflating MOS Wind Speeds

If the entire truth be told, the predicted wind speeds that you observe on MOS tables have already been "statistically adjusted" by a process called "inflation." MOS developers’ original stated goal was to accurately predict the mean (average) wind-speed event from the range of possibilities predicted by the dynamic models, meaning that observed wind speeds would sometimes be under-forecasted and sometimes over-forecasted. For example, odds were that MOS would under-predict a sustained wind of, say, 35 knots. At the time, it was not rocket science to anticipate that the public would definitely notice such “under-forecasts” of strong winds.

But, if MOS predictions were to be of any real use to forecasters and the general public, MOS developers needed to capture high-wind events, so they initiated a technique called "inflation" to post-process raw MOS forecasts (you can read the details of the technique in this seminal paper on inflation (opens in a new window), if you're so inclined).

Trees blowing in the wind.

MOS developers designed the technique of inflation to post-process raw forecasts from MOS regression equations. This technique statistically inflates higher MOS wind speeds to more accurately predict windy days like this one.

The goal of the inflation technique is to increase the standard deviation of the wind-speed forecasts away from the climatological value (mean wind speed). In other words, inflation tends to increase MOS wind speeds that are higher than the climatological value. For the record, the Meteorological Development Lab used to deflate MOS wind speeds that were lower than the climatological value, but they ended this technique in 2006.

Here's a specific example of inflation. Suppose the climatological wind speed at 18Z at a given site is 12 knots and that the raw MOS regression equations predict 18 knots. After the technique of inflation is applied, the final "inflated" MOS forecast for 18Z might be 22 knots. So, the winds predicted at the synoptic times might end up being closer to the true maximum value for the day than they otherwise would be, especially if the dynamic models are predicting a windy day (winds notably above climatology).

Points to remember...

  • When forecasting maximum sustained wind speed, remember that MOS does not forecast this exact variable. It only predicts sustained wind speeds at the synoptic times.
  • To compensate for the fact that maximum sustained wind speeds often occur in between synoptic times, a good starting point for estimating a maximum sustained wind is often to take MOS's highest wind speed prediction and add a few knots.
  • Be careful when winds are predicted to be dramatically higher than climatology. The technique of inflation is already working to increase MOS's predicted wind speeds, so you need to have a sound reason for dramatically adjusting MOS's predictions upward (we'll cover some possible sound reasons later).

The Bottom Line

Keep in mind that my suggestion to add a couple of knots to MOS wind speeds in order to predict a maximum for a period doesn't imply that MOS routinely underestimates wind speed. It really doesn't (even though it might appear that way if you need to regularly forecast a maximum sustained wind). MOS regression equations just don't account for any peak winds that occur between the three-hour time intervals listed on standard MOS tables.

If you're predicting a maximum sustained wind speed, just adding a few knots sometimes doesn't cut it, however. Later in the course we'll cover some reasons and situations where you might need to make dramatic adjustments. MOS also doesn't predict wind gusts, so if you need to predict gusts, MOS won't be much help. We'll talk a bit more about tools and strategies wind gust forecasting later in the course, as well.

MOS Precipitation Forecasts

Image showing the various regions for which MOS precipitation equations are developed

The color coding of the individual MOS stations indicates the different regions used for regionalized MOS equations, such as PoP.
Credit: National Centers for Environmental Prediction

As you know, MOS regression equations for temperature and wind are developed for each unique station, but that's not the case with the regression equations for Probability of Precipitation (PoP). The MOS PoP equations are regionalized. To get an idea of what the regions look like, check out the image on the right, showing the GFS MOS stations. The dots representing each individual MOS station are grouped into different colors, each color corresponding to a region over which data are pooled to develop regionalized PoP regression equations.

Using a regionalized approach is important when dealing with precipitation because it helps MOS better forecast events that occur very infrequently at an individual station. For example, at any individual station, it might be very rare for two inches of liquid precipitation to fall in a 12-hour period, so it's hard to represent such cases in the statistical sample. But, when we pool the single-site data over regions, we can add more cases to the sample and better predict such events with MOS's statistical approach. Furthermore, regionalization allows MOS to create precipitation forecasts for stations without regular, reliable observations.

Typical MOS PoP Predictors

Frequently chosen PoP predictors are:

  • dynamic model predictions of mean relative humidity between 1000 mb and 500 mb
  • dynamic model QPF
  • dynamic model predictions of vertical velocity at 850, 700, and 500 mb.
  • dynamic model predictions of wind direction at 10 meters

But, developing regression equations for variables like PoP is a bit different than it is for temperature. Precipitation, by its very nature, is a binary variable. Either it precipitates or it doesn't. There's no in between. That's not the way it is with, say, temperature, which is a continuous variable (we always have a temperature...the value just changes with time). That difference is at the root of why the process of developing PoP regression equations is different.

Indeed, a binary system (0 or 1) is used for scatterplots (0 corresponds to any case when no measurable precipitation fell; 1 represents any case where at least 0.01 inches of liquid-equivalent precipitation fell). Below is an example binary scatterplot of measurable precipitation (0.01 inches or greater) that fell in the 12-24 period after 00Z as a function of predicted mean 1000-500-mb relative humidity (12 to 24-hour forecast based on a dynamic model's 00Z runs) at Columbus, Ohio, from December 1 through March 5. From this binary scatterplot, a linear regression equation could be obtained (although doing so requires advanced statistical concepts like reduced variance, which are outside the scope of this course). At any rate, a more statistically significant sample size at Columbus would ultimately be "pooled" with other single-site data in the region to yield regional equations for PoP.

A sample binary scatterplot of measurable precipitation

The binary scatterplot of measurable precipitation (0.01 inches or greater) that fell in the 12-24 period after 00Z as a function of predicted mean 1000-500-mb relative humidity (12 to 24-hour forecast based on a dynamic model's 00Z runs) at Columbus, Ohio, from December 1 through March 5.
Credit: NOAA

You may have noticed that all of the MOS PoP predictors are predictions from the dynamic models (no "observation-based" predictors are used like they are with temperature). So, not surprisingly then, in situations where the dynamic models traditionally struggle (such as warm season convective events or any precipitation driven by mesoscale processes, really), MOS precipitation forecasts can be expected to struggle, too.

One final note about MOS precipitation forecasts. You'll surely encounter situations where MOS predicts a PoP less than 100% over a six- or twelve-hour period (perhaps even a low PoP, like 30%), yet also indicates that liquid precipitation will accumulate. How do we reconcile the two? Well, technically, MOS QPF forecasts are conditional, meaning that the QPF is based on the condition that at least 0.01 inches of liquid precipitation falls. So, MOS isn't really contradicting itself by forecasting a low PoP and possibly significant QPF. Such output simply means that the probability of precipitation is low, but if it does precipitate, MOS "thinks" the precipitation could be heavy.

Now that we've looked more in depth at how MOS forecasts wind and precipitation, we have to grapple with a key issue -- since the MOS approach to statistical forecasting relies on historical weather data, how does it perform when rare or unprecedented weather might occur? Read on.

sas405

MOS and Climatological Extremes

MOS and Climatological Extremes

Prioritize...

When you're finished with this page, you should be able to discuss why MOS tends to struggle more often when the weather is near climatological extremes, but doesn't always struggle in such situations. You should also be able to discuss why MOS's forecasts can be highly suspect when uncertainty is great.

Read...

As you know, MOS uses observed weather data and dynamic model forecasts over a span of several years to help create its statistical forecasts of temperature, precipitation, wind, and other variables. But, what happens when the weather starts to vastly depart from "average" and nears (or exceeds) climatological extremes?

Your first thought might be that MOS will always struggle because it's likely such situations are underrepresented in MOS's statistical sample. Record-breaking weather doesn't happen very often at any one location, so MOS regression equations are naturally developed with few data points (or none) near climatological extremes. Such thoughts have some merit, but it's not quite that simple. Let's investigate.

A Partial Myth

Ultimately, the thought that MOS always struggles when the weather greatly departs from climatology is a partial myth. MOS does tend to struggle more frequently in extreme weather situations, but sometimes it can perform quite well when weather conditions near the extremes (MOS can sometimes accurately predict a record high temperature, for example). One big determining factor in how MOS will perform in these situations is how the MOS predictors behave near the extremes. For example, in forecasting wind speed, a key MOS predictor is the predicted wind components (based on the pressure gradient) from the dynamic models. Because the relationship between wind-component predictions from the dynamic models and observed wind speed is roughly linear, we can extrapolate it successfully to predict wind speeds even when winds blow much faster than average (or even faster than any of the observations in MOS's data sample). All of this assumes, of course, that the forecasts for the predictors from the dynamic model going into the MOS equations are reasonably accurate.

But, such linear relationships don't exist for all forecast variables. Indeed, MOS runs into trouble when the linear relationship breaks down in weather conditions that depart from climatology. For example, when the dynamic models predict upward motion increasing from -1 to -2 microbars per second, MOS's prediction for rainfall will probably be in the ballpark. But, when dynamic models predict upward motion increasing from -10 to -11 microbars per second, MOS might not handle this very well because the linear relationship breaks down at these higher values for upward motion.

Enhanced IR image of Hurricane Florence making landfall in North Carolina
Enhanced infrared satellite image of Hurricane Florence making landfall on the morning of September 14, 2018.
Link to animated loop of the image. (opens in a new window)
Credit: National Weather Service

Let's take a look at a concrete example. I think we can all agree that a hurricane making landfall qualifies as an extreme event. Likely any given coastal city has few (if any) such cases in its MOS statistical sample. Can MOS give us decent wind predictions? Let's see how MOS did at Wilmington, North Carolina (KILM) on September 14, 2018 as Hurricane Florence (a Category 1 storm) made landfall nearby. Here's the 06Z GFS MOS forecast (opens in a new window) from the previous day, which predicted wind speeds ranging from 27 knots to 44 knots at the synoptic times between 03Z and 18Z on September 14 (the hours surrounding landfall). Wait a second! Those predictions seem awfully low for having a landfalling hurricane nearby, don't they? Shouldn't we expect maximum sustained winds near minimum hurricane strength (64 knots on the Saffir-Simpson Scale (opens in a new window)) at least?

Not necessarily, and it turns out that MOS's seemingly "low ball" predictions performed fairly well. The METARS from 03Z, 06Z, 09Z, 12Z, 15Z, and 18Z (opens in a new window) showed sustained winds between 29 knots and 53 knots, and MOS was within a few to several knots at most of these times (some of its forecasts were actually too high). Its worst forecast was at 12Z, when it was 11 knots too low (a 42-knot forecast, with an observation of 53 knots), but overall, its forecasts were in the ballpark.

Of course, maximum winds are of great interest when a hurricane makes landfall. Since MOS does not predict the actual maximum sustained wind for a period of time, a forecaster looking at MOS's forecasts and wanting to estimate a maximum sustained wind might have come up with something around 47 knots just by adding a few knots to the highest forecast at the synoptic times. As it turns out, such a forecast would have been 10 knots too low, as the maximum sustained wind measured at KILM from Florence was 57 knots (which was not captured by any METAR since it occurred in between observations).

But, hurricanes have significant variations in wind speeds over short time periods (much more so than an "average" day), so one could credibly argue that the "add a few knots" approach to MOS's forecasts wouldn't be sufficient here. A larger adjustment might be needed to estimate a maximum sustained wind speed. But, if MOS had a special predictand that forecasted maximum sustained wind speed in a 24-hour period, then it would have done even better with Hurricane Florence. So, even though conditions departed greatly from climatology, the linear connection between increased pressure gradient force and observed wind speed held up pretty well, even in a hurricane! By the way, the maximum wind gust at KILM from Florence was 91 knots, but MOS does not predict gusts, so it wasn't much help in that department.

Meteorologically, why was the maximum sustained wind so "low" at Wilmington? Actually, this scenario is fairly common: The maximum sustained wind speed at a given city can be surprisingly low in the face of a landfalling hurricane, unless, of course, the eye wall passes over the forecast area. Keep in mind that the eye wall doesn't cover very much territory, so only a relatively small strip of land is subjected to the fiercest winds of a hurricane.

Of course, GFS MOS performed fairly well here in part because the dynamic model provided reasonably good forecasts of storm track, timing, and intensity near the time of landfall. In cases where dynamic model forecasts show large uncertainty (a large spread in track or intensity forecasts), or in cases where the dynamic models typically struggle (rapid strengthening or rapid weakening scenarios), it's likely that MOS's forecasts will be highly questionable, too.

ECMWF ensemble track forecasts for Hurricane Florence, initialized at 12Z on September 13, 2018

ECMWF ensemble track forecasts showed relatively little spread in Florence's track leading up to landfall (purple oval). The purple star marks Wilmington, North Carolina.
weathernerds.org

By the time the 06Z GFS MOS forecast was issued (within about 30 hours of landfall), the uncertainty in Florence's forecast up to landfall was rather small, as is often the case in the last day or so leading up to landfall. Indeed, the clustering of tracks within the purple oval in the image above showed high confidence that the center of Florence would likely pass just south of Wilmington (marked by the purple star), which is what happened (check out the radar loop of Florence's landfall (opens in a new window)). Ensemble model forecasts showing Florence slightly weakening (opens in a new window) leading up to landfall. Since the dynamic models were handling Florence's forecast pretty well (no major changes in intensity and relatively high confidence in the track forecast in the 24-hours before landfall), the GFS MOS forecast performed reasonably well. But, how does MOS handle situations where uncertainty is larger?

How Does MOS Handle Uncertainty?

Truth be told, MOS doesn't really "know" anything about uncertainty, but yet uncertainty is factored into MOS forecasts, in a sense. How? Well, MOS always predicts the mean expected event out of a range of possibilities because the mean will minimize forecast error over the long haul. For example, let's assume that, on a given night at a given city, the sky coverage forecast is extremely uncertain, and the range of possible outcomes for cloud cover spans from overcast to clear. If cloud cover persists, MOS might predict a low of 30 degrees. If the sky clears, however, MOS might predict a (record) low of 18. Given this range of possible outcomes, MOS would "officially" predict 24. In reality, MOS doesn't actually "think" quite this way, but I hope you get the general idea. If you're interested in investigating the details of predicting the mean expected event from a range of possibilities, check out the Explore Further section below.

So, what does this all mean? The fact that MOS predicts the mean expected event makes it hard for MOS to predict record-breaking weather if the forecast uncertainty is large (thus making for a large range of possible outcomes), because the mean expected event will naturally draw MOS's forecasts away from the extremes a bit. MOS takes the "middle of the road," making it hard for MOS to perform well when the weather actually nears or exceeds climatological extremes and forecast uncertainty is large. But, it is possible for MOS to predict record-breaking weather. When uncertainty in the forecast is smaller, the range of possible outcomes is smaller. That gives MOS a better chance to perform well even when record-breaking weather is on the horizon.

Points to remember...

  • As a general rule, MOS is more likely to struggle when the weather greatly departs from climatology, but MOS will not always struggle in these situations.
  • When forecast uncertainty is small, the regression equations can be linearly extrapolated near the climatological extremes (as is often the case with temperature and wind speed), and the dynamic models provide MOS with reasonably good forecasts for MOS predictors, then MOS will likely perform fairly well, even though the weather is near climatological extremes.
  • When forecast uncertainty is large and the weather will be near climatological extremes, MOS forecasts are likely to be highly questionable. The fact that MOS predicts the mean expected event will tend to drag MOS's forecasts away from the extremes in these cases.
  • When the regression equations cannot be linearly extrapolated near climatological extremes (heavy precipitation events, for example), MOS forecasts are more likely to struggle.
  • Don't forget to always connect MOS's forecasts to the big picture and climatology so that you can assess the degree of uncertainty in a forecast, as well as whether the situation could truly be record breaking (or nearly so).

Okay, we've covered a lot of details about how MOS really works. Try to keep the "points to remember" and key philosophies about MOS that you've seen throughout the lesson in mind as you go on throughout the rest of the course. As a forecaster, MOS is a very useful tool for getting short-range forecasts (out to a few days) in the ballpark, but sometimes it's helpful to have forecast data available at intervals shorter than three hours (as is the case in standard MOS tables). As it turns out, there's a MOS-based product for that! Read on.

Explore Further...

Forecasting the Mean Event: A Range of Possibilities

How does MOS know that there is a range of possibilities? In the literal sense, MOS doesn't really "know" anything, of course -- it's just a set of regression equations. For starters, MDL developers used a least-squares regression scheme to create MOS. What does that mean? In a nutshell, the coefficients in the regression equations were chosen so that the sum of the squared differences between the forecasts and the observations was minimized (in other words, the sum of the squared errors between forecasts and observations).

As it turns out, the mean of a population of observations minimizes the sum of the squared differences between the mean forecast and each of the observations. So, if the regression scheme minimizes the squared error between forecasts and observations, then, by a roundabout argument, the value that MOS predicts is the number MOS "thinks" will be the mean forecast in a range of possibilities.

Let's look at a simple example. Consider the following numbers: 0, 1, 0, 0, 1

Just think of them as binary observations of precipitation.  Did precipitation occur? (1=yes and 0=no)

The mean of these numbers is 0.40. If we sum the squared differences between 0.40 and each of the numbers, we get:

(0.4 - 0)2 + (0.4 - 1)2 + (0.4 - 0)2 + (0.4 - 0)2 + (0.4 - 1)2 = 1.2

For the record, there isn't any other number, except 0.40 (the mean), that produces a lower sum of the squared differences (1.2). Feel free to try other values for yourself.

The bottom line here is that MOS uses a regression scheme that minimizes the sum of the squared errors. In order to minimize this sum, MOS must choose the expected mean forecast. In the context of precipitation in my simple example, 0.40 corresponds to 40%, which just happens to be the probability of precipitation.

Back to the question at hand. How does "MOS" know that there's a range of possibilities? Well, the folks at MDL who created MOS developed a simple but elegant approach: They use binary predictors. For example, in GFS MOS, if the mean relative humidity between 1000 mb and 500 mb is less than 70%, the binary predictor for clouds = 0. On the flip side, if the mean relative humidity between 1000 mb and 500 mb is greater than 70%, the binary predictor = 1.

If that's the case, are there only two possible outcomes at a given city? Well, not exactly. You see, MDL also incorporates predictors from surrounding grid points. Let's suppose that there's a front just to the west of the city. MOS will pick up on its presence at some of the neighboring grid points and allow this associated uncertainty to enter into the forecast at the city. In other words, there can be a wider range of possible outcomes.

MDL developers also used time-offset predictors. For example, suppose we wanted to forecast the 36- to 48-hour PoP. Developers at MDL might have included the 30-hour and 60-hour mean-layer relative humidity forecasts as predictors. So, if precipitation is likely to occur just before or after the forecast period, then MOS will "know" to bump up the PoP forecast slightly, just in case the speed of the dynamic model is too slow or too fast. Thus, the range of possible outcomes gets handled by the way MDL developers code the predictors.

sas405

LAMP Guidance

LAMP Guidance

Prioritize...

On this page, your focus should be on the utility of LAMP guidance in forecasting (and when it has an advantage over GFS MOS, on average).

Read...

MOS guidance is very useful for providing forecasts that are typically "in the ballpark," but the forecasts in MOS tables are only available in three-hour increments, and MOS itself is only run a few times a day (coinciding with the schedule that the dynamic models are run). In order to "fill in the gaps" and provide more frequently updated guidance, the Meteorological Development Lab created another MOS-based forecasting tool, called LAMP. Without further delay, let's shed some light on LAMP.

Shedding Light on LAMP

When making short-term forecasts, there's a tool that outperforms GFS MOS over the long haul -- the Local Aviation MOS Program (opens in a new window) (LAMP), which is meant to improve on GFS MOS in the first 12 hours of the forecast period. After that, LAMP and GFS MOS have basically the same forecasting skill (see the skill plot below). So, if your forecast interests include details of the next 12 hours, LAMP guidance is worth consulting.

Plot comparing the skill of GFS MOS and LAMP guidance

A plot of forecasting skill (vertical axis) versus forecast time (in hours along the horizontal axis) between LAMP, GFS MOS, and persistence. Over the first 12 hours of the forecast, LAMP tends to have more skill than GFS MOS.
Credit: NOAA

The idea behind LAMP output is simple. Hourly observations are used to "update" GFS MOS output and help improve it over the short term. Since current observations hold less predictive value as we go out further in time, the improved skill of LAMP guidance compared to GFS MOS really starts to decline after 12 hours (and is basically gone after 20 hours).

LAMP guidance is produced every hour according to this LAMP update schedule (opens in a new window). So, for example, the 12Z GFS MOS run is used to generate the 16Z through 21Z LAMP runs, while the 18Z GFS MOS is used to generate the 22Z through 03Z LAMP runs. One important point about the format of LAMP guidance is that specific predictions for maximum and minimum temperature aren't present (unlike standard MOS output). However, LAMP guidance gives you hourly predictions (LAMP for KUNV) (opens in a new window), which is a better temporal resolution than MOS's predictions for temperatures every three hours.

LAMP was originally developed in the late 1980s to aid in aviation forecasting (hence the name "Local Aviation MOS Program"), and output was available for individual stations. Just like with MOS, these forecasts were not spatially continuous (they were developed and created individually for specific stations). But, obviously, not all weather that impacts aviation occurs right at a particular airport, so in order to create a spatially continuous version of LAMP, MDL created "gridded LAMP" (or GLMP) which essentially takes LAMP forecasts, interpolates them (adjusting for elevation, etc.), and applies them to a model grid.

The bottom line is that LAMP (and gridded LAMP) provides forecasters with typically reliable hourly forecasts of critical weather variables, including probabilistic forecasts for various thresholds of horizontal visibility and cloud ceiling height (which are of great interest to aviators). For example, check out the one-hour probabilistic forecast for ceilings less than 1,000 feet valid at 19Z on May 18, 2022 (below).

One-hour GLMP forecast for the probability of ceiling height less than 1,000 feet at 19Z on May 18, 2022.

The one-hour GLMP probabilistic forecast for ceiling heights of less than 1,000 feet at 19Z on May 18, 2022 showed that the most likely area in the contiguous U.S. to have such low ceilings was in northern Illinois, near Chicago.
Credit: NOAA MDL

Indeed, this gridded LAMP forecast showed that areas of northern Illinois, near Chicago were particularly likely to have ceiling heights below 1,000 feet at 19Z on May 18, 2022. The regional surface observations (opens in a new window) at that time showed a number of stations with such low ceilings (marked by the red station models, indicating Instrument Flight Rule conditions (opens in a new window) with ceilings less than 1,000 feet).

Ultimately, if you understand how MOS works based on what we've studied in this lesson, you should have a pretty good idea of how LAMP works, but if you want to get into the nitty-gritty details, check out this presentation about LAMP (opens in a new window). This LAMP Documentation site (opens in a new window) has a slew of informative links, too, if you're interested. To access and decode LAMP forecasts, I recommend the links on the box below (there are some similarities to MOS, but some unique categories, too).

Key Data Resources

LAMP Guidance

The folks at the Meteorological Development Lab have dedicated tremendous efforts over decades to develop statistical tools using the "MOS" approach of linear regression. But, these tools rely on forecast output from the dynamic models, and in an era of increasing computer power and frequent model changes, that can cause problems. We'll examine the reasons why in the next section.

sas405

Case Study: Model Changes and MOS Impacts

Case Study: Model Changes and MOS Impacts

Prioritize...

On this page, focus on how changes to dynamic models can impact MOS, and discuss why "hybrid" MOS systems are not ideal.

Case Study...

By now, you know that MOS relies on a linear regression approach that connects dynamic model forecasts to historical weather observations. When this approach was developed (in the 1970s), dynamic models were relatively static, meaning that they were rarely updated or changed significantly (by increasing resolution, improving parameterizations, etc.). The available computing resources put severe limits on what upgrades modelers could implement. But, as computing power has grown exponentially, modelers are more frequently able to make significant changes to the dynamic models. Nowadays, it's not uncommon to see dynamic models receive minor updates multiple times per year and major updates every couple of years (if not more frequently).

This reality creates a big challenge for MOS developers. After all, if you develop regression equations based on data from a specific configuration of a dynamic model, and then change the model (causing it to behave differently), the old regression equations shouldn't "fit" the data from the new dynamic model quite as well. So, meteorologists at the Meteorological Development Lab (MDL) must evaluate MOS's performance with the upgraded model to see if they need to develop new regression equations. Remember that MOS has more than 10 million regression equations since each site has a set of unique equations, so the decision to develop new equations can't be taken lightly (it's time consuming, tedious, and requires a lot of data). When model upgrades came infrequently, this wasn't a big deal, but with the ever quickening pace of model upgrades, developers must assess whether the dynamic model changes will hurt MOS performance enough to warrant redeveloping the regression equations.

Let's look at an example of a major upgrade to the GFS that occurred in 2019 to see what the impacts were for GFS MOS.

From GFS to GFS-FV3

On June 12, 2019, a new version of NCEP's GFS model became operational -- the GFS-FV3. Prior to this time, the GFS was a spectral model. As you may recall from previous studies, spectral models describe the present and future states of the atmosphere by solving mathematical equations whose graphical solutions consist of a series of waves. Spectral models are not based on grids (opens in a new window), like other grid-point models. But, with its 2019 upgrade, the era of the GFS as a spectral model came to an end, as the GFS-FV3 was implemented. The "FV3" in the name stands for "Finite-Volume Cubed Sphere." This dynamic core for the model employs a flexible grid scheme, which can be "stretched" to provide higher resolution forecasts in certain areas of the world where forecasters want to see more detail (at the expense of resolution in other areas).

The change to the FV3 dynamic model core brought several advantages: The FV3 can be run faster, using less computing resources because of its higher calculation efficiency, it provides better simulations of small-scale vertical air motions, and better depictions of weather features with irregular shapes and / or large gradients (which were not simulated well by spectral models).

Schematic showing variable grid spacing in the FV3.

The FV3 dynamic core allows for variable grid spacing, a major change from the GFS model's spectral roots.
Credit: Geophysical Fluid Dynamics Laboratory

The change to the FV3 dynamic model core represented the most significant change to the GFS model since its inception, and such a change would impact the model's behavior and performance. This change would seem like an ideal time to redevelop the GFS MOS equations, but with further significant model upgrades planned for the few years after the 2019 upgrade (such as increased vertical resolution, a change in cloud microphysics parameterization scheme, and changes to the model data initialization scheme), MOS developers had to decide whether redeveloping the equations in 2019 was "worth it" only to potentially have to update them again (perhaps multiple times) fairly quickly.

Developers ran experimental tests using the existing GFS MOS equations (which had been developed for the spectral version of the model implemented in 2015) and forecast output from the new FV3 version of the GFS and determined that while GFS MOS performance was degraded somewhat, its performance was still "good enough" to get by. So, with the 2019 GFS dynamic model upgrade, GFS MOS became a "hybrid" MOS system, in which the equations were developed for a previous version of the dynamic model. Such hybrid MOS systems are not ideal (MOS performance is not optimized), but that's not to say that their forecasts are necessarily inaccurate. In 2023, MDL finally upgraded the GFS MOS equations (opens in a new window) to base them on the updated version of the dynamic model. Ultimately, the 2019 GFS upgrade wasn't the first time a hybrid MOS system existed (MDL had used similar approaches with past major model changes, resulting in minor degradation in forecast quality), as developers have to weigh the time and expense of developing new equations right away pending planned future model changes.

Ultimately, modern-day weather forecasters can't get too comfortable with the behavior of a particular model or MOS, because the models are regularly being upgraded as computer power increases, and those changes can impact MOS performance. While most changes end up having largely "invisible" impacts, periods of time do exist (sometimes for a few years) where MOS performance is not optimized because the regression equations do not match the current version of the dynamic model. So, it's always wise to be on your toes! By the way, if you're interested in keeping tabs on changes to NCEP products, you can follow the product change notices (opens in a new window). These notices can be useful because they provide forecasters with a heads-up about what model changes are occurring and what the expected impacts are.

Clearly, however, the rapid pace of dynamic model upgrades that is common today presents a real challenge for the linear regression approach employed by MOS. Another reliable statistically post-processed tool produced by MDL does not rely on linear regression, however. Up next, we'll turn our attention to the National Blend of Models (NBM).

sas405

The National Blend of Models

The National Blend of Models

Prioritize...

When you've completed this page, you should be able to describe the National Blend of Models (NBM), and be able to extract information about basic weather elements, such as temperature, dew point, sky coverage, wind directions, sustained wind speeds and gusts, PoP, and QPF from an NBS table. You should also be able to interpret box-and-whisker plots of NBM forecasts.

Read...

In the wake of Hurricane Sandy in 2012, the United States Congress mandated that the National Weather Service adopt a national blended model approach to forecasting, which became operational in 2016 as the National Blend of Models (NBM). The NBM is essentially a huge ensemble of model guidance (including dynamic models, AI models, and MOS) that is blended together to create reliable forecasts for temperature, wind, sky cover, precipitation, ceiling, visibility, etc. But, as I've alluded to before, the NBM is more than just a huge "superensemble" of model guidance.

Recall that all of the dynamic model, AI model, and MOS forecasts in the NBM are first bias corrected, which means that their forecasts are adjusted based on their recent errors. In the case of the NBM, the average error for each model over the past 60 days is calculated and adjusted for. If a particular model, for example, has been 3 degrees Fahrenheit too low on average over the past 60 days, we would add 3 degrees to the model's forecast for tomorrow to correct for the recent bias. The accuracy of each bias-corrected ensemble member over the past 60 days is then used to determine its weighting in creating a final consensus forecast. Those members that have performed better in the last 60 days are weighted more heavily, while those members with less accurate forecasts over the last 60 days are weighted less in the final consensus forecast. The flow chart below may help you visualize the process of how NBM forecasts are created.

Flowchart showing the development of NBM forecasts

National Model Blend forecasts are created by bias-correcting a variety of dynamic model, AI model, and gridded MOS forecasts, and creating a consensus based on recent model performance.
Credit: National Weather Service

This approach to bias correction marks a major contrast to the MOS approach to bias correction, which relies on linear regression. MOS takes a "static" approach to bias correction (the regression equations do not change until they are redeveloped after a significant model upgrade), but the NBM's formula for bias correction is always changing depending on recent model performance. In an era of frequent model upgrades, the NBM's approach to bias correction has a big advantage. Maintaining the NBM is simpler than maintaining MOS. With every dynamic or AI model upgrade, the new features of the models are already "built in" to the NBM system, and the bias corrections are relatively simple to calculate because they're determined over the most recent 60 day period. Furthermore, NBM forecasts are spatially continuous. They aren't specifically developed for single stations (like MOS).

The saying "there's power in numbers" comes to mind with regard to the NBM. While the exact components of the blend change a bit with every upgrade, generally speaking, the NBM includes forecasts from just about every major global dynamic and AI model (and their ensembles), along with, MOS, LAMP, some high-resolution models like the RRFS and HRRR, and more. To get an idea of the full scope of the number of members, here's the list for version 4.1 of the NBM (opens in a new window). It's a big list! Even ignoring the models specifically for maritime / wave forecasting, there's more than 200 members blended together to create NBM forecasts!

NBM forecasts are updated every hour (although not all of its components are "new" each hour), and its forecasts are available in a variety of formats, including tables with hourly forecasts much like LAMP (called the "NBH"), tables with three-hourly forecasts out to a few days, much like MOS (called the "NBS"), as well as extended forecasts going out a week or more (called the "NBE" and "NBX"). These tabular NBM forecasts are available for thousands of stations around the U.S., and they resemble MOS tables in many of their conventions, although some key distinctions exist. Before you leave this section, you should make sure you can decipher major weather elements like temperature, dew point, wind directions, sustained wind speeds and gusts, PoP, and QPF from an NBS table (see the Key Skill box below).

Finally, because the NBM is a superensemble, it predicts a range of outcomes, so statistics like standard deviation are readily available, and there's more probabilistic information. For example, check out the KHOU NBM Table — standard deviations of temperature forecasts ("TSD") (opens in a new window) at Hobby Airport in Houston in the NBS table from 13Z on February 2, 2022. Standard deviations of 4 to 5 degrees Fahrenheit for less than a 24-hour forecast indicate a wide range of solutions in the blend members, indicative of a very uncertain short-range forecast. For the record, standard deviations of three degrees or less are pretty common for short-range forecasts. If you see larger standard deviations, that's a sign of a larger spread in NBM members, and more uncertainty than is typical for a short range forecast.

In addition to tabular formats, forecasters also use tools like percentiles and box-and-whisker plots to help them visualize the spread of bias-corrected NBM forecasts and gauge uncertainty, as in the example below showing maximum and minimum temperature forecasts for University Park, Pennsylvania, from the 13Z run of the NBM on April 30, 2021.

Box-and-whisker forecast plot for maximum and minimum temperatures from the NBM.

A box-and-whisker plot of maximum and minimum temperature forecasts from the NBM run at 13Z on April 30, 2021. The tops and bottoms of the boxes represent the 75th and 25th percentile solutions, respectively, while the tops and bottoms of the "whiskers" extending from the boxes represent the 90th and 10th percentile solutions, respectively. The black dot represents the "blended" deterministic forecast. If you need a closer look, here's a larger version (opens in a new window).
Credit: National Weather Service

The box-and-whisker plot above shows maximum and minimum temperature forecasts for University Park, Pennsylvania, from the 13Z run of the NBM on April 30, 2021. The "blended" deterministic forecasts are marked by the black dots, while the boxes represent the range of solutions between the 25th and 75th percentiles. The "whiskers" that extend from each box end at the 10th percentile solution (bottom whisker) and the 90th percentile solution (top whisker). So, with this one plot, we can see the NBM's bias-corrected, weighted, "blended" deterministic forecast and gauge the spread of the bias-corrected solutions. Note that on some days, the deterministic forecast doesn't fall in the middle of the spread because of the weighting scheme used (favoring models that have performed better recently in the calculation).

The suite of NBM products accessible to the general public is continuing to expand, and I've collected some links for NBM forecasts in the Key Data Resources box below.

Key Data Resources

You'll need to be able to extract basic information from an NBS table, so make sure to check out the Key Skill section below, and if you are interested in further reading about the details of the NBM, check out the links I've gathered in the Explore Further section below. While the information in this section should give you a basic idea of how the NBM works, in order to understand better how to use its products, we have to dig a little deeper into how it does its bias correction. We'll tackle that in the next lesson.

Key Skill...

Decoding NBM Tables

NBM tables come in a variety of formats, as mentioned above (NBH, NBS, NBE, and NBX), but we're going to focus on NBS tables since they're the most applicable to forecasting on a one- to three-day time scale. The available forecast variables and their format changes a bit with each new NBM upgrade, so I recommend that you check out the current NBM Text Bulletin Key (or "NBM Card") available on the text product information page (opens in a new window). Use this key to become familiar with at least the top half of the NBS table, which includes forecasts for major weather variables (temperatures, sky coverage, wind direction and speeds, and precipitation). The rest of the table includes a variety of other variables that may be of interest, but we won't be using them regularly in class.

By using the text bulletin key for NBS tables, you should be able to answer the questions below, based on the 13Z NBS table for Raleigh, North Carolina (KRDU) on May 23, 2022.

KRDU NBS Table from the 12Z run on May 23, 2022.

Complete NBS table for Raleigh, North Carolina (KRDU) from the 13Z run on May 23, 2022.
Credit: NOAA MDL

Explore Further...

More on the NBM

While you should have a good general idea of how NBM forecasts are created, I glossed over a lot of the statistical details which are outside the scope of this course. If you have interest in learning more of the details of the NBM, I recommend the links below:

sas405

Bias Correcting the NBM

Bias Correcting the NBM

Prioritize...

When you've completed this page, you should be able to discuss the reliance of NBM performance on the Unrestricted Mesoscale Analysis (URMA), as well as how errors in the URMA can impact the quality of NBM forecasts.

Read...

One important advantage of the NBM is that it provides spatially continuous, bias-corrected forecasts. But, how does this bias correction occur? Most of the land area within the United States (or any nation, for that matter) is not sampled directly by the surface observing network. How can we get a bias-corrected forecast for locations where there are no observations to use as "ground truth" in checking for model errors?

Locations with METAR observations in the western U.S.

In many areas, METAR observations are sparsely located.
Credit: University of Wyoming

Indeed, huge swaths of land are uncovered by surface observations. The image on the right, for example, shows the locations with METAR observations in the western U.S. While some urban areas have numerous observation sites, more rural areas may have hundreds of miles in between observation sites.

Fortunately, the network of stations providing METAR observations isn't the only network of observations available. As you may recall from previous studies, some states operate their own mesonets, some state departments of transportation operate observation networks associated with major roadways, and other networks such as the Remote Automatic Weather Station (RAWS) Network (opens in a new window), and the Citizen Weather Observer Program (CWOP) (opens in a new window) help fill in the gaps. When including some of these other observational networks, the gaps are smaller (check out this map of Nevada surface observation sites (opens in a new window) and compare to the METAR locations shown on the right above), but considerable gaps still exist.

In order to create a spatially-continuous analysis of weather variables, NCEP created the Real-Time Mesoscale Analysis (RTMA), which includes surface observations from various networks and fills in the remaining gaps with a blend of model analyses from mesoscale models. So, not everything in this "analysis" truly comes from a real observation. Data from areas with no observations is actually coming from mesoscale models. The end result is a high-resolution analysis of weather data, such as the RTMA 2-meter temperature analysis from 18Z on May 26, 2022 below. The striking curlicue of relatively cool air from the Upper Midwest to the Central Plains coincided with low clouds (opens in a new window) in the cold conveyor belt of a well-developed mid-latitude cyclone.

18Z RTMA 2-m temperature analysis on May 26, 2022.

RTMA 2-meter temperature analysis at 18Z on May 26, 2022.
Credit: Pivotal Weather

In order to assess the overall quality of such analyses, researchers have used a process called "data denial," which involves removing actual observations from the analysis to see if the analysis method can "recreate them" accurately. Such studies show that the RTMA does a very good job overall with variables like temperature and pressure, but not quite as well with other variables like visibility, ceiling, or winds.

To improve the quality of the analysis further, another version is created six hours later, which incorporates the most actual surface observations possible (some arrive too late to be included in the real-time analysis). This refined product is called the Unrestricted Mesoscale Analysis (URMA), which serves as the version of the "ground truth" that the NBM is bias corrected against. The spatially-continuous nature of the URMA allows NBM bias-corrected forecasts to also be spatially continuous, but the analysis isn't entirely comprised of observations, and it isn't always perfect. The mesoscale model data used to fill in the gaps isn't perfect, nor are all of the actual surface observations included (instruments occasionally take inaccurate measurements). Usually these errors are small and insignificant, but sometimes they are not!

To see what I mean, consider this example. In May, 2018, forecasters at the National Weather Service Forecast Office in Huntsville, Alabama noticed a persistent warm bias in the NBM forecasts for a portion of the Tennessee River Valley (opens in a new window) that runs through northeastern Alabama. Repeatedly, forecasts for that area were a few to several degrees too warm (highlighted in the box below).

NBM forecast with erroneously high temperature predicted for the Tennessee Valley

NBM forecast from 17Z on May 21, 2018 with erroneous warmth in parts of the Tennessee Valley marked in the black box.
Credit: National Weather Service / Brian Carcione

Indeed, on some days, the NBM predicted the area to be the hottest in the entire region, like in this NBM run from 23Z on May 14, 2018 (opens in a new window), which raised red flags. From climatology, forecasters knew that the area in question is usually warmer than its immediate surroundings, but not by as much as what the NBM was consistently predicting. The source of the problem, however, wasn't the many members in the NBM superensemble. The problem was in the analysis used for bias correction. A faulty thermometer near the Alabama / Tennessee border kept reporting temperatures that were too high. Those observations caused the URMA to "think" it was warmer in the area than it actually was, which meant that the NBM was bias correcting against erroneous warmth. So, the bias correction scheme kept trying to adjust the NBM forecasts to make them warmer than they actually needed to be.

In this particular case, very observant forecasters applied their knowledge of local climatology to notice a problem with the NBM (caused by the URMA), which they reported to NCEP, and it was eventually fixed. This case illustrates a key lesson about the NBM: Errors or biases in NBM forecasts are a "moving target" of sorts, so regularly using NBM forecasts (and assessing them) can help forecasters pick up on situations where its forecasts might be erroneous (although these situations may not last for long periods of time). Had forecasters not been closely watching its forecasts on a daily basis, they likely would not have noticed or investigated the warm bias in the Tennessee Valley, and it would have persisted for a longer period of time (as long as the faulty thermometer kept recording temperatures that were too high).

Lessons Learned

Keep these lessons in mind when using NBM forecasts:

  • NBM forecast errors can result from either model performance (or the blending scheme) or errors in the analysis used for bias correction.
  • While NBM forecasts are reliable overall, their errors or biases can be a "moving target" because the bias-correction scheme is regularly updating based on recent model performance compared to the URMA.
  • Regularly using and assessing NBM forecasts, along with local knowledge of climatology and the current weather pattern, can help forecasters gauge NBM performance; however, particular errors or biases may only be present for short periods of time.

Bias correcting against the URMA can cause additional complications for other forecast variables, much more so than temperatures. In the early versions of the NBM, its wind speed forecasts were consistently much too low. Why? Recall that standard "surface" wind measurements are actually taken at an altitude of 10 meters. But, the number of actual wind observations in the URMA that are taken at a standard height of 10 meters (such as at ASOS sites at airports), is relatively small compared to the number of observations taken from other networks at non-standard heights (usually below 10 meters). For example, most observations in the CWOP network are in urban or suburban areas where local ordinances may prohibit measurements to be taken that high, or measurements may be impacted by nearby buildings or trees.

Because of nearby obstructions and / or increased friction due to measurements being taken closer to the ground, many of these "other" wind observations are slower than those taken at standard height, free of nearby obstructions. Therefore, the wind speeds analyzed by the URMA were too slow overall, causing the NBM to bias correct against speeds that were too slow. So, the bias-correction scheme made NBM wind forecasts consistently too slow. Much work has been done to improve the wind speed forecasts in the NBM, ranging from flagging and omitting wind observations that appear to be too low, to adding an additional layer of statistical post-processing (a MOS-like regression) to "inflate" NBM forecasts and bring them in line with standard observations. As a result, the wind forecasts have significantly improved since the first several versions of the NBM, thanks to a savvy use of statistics.

If you're interested in further reading on the RTMA / URMA or accessing RTMA graphics, I've included a few links of note in the Explore Further section below. Otherwise, we're almost done with our study of statistical models. Let's wrap up the lesson with a few final thoughts and other forecasting applications that may be of interest to you.

Explore Further...

Are you looking for RTMA graphics or do you want to read more about the RTMA / URMA? You may be interested in the links below:

sas405

Final Thoughts and the Future

Final Thoughts and the Future

Prioritize...

This lesson includes a lot of information about statistical models! This section is meant to summarize key points that you can take away from this lesson, as well as touch on other ways that statistical forecasting is increasingly being used in meteorology.

Read...

This lesson was packed with information about statistical models. For general day-to-day forecasting, we explored two commonly-used approaches to developing statistical forecasts -- linear regression (the basis for MOS), and the weighted bias correction of a large superensemble (the basis for the NBM). Both approaches have strengths and limitations, and we'll continue to explore how those apply to forecasting practice in the coming lessons. But, before we end the lesson, I want to emphasize a key message that can sometimes get lost on newer forecasters: Statistical models exist for a reason!

Meteorologists invest considerable time and resources in statistical post-processing because doing so results in superior forecasts for surface weather variables, on average. Statistically post-processed forecast products like MOS and the NBM were designed to improve upon raw model forecasts for surface weather variables. These forecasts aren't perfect, but MOS and the NBM will get you in the ballpark the vast majority of the time. Remember the stats about MOS and NBM accuracy for 24-hour high- and low-temperature forecasts mentioned earlier in the lesson: They're getting within 5 degrees Fahrenheit around 90 percent of the time! Dynamic and AI models, even with their increasing sophistication, can't boast a track record like that. Their forecasts aren't always poor, but they are less consistently "good." Statistical models are built to be consistently in the ballpark by correcting for biases.

120-hour ECMWF forecast for 2-m temperatures valid at 12Z on April 13, 2022.

Very detailed and colorful maps (like this ECMWF 2-m temperature forecast) displaying dynamic or AI model forecasts are readily available online, but for surface weather variables, they aren't the best forecast data, on average.
Credit: Pivotal Weather

With that in mind, here are some forecast tips when using statistically post-processed guidance like MOS and the NBM:

Forecast Tips

  • Statistically post-processed guidance is usually "in the right ballpark" for surface weather variables and often serves as a good "first guess" for deterministic forecast numbers. For short-term forecasts (1-2 days), it's relatively rare that you'll need to make large changes to these forecasts.
  • Always keep your forecast context in mind when interpreting these forecasts, and remember that for some variables (like wind speed), the available forecasts may not directly answer key forecast questions (like what the maximum sustained wind or wind gust is over a period of time).
  • Statistically post-processed guidance isn't perfect, so it's always up to the forecaster to have keen "situational awareness" of current weather and the upcoming weather pattern in order to make decisions about whether to adjust forecast guidance, and by how much (we'll talk more about this in upcoming lessons).
  • All forecasts from statistical models aren't created equal. Forecasts for variables like temperature and wind tend to be more reliable than those for precipitation.
  • As a bias-corrected superensemble, probabilistic products from the NBM can be great resources for thinking probabilistically about a forecast (even if you have to settle on a deterministic forecast).
  • It's generally not a sound forecasting practice to view dynamic or AI model forecasts for surface variables as "equals" to statistically post-processed guidance. On average, the statistically post-processed guidance performs better. In other words, weighing dynamic or AI model guidance (like raw 2-meter temperatures or 10-meter winds from the a single model or an ensemble mean, for example) in a consensus with statistical guidance is not a sound strategy and will tend to make your forecasts worse over time.

As a forecaster, try to leverage the strengths of dynamic/AI models and statistical models. Dynamic/AI models are most useful for assessing the weather pattern (upper-air features, surface features, gradients, vertical motion, etc.). Statistical models are most useful for forecasts of surface weather variables (temperature, dew point, wind, etc.). Both should be part of your forecast routine.

Statistical forecasting methods and bias correction have been important in meteorology since the 1970s with the invention of MOS, and it's likely that their importance will continue even as dynamic and AI models continue to improve. Bias correction of raw model forecast tends to produce superior results! In the future, AI may be able to help with the bias correction, which is just one of the many ways that AI may affect weather prediction. The integration of AI into weather modeling is leading to rapid changes in the field, with applications ranging from general day-to-day weather forecasting to tackling very specific forecast problems (like hurricane track forecasting or whether or not an individual thunderstorm might produce severe weather). If you're interested in reading more about the role of AI in modeling and accessing more statistical forecasting tools (particularly for high-impact weather), check out the Explore Further section below.

In the coming lessons, we'll be examining major controllers of common forecast variables like temperature, wind, and precipitation (we'll be revisiting a number of concepts from your previous course work), and we'll cover some advanced strategies for forecasting these variables using the tools we've already covered.

Explore Further...

Further reading:

Other statistical forecasting tools:

sas405

Lesson 4. Forecasting Temperature: Controllers and Strategies

Lesson 4. Forecasting Temperature: Controllers and Strategies

Motivate...

Photograph of The World's Tallest Thermometer

The "World's Tallest Thermometer" (really just a large electric sign standing 134-feet tall) at Baker, California registered 101 degrees Fahrenheit at 10 PM PDT on June 6, 2010. Even when thermometers aren't this big, the public pays a lot of attention to temperature forecasts.

Recall that the accuracy of dynamic model forecasts of two-meter temperatures is somewhat inconsistent because the nonlinear processes that occur near the Earth's surface (turbulence and radiation fluxes that vary on the microscale, etc.) are too complicated to mathematically model accurately. Instead, they're simulated by imperfect parameterizations that introduce error into the model. That's why forecasters created more reliable statistical forecasts for variables like two-meter temperature, but even temperature forecasts from statistical models aren't perfect. 

Still, despite their imperfections, sticking with statistical forecasts like MOS or the NBM (or a consensus of them) can be the best strategy, as long as you have a sound meteorological reason for doing so. However, sometimes it's wise to tweak these forecasts, and once in a while, forecasters need to make larger adjustments. How do forecasters make decisions about how to treat MOS or NBM forecasts on any given day? In this lesson, we'll explore important factors to consider in making such decisions and, along the way, hopefully help you minimize errors in your temperature forecasts. In the operational forecasting world, the general public certainly notices when temperature forecasts are way off target!

First, it's important to realize that we can't forecast temperature in a vacuum. Temperature (like all forecast variables) is intimately tied to other atmospheric conditions. Think about it. Can we create accurate temperature forecasts if we have no idea what the sky conditions are going to be? Or from what direction the wind might blow? Or whether it's going to rain? Because of the interconnected nature of forecast variables, it stands to reason that the first step in creating a good temperature forecast is to have a solid grasp on the big picture situation every day.

Being "one with the atmosphere" and having a good handle on the day's big-picture setup is only the beginning, however. We also need to keep in mind the major controllers of temperature, so that we can focus in on the variables that will have the greatest impact on our forecast. Thus, in this lesson, we need to address the important controllers of temperature. I've listed the main controllers below, which you've already encountered in previous certificate courses:

  • long-term "climate factors" (a location's latitude, altitude, and proximity to water, etc.)
  • local energy transfer (radiation, convection, and conduction)
  • wind-related controllers (temperature advection; orographic effects)
  • "latent-heat effects" (related to phase changes of water, like evaporational cooling)

If you want to review complete treatments of these topics, check out the appropriate sections in METEO 101 (opens in a new window) (login required). In this lesson, however, we'll hit the high points of these topics again just to refresh your memory and introduce specific applications to making temperature forecasts. We'll also go into more depth regarding weather patterns that often cause MOS or NBM forecasts to struggle, and cover a few advanced techniques for using dynamic model output to help you make better forecasts. Let's get started!

sas405

Climate Factors

Climate Factors

Prioritize...

After completing this page, make sure that you can describe the effects of latitude, altitude, and proximity to large bodies of water on large-scale, seasonal temperature trends.

Read...

In the world of buying and selling real estate, a common saying is that everything boils down to "location, location, location." Over the long haul, the same can be said for temperatures. A particular town's location on the planet Earth helps determine its long-term temperature ranges. Particularly, a location's latitude, altitude, and proximity to water are major factors that impact annual temperature trends.

Latitude

You may recall that the intensity of incoming solar radiation at any given point is determined by the angle at which it strikes the Earth. More direct angles lead to more intense incoming solar radiation, while increasingly oblique angles lead to reduced intensity. These side-by-side images of flashlights aimed at a wall (opens in a new window) (credit: David Babb) give a visual depiction of this idea. So, what factors determine the angle that incoming solar radiation strikes the surface? The tilt of the Earth's axis of rotation (opens in a new window) and a given location's latitude, of course. The tilt of Earth's axis is the underlying reason for the existence of seasonal temperature variations as the planet revolves around the sun, and the latitude of a given town helps determine the details of the local seasonal variations.

As latitude increases, the angle that incoming solar radiation strikes the earth becomes increasingly oblique, resulting in less solar heating power. Let's examine the implications of this at two cities with roughly the same altitude, both located near the middle of the United States. Using NOAA's solar calculator (opens in a new window), I ran some numbers for Bismarck, North Dakota (opens in a new window) (46.8 degrees north latitude) and Oklahoma City, Oklahoma (opens in a new window) (35.5 degrees north latitude).

On December 21 at local noon (near the time of the winter solstice), the sun's radiation strikes the surface at an angle of roughly 19.8 degrees at Bismarck, while at Oklahoma City, the sun's radiation strikes the surface at an angle of 31.1 degrees (the sun is "higher in the sky" at Oklahoma City). If we perform the same calculations for June 21 at local noon (near the time of the summer solstice), we find that the radiation strikes the earth at these locations at angles of 66.6 degrees (Bismarck), and 78.0 degrees (Oklahoma City).

So, what does this mean in terms of the amount of solar radiation received by each location? We have to invoke some trigonometry (I'll spare you the details), but we can calculate that at Bismarck on December 21, the incoming radiation is only about 34 percent of the maximum possible (often called the "direct beam"). By June 21, however, Bismarck is experiencing around 92 percent of the radiation associated with the direct beam. Oklahoma City, on the other-hand, experiences a range of 52 percent to 98 percent of the sun's direct beam. The net result is that Oklahoma City consistently receives more radiation than Bismarck throughout the year. Note, however, that the higher-latitude city (Bismarck) experiences a greater variation in radiation amounts than does Oklahoma City. These factors lead us to the annual plots of daily mean maximum temperature at both cities, shown below.

Graph of the annual variation of average daily high temperatures at Oklahoma City, OK and Bismarck, ND.

The annual variation of average daily high temperatures at Oklahoma City, OK (green curve) and Bismarck, ND (purple curve). The difference in latitude is the primary factor in the temperature difference between the two cities.
Credit: Constructed from data supplied by the Earth System Research Laboratory

Not surprisingly, Oklahoma City is, on average, warmer than Bismarck (because it receives stronger incoming solar radiation throughout the year thanks to its lower latitude). In addition, we notice that Bismarck has a much wider range in average high temperatures than does Oklahoma City. Average high temperatures in Bismarck increase from about 20 degrees Fahrenheit in January to about 85 degrees Fahrenheit in July (a range of 65 degrees Fahrenheit), while Oklahoma City's average high temperatures increase from about 44 degrees Fahrenheit in January to 95 degrees Fahrenheit in July (a range of 51 degrees Fahrenheit). The larger range at Bismarck results from the fact that there's a larger variation in incoming solar radiation throughout the year at higher latitudes.

Lesson learned: All else being equal, the greater a location's latitude (farther away from the equator), the lower its average high temperatures will be and the more extreme the variation between summer and winter.

Altitude

Now that we understand the impacts of latitude on temperature, let's look at altitude. We know that temperature usually decreases with increasing height in the troposphere. The reason for that typical decrease, as you may recall, is that the primary heat source in the troposphere is the Earth's surface. That's right: The earth's atmosphere is relatively transparent to solar radiation. Only the surface absorbs the lion's share of this radiation. The absorption of solar radiation, of course, warms the ground which then transfers its heat to the atmosphere via conduction, convection, and the emission of long-wave radiation. The end result is that, in the troposphere, the farther away from the Earth's surface you are, the colder the surrounding air.

But then, why are high altitude cities and towns colder than their valley counterparts (all else being equal)? After all, the mountain tops still qualify as the surface of the earth and thus should have similar heating properties as lower elevation surfaces. Indeed, that's true, but remember that as mountain tops warm, the surrounding environmental air remains cooler. This generates the pressure differences that drive mountain-valley circulations (opens in a new window) and high-level heat source convection, which transport the relatively warm air away from the mountain top, keeping it cooler than it would be without such circulations.

As an example, consider two cities — Columbia, Missouri (opens in a new window) and Eagle, Colorado (opens in a new window). Both cities lie at similar latitudes and are located near the middle of the United States. The difference between these two cities is elevation. Columbia's elevation is 705 feet above sea level; Eagle's elevation is 6,600 feet above sea level. Below is a comparison of the mean high temperatures for these two cities.

Graph of the annual variation of average daily high temperatures at Eagle, CO and Columbia, MO.

The annual variation of mean maximum temperatures at Eagle, Colorado (purple curve), and Columbia, Missouri (green curve). Note that Eagle's higher elevation of 6,600 ft causes it to have a lower average high temperature compared to Columbia (located at 705 ft).
Credit: Constructed from data supplied by the Earth System Research Laboratory

As you can see in the graph, Eagle's higher elevation causes it to have a slightly lower average high temperature than the lower-elevation city of Columbia. The two cities, however, have similar annual ranges in temperature because of their similar latitudes.

Lesson learned: All else being equal, over the long haul, as elevation increases average high temperatures are lower.

Proximity to Bodies of Water

You should also recall that bodies of water warm or cool particularly slowly compared to land because water has a much higher heat capacity. In other words, it requires a great deal of energy to change the temperature of water even a fraction of a degree. Therefore, locations that are near (particularly downwind) of large bodies of water tend to have average temperature curves that exhibit smaller ranges between the seasons.

Examine the plot of daily mean high temperatures at San Francisco, California (opens in a new window) and St. Louis, Missouri (opens in a new window) below (both cities lie at approximately the same latitude). At San Francisco, prevailing winds blow in from the ocean most of the time, and you can observe the moderating influence of the Pacific Ocean in San Francisco's annual temperature curve below (note its relative "flatness"). Meanwhile, St. Louis is landlocked, and without the moderating influence of the ocean, it experiences a much larger variation in temperature from season to season.

Graph of the annual variation of average daily high temperatures at San Francisco, CA and St. Louis, MO.

The annual variation of mean maximum temperatures at San Francisco, California (green curve), and St. Louis, Missouri (purple curve). Note that the annual variation at San Francisco is smaller than St. Louis, indicating the moderating effects of the Pacific Ocean.
Credit: Constructed from data supplied by the Earth System Research Laboratory

So, the residents of San Francisco can thank the Pacific Ocean for their relatively mild winters and relatively cool summers (compared to St. Louis). Oceans do not have a monopoly on moderating temperatures, however. To a lesser extent, large lakes, rivers, and seas modify air temperatures, too.

Lesson learned: All else being equal, cities downwind of large bodies of water tend to experience a smaller annual range in temperatures than do landlocked cities.

The Bottom Line

You may not realize it, but you deal with the major factors outlined on this page any time you "get to know" a city by researching its climatology. By studying a station's location, elevation, surrounding geography, etc., you gain an understanding of how latitude, altitude, and proximity to bodies of water may impact temperatures at your city. Of course, surrounding topography is important, too, which we'll talk a bit more about later. But, for now, let's move on to a major temperature controller on shorter time-scales — clouds.

sas405

Clouds and Temperature Forecasting

Clouds and Temperature Forecasting

Prioritize...

When you've finished this section, you should be able to describe the sky coverage factors that play important roles in both daytime and nighttime temperature forecasts, and apply these principles to your own weather forecasts.

Read...

I'm sure it's obvious to you that, on any given day, the presence (or lack) of clouds has important impacts on temperature. That's because clouds affect the local radiation budget. If the ground absorbs more radiation than it emits, ground temperature increases. But, if the ground emits more radiation than it absorbs, ground temperature decreases. During the day, clouds act as a "natural sun block" by reducing the amount of incoming solar radiation. Clouds are like "double agents," though, in that they also emit their own infrared radiation. This fact is particularly important at night because this downwelling infrared radiation helps keep the Earth's surface warmer than it otherwise would be. If you want to review the details of radiation budgets, please review Lesson 4 of METEO 101 (opens in a new window) (login required) ...and remember, clouds don't act like blankets!

Seems simple enough, right? Those general principles are good enough for making qualitative comments like "all else being equal, a cloudy night will be warmer than a clear night," but a purely qualitative approach only goes so far in coming up with specific temperature forecasts. Simply describing the sky as "cloudy" doesn't cut it. What kinds of clouds are they? What's the overall cloud coverage? How thick are the clouds (qualitatively)? Are they high, middle, or low clouds? These are all questions that operational weather forecasters must consider when making temperature forecasts.

Daytime Concerns

Sun shining through cirrostratus clouds

Cirrostratus clouds let much more solar radiation pass through to the surface compared to more optically thick clouds, like stratus.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

During the daytime, the amount of incoming solar radiation is a major temperature controller, so we're primarily concerned with how much incoming solar radiation will be reduced by cloud cover. Which type of cloud will cause a bigger reduction in solar radiation that reaches the surface — a low overcast of stratus clouds (opens in a new window), or a high overcast of cirrostratus clouds (like those on the right)? Obviously, the low overcast of stratus clouds will block more incoming solar radiation and suppress daytime temperatures more than a thinner veil of cirrostratus clouds. That's because the deck of low stratus clouds has a greater "optical thickness" than the cirrostratus clouds.

The formal definition of optical thickness (opens in a new window) is mathematically complicated, but you can think of optical thickness as a way to assess of the amount of radiation that passes through an object. Clouds with a greater optical thickness allow less solar radiation to pass through to the surface (more of the radiation is absorbed or scattered by the cloud). Meanwhile, clouds that have a lower optical thickness allow more solar radiation to pass through to the surface (because they absorb and scatter less). 

Several factors help determine the optical thickness of a cloud, such as the composition of the cloud (water droplets versus ice crystals, etc.), the concentration of particles in the cloud, and its vertical depth. So, if we can accurately predict the cloud types that will exist, we'll be able to better fine tune our temperature forecasts. In general, clouds that look thicker to us on the ground (low stratus or nimbostratus, towering cumulus, cumulonimbus, etc.) are "optically thick." Meanwhile, clouds that look thinner to us (cirrus, altostratus, etc.) tend to be relatively "optically thin." Ultimately, for daytime temperature forecasting, our primary concerns are about the optical thickness of the clouds present and the sky coverage (portion of the sky covered by clouds).

Nighttime Concerns

At night, we're mostly concerned with the amount of downwelling radiation emitted by clouds to the surface. What kinds of clouds emit the most downwelling radiation? "Warm" clouds, of course. Recall that the warmer an object, the more radiation it emits. Since temperature typically decreases with height in the troposphere, low clouds (being the warmest) are the most notable emitters of downwelling infrared radiation, and thus, have the greatest potential to keep nighttime temperatures elevated. Meanwhile, colder clouds (cirrus, for example) emit less downwelling radiation. So, at nighttime, our primary concerns are the altitudes of the clouds present and the sky coverage.

Forecasting Tools

What tools do we have our disposal to help us get a bit more specific with our cloud forecasts? Here are some useful tools that can help. You'll need to explore the menus on each site to find the products you're looking for:

  • MOS tables: Not only does MOS predict sky coverage (categorically), but it also predicts cloud ceiling and obstructions to visibility at three-hour intervals.
  • NBM tables: NBM also predicts sky coverage (as a percent cloud cover), as well as cloud ceiling and obstructions to visibility.
  • Forecast soundings: By looking for layers predicted to have high relative humidity, we can spot potential layers of clouds and assess their approximate heights and thicknesses. Some good sites for forecast soundings are the Air Resources Forecast Laboratory (opens in a new window), Pivotal Weather (opens in a new window), Tropical Tidbits (opens in a new window) and the College of DuPage (opens in a new window) (for Pivotal Weather, Tropical Tidbits, and the College of DuPage, you can get forecast soundings by clicking on a point on your model forecast map).
  • Forecasts for relative humidity at specific atmospheric levels: An extensive selection exists at Pivotal Weather, College of DuPage, and WeatherNerds (opens in a new window).
  • Simulated satellite imagery / cloud fraction or percentage forecasts: Various products are available at Pivotal Weather, College of DuPage, Tropical Tidbits, and WeatherNerds.
  • REFS mean low, mid, and high cloud forecasts: Available on the SPC REFS page (opens in a new window) (look in the "Synoptic" menu).

An Example

Let's look at an example to see some of these cloud forecasting tools in action. For starters, check out the GFS simulated infrared satellite image (below) valid at 12Z on June 14, 2022 (from the model initialized 12 hours earlier at 00Z). The model forecast predicted cold cloud tops over much of Pennsylvania, as evidenced by the very low cloud-top temperatures.

GFS simulated IR forecast valid at 12Z on June 14, 2022.

The GFS simulated infrared image valid at 12Z on June 14, 2022 (from the run initialized 12 hours earlier at 00Z) showed a canopy of high cloud tops over Pennsylvania.
Credit: Weathernerds.org

But, can we get a more complete picture of the predicted cloud coverage at this time? Absolutely -- by integrating other cloud forecasting tools. Forecast soundings from the GFS (opens in a new window) and ECMWF (opens in a new window) (also from the 00Z run) showed a deep layer of saturation or near saturation throughout the troposphere at University Park, Pennsylvania (right near the center of the state), suggesting that clouds would be very thick and extend down into the lower troposphere. Furthermore, the high-resolution ensemble mean forecast for percentage of cloud cover (opens in a new window) also suggested the presence of low, middle, and high clouds over the area (the red shading, indicative of high clouds, is most obvious, but it overlaps with green and blue shading, indicative of some low- and mid-level clouds, respectively). So, multiple products from various dynamic models suggested the presence of thick cloud cover over western and central Pennsylvania.

What about statistically post-processed guidance? GFS MOS (opens in a new window) at University Park, Pennsylvania predicted a stretch of overcast skies ("OV") and the ceiling forecasts ("CIG" row) predicted a ceiling category of "6" at 12Z, which equates to ceiling heights between 3,100 and 6,500 feet. Three hours later at 15Z, GFS MOS predicted an even lower ceiling ("4" = a cloud ceiling between 1,000 and 1,900 feet). The 01Z NBM forecast for University Park (opens in a new window) predicted cloud coverage of 79 percent ("SKY" row), with a ceiling of 7,000 feet (the "CIG" row has ceilings in hundreds of feet), so its message of low cloud cover wasn't as strong. However, the standard deviation of 26 percent indicated that most forecasts were somewhere between about 50 percent sky coverage and overcast.

How did these forecasts verify? Here are the morning METARS at University Park (opens in a new window), which reported a broken layer of clouds at 4,500 feet (marking the ceiling) at 12Z, as well as overcast clouds at 6,000 feet. So, the idea of clouds extending down into the lower troposphere was on target, and GFS MOS nailed the ceiling forecast. Ceilings later descended below 2,000 feet, although they were back up above 2,000 feet by 15Z. Note that the multi-model mean of the NBM was not the best forecast here because inevitably some constituents of the NBM disagreed on their forecasts for cloud cover in the area, and averaging out all of those bias-corrected forecasts led to a sky coverage prediction that was below 100 percent even though skies ended up being overcast.

With such optically thick clouds present throughout the morning (and rain for part of the morning), temperatures hardly budged from the low 60s until clouds broke around midday, as shown on the meteogram (opens in a new window). So, getting the cloud forecast right was a key component to understandng that it would remain cool into the late morning in central Pennsylvania. The take-away message here is that using a variety of tools for cloud forecasting will aid you in getting more specific with your cloud forecasts. Of course, these tools are imperfect; they won't verify as well as they did here all the time. You must always apply your conceptual models of how the atmosphere works to understand why clouds might be forming (what are the lifting mechanisms?) and look at observations to judge the validity of the model forecasts.

In the final analysis, getting a more specific handle on your cloud forecasts will help you fine-tune your temperature forecasts. Now, it's time to look at how winds can affect temperatures.

sas405

Wind and Temperature Forecasting

Wind and Temperature Forecasting

Prioritize...

Upon completion of this page, you should be able to discuss the primary ways that winds can impact temperature (through mechanical mixing, advection, and orographic effects), and apply these principles to your weather forecasts.

Read...

Of the numerous variables that can affect temperature, another important one is wind, and you should already be familiar with the major ways that wind can impact temperature (these were mainly covered in Lessons 4 and 6 of METEO 101 (opens in a new window) (login required)):

  • mechanical mixing via eddies caused by the wind
  • temperature advection
  • orographic effects

On this page, we'll review the basics of these concepts and discuss some specific strategies for incorporating your knowledge of these concepts into your forecasts.

Mechanical Mixing

Recall from METEO 101 that on a sunny day (particularly in the warm season), convective eddies develop, which transport increasingly warm surface air upward and cooler air from aloft downward to the ground. The end result is that the planetary boundary layer becomes "well mixed" and the thermal profile becomes dry adiabatic (opens in a new window). Winds in the boundary layer can cause a similar mixing process to occur thanks to presence of friction at the surface, which causes turbulent eddies (opens in a new window) to develop. A well-mixed boundary layer during the daytime comes with its own considerations for daytime temperature forecasting (we'll get into these later). For now, though, I want to concentrate on the impact of eddies on nighttime temperatures. For a complete treatment of this topic, you'll want to review Lesson 4 of METEO 101.

Schematic of eddies mixing colder air upward and warmer air downward at night.

At night, eddies mix colder air in contact with the cooling ground upward, while also circulating slightly warmer air toward the ground from higher up.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

You'll likely recall that nights with calm winds are cooler than windy nights (all else being equal and assuming neutral temperature advection, of course). That's because mechanical eddies caused by the wind "stir up" the lower troposphere, mixing cold air next to the ground with warmer air higher up. As the speed of the wind increases, eddies become more turbulent and more vigorously circulate the air through a layer several thousand feet thick. This process helps keep two-meter temperatures higher than they otherwise would be at night.

What about when winds are calm (or nearly so) at night? Well, without notable winds, the middle and upper parts of the boundary layer decouple from the lowest part (just above the surface).  When decoupling occurs, and eddies no longer mix air throughout the boundary layer, the processes of radiation and conduction dominate low-level temperature trends. The ground cools rapidly via a net loss of radiation, and a thin layer of air next to the ground cools via conduction. If these conditions persist, in time, a nocturnal inversion forms as the layer of "cool air" near the ground thickens. Of course, the conditions for the development of nocturnal inversions are perfected when skies are also clear (no clouds to provide downwelling infrared radiation to the ground). 

Dynamic model forecasts of two-meter temperatures tend to struggle when conditions favor the development of nocturnal inversions because radiative fluxes and conductive properties are modeled poorly (and they can vary greatly over small distances). Of course, if the dynamic model predictions of two-meter temperatures highly questionable, statistical post-processing of the forecasts might help, but various bias-correction methods used by MOS and the NBM may not perform optimally with such an erroneous starting point, so forecasters should be aware that even normally reliable guidance might need some adjustments on nights with prime nocturnal cooling conditions.

Forecasting Tip

How do we go about forecasting whether decoupling will occur? One method is to evaluate forecast soundings, and look for the development of a nocturnal inversion (especially a strong inversion). If you see well-developed nocturnal inversions on forecast soundings around sunrise, that's a sign that the model is predicting conditions to favor decoupling.

Of course, it's critical to keep the big picture in mind. You should also look for conditions that favor calm winds overnight (weak synoptic pressure gradient), as well as clear skies (or at least a lack of low cloud cover). Those conditions, along with the consensus among the dynamic model soundings that nocturnal inversions will form, should give you some confidence that decoupling and enhanced nocturnal cooling will occur. We'll get a bit more specific about how to deal with MOS's predictions in such situations later in the lesson.

Temperature Advection

You should recall that advection is just the horizontal transport of some quantity by the wind, so with temperature advection, we're dealing with the transport of air with certain temperature characteristics. When forecasting, it's important to recall the big three controllers of temperature advection:

  • the speed of the wind
  • the magnitude of the temperature gradient
  • the angle that the wind crosses local isotherms

Temperature advection is maximized when fast winds blow at right angles to local isotherms marking a strong temperature gradient. Temperature advection can be cold (colder air is arriving via the wind), warm (warmer air is arriving via the wind), or neutral (winds are calm, or they blow parallel to local isotherms).

Chances are, though, that you've probably gotten into the habit of thinking that northerly winds (or northwesterly / northeasterly winds) will cause cold advection, while southerly winds (or southwesterly / southeasterly winds) will cause warm advection in the Northern Hemisphere. Such broadbrush descriptions are correct sometimes, but when making specific weather forecasts, it's not safe to rely on them. It's important to actually examine the specific factors that control advection (the three bullet points above). 

Take the circled station model (Dodge City, Kansas) in the image below from 1643Z on March 21, 2012 as an example. Winds were clearly blowing from the north-northeast at around 20 knots. It's a case of cold advection, right?

A plot of station model observations at 1643Z on March 21, 2012.

A plot of station model observations collected at 1643Z on March 21, 2012. What is the temperature advection at the circled station?
Credit: NCAR

Wrong. Look at the temperature "upwind" of the circled station (to the north-northeast, where the wind was coming from). The air to the north-northeast was warmer than the air at the circled station (temperatures were in the 50s), meaning that these northerly winds were causing warm advection. To confirm, check out the 17Z chart of temperature advection (opens in a new window), which shows that warm advection of roughly 1 degree Fahrenheit per hour was occurring over southwest Kansas. So, it's really not safe to automatically equate cold advection with northerly winds (or warm advection with southerly winds). You should always evaluate the three factors that control temperature advection. It's the only way to really assess the type and significance of temperature advection (strong or weak). Otherwise, you're bound to get fooled at some point.

Key Data Resource

You can get hourly analyses of temperature advection at the coolwx.com (opens in a new window) site, but when looking at model forecasts, you often need to evaluate the wind direction, speed, and orientation of the isotherms to accurately assess temperature advection (qualitatively).

Surface station models in the vicinity of Lake Ontario.

Note the land trajectory of the surface wind at Watertown, New York (in green) and compare it to the cooler lake trajectory at Toronto, Ontario (in red). As a result, temperatures differed by nine degrees at this time.
Credit: NCAR

The most common culprits for major temperature advection are the cold and warm fronts (and the temperature gradients that come with them) associated with mid-latitude lows. But, even in the absence of fronts, weather forecasters still have to consider advection in the form of "air trajectories." The air trajectory is simply the path that the air takes along its way to your forecast location. To get a sense of what I mean, look at the example on the right and compare the observed weather at Toronto, Ontario (colored red) and Watertown, New York (colored green). These observations are from an April day when the waters of Lake Ontario were relatively chilly. The southeasterly winds at Toronto mean that the air had to travel over the chilly lake before reaching Toronto. As the air was modified by the underlying lake waters, it cooled, and not surprisingly, Toronto's temperature was a relatively chilly 39 degrees Fahrenheit. That's a full 9 degrees lower than at Watertown, thanks to the different air trajectories at the two cities (air arriving at Watertown did not pass over the lake first).

Accurately predicting the impact of air trajectories on temperature requires making accurate forecasts of wind direction. If air trajectories are an important part of the forecast, be aware that if the dynamic models incorrectly forecast the wind direction (perhaps causing MOS or the NBM to do the same), conditions may be ripe for a large errors in the temperature forecast. For example, take a look at this zoomed-in 18Z surface map (opens in a new window) and note the occluded front (associated with a 997 mb low) parked just to the south of Erie, Pennsylvania (opens in a new window). Dynamic models had predicted that this front would move north of Erie, and southeasterly winds would develop in its wake during the afternoon. It never happened, and chilly northeasterly winds blew off the cold lake at Erie (Erie, PA meteogram (opens in a new window)) throughout most of the day. Because the air trajectory was from a much colder source than what forecast guidance predicted, temperature forecasts were 5-10 degrees too high.

So, forecasters must always be aware of air trajectories, and the potential consequences for temperature. Does the air pass over a snowpack before reaching your station? The relatively hot downtown area of a city (an urban heat island)? A cold body of water? An area with vastly different sky cover than your station? It's always wise to look upstream from your station and be aware of conditions in the surrounding area.

Orographic Effects

Finally, you may recall that the interaction of wind and topography can impact temperature, too. Air that is forced to ascend up a mountain slope, for example, cools. Meanwhile, as air descends down the leeward slopes, it warms via compression. So, it's important to be aware of the topography that surrounds your forecast location. That knowledge will help you understand the impacts of the air trajectories that may exist on any given day. Will the air downslope into your city? Will it upslope? The answers to these questions may also have implications for clouds, which further the impact on temperature (upslope cooling can lead to cloud development, while downslope warming can cause clouds to dissipate).

In the words of a wise old forecasting professor, when you're forecasting for a location with widely varying regional topography, the most important forecast question is often, "From which direction will the wind blow?" Answering that question correctly is a big first step toward a successful forecast in areas with rough terrain.

Forecasting Tips

  • Statistically post-processed guidance can handle the impacts of air trajectories on temperatures fairly well as long as the situation isn't extremely unusual and the dynamic models have a good handle on the wind forecast.
    • Some issues related to air trajectories are "built in" to climatology (the common development of sea breezes at a coastal station, for example, or downsloping winds from a particular wind direction at a station in the mountains). So, statistical models based on a linear regression approach (like MOS) will often handle these issues fairly well because they're built into the historical set of observations used to create the regression equations. I don't mean that MOS knows about sea breezes or downsloping winds, but it does know what happens to temperature when winds blow from a certain direction.
    • If the bulk of the model guidance that comprises the National Blend of Models has a good handle on wind direction and speed, its temperature forecasts can perform well even when factors like sea breezes and upsloping / downsloping are an issue.
  • Don't automatically assume that you have to make drastic changes to statistically post-processed guidance just because of an air trajectory. You should first ask yourself, "Do I believe the dynamic models are handling the wind forecast well?" And, in the case of MOS, "Is this situation well-represented in the regression equations? Is it historically common?"

Okay, that wraps up our discussion on wind and temperature forecasting. We'll revisit some of these ideas later in the lesson, but in the meantime, it's time to move on to water -- another factor that affects temperature forecasts.

sas405

Water and Temperature Forecasting

Water and Temperature Forecasting

Prioritize...

Once you've finished this section, you should be able to discuss the impacts that dew points and latent-heat effects can have on temperature forecasts, and apply these principles to your own weather forecasts.

Read...

We've already talked a bit about the importance of a station's proximity to water for making temperature forecasts (in terms of climatology and specific air trajectories on short time scales). Now, though, I want to focus on different aspects of water as a controller of temperature: its radiative properties and "latent heat" effects (energy release or consumption resulting from water's phase changes). In this case, when I say "radiative properties" of water, I'm specifically talking about water in its gaseous form -- water vapor. 

Dew Points and Nighttime Temperatures

As you know, water vapor is an efficient absorber and emitter of infrared radiation. You also know that low clouds (comprised of liquid water droplets) emit downwelling infrared radiation, which the ground absorbs. In a similar way, invisible water vapor also emits downwelling infrared radiation. Like with clouds, this property of water vapor is especially important at night (when incoming solar radiation isn't a factor in the radiation budget). 

How do we assess water vapor concentration near the surface? Dew point temperature, of course! On sultry, very humid nights during the summer, dew points and overnight low temperatures are often in the 70s. However, in arid desert regions, nighttime temperatures can fall considerably. For example, during the Gulf War in 1991-92, the military endured searing desert heat by day. Away from the tropical Persian Gulf, dew points were much lower, setting the stage for hot days to be followed by nights so chilly that soldiers needed blankets to stay warm. To visualize this concept, check out the image below showing global average temperatures from January 1979, and focus in on the Sahara Desert in northern Africa. Note that the Sahara was hot during the day (top map) and cool at night (middle map), making this desert and others the areas of the world with the greatest diurnal change (bottom map).

Global average daytime temperatures (top), nighttime temperatures (middle) and day-night temperature difference (bottom).

Global average daytime temperatures (top image), nighttime temperatures (middle image), and day-night temperature difference (bottom image). These images indicate that the Sahara Desert (and other deserts) are characterized by large diurnal temperature differences, in part because of low dew points.
Credit: NASA Jet Propulsion Laboratory

To isolate the impact of water vapor on surface temperatures, let's ignore other factors (we'll assume the night is clear with calm winds). All else being equal, on a night with high dew points, ample water vapor emits infrared radiation to the readily absorbing ground, helping to slow its cooling rate. In turn, the ground, radiating at an elevated temperature courtesy of water vapor emissions, now provides a boost in infrared energy for water vapor to absorb and warm up. This radiative synergy between the ground and water vapor keeps the ground and the overlying air warmer at night, resulting in elevated low temperatures.

On nights with relatively low dew points, however, reduced concentrations of water vapor limit the amount of infrared radiation that the ground receives from the atmosphere, causing the ground to run a much larger radiation deficit than when dew points are high. The result is that temperatures can fall like a rock, toward the dew point. Of course, when the temperature reaches the dew point, net condensation occurs and fog forms. The process of net condensation releases latent heat which limits further declines in temperature (for all practical purposes, the temperature cannot measurably fall below the dew point). We'll talk more about other such "latent heat" effects shortly, after we address a couple of forecasting tips.

Forecasting Tips

  • A "crusty old forecasting rule" is that, on clear nights with relatively calm winds, afternoon dew points serve as a reasonable bound for the overnight low temperature. This rule assumes, of course, that because conditions are clear and calm, dew points won't notably change overnight. That assumption isn't always valid, however, such as in cases where a snowpack exists.
  • When faced with a difficult low temperature forecast (perhaps a frontal passage will mark the arrival of a different air mass), examining predicted trends in dew points can help us establish boundaries for what temperatures might do overnight. For example, if you expect the advection of warm, moist air to occur at your forecast location overnight, you should consider looking at MOS / NBM dew point forecasts and ensemble dew point plumes for trends and ranges to help us make more informed forecast decisions.

Latent-Heat Effects

As I just mentioned above, when net condensation occurs, latent heat of condensation is released, which limits additional temperature drops. That's not the only way that latent heat can impact temperature, however. Recall that any time water changes phase, energy is either required (which cools the air) or is released (which warms the air). The familiar "energy staircase" (below) for the three phases of water should jog your memory.

The energy staircase for the phase changes of water.

The energy levels associated with ice, water, and water vapor can be thought of as a set of steps. Changing from one phase (solid, liquid, or gas) to another requires either an addition of energy (stepping up) or a release of energy (stepping down).
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

Any time water moves "up the staircase" (changes from ice to liquid, or liquid to vapor), energy is required (which cools the air) to break the bonds between molecules. On the flip side, moving down the staircase (phase changes from vapor to liquid, or liquid to ice) results in the release of energy.

So, it's not only the release of latent heat of condensation that impacts surface temperatures. Cooling from evaporation of raindrops can be an important factor, too, particularly when relative humidity values in the lower troposphere are initially low (the air is not close to saturation). When snowflakes are falling, they can transition directly from ice into vapor (sublimation), which requires even more energy than evaporation, leading to greater cooling. But, evaporational cooling is not limited to falling precipitation. Wet ground depletes the heating power of solar radiation because some energy goes into evaporating water instead of directly heating the ground. So, the phase changes of water near the ground are important temperature controllers!

Forecasting Tips

  • In situations where evaporational cooling of precipitation is going to play an important role in the forecast, you should take time to think about wet-bulb temperatures. Recall from your previous course work that the wet-bulb temperature is the lowest reading to which temperatures can fall by continually evaporating moisture into the air (assuming that pressure remains constant during the process). In this context, "lowest reading" translates to the temperature at saturation. You may want to bookmark the handy wet-bulb calculator (opens in a new window) for situations when you want to get a feel for the potential effects of evaporational cooling.
  • Knowing the recent history of precipitation at your forecast location can be helpful when making a forecast. Have conditions been abnormally wet or dry lately? If so, soil moisture may also be abnormally high or low, which may justify tweaks to your temperature forecasts. For MOS, such adjustments may only be notable in extreme situations (because few, if any, similar situations would be present in the sample used to develop MOS equations). So, in the midst of a record-breaking drought, for example, the abnormally low soil moisture means less groundwater for evaporation, which maximizes solar heating power compared to the set of days in MOS's sample. On the flip side, a big change in soil moisture could bring challenges for the NBM because it's bias corrects against recent weather.
    • Similarly, short-term weather events that the models don't predict well (like an isolated thunderstorm dumping heavy rain over your forecast location in the evening) can change the local "water situation" by making the ground very wet and perhaps boosting dew points locally (which may affect temperatures overnight). So, be on your toes for mesoscale situations that impact local dew points and ground wetness.
  • Naturally, judging the relative importance of these factors (or any particular temperature controller) for a given forecast requires a solid grasp of the big picture situation. For example, there's no need to split hairs over how an isolated shower might change dew point temperatures temporarily if a strong cold front is slated to pass your city before the night is over. See what I'm getting at? I know it probably sounds like a broken record, but the big picture is always crucial!

Water has a number of very important implications for temperature forecasting, and this was merely a brief review. If you feel that you need a more complete review of these topics and other controllers of temperature, I recommend rereading Lesson 4 of METEO 101 (opens in a new window) (login required). In the meantime, we'll now shift out of "review" mode and look more specifically at applying our knowledge of temperature controllers to actual forecast data in order to develop some strategies for improving our temperature forecasting. Keep reading!

sas405

Forecasting Temperature Extremes

Forecasting Temperature Extremes

Prioritize...

After you've completed this section, you should be able to discuss how extreme temperatures are typically associated with extreme synoptic weather patterns and identify whether a particular pattern might lend itself to producing record warmth or cold.

Read...

Occasionally, temperature forecasts from dynamic models and / or statistically post-processed guidance may suggest record-breaking (or near record-breaking) warmth or cold. By its very nature, record-breaking weather is relatively rare at locations that have long observational periods of record since observations outside of historical bounds (records) have never been measured.

Given how unusual such observations have been historically, a forecaster should always question whether record-breaking forecasts (or record-challenging) forecasts have merit. How does a forecaster determine whether the forecasts for record heat or cold are legitimate? Doing so begins first with a knowledge of local and regional climatology: What are the daily (and monthly) record temperatures? How long is the period of record in this area? In other words, how common are the temperatures that are being predicted? Answering those questions is a first step in assessing how "strong" the records are and how hard they may be to break.

After developing a good understanding of climatology, the key is to perform a thorough assessment of the big-picture pattern. But, what is a forecaster looking for that will help determine whether record temperatures might be in play? Well, it boils down to this simple rule: Extreme temperatures are often associated with extreme synoptic weather patterns. What does that mean?

As an example, a very deep 500-mb trough may signal an outbreak of record cold, or a high-amplitude ridge may signal the potential for record heat. Recall that dynamic model skill in predicting conditions above the planetary boundary layer is better than skill in predicting surface conditions, so forecasters can look to dynamic model forecasts for signs of extreme temperatures in conditions aloft. The presence (or lack of) extreme conditions aloft (say in 500-mb heights or at 850-mb temperatures), can either confirm or call into question the possibility of record-breaking temperatures.

To see what I mean, check out the short video below (3:39) showing the synoptic signs of record-breaking heat in Death Valley.

Death Valley Record Heat (3:39)

Death Valley Record Heat
Transcript: Death Valley Record Heat (3:39)

Pattern recognition is a key tool for weather forecasters, and it’s important that forecasters can recognize patterns that may lead to extreme weather. After all, record-breaking temperatures often occur in concert with extreme synoptic patterns. As an example, check out this 3-day GFS forecast valid at 00Z on August 17, 2020 for 500-mb heights, which are shaded, along with mean sea-level pressure. The feature that jumps out immediately is the big upper-level high over the western United States, with heights in the dark reddish area between 594 and 600 dekameters. Such values are commonly associated with very hot weather. One of the hottest areas in the United States, Death Valley, California, marked by the star, was in this area, so the key forecast question was whether it was just going to be typically hot for summer, or would the heat challenge or break records.

To determine whether this pattern could be associated with unusually hot weather, let’s first look at the corresponding GFS forecast for 500-mb heights and anomalies, which are shaded. It’s clear that 500-mb heights were above average over much of the western U.S. where the high-amplitude ridge was located. The greatest difference from normal is here over southeastern British Columbia, where heights were 12-15 dekameters above normal in the GFS forecast. The anomalies farther south were not as large.

But, we need climatological context for those anomalies, so let’s check out the GFS forecast for normalized anomalies. The normalized anomalies make it clear that the most climatologically significant area of anomalies was predicted to be near the California / Nevada border into southern Nevada and Arizona. In this area, the predicted 500-mb heights were between 2 and 3 standard deviations above normal. So, they were effectively somewhere in the top 5 percent of climatology – a red flag that unusually hot weather, even for August in parts of the desert southwest U.S. could be on tap.

Ensemble forecasts provided further support. The North American Ensemble Forecast System mean forecast showed some predicted 500-mb heights that were the highest in the climatological sample for this time of year, and an even larger area in the top 1 percent.

A similar signal appeared in the NAEFS mean 850-mb temperature forecasts. This large area is gray because high elevations mean the 850-mb surface lies below the ground in much of the western U.S., but we can still see an area in California, southern Nevada, and Arizona where the predictions for 850-mb temperatures were the highest or nearly the highest in the climatological sample for this time of year.

Recognizing these signals in the synoptic pattern gave forecasters confidence that extremely hot weather was in store, even by Death Valley’s standards. Daily record highs during the month of August in Death Valley are generally in the 120s Fahrenheit, but this pattern suggested that such temperatures could be challenged or exceeded, and indeed, not only was a daily record high temperature of 130 degrees set in Death Valley on August 16, but it was the first time since records began in 1911 that the temperature hit 130 degrees in August. Overall at the time, August 16, 2020 ended up tied as the 3rd highest temperature on record in Death Valley, and the possibility of such extreme heat was telegraphed well by extremely high 500-mb heights and 850-mb temperatures predicted over the area.

Credit: Penn State

Note how this type of analysis requires the forecaster to analyze the overall synoptic pattern and also apply knowledge of climatological statistics (percentiles, normalized anomalies, etc.). For the other end of the spectrum, check out this short video (4:16) showing the synoptic signs of record-breaking cold in the southern Great Plains. You may need to expand the video to full screen if you're having trouble viewing the text forecasts.

Southern Plains Record Cold (4:16)

Southern Plains Record Cold
Transcript: Southern Plains Record Cold (4:16)

In February, 2021, model guidance suggested an outbreak of extreme cold over the southern Plains in the U.S. These National Blend of Models forecasts from the 13Z run on February 14th predicted jaw-dropping cold for Dallas-Ft. Worth, Texas and Oklahoma City, Oklahoma on the mornings of February 15th and 16th. The NBM predicted low temperatures of 4 degrees Fahrenheit and 2 degrees Fahrenheit in Dallas on those mornings, but compare those forecasts to the daily record lows. These forecasts were 10 or more degrees lower than the daily records! Similarly in Oklahoma City, the predicted lows on the mornings of the 15th and 16th were -6 degrees and -10 degrees, respectively. Those forecasts were almost 15 degrees lower than the daily records! In both cities, the long-term period of record went back before 1900, so crushing daily records by 10 or more degrees would be a huge deal.

When model guidance predicts record challenging or record breaking weather, forecasters can start assessing whether such forecasts are realistic by analyzing the synoptic pattern. The GFS forecast initialized at 00Z on February 14 and valid at 06Z on February 15 showed a deep 500-mb trough entering the central U.S. We would expect such a trough to be associated with cold air columns. At the surface, a strong, Arctic high was moving into the Upper Midwest, with the general clockwise flow around it causing northerly flow and cold advection into the southern Plains. Note also that snow was predicted to be occurring, and in fact parts of northern Texas and Oklahoma ended up receiving a few to several inches of snow.

The snow depth analysis from 06Z on February 15 showed widespread snow cover across Texas and Oklahoma with more than 4 inches in some areas of central Oklahoma.

But, how significant was the 500-mb trough climatologically? The North American Ensemble Forecast System mean 500-mb height forecast valid at 12Z on February 15 was between 2 and 3 standard deviations below normal, as marked by the blue shading over Oklahoma City, Dallas, and the surrounding areas. So, the heights predicted in this region were somewhere in the bottom 5 percent of the climatological sample.

To get more specific, the NAEFS mean forecast for 500-mb heights in much of Texas and Oklahoma was actually in the bottom 1 percent, as shown on this forecast map of mean 500-mb heights and climatological percentiles. So, this was a climatologically significant trough! Keep in mind that these are not based on any single model, or the most extreme model forecast. These are based on the mean ensemble forecast, so inevitably some members had more extreme forecasts than this!

The NAEFS mean 850-mb temperature forecasts for the morning of February 15 were also extreme, with a large section of Texas and Oklahoma, marked in purples, predicted to have 850-mb temperatures that were the lowest in the climatological data set. So, the idea that these mornings could have record-breaking chill had support from the synoptic pattern. The extreme chill in this air mass really was exceedingly rare.

So, what happened? The NBM forecasts were pretty good! It nailed the low temperature forecasts on the morning of the 15th at both Dallas and Oklahoma City. Its forecasts were several degrees too high on the morning of the 16th, however. The actual low of -2 degrees Fahrenheit in Dallas and -14 degrees in Oklahoma City both represented the second lowest temperatures on record for any date, with records going back prior to 1900 at each city! This really was a remarkable event. Also take note that the morning of the 16th was much colder than the morning of the 15th even in the same air mass, and those differences depend on temperature controllers like cloud cover and wind speed. While the synoptic pattern signaled that record cold was very possible, the other details of the big picture determine just how cold it can get.

Credit: Penn State

Identifying an extreme synoptic pattern can alert you to the possibility of record-challenging or record-breaking temperatures (whether post-processed temperature guidance predicts them or not). But, I should caution you that record-breaking temperatures at the surface do not require record-breaking conditions aloft. In other words, you may encounter situations where conditions aloft are near the extremes, but not actually record breaking, and records still get broken at the surface. On the other hand, you may encounter situations where record-breaking conditions exist aloft and records are not broken at the surface.

How can this be? Well, the devil in any forecast is in the details, and forecasters have to consider the entire big picture. For example, record-breaking cold at 850 mb may not correspond to a record low temperature at nighttime if there's a persistent low overcast, which could provide enough downwelling infrared radiation to stave off record-breaking temperatures (even if it's still a chilly night!). On the other hand, record heat might occur even under a 500-mb ridge that isn't record breaking if the area is experiencing a drought and solar heating power is maximized each day. So, as a forecaster, you must be aware of the entire big-picture situation before making your final forecast decision, but identifying whether aspects of the synoptic pattern are near extremes can provide signals to forecasters about the prospects for record-breaking temperatures.

Forecasting Tips

  • To see if extreme warmth or cold is in the cards, look for extreme aspects of the synoptic pattern. Are 500-mb heights extremely high or low? Are 850-mb temperatures extremely high or low? Answering those two questions can provide clues about how real the risk of extreme temperatures is in any given situation. The NWS Situational Awarenes Tables (opens in a new window) can be a great tool for identifying extreme aspects of the synoptic pattern. By selecting a region of the U.S., and either standardized anomalies or percentiles for your output, you can quickly identify areas where the synoptic pattern is predicted to be extreme. Note that depending on elevation, other pressure levels may be more appropriate for analysis -- you may want to look at 700-mb temperatures if your location is very near (or even higher than) the 850-mb level, for example).
  • Whether or not extremes actually occur at the surface also depends on other temperature controllers (cloud cover, precipitation, soil moisture, advection / air trajectories, etc.). So, you must evaluate the entire big picture.
    • Do all aspects favor warmth or cold? If you find aspects of the big picture that run contrary to the overall message of extreme warmth or cold, record-breaking temperatures may not occur (or you should at least be open to the possibility that they won't). On the other hand, if all aspects favor a particular extreme, that increases forecast confidence that records may be broken (and in some cases, perhaps by a lot).

Ultimately, connecting statistical forecast guidance to the big picture is critical for verifying whether record breaking temperatures are a legitimate possibility. But, awareness of the big picture is also critical in identifying situations when statistical forecast guidance may struggle with its temperature forecasts. Up next, we'll explore situations in which dynamic model errors lead to poor forecasts from statistical models.

sas405

Model Misses

Model Misses

Prioritize...

When you've completed this page, you should be able to discuss and identify how errors in dynamic or AI model forecasts can negatively impact forecasts from statistical models.

Read...

Statistical models excel at getting forecasters reliably "in the ballpark" for their temperature forecasts. But, remember that the primary statistical models in use rely on dynamic or AI model output. So, if dynamic or AI models wildly miss on an important aspect (or aspects) of the big picture, those errors can negatively impact forecasts from statistical models like MOS or the NBM. Therefore, it's imperative that forecasters always keep close tabs on current observations and the big picture so that they can identify situations in which dynamic or AI model "misses" may cause statistical models to struggle.

One scenario that can cause these struggles can occur when dynamic or AI models simply do not predict a variable that plays a critical role in the forecast. For example, in September 2020, widespread wildfires raged across the western United States. These fires burned more than 3.5 million acres during the month (opens in a new window), which was the most on record at the time. Skies were regularly obscured by thick smoke, creating eerie reddish skies and darkness during the day. This collection of photos in The Atlantic (opens in a new window) chronicles the almost apocalyptic scenes during the event. The GOES geocolor satellite loop below, from the morning of September 9, shows the stunning extent of thick smoke across western Oregon and much of western and central California.

The GOES geocolor satellite loop from the morning of September 9, 2020 shows the stunning extent of wildfire smoke across western Oregon and western and central California.
Text description of the GOES geocolor satellite loop (0:02)

The loop shows a satellite view covering a portion of the western United States, the eastern Pacific Ocean, and part of Canada. The upper part of the image displays dense clouds over the ocean, with a swirling formation moving inland towards the North American coast. The cloud cover extends from the Pacific Ocean over Oregon, Washington, and British Columbia, moving eastward across parts of Idaho and Montana with visible green areas indicating forested regions. The right third of the image reveals a mix of cloud cover and clear skies over the Great Basin and the southwestern U.S. The mid-to-lower portion of the image is dominated by a large, distinct area of grayish-brown, indicating smoke likely from wildfires, stretching from the Pacific coast inland across California and beyond. The date and time "2020-09-09 15:30:31 UTC" is shown in the bottom left corner. In the bottom right corner, logos for CIRA and RAMMB are visible, representing the Cooperative Institute for Research in the Atmosphere and Regional and Mesoscale Meteorology Branch.

Credit: RAMMB / CIRA

With a large 500-mb high anchored just off the West Coast (check out the 12Z analysis of 500-mb heights on September 9 (opens in a new window)), the upper-air pattern certainly pointed to hot weather along the West Coast. Temperature forecasts from statistical models followed suit. MOS maximum temperature forecasts for September 9 (opens in a new window) from the 12Z run on September 8 predicted widespread highs in the 90s away from the coast, and GFS MOS even had some triple-digit highs in northern California and Oregon.

If you're thinking that such high temperatures seem out-of-touch with the big picture, you're correct! Thick smoke from wildfires would really limit incoming solar radiation (a great deal of solar radiation was scattered before ever reaching the surface). In effect, the thick smoke had similar impacts to very optically thick clouds, but all of the ways that MOS typically accounts for clouds in its temperature predictions (like mean-layer relative humidity) actually suggested clear skies. For example, check out the forecast sounding for Sacramento, California (KSAC) below, valid at 21Z on September 9 (from the model initialized just three hours earlier at 18Z).

The 3-kilometer NAM forecast sounding valid at 21Z on September 9, 2020.

This forecast sounding valid at 21Z on September 9, 2020 at Sacramento, California (from the 18Z run on the same day) showed low relative humidity throughout the troposphere.
Credit: Tropical Tidbits

Such low relative humidity throughout the troposphere signaled to MOS that skies would be clear, and under the premise of full sunshine, MOS predicted a very hot day. But, the thick smoke, which was unaccounted for, blocked a significant fraction of incoming solar radiation, and the day ended up being much cooler. At 21Z, MOS forecasts were MOS forecasts were running in excess of 20 degrees Fahrenheit too high (opens in a new window) across a number of stations in California and Oregon (with other stations being 10 to 20 degrees too high). Those are huge errors (especially for a one day forecast)!

Note that the predicted surface (two-meter) temperature in the Sacramento forecast sounding above was a toasty 93 degrees Fahrenheit, since the dynamic model also predicted full sunshine. Indeed, in 2020 no operational dynamic models had the ability to explicitly predict smoke (and even now, most dynamic and AI models still do not), and that presented a problem for the NBM. Its bias correction and blending techniques were unlikely to provide an accurate forecast because its members were forecasting temperatures under the false premise of mainly clear skies.

At Sacramento, the 12Z run of the NBM from September 8 predicted a high of 94 degrees (NBM forecast (opens in a new window)) Fahrenheit on the September 9. The actual high was 81 degrees (opens in a new window)! Like MOS, the NBM forecast was definitely not in the ballpark in this case! Keep in mind that such large errors are relatively rare in the scheme of things, but being aware of the wildfire situation and knowing that models did not have the ability to predict smoke could have led forecasters to make forecasts that were significantly better than forecast guidance. This example was a bit unusual and very extreme, but it emphasizes the importance of keeping in touch with the big picture and not treating statistical forecast guidance as a "black box" that just magically spits out excellent forecasts. Knowing a bit about how the forecasts are generated and what models can't predict helps a lot!

More commonly, forecasters must be aware of situations where dynamic or AI models produce erroneous cloud forecasts (predicting clouds when none actually exist or the opposite — predicting sunny skies when substantial cloud cover exists), as such errors can drastically impact temperature forecasts. The same goes for situations with erroneous wind direction forecasts, especially in areas near bodies of water or in areas of rough terrain. All of these errors can negatively impact temperature guidance. But, let's turn our attention to a weather feature that frequently causes model guidance to struggle with temperature forecasts.

A General Temperature Forecasting Problem — Fronts

Model guidance of all types can have problems with temperature forecasts in the vicinity of fronts, especially strong fronts which are associated with large temperature gradients. Why are large temperature gradients problematic? Well, a large temperature gradient means that there's a large change in temperature over a short distance. So, if a dynamic or AI model's forecast for the timing of a frontal passage (or the location of a front at a specific time) is wrong, it can lead to large errors in its temperature forecasts.

In terms of the NBM, bias correcting and blending a bunch of model solutions doesn't always produce the best temperature forecast. Even subtle disagreements in the predicted location of a front (and its sharp temperature gradient) can lead to "washed out" temperature gradients when all of the solutions are blended together. By itself, this might lead to modest errors, but when most models "miss" on the timing of a frontal passage, forecast errors can get pretty big. To see what I mean, let's look at an example of a strong, arctic cold front moving through Houston, Texas. As you can see from the animation of dynamic model forecasts valid at 06Z on February 3, 2022 (below), this front had a very sharp temperature gradient associated with it. Locations along the Gulf Coast were predicted to be in the 60s at this time, while temperatures in the teens and 20s were predicted to be as far south as central Texas.

Animation of dynamic model 2-m temperature forecasts valid at 06Z on February 3, 2022, showing disagreement in the location of a strong front.

Dynamic model two-meter temperature forecasts valid at 06Z on February 3, 2022 (from the 18Z run on February 2) showed minor differences in the location of a strong cold front causing big differences in temperature forecasts around Houston, Texas.
Credit: Pivotal Weather

Also note that the models differed on the exact location of the front (the leading edge of advancing colder air). The differences weren't large, but for locations (like Houston) very near the front, the impacts on predicted temperatures were significant. For example, the GFS two-meter temperature prediction at Houston Hobby Airport (KHOU) at the time was 54 degrees Fahrenheit, while other dynamic models were as much as 10 degrees higher. Most (but not all) dynamic model guidance had KHOU on the warm side of the front at 06Z. But, a slightly faster arrival of the front could mean that the "warmer" forecasts would be way too warm at this time.

The 19Z NBM temperature prediction for 06Z at KHOU (opens in a new window) was 64 degrees Fahrenheit, so the bias-corrected weighted mean was toward the warmer end of raw dynamic model guidance (likely because so many dynamic models had the front just west of KHOU at 06Z). What actually happened? Well, the front passed through Houston before 06Z (opens in a new window), and the KHOU METARS (opens in a new window) tell the story. At 05Z, the temperature was still 64 degrees Fahrnheit (17.8 degrees Celsius in the T-group), and the front hadn't passed yet. But, in the 0537Z observation, the "FROPA" remark appears, indicating that the front had passed. Winds were now out of the northwest and temperatures had plunged to 54 degrees Fahrenheit (12.2 degrees Celsius in the T-group). By 0553Z (the official time of the 06Z observation), the temperature was 53 degrees Fahrenheit — more than 10 degrees lower than what the NBM had predicted. So, this huge error was caused by the bulk of dynamic models being perhaps an hour too slow with the arrival of the cold front — a small error that had huge consequences for the high temperature on February 3 at KHOU (opens in a new window) (since 06Z was local midnight).

Meanwhile, the MOS approach didn't fair much better here. MOS temperature predictions (opens in a new window) were in the low 60s at 06Z — still much too warm. MOS has its own problems in situations with large temperature gradients, even if dynamic models get the location of the front right. That's because the two-meter temperatures from the dynamic models, which are frequently used as MOS temperature predictors, aren't directly substituted into the MOS regression equations. Rather, the actual predictor is a smoothed value of two-meter temperatures over a number of grid points. These "smoothing schemes" were introduced to eliminate spurious "noise" from patterns of MOS data. Most of the time, they don't cause problems, but in the vicinity of fronts, the smoothed two-meter temperatures do not have the same sharpness that the raw (unsmoothed) patterns of two-meter temperatures have. The temperature gradient gets "washed out" a bit.

The smoothing of two-meter temperatures is particularly a problem in the immediate wakes of cold fronts, where the largest temperature gradients prevail. Smoothing can also present problems, though, just ahead of cold fronts. Smoothing ends up artificially lowering the values of the thermal predictors that go into the equations, causing MOS's forecasts to commonly be too low, especially if there's no precipitation ahead of the front.

Forecast Tips

In the vicinity of fronts (strong fronts, in particular), dynamic and AI model difficulties in predicting the exact location of the front along with smoothing issues of MOS predictors may cause problems for statistical models. Keep these forecasting tips in mind:

  • Always view temperature guidance with an extra critical eye when frontal passages are in play.
  • An ensemble mean (bias-corrected or not) may not realistically portray the temperature gradient with the front (gradients can get washed out), which means statistical guidance may underestimate the rate of cooling immediately in the wake of strong cold fronts.
  • MOS is notorious for predicting high temperatures that are too high in the wake of Arctic fronts; however, its smoothing scheme makes it likely that MOS will underestimate the cooling in the immediate wake of a sharp cold front at any time of day.
  • When forecasting for locations of relatively open, flat terrain, dynamic models often underestimate the forward speed of strong cold fronts, which can exacerbate errors in temperature forecasts. However, the opposite can be true in areas of more rugged terrain — dynamic model forecasts for strong cold fronts can move them along too fast.

The bottom line is that forecasters must be in tune with the synoptic pattern to identify situations where fronts might cause problems in the forecast or to identify unusual situations in which the forecast relies on a variable that's not predicted very well (or at all). Overall, recognizing patterns that often give models "fits" is a key skill for forecasters to possess, and we'll explore common patterns in the next section. Read on.

sas405

Problematic Patterns — Daytime

Problematic Patterns — Daytime

Prioritize...

Upon completion of this page, you should be able to discuss and identify weather patterns that often cause problems for even normally reliable daytime high temperature forecasts from statistical models. Upon identifying one of these patterns, you should be able to apply this knowledge to your own weather forecasts.

Read...

In the previous section, we focused on scenarios where temperature forecast errors are rooted in the dynamic or AI models "missing" on some big-picture aspect of the forecast. But, in certain weather patterns, statistical guidance may struggle even when dynamic or AI models have a good handle on the overall big picture pattern.

Graph showing a simplified linear regression example.

The linear regression approach typically produces reliable temperature forecasts, but can struggle with statistical outliers.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

In such patterns, both MOS and NBM forecasts may struggle, but for different reasons. For MOS, the problems mainly arise from deficiencies of using a linear regression scheme. These problems lead to MOS having established biases (forecasts typically too warm or too cold) in specific weather patterns because they are statistical outliers or because MOS's predictors don't capture a particular factor in the forecast very well. On the other hand, the NBM is also susceptible to problems in these patterns, but the problems are less consistent and pinning them down is more of a moving target because the bias corrections in the NBM are based on recent model performance. But, ultimately, the multi-model average isn't a great forecast in some weather patterns (even after the bias correction and blending process occurs). 

With the NBM, forecasters also have to be aware of how upcoming weather may differ greatly from recent weather if a particular pattern has dominated recent weeks. For example, if several weeks of sunny, dry weather has occurred, the NBM's daytime high temperature forecasts may struggle a bit if suddenly you're faced with an all-day rain because the bias corrections are based on model performance mostly on sunny days, dry days. So, forecasters should be also aware of how big changes in the weather pattern. Mind you, the NBM errors in these scenarios may not be large, but a good situational awareness can help forecasters make modest adjustments in the right direction.

Points to remember...

  • Limitations in the linear regression approach used by MOS lead to specific weather patterns in which MOS is fairly consistently too warm or too cool
  • These same weather patterns can make NBM forecasts also more susceptible to error, but the errors are less consistent than those of MOS
  • Be aware of major changes in the weather pattern compared to recent days, as the bias correction and blending scheme of the NBM may not work as well temporarily when the pattern shifts

With that general discussion out of the way, let's get right to the patterns that can be problematic.

Problematic Patterns for Forecasting Daytime High Temperatures

I recommend memorizing the patterns below. When you identify one of them, keep in mind that MOS often predicts daytime maximum temperatures that are too high (an established bias). You should also view NBM forecasts with increased scrutiny, but its errors are less consistent.

Forecast Tips

Be alert for the following patterns, and give statistical guidance increased scrutiny (MOS specifically tends to predict daytime temperatures that are too high):

  • if you expect protracted precipitation to fall, often on the precipitating, cold side of an anafront (warm or stationary front)
  • when you expect cold-air damming
  • when a persistent, relatively thin, but optically thick layer of clouds forms in the lower troposphere (especially during the warm season)
  • when there's snow cover, particularly in non-forested areas
  • when shallow cold air arrives and persists (opens in a new window) (especially shallow, Arctic air during winter)
  • when sea breezes associated with relatively cold offshore waters develop and replace an unusually hot air mass within 100 kilometers (or so) of the coast (most commonly observed in the spring)
  • when you expect a cold air mass to break records

You've encountered all of these patterns at some point in the certificate program, but I want to talk a bit more about one of the "sneakier" patterns that can cause forecasters fits:

Thin Layers of Optically Thick Clouds

Earlier in this lesson, I talked about the importance of getting specific about your cloud forecasts (in terms of cloud type, height, thickness, etc.), and gave you a handful of tools to help determine such details. Now, it's time to look at an example scenario where such knowledge will come in handy. Recall that during the day, the optical thickness of clouds is an important temperature controller because it determines how much incoming solar radiation ultimately passes through to the earth's surface. Well, a persistent, relatively thin, but optically thick, layer of clouds in the lower troposphere can frequently cause temperature guidance to be too warm during the daytime (MOS more consistently than the NBM, however). This MOS shortcoming is primarily a warm-season deficiency because summer days with a persistent, optically thick low overcast tend to be statistical outliers.

The most consistent offender in these cases is GFS MOS. Why? As the schematic below shows, one of its common temperature predictors is the 1000 mb to 500 mb mean relative humidity, which doesn't capture the true nature of cloud cover in these instances.

Schematic showing why GFS MOS in particular tends to struggle on days with a thin layer of optically thick clouds.

Given that the 1000-500 mb mean relative humidity (or some other layer) is a predictor for GFS MOS temperatures (to account for clouds), a relatively thin layer of persistent, optically thick clouds (which back-scatters a lot of sunlight) dramatically increases the potential that GFS MOS will be too high, particularly during the warm season.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

When a physically thin, but optically thick layer of clouds persists over a region (particularly during the warm season), the mean relative humidity from 1000 mb to 500 mb (or some other layer) would be rather low because there's a lot of unsaturated air above (and some below) the overcast. So, GFS MOS would "think" that low mean relative humidity translates to limited clouds, ample sunshine and higher temperatures. But, even relatively thin stratus decks can be relatively "opaque," blocking large amounts of incoming solar radiation, leading to a cooler day than MOS predicted.

What's interesting about this problematic pattern is large errors in high temperature guidance can still occur even when the dynamic models accurately predict a persistent, relatively thin layer of optically thick clouds! Let's look at an example from September 12, 2020. The short video below (2:34) outlines the big picture set up.

The visible satellite image from just before 21Z on September 12, 2020 showed a deck of stubborn low, optically thick clouds over central Pennsylvania, which hung around throughout the afternoon. The prominent appearance of these clouds on visible satellite imagery tells us that they were optically thick, with a significant fraction of sunlight scattered back to the satellite.

But, the METAR observations from University Park Airport right near the middle of the state, spanning from 15Z to 23Z – or about midday to early evening in central Pennsylvania – confirm that these were also low clouds, with overcast skies below 2,000 feet being reported.

The presence of these low clouds wasn’t really a forecast surprise because they were telegraphed very well by the big picture. The 24-hour GFS forecast for mean sea-level pressure and 10-meter winds valid at 18Z on September 12 showed a 1028 millibar high anchored just off the coast of Maine, with clockwise flow around it leading to moist, southeasterly winds into central Pennsylvania.

At 925 mb, GFS forecasts valid from 00Z on the 12th through 06Z on the 13th show dew points increasing throughout central Pennsylvania in concert with the low-level southeasterly flow, a sign of moisture being transported into the area.

As these moist winds flowed toward central Pennsylvania, regional terrain provides the natural lifting mechanism of orographic lift, as southeasterly winds have to upslope into central Pennsylvania. The upsloping of moist air arriving from the east or southeast in central Pennsylvania is a tried and true recipe for low clouds.

Forecast soundings valid at 18Z on September 12 from models run 24 hours beforehand sent a clear message that a thin layer of optically thick clouds would form in the lower troposphere. Note the layers of high relative humidity in the GFS and 3-kilometer NAM forecast soundings. So, dynamic models were on top of the potential for a thin layer of optically thick low clouds to develop.

Under these optically thick low clouds, temperature hardly budged through the afternoon, crawling upward from the low 60s near midday to 65 degrees Fahrenheit late in the evening. Temperatures were in the low 60s for much of the afternoon, which is a testament to how effective thin layers of optically thick low clouds can be at reducing incoming solar radiation and stifling temperature increases.

Credit: Penn State

Given how GFS MOS accounts for clouds in its temperature regression equations, seasoned forecasters should have expected its maximum temperature forecasts to be too high because the mean relative humidity between 1000 mb and 500 mb was still pretty low despite the presence of the low clouds. Indeed, the GFS MOS high temperature forecast of 69 degrees Fahrenheit was 4 degrees Fahrenheit too high on this day. But, GFS MOS wasn't the only statistical model that struggled. The 06Z run of the NBM run on September 12 (opens in a new window) predicted a high of 72 degrees Fahrenheit! The bias-corrected, weighted mean of a superensemble was much too warm here. So, the GFS MOS and NBM forecasts were pretty poor.

Lesson learned: Recognizing thin layers of optically thick clouds is important to your temperature forecast! Make sure you have a firm grip on the big picture and that you've thoroughly vetted the forecast soundings and eliminated the possibility of a thin layer of optically thick clouds persisting for much of the daytime hours (especially during the warm season).

In the above scenarios, guidance (especially MOS) tends to be too warm with its high temperature forecasts. What about patterns when the opposite is true? In the patterns below, MOS often predicts daytime maximum temperatures that are too low. Admittedly, the variety of situations is smaller compared to warm-bias situations. You should also view NBM forecasts with increased scrutiny, but its errors are less consistent.

Forecast Tips

Be alert for the following patterns, and give statistical guidance increased scrutiny (MOS specifically tends to predict daytime temperatures that are too low):

Note, on the template of a mid-latitude cyclone (the link above in the blue box), that the area where there's a tendency to be too cool lies well in advance of the cold front. Clouds, showers, or thunderstorms tend to arrest the daytime increase in temperatures along and just ahead of the cold front, disrupting the formation of the classic thermal ridge that should theoretically develop ahead of the cold front.

In many cases, however, the highest daytime temperatures occur in an area that lies farther east, where clouds and showers associated with the approaching cold front do not reach the region until later in the day. This delay allows sunshine and strengthening south-southwesterly winds to boost temperatures. So, if the warm sector is "dirty" (meaning that it's contaminated with lots of cumulus clouds and showers well ahead of the front), this cool bias probably won't show up in a major way.

Now that we've covered common "problem patterns" for the daytime, let's go over to "the dark side" and examine some problematic patterns at night in the next section.

sas405

Problematic Patterns -- Nighttime

Problematic Patterns -- Nighttime

Prioritize...

After completing this page, you should be able to discuss and identify weather patterns that often cause problems for even normally reliable nighttime low temperature forecasts from statistical models. Upon identifying one of these patterns, you should be able to apply this knowledge to your own weather forecasts.

Read...

Weather patterns that are more likely to cause problems for statistical guidance aren't just limited to daytime high temperature forecasts. Indeed, specific weather patterns make it more likely that nighttime low temperature forecasts from statistical models might have relatively large errors. However, much like with the weather patterns discussed previously for daytime highs, MOS and NBM forecasts may struggle for different reasons. For MOS, the problems mainly arise from deficiencies of using a linear regression scheme. These problems lead to MOS having established biases (forecasts typically too warm or too cold) in specific weather patterns because they are statistical outliers or because MOS's predictors don't capture a particular factor in the forecast very well.

Problems experienced by NBM forecasts in these patterns are not as consistent because its bias-corrected forecasts rely on recent model performance. But, again, the multi-model average isn't a great forecast in some weather patterns (even after the bias correction and blending process occurs). Forecasters must also be aware of how upcoming weather may differ greatly from recent weather if a particular pattern has dominated recent weeks (perhaps snowless ground is suddenly covered by a deep snow pack in the wake of a winter storm, for example). Allow me to reiterate the important "points to remember" below, because I think they bear repeating:

Points to remember...

  • Limitations in the linear regression approach used by MOS lead to specific weather patterns in which MOS is fairly consistently too warm or too cool
  • These same weather patterns can make NBM forecasts also more susceptible to error, but the errors are less consistent than those of MOS
  • Be aware of major changes in the weather pattern compared to recent days, as the bias correction and blending scheme of the NBM may not work as well temporarily when the pattern shifts

With that general discussion out of the way, let's get right to the patterns that can be problematic.

Problematic Patterns for Forecasting Nighttime Low Temperatures

For predicting nighttime low temperatures, forecasters really need to be aware of patterns in which multiple factors conspire to either greatly favor nocturnal cooling (favoring a chilly night), or conspire to work against nocturnal cooling (keeping the night relatively warm or mild). We'll start with the first scenario: When several factors come together to maximize nocturnal cooling, MOS has an established bias of predicting nighttime temperatures that are too high (so you should consider predicting something on the low side of MOS or cooler). The NBM can have its own struggles in these situations, but again, not as consistently.

Forecast Tips

Be on the lookout for the following conditions (I recommend committing them to memory), which conspire to maximize nocturnal cooling...

  • the nighttime sky is clear
  • winds are light
  • dew points are relatively low
  • there's a snow cover

To see these conditions in action, check out the short video below (2:23), which outlines the big picture set-up for a very cold night in Iowa.

Prime Nocturnal Cooling Conditions (2:23)

Prime Nocturnal Cooling Conditions
Transcript: Prime Nocturnal Cooling Conditions (2:23)

Let’s look at an example of a big-picture set up that strongly favored nocturnal cooling, and a very cold night overall. To do so, we’ll focus on a late December night over the state of Iowa. Dynamic model guidance was unanimously showing a strong high-pressure system to be anchored right over Iowa at 12Z on December 29 – this happens to be a GFS model forecast for mean sea-level pressure, 850-mb temperatures, and 6-hour precipitation. Also note the southward plunging of the 850-mb isotherms over the eastern U.S., with the -10 degree Celsius isotherm as far south as the North Carolina border, indicating a very cold air mass plunging into the eastern U.S. on the eastern flank of the high. Under the center of the high, we often expect subsidence to lead to clear skies, and a weak pressure gradient suggests light or calm winds.

Looking at the surface map from 12Z on the 29th, it’s very clear that this was a cold, dry air mass. Temperatures over Iowa ended up near or below zero, with dew points running a bit below zero. Such low dew points in the region are the hallmark of a continental Arctic air mass.

To add another piece of the big picture puzzle, the region was also covered with snow, as shown on this North American snow and ice chart from the morning of December 29. Much of the northern half of the United States had at least some snow cover.

And, the snow cover over Iowa was particularly deep. The morning snow depth analysis showed a snow pack over the western half of the state that was largely around 10 inches deep at least, with some areas of northwestern Iowa having a snow depth of 20 inches or more. But, the entire state had a snow pack that was at least a few inches deep. So, recapping the big picture, there was an unseasonably strong high-pressure system centered over Iowa, setting the stage for light winds and clear skies. Dew points were markedly lower than average, which is symptomatic of an arctic air mass. And, there was an unusually deep snow pack over the region.

This is an ideal recipe for a cold night, and for the development of nocturnal inversions. This model analysis sounding from 12Z on December 29 at Ames, Iowa, located near the middle of the state shows that a deep nocturnal inversion is exactly what developed. Astute forecasters in this situation could have expected net radiational cooling to really drive temperatures downward.

Credit: Penn State

Clear skies, light winds, and low dew points should cause forecasters to expect net radiational cooling to really drive temperatures downward, but the presence of a thick snow pack really enhanced the nocturnal cooling potential by further deepening the radiation losses in a couple of ways. First, the top of the snow pack can be get quite cold since snow is an effective insulator (a snow pack greater than about two inches seals off any emissions from the ground). Secondly, the snow pack can cause dew points to decline overnight as water vapor in the layer of air near the ground deposits (changes phase from vapor to ice) onto the snow pack. While net deposition causes the release of some latent heat (remember the energy staircase (opens in a new window)), it's not enough to counteract the fact that less water vapor in the air means even less infrared radiation for the snow to absorb. In the Iowa case, all of these factors screamed "major nocturnal cooling" and the main big-picture message was "cold, cold, cold!"

Forecasters predicting low temperatures for the morning of December 29 at Ames, Iowa in this case were faced with a pretty large range in statistical guidance -- low temperature predictions range from 2 degrees Fahrenheit to -6 degrees Fahrenheit. Given the large range in forecast lows, considerable uncertainty existed, but with the big-picture set up, a wise initial "take" would have been to favor guidance predicting a colder night and hedge forecasts closer to the lower end of guidance (closer to -6 degrees Fahrenheit).

The early morning METARS at Ames (opens in a new window) on December 29 reveal that the low temperature was -20.6 degrees Celsius (-5 degrees Fahrenheit), which occurred some time between 12Z and 18Z. If you're wondering why the low happened after 12Z, keep in mind that sunrise at Ames didn't occur until 7:42 A.M. CST (1342Z) on this date, so forecasters that hedged toward the low end of statistical guidance did very well. For what it's worth, a forecast lower than -7 degree Fahrenheit could have been considered somewhat risky since that would have been more than 10 degrees Fahrenheit below the warm end of guidance (and MOS / NBM Day 1 temperature forecasts only miss by more than 10 degrees about 1-2 percent of the time). Could the temperature have dropped to -8 degrees or lower? Sure, but it wasn't a likely scenario.

In addition, if you look at the trend in dew points on the relevant meteogram for Ames, IA (opens in a new window) (the green trace on the meteogram's first plot below), you'll see signs of net deposition overnight. Dew points bottomed out around 14Z on December 29, despite the fact that winds were light or calm for much of the night (meaning little dry advection, consistent with the town's location near the center of a high pressure system). Ultimately, this situation marked an exception to the "crusty old forecasting rule" that dew points near sunset on the previous day would be a reliable indicator of overnight temperature. Net deposition of water vapor onto the snow pack helped lower dew points overnight, paving the way for additional cooling.

The presence of snow cover at night can cause other forecast complications that merit consideration. A thick snow pack obviously favors more nocturnal cooling when clear skies, light winds, and low dew points are present, but snow cover can have very different impacts on MOS and NBM forecasts. With MOS, you have to consider a city's climatology (because MOS equations were developed based on past observations). You may not want to go lower than MOS when all four of the above conditions are in play at a site where snow cover is "climatological," because lots of cases with a snow pack would be in the data set used to derive the MOS equations. For example, at which city would all of the four nighttime conditions lead to the bigger MOS error for the observed low temperature in January -- St. Louis, Missouri, or Hibbing, Minnesota (opens in a new window) (a town very close to International Falls, the "Icebox of the Nation")?

Table containing the long-term average number of days with snow cover of at least one inch at St. Louis, MO and Hibbing, MN superimposed on a photograph of the snow-covered landscape around the Gateway Arch in St. Louis

The snow-covered landscape around the Gateway Arch at St. Louis on December 9, 2005. (Inset) The long-term average number of days with snow cover greater than, or equal to, one inch at St. Louis, Missouri, and Hibbing, Minnesota.
Credit:  NOAA

The answer, of course, is St. Louis, because snow cover is not climatological there (check out the climatology of snow cover at St. Louis and Hibbing above). Thus, there may be winter nights at Hibbing when all four conditions are present, but MOS has a pretty good handle on the nighttime low. In contrast, you should probably lower MOS's nighttime minimum at St. Louis when the sky is clear, dew points are low, winds are light and there's at least one inch of snow on the ground. The bottom line here is that at locations where snow cover is not climatological the historical records probably contain too few cases of snow cover for MOS regression equations to handle these kinds of statistical outliers. It pays to know your climatology!

NBM performance can also suffer with snow cover because dynamic and AI model two-meter temperature forecasts can vary widely (some can be too warm, some can be much too cold), sometimes yielding a large range of solutions, and the bias-corrected, weighted mean of all the forecasts based on their recent performance might not be the best forecast. This is especially true if the snow cover is fresh (where most or all of the previous month had no snow cover). So, while we can't generalize and say that the NBM tends to be too warm in snow cover situations (anecdotally, I've actually seen some situations where it was several degrees too cold), its forecasts still merit some extra scrutiny if the rest of the big picture favors good nocturnal cooling conditions.

Of course, sometimes snow cover isn't in the picture -- especially in the warm season. Still, the presence of clear skies, light winds, and low dew points might yield a night colder than guidance predicts, depending largely on dew points and the daytime high observed earlier during the previous afternoon. Unseasonably low dew points and a below-average high temperature could result in a cooler night than expected, assuming the sky is clear and winds are light. In some situations, you might be able to look upstream to find an analogous city to see how temperature guidance performed under similar conditions.

Looking at forecast soundings valid just before sunrise can be a useful way to assess the impacts of nocturnal cooling conditions. Remember one of the forecast tips from earlier in this lesson: If you see well-developed nocturnal inversions on forecast soundings around sunrise, that's a sign that the model is predicting conditions to favor decoupling, which goes along with significant nocturnal cooling. But, you should always perform an internal check and ask yourself, "Does the big picture support nocturnal cooling conditions that would create strong nocturnal inversions before sunrise?"

Now, what about situations in which big-picture factors conspire to work against nocturnal cooling? When conditions work strongly against nocturnal cooling, MOS has an established bias of predicting nighttime temperatures that are too low (so you should consider predicting something on the high side of MOS or warmer). The NBM can have its own struggles in these situations, but again, not as consistently.

Forecast Tips

Be on the lookout for the following conditions (I recommend committing them to memory), which conspire to minimize nocturnal cooling...

  • when the sky is cloudy, but no precipitation occurs--particularly on nights with a low overcast
  • when it's breezy or windy, but no precipitation occurs--particularly when winds have a southerly component

I emphasize here that either one of these bullets, acting alone, can also result in small model errors (with MOS's nighttime temperatures being slightly too low as long as it doesn't precipitate). Yes, a cloudy night is warmer than a clear night, and a windy night is warmer than a calm night, all other factors being equal (clouds are like space heaters, and mechanical eddies mix warmer air toward the ground), so you might consider tweaking statistical guidance upward on a windy night with no precipitation or a cloudy night with no precipitation. The major potential for large errors exists, however, when both bullets are at work. MOS definitely has a cool bias on a cloudy, windy night with no precipitation (MOS forecasts are typically too low).

For example, let's look at a very mild April night across the Ohio Valley and eastern Great Lakes. The short video below (2:25) describes the big picture.

Cloudy, Breezy, Mild Nights(2:25)

Cloudy, Breezy, Mild Nights
Transcript: Cloudy, Breezy, Mild Nights (2:25)

Let’s look at a very warm night by April’s standards in the Ohio Valley and eastern Great Lakes. The 09Z surface analysis from April 7 showed that this region was located in the warm sector of a pair of relatively weak low-pressure systems centered over Missouri and Iowa. The pressure gradient was rather modest in the warm sector, but still enough to create steady southwest breezes overnight across Ohio.

We’ll focus in on the conditions at Cleveland, Ohio, with the overnight METARS showing clouds gradually increasing. The night started with just scattered high clouds, but by 05Z we had an instance of broken mid-level clouds at 12,000 feet, and a few hours later, clouds were broken at 6,000 feet and below, with a high overcast. So by dawn, cloud ceilings had become pretty low.

The meteogram from 13Z on April 6 to 13Z on April 7 also showed persistent southwesterly breezes, with gusts regularly more than 20 knots. With increasing cloud cover, lowering cloud bases, and steady southerly breezes, the overnight low temperature was a balmy 73 degrees Fahrenheit, which is extremely warm for an April night in Ohio.

Another key ingredient in keeping the night so warm was a lack of precipitation. 24-hour forecasts valid at 12Z on the morning of the 7th consistently showed that rain with the front would stay to the west of Cleveland in the 6 hours leading up to 12Z – here’s the NAM and corresponding GFS forecasts. The lack of precipitation allowed the warming influence of clouds and southerly winds to dominate.

To show the impacts that precipitation can have on what would otherwise be a very mild night, check out this meteogram from Guthrie, Oklahoma taken about a week before the Cleveland case. This was another case with considerable southerly breezes, with 20-30 knot gusts overnight, along with increasing cloud cover. Temperatures slowly trickled down to 69 degrees and remained steady for a few hours overnight…until a heavy thunderstorm rolled through and temperatures plummeted a full 20 degrees down to 49 degrees Fahrenheit at 12Z. Last-minute evaporational cooling of falling precipitation really changed the tenor of the night!

Credit: Penn State
Lesson Learned: Before making any significant adjustment to statistical guidance (like forecasting nighttime temperatures higher than MOS on a cloudy, breezy night), make sure precipitation is out of the picture. When precipitation enters the picture, all bets are off.

In the Cleveland example shown in the video, the actual overnight low temperature was 73 degrees Fahrenheit, but forecasts from available statistical guidance run at 12Z on April 6 were in the low 60s! A double-digit error. Yikes! The fact that gusty winds had a notable southerly component and cloud ceilings became increasingly low (increasing downwelling infrared) made for prime conditions to keep temperatures elevated overnight. But, I can't overemphasize the importance of a lack of precipitation. The Guthrie, Oklahoma example in the video (where a thunderstorm arrived before dawn and temperatures plunged by 20 degrees Fahrenheit) made for a scenario where statistical guidance (which had been predicting a low temperature of 60 - 62 degrees Fahrenheit) looked like it was going to be much too cool, but thanks to last-minute evaporational cooling, statistical guidance quickly became much too warm as temperatures plunged into the upper 40s!

As we wrap up our look at problematic patterns for statistical guidance, it's important to keep in mind that the patterns described in this section and the previous section are merely those that are the most likely to result in large forecast errors. MOS has established biases in these patterns (usually too warm or too cool), whereas the NBM does not, even though it's forecasts may also struggle. So, as a forecaster, it's critical that you stay alert for the patterns I've described, as well as think through your temperature forecasts in terms of major temperature controllers (discussed earlier in this lesson).

But, overall, statistical forecasts from the NBM and MOS tend to be very reliable. If all aspects of a particular problematic pattern are present, that's when larger forecast errors are on the table, and forecasters might consider large adjustments to statistical guidance. In the scheme of things, these cases are relatively rare (occurring in about 10 percent of short-range forecast situations or less), but they provide opportunities for forecasters to make significantly better forecasts than statistical guidance. More commonly, some, but not all, characteristics of a troublesome pattern are present, and it's less likely that statistical guidance will have a large error. So, it's much more common that a small or modest adjustment might be needed at most.

Now that we've covered weather patterns when statistical guidance might be highly questionable, we'll wrap up the lesson with a few "old school" forecasting techniques that can help you make smaller tweaks to improve upon statistical guidance in situations when it typically performs fairly well. Up next, we'll investigate the "Delta Method." Read on.

sas405

The Delta Method

The Delta Method

Prioritize...

When you've finished this page, you should be able to identify situations when using the Delta Method for temperature forecasting is appropriate, and you should be able to execute the Delta Method properly using dynamic model grid interpolations (including determining whether you need to incorporate 850-mb temperatures into your method based on the depth of the well-mixed boundary layer on a skew-T).

Read...

You've heard the message many times by now that specific forecast variables (two-meter temperatures, dew points, etc.) are not consistently well forecast by the dynamic models. That's why meteorologists have spent decades developing statistical forecasting products like MOS and the NBM! But, here's a bit of a revelation: Dynamic models tend to predict the change in crucial meteorological variables more accurately than they forecast the variable itself. A strategy called the "Delta Method" takes advantage of this capability. By the way, delta is the Greek letter that meteorologists use to represent a change in a variable. A change in temperature, for example, would be written ΔT, which is pronounced "delta T".

The Delta Method is merely a quantitative form of modified persistence, a technique you initially learned about in your previous studies. If you recall, modified persistence forecasts are useful when you only expect subtle changes in day-to-day weather (the same air mass remains in place, etc.). Quantitatively speaking, we start with persistence (today's high temperature, for example) and add any changes (sometimes called the correction factor) that we expect to occur tomorrow. If we're forecasting high temperatures, for example, the possible correction factors that meteorologists typically weigh include adiabatic changes (upsloping / downsloping winds), diabatic changes (less / more solar heating), and advection. To properly apply the Delta Method, however, it's critical that the situation call for a modified persistence approach.

When to use the Delta Method ...

Executing the Delta Method isn't hard, but it's important that we make good decisions about when to apply it. We have to find situations when a modified persistence approach is called for. To apply the Delta Method, you should make sure that all of the following conditions are met:

  • Sky cover should be very similar on both days. Otherwise, you're simply comparing apples and oranges. Sky cover must be similar so that the "radiative character" of both days is similar (amount of incoming solar radiation, downwelling infrared radiation from clouds, etc.). Take a look at these two visible satellite images from consecutive August days and focus in on Des Moines, Iowa (opens in a new window).

The visible satellite images from 21Z on August 3 and 4, 2006.

The visible satellite images from 21Z on August 3 and 4, 2006 show that the "radiative character" of the afternoon at Des Moines, Iowa was similar on both days (sunny skies prevailed).
Credit: NCAR

Clearly, the "radiative character" was remarkably similar in Des Moines on these two days. Significant cloud cover was really nowhere nearby on either afternoon. Skies need not be sunny on both days (as in this example), but sky cover should be similar (roughly similar percentages of sunshine during the day).

  • On both days, the wind directions should be similar at coastal locations (or, to a lesser degree, near mountains). Probably the most egregious breakdown of the Delta Method occurs near the coast, where a fairly abrupt shift to onshore or offshore winds could be a problem. That's because the models don't always accurately capture the influence of maritime air at low levels, so be very cautious when you apply the Delta Method near the coast. A shift in wind direction near mountains can have some adverse effects on the reliability of the Delta Method, although high-resolution computer models tend to do fairly well in situations when there's a transition from upslope to downslope flow (or vice versa).

In general, I would shy away from using the Delta Method during frontal passages. I'm not saying that the Delta Method will always fail in frontal passage situations, but frontal passages make it much less likely that the two above criteria will be satisfied. Thus, I'd be hesitant to apply it in such situations without seriously weighing the pros and cons, which requires some experience.

How do we execute the Delta Method? In order to successfully execute it, I need to introduce a dynamic model product called a grid-point interpolation. In a nutshell, model output is "interpolated" (estimated) from nearby model grid points to exact locations of interest, like specific airports. Let me emphasize: the data presented in grid interpolations comes directly from the dynamic models. It's the same data that's visually displayed on model forecast maps, but grid interpolations eliminate the guesswork needed to visually estimate specific values at specific points on a map. The model forecast data is simply interpolated from the nearest model grid point to a nearby location (usually an airport).

The formats of grid interpolations can vary from website to website. For example, compare this GFS grid interpolation for New Orleans (KMSY) (opens in a new window), which has forecast time increasing from left to right, with the grid interpolation below for Baltimore, Maryland (KBWI), which has forecast time increasing from top to bottom. But, if you take a moment to get your bearings, model grid interpolations are fairly self-explanatory to interpret, as long as you pay attention to the forecast parameters and units.

The GFS grid interpolations for temperature at Baltimore, Maryland, based on the 12Z run on June 21, 2022.

The GFS grid interpolations for Baltimore, Maryland, based on the 12Z run on June 21, 2022.
Credit: @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

On the grid interpolation above, the "VALID" column has the date / time (in UTC) that forecasts are valid, so for example "21/15" means the forecast was valid on June 21 at 15Z. There's a lot of data available in grid interpolations, but I'll briefly highlight the variables we'll most commonly use this semester: forecasts of two-meter temperatures (opens in a new window) (in degrees Fahrenheit), 850 mb temperatures (opens in a new window) (in degrees Celsius), and three- and six-hour precipitation (opens in a new window) (in inches). For example, at 18Z on June 22, the GFS predicted two-meter temperature at KBWI was 80 degrees Fahrenheit, the predicted 850-mb temperature was 17 degrees Celsius, and the three- and six-hour precipitation totals ending at that time were 0.06 inches and 0.23 inches, respectively (see annotations (opens in a new window)). It's important to remember that the 3- and 6-hour precipitation totals you see are for the 3- and 6-hour periods ending at the valid time.

There's one quirk to model grid interpolations that you should be aware of. Different websites may have slightly different grid interpolation forecasts for a given airport at any given time. How is that possible? NCEP distributes raw model data at varying resolutions, all the way from the maximum resolution of the model, to reduced resolutions often used by those who want to save computer processing power. If two different websites are processing the raw model data at different resolutions, the values in the grid interpolation will be different. Don't worry too much, though --the differences are usually pretty minor.

The Delta Method for Daytime High Temperature

If the necessary criteria discussed previously are met, follow these steps to make a sound forecast based on the Delta Method in a modified persistence pattern:

  • Estimate the high temperature for today (obviously, you should ideally wait until late afternoon for the best estimate).
  • Identify the maximum two-meter temperatures predicted for today and tomorrow in the GFS grid interpolations (usually at 18Z, 21Z, or 00Z depending on time zone and season). Calculate the difference (the "delta") between the two (tomorrow's GFS maximum two-meter temperature minus today's predicted GFS maximum two-meter temperature), and add this value to today's estimated high temperature.
  • Repeat this process using the two-meter temperatures from any other corresponding grid interpolations available. If no others are available, you could perform the same technique using the mean two-meter temperature forecasts from ensemble plumes.
  • Find the average of these results (consensus).

The consensus you determined in the last bullet point is the Delta Method forecast if the planetary boundary layer (PBL) won't be well mixed up to 850 mb. However, if forecast soundings valid for tomorrow afternoon around the time of maximum temperature show a well-mixed PBL that reaches to at least 850 mb, then you must also perform the Delta Method at 850 mb. In cases where you perform the Delta Method at 850 mb, too, you should average your consensus results using two-meter temperatures and 850-mb temperatures to arrive at your final "Delta Method" result. Of course, you then should compare the Delta Method's forecast with statistical guidance like MOS and the NBM. Base your final decision whether to tweak statistical guidance on the overall big picture.

If you're wondering why we have to perform the Delta Method for two-meter temperatures and 850-mb temperatures when the PBL is well-mixed up to 850 mb (or beyond), remember that layers become well-mixed (opens in a new window) as convective and mechanical eddies mix air near the top of the PBL toward the ground. Such mixing means that changes in 850-mb temperatures can get transferred to the Earth's surface as environmental lapse rates tend toward dry adiabatic. Thus, we can identify the depth of the well-mixed PBL by looking at soundings (or forecast soundings), as in the examples below.

Idealized soundings showing a well-mixed PBL up beyond 850 mb (left) and a well-mixed PBL that doesn't extend up to 850 mb (right).

The image on the left shows a sounding in which the well-mixed PBL extends up beyond 850 mb (note the layer with a dry adiabatic lapse rate). In this case, we'd want to perform the Delta Method using both two-meter and 850-mb temperatures. The image on the right, however, shows a shallower well-mixed boundary layer that does not extend up to 850 mb, so we'd avoid performing the Delta Method with 850-mb temperatures. 
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

When your forecast soundings show a dry adiabatic layer (or superadiabatic layer) that extends from the surface up to 850 mb or beyond, you'll need to incorporate 850-mb temperatures into your Delta Method. Generally speaking, you'll less often incorporate 850-mb temperatures into your Delta Method in the winter time because the boundary layer extends up to 850 mb less often. In such cases, if you're looking for a second opinion, you may execute the Delta Method at, say, 925 mb, if the boundary layer is well-mixed up through that level.

The Delta Method at Des Moines, IA

Let's look at an example of the Delta Method in action at Des Moines, Iowa from back-to-back August days, in the short video (4:41) below:

Let’s run through an example of the Delta Method for daily high temperature. For this example, we’ll focus on Des Moines, Iowa on back-to-back August days. On both August 3rd and August 4th, the radiative character was similar because skies were sunny across central Iowa, as these visible satellite images from each afternoon show.

Des Moines happened to be near the center of a continental Polar air mass, marked by Canadian high pressure on both days – here’s the 21Z surface analysis on August 3rd.

And here’s the surface analysis from 24-hours later on August 4th. So, this was a great set-up for a modified persistence approach and the Delta Method – skies were sunny on both days, and there was no change in air mass. We also don’t have any concerns in Des Moines about wind directions along the coast or in mountainous areas.

If we were making a forecast for August 4th late in the afternoon on August 3, the best estimate we could make for August 3rd’s high temperature would be from the hourly METAR observations. The highest temperature we can find is 29.4 degrees Celsius, which happens to be 85 degrees Fahrenheit. Just based on the data we have here, we don’t know for sure whether that’s the actual high on August 3rd, but it’s the best estimate we can make, so we’ll use it for the basis of our Delta Method calculation.

Now we need to calculate our Delta, or our predicted change between the 3rd and the 4th. For brevity, I’m only going to use one model for this calculation, but whenever possible, you should use more than one model if you can. We’ll use these grid interpolations, which were run at 12Z on August 3.

The highest 2-meter temperature predicted for August 3rd was 84 degrees, which shows up at both 18Z on the 3rd and 00Z on the 4th.

Meanwhile, the model’s maximum predicted 2-meter temperature for the 4th was 88 degrees Fahrenheit, which shows up at 18Z on the 4th. That marks a change of 4 degrees Fahrenheit between the two days.

So, to predict our high temperature on the 4th with the Delta Method using 2-meter temperatures, we would start with our estimated high from August 3rd, which was 85 degrees Fahrenheit, and we would add the predicted change in 2-meter temperatures, which was 4 degrees, and we end up with a prediction of 89 degrees.

Now, since these were sunny days during the warm season, there’s a good bet that the boundary layer would become well-mixed up to 850 mb. We’ll use the observed soundings from Omaha, Nebraska to take a look, since Des Moines is not an upper-air observation site. The 00Z skew-T from 00Z on August 4 shows a slightly superadiabatic layer from the surface up to about 925 mb, with a dry adiabatic profile up to 850 mb. This would suggest vigorous mixing up to 850 mb.

The 00Z sounding on August 5 shows a similar picture, although the dry adiabatic profile extends even a bit higher, up to about 825 mb. So, since boundary-layer mixing extended up to at least 850-mb on each day, we need to execute the delta method at 850 mb.

Looking back at our model grid interpolations, the highest 850-mb temperature on August 3rd was 18 degrees Celsius, here at 00Z on the 4th. The highest value for the afternoon into the evening of the 4th was 19 degrees Celsius. That marks a change of 1 degree Celsius, which is equivalent to 1.8 degrees Fahrenheit. We’ll round it up and call it 2 degrees Fahrenheit.

So applying the Delta Method using 850-mb data looks like this: We start again with our observed high temperature of 85 degrees Fahrenheit, and add our model-predicted change in 850-mb temperature, which was 2 degrees Fahrenheit, to get a prediction of 87 degrees Fahrenheit.

To recap, our Delta Method based on 2-meter temperatures resulted in a forecast of 89 degrees Fahrenheit, while our 850-mb Delta Method resulted in a forecast of 87 degrees. To get our overall forecast, we average those two results, to get a Delta Method forecast of 88 degrees.

Again, I only performed this method using one model, and in reality, you would want to use two or more if they’re available. And, you wouldn’t have observed soundings at your disposal for your forecast day, so you would need to use forecast soundings to assess how high boundary-layer mixing will extend. But, hopefully, this example shows you how to execute the Delta Method.

Credit: @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

As executed in the video, the Delta Method produced a high temperature forecast of 88 degrees Fahrenheit on August 4. So, what happened? Well, here's the climate summary for August 4 (opens in a new window). The forecast worked out pretty well, although we were two degrees too high. Still, that's a pretty good forecast considering we didn't use any statistical forecast guidance! For what it's worth, the available statistical forecasts on this day ranged from 88 to 90 degrees Fahrenheit; so, the Delta Method correctly pointed us toward the lower end of that range and would have helped us minimize error.

Forecast Tips

  • In the final analysis, you should blend the high temperatures that you get from using the available grid interpolations at two meters (and 850-mb, if necessary). Then compare these results to statistical temperature guidance, and blend the message from statistical guidance with the message from the Delta Method.
  • In situations where day-to-day changes in the 850-mb temperatures are large (on the order of several degrees Celsius) you might consider averaging the 850-mb temperatures for each daytime period and then compare the averages to arrive at a ΔT. Feel free to experiment, but I would keep a diary and record your results.
  • You may occasionally find performing the Delta Method for dew points to be helpful (just use two-meter dew point forecasts (opens in a new window)). For example, the arrival of drier air at night may have implications on the overnight low temperature.
  • Grid interpolations are available via the Texas A&M Weather Interface (opens in a new window) (choose "Model Data" as your Product Category, and then grid interpolations as your Data Type). Grid interpolations for cities of interest this semester are also available on the Penn State Grid Interpolations (opens in a new window) page.

The Delta Method at Night

Assuming it's a modified-persistence forecast, and assuming weather conditions early tomorrow morning will be similar to what they were this morning, we can also use the Delta Method to predict the nighttime low temperature for tonight.

The Delta Method for Nighttime Low Temperatures

  • Determine this morning's low temperature by looking at the relevant METARS.
  • Perform the Delta Method using two-meter temperatures from available grid interpolations (usually at 12Z, depending on the time zone and the season). You could also try using mean two-meter temperature forecasts from ensemble plumes.
  • Average your results to come up with a consensus.

As a general rule, we do not use 850-mb temperatures on the model grid interpolations when we perform the Delta Method at night, unless, of course, forecast soundings indicate that there will be a well-mixed PBL that reaches to at least 850 mb before dawn tomorrow morning (which rarely happens).

The Delta Method might seem a bit mechanical, but its main advantage is that you don't need to know or address specific model issues such as the pitfalls of the parameterizations, modeling the influence of maritime air, etc., to accurately predict temperatures. All you really need to know is "what happened" (today's or yesterday's high) and the correction factor. The most important issues are knowing when to apply it, and of course, performing the arithmetic correctly.

Up next, we'll look at one last temperature-forecasting strategy, which uses thickness forecasts from the dynamic models. But, before you move on, check out the Quiz Yourself section below to get some practice in executing the Delta Method.

Quiz Yourself...

In order to get some practice with the Delta Method, check out this case from June 9-10, 2015 at Little Rock, Arkansas (KLIT) (opens in a new window). On the afternoon of June 9, Little Rock was located near the center of a weak high pressure system (check out the 21Z surface analysis (opens in a new window)), which marked the center of a continental Polar (cP) air mass. The air mass had been modifying (warming) over the southern tier of the U.S. thanks to strong June solar heating, so it was far from chilly (temperatures were in the 90s). Near the center of the high, the afternoon of June 9 was fairly calm and mostly sunny in Little Rock, as evidenced by the 2045Z visible satellite image (opens in a new window). Only scattered boundary-layer cumulus clouds had formed over central Arkansas.

The center of the high was predicted to drift eastward on the 10th, and indeed, that's what happened (here's the 21Z surface analysis from June 10 (opens in a new window)). Still under the influence of high pressure, sky conditions looked to be similar on the 10th, which was supported by the forecast soundings from the 18Z run on June 9, valid at 21Z on June 10 (NAM forecast sounding (opens in a new window); GFS forecast sounding (opens in a new window)). Note that relative humidity was predicted to be low throughout the troposphere, with the exception of a thin layer where temperatures and dew points came close together near the top of the boundary layer. This pattern (a dry, well-mixed boundary layer with a thin layer of high relative humidity near its top) is a common sign that the models are predicting some boundary layer cumulus clouds.

To summarize, on both days, Little Rock was under the influence of the same high-pressure system (no change in air mass). Both days had a weak pressure gradient, which corresponded to light winds, and the 10th was likely to be a mostly sunny day similar to the 9th. This information should tell you that a modified persistence approach was appropriate here, and therefore, we can use the Delta Method to forecast the high temperature at KLIT on June 10. In addition to the information I've already given you, here's the rest of the data you'll need:

Question #1

What was the high temperature observed at KLIT on the afternoon of June 9?

Answer: 34.4 degrees Celsius (94 degrees Fahrenheit) as evidenced by the 1-group in the 00Z observation (10344).


Question #2

Perform the Delta Method for the high temperature on June 10 using two-meter temperatures from the GFS and NAM grid interpolations from the 12Z run on June 9. What are your results?

Answer:  96 or 97 degrees Fahrenheit. The maximum two-meter temperature predicted by the GFS for June 9 was 90 degrees (at 18Z). The maximum on June 10 was predicted to be 93 degrees Fahrenheit (at 18Z).

GFS 2-m Delta: 93 degrees - 90 degrees = 3 degrees

The maximum two-meter temperature predicted by the NAM for June 9 was 89 degrees (at 18Z), while the maximum on June 10 was predicted to be 91 degrees (at 18Z). 

NAM 2-m Delta: 91 degrees - 89 degrees = 2 degrees

If we average our two "deltas" we get 2.5 degrees, so adding that to 94 degrees (the high from June 9) gives us 96.5 degrees. Thus, the Delta Method using two-meter temperatures supported a forecast of 96 or 97 degrees Fahrenheit.


Question #3

In this situation, do you also need to perform the Delta Method using 850-mb temperatures? Why or why not? Where is the top of the boundary layer on each forecast sounding?

Answer:  Yes, we need to perform the Delta Method using 850-mb temperatures because both forecast soundings show a well-mixed boundary layer extending up to at least 850 mb. On the NAM forecast sounding, the dry adiabatic (well-mixed) layer extends up to about 825 mb. On the GFS forecast sounding, the dry adiabatic (well-mixed) layer extends up to about 800 mb.


Question #4

Perform the Delta Method for high temperature on June 10 using 850-mb temperatures from the GFS and NAM grid interpolations. What are your results?

Answer:  96 degrees Fahrenheit. At 00Z on the 10th (early evening on June 9), the GFS predicted 850-mb temperatures of 18 degrees Celsius. At 00Z on the 11th (early evening on June 10), the GFS predicted 19 degrees Celsius at 850 mb.

GFS 850-mb Delta: 19 degrees Celsius - 18 degrees Celsius = 1 degree Celsius (equivalent to 1.8 degrees Fahrenheit, which we'll round to 2 degrees Fahrenheit).

At 00Z on the 10th (early evening on June 9), the NAM predicted 850-mb temperatures of 19 degrees Celsius. At 00Z on the 11th (early evening on June 10), the NAM predicted 20 degrees Celsius at 850 mb. 

NAM 850-mb Delta: 20 degrees Celsius - 19 degrees Celsius = 1 degree Celsius (equivalent to 1.8 degrees Fahrenheit, which we'll round to 2 degrees Fahrenheit).

If we average our two "deltas" we get 2 degrees Fahrenheit, so adding that to 94 degrees (the high from June 9) gives us 96 degrees. Thus, the Delta Method using 850-mb temperatures supported a forecast of 96 degrees Fahrenheit.


Question #5

Determine a final Delta Method forecast based on a consensus of your results from Questions 2 and 4. What is the final high temperature forecast based on the Delta Method?

Answer:  96 degrees Fahrenheit. The Delta Method based on two-meter temperatures yielded 96.5 degrees Fahrenheit (question 2), while the Delta Method for 850-mb temperatures yielded 96 degrees Fahrenheit (question 4). Averaging these two results gives us 96.25 degrees Fahrenheit, which we'll round to 96 degrees Fahrenheit.

For what it's worth, the Delta Method wasn't perfect for June 10, but it worked out pretty well. The high in Little Rock was 95 degrees, so we were 1 degree too high. Not bad! Here's the KLIT METARS from the afternoon of June 10 (opens in a new window). The maximum temperature can be found in the 00Z observation, where the 1-group, "10350" appears. 35 degrees Celsius translates to 95 degrees Fahrenheit. Note also that winds were light (6 knots or less) as they had been on June 9, and sky coverage was consistently "scattered" or "few" with cloud bases of 5,500 - 6,500 feet, Thus, our interpretation of the forecast soundings was right on (sky coverage on the two days was quite similar).

sas405

Analogous Thickness

Analogous Thickness

Prioritize...

By the time you're finished with this page, you should be able to identify situations when using the Analogous Thickness Method for temperature forecasting is appropriate, as well as execute the Analogous Thickness Method in those situations.

Read...

It's been a long journey learning about the controllers of temperatures and specific forecasting strategies. Your journey started way back at the beginning of this program, when you learned about the main controllers of temperature and some basic forecasting strategies like the 850-mb method, climatology, persistence, and modified persistence forecasting (if you need to review any of these, I recommend checking out Lesson 11 of METEO 101 (opens in a new window) (login required)).

In this lesson, we've reviewed some of the main temperature controllers and applied them to specific forecasting situations. We've also gone into more depth about recognizing patterns when statistical guidance may struggle to provide accurate forecasts and strategies for dealing with those patterns (a good analysis of the big picture is essential). As we round out our lesson on temperature forecasting, we have one last specific approach to cover -- the Analogous Thickness Method.

Photograph of Lee Grenci extrapolating the movement of a tropical cyclone.

Lee Grenci employs extrapolation ("the finger method") to track a tropical cyclone on a computer monitor from a bygone era.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

Before I get into the specifics of the Analogous Thickness Method, I want to briefly explain extrapolation (informally, the "finger method"). For example, suppose you noted a line of thunderstorms on a loop of radar reflectivity and you want to forecast its future position. Although it sounds crude, you can determine the distance the line moved by placing two fingers on the computer screen (start and end points) and noting the elapsed time. Then, just calculate an estimated speed and extrapolate the line's future position (here's a finger-measuring example (opens in a new window) involving a cold front). Extrapolation may help you estimate what time a cold front or a line of thunderstorms will reach your city, for example.

I mention extrapolation here not only because it's something that may be useful to you, but it actually has a connection to the Analogous Thickness Method. Both methods require you to track the position of your forecast city relative to particular weather features as they move.

The Analogous Thickness Method

By way of introduction, forecasters typically use the Analogous Thickness Method when the overall pattern is progressive (air masses move progressively from west to east). For starters, forecasters look at tomorrow's progs, noting the location of their forecast city relative to a nearby front, low-pressure system, or high-pressure system. From the progs, forecasters read off the predicted values of the local isobars and 1000-500-mb thickness lines (recall that thickness is a measure of mean temperature in a layer). Then, forecasters look upstream from the forecast city for a location that currently lies (or just recently lay) in a similar position relative to the same weather feature slated to affect the forecast city tomorrow. Essentially, this upstream location serves as an "analog" for the forecast city, and the high (or low) temperature recorded at the analog location can serve as guidance for tomorrow's maximum (or minimum) at the forecast city.

The Analogous Thickness Method for High Temperature

To make a sound forecast based on the Analogous Thickness Method, follow these steps:

  • Estimate the values of sea-level pressure and 1000-500-mb thickness at your forecast location on a forecast prog valid near the likely time of tomorrow's high temperature (usually 18Z or 00Z, depending on time zone and season), and note your city's position relative to the nearest major weather system(s).
  • Find your analog position by determining the region upstream that corresponds to the air mass's recent past or current position. Then use sea-level pressure and 1000-500-mb thickness to pinpoint the analog position. If you're forecasting at a time before today's high temperature occurs, you'll want to find an analog based on yesterday's weather. If you're forecasting after today's high temperature occurs, you can find an analog based on today's weather. The PSU e-Wall has the most recent five days worth of Rapid Refresh model analyses (opens in a new window), which can be useful for finding your analog location.
  • Find the high temperature at your analog position (for yesterday or today, depending on the time of the analysis you're using). If your analog position doesn't correspond to a city with weather records, find the high temperature for the nearest city and make minor adjustments, if necessary.
  • If the elevations of your forecast city and analog city are similar, the high temperature at your analog position is your Analogous Thickness Method forecast. If not, you'll have to adjust for elevation. Assume a standard environmental lapse rate of 3.5 degrees Fahrenheit per 1000 feet.

Not surprisingly, several caveats exist for using the Analogous Thickness Method, and I'll get to them after we look at an example.

Let's suppose that on April 4, your responsibility was to predict tomorrow's high temperature at Pittsburgh, Pennsylvania (the high temperature on April 5). At the time, a ridge of high pressure was slated to move eastward from the Middle West (check out the 00Z model run on the 4th (opens in a new window) and move your cursor over the forecast times at the top of this loop). The prog below (the 18-hour forecast from the 00Z run of April 5, 2005) shows that the predicted 1000-500-mb thickness at Pittsburgh was roughly 554 dekameters, and the sea-level pressure was approximately 1019 mb.

The 18-hour forecast from the 00Z run of the ETA on April 5. 2005.

The 18-hour prog from the 00Z run of the ETA on April 5, 2005. At Pittsburgh, Pennsylvania, the sea-level pressure and 1000-500-mb thickness were, approximately, 1019 mb and 554 dekameters, respectively.
Credit: @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

If we were forecasting in the afternoon on the 4th (or earlier), we'd next look upstream to find yesterday's (April 3, in this case) analog location because the high temperatures for today (April 4, in this case) often haven't yet occurred when you're making your forecast. If we were forecasting later in the day (near 00Z or later), we could find today's analog, however.

In this case, we'll use a model analysis from 18Z on April 3 (below) to find the analog city. The first step is to locate the ridge of high pressure that was ultimately predicted to move near Pittsburgh on the 5th, and find a city that had the same 1000-500-mb thickness value and sea-level pressure on the 3rd at 18Z. In this case, the analog location looks to be west of Cape Girardeau (opens in a new window) in southeast Missouri, where the sea-level pressure is roughly 1019 mb and the 1000-500-mb thickness is 554 dekameters). Next, we look up the high temperature at Cape Girardeau on April 3 (opens in a new window), which was 61 degrees. Since our analog location wasn't exactly in Cape Girardeau, it's possible that the high at our analog location was slightly higher, since the chart of high temperatures suggests that it was slightly warmer to the south and west of Cape Girardeau. We'll call it 62 degrees, although sticking with 61 degrees might still be reasonable.

The RUC analysis at 18Z on April 3, 2005.

The RUC analysis at 18Z on April 3, 2005. At this time, Pittsburgh's analog was located in southeast Missouri (purple dot), near Cape Girardeau (KCGI).
Credit: @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

We're not done yet, though. We have to take elevation into account. The elevation of Cape Girardeau is 335 feet, and the elevation at Pittsburgh is 1223 feet. If we assume an average environmental lapse rate (3.5 degrees Fahrenheit per 1000 feet), then we need to subtract three degrees from the analog location's high of 62 degrees to arrive at the analogous thickness forecast for Pittsburgh on April 5. So in this case, our Analogous Thickness forecast is 62 degrees - 3 degrees = 59 degrees. How did it work out? The official high temperature at Pittsburgh on April 5 (opens in a new window) was 58 degrees. Not too shabby, especially considering that we came to this forecast without using any specific temperature guidance!

Before you get too carried away, the Analogous Thickness Method isn't appropriate for every situation. There are some rules and procedures you must always use or keep in mind when you apply the method of analogous thickness.

Forecast Tips

When applying the Analogous Thickness Method, make sure to keep these caveats and rules in mind:

  • When forecasting the daytime high temperature, compare 18Z to 18Z or 00Z to 00Z.
  • When forecasting the nighttime low temperature, compare 12Z to 12Z.
  • Make sure to note the analog's position relative to synoptic-scale features (low- and high-pressure systems, fronts, etc.). If features intensify or weaken significantly, then locations with equal values of SLP and thickness may not be analogous. Finding true analogs will be difficult and you may want to avoid the technique entirely.
  • Always keep in mind that air masses can moderate (move over snow cover, move over large bodies of water like the Great Lakes, etc.), so you need to adjust accordingly if the situation dictates.
  • In general, you should avoid the technique near strong fronts. Temperature gradients associated with strong fronts are large, so determining the true analogous site may be fraught with difficulty because temperature changes so rapidly over relatively small horizontal distances.
  • Of course, to calculate a consensus, repeat this process using at least two dynamic models, and average your results.

As with the Delta Method, make sure you compare your analogous thickness forecast with statistical forecasts from MOS or the NBM. Then make a final decision based on the overall big picture about whether to tweak statistical guidance or not.

Okay, that wraps up our lesson on temperature forecasting. You now have a complete bevy of tools and strategies at your disposal, and you'll have some opportunities this semester to put them to use. The key is to make good decisions about when to use each technique, which takes some skill and experience in pattern recognition. 

Up next, we'll move on to another critical forecast variable — wind.

sas405

Lesson 5. Forecasting Wind: Controllers and Strategies

Lesson 5. Forecasting Wind: Controllers and Strategies

Motivate...

A Nittany Lion wind vane

A Nittany Lion Weather Vane sits atop the south end zone upper deck at Beaver Stadium on Penn State's University Park Campus. Weather vanes are instruments for showing wind direction.

In his 1965 song, "Subterranean Homesick Blues (opens in a new window)," Bob Dylan famously quipped, "you don't need a weatherman to know which way the wind blows." I'm sure he didn't know at the time that this lyric would become one of the most influential in modern music history (it's apparently a favorite among judges and lawyers (opens in a new window) in legal proceedings). When it comes to weather forecasting, Bob is right -- kind of. You don't need a weather forecaster to know which way the wind blows right now.  Just about anyone can take their own observation, or look at a weather vane (see photograph at the right) if one happens to be nearby. But, you do need a competent weather forecaster to consistently tell you which way the wind will blow in the future.

Wind forecasting is important in many settings. First, think back to the last lesson when we talked about controllers of temperature -- several of them were based on the wind (mechanical mixing of eddies caused by the wind, advection and air trajectories, and orographic effects). So, if we make significant errors in our wind forecasts, the accuracy of our temperature forecasts can be at risk, as well.

Furthermore, wind forecasting is pivotal in many other contexts: Wind is increasingly being used as an energy source via the generation of electricity from wind farms (opens in a new window). Sailors and pilots rely on accurate wind forecasts for obvious reasons. Operators of tall, heavy construction equipment may not be able to use their machinery when wind speeds reach certain thresholds, and even seemingly innocuous wind gusts can pose a danger to children playing in bounce houses (opens in a new window). Of course, when winds really roar, they can pose a risk to life and property, and cause widespread power outages. These are just a few examples, but in a nutshell, being able to predict wind direction and speed accurately is beneficial in many settings. Indeed, we do need weather forecasters to know which way the wind will blow.

In previous certificate courses, you've covered the large-scale forces that determine wind direction and speed (namely, the pressure-gradient force, the Coriolis Force, and friction). You've also studied mesoscale processes that affect the wind, such as sea-breeze circulations and terrain effects. Until now, though, you haven't had to incorporate those concepts into creating specific wind forecasts. That, naturally, will change in this course.

In this lesson, we'll be reviewing some of the basic controllers of wind and examining advanced strategies for incorporating your knowledge of these controllers into making specific wind forecasts. Although our focus in this lesson will be on forecasting wind speeds, you'll soon see that wind direction and wind speed aren't entirely exclusive. Therefore, we will also need to pay some attention to wind direction, even though it's not our main focus. First up, we'll discuss the importance of climatology. Let's get started!

sas405

Clues from Climatology

Clues from Climatology

Prioritize...

In this section, you should be able to discuss the use of wind climatology data (specifically wind roses and monthly record wind speeds) in making prudent forecast decisions for maximum sustained wind speeds. You should also be able to use such data appropriately in your wind forecasts.

Read...

As I've discussed before, having a good handle on a location's climatology will really help you get to "know" your forecast area. Think about it: By collecting data such as a monthly wind rose and the maximum monthly wind speeds, you can learn a lot about typical wind speeds and prevailing wind directions at a particular location during various times of the year. As with temperature, much of the climatology dealing with winds boils down to location, location, location. For example, let's think of some common things that wind roses can reveal:

  • a dominant wind direction (or directions) caused by local terrain effects at a station in a mountainous region
  • dominant wind directions caused by daily land-breeze and sea-breeze circulations at a coastal station
  • a prevailing set of wind directions caused by a fairly regular storm track at a station on the Plains (or anywhere with flat terrain). For example, a predominance of southwesterly winds can reveal that the station spends significant amounts of time in the warm sectors of mid-latitude cyclones that pass to its north during a given month.

So, you can learn a lot about the local meteorology of a station just by looking at the wind rose. As an example, let's look at Juneau, Alaska and its wind rose.

Winds at Juneau International Airport (PAJN)

Before we look at the wind rose for Juneau International Airport (PAJN), we should get a look at the lay of the land. Juneau, Alaska (opens in a new window) is located in the state's panhandle, which borders British Columbia, Canada. Specifically, Juneau International Airport is located northwest of downtown Juneau (opens in a new window). The airport is located at just 22 feet above sea level near the northwest end of the Gastineau Channel (opens in a new window), which separates the mountainous mainland from mountainous Douglas Island.

Not surprisingly, the surrounding topography plays a big role in the winds experienced at PAJN, and a wind rose will help us get a handle on the specifics. Check out the November wind rose (below) for PAJN as an example.

November wind rose for Juneau, Alaska

The November wind rose for Juneau International Airport (PAJN) in Alaska shows that east-southeasterly winds prevail more than any other direction.
Credit: Iowa Environmental Mesonet

What can we take away from this plot? Lots of things, actually. First, it's pretty clear that east, east-southeasterly, and southeasterly winds are more common than any other directions in November. If you just look at the three most common wind directions (the three longest "spokes") corresponding to 100 degrees, 110 degrees, and 120 degrees, they account for a total of nearly 30 percent of the wind observations during the month. Given the local topography around the airport, that's not shocking. Remember that the airport is located in a valley at the northwest end of the Gastineau Channel with high mountains to the northeast and south, particularly. So, winds commonly funnel along the Gastineau Channel toward PAJN from the southeast. Furthermore, it's clear from the wind rose that when fast winds occur (say, faster than 20 miles per hour), they almost always blow from the east, east-southeast, or southeast (marked by the largest maroon sections on the spokes). As air zips right up the channel to the airport, it can maintain its speed better than it can on approach from other directions over the rough terrain.

Will any of this help us come up with a specific wind speed forecast on one particular day? Not really. But, it does give us some important context when making wind forecasts. Knowing the basics from your location's wind rose can help you assess the likelihood of various scenarios occurring, which is especially helpful when great uncertainty exists in a forecast.

A Windy November Day at PAJN?

Forecasters who had to make sustained wind speed predictions for November 17, 2011 at PAJN were faced with a bit of a conundrum. The big picture showed that a surface low was forecast to deepen offshore, south of Juneau. The 30-hour GFS forecast (opens in a new window) of MSLP and six-hour precipitation showed the predicted set up at 18Z on November 17. Note that the forecast prog showed a large pressure gradient around the low (including in the Juneau area). So, your first impression was that it would probably be a fairly windy day.

Dynamic model forecasts largely supported that impression. For example, this 24-hour ensemble forecast (opens in a new window) (valid at 09Z on November 17) showed a probability near 50 percent of winds greater than, or equal to, 20 knots around PAJN at that time. At the extreme, the Uncoupled Surface Layer (USL) model (a proprietary, microscale dynamic model with great skill in wind forecasting) predicted maximum sustained winds of 45 knots in its 12Z run (opens in a new window) and 39 knots in its 22Z run (opens in a new window). Both runs predicted wind directions to blow primarily out of the east-northeast and northeast.

So, the dynamic models sent a strong message for a very windy day, but these extreme predictions should have thrown up red flags immediately. First, the climatology for PAJN told forecasters that the monthly maximum sustained wind speed in November at PAJN was 42 knots, so some dynamic model forecasts were record breaking. By itself that might not be concerning, but it should have been in the context of the wind rose, which showed a lack of historical observations with such fast speeds from the predicted wind directions (east-northeast and northeast). Indeed, the wind rose (opens in a new window) indicated that when winds blow fast (greater than 20 miles per hour), they almost always blow from the east, east-southeast, or southeast (up the Gastineau Channel toward the airport).

What about statistical models? Well, the NBM did not yet exist in 2011, so we'll have to stick with MOS. Check out the GFS MOS table from the 12Z run on November 16 below.

GFS MOS for Juneau International Airport (PAJN) from the 12Z run on November 16, 2011

GFS MOS for Juneau International Airport (PAJN) from the 12Z run on November 16, 2011 predicted very low wind speeds (2 knots or less) at PAJN at synoptic times from 06Z November 17 to 06Z November 18.
Credit: NCEP

Meanwhile, the now defunct NAM MOS (opens in a new window) from the 12Z run on November 16 also predicted modestly light winds (a maximum at synoptic times of 8 knots). Recall that MOS does not predict a maximum sustained wind speed, so forecasters would need to start by adding a few knots to the highest forecast at the synoptic times as a starting point for predicting maximum sustained wind speed. But, even accounting for that, these predictions (especially GFS MOS) seemed to be at odds with the big picture set up of a deepening low pressure system with a fairly strong pressure gradient centered south of Juneau.

So, how does a forecaster make sense out of all of this? Well, MOS's highest wind speed predictions were between 2 knots and 8 knots at the synoptic times, while the dynamic models were predicting much faster speeds. Obviously, you know that over the long haul, statistical models tend to be better performers. But, the difference between MOS and the dynamic models here was huge, and forecasters needed to focus on the big picture and climatology to lead them to a forecast decision.

Given the position of the deepening low south of Juneau, we'd probably expect winds to have some sort of northerly component, right? Remember, winds blow counterclockwise around areas of low pressure in the Northern Hemisphere, crossing local isobars in toward the low at greater crossing angles with greater friction. So, looking back at the GFS forecast valid at 18Z (opens in a new window) on the 17th, the pattern appeared to signal some sort of northeasterly wind blowing at PAJN, likely blowing fairly fast because of a strong (and intensifying) pressure gradient.

The Gastineau Channel

The orientation of the Gastineau Channel has a profound effect on wind speeds and directions at Juneau International Airport.
Credit: Wikimedia Commons

But, climatology sent a strong message that at the airport, fast northeasterly winds (greater than 20 miles per hour) are virtually unheard of because of the local terrain. Right off the bat then, if forecasters believed winds would blow from any direction other than southeasterly (or east-southeasterly), it became clear that the really fast winds supported by the dynamic models seemed really unlikely. Local topography just makes it really difficult for fast winds at PAJN when the trajectory isn't from the southeast or east-southeast. Yet, at the same time, MOS's forecasts were really low for a situation when a strengthening low was passing south of Juneau. So, starting with the high end of MOS and tacking on a few or several knots might be a reasonable approach -- perhaps a forecast of 10, 11, or 12 knots. Such speeds would be consistent with climatology based on the local effects of terrain at the airport and the fact that we did have a strengthening low to the south. Ultimately, winds would probably be on the higher end of climatology for winds with a northerly component. Even forecasts in the low or mid teens wouldn't have been unreasonable.

So, what was ultimately verified? The maximum sustained wind speed on November 17 at PAJN was 11 knots. Forecasters that bought the dynamic models hook, line, and sinker had horrendous forecasts. But, forecasters who studied their wind roses and saw a pattern that didn't favor southeasterly winds did much better! Ultimately, weighing the big picture against climatology was a combination that was hard to beat here, and the (no longer operational) NAM MOS forecast ended up being great! The fact that MOS forecasts were very insightful here shouldn't have been surprising because of their incorporation of past observations into developing the regression equations (so they have a keen "sense" of the impacts of terrain). This example also shows that wind direction and speed aren't totally exclusive, especially in locations where local terrain effects dominate.

Forecast Tips

  • When making your forecasts, always have the climatology of wind directions and speeds in the back of your mind. For maximum monthly sustained wind speeds and gusts, Local Climate Data documents (opens in a new window) can help you get a feel for the upper bound of historical observations at different times of year. It won't pinpoint you to a specific forecast, but it will allow you to assess the likelihood of various scenarios.
  • Keeping the general messages from wind roses and monthly record wind speeds in mind can really help in situations where great uncertainty exists and / or where the weather situation may near climatological extremes. Remember, conditions not depicted on a wind rose can happen, but it obviously takes a rarely-seen set of circumstances to do so.

Don't underestimate the value of knowing your wind climatology! It won't point you exactly to the right forecast, but many times climatology sends us general messages that can help us with our forecasts. Now, it's time to turn back the clock a bit and talk about the fundamental forces that determine wind speed. Read on.

sas405

Back to Basics

Back to Basics

Prioritize...

When you've finished this page, you should be able to discuss why forecasters can't just simply determine a specific wind speed based on the size of the pressure gradient. You should also be able to discuss why a weak pressure gradient may not equate to a day with calm or light winds (especially at low latitudes), and incorporate this idea into your wind forecasts.

Read...

Way back in Lesson 5 of METEO 101 (opens in a new window) (login required), you investigated the major forces that determine wind direction and speed, namely the pressure-gradient force (PGF), the Coriolis Force, and friction. I won't go over the details of all those forces again here, but I do suggest you review that lesson if you need to brush up on the basics. 

Our knowledge of these basic forces is crucial to getting a handle on the wind. For example, it allows us to know that winds will blow parallel to local isobars if the wind is in geostrophic balance (PGF and Coriolis Force are equal in magnitude). It also allows us to make quick big-picture analyses based on patterns of surface isobars, by following general rules like "the wind blows counterclockwise around low pressure in the Northern Hemisphere, crossing local isobars in toward lower pressure because of friction." If you're rusty on making such analyses, you might want to review this interactive wind direction tool (opens in a new window), which allows you to estimate the surface wind at a point based on the pattern of isobars, and then do a self check by clicking on the point. Try it out a few times. If you find yourself struggling, you may need to brush up on your fundamentals, or ask for some clarification.

Of the "big three" forces for and against the wind, it's important to remember that the PGF and friction help determine wind speed (stronger PGF equals faster winds, while friction works to slow the wind). Meanwhile, the strength of the Coriolis Force is determined by the speed of the wind and latitude. Faster wind speeds and higher latitudes yield a stronger Coriolis Force.

Armed with knowledge of these three major forces, a lot of "rookie" weather forecasters wonder, "why can't I just look at a map of surface isobars and come up with a specific forecast for wind speed, since ultimately, the PGF is the driver of the wind?" Several reasons make doing so impossible, one of which we've already covered -- friction. But, you've also learned about other mesoscale factors that can impact wind speed (vertical mixing and momentum transfer, local terrain effects, etc.). In short, the PGF is the main large-scale driver of wind speed, but it's not the only driver.

Determining wind speed is, in fact, more complicated than simply thinking about the PGF and friction. To see what I mean, check out the short video (2:22) below. The interactive force diagram used in the video shows that wind speeds can change merely by changing latitude, without changing the strength of the pressure-gradient force or friction.

I’m going to use this interactive force diagram to show you how wind speed can be affected by factors other than the pressure-gradient force or friction. This tool allows me to adjust the strength of the pressure-gradient force, the Coriolis Force by adjusting latitude, and friction.

We’ll start with a parcel in geostrophic balance, which means that we’ll ignore friction for a moment. This air parcel would move parallel to local isobars with lower pressure on its left thanks to the balance of the pressure gradient force and the Coriolis Force. We know we could increase the parcel’s speed by increasing the pressure gradient force, and that’s marked by the velocity vector getting larger to indicate a faster speed.

But, determining wind speed is more complicated than simply thinking about the pressure-gradient force. If I hold the pressure-gradient force constant, and change latitude look what happens to the velocity vector. The velocity vector gets larger equating to faster wind speeds at lower latitudes, and gets smaller, equating to slower wind speeds, at higher latitudes, all without changing the pressure-gradient force.

This occurs because the Coriolis Force depends on both latitude and a parcel’s velocity, so there are two ways to boost the Coriolis Force – by increasing latitude or increasing wind velocity. But, if you're analyzing parcels at a fixed latitude, then the only way to increase the magnitude of the Coriolis Force is to increase velocity. Therefore, at lower latitudes, for parcels to achieve geostrophic balance, air parcels must move faster than they would at higher latitudes in order to compensate for latitude's smaller contribution to the overall magnitude of the Coriolis Force.

This holds true even if we activate friction. With friction added to the mix, our parcel is no longer in geostrophic balance and crosses local isobars in toward lower pressure. But, if we hold friction and the pressure-gradient force constant, the velocity increases at lower latitudes and decreases at higher latitudes.

The practical application of this relationship is if all other forces are equal, cities at different latitudes will experience different wind speeds.

Credit: Penn State

How is this possible? Remember that the magnitude of the Coriolis Force is determined by a parcel's velocity and latitude. So, two ways exist to boost the magnitude of the Coriolis Force -- increased latitude or increased wind velocity. But, if you're analyzing parcels that exist at a fixed latitude, then the only way to increase the magnitude Coriolis Force is to increase velocity. Therefore, at lower latitudes, for parcels to achieve geostrophic balance (PGF = Coriolis Force), air parcels must move faster than they would at higher latitudes in order to compensate for latitude's smaller contribution to the overall magnitude of the Coriolis Force. The bottom line is that if you picked two cities, one at 25 degrees North latitude and another at 40 degrees North latitude, for example, and all forces were equivalent (PGF, friction, etc.), the wind speeds would not be the same at the two cities. Indeed, the wind speed at the city with a lower latitude would be faster.

An Example

Let's look at an example from at Fort Lauderdale, Florida (KFLL) (opens in a new window) to see how you can use this idea to your advantage as a weather forecaster. The surface analysis (below) showed a broad high pressure system over the eastern U.S., centered along the Ohio River.  Florida was on the southern flank of this sprawling high, and a very modest pressure gradient existed over the region. One might even be tempted to describe the pressure gradient over South Florida as fairly "weak."

The 15Z surface map on October 5, 2011.

The 15Z surface map on October 5, 2011 showed a sprawling area of high pressure over the eastern U.S. and an unremarkable pressure gradient over Florida.
Credit: Weather Prediction Center

Just glancing at the surface map may make you think that with a modest (or even "weak") pressure gradient over Florida, winds in Fort Lauderdale would probably be fairly weak around this time. But, not so! Check out the METARS from KFLL (opens in a new window) from 15Z on October 5 through 01Z on October 6. At 15Z, winds were sustained out of the east-northeast (70º) at a whopping 20 knots (23 mph), gusting to 24 knots (28 mph). Throughout the afternoon, sustained winds were pretty constant between 13 knots and 16 knots (15 mph to 18 mph). That's a pretty stiff breeze for such an unremarkable pressure gradient. However, at locations in the tropics and subtropics, such scenarios are common.

Forecast Tips

  • When forecasting for locations in the tropics or even subtropics (including the southern U.S.), don't be lulled into expecting weak winds just because the pressure gradient appears to be weak or modest. You get more "bang for your buck" out of a given pressure gradient at lower latitudes because the increased velocities are needed to strengthen the Coriolis Force.
  • There's no need to make extra adjustments to statistical guidance for this effect, unless you believe that the dynamic or AI models are grossly misrepresenting the surface pressure gradient. Just don't be surprised if the wind speeds forecast by statistical models for southern locations are higher than you'd expect given a modest or weak pressure gradient.
  • Weak synoptic pressure gradients can allow for mesoscale processes to drive wind speeds (sea-breeze circulations, etc.), so again, don't be lulled into thinking that a weak pressure gradient always means weak winds. Yes, winds can be weak when synoptic pressure gradients are weak, but we have to account for other factors.

Ultimately, the synoptic pressure gradient is an important driver of wind speed, but it's not the only factor that matters. Keeping this in mind will help you make sense of some of the model forecasts that you see. In the next few sections, we'll look at some of the major mesoscale processes that can impact wind speed, starting with vertical mixing and momentum transfer. Keep reading!

sas405

Vertical Mixing and Wind Forecasting

Vertical Mixing and Wind Forecasting

Prioritize...

Upon completion of this page, you should be able to use forecast skew-T diagrams to help forecast maximum sustained wind speeds and maximum possible wind gusts when the planetary boundary layer will become well mixed. This process includes being able to identify a well-mixed boundary layer and the wind profile within that layer on a skew-T diagram. You should also be able to identify mixing situations in which statistical guidance seems to have a low bias in its sustained wind speed predictions (it underestimates wind speeds).

Read...

The concept of vertical mixing is one that you should be well acquainted with by now. Remember that convective and mechanical eddies in the boundary layer mix surface air upward while bringing air from the top of the boundary layer down toward the surface. This process has consequences for the thermal structure of the boundary layer that we've already talked about, but now we need to consider how this process impacts wind speed.

You may recall from previous courses that when the lower troposphere destabilizes and vertical mixing deepens, momentum from winds aloft can be drawn down to the surface, resulting in gusty winds. That's why there's a "crusty old forecasting rule" that 850-mb wind speeds serve as a reasonable estimate for maximum wind gusts at the surface. This guideline is valid when a well-mixed boundary layer will develop up to 850-mb and momentum from speedy 850-mb winds gets transferred down to the surface by eddies. But, in any situation with a well-mixed boundary layer, the fastest wind speeds in the boundary layer (usually near the top of the boundary layer) serve as a proxy for the maximum possible wind gust.

But, this rule only applies to wind gusts. How can we quantify the impact of vertical mixing on sustained wind speeds at the surface? As it turns out, the mean wind in the well-mixed boundary layer serves as a reasonable estimate for the maximum sustained wind speed at the surface. Forecasters often take 850 mb to be the average top of a well-mixed boundary layer, but the top of the well-mixed layer can be higher or lower depending on the prevailing weather pattern and the season (it tends to be higher in the summer and lower in the winter). How can we incorporate this knowledge into our forecasts? First, we have to be able to assess two things --  the height of the well-mixed layer and the wind speed profile in that layer. Sounds like a great time to look at forecast soundings, doesn't it? Let's look at an example.

A Breezy Day in Minneapolis, MN

Let's look at an example in which downward transfer of momentum in a well-mixed boundary layer played a big role in the surface wind speed forecast. The short video (3:35) below outlines the big picture and the signs that forecasters could see on forecast soundings for a breezy day in Minneapolis, Minnesota.

Let’s check out a case that shows how assessing the potential for vertical mixing is critical to a wind speed forecast. We’re going to focus on winds at Minneapolis, Minnesota on a warm October day. The WPC surface forecast map valid at 18Z, which is early afternoon in Minneapolis, showed Minneapolis in the warm sector of a mid-latitude cyclone, with a fairly robust pressure gradient. This was an initial sign of a breezy afternoon at Minneapolis. But, how breezy? If we needed to predict the maximum sustained wind speed and maximum wind gust in the afternoon, we would need to assess whether downward transfer of momentum in the boundary layer via vertical mixing could be an issue.

A quick look at the GFS forecast for 700-mb relative humidity and streamlines valid at 18Z shows areas of very high relative humidity indicative of likely cloud cover well west of Minneapolis with the cold front and another area southeast of Minneapolis.

The corresponding GFS forecast for 500-mb heights and relative vorticity supports the idea of clouds well west of Minneapolis early in the afternoon downwind of the axis of a trough centered over the western Dakotas. But, the main region of upward motion would still be well west of Minneapolis at this time.

So, the big picture suggested some afternoon sunshine at Minneapolis, and this is backed up by the GFS simulated IR image forecast for 19Z – the bulk of the cloudiness associated with the mid-latitude low and cold front was well west of Minneapolis. With some sunshine, and a fairly robust pressure gradient, the big picture supported the development of both convective and mechanical eddies that would create a well-mixed boundary layer.

At 925 mb, the GFS forecast showed some pretty speedy winds by mid-afternoon, with wind speeds in excess of 30 knots over Minneapolis. Speedy winds not all that far above the surface certainly suggested that downward transfer of momentum would be a factor.

The GFS forecast sounding valid at 21Z – late afternoon in Minneapolis showed a deep well-mixed layer up to 800 mb would develop. The well-mixed layer is highlighted on this zoomed in sounding – there’s a thin slightly superadiabatic layer near the surface, and the environmental temperature profile is dry adiabatic above that up to 800 mb. We can approximate the maximum sustained winds by averaging the winds in this layer, and if we do so, we get an average of about 30 knots. The highest wind in the boundary layer is 36 knots, which is a good proxy for the maximum possible gust around this time.

Comparing with the corresponding NAM forecast sounding – it showed a well-mixed boundary layer that wasn’t as deep, only up to about 875 mb. But, the mean wind calculation was only slightly lower – about 27 or 28 knots. The maximum wind predicted in the boundary layer at this time by the NAM was 34 knots, which is only slightly less than what the GFS predicted.

If we take a consensus of both models, a maximum potential gust forecast would be around 35 knots, since the GFS predicted a maximum wind near the top of the boundary layer of 36 knots, while the NAM predicted 34 knots. A maximum sustained wind speed forecast based on the mean wind speed in the boundary layer would be 28 or 29 knots based on the average of the GFS and NAM forecasts.

Credit: Penn State

So, what happened? The maximum sustained wind speed for the forecast period verified at 25 knots (29 mph), and the maximum gust was 36 knots (42 mph) as shown toward the bottom of the NWS Daily Climate Report for the date (opens in a new window). So, our consensus forecasts using forecast soundings in the video above (maximum sustained winds of 28 or 29 knots and maximum gusts of 35 knots) weren't perfect, but they were pretty good! And, most importantly, they pointed us to the idea that the "first guess" approach for predicting maximum sustained winds that you learned previously (adding a few knots to statistical guidance to account for what happens in between synoptic times) would be insufficient because of the downward transfer of momentum. Statistical guidance for this date (opens in a new window) showed highest sustained wind speeds at synoptic times of only 16-19 knots. Our typical adjustment of adding a few knots to calculate a maximum sustained wind would result in a maximum sustained wind speed forecast of perhaps 20-22 knots at the most, which was too low. In fact, taking a consensus of a "statistical guidance + a few knots" approach and the forecasts from the video would have been just about perfect.

The maximum gust in Minneapolis (36 knots) also verified much higher than the highest gust shown in the NBM table (28 knots), so while the NBM sent the message that winds would be gusty, it didn't capture the maximum gust very well. In this case, the impact of downward momentum transfer on sustained wind speeds at the surface was optimized because winds in the well-mixed boundary layer were "aligned" (blowing from approximately the same direction). Indeed, statistical guidance seems to have a bias (underestimating maximum wind speeds) in situations when there's a significant downward transfer of momentum, especially when winds are aligned in the boundary layer.

It's important to note that this apparent bias is based on anecdotal evidence from personal forecasting experience. Remember that MOS and the NBM don't predict wind speeds between synoptic times, which accounts for some of the underestimates, but in these circumstances, the typical adjustment of just adding a few knots to come up with a maximum sustained wind speed often doesn't work very well. Ultimately, the degree of the bias tends to vary from place to place and time to time, which is why including an analysis of forecast soundings as part of your big-picture analysis is so important.

Forecast Tip

Given the apparent shortcomings of statistical guidance in some mixing situations, when the big picture indicates that downward transfer of momentum may be significant, adding a few knots to MOS / NBM forecasts to predict maximum sustained winds may be an insufficient approach. Still, even in an optimum mixing situation, don't throw MOS and the NBM out with the bath water. Instead, I suggest this approach:

  • Take note of the wind speed predictions of MOS and the NBM.
  • Consult forecast soundings from dynamic models to assess the depth of the well-mixed layer and wind speeds throughout the layer. Calculate a mean layer wind speed for each model.
  • Come to a consensus based on MOS / NBM and your analysis of the forecast soundings. If wide disagreement exists between MOS / NBM and mean boundary-layer wind speeds on forecast soundings, hedge your bets in light of the big picture.

Getting a firm grasp on the potential for downward momentum transfer requires a detailed analysis of forecast soundings, so it's important that we have some good sources for forecast soundings at our disposal.

Key Data Resource

I already provided some links to various sites for forecast soundings previously, but I want to introduce a few features of the most extensive site that may help you with wind forecasting.

In my opinion, NOAA's Air Resources Laboratory (opens in a new window) is the best website to access forecast soundings, but it's not the easiest site to navigate. Still, this site has some significant advantages, such as the following:

  • You can create animated images showing forecasts at three-hour time steps spanning a specific time period, which is very helpful in quickly assessing how the boundary layer will evolve in time. Looking at a single time can leave a lot to be desired.
  • Sometimes reading the specific wind data in the boundary layer can be challenging (lots of wind barbs crammed into a small space). This site, however, allows you to "access text data" in order to clearly read wind speeds. Be careful, though. The text data wind speeds are expressed in m/s, not knots. So, you'll have to convert to knots (opens in a new window) if you go the text route.

Creating soundings on this site requires several steps. First, enter the three-letter identifier of the station in which you're interested and click "Continue." To the right of "Sounding" in the menu that appears, choose your model. I recommend looking at both the RRFS and GFS in order to cover all your bases (for short-term forecasts, forecast soundings from the HRRR are also available). Once you select a model, you can select the model run on the next page. Finally, you can choose your display type. I recommend making an animation (either .gif or JavaScript) that starts at the beginning of your period of interest and lasts until the end. If you have any problems, feel free to ask.

Situations in which the downward transfer of momentum plays a role are fairly common, and identifying them can help you stay one step ahead of statistical guidance. Now, you might be asking yourself the question, "What if there's a well-mixed boundary layer with mean wind speeds that are slower than statistical forecasts?" Should you predict a maximum sustained wind speed lower than MOS / NBM forecasts? Probably not. In such cases, the maximum speed likely won't be driven by mixing, so the "statistical guidance + a few knots" approach usually works fine. But, you can't assess the potential for downward transfer of momentum without looking at forecast soundings, so I highly recommend incorporating forecast soundings into your wind forecasting routine every day.

After getting a little practice calculating the mean wind in the boundary layer in the Quiz Yourself section below, we'll move on to another process that can impact wind speeds on the mesoscale -- sea-breeze circulations.

Quiz Yourself...

Incorporating information from forecast soundings into your wind forecasts is a critical skill. To give you some practice, here's a GFS forecast sounding for Oklahoma City, Oklahoma valid at 21Z on October 11 in an unknown year. The questions below refer to this sounding. Note that the bottom two wind speeds are a hard to read (the numbers are practically on top of each other), but there's a "17" and an "18" below the "23" that is easier to read.

GFS forecast sounding valid at 18Z on October 11 in an unknown year.

A GFS forecast sounding valid at 21Z on October 11 in an unknown year.
Credit: NOAA / Air Resources Laboratory

sas405

Sea / Lake Breezes and Wind Forecasting

Sea / Lake Breezes and Wind Forecasting

Prioritize...

When you're finished with this section, you should be able to discuss how sea breezes impact wind speed forecasts, and be able to apply your knowledge in forecasting situations.

Read...

In METEO 361, you learned all about the meteorology of sea breeze circulations. I'm only going to give a quick broad-brush treatment of their formation and structure here, and instead, I want to focus on the aspects of sea breezes that will help with forecasting maximum sustained wind speeds.

The first step in using your knowledge of sea and land breezes to help with your wind forecasts is to recognize the weather patterns in which sea and land breezes commonly form. You may recall that in order for sea breezes to form: 

  • the synoptic-scale pressure gradient must be weak
  • the temperature difference between land and the adjacent water must be at least 6 degrees Fahrenheit (this difference is based on empirical studies)

The synoptic-scale pressure gradient must be weak because a stronger synoptic-scale pressure gradient creates stronger synoptic-scale winds, which disrupt the development of sea-breeze circulations. The temperature difference between land and water must be at least 6 degrees Fahrenheit because it's the differential heating between land and water that drives the circulation (larger temperature differences make for stronger sea breezes). 

As a brief refresher, when these conditions are present, the formation of a sea breeze occurs because of differential heating between land and water (land warms more quickly and to a greater extent). Pronounced warming over land in the morning causes heights to rise on constant pressure surfaces near 925 mb (Step #1 in the schematic below). With higher heights (higher pressure) aloft, air flows toward the sea at higher altitudes, creating weak divergence that helps to lower the surface pressure over the land (Step #2). Air aloft moving away from land eventually converges over the sea, causing surface pressure to increase over coastal waters (Step #3). The resulting mesoscale pressure gradient near the earth's surface now drives cooler maritime air toward land, creating the faithful sea breeze (Step #4). In reality, the facets of the mesoscale circulation you see here develop simultaneously, but I presented the sea-breeze circulation as a series of steps for simplicity.

Schematic showing the development of a sea breeze.

A schematic showing the development of a sea breeze.  Differential heating between land and water drives the circulation as described in the paragraph above.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

What's the practical implication of all this for sustained wind speed? Primarily, the necessary synoptic set-up means that we can't let ourselves be fooled by weak synoptic pressure gradients. If sea breezes are going to be a factor, conditions can be quite breezy, despite the weak synoptic pressure gradient. In fact, sustained winds associated with the passage of a sea-breeze front are commonly between 5 and 10 meters per second (10 and 20 knots). In general, sea-breeze fronts only penetrate about 25 miles inland (or less), but when they get a boost from onshore synoptic scale winds (that aren't too strong, of course), they can travel up to 100 miles inland or more.

How weak must the synoptic-scale pressure gradient be for a sea breeze to form? That's a tough question to quantify, because it ultimately depends on the temperature difference between land and sea. As a general rule, the synoptic pressure gradient must be sufficiently weak that it generates synoptic-scale winds of 10 knots or fewer (this rule usually works well for coasts in the Southeast and Middle-Atlantic regions of the U.S.). When the synoptic-scale pressure gradient generates winds faster than 10 knots, one of two things usually happens. If the synoptic pressure gradient causes winds to blow strongly offshore, the sea breeze can't advance inland. On the flip side, if the synoptic pressure gradient causes winds to blow strongly onshore, the sea-breeze front gets "lost" to heating and mixing as it races inland. 

An Example

Let's take a look at the case of a sea breeze in Miami, Florida (opens in a new window) on a September day. As you can see from the 18Z surface analysis below, the pressure gradient over the entire Southeast U.S. was quite weak.

The 18Z surface analysis from September 14, 2011.

The 18Z surface analysis from September 14, 2011 showed a very weak pressure gradient over the Southeast U.S.
Credit: Weather Prediction Center

The contour interval on this map is four millibars (which is standard), but "intermediate" isobars at two-millibar intervals (like the wavy, dashed one that crosses Florida) have been drawn over the Southeast U.S. because the pressure gradient was so weak. That satisfies one of our criteria for a sea breeze. What about the land / sea temperature contrast? It was a hot afternoon in Miami, with the 18Z KMIA METAR (opens in a new window) showing a temperature of 32.8 degrees Celsius (91 degrees Fahrenheit). A glance at the "1-group" shows that some time before 18Z, the temperature had spiked up to 33.9 degrees Celsius (93 degrees Fahrenheit). Compare that to water temperatures (opens in a new window) of around 30 degrees Celsius (86 degrees Fahrenheit), and we're right in the ballpark of the temperature difference needed for sea breeze development.

You may have noticed on the 18Z METAR, however, that winds were calm at KMIA.  Not to worry, the trusty sea breeze was on its way! By 20Z, winds were blowing out of the east-northeast at 9 knots, and an hour later at 21Z, sustained winds were blowing at 13 knots (see 20Z - 22Z KMIA METARS (opens in a new window)). In fact, the maximum sustained wind for the afternoon ended up verifying at 14 knots, even with a pathetically weak synoptic pressure gradient, thanks to the sea breeze. You also may have noticed that as winds came onshore from the east-northeast, temperatures dropped to 85 degrees Fahrenheit through the afternoon (daytime highs are often reached before the sea breeze kicks in).

How did statistical guidance handle the situation? Check out the 12Z GFS MOS (opens in a new window) run from the day before (September 13). GFS MOS predicted a maximum of 11 knots (at synoptic times). Adding a few knots to account for the fact that MOS only predicts winds at the synoptic times would have given us a forecast of 13 or 14 knots, which would have been an excellent forecast! The fact that MOS performed well here shouldn't be a big surprise. Sea breezes are fairly common in Miami in the warm season, so lots of instances of sea breezes exist in the data set used to derive the MOS wind equations for Miami. When the big picture supports the formation of a sea breeze at a climatologically favored time, MOS should do fairly well as long as it's getting reasonably accurate forecasts for predictors.

Dynamic and AI models models are also run at sufficiently high resolutions that most can pick up on the presence of a sea breeze if the synoptic set-up favors it. So, the bias-corrected, blended forecasts from the NBM typically perform reasonably well, too. Of course, this all assumes that the dynamic and AI models are correct about the overall nature of the big picture favoring a sea breeze.

Forecast Tips

Assuming the big picture supports the development of a sea breeze, consider the following issues in your forecast:

  • Maximum sustained winds will likely be faster than the weak synoptic pressure gradient suggests.
  • MOS and the NBM typically handle such situations pretty well, as long as sea breezes are climatological and the dynamic/AI models have a reasonably accurate handle on the big picture.
  • Because days with sea breezes can be somewhat gusty (larger variation in wind speed with time), you might consider tacking on an extra knot or two on top of "statistical guidance + a few knots" as a starting point. The larger variations in wind speed often mean that 3-hourly snapshot forecasts are a little less likely to capture some of the spikes in sustained wind speed.
  • Days on which sea breezes form are frequently days with deep, well-mixed boundary layers, so make sure to consult forecast soundings to examine the forecast wind profiles in the boundary layer to help you adjust statistical forecasts (if necessary).
  • Sea breezes aren't just an issue at stations along the coasts of major oceans (or, say, the Gulf of Mexico). Don't forget that other, smaller, bodies of water can generate similar breezes (like the Great Lakes, for example).

You may have noticed that I didn't talk about "land breezes" on this page (the nighttime counterparts to sea breezes that form when the land cools off rapidly and water becomes relatively warmer). That's because the differential between land and water temperatures at night is typically smaller, resulting in breezes that are usually weaker than those experienced with the passage of a daytime sea breeze front. Similarly, however, statistical guidance usually handles these situations pretty well as long as land breezes are climatological, and the dynamic/AI models have a good handle on the big picture.

Let's move away from the water now and talk about another mesoscale process that can impact sustained winds in rugged, mountainous terrain -- the mountain-valley circulation. Read on.

sas405

Mountain-Valley Circulations and Wind Forecasting

Mountain-Valley Circulations and Wind Forecasting

Prioritize...

In this section, you should focus on the synoptic-scale conditions that can cause mesoscale mountain-valley circulations to develop. You should also be able to incorporate the impacts of mountain-valley circulations on wind speed and temperature into your forecasts.

Read...

As with sea breezes, you learned about the meteorological details of mountain-valley circulations in METEO 361. I don't intend to go over all of those details again here. Instead, I want to focus on the aspects of mountain-valley circulations that will help with forecasting maximum sustained wind speeds.

Similar to sea breezes, mountain-valley circulations develop thanks to differential heating and cooling of the earth's surface. They form in environments with a weak synoptic-scale pressure gradient, and, as with sea breezes, a strong synoptic-scale pressure gradient can disrupt the development of mesoscale mountain-valley circulations. Remember, mountain-valley circulations develop as mountain slopes warm much more during the day than the surrounding "free atmosphere," leading to an air flow up the mountain slope (opens in a new window) during the daytime. At night, the slopes cool to a greater degree leading to a flow of air down the mountain slope (opens in a new window), causing cold air to collect along the valley floor.

You may also recall that when the valley floor is not flat, such mountain-valley circulations can also create up and down-valley winds that result from low-level divergence along the valley floor during the daytime and low-level convergence that occurs along the valley floor at night (see the schematic below).

Schematic showing the development of up- and down-valley winds.

A down-valley wind at night (left) and an up-valley wind during the daytime (right) can develop because of patterns of convergence and divergence (respectively) that form along the valley floor in mountain-valley circulation situations.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

The practical implication for wind-speed forecasting is that, yet again, we can't let ourselves be fooled by a weak synoptic pressure gradient. Mountain-valley circulations (and resulting up-valley and down-valley winds) can cause relatively breezy conditions when notable differential heating or cooling is likely (a mostly sunny day / clear skies at night). These circulations can occur at any time of the year in mountainous areas, but the details at any given city are governed by the local lay of the land and the specific station location with respect to surrounding topographic features.

How do we know whether a particular location is prone to impacts from up-valley and down-valley winds? A good understanding of a station's climatology can help, since climatology can reveal the impact of terrain on local winds (both in terms of mountain-valley circulations and other factors). For example, Knoxville, Tennessee (opens in a new window) lies in a broad valley between the Cumberland Mountains and the Great Smoky Mountains, and furthermore, the valley floor is not flat. Elevations to the northeast of Knoxville are higher than those to the southwest. If the southwest-to-northeast orientation of the valley impacts the local distribution of winds at Knoxville, we might expect to see that the most frequent wind directions run along the orientation of the valley (either from the southwest or the northeast).

Is that the case? The Knoxville wind rose for June below suggests it is -- southwesterly and west-southwesterly winds are the most frequent during the month, with a secondary relative maximum from the northeast. This corresponds to winds blowing from southwest up the valley floor during the daytime. However, as we move into the cold season (click the slider to see the October wind rose), and nights get longer, northeasterly winds become more frequent and the secondary maximum is from the southwest. This corresponds to nighttime down-valley winds becoming more prominent with the longer nights and shorter days of the cold season.

Image slider (click the arrow to change the image) showing wind roses for June and October in Knoxville, Tennessee.
Credit: Natural Resources Conservation Service

A location that is frequently impacted by up- and down-valley winds will often have a wind rose showing a pattern like the one at Knoxville above -- a maximum and secondary maximum frequency in wind direction along the valley floor with the up-valley direction favored in the warm season when days are longer and the down-valley direction favored during the cold season when nights are longer. These climatological patterns aren't universal, however, because they depend on how often the circulations occur at a given location. Prominent regional storm tracks and frequently strong synoptic pressure gradients can "muddy" the picture, even if up- and down-valley winds do occur occasionally. Still, I recommend studying a location's wind roses for clues. Now, let's look at an example of the effects of mountain-valley circulations at a specific station.

Winds at Grand Junction, Colorado (KGJT)

For a specific example, let's focus on Grand Junction, Colorado (opens in a new window), which lies in the Grand Valley (opens in a new window) in the western part of the state. Please note that the Grand Valley has a northwest-to-southeast orientation, which clearly impacts the distribution of local wind directions (see the KGJT wind rose for November (opens in a new window), for example). The maximum frequency of southeasterly winds and the secondary relative maximum in northwesterly winds is a result of the Grand Valley's orientation.

Let's look at a late November night. An area of high pressure was situated over the central Rockies and the eastern half of the central Great Basin (check out the 12Z surface analysis (opens in a new window)). Because of the varying terrain in the region, we should also look at 850 mb to get a true sense for the strength of the pressure gradient, and given the corresponding 850-mb analysis at 12Z (opens in a new window), I hope you agree that the synoptic-scale pressure (height) gradient was relatively weak. At face value, it appeared that winds overnight should have been pretty light.

But, that wasn't the case at all.  Take a look at these METARS for Grand Junction (KGJT) from 04Z (bottom) to 13Z (top):

KGJT METARS from 04Z to 13Z on November 13, 2006.

The METARS from Grand Junction, Colorado from 04Z (bottom) to 13Z (top) on November 23, 2006 show the presence of nighttime down-valley winds.
Credit: NCEP

Note the calm wind at 04Z (9 P.M. local time), which was consistent with the prevailing weak synoptic-scale pressure gradient. As the evening progresses, the down-valley wind at Grand Junction starts to blow from the northeast early, but then turns out of the east and then becomes southeasterly during the night. As you can see from the METARS, the down-valley wind really kicked in by 08Z (1 A.M. local time) --  note the east-southeasterly wind direction of 100 degrees and the wind speed of 11 knots. In fact, sometime during the night, the down-valley wind reached a peak sustained speed of 17 mph (15 knots), obviously between observation times. Not too shabby for a weak synoptic-scale pressure gradient, eh? Such is the power of the down-valley wind associated with the mesoscale circulation of air generated by differential cooling at night. Further note in the observations above that the sky was clear at night, which constitutes ideal conditions for down-valley winds to develop when there's an overall weak synoptic-scale pressure gradient.

On the flip side, during the daytime, the up-valley wind is a forecast issue at KGJT. Using an early December day as an example, a weak pressure gradient associated with high pressure (see the 21Z surface analysis (opens in a new window) and 00Z 850 mb analysis (opens in a new window) from three hours later) set the stage for an up-valley wind at Grand Junction. Note the developing northwesterly breeze (from 300 degrees) in the METARS below. Not surprisingly, the METARS also show that the daytime sky was clear on this day, providing optimal conditions for relatively pronounced differential heating of the Earth's surface.

KGJT METARS from 17Z to 22Z on December 6, 2006.

The METARS from Grand Junction, Colorado from 17Z (bottom) to 22Z (top) on December 6, 2006 show the presence of daytime up-valley winds.
Credit: NCEP

Now that we've seen the impacts of mountain-valley circulations in action, what are some general guidelines that we can apply to our wind forecasts?

Forecast Tips

In forecasting for stations located in mountainous regions, I suggest keeping the following tips in mind:

  • Look for a weak synoptic pressure gradient and mostly sunny skies (during the daytime) or clear skies (at night) to recognize patterns in which mountain-valley circulations may be an issue.
  • Know your climatology! The nature of the local impacts of mountain-valley circulations are often revealed in the climatological data.

When you identify big-picture patterns that support the development of mountain-valley circulations, expect:

  • stronger daytime and nighttime winds than the weak synoptic pressure gradient suggests.
  • higher nighttime temperatures than you might expect on a clear night with a weak pressure gradient (the winds may not be calm, which slows lower tropospheric cooling). You may not want to forecast below statistical guidance even if the formation of a nocturnal inversion appears on forecast soundings (as you learned in previous lessons).
  • typical daytime wind speeds of 6-12 knots, and typical nighttime wind speeds of 2-10 knots are observed with mountain-valley circulations. Drainage winds can exceed 15 or 16 knots at night depending on the steepness of the local terrain and prevailing weather conditions.
  • Statistical guidance should typically perform reasonably well with wind speeds as long as mountain-valley circulations are climatological and the dynamic models have a good handle on the situation (which is not necessarily a given). Of course, in addition to evaluating statistical guidance, viewing forecast soundings will help you quantify needed adjustments.

Patterns that lend themselves to mountain-valley circulations also tend to support mountain-top convection during the daytime (which could impact precipitation forecasts), and light turbulence for small aircraft. So, mountain-valley circulations don't just impact wind speed. These circulations impact wind direction, and can affect temperature and even precipitation, which underscores the importance of getting a handle on the wind as a forecast variable. In our next section, we'll look at another mesoscale process that can impact wind speeds and other forecast variables, too! Keep reading.

sas405

Effects of the Isallobaric Wind

Effects of the Isallobaric Wind

Prioritize...

When you've finished this page, you should be able to describe the isallobaric component of the wind and identify forecasting situations when it may be a significant factor. You should also be able to discuss its possible impacts on wind speed, wind direction, temperature advection, and precipitation, and apply them appropriately in making your own forecasts.

Read...

As we continue our look at mesoscale factors that can impact wind speeds, let's look at the isallobaric wind. You encountered the isallobaric wind briefly in METEO 361 in the context of cold-air damming. Here, we'll take a little deeper look at it, but I will keep the focus on its practical forecasting implications.

First off, the isallobaric wind is a component of the wind that blows toward areas of maximum negative pressure tendency (or away from areas of maximum positive pressure tendency), with greater pressure tendencies causing greater contributions from the isallobaric wind. In the grand scheme of things, the isallobaric wind approximates the non-frictional component of the ageostrophic wind at the surface (both friction and the isallobaric wind cause the wind to deviate from geostrophy).

For an idealized example of the isallobaric wind, recall that a center of negative pressure tendencies exists to the north and east of a traveling and developing mid-latitude cyclone (caused by warm-air advection and upper-level divergence). In the schematic below, the wind vectors represent the isallobaric wind. Note that they blow perpendicularly to the local isallobars (contours of constant pressure tendency) toward increasingly negative values of pressure tendency.

Schematic showing the idealized isallobaric wind field associated with a center of maximum pressure falls

The idealized isallobaric wind field associated with the center of maximum pressure falls to the east and north of a classic mid-latitude low-pressure system.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

When you look at a map of surface winds, you won't see the isallobaric wind by itself. Remember, it's just a component of the observed wind, so it contributes to the wind direction and speed. Many times, pressure tendencies are small enough that the isallobaric wind is hardly noticeable, but when pressure tendencies become larger, the isallobaric contribution becomes more obvious. Take, for example, the 18Z surface analysis from December 23, 2022 (opens in a new window), when a very strong 974 mb low-pressure system was centered over southern Quebec. This storm led to widespread blizzard conditions, and the worst blizzard on record in Buffalo, New York (opens in a new window). Notice some of the very large crossing angles of the surface winds behind the cold front in the circled area in particular (note that it does not include all stations being impacted by the isallobaric wind).

Recall that, on average, friction causes surface winds to blow across local isobars in toward lower pressure at about a 30-degree angle. Terrain isn't flat in parts of this region, so stronger friction accounts for some of the increase in crossing angle, but not all of it. Such a strong low-pressure system brought significant pressure falls and pressure rises, as shown on the 18Z analysis of 3-hour pressure change below. Pressure was falling at a rate in excess of 9 millibars per 3 hours over Quebec (which is a very large negative pressure tendency), while pressure was rising at a rate of more than 4 millibars per 3 hours over the Mid-Atlantic behind the cold front.

Analysis of 3-hour pressure change valid at 18Z on December 23, 2022, showing huge negative pressure tendancies over eastern Canada and New England.

Analysis of 3-hour pressure tendencies at 18Z on December 23, 2022. A large area of negative pressure tendencies was located over eastern Canada and New England ahead of a very strong low-pressure system.
Credit: College of DuPage

Now look back at the corresponding 18Z surface analysis (opens in a new window). The isallobaric wind helped turn the wind direction more toward the northeast in the wake of the cold front (toward the center of pressure falls and away from the center of pressure rises), contributing to the large angles that winds crossed local isobars. The isallobaric component of the wind also boosted wind speeds (in addition to the large synoptic pressure gradient overall). Statistical guidance suggested it would be a windy day in the wake of the front (check out the 18Z GFS MOS and 23Z NBM forecasts from December 22 for Buffalo, New York (opens in a new window)), with sustained winds predicted by GFS MOS to increase to 36-37 knots by local midnight on the 23rd. The NBM had sustained winds at synoptic times that were notably lower -- "only" 28 knots by local midnight on the 23rd.

Even if a forecaster took the more aggressive GFS MOS numbers, and tacked on a few knots to account for what might occur between synoptic times, they might have predicted something like 40 knots as a maximum sustained wind speed at Buffalo on December 23. If they took a consensus, their forecast would have been lower. What was the actual maximum sustained wind measured at Buffalo on December 23? 42 knots (48 miles per hour). The GFS MOS forecast wasn't actually too bad, but an additional hedge up to account for the isallobaric wind would have made for a great wind speed forecast.

Forecasts in which the isallobaric wind plays a major role can cause headaches for forecasters because uncertainty is usually fairly large. The GEFS plumes for Buffalo (opens in a new window) from the 12Z run on December 22 showed a roughly 20-knot spread in 10-meter wind speeds as wind speeds were predicted to spike on the 23rd! Why the uncertainty? The physics of the isallobaric wind are incorporated into dynamic models, but the isallobaric wind is complex and non-linear. In practical terms, if a dynamic model makes a small error in the predicted pressure field, but the pressure tendencies are small, the model error in the wind forecast should be minor. But, if a dynamic model makes a small error of the same magnitude and the pressure tendencies are large, the error in the wind forecast will be more significant. So, small differences in the predicted pressure pattern when large pressure tendencies are in the picture can lead to large disagreements in the wind forecast.

The bottom line is that when a strong (or strengthening) high- or low-pressure system passes near your forecast city, you should consider possible impacts of the isallobaric wind. Also consider the isallobaric wind any time pressures at your city will rapidly rise or fall (such as with the passage of an Arctic cold front).

Key Data Resources

For current and recent analyses of 3-hour pressure tendencies, I recommend:

Forecasts of pressure tendency tend to be very hard to find, so forecasters need to recognize general patterns (described above) when pressure rises or falls could be large.

The impacts of the isallobaric wind can bring other consequences, too, so make sure to consider these forecasting tips when faced with a situation where the isallobaric wind may play a notable role:

Forecast Tips

  • In cases where the isallobaric wind will play an important role, the day may be windier than predicted by statistical guidance. Specific adjustments depend on the specific big picture situation (the location, strength, and timing of the pressure falls or rises).
  • Because the isallobaric wind also can impact wind direction in the vicinity of mid-latitude lows (where fronts and temperature gradients naturally exist), it's important to be conscious of any potential impacts on temperature advection. When the isallobaric component causes winds to cross isotherms at greater angles at faster speeds, temperature advection can be stronger than anticipated, and may not be modeled well. This tends to be a bigger issue with cold advection. So, when the isallobaric wind may be an important player, be sure to view MOS and NBM temperature forecasts with a critical eye.
  • Changes in wind direction can enhance low-level convergence, which can enhance precipitation totals (although pinpointing this in detail even a day in advance can be difficult).

In most cases, when you feel it's necessary to make adjustments based on the isallobaric wind, adjustments should be relatively small. But, once in a while, the isallobaric wind can really throw off MOS or NBM wind or even temperature forecasts (generally when the isallobaric wind ends up being significant). The best you can do is to qualitatively assess what the impacts of the isallobaric wind will be (in terms of wind speeds and direction and possible other consequences) and hedge your bets according to your analysis.

If you want to see an example of the isallobaric wind aiding in low-level convergence and thunderstorm development, check out the Explore Further section below. Otherwise, we have one other mesoscale factor to cover, which can throw a wrinkle into our wind forecasts -- deep convection and precipitation. Read on.

Explore Further...

When blowing toward the center of maximum pressure falls, the isallobaric wind can enhance low-level convergence and cyclonic vorticity in the region. For example, take a look at this 18Z analysis of MSLP and surface winds (opens in a new window) over the Middle Atlantic region from July 22, 2006. Note the surface low located over northeastern Pennsylvania at the time, and also notice the pattern of isobars and surface winds surrounding the low (especially over eastern Pennsylvania and New York). Over extreme southeastern Pennsylvania, the wind barbs show winds crossing local isobars at very large angles, which friction alone doesn't explain. Over eastern New York, we see the same thing -- we can't explain the large crossing angles with friction alone.

Now take a look at the 18Z analysis of three-hour surface pressure tendency (below) and surface wind observations.  Do you see the area of maximum pressure falls (the "-24" closed isallobar in eastern New York)? The "-24" indicates a pressure fall of 2.4 millibars in the previous three hours.

Three-hour pressure tendencies at 18Z on July 22, 2006 (expressed in tenths of a millibars).

Contours represent three-hour pressure tendencies at 18Z on July 22, 2006 (expressed in tenths of millibars). The wind barbs denote the 18Z observed surface winds and the number and sign to the right of each station model represents the pressure tendency. Note how winds near the center of maximum pressure falls were blowing almost directly toward it.
Credit: @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

The area of negative pressure tendency over eastern New York was turning observed winds a bit toward the center of maximum pressure falls. The winds nearest the center of pressure falls were blowing almost directly toward it! Farther away, the effect isn't as dramatic, but it's still there. It helps explain why winds over southeastern Pennsylvania were out of the south-southwest, instead of being southwesterly (or maybe even west-southwesterly), which is what we'd expect just based on the pattern of local isobars from the 18Z MSLP analysis (opens in a new window).

Altering the observed wind direction by the isallobaric wind component had a couple of important effects in this case. First, it enhanced low-level convergence over southeastern Pennsylvania (clearly winds were coming together in extreme southeastern Pennsylvania with west-southwesterly flow farther west meeting up with south-southwesterlies near the New Jersey border). Furthermore, it enhanced the magnitude of wind shear in the lower troposphere, which made the environment more favorable for supercell thunderstorms (see the 18Z radar image (opens in a new window)). These thunderstorms spawned some damaging wind reports in New Jersey and one tornado in northeastern Pennsylvania (SPC storm reports (opens in a new window)). In case you're wondering, the fine forecasters at SPC weren't caught off guard in this situation; they had a good handle on it as evidenced by Mesoscale Discussion #1606 (opens in a new window) issued just after 14Z (nearly four hours before the thunderstorms erupted). Note the specific reference to "isallobaric forcing."

The big tip off to SPC forecasters on this date was that the surface low would be strengthening as it moved toward the northeast, and indeed, that's something forecasters often look for when trying to evaluate significant isallobaric contributions to the wind (the isallobaric wind becomes more prevalent when pressure falls or rises become stronger).

sas405

Complications from Convection and Precipitation

Complications from Convection and Precipitation

Prioritize...

From this section, you should be able to discuss the additional complications that convection and precipitation can bring to a wind forecast, and be able to apply reasonable adjustments to your wind forecasts, such as the ones demonstrated in the examples.

Read...

We have a final mesoscale "headache" for wind forecasting that I want to discuss -- deep convection, and more generally, precipitation. For starters, the presence of deep, moist convection can wreak havoc on a wind speed forecast. You may recall from your previous studies that in extreme cases, convective storms can result in damaging winds from microbursts (opens in a new window), as negatively buoyant downdrafts accelerate to the surface and spread out. But, downdrafts and the associated outflow from showers and thunderstorms (even in non-severe microburst cases) can dramatically alter local sustained wind speeds and directions (the details of which are not predicted well by dynamic models, AI models, or statistical models). Since forecasting the details of deep, moist convection beyond even 6 or 12 hours can bring huge challenges, wind forecasts involving deep, moist convection can be extremely difficult.

Yes, we can usually tell if the overall environment favors shower or thunderstorm development in a particular region, but we don't know exactly where the showers or thunderstorms will be or exactly when they'll develop. Will one strike our location exactly, or at least be close enough that its outflow could make for a brief period of gusty winds? It's those details that will really determine the impact that deep convection will have on maximum sustained winds and gusts.

Clearly, such circumstances bring great uncertainty to a forecast. About the best we can do is recognize the fact that the big picture supports shower or thunderstorm development in the area and make some adjustments. If you feel that showers or thunderstorms may develop in or around your forecast location, you may want to adjust your wind forecasts upward slightly just to hedge your bets. But, swinging for the fences based on low-probability scenarios can be really risky -- an isolated shower or thunderstorm may not get close enough to your location of interest to matter. To see how deep, moist convection can complicate a wind forecast, and how a forecaster can reasonably deal with such a situation, let's look at an example.

Briefly Breezy in Astoria, Oregon

The video below (2:20) walks through a quick example at Astoria, Oregon (opens in a new window) (KAST), in which localized convection really threw a wrench into the wind forecast.

Let’s look at an example of convection really throwing a wrench into a local wind forecast. Let’s imagine that we had to forecast maximum sustained wind speed at Astoria, Oregon on a November day. Astoria is located in the far northwest corner of Oregon, right near the coast and the Washington State border. The 36-hour GFS 4-panel prog valid at 00Z on November 10 showed a fairly weak sea-level pressure gradient along the northwest coast of Oregon beneath a deep 500-mb trough as shown in the upper-left panel here.

The corresponding NAM forecast showed the same overall picture with some subtle differences. It also showed a weak synoptic pressure gradient with a weak surface low along the northwest coast of Oregon, and a deep 500-mb low that was actually closed off. The weak pressure gradient initially sent the message that winds would likely be pretty weak on this day, but a wise forecaster could note that Astoria’s position beneath a deep 500-mb low could play a role in the forecast. Deep 500-mb troughs or lows suggest relatively cold air aloft, and such patterns often lend themselves to the development of scattered afternoon showers and thunderstorms because the cold air aloft steepens lapse rates, especially along the relatively mild Pacific Coast during the cold season.

This date was no different. Scattered showers and thunderstorms developed across Oregon, but didn't have any immediate impact on Astoria until just before 00Z. This regional radar mosaic from 2327Z on November 9 shows a relatively isolated thunderstorm that popped up right near Astoria. It turns out that only two hundredths of an inch of rain fell at Astoria Airport, but that didn’t matter much for the wind forecast.

The METARS from the airport from 23Z on November 9 to 02Z on November 10 show that winds went from being relatively light at four knots at 23Z to gusty very quickly -- gusting as high as 23 knots at 2323Z. Note that it was windy at the site even when it wasn’t actually raining right at the site – the METARS reported a thunderstorm in the vicinity, and outflow from that thunderstorm helped boost the winds. After the storm passed after 00Z, winds abruptly calmed back down, but during a brief breezy time, sustained winds as high as 17 knots were measured in between METAR observations.

Credit: Penn State

A cursory look at the surface pressure gradient gave the impression that winds would be weak on this day, and statistical guidance gave the same impression, predicting highest sustained wind speeds at the synoptic times of 5-6 knots. Adding a few knots to predict a maximum sustained wind speed would have led to a forecast of just 7 or 8 knots. Frankly, there was nothing in the wind guidance for this day that suggested anything close to the 17 knots that verified for maximum sustained wind speed (or gusts over 20 knots, for that matter).

Realistically, what can forecasters do in a situation like this? Was there any way to know for sure that the maximum sustained wind at KAST would be so high? Not really. Had the culprit thunderstorm formed just 50 miles away, the typical approach of adding a few knots to statistical guidance predictions likely would have worked out great. But, that's not the way it worked out.

A forecaster in tune with the big picture would have known that gusty winds were possible because of pop-up showers and thunderstorms. While forecasters couldn't know for sure whether one would impact Astoria a day in advance, or know precisely how big of a boost it might give wind speeds, adding a few knots to statistical guidance, and then hedging up a bit more to account for the possibility of gusty winds with a pop-up shower or thunderstorm would have made for a pretty good forecast in the scheme of things. Even a forecast of something in the range of 9-12 knots for maximum sustained wind speed would have been too low, but it would have been significantly better than model guidance, and pushed your forecast in the right direction. At the same time, by not swinging for the fences, such a forecast wouldn't have been drastically too high had a thunderstorm not impacted Astoria.

Photograph of a single-cell thunderstorm near Richfield, Utah.

The details of thunderstorm development, like this single-cell thunderstorm that developed near Richfield, Utah on August 16, 2006, are difficult to predict, and can wreak havoc on wind forecasts.
Credit: Steve Seman

In such situations, keep in mind that the shower or thunderstorm need not occur right at the observing site in order for it to impact wind speeds there. Even outflow from a nearby shower or thunderstorm can boost wind speeds at a particular site, even if it never actually rains right at the site. That's something to keep in mind in situations where probabilities of precipitation at a particular site might be low, but the big picture supports deep convection in the area.

So, while convective precipitation can have hard-to-predict impacts on local wind speeds, more generally, even non-convective precipitation can sometimes impact local wind speeds. Forecasters need to be aware of these possibilities in the following situations:

Forecast Tips

When you expect precipitation to occur, be on the look out for either of these scenarios:

  • precipitation falling into an initially dry lower troposphere (low relative humidity), favoring evaporational cooling and negatively buoyant downdrafts
  • low-level lapse rates that aren't highly stable (they need not be dry adiabatic, but the closer the better) with wind speeds in the "relatively unstable" layer that are relatively fast compared to those at the surface

The reason why evaporatively-driven downdrafts are important should be fairly intuitive -- it's similar to what can happen in convective situations (negatively buoyant downdrafts cause localized gusty winds when they spread out at the surface). But, even without the potential for significant evaporational cooling, falling precipitation exerts drag on the air and can force air down to the surface. When wind speeds in the lower troposphere are fast, and low-level lapse rates aren't highly stable, momentum can be brought down to the surface, boosting wind speeds. Why are low-level lapse rates important? A highly stable lower troposphere (say, near isothermal or a temperature inversion) hinders the downdrafts actually reaching the ground (opens in a new window), while less stable layers make it more likely that downdrafts (and the momentum they carry) can actually reach the ground.

To see an example, let's look at the wind speed forecast for Milwaukee, Wisconsin on October 25, 2022, when a cold front was approaching Milwaukee, and rain was expected ahead of and behind the front (check out the WPC surface forecast map valid at 06Z on October 25 (opens in a new window)).

Windy in Milwaukee

At 925 mb, a low-level jet stream was evident ahead of the approaching front, and wind speeds over Milwaukee were predicted to be near 40 knots at that level on the GFS forecast valid at 06Z on October 25 (from the 12Z run on the 24th). Such speedy winds at 925 mb, combined with the prospects for precipitation tipped off alert forecasters that it could get quite windy at Milwaukee around 06Z.

GFS forecast for 925-mb heights and winds, showing a low-level jet stream and 40 knots winds predicted over Milwaukee.

The 18-hour GFS forecast for 925-mb heights and winds showed a low-level jet stream with 40-knot winds over Milwaukee at 06Z on October 25, 2022.
Credit: College of DuPage

Was the environment in the lower troposphere favorable for precipitation to impact surface wind speeds, though? Let's turn to forecast soundings to see. The GFS forecast soundings valid at 06Z and 09Z on October 25 (opens in a new window) (from the 12Z run on the 24th) showed an initial layer of dry low-level air, which favored evaporation of falling raindrops and downdrafts. The 09Z sounding shows the predicted impacts of evaporation -- higher relative humidity and cooling in the lower troposphere. Furthermore, low-level lapse rates were not highly stable, suggesting that forced downdrafts via precipitation could transport momentum to the surface.

Statistical guidance on this day did show breezy conditions around 06Z (opens in a new window), but nothing higher than 16-knot sustained winds at the synoptic times (and gusts no higher than 22 knots on the NBM). Given the potential for evaporation to encourage downdrafts with the onset of precipitation (which can cause gusty winds by themselves) and the potential for downward transfer of momentum through the lower troposphere, the potential was there for much higher wind speeds. Indeed, sustained winds topped out at 25 knots (29 miles per hour) in the hours just after 06Z, and winds gusted as high as 40 knots (opens in a new window) (46 miles per hour)!

Quantifying just how high sustained wind speeds and gusts can get in these situations is tricky, because it's not obvious from what altitude parcels can bring momentum down to the surface. When the boundary layer is well-mixed, we have sound quantitative strategies to estimate maximum sustained wind speeds and gusts. But, when the boundary layer isn't perfectly well mixed (as was the case here) and evaporatively-driven downdrafts and downward momentum transfer are originating from unknown levels in the lower troposphere, quantifying the results becomes much harder. But, given the fact that wind speeds were near 40 knots at 925 mb, and lapse rates between the surface and that level were not highly stable, the approach of adding a few knots to statistical guidance in this time period (giving a forecast of perhaps 18 or 19 knots for maximum sustained wind speed) was likely to be too low. An additional hedge up of a few knots (perhaps giving a forecast of 21 or 22 knots for maximum sustained wind) would have been a reasonable hedge in the right direction.

Thanks in part to the many mesoscale factors that can impact a wind forecast, it should be pretty clear by now that wind forecasts can be filled with uncertainty, and hedging our bets in the right direction on a consistent basis is the best we can hope for. To wrap up the lesson, we'll cover some final tools and strategies for dealing with uncertainty in wind forecasts.

sas405

Summary and Final Strategies

Summary and Final Strategies

Prioritize...

From this section, you should be able to use the additional tools given for dealing with forecast uncertainty, as well as execute the Delta Method for winds in appropriate situations. Finally, make sure to pay attention to the "Summary" provided, which includes many of the major themes from the lesson.

Read...

Before we wrap things up with the lesson summary, I want to touch on a few final tools and approaches for dealing with forecast uncertainty.

With all of the mesoscale details and challenges involved in a wind forecast, model uncertainty can rear its ugly head on any given day. You'll surely encounter situations in which the models (statistical or dynamic/AI) just simply disagree. Sure, you may be able to judge which model is more in touch with the big picture and hedge your predictions accordingly, but don't forget about the role of ensemble forecasts in helping us gauge uncertainty. Indeed, ensembles have a number of products that can help us with our wind forecasts, such as the plume diagram for 10-meter winds below.

GEFS plumes of 10-m wind at Buffalo, New York showing a period of high forecast uncertainty.

The GEFS plumes for 10-meter wind speeds at Buffalo, New York (KBUF) from the 12Z run on December 22, 2022 revealed substantial forecast uncertainty for wind speeds late on December 23 into December 24.
Credit: NOAA EMC

These GEFS plumes from the 12Z run on December 22, 2022 showed a very large spread in wind forecasts for late on December 23 into December 24 at Buffalo, New York, indicative of large uncertainty in the forecast. Plumes can serve as a great visual tool to discern times when the wind forecast may be more uncertain, as well as for discerning general trends in wind speeds. Gauging uncertainty and general trends in wind speeds from plumes is typically more valuable than the specific values on the plumes themselves (which come from dynamic model forecasts, and remember that dynamic model wind forecasts don't have the best track record for reliability).

For a statistically post-processed look at uncertainty, the standard deviations available from the NBM can be useful for identifying times when model spread is large. The 23Z NBM forecast for Buffalo from December 22 (opens in a new window) showed fairly large standard deviations for sustained wind speed (4-5 knots) and gusts (as high as 12 knots) during the timeframe highlighted in the plumes above. At forecast lead times of 1-2 days, these are large standard deviations (indicative of a large bias-corrected model spread) indeed!

You encountered a complete list of ensemble forecasting resources previously, but I want to point you to a few specific tools here that may be of help.

Key Data Resources

  • SPC REFS Page (opens in a new window): In the "Fire" menu, you can access forecasts for REFS hourly mean 10-m sustained wind speeds and gusts. The hourly minimum, mean, and maximum member forecasts are available. 
  • DESI Viewer (opens in a new window): You can generate maps showing probabilities of various wind speed thresholds from a number of different ensemble systems, including the bias-corrected NBM. Via the "Surface" menu, you can also generate other visualizations, such as ensemble mean forecasts. 
  • GEFS Plumes (opens in a new window): You can generate plumes for 10-m winds for any of the stations on the map. Just keep in mind that these forecasts come from dynamic models, so their individual forecasts for 10-m wind speed usually perform worse than statistical guidance (so don't focus on specific wind speeds or member solutions. Look for timing, trends, and overall uncertainty.

The Delta Method for Winds

Remember the Delta Method from our lesson on temperature forecasting? Well, you can conduct the Delta Method for wind forecasts, too! Doing so is just as straightforward as using the Delta Method to predict high temperature. All you need to do is determine today's maximum sustained wind at your forecast city by looking at the hourly METARS (given that the maximum sustained wind may occur between hourly observations, you might need to make a reasonable estimation). Next, using the model's grid-point interpolations (like the example below), determine the correction factor for tomorrow at 10 meters or 850 mb and then add it to today's observed maximum sustained wind speed. When you expect tomorrow to be well-mixed to 850 mb, I suggest you use the interpolations at 850 mb (instead of 10 meters).

A sample of wind output in a dynamic model grid interpolation.

A sample of a wind forecast in a dynamic model grid-point interpolation.
Credit: Texas A&M University

The wind forecasts in grid interpolations are commonly expressed as those in the example above. Wind direction is expressed in tens of degrees, followed by wind speed in knots. For example, the very first 10-meter wind forecast above, "33/006" corresponds to a forecast of winds from 330 degrees (northwest) at 6 knots.

But, when applying the Delta Method for winds, there's a big caveat to always keep in mind: Low-level stability should be similar on both days.

Clearly, the degree of the downward mixing of momentum by eddies will influence the maximum sustained wind speed observed at the earth's surface, so applying the Delta Method on days with similar low-level stability should make a lot of sense to you. If low-level stability is notably different, you're really comparing apples and oranges. Perform the Delta Method with more than one model to come to a consensus if you can, which you can use in conjunction with statistical guidance and the rest of your analysis to come to a final prediction.

Summary

Finally, I want to summarize the main issues that you should keep in mind when making wind forecasts:

  1. While the synoptic pressure gradient is a major driver of wind speed, other factors (latitude and various mesoscale factors like downward mixing of momentum, sea-breeze circulations, mountain-valley circulations, and the isallobaric wind) can also affect wind speeds. Thus, don't be fooled by weak synoptic pressure gradients. Wind speeds may not be as weak as your initial impression.
  2. Remember that statistical guidance will get you in the ballpark most of the time. But, since MOS and the NBM only forecast wind speeds at the synoptic times (they do not specifically forecast a maximum sustained wind speed in a given period), if you're forecasting maximum sustained winds, it's often a good idea to add a few knots to their highest predictions to account for what happens between synoptic times (note that this is more reliably true for MOS than the NBM).
  3. Simply adding a few knots to statistical guidance may be an insufficient approach in certain cases, especially when you recognize that mesoscale processes may be important in the wind forecast (such as downward transfer of momentum or a notable isallobaric component to the wind).
  4. Evaluating forecast soundings should be a regular part of your forecasting routine. They'll allow you to evaluate lower tropospheric stability, the depth of the well-mixed layer, and the wind profile in the well-mixed layer. Remember, the mean wind speed in the well-mixed boundary layer serves as a reasonable proxy for maximum sustained wind speed, and the fastest winds (usually near the top of the boundary layer) serve as a proxy for maximum surface gusts. Use your analysis of forecast soundings along with statistical guidance to come to a consensus.
  5. If you feel that a particular mesoscale process (like the isallobaric wind or the presence of convection / precipitation) may not be predicted accurately by the dynamic and/or AI models, statistical guidance may suffer, too. Hedge your bets in the direction supported by your analysis of forecast soundings and other factors.
  6. When uncertainty is a big part of the forecast, don't forget the value of ensemble products to help you gauge the uncertainty and hedge your bets (in accordance with the big picture, of course).
  7. Don't forget the value of climatology. Knowledge of a location's climatology can clue you into the spirit of sea breezes, mountain-valley circulations, and other local terrain effects that can pay big dividends when the models are forecasting something that departs greatly from climatology.

Keeping these seven points in mind as you create your wind forecasts will serve you well. At this point, you have a tool box that's full of strategies for wind forecasting. These forecasts can be quite challenging, but if you have a good handle on the big picture and climatology (which should always be a priority), you can make good forecast decisions about when statistical guidance might need an adjustment.

Now, it's time to move on to our final forecast parameter -- precipitation. We'll investigate in the next lesson.

sas405

Lesson 6. Forecasting Precipitation: Controllers and Strategies

Lesson 6. Forecasting Precipitation: Controllers and Strategies

Motivate...

We'll round out our discussion on forecasting specific variables with what is typically the most challenging variable -- precipitation. Forecasting precipitation is very challenging for even experienced forecasters. Take a look at the Weather Prediction Center's threat scores for 0 - 6 hour areal forecasts of at least 0.25 inches during the period from November 2021 through November 2022 (below). The first important message from this graph is that WPC forecasters beat the models (they have higher threat scores).

WPC's threat scores compared to the operational NAM, GFS, and ECMWF models for six-hour forecasts for areas with precipitation totals of at least 0.25 inches from November 2021 through November 2022..

The Weather Prediction Center's threat scores compared to the operational NAM, GFS, and ECMWF models for 0 - 6 hour areal forecasts of at least 0.25 inches from November 2021 through November 2022. WPC forecasters consistently scored better than the models, but threat scores vary widely throughout the year.
Credit: Weather Prediction Center

The second important message comes from the scores themselves and their large variation throughout the year. During the month with the highest threat scores in this period (January), WPC's threat scores peaked near 0.60, which equates to forecasters getting about 75 percent of the predicted area of at least 0.25 inches correct. During the month with the lowest threat scores (August), WPC's threat scores were near 0.26, which equates to forecasters getting about 41 percent of the forecast area correct. And, remember, these statistics are just for the first six hours after the forecasts were issued! If we look at the corresponding threat scores for 66-72 hour forecasts (opens in a new window), the scores are much lower. At a three-day lead time, WPC six-hour precipitation forecasts had a threat score of just 0.11 in August, which equates to forecasters getting about 20 percent of the predicted area of at least 0.25 inches correct. Yikes!

These stats are not meant to be a criticism of the fine forecasters at WPC. They're some of the best! But, quantitative precipitation forecasting can be really difficult for a few reasons. First, the microphysics that relate to precipitation production are enormously complicated and are modeled imperfectly. Secondly, precipitation can be stratiform or convective. Stratiform precipitation offers its own forecasting challenges because it depends on accurately forecasting so many other meteorological features (frontal positions, mid-latitude cyclone tracks, etc.). But convective precipitation is another story entirely. Dynamic models struggle with small-scale convective processes (they're parameterized in most models), and global AI models are trained on data that doesn't capture the fine details of convection very well, so that really limits the skill of precipitation forecasts. Note that in the graph above, the lowest threat scores occur in the warm season, when convective precipitation is more common.

To make matters worse, forecasting precipitation for a single point comes with its own set of enormous challenges, especially when model uncertainty is great and / or precipitation will be convective. If you're forecasting a liquid precipitation total for a specific location, and scattered thunderstorms are in your forecast, the error potential is huge! To see what I mean, imagine you were tasked with making a one-day forecast for a town in Northwest Ohio on a June day. The next day, parts of the region were deluged with rain (see the map of estimated rainfall (opens in a new window)). But, look at the variation in rainfall totals! Within 50 miles or so, totals ranged from more than four inches to less than one inch. As a forecaster, could you have known a day in advance exactly where the heaviest rain would fall, or exactly how much would fall in any given spot? No way.

Nevertheless, despite these challenges, deterministic, point forecasts for precipitation are pervasive on weather apps and weather broadcasts. So, even in the face of great uncertainty, it behooves us to investigate the challenges associated with making these forecasts and identifying strategies for making the best forecasts possible. Ultimately, I think it's important to follow the advice I gave you previously: Think probabilistically while forecasting deterministically. Keep your wits about you and have a forecast plan and reasonable expectations (don't expect to have perfect precipitation forecasts). In this lesson, we'll talk about the important aspects of a prudent forecast plan, namely:

  • identifying potential sources of upward motion (synoptic or mesoscale) based on a solid grasp of the big-picture pattern,
  • wisely using forecast data (dynamic and statistical models), and recognizing the uncertainty in every situation through using ensembles, and
  • checking your forecast against climatology to make sure that, if your forecast nears (or exceeds) a climatological extreme, you have a very good reason for doing so.

We'll cover these issues over the course of this lesson and review some critical tools for assessing the big picture situation and model QPF.  Let's start with our common starting point -- the big picture.

sas405

The Big Picture and Upward Motion

The Big Picture and Upward Motion

Prioritize...

Once you've completed this page, you should be able to identify key features at the major analysis levels (the surface, 850 mb, 500 mb, and 300 mb) that can induce upward motion, and incorporate your knowledge of these features into your weather forecasts.

Read...

By this point in the program, you've heard the phrase "focus on the big picture" many times (it seems like thousands, I'm sure). Well, that broken record is going to repeat itself a few more times in this lesson, because the message really is that important. When it comes to forecasting precipitation, before we worry about detailed quantitative precipitation forecast (QPF) output from dynamic, AI, or statistical models, it's critical to evaluate the big picture. Most importantly, we have to assess sources of upward motion. Of course, upward motion over our forecast location doesn't guarantee that precipitation will fall. But, we won't get precipitation without upward motion!

So, when forecasting precipitation, we first have to look for synoptic sources of upward motion. You've learned about all of these in previous certificate courses (for detailed readings, you may want to review Lessons 8 and 9 in METEO 101 (opens in a new window) (login required) and Lesson 3 in METEO 361 (opens in a new window)). On this page, I am going to quickly review the important features for you to look for when conducting a synoptic analysis. Let's start at the ground and work our way up.

The Surface

Simply put, at the surface, we're looking for sources of convergence: fronts, surface troughs, or regions of surface confluence. Although we often tend to focus on fronts, keep in mind that we need not have a front to have convergence at the surface (troughs and regions of confluence can sneak up on you, if you're not careful). In addition to studying surface forecast maps (which show MSLP isobars and perhaps fronts) to identify likely zones of convergence, forecasts of surface streamlines (like the example below) are another useful tool.

The 24-hour forecast for surface streamlines from the NAM model initialized at 18Z on September 17, 2012, valid at 18Z on September 18.

The 24-hour forecast for surface streamlines showed an obvious axis of convergence from West Virginia north through western Pennsylvania and the eastern Great Lakes. Areas of convergence aren't always this obvious, but streamlines can help you identify even more subtle areas of surface convergence.
Credit: University of Wyoming

Other mesoscale boundaries (like outflow boundaries or dry lines, for example) can provide low-level convergence, too. But, those details are certainly hard to forecast a day or more in advance. Remember that low-level convergence is merely an ingredient that gives us upward motion. One caveat is in order, however. Warm fronts and stationary fronts can be sources of surface convergence, as well, but they're more important because regions of overrunning (formally classified as "isentropic lift" as you may recall) often occur on their cool sides. The topic of overrunning is connected to our next analysis level -- 850 mb.

Key Data Resources

Since surface streamlines can help us identify areas of synoptic-scale convergence, these sites may be helpful:

850 mb

The composite of radar reflectivity at 14Z (9 A.M. local time) on July 14, 2010.

This composite of radar reflectivity at 14Z (9 A.M. local time) on July 14, 2010, shows an elevated mesoscale convective system over northern Minnesota.
Credit: WSI Corporation / @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

As you'll recall, overrunning occurs when warm air glides up over a wedge of cool air (the surface boundary between the two air masses is often marked by a warm or stationary front). Not surprisingly, then, the footprint of overrunning at 850 mb is an area of warm advection. Thus, assessing temperature advection at 850 mb also provides us with clues about upward motion. Check out this example showing strong warm advection (opens in a new window) within a low-level jet stream importing moisture into northern Minnesota, north of a surface warm front (see the 15Z surface analysis (opens in a new window). The end result was an elevated mesoscale convective system, as shown in the radar image on the right. Although the strongest overrunning typically occurs on the cool side of a warm or stationary front, a surface front isn't always necessary for overrunning, so keep an eye out for warm advection aloft!

Also, be sure to keep an eye out for low-level jet streams (or nocturnal low-level jets). These regions of fast winds at 850 mb can maximize transport of moist air in the lower troposphere. Of course, a low-level jet stream isn't present every time it rains, but the heaviest precipitation events (not solely tied to deep, moist convection or terrain effects) do tend to have contributions from low-level jet streams that continually transport moisture into the region. Let's keep moving upward now and take a look at important 500-mb features.

500 mb

Remember tracking air parcels through shortwave troughs (opens in a new window) in your previous studies? Ah, memories... Anyway, that critical animation sums up the major features forecasters look for at 500 mb when looking for sources of upward motion -- shortwave troughs and vort maxes. After air parcels blow through the base of a shortwave trough (through the vort max), their rate of cyclonic spin decreases and their horizontal area increases (to conserve angular momentum). That means upper-level divergence and upward motion occur downwind of a 500-mb vort max.

As an experienced forecaster, when analyzing 500-mb forecast progs, it's important to pay attention to detail; you can't just quickly glance and look for vort maxes marked by an "X." Not every trough will have a distinct vort max marked by an "X", and sometimes subtle vort maxes exist in "sneaky" shortwaves that are hard or impossible to see in the height field. For example, do you see the vort max predicted to be over eastern Ohio on this GFS forecast prog (opens in a new window)? There's no "X" marking the vort max, but the small area of yellow shading indicates an area with higher absolute vorticity than its surroundings (in other words, a vort max). So, pay close attention to the specific location and relative strength of vort maxes you see at 500 mb. Even subtle vort maxes that aren't marked with a nice, tidy "X" can cause gentle upward motion downwind.

Keep the overall pattern of 500-mb heights (troughs and ridges) in mind, too, along with trends in heights over your region. Falling heights go along with cooling in the middle troposphere (which increases mid-level lapse rates), while rising heights go along with warming in the middle troposphere. So, simply by watching the pattern of 500-mb heights, you can pick up general tendencies for destabilization (or stabilization) in the middle troposphere. Now, let's move up near the top of the troposphere.

300 (or 250) mb

Near the top of the troposphere, forecasters focus on wind patterns at 300 mb (or perhaps 250 mb during the warm season). They look at the polar and subtropical jet streams, and specifically, jet streaks. Remember the four-quadrant conceptual model (opens in a new window) of a straight jet streak? In this model, we expect upper-level divergence (and upward motion) to occur in the right-entrance and left-exit regions. But, the key point about this model is that it applies to straight jet streaks. Yes, some jet streaks are straight (or close to being straight), but most are curved. When we're dealing with curved jet streaks, the four-quadrant model breaks down, and a two-quadrant model is more appropriate (see schematic for a cyclonically-curved jet streak below).

The two-quadrant model of a cyclonically curved jet streak.

The two-quadrant model of a cyclonically curved jet streak (shades of blue indicate wind speeds around its core). Note that upper-level divergence and the associated upward motion in the left-exit region can "bleed" into the right-exit region.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

So, with most jet streaks, the right-entrance and left-exit regions may not be the sole locations of upper-level divergence and upward motion. When jet streaks are cyclonically curved, be prepared that some upper-level divergence and upward motion may "bleed" into the right exit region (while upper-level convergence and downward motion may bleed into the right entrance). Ultimately, no single conceptual model works for all jet streaks, so as a forecaster, be on your toes and look for jet streaks passing over or near your forecast location. You may be able to confirm that the jet streak will surely contribute to upward motion, or at least be open to the possibility that it could (sometimes it's hard to tell for sure). Jet streaks aren't the only feature aloft that can have ambiguities, however.

Diffluence / Confluence

I already mentioned surface confluence earlier on this page (surface confluence goes hand in hand with surface convergence), but we can have confluence and diffluence (the spreading apart of streamlines) aloft, too, and the impacts aren't always so clear.

Sometimes diffluence aloft goes along with divergence aloft, but not always. To explain why, focus your attention on the first row of this idealized schematic of confluence and diffluence (opens in a new window). For starters, note the spreading of the height lines from west to east. This increased spacing between height lines from west to east indicates directional divergence (the horizontal spreading of a stream of air). In the second column, wind speed increases from left to right, so there's speed divergence at work. Think of cars accelerating away from a stop light from left to right. There would be "car divergence" as the cars spread out. In this case, there's little doubt that the diffluent pattern is indeed divergent.

In the third column, it looks like there's speed convergence occurring in the diffluent zone. Indeed, air tends to pile-up because wind speeds decrease from left to right. Think of cars approaching a red light: there's "car convergence" as cars slow down and approach the cars that have already stopped for the light. In this case, the diffluent pattern is convergent.

In more realistic diffluent patterns aloft, it's difficult to determine at face value whether diffluent patterns are convergent or divergent. So, we have to look at other big-picture clues (the broader pattern overall) to determine the overall message.

Final Thoughts

Before moving on, it's important to remember a few key points. First, having a good handle on the big picture won't point you to a specific "correct" numerical QPF. Instead, having a good handle on the big picture will help you make good decisions when evaluating the specific QPF output of dynamic, AI, or statistical models (it should help point you in the right direction and help you limit your error). Also, all of these features won't be present every time it precipitates (you may not have a jet streak or a low-level jet stream present, for example, and still have precipitation in the forecast). Still, heavier synoptic precipitation events do tend to have most or all of the features above involved, and forecasters should look vertically throughout the atmosphere at the key levels on this page to assess the presence (or lack of) features that induce upward motion and favor moisture transport.

Finally, I didn't focus on the development of deep, moist convection on this page, but all of these big-picture factors can play a role its initiation (by either changing environmental lapse rates or lifting parcels to the level of free convection). So, if the big picture supports deep, moist convection, it's important to consider that aspect as well (the details of which were the focus of Lesson 3 in METEO 361).

With our brief review of important big-picture factors out of the way, let's look at a specific example of how a forecaster might incorporate the big picture into a QPF forecast. An important Case Study is up next! Read on.

sas405

Case Study: Incorporating the Big Picture

Case Study: Incorporating the Big Picture

Prioritize...

On this page, your focus should be on how a forecaster evaluates sources of upward motion, along with other data (such as dew points and precipitable water) to help with a quantitative precipitation forecast.

Case Study...

As I mentioned on the previous page, incorporating the big picture into your forecast won't point you specifically to the correct QPF on any given day. But, it should point you in the right direction and help you limit your error. To help you see how a forecaster might incorporate the big picture into a precipitation forecast, let's look at an example for New Orleans, Louisiana (KMSY) from a mid-December day. The short video below (5:31) outlines the big picture signs that were available to forecasters on December 14 for the next day.

Imagine you were forecasting for New Orleans, Louisiana for the period from 06Z on December 15 to 06Z on December 16 – effectively the calendar day of December 15 in New Orleans. If you were making your forecast on the afternoon of December 14, your first task would be to examine the big picture. I'll start with the 12Z surface analysis from December 14. The most prominent surface feature near New Orleans was the stationary front draped along the Gulf Coast, which separated maritime-Tropical air over the Gulf of Mexico from a Pacific air mass over the Deep South. Meanwhile, a cold front marking the leading edge of an arctic air mass was plunging southward over the Plains and Midwest from western Canada.

A closer look at the stationary front at 12Z shows large differences in dew points south and north of the front, with dew points in the 70s over the Gulf of Mexico, and dew points in the mid 50s over central Louisiana. At the time, the stationary front was just barely to the south of New Orleans. But the developing southerly and southwesterly winds ahead of the approaching cold front meant that cooler, less humid Pacific air would start to retreat northward, allowing the stationary front to transition to a warm front. Once that happened, the tropical air over the Gulf of Mexico could press inland over New Orleans.

Indeed, if we look at the 15Z surface analysis, we can see that the stationary front was now being analyzed as a warm front.

And, by 18Z on December 14, it had made clear northward progress, thanks to the southerly flow on its cool side causing the cooler, less humid air to retreat.

By 18Z on the 14th, dew points were approaching 70 degrees along the Gulf Coast as the maritime-Tropical air mass forged farther inland, and there was ribbon of moisture convergence, which is a combination of moist advection and surface convergence, marked by the red contours here, along the Gulf Coast.

So, if you were forecasting on the afternoon of the 14th, a good big picture analysis showed the increasingly moist low-level air mass present over New Orleans, which was increasingly favorable for deep, moist convection. The 20Z mesoanalysis of mixed-layer CAPE and CIN showed this favorable environment ahead of the approaching cold front. Furthermore, this "juicy" air mass would remain in place until the cold front passed. All that would be needed for a more organized outbreak of numerous showers and thunderstorms was an organized lifting mechanism, which was on its way in the form of the cold front.

The WPC surface forecast valid at 12Z on the 15th showed that by that time, the front would be on the doorstep of New Orleans early on the 15th.

But, to further complicate matters, the GFS forecast prog valid at 06Z on the 15th from the 12Z run on the 14th showed that winds at 500 mb and 700 mb would be from the west-southwest over New Orleans – largely parallel to the approaching front. These winds suggested that the front would slow down, and that any showers and thunderstorms forming along and ahead of the front could be recurrent at New Orleans.

So, signs were clear that the environment was favorable for numerous showers and thunderstorms around New Orleans early on December 15th as the front approached. We had increasingly moist low-level air in place, and were just awaiting the arrival of organized low-level convergence to trigger more organized deep, moist convection. But, how moist was this air mass? Precipitable water can give us an idea. You may recall that precipitable water is the amount of liquid precipitation that would accumulate if all the water vapor in a local air column of unit cross-sectional area were forced to condense and fall to the earth's surface as rain. There are some important caveats to using precipitable water data, though, namely that it doesn't account for water vapor added to air columns from convergence. Still, the 18Z model analysis of precipitable water from December 14 showed that precipitable water at New Orleans was already 1.75 inches! And, remember, the air mass was continuing to moisten, so that number would likely increase later on the 14th and into the 15th ahead of the front.

The sounding climatology for New Orleans shows that a precipitable water of 1.75 inches in mid-December is remarkable – near the daily records for the time of year. Such values were more in line with typical values for the middle of summer!

Given the extreme moisture in the air mass, and a set-up that favored numerous showers and thunderstorms, the big picture favored a very wet scenario, and that’s exactly what happened – recurrent showers and thunderstorms with heavy rain developed along and head of the front, as shown on this regional radar loop from 06Z -23Z on December 15.

Exactly how much rain fell? A lot. For the six-hour period ending at 12Z on December 15, 3.97 inches of rain fell in New Orleans. It kept raining hard because 1.67 inches of rain were measured in the six hour period ending at 18Z on December 15. The rain was all but over after 18Z because only 0.14 inches fell in the remainder of the period. If you do the math, that's 5.78 inches of rain, making this one of the wettest December days on record in New Orleans.

Credit: Penn State

One big takeaway from this video is that precipitable water is not indicative of a specific quantitative precipitation forecast, but it is helpful in characterizing the moisture in the air mass. In this particular case, the air mass at New Orleans had near record moisture for mid-December, which supports the idea that rainfall could be extreme if sufficient lifting mechanisms were present (and they were here, in the form of a convergence with a slow-moving front helping to initiate deep, moist convection).

What about actual QPF numbers from dynamic models? A grid interpolation from a regional model (opens in a new window) from the 12Z run on December 14 predicted 0.40 inches of precipitation from 06Z December 15 to 06Z December 16. Meanwhile, the GFS grid interpolation (opens in a new window) was much, much wetter. It predicted 2.28 inches of precipitation for the same time period.

Talk about uncertainty! While averaging these numbers to get a consensus forecast (1.34 inches, in this case) often isn't a bad initial strategy, the big picture pointed to the potential for an extreme rain event, and a "middle of the road" forecast of 1.34 inches wouldn't have worked out very well. A hedge toward the wetter side of guidance was a reasonable forecast decision, and if thunderstorms trained right over New Orleans, even the GFS prediction might be too low -- a possibility forecasters needed to consider, given the winds aloft.

But, there's a brutal truth in this case: As a point forecaster, you never could have nailed down the fact that exactly 5.78 inches of rain would fall during this 24-hour period (without pure luck). Even predicting a total within one inch of what eventually fell would have been a phenomenal forecast, so make sure you have realistic expectations about your QPF performance. In big rain events, sometimes even forecasts that perform rather well end up with errors that might seem large. Overall, issuing a QPF range or a probability of meeting particular thresholds would be better approaches if the circumstances of your forecast allow for these formats.

Ensembles to the Rescue?

If you're thinking that ensembles can help forecasters better understand the range of possibilities, and help forecasters make intelligent hedges in light of the big picture, you're correct...most of the time. Ultimately, no ensemble system truly depicts all real possibilities, and occasionally the amount of precipitation that falls is completely outside of the range of ensemble solutions.

Take for example, a case of some surprisingly heavy rain in Greensboro, North Carolina in March, 2021. The GFS forecast for 850-mb temperatures, winds, temperature advection, and frontogensis (opens in a new window), showed a clear signal for warm advection over Greensboro at 06Z on March 18. This warm advection signaled overrunning, which was occurring on the cool side of a warm front extending southeastward from a low-pressure system in the Mississippi Valley (09Z surface analysis (opens in a new window)). Increasing low-level southerly flow over the region also brought ample moisture into North Carolina (GFS forecast for precipitable water valid at 06Z March 18 (opens in a new window)).

Furthermore, GFS forecasts initialized at 12Z on the 17th showed the approach of a deep 500-mb trough over the Mississippi Valley (opens in a new window) (forecast valid at 06Z March 18), which coincided with falling heights and the arrival of cooler air aloft, steepening mid-level lapse rates and generating a small amount of MUCAPE (check out the 09Z mesoanalysis of MUCAPE and lifted-parcel level (opens in a new window)). It turned out that precipitation that fell in the hours after 06Z had a convective element (check out the 0656Z radar image from KRAX (opens in a new window) -- Greensboro is marked by the star), and a fairly quick 0.53 inches of fell in Greensboro by 12Z. Did ensembles give forecasters a heads up about such a possibility? Check out the ensemble plumes from the 15Z run on March 17 for Greensboro below:

SREF plumes for Greensboro, NC, showing that all ensemble members predicted too little rain by 12Z on March 18.

Ensemble plumes for Greensboro, North Carolina from the 15Z run on March 17, 2021. All ensemble members predicted less than 0.20 inches of rain by 12Z on March 18, but in reality, 0.53 inches actually fell. Full-sized image (opens in a new window).
Credit: Storm Prediction Center

Ouch! The 0.53 inches of rain that fell by 12Z was way more than any ensemble member predicted! The amount of rain that fell was outside of the ensemble spread, which unfortunately does happen once in awhile (especially when convection is in the picture). And, this wasn't a simple timing error (note that most members had significant rain after 18Z), because more heavy rain fell later in the day, and the daily total was more than 1 inch, which was outside the ensemble range for the whole day. Essentially, all ensemble members really missed badly on the early round of precipitation.

So, while forecasters were hard-pressed to forecast specific QPF numbers that were close to what actually happened in these cases, connecting to the big picture pointed forecasters in the right direction by suggesting that forecasts toward the wetter side of guidance, or even possibly wetter than the existing guidance, were reasonable. Without connecting to the big picture, forecasters are stuck guessing, or simply playing with numbers and probably taking a consensus forecast. While consensus is a powerful strategy for predicting precipitation, it isn't always the best forecast, and a good analysis of the big picture can help forecasters find times when they can do better.

Forecast Tip

In addition to assessing the big picture for sources of upward motion, it's important to assess the moisture characteristics of the local air mass, too. I suggest the following:

  • Look at forecast soundings to look for deep saturated layers, or alternatively, very dry layers in the lower troposphere that could initially limit precipitation totals through net evaporation.
  • Look at surface dew points as an initial guide for how moist the low-level air mass is (of course, these judgments will be relative to location and season).
  • Check precipitable water values and normalized anomalies or percentiles to get an overall feel for the character of the air mass. Keep in mind, though, that precipitable water values do not serve as exact QPF predictions (all of the water vapor in a column may not condense and fall as precipitation, or additional water vapor may continuously be added to local columns through convergence).

I hope these examples help you to see how useful the big picture can be, even though it doesn't get us exactly to the "right" prediction. At the end of the day, successfully incorporating big picture observations into your forecasts (for precipitation and other variables) is one of the things that separates good forecasters from the rest of the pack. Doing so, however, requires experience, and a sound knowledge of meteorology.

Now let's shift gears from the big picture to some specific quantitative fields predicted by the models that can help us with precipitation forecasting. Keep reading!

sas405

Getting Quantitative with Upward Motion

Getting Quantitative with Upward Motion

Prioritize...

Upon completion of this page, you should be able to discuss omega and interpret it correctly on forecast progs. You should also have a feel for the relative magnitudes of synoptic and mesoscale vertical motion, and be able to incorporate your knowledge of omega into your weather forecasts.

Read...

So far in this lesson, we've been concentrating on the big picture as it relates to precipitation forecasting, especially recognizing the importance of identifying sources of upward motion. Recognizing synoptic features that generate upward motion is a key part of a good big-picture analysis, but just recognizing a shortwave trough or a jet streak only allows us to make qualitative judgments about upward motion (like "strong," "modest," or "weak"). In precipitation forecasting, though, being a bit more quantitative in assessing upward motion can be helpful.

You learned about quantitative assessments of upward motion previously, but I want to jog your memory. First, remember that on the synoptic scale, vertical motions tend to be small in magnitude because on large scales, the atmosphere is approximately in hydrostatic balance (the upward-directed pressure gradient force balances the downward-directed force of gravity (opens in a new window)). So, most large-scale vertical motions are only a few centimeters per second.

Believe it or not, on the synoptic-scale, even when we have some sources of synoptic upward motion, upward vertical velocities will still likely only be a few to several centimeters per second. That means light or moderate stratiform rain can be caused by upward motion as small as a few centimeters per second! Even when several strong synoptic sources of upward motion exist, upward vertical velocities may only be 10-20 centimeters per second (that's less than a half mile per hour). So, how do we put this knowledge to use in forecasting? Do forecasting products exist that show us predicted magnitudes of vertical motion? Of course!

Omega

As you may remember, vertical motion usually isn't expressed in centimeters per second on forecast progs. Instead, omega is plotted. The equation for omega (opens in a new window) might look daunting, but don't sweat it. The dynamic models are performing those calculations for us (thank goodness). For our practical purposes, we just need to know that omega is vertical velocity expressed in pressure coordinates, typically microbars per second or decipascals per second. Omega is typically expressed in microbars per second or decipascals per second because, conveniently, 1 microbar per second equals 1 decipascal per second, which is roughly equal to 1 centimeter per second. Omega is most commonly plotted at 700 mb because synoptic-scale sources of upward motion (low-level convergence, overrunning, divergence from 500-mb shortwaves, etc.) tend to generate a maximum response near 700 mb. So, the 700-mb level serves as a convenient place to detect upward motion from those processes.

There's an important detail about omega, however, that you may not recall: The signs are not intuitive. Negative values of omega equate to upward vertical velocities, while positive values of omega equate to downward vertical velocities. To reconcile this seemingly contrary sign convention, remember that omega is expressed in pressure coordinates. Consider a parcel of rising air. Over time, the parcel's pressure decreases as it ascends higher and higher into the atmosphere. For subsiding parcels, pressure increases during descent. Thus, the sign convention corresponds to the pressure change of the parcel with time -- negative for rising air and positive for sinking air. 

So, by standard convention, positive omega values show us areas of downward motion and negative omega values show us areas of upward motion.  Most "mainstream" websites stick to this convention, as in the GFS initialization of 700-mb heights and vertical velocity (omega) from 06Z on December 19, 2022, below. Note that the negative values of omega marked over Kansas, for example, indicate upward motion.

GFS model analysis of 700-mb heights and vertical velocity (omega) from 06Z on December 19, 2022.

The GFS initialization of 700-mb heights and vertical velocity from 06Z on December 19, 2022. The negative values of omega marked over Kansas (orange and red colors) correspond to upward motion. Positive values of omega (increasingly dark shades of gray) mark areas of downward motion
Credit: Pivotal Weather

But, to further complicate matters, you may come across some websites that reverse this convention. For example, check out the same GFS model initialization for 700-mb heights and omega (opens in a new window) as the one above, just from a different website. On this graphic, the area of upward motion over Kansas is marked by positive values of omega! If you look closely, the units are negative microbars per second. Including the negative sign in the units allows them to plot upward motion as positive numbers.

The bottom line is that you must carefully look at labels, units, contour intervals, etc. since plotting conventions vary from site to site. If you're unsure of a site's convention, look for areas of predicted precipitation (we can't have precipitation without upward motion), or regions of high 700-mb relative humidity on other progs to get your bearings. The fact that rising air cools (causing relative humidity to increase) allows us to often connect the patterns of 700-mb omega and relative humidity.

Key Data Resources

Pivotal Weather (opens in a new window), the College of DuPage (opens in a new window), and Weathernerds (opens in a new window) all are good resources for omega forecast progs (labeled "Vertical Velocity").

Synoptic vs. Mesoscale Upward Motion

You may also recall from your studies of mesoscale forecasting that magnitudes of upward motion in convective updrafts are much larger than synoptic-scale upward motion. In fact, thunderstorm updrafts can be as fast as tens of meters per second. That's hundreds of times faster than the omega values you'll see on forecast progs from the dynamic models. So, you're not going to see accurate representations of convective upward motion on the progs, which also means that even the omega probabilities from ensembles may not be helpful in convective situations.

The small-scale nature of convective processes doesn't just wreak havoc on predictions of omega. You may also recall that there's a statistical correlation between stratiform precipitation and sizable areas of relative humidity greater than 90 percent at 700 mb. That general guideline is only useful in stratiform precipitation cases, when the processes that are creating the precipitation are occurring on the synoptic (or perhaps meso-alpha) scale. In convective precipitation situations, smaller mesoscale processes are occurring below the grid scale of the model, which means that the 700-mb relative humidity patterns in the vicinity of thunderstorms are often depicted inaccurately. The schematic below shows why.

Schematic showing how convective processes are often smaller scale than the spacing between grid points on most computer models.

Convective processes are often smaller scale than the spacing between grid points on most computer models. So, even though the relative humidity inside a thunderstorm is approximately 100 percent, the average 700-mb relative humidity inside the computer model's grid box that contains the thunderstorm might be only 75 percent, for example.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

The fact that relative humidity can vary greatly over short distances in and around convective clouds means that the overall average relative humidity within a given grid box may not be representative. In the example above, the relative humidity within the cumulonimbus cloud is near 100 percent (and the upward motion is probably on the order of 10-20 meters per second), but the model only knows that the average relative humidity in the grid box is 75 percent because of all the relatively dry areas surrounding the thunderstorm. So, looking for blobs of 700-mb relative humidity greater than 90 percent won't necessarily help us identify predicted areas of convective precipitation.

Ultimately, the omega and relative-humidity patterns on 700-mb forecast progs can help us to identify possible areas of precipitation. But, these fields are more useful in stratiform precipitation situations. In convective situations, the omega and relative humidity patterns you see on 700-mb progs will often be misleading because of the importance of small-scale processes, which aren't modeled well. So, it's really not surprising that convective precipitation is predicted so poorly by the dynamic models (AI models, too, for that matter because of limitations in their training data).

But, even stratiform QPF isn't forecast perfectly, either. We'll briefly explore why on the next page, and talk more specifically about problems with model QPF. Read on.

sas405

The Flaws of Model QPF

The Flaws of Model QPF

Prioritize...

On this page, you should focus the reasons why models don't give perfectly accurate QPF predictions, even when they have a good handle on the big picture. Furthermore, focus on the problem of model convective feedback and how it can create erroneous QPF output. You should also be able to properly interpret output from mesoscale "convection-allowing models."

Read...

Now it's time to shift gears from controllers and assessment methods of upward motion to the real heart of precipitation forecasts -- model QPF. I'll start off with a fairly depressing thought: Of all the commonly-used forecast variables predicted by the dynamic and AI models, the least accurate over the long haul is QPF. That, of course, is a major reason behind why forecasting precipitation is so challenging. 

Even model QPF for stratiform precipitation, which is mainly driven by synoptic-scale factors, is flawed. Yes, the dynamic and AI models have a much better chance of reasonably predicting the omega and moisture fields in stratiform cases, and indeed, model QPF does tend to verify better when we're dealing with stratiform precipitation. But, even when the dynamic/AI models have a good handle on the big picture (they provide reasonably accurate forecasts of synoptic features, along with the omega field, etc.), they almost certainly won't give you a perfect depiction of QPF. At best, they'll be in the ballpark. That's right, even when models are handling the big picture well, you should be wary of specific model QPF values.

What's the problem with model QPF for stratiform precipitation? Even in large areas of stratiform precipitation, there are still small-scale processes occurring within the clouds (minor variations in vertical motion and relative humidity) that aren't modeled well. In addition, the microscale physical processes that actually create hydrometeors which fall to the earth as precipitation are parameterized (oversimplified) in dynamic models. So, when we consider all of the factors we can't model very well, we just can't expect perfect QPF predictions from dynamic/AI models. Simply getting within the ballpark is about the best we can ask.

As I've already discussed, convection is another issue entirely. Since the dynamic/AI models cannot adequately depict the omega and relative humidity fields in and around areas of deep, moist, convection, the reasons why model convective QPF predictions are so shaky should be fairly obvious. If the model can't tell how fast the air is moving upward and how moist the air is, good luck to it with its QPF prediction!

Unfortunately, many inexperienced weather forecasters, in a rush to see how much precipitation a particular model might predict, often fixate on a model's QPF depiction because it's easy to interpret. But, that's just about the worst approach to take. Remember, the model has a much better chance of predicting the big picture reasonably well than it does of giving a perfect QPF depiction, so I recommend studying the big picture before trying to dial in on a model's QPF. In a worst-case scenario, really wacky things can appear in a model's QPF prediction, and forecasters must beware.

Models Gone Wild

Indeed, the models can go wild! Occasionally, a bizarre mesoscale bull's eye of dramatically high QPF appears on the progs like the local maximum in six-hour precipitation near Orlando, Florida, on the GFS forecast below, valid at 18Z on September 19, 2006. Such bull's eyes should make you suspicious. They're apparently the footprints of thunderstorms predicted to erupt within an mT air mass (or near its boundaries).

The lower-right panel of the 30-hour GFS forecast from the 12Z run on September 18, 2006 (valid at 18Z on September 19).

The lower-right panel of the 30-hour GFS forecast from the 12Z run on September 18, 2006 (valid at 18Z on September 19), showed a spurious mesoscale bull's eye of high six-hour precipitation totals near Orlando, Florida.
Credit: Penn State University

Sometimes, such small-scale bull's eyes are the footprint of latent heat feedback (also sometimes called "convective feedback") within computer models. In such cases, "buyer should beware", because such predicted precipitation totals can be spurious. For example, a forecaster at the National Weather Service in Melbourne, Florida, made a veiled reference to this latent heat feedback in his afternoon Area Forecast Discussion (opens in a new window) on September 18, 2006.

These extreme bull's eyes result from the flaws of model parameterizations attempting to simulate convection and other sub grid-point processes. When models develop deep convection, they can sometimes overdo the convection, which releases too much latent heat into local air columns (particularly the GFS, based on anecdotal evidence). The erroneously warm air columns can artificially reduce surface pressures, encouraging low-level convergence, and promoting further deep convection, which prompts the release of more spurious latent heat. This sets up a feedback loop that results in a mesoscale bull's eye of unrealistically high QPF (like the one shown above).

Yes, slow-moving thunderstorms can drop several inches of rain over small areas in reality, so the small-scale areas of high QPF that sometimes appear on computer progs might look real to you. But, occasionally, they are spurious features generated by model-induced, latent-heat feedback loops associated with imperfect convective parameterizations. The end result is QPF totals that are too high. Ultimately, problems with latent-heat feedback have improved over the past couple of decades as convective parameterizations have improved, but they still pop-up from time to time, so I suggest viewing mesoscale bulls-eye's of high QPF with an extra critical eye.

Convection-Allowing Models...to the rescue?

With the troubles the dynamic/AI models have with QPF, perhaps you're wondering, "what about convection-allowing models (CAMs)...are they any better?" Well, CAMs (which sometimes get referred to generically as "mesoscale models"), as you've learned, certainly have higher spatial resolutions, and some have a higher temporal resolution than coarser global and regional-scale models. The High-Resolution Rapid Refresh (HRRR), for example, is initialized every hour instead of every six hours like most global models. The higher spatial and temporal resolutions certainly have their advantages for forecasting mesoscale weather.

Convection-allowing models are also very different than models like the GFS or ECMWF (or their AI counterparts) because they actually attempt to realistically simulate convection. The higher resolution of CAMs (generally four kilometers or better) means that convection does not have to be parameterized, and these models can create more detailed, realistic convective patterns and structures. However, "realistic-looking" and detailed doesn't necessarily equate to "accurate."

Models like the HRRR and the RRFS have the ability to realistically capture convective structures that the coarser models couldn't dream of. That can be a help when convective precipitation is on tap. As an example, check out the six-hour forecast for 1 km AGL radar reflectivity from the HRRR initialized at 17Z on April 27, 2011 (valid at 23Z) below.

The six-hour forecast of reflectivity from the 17Z run of the High Resolution Rapid Refresh Model on April 27, 2011 (valid at 23Z).

The six-hour forecast of reflectivity from the 17Z run of the HRRR on April 27, 2011 (valid at 23Z), accurately predicted the timing, location, and structure of a squall line moving eastward across northern Pennsylvania and western New York.
Credit: ESRL

The detail of the predicted squall-line in northern Pennsylvania is certainly impressive! What's more impressive is that the prediction for the feature was quite accurate! Check out the corresponding 23Z radar image (opens in a new window). Note, however, that the model missed the scattered thunderstorms ahead of the line, and completely dropped the ball on the precipitation in northern Ohio. So, some aspects of this HRRR forecast were scarily accurate, while others weren't accurate at all (such "mixed results" are fairly common). On the other hand, this CAM forecast from 12Z on August 9, 2020, valid at 18Z on August 10 (opens in a new window) completely missed a raging derecho in Iowa (18Z regional radar mosaic (opens in a new window)). Ouch!

So, CAMs predict realistic and detailed convective structures, which may or may not be accurate. Generally, it's critical to view CAM guidance in the context of the big picture (does the convective structure seem consistent with the big picture?) and look for some run-to-run consistency before giving added weight to their output.

While QPF predictions from dynamic and statistical models give us precision to the nearest hundredth of an inch, we just can't expect them to do so accurately on a consistent basis. The overall message here is that the best we can expect them to do is get us in the ballpark when they have a good handle on the overall weather pattern. When the dynamic models don't have a good handle on the overall pattern...yikes! The uncertainty can seem overwhelming!

How is a forecaster to make quality QPF predictions in a world filled with forecast uncertainty? We'll explore on the next page.

sas405

Navigating Model QPF

Navigating Model QPF

Prioritize...

When you've finished this page, you should be able to incorporate QPF predictions from dynamic/AI models, statistical models, and ensembles to come up with a reasonable precipitation forecast in the context of the big picture.

Read...

So far in this lesson, we've reviewed the importance of the big picture and concentrated on meteorological factors that plague the accuracy of model QPF. Indeed, as I've commented before, it's important to have realistic expectations when making a precipitation forecast. You're rarely going to nail the forecast perfectly when it precipitates. Your forecast goal should be to get in the ballpark and limit your errors.

Photograph of forecasters gathered at the old map wall in the Penn State Weather Station

Forecasters gather at the old map wall in the weather station (opens in a new window) at Penn State. Through discussing (and sometimes arguing!) various aspects of the forecast, they try to come to a consensus, which is a particularly important strategy when forecasting precipitation.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

The strategy of consensus forecasting can help you limit your error in forecasting any variable, and it's one of the best strategies we have at our disposal when forecasting precipitation. When operational forecasters are on the job, they frequently discuss differences in model guidance and various aspects of the forecast to help them come to a consensus. Studies going back decades have shown that, over time, the consensus of a group of human forecasters tends to be more skillful than most of the individual forecasters themselves.

Even when we're not forecasting with a team of colleagues, we can still find consensus in model QPF output, which is very helpful because unlike with forecasting temperature or wind speed, we don't have other specific methods (like the Delta Method or Analogous Thickness) at our disposal. So, consensus QPF, in light of the day's big picture situation, really is a viable strategy. Skilled forecasters realize that it's not always the best forecast, but over the long haul, it's hard to beat.

Thus, I think it's important to review the key sources of QPF predictions, from which we can come to a consensus:

  • QPF maps from operational dynamic/AI models (GFS, ECMWF, their AI counterparts, RRFS, etc.), from which you need to estimate values for your specific forecast point
  • Dynamic model grid interpolations, which give you a model interpolated value for your specific forecast point (no need for visual estimates)
  • QPF forecasts from statistical models (MOS and NBM)
  • Ensemble QPF forecasts, probability maps, plumes, etc.

That's really quite a bit of QPF data to deal with on any given day, and it's important to synthesize that data before coming up with a forecast. How do we make sense of it all?  Well, here's how I work my way through QPF data if I'm making a point QPF prediction:

Forecast Tips

  1. After assessing the big picture, I look at the accumulated precipitation maps from dynamic/AI models, not only to get an initial sense of the range those models predict for my forecast location, but also at the overall QPF pattern. Is my forecast location near a large QPF gradient? That's an initial sign that uncertainty might be a big issue.
  2. To dial in more specifically on my location, I look at grid interpolations and make notes on their exact predictions for my forecast location.
  3. I follow that up by looking at the QPF from statistical models. MOS predicts a range, which I can compare to the deterministic forecast from the NBM (a bias-corrected, weighted mean).
  4. Next, I look at ensemble data. I like to start with forecast maps of accumulated precipitation (if available for my region), noting both the ensemble mean and individual solutions to get an initial sense of gradients and the overall level of agreement among the members.
  5. After that, I look at ensemble plumes for my location, if they're available. Plumes help me assess the overall range of the solutions (and their distribution), as well as the value of the mean forecast. If convective precipitation is in the picture, I may look at plumes for three-hour precipitation instead of total accumulated precipitation because the individual solutions for total precipitation can get pretty wild. The plumes for three-hour precipitation allow me to more easily see when various members predict heavy bursts of convective precipitation. Plumes also help me assess the uncertainty in the timing of precipitation.
  6. Based on statistical guidance and ensembles, I come up with my QPF prediction, either choosing to stay close to a consensus of ensemble means and statistical guidance, or to deviate from it based on compelling aspects of the big picture.
  7. Finally, I study ensemble probability maps (showing me probabilities of various precipitation thresholds), if available. I use this as a check on my QPF prediction. For example, if I predict 1.05 inches of rain, but only 10 percent of ensemble members predict at least one inch of precipitation, then I'll re-evaluate the big picture before deciding whether I really want to go with what ensembles suggest is a low probability scenario.

An Example

In real-life forecasting, you'll certainly encounter scenarios where clear-cut answers are hard to come by. We've explored some examples when there were clear big picture messages that strongly favored extreme rainfall. But, what about cases in which the signs weren't so clear? Let's look at an example covering a 24-hour time period (from 06Z February 1 through 06Z February 2) at New Orleans, Louisiana (opens in a new window) (KMSY). The short video below (5:22) briefly describes the big picture and a variety of QPF guidance available.

Let’s imagine that we were tasked with coming up with a QPF prediction for New Orleans, Louisiana from 06Z on February 1 – 06Z on February 2, so we’re forecasting for a 24-hour period – essentially the calendar day of February 1 in New Orleans. The surface forecast from WPC valid at 12Z on February 1 showed that New Orleans was predicted to be in the warm sector of a mid-latitude cyclone and on the western periphery of a high pressure system during the period. Such a set up supports onshore flow at the surface, and WPC forecasters also believed that widespread showers and thunderstorms would develop in the vicinity of New Orleans.

Indeed, SPC forecasters noted a chance of thunderstorms in the region, too, as evidenced by their convective outlook issued on the morning of January 31, valid from 12Z on February 1 until 12Z on February 2.

At 500-mb, a weak shortwave was predicted to approach from the west, but this feature was continuing to weaken as it moved closer to New Orleans. There’s only a subtle dip in the height contours to the west of New Orleans on these 36-hour 500-mb height and vorticity forecasts valid at 18Z on February 1.

However, at 850-mb, weak southwesterly flow was predicted to induce weak warm advection as you can see on these 850-mb height, temperature, and wind forecasts valid at 18Z on February 1. The southwest flow was predicted to cross isotherms, transporting warmer air toward New Orleans. So, overall, there was a large amount of agreement in the basics of the big picture: New Orleans would be in the warm sector of a mid-latitude cyclone, and southerly flow in the lower troposphere would help supply the region with warm, moist air, while weak warm advection would support weak upward motion. There were no strong signals that heavy rain was a guarantee for New Orleans, even though the prospects for convection made it possible.

Now let’s dive into the details of QPF guidance. For starters, the NAM 24-hour precipitation forecast for the period ending at 06Z on February 2 has New Orleans in the dark green shading, which means 0.25 – 0.50 inches of rain.

The corresponding GFS forecast had New Orleans very near the 0.25-inch contour, so its prediction was likely very close to 0.25 inches. Overall these progs were pretty similar, and of note, there was a concerning signal in both models that heavier rain would be just southeast of New Orleans. In the GFS, totals more than 1 inch weren’t that far away.

With grid interpolations, we can specifically see what the models were predicting for New Orleans. The NAM was predicting 0.45 inches, and the GFS was predicting 0.23 inches, which is in line with the estimates we made from the QPF maps.

Looking at MOS forecasts from the 12Z runs on January 31, if we sum the Q06 forecasts for this 24-hour period, we get a NAM MOS predicted range of 0.36 - 0.82 inches and a GFS MOS predicted range of 0.21 - 0.57 inches. If we take a consensus of the two ranges, we get 0.29 - 0.70 inches approximately.

What about ensembles? The 3-hour precipitation plumes from the 09Z run ensemble run on January 31 allow me to better visualize the individual members' predictions for heavy convective precipitation. Here, the spikes in the individual members are noteworthy -- a few members predicted three-hour totals near an inch (or higher)! Those spikes are the hallmark of convective precipitation, and knowing that they're part of the mean tells us that the mean is being influenced by predicted bursts of convective rains that may or may not happen right at the airport in New Orleans in reality. In this case, if we sum together the mean three-hour forecasts, we get a total of about 1.25 inches during the period. So, the ensemble mean was much higher than the MOS consensus range because of some members predicting heavy bursts of convective precipitation. If those heavy bursts don't pan out, the ensemble mean would certainly be too high. But, heavy rains from convection could tip totals up toward the ensemble mean or even higher.

Finally, let’s think about things probabilistically with ensemble probabilities. The ensemble probability of at least 0.25 inches in the 24-hour period ending at 06Z on February 2 was over 90 percent.

The probability of at least 0.50 inches was between 70 percent and 90 percent, so the ensemble suggested that such totals were likely, but not guaranteed..

The probability of at least 1 inch was between 50 and 70 percent.

And finally, the probability of at least 2 inches was just over 10 percent. So, the ensemble probabilities suggested that such high totals were pretty unlikely. Overall, the ensemble probabilities sent a clear message that at least a quarter of an inch was highly likely, but the prospects of totals above 1 inch were increasingly questionable, and it’s up to forecasters to synthesize all of this QPF guidance in light of the big picture.

Credit: @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

What important messages can you take from this video? Well, this was a case where the signals for heavy precipitation were not strong, but heavy precipitation was still possible. Indeed, the possibility of convective precipitation can strike fear into any point forecaster's heart.

In this case, taking a consensus of statistical guidance (0.29 inches - 0.70 inches) would have led to a notably lower forecast than an ensemble mean (about 1.25 inches), but it's important to note why the ensemble mean was so much higher: A small number of members predicted very heavy convective rain. Such heavy rains were possible, but not likely, so how could a forecaster reconcile all of this information to come up with a consensus forecast in light of the big picture?

Views on how to tackle that question might vary a bit from forecaster to forecaster, and some might be tempted to go with something greater than one inch noting that ensemble probabilities had more than a 50-50 shot at totals of at least one inch. However, the fact that the heavy bursts of rain depicted on ensemble plumes were dependent on deep, moist convection occurring right over KMSY might suggest going lower. Heavy, convective rain was simply not a sure bet, given the lack of strong synoptic upward motion; therefore, I'd strongly consider starting with the statistical guidance and hedging up a bit for the possibility of heavier rain. A forecast toward the higher end of what statistical guidance suggested might make sense in an effort to limit error -- perhaps something like 0.60 inches. Such a forecast doesn't go "all in" on heavy, convective rain, but hedges in that direction based on the possibility.

So, what happened? Officially, 0.35 inches verified at KMSY (opens in a new window) (all measurable precipitation fell between 12Z and 00Z, actually). So, a forecast of 0.60 inches was too high. There were a few periods of heavy rain, but they were brief -- too brief to really boost total rainfall appreciably. Even though our forecast was too high by a couple of tenths, it avoided disaster by not being overly influenced by the high ensemble totals.

Finally, even though convective scenarios are not a strength of statistical guidance, the MOS consensus range wasn't a bad forecast here. In other words, statistical guidance can still be a useful starting point for your forecast, even though you might decide that an adjustment is warranted. All of the other data at your disposal should help you adjust statistical guidance (to hedge toward a wetter or drier scenario), if necessary, based on the big picture. At the end of the day, though, a consensus of statistical guidance in tandem with the big picture is an approach that's hard to beat over the long haul (even though it won't point you exactly to the right forecast on any given day).

Of course, forecasters have one more tool at their disposal when it comes to forecasting precipitation -- climatology.  Let's investigate.

sas405

Understanding Precipitation Climatology

Understanding Precipitation Climatology

Prioritize...

Upon completion of this page, you should be able to interpret precipitation climatology information, including the evaluation of daily normals in the context of the station's overall precipitation frequency, as well as the "strength" of the daily records.

Read...

As I've emphasized a number of times, part of being a good forecaster is having a good handle on a location's climatology. In terms of precipitation forecasting for a point, it's useful to have basic statistics like your location's normal daily precipitation and record daily precipitation in the back of your mind. Not only can such statistics give us a basic feel for the location's climate during a particular time of year, but they can also alert us to situations when the weather is departing dramatically from climatology, possibly sending up red flags about QPF guidance.

However, it's wise to view precipitation statistics like daily normals and extremes a bit differently than their temperature counterparts because precipitation tends to have greater variability. Knowing a station's normal daily high temperature, and a bit about its location (altitude, latitude, proximity to water, etc.) gives you some clues about how temperature "behaves" there. But, drawing similar generalizations about precipitation from normals isn't as easy.

For example, check out this graph of daily normal and raw average high temperatures (opens in a new window) at State College, Pennsylvania. While the "normal" curve is smoother than the raw averages, it's easy to follow the influence of the seasons on the temperature graph (neither curve has any drastic day-to-day variation). Even if we look at this plot of daily record maximum and record lowest maximum temperatures (opens in a new window) at State College, the data are fairly "well behaved." The red and blue dots, which mark the record highest and lowest maximum temperatures, respectively, for each date, have a few outliers here and there, but overall form fairly smooth curves that mimic the shape of the normal high temperature curve very well.

If we were to look more closely at these graphs, we can see that the high temperature nearly always falls within 25 degrees or so of the normal high for the date (the extremes are rarely more than 25 degrees above or below the normal high for the date, give or take). There are a few exceptions, but for the most part, daily high temperatures tend to cluster somewhat close to the day's average high temperature. This clustering around the mean tends to be even tighter for locations that are moderated by bodies of water, like Eureka, California (opens in a new window), for example.

Precipitation can be a different story. Take a look at State College's graph of unsmoothed daily average precipitation (orange trace) and normal precipitation (gray trace) below. The daily average precipitation is all over the place! It's a very jagged line, although there is a signal showing generally higher average daily precipitation values in the warm season. Beyond that, there's pretty big variation over daily and weekly time scales

The plot of daily average and daily normal precipitation at State College, PA.

The plot of daily average and daily normal precipitation at State College, Pennsylvania shows evidence of seasonal variation, but the shorter-term variation is much larger compared to that of average temperature data.
Credit: Steve Seman (Data from ACIS)

Just take the month of August, for example. Average daily precipitation values range from less than 0.05 inches to near 0.30 inches! Why the big variation? If you guessed that convection is the fly in the ointment, you're right. Over a 30-year period of time (the length of time over which climatological "normals" are calculated), just a few very heavy convective rain events can skew the average for any given date. Smoothing the average values to get the official normals helps a lot, however. Notice how much less day-to-day variation there is in the official normal curve. The official normals paint a "less noisy" picture of seasonal variation -- showing slightly higher daily normal values in the warm season compared to the cold season.

But, raw average or official normal precipitation values can be misleading sometimes. As a hypothetical example, say you come across a city with a daily normal precipitation value of 0.10 inches. Alone, that statistic doesn't really help us with our forecasts. That daily normal could have come from a data set that had many years with no precipitation, but a few heavy rain events, or it could come from a data set that has measurable precipitation most years, but relatively low totals. To see what I mean, consider an imaginary 10-year precipitation climatology for August 1 at two cities, "A" and "B" (I'm just using 10 years to simplify the point).

Ten year precipitation climatology for August 1 at two cities (A and B).
DateCity A Precipitation (in.)City B Precipitation (in.)
08/01/201100.05
08/01/201200.15
08/01/201300
08/01/20140.590.10
08/01/201500.19
08/01/201600.11
08/01/20170.410.06
08/01/201800.07
08/01/201900.10
08/01/202000.17
08/01 AVERAGE0.100.10

Yes, the average precipitation at Cities A and B for August 1 is the same, but wouldn't you agree that the character of the precipitation climatology is vastly different for the date? City A experienced few days with precipitation in the sample, but when it did precipitate, the amounts were higher. On the other hand, City B experienced many days with precipitation in the sample, but the individual amounts tended to be lower. Ultimately, there are a lot of different ways to arrive at a particular climatological normal precipitation value.

Lesson learned: Don't view daily normal precipitation values in a vacuum. Consider them in the context of other data such as daily extreme precipitation to paint a more complete picture of the station's character. Is the station prone to occasional heavy precipitation events that are skewing the normals even though precipitation is infrequent on a particular date? Is precipitation relatively frequent but usually light? A station's location with respect to surrounding bodies of water and terrain, etc., can also provide helpful clues about the character of a station's precipitation climatology.

"Weak" Records

Heavy convective precipitation can also cause great day-to-day variability in daily record precipitation values at a given city. Take, for example, the record daily precipitation values at Pensacola, Florida (opens in a new window), for this 10-day period in late September and early October.

Daily Record Rainfall at Pensacola, Florida.
DateRecord Daily Precipitation (in.)
09/265.63
09/279.10
09/284.48
09/294.37
09/302.96
10/13.02
10/22.57
10/35.00
10/410.45
10/51.71


The fact that there's huge day-to-day variation even over this short time period should immediately jump out at you  The daily record precipitation for October 4 is 10.45 inches, but the very next day, the daily record is only 1.71 inches! Not only is Pensacola prone to convective rains, but it can also fall in the path of land-falling tropical cyclones during this time of year. Not surprisingly, the 10.45 inch daily record for October 4 did, indeed, come from a land-falling tropical cyclone (Hurricane Opal (opens in a new window) in 1995). I like to call such records "strong" records, because they're quite extreme, even in the collection of records! Could they possibly be broken? Sure. But, it would take a very unusual and spectacular meteorological event to do so.

On the flip side, the very next day (October 5) has what I call a "weak" record of 1.71 inches. If you just considered that daily record alone, you might be hesitant to forecast two or three inches of rain (or more) for October 5 because it would shatter the record. But, there's really nothing meteorologically different between October 4 and 5 at Pensacola. Just by chance, a very heavy rain event has never occurred on October 5. So, if you happened to be forecasting for Pensacola on October 5, and suspected that a very heavy rainfall event was in the works, you shouldn't be shy about predicting record-breaking rainfall, as long as it's supported by the big picture. Such a "weak" record is relatively easy to break, and glancing at the other records on the preceding days shows that rainfall of 4 inches or more has occurred previously on most of them. Thus, the 1.71-inch record on October 5 is a bit of an anomaly. If two inches of rain falls on October 5, it's a record breaker. On any of the previous dates in the table, it's just another soggy day.

Lesson learned: Not all daily record precipitation totals are created equal! The large variability largely caused by convection gives us some records that are much harder to break than others. So, it's always a good idea to judge the relative "strength" of a record by glancing at other daily records on surrounding dates.

The Bottom Line

The bottom line is that precipitation climatology data (in terms of normals and records) still provide you useful context as a forecaster, but to get the most value from them, it's important to view them in proper context (don't view any single statistic in a vacuum). Studying a station's climatology as a whole (impacts of geography, seasonal variations, etc.) will help give you a more complete picture.

With the shortcomings of QPF and an understanding of climatology fresh in your mind, let's examine a Case Study detailing the factors that forecasters need to think about when assessing whether extreme precipitation might be in the cards. Read on.

sas405

Forecasting Extreme Precipitation

Forecasting Extreme Precipitation

Prioritize...

After you've completed this section, you should be able to discuss the big picture weather patterns typically associated with extreme precipitation events, and identify whether a particular pattern might lend itself to producing record precipitation.

Read...

With model QPF struggles and precipitation climatology fresh in your mind, let's look at how the two come together in extreme precipitation events. When extreme precipitation is in the picture, the consequences of a poor precipitation forecast can be extremely disruptive or even dangerous.

What should forecasters look for when thinking about extreme precipitation? Ultimately, if you recall the "First Law of QPF": The greatest precipitation occurs where it rains the hardest for the longest, then forecasters are looking for synoptic or mesoscale factors that cause precipitation to be very intense (which requires strong upward motion and ample moisture) and/or last for a long time (which is often related to the movement of weather systems).

In more concrete terms, forecasters should keep an eye out for synoptic patterns that:

  • contain multiple synoptic lifting mechanisms to produce strong synoptic lifting through a deep layer of the troposphere, and/or are prone to deep, moist convection
  • have anomalously high moisture present (precipitable water values well above normal for local climatology, or approaching historical extremes),
  • cause precipitation to recur over the same area or move very slowly over a given area (often marked by weak upper-level winds or upper-level winds largely parallel to a surface boundary).

Patterns that cause extreme precipitation may not have all of these characteristics, but they typically have at least some of them. Patterns that lack strong synoptic lifting, for example, may still produce mesoscale areas of extreme precipitation in a very moist air mass if mesoscale lifting and convection cause precipitation to recur over the same area or if precipitation moves very slowly.

To help you see these factors in action, let's look at the factors that led to serious flooding in parts of eastern Pennsylvania, New Jersey, and New York as the remnants of Hurricane Ida (opens in a new window) moved through the Mid-Atlantic in early September 2021. Ida's remnants dropped a narrow swath of 6 - 8 inches of rain (opens in a new window) (or more, locally) from southeastern Pennsylvania through New York City, most of which fell in just a several hour period. In fact, flooding from the remnants of Ida caused more deaths in the Mid-Atlantic than its landfall in Louisiana, and it produced terrifying footage of flooding in basements and underground apartments like this one from Cranford, New Jersey (opens in a new window). To see the synoptic set-up associated with Ida's prodigious rainfall, check out the Case Study below.

Case Study...

The short video below (5:04) describes the aspects of the pattern associated with the remnants of Hurricane Ida that favored extreme rainfall in the Mid-Atlantic.

Let’s look at the synoptic signs that indicated that the remnants of Hurricane Ida would produce extreme rainfall in parts of the Mid-Atlantic in early September, 2021. At the surface, WPC surface analyses starting at 21Z on September 1 showed a strong surface low with attendant fronts tracking from Maryland to the southern New England Coast by 09Z on September 2. Specifically, note the warm front extending eastward from the low’s center.

The cyclonic flow around post-tropical cyclone Ida brought moist, tropical air into the Mid-Atlantic, and GFS forecasts for precipitable water initialized at 00Z on September 1 and valid through 06Z on September 2 showed a surge of very moist air just ahead of the low. These pinkish / purplish shaded areas show precipitable water values in excess of 2 inches streaming into southeastern Pennsylvania, New Jersey, the New York City area, and southern New England from about 21Z on the first through 06Z on the second.

That’s really moist air, but for context, let’s see how such values would compare historically. This is the Storm Prediction Center’s sounding climatology for Upton, New York, located on Long Island, showing precipitable water values throughout the year since the 1950s. The light blue curve shows the moving average of the 10th percentile observations, while the orange curve shows the moving average of the 90th percentile observations. Values greater than 2 inches in early September are way up here by the red curve, which marks daily maximum values of precipitable water on observed soundings. In fact, the 00Z sounding on September 2 had a precipitable water value of 2.08 inches, so the precipitable water values were near historical extremes.

The GFS forecast for 850-mb heights and winds valid at 00Z on September 2 showed a notable low-level jet stream into southeastern Pennsylvania, New Jersey, and New York City soon thereafter, with peak wind speeds near 50 knots.

The low-level jet stream would also lead to strong warm advection north of the surface warm front, shown on this 850-mb forecast from the 3-km NAM valid at 03Z on September 2. This warm advection signaled overrunning, and these purple contours indicated strong frontogenesis, which can lead to narrow bands of strong upward motion within the broader area of overrunning.

At 500-mb, the axis of a trough was predicted to run through central Pennsylvania down into the Southeast at 00Z on September 2, placing locations to its east in a region prone to upper-level divergence and upward motion.

Finally, at 250-mb, the GFS predicted a potent jet streak with winds in excess of 140 knots at its core north of the region at 00Z on September 2, putting eastern Pennsylvania over to southern New England near the periphery of the right entrance region, which is a favorable region for upper-level divergence and upward motion.

So, just to summarize the big picture, eastern Pennsylvania over to southern New England had near-record moisture present, a near-record strength low-level jet stream present, contributing to strong overrunning, or isentropic lift, north of the warm front, as well as strong frontogenesis, which can cause narrow bands of strong upward motion. And, to top things off, the region was in a favorable position with respect to a 500-mb trough axis and a 250-mb jet streak for upper-level divergence and upward motion. So, the environment favored strong lifting through a deep layer of the troposphere, and anomalously high moisture was present – both can be key ingredients in an extreme precipitation event.

The end result was a band of very heavy rain that fell on the evening of September 1 into the wee hours of September 2. These radar estimates show that the heaviest rains occurred from southeastern Pennsylvania into northern New Jersey and the New York City metropolitan area, then into southern New England. The data points here are actual gauge reports, which show observations of 6-8 inches or more from the northwestern suburbs of Philadelphia over toward New York City – likely in the zone where frontogenesis banding occurred.

Just looking at a few select single locations, Newark Airport measured 8.31 inches of rain, and more than 6 inches of that fell in the 6-hour period from 00Z – 06Z on September 2. Central Park in New York City and LaGuardia Airport also measured totals near 6 inches in that 6 hour period. Central Park ended up measuring more than 7 inches in total, and LaGuardia measured just shy of 7 inches. All of this rain fell over an 8 to 9 hour period, resulting in tremendous flash flooding.

Credit: Penn State

As illustrated in the video, this case had the synoptic hallmarks of an extreme precipitation event -- multiple mechanisms to produce strong lifting through a deep layer, and near-record moisture present. Just how extreme was the rainfall in this case? Well the remnants of Ida easily crushed rainfall records for the date, and resulted in a top-5 wettest day overall at Central Park in New York City, LaGuardia Airport, and Newark Airport at the time (with records at Newark and Central Park going back to before 1900). So, this was a really significant rainfall with deadly impacts.

Publicly available short-term forecasts before the event did point to the likelihood of a truly rare event. The 12-hour QPF map from WPC (opens in a new window) issued at 1748Z on September 1, valid from 18Z on September 1 through 06Z on September 2 showed a widespread area with more than 4 inches of rain from southeastern Pennsylvania to southern New England, with a maximum of 5 to 7 inches. The forecast issued the day before (opens in a new window) covering the same 12-hour period carried a similar message, although the details differed.

Although WPC's forecasts underestimated the highest rainfall totals a bit (some areas received in excess of 8 inches of rain), its forecasts were pretty good considering the QPF guidance available at the time. As the GEFS forecasts for LaGuardia Airport (below) from 00Z on September 1 show, WPC's forecasts of 4 to 7 inches of rain in the region were on the high end, or even outside the range of ensemble solutions for the event. The wettest GEFS member predicted about 4.25 inches of rain at LaGuardia. The GEFS mean was just shy of 2 inches.

GEFS plumes for KLGA from the 00Z run on September 1, 2021.

GEFS plumes for LaGuardia Airport (KLGA) from the 00Z run on September 1, 2021 showed that all ensemble members underestimated the amount of rain that would fall from the remnants of Ida.
Credit: NCEP

Nearly 7 inches of rain fell at LaGuardia Airport, so what actually fell was way outside of the range of GEFS forecasts. Short-Range Ensemble Forecasts weren't much better for KLGA. The wettest member of the 03Z run on September 1 (opens in a new window) predicted about 5.25 inches of rain, and the mean was near 2.50 inches. Notably, many members also had the rain falling over a longer period of time that what actually occurred.

Overall, hugging the ensemble mean wasn't a good forecast strategy in the areas where the best synoptic and mesoscale lifting occurred. Bias-corrected, weighted mean forecasts from the NBM were all much too low, as well. During the 6-hour period when Central Park, LaGuardia Airport, and Newark Airport all received near 6 inches of rain or more, the NBM predicted less than 2.5 inches of rain (opens in a new window). The NBM's overall rain forecasts were also spread out over a longer period of time than what actually happened and were still much too low overall.

While predicting the detailed nuances of the location and exact amounts of heaviest rainfall were a huge challenge (as they always are), the big picture (a combination of strong synoptic and mesoscale lifting mechanisms and historically significant moisture present) sent a strong message that ensemble mean forecasts likely wouldn't cut it in this case. Forecasters were able to hedge their forecasts up toward the high end of ensemble guidance, or in some cases even higher, and such forecasts sent a message that a major and dangerous rain event was on the way. (opens in a new window) It's unrealistic to expect such forecasts to have all the precise QPF details correct given everything you've learned, but the forecasts still had value.

I hope this example shows you how forecasters can recognize patterns that may lend themselves to producing extreme rainfall, and how recognizing the signs can help forecasters hedge appropriately away from ensemble mean forecasts. Forecasters also have ways of quantifying some patterns that may lead to extreme rainfall, which we'll cover next. Keep reading!

sas405

Atmospheric Rivers

Atmospheric Rivers

Prioritize...

After you've completed this section, you should be able to describe atmospheric rivers, integrated water vapor transport (IVT), and identify the meteorological conditions that produce higher IVT values. Finally, you should be able to identify atmospheric rivers using forecasts of IVT and classify an atmospheric river as primarily beneficial or primarily hazardous.

Read...

Photograph of the Mississippi River at Vicksburg, MS

The mighty Mississippi River (shown here at Vicksburg, Mississippi) is a prolific mover of water.
Credit: EarthCam

Rivers do the job of carrying water from one place to another, and the mightiest rivers carry a lot of water from one place to another. The Mississippi River is a great example, acting as the second-longest river in the largest river drainage basin in the continental United States. After originating in Minnesota, the river flows south for more than 2,300 miles before emptying into the Gulf of Mexico. On average, the mighty Mississippi discharges a whopping 593,003 cubic feet of water (16,792 cubic meters) per second into the Gulf according to the National Park Service!

But, the rivers of liquid water that snake across the surface of the earth aren't the only rivers that meteorologists are concerned with. Another type of river has a similar job of transporting a lot of water over comparatively great distances in a short amount of time through the atmosphere. While these "atmospheric rivers" transport water (in the form of water vapor) that can bring beneficial rains, when they become extreme, their impacts can sometimes be destructive and even deadly.

Atmospheric Rivers (ARs) are simply long, narrow corridors in the atmosphere (250-375 miles wide, on average) that transport copious amounts of water vapor from the tropics and subtropics to the higher latitudes. At any given time, multiple atmospheric rivers are flowing across the globe, often with very little fanfare. In the image below showing satellite-derived total precipitable water, several narrow channels of high precipitable water are drawing copious moisture from the tropics into the mid-latitudes as part of the typical circulation around strong mid-latitude cyclones. While atmospheric rivers have become an increasingly hot topic of research, they are not a new phenomenon. For example, forecasters have for decades colloquially referred to a long, narrow corridor of deep moisture stretching from near Hawaii toward the west coast of North America (like this example from January 31, 2024 (opens in a new window)) as a "pineapple express."

A large scale image of precipitable water patterns reveals several potential atmospheric rivers.

The Morphed Integrated Microwave Imagery from CIMSS (MIMIC) depiction of precipitable water at 01Z on January 10, 2024 reveals several narrow channels of high moisture content being drawn into the middle latitudes from the tropics, including one near the U.S. East Coast.
Credit: CIMSS

But, does any narrow channel of high precipitable water qualify as an atmospheric river? Not exactly. Atmospheric rivers are classified by a unit of measure called Integrated Water Vapor Transport (often abbreviated IVT). IVT combines the total amount of water vapor in the atmosphere with the total amount of wind that is moving the vapor around. To qualify as an atmospheric river, the narrow channel of moist air must have a maximum IVT magnitude of at least 250 kilograms (of water vapor) per meter per second. So, keep in mind that atmospheric rivers don't just depend on the amount of water vapor present; they depend on water vapor transport, which means both the amount of water vapor and wind speeds are important. IVT values are maximized with high precipitable water and fast wind speeds. Areas with high precipitable water, but little wind, would yield lower IVT values.

The Center for Western Weather and Water Extremes has conducted much research into atmospheric rivers, and their work has resulted in a scale that classifies their strength and helps forecast their potential impacts days before they actually occur. The scale spans five categories from weak to exceptional (opens in a new window). Categories 1 and 2 represent a primarily-beneficial AR. Category 3 represents an AR that is a roughly equal mix of beneficial and hazardous impacts, while categories 4 and 5 represent a primarily-hazardous AR. While this scale was developed specifically with the western U.S. in mind, when IVT values begin to surpass 1,000 kilograms per meter per second for 24 hours or more in other land areas, the AR is highly likely to cause damage.

Note that it's not just the IVT value that determines the category on the scale, however. The duration of the event also plays a really important role. For example, an atmospheric river with a maximum IVT of 1,100 kilograms per meter per second that lasts less than a day over a particular area would be a Category 3, but if the same atmospheric river impacts the same area for more than two days, then it's a Category 5 event that will almost certainly cause damage. Ultimately, an atmospheric river can transition from being mostly beneficial to mostly hazardous via three factors:

  • transporting an extremely large amount of water vapor (very high IVT values)
  • the atmospheric river impacts a particular area for an unusually long time (2-3 days or more)
  • the direction of water vapor transport aligns perpendicularly to a coastline or mountainous terrain (which favors persistent orographic lift (opens in a new window)).

Forecasters identify potential atmospheric rivers by looking at forecasts of IVT, like the GFS initialization for IVT and mean sea-level pressure valid at 00Z on November 5, 2022. The vectors on the prog illustrate the direction of water vapor transport, while the colors represent IVT values. This particular prog shows an atmospheric river pointed right at the northwest Pacific coast with maximum IVT values in excess of 1,000 kilograms per meter per second.

GFS forecast of IVT for at atmospheric river event impacting the northwest Pacific Coast in November, 2022.

The GFS analysis (0-hour forecast) for IVT and mean sea-level pressure valid at 00Z on November 5, 2022. The vectors on the prog illustrate the direction of the vapor transport, and the colors represent the IVT value.
Credit: Center for Western Weather and Water Extremes

This particular atmospheric river brought together a couple of ingredients that pushed it to become primarily hazardous. The tremendous water vapor transport classified this as a Category 4 atmospheric river, and the transport was aimed perpendicularly to the Washington and Oregon coastline and the Cascade Mountains (opens in a new window), ensuring persistent orographic lift. The result was 2-10 inches of liquid throughout the Cascades and coastal mountain ranges (opens in a new window) in a 24-hour period, which caused flooding and debris flows, knocked out power to hundreds of thousands of customers, and led to more than a dozen fatalities (check out a map of power outages and damage photos from Washington (opens in a new window)).

Along the U.S. West Coast, nearly all flood-related damage is tied to atmospheric rivers, and when averaged across the entire western U.S. (including inland areas), 84 percent of all flood-related damage is attributable to atmospheric rivers. As this map showing the proportion of National Flood Insurance Program insured losses due to atmospheric rivers (opens in a new window) shows, the percentage of flood damage attributable to atmospheric rivers is significant even in some areas of the Intermountain West.

Atmospheric rivers make news most frequently when they are severe enough to cause damage; however, quite a few locations rely on atmospheric rivers for a majority of their annual water supply. Atmospheric rivers deliver more than 65 percent of the average yearly precipitation to the west coasts of continents, which means that they can single-handedly make or break the wet season in locations that depend on their moisture. An unusually low number of atmospheric rivers impacted California from 2020-2022, for example, allowing drought to envelop 100 percent of the state by summer 2022 (August 2, 2022 drought monitor (opens in a new window)). Then, an onslaught of atmospheric rivers from November 2022 through April 2023 had the opposite effect -- filling up reservoirs and cutting the percentage of the state in drought to less than 5 percent by summer 2023 (July 4, 2023 drought monitor (opens in a new window)).

The U.S. West Coast is one of the most prominent areas of the world for atmospheric rivers with places like California typically experiencing between one and two dozen atmospheric rivers per year; however, many other coastlines experience multiple atmospheric rivers per year, on average. This map of atmospheric river frequency (opens in a new window) shows that even Antarctica and places above the Arctic Circle typically experience a few atmospheric rivers a year! To see an example of an atmospheric river impacting the U.S. East Coast, check out the Case Study below.

Case Study...

Many eastern U.S. residents got a taste of an atmospheric river barrage in the winter of 2023-2024. This culminated on January 9, when a Category 5 atmospheric river was forecast to impact the eastern U.S. The 12-hour GFS forecast for IVT values valid at 00Z on January 10 (opens in a new window) exceeded an astounding 1,500 kilograms per meter per second from Florida to Virginia, while values greater than 1,000 kilograms per meter per second extended as far north as central New England.

Such high values of integrated water vapor transport existed because this atmospheric river had connections to the tropics, and not just to origins in the Western Caribbean. Using the GFS forecast for precipitable water valid at 00Z on January 10, 2024 (below), we can arguably trace the January 9 East Coast atmospheric river from New England all the way back to the tropical Pacific Ocean -- a moisture feed of over 4,000 miles!

GFS forecast for precipitable water showing a continuous chain of high moisture content from New England back to the tropical Pacific.

The GFS forecast for precipitable water valid at 00Z on January 10, 2024 (initialized 12 hours earlier) shows moisture feeding into the atmospheric river encroaching on New England extended all the way back to the tropical Pacific.
Credit: Tomer Burg / polarwx.com

This storm came on the heels of a significant snow event that dumped 6 to 16 inches of snow on the interior northeast just three days prior. The combination of a rapid melting of the snow pack, along with the expectation of a very strong atmospheric river allowed the Weather Prediction Center to issue a moderate risk (level 3 out of 4) for flash flooding with a lead time of five days (opens in a new window). That amount of “heads up” for potential flooding is pretty rare, but it speaks volumes about the utility of identifying extreme atmospheric rivers (especially in the context of antecedent conditions like a considerable melting snow pack). The flooding from this event was widespread in the Mid-Atlantic and Northeast (opens in a new window), and hundreds of thousands were left without power.

Lesson learned: when a moisture-laden storm appear likely, adding IVT model progs to your toolbox can help reveal if an atmospheric river is the culprit, and help you assess whether it may bring beneficial precipitation or might rise to the level of something much more severe. Knowledge of antecedent conditions, like recent heavy rains or a snow pack that may melt, helps, too! Speaking of snow, we'll shift our focus to wintry precipitation from here on out, but I encourage you to check out the forecasting resources in the Explore Further section below before moving on, if you're interested.

Explore Further...

Key Data Resources

Looking for forecast resources to track atmospheric rivers? Model forecasts of precipitable water can be a starting point, but to get a more detailed look with IVT, these resources may be helpful:

A Deeper Dive

sas405

Snow Forecasting Products and Problems

Snow Forecasting Products and Problems

Prioritize...

When you've finished this page, you should be able to discuss the effective use of model snow maps and their limitations.

Read...

Near the end of METEO 101, you learned about snow forecasting (opens in a new window) (login required) (and how challenging it can be). After our deeper look at liquid precipitation forecasting in this course, the challenges of snow forecasting should be even more apparent. For a computer model to produce a snow forecast that is accurate, it needs to have a good handle on the amount of liquid precipitation that a storm will produce (an iffy proposition, as you know) and the correct snow-to-liquid ratio must be applied (which is not always straightforward). Ultimately, many of the model snow maps that are available on the web are based on either faulty liquid QPF predictions, incorrect snow-to-liquid ratios, or both.

If you're rusty on the fundamentals of snow forecasting, I recommend refreshing yourself on the METEO 101 section linked above. But, as a reminder here, the snow-to-liquid ratio is a critical part of any snow forecast and it depends on a few main factors:

  • temperatures in the boundary layer and at the ground
  • temperatures and humidity where the snowflakes form
  • dominant type(s) of snowflakes

But, even though snow-to-liquid ratios vary in time (even within a single storm) and space, perhaps the most commonly found model snowmaps apply a fixed ratio of 10-to-1 (10 inches of snow comes from 1 inch of liquid), which is nothing more than a first approximation. In fact, in a large sample collected from nearly 8,000 stations in the United States over a 30-year period (distribution shown below), the median snow-to-liquid ratio was actually 12-to-1, with considerable variations geographically and by season.

Distribution of snow-to-liquid ratios in a large sample collected from nearly 8,000 stations across the United States over a 30-year period.

Observed snow-to-liquid ratios taken over a 30-year period from nearly 8,000 stations in the contiguous U.S. show that ratios between 10-to-1 and 12-to-1 are the most common, with a median close to 12-to-1 (long dashed line) and a mean greater than 13-to-1 (short dashed line). The solid lines represent the 25th percentile (about 9-to 1) and 75th percentile (almost 17-to-1).
Credit: Weather and Forecasting 20, 5; 10.1175/WAF856.1 (opens in a new window)

The 75th percentile snow-to-liquid ratio in the sample was almost 17-to-1, meaning that about 25 percent of snow cases had a snow-to-liquid ratio that was almost double (or more!) the 10-to-1 ratio that many model snow maps are based on. Yikes! Fortunately, meteorologists have developed methods to more "dynamically" estimate snow-to-liquid ratios based on model predictions for variables that impact the actual ratio in a given situation. By "dynamically" I just mean that the ratio can change throughout an event based on atmospheric conditions.

One of the more common alternatives is the "Kuchera Ratio," named after Air Force Meteorologist Evan Kuchera, who developed the technique (many models sites include "Kuchera" snowfall maps). The Kuchera snowfall ratio is based on the model predicted maximum temperature between the surface and 500 mb in local air columns. This plot showing the Kuchera Ratio versus maximum temperature in the air column up to 500 mb (opens in a new window) gives you the basic idea that colder columns will yield higher snow-to-liquid ratios and warmer columns will yield lower ratios (all the way down to zero when the maximum column temperature gets to about 4 degrees Celsius). The plot also shows negative snow-to-liquid ratios in even warmer columns, but this is just an extrapolation and such air columns typically can't produce snow anyway.

Air column temperatures are a really important factor in determining snow-to-liquid ratios, but they're not the only factor. So, the Kuchera ratio has some limitations. First, obviously it's possible that the model-predicted maximum temperature predicted in an air column could be wrong (air columns could end up warmer or colder than a model predicts). Secondly, the method doesn't take into account exactly where in the air column the maximum temperature occurs. If it's found right near the surface, the outcome could be very different compared to a case where it's found few thousand feet above ground, for example. Finally, it doesn't take into account the dominant types of snowflakes that will form. For example, in a relatively cold column, the high ratios predicted by Kuchera may or may not come to fruition depending on the dominant snowflake type (opens in a new window).

Kuchera's method isn't the only "dynamic" method for estimating snow-to-liquid ratios, and if you're interested in learning about a few others, check out the Explore Further section below. Overall, which is the "best" approach to estimating snow-to-liquid ratios? That question doesn't have a single answer. In a recent study of a few common approaches verified at CoCoRAHS sites (opens in a new window), no single snow-to-liquid ratio universally performed the best. On most days in the sample, the Kuchera Ratio verified the best at the highest percentage of observation sites, but on most of those days it verified the best at less than half of the sites (it just happened to have the largest percentage). On any single day, each ratio in the study performed the best at some stations while performing the worst at other stations.

Ultimately, as a forecaster, it's important to keep in mind that no single snow ratio works best in all circumstances. Even over the geographic scope of a single snow storm, a 10-to-1 ratio might work out the best in some areas, while the Kuchera Ratio might be best in other areas, and so on. Therefore, any large-scale snow map that applies one of these approaches is bound to have flaws. But, even beyond the snow-to-liquid ratio used for a particular model snow map, snow maps can differ in other ways.

All Snow Maps are Not Created Equal

The first thing to remember when looking at a model snow map is that all snow maps are not created equal. Even corresponding snow maps from the very same model, same initialization time, and same valid time might give you a different snow forecast! To see what I mean, check out the two images below and focus on the forecasts along the southern periphery of the area of snow in northern Indiana, northern Ohio, Pennsylvania and New Jersey.

Corresponding GFS snowfall forecasts with a 10:1 snow ratio and a dynamic ratio.

GFS forecasts with a 10-to-1 snow-to-liquid ratio (left) and a dynamic ratio (right) initialized at 06Z on February 20, 2023 and valid at 00Z on February 25, 2023 show very different solutions. Larger images (opens in a new window).
Credit: Tropical Tidbits / Weathernerds.org

Both of these forecasts are from the GFS and cover the same time period (forecasts were initialized at 06Z on February 20, 2023 and valid at 00Z on February 25, 2023), so why are the forecasts so different (especially along the southern periphery of the predicted area of snow (opens in a new window))? For starters, they differ in the snow-to-liquid ratio applied (the map on the left employs a 10-to-1 ratio, while the map on the right uses a dynamic ratio based on atmospheric conditions -- more on this particular approach in the Explore Further section below, if you're interested), but the differences don't stop there. Notice that some places showing accumulating snow on the map on the left in northern Ohio, for example (up to several inches) don't have any snow in the map on the right!

How can that be possible? Well, the map on the left includes sleet in its snowfall totals (at an unrealistic 10-to-1 ratio, no less), while the map on the right has filtered out frozen precipitation types that are not snow. That makes for some major differences in areas where the precipitation type may not be snow throughout the duration of a storm! And, keep in mind that these differences are determined largely by folks who create the graphics for various model websites. They can take the same model data and produce different looking forecasts with it based on decisions about snow-to-liquid ratio, precipitation-type filtering, etc. Unfortunately, the differences are not always transparent!

Both of the maps above show "total snowfall," which on most sites refers to the total amount of snowfall (or in some cases sleet, too), that falls to the Earth's surface, which isn't the same thing as how much snow actually accumulates. To get a better feel for how much snow might actually accumulate, forecasters sometimes turn to forecasts that show snow depth change instead of total snowfall. Snow-depth change forecasts attempt to account for processes like melting and compaction which determine how much snow actually accumulates on the ground. The actual modeling of these processes is imperfect and the details vary from model to model, but not surprisingly, forecasts of total snowfall and snow depth change can be quite different. Check out the same GFS 10-to-1 total snowfall map above compared to the corresponding GFS forecast for positive snow depth change (opens in a new window) as an example (use the slider to see the differences between the two images).

In many areas, the positive snow depth change forecast is a few (or even several) inches less than the total snowfall forecast. But, even forecasts for snow depth change can differ in the same ways that total snowfall maps differ (precipitation type filters, snow-to-liquid ratios, etc.). For example, check out the corresponding GFS forecasts for snow depth change from the model initialization at 06Z on February 20, 2023 through its valid time at 00Z on Februrary 25 below.

Corresponding GFS snowfall forecasts for positive snow depth change

Even maps from the same model, model run, and valid time for snow depth change don't always produce the same forecast. How do these two GFS forecasts for snow depth change differ? Larger images (opens in a new window).
Credit: Tropical Tidbits / PolarWx

Much like with the total snowfall forecast graphics, the snow depth change forecasts differed greatly along the the southern periphery of the area of snow (opens in a new window) in northern Indiana, northern Ohio, Pennsylvania and New Jersey (and even a bit farther north). Even though the image on the right is labeled "Total Snow" it actually is a forecast of snow depth change with filtering that excludes frozen precipitation that isn't snow. The snow depth forecast on the left is still including other precipitation types like sleet in its snow depth change forecast, leading to dramatic differences along the southern periphery of the area of snow. But, they both tend to show lower amounts than the total snow forecasts described earlier (opens in a new window) because they attempt to account for processes like melting and compaction that affect how much snow piles up on the ground.

In this case, the products with dynamic ratios and better precipitation-type filtering (excluding sleet, etc.) tended to perform much better on the southern periphery of the storm (verification (opens in a new window)). Standard 10-to-1 snow maps and those that included sleet in the total snowfall tended to be much too high. Clearly, forecasters need to be really careful when incorporating various model snow forecast maps, which leads me to the following forecast tips:

Forecast Tips

When using model snow forecast products, keep the following in mind:

  • Note the snow-to-liquid ratio used for the map. 10-to-1 maps are often not your best choice! "Dynamic" approaches to estimating snow-to-liquid ratio (like Kuchera) will often (but not always) give better results.
  • Never take snow totals at face value without doing a complete analysis of the big picture. Snow maps can be useful in getting a quick look at where a swath of snow may occur (and were higher amounts may occur generally).
  • Be careful with any total snowfall map early or late in "snow season." A snow event in fall or early spring, for example, may have significantly lower accumulation because of marginal boundary layer temperatures and melting. Snow depth change forecasts may be much more useful than total snowfall forecasts in these situations. The same idea applies to any forecast in which precipitation type may change throughout the storm -- the potential for error on any particular snow map is quite large.
  • The details of any snow forecast product (whether sleet is included or filtered out, for example) can vary from website to website, and they're not always transparent. You may need to compare forecasts from more than one site to get your bearings.

The Bottom  Line

The bottom line is that I firmly believe the advice given in METEO 101 still applies: Even with increasingly sophisticated snow forecast products available, an "old fashioned" approach to snow forecasting still has merit. Remember, model snow maps are can be flawed because of flaws in the snow-to-liquid ratio used or simply because the model errs greatly in how much liquid that a storm will produce. Therefore, I recommend completing a solid big picture analysis to estimate the amount of liquid that the storm may produce, and estimate a snow-to-liquid ratio based boundary layer temperatures, as well as the conditions in the cloud that could affect snowflake type. Such an approach will allow you to evaluate various model snowfall products to see which ones might be the most reasonable in any given scenario. This approach will also get you in tune with other issues like precipitation type changes or melting at the surface that may complicate the forecast and cause model snow forecast products to perform worse than usual.

Speaking of precipitation type issues, I want to wrap up the lesson with a few thoughts about tricky forecast situations involving frozen precipitation that isn't snow. Read on.

Explore Further...

Dynamic approaches to estimating snow-to-liquid ratios vary in complexity. For example, some are based on boundary layer temperatures, some are based on air column thickness, and the Kuchera Method described on this page focuses on the maximum temperature in local air columns between the surface and 500 mb. Other approaches use more complex methods (none of which are perfect), but here's some background information on approaches you may encounter:

  • Cobb Method (opens in a new window): Primarily focuses on vertical motion, temperature, and ice crystal growth in its estimates, so it's a more complex approach than Kuchera; however, it requires high resolution and high quality forecasts of vertical motion and temperature, so it can be vulnerable to model errors in these predictions.
  • WeatherNerds Snow-to-Liquid Ratio (opens in a new window): A brief, readable description by Matt Onderlinde of how snowfall is derived for graphics on weathernerds.org (opens in a new window). This dynamic approach incorporates temperature, relative humidity, and vertical motion and scales upward or downward from a 10:1 ratio based on those factors and a compaction factor.

You'll likely find that most model websites include 10-to-1 snowfall graphics, but from there the offerings vary from site to site. Methods for estimating snow-to-liquid ratios can be developed and applied by anyone producing model graphics online, so other "proprietary" methods exist with methodologies that are not always clear. However, they typically attempt to account for the main factors outlined in this section.

sas405

Freezing Drizzle and Freezing Rain

Freezing Drizzle and Freezing Rain

Prioritize...

When you've finished this page, you should be able to identify freezing drizzle profiles on skew-T diagrams and discuss the limitations of freezing rain accumulation maps.

Read...

Near the end of METEO 101, you learned about precipitation types (opens in a new window) (login required) and precipitation type forecasting (opens in a new window) (login required). One of concepts you learned was that when temperatures throughout the entire troposphere are lower than the melting point of ice and when clouds are sufficiently cold enough so that the Bergeron-Findeisen process can take place, only snow can reach the ground.

There's an important caveat in the previous sentence: For snow to reach the ground, the clouds must be sufficiently cold enough so that the Bergeron-Findeisen process can take place. What happens when an entire air column is colder than the melting point of ice and precipitation falls, but clouds aren't cold enough for the Bergeron-Findeisen process to take place?

When is a "snow sounding" not really a snow sounding?

Recall from your previous studies that 0 degrees Celsius is the melting point of ice, but not always the freezing point of water. Pure water freezes at 0 degrees Celsius, but water in the real atmosphere has impurities that can cause it to freeze at temperatures less than 0 degrees Celsius. Therefore, when clouds form at temperatures less than 0 degrees Celsius, they aren't automatically composed of ice crystals. In fact, between 0 degrees Celsius and about -6 degrees Celsius, clouds consist mostly of supercooled water droplets.

The temperature at which ice crystals begin to grow via the Bergeron-Findesen process (opens in a new window) (login required) can vary a bit based on the types of freezing nuclei present (freezing nuclei are like cloud condensation nuclei, but for ice crystals), but the Bergeron-Findeisen process typically takes off and ice crystal growth dominates at temperatures less than -10 degrees Celsius. So, when clouds form at temperatures higher than about -10 degrees Celsius, relatively few, if any, ice crystals (snowflakes) grow in the cloud. With that in mind, check out the observed sounding from North Platte, Nebraska from 00Z on March 8, 2023 below. I've highlighted the 0 degree Celsius isotherm for emphasis.

The 00Z sounding from North Platte, NE on March 8, 2023 showed a temperature profile entirely below 0 degrees Celsius.

The 00Z sounding at North Platte, Nebraska from March 8, 2023 shows a temperature profile that's entirely below 0 degrees Celsius.
Credit: University of Wyoming

Clearly, the entire temperature profile is less than 0 degrees Celsius, but note that the layer in the lower troposphere near saturation (high relative humidity -- indicative of clouds) also has temperatures higher than -10 degrees Celsius (check out the annotated sounding (opens in a new window)). If precipitation fell in this environment, what type would it be? It probably wouldn't be snow because at these temperatures, the majority of the clouds consisted of supercooled liquid droplets.

Precipitation production in such clouds occurs via a "warm-cloud process" involving the collision and coalescence of small droplets that grow into drizzle drops (smaller than raindrops -- typically having diameters of 0.2 to 0.5 millimeters). So, any precipitation particles in this environment would be drizzle, but since surface temperatures were less than 0 degrees Celsius, drizzle drops would freeze on contact, meaning the end result would be freezing drizzle!

Indeed, the METARS between 02Z and 03Z on March 8 from Broken Bow, Nebraska (opens in a new window) (about 60 miles from North Platte) reported light freezing rain, although in reality it was freezing drizzle (automated sensors at this station did not have the capability of discerning drizzle from rain). Also note that even though the hourly measurable precipitation total was only a trace (typical of drizzle situations), the one-hour ice accretion reported at 03Z was 0.02 inches.

Freezing drizzle can also occur when the clouds are even warmer (temperatures higher than 0 degrees Celsius in part or all of the cloud layer) as long as temperatures at the surface are less than 0 degrees Celsius, as in this example from Dallas-Fort Worth, Texas from February 11, 2021 (opens in a new window). As a forecaster, you should be on the look out for sneaky freezing drizzle set-ups, as they sometimes don't get the public's attention (compared to heavier winter precipitation events), and can lead to dangerous and slippery travel even though the amount of ice accretion is typically quite small. To help you identify freezing drizzle situations, check out these forecast tips:

Forecast Tips

Be aware of the risk for freezing drizzle when you see:

  • Forecast soundings with relatively shallow saturated (or nearly saturated) layers in the lower troposphere (typically with much drier air above) indicative of low stratus clouds with temperatures above -10 degrees Celsius
  • Surface temperatures less than 0 degrees Celsius (so drizzle drops can freeze on contact)
  • A pattern that favors weak upward motion in the lower troposphere (often via upslope flow, low-level convergence, or overrunning) to favor precipitation production via warm-cloud processes.

So, keep in mind that all soundings in precipitation environments that are completely below 0 degrees Celsius through the troposphere don't always produce snow! If ice crystals never form in the cloud to begin with, freezing drizzle will be the result, so keep an eye out for these sneaky (and potentially dangerous) events! When the potential for ice build up from freezing rain is greater, forecasters are turning to an increasing number of freezing rain forecast products, so I'll wrap up this section with a few comments about them.

Freezing Rain Accumulation Maps

An increasing number of model websites are including freezing rain accumulation maps in their array of available products. For example, check out the 108-hour GFS forecast for freezing rain accumulation valid at 00Z on January 2, 2021 (initialized at 12Z on December 28, 2020) below.

108-hour GFS forecast for freezing rain accumulation valid at 00Z on January 2, 2021.

The 108-hour GFS forecast for freezing rain accumulation valid at 00Z on January 2, 2021 (initialized at 12Z on December 28, 2020).
Credit: Pivotal Weather

These products are typically just a plot of the amount of liquid precipitation predicted to fall while the model thinks the precipitation type will be freezing rain. In other words, it shows a 1-to-1 ratio of freezing rain to freezing rain accumulation. But, if you're using these maps to make a forecast (or communicate forecast information) that's problematic for a couple of reasons:

  • The model's forecasts for amount of liquid precipitation and precipitation type may be erroneous.
  • When freezing rain falls, it typically does not freeze on contact instantly. Some liquid drips off of elevated surfaces (tree branches, power lines, etc.) or otherwise runs off before it can freeze.

Even if we assume that the model's QPF prediction and precipitation type forecasts are reasonably good (which often may not be true, especially over the course of several days as in the example above), the second point above is very important because it determines how much ice will accumulate on various surfaces (which impacts road conditions as well as the potential for power outages, tree damage, etc.). In other words, even with a perfect QPF and precipitation type forecast, maps like the one above will tend to overestimate ice accumulation.

Research based on a large sampling of freezing rain cases has shown a median ratio of 0.72-to-1 (0.72 inches of ice accumulates for every one inch of liquid that falls as freezing rain). But, the amount of freezing rain that actually accumulates can depend on a number of factors, including drop size, precipitation intensity, wind speed, and the temperature of the ground. How can these factors impact freezing rain accumulation?

  • Precipitation intensity and drop size: Freezing rain accumulation tends to decrease as precipitation intensity and drop size increase (because of greater runoff and larger drops can take longer to freeze)
  • Wind speed: Freezing rain accumulation tends to increase as wind speeds increase (faster winds may increase the number of drops that encounter an elevated surface like a tree branch or power line, and faster winds can lead to faster freezing via evaporational cooling)
  • Ground temperature: Freezing rain accumulation tends to increase with lower ground temperatures (drops freeze more quickly)

So, a scenario with lighter precipitation, smaller drops, breezy conditions, and a ground temperature well below 32 degrees Fahrenheit (0 degrees Celsius) tends to maximize freezing rain accumulation. A scenario with heavier precipitation, larger drops, calm or very light winds, and a ground temperature barely below 32 degrees Fahrenheit tends to minimize it. To attempt to account for variables that can impact freezing rain accumulation, meteorologists have developed a statistical method called the Freezing Rain Accumulation Model (FRAM), which can be applied to model output for improved forecasts. The size-by-side comparison of the model prog above with a version based on the GFS with FRAM applied shows that the differences can be quite striking!

(Left) 108-hour GFS forecast for freezing rain accumulation valid at 00Z on January 2, 2021. (Right) Corresponding GFS forecast with the FRAM applied.

(Left) The 108-hour GFS forecast for freezing rain accumulation valid at 00Z on January 2, 2021 (initialized at 12Z on December 28, 2020). (Right) The corresponding GFS forecast with the FRAM applied, resulting in lower predicted ice accumulations in most areas.
Credit: Pivotal Weather

By applying the FRAM to the same GFS forecast, predicted ice totals were nearly 50 percent lower (opens in a new window) in some of the areas predicted to get the most ice, so applying the FRAM made a big difference! Is the FRAM perfect? Definitely not! It's merely a statistical technique based on more than 1,000 hours of freezing rain observations. It can't perfectly account for all factors that impact freezing rain accumulation, and FRAM progs may still be based on flawed QPF or precipitation type forecasts. So, you should never use FRAM-based progs without conducting a good big picture analysis to estimate the amount of precipitation that might fall as ice to begin with. But, freezing rain model maps based on the FRAM will tend to give more realistic depictions of how much ice may accumulate than ice accumulation maps based on "raw" QPF falling as ice.

That wraps up our look at precipitation forecasting. If you're looking for more information on the FRAM or sources where you can find FRAM-based model forecasts, check out the Explore Further section below.

Explore Further...

If you're looking for more information on the Freezing Rain Accumulation Model (FRAM), I recommend checking out this insightful PowerPoint presentation (opens in a new window) and the seminal paper (opens in a new window) on the topic.

Key Data Resources

While an increasing number of model websites are including raw freezing rain accumulation maps, a smaller number include FRAM-based forecast maps (and some sites require a subscription to access FRAM-based maps). The sites below, however, offer FRAM-based ice accumulation maps for free:

Freezing rain forecasts from the NBM also use a FRAM-type approach, as do freezing rain forecast products from the Weather Prediction Center (opens in a new window).

sas405

Lesson 7. Introduction to Medium- and Long-Range Forecasting

Lesson 7. Introduction to Medium- and Long-Range Forecasting

Motivate...

Although our focus in METEO 410 (and throughout much of the certificate program) has been on short-range forecasting, I think it's important to cover some basic topics relating medium- and long-range forecasting before you finish the program. This lesson, however, isn't meant to be an exhaustive treatment. In fact, at Penn State, the Department of Meteorology used to offer an entire semester-long course on just medium- and long-range forecasting! So, we can't thoroughly cover these topics in just one lesson. Instead, I'm just going to scratch the surface by exposing you to some basic tools and strategies for forecasting out beyond two days, and provide some additional resources for those who want to do more investigating on their own.

First, let's get some definitions out of the way. Earlier this semester, I defined medium-range forecasting as covering the three to seven day time period. The AMS Glossary, for example, still uses this traditional definition (opens in a new window), but, no single agreed-upon definition for medium range forecasting exists. With a greater availability of computer guidance that makes predictions more than a week into the future, the "working definition" of medium-range forecasting is more frequently including 10-day or even two-week forecasts. So, whenever you hear "medium-range forecast" just realize that it could refer to any forecast from three days to about two weeks into the future.

What's long-range forecasting, then? Well, it depends on where one decides to cut-off their definition of medium-range forecasting. The AMS Glossary definition of long-range forecasting (opens in a new window) states that any forecast for a period more than seven days into the future is "long range." But, again, the definition is murky, depending on how one defines the medium-range. For all practical purposes, long-range forecasts are those for any period in the future beyond the medium range. That's a pretty broad definition! In this lesson, we'll consider long-range forecasts to be those from about two weeks into the future to those covering a single season (a "summer" -- June, July, August forecast, for example).

A tree obscured by mist

It's impossible to see the details in the "misty, distant future," so good forecasters adjust their forecasting goals accordingly when making medium- and long-range forecasts.

You're already familiar with the many uncertainties that play a role in making short-range forecasts, and in the medium and long ranges, those uncertainties are greater. A colleague here in the Department of Meteorology likes to refer to the medium and long range as "the misty, distant future" because the details are always unclear. So, are medium- and long-range forecasts even valuable, or are they a big waste of time and effort?

The truth of the matter is that medium- and long-range forecasts can be valuable if forecasters don't lock themselves into the same types of forecasts they make in the short range. When making predictions for the "misty, distant future" we still must start with the same basic questions that we used to determine our short-range forecast goals:

  1. What do I want to know?
  2. What time period am I interested in?
  3. What geographic region am I interested in?

However, the "realistic" answers to these questions are typically different in the medium and long ranges compared to the short range. In the short range, we might want to forecast a specific high temperature for tomorrow at a given city, and in many cases, we stand a pretty good chance of getting within a few degrees of what verifies. But, a forecast for a specific high temperature on a day two weeks into the future stands a much lower chance of getting within a few degrees. So, it behooves us to rethink our forecast goals, as well as the tools and strategies that we use when making medium- or long-range forecasts. Not surprisingly, thinking probabilistically, focusing on the "big picture," and knowing (and comparing to) climatology are all critical parts of the process.

Let me be clear -- making medium- and long-range forecasts is a huge challenge, but we can create valuable forecasts if we approach the task with realistic expectations and adjust the format of our forecasts to appropriately convey uncertainty. Indeed, forecast format is a critical piece of creating valuable medium- and long-range forecasts. In this lesson, we'll discuss various forecast formats, as well as the computer guidance tools and other methods (such as using teleconnections (opens in a new window)) that forecasters use to get a "beat" on the evolution of the atmosphere weeks or months into the future. Let's get started!

sas405

Formatting the Forecast

Formatting the Forecast

Prioritize...

On this page, focus on the differences in medium- and long-range forecast formats compared to commonly published "tombstone" and other point forecasts.

Read...

The first step of creating a valuable medium- or long-range forecast is to consider what forecast format is most appropriate. You probably already realize that the most commonly consumed type of forecast (a deterministic point forecast) is horribly suited to medium- and long-range forecasting. But, that doesn't stop many purveyors of weather information from clinging to that very format! Often, these forecasts take on the ”tombstone“ display of a seven or ten-day forecast (opens in a new window) on TV or some similar format online.

Why are deterministic point forecasts so problematic for medium- and long-range forecasting? As you know, creating reasonably accurate deterministic point forecasts is challenging, even in the short term. Just think of all the things that can cause a forecast to go awry, like a lone isolated thunderstorm passing over your station, that are difficult to predict in the short range. Even with a solid grasp on the big picture, the mesoscale details can be hard to pin down.

In the medium or long range, we have absolutely no skill in trying to predict mesoscale details, which makes creating accurate deterministic point forecasts next to impossible. Even in cases where we do have a handle on the big-picture weather features that will play an important role in a medium-range forecast, timing them accurately presents more problems for the deterministic point forecast format. Say that you're confident a strong cold front will pass your city next week, so you create your forecast accordingly. Merely being off on the timing of the front by one day (which happens regularly on these time scales) will throw your medium-range point forecast into disarray. Your predictions for one or more days will likely be very, very wrong!

Just how poorly do deterministic medium- and long-range forecasts for single points actually verify? For several years now, freshman meteorology students at Penn State have verified deterministic point forecasts for selected cities from The Weather Channel and AccuWeather (the two largest private forecasting outlets in the U.S.), and the results are remarkably similar every year. On average, after about a seven to ten days at most cities, deterministic point forecasts perform worse than climatology. That's right — beyond seven to ten days, you'd be better off simply looking up a city's climatology for the date and using that as your forecast. If you want to see the actual verification data and some additional discussion of this topic, check out the Explore Further section below.

Broadening our alternatives

When forecasts lose to climatology, what's the point of making the forecast at all? Surely we must have some alternatives that allow us to create more valuable medium- and long-range forecasts. Since forecast details present serious problems in the medium and long range, I'd advocate using a less detailed approach in our forecast formats. We can do so by losing the deterministic point forecast format altogether and instead forecast for larger areas (entire states, regions, or countries, for example) with ranges, probabilities, and risk levels. The Department of Meteorology utilizes such an approach in their medium range forecasting segments on the show Weather World (opens in a new window).

Below, you'll see two forecast graphics. On the left is a forecast graphic for the "Extended Forecast" segment (covering three to seven days into the future — a "traditional" medium-range forecast). On the right is a forecast graphic for the "12-Day Trends" segment (which covers the period 7 - 12 days into the future).

Weather World forecast graphics showing alternative forecasting formats

(Left) A sample Extended Forecast graphic for Pennsylvania featured on Weather World. Note that the graphic features statewide ranges of maximum temperatures, qualitative risk of precipitation, and various text descriptions to add detail. (Right) A sample 12-Day Trends forecast graphic for Pennsylvania. This graphic shows probabilities of above and below average temperatures,and areas of heavy precipitation.
Credit: Penn State Department of Meteorology

These graphics might seem unconventional, but they convey loads of useful information, and aren't too difficult to decipher. Let's start with the Extended Forecast graphic on the left above (larger version of the Extended Forecast graphic (opens in a new window)). The first thing to note is that the graphic includes statewide ranges for daily high temperatures and qualitative risk of precipitation with various text descriptions to add detail (it does not include forecasts for low temperature or exact probabilities of precipitation). For Monday's forecast, as an example, we can tell that high temperatures were predicted to range from the mid-60s to upper 70s across Pennsylvania, with decreasing probabilities of precipitation later in the day. By the end of the forecast period (Thursday and Friday), warmer conditions were predicted to prevail, with afternoon highs ranging from about 70 degrees to near 80 degrees. Thursday and Friday afternoons each would bring some increased risk of precipitation (note the peaks in the trace showing precipitation risk).

Note that the Extended Forecast graphic also makes use of brief text descriptions to help add details that we can predict in the medium range. For example, forecasters were confident that Tuesday would have low humidity and a cool overnight period. Thursday and Friday were predicted to have increasing humidity, and the afternoon spotty showers would mainly be in western Pennsylvania. So, while avoiding the deterministic point forecast format, this three to seven day forecast actually gives quite a bit of useful information about the weather in Pennsylvania—information that focuses on variables and aspects of the weather that are more likely to be predicted reasonably well.

Beyond seven days, Weather World adopts an approach strictly based on probabilities and risk in its 12-day Trends segment (larger version of 12-Day Trends graphic (opens in a new window)). This particular forecast was created on August 23, 2013 and covered the period from August 30 through September 4 (7 - 12 days from when the forecast was issued). Note that this graphic displays qualitative probabilities of above normal temperatures, below normal temperatures, and areas of heavy precipitation across Pennsylvania, so it avoids deterministic values entirely. Probabilities are color-coded in terms of "risk" (low risk = green, medium risk = yellow, high risk = dark orange, very high risk = dark red). 

From this graphic, we can easily see that forecasters predicted the first part of the period to be warmer than average, with the "peak" of the heat to last from Saturday through Monday. By the end of the period, however, below average temperatures were more likely. The only days that forecasters believed would have an increased risk of heavy precipitation were Monday and Tuesday, because of thunderstorms across the state. The graphic also includes some text to summarize the main themes of the forecast — in this case, that it would be warm and humid for the end of meteorological summer, but by the end of the period a cooler air mass from Canada would arrive.

Both of these graphics convey uncertainty much more realistically than is possible with a deterministic point forecast. They make no bold claims about specifics, which have little chance of verifying correctly, while still providing useful forecast information. When forecasters have more confidence in a particular aspect of the forecast, they can display greater levels of risk, and when they have less confidence, they can adjust appropriately by not going "all in." If you're wondering how this forecast panned out, the daily temperature anomalies from August 31 through September 2 (opens in a new window) (when forecasters predicted the greatest likelihood of hotter than average weather) show that temperatures in Pennsylvania ran 2ºC to 4ºC above average for those days (roughly 4ºF to 7ºF above average). And, indeed, cooler weather followed, as the map of temperature anomalies for September 4 (opens in a new window) shows. Not too shabby for a 12-day forecast!

Other alternatives

The two examples above from Weather World certainly aren't the only ways that medium- and long-range forecasts can be tackled, but they do show that with some creativity, meteorologists can create useful medium- and long-range forecast products. A more commonly used format is an areal forecast with probable departures from climatology plotted. That's the basic approach that the Climate Prediction Center (opens in a new window) (CPC) uses in its forecasts, which range from six days out to three months. Let's take a look at a couple of their products.

A 6-10 day temperature forecast from the Climate Prediction Center

The 6-10 day forecast issued September 5, 2013 and valid September 11 - 15, 2013 shows that warmer than average weather was likely to dominate the United States. The Western U.S. in particular had a high probability of above average temperatures.
Credit: Climate Prediction Center

CPC forecast products are probabilistic and are based on the idea that three possible forecast outcomes exist — average, above average, or below average temperatures (or precipitation). If these were random events, they would all have equal probabilities (a 1 in 3 chance exists for each possible outcome). But, of course, particular weather patterns make some outcomes more likely than others. If a particular weather pattern means a 40% chance of above average temperatures exists, for example, that tips the scales toward it being the most likely outcome, since it reduces the chances of the other outcomes. 

In the 6-10 day forecast map above, issued on September 5, 2013, clearly the abundance of tan, orange, and red shadings suggests that above average temperatures were the most likely outcome for much of the country (climatological average temperatures are contoured with dashed lines for reference). But, note that the shading starts at probabilities greater than 33%, suggesting that for many areas (especially in the eastern half of the U.S.), above average temperatures are far from a sure thing (the probabilities of normal or below average temperatures may only be slightly lower). The interior Northwest, however, had over a 70% chance of above average temperatures, so the likelihood of warmer than average weather there was high. CPC's graphics for precipitation take on a similar feel (here's the corresponding 6-10 day precipitation forecast (opens in a new window) for reference).

A Nittany Lion wind vane

The three-month forecast for temperature, valid during September, October, and November, 2013, issued on August 15, 2013.
Credit: Climate Prediction Center

On longer time scales, CPC provides similar products. For example, check out their three-month forecast for temperature (on the right) and three-month forecast for precipitation (opens in a new window), valid for the months of September, October, and November 2013, issued on August 15, 2013. The labels "EC" might seem curious to you, but they merely stand for "Equal Chances," meaning that average, above average, and below average conditions each have a 1 in 3 chance of occurring. You'll rarely see very high probabilities on these monthly forecasts because the uncertainties are so great. On the time scale of three months, for example, forecasters can't be extremely confident that conditions will be average, above average, or below average. Think about it: even if most of the days during a three-month period are cool compared to climatology, one very intense heatwave that lasts a week or two could entirely cancel out the cool weather so that the three-month average was right at normal! And, of course, nailing down the details of individual warm and cool spells months in advance is impossible.

The Bottom Line

The most common mistake that forecasters make in creating medium- and long-range forecasts is clinging to the same formats often used for short-range forecasts (namely, deterministic point forecasts). But, with some effort and creativity in conveying varying levels of confidence in a forecast, and creating forecasts based on probabilities for larger areas, forecasters can still provide helpful medium- and long-range forecast products. This approach can be summarized by some sage old forecasting advice: emphasize what you know (probabilities, levels of confidence, departures from climatology), not what you don't know (specific temperatures, precipitation amounts, exact timing, etc.).

Now, the question is, "what tools and strategies do forecasters use to actually create valuable medium- and long-range forecasts?" For the rest of this introductory lesson, we'll shift gears to answer that very question. We'll start by looking at computer guidance that medium- and long-range forecasters use. Read on.

Explore Further...

Above, I summarized the verification of medium- and long-range deterministic point forecasts from The Weather Channel and AccuWeather. For those who are curious, here's some actual verification data (opens in a new window) for maximum temperature at selected cities from August through November of 2009. In subsequent years, the results obtained by freshman meteorology students have been remarkably similar — such forecasts typically are worse than climatology after seven to ten days. Since deterministic point forecasts inappropriately convey the exact same certainty about medium and long-range forecasts as they do short-range forecasts, it's pretty clear that they send misleading messages.

Dr. Jon Nese, the professor whose students conduct the analysis each year, wrote up a summary of the results in the Washington Post's Capital Weather Gang blog, which stirred much discussion in the meteorological community. I've included the link to his write-up below, along with a response by Evan Myers, the Chief Operating Officer of AccuWeather. Both pieces make for some interesting reading, and give insights into how such forecasts are viewed in the academic and private sectors of meteorology:

A thought to ponder: If forecasts are horribly inaccurate, yet the public continues to consume them (click on them), is the forecast product ultimately successful? Certainly meteorologists in the academic sector and private sector have different answers!

sas405

Medium and Long-Range Computer Guidance

Medium and Long-Range Computer Guidance

Prioritize...

In this section, focus on the models that forecasters use in creating medium- and long-range forecasts, as well as what fields are helpful in making such forecasts.

Read...

As you already learned this semester, dynamic and AI computer models don't predict important forecast variables like surface temperature, dew point, liquid precipitation etc., very well, even for short-range forecasting. On average, they are better at predicting fields like 500-mb heights, which don't require parameterizations to approximate complicated radiative fluxes, turbulence, etc. We can use this to our advantage in using computer guidance in making medium-range forecasts. Notice I didn't include "long-range" forecasts there, too. That's because computer guidance for long-range forecasting (on the scale of months or longer, anyway) is a bit more limited and has even less skill. We'll start off by briefly investigating computer models used by forecasters for medium-range forecasts, and then touch on long-range guidance.

Going Global

Medium-range forecast guidance falls almost exclusively to the global models, like the GFS, ECMWF, and their AI counterparts. Why? Well, in order to look further into the future, we have to be able to look farther away in space, too. We have to be able to track features that might literally be "half a world away." Models like the RRFS, which only have a North American domain, or other mesoscale models, which have smaller domains, don't allow us to do that (not to mention the fact that their forecasts generally only go out to the early part of the medium range, at best). Global models, on the other hand, allow us to track important weather features across the entire globe, and their forecasts go out at least ten days or so. 

How good are the global models?

Just how useful are the predictions from these global models? It depends on what forecast variables you're interested in. A 10-day forecast of 2-meter temperatures isn't likely to be very good. The predictions for other variables, however, tend to be more successful (although still imperfect). One way that forecasters assess the accuracy of model forecasts is to calculate the anomaly correlation coefficient (opens in a new window), which essentially compares the predicted anomaly (based on climatology) to the observed anomaly. Higher correlation coefficients (the maximum is 1) correspond to better forecasts. For reference, an anomaly correlation coefficient of 0.5 represents an error equal to a forecast of the climatological average. Anomaly correlation coefficients of 0.6 correspond to forecasts that have some skill in predicting the largest-scale weather features, while coefficients of 0.8 represent some skill in predicting typical synoptic features.

Chart displaying anomaly correlation coefficient for the global models from August 11 - September 10, 2013.

Anomaly Correlation Coefficients for 120-hour forecasts from the 00Z runs of the major global models verified between August 11 and September 10, 2013 indicate that the ECMWF had the most accurate forecasts of 500-mb height anomalies on average during this period, but most models demonstrated reasonable skill at predicting typical synoptic-scale features (anomaly correlation coefficients greater than 0.8).
Credit: NCEP Environmental Modeling Center.

The chart above shows anomaly correlation coefficients for 120-hour forecasts of 500-mb height anomalies from the 00Z runs of the major global models, verified from August 11 through September 10, 2013, and it allows us to draw some important conclusions. During this time period, the ECMWF's forecasts were the best for this variable on average (anomaly correlation coefficient = 0.877), followed by the model from the UK Met Office (opens in a new window) (0.852), and the GFS (0.843). Other models with lower scores were the CMC, the NCMWRF (a global model run by the Ministry of Earth Sciences in India), the NAVGEM (denoted by "FNO" on this chart), and the "Climate Forecast System" (CFSR on this chart). More on the Climate Forecast System in just a bit.

The ECMWF had the best average anomaly correlation coefficient here, so is it the best global model, case closed? Well, for predicted 500-mb height anomalies during this period, on average, it was. And, in general, the ECMWF often scores best, on average, for 500-mb height anomaly correlation coefficient. But, the key phrase is on average, meaning it's not the best every single day. If you look closely at the chart above, on several individual days, its forecasts were third or fourth best in the group.

Indeed, the mantra that you should avoid choosing a model of the day applies to medium-range forecasting, too. Clearly, on any given day, any of the global models could outperform the others. Also note that, around August 30, the scores of all the models took a sharp dive, and a few of the models had scores near 0.6 (so, they were barely skillful in predicting even the largest weather features at five days), but in general, at five days, the models have some skill in predicting typical synoptic features most of the time. I should note that these scores were taken from the warm season, when anomaly correlation coefficients tend to be lower than they are during the cold season (it's not uncommon for some models to score above 0.90 during the cold season for five-day forecasts of 500-mb height anomalies).

How about further into the future at 10 days? Below are the anomaly correlation coefficients for the ECMWF and GFS for 240-hour forecasts of 500-mb height anomalies over the same time period. Two important things stand out. First, the GFS performed better on average during this period for 10-day forecasts (although that's not always the case), and secondly, the scores are much lower than they were for five-day forecasts. Indeed, during this period, on average, the models did not have skill compared to climatology in predicting 500-mb height anomalies.  It's fairly common for medium-range computer guidance to lose all skill compared to climatology by about 10 days, especially in the warm season. Occasionally though, they can have some level of skill out to two weeks (more likely in the cold season).

Chart showing anomaly correlation coefficients for 10-day 500-mb height anomaly forecasts from the GFS and ECMWF.

Anomaly Correlation Coefficients for 240-hour forecasts from the 00Z runs of the GFS and ECMWF models from August 11 through September 10, 2013 indicate that the GFS had the most accurate forecasts of 500-mb height anomalies on average during this period, but neither model demonstrated reasonable skill at predicting even the largest-scale features, on average (anomaly correlation coefficients less than 0.6).
Credit: NCEP Environmental Modeling Center.

Keep in mind that 500-mb heights (and anomalies) are one of the better forecasted variables. So, if global models lose their skill around ten days for the "easier"-to-predict variables, one can imagine how well other variables (2-m temperatures, 10-m winds, etc.) fare. Ultimately, forecasters aren't (or shouldn't be) focusing on very specific details when evaluating computer guidance for medium-range forecasting.

What do forecasters look for in the models?

When looking at medium-range computer guidance from the global models, what are forecasters looking for? Largely, forecasters are looking to get a handle on the big picture weather pattern because, as we discussed above, the global models stand a better chance of simulating larger-scale features relatively well than smaller-scale ones, at least out to 7-10 days. After that, the prospects of getting a skillful forecast for even large-scale features are remote. In general, forecasters are looking for:

  • the location and strength of 500-mb troughs and ridges, which helps them to identify large areas of relatively cool and warm air (respectively) compared to climatology, as well as any instigators of cyclogenesis (namely, shortwave troughs).
  • the corresponding MSL pressure pattern (location and strength of surface highs and lows, troughs, etc.) -- connecting it to the 500-mb pattern whenever possible. Keep in mind that weak 500-mb shortwaves may not generate well-developed surface cyclones, and instead may only create surface troughs (no closed isobars).
  • air mass boundaries (cold and warm fronts) using 1000-500 mb thickness (areas of tight packing indicate large temperature gradients) and 850-mb temperatures.
  • the general "nature" of the 500-mb flow. Is it fast and largely west-east (no major troughs and ridges)? If so, that favors numerous weak, fast-moving shortwaves which can be incredibly difficult to time and place accurately. The weather in such a pattern tends to be changeable, and confidence in timing the changes is low. If the pattern is characterized by large troughs and ridges, confidence in timing is higher because weather systems tend to be stronger and slower-moving.

For some online resources where you can access medium-range computer guidance and verification stats, check out the Explore Further section below.

Long-Range Computer Guidance

Before we move on, I want to quickly touch on computer guidance that meteorologists use for longer-range predictions (weekly, monthly, and seasonal predictions). In the long-range, forecasters typically aren't interested in the forecast solutions from the models at a single specific time (and with good reason, given the stats above about model skill in the medium range). Instead, they are more interested in departures from climatology averaged over weeks, months, or seasons.

The Climate Forecast System (opens in a new window) (CFS) is NCEP's flagship long-range model. The CFS is run four times per day, with forecasts going out as long as nine months, and its products serve as one puzzle piece that CPC uses when creating its monthly and seasonal forecasts. It provides a wide array of products, but the most commonly used ones are monthly and three-month anomalies of fields like 500-mb heights, 850-mb temperatures, 2-m temperatures, precipitation, SSTs, etc. ECMWF also produces some publicly available long-range computer guidance (predicted monthly anomalies), as well as additional products that they offer only to paying subscribers.

I won't get into the details of such long-range guidance or its verification, but instead encourage you to do your own exploring, if you wish. I've included some links below in the Explore Further section to get you started in accessing CFS forecasts, verification, and model documentation, along with publicly-available ECMWF products.

Computer models aren't the only tools that forecasters have at their disposal for medium- and long-range forecasts, however. Large-scale atmospheric patterns can also provide helpful clues by way of teleconnections. We'll explore up next. Keep reading!

Explore Further...

Computer Models for Medium-Range Forecasts

Computer Models for Long-Range Forecasts

For those interested in learning the highly-technical details of the CFS, check out the page on CFS Documentation (opens in a new window).

sas405

Teleconnections

Teleconnections

Prioritize...

When you've finished this page, you should be able to discuss the basic definition of teleconnections and the example of how ENSO can impact weather across the globe through teleconnections.

Read...

After reading the last few sections, I hope you've gotten the message that forecasters have to use computer guidance for medium and long-range forecasts very carefully. On average, the models do have some skill in predicting the largest-scale weather systems out to about seven to ten days into the future (and occasionally, a few days longer). But, the uncertainties grow in time, and especially beyond about ten days, the models aren't much help. So, what's a forecaster to do? What other tools do we have to improve on the models' questionable depictions and get an edge in making their medium and long-range forecasts?

Since the "details" of prevailing weather patterns are not predictable in the medium- and long-range periods, forecasters must look for other, more predictable "atmospheric rhythms" called teleconnections. Most generally, the AMS Glossary defines a teleconnection as a "linkage between weather changes occurring in widely separated regions of the globe." More specifically, teleconnections have a few more attributes that may better help you see why they can be helpful in forecasting. One of the more common definitions of teleconnections comes from a 1981 paper by John Wallace and David Gutzler (for the full paper, check out the Explore Further section at the end of the page). They identified some additional characteristics of teleconnections, namely that they:

  • provide evidence for the transitory behavior of planetary-scale waves,
  • fluctuate with time scales on the order of a few weeks to many weeks,
  • suggest preferred regions of standing wave "nodes" and "antinodes."

In other words, teleconnections give helpful clues about the nature of the longwave pattern across the globe, and they tend to change reliably on the order of weeks. What does "suggest preferred regions of standing wave 'nodes' and 'antinodes' mean?" Well, first, in this context, a node refers to a correlation of similar features, while "antinode" refers to a correlation of dissimilar features. So, for example, if we can identify a pattern that favors a prevailing trough over a particular region, that has consequences for waves at other specific locations. Based on the position of the trough, other regions will exist where troughs are similarly favored (nodes) and where ridges will exist (antinodes).

In general, teleconnections provide forecasters with tools to identify likely scenarios for the large-scale flow pattern on the order of weeks (or even months), which can be very helpful in a sea of uncertain medium- and long-range model guidance. Many atmospheric teleconnections exist (far too many to cover in this brief lesson), and most are referred to by acronyms (like AO, NAO, PNA, MJO, ENSO, etc.), so it can be a bit challenging to keep them all straight. In this lesson, we're only going to cover a handful of the most prominent teleconnections, and we'll start with a quick review of one that you're already familiar with -- The El Niño Southern Oscillation (ENSO).

The El Niño Southern Oscillation (ENSO)

Photograph of ripples on a pond

ENSO is like a rock thrown into an atmospheric "pond." It makes its own waves, but other factors make waves, too, creating a complex pattern of ripples that can't be assigned back to any single "rock" thrown into the pond.
Credit: Wave Fronts Triggered By Falling Raindrops / CC BY 2.0 (opens in a new window)

The El Niño Southern Oscillation, as you may recall, refers to sea-surface temperature anomalies in the central and eastern equatorial Pacific. So, teleconnections based on ENSO really have their roots based in oceanic cycles (as do some other teleconnections). You may recall that I likened the impacts from ENSO to throwing "rocks" into an atmospheric "pond." The ENSO "rock" makes its own waves, but other factors ("rocks") in the atmosphere and ocean make waves, too, so it's impossible to say that ENSO is solely responsible for any single weather event. If you want to review the details of ENSO, I recommend checking out Lesson 6 in METEO 241 (opens in a new window) (login required), but I'll briefly recap the gist of how the teleconnection works here.

You may recall that the Climate Prediction Center quantifies the strength of ENSO with the Southern Oscillation Index (SOI), which is based on the difference in surface pressure anomalies at Tahiti (French Polynesia) and Darwin, Australia. Negative SOI values correspond to the El Niño phase, while positive SOI values correspond to the La Niña phase. We'll see this general approach (quantifying oscillations with mathematical indices) with other major oscillations, too, because these calculations make tracking the phase of the oscillations relatively simple.

In the El Niño (warm) phase of the oscillation, additional warmth from above average sea-surface temperatures in the central and eastern tropical Pacific gradually translates into anomalously warm air columns over the region. The warm air columns lead to higher heights aloft, and an increased height gradient over the Pacific (check out the 200-mb height anomalies (opens in a new window) from winter 1997-1998 during one of the strongest El Niño phases on record). The increased height gradient helps create a stronger-than-average subtropical jet (check out the comparison between 200-mb wind climatology and 200-mb winds from January - March of 1998 (opens in a new window)).

The strengthened subtropical jet over the Pacific and southern tier of the United States alters patterns of cyclogenesis (thanks to more jet streaks passing by, etc.) and leads to more (and stronger) mid-latitude lows traversing the southern United States. The end result is that cool and wet anomalies are typical over the region during the winter. Research has also shown common anomalies in temperature and precipitation that develop in other regions, as shown in the images below.

Global anomalies associated with El Nino during typical winter and summer seasons.

Typical global anomaly patterns associated with moderate to strong classic El Niño phases during meteorological winter (top) and meteorological summer (bottom).
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

During the cold phase of ENSO (La Niña), the subtropical jet is weakened (here's a look at the 200-mb winds from January -March 1999 (opens in a new window) for reference), which tends to lead to warm, dry weather over the southern United States. In general, the pattern of anomalies generated by a moderate to strong La Niña (opens in a new window) is quite different than those generated by an El Niño.

Of course, using the ENSO phase in a forecast isn't as simple as just looking at the above maps of seasonal anomalies. Not even close! First, it's important to remember that the phase of ENSO is not singularly responsible for any individual storm or weather event (it's not the only "rock" being thrown into the atmospheric "pond"). Second, teleconnections based off of ENSO are most useful for seasonal forecasts, since important fluctuations in ENSO tend to occur on the timescale of months. In other words, forecasters may weigh ENSO heavily in a seasonal forecast, but not as heavily in a medium-range forecast.

Also, the strength of the sea-surface temperature anomalies is very important. Strong El Niño and La Niña phases (based on large sea-surface temperature anomalies), are more likely to "leave their mark" on seasonal weather patterns. As we're about to explore, other important atmospheric teleconnections exist, and their influences may contrast sharply from those based on ENSO. When the ENSO phase is strong, it's more likely to "overwhelm" other atmospheric factors and lead to patterns of anomalies like those above. But, when the ENSO phase is weaker, other atmospheric factors can dominate, and seasonal anomalies may look nothing like the ones above.

Furthermore, the location of the maximum sea-surface temperature anomalies matters, too. So-called El Niño and La Niña "modokis," (opens in a new window) which have maximum sea-surface temperature anomalies in the central equatorial Pacific instead of the eastern Pacific, have very different teleconnections compared to "classic" ENSO events.

So, merely knowing the ENSO state by itself doesn't necessarily give clear-cut answers about seasonal weather patterns, and forecasters at the Climate Prediction Center (and other long-range forecasters) have to weigh other teleconnections in their forecasts. It's no wonder that forecasters at CPC stick with a probabilistic approach (which you may want to review in Lesson 6 of METEO 241), because large uncertainties can still exist even with a sound knowledge of the ENSO state and its impacts.

With our brief ENSO refresher out of the way, it's time to move on to another teleconnection with tropical roots (and another one you encountered, albeit very briefly, in METEO 241)--the Madden-Julian Oscillation. Read on.

Explore Further...

Key Data Resources

If you're looking for information on ENSO current conditions or forecasts, you'll be interested in these links:

Further Reading

If you're interested in reading more about various teleconnections and their history, here's Wallace and Gutzler's complete 1981 paper (opens in a new window) on the topic. It also has some additional technical details about some teleconnections which we're about to cover (and some we won't formally cover).

sas405

The Madden-Julian Oscillation: Part I

The Madden-Julian Oscillation: Part I

Prioritize...

When you've finished this page, you should be able to discuss what the Madden-Julian Oscillation is and the basics of how it's monitored.

Read...

On the previous page, we reviewed ENSO and some of its implications for seasonal forecasting. The teleconnections based on ENSO are perhaps the best known and most frequently used in seasonal forecasting. ENSO's fame is likely a result of its overall importance and the fact that its phases were first observed hundreds of years ago, and it's been systematically studied since the early 20th century. Indeed, Sir Gilbert Walker coined the term "Southern Oscillation" in the 1920s, so atmospheric scientists have been building the knowledge base on ENSO for a fairly long time.

Our next topic, however -- the Madden-Julian Oscillation (MJO) is a mere baby in comparison. It was first identified in the early 1970s by Roland Madden and Paul Julian. So, in the scheme of things, the understanding of the MJO and its teleconnections is in its infancy compared to ENSO. In the past couple of decades, however, researchers have come to view the MJO as the dominant driver of intraseasonal weather variations (on the time scale of weeks to a few months) in the tropics. So, the MJO is really important, and the atmospheric science community is only in its very early stages of understanding it. In fact, a universally accepted definition of the MJO along with a commonly accepted way to monitor it wasn't really established until 2004.

Much like ENSO, the origins of the MJO are complex. Indeed, the details of the MJO could easily comprise an entire lesson. But, for starters, on this page, I want to focus on a basic summary of what the MJO is, and give you a general sense for how it's tracked.

What is the MJO?

Formally, the MJO is an intraseasonal wave that occurs in the tropics, which manifests itself in the form of alternating areas of enhanced and suppressed rainfall (compared to climatology) that migrate eastward from the Indian Ocean to the central Pacific Ocean. This eastward migration occurs slowly, at an average pace of about five meters per second. The anomalously high rainfall is often first evident over the Indian Ocean as an area of large-scale deep convection that can be tracked eastward in time using various fields (satellite imagery, rainfall anomalies, outgoing longwave radiation anomalies, zonal wind anomalies, etc.).

You may recall that the large-scale cyclonic circulations that correspond to the long-lived clusters of thunderstorms associated with the MJO create low-level westerly winds near the Equator, which current theories suggest play a role in the onset of El Niño episodes (so, the MJO and ENSO are not completely independent of each other). The schematic below is highly simplified, but it will give you a visual appreciation of the MJO as areas of enhanced and suppressed convection moving eastward through the tropics.

Schematic showing the MJO as alternating areas of enhanced and suppressed rainfall migrating eastward through the tropics.

This schematic shows the large-scale nature and eastward propagation of the MJO in time. Sun icons represent areas of suppressed convection, while the "rainy" icons represent areas of enhanced convection. The period of time covered by the schematic is roughly 30-60 days.
Credit: NOAA and the Western Water Authority

Wheeler and Hendon (2004) defined eight phases of the MJO (numbered 1-8), with each phase being described by the general location of the main large-scale area of enhanced convection. On average, the MJO spends about six days in each phase, but large variability exists in the speed of the MJO's progression. On average, a complete cycle takes about 30-60 days, but can take as long as 90 days. The intensity of the MJO is also highly variable; at times the convective signal is fairly weak, while other times it's much stronger.

To better visualize the MJO, I highly recommend studying this animation of precipitation anomalies (opens in a new window), courtesy of Adrian Matthews (opens in a new window), University of East Anglia. This animation was created to illustrate a typical 48-day MJO cycle in Northern Hemispheric winter. The data set used to create the animation has a fairly low resolution, so it doesn't show the details of smaller-scale convection, but it's very useful for understanding the progression of the MJO. Blue shaded areas represent areas of enhanced rainfall rate (at least 1 mm/day more than climatological average), while orange shaded areas represent areas of anomalously low rainfall rate (at least 1 mm/day less than climatological average). Using this animation, it's easy to track the eastward progression of the MJO as it passes through each of the eight phases (note the phase tracker in the upper right corner).

At the beginning of the cycle (Phase 1), an area of enhanced convection leads to wet conditions (blue shading) over extreme eastern Africa and the western Indian Ocean. This area then gradually moves eastward across the Indian Ocean in Phases 2 and 3. When the MJO reaches Phase 4, the enhanced rainfall is located over what's referred to as "the maritime continent (opens in a new window)," a series of islands in Southeast Asia (opens in a new window) including Indonesia and the Philippines, among others. The region of enhanced rainfall continues across the maritime continent in Phase 5. In Phases 6, 7, and 8, the area of enhanced rainfall moves out over the western and central Pacific Ocean, where it fades.

As the main area of convection fades in the central Pacific, a new area of convection begins to blossom over eastern Africa and the western Indian Ocean, and the cycle begins again. Also note that as the area of enhanced rainfall moves out of the Indian Ocean, an area of suppressed rainfall (orange shading) occurs behind it, which trails it across the maritime continent into the Pacific.

With a very brief background on the MJO complete, let's turn our attention to more details of how forecasters keep track of the MJO and how they use it in medium-range forecasts on the next page.

sas405

The Madden-Julian Oscillation: Part II

The Madden-Julian Oscillation: Part II

Prioritize...

When you've finished this page, you should be able to track the MJO with a phase diagram, and discuss its basic impacts on mid-latitude and tropical weather.

Read...

Research has shown that, on average, predictions of the MJO phase have skill out to about two weeks or a bit longer, which makes it a useful source of teleconnections for medium-range forecasts. So, how do forecasters track the current state of the MJO? One of the most commonly used ways is with phase diagrams, which are divided into eight sections, one for each phase of the MJO. Points are plotted on the diagram based on the location and intensity of the MJO, which are determined by the "Real-Time Multivariate MJO (RMM) Index -- a mathematical index for analyzing cloud cover and wind field anomalies developed by Wheeler and Hendon in their 2004 research.

You can see an example phase diagram below, which tracked the MJO from August 18, 2013 through September 26, 2013. In general, the closer the plotted points are to the edges of the diagram, the stronger the MJO signal. When the index is in the circle at the center of the diagram, the MJO signal is weak (and difficult to track). From the diagram below, we can tell that the MJO started the period in Phase 7 (note the "Start" label on the diagram), which means the large-scale area of deep convection was over the Western Pacific Ocean. Within a few days, however, it moved into Phase 8, and by August 30th, it had moved to Phase 1 (the numbers along the trace indicate the day of the month with the green line representing August, and the blue line representing September). As it moved into Phase 2, however (large-scale convection over the Indian Ocean), the MJO became quite weak around September 9, as indicated by the trace moving through the center circle. It continued to be quite weak as it moved through Phases 3 and 4, but strengthened in Phase 5 around September 14. From there, the MJO continued through Phase 5 and into Phase 6 by September 26 (the last day for the plot). So, in 40 days, the MJO was not quite through one complete cycle.

MJO phase diagram covering the period from August 18, 2013 through September 26, 2013

This MJO Phase Diagram covers the time period from August 18, 2013 through September 26, 2013, and plots the index developed by Wheeler and Hendon (2004) to track the MJO. The MJO was in Phase 7 at the beginning of the period ("START" in the diagram), and moved through Phase 8 into Phase 1 by the end of August.
Credit: Australian Bureau of Meteorology

In general, the progression of the MJO can be tracked in an uneven counterclockwise fashion around these phase plots. Note that the trace is somewhat jagged because of the varying speed of its movement, and its varying intensity can cause some challenges in tracking it with a mathematical index. For more on these challenges, check out the discussion in the Explore Further section at the end of the page.

These plots also make for a convenient mechanism for viewing MJO forecasts. A common format is a plume diagram of sorts, like this forecast from the GEFS (opens in a new window) for September 29 through October 13, 2013. The plot shows the last 40 days of the MJO's evolution, along with the ensemble forecast for the next 15 days. The individual yellow lines are individual ensemble member forecasts, and the green line is the ensemble mean. The light gray shading marks the area within which 90% of the ensemble forecasts were plotted, and the darker gray shading marks the area within which 50% of the ensemble members were plotted. The take home message from this forecast from the GEFS is that the MJO would likely progress through Phases 6, 7, and 8 in the next 15 days, but had the potential to be fairly weak.

MJO Teleconnections

Being able to predict the MJO out to two weeks or so can really be beneficial for medium-range forecasting because of the MJO's many teleconnections. The list of teleconnections associated with the MJO is long, and is seemingly ever growing. For practical purposes, I'm going to discuss some main themes here and provide some links for you to do some more exploring on your own if you wish (see the Explore Further section at the end of the page).

First, like with ENSO, using the MJO as part of your forecast is not as simple as just memorizing a bunch of teleconnections. The MJO is just another "rock" being thrown into the atmospheric "pond." When it's intense, it can be a large rock, and can be a dominant signal, but when it's weaker, its waves are more likely to be overcome by those from other, bigger "rocks." Even though the signal of the MJO is most obvious over the Indian Ocean and the Pacific, it sends out waves that can impact weather throughout the global tropics, and even the mid-latitudes.

Tropical Connections

Forecasters have suspected for years that the MJO acts as a modulator on global tropical cyclone activity, and research over the past couple of decades bears that out. The tendency for brief periods of active tropical cyclone development to occur, followed by lulls, in a given basin can be connected to the MJO. In the Atlantic Basin, researchers have documented that tropical activity is enhanced when the MJO is in Phases 1 and 2 (large-scale convection over the Indian Ocean with suppressed convection over the Pacific). While the MJO is in Phases 1 and 2, the "ripples" from the MJO help cause reduced vertical wind shear and higher relative humidity in the middle troposphere over the Atlantic (both of which favor tropical cyclogenesis).

On the flip side, when the MJO is in Phases 6 and 7, and the large-scale convection is focused over the Western Pacific, tropical cyclone activity tends to be suppressed over the Atlantic. And, when tropical cyclones do form in the Atlantic during Phases 6 and 7, they tend to be far weaker and shorter-lived compared to Phases 1 and 2.

Mid-Latitude Connections

Forecasters have documented a number of teleconnections between the MJO and mid-latitude weather, as well. Naturally, when the MJO is more intense, the teleconnections are more robust. Indeed, researchers have identified a 40% increase in extreme rainfall events globally when the MJO is strong compared to when it is weak. Here, I'll briefly touch on some major teleconnections for weather in the United States, such as those shown in the composite temperature anomaly maps below.

Composite temperature anomalies associated with various MJO phases across the U.S.

Composite temperature anomalies across the U.S. associated with the eight phases of the MJO during meteorological winter (December - February).
Credit: Climate Prediction Center

Keep in mind that these composite temperature anomalies are only for meteorological winter (December - February). While spells of anomalously warm and cold weather, and outbreaks of heavy precipitation, can all be connected to the MJO's various phases, these teleconnections vary by season (they tend to be strongest in Northern Hemispheric winter) and with the intensity of the MJO. In other words, when the MJO is in Phase 1, for example, the majority of the United States is not always cool.

For reference, here are the corresponding precipitation anomalies (opens in a new window) associated with the eight MJO phases during meteorological winter. Note that these anomalies (for both temperature and precipitation) cannot be entirely attributed to the MJO phase, but some signals still emerge. Phases 1 and 8 (when the active area of MJO convection fades in the Central Pacific and a new area develops over East Africa and the Western Indian Ocean) correspond to anomalously cold weather in the eastern half of the U.S. during meteorological winter. That's because those phases tend to teleconnect to upper-level ridges over the Western U.S., which allows troughs to dig into the East. Of course, deep troughs in the eastern U.S. provide ample cold air for East Coast snow storms, if a favorable shortwave happens to come along. Meanwhile, phases 4, 5, and 6 (as MJO convection travels from the Maritime Continent into the Western Pacific), tend to correspond to warmer than average winter weather in much of the United States (but especially the eastern half).

The MJO can also be teleconnected to extreme precipitation events along the West Coast. Winter-time floods in the Northwest are most common when MJO convection is active over the Indian Ocean (phases 2 and 3), but severe flooding in the Northwest is actually most common in Phase 6 (when MJO convection is active over the Western Pacific). However, the rest of the West Coast tends to be relatively dry and warm in MJO phases 4 - 7.

These impacts on U.S. weather are really just the tip of the iceberg when it comes to MJO impacts. And, remember, the teleconnections tend to be more robust when the MJO is intense. When it is weak, discerning clear teleconnections can be difficult, or impossible, because the MJO signal is overwhelmed by other atmospheric features. But, with medium-range MJO forecasts at our disposal, we can use some of these teleconnections to help us identify favorable patterns for various types of weather (even though the details may be elusive). If you're interested in accessing MJO data and forecasts, as well as reading more about the teleconnections around the globe and in different seasons (since we focused on Northern Hemisphere winter in the U.S. here), check out the Explore Further section below.

Certainly, the MJO is physically complex, and its teleconnections are numerous (the MJO could easily take up an entire lesson or more). I've just scratched the surface here, but hopefully you now have some basic tools at your disposal to start observing the MJO and using it as a forecasting tool. Next, we'll move into some simpler teleconnections -- the North Atlantic Oscillation and the Arctic Oscillation. Keep reading!

Explore Further...

Key Data Resources

If you're looking for information on MJO current conditions or forecasts, you'll be interested in these links:

Further Reading

If you're interested in reading more about the MJO and its teleconnections, I recommend the following resources:

sas405

Oscillations with Arctic Connections

Oscillations with Arctic Connections

Prioritize...

Upon completion of this page, you should be able to discuss the implications of the positive and negative phases of the Arctic Oscillation and North Atlantic Oscillation on the westerly flow in the Northern Hemisphere mid-latitudes.

Read...

So far, the teleconnections we've covered have been based in the tropics, and have strong ties to the ocean. Now, we're going to shift gears to teleconnections based on atmospheric oscillations. Two really important oscillations for weather in North America (especially during the cold season) are the North Atlantic Oscillation (NAO) and the Arctic Oscillation (AO). These oscillations are somewhat related, in that they compare pressures at high latitudes to those in the mid-latitudes in an attempt to gauge the nature of mid-latitude westerlies. Let's explore, starting with the NAO.

The North Atlantic Oscillation (NAO)

The NAO is basically a comparison of pressure near Iceland and near the Azores Islands in the Atlantic Ocean. Numerous methods for tracking the NAO exist (some compare surface pressure anomalies in Iceland to those in the Azores, while others focus on comparisons of 500-mb height anomalies over those regions), but the Climate Prediction Center utilizes 500-mb heights for the basis of calculating their NAO index. Like ENSO, the NAO can only be positive, negative, or neutral. But, also like other oscillations, the strength of the NAO signal varies, and when it is weak, it can easily be overwhelmed by other events going on in the atmosphere. Still, to a large degree, variations in the NAO explain much of the winter-time variability in Eastern U.S. weather.

In the positive phase of the NAO, 500-mb heights over Iceland are anomalously low, while 500-mb heights over the Azores are anomalously high. This set-up creates an enhanced north-south height gradient over the Atlantic, which favors strong, zonal (west-east flow) over the basin from North America toward Europe, as shown in the schematic below.

Schematic showing the positive phase of the NAO resulting in fast zonal flow over the Atlantic

An idealized representation of the positive phase of the NAO. Anomalously low 500-mb heights near Iceland and anomalously high 500-mb heights over the Azores create an enhanced height gradient over the Atlantic Ocean, resulting in fast zonal flow.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

The fast zonal flow has some important important implications. First, the tendency for zonal flow to dominate limits large pools of cold air (large 500-mb troughs) from diving into the Eastern U.S. And, when cold air does invade the eastern U.S., it tends to exit quickly offshore.  Therefore, when a positive NAO prevails, conditions in the eastern U.S. tend to be milder, and the zonal flow leads to weaker storms (and less snow). Northern Europe also tends to be milder, but wetter than average, while Central and Southern Europe tend to be drier.

The negative phase of the NAO is characterized by a reversed pattern of height anomalies (compared to the positive phase). When the NAO is negative, 500-mb heights near Iceland are anomalously high (usually thanks to a large 500-mb ridge or closed high over the region), while 500-mb heights near the Azores are anonymously low (which translates to a relatively weak subtropical high). This orientation of height anomalies tends to weaken the westerlies over the Atlantic because of a weakened height gradient, as shown in the schematic below.

Schematic showing the negative phase of the NAO and weakened westerly flow over the Atlantic

An idealized representation of the negative phase of the NAO. Anomalously high 500-mb heights near Iceland and anomalously low 500-mb heights over the Azores (a weak subtropical high) create a weakened height gradient over the Atlantic, resulting in slower, "wavier" flow.
Credit: David Babb @ Penn State is licensed under CC BY-NC 4.0 (opens in a new window)

The weaker height gradient results in "wavier" flow aloft (more troughs and ridges), which favors more cold outbreaks in the Eastern U.S. as the strong high over the North Atlantic acts as a block, and slows the progression of weather systems across the U.S. With deeper troughs, the negative phase of the NAO also favors the development of stronger mid-latitude cyclones, which can lead to major East Coast snow storms when cold air is present over the East. When a winter season is characterized by a predominantly negative NAO, the winter tends to be cold and snowy across much of the Eastern U.S. and Europe. Indeed, deeply negative phases of the NAO usually get snow-lovers in the Eastern U.S. all worked up!

Observing and Predicting the NAO

Like with the MJO, the NAO is tracked with a mathematical index over time. This graph shows the three-month average state of the NAO (opens in a new window) from 1950 through September of 2013. While the phase of the NAO tends to fluctuate on time scales of a few weeks, one particular phase of the NAO can dominate for entire seasons or longer (as the graph suggests). For example, during the winter of 2009-2010, the NAO was overwhelmingly negative, and included the strongest negative signal on record. Not surprisingly, much of the eastern U.S. (and the entire country, in fact) experienced a colder than average winter (check out the seasonal anomalies below), and some major East Coast snow storms (like the early February ”Snowmageddon“ storm (opens in a new window)) occurred while the NAO was highly negative (ARCHIVED) (opens in a new window).

Historical rankings of temperature by climate division for the period from December, 2009 through February, 2010.

Climate divisional temperature rankings show that December, 2009 through February, 2010 was colder than average in much of the eastern U.S. (and the nation overall). This period was characterized by a strongly negative NAO phase.
Credit: National Centers for Environmental Information

While the exact reasons for the fluctuations in the NAO are unknown, global models have shown good skill in predicting its phase up to a week into the future, and modest skill in predicting it out to two weeks. So, the NAO is another useful medium-range forecasting tool. If we have an idea of the future state (and strength) of the NAO signal, we can have more confidence that a particular large-scale weather pattern will prevail.

Generally, ensemble forecasts of the NAO are most useful, because they give us a feel for the degree of confidence in a future NAO state.  Here's an example of a GEFS NAO forecast from October 3, 2013 (opens in a new window). The black line on each graph shows the state of the NAO index for the previous 120 days. On the top graph, the red lines represent individual ensemble member forecasts. The three graphs below the top one show the ensemble mean 7-day, 10-day, and 14-day forecasts, represented by the blue line (so that you can track its performance over the last 120 days, and see the ensemble mean prediction for the next 7, 10, and 14 day periods). The yellow shading represents forecasts within one standard deviation of the mean, while the red lines indicate the maximum and minimum ensemble forecasts. 

As you can see on this particular set of graphs, the 7-day forecasts had been performing well (a correlation of 0.8823) over the last 120 days, but the 14-day forecasts weren't (a correlation of 0.4893 with observations). At the time the forecasts were initialized, the NAO index was weakly negative, but was predicted to trend back toward neutral or slightly positive by most ensemble members in the coming week or two.

The Arctic Oscillation

If you understand how the NAO acts as a tool to gauge the behavior of the westerlies over the Atlantic, then you already understand the basic gist of the Arctic Oscillation (AO, but also sometimes called referred to as the NAM -- Northern Annular Mode, not to be confused with the now defunct NAM computer model). Instead of focusing on specific regions near Iceland and the Azores, the AO compares pressures in the Arctic with those in the northern mid-latitudes (roughly 37-45ºN).

The AO is in its negative phase when pressures are anomalously high in the Arctic (which translates to a weaker than normal polar low), while pressures in the mid-latitudes are anomalously low. In the negative phase, the westerlies are weakened because of a weakened pressure gradient (like the negative phase of the NAO), which helps pave the way for cold air to plunge into the middle of the U.S. and western Europe. In the positive phase, pressures are anomalously low in the Arctic (a stronger than normal polar low), and anomalously high in the mid-latitudes, yielding a larger pressure gradient and stronger westerlies (like the positive phase of the NAO). The positive phase of the AO, then, often corresponds to milder weather in much of the U.S., and weaker storms. During the winter of 2009-2010 (briefly discussed above), the AO was also highly negative (opens in a new window), just like the NAO.

Also like the NAO, the AO is a medium-range forecast tool, in that we have some skill in predicting its state out to about two weeks. If we have a handle on the state of the AO, we have some clues about the large-scale weather pattern a week or two into the future (especially when the AO signal is strong). CPC uses 1000-mb height anomalies to calculate its AO index, and given the similarities between the NAO and the AO, you're probably not surprised to find that their phases are often (but not always) the same.

You may be asking yourself "Why do forecasters look at both the NAO and AO if they're so similar? Which one is a better indicator of the strength of the westerlies?" Those are great questions, and you're not alone in asking them. Some researchers believe the NAO is merely part of the AO (and prefer the AO as a diagnostic tool). Other research, however, has suggested that we can make more physically meaningful measurements of the NAO, and thus, it's more representative of the true nature of the westerlies. So, some disagreement exists, and I included both oscillations here because both are commonly referred to in medium-range forecasting discussions.

Those who believe that the NAO is a better diagnostic tool advocate assessing it along with another oscillation focused in the Pacific to more truly gauge the nature of the westerlies. This oscillation, called the Pacific-North American (PNA) pattern, is our next topic. In the meantime, if you're looking for resources to track and forecast the NAO and AO, along with some further reading about both, check out the Explore Further section below.

Explore Further...

Key Data Resources

If you're looking for information on AO and NAO current conditions or forecasts, you'll be interested in these links:

Further Reading

If you're interested in reading more about the AO, NAO, and their teleconnections, I recommend the following resources:

sas405

The Pacific-North American Pattern

The Pacific-North American Pattern

Prioritize...

Upon completion of this page, you should be able to discuss the implications of the positive and negative phases of the Pacific-North American pattern on the upper-air pattern across North America.

Read...

The last oscillation that we'll cover is the Pacific-North American pattern (PNA), which you can think of as a complement to the NAO. The PNA helps forecasters to gauge the state of the westerlies over the Pacific and North America (the name makes sense, then, doesn't it?). Just like the NAO, the PNA has a positive phase and a negative phase, although each phase is characterized by a "quadripole" (four regions) of nodes and antinodes. Height anomalies of similar sign (nodes) appear south of the Aleutian Islands and in the Southeast U.S., while antinodes of the Aleutian anomalies appear near Hawaii and in the Intermountain region of the Western U.S. and Canada.

What does that really mean? When heights are anomalously low south of the Aleutian Islands, they also tend to be low in the Southeast U.S., while they tend to be anomalously high near Hawaii and in the mountains of the Western U.S. This set-up describes the positive phase of the PNA (a ridge over the Intermountain region of the West and a trough over the Southeast U.S. that teleconnect to the anomalies over the Pacific).

On the other hand, in the negative phase of the PNA, heights south of the Aleutian Islands tend to be anomalously high, as do heights in the Southeast U.S. Heights over Hawaii, meanwhile, tend to be anomalously low, as do heights over the Intermountain West. In other words, the negative phase of the PNA tends to favor a trough over the western U.S. (or at least a weaker than average ridge), and a ridge over the Southeast U.S., which teleconnect to the height anomalies over the Pacific.

To help you visualize the PNA in its positive phase, check out the four images below. They show the typical pattern of height anomalies associated with the positive phase of the PNA during January, April, July, and October. Note that the quadripole pattern discussed above is strongest in January (Northern Hemispheric winter), but is evident in all months shown, except for July. Indeed, the summer-time PNA signal is typically quite weak or non-existent. From these graphics, you can clearly see the below average heights south of the Aleutian Islands and in the Southeast U.S., as well as the above average heights near Hawaii and in the Intermountain West.

Maps showing the typical orientation of height anomalies associated with the PNA during January, April, July, and October.

Typical patterns of height anomalies associated with the positive phase of the PNA in January (upper-left), April (upper-right), July (bottom-left), and October (bottom-right). Note that the quadripole of height anomalies is strongest in January (meteorological winter) and weakest (nearly non-existent) in July.
Credit: Climate Prediction Center

Impacts of the PNA

The quadripole pattern of the PNA has some important implications for weather in the Pacific and over North America. In the positive phase, the anomalously low heights south of the Aleutian Islands and the anomalously high heights to the south near Hawaii create an enhanced height gradient aloft over the Pacific, strengthening the upper-level westerly flow over the basin (as the positive NAO does for the westerlies over the Atlantic). The exit region of these fast westerlies ends up near the West Coast of North America, enhancing cyclogenesis in the region and creating above average precipitation along the West Coast of Canada and in the Pacific Northwest of the U.S.

Meanwhile, drier than average conditions tend to persist in the Upper-Midwest during the positive phase of the PNA. Thanks to the ridge in the Intermountain West, temperatures tend to be above average in the Western U.S. during the positive phase, while they tend to be below average in the Eastern U.S. (especially in the Southeast U.S., where the core of the negative height anomalies sets up).

If the idea of an enhanced height gradient and stronger westerlies over the Pacific sounds familiar, recall what you learned about ENSO. This pattern of 200-mb height anomalies (opens in a new window) over the Pacific during the El Niño of 1997-1998 seems similar to the signal of a positive PNA over the Pacific, doesn't it? You might be asking yourself, "are ENSO and the PNA related?" Indeed, research suggests they are. While the PNA fluctuates on shorter time-scales than ENSO, the positive phase of the PNA tends to dominate during the El Niño phases of ENSO, and negative PNA patterns tend to be more common during La Niña episodes.

When the PNA is in its negative phase, the height gradient over the Pacific is weakened (with higher than average heights south of the Aleutian Islands and lower than average heights over Hawaii). With weakened westerly winds aloft, blocking tends to develop at high latitudes over the Pacific (much like the high-latitude blocking in the negative phase of the NAO (opens in a new window)), often leading to split flow over much of the basin. With a weakened ridge over the Intermountain West (negative height anomalies) in the negative phase, flow tends to be more zonal across North America, making it more likely that Pacific air can flood across the U.S. (which limits outbreaks of cold weather and the development of major cyclones).

Overall, in the negative phase, the Western U.S. tends to be on the cool side (with below average heights), while the Eastern U.S. tends to be mild (especially the Southeast U.S.). Precipitation tends to be above normal in the Mississippi and Ohio Valleys (opens in a new window) in the during a negative PNA episode, while the Southeast U.S. has drier than average weather; however, research has shown that the precipitation anomalies associated with the PNA are not as clear cut as the typical temperature anomalies.

Observing and Forecasting the PNA

Like our other teleconnections, the phase of the PNA is determined with a mathematical index (the one CPC uses is based on 500-mb height anomalies), so it can easily be tracked in time (check out the time-series of the three-month mean PNA index from 1950 through September 2013 (opens in a new window)). Forecasts of the PNA also take on similar formats to those of the NAO and AO, as you can tell from the GFS ensemble forecast for the PNA issued on October 10, 2013 (below).

GFS Ensemble forecast for the PNA issued on October 10, 2013.

GFS Ensemble forecasts for the PNA, issued on October 10, 2013. The black line on each graph shows the state of the PNA index for the previous 120 days. On the top graph, the red lines represent individual ensemble member forecasts. The three graphs below show the ensemble mean 7-day, 10-day, and 14-day forecasts, represented by the blue line. The yellow shading represents forecasts within one standard deviation of the mean, while the red lines indicate the maximum and minimum ensemble forecasts.
Credit: Climate Prediction Center.

The top graph indicates that the PNA had been negative in the final days leading up to October 10 (the day the forecast was issued), but had been trending toward neutral. The mean ensemble forecasts (the blue line on the lower three graphs) suggest a continued trend toward neutral with perhaps turn toward a positive PNA pattern 7-10 days after the forecast was issued. You'll also note that there was considerable disagreement amongst the ensemble members, with one outlying solution turning highly negative.

The performance of the ensemble mean forecast for the PNA during the 120-day period leading up to October 10 (for 7, 10, and 14-day forecasts) wasn't as good as the NAO forecasts we reviewed on the last page. Why might that be? Well, this 120-day period includes almost all of meteorological summer -- the period of the year when the PNA signal is very weak or non-existent (it's often totally non-existent especially in July and August). So, it's really not surprising that the ensemble forecasts for the PNA weren't very accurate during the summer. This idea is most obvious in the verification of the 14-day forecasts, which had a slightly negative correlation (-0.0805) with observations! Yikes! Forecast performance during other times of the year is typically higher -- high enough that the PNA is a useful medium-range forecasting tool, just like the NAO and AO (research suggests that the PNA phase is predictable out to 1-2 weeks, on average).

If you're interested in accessing current data and forecasts for the PNA, check out the Explore Further section below. It also includes some additional readings about the PNA and a more comprehensive overview of its teleconnections for U.S. weather. Otherwise, it's time to wrap up our lesson on medium-and long-range forecasting with a few final comments and a Case Study. Read on.

Explore Further...

Key Data Resources

If you're looking for information on PNA current conditions or forecasts, you'll be interested in these links:

Further Reading

If you're interested in reading more about the PNA and its teleconnections, I recommend the following resources:

sas405

Final Comments and Case Study

Final Comments and Case Study

Prioritize...

As we wrap up this lesson, focus on how the major teleconnections that we've covered don't always give a clear forecast picture. They sometimes send conflicting messages, and forecasters have to make difficult decisions about which aspects of the global weather pattern will dominate.

Read...

We've covered teleconnections based on a veritable alphabet soup of oscillations (ENSO, MJO, NAO, AO, and PNA), and we've truly only scratched the surface. Many more atmospheric teleconnections exist that are helpful for forecasting on various medium- and long-range time scales. But, even with knowledge of all of these oscillations and teleconnections, medium- and long-range forecasting can sometimes be quite challenging.

Oscillations sometimes send conflicting messages, and forecasters have to determine which signal will dominate, if any. The answers aren't always clear, and, yes, seasonal forecasts can be horribly wrong. Let's face it: The atmosphere is complicated. The oscillations and teleconnections in this lesson, however, (along with the discussion on computer guidance) should help you better understand the medium- and long-range forecasts and discussions that you encounter, and get you started in experimenting with your own medium- and long-range forecasts.

To see an example of how teleconnections can send conflicting signals, and how the weather doesn't always follow our "textbook" teleconnections, check out the Case Study below, which highlights the winter of 2014 - 2015.

Case Study...

The winter of 2014 - 2015 was a season of temperature extremes in the United States. On average, from December 2014 through February 2015, the western U.S. was much warmer than average, with many areas experiencing record warmth for the winter season. Meanwhile, the eastern U.S. was colder than average.

Climate Divisional average temperature ranks for winter 2014 - 2015.

Climate divisional temperature ranks for December 2014 through February 2015 (meteorological winter) show that the winter was characterized by record warmth in the West, with below average temperatures in the East.
Credit: National Centers for Environmental Information

This general pattern of warmth in the West and cold in the East was particularly amplified during the month of February (check out the climate divisional temperature rankings for February, 2015 (opens in a new window)). Most of the eastern U.S. was much below average during February, with a few climate divisions having their coldest February on record. It was a brutally cold month in State College, PA, where the mean temperature for the month was only 17.3 degrees Fahrenheit, which was not only the coldest February on record, but also the fourth-

 coldest month on record overall!

The winter of 2014 - 2015 was also very snowy in much of the Northeast (opens in a new window). Boston, Massachusetts had over 110 inches of snow during its cold season, which was more than 70 inches above average and an all-time record. Some of the snow piles in Boston actually lasted into July (opens in a new window)! It was a memorable winter, indeed.

With persistent cold in the eastern U.S., especially in mid-to-late winter, you might be thinking to yourself, "sounds like a case with a negative NAO." Remember that when the NAO is negative (opens in a new window), a blocking high over the North Atlantic slows the progression of weather systems across North America, and allows deeper troughs to plunge into the eastern U.S., which can lock in cold air there. Furthermore, the deeper troughs that characterize a negative NAO pattern are more likely to spawn stronger surface cyclones that can bring heavy snow to the Northeast.

However, if we look at the historical three-month running mean of the NAO index (opens in a new window), we find that the NAO was overwhelmingly positive during the winter of 2015. For what it's worth, the AO was overwhelmingly positive, as well. Remember that the positive phase of the NAO (opens in a new window) and the AO tends to favor a faster progression of air masses across North America thanks to faster westerly flow aloft (and smaller ridges and troughs). When the NAO is positive, the eastern U.S. is typically fairly mild.

So, forecasters that locked in on the positive NAO for most of the winter got exactly the opposite of what they expected. So, what went "wrong?" We can get some insight by shifting our focus westward and examining what was going on in the Pacific. Off the West Coast of North America, a pool of anomalously warm water in the Pacific Ocean (opens in a new window) persisted throughout the winter. This pool of relatively warm water warmed air columns in the region over time, creating a persistent ridge near the West Coast.

The image below, showing mean 500-mb heights during February 2015 (when the warmth in the west and the cold in the east was perhaps most amplified), shows the huge ridge near the West Coast, and a trough over the East. Note that this pattern is also characteristic of a positive PNA. Recall that a positive PNA favors anomalously low heights heights south of the Aleutian Islands and over the southeastern United States. Meanwhile, anomalously high heights occur near Hawaii and over the Intermountain West. Looking again at the map below, we see that the trough that digs down into the Southeast U.S. teleconnected to a "node" (another trough) that extends south of the Aleutian Islands. In between those two troughs, we had the huge ridge over the West.

Mean 500-mb heights during February 2015 over North America.

Mean 500-mb heights during February 2015 show a trough extending south from the Aleutian Islands, a big trough in the eastern U.S., and a ridge over the western U.S. All are characteristic of a positive PNA.
Credit: Earth System Research Laboratory

If we look at the corresponding map of 500-mb height anomalies (opens in a new window), we can see just how significant the trough over the eastern U.S. and the ridge over the western U.S. were. With significant anomalies like these averaged over a month, no wonder records were broken in both the western and eastern United States!

As a check on our observation that this 500-mb pattern was characteristic of a positive PNA, take a look at the historical three-month running mean of the PNA index (opens in a new window). It shows that a positive PNA dominated during the winter of 2015, and our typical expectations with a positive PNA in the winter (cold in the East, warm in the West) fit much better with what actually happened.

So, even though the NAO explains much of the variability during winter in the eastern United States, it did not hold all the answers in this case. Usually, for deep chill to persist in the eastern U.S., blocking must occur over the North Atlantic, allowing a deep trough to dig into the eastern U.S. and persist for long periods of time (characteristic of a negative NAO). But, in this case, we didn't have blocking over the North Atlantic. Remember that the NAO was overwhelmingly positive, and it was still cold in the East.

Air masses did move quickly across eastern North America as we would expect with a positive NAO, but because of the large-scale pattern driven by the water temperatures in the eastern Pacific (and characteristic of a positive PNA), the air masses entering the eastern U.S. were bone-chilling air masses coming from northern Canada and the Arctic. In other words, the record-breaking chill in the East occurred because frigid air mass after frigid air mass raced through the East. As soon as one frigid air mass left, another one replaced it! This was not a case where a single frigid air mass took hold of the East and didn't release its grip for weeks (as happens in negative NAO set-ups).

Any way you slice it, February 2015 was brutally cold in the East, and simply knowing the NAO phase by itself was of little help to forecasters because the pattern in the Pacific overwhelmed other atmospheric teleconnections. I hope this example shows you that forecasters should never look at any single oscillation in a vacuum. These oscillations sometimes send conflicting messages, as they did in this case (the positive NAO suggested mild weather in the East, while the positive PNA suggested cold weather). It's critical to weigh as many pieces of the puzzle as you can and keep your eye on the big picture (SST patterns, locations of large troughs and ridges, etc.). Even with knowledge of all of these oscillations, and access to computer-model guidance, medium- and long-range forecasts still pose an enormous challenge!

sas405