Upon completion of this page, you should be able to describe the advantages of models like the Rapid Refresh (RR) and High-Resolution Rapid Refresh (HRRR) in mesoscale forecasting. You should also be able to discuss their limitations and the importance of looking for consistency in successive solutions.
On February 10, 2009, supercells erupted over parts of the Southeast States. The 2238Z radar reflectivity (below) from Maxwell Air Force Base (KMXX) indicates the rather small coverage of the severe thunderstorms over eastern Alabama and western Georgia. Only the most favorable local environments supported deep, moist convection at this time. Of course, there was no way to predict exactly where these supercells would have developed, but accurately identifying the general area (Alabama, Georgia and parts of the surrounding states) where storms were likely to "initiate" on this day would have been a pretty good forecast. It turns out that these storms spawned several reports of tornadoes and numerous reports of large hail across the region.
To successfully identify regions at risk for severe thunderstorms, forecasters first assess the background synoptic-scale pattern by looking at progs from models like the ones you learned about in your previous studies (the GFS, NAM, or others). Assessing the "big picture" from these models is a crucial step in the forecasting process. But, for outbreaks of thunderstorms like the one shown above, these models have some serious flaws. One is that important convective processes are occurring on spatial scales that are smaller than the model's grid-point scheme. The end result is that convection in these models is greatly oversimplified (formally, "parameterized"), which leads to struggles with forecasts for convective precipitation.
Another major problem stems from the fact that, as you just learned, many mesoscale weather features have a relatively short duration. Supercell thunderstorms typically last one to four hours before dissipating (some other types of thunderstorms last less than one hour). But, models like the NAM and GFS are only initialized every six hours (00Z, 06Z, 12Z, and 18Z).
In terms of mesoscale weather, a lot can change in six hours! This relatively long time lag between successive runs, in addition to the inability to infuse hourly observations into the operational GFS and NAM, make these two models less viable for predicting the changing, smaller-scale environments that might favor the initiation of thunderstorms in the next hour (or even a couple of hours).
Forecasters require "mesoscale" models, with a fine spatial resolution, that are continually updated with timely weather observations so that they can more reliably refine and update their forecasts as weather conditions change in time. Do such models exist? Indeed they do. In 2012, NCEP implemented the Rapid Refresh Model (RR), a short-range model that incorporates NAM forecast data and an analysis / assimilation system to update the model with hourly observations. The Rapid Refresh Model runs every hour, providing crucial short-range forecasts. Forecasters at the Storm Prediction Center, as well as forecasters in the aviation community, frequently incorporate RR analyses (0-hour forecasts) and predictions into their forecasting routines.
The RR model provides data that have a relatively high resolution in space and time (forecasts are available at one-hour intervals). There's also a high-resolution version of the Rapid Refresh that mesoscale forecasters use operationally (the High-Resolution Rapid Refresh or, more simply, the HRRR). For the record, the HRRR model has an even higher spatial resolution, and offers forecasts at 15-minute intervals (read more about the details of the HRRR, if you're interested).
Models like the RR and HRRR have a couple of key advantages. First, because they're initialized every hour, they're more "in touch" with rapidly changing weather situations than models that are initialized every six hours (like the GFS and NAM). Second, with forecast intervals of an hour or less, the RR and HRRR are able to depict the evolution of mesoscale weather systems with greater detail than models having longer forecast intervals.
Furthermore, convection in the the HRRR is not parameterized. It has a sufficiently high spatial resolution that it can actually simulate real convection. Such models are called "convection allowing" models and need to have a grid spacing no larger than four or five kilometers. Because it doesn't have the great oversimplifications that come with convective parameterizations in coarser models, the HRRR is able to depict much more realistic convective structures. As you can see from the HRRR forecast below, its prediction of radar reflectivity looks pretty realistic, doesn't it?
In the six-hour forecast of radar reflectivity from the 17Z run of the HRRR on April 27, 2011, valid at 23Z (shown above) note the placement and structure of the narrow squall line in western New York and northern Pennsylvania. Now, compare the forecast to the actual 23Z mosaic of composite reflectivity. As you can clearly see, the HRRR had an awesome forecast, capturing the timing and structure of the squall line really well. On the other hand, the HRRR didn't predict the severe storms that formed out ahead of the squall line at all. The HRRR forecast also had problems in Maryland, Ohio and West Virginia, so this forecast was far from perfect.
I hope this example makes it clear that even though such "convection-allowing" models create detailed, realistic-looking convective structures, that does not mean their solutions are always accurate. Indeed, while such models are skillful in predicting the mesoscale details and structure of convection, they do not show consistent skill in predicting the exact timing or location of individual convective cells.
Another problem with "convection-allowing" mesoscale models is that they are prone to huge run-to-run variability (successive solutions may look nothing alike). To combat the large run-to-run variability, forecasters often look for a degree of consistency in three consecutive runs of the HRRR. If the model's solution is similar for three runs in a row, then forecasters have a bit more confidence in the solution. Researchers involved in the Vortex2 project routinely weighed HRRR forecasts to help them formulate plans to intercept storms. If the HRRR was producing consistent solutions for three consecutive runs, chasers would adjust their intercept plans accordingly.
Because these mesoscale models require great computer power to run, they are only run over a short forecast period (a day or less). Furthermore, their performance is somewhat at the "mercy" of the NAM model's initialization. Remember that the NAM feeds its initial conditions into the Rapid Refresh, so any major errors in the NAM initialization will be transferred into the Rapid Refresh, which can wreak havoc on its forecast accuracy.
Regardless of these limitations, the analyses and forecasts based on the Rapid Refresh Model are still often useful for timely short-range mesoscale prediction. For much of our work in this course, we'll focus on real-time mesoanalyses from the Rapid Refresh model available on SPC's Web site. As outbreaks of severe weather unfold you can rely on these SPC analyses to gain insight about the background synoptic and mesoscale environments.
To give you an example of the types of analyses that are available, check out the SPC mesoanalysis of vertical wind shear between the ground and an altitude of six kilometers over Deep South at 23Z on April 27, 2011. Vertical wind shear refers to a change in wind speed and / or direction with increasing altitude, and it's an important variable in determining the organization and longevity of thunderstorms that develop. On this particular date, very strong vertical shear existed over the Deep South, which played a role in one of the biggest tornado outbreaks in U.S. history that occurred over the region.
Later on, we'll get into the basics on how you can interpret these and other mesoanalysis images, and discuss their connections to the development of deep, moist convection. If you're interested in seeing more about this outbreak, and getting some links where you can access RR and HRRR forecasts, check out the Explore Further section below. Before we end this lesson, however, allow me to introduce the 4-kilometer NAM, which also has some utility for creating short-term mesoscale forecasts. Read on.
April 27, 2011
The mesoanalysis of vertical wind shear between the ground and six kilometers above came from April 27, 2011, the date of one of the biggest tornado outbreaks in U.S. history. We'll encounter this outbreak again later in the course, but for now, I thought you might be interested in a few tidbits about this outbreak:
- 23Z composite of radar reflectivity, showing swarms of supercell thunderstorms over the Southeast
- SPC storm reports for the date
- YouTube video of an EF-5 tornado in Philadelphia, Mississippi
Key Data Resources
If you're looking for forecasts from the Rapid Refresh or High-Resolution Rapid Refresh, you may be interested in the following links. They'll give you an idea about the variety of forecast variables available from these models, some of which you may already be familiar with. We'll cover some others this semester, but some are beyond the scope of the course.
- Rapid Refresh model fields
- High-Resolution Rapid Refresh model fields
- Rapid Refresh soundings: soundings from other models are available, too. Select one of the "RAP" options to get a sounding from the Rapid Refresh. Select your valid time, the three-letter airport ID for the station you want, and choose your output type. Most output types are interactive, but may take up to 30 seconds to load.