Computer models used to forecast hurricanes

The behavior of the atmosphere is governed by physical laws which can be expressed as mathematical equations. These equations represent how atmospheric quantities such as temperature, wind speed and direction, humidity, etc., will change from their initial current values (at the present time). If we can solve these equations, we will have a forecast. We can do this by sub-dividing the atmosphere into a 3-D grid of points and solving these equations at each point. These models have three main sources of error:

1) Initialization. We have an imperfect description of what the atmosphere is doing right now, due to lack of data (particularly over the oceans). When the model starts, is has an incorrect picture of the initial state of the atmosphere, so will always generate a forecast that is imperfect.

2) Resolution. Models are run on 3-D grids that cover the entire globe. Each grid point represents of piece of atmosphere perhaps 40 km on a side. Thus, processes smaller than that (such as thunderstorms) are not handled well, and must be "parameterized". This means we make up parameters (fudge factors) that do a good job giving the right forecast most of the time. Obviously, the fudge factors aren't going to work for all situations.

3) Basic understanding. Our basic understanding of the physics governing the atmosphere is imperfect, so the equations we're using aren't quite right.

Types of hurricane forecasting models
The best hurricane forecasting models we have are "global" models, that solve the mathematical equations governing the behavior of the atmosphere at every point on the globe. Models that solve these equations are called "dynamical" models. The four best hurricane forecast models--GFDL, GFS, UKMET, and NOGAPS--are all global dynamical models. These models take several hours to run on the world's most advanced supercomputers. There are also dynamical models that cover just a portion of the globe. These are less useful, unless the hurricane happens to start out inside the domain the model covers and stay there. Hurricanes moving from outside the model domain into the model domain are not well handled. An example of this kind of model is the NAM model covering North America and the surrounding waters, run by the National Weather Service (NWS). Another type of hurricane model is a statistical model. These models do not try to solve mathematical equations on a grid. The advantage of these statistical models is that they are fast to run and can provide output in a few minutes. There are also hybrid statistical/dynamical models, and simple trajectory models.

A summary of the top six models:

GFDL: The NWS/Geophysical Fluid Dynamics Laboratory model. This model has been the best overall performing model the past three years. An upgrade to the model this spring has produced a 10-12% improvement in track forecasts and a 30% improvement in intensity forecasts when re-run on hurricanes from the 2005 season. The GFDL and HWRF models are the only models that provide specific intensity forecasts of hurricanes. Wunderground.com makes these graphics available on our computer models page. More deatiled GFDL graphics are available at NOAA/NCEP. See the "GHM" model under the heading, Hurricane Graphics.

GFS: The Global Forecast System model run by the NWS. Excellent graphics are available on the web from the National Center for Environmental Prediction. I like their long surface pressure/precipitation loops. Wunderground.com also has GFS plots. I like the Tropical Atlantic imagery. If you select "Shear" from the "level" menu, then click on "Add a Map", you'll get plots of the wind shear that I talk so much about.

UKMET: The United Kingdom Met Office model. Data from this model is restricted from being redistributed according to international argreement, and graphics from the UKMET are difficult to find on the web. Only paying subscribers are supposed to have access to the data.

NOGAPS: The U.S. Navy's Navy Operational Global Prediction Center System. Graphics are available at the Navy web site.

The European Center for Medium-Range Weather Forecasting (ECMWF) model. This is the premier global model in the world for medium range weather forecasting in the mid-latitudes; the 4-day ECMWF forecasts are as good as the 3-day GFS forecasts. Beginning in 2006, the ECMWF made improvements that starting producing very accurate hurricanes forecasts. Output from this model is also restricted via international agreement, and is available on the web only at the ECMWF web site.

HWRF: The NWS/Hurricane Weather Research Model. This model, new for 2007, is scheduled to replace the GFDL by 2009. HWRF is a non-hydrostatic a coupled ocean-atmosphere model, will utilize highly advanced physics of the atmosphere, ocean and waves in one prediction system, providing unparalleled understanding of the science of tropical cyclone evolution. Its output gives meteorologists an analysis of the hurricane in three-dimensions from real-time airborne Doppler radar. It will make use of a wide variety of observations from satellites, data buoys, and hurricane hunter aircraft. No other hurricane model accesses this wide of a range of meteorological information. The GFDL and HWRF models are the only models that provide specific intensity forecasts of hurricanes. Detailed HWRF graphics are available at NOAA/NCEP. See the "HWRF" model under the heading, Hurricane Graphics.

One other model worth looking at, but not as good as the other six is the Canadian GEM model.

Non-global models
The BAMM model (Beta and advection model, medium layer) is included on wunderground.com's computer model page. The BAMM is a simple trajectory model that is very fast to run, and did the best of any individual model at 3-5 day track forecasts in 2005. Since this model is always available, we have included it along with the "big four". In general, one should not trust the BAMM model for the 1-2 day time period when output from "the big four" are available. "The big four" are generally not available for tropical disturbances, and for these situations we post plots of a number of other non-global models such as the LBAR, A98E, etc. All of these models are described in detail on NHC's web site.

Model performance in 2005
The National Hurricane Center issues annual verification reports comparing model performance to the official NHC forecast. The 2005 report found that the official NHC forecast was usually the best forecast, but was closely matched by taking an average of the "big four" models to come up with a consensus forecast. There are several techniques used to come up with these consensus model forecasts. The three best techniques are called the GUNA, CONU, and Florida State University Superensemble (FSSE). The FSSE model was developed by FSU with funding from the private company Weather Predict, and is not available to the public. The performance of the "big four", official forecast, and consensus models are plotted below. Among individual track models, the GFDL did the best at 1-2 day forecasts, and the UKMET and BAMM (not shown in the plot) did very well at 4-5 day forecasts. For intensity, the SHIPS model (which we post in the lower right corner of the wunderground.com computer model forecast image) was the best performer. The SHIPS model is run using input from the GFS model.

Figure 1. Forecast performance in 2005, compared to a simple "Climatological and Persistence' (CLIPER) model. OFCL=Official NHC forecast. The "big four" global dynamical models are the GFDL, GFS, NOGAPS, and UKM (UKMET). Three methods of averaging the "big four" and coming up with a consensus model forecast are CONU, GUNA, and FSSE (Florida State Super Ensemble). The Official forecast and the three consensus forecasts did the best at all time periods. Among individual models, the BAMM model(not shown) did the best at the 3-5 day range, followed closely by the UKMET. The GFDL did the best in the 1-2 day range. Image credit: National Hurricane Center.

Jeff Masters, Director of Meteorology for The Weather Underground