Tag Archives: Weather Models


The main model used in US weather forecasting is the GFS (Global Forecast System), a model that continues to be developed and improved. The last major upgrade to the GFS model occurred in June 2019 with the release of the GFS-FV3 which was “version 15” in its evolution.

I wrote a brief post about this model back then, describing the changes in “FV3” ( Finite-Volume Cubed Sphere Dynamical Core) geometry that the model uses to divide up the atmosphere into geometric units that can be handled mathematically. The complexity of these models is amazing and the NOAA scientists that create these models can only be admired.

The new version of the GFS model continues with the FV3 geometry (in the long line of this model evolution, this will be GFS version 16) and it has just become available this past week on a full time NOAA server. It’s available in what is referred to as a test “parallel mode”, meaning that it is running at the same time the current operational version (v 15) is running.

If you’re into technical stuff, or just want to be dazzled by the science of these models, check out this early presentation from the NOAA Model Evaluation Group. It gives a brief history of the GFS model development and also presents the improvements, short-comings and challenges of the new GFS v16 model. The official GFS v16 evaluation website can be found here with plenty of additional information.

If all goes according to plan, it should become the GFS model early February 2021.

I would expect general improvements in weather forecasts with its release. Additionally, the GFS model is used to set the boundaries and initial conditions for several other models—


So several other models’ accuracy will benefit when the GFS v16 becomes operational early 2021.

What could be a better example of the new GFS model’s improvement than today’s cloud forecast? Last night’s operational GFS had forecast for it to be sunny here at 3 PM today, keeping the clouds to our west —

Current GFS model showing clouds stay to our west at 3 PM (Click on image for a larger view.)


The new, soon to become operational GFS version 16, had a correct cloud forecast for today at 3PM—

New GFS (version 16) had the correct cloud forecast for this afternoon, showing clouds overspread much of the area.  (Click on image for a larger view.)

Last night’s indecision about cloud cover for today would have been less of a dilemma had I used the new GFS.

One of the issues of the new GFS is that the new model’s data size is almost twice as large as the existing GFS. NOAA is concerned about the data download size issue and it is expected to limit data download rates within the month for all their models. We’ll see how that plays out.

Another minor issue— the higher complexity of the new GFS takes more computational time. As a result, the new model take 7-12 minutes longer to complete. As an example, the first 24 hours of the current GFS become available about 10:37 PM EST. With the new GFS model, the first 24 hour forecast data will become available about 10:46 PM EST.

Those 11 PM TV weather entertainers are going to have even less time to update their forecasts during Eastern Standard Time.

The new model is an overall improvement, but not without its problems. It tends to handle tropical systems not so well, sometimes developing systems that don’t actually form. Certain parameters used for forecasting thunderstorms are understated, especially when soil moisture is low. Forecast temperature biases, while improved, also have their issues. You can be sure that scientists are already hard at work with a future version 17.

One more thing— the HRRR and RAP models were updated earlier this month and have shown improved forecasting accuracy and an increase in model forecast length. The HRRR has become a full 48 hour model and the RAP a 51 hour model. These same models are expected to start using the same FV3 core geometry used by the GFS in their own next iterations.

Several other high resolution models (HIRESW and HIREF) have been upgraded to the FV3 core geometry and are planned for release in the next few months. Their output is now available as “parallel” models.


My regular forecast followers may have noticed that I’m spending more time using  some new, high resolution models for these forecasts— these models are referred to as HIRESW  (High RESolution Window) models.  The models are run by NOAA/NCEP. I’ve been really impressed with these models;  I have only recently gotten access to them.

These models were developed through open-source university development of the WRF (Weather Research Forecast) Model around 2002 and later enhanced and further developed.  (In fact, the old ETA model starting using the WRF physics packages about 2005; subsequently, the ETA then became known as the NAM.  

–> There are two major forks and development paths of the HIRESW models, an ARW (Advanced Research Weather version) and an NMM version (a Non-hydrostatic Mesoscale Model version). 

The ARW and NMM versions use different “physics packages” and different “initializations”.     (There are different sets of equations  or “packages” used in each that make different assumptions and approximations about the atmosphere.  These equation packages are used to predict things such as rain, clouds etc..  The “packages” are referred to by the name of the researcher who developed these advanced, applied equations.)

It’s complicated, but in a loose way,  the HIRESEW-ARW  version is closer to the physics of the GFS model and the HIRESW-NMM version is closer to the physics of the NAM model. 

In recent years, there have been further improvements in both versions (ARW, NMM) of these HIRESW models.   These improvement included increases in resolution, (now 3 kilometers).

More interestingly, new statistical versions have been developed for each model — groups (“ensembles”) of models with intentionally introduced known errors called “perturbations. These allow forecasters to see how known errors statistically affect the computed forecast outcomes.

Out of these ensembles, it emerged that two perturbed versions (called “members”) of each seemed to offer improved forecasts:

  • HIRESW-ARW-MEM2   (MEM 2 refers to member two)
  • HIRESW-NMMB2      (B2  refers to grid B, member two)

Both of these versions are run twice a day at 8 AM and 8 PM EDT., for the continental US by the National Center for Environmental Prediction (NCEP-NOAA).  These high resolution models forecast up to 48 hours out in time.

The forecast output of these models becomes available between 10:25 AM/PM  EDT for the first 24 hour forecast and about 10:58 AM/PM EDT for the next 24 hour forecast.) 

That’s why I’ve been holding back posting my forecasts until after 11 PM in recent days.  (The model data available at 11 PM  Friday forecasts out to 8 PM Sunday. )

Sincere there remain differences in the forecasts of both version’s ensembles, you’d think there would be a model that combines them.  There is.  It’s called the HIREF.  I haven’t found it very useful to date.

The latest development for the HIRESW models still in experimental stages, is use of a new grid arrangement, the FV3 geometry, that is being used for the latest GFS model released in May 2019.