Upgraded HWRF and GFDL Hurricane Models Excelled During Hurricane Arthur

Today's guest post is by Dr. Morris Bender of NOAA's Geophysical Fluid Dynamics Laboratory (GFDL) in Princeton, New Jersey, and Dr. Vijay Tallapragada of NOAA's National Center of Environmental Prediction Environmental Modeling Center (NCEP/EMC), with help from GFDL's Timothy Marchok. They outline some very encouraging news on the ability of the latest versions of the GFDL and HWRF hurricane models to significantly improve hurricane track and intensity forecasts.

- Jeff Masters

The landfall last week of Hurricane Arthur, the first named tropical system in the Atlantic for 2014, brought a quick start to this year’s hurricane season. Perhaps lost in the predictions and preparations for Arthur’s landfall was the fact that there have been major upgrades this year to the two operational National Weather Service (NWS) regional hurricane prediction systems, the GFDL and HWRF models. Here we will provide background on each of those models and highlight the forecast improvements achieved from recent upgrades to both models.

Since 1995, the GFDL hurricane model has been an official operational product of the NWS, providing forecast guidance on track and intensity for the National Hurricane Center (NHC). The model was originally developed as a research tool, by scientists at NOAA’s Geophysical Fluid Dynamics Laboratory in Princeton (GFDL), to help understand the behavior and structure of tropical cyclones, including hurricane formation, decay and intensification. To adequately represent the structure of the hurricane and its inner core, the GFDL hurricane model required high resolution (distance between the individual grid points where the atmosphere equations of motion are solved), compared to other models of the atmosphere that define processes over the entire globe (typically called general circulation or global models--for example, the GFS and European models). Also, advanced physics were required to correctly reproduce the processes that occur in the hurricane core, as well as the interaction with the ocean below (Figure 1).


Figure 1. Inner core structure of Hurricane Katrina of 2005 simulated from the GFDL hurricane forecast model. Sea Surface Temperatures (SST) are denoted by the color shading, with the darker colors of blue showing the cooling of the SSTs due to the hurricane winds mixing the cooler waters from below to the surface.

In the early 2000s scientists at the NWS National Center for Environmental Prediction (NCEP) began to develop a new state of the art hurricane model using the most advanced numerical techniques available, to more accurately solve the mathematical equations that represent the processes of the atmosphere. This model (named HWRF, or Hurricane Weather Research Forecast model) became operational in 2007, as an official product of the National Weather Service. Since then, improvements have been made to the HWRF modeling system every year, resulting in a steady reduction in track and intensity forecast errors. The recently upgraded HWRF model implemented for the 2014 hurricane season has shown much reduced track forecast errors compared to the 2013 version of HWRF for a large sample of North Atlantic basin tropical cyclones, with performance comparable to the NHC Official Forecasts (Figure 2).


Figure 2. Track forecast errors from 2014 HWRF upgrades (H4FI, red) compared to previous year’s operational HWRF (2013 version, H3FI, blue) and NHC Official Forecasts (OFCL, purple) shown for a large sample of North Atlantic storms from 2008 to 2013.

A major accomplishment is the significant reduction of intensity errors from the HWRF model in the past three years since the model was upgraded to run using cloud-permitting, 3 km resolution nests (see Figure 3), making it a primary model for intensity forecast guidance for NHC. Much of the increased skill seen in the HWRF model over the past 3 years was due to the successful collaboration between agencies within NOAA (GFDL, NCEP, AOML, ESRL) and partners within the academic community (such as the University of Rhode Island), that was made possible through the coordinated efforts and support from NOAA’s Hurricane Forecast Improvement Project (HFIP). Apart from providing operational forecast guidance to the NHC for the Atlantic and Eastern Pacific basins, the HWRF model is also run in real-time for all global oceanic basins, providing forecast guidance to the US Navy’s Joint Typhoon Warning Center (JTWC). All real-time forecast products can be accessed from the HWRF website.


Figure 3. HWRF model intensity forecast improvements from 2011 to 2013 for North Atlantic basin. The intensity errors shown here are collected from hundreds of retrospective simulations for each upgraded HWRF configuration since 2011. The 2011 version of HWRF (blue) was run at 9km resolution, while the model was upgraded to run at 3km resolution in 2012 (purple). The 2013 HWRF (red) was able to meet or exceed the HFIP intensity error baseline whereas the 2014 HWRF (green) further reduced the intensity errors, approaching the HFIP 5-year intensity error goal.

At the same time, scientists at GFDL have also upgraded the GFDL hurricane modeling system, with major improvements made operational in 2014, particularly to improve the prediction of hurricane intensity as shown in Figure 4. Note that the improvements made to the GFDL hurricane model reduced the error in the prediction of the storm maximum wind by about 15% in the 3 to 5 day forecast time period, for a set of forecasts rerun from the 2008, 2010, 2011, and 2012 Atlantic hurricane seasons, using both the 2013 version of the GFDL model and the newly upgraded model.

The National Hurricane Center continues to support both of these operational regional hurricane models (HWRF and GFDL) since the forecast error of both models is often not correlated (individual model errors often go in different directions). Numerous scientific studies suggest that the average forecast from models that are well behaved produce errors that are less than those from the individual models. This has led to an increase in the use of model ensembles (models with slightly different initial conditions or different physics). For example, as shown in Figure 5, a consensus made up of the average of the intensity forecasts from the 2014 versions of the GFDL and HWRF (solid black line) results in an intensity forecast error that is significantly less than either the HWRF or GFDL model at every forecast lead time. Note that the average intensity forecast error of the 2-model consensus at days 4 and 5 is even less than the HFIP 5-year goal established in 2009.


Figure 4. Comparison of the intensity forecasts errors (in knots) from 12 to 120 hours in the future, for Atlantic storms rerun from the 2008, 2010, 2011 and 2012 Atlantic hurricane season. Plotted are the forecast errors for the version of the GFDL hurricane model used in 2013 (black line), compared to the version recently made operational in 2014 (red line).


Figure 5. Intensity forecast errors at hours 12 through 120, for the 2014 versions of the GFDL (blue) and HWRF (red) models for over 800 forecasts from the 2008, 2010, 2011 and 2012 Atlantic hurricane seasons, compared to the forecast error for the consensus model made up of the average intensity prediction of GFDL + HWRF.

Both the upgraded GFDL and HWRF modeling system did well for track and intensity forecasts for Hurricane Arthur, the first hurricane of the 2014 Atlantic hurricane season (Figure 6). The new GFDL and HWRF had very low track errors although the sample size was small, with the average intensity errors comparable to the other two top NWS intensity prediction models (Decay SHIPS and the LGEM).

With continuous advancements to the NCEP hurricane models supported by HFIP, and enhanced computational resources available for operational models, we anticipate further improvements in track and intensity forecasts through improved hurricane physics, advanced inner core data assimilation, and increased horizontal and vertical resolutions. Apart from coupling the atmospheric model to the ocean model, future efforts also include coupling to wave, land surface, hydrology, and surge and inundation models for improved prediction of land falling storms.



Figure 6. Average track forecast error (top) and average errors in the forecast maximum surface winds (bottom) for the upgraded GFDL model (green), the upgraded HWRF (red), compared to the official forecast of the National Hurricane Center (black), and other NWS operational models, for all forecasts of Hurricane Arthur (2014).

Morris Bender and Vijay Tallapragada