Fully explicit real-time WRF forecasts of Frances, Ivan and Jeanne (2004)

Real-time forecasts of three of the major Florida hurricanes from 2004 using the Advanced Research Weather Research and Forecasting (ARW) model are examined. The forecasts were integrated on both 12 km and 4 km grids beginning at 00 UTC during the time when each hurricane threatened landfall. For ARW forecasts on the 4-km domain, no cumulus parameterization was used. Both domains were initialized directly from the GFS model with no additional data assimilation or balancing. In most cases, even on the 4-km grid, spin-up was complete by about 6 h.

In all three cases, the 4-km ARW was able to predict the correct sign and approximate magnitude of the wind intensity change within 48 h of landfall. In both Frances and Ivan, the prediction of maximum 10-m wind from the ARW was more accurate than the official forecast issued 3 h after the model initialization time. ARW captured weakening and subsequent strengthening of Frances prior to landfall, and captured the weakening of Ivan just prior to landfall. For Jeanne, the ARW predicted a storm too weak at landfall, however, the intensity change forecast was more accurate than that of the official forecast.

Whereas the tracks of all storms were broadly similar to other models and to the official guidance, there was one noteworthy exception. In Frances, the ARW correctly delayed landfall relative to the corresponding official forecast. Because the ARW is not coupled to the underlying ocean, the absence of upwelling could have contributed to the positive intensity bias in that forecast.

Regarding storm structure, there were numerous rainbands in each storm that were realistic in terms of timing, number and quadrant of the storm. In Frances, a series of rainbands was predicted with remarkable accuracy. The eyes of each simulated storm were too wide compared to reality. In Ivan the error was at least a factor of two, and seemed to occur from the beginning of the simulation. The central sea-level pressure was also too high. Implicit in this result is the need for improved assimilation of initial storm structure.

Overall, the results suggest that the WRF model has considerable potential for hurricane applications, both as a research tool and for improving forecasts of structural features. Fundamental problems of under-resolved core structure are addressed using a 4-km grid and the advanced numerics in ARW and foregoing convective parameterization in favor of explicit cloud schemes. We are particularly hopeful for significant improvements in the landfall forecasts, especially if additional, underutilized, information, such as radar data and the HRD surface wind analyses can be explicitly included at the assimilation stage.