May 21, 2009 Meeting Summary
Bob Tuleya presented a summary of preliminary results from the May 7-8 hi-res. workshop at NHC, which was a modified version of the presentation given by DTC's Louisa Nance and Ligia Bernardet. The goals of the high resolution hurricane forecast test (HRH) were to evaluate what kind of effect an increased horizontal resolution, for a given model, would have on intensity forecasts. This would hopefully lead to the production of a data set for use in a multi-model ensemble. Bob mentioned that for this study, comparison with operational models "was not emphasized." Ten storms from the 2005 and 2007 seasons were selected for a total of 69 cases. To evaluate model performance, DTC output grib files in 30 minute increments, then ran a modified GFDL tracker which created a modified a-deck. An averager was then used to compute a running mean of the maximum wind over a 2h window.
Bob then gave a summary of results from AOML, MMM, GFDL, and PSU. The AOML cases used the HWRFx model with no ocean coupling, GFDL initial conditions and a resolution of 9km versus hi-res. runs at 3km. Track absolute error (TAE) showed an improvement for the 3km resolution. Intensity error or bias plots showed that the 9km resolution under predicted intensity while the hi-res runs had less bias. The intensity absolute error (IAE) plot showed that the hi-res runs produced a better intensity forecast. For rapid intensification (RI), the probability of detection (POD) for hi-res. was higher than that for low-res.
For the MMM cases, a 1-d mixed layer ocean and low-res EnKalman filter was used for initialization. Low-res. runs used 12km versus 1.33km hi-res. TAE showed a more degraded track at hi-res., but the bias was less for hi-res. compared to low-res. There was found to be no significant difference between hi- and low- res. for the IAE plot. Once again, the hi-res. showed a larger POD value for RI.
The GFDL cases used POM ocean coupling and GFDL bogusing for it's low-res. (1/12 degree) and hi-res. (1/18 degree) runs. For TAE, bias, and IAE, there was not really a statistically significant difference between the hi-res. and low-res. values. The PSU cases used the ARW model and radar data at initialization via the EnKalman filter. Instead of two resolutions, they used three resolutions: 13.5km, 4.5km, and 1.5km. It was found that increasing the resolution improved the IAE and reduced systematic under prediction of the bias. Overall, use of individual models shows no better performance than operational models.
Next, Zhan Zhang presented his work on a regression model for hurricane ensemble forecasts. In his regression model, the regression coefficients are determined using least squares fit for every 6h. So far, only the HWRF and GFDL models were used, but more models can be added. To get the ensemble output, Zhan asked the question, was the number of records in the existing database > 30?. If no, then HWRF output from past years could be used as input into the regression model. If yes, then the regression coefficient could be computed using GFDL output as well.
Next, Zhan showed intensity error plots for 2008 season Atlantic hurricanes. Overall, the ensemble (HENS) showed lower errors than H209, HWRF (2008 oper.), GFDL, and the average (HAVE) at all forecast times but 120h. For Hurricane Fay, HENS showed lower errors except from 72h onward. For Gustav, HENS showed lower error values except at 48h compared to HWRF and H209. Hurricane Hanna's intensity error values were lower for HENS compared to HWRF, except at 24h. For Ike, HENS showed low error values except at 120h.
Plots showing the regression coefficient values for HWRF (in red) and GFDL (in green) at 12, 48, and 72 hours show that the coefficients level off around 200 cases. Zhan mentioned that when calculation the coefficients, he used the constraint that the coefficient values could never be less than zero or greater than 1. The wind-pressure relationship for the ensemble shows values closer to observed, but the slope is not improved.
Zhan mentioned that calculating track error was complicated because latitude and longitude are related and cannot be regressed separately. Thus, he used forecast direction and displacement. The plot showing track error for all 2008 Atlantic hurricanes shows lower HENS error values than H209 and HWRF but higher values than HAVE and values comparable to GFDL.