January 14, 2010 Meeting Summary
Zhan Zhang gave a presentation titled, "Preliminary Results and Work Plan for HWRF Ensemble Forecast". First Zhan gave an overview of his method and what is required for an ensemble system. He mentioned that ensemble perturbations were needed which could be either initial large-scale flow-based perturbations or physics-based perturbations. There could also be combined perturbations which would be a mix of the first two, or a multi-model ensemble. Zhan also mentioned that post-processing for an ensemble is very important to get a single ensemble forecast. Different methods included an equally weighted ensemble mean, a weighted average, or a hierarchical cluster analysis. For this presentation and his work, Zhan focused on the large-scale flow-based perturbation and the hierarchical cluster analysis.
Next Zhan discussed in more detail the large-scale flow-based perturbations. The experiment created was designed to look at the sensitivity of the HWRF model forecasts to the initial large-scale flows and the lateral boundary conditions. Two sets of GEFS data was used: one with T126 and 28 vertical levels and one with T190 and 28 vertical levels. There were 21 ensemble members of which 1 was the control (CNTL). Zhan included that the transform of the forecast error led to the initial perturbation. Results from this experiment showed that hurricanes dissipated in some of the ensemble members, which could have been due to the lack of vertical resolution. Zhan also noted that there is a different vertical resolution for GFS data in the ensemble and operational model. Based on experiment results, the suggested configuration for perturbations included T382 with 64 vertical levels (same as the current GFS) plus T190 and 28 vertical levels. Next, a sample ensemble forecast for Gustav at 2008082600 was shown with all ensemble members. For track and intensity, there was a large spread. This is further illustrated by ensemble spread plots for track and intensity. The track spread grew steadily with time while that for intensity grew quickly at first and then slightly leveled off. Zhan noted that for this specific case, the track spread was the same magnitude as the forecast error without the ensemble.
Then Zhan presented a more detailed description of the ensemble post processing methods. For an equally weighted ensemble average, which is the most common method, uncertainties in the model initial conditions are removed. For the weighted ensemble-average, model biases are eliminated based on past performance. Finally, for a cluster analysis, ensemble members are grouped into several clusters in a case-based scenario. Zhan expanded on the cluster analysis method by detailing a hierarchical cluster analysis. First off, distance (or similarity) is computed between each ensemble member. Then, each member is treated as a cluster, initially. Next, the two closest clusters are joined to form a new cluster, and the process is repeated until only one cluster remains. Zhan then showed two examples where clusters were compared to the ensemble mean using Fay. Then, the Gustav example mentioned above was analyzed using the cluster method. For that case, there were three large clusters compared to the total mean. The ensemble track spread was also shown for Gustav. Zhan noted that the spread was reduced, but the results were not necessarily better after analysis.
Zhan concluded by presenting his work plan. First he planned to find the optimum perturbation for the hurricane ensemble forecast. This involved deciding on a choice of initial flow-based perturbations (which included GEFS resolution) and perturbation vertical profile, a choice of physics-based perturbations, the number of ensemble perturbations, and how to combine the initial flow-based and physics-based perturbations. Zhan would also have to decide whether to develop ensemble transform (ET) -based perturbations for the HWRF system. For the post-processing system, there would be a choice of a weighted average method if a physics-based ensemble was used, a hierarchical cluster analysis for a case-dependent ensemble or to apply cluster analysis to a multi-model ensemble.
Next, Vijay Tallapragada presented an update on HWRF FY2010 experiments including results from all cases and preliminary results from combined upgrades. As a reminder from last week's presentation, Vijay mentioned that the results shown in this presentation included the HWRF (operational HWRF), H050 (the new baseline runs including bug fixes and the new GFS), H051 (which uses H050 plus initialization changes involving additional GSI data near the storm), H052 (which uses H050 plus surface physics changes involving new Cd/Ch coefficients calculated based on observations), and H053 (which uses H050 plus gravity wave parameterization). The test cases shown in these results include Cristobal, Dolly, Fay, Gustav, Hanna, Ike and Omar from the 2008 Atlantic season; Elida, Fausto, Genevieve, Marie, and Norbert from the 2008 East Pacific season; Bill, Claudette, Danny, Erika, Fred, Henri, and Ida from the 2009 Atlantic season; and Felicia, Guillermo, Hilda, Ignacio, Jimena, Linda, Olaf and Rick from the 2009 East Pacific season. All runs for H050, H051, H052, and H053 have been completed. The next round of testing, which is currently underway for combined upgrades, includes the HYC1 experiment, which uses HYCOM, and H054 (which combines H051+H052+H053). Future experiments include HYC2 (HYC1 plus H054), H055 (which uses H050 plus the newest GFS phase-2 implementation), HYC3 (which uses H050 plus the GFS phase-2 data plus HYCOM) and then the final HWRF configuration will be decided upon, called H210.
Next Vijay showed track and intensity error plots from various cases. Track errors for the 2008-2009 Atlantic storms showed H051 (in purple) with the lowest errors through 96h. H051 also showed a 5-6% improvement in track error compared to HWRF and an 8-9% improvement compared to H050 for the 2008-2009 Atlantic. Intensity errors from the 2008-2009 Atlantic showed H052 (in pink) with the lowest values. This was also true for standard deviation. H052 also had the lowest intensity errors relative to HWRF and H050. For the 2008-2009 East Pacific, the track errors were clustered with H051 and H053 (in orange) having the lowest values. Intensity errors for the 2008-2009 East Pacific showed the same findings as for the Atlantic with H052 having the lowest errors. These results were in line with those from only the priority testing cases.
Vijay then showed some preliminary results from the combined upgrades for priority cases Fay, Gustav, Hanna, and Ike for the H054 experiment. Vijay mentioned that extended testing for this experiment was in progress. Thus far, it looked like H054 track errors retained the skill of H051. The intensity errors for H054 improved over H052 through 96h but degraded at 120h, which was most likely due to Gustav. The track error plot for the priority cases showed H051 and H054 (in light blue) with very similar values and the lowest errors through 96h. Vijay also noted that H054 also reduced the northward bias (shown in the Y track bias plot). For track error relative to H051, H054 was better than H051 up to 48h before showing a slight degradation. Individual storm track errors showed H051 and H054 very close for Fay, H051 slightly better from 72h onward for Gustav, H054 improved over H051 for Hanna, and H051 and H054 the same for Ike. Vijay mentioned that for Gustav, removing GWD might improve track error.
The intensity error for all 2008 priority cases showed H054 with similar errors to H052 through 96h. For intensity error relative to H052, H054 performed better at 36h, and otherwise was very similar to H052, except at 120h where it was worse. For individual storm intensity errors, H054 was comparable to the other experiments for Fay. H054 was better at 72 and 96h before degrading at 120h for Gustav, and H054 was again comparable for Hanna. For Ike, H054 and H052 were very similar with the lowest error values. Vijay concluded by making the following suggestions for the model: adding two more satellite data sets (GOES and AIRS) and upgrading the hwrf_gsi to current operational regional (NAM) GSI. He also suggested withdrawing H053 and using only H051+H052. This configuration is currently being run for Gustav. Vijay also mentioned that diagnostics for large-scale fields is ongoing. This is particularly related to the simulation of the sub-tropical high mentioned last week, mid-latitude troughs and ridges and how they interact with the model storm circulation. Also, for the more than 4000 runs involved with this testing, Vijay expressed the need for several people to look into experiment forecasts. He also noted the need to develop some advanced tools to diagnose model output aside from track and intensity error. Future work involves focusing on these issues.