Every year hurricanes strike the United States with the devastating impact on lives and public and private properties. Enhancement of community resilience requires improving the quality of hurricane track forecasts and developing reliable tools for quantitative assessment of the level of confidence in issued forecasts to support decisions on storm-response activities. Among various forecasting techniques, multimodel forecasts, which combine forecasts produced by individual models for weather prediction, appear to be generally more reliable in predicting hurricane tracks than individual forecasts.

Previously, we developed (Poroseva, Letschert & Hussaini, 2006, 2007) a new multimodel approach based on the Dempster-Shafer theory of evidence (Shafer, 1976) for forecasting hurricane tracks. Mathematical foundation of the approach differs completely from currently existing prediction tools. This approach allows one to fuse individual forecasts produced by different models along with the quantified information on the past performance of the models. At each forecast period, the multimodel prediction is given in a form of areas where the hurricane position is likely to occur. Each area has the quantitative assessment of the confidence level in the prediction. Areas with the highest level of confidence at each forecast period form the final hurricane/typhoon track forecast. The approach was successfully validated using two models produced by the U.S. Navy Operational Global Atmospheric Prediction System (NOGAPS) and the European Center for Medium-Range Weather Forecasts (ECMWF) and the best track database for 70 tropical cyclones in the North Pacific and South Basins of the year 2000.

Currently, we are studying how the number of individual models included in the multimodel forecast influences its accuracy. Computational algorithm, which allows one to combine up to ten individual forecasts, has been developed for this purpose. The results of multimodel forecasts including three and four individual forecasts for the year 2000 will be discussed in the final paper and compared with those obtained with two models – NOGAPS and ECMWF – only. Additional two models currently included in the multimodel forecast are the NCEP Medium Range Forecast (MRF) model and the Aviation Run of the MRF model (AVN). Notice that any model for which the data are available can be a part of the multimodel approach.

Further improvements in the approach are introduced, which stem from overcoming some limitations imposed by Dempster's rule application. In the multimodel approach, Dempster's rule is used for combining individual forecasts along with the quantified information on the models' past performance. Valid use of Dempster's rule requires independence of combined evidence. In hurricane/typhoon track forecasting, evidence comes from the predictions of models and best track data. To satisfy this requirement, past performance of different models was evaluated in Poroseva, Letschert & Hussaini (2006, 2007) using the forecasts produced by different models and different data for best track positions of different tropical hurricane/typhoons. If data is limited (which is the case at the inception of the hurricane season), one would like to use database in a more effective manner. The full paper will include a detailed discussion on the definition of evidence independence, its application to hurricane/typhoon track, and as to how one can optimize the use of available data without jeopardizing mathematical rigor of Dempster's rule.