Probabilistic prediction of tropical cyclones. Part I: Position

A new objective aid for operational prediction of the positions of tropical cyclones is presented. Its method is based on a simple analysis of the performance of all operationally available numerical models during a training period. In a subsequent forecast period, the results of this analysis are used to produce both high-quality deterministic (as by-product) and probabilistic storm position forecasts in the form of geographical maps of strike probability for all prediction times out to 120 h from a given base date and time. The model was developed using operationally available position predictions of all global tropical-cyclone events of the years 2000, 2001, and 2002 as provided in the U.S. Navy's Automated Tropical Cyclone Forecasting System (ATCF). Forecasts have been carried out for the years 2001 and 2002, with corresponding training periods 2000 and 2001. The global annual mean deterministic position errors for the years 2001 (2002) were 148 (136), 266 (235) 393 (329), 541 (435), and 733 (554) km at 24-, 48-, 72, 96-, and 120-h prediction times, respectively. The deterministic mean errors and their corresponding standard deviations were found to be lower than those of most statistical and dynamical operational models and approximately equal to those of all consensus approaches currently in operational use. The main feature of the new method is the automatic production of geographical strike probability maps. For all tropical-cyclone events during the year 2001 (2002), the mean annual diameters of, for example, the 66% strike probability regions (i.e., the regions inside which future storm positions can be expected with a probability of 66%) at 24-, 48- 72-, 96-, and 120-h prediction time were found to be 290 (280) 490 (480) 710 (660) 1020 (780), and 1310 (1020) km, respectively. At all prediction times, the predicted sizes of areas of given strike probability represent conservative estimates in that the observed percentages of storm positions inside these areas are larger than the corresponding expected percentages.