Extended Abstract (96K)P1.91
Deep layer of upper air and multivariate time series models to predict hurricane tracks (Paper Formerly Number 5C.5)
Joan Manuel Castro, University of Puerto Rico, Mayaguez
A multivariate time series model is used to predict hurricane tracks in the North Atlantic basin. Two types of data sets are developed to build the prediction scheme. The first data set is used to identify analog hurricanes to the current storm and the second data set uses steering and synoptic observations along the storm to predict the hurricane displacements. The first data set includes climatology and persistence information from the last 20 years and the data sources are the Hurricane Best Track, NCEP Reanalysis, and Radiosonde data. The Best Track and NCEP reanalysis data provide historical information while Radiosonde includes archived and current information. The first data set and a self organized neural network (NN) with the Kohonen learning rule are used to identify the storm analogs, which are based on: Julian date, Eastward and Northward displacement, passed intensity, storm direction and sea surface temperature. Once the analog hurricanes are identified the second set of data is developed based on the upper air of the analogs. The upper air for every analog is obtained from NCEP reanalysis at 17 vertical levels with origin located on the center of the storm. The horizontal area covers 20x20 degrees and is obtained at every six hours and spanned a long the storm track. Spatial interpolation algorithm is used to obtain estimation at one degree of resolution of the following variables: relative humidity, geopotential height, air temperature, and wind speed components. A deep layer for every variable is computed for the analogs. Radiosonde data and deep layer analogs are used to estimate the deep layers for the current storm. Dimensionality reduction is accomplished using the fist 10 principal components, which represents more than 90% of the total variance of the considered variables. An algorithm is used to identify the optimal lags and the best three variables that best explain the hurricane displacement. The best variables and the hook and Jeeves algorithm are used to identify the structure of the multivariate time series model and the hurricane displacement is predicted. A different prediction model is built at every 6 hours, implementing the described methodology. Aircraft reconnaissance and satellite data are under exploration to be included in the data sets to improve prediction capabilities.
It has been shown that the multivariate time series model predicts the storm displacements with small prediction errors, especially for the lead times of 12 and 24 hours. More exploration is required to claim that the proposed method is better than the existing prediction scheme. However, preliminary results show that the prediction scheme is a potential tool to increase the accuracy of predicting hurricane displacements.
Poster Session 1, Posters
Wednesday, 5 May 2004, 1:30 PM-1:30 PM, Richelieu Room
Previous paper Next paper