Wavelet Support Vector Machines for Forecasting Precipitation in Tropical Cyclones: Comparisons with GSVM, Regression, and MM5

Abstract

Abstract This study presents two support vector machine (SVM) based models for forecasting hourly precipitation during tropical cyclone (typhoon) events. The two SVM-based models are the traditional Gaussian kernel SVMs (GSVMs) and the advanced wavelet kernel SVMs (WSVMs). A comparison between the fifth-generation Pennsylvania State University–National Center for Atmospheric Research (PSU–NCAR) Mesoscale Model (MM5) and statistical models, including SVM-based models and linear regressions (regression), was made in terms of performance of rainfall prediction at the Shihmen Reservoir watershed in Taiwan. Data from 73 typhoons affecting the Shihmen Reservoir watershed were included in the analysis. This study designed six attribute combinations with different lag times for the forecast target. The modified RMSE, bias, and estimated threat score (ETS) results were employed to assess the predicted outcomes. Results show that better attribute combinations for typhoon climatologic characteristics and typhoon precipitation predictions occurred at 0-h lag time with modified RMSE values of 0.288, 0.257, and 0.296 in GSVM, WSVM, and the regression, respectively. Moreover, WSVM having average bias and ETS values close to 1.0 gave better predictions than did the GSVM and regression models. In addition, Typhoons Zeb (1998) and Nari (2001) were selected for comparison between the MM5 model output and the developed statistical models. Results showed that the MM5 tended to overestimate the peak and cumulative rainfall amounts while the statistical models were inclined to yield underestimations.