Forecasting surface wind speeds over offshore islands near Taiwan during tropical cyclones: Comparisons of data-driven algorithms and parametric wind representations

Tropical cyclones often affect the western North Pacific region. Between May and October annually, enormous flood damage is frequently caused by typhoons in Taiwan. This study adopted machine learning techniques to forecast the hourly wind speeds over offshore islands near Taiwan during tropical cyclones. To develop a highly reliable surface wind speed prediction technique, the four kernel-based support vector machines for regression (SVR) models, comprising radial basis function, linear, polynomial, and Pearson VII universal kernels were used. To ensure the accuracy of the SVR model, traditional regressions and the parametric wind representations, comprising the modified Rankine profile, Holland wind profile, and DeMaria wind profile were used to compare wind speed forecasts. The methodology was applied to two islands near Taiwan, Lanyu, and Pengjia Islets. The forecasting horizon ranged from 1 to 6h. The results indicated that the Pearson VII SVR is the most precise of the kernel-based SVR models, regressions, and parametric wind representations. Additionally, Typhoons Nanmadol and Saola which made landfall over Taiwan during 2011 and 2012 were simulated and examined. The results showed that the Pearson VII SVR yielded more favorable results than did the regressions and Holland wind profile. In addition, we observed that Holland wind profile seems applicable to open ocean but unsuitable for areas affected by topographic effects, such as the Central Mountain Range of Taiwan.