Comparison of SVM and RF for Hourly Typhoon Rainfall Forecasting

Kun-Xiang Lin

This study aims to develop and compare the rainfall forecasting models based on two machine learning (ML) methods, support vector machines (SVMs) and random forests (RFs). Furthermore, an optimization algorithm named non-dominated sorting genetic algorithm (NSGA-II) is applied to construct a optimizing program for searching the optimal predictor set to improve the performance of forecasting. Firstly, four predictor sets: (1) antecedent rainfalls, (2) antecedent rainfalls and typhoon characteristics, (3) antecedent rainfalls and meteorological factors, and (4) antecedent rainfalls, typhoon characteristics and meteorological factors, respectively, were optimized by using NSGA-II to construct 1- to 6-hour ahead rainfall forecasting and the model performances were also investigated. The results reveal that the ML-based models with predictor set #4 as inputs show a significant improvement when compared to predictor set #1 especially for the long lead time forecasting. Finally, the performances of the SVM-based and RF-based model using the optimal predictors from set #4 were further compared. It is found that the RF-based model is superior to the SVM-based model in hourly typhoon rainfall forecasting.

Keywords: hourly typhoon rainfall forecasting, predictor selection, support vector machines, random forests.