Predictor selection method for the construction of support vector machine (SVM)-based typhoon rainfall forecasting models using a non-dominated sorting genetic algorithm

Abstract

This study proposes a predictor selection method for constructing a support vector machine (SVM)-based typhoon rainfall forecasting models using a fast elitist Non-dominated Sorting Genetic Algorithm II (NSGA-II). Based on SVMs, four rainfall forecasting models with different combinations of the three types of input variables (i.e. antecedent rainfalls, typhoon characteristics and local weather factors) were constructed for 1–6 hr-ahead forecasting. An application to three rain gauge stations in the Yilan River basin, northeastern Taiwan, was conducted to demonstrate the superiority of the proposed predictor selection method. The results showed that the optimal combination of predictors for each SVM-based rainfall forecasting model can be automatically and effectively determined by the proposed predictor selection method. The rainfall forecasting model using all three types of input variables performed better than the other three models, especially for long lead-time forecasting. The construction of rainfall forecasting models is helpful to extend the lead time of flood forecasting. The optimal rainfall forecasting model can be further integrated with river hydraulic models or flood inundation models for flood forecasting to assist floodplain managers to take suitable precautionary measures during typhoon landfall.