A modified support vector machine based prediction model on streamflow at the Shihmen Reservoir, Taiwan

Abstract

The uncertainty of the availability of water resources during the boreal winter has led to significant economic losses in recent years in Taiwan. A modified support vector machine (SVM) based prediction framework is thus proposed to improve the predictability of the inflow to Shihmen reservoir in December and January, using climate data from the prior period. Highly correlated climate precursors are first identified and adopted to predict water availability in North Taiwan. A genetic algorithm based parameter determination procedure is implemented to the SVM parameters to learn the non-linear pattern underlying climate systems more flexibly. Bagging is then applied to construct various SVM models to reduce the variance in the prediction by the median of forecasts from the constructed models. The enhanced prediction ability of the proposed modified SVM-based model with respect to a bagged multiple linear regression (MLR), simple SVM, and simple MLR model is also demonstrated. The results show that the proposed modified SVM-based model outperforms the prediction ability of the other models in all of the adopted evaluation scores. Copyright (C) 2009 Royal Meteorological Society