Abstract: An ensemble model involving different impact factors (input vectors) based on support vector regression (SVR) is put forward to improve runoff prediction accuracy and generalization ability. The runoff at Nanpanjiang west bridge station in Yunnan from 1961 to 2007 is taken as a case study. First, a number of impact factors for annual runoff forecast are selected to build different models for the study of a single instance of SVR, and the corresponding RBF models are built as a comparison. In subsequence, the results of single models (which are accurate and complementary) are integrated by using weighted average and simple average respectively. Results showed thatthe average relative absolute error of weighted average and simple average ensemble model based on SVR was respectively 1.27% and 1.54%, and the maximum relative absolute error is 2.99% and 2.74%. The accuracy and generalization capabilities are significantly superior to the single models as well as the weighted average and simple average ensemble model based on RBF models. The weighted average ensemble model based on SVR has better accuracy and generalization capability than simple average because it gives more weight to the models with good prediction result.

Key words: runoff forecasting, ensemble model, SVR, weighted average, simple average