Abstract: In order to improve the accuracy of runoff forecast and generalization ability, a regression support vector machine (SVR) integrated annual runoff forecasting model is developed based on multivariate combinations, and the annual runoff forecasting of Longtan hydrologic station in Yunnan Province is taken as an example for the case studies. First, the average monthly discharge from January to October is taken as predictor, and correlation coefficients of the predictior and the average annual runoff are determined by a correlation analysis method. Then, the predictor is sequentially selected, according to the correlation coefficients from the maximum to the minimum, to develop nine SVR models for 2-D to 10-D input variables, and annual runoffs of next 12 years are forecasted by the models. At last, the simple average (SA) and the weighted average (WA) methods are applied to forecasting comprehensive integration for seven SVR models with high accuracy. The analysis research results show that: ①the prediction accuracy of SVR model improves significantly with the increase of imput variables dimension; ②for annual runoff predictions of next 12 years, the absolute average relative errors of SA-SVR and WA-SVR models are 1.73% and 1.79%, and the absolute maximum relative errors are 6.34% and 6.47%, which indicates that the accuracy and generalization ability of the SA-SVR and WA-SVR models are better than that of other SVR models, therefore SA-SVR model is better than WA-SVR model slightly.