Abstract: This paper presents an annual runoff forecasting model using the heuristic algorithms of PSO-SVR and GA-SVR to overcome the practical difficulties in selection of the penalty factor and kernel parameters of support vector machines for regression (SVR) and to reduce the influence of selection on model performance, including a case study of the Luoque station in Yunnan. This method first determines the input vector by selecting the impact factor of annual runoff using the SPSS code, and then constructs PSO-SVR and GA-SVR-multivariate models of annual runoff forecasting, based on the basic principles of particle swarm optimization (PSO) and genetic algorithm (GA), using the penalty factor and nuclear function parameters of PSO and GA optimization SVR. For comparison, a combined GS-SVR model is constructed with the algorithms of grid search (GS) and cross-validation (CV). Application of these two multivariate models to one practical example shows that the predictions of a runoff series of 16 years by PSO-SVR and GA-SVR (averaging five random predictions) have average relative errors of 2.53% and 2.79%, maximum relative error of 7.11% and 6.64%, and average absolute errors of 0.14 and 0.15, respectively. Their prediction accuracy and generalization ability are better than GS-SVR model, showing that PSO and GA can optimize the penalty factor of SVR and its nuclear function parameters. In addition, the two models are superior in robustness and performance. Relatively, PSO-SVR model is slightly better than GA-SVR.