Aimed at the shortage that the best algorithm parameters of support vector machine (SVM) deficiency is difficult to determine, the paper used the simulation annealing algorithm (SA) to search SVM learning parameters and put forward SA-SVM prediction model.Based on comparison between genetic algorithm (GA) and GA-SVM model in searching SVM learning parameters,it took the runoff prediction in dry season of January to march at LongTan station in Yunnan Province as an example.It trained and forecasted the model by using data of the example before 43 years and after 10 years. The results show that the absolute values of average relative errors of prediction runoff in 3 months of dry season by SA-SVM model after 10 years of instance were 3.11%, 4.93% and 6.75%, the accuracy is better than that by GA-SVM model, which showed that SA-SVM model has higher prediction accuracy and generalization ability. The sudden jump time-varying and final trend to zero probability by SA algorithm through giving the search process can effectively avoid the algorithm falls into local extreme and tends to the global optimal.