In order to overcome the lack of learning parameters of least square support vector machine (LSSVM) which is chosen by human experience, the paper used the genetic algorithm (GA) to select LSSVM penalty factor c and kernel parameter and construct GA-LSSVM annual runoff forecasting model, and set up LSSVM, GA-BP and traditional BP model as the comparison.Taking Yunnan province river hydrological station annual runoff prediction as a case study , it used the data before 30 years and after 22 years to train and predict every model. The results show that the absolute value of the average relative error of annual runoff and the maximum absolute value of relative error predicted by GA-LSSVM model after 22 years are 3.13%, 8.66%, the prediction accuracy of GA-LSSVM model is better than that of LSSVM, GA-BP and traditional BP model. The global optimization ability of GA algorithm is strong. The LSSVM learning parameters gotten by optimization of GA algorithm can effectively improve the prediction accuracy of LSSVM model and the generalization ability.