Taking Gelei station of the Qingshui River in Yunnan for example, the paper researched the cross validation (CV) SVR prediction model of annual runoff. It chose the influence factor of annual runoff and determined the input vector by using SPSS software; CV method was used to search the SVR penalty factor and kernel parameter, and constructed the prediction model of annual runoff of multivariate CV-SVR, and GA-BP model, traditional BP model as a comparable model. The results show that the absolute value of the average relative error and the maximum absolute value of relative error of prediction of 15 annual runoff example by CV-SVR model are 3.4596%, 9.3035% respectively. the prediction accuracy and generalization ability of CV-SVR are better than that of GA-BP and traditional BP model. CV method can effectively search the SVR penalty factor and kernel parameter. CV-SVR model has many characteristics of high forecast precision, strong generalization ability and stable algorithm method.