Based on T213 NWP(Numerical Weather Prediction）model outputs and prec ipitation observations, crossvalidation is performed with random samples to fi n d the samples with best predictors and optimal parameters. The forecast models o f precipitation are established at 72 meteorological stations in China by the SV M (Support Vector Machine) statistical method. The models are verified with inde pendent samples. The predictors are selected and the precipitation forecast mode ls are improved by pressing close degree. Forecast experiments show that the imp roved models are better. The precipitation forecasted by SVM models is superi or to the precipitation of T213 DMO (direct model output) in real-time experimen ts.