Abstract: A multi-model super-ensemble forecasting system with 15 ensemble members generated based on different cumulus parameterization and planetary boundary layer (PBL) schemes by means of changing the AREM, MM5,and WRF three limitedarea Mesoscale models’ configuration is employed to investigate the potential of multimodel shortrange ensemble forecasting in the Huaihe Valley during the rainy season in China. Weighted average by correlation, multivariate linear regression and support vector machines (SVM) for regression are respectively combined with crossvalidation in this system in this paper to research the weather elements during the rainy season in July in 2003, which mainly contain the 24hour accumulated precipitation and streamline field on 700 hPa. The experiment results compared with the traditional ensemble average of the 15 ensemble members and forecast by the T213 model shows that: (1) for the 24hour accumulated precipitation, the rootmeanssquare error (RMSE) and TS score of the results made by support vector machines for regression is much better than that of the multimodel traditional ensemble average, and also better than that of other two methods; (2) for streamline field on 700 hPa, the first mode and the second mode created by means of vector EOF expanding all the prediction results indicate that the ensemble average decrease mainly the magnitude of wind field and multivariate linear regression and support vector machines for regression can however present a good intensity distribution map of wind vectors with the total effect made by support vector machines for regression being the best among the four means as mentioned above; (3) the results from the multimodel superensemble prediction are obviously superior to the simultaneous results created by the T213 model for the wind speed on 700 hPa so that superensemble results by SVM can be used for dicisionmaking in advance of the T213 model at least 12 hours in the sight of prediction rootmeanssquare error.