In recent years, SVM (support vector machine) has been widely used in meteorological field. Single modeling is the most common approach for this type of application which just looks for a large, generic prediction mode to forecast surface rainfall. In this study, individual meteorological stations were modeled dynamically through multitime scale SVM. So we established a dynamic shortterm rainfall forecasting model and fully considered the difference of meteorological elements at different time stamps of different sites, solving the problem that the single fixed global model is concerned with the whole law too much and neglects the difficiency of local meteorological changes at different sites and different times. Therefore, our method has the ability of improving the accuracy of shortterm precipitation forecast. In our study, the prediction for higher density and finer rainfall in geographical space was basically achieved, the temporal resolution was 1 h, and the TS score was always kept at a high level. As a result, the average TS score of 1 h forecast is more than 40%, and for some sites, it is close to 50%. Thus, the prediction accuracy of the model has certain stability and reference value.