Abstract: Based on daily precipitation data from 342 stations in Gansu province and ECMWF＿ERA5reanalysis data from 2008 to 2019,the circulation characteristics of different types of rainstorm weather events were summarized.A “batching method” for rainstorm prediction was established that considers multiple sampling at each forecast time for rainstorm days and was precise to model prediction at each grid point.The method was validated using 5 rainstorm events in 2020.The results showed that heavy rain in eastern Gansu can be classified into three types:mesoscale disturbance at the edge of the subtropical high,cold trough shear,and northward convection.Due to the differences in the circulation characteristics of the different types of heavy rain,the precipitation range and duration as well as the key physical quantities and thresholds that represent them also vary.The grid-based batching method improves the process prediction accuracy by 20% and the site prediction accuracy by 7.8%,compared to ECMWF＿THIN.The performance of the batching method is better than that of the ECMWF＿THIN model for heavy rain forecasting.The batching method is especially effective for small-area prediction,and the accuracy of the forecast decreases as the range of the forecast increases.