Abstract: Numerical forecast of rainfall is an important means to prevent disasters caused by extreme weather, and its forecast accuracy is the basis for the application of forecasting products. In this study, taking the Bashan Reservoir watershed as the research object, the control forecast and ensemble forecast of ECMWF, NCEP, CMA, JMA and UKMO in TIGGE are selected. From the view of rainfall classification forecast test, rainfall forecast test, as well as Brier score and Talagrand distribution, the accuracy of control rainfall forecast and ensemble rainfall forecast from May to September in 2007 to 2019 is analyzed. The results show that: (1) For the rainfall classification forecast, there is a little difference between the control forecast and the ensemble average forecast of the five forecasting models for the no rain, light rain and moderate rain; for the heavy rainfall, the control forecast accuracy of JMA and NCEP is the best, and the ensemble average forecast accuracy of JMA and UKMO is the best; for the heavy rain and above, the control forecast accuracy and ensemble average forecast accuracy of NCEP are the best. (2) For rainfall forecast, the control forecast accuracy and ensemble forecast accuracy of UKMO are the best. (3) The Brier score of no rain is the largest, and with the increase of rainfall grade, the Brier score decreases, indicating that the Brier score is better in extreme weather; the Brier score of CMA is the best in the no rain, light rain, moderate rain and heavy rain. (4) The Talagrand distributions of the 5 forecasting models all show a rough “U” shape, indicating that the ensemble members of the 5 forecasting models are not dispersed enough. Generally speaking, the forecasting accuracy of UKMO and NCEP is the best. Based on this study, high-precision rainfall forecasting products can be selected.