In order to improve the accuracy and resolution of precipitation forecasts, taking Jiaojiang river basin in Zhejiang province as a case study, a precipitation post-processing model with the deep learning (CNN-LSTM) was proposed by using CMA-CMORPH precipitation grid dataset and ECMWF numerical precipitation predictions. The accuracy of precipitation forecast before and after the post-processing for different lead times were explored and the ability of precipitation forecasts to predict typical rainstorm events was evaluated. Results showed that CNN-LSTM can significantly improve the accuracy of the raw precipitation forecasts, the root mean square error decreasing from 6.0 mm to 3.0 mm, and the coefficient correlation was improved from 0.6 to 0.9. After the post-processing of two typhoon events, the precipitation forecast error of 6-hour accumulated areal rainfall forecasts after post-processing in Jiaojiang river basin were confined within 10%. The TS score during rainy season after post-treatment was generally near 0.90. The TS score of light rain was increased from less than 0.80 to 0.91, and the TS score of moderate rain increased from less than 0.50 to 0.60.