By using the precipitation observations (including the encrypted automatic stations) and the ECMWF forecasts for Sichuan during June to August from 2016-2018, based on the probability forecast of short-time heavy rain every 3h derived by the "ingredients method", we calculate the times of various transition probability threshold at each grid, and then a method to improve forecast accuracy of heavy rainfall of 24h cumulative precipitation forecast in Sichuan was developed. We then applied it to daily test of ECMWF model precipitation forecast for 24-72 h on the precipitation concentration period from June to August of 2018. The results show that:1) According to the composite experiment indicators of precipitation above 25 and 50mm, the hit rate, missing rate and TS score of each period validity have been significantly improved after the calibration, and with the extension of the period validity, the value of each index has been further increased greatly. Although the false alarm rate increases a bit of 0-24 h and 24-48 h forecast, the increase of false alarm rate is far less than the decrease of missing rate. 2) According to the experiment of individual cases, the precipitation value increased significantly, the effect of precipitation area forecast above 50mm is greatly improved, which means that the distribution of precipitation area after the calibration is basically consistent with the actual situation, especially for the 0-24h cumulative precipitation forecast.