Abstract: In order to evaluate the influence of the assimilation of different observational data such as conventional ground observations, radiosonde and radar radial wind on the meso-scale model of heavy rain forecast in Sichuan Province, a heavy rainstorm process in Sichuan from 14 to 18 June, 2020 is used as an example. Using Weather Research And Forecasting (WRF) model and Grid Point Statistical Interpolation (GSI) assimilation system, we assimilated the conventional and radar data respectively and simultaneously, and compared the results of three assimilation experiments qualitatively and quantitatively. The results show that the WRF model combined with the GSI assimilation system can simulate the rainstorm well. For the 21-h cumulative precipitation forecast, assimilating conventional observation data can better improve the trend of rain belt and the fall area of the rainstorm. The assimilated radar data showed better performance in precipitation intensity, rainstorm range and the light to moderate rain forecast, The average ETS score of the light to moderate rain was increased by 0.05. Assimilation of both the conventional observation and radar data improved ETS, POD, FAR and BIAS scores for heavy rain. For the12-h cumulative precipitation forecast, the simulation performance of the precipitation trend is the best with the assimilation of radar data, and the experiment involving the assimilation of radar data has better improvement on the precipitation area. For the 3-h cumulative precipitation forecast, the assimilation experiment improved the precipitation evolution, and the assimilation of radar data showed the best performance. The simulation of precipitation at night was generally better than that in the daytime, and the improvement period of assimilation experiment was mainly concentrated in the nighttime, and the assimilation of conventional observation data showed significant performance improvement. Based on the scores of 21-h, 12-h and 3-h cumulative precipitation forecast, the precipitation forecast effect of assimilating multiple data is not absolutely better than those of assimilating only one data, but the assimilation of multiple data can achieve better scores than those of assimilating only one data.