In order to further improve the precipitation forecast skill of the GRAPES Regional Ensemble Prediction System (GRAPES-REPS), this paper uses the high resolution assimilation analysis field of GRAPES-Meso model to generate the deterministic initial condition of GRAPES-REPS by dynamic up-scaling scheme (hereinafter referred to as the GRAPES-M-US scheme). On this basis, a series of 10 d ensemble prediction experiments were carried out. Then the results were compared with the deterministic initial condition which was generated by dynamic down-scaling scheme from the T639 global model assimilation analysis initial filed (hereinafter referred to as T639-G-DS scheme) and the corresponding ensemble forecast results. The test results of the precipitation forecasts were more focused on. The results show that the deterministic initial condition based on the GRAPES-M-US scheme has more small- and mesoscale information at lower levels compared with deterministic initial condition based on the T639-G-DS scheme. Low-level continuous variable forecast skill is better, the root mean square error (RMSE) and continuous ranked probability score (CRPS) of 850 hPa geopotential height and temperature show some improvements, but the improvement of middle- and high-level elements is not significant. The RMSE and CRPS of 10 m wind speed have obvious improvement, and the RMSE and CRPS of 2 m temperature are basically equal. For the precipitation forecast, the TS score, Brier score and AROC of the 24 h forecast are improved to some extent for light rain, moderate rain and heavy rain, and the rest of the forecast lead time is generally equivalent or slightly negative. In the case of severe rainfall on 7 August 2017, the forecast of rainfall area and intensity of severe precipitation show a certain “refinement” and “corrective” effect. In general, the GRAPES-M-US scheme shows a certain advantage over the T639-G-DS scheme, especially in the short-range precipitation forecast.