Abstract: The goal of time downscaling for numerical model precipitation forecast is to downscale the model precipitation forecast with lower time resolution to higher time resolution through statistical methods, which serve as an approximate substitution of model precipitation forecast with higher time resolution. Currently, the steps of split by point with cubic spline interpolation, position calibration, optical flow dynamic reconstruction, frequency matching, and total amount constraint by point (hereinafter referred to as"existing method") are applied to the existing time downscaling method for grid precipitation forecast. However, when applying them to uncalibrated numerical model precipitation forecast, the calibration steps are limited by the total precipitation amount constraint, leading to obviously unreasonable fluctuation in intensity for the generated precipitation forecast time downscaling product. Therefore, an improved time downscaling method for numerical model precipitation forecast is designed (hereinafter referred to as"improved method"). The"improved method"includes the steps of split by point based on piecewise cubic Hermite interpolation, pyramid LK optical flow dynamic reconstruction, and balanced total amount constraint by point, which does not depend on calibration with precipitation observation and retains the systematic deviation of the original model. In this study, the 1 h precipitation forecast field and the cumulative 3 h precipitation forecast field of the CMA-SH9 model from April to October 2020 were used as the validation field and input field of the numerical model precipitation forecast time downscaling method, respectively. The effects of time downscaling methods of simple split by point, "existing method", and"improved method"were evaluated, compared, and verified by the ideal model of a circular rain belt with constant speed and real case analysis. The results are as follows. (1) The 1 h precipitation forecast generated by the"improved method"is more balanced and reasonable in intensity distribution. The amplitude of fluctuation of average TS and BIAS for short-time heavy precipitation is only 27% and 9% of the"existing method", respectively. (2) The 1 h precipitation forecast generated by the"improved method"is more similar to the 1 h precipitation forecast field generated by the real numerical dynamic model. The average TS and BIAS improved by more than 13% and reduced by more than 31%, respectively, and the root mean square error generally reduced by 10% compared with the"existing method"in the ideal model. The improved numerical model time downscaling method for precipitation forecast meets the basic requirements of reasonable movement, reasonable intensity fluctuation, and total precipitation constraint at the same time, and can be generally used for time downscaling of any grid precipitation forecast.