In order to improve the accuracy of grid precipitation forecast, aiming at the phenomenon that there are still much empty precipitation forecasts in the numerical model grid precipitation forecast, based on the live grid precipitation, applying the dynamic modelling and machine training, with the monotonicity of normalized new test parameters, and the computational convenience of the new precipitation TS scoring formula, the dynamic optimal elimination of grid-by-grid precipitation is researched. The research show that the empty precipitation threshold is suppressed in the two-step method, and the process of threshold training selection is more direct in the normalized method. With these two methods, the rain or shine accuracy is all increased. Among them, ECMWF (European Centre for Medium-Range Weather Forecasts) increases by 2.39%-4.76%; the TS score of ECMWF improves the most, increasing 2.98%-3.64% during the day and increasing 1.61%-1.78% at night. However, the TS scores in CMA-SH9 and CMA-BJ decline. The normalized method increases the clear/rain forecast accuracy the most during the day. The analysis shows that the drop in the false alarm rate is significantly greater than the missing forecast rate. As a result, the FAR drops significantly, and the clear/rain forecast accuracy also increases significantly.