A new method for probabilistic precipitation forecasts with dynamic weights is proposed. First, a scoring model for weight distribution is established and the forecast accuracy calculated from radar extrapolation and numerical model prediction is evaluated. Second, an improved method based on the Brier scoring is presented, which takes into account the size of the precipitation area and precipitation, and reduces the sensitivity to the number of samples. Third, according to the scores of different lead times, the weights of two kinds of forecasting methods are dynamically allocated in different forecast lead time. In the experimental part, the Brier and other ratings show that the prediction of each lead time is similar to the radar extrapolation or the numerical model, and even has a higher technical score.