Based on the TIGGE datasets from the European Centre for Medium Range Weather Forecasts (ECMWF), the United Kingdom Met Office (UKMO), the China Meteorological Administration (CMA), and the Japan Meteorological Agency (JMA), and its multi-center ensemble systems, and observations in the Qingjiang River Basin, Bayesian model averaging (BMA) probability forecast models were established. The results showed that the optimal length of the training period is about 40 days, and the BMA models for multi-center ensemble outperform those for single center system for lead times of 24 h. The mean absolute error (MAE) and continuous ranked probability score (CRPS) skills of the BMA models are improved approximately 11% and 15%, respectively, compared with those of raw ensemble forecasts. In operation, when the BMA 90 percentile predicted precipitation is extreme precipitation ［［50-100) mm·(24 h)-1］, the 75-90 percentiles predicted precipitation could be used as the forecast reference, and the heavy precipitation warning could be carried out. For the forecast of severe precipitation ［［50-100) mm·(24 h)-1］, the forecast result of the 50-75 percentile predicted by BMA can be taken as a reference, while for the general precipitation ［≤25 mm·(24 h)-1］, the reference of BMA deterministic forecast is relatively strong. BMA probability forecast could give both the PDF curve with full probability and the probability greater than a certain precipitation intensity, which could provide the basis for the probability forecast in operation. However, the small probability value is often ignored, resulting in omission. So how to capture more useful information through the probabilistic prediction method and increase the accuracy of the prediction of extreme weather events will be a challenge for the probabilistic prediction technology.