Using the precipitation observations and SWCWARMS forecasts from May to October, from 2014 to 2015, based on the probability matching method, we analyze the characteristics of mean rain forecast deviation. An experiment is performed to correct the 12hour cumulative precipitation forecast from June to August of 2016 by taking the division and pointtopoint matching plans. The results show that: (1) After calibration, the mean (absolute) error reduces, and the rain area and average intensity are more similar to the observed. (2) The effectiveness of correction is more obvious with larger forecast deviation, and the method performs better at night than in the day time. (3) The division plan works better than the other one for the model systematic deviation; that is to say, a reasonable division to increase statistical samples can improve the effectiveness of correction.