Abstract: Based on the observations from conventional weather stations and the rainfall forecasts respectively obtained from four numerical weather prediction models of European Centre for Medium-Range Weather Forecasts (ECMWF), Japan Meteorological Agency (JMA), Mesoscale Model Version 5 (MM5) and Weather Research and Forecasting model (WRF) run by the Anhui Meteorological Observatory during flood season from 2011 to 2012, we evaluate the forecasting performances of the four numerical models mentioned above for the objective areal rainfall of 15 sub-basins in Huaihe River Basin. The forecasted areal rainfall of the 15 sub-basins is calculated by the grid arithmetic average method, while the empirical areal rainfall is calculated by Thiessen polygon method. Moreover, the scoring methods used to evaluate the performance include Mean Absolute Error (Ea), Fuzzy Grading (MP), Percentage Correct (PC) and Treat Score (TS). The evaluation results are as follows. 1) ECMWF performs better than other models as a whole; it clearly outperforms over others for light to heavy rain categories. Among the other 3 models, the performance decreases in order of JMA, MM5 and WRF. 2) The performances of the four models all decrease with the increasing grade of precipitation in order of light, moderate, heavy and torrential rain. 3) The performances of the four models all gradually decline with the extension of period of validity (i.e., 24 h, 48 h and 72 h). 4) By analyzing typical rain events, it is found that the grade of areal rainfall forecasted by ECMWF, JMA and WRF overestimate the weak precipitation while they underestimate the strong precipitation. MM5 overestimates both weak and strong precipitations.