In order to make the numerical forecast products achieve the best application effect in Huaihe River Basin meteorological operation, and to provide decision support for flood control, the forecasting performances of four numerical weather predictions (EC, JMA, WRF and INCA) for 24 h, 12 h, 6 h, 3 h and 1 h time-scale rainstorms in Huaihe River Basin during flood seasons from 2015 to 2017 were evaluated by means of TS score, false alarm rate and missing forecast respectively. The lead time of usable forecast of four models for various time scale rainstorms was compared and analyzed as well. On this basis, EC precipitation forecasting with 24 h time resolution was selected to drive distributed hydrological model CREST and construct a meteorological-hydrological coupling flood forecasting model for the upper reaches of the Huaihe River. Moreover, the real-time operation results of June-August 2016 and May-August 2017 were evaluated, and the predictability of flood in the Huaihe River basin was discussed. The main conclusions are as follows: (1) EC has the best forecasting performance and the longest usable time for 24 h and 12 h rainstorms. When the time scale of rainstorm is 6 h, the WRF’s forecasting performance and usable time are both better than EC. When the rainstorm occurs within 3 h, the advantage of WRF is more obvious. JMA has the worst forecast performance for all time scale rainstorms. （2） EC and WRF show abnormally high missing forecast rate and low TS when the lead time is less than 3 h. So the two NWPs cannot provide valuable reference for rainstorm nowcasting. However, INCA shows perfect performance in rainstorm nowcasting with TS score of 54% in 1 h forecast, and the lead time of usable forecast is 3 h. （3） The performances for 6 h and 3 h time scale rainstorm forecasting of EC, JMA and WRF have obvious features of diurnal variation. （4） The lead time of usable forecast for flood forecasting in the upper reaches of Huaihe River is 108 h.