The radar data and GRAPES model data and convective weathers such as thunderstorm wind gusts, hails and tornados simultaneously from surface automatic meteorological stations or human observations from July to October in 2007 and 2008 were collected. Outputs of mesoscale numerical weather prediction (NWP) model of Guangzhou Institute of Tropic and Marine Meteorology of CMA were used to calculate the thunderstorm environment characteristics such as CAPE, winds, temperatures, instabilities etc. NWP variables, radar data together with convective events were incorporated into a statistic development dataset. Based on this dataset and the multiple regression approach, a convective weather potential forcasting method was developed to forecast the probability of a convective cell that could produce convective weathers within 0-1 hour. Scores as POD (probability of detection), FAR (false alarm ratio) and CSI (critical success index) were used to evaluate the performance of the method. The 12 of 31 predictors were chosen into the multiple regression equation. For a threshold value of 0.26, the method for a sample of 5540 cases produces forecast scores as follows: POD=0.73, FAR＝0.61, CSI＝0.338, which are better than results from the early rainy season of Guangdong （April to June） in 2004. This indicates that the convective weather potential forcasting can be a tool to the nowcasting of severe convective weathers in the later flood season of Guangdong.