A method of forecasting rainfall based on numerical prediction products is prese nted. According to the idea of artificial intelligence classification and integr ation, numerical prediction products of T213, Japanese and German models are int egrated together by using the BackPropagation neural network, and it will cont ribute advantages of various means and form an integrated forecast system. On th is basis, selforganizing neural network is used to classify the weather type a ccording to the situation of height field and temperature and humidity on the su rface layer in forecasting area. Then different forecast elements are selected a nd different forecast models are established for different weather types. Using the method mentioned above, the forecast model is built by using the data from M ay to September in 2003 to 2005, and is tested by forecasting rain fall from May to September in 2006 to 2007 at 68 stations in the ChangjiangHua ihe River basin. The result shows that the method is quite practicable.