Abstract: Quantitative precipitation forecast is much difficult in objective forecast. It is because that the precipitation is not a continuous variable, either not a normal distribution variable. Except that the precipitation has an obviously different characteristic to other elements that zero precipitation include many different situations. For example, black clouds blot out the sky but precipitation amount is zero. There are not many clouds on the sky and precipitation amount is also zero. These two situations are different, but the difference is not exhibited in the amount of precipitation. So the effectiveness of the forecast model is affected.Preprocessing has important effects on forecast results. For the purpose of improving quantitative precipitation forecast effectiveness, a reasonable preprocessing scheme needs to be developed. Precipitation observation is preprocessed using relative humidity before modeling. The amount of precipitation is changed to different negative values when relative humidity is smaller than usual and precipitation amount equals zero. Based on the summer data of 2003 and 2004, forecast model of preprocessing precipitation and direct precipitation are developed respectively.BP network method is used in the study, which is a kind of artificial neural network. The BP network is a back propagation network. It contains input layer, implicit layer and output layer. There can be one or more implicit layers on the BP network. Joint on the implicit layer is named implicit node. Input signals propagate to implicit nodes. Then signals of implicit nodes propagate to the next layer after the disposal of weights and operating functions. At last, the value on the output nodes is gotten. BP networks can be considered as a nonlinear projection from input layer to output layer. Sigmond function is always taken as the operating function. Network weights of different nodes are obtained by training. The BP algorithm is used in training process.Predictors are corresponding to precipitation amount forecast of operational global model of China National Meteorological Center, operational models of German Meteorological Administration and Japan Meteo rological Agency. In order to avoid errors caused by interpolating, to every model, precipitation forecast on four grids around the stations are used as predictors instead of interpolated model precipitation forecasts at stations.Test of different forecast results during 2005 summer of Beijing and other 5 stations indicate that precipitation forecasts are improved after preprocessing. Especially absent forecasts are reduced. So it can be concluded that this reprocessing method is simple and effective. But it only provides some ideas. In real settings, different element also can be used to reprocess precipitation amount according to different areas or different seasons in a year.