客观定量预报中降水的预处理

Preprocessing Precipitation in Objective Quantitative Forecast

建立预报模型前,对降水量进行一定的处理会对预报效果有较大的影响。对于降水量为0的样本,根据对应的相对湿度情况分别赋予0或不同的负值,并通过神经网络方法,以中国国家气象中心T213模式、德国气象局业务模式和日本气象厅业务模式相应的降水量预报结果作为预报因子,利用2003年和2004年夏季资料分别建立了处理后降水量以及未经处理降水量的预报模型。以北京等站为例,2005年夏季试报结果的对比分析表明:通过相对湿度对降水量进行适当处理后,预报结果从TS评分、空报率、漏报率及预报偏差来说,不论是与不进行处理的预报结果还是与模式直接的预报结果相比都有提高,尤其是减少了空报的情况。该处理方法简单可行,并且对降水预报效果提高明显。

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 Meteorological 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.