交叉熵神经网络及其在闽北大雨以上降水预报中的应用

基于误差平方和最小化准则的BP神经网络(ANN-MSE)并不适合解决小概率天气事件的预报问题,引进一种改进的以交叉熵函数为目标函数的神经网络方法(ANN-CE),该法是一个三层反向传播神经网络,其输出层只用一个节点.利用2003-2008年的ECMWF预报场资料,把该法用于福建省南平市4-6月部分大雨或以上降水96h预报中,分别用原始因子和PCA降维后的主因子建立了ANN-CE预报模型和ANN-MSE预报模型,用这些模型对2009-2010年独立样本进行了试报.测试结果显示主因子预报模型TS评分比原始因子预报模型高且漏报次数少,其中,主因子ANN-CE预报模型的TS评分和漏报率分别是0.51和0.17,其性能是所有模型中最好且最为稳定的,是一种适合于小概率事件预报的方法.

Application of BP neural network using cross-entropy to 96 hours forecast of heavy precipitation event in northern Fujian province

As a neural network based on MSSE,ANN-MSE is not an appropriate solution to the problem of predicting rare weather event.In this paper,an improved neural network method,ANN-CE is presented,which is a three layered back-propagation neural network with one output unit.The error function of ANN-CE is a cross entropy function.Then utilizing ECMWF forecast fields data,this method was applied to 96 hours forecast of heavy precipitation event in northern Fijian province.The ANN-CE model and the ANN-MSE model based on original factor and principle component after PCA reducing dimensions were respectively built.These models were applied to independent samples in 2009-2010,and the test results are as following:TS grade for model based on principal component is higher than that of model based on original factors;miss-rate for the ANN-CE model is lower than that of the ANN-MSE model.All in all,ANN-CE model based on principal component has best performance and stability,whose TS grade and miss-rate was respectively 0.51 and 0.17,so it was suited for forecasting rare event.