基于CNN和RF算法的ECMWF降水分级订正预报方法

ECMWF Precipitation Grading Correction Forecast Method Based on CNN and RF Algorithm

利用卷积神经网络(CNN)和随机森林回归模型,提出了一种新的欧洲中期天气预报中心(ECMWF)降水订正预报方法。该方法首先根据ECMWF模式对站点雨量预报值所属的等级进行划分,再计算出不同等级相对应的高相关因子矩阵。进一步利用CNN模型对高相关矩阵进行综合特征提取的学习和训练。最后对CNN模型最终输出的特征因子中,选取若干个与预报站点相关性高的特征,并与ECMWF降水量场插值到预报站点的因子一起,作为随机森林回归模型的输入因子进行预报建模。通过对10个预报试验站点未来24h降水量的分级和不分级订正预报试验,结果表明:(1)ECMWF降水量分级订正预报方法的平均绝对偏差和均方根误差分别比利用ECMWF插值到站点的预报方法减小了20%和15%;(2)24h暴雨及以上的降水分级订正预报方法的平均TS评分为0.32,也显著高于EC插值的0.19;(3)与利用同样的预报模型对全样本(不分级)的传统数值预报模式产品订正预报方法相比,本文提出的分级订正预报方法在总体预报精度和暴雨及以上的强降水预报TS评分上均有更高的预报技巧。

A new precipitation correction forecast method for ECMWF is proposed by using Convolution Neural Network(CNN)and Random Forest Regression(RFR)models.In the new method,the grades of rainfall forecast values of stations are divided according to the ECMWF model,and then the high correlation factor matrix corresponding to different grades is calculated.The CNN model is used to train the comprehensive features of the high correlation matrix.Finally,the features highly correlated with the forecast stations and the factors by the ECMWF precipitation field interpolating to the forecast stations are served as the RFR forecast modeling inputs.The graded and non-graded 24 hour precipitation correction forecast experiments for 10 stations are conducted.The results show that the MAE and RMSE errors of the ECMWF model precipitation grading correction forecast method proposed in this paper are reduced by20%and 15%,respectively,when compared with the forecast method of ECMWF rainfall forecast field interpolating to stations.Meanwhile,the TS(Threat Score)of 24 hour rainstorms at 10 stations is 0.32,which is significantly higher than that of the EC interpolation(0.19).In addition,compared with the traditional numerical prediction model product correction method using the same forecast model for the full sample(non-graded),the grading correction forecast method proposed in this paper shows a higher forecast sill in the overall forecast accuracy and the heavy rainfall forecast TS score.