机器学习方法在湖南夏季降水预测中的应用  

Prediction of summer precipitation in Hunan based on machine learning

作　　者：黄超 李巧萍[3] 谢益军[1,2] 彭嘉栋 HUANG Chao;LI Qiaoping;XIE Yijun;PENG Jiadong(Hunan Climate Center,Changsha 410118,China;Hunan Key Laboratory of Meteorological Disaster Prevention and Reduction,Changsha 410118,China;CMA Earth System Modeling and Prediction Centre(CEMC),Beijing 100081,China)

机构地区：[1]湖南省气候中心,湖南长沙410118 [2]气象防灾减灾湖南省重点实验室,湖南长沙410118 [3]中国气象局地球系统数值预报中心,北京100081

出　　处：《大气科学学报》2022年第2期191-202,共12页Transactions of Atmospheric Sciences

基　　金：湖南省气象局研究型业务预报预测专项(XQKJ21C011);中国气象局预报员专项(CMAYBY2020-087);国家重点研发计划项目(2018YFC1505806)。

摘　　要：利用湖南97个国家站的逐月降水资料、国家气候中心130项气候指数集以及国家气候中心和美国国家环境预报中心两套季节预测模式的降水预测资料,采用递归特征消除法确定预测因子并使用多层前馈神经网络、支持向量回归和自然梯度提升三种算法建立了两种湖南夏季降水统计预测方案的模型,检验了预测效果。结果表明:基于机器学习的预测模型对湖南夏季雨型分布有较好的预测能力,两种统计方案提前1~6 mon起报的夏季降水平均距平相关系数分别为0.15和0.19,相比于NCEP和NCC模式有较大提升,平均PS评分分别为69.3和69.2,高于NCC模式的63.1,略低于NCEP模式的71.5;进一步分析表明,3—5月起报的机器学习模型的预测技巧可能来源于前冬极地和中高纬环流,12—2月起报的模型预测技巧则可能来自海温的前兆信号。Against the background of global warming,summer extreme precipitation in Hunan has increased significantly.Therefore,improving the prediction accuracy of precipitation is of great practical significance for disaster prevention and mitigation in Hunan Province.Using the monthly precipitation data from meteorological stations in Hunan,the climate index sets from the National Climate Center(NCC)and the precipitation data from the hindcast experiments are performed using seasonal prediction models of NCC and NCEP(National Centers for Environmental Prediction).The recursive feature elimination(RFE)method is used to determine the key factors,and two statistical prediction schemes of summer precipitation in Hunan are established by three algorithms:multilayer feedforward neural network(FNN),support vector regression(SVR)and natural gradient boosting(NGBoost).The results show that the prediction model based on machine learning(ML)has superior ability to predict the distribution pattern of summer precipitation in Hunan.The respective average ACC skills of the two statistical schemes with lead times of1to6months are0.15and0.19,which is a great improvement compared with the dynamic model.The respective average PS scores are69.3and69.2,which are higher than the NCC model.The further analysis indicates that the preceding winter polar and mid-and high-latitude latitude circulation may be the main predictability sources of ML models with lead times of1to3months.Finally,the prediction skills of models with lead times of4to6months are likely derived from the precursory signal of sea surface temperature.

关 键 词：机器学习 夏季降水 预测 

分 类 号：P457.6[天文地球—大气科学及气象学]