基于卷积神经网络的雷电临近预警模型  

Lightning Nowcasting Early Warning Model Based on Convolutional Neural Network

作　　者：张烨方 冯真祯 刘冰 ZHANG Yefang;FENG Zhenzhen;LIU Bing(Fujian Key Laboratory of Severe Weather,Fuzhou 350001;Fujian Meteorological Disaster Prevention Technology Center,Fuzhou 350001)

机构地区：[1]福建省灾害天气重点实验室,福州350001 [2]福建省气象灾害防御技术中心,福州350001

出　　处：《气象》2021年第3期373-380,共8页Meteorological Monthly

基　　金：福建省科技厅社会发展引导性(重点)项目(2019Y0063);福建省气象局研究型业务专项项目(2020YJ08)共同资助。

摘　　要：从研究人工智能雷电临近预警模型的目的出发,以卷积神经网络模型为基础,结合多个时间序列的雷达产品(组合反射率、液态水含量、回波顶高)与闪电数据,对雷电临近预报方法进行基于卷积神经网络结构的应用,以福建省2017—2018年雷达、闪电数据为样本完成了模型的训练与预测研究。训练结果显示,15~30 min模型训练样本测试集准确率为0.7985;选取福建省2019年20个雷电过程验证分析表明,15~30 min模型对动力抬升型雷电过程预警TS评分为0.716,夏季局地热雷暴预警TS评分为0.694,与常规采用雷达、闪电阈值控制的雷电预警算法相比,准确率有一定的提高,具有一定的实践意义。For the purpose of studying the lightning nowcasting early warning model of artificial intelligence,by relying on the convolutional neural network model and combining the radar data(MCR,VIL,ET)and lightning data of multiple time series,we conduct the application of the lightning nowcasting prediction method based on the structure of convolutional neural network.In addition,taking the radar and lightning data of Fujian Province in 2017 and 2018 as samples,we also finish the training and prediction research of the model.The training results show that the test set accuracy of 15-30 min model training samples is 0.7985.The verification analysis of the 20 lightning processes in Fujian Province in 2019 indicates that the TS score of the 15-30 min model for the nowcasting early warning of the dynamic-lift lightning process is 0.716,and the TS score of the localized thermal thunderstorm nowcasting warning in summer is 0.694.Compared with the conventional lightning warning algorithm which uses radar and lightning threshold control,these values have a certain improvement in accuracy,so having certain practical significance.

关 键 词：卷积神经网络 雷电临近预警 人工智能 

分 类 号：P456[天文地球—大气科学及气象学] P457