基于卷积神经网和SVM雷电监测预警  

Lightning nowcasting and warning based on convolutional neural network and SVM

作　　者：王志斌[1] 肖艳姣[1] 王珏[1] 吴涛[2] WANG Zhibin;XIAO Yanjiao;WANG Jue;WU Tao(Hubei Key Laboratory for Heavy Rain Monitoring and Warning Research,Institute of Wuhan Heavy Rain,China Meteorological Administration,Wuhan 430205,China;Wuhan Central Meteorological Observatory,Wuhan 430074,China)

机构地区：[1]中国气象局武汉暴雨研究所,暴雨监测预警湖北省重点实验室,湖北武汉430205 [2]武汉中心气象台,湖北武汉430074

出　　处：《自然灾害学报》2022年第1期219-225,共7页Journal of Natural Disasters

基　　金：国家重点研发计划“重大自然灾害”专项(2018YFC1507503);国家自然基金项目(42075143);湖北省气象局科研项目(2022Y04)。

摘　　要：利用天气雷达三维网格拼图资料,并针对正负样本数据不平衡问题,使用SMOTE方法进行处理,通过设计多层卷积神经网络,在网络末端加入SVM判别器,实现对雷电监测;使用改进的变分光流方法反演风场,利用半拉格朗日方法对雷达回波进行外推,使用设计的卷积神经网络进行雷电预报;实验表明通过卷积神经网络识别雷电的准确率为76.4%,外推30 min的预报准确率为54.0%,比传统的机器学习方法预报准确率有一定提高,对业务应用有一定的帮助。Based on the 3 D mosaic data of weather radar,the SMOTE method is applied to resolve the imbalance of positive and negative sample data,then a multilayer convolutional neural network is designed with adding SVM discriminator at the end of the network to identify and monitor the lightning.Moreover,the wind field is retrieved by an improved variational optical flow method,and the radar echo is extrapolated by the semi-Lagrangian method,after that the lightning is predicted by the designed convolutional neural network. Our results show that the accuracy of identifying lightning by convolution neural network is 76.4%,and the prediction accuracy of lightning with extrapolation of 30 minutes is 54.0%,which is higher than the prediction accuracy by traditional machine learning methods. This study shows that convolutional neural network is useful for lightning monitoring and prediction.

关 键 词：卷积神经网 变分光流 SVM 天气雷达 SMOTE 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构] X16[自动化与计算机技术—计算机科学与技术]