基于ConvLSTM的气象雷达回波外推  

Radar echo extrapolation of weather based on ConvLSTM

作　　者：李英睿 李琼[2] 魏加华[2,3] 乔禛 申惟文 LI Yingrui;LI Qiong;WEI Jiahua;QIAO Zhen;SHEN Weiwen(School of Civil Engineering,Qinghai University,Xining 810016,China;State Key Laboratory of Plateau Ecology and Agriculture,Qinghai University,Xining 810016,China;State Key Laboratory of Hydroscience and Engineering,Tsinghua University,Beijing 100084,China)

机构地区：[1]青海大学土木工程学院,青海西宁810016 [2]省部共建三江源生态与高原农牧业国家重点实验室,青海大学,青海西宁810016 [3]清华大学水沙科学与水利水电工程国家重点实验室,北京100084

出　　处：《青海大学学报》2021年第1期93-100,共8页Journal of Qinghai University

基　　金：科学技术部国家重点研发计划项目(2017YFC0403600);清华大学水沙科学与水利水电工程国家重点实验室——宁夏银川水联网数字治水联合研究院联合基金(sklhse-2020-Iow01)。

摘　　要：针对传统雷达回波外推方法存在数据利用率低、预报精度不高等局限,利用定标后的青海大学校内X波段天气雷达监测数据,建立ConvLSTM深度学习模型进行回波外推,并与传统光流法的外推结果进行对比。结果表明:(1)在不同的输入时长和外推时间中,ConvLSTM的精度都明显优于光流法,在第1.5小时内的平均预测误差比光流法低69%。(2)模型输入时间恒定为30 min时,误差随外推时间的增加随之增大;在空间分布中,误差随着高度层的增加呈上升趋势,且回波与700 hPa的气压有较好的相关性。因此,与传统雷达回波外推方法相比,ConvLSTM模型的预测精度高,并且能有效延长外推时长。In order to solve the problems of low data utilization and low accuracy of the traditional radar echo extrapolation,this study establishes a deep learning model ConvLSTM(Convolutional Long Short-Term Memory)by using the calibrated X-band monitoring data of weather radar in Qinghai University so as to compare the results of traditional optical flow extrapolation.The results indicate that:(1)In different input time and extrapolation time,the accuracy of ConvLSTM is much better than that of optical flow method,and the average prediction error within 1.5 h is 69%lower than that of optical flow method.(2)The error increases with the increase of extrapolation time when the input time of the model is constant to 30 min;the error increases with the increase of altitude in the spatial distribution,and the echo has a great correlation with the pressure of 700 hPa.Therefore,compared with the traditional radar echo extrapolation,ConvLSTM can effectively extend the extrapolation time with higher-level prediction accuracy.

关 键 词：临近预报 雷达回波外推 深度学习 ConvLSTM 光流法 

分 类 号：P456.1[天文地球—大气科学及气象学]