基于循环神经网络改进雷达定量估测强降水  

Improved radar heavy precipitation estimation based on RNN

作　　者：郑玉[1] 魏鸣[1] 李南[1] Abro Mohammad ILYAS ZHENG Yu;WEI Ming;LI Nan;Abro Mohammad ILYAS(Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters, Nanjing University of Information Science & Technology, Nanjing 210044, China;Education and Literacy Department, Government of Sindh, Karachi, Sindh 74000, Pakistan)

机构地区：[1]南京信息工程大学气象灾害预报预警与评估协同创新中心,南京210044 [2]信德省文化教育部,巴基斯坦卡拉奇74000

出　　处：《中国科技论文》2020年第5期585-592,共8页China Sciencepaper

基　　金：国家自然科学基金资助项目(41675029,41675046)。

摘　　要：为了解决天气雷达传统定量估测强降水有较大偏差的问题,基于循环神经网络(recurrent neural network,RNN)模型,提出了一种有效的雷达估测强降水方法。基于对降水演变规律的认识,重新设计RNN结构,实现了对前3个时次雨量计降水的耦合,从而改善天气雷达估测强降水的效果。利用连续观测的雷达资料,结合前3个时次雨量计对地面降水量进行估测,解决了利用Z-R关系估测强降水时存在较大偏差的问题。利用2015年夏季降水测试数据集进行验证,并与滚动法建立的Z-R关系方法进行了对比,结果表明,在≥30 mm/h的强降水下,RNN估测降水方法相比Z-R关系方法的均方根误差(root mean square error,RMSE)降低了24.28%,中位绝对误差(median absolute error,Median AE)降低了32.83%。因此,循环神经网络估测降水方法对于定量估测强降水效果显著,具有业务应用价值。Aiming at the large deviation of radar heavy precipitation estimation,we propose an effective radar heavy precipitation estimation method based on recurrent neural network(RNN).Based on the understanding of precipitation evolution,we redesigned RNN network structure to merge previous rain gauge measurements in order to improve quantitative precipitation estimation.The weather radar combined with previous three-times rain gauge measurements can effectively reduce large deviations in heavy precipitation estimation by Z-R relationship.In this paper,the RNN and the Z-R relationship method are compared in the 2015 summer precipitation test dataset.The root mean square error(RMSE)of the RNN method is reduced by 24.28%,and the median absolute error(Median AE)is reduced by 32.83%compared to the Z-R relationship by real-time calibration method where rainfall intensity is greater than or equal to 30 mm/h.Therefore,estimating precipitation by RNN method has a significant effect on quantitatively estimating heavy precipitation,and has business application value.

关 键 词：天气雷达 强降水估测 循环神经网络 耦合雨量计模型 

分 类 号：P412.25[天文地球—大气科学及气象学]