一种基于迁移学习和长短期记忆神经网络的降水预报方法  

A precipitation forecast method based on transfer learning and Long Short Term Memory

作　　者：黄天文 焦飞[2] 伍志方[3] HUANG Tianwen;JIAO Fei;WU Zhifang(Zhaoqing Meteorological Bureau,Zhaoqing 526040;Zhaoqing University,Zhaoqing 526061;Guangdong Meteorological Observatory,Guangzhou 510641)

机构地区：[1]广东省肇庆市气象局,肇庆526040 [2]肇庆学院,肇庆526061 [3]广东省气象台,广州510640

出　　处：《暴雨灾害》2024年第1期45-53,共9页Torrential Rain and Disasters

基　　金：国家重点研发计划项目(2019YFC1510400);广东省气象局科学技术研究项目(GRMC2020M21);广东省肇庆市气象局科技创新团队(C202202);广东省气象局科技创新团队(GRMCTD202004)。

摘　　要：为给智能网格强降水预报提供客观参考,提出了一种基于迁移学习和长短期记忆神经网络(LSTM)的降水预报方法。迁移学习是一种机器学习方法,可将源域学习到的知识迁移到目标域中以应用;LSTM是一种可以处理序列数据中的长期依赖关系的深度学习模型。基于2009—2022年广东省肇庆市6个国家气象观测站逐小时雨量、气温、气压、相对湿度、风向和风速的观测资料,以肇庆高要国家气象观测站作为目标域,其它5个国家气象观测站作为源域,利用迁移学习方法对目标域有缺失值的观测资料进行订正,使目标域形成完整的训练样本;然后,利用深度学习方法,对目标域分别建立单变量LSTM日雨量预报模型和多变量LSTM小时雨量预报模型;最后,对目标域2022年日雨量与小时雨量进行预报,与对应实况对比。结果表明:(1)单变量LSTM预报方法在1—2月、6月、10—12月的24 h晴雨预报准确率在80%以上,多变量LSTM预报方法在3月、6月、8月、12月的1 h晴雨预报准确率在80%以上。(2)单变量LSTM预报方法仅6月能预报出24 h雨量在50 mm以上的强降水,多变量LSTM预报方法在3月、5月、6—8月能预报出1 h雨量在20 mm以上的强降水,其中3月和6月的小时雨量预报TS评分高于25%。A precipitation forecasting method based on transfer learning and Long Short-Term Memory(LSTM)is proposed to provide an ob⁃jective reference for intelligent grid heavy precipitation forecasting.Transfer learning is a machine learning method that can transfer knowl⁃edge learned from the source domain to the target domain for application.LSTM is a deep learning model that can handle long-term depen⁃dencies in sequence data and can remember long and short periods.In this study,the hourly observation data(rainfall,temperature,air pres⁃sure,relative humidity,wind direction,wind speed)from 2009 to 2022 of 6 meteorological observation stations in Zhaoqing City is used.The Gaoyao National Meteorological Observatory is selected as the target domain and the other 5 national meteorological observatories as the source domain,and the transfer learning method is used to transfer the source domain and correct missing values in the target domain.Then the complete training samples are classified in the target domain.Then,the deep learning methods are applied to establish the univariate LSTM daily rainfall prediction models and the multivariate LSTM hourly rainfall prediction models for the target domain,respectively.The daily and hourly rainfall forecast in the target domain for the year 2022 is compared with the actual observations.The results are as follows:(1)For the daily precipitation forecast of clear rain,the univariate LSTM method from January to February,June,and October to December can achieve an accuracy of over 80%,while for the hourly precipitation forecast of clear rain,the accuracy of the multivariate LSTM precipi⁃tation forecast method in March,June,August,and December can be over 80%.(2)The univariate LSTM method can only forecast precipita⁃tion with a 24-hour rainfall over 50 mm in June.The multivariate LSTM method can forecast precipitation with a 1-hour rainfall over 20 mm in March,May,and June to August,with the TS score in March and June being higher than 25%.

关 键 词：降水预报 时间序列 迁移学习 LSTM 深度学习 

分 类 号：P457.6[天文地球—大气科学及气象学]