基于深度学习的改进ERRIS径流预报实时校正模型  

Improved ERRIS model for real-time correction of streamflow forecast based on deep learning

作　　者：刘莉 梁霄 WANG Quanjun 许月萍[1] LIU Li;LIANG Xiao;WANG Quanjun;XU Yueping(College of Civil Engineering and Architecture,Zhejiang University,Hangzhou 310058,China;Faculty of Engineering and Information Technology,University of Melbourne,Melbourne 3052,Australia)

机构地区：[1]浙江大学建筑工程学院,浙江杭州310058 [2]墨尔本大学工程与信息技术学院,维多利亚墨尔本3052

出　　处：《水资源保护》2024年第6期155-164,共10页Water Resources Protection

基　　金：浙江省重点研发计划项目(2021C03017);浙江省自然科学基金项目(LQ22E090004);国家自然科学基金项目(52309038);浙江省自然科学基金重点基金项目(LZ20E090001);中国博士后科学基金项目(2023M733117)。

摘　　要：为提高径流预报精度,基于长短期记忆网络(LSTM)改进ERRIS模型,构建了径流预报实时校正的ERRIS-LSTM模型,以雅鲁藏布江流域和椒江流域为例进行对比分析。结果表明:与ERRIS模型相比,ERRIS-LSTM模型使雅鲁藏布江流域和椒江流域径流预报的纳什效率系数分别提升了4.1%和1.1%,均方根误差分别减小了67.7%和5.7%,使雅鲁藏布江流域中、低水流量的百分比偏差分别降低了75.5%和79.1%,椒江流域低水流量统计指标均改善超过20%;ERRIS-LSTM模型能够充分获取误差序列的序贯相关性,生成的集合预报比ERRIS模型预报的整体精度更高,连续排序概率评分降低了75%以上,不确定性更小,可靠性更强;相比于LSTM模型的校正结果,ERRIS-LSTM模型可以额外提供校正结果的不确定性信息,在业务预报和防洪决策中具有重要的应用前景。In order to improve the accuracy of streamflow forecast,the ERRIS model was improved based on LSTM,and the ERRIS-LSTM model was constructed for real-time correction of streamflow forecast.The Yarlung Zangbo River and Jiao River basins were taken as examples for comparative analysis.The results showed that,compared with the ERRIS model,the ERRIS-LSTM model increased the Nash-Sutcliffe efficiency coefficient by 4.1% and 1.1%,decreased the root mean squared error by 67.7% and 5.7% in streamflow forecast of the Yarlung Zangbo River and Jiao River basins,respectively.Especially for medium and low flows of the Yarlung Zangbo River Basin,the values of percent bias of streamflow forecast obtained by the ERRIS-LSTM model were reduced by 75.5% and 79.1%,respectively,and the statistical indexes of low flow in the Jiao River Basin obtained by the ERRIS-LSTM model were improved by more than 20%.The ERRIS-LSTM model could fully capture the continuity of the error series,and the ensemble forecasts generated by the ERRIS-LSTM model were more accurate,less uncertain,and more reliable than those of the ERRIS model,with the continuous ranked probability score reduced by more than 75%.In comparison with the deterministic corrected results of the LSTM model,the ERRIS-LSTM model can provide additional uncertainty information,which is promising in operational forecasting and decision-making in flood control.

关 键 词：径流预报 实时校正 深度学习 ERRIS模型 LSTM模型 雅鲁藏布江流域 椒江流域 

分 类 号：P338[天文地球—水文科学]