基于循环神经网络的水沙过程预测方法及其应用——以黄河下游为例  

Prediction method of runoff and sediment processes based on RNN and its application:a case study of the Lower Yellow River

作　　者：徐睿霖 夏军强[1] 李志威 缪驰远 周美蓉[1] 陈翠霞 鲁俊 XU Ruilin;XIA Junqiang;LI Zhiwei;MIAO Chiyuan;ZHOU Meirong;CHEN Cuixia;LU Jun(State Key Laboratory of Water Resources Engineering and Management,Wuhan University,Wuhan 430072,China;Faculty of Geographical Science,Beijing Normal University,Beijing 100875,China;Yellow River Engineering Consulting Co.,Ltd.,Zhengzhou 430003,China)

机构地区：[1]武汉大学水资源工程与调度全国重点实验室,湖北武汉430072 [2]北京师范大学地理科学学部,北京100875 [3]黄河勘测规划设计研究院有限公司,河南郑州450003

出　　处：《武汉大学学报（工学版）》2024年第5期571-581,共11页Engineering Journal of Wuhan University

基　　金：国家自然科学基金项目(编号:U2243237;U2243238);国家重点研发计划项目(编号:2023YF3209304)。

摘　　要：精准快速地预测水沙过程对于高效制订防汛应急方案具有重要意义。选取可记忆时段信息的循环神经网络算法,针对黄河下游游荡河段,按照进口花园口站来水量及来沙量对各年份数据进行分类,分别构建了适用于不同来水来沙类型的循环神经网络水沙预测模型用以预测高村站的水沙过程。模型训练完成后,输入花园口站的水沙数据即可输出对应时段内高村站的水沙数据预测值。预测结果表明:1)数据集预处理时,按照河段进口来水来沙类型进行划分可以提高预测精度,相较不划分方式,洪峰与沙峰的预测精度最高可提升50%以上;2)预测的水沙过程与实测数据符合良好,纳什效率系数(Nash-Sutcliffe efficiency coefficient,NSE)最优分别可达到0.99与0.92,年内汛期与非汛期流量过程的预测效果相近,NSE为0.97左右,而含沙量过程汛期预测结果的NSE为0.88左右,优于非汛期;3)循环神经网络模型对流量过程的预测精度能够达到甚至超过马斯京根法,并且可以弥补马斯京根法不能预报含沙量过程的不足。总体来看,循环神经网络水沙预测方法的精度较高,适用于黄河下游游荡段不同类型水沙过程的预测。Accurate and fast prediction of runoff and sediment processes is crucial for the efficient formulation of flood control schemes.A recurrent neural network(RNN)algorithm was selected in this study,which can memorize timing information.The study applied the RNN algorithm to the braided reach of the Lower Yellow River.Data of different years were classified based on the flow and sediment from the entrance station named Huayuankou Station.According to the classification result,respective RNN runoff-sediment processes prediction models suitable for different types of inflows of water and sediment to predict the runoff and sediment processes of Gaocun Station.After the model training was completed,inputting the runoff and sediment data at Huayuankou Station could output the predicted runoff and sediment data at Gaocun Station in the corresponding period.The prediction results show that:1)Pre-processing the datasets by classifying the data according to incoming runoff and sediment types can improve the prediction accuracy.Compared with the non-classification method,the prediction accuracy of flood peak and sediment peak may be improved by more than 50%.2)The predicted runoff and sediment sequences are in good agreement with the measured data,and the optimal NashSutcliffe efficiency coefficient(NSE)can reach 0.99 and 0.92,respectively.The prediction effect of the annual flow process in flood season and non-flood season is similar,with NSEs around 0.97.In contrast,the prediction accuracy of sediment content process is better during the flood season(with the NSE around 0.88),as compared with that during the non-flood season.3)The prediction accuracy of the RNN model for the runoff process can reach or even exceed the Muskingen method,and can make up for the deficiency of the Muskingen method in predicting sediment process.Overall,the RNN runoff-sediment prediction method exhibits high accuracy,and it is suitable for predicting different types of runoff and sediment processes in the braided reach of the Lower Yellow River.

关 键 词：水沙过程预测 机器学习 循环神经网络 游荡段 黄河下游 

分 类 号：P338[天文地球—水文科学] TV121[水利工程—水文学及水资源]