摘要
洞庭湖是长江中游的重要湖泊,准确模拟其各输入、输出站点的径流响应关系,对湖区防汛抗旱和生态保护至关重要。针对复杂水力连接下洞庭湖流域多站点径流过程的时空非线性关联特性,本研究提出了一种基于图神经网络的多输入-多输出径流响应模型。首先,利用长江、洞庭湖和四水的流域拓扑空间结构,将各站点的原始观测序列转化为图结构数据,以表征多站点之间的空间关联特性;然后,通过互相关分析法研究各站点观测变量之间的时滞关系,确定模型的输入特征步长;最后,利用图神经网络对数据中的站点特征进行聚合与更新,以捕捉关键控制站点间的复杂时空依赖性,提高多站点径流模拟的准确性和可靠性。结果表明:在洪水事件中,图神经网络的纳什效率系数和平均绝对误差相比前馈神经网络和长短期记忆神经网络模型均提高5%以上,且相关性系数均超过0.97;在枯水断流事件中,召回率和精度普遍超过0.96。图神经网络在洪水和枯水断流等水文事件模拟方面具有明显优势,可为洞庭湖防汛抗旱和生态治理提供科学依据。
Lake Dongting,located in the middle reaches of the Yangtze River,is a crucial body of water.Accurately modelling the relationships between the runoff and the various input and output stations is essential for regional ecological protection,flood control,and drought defense.In order to address the complex relationships between runoff and the various input and output stations in the Lake Dongting Basin,this study proposes a multiple-input and multiple-output runoff response model based on graph neural networks.Firstly,the model utilizes the topological spatial structure of the Yangtze River,Lake Dongting and Sishui Basins to transform the original observation sequences at each station into graph-structured data,thereby characterizing the spatial features of the basins.Secondly,the mutual correlation analysis method is used to identify the time lag relationship between the observed variables at each station,determining the input feature step of the model.Finally,graph neural networks are employed to aggregate and update the features,capturing the complex spatial and temporal dependencies among the control stations and enabling runoff simulation at multiple stations.The results show that,compared with backpropagation neural networks and long short-term memory neural networks,the graph neural network(GNN)model can achieve improvement rates of over 5%for the Nash-Sutcliffe efficiency coefficient and mean absolute error indicators.The correlation coefficient is also greater than 0.97.In dry water cutoff events,the true positive rate and precision are generally greater than 0.96.GNNs have significant advantages in simulating hydrological events such as floods and droughts,and can provide scientific support for the ecological protection of Lake Dongting and its flood control and drought resistance measures.
作者
韦溢龙
周研来
罗宇轩
Wei Yilong;Zhou Yanlai;Luo Yuxuan(State Key Laboratory of Water Resources Engineering and Management,Wuhan University,Wuhan 430072,P.R.China)
出处
《湖泊科学》
北大核心
2026年第1期328-338,共11页
Journal of Lake Sciences
基金
国家重点研发计划项目(2021YFC3200303)资助。
关键词
洞庭湖
图神经网络
径流响应模型
多输入-多输出
时空关联特征
Lake Dongting
graph neural network(GNN)
runoff response modelling
multiple-input and multiple-output
spatio-temporal correlation
