基于图神经网络的地下水位动态模拟模型  

Dynamic Model of Groundwater Level Based on Graph Neural Networks

作　　者：许明家[1] 孙龙[1] 李爽 鲁程鹏[2] XU Mingjia;SUN Long;LI Shuang;LU Chengpeng(Information Center,Ministry of Water Resources,Beijing 100053,China;College of Hydrology and Water Resources,Hohai University,Nanjing 210098,China)

机构地区：[1]水利部信息中心,北京100053 [2]河海大学水文水资源学院,江苏南京210098

出　　处：《水文》2025年第1期30-36,共7页Journal of China Hydrology

基　　金：国家重点研发计划项目(2021YFB3900604,2021YFC3200501)。

摘　　要：地下水位的模拟精度在可持续的地下水资源利用和管理中起着重要的作用。机器学习方法可以捕获输入变量和目标变量之间的非线性关系,在地下水位模拟中得到了广泛的应用。然而,传统的机器学习方法没有考虑站与站之间的空间关系。本文使用图神经网络(GNN)模拟地下水位动态变化,以地下水水位监测站为节点,通过邻接矩阵连接节点;选择河北省典型漏斗区的监测数据对模型进行应用和评价。与三个对照模型:随机森林(RF)、支持向量机(SVR)和多层感知机(MLP)相比,所提出的模型在所定义的评估指标方面均表现更好。此外,所提出的模型可同时模拟建模系统中所有监测站的地下水位变化,相比单站模型具有更高的数据利用率。The precision of groundwater level simulation is crucial for the sustainable utilization and management of groundwater resources.Machine learning method,which is widely used in simulating groundwater level,can capture nonlinear relationships be-tween input and target variables.However,traditional machine learning approaches often overlook the spatial relationships between monitoring stations.In this study,Graph Neural Networks(GNN)was used to model the dynamic changes in groundwater levels,treating each monitoring station as a node within a graph interconnected through an adjacency matrix.A typical groundwater de-pression area in Hebei Province was taken as a study case to apply and evaluate the model.Compared to the other three models(i.e.Random Forest,Support Vector Machine,and Multilayer Perceptron),the proposed GNN model performs better in defined evaluation metrics.Moreover,the GNN model can simulate the groundwater level at all monitoring stations within the system simul-taneously,which is different from single-station models.

关 键 词：地下水位模拟 图神经网络 非平稳 时间序列 

分 类 号：P345[天文地球—水文科学]