Groundwater level prediction in a karst aquifer using spatio-temporal graph convolutional networks: a case study of Jinan, China

Abstract

Workflow of a spatio-temporal graph convolutional network predicting karst groundwater levels, with a Random Forest ranking driving factors. Accurate groundwater level prediction in karst aquifer systems remains challenging due to their complex hydrogeological characteristics, including rapid recharge responses, preferential flow paths, and strong spatial heterogeneity. This study develops a spatio-temporal graph convolutional network (STGCN) framework specifically designed for the karst spring system in Jinan, China. The framework treats the monitoring network as an interconnected graph, with well-to-well relationships encoded through a hydraulically-informed adjacency matrix. Multiple environmental and anthropogenic driving factors – precipitation, ecological water replenishment, sea-level pressure, pumping, air temperature, and wind speed – are integrated to comprehensively capture karst groundwater dynamics. Three years of daily monitoring data (2021–2023) were utilized for model development and validation. The proposed model achieved a root mean squared error of 0.2077 m, reducing prediction errors by approximately 40% compared to conventional approaches. Feature importance analysis revealed ecological water replenishment (46.1%) and precipitation (37.4%) as dominant drivers of groundwater variations. The model successfully captured both long-term seasonal trends and rapid short-term responses to rainfall events characteristic of karst systems. This framework provides an interpretable and accurate tool for groundwater management in complex karst terrains, supporting sustainable water resource decisions in urbanized karst regions.