Author(s): Zhenyue Han; Fawen Li; Yong Zhao
Linked Author(s):
Keywords: Groundwater level Deep learning Spatial-Temporal Graph Convolutional Networks
Abstract: Groundwater level (GWL) prediction is a critical issue in water resource management. With the development of deep learning (DL) models in recent years, many GWL prediction models built using influencing factors such as precipitation have achieved accurate predictions. However, data-scarce areas still face challenges in high-quality meteorological data collection. Given the hydraulic connectivity and spatial correlation of groundwater systems, this study proposes establishing multiple-well GWL prediction models based on historical GWL data from adjacent wells. This study uses the Spatial-Temporal Graph Convolutional Networks (STGCN) model to construct a GWL prediction network, where GWL observation points are represented as nodes and the connections between the points are represented as edges. Meanwhile, the LSTM (both single-well and mulit-well) model is built to compare with the STGCN groundwater level prediction model in terms of computation time and prediction accuracy. The results show that the single-well LSTM model has the best prediction accuracy, followed by the STGCN model, which outperforms the multi-well LSTM model. Although the prediction accuracy of the STGCN model is slightly lower than that of the single-well LSTM model, it has the fastest computation speed and can meet the application needs for rapid multi-well prediction. This study provides new insights for multi-well GWL prediction and serves as a reference for the application of DL models in the field of groundwater.
Year: 2025