Regional Multi-Well Groundwater Level Forecasting in Sensor-Based Monitoring Networks Using a GCN-LSTM Framework

Groundwater level forecasting is essential for groundwater monitoring and sustainable water resource management in urban regions. This study proposes a graph-based spatiotemporal forecasting framework for regional multi-well groundwater prediction using a Graph Convolutional Network Long Short-Term Memory (GCN-LSTM) model. Daily sensor-derived observations from 56 monitoring wells in Chengdu, China, collected from 2019 to 2023, are used to construct a graph representing inter-well spatial relationships. The GCN component captures spatial dependencies among wells, while the LSTM component models temporal variations in groundwater levels. Two data-availability scenarios are evaluated: a reduced-input setting and a multi-factor setting incorporating rainfall and water temperature. Results demonstrate that the proposed framework provides reliable groundwater level forecasts across multiple wells under both scenarios, and the inclusion of additional temporal variables improves the prediction of groundwater fluctuations. A network-wide controlled comparison under the S+W setting further shows that GCN-LSTM outperforms the conventional LSTM baseline across the 56-well monitoring network, supporting the contribution of inter-well spatial modeling to regional groundwater level forecasting. The findings highlight the potential of graph-based deep learning for regional groundwater forecasting and intelligent sensor-based monitoring networks.