Abstract: To address the limitations of current numerical models and deep learning approaches in land subsidence prediction, this study proposes a spatio-temporal graph convolutional neural network (STGCN) combining graph convolutional networks (GCN) and gated temporal convolutional networks to integrate both spatial and temporal features of land subsidence. The model enables simultaneous prediction of subsidence at multiple monitoring nodes across a study area. Using land subsidence data from Shanghai spanning January 2008 to March 2023 for training and validation, the results demonstrate that the STGCN model achieves a mean absolute error (MAE), root mean square error (RMSE), and normalized root mean square error (NRMSE) of 0.93 mm, 1.05 mm, and 0.49, respectively. These metrics significantly outperform control models that consider only spatial dependencies (GCN and graph attention network, GAT) or temporal dependencies (gated recurrent unit, GRU; long short-term memory, LSTM). The STGCN model presents higher prediction accuracy for nodes with minor subsidence, rebound, or significant settlement. The prediction study based on the STGCN model shows that under the current pattern of groundwater extraction and recharge, the subsidence rates at various nodes in Shanghai from 2023 to 2028 will range from −0.08 to 5.59 mm/a (negative values indicate rebound), with an average subsidence rate of 2.92 mm/a across all nodes. By fusing spatio-temporal features within a deep learning architecture, the proposed STGCN provides a robust scientific framework for precise urban land subsidence forecasting.