PEGNet: Integrating Peridynamics and Emergent Features into a Spatio-Temporal Graph Neural Network for Land Subsidence Modeling

Land subsidence is a widespread geological hazard in urban areas, yet accurate prediction remains challenging due to the complex, non-linear spatio-temporal dynamics governing subsurface deformation. In this work, the PEGNet (Peridynamics-Emergent-GAT-GRU Network) method integrates Graph Attention Networks (GAT) with Gated Recurrent Units (GRU) under a three-level embedding strategy to predict land subsidence at monitoring points. Peridynamics gradient and emergent velocity features are fused with multi-source data at the feature level; multi-head graph attention captures spatial dependencies among monitoring points at the structure level; and a Peridynamics-based loss regularizer penalizes unrealistic spatial gradients while an attention entropy regularizer encourages focused attention patterns at the constraint level. Evaluated in Tongzhou District, Beijing over 60 months, the model achieves an RMSE of 5.16 mm, outperforming the GAT-GRU baseline by 25.1% and reducing RMSE by 51.7% relative to GCN-LSTM. Comprehensive ablation studies and baseline comparisons validate each component, showing that the dual-constraint framework effectively suppresses large-magnitude outliers. This work demonstrates that coupling macro-micro priors with spatio-temporal deep learning provides an effective paradigm for integrating physical and geoscience principles into data-driven subsidence prediction.