Multilayer Residual Spatiotemporal Graph Network Integrating Cross-Modal Fusion and Hydrological Lag for Dam Displacement Prediction

Abstract

Accurate prediction of dam deformation is crucial for disaster prevention and ensuring engineering safety. However, traditional methods are limited when predicting dam deformation with complex spatiotemporal dynamics and multiscale features due to the nonlinear effects of factors like water level and temperature. This study introduces a novel model based on spatiotemporal graph networks called Multiscale Cross-Modal HydroLag Residual Spatiotemporal Graph Network (MC-HRSTGN). The MC-HRSTGN model (1) employs a hydrological lag-aware dynamic graph module to adaptively capture spatiotemporal correlations across multiple monitoring points driven by water level changes; integrates a multiscale learning module to identify deformation dependency patterns from short-term fluctuations to long-term trends; (2) features a cross-modal attention mechanism to combine heterogeneous data from deformation and environmental sensors, enhancing feature representation; and (3) uses a multilayer residual spatiotemporal graph convolutional network to improve deep feature learning and mitigate the vanishing gradient problem. The model was applied to forecast the displacement of the Shiziping rockfill dam in Sichuan, China. Results demonstrated that MC-HRSTGN surpasses models such as multilayer perceptrons (MLP), bidirectional long short-term memory (BiLSTM), gated recurrent units (GRU), and spatiotemporal graph convolutional network (STGCN) in accuracy and robustness. Ablation experiments verify the significance of the hydrological lag-aware dynamic graph, the multiscale framework, and the cross-modal attention. This study provides an advanced and reliable method for predicting rockfill dam displacement and disaster prevention and mitigation.