A TCN-LACA-GRU Model for Interpretable Surface Subsidence Prediction with Rainfall Lag Effects

Surface subsidence poses a persistent hazard in coastal and reclaimed areas, where long-term ground deformation can jeopardize buildings, roads, pipelines, and other infrastructure. In these environments, precipitation serves as a significant external driver, as rainfall infiltration can modify groundwater conditions, pore-water pressure, and soil consolidation. However, the impact of rainfall on deformation is seldom immediate; it is typically nonlinear, cumulative, and delayed, complicating accurate predictions. Traditional statistical methods and many data-driven models often overlook rainfall or treat it merely as a secondary input, thereby constraining both predictive performance and physical interpretability. To address this limitation, this study introduces a TCN-LACA-GRU framework for short-term multi-step surface subsidence prediction. The model employs a Temporal Convolutional Network to extract local rainfall features, a Lag-Aware Cross-Attention module to account for the varying influence of antecedent rainfall on current deformation, and a GRU predictor to forecast future subsidence values. This methodology is applied to the Shanwei campus of South China Normal University, utilizing Sentinel-1 deformation observations and NOAA precipitation data from the Shanwei station. Results demonstrate that the proposed model significantly outperforms traditional baselines and achieves accuracy comparable to the best pure sequence deep-learning model. Notably, the attention weights reveal that the effects of rainfall are concentrated in earlier periods rather than at the current time step, aligning with the delayed infiltration and consolidation processes. The framework therefore provides both strong predictive performance and a more interpretable understanding of rainfall-driven subsidence evolution.