A BiGRU-Based Spatiotemporal Prediction Method for Surface Displacement in Coal Mining Areas Considering Multiscale Seasonal Characteristics

Accurate prediction of surface displacement in coal mining areas is essential for enabling intelligent early warning systems and effective risk mitigation of mining-induced geological hazards. However, current deep learning models often fall short in capturing the complex and dynamic characteristics associated with seasonal displacement patterns. To address this limitation, this study presents a bidirectional gated recurrent unit (BiGRU)-based spatiotemporal prediction framework that incorporates multiscale seasonal features, aiming to enhance displacement forecasting accuracy in coal mining regions. The proposed approach begins by integrating the time-series decomposition module from the autoformer model to separate interferometric synthetic aperture radar-derived displacement sequences into trend and seasonal components, thereby enabling targeted modeling of long-term trends and periodic fluctuations. Given the complex periodic structures and nonlinear volatility exhibited by the seasonal component, a multiscale feature generation strategy is employed to extract both short-term (microscale) and long-term (macroscale) fluctuation patterns. These features are then fused to construct composite input representations that reflect complex spatiotemporal evolution patterns, thereby improving the model capacity to capture seasonal variability. For the trend component, an autoregressive integrated moving average model is utilized owing to its characteristics. In contrast, the BiGRU model is adopted for the seasonal component to learn nonlinear and intricate temporal features. The proposed method is validated using two representative mining sites with distinctly different operational histories. Experimental results demonstrate that the method achieves robust predictive performance across displacement scenarios with substantial variation, yielding reductions of at least 30.2% in MAE, 6.8% in RMSE, and 29.8% in MAPE. This work demonstrates the ability of the model to capture complex spatiotemporal dynamics in seasonal displacement, offering a promising solution for high-precision displacement prediction that supports real-time mine monitoring and risk management.