Coupling sPCA-Based Statistical Modeling With Deep Residual Networks Considering Thermal Effect for Deformation Forecasting in High Dams

Accurate prediction of deformation under thermal influences is critical for the safety assessment and long-term performance of high dams. This study proposes a novel two-stage prediction framework that integrates statistical modeling with deep learning to enhance the interpretability and accuracy of dam deformation forecasting. In the first stage, sparse principal component analysis (sPCA) is employed to extract dominant features from high-dimensional thermometer data. These features are then used to construct an interpretable dam deformation monitoring model using multiple linear regression (MLR), referred to as the HT_sPCAT-MLR model. In the second stage, the multilayer bidirectional gated recurrent unit (multi-Bi-GRU) network is developed to model the residuals of the HT_sPCAT-MLR framework, leveraging advanced gating mechanisms and bidirectional temporal learning to improve long-term prediction accuracy. Furthermore, the adaptive genetic algorithm (AGA) is utilized to optimize the hyperparameters of the multi-Bi-GRU model, enhancing the robustness and generalization of the residual correction module. The proposed methodology is validated using real-world monitoring data from an ultra-high arch dam. Quantitative evaluation at four representative measurement points shows that the proposed model consistently outperforms baseline methods across all key metrics. Specifically, it achieves R² values above 0.99, mean absolute error reductions of over 80% compared to traditional models, and the lowest sMAPE across all cases. The experimental results demonstrate model’s superior prediction accuracy, robustness, and practical applicability for dam deformation. The integrated framework offers a reliable and interpretable solution for thermal deformation forecasting in high dam structures.