A deep learning-based method for enhanced prediction of long-term diaphragm wall deflections using multi-source data

Abstract Machine learning-based autoregressive models are widely applied to deformation prediction of braced excavations, yet they often suffer from pronounced error accumulation and deteriorated accuracy in long-term forecasting. To address these limitations, this study proposes a multi-source information fusion framework for predicting diaphragm wall deformation that integrates construction information with historical wall-deflection data. By incorporating construction-related dynamic features as auxiliary inputs, the proposed model effectively strengthens long-term dependency learning and reduces the propagation of autoregressive errors. A depth-weighted loss coefficient is further introduced to prioritize accuracy in shallow regions where deformation is most critical, and a 1D U-Net decoder is employed to enhance profile-level reconstruction through multi-scale spatial feature extraction. The framework is validated using two deep excavation projects on Nanjing Metro Line 11. For excavation-induced deformation, the model demonstrates stable and accurate multi-step prediction performance for diaphragm wall deflection, even under long forecasting steps. Comparative analyses further confirm that the proposed fusion architecture provides superior long-term robustness under limited monitoring data. The results highlight the method’s potential as a practical tool for deformation forecasting and construction risk management in foundation pit engineering.