Explainable deep learning-based dynamic prediction of surface settlement considering temporal characteristics during deep excavation

Abstract

Accurate prediction of surface settlement (SS) induced by deep foundation pit (DFP) excavation is challenging considering the complicated soil distributions and uncertain construction situations. Therefore, a Recurrent Gated Unit (GRU) neural network model incorporated with Variation Mode Decomposition (VMD) was developed for SS prediction. By data denoising with VMD and hyperparameter tuning using Bayesian Optimization (BO), the proposed BO-GRU model achieved dynamic and accurate prediction of SS caused by DFP excavation. Shapley Additive exPlanations (SHAP) analysis was then performed to enhance the interpretability of the BO-GRU model. The presented BO-GRU model was validated with a case study of Wuhan Metro Line 12 Shiqiao Station. The results indicate that: (1) The BO-GRU models with VMD data denoising technique achieved accurate prediction with the average RMSE of 0.0244, MAE of 0.0189, MAPE of 0.0017, R2 of 0.9894, and VAF of 99.1206. (2) The BO-GRU model outperformed the other five state-of-the-art models in predictive performance with robustness improvement of 13.5%, 12.0%, 50.0%, 29.2%, and 27.3%, separately, compared to the other five models. (3) The historical values of the SS make the most significant contributions to the outcomes of the BO-GRU models. In short, this study enhances the accurate and dynamic prediction of SS caused by DFP excavation, contributing to the safe execution of DFP projects.