A hybrid deep learning model for accurate prediction of subway deep excavation settlement

Abstract Settlement predictions during deep excavation construction are crucial for ensuring engineering safety and structural stability. However, traditional linear regression prediction methods show significant limitations when dealing with complex nonlinear and long-term settlement data, failing to meet high accuracy requirements. To address this issue, this paper proposes a prediction model called DA-BiLSTM-AdaBoost, which combines a data augmentation (DA) technique, a bidirectional long short-term memory network (BiLSTM), and an ensemble learning algorithm, specifically AdaBoost (Adaptive Boosting). This model aims to enhance the accuracy and stability of settlement predictions by combining the bidirectional dependencies of BiLSTM, the diversity expansion of data augmentation, and the robustness of the AdaBoost ensemble algorithm. In an empirical study conducted on the Wuxi Metro, compared with results of the BiLSTM, the proposed model demonstrates superior performance in both single-step and multi-step prediction. Specifically, the root mean square error (RMSE) for single-step, two-step, and three-step predictions are 0.48 mm, 0.65 mm, and 0.79 mm, respectively, representing improvements of approximately 26.2%, 22.6%, and 21.8%. These results indicate that the proposed model exhibits strong adaptability and generalization capabilities, effectively meeting settlement prediction needs in deep excavation construction.