Prediction of shield tunneling-induced ground settlement using LSTM architecture enhanced by multi-head self-attention mechanism

Abstract

Numerous advanced deep learning models have been applied to forecast shield tunneling-induced ground settlement to mitigate the adverse impacts of excavation on surface infrastructures. However, most existing studies neglect the spatiotemporal correlations in raw settlement data. This study proposes data-driven LSTM model enhanced by multi-head self-attention (MHSA) mechanisms, which proficiently capture spatiotemporal features and extract vital information from the data. Two different datasets were employed in this study. In Case A, the optimal model architecture was identified, and the performance of MHSA-Bi-LSTM model were compared with SA-Bi-LSTM, Bi-LSTM, GRU, and RF models. Overall, MHSA-Bi-LSTM model exhibited superior performances with the average MSE, and MAE, MAPE values of 0.089 mm, 0.227 mm, 8.02 % in Case A and showcased remarkable generalization in Case B without architecture modifications, achieving the corresponding metrics of 0.213 mm, 0.426 mm, and 16.72 %. Ablation studies revealed that adjacent MHSA layers possessed interdependencies and significantly influenced predictive accuracy. Moreover, the utilization of dropout and sliding window techniques is essential for precise time-series ground settlement prediction.