A CEEMDAN–SVR–PSO-LSTM Hybrid Model for Construction-Induced Displacement Prediction of Metro Deep Excavations

Accurate prediction of retaining-pile displacement is important for deformation control during staged deep-excavation construction, where monitoring series often exhibit pronounced nonstationarity and multiscale temporal variation. To account for the distinct temporal characteristics of the trend and fluctuation components, this study proposes a CEEMDAN–SVR–PSO-LSTM hybrid framework, termed CSPL. Complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is first used to decompose the monitored displacement series into a slowly varying residual and oscillatory intrinsic mode functions (IMFs). Support vector regression (SVR) is employed to predict the trend component, whereas particle swarm optimization (PSO)-optimized long short-term memory (LSTM) is used to predict the fluctuation components. The component-wise predictions are then reconstructed to obtain the final displacement prediction. The proposed model is validated using monitoring data from two Zhengzhou Metro projects. For Case 1, the model achieves average R2, RMSE, and MAPE values of approximately 0.94, 0.33 mm, and 2.8%, respectively, across different monitoring depths, showing better overall predictive performance than BP, EMD-LSTM, and VMD-GRU. For Case 2, external validation using 90 depth-wise monitoring points over six construction stages yields millimeter-level errors, supporting stable predictive performance under different geological and support-system conditions. The results further indicate that the late construction stages and the upper pile segments deserve particular attention in deformation control. The proposed model provides a data-driven tool for construction-stage displacement prediction and deformation control in underground geotechnical engineering.