Abstract: To enhance the accuracy and generalization of surface final settlement prediction models in shield tunneling, the authors establish a principal component analysis (PCA)-stacking model for surface settlement prediction, integrating the PCA method with a multi-layer Stacking ensemble algorithm. The PCA algorithm processes the extensive data generated during shield tunneling, reducing feature dimensions and extracting critical information. Simultaneously, the multi-layer Stacking algorithm integrates multiple heterogeneous models, improving predictive performance while avoiding optimization comparisons among sub-models. Based on a case study of an ultra-large diameter shield tunnel from the Shanghai Beiheng tunnel project, multi-source data from shield tunneling engineering are analyzed. The performance of the Stacking model before and after PCA processing is compared, and the PCA-Stacking model is benchmarked against random forests (RF) and extreme gradient boosting (XGBoost) models. The research findings are as follows: (1) The R² values of the Stacking model before and after PCA processing are 0.792 and 0.831, respectively, demonstrating that PCA improves the stacking model′s performance. (2) After hyperparameter optimization, the R² values for the RF and XGBoost models are 0.748 and 0.612, respectively, showing inferior performance compared to the PCA-Stacking model without hyperparameter optimization. (3) The PCA-Stacking model exhibits robust predictive capabilities for both ground uplift and subsidence variations. (4) The PCA-Stacking algorithm, employing heterogeneous sub-models, outperforms ensemble algorithms based on homogeneous sub-models in predicting ground settlement during shield tunneling.

Key words: shield tunnel, surface settlement, machine learning, Stacking ensemble learning, principal component analysis