Ground surface settlement caused by shield tunneling is influenced by various factors, complicating predictive efforts. While machine learning models can forecast settlement, their lack of interpretability hinders understanding of input parameters and the implementation of control measures. This study introduces a novel approach that combines ensemble learning algorithms with SHAP (SHapley Additive exPlanations) to predict ground surface settlement and perform both global and local analyses of the predictive model. Selected parameters are adjusted to modify shield-driven variables to effectively control surface settlement. The dataset, derived from a shield tunnel of the Shenzhen Metro, initially contained 146 samples (four of these are data samples pertaining to connection passage). To mitigate small-sample bias and data scarcity, the Synthetic Minority Over-sampling Technique (SMOTE) was employed to generate 142 synthetic samples, expanding the dataset to 288 samples. The employed ensemble learning models include Natural Gradient Boosting (NGBoost), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LGBM), with hyperparameters optimized using Particle Swarm Optimization (PSO) and Adaptive Particle Swarm Optimization (APSO). Results indicate that APSO effectively avoids local optima compared to PSO. On the test set, LGBM achieves a predictive accuracy R 2 of 0.891, outperforming NGBoost (0.853) and XGBoost (0.824). By integrating the predictive model with SHAP, the decision-making process is visualized, facilitating the selection and adjustment of parameters. This approach reduces the impact of shield tunneling on the ground surface environment and effectively controls ground surface settlement. • Developed an adaptive particle swarm-optimized ensemble model achieving high-accuracy ground settlement prediction. • Identified grouting pressure and tunneling speed as dominant settlement factors through SHAP-driven interpretability. • Reduced ground settlement by 43% in field cases via data-driven parameter adjustments, validated on metro tunnel data.
Liu et al. (Sun,) studied this question.