Development of the Optuna-NGBoost-SHAP model for estimating ground settlement during tunnel excavation

Abstract

This study aims to develop and evaluate a natural gradient boosting (NGBoost) model optimized with Optuna for estimating ground settlement during tunnel excavation, incorporating Shapley additive explanations (SHAP) to perform interpretability analysis on the model’s estimation results. The model’s predictive performance was comprehensively assessed using datasets from two earth pressure balance shield tunneling projects in Changsha and Zhengzhou, China. Comparative analyses demonstrated the superior accuracy and generalization capability of the Optuna-NGBoost-SHAP model (training set: R² = 0.9984, MAE = 0.1004, RMSE = 0.4193, MedAE = 0.0122; validation set: R² = 0.9001, MAE = 1.3363, RMSE = 3.2992, MedAE = 0.3042; test set: R² = 0.9361, MAE = 0.9961, RMSE = 2.5388, MedAE = 0.2147). SHAP value analysis quantitatively evaluated the contributions of input features to the model’s estimations, identifying geometric factors (distance from the shield machine to the monitoring section and cover depth) as the most important features. The findings provide robust decision support for safety management during tunnel construction and demonstrate the reliability and efficiency of the Optuna-NGBoost-SHAP framework in estimating complex ground settlement scenarios.