Prediction of soil conditioner dosages for shield tunneling in sandy soil based on machine learning

Inadequate soil conditioning during Earth Pressure Balance shield (EPBS) tunneling in sandy strata frequently causes operational issues. This study developed a data-driven framework integrating 15 operational and geological parameters from Shenyang Metro Line 4. Using principal component analysis for dimensionality reduction and quantitative TPI/FPI criteria for dataset selection, and constructed SVR, XGBoost, and LightGBM models optimized via PSO, BO, and Optuna. The LightGBM-Optuna ensemble demonstrated superior performance, reducing prediction errors by 45–58 % compared to baseline configurations. Application to suboptimal datasets validated a 13.15 % reduction in specific energy consumption, demonstrating significant potential for intelligent tunneling control.