Curtain grouting volume prediction using a Bayesian-optimized stacking ensemble model with SHAP analysis

Curtain grouting is widely used to control seepage in large-scale water conservancy and hydropower projects, and accurate prediction of grouting volume is essential for ensuring construction quality and cost control. This study proposes a grouting volume prediction model combining Bayesian optimization (BO) with a stacking ensemble learning framework. The model was developed using a dataset of 778 valid grouting records and incorporated seven key input features: hole sequence, hole depth, section length, hole diameter, pre-grouting permeability, initial water-cement ratio, and grouting pressure. Within this framework, XGBoost, LightGBM, and Random Forest were employed as base learners, with BO applied for global hyperparameter optimization. Ridge regression served as the meta-learner to construct the Bayesian-optimized stacking ensemble (BO-Stacking) model. SHapley Additive exPlanations (SHAP) analysis was used to quantify feature contributions and enhance model interpretability. The results show that the BO-Stacking model outperformed the benchmark models, achieving a coefficient of determination (R²) of 0.92, a mean absolute error (MAE) of 70.19 L, and a root mean square error (RMSE) of 187.07 L. Scatter analysis further indicated strong agreement between predicted and measured values. SHAP analysis quantified the relative contributions of geological conditions, construction parameters, and slurry properties to grouting volume. Overall, the proposed approach improves predictive performance of grouting volume under complex geological conditions and provides support for construction planning and quality management.