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

Area of Science:

• Civil Engineering
• Geotechnical Engineering
• Data Science

Background:

• Accurate grouting volume prediction is crucial for large-scale water conservancy and hydropower projects to manage seepage, ensure construction quality, and control costs.
• Existing methods may lack precision, particularly under complex geological conditions.

Purpose of the Study:

• To develop and validate a novel grouting volume prediction model using a stacking ensemble learning framework combined with Bayesian optimization (BO).
• To enhance the accuracy and interpretability of grouting volume predictions for improved project management.

Main Methods:

• A dataset of 778 grouting records was utilized with seven input features including geological and construction parameters.
• XGBoost, LightGBM, and Random Forest were used as base learners, optimized globally by BO.
• Ridge regression acted as a meta-learner, forming the Bayesian-optimized stacking ensemble (BO-Stacking) model.
• SHapley Additive exPlanations (SHAP) analysis was employed for feature contribution assessment and model interpretability.

Main Results:

• The BO-Stacking model demonstrated superior performance over benchmark models.
• Achieved a coefficient of determination (R²) of 0.92, Mean Absolute Error (MAE) of 70.19 L, and Root Mean Square Error (RMSE) of 187.07 L.
• SHAP analysis revealed the significant influence of geological conditions, construction parameters, and slurry properties on grouting volume.

Conclusions:

• The proposed BO-Stacking model significantly improves the prediction accuracy of grouting volume, especially in complex geological settings.
• This approach offers valuable support for construction planning, quality management, and cost control in major water infrastructure projects.