A stacking ensemble approach for blast vibration prediction incorporating Yeo–Johnson normalization

Blasting operations in mining often generate ground vibrations that may endanger nearby structures and negatively impact the surrounding environment and community. Effective vibration control is therefore essential for advancing green mining practices, as it reduces environmental disturbance and enhances community safety. Peak Particle Velocity (PPV), the primary indicator of blast-induced ground vibration, plays a central role in vibration assessment and in ensuring safe and optimized blast design. Despite growing interest in machine learning (ML)-based PPV prediction, existing research has not fully leveraged the strengths of ensemble algorithms or systematically incorporated geotechnical factors and advanced normalization techniques. To address these gaps, this study evaluates and compares the performance of three ensemble ML models, Extreme Gradient Boosting (XGBoost), AdaBoost, and Random Forest (RF), enhanced using the Bayesian Optimization Algorithm (BOA) for hyperparameter tuning. In addition, a Stacking Ensemble (SE) model is developed that integrates the predictions of the three optimized models to further improve accuracy. The models are trained on a comprehensive open-pit mine dataset incorporating parameters such as total charge, maximum charge weight per delay (MCWD), distance from blast site, hole depth, burden, spacing, rock quality designation (RQD), and penetration rate. To address data skewness and enhance model performance, the Yeo-Johnson power transformation is applied during preprocessing. Model performance is evaluated using coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). Results confirm the effectiveness of the transformation approach and show that the proposed SE model outperforms both individual ML algorithms and conventional empirical equations, achieving an R² of 0.9292, an RMSE of 0.0076, and an MAE of 0.0058 on the test dataset.