Multi-response prediction and uncertainty quantification of blasting vibration using WOA-optimized XGBoost and conformal inference

Existing studies on blast-induced vibration prediction predominantly focus on a single response indicator, leaving multi-parameter characterization and predictive uncertainty quantification insufficiently addressed. This study proposes a unified framework that integrates Kennard–Stone (KS) sample partitioning, Whale Optimization Algorithm (WOA)-tuned Extreme Gradient Boosting (XGBoost), and out-of-fold (OOF) calibrated adaptive conformal prediction. Separate WOA-XGBoost models are developed to predict peak particle velocity (PPV), dominant frequency (f), and vibration duration (T). To quantify prediction uncertainty, a Random Forest-based error-scale model is embedded within the conformal prediction framework, enabling input-dependent interval widths without requiring a separately held-out calibration subset. On an independent hold-out test set, the proposed framework achieves coefficients of determination (R 2 ) of 0.899, 0.880, and 0.967 for PPV, f, and T, respectively, outperforming benchmark models in RMSE and MAE. At a nominal 90% confidence level, the prediction interval coverage probabilities are 92.86% for PPV, 92.86% for f, and 89.29% for T, with normalized interval widths and coverage-width-based criteria revealing distinct, target-dependent uncertainty structures. External validation using an independent dataset from a Xinjiang iron mine yields R 2 values of 0.865, 0.915, and 0.921, confirming robust generalization across varying geological conditions. The proposed framework provides an engineering-oriented, interpretable, and uncertainty-aware decision-support strategy for blasting vibration risk assessment, demonstrating the practical applicability of optimized ensemble learning combined with conformal inference.