Interpretable prediction of peak particle velocity in tunnel blasting based on feature-enhancement-weighted WOA-XGBoost and SHAP analysis

To improve the prediction accuracy of peak particle velocity ( PPV ) in tunnel blasting, enhance the model's sensitivity to high- PPV samples, and improve model interpretability, this study develops a hybrid prediction framework, FEW-WOA-XGBoost, that integrates feature enhancement, high- PPV sample weighting, and Whale Optimisation Algorithm (WOA)-based hyperparameter optimisation. Based on 125 sets of field blasting vibration monitoring data from the Chongli Tunnel, the distance from the blast centre, elevation difference, single-hole charge, total charge, and maximum charge per delay were selected as the original input variables. In addition, scaled-distance terms, logarithmic transformation terms, and charge-distance coupling terms were constructed to form an 11-dimensional input feature system. This study employs min–max normalisation to unify the feature scales and introduces a high- PPV sample-weighting strategy to enhance the model's sensitivity to high- PPV samples. On this basis, the FEW-WOA-XGBoost prediction framework is developed using XGBoost as the base learner, with WOA employed to globally optimise its key hyperparameters. To verify its performance, the proposed model was compared with benchmark models and empirical methods, including GBO-XGBoost, GWO-XGBoost, RUN-XGBoost, PSO-XGBoost, standard XGBoost, the modified Sadowski empirical formula and SVR. A comprehensive evaluation was further conducted using regression performance metrics, TOPSIS-based ranking, Taylor diagrams, PPV response-level classification, SHAP analysis, and partial dependence plot (PDP) analysis. The results show that FEW-WOA-XGBoost achieves R 2 , RMSE, MAE, VAF, and MBE values of 0.971, 0.134, 0.084, 97.16%, and 0.024, respectively, on the independent test set. Under the dataset and evaluation scheme used in this study, the model demonstrates better overall prediction performance than the benchmark models. In PPV response-level classification, the model achieves accuracy, precision, recall, F1 score, and Cohen's kappa coefficient values of 0.888, 0.911, 0.884, 0.895, and 0.843, respectively, indicating that it has good capability for PPV response-level classification. The SHAP analysis results show that the maximum charge per delay, the scaled distance based on total charge, the elevation difference, the ratio of total charge to blast centre distance, and the blast centre distance are the main factors affecting PPV prediction, with relative contributions of 27.5%, 21.8%, 20.9%, 13.8%, and 13.5%, respectively. The results show that, under the dataset and evaluation scheme used in this study, the integrated FEW-WOA-XGBoost framework can effectively characterise the nonlinear relationships between the selected blasting-related features and PPV. The developed model can provide a reference for blast-induced vibration prediction in tunnel blasting, PPV response-level classification, and construction-parameter optimisation.