Rockburst prediction is vital for ensuring safe underground construction. Machine learning (ML) has advanced rockburst prediction, yet imbalanced datasets, particularly for specific intensities, often causes biased models or overfitting. This study introduces a novel ML framework for rockburst intensity prediction, utilizing six key indices: the maximum tangential stress ( σ θ ), uniaxial compressive strength ( R c ), uniaxial tensile strength ( R t ), stress concentration factor (SCF), brittleness index (BI), and strain energy storage index ( W et ). A global dataset of 570 rockburst cases was compiled, with the synthetic minority oversampling technique (SMOTE) applied to address class imbalance, improving the representation of underrepresented grades. An improved IVY algorithm (IIVY) optimizes hyperparameter tuning. It integrates into an IIVY-AdaBoost model that combines decision trees (DT) and support vector machines (SVM) as base learners. The model achieves superior performance, with accuracy of 90.17%, precision of 0.89, recall of 0.90, and F1-score of 0.90. Compared with the base learners (DT, IVY-DT, IIVY-DT, SVM, IVY-SVM, IIVY-SVM, AdaBoost, IVY-AdaBoost, PSO (PARTICLE SWARM Optimization)-AdaBoost, GWO (Grey Wolf Optimizer)-AdaBoost), the IIVY-AdaBoost demonstrates an enhanced accuracy and robustness. The SHapley Additive exPlanations (SHAP) analysis highlights W et and σ θ as the key predictors. Validation using field data from Jinping II Hydropower Station, Sanshandao Gold Mine, and Xincheng Gold Mine confirms the model’s robust engineering applicability.
Peng et al. (Sun,) studied this question.