A novel cluster-based ensemble learning method for long-term rockburst risk prediction and its application

Accurately predicting the long-term rockburst risk is essential to ensure the underground excavation safety. In this study, a cluster-based ensemble learning (CEL) model was developed by aggregating balanced iterative reducing and clustering using hierarchies (BIRCH) and random forest (RF) algorithms to predict the long-term rockburst risk. A total of 259 historical rockburst cases with six indicators were collected to verify the feasibility of the proposed CEL-RF model. To improve the reliability of the model, the Bayesian optimization (BO) and five-fold cross validation (CV) approaches were combined to search the optimal hyperparameters and weights of RF classifiers. The comprehensive performance of models was compared and evaluated by five metrics (accuracy, Cohen’s Kappa, macro average of the precision, recall and F₁-score). The results indicated that the CEL-RF model performed best with the accuracy of 0.885. In addition, the CEL-RF model was applied to predict the probability of rockburst risk in four underground gold mines and the results were consistent with the field conditions. The Shapley additive explanations (SHAP) method revealed that the elastic energy index W_et was the most important indicator. Overall, the proposed CEL-RF model is a promising model for long-term rockburst risk prediction.