Short-term rockburst risk prediction using ensemble learning methods

Abstract

Short-term rockburst risk prediction plays a crucial role in ensuring the safety of workers. However, it is a challenging task in deep rock engineering as it depends on many factors. More recently, machine learning approaches have started to be used to predict rockbursts. In this paper, ensemble learning methods including random forest (RF), adaptive boosting, gradient boosted decision tree (GBDT), extreme gradient boosting and light gradient boosting machine were adopted to predict short-term rockburst risk using microseismic data from the tunnels of Jinping-II hydropower project in China. First, labeled rockburst data with six indicators based on microseismic monitoring were collected. Then, the original rockburst data were randomly divided into training and test sets with a 70/30 sampling strategy. The hyperparameters of the ensemble learning methods were tuned with fivefold cross-validation during training. Finally, the predictive performance of each model was evaluated using classification accuracy, Cohen’s Kappa, precision, recall and F-measure metrics on the test set. The results showed that RF and GBDT possessed better overall performance. RF obtained the highest average accuracy of 0.8000 for all cases, whereas GBDT achieved the highest value for high (moderate and intense) risk cases with an accuracy of 0.9167. The proposed methodology can provide effective guidance for short-term rockburst risk management in deep underground projects.