Rockburst Prediction and Evaluation Model for Hard Rock Engineering Based on Extreme Gradient Boosting Ensemble Learning and SHAP Value

Rockburst prediction is the basis of rockburst prevention and construction guidance. However, the complexity of the rock burst occurrence mechanism and inducing factors and the suddenness and randomness of rock burst behavior make the accurate prediction of rock bursts very difficult. In this study, the eXtreme Gradient Boosting (XGBoost) algorithm is used to learn and predict the rockburst intensity of a database including 341 rockburst cases worldwide. A procedure for parameter optimization of XGBoost combined with grid search and cross validation methods is proposed. It improves the prediction performance, effectively avoids overfitting and also improves the operation efficiency. The model predicted 7 typical rockburst cases that occurred at Jinping II Hydropower Station, and the results showed that the GC-XGBoost model performs well in predicting rockburst intensity. In addition, compared with typical supervised learning models (SVM and RF), the model showed improved prediction performance. SHapley Additive exPlanations (SHAP, a game theoretic approach) was used to study the importance of feature parameters. The SHAP values showed that W-et and sigma(0). are the two most important feature parameters for predicting rockburst intensity.

Automated summary

Rockburst prediction is difficult due to the complexity and randomness of rock burst behavior. This study uses the XGBoost algorithm to predict rockburst intensity and proposes a parameter optimization procedure to improve prediction performance. The model performs well in predicting rockburst intensity and identifies important feature parameters. The results show improved performance compared to other supervised learning models.