Interpretable Tree-Based Models for Predicting Short-Term Rockburst Risk Considering Multiple Factors

Abstract

Rockbursts are typical dynamic disasters in underground coal mines. Their formation and occurrence are influenced by multiple factors, including geological conditions, mining disturbances, and microseismic activities. Machine learning models can support early warning by forecasting high-energy microseismic (MS) activity associated with elevated rockburst risk. However, many of these models are black boxes that are difficult for humans to understand, which may diminish their credibility and constrain their practical applicability. To establish interpretable models for short-term rockburst prediction considering multiple factors, we assess five tree-based models (decision tree, random forest, LightGBM, XGBoost, and CatBoost) using field databases from Gao Jiapu coal mine and utilize the model explanation method SHAP to interpret the models. A day is labeled as rockburst risk if its daily maximum MS energy exceeds 1 × 10⁵ J, otherwise no risk. The results revealed that CatBoost exhibited the highest overall performance, followed by LightGBM, and both the random tree and XGBoost performed better than the decision tree. When considering the influence of time window length, all models displayed optimal performance under a 3-day time window. Specifically, CatBoost achieved F1 scores of 0.856, 0.778, and 0.807 for the 3-day, 5-day, and 10-day time windows, respectively. Global feature importance analysis showed that the average energy, energy deviation, and cumulative energy were the three most important predictors. Furthermore, local SHAP values were drawn to reveal the complex underlying relationships between the 12 factors and model predictions. The results show that factors such as energy-related factors, coal depth, and working face advance positively contribute to rockburst risk prediction. Conversely, fold distance, total frequency, and source concentration degree negatively impact the model outcomes. The findings suggest that incorporating CatBoost and the SHAP method holds great potential in developing interpretable models for accurate rockburst risk prediction.