Advancing water-conducting fracture zone height prediction in coal mining via hybrid XGBoost models with interpretability analysis

Abstract

The height of the water-conducting fracture zone (WCFZH) is a crucial indicator for assessing mining-induced damage in coal mines. Accurate and interpretable prediction of WCFZH is essential for disaster prevention and safe mining. Therefore, this paper combines extreme gradient boosting (XGBoost) with heuristic intelligent algorithms to develop hybrid intelligent models for WCFZH prediction. Accordingly, a database including six features (geological conditions, mining parameters, and coal seam attributes) and one target label was established for model training and validation. Evaluation results show that the hybrid model exhibits superior predictive performance and generalization ability. For the optimal hybrid model, the coefficient of determination (R2) and variance accounted for (VAF) reached 0.9654 and 96.6232, respectively, while the root mean square error (RMSE) and the mean absolute error (MAE) were 6.2489 and 5.0324, respectively. Finally, interpretability analysis via Shapley Additive exPlanations (SHAP) reveals the ranking of input feature importance: N, H, L, s, m, and k. Moreover, interactions between H and m, L and m, and H and N significantly influence model predictions.