Abstract: The low calorific value (LCV) of coal is a key parameter for pricing, mining and stripping planning, and trade settlement in open-pit coal mines. However, traditional kriging estimation relies on assumptions of stationarity, making it difficult to capture the nonlinear spatial variation of coal quality. Although existing machine learning methods have improved prediction accuracy, their “black-box” nature limits the understanding of underlying coupling mechanisms among indicators, thereby restricting engineering applications in real-time blending during extraction. To balance accuracy and interpretability, this study constructs an explainable LCV prediction model that quantifies the contribution patterns of various factors to low calorific value. Using 1 992 exploration and production coal quality data from 21 coal seams at the Dainanhua No. 2 Open-pit Mine in Xinjiang, we build stratified feature sets based on sample size—14-dimensional full features for major seams and 10-dimensional simplified features for low-sample seams. An extreme gradient boosting (XGBoost) regression model is established, optimized via Bayesian optimization combined with fivefold cross-validation across nine hyperparameter configurations, and enhanced with conformal prediction to provide 95% confidence intervals. Three experimental groups are designed: accuracy comparison, interpretability analysis, and ablation studies. Feature importance mechanisms are analyzed using SHapley Additive exPlanations (SHAP) and accumulated local effects (ALE). Results show that XGBoost achieves the highest accuracy among nine regression models, with a root mean square error (RMSE) of 0.760 3 MJ/kg and a coefficient of determination R² of 0.894 5 under spatial block cross-validation, representing a 3.1-fold improvement over conventional kriging (RMSE = 2.367 3). Ablation experiments confirm industrial analysis indicators—ash content, volatile matter, and moisture—as absolute core features; their removal increases RMSE by 151%, indicating that these three parameters alone can achieve engineering-grade accuracy for this scenario. SHAP attribution reveals ash as the primary driver, contributing approximately 65.5%. The ALE-derived ash effect slope (−0.27 MJ/(kg·%)) aligns closely with empirical coal chemistry formulas and highlights strong negative interaction between high-ash (> 25%) and high-volatile (> 46%) zones. The LCV distribution characteristics of Dainanhua No. 2 Mine show predicted values ranging from 22.21 to 25.66 MJ/kg, with the optimal “calorific column” located in the central section at depths of 150–280 m, exhibiting a three-stage vertical differentiation pattern: shallow depletion, mid-depth enrichment, and deep differentiation. Dual diagnostics using Bayesian optimization and learning curves indicate that current accuracy approaches the theoretical upper limit given available data, confirming ash dominance and strong negative interaction between high-ash and high-volatile regions as the main controls on spatial LCV enrichment. Based on these findings, a two-stage blending strategy is proposed: coarse blending at the working face guided by spatial patterns, followed by fine adjustment in stockpiles. Additionally, an uncertainty-driven indicator-based exploration mechanism is introduced. Predicted LCV values are translated into executable blending ratios (e.g., monthly blended low calorific value of 19.55 MJ/kg with 29.6% low-quality coal), forming a closed-loop coal quality management system from prediction to blending. This approach directly supports optimization of mining sequences, segregated extraction and storage control, and dynamic blending decisions, enabling open-pit mines to transition from post-hoc laboratory adjustments to proactive spatial forecasting, providing an interpretable and uncertainty-controllable technical solution for refined coal quality management.