Reliable estimation of the height of the caving-fracturing zone (HCFZ) is indispensable for strata control design and production scheduling in longwall operations, yet it remains a technically demanding task because the process is mediated by a web of nonlinear interactions among lithological and geometric variables. To capture this complexity, this study compiles a rigorously screened database of 83 longwall panels drawn from Iranian coalfields and earlier case studies, covering a broad spectrum of overburden depths (H = 130–650 m), panel widths (L = 40–330 m), and rock mass properties (elastic modulus E = 1.13–18.85 GPa, Poisson’s ratio ν = 0.18–0.33, etc.). Preliminary diagnostics, including Pearson correlation matrices and variance inflation factors, confirm the absence of critical multicollinearity, thereby justifying the simultaneous inclusion of these variables in multivariate modeling. Five representative machine learning regressors are benchmarked as base learners, and the best performer, Extreme Gradient Boosting (XGBoost), is further tuned with three particle swarm optimization (PSO) variants. Under the existing data conditions, the XGBoost optimized by random weight PSO (RandWPSO–XGBoost) emerges as the most accurate configuration, achieving a determination coefficient of 0.9452, a root mean square error of 10.8440 m, and a mean absolute error of 8.2721 m, outperforming both the untuned baseline and competing PSO variants. Sensitivity analysis using Shapley Additive Explanations (SHAP) values reveals that elastic modulus (46.08%), overburden depth (31.22%), and panel width (14.18%) collectively dominate HCFZ predictions, accounting for 91.48% of global influence. This study combines data-driven modeling with geological mechanics, offering a reliable framework for engineering applications in complex mining environments.Highlights<UnorderedList Mark="Bullet"> <ItemContent> Compile a high-quality dataset that pairs dominant longwall mining causative factors with quantitative footprints of failure and disturbance. </ItemContent> <ItemContent> Develop a random weight particle swarm optimization-tuned framework for data-driven prediction of the height of the caving-fracturing zone in longwall mining. </ItemContent> <ItemContent> Provide global and local model interpretability, elucidating factor sensitivities and dominant controls. </ItemContent> <ItemContent> Develop an interactive visualization tool integrating residual diagnostics to support rapid analysis and decision-making. </ItemContent> </UnorderedList>

Hybrid RandWPSO–XGBoost with SHAP-Based Interpretability for Prediction of Caving-Fracturing Zone Height in Longwall Mining

期刊论文
发表日期：

2026年4月9日

查看全文

Shuai Huang,
Yulin Zhang,
Mohammad Rezaei,
Jian Zhou

摘要

Reliable estimation of the height of the caving-fracturing zone (HCFZ) is indispensable for strata control design and production scheduling in longwall operations, yet it remains a technically demanding task because the process is mediated by a web of nonlinear interactions among lithological and geometric variables. To capture this complexity, this study compiles a rigorously screened database of 83 longwall panels drawn from Iranian coalfields and earlier case studies, covering a broad spectrum of overburden depths (H = 130–650 m), panel widths (L = 40–330 m), and rock mass properties (elastic modulus E = 1.13–18.85 GPa, Poisson’s ratio ν = 0.18–0.33, etc.). Preliminary diagnostics, including Pearson correlation matrices and variance inflation factors, confirm the absence of critical multicollinearity, thereby justifying the simultaneous inclusion of these variables in multivariate modeling. Five representative machine learning regressors are benchmarked as base learners, and the best performer, Extreme Gradient Boosting (XGBoost), is further tuned with three particle swarm optimization (PSO) variants. Under the existing data conditions, the XGBoost optimized by random weight PSO (RandWPSO–XGBoost) emerges as the most accurate configuration, achieving a determination coefficient of 0.9452, a root mean square error of 10.8440 m, and a mean absolute error of 8.2721 m, outperforming both the untuned baseline and competing PSO variants. Sensitivity analysis using Shapley Additive Explanations (SHAP) values reveals that elastic modulus (46.08%), overburden depth (31.22%), and panel width (14.18%) collectively dominate HCFZ predictions, accounting for 91.48% of global influence. This study combines data-driven modeling with geological mechanics, offering a reliable framework for engineering applications in complex mining environments.

Highlights

Compile a high-quality dataset that pairs dominant longwall mining causative factors with quantitative footprints of failure and disturbance.

Develop a random weight particle swarm optimization-tuned framework for data-driven prediction of the height of the caving-fracturing zone in longwall mining.

Provide global and local model interpretability, elucidating factor sensitivities and dominant controls.

Develop an interactive visualization tool integrating residual diagnostics to support rapid analysis and decision-making.