Abstract
Underground coal mining presents significant environmental challenges, particularly land subsidence, which can lead to severe ecological and structural consequences. This phenomenon alters surface topography, disrupts groundwater flow, and poses risks to infrastructure, necessitating accurate predictive models for effective mitigation. Longwall mining, a widely used extraction method, is especially prone to inducing ground subsidence due to its caving process. This study leverages advanced machine learning techniques to enhance subsidence prediction accuracy and inform sustainable mining practices. Three hybrid models-biogeography-based optimization with gene expression programming (BBO-GEP), gray wolf optimizer with gene expression programming (GWO-GEP), and salp swarm algorithm with gene expression programming (SSA-GEP)-are applied to assess subsidence risks. Three hybrid models-BBO-GEP, GWO-GEP, and SSA-GEP-were developed and tested to enhance prediction accuracy and reduce model uncertainty. A comprehensive dataset comprising 11 key geotechnical and mining parameters, including seam thickness, mining depth, and various rock properties, was collected from 14 coal mines. Results indicate that the BBO-GEP model achieved the highest predictive accuracy, with a correlation coefficient of 0.99. Sensitivity analysis revealed that mining depth is the most influential factor in subsidence occurrence, whereas density has the least impact. These findings contribute to environmental risk management in mining by providing a robust predictive framework that aids in proactive subsidence mitigation strategies. The proposed models support decision-making for policymakers and industry stakeholders, fostering more sustainable mining operations with minimized ecological disruption.