Assessment of coal spontaneous combustion potential using statistical and machine learning approaches

Spontaneous combustion of coal remains one of the most critical safety challenges in underground coal mining, posing significant risks to personnel, operational continuity, and the environment. Reliable prediction of coal self-heating potential is therefore essential for effective risk management throughout the mine life cycle. This study presents a data-driven framework for evaluating coal spontaneous combustion potential based on intrinsic coal properties derived from proximate and maceral analyses. A comprehensive dataset comprising 165 coal samples was compiled from published experimental studies, representing a wide range of coal characteristics. The crossing point temperature was adopted as a quantitative indicator of spontaneous combustion tendency. To investigate the complex relationships between coal properties and crossing point temperature, a combination of statistical and intelligent modelling approaches was employed, including nonlinear regression, artificial neural networks (multilayer perceptron and radial basis function), and a decision tree classifier. The results demonstrate that the relationship between intrinsic coal parameters and spontaneous combustion potential is inherently nonlinear and multivariate. While nonlinear regression achieved acceptable predictive performance (R² = 0.82), its mathematical complexity limits practical applicability. Among the developed intelligent models, the radial basis function neural network achieved the lowest prediction error (approximately 5.7%). In parallel, the decision tree model achieved a classification error of about 8% while providing a transparent and interpretable structure for practical decision-making. A key contribution of this study is the development of an integrated and interpretable classification framework based on readily available proximate and maceral analysis parameters, addressing the practical limitations associated with the lack of detailed elemental data in many mining operations. The developed framework was further evaluated using independent field samples from the Parvadeh coal mines (Tabas coalfield, Iran), confirming its applicability under real mining conditions. Unlike previous studies that mainly focused on individual algorithms or laboratory-scale datasets, this study establishes a comparative modelling framework by integrating multiple statistical and intelligent approaches using a comprehensive database of 165 coal samples and independent field validation. The proposed approach provides a practical and interpretable tool for assessing coal spontaneous combustion potential, with potential applications in mine design, operational planning, and post-closure risk management.