An optimized machine learning framework for prediction of coal abrasive index: Leveraging supervised learning, metaheuristic optimization, and interpretability analysis

•The study compares 18 supervised learning models for Abrasive Index (AI) prediction using 129 coal samples from South Africa's Witbank Coalfield.•Eight metaheuristic optimization algorithms (MOAs) are used to optimize hyperparameters and improve model predictive performance.•Uncertainty and fe...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Fuel (Guildford) Ročník 403; s. 136065
Hlavní autoři: Qi, Hongning, Zhou, Jian, Khandelwal, Manoj, Onifade, Moshood, Lawal, Abiodun Ismail, Li, Chuanqi, Bada, Samson Oluwaseyi, Genc, Bekir
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier Ltd 01.01.2026
Témata:
Abrasive index
Artificial intelligence
Coal
Metaheuristic optimization algorithms
Supervised learning
Metaheuristic optimization algorithms
Supervised learning
Artificial intelligence
Abrasive index
Coal
ISSN:0016-2361
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:•The study compares 18 supervised learning models for Abrasive Index (AI) prediction using 129 coal samples from South Africa's Witbank Coalfield.•Eight metaheuristic optimization algorithms (MOAs) are used to optimize hyperparameters and improve model predictive performance.•Uncertainty and feature importance are analyzed using MCS, SHAP, and ICE to provide insight into model reliability and key factors.•The DBO-RF model outperforms all other models, achieving R² = 0.94 and demonstrating its potential for coal processing applications. Abrasive Index (AI) is crucial in evaluating coal’s impact on mechanical equipment wear, influencing operational efficiency and maintenance costs. Accurate AI prediction is essential for optimizing coal utilization and reducing economic losses in industrial applications. This study systematically evaluates AI prediction using 18 supervised learning (SL) models on a dataset of 129 coal samples from the Witbank Coalfield, South Africa. The most suitable model is identified based on performance comparisons, followed by hyperparameter optimization using eight metaheuristic optimization algorithms (MOAs). To further ensure robustness and interpretability, Monte Carlo Simulation (MCS), SHapley Additive exPlanations (SHAP), and Individual Conditional Expectation (ICE) methods are employed for uncertainty quantification and feature importance analysis. The results indicate that the Dung Beetle Optimizer-optimized Random Forest (DBO-RF) model outperforms all other hybrid models, achieving an R2 of 0.94, RMSE of 26.583, MAPE of 0.202, and VAF of 94.178% on the test set. Feature interpretation analysis reveals that quartz content significantly impacts AI predictions, while volatile matter contributes the least. This study provides a comprehensive AI prediction framework that integrates SL, optimization, and interpretability analysis, offering valuable insights for coal selection, equipment maintenance, and industrial process optimization.
ISSN:0016-2361
DOI:10.1016/j.fuel.2025.136065