Coke reactivity index: application of robust machine learning algorithms for prediction capabilities

Abstract

The coke reactivity index (CRI) significantly impacts metallurgical coke quality, influencing blast furnace efficiency and environmental emissions. This study focusses on leveraging advanced machine learning algorithms to predict CRI based on key input features such as volatile matter, sulphur (%), ash (%), maximum fluidity (ddpm), plastic thickness (mm) and basicity index. A comprehensive dataset of 636 coal samples from diverse origins, measured according to ASTM standards, was analysed. Rigorous preprocessing, including outlier detection using the random forest algorithm, ensured data reliability. Predictive modeling employed random forest in combination with five metaheuristic optimisation algorithms, including particle swarm optimisation (PSO), genetic algorithm (GA) and grey wolf optimiser (GWO), with k-fold cross-validation to ensure robustness. Evaluation metrics such as R², MSE and AARE% demonstrated RF-PSO's superiority, achieving R² of 0.846 and the lowest AARE% during testing, coupled with a significant reduction in computational runtime. SHAP analysis revealed volatile matter as the most critical feature affecting CRI. These findings underline the potential of hybrid machine learning models to enhance predictive accuracy and operational efficiency in industrial coke production.