Modeling coking coal indexes by SHAP-XGBoost: Explainable artificial intelligence method

Abstract

Coking coal is still on the list of critical raw materials in many countries since it is the main element integrated into the blast furnace. While the energy consumption and steelmaking efficiency in the furnace depends on the coke quality, understanding and modeling coking indexes based on their coal parent properties would be a substantial approach for the steelmaking industry. As an innovative approach, this short comminucation has been considered explainable artificial intelligence (XAI) for modeling coal coking indexes (Free Swelling index “FSI” and maximum fluidity “Log (MF)”). XAIs can convert black-box models into human basis systems and develop a significant learning performance and estimation accuracy. SHapley Additive exPlanations (SHAP), as one of the most recently developed XAI models in combination with eXtreme gradient boosting (XGBoost), were used to model coal samples from Illinois, USA. For the first time, FSI and Log (MF) treat as ordinal variables for modeling. Modeling outcomes relieved that SHAP-XGBoost could accurately show interdependency between features, demonstrate the magnitude of their multi relationships, rank them based on their importance, and predict the coking index quite accurately compared with conventional machine learning methods (random forest and support vector regression). These significant results would be opened a new window by applying XAI tools for controlling and modeling complex systems in the energy and fuel sectors.

Keywords

• Free swelling index
• Gieseler plastometer
• Coal
• Explainable artificial intelligence
• Machine learning
• Modeling