SHAP-based interpretation of an XGBoost model in the prediction of grindability of coals and their blends

Abstract

ABSTRACT The Hardgrove Grindability Index (HGI) is a measure of coal’s resistance to crushing. HGI is influenced by many factors due to the complex structure of coal. This study examines the effect of the proximate and ultimate analysis and maceral content on HGI, based on 329 samples of Polish coals. In this study, a machine learning technique XGBoost (Extreme gradient boosting regressor) was used to develop a predictive model of HGI with satisfactory accuracy (R2 = 0.86). The Shapley additive explanations (SHAP) technique was used to explain the relationship between the predicted value and the input data. Studies have shown that the moisture (negative impact), carbon (positive), volatile matter (negative), vitrinite (positive) and liptinite (negative) content have the greatest impact on HGI. Additionally, three experimentally obtained coal blends that differ significantly in the degree of grindability were selected for testing the model. The results confirmed the effectiveness of the method when used also for blends. The influence of individual coal parameters on the predicted grindability of the blends has been thoroughly examined.