Experimental study and extreme gradient boosting (XGBoost) based prediction of caking ability of coal blends

摘要

The caking ability sets coking coals apart from other coals. In the industry, the practice of blending coals with different physicochemical parameters is often used to obtain the fuel with the desired properties. Caking ability is a non-additive parameter, which means that it cannot be calculated from the weighted average of the components of the blend. In this study, a model for predicting the caking ability of coal in blends was developed, based on the XGBoost regressor machine learning method. Experiments involving blending eight coals in five different proportions were carried out to obtain a training dataset. The coals were selected to reflect the widest possible range of types and caking ability levels. The Roga Index (RI) was chosen as a measure of caking ability. The experiments proved the non-additivity of caking ability of coal in blends, in particular for low-rank coals and good quality coking coals. The Feature Importance and Partial Dependence Plots techniques were used to determine the impact of model input parameters on the predicted value. Our study demonstrated that the carbon, moisture, volatile matter and vitrinite contents had the greatest impact on predicted RI. The prediction efficiency of the model was R2 = 0.95, confirming the effectiveness of the method used.