Understanding the impact of coal blending decisions on the prediction of coke quality: a data mining approach

Abstract

Abstract The accurate prediction of coke quality is important for the selection and valuation of metallurgical coals. Whilst many prediction models exist, they tend to perform poorly for coals beyond which the model was developed. Further, these models general fail to directly account for physical interactions occurring between the blend components, through the assumption that the aggregate properties of the blend are suitably representative of the overall behavior of the blend. To study this assumption, a parameter termed the vitrinite distribution category was introduced to directly account for the distribution of one of these commonly aggregated parameters, the vitrinite reflectance. The introduction of this parameter in a regression model for coke quality prediction improved the model fit. The vitrinite distribution category was demonstrated to provide new information about coal blending decisions, and was found to be capable of providing insight into the behavior of different blending structures. Residual analysis was applied to explore the behavior of the coke quality prediction model, with the vitrinite distribution category found to explain more than just the presence or absence of coals within a blend. This work provides the foundation of future studies in examining coal blending decisions, with the proposed parameter having the potential to be applied as part of a coke quality prediction model to optimize coal blending decisions.

Keywords

• Coke
• Quality
• Prediction
• Self-organizing maps
• Vitrinite reflectance distribution