Prediction of Gray-King coke type from radical concentration and basic properties of coal blends

Metallurgical coke is mainly produced from coal blends. The coke qualities have been related with or predicted by numerous polynomials with one or a few coal parameters from the proximate and ultimate analyses, maximum vitrinite reflectance (R-max) and quantity of plastic matters. More fundamental and intrinsic prediction of coke quality, such as that required by artificial intelligence in the future, calls for relations between coke quality and its intermediate state, such as the coke type (CT) determined by the well-known Gray King (GK) assay, and consequently the relations between GKCT with the basic properties of coal blends. This work studies the GKCT of 68 coal blends and predicts the G type coke (G-coke), the best coke form defined by GK, with the parameters from the ultimate and proximate analyses, and the radical concentration (Cr) of coals because Cr is found correlating well with Rmax. The prediction methods include the traditional single- and multi-parameter range (MPR) methods and 3 machine learning models, namely K-Nearest Neighbors (KNN), Linear Discriminant Analysis (LDA), and Support Vector Machine (SVM). It is found that the readily measurable Cr of coals is an important parameter in GKCT prediction. MPR, KNN, LDA and SVM are capable to predict G-coke with no more than 5 parameters, and SVM is more effective than other models.