RAPID PREDICTION OF COAL QUALITY PARAMETERS FROM DIGITAL IMAGES USING MULTIPLE REGRESSION AND ARTIFICIAL NEURAL NETWORKS: A COMPARATIVE STUDY ON LAKHRA LIGNITE

Abstract

Coal quality assessment governs the acceptance, blending, and combustion control of feedstock in coal-fired power plants, yet the conventional proximate and calorific analyses on which it relies are slow, equipment-intensive, and ill-suited to real-time process control. This study develops and compares two data-driven models that predict the three most decision-relevant coal-quality parameters, namely ash content, fixed carbon (FC), and gross calorific value (GCV), directly from a digital image of the coal, eliminating the day-long laboratory turnaround. Thirty-six lignite samples were collected from the Lakhra coalfield (Sindh, Pakistan); each sample was imaged in a controlled illumination box, segmented by K-means clustering, and characterised by thirty-five colour and texture descriptors drawn from colour moments and the GLCM, Haralick, and Tamura families. The same samples were analysed by a thermogravimetric analyser (TGA-701) and a calorific-value analyser (AC-500) to obtain reference values. A multiple-regression model (MRM) and a two-hidden-layer artificial neural network (ANN) trained with Bayesian regularisation were then fitted to map image features onto coal quality. The ANN clearly outperformed the regression baseline, raising the coefficient of determination from 0.726, 0.705, and 0.645 (MRM) to 0.963, 0.944, and 0.928 (ANN) for ash, FC, and GCV respectively, while reducing mean square error by one to two orders of magnitude. The results confirm that ash and fixed carbon are more tightly coupled to surface colour and texture than GCV, and that a compact neural network can deliver laboratory-comparable estimates within seconds, offering an inexpensive screening tool for indigenous lignite deposits.