Accurate analysis of coal calorific value using NIRS-XRF: Utilizing RF classification and PLSR subtype modeling

Rapid and precise measurement of the calorific value of coal is crucial for coal chemical enterprises. However, due to the wide variety of coal sources and the diverse types of coal, the matrix effect significantly reduces the accuracy of spectroscopic rapid detection techniques for coal quality. To address this issue, this study conducts research on a coal calorific value detection method based on the combined use of near-infrared spectroscopy (NIRS) and X-ray fluorescence spectroscopy (XRF), exploring the feasibility of identifying coal types first and then predicting calorific value based on corresponding subtype models. In the aspect of spectral preprocessing, fusion of the dual spectra is achieved using techniques such as truncation, smoothing, Standard Normal Variate (SNV), and normalization. For coal type identification, the identification performance of five machine learning algorithms is compared, and the random forest (RF) algorithm is used to improve the identification accuracy to 98.43 %. In terms of prediction modeling, a comparison is made between traditional holistic models based on Partial Least Squares Regression (PLSR) and subtype models built according to coal types, revealing that the latter demonstrating stronger generalization ability and robustness. Accordingly, a precise prediction strategy for coal calorific value based on the combined use of NIRS and XRF in complex coal types is proposed: first, the RF algorithm is used to automatically identify coal types based on the fusion spectrum of NIRS and XRF, and then the calorific value of the corresponding coal type is accurately predicted using the PLSR subtype model selected according to the identification results. Industrial test results indicate that the average absolute error (AAE), root mean square error of prediction (RMSEP), average relative error (ARE), and standard deviation (SD) for coal calorific value prediction are 0.20 MJ/kg, 0.24 MJ/kg, 0.79 %, and 0.04 MJ/kg, respectively, meeting the requirements of enterprises. The research findings provide a practical technical method for timely and accurate acquisition of coal calorific value data, addressing the interference problem of matrix effects in complex coal type scenarios and holding broad application prospects in the coal-based energy industries.