PREDICTION OF PROXIMATE ANALYSIS AND PROCESS TEMPERATURE OF TORREFIED AND PYROLYZED WOOD PELLETS BY NEAR-INFRARED SPECTROSCOPY COUPLED WITH MACHINE LEARNING

Abstract

Near-Infrared (NIR) Spectroscopy is a time and cost-effective method to characterize the materials in the food, petrochemical, pharmaceutical, and agricultural industries. Proximate analysis of the carbon-containing materials and investigating the effectiveness of the heat treatments on the material are a particularly time-consuming process. This work presents the four regression methods, i.e., decision tree regression, support vector regression and two versions of ensembles of decision trees to predict the proximate analysis of biomass and heat treatment temperature. Thus, effective method has been proposed to reduce experimental effort and present the characterization of heat-treated biomass feedstock theoretically. Prediction results show that SVR and ENS2 regression methods calibrating the NIR spectra to the values of wood pellet properties achieved good performance with the coefficient of determination (R2) of 0.880- 0.984 and RMSE of 0.444- 5.308 for ash and volatile matter. This study suggests that machine learning-based regression methods with integrated NIR spectroscopy of biomass is promising as an alternative method for rapid characterization. Another possible application of the current study is that it can be used for processed fuel recognition prior to a fully automated fuel quality assessment system in the biomass industry.