General Models for Predicting Torrefaction Solid Yield for Any Type of Biomass and Reactor: A Comparison of Empirical, Kinetic and Machine Learning Methods

Torrefaction is a pre-treatment that can be applied to improve the fuel properties of biomass, reducing solid mass while simultaneously increasing carbon content, reducing oxygen content, and increasing higher heating values. Solid yield from torrefaction is particularly important as it affects many other properties. It is challenging to obtain a generalized model that can predict solid yield for any type of biomass with any type of torrefaction reactor. A comparison is made here between empirical, kinetic, and machine learning models for predicting solid yield. The highest-accuracy general model, based on a dataset of 400 points representing the torrefaction of a diverse range of biomass types and reactor types, is a gradient boosting method with a testing accuracy R2 of 0.925 and an average absolute error of 3.4%. However, for data outside the training range (e.g., high moisture or higher carbon content biomass), the average absolute error for this method increases to 11.7%, while a new optimized 2-step kinetic model is shown to give average absolute errors from 7.6% to 9.4% for data inside and outside the training range, offering a simple alternative that is more accurate than existing kinetic models and correlations.