Machine Learning Prediction of Torrefied Biomass Properties From Diverse Agricultural Residues

Torrefaction has emerged as a promising thermochemical pretreatment to enhance the fuel properties of agricultural biomass, yet predicting product characteristics remains challenging due to feedstock variability. This study developed and compared four machine learning (ML) models, support vector regression (SVR), artificial neural network (ANN), gradient boosting machine (GBM), and random forest (RF), to predict the higher heating value (HHV), mass yield (MY), and energy yield (EY) of torrefied biomass. A comprehensive dataset comprising 117 experimental samples from 11 biomass types torrefied at 200–280°C for 20–60 min was utilized. Input variables included the contents of proximate composition (ash, fixed carbon, and volatile matter), ultimate composition (C, H, N, O, S), and process conditions (temperature and residence time). SHapley Additive exPlanations (SHAP) analysis revealed that torrefaction temperature was the most influential parameter, followed by volatile matter and ash content. Results demonstrated that GBM consistently outperformed the other models, achieving an acceptably high coefficient of determination, with minimal prediction errors across different input configurations. Notably, simple proximate‐analysis data provided performance comparable to that of more detailed and instrument‐intensive ultimate analysis (UL), highlighting the practical value of GBM for operators handling diverse biomass feedstocks with limited access to specialized analytical instruments.