Multi-Stage Machine Learning for Physically Consistent Prediction of Fuel Properties in Thermally Upgraded Lignocellulosic Biomass
BioEnergy Research · 24.08.2026 · 19 (1)
DOI: 10.1007/s12155-026-11045-9 ↗Kurz gefasst
Fachartikel in der Zeitschrift BioEnergy Research, online erschienen am 24.08.2026. Verfasst von Sunyong Park und 2 weiteren Autor:innen. Die Arbeit nennt 92 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: National Research Foundation of Korea, Ministry of Agriculture, Food and Rural Affairs.
Abstract (Originalsprache)
Torrefaction and mild pyrolysis are widely used to improve the fuel quality of lignocellulosic biomass, but these processes induce coupled changes in mass yield, elemental composition, proximate composition, and higher heating value. Conventional single-stage machine-learning models can achieve high global prediction accuracy, but they may generate physically infeasible combinations of biomass fuel properties. In this study, a physically constrained multi-stage machine-learning model was developed to predict the fuel properties of thermally upgraded lignocellulosic biomass. The proposed model consisted of three sequential stages: mass yield prediction, elemental composition and higher heating value prediction using logarithmic enhancement factors, and proximate composition prediction with deterministic closure. A single-stage baseline model showed high apparent accuracy but produced proximate closure violations, mass-yield boundary violations, and directional inconsistencies in higher heating value. In contrast, the proposed multi-stage model maintained physically feasible predictions by construction, with cross-validated R2 values of 0.8992 for Stage 1, 0.8772 for Stage 2, and 0.8258 for Stage 3. The corresponding RMSE values were 6.3682%p for mass yield, 0.0729 in the Stage-2 logEF domain, and 0.1904 in the Stage-3 logEF domain. The unconstrained baseline produced proximate-closure violations in 40.24% of the evaluated samples, whereas the proposed framework eliminated such violations through deterministic closure. External validation using literature-derived biomass datasets showed that the model reproduced the major trends of carbon enrichment, oxygen removal, volatile matter reduction, and fixed carbon increase. These results indicate that physically constrained machine learning can provide a useful screening tool for estimating solid biofuel properties of torrefied and mildly pyrolysed lignocellulosic biomass.
Förderung
- National Research Foundation of Korea
- Ministry of Agriculture, Food and Rural Affairs
Zitieren
APA
Park, S., Oh, K. C., & Kim, D. (2026). Multi-Stage Machine Learning for Physically Consistent Prediction of Fuel Properties in Thermally Upgraded Lignocellulosic Biomass. BioEnergy Research, 19(1). https://doi.org/10.1007/s12155-026-11045-9
Vancouver
Park S, Oh KC, Kim D. Multi-Stage Machine Learning for Physically Consistent Prediction of Fuel Properties in Thermally Upgraded Lignocellulosic Biomass. BioEnergy Research. 2026;19(1). doi:10.1007/s12155-026-11045-9
BibTeX
@article{Park2026,
author = {Park, Sunyong and Oh, Kwang Cheol and Kim, DaeHyun},
title = {Multi-Stage Machine Learning for Physically Consistent Prediction of Fuel Properties in Thermally Upgraded Lignocellulosic Biomass},
journal = {BioEnergy Research},
year = {2026},
volume = {19},
number = {1},
doi = {10.1007/s12155-026-11045-9}
}
EUVIMED