Machine learning prediction of the yield and oxygen content of bio-oil via biomass characteristics and pyrolysis conditions

The bio-oil produced from biomass pyrolysis offers an important potential alternative to fossil fuels, but the yield and composition of pyrolysis product are impacted by many conditions. This work aims to predict the yield and oxygen content of bio-oil via machine learning tools based on biomass characteristics and pyrolysis conditions. For this purpose, the Random Forest (RF) algorithm is introduced and successfully applied. The performances of trained prediction models are assessed based on the regression coefficient (R2) for the test data. The results shows that the Proximate-Yield model (R2 = 0.925) has the best performance for predicting bio-oil yield, and the Ultimate-O model (R2 = 0.895) has the best performance for predicting the oxygen content of bio-oil. According to feature importance analysis, the heating rate occupied the biggest importance for predicting bio-oil yield, and the internal information of biomass is more important than that of pyrolysis conditions for predicting the bio-oil oxygen content. Besides, the modes of each variable affecting the bio-oil yield and oxygen content are described by partial dependence analysis. This work will provide a new insight for controlling the yield and oxygen content of bio-oil, which is helpful to facilitate the process optimization in engineering application. (c) 2022 Elsevier Ltd. All rights reserved.

Automated summary

This study aims to predict the yield and oxygen content of bio-oil using machine learning tools, with the Random Forest algorithm showing successful results. The Proximate-Yield model and Ultimate-O model exhibit the best performance in predicting bio-oil yield and oxygen content, respectively. Feature importance analysis highlights the significance of heating rate in predicting bio-oil yield, and the internal information of biomass in predicting bio-oil oxygen content over pyrolysis conditions.