Machine learning prediction of pyrolytic products of lignocellulosic biomass based on physicochemical characteristics and pyrolysis conditions

This study predicts pyrolytic product yields via machine learning algorithms from biomass physicochemical characteristics and pyrolysis conditions. Random forest (RF), gradient boosting decision tree (GBDT), eXtreme Gradient Boosting (XGBoost), and Adaptive Boost (Adaboost) algorithms are comparatively analyzed. Among these algorithms, the RF algorithm is the best modeling algorithm and performs best in predicting the bio-oil yield and performs well in predicting biochar and pyrolytic gas yields. The moisture content, carbon content, and final heating temperature are the most important factors in predicting pyrolysis product yields, and biomass characteristics are more important than pyrolysis conditions. Furthermore, the carbon content positively affects the bio-oil yield and negatively affects the biochar yield, and the final temperature positively affects the pyrolytic gas yield and negatively affects the biochar yield. This work provides new insight for controlling the yields of pyrolytic products via the RF algorithm, which can facilitate the process optimization in engineering applications.