Neural network-based prediction of levoglucosan yield: A novel modeling approach

Levoglucosan, as a high-value-added product from biomass pyrolysis, has become a focal point in biomass energy utilization due to its significant potential in various applications. In this study, nine variables were identified via Pearson analysis, and five neural network models predicted levoglucosan yield under various pyrolysis conditions. Results showed that the LM-BPNN fit exhibited excellent prediction accuracy (R2>0.9, MAE<3, RMSE<5, MSE<25). Through SHAP analysis, the importance and correlation between experimental parameters and levoglucosan yield were clarified, with cellulose, as the main material for levoglucosan production, having the most significant positive impact on its yield. In addition, partial dependence analyses further explored the synergistic effects of biomass characteristics, reaction conditions and biomass pretreatment on levoglucosan yield. Among them, pyrolysis temperature, pyrolysis time, and biomass particle size all showed negative feedback with yield after reaching their peak ranges. In these analyses, the interactions between input features and output targets of the LM-BPNN model were explored to analyze in depth the regulation of experimental parameters for efficient levoglucosan production. This study highlights the potential of algorithm-enhanced neural networks in predicting high-value pyrolysis products, offering innovative pathways and scientific support for efficient solid waste utilization and advancing sustainable waste-to-energy technologies.