关键词: 生物质, 生物炭, 热解气, 焦油, 算法, 预测, 机器学习

Abstract:

To achieve the directional regulation of biomass pyrolysis products, a machine learning-based model for yield prediction and process optimization was developed. A dataset containing 578 valid samples was established by integrating literature and experimental data, and data preprocessing and feature selection were performed using boxplot and clustering algorithms. On this basis, Bayesian optimization was applied to tune four base learners, including Random Forest and Gradient Boosting Decision Tree models, and a high-accuracy yield prediction meta-model was constructed through a stacking ensemble strategy. The results showed that the coefficients of determination (R²) for biochar, tar, and pyrolysis gas yields reached 0.92, 0.90, and 0.94, respectively. Feature importance analysis using SHAP and GINI indicated that biochar yield was mainly influenced by pyrolysis temperature and carbon content, tar yield was primarily affected by lignin and hemicellulose contents, and gas yield was dominated by hemicellulose content and pyrolysis temperature. Furthermore, partial dependence plot (PDP) analysis revealed the optimal operating regions for product formation: low temperature and slow heating (500℃, HR < 20℃/min) favored biochar production, moderate temperature and short residence time (600℃,40 min < RT < 50 min) promoted tar enrichment, while high temperature with moderate heating rate (>750℃, 15℃/min < HR < 25℃/min) maximized gas yield, providing an efficient predictive tool and theoretical basis for the intelligent control and industrial application of biomass pyrolysis.

Key words: biomass, biochar, pyrolysis gas, tar, algorithm, prediction, machine learning