Abstract: By mining the experimental data in the literatures of fast pyrolysis of lignocellulosic biomass in a bubbling fluidized bed and establishing a random forest(RF) regression model, the yield of bio-oil, biochar, and gas via biomass pyrolysis was predicted based on biomass feedstock characteristics and pyrolysis conditions. Fifteen feature variables were sorted out from five key factors influencing the distribution of biomass pyrolysis products, and seven models were obtained by combining the input variables. All models showed good prediction performance for the three-state products from biomass pyrolysis, with a regression coefficient(R²) greater than 0.9. Model 6 had the fewest input variables and the highest accuracy, with R² values of 0.942 8, 0.956 1, and 0.939 1 for the yield predictions of biochar, bio-oil, and biomass pyrolysis gas, and the root mean square errors(RMSE) were 2.679 1, 2.939 5, and 3.108 3, respectively. Contribution analysis of the models revealed that pyrolysis conditions(Ⅴ) were the most important factors affecting the pyrolysis products yield, with contributions degree of 0.332 7, 0.220 4, and 0.214 7 for biochar, bio-oil, and gas yield predictions, respectively. Partial dependence plots(PDP) combined with the distribution boxplots analysis of each feature variable showed that pyrolysis temperature(HT), lignin mass fraction(Lig), and particle size(PS) were the main factors affecting biochar yield. Bio-oil and biomass pyrolysis gas yields were determined by HT, cellulose mass fraction(Cel), hemicellulose mass fraction(Hem), feed rate(FR), and gas flow rate(GFR), which were less affected by Lig and PS. The bio-oil yield could be improved by selecting biomass feedstock with high-quality fractions of cellulose and hemicellulose and increasing gas flow rate appropriately. In addition, the regression models of extreme gradient boosting(XGBoost), support vector machine(SVM), and artificial neural network(ANN) were established, which were compared with the RF regression model. The RF model showed the highest accuracy and good generalization ability for predicting the yield of the three-state products. The research findings promoted a comprehensive understanding of the biomass pyrolysis process and provided theoretical guidance for the control of the of the three-state products yield in biomass fast pyrolysis.