Machine learning-based modeling approaches for estimating pyrolysis products of varied biomass and operating conditions

The pyrolysis products of different biomass are difficult to predict due to the complex biomass properties and wide range of operating conditions. In this study, machine learning techniques based on artificial neural networks, gradient boosting, decision trees, random forest, K-nearest-neighbors, bagging regressor, and lasso regression were employed to develop different predictive models for char, liquid/bio-oil, and gas product yields estimation. The performance of these models was evaluated by R-2 score. All models performed well (R-2 approximate to 0.90) except the lasso model (R-2 < 0.90). Gradient boosting gave better predictions with R-2 > 0.90 in single output models. The bagging regressor showed the best performance in multiple output models. Relative importance analysis showed that cellulose content, ash content, carbon element, and temperature had significant effects on pyrolysis products. The results show that the machine learning-based approach is a viable alternative for predicting the product yields from varied biomasses and operating conditions.