Precision biochar yield forecasting employing random forest and XGBoost with Taylor diagram visualization

Abstract Waste-to-energy conversion via pyrolysis has attracted increasing attention recently owing to its multiple uses. Among the products of this process, biochar stands out for its versatility, with its yield influenced by various factors. Extensive and labor-intensive experimental testing is sometimes necessary to properly grasp the output distribution from various feedstocks. Nonetheless, data-driven predictive models using large-scale historical experiment records can provide insightful analysis of projected yields from a variety of biomass materials, hence overcoming the challenges of empirical modeling. As such, five modern approaches available in modern machine learning are employed in this study to develop the biochar yield prediction models. The Lasso regression, Tweedie regression, random forest, XGBoost, and Gradient boosting regression were employed. Out of these five XGBoost was superior with a training mean squared error (MSE) of 1.17 and a test MSE of 2.94. The XGBoost-based biochar yield model shows excellent performance with a strong predictive accuracy of the R 2 values as 0.9739 (training) and 0.8875 (test). The mean absolute percentage error value was only 2.14% in the training phase and 3.8% in the testing phase. Precision prognostic technologies have broad effects on sectors including biomass logistics, conversion technologies, and effective biomass utilization as renewable energy. Leveraging SHAP based on cooperative game theory, the study shows that while ash and moisture lower biochar yield, FPT, nitrogen, and carbon content significantly boost it. Small variables like heating rate and volatile matter have a secondary impact on production efficiency.