Machine learning-based prediction and optimization of biomass pellet strength for sustainable bioenergy production

Abstract

Against the backdrop of growing emphasis on sustainable development and resource efficiency, biomass pellets have garnered significant attention due to their critical role in multiple fields. Accurate strength prediction is essential for optimizing processing techniques and improving resource utilization. However, current strength prediction methods still exhibit certain limitations. This study focuses on mixed pellets composed of corn stover (Zea mays L.) and potato (Solanum tuberosum L.) residue, employing machine learning for strength prediction. Through data collection, preprocessing, and analysis, various models were compared using the Particle Swarm Optimization (PSO) algorithm, ultimately selecting PSO-XGBoost as the base model. The model achieved an R² of 0.72 on the test set, demonstrating strong generalization capability. To further enhance predictive performance, a novel SVBoost model was developed, which improved the R² from 0.72 to 0.93 (an increase of 0.21) and reduced the prediction error by 34.78 %. Additionally, the interpretability of the SVBoost model was analyzed using SHAP and LIME methods to elucidate the influence mechanisms of different experimental conditions on pellet strength. The optimal forming process parameters were determined as follows: force (30–35 kN), compression speed (10 mm/min), mixing ratio (1:3), temperature range (80–120 °C), feed amount (35 g), and holding time (1200 s). These findings provide theoretical support for optimizing biomass pellet forming conditions and quality control.

Topics & Concepts