Machine learning prediction of pyrolytic sulfur migration based on coal compositions

Abstract

Understanding the sulfur migration during pyrolysis of coals especially high-sulfur coals is important. However, structural complexity and diversity of coals make it face huge challenge. In this study, a predictive model for morphological sulfur migration was developed using machine learning based on proximate analysis, ultimate analysis, sulfur forms of raw coal, ash composition, and blending ratio of coal. Three algorithms, i.e., Random Forest, XGBoost, and LightGBM were introduced and compared. The results show that six features are sufficient to accurately predict the products (R2 > 0.9, RMSE < 3.01%). LightGBM model has the advantages of better accuracy, generalization, efficiency, and performance, and Hyperopt has a higher upper limit than Grid-search. H content has a significant effect on S content in chars (St,dchar) and increasing H content from 5.0–5.3 wt% facilitates desulfurization. In addition, CaO, K2O and Fe2O3 also have remarkable effects on St,dchar. Higher H and volatile contents have a greater effect on thiophene removal in char. This work can provide a new approach to explore the sulfur migration in coal blending for coking.