机器学习模型预测煤热解产物分布研究OA北大核心CSTPCD

Pyrolysis technology is one of the primary methods for clean and efficient con-version and utilization of coal.However,due to the complex composition of coal and the challen-ges in controlling pyrolysis reactions,extensive time and effort are required to repeatedly investi-gate the pyrolysis product distribution for different types of coal.Taking coal composition and fi-nal pyrolysis temperature as input conditions and yield of coal pyrolysis products as output condi-tion,three machine learning models with adjusted parameters,that is,random forest(RF),support vector machine(SVM)and XGBoost model,were utilized to predict the yield of coal py-rolysis products.The importance of various input features on the yield of pyrolysis products was analyzed,and the effects of these input features were also quantified via two-factor partial de-pendence(PDP).The results show that the XGBoost model have the best performance among these three models for predicting the yield of coal pyrolysis products,and its decision coefficient and root mean square error for tar yield reach the highest 0.95 and the lowest 0.86,respectively.The important input features for coke yield are fixed carbon,volatiles,oxygen and pyrolysis final temperature,and their importance accounts for 84%.The tar yield are mainly affected by vola-tiles,hydrogen,nitrogen and pyrolysis final temperature,accounting for 83%of total impor-tance.The PDP analysis indicates that an increase in mass fraction of hydrogen and volatile in coal,the yield of coal pyrolysis tar will increase.According to the XGBoost model,the yield of pyrolysis tar exceeds 9%when the mass fraction of hydrogen is between 5.0%and 6.0%in coal,coupled with mass fraction of volatile between 30%and 40%.When mass fraction of the fixed carbon is greater than 50%,a decrease in mass fraction of hydrogen and volatile in coal,as well as a decrease in the final pyrolysis temperature,would result in a decline in the yield of py-rolysis gas.