机器学习模型预测煤热解产物分布研究  

Prediction of coal pyrolysis product distribution using machine learning model

作　　者：王敏欣 师印光 刘长波 吴雷[2] 周军[2] 蒋绪[3] WANG Minxin;SHI Yinguang;LIU Changbo;WU Lei;ZHOU Jun;JIANG Xu(Energy Saving and Environmental Protection Company Limitied,MCC Group,100088 Beijing,China;School of Chemistry and Chemical Engineering,Xi’an University of Architecture and Technology,710055 Xi’an,China;Research Institute of Energy and Chemical Industry,Xianyang Vocational Technical College,712000 Xianyang,China)

机构地区：[1]中冶节能环保有限责任公司,北京100088 [2]西安建筑科技大学化学与化工学院,西安710055 [3]咸阳职业技术学院能源化工研究所,陕西咸阳712000

出　　处：《煤炭转化》2024年第4期11-22,共12页Coal Conversion

基　　金：陕西省自然科学基础研究计划一般项目(2024JC-YBQN-0090);陕西省教育厅服务地方专项计划项目(22JC045)。

摘　　要：热解技术是煤清洁高效转化利用的主要技术之一,然而由于煤成分复杂、热解反应难以控制,对于各煤种都需要消耗大量时间和精力重复进行热解产物分布研究。以煤的组成和热解终温为输入条件,煤热解产物收率为输出条件,采用机器学习模型调参后的随机森林(RF)、支持向量机(SVM)和极限梯度提升(XGBoost)算法对煤热解产物收率进行了预测,分析了不同输入特征对热解产物收率的重要性,并使用双因素部分依赖分析(PDP)评估了输入特征对热解产物收率的影响。结果表明:三种算法模型中,XGBoost算法对煤热解产物收率预测的性能最佳,其对焦油收率的决策系数(R^(2))和均方根误差(R_(MSE))分别达到最高的0.95和最低的0.86;焦炭收率主要受到固定碳、挥发分、氧元素和热解终温的影响,占比为84%,而焦油收率主要受挥发分、氢元素、氮元素和热解终温的影响,占比为83%;PDP分析表明,随着煤中氢原子和挥发分质量分数的增加,煤热解焦油收率会随之上升;根据XGBoost算法计算可知,当煤中氢元素的质量分数在5.0%~6.0%,挥发分的质量分数位于30%~40%时,热解焦油收率可达到9%以上,当固定碳的质量分数大于50%时,煤中氢元素和挥发分的质量分数减少以及热解终温下降,都会引起热解气收率的降低。Pyrolysis technology is one of the primary methods for clean and efficient conversion and utilization of coal.However,due to the complex composition of coal and the challenges in controlling pyrolysis reactions,extensive time and effort are required to repeatedly investigate the pyrolysis product distribution for different types of coal.Taking coal composition and final pyrolysis temperature as input conditions and yield of coal pyrolysis products as output condition,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 pyrolysis 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 dependence(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 volatiles,hydrogen,nitrogen and pyrolysis final temperature,accounting for 83% of total importance.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 pyrolysis gas.

关 键 词：煤热解 产物收率 机器学习 重要特征性分析 部分依赖性分析 

分 类 号：TQ530.2[化学工程—煤化学工程] TP181[自动化与计算机技术—控制理论与控制工程]