基于深度学习的重质馏分油分子层次组成预测模型  

A Deep Learning Model for Predicting Molecular Level Composition of Heavy Distillate Oil

作　　者：袁壮 王源 杨哲 徐伟 周鑫 赵辉[2] 陈小博[2] 杨朝合[2] 林扬 YUAN Zhuang;WANG Yuan;YANG Zhe;XU Wei;ZHOU Xin;ZHAO Hui;CHEN Xiaobo;YANG Chaohe;LIN Yang(State Key Laboratory of Chemical Safety,SINOPEC Research Institute of Safety Engineering Co.,Ltd.,Qingdao 266000,China;State Key Laboratory of Heavy Oil Processing,China University of Petroleum,Qingdao 266580,China;College of Chemistry and Chemical Engineering,Ocean University of China,Qingdao 266100,China)

机构地区：[1]中石化安全工程研究院有限公司化学品安全全国重点实验室,山东青岛266000 [2]中国石油大学(华东)重质油全国重点实验室,山东青岛266580 [3]中国海洋大学化学化工学院,山东青岛266100

出　　处：《石油学报(石油加工)》2025年第2期362-370,共9页Acta Petrolei Sinica(Petroleum Processing Section)

基　　金：国家自然科学基金青年基金项目(22108307);化学品安全全国重点实验室开放课题资助。

摘　　要：随着工业大数据时代的到来,基于深度学习建立的原油分子组成预测模型具有适用范围广、构建快捷、准确性高等优点。然而,石油馏分分子层次信息标签获取困难,难以满足深度学习模型训练需求。为解决上述问题,基于商业流程模拟软件Aspen HYSYS与GC-MS×MS全二维气相色谱-飞行时间质谱联用仪提出了一种创新方法,建立足够规模的训练数据库。采用深度神经网络(DNN)建立了重质馏分油分子层次结构组成预测模型,该模型以炼油厂易测得的油品物理化学性质为输入,分子层次结构信息为输出,针对某炼油厂的催化裂化原料油进行分子组成预测,通过SHAP(SHapley Additive exPlanation)方法对模型进行可解释分析。结果表明,基于深度学习的重质馏分油分子组成预测模型能够准确地预测油品分子层次结构信息,目标装置原料分子组成预测平均相对误差小于8%。该模型不仅可对其他炼化装置的原料油性质进行软测量,也可为石油分子层次模型的开发提供准确的重油原料分子信息模型。With the advent of the industrial big data era,the prediction model for the molecular composition of crude oil,which is established based on deep learning techniques,has the advantages such as wide range of application,rapid construction,and high accuracy.However,it is challenging to obtain molecular level information labels for petroleum fractions,and thus it is difficult to meet the training requirements of deep learning models.To address these issues,an innovative method that utilizes commercial process simulation software,Aspen HYSYS,in conjunction with GC-MS×MS full two-dimensional gas chromatography-time-of-flight mass spectrometry(GC-TOF MS),was proposed to create a sufficiently large training database.A molecular level composition prediction model for heavy distillate oils was developed based on deep neural network(DNN).This model utilizes the physicochemical properties of easily measurable oils in the refinery as input and the molecular level information as output.The molecular compositions of catalytically cracked feedstock oils in a refinery were predicted,and an interpretable analysis was conducted on the model through SHAP(SHapley Additive exPlanation)method.The results indicate that the deep learning-based molecular composition prediction model for heavy distillate oils can accurately forecast the molecular level structure information of oils,with an average relative error of less than 8%in the molecular composition prediction for feedstocks in the target unit.In conclusion,this model not only facilitates the soft measurement of feedstock oil properties in other refining units,but also provides an accurate molecular information model of heavy oil feedstocks for the development of petroleum molecular level models.

关 键 词：重质馏分油 分子组成 深度学习 SHapley Additive exPlanation(SHAP)解释 分子管理 

分 类 号：TE624[石油与天然气工程—油气加工工程]