文献类型：期刊文献

中文题名：基于滑动窗决策树的加氢裂化装置过渡状态识别

英文题名：Transition State Identification of Hydrocracking Unit Based on a Sliding Window Decision Tree

作者：曹跃[1];余冲[3];纪晔[4];杨明磊[1,2];李智[1]

机构：[1]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237;[2]华东理工大学过程系统工程教育部工程研究中心,上海200237;[3]上海交通大学溥渊未来技术学院,上海200240;[4]中国石油天然气股份有限公司规划总院,北京100083

年份：2025

卷号：41

期号：1

起止页码：187

中文期刊名：石油学报(石油加工)

外文期刊名：Acta Petrolei Sinica(Petroleum Processing Section)

收录：;EI(收录号：20250317712933);北大核心:【北大核心2023】；

基金：国家重点研发计划政府间国际科技创新合作项目(2021YFE0112800);国家自然科学基金项目(62203173);全国博士后创新人才支持计划项目(BX20220105);上海市浦江人才计划(2022PJD018)资助。

语种：中文

中文关键词：加氢裂化装置;过渡状态;滑动窗口;特征矩阵;决策树分类;可解释性

外文关键词：hydrocracking unit;transition state;sliding window;feature matrix;decision tree classification;interpretability

摘要：加氢裂化生产装置处于多工况运行状态,而不同工况间切换存在过渡状态,操作员会根据装置所处状态进行相应的操作和调整。然而,装置所处的过渡状态难以识别,需要长期操作学习并积累经验。为此,提出了一种基于滑动窗决策树的加氢裂化装置过渡状态识别方法。加氢裂化装置工业数据经去噪、降维等预处理后,使用滑动窗口保留窗口内的数据局部动态时序特征,并建立特征矩阵,再利用精细决策树发掘复杂过程变量之间的关系,可视化地描述了决策树结构,体现其可解释的优势,最终实现加氢裂化装置过渡态的快速、准确识别。基于F1分数,对比了高斯朴素贝叶斯、精细高斯支持向量机、粗略树、中等树、精细树、可优化决策树对加氢裂化装置过渡态的综合识别性能,10次五折交叉验证后,基于精细树的F1分数均值可达0.9896,训练时间均值为3.028 s。
The hydrocracking unit often runs under multi-operating conditions,and there exist transition states during mode switching,in which case workers will make corresponding operations and adjustments according to the current state of the unit.However,it is difficulty to identify the transition state of the unit,which relies on long-term practical operation and learning as well as experience accumulation.Therefore,a method for identifying the transition state of the hydrocracking unit based on a sliding window decision tree is proposed.After the industrial data of the hydrocracking unit has been preprocessed by denoising and dimensionality reduction,a sliding window is used to retain the local dynamic time series characteristics of the data within the window,and a feature matrix is established.Then a fine decision tree is used to explore the relationships between complex process variables and visualize its structure to represent its interpretable advantage.Finally,the transition state of the hydrocracking unit can be identified rapidly and accurately.Based on F1 score,the paper compares the comprehensive identification performances of Gaussian naive Bayes,fine Gaussian support vector machine(SVM),coarse decision tree,medium decision tree,fine decision tree,and optimizable decision tree for the transition state of the hydrocracking unit.After 5-fold cross validation for 10 times,the average F1 score based on the fine decision tree can reach 0.9896,and the average training time is 3.028 s.