石油学报(石油加工) ›› 2020, Vol. 36 ›› Issue (4): 756-766.doi: 10.3969/j.issn.1001-8719.2020.04.013

• 研究报告 • 上一篇    下一篇

基于LightGBM的催化重整装置产品预测及操作优化相关性分析

刘禹含, 曹萃文   

  1. 华东理工大学  化工过程先进控制和优化技术教育部重点实验室,上海 200237
  • 收稿日期:2019-05-17 修回日期:2019-12-14 出版日期:2020-07-25 发布日期:2020-09-23
  • 通讯作者: 曹萃文,女,副教授,博士,从事复杂工业系统建模,分析与控制 ,流程工业生产计划与生产调度技术,供应链管理与优化,工业系统可靠性分析等研究,Tel:021-64251607,E-mail:caocuiwen@ecust.edu.cn E-mail:caocuiwen@ecust.edu.cn
  • 作者简介:第一作者:刘禹含,女,硕士研究生,从事复杂工业系统建模,分析与控制,Tel:021-64251607,E-mail:18721379372@163.com
  • 基金资助:

    国家自然科学基金项目(61673175,61573144)资助

Product Prediction Technology and Optimal Operation Correlation Analysis for Catalytic Reforming Unit Based on LightGBM

 LIU Yuhan, CAO Cuiwen   

  1.  Key laboratory of Ministry of Education of Chemical Process Advanced Control and Optimization Technology, East China University of Science and Technology, Shanghai 200237, China
  • Received:2019-05-17 Revised:2019-12-14 Online:2020-07-25 Published:2020-09-23
  • Supported by:
     

摘要: 基于Aspen HYSYS软件建立了与某炼油厂有限实际生产数据相吻合的连续催化重整装置的机理模型;然后在此模型中考虑多种生产可能性,扩展数据范围得到完整的装置产品预测数据集;与常用的BP神经网络作对比,采用训练速度快、预测精度高、适合非线性过程建模的LightGBM决策树模型对该催化重整装置以4个反应器的温度和循环氢流量为特征变量,分别以戊烷、二甲苯、C6、重整汽油、氢气的流量和氢气纯度为目标建立了6个单目标数据驱动产品预测模型。通过对特征变量和目标之间的相关性分析,进行10折交叉验证,得到了特性变量的重要度排序,从而针对不同生产目标找出影响最大的操作变量。结果表明,使用LightGBM建立模型的预测准确度比BP神经网络的预测准确度有大幅度提升。

关键词: 催化重整;LightGBM决策树模型, 产品预测;特征重要度;相关性分析

Abstract: Based on the Aspen HYSYS software platform, the mechanism model of a continuous catalytic reforming unit was established in accordance with the limited real production data from a real-world refinery. Considering various production status on this model, the range of the production data were extended to a more complete data set for product prediction. Furthermore, the LightGBM decision tree model, which is suitable for nonlinear process modeling, was used to model the catalytic reforming unit. Compared with the BP neural networks commonly used in current similar studies and with the same training speed, the LightGBM models have shown the higher prediction accuracy through the 10-fold cross validation tests. The temperatures of the four reactors and the recycle hydrogen consumption were selected as characteristics and operational variables, and the products’ flow of pentane, xylene, C6, gasoline, hydrogen and the purity of hydrogen were objectives. Six single objective data-driven product prediction models are established. Finally, the ranks of the feature importance were obtained by the correlation analysis between the characteristics and the objectives. The obtained research results could provide decision reference for the refinery’s online optimal operations.

Key words: catalytic reforming, lightGBM decision tree model, product prediction, feature importance, correlation analysis

中图分类号: