石油学报(石油加工) ›› 2020, Vol. 36 ›› Issue (1): 179-187.doi: 10.3969/j.issn.1001-8719.2020.01.0.22

• 研究报道 • 上一篇    下一篇

基于GBDT和新型P-GBDT算法的催化裂化装置汽油收率寻优模型的构建与应用

王伟1, 汪坤1,杨帆2,戴超男2,金继民2,金宝宝2   

  1. 1中国石化武汉分公司,湖北 武汉 430082;2联想大数据智能应用实验室,四川 成都 610041
  • 收稿日期:2018-11-12 修回日期:2019-05-31 出版日期:2020-01-25 发布日期:2020-04-01
  • 通讯作者: 杨帆,男,研究员,硕士,从事大数据及工业智能相关研究工作,Email:yangfan24@lenovo.com E-mail:yangfan24@lenovo.com

Construction and Analysis of Gasoline Yield Prediction Model for Fluid Catalytic Cracking Unit (FCCU) Based on GBDT and P-GBDT Algorithm

WANG Wei1, WANG Kun1, YANG Fan2, DAI Chaonan2, JIN Jimin2, JIN Baobao2   

  1. 1SINOPEC Wuhan Company, Wuhan 430082, China; 2Data Intelligence Application Laboratry,Lenovo Group, Chengdu 610041, China
  • Received:2018-11-12 Revised:2019-05-31 Online:2020-01-25 Published:2020-04-01

摘要: 催化裂化装置是一个高度非线性和相互强关联的多变量系统,基于数据挖掘技术的分析方法是优化该工艺过程的一类有力工具。笔者利用某石油化工企业集散控制系统(Distributed control system, DCS)和实验室信息管理系统(Laboratory information management system, LIMS)的工业生产实时数据,分别从指标与汽油收率的正负相关性、工业经验以及模型重要性筛选等方面选取了182个关键影响参数,利用梯度提升决策树(GBDT)算法构建催化裂化汽油收率的预测模型,预测相应的汽油收率。基于GBDT集成学习框架构建了P-GBDT模型,引入了特征扰动和特征权重,增大经验可控参数的权重,解决了普通GBDT模型对特征缺乏偏好、经验可控参数特征的权重较小的问题。结果显示,由P-GBDT算法构建的汽油收率预测模型预测结果的准确率、R2、均方根误差等指标相比由GBDT算法构建的基准模型的预测结果明显更好,对真实收率的拟合效果更为接近,对优化改进实际可控装置操作条件具有更好的指导意义。

关键词: P-GBDT算法, 催化裂化, 收率寻优模型, 人工智能算法

Abstract: Catalytic cracking unit is a highly nonlinear and strongly correlated system. Currently, the data mining techniques are powerful analytical methods for optimizing this process. Based on the industrial production data collected by the laboratory information management system and the distributed control system, 182 key indicators are selected and a gasoline yield prediction model is built based on gradient-growth decision tree (GBDT algorithm). Afterwards, a PGBDT model is constructed with reference to the GBDT framework, by introducing feature disturbances and feature weights, and increasing the weight of empirically controllable parameters. The results show that P-GBDT has significant higher accuracy with smaller R2 value and the root mean square error.

Key words: P-GBDT algorithm, fluidic catalytic cracking, product yield optimization model, artificial intelligence algorithms

中图分类号: