炼油技术与工程 ›› 2021, Vol. 51 ›› Issue (12): 49-53.

• 计算机应用 • 上一篇    下一篇

基于RF-XGBoost算法的汽油辛烷值损失预测模型

陈亚丽1,苟苗苗1*,邵露娟1,邓丹2   

  1. 1.西南石油大学理学院，四川省成都市 610500； 2.西南石油大学计算机科学学院，四川省成都市 610500
  • 出版日期:2021-12-15 发布日期:2021-12-15
  • 通讯作者: 苟苗苗,E-mail：1713866471@qq.com
  • 作者简介:陈亚丽，教授，硕士研究生，从事在线与移动交互学习环境构建、教育大数据分析与应用研究

Prediction model of octane number loss  based on RF-XGBoost algorithm

Chen Yali1, Gou Miaomiao1, Shao Lujuan1, Deng Dan2   

  1. 1.School of Science, Southwest Petroleum University, Chengdu， Sichuan 610500; 2.School of Computer Science, Southwest Petroleum University, Chengdu， Sichuan 610500
  • Online:2021-12-15 Published:2021-12-15

PDF

摘要： 对某炼油企业催化裂化汽油精制脱硫装置的354个操作特征进行定性降维，得到与辛烷值损失相关的39个操作特征变量；再利用随机森林(RF)算法进行定量降维，得到28个辛烷值损失评价指标体系；最后使用极端梯度提升(XGBoost)算法及调优策略，建立基于RF-XGBoost算法的汽油辛烷值损失预测模型。结果表明：通过充分降维，得到的操作特征比较全面，能够作为刻画辛烷值损失的指标，并且符合企业实际优化条件，可以用于实际生产操作调控。建立的预测模型通过对相关参数调优，提高了预测精度。经过可视化分析，发现辛烷值的预测值和真实值较为接近，模型测试集的误差在0.13~0.27，决定系数约78.3%，需要进一步优化算法来提高测试集的预测精度。

关键词: RF-XGBoost算法, 汽油, 辛烷值损失, 预测模型, 操作特征, 调优策略

Abstract: The qualitative dimensionality reduction of 354 operating characteristics of FCC gasoline refining desulfurization unit in an oil refinery is carried out, and 39 operating characteristic variables related to octane number loss are obtained. Then the RF algorithm is used to quantitatively reduce the dimension, and 28 octane number loss evaluation index systems are obtained. Finally, the prediction model of gasoline octane number loss based on RF-XGBoost algorithm is established by using XGBoost algorithm and optimization strategy. The results show that the operation characteristics obtained by full dimensionality reduction are comprehensive, can be used as an index to describe the octane number loss, and meet the actual optimization conditions of enterprises, and can be used to regulate and control actual operation. The prediction accuracy of the model is increased by optimizing the relevant parameters. Through visual analysis, it is found that the predicted value of octane number is close to the real value, the error of the test set of the model is between 0.13~0.27, and the determination coefficient is about 78.3%. It is necessary to further optimize the algorithm to improve the prediction accuracy of the test set.

Key words: RF-XGBoost algorithm, gasoline, loss of octane number, prediction model, operating characteristics, adjust and optimize strategy