融合可解释机器学习的成品汽油调和配方质量预测评价与致因分析  

Predictive Evaluation and Cause Analysis of Finished Gasoline Blending Formulation Quality by Interpretable Machine Learning

作　　者：李炜[1,2,3] 郑明杰 李亚洁[1,2,3] 梁成龙 LI Wei;ZHENG Mingjie;LI Yajie;LIANG Chenglong(College of Electrical and Information Engineering,Lanzhou University of Technology,Lanzhou 730050,China;Key Laboratory of Gansu Advanced Control for Industrial Processes,Lanzhou 730050,China;National Demonstration Center for Experimental Electrical and Control Engineering Education,Lanzhou University of Technology,Lanzhou 730050,China;Oil Storage and Transportation Plant,PetroChina Lanzhou Petrochemical Company,Lanzhou 730060,China)

机构地区：[1]兰州理工大学电气工程与信息工程学院,甘肃兰州730050 [2]甘肃省工业过程先进控制重点实验室,甘肃兰州730050 [3]兰州理工大学电气与控制工程国家级实验教学示范中心,甘肃兰州730050 [4]中国石油兰州石化分公司油品储运厂,甘肃兰州730060

出　　处：《石油学报（石油加工）》2024年第1期126-136,共11页Acta Petrolei Sinica(Petroleum Processing Section)

基　　金：甘肃省青年博士基金项目(2021QB-044)资助。

摘　　要：受成品汽油调和配方需“先验”评价与修正的驱动,本研究将轻量级梯度提升树(LightGBM)与可解释机器学习(SHAP)方法相结合,兼顾复杂模型精度高与后验SHAP可解释性强的各自优势,提出了一种调和配方质量预测评价及致因分析方法。该方法先引用改进遗传算法(IGA)优化LightGBM的超参数,建立了可同时预测成品汽油性能和环保指标的模型,并结合汽油国ⅥA标准与企业生产实际制定了配方质量评价标准,实现配方“先验”评价;再基于SHAP的全局和局部致因分析,对缺陷配方给出了易于操作的单变量定性修正建议。实验结果表明:相比于传统BP网络和随机森林(RF)、以及采用随机搜索和GA优化参数的LightGBM等模型,IGA_LightGBM模型可得到更全面和精准的预测指标,SHAP致因分析可给出契合实际的修正建议。该方法是智能算法代替人工的有益探索。Driven by prior evaluation and correction of finished gasoline blending formula,this study proposes a predictive evaluation and causal analysis method for gasoline blending formulation quality based on combining the light gradient boosting machine(LightGBM)with Shapley additive explanation(SHAP)interpretable machine learning and considering the advantages of high precision of complex models and strong post-hoc-SHAP interpretability.This method first optimizes the hyper-parameters of LightGBM by introducing the improved genetic algorithm(IGA),establishes a model that can simultaneously predict the performance and environmental indicators of finished gasoline,and then formulates the formula quality evaluation standards in light of nationalⅥA standard and the actual production in factories to realize the prior evaluation of formula.Besides,the easy-to-operate univariate correction scheme for defective blending formula is proposed based on the global and local causes analysis of SHAP.The experimental results show that the IGA_LightGBM-based model can present more comprehensive and accurate predictors as compared with the traditional back propagation(BP)and random forest(RF)based model,and the LightGBM model with hyperparameters optimazed by random search and normal GA.The SHAP causal analysis can provide practical correction schemes.This method can be considered as a helpful exploration in applying the intelligent algorithms instead of human experiences.

关 键 词：成品汽油调和 配方质量评价 可解释机器学习 预测建模 致因分析 参数优化 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]