文献类型：期刊文献

中文题名：Hybrid model of multimodal based on data enhancement and lumped reaction kinetics: Applying to industrial ebullated-bed residue hydrogenation unit

作者：Jian Long[1];Mengru Zhang[1];Anlan Li[2];Cheng Huang[1];Dong Xue[1]

机构：[1]Key Laboratory of Smart Manufacturing in Energy Chemical Process,Ministry of Education,East China University of Science and Technology,Shanghai 200237,China;[2]Sinopec Engineering Incorporation,Beijing 100101,China

年份：2025

卷号：78

期号：2

起止页码：284

中文期刊名：Chinese Journal of Chemical Engineering

外文期刊名：中国化学工程学报(英文版)

基金：supported by National Natural Science Foundation of China(Basic Science Center Program:61988101);National Natural Science Foundation of China(62394345,62373155,62173147);the Major Science and Technology Project of Xinjiang(No.2022A01006-4).

语种：英文

中文关键词：Mixed modeling;Generative adversarial network;Lumped kinetic model;Multi-modal learning;Ebullated-bed residue hydrogenation

摘要：Industrial ebullated-bed is an important device for promoting the cleaning and upgrading of oil products. The lumped kinetic model is a powerful tool for predicting the product yield of the ebullated-bed residue hydrogenation (EBRH) unit, However, during the long-term operation of the device, there are phenomena such as low frequency of material property analysis leading to limited operating data and diverse operating modes at the same time scale, which poses a huge challenge to building an accurate product yield prediction model. To address these challenges, a data augmentation-based eleven lumped reaction kinetics mechanism model was constructed. This model combines generative adversarial networks, outlier elimination, and L2 norm data filtering to expand the dataset and utilizes kernel principal component analysis-fuzzy C-means for operating condition partitioning. Based on the hydrogenation reaction mechanism, a single and sub operating condition eleven lumped reaction kinetics model of an ebullated-bed residue hydrogenation unit, comprising 55 reaction paths and 110 parameters, was constructed before and after data augmentation. Compared to the single model before data enhancement, the average absolute error of the sub-models under data enhancement division was reduced by 23%. Thus, these findings can help guide the operation and optimization of the production process.