摘要
深度学习凭借其高效的特征提取能力在工艺流程建模中得到广泛研究应用,然而对于工艺流程知识与深度学习融合及多尺度特征提取方面尚未有系统研究。为解决上述问题,构建了一种可容纳局部子流程单元操作的全流程矩阵以实现知识融合,并基于扩展的Inception卷积块提出了多尺度卷积神经网络(MSCNN)。柴油加氢裂化工业实验结果表明:基于知识融合的MSCNN对加氢精制催化剂温度和氢气耗量的预测均方根误差(RMSE)分别仅有0.75和1053,与传统卷积网络(CNN)、全连接网络(BPNN)相比显示出优越的预测性能。提出了一种用于特征评价的GMM-t-SNE框架,其可视化结果表明,MSCNN所提取特征的t-SNE分布与预测目标的GMM聚类分布一致,表明MSCNN提取的特征合理,显著提升了模型的预测性能。
Deep learning has been widely studied and applied in process modeling thanks to its efficient feature extraction capability.However,there is still less systematic research on the integration of process knowledge and deep learning as well as the multi-scale feature extraction.To address this issue,a full process matrix has been constructed to accommodate the local sub-process unit operation for the knowledge integration.Thus,a multi-scale convolutional neural network(MSCNN)has been proposed based on the extended-Inception convolutional block.The experimental results of diesel oil hydrocracking industry show that the root mean square error(RMSE)for predicting the hydrotreating catalyst temperature and hydrogen consumption by use of the knowledge fusion-based MSCNN is only 0.75 and 1053,respectively,which demonstrates superior forecasting accuracy as compared with the traditional convolution network(CNN)and fully connected network(BPNN).A GMM-t-SNE framework has also been proposed for feature evaluation.The visualization results show that the t-SNE distribution of features extracted by MSCNN is consistent with the GMM cluster distribution of prediction targets,revealing that the features extracted by MSCNN are reasonable,which have significantly improved the forecasting accuracy.
作者
王晨
罗文山
刘建华
陆鹏飞
曹晓红
蓝新志
陈晗冰
WANG Chen;LUO Wenshan;LIU Jianhua;LU Pengfei;CAO Xiaohong;LAN Xinzhi;CHEN Hanbing(CNOOC Huizhou Petrochemical Company,Huizhou 516086,China)
出处
《石油学报（石油加工）》
EI
CAS
CSCD
北大核心
2023年第3期532-543,共12页
Acta Petrolei Sinica(Petroleum Processing Section)
基金
中海油惠州石化科研项目(E-2421E002)基金资助。
关键词
加氢裂化
知识融合
多尺度卷积
特征评估框架
hydrocracking
knowledge fusion
multi-scale convolution
features evaluation framework
