基于VMD-GRU的润叶过程片烟水分集成预测方法  

Dynamic prediction method of tobacco leaf moisture in the leaf wetting process based on VMD and GRU network integration

作　　者：张雷 任国峰 洪斌斌 邹泉 郑红艳 赵云川 徐大勇[2] 堵劲松[2] 李银华 苏子淇 熊开胜 ZHANG Lei;REN Guofeng;HONG Binbin;ZOU Quan;ZHENG Hongyan;ZHAO Yunchuan;XU Dayong;DU Jinsong;LI Yinhua;SU Ziqi;XIONG Kaisheng(School of Electrical and Information Engineering,Zhengzhou University of Light Industry,Zhengzhou 450001,China;China Tobacco Corporation,Zhengzhou Tobacco Research Institute,Zhengzhou 450001,China;China Tobacco Yunnan Industry Co.,Ltd.,Kunming 650231,China)

机构地区：[1]郑州轻工业大学电气信息工程学院,郑州市金水区东风路5号450001 [2]中国烟草总公司郑州烟草研究院,郑州国家高新技术产业开发区枫杨街2号450001 [3]云南中烟工业有限责任公司,昆明市世博路6号650231

出　　处：《中国烟草学报》2025年第2期58-65,共8页Acta Tabacaria Sinica

基　　金：国家烟草专卖局科研大数据重大科技项“卷烟制丝加工大数据关键技术研究与应用”(110202101083(SJ-07));河南省科技攻关计划项目(242102220028);河南省博士后科研项目(202102087)。

摘　　要：润叶过程中出口烟叶的水分是重要质量指标,然而润叶过程具有多变量、非线性、非平稳等特点,给水分预测带来了巨大挑战,本研究提出了一种基于变分模态分解(Variational Mode Decomposition, VMD)和门控循环单元(Gated Recurrent Units, GRU)的集成预测方法。首先,利用VMD对烟叶水分含量进行分解,得到若干本征模态函数(Intrinsic Mode Function, IMF)。然后,针对不同尺度的模态分量,建立相应的GRU网络以提取多尺度特征。同时,设计并行GRU网络提取过程变量与烟叶水分之间的复杂时序依赖关系。最后,将所有GRU网络的输出隐藏状态进行拼接,并通过全连接层进行进一步特征提取和水分预测。研究结果表明,在某复烤厂实际生产数据集上,VMD-GRU的预测结果较传统预测方法提高了平均40%的预测准确率,特别是在多步预测上精度优势明显,证明了算法的有效性和优越性。The moisture content of outlet tobacco leaves in the leaf wetting process is an important quality indicator.However,the wetting process is characterized by multivariable,nonlinear,and non-stationary features,posing significant challenges to moisture prediction.This study proposes an integrated prediction method based on Variational Mode Decomposition(VMD)and Gated Recurrent Units(GRU).First,VMD is utilized to decompose the moisture content of tobacco leaves,obtaining several Intrinsic Mode Functions(IMF).Then,GRU networks are established for modal components at different scales to extract multi-scale features.Concurrently,parallel GRU networks are designed to capture the complex temporal dependencies between process variables and tobacco leaf moisture.Finally,the hidden states output by all GRU networks are concatenated and further feature extraction and moisture prediction are performed through a fully connected layer.On a real production dataset from a rotary kiln factory,the results indicate that the VMD-GRU method improves the prediction accuracy by an average of 40%compared to traditional prediction methods,especially demonstrating significant precision advantages in multi-step predictions,thereby proving the algorithm's effectiveness and superiority.

关 键 词：烟叶水分预测 变分模态分解 门控循环单元网络 润叶过程 软测量 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.4[自动化与计算机技术—控制科学与工程]