基于ESN的化工过程软测量建模研究

发布时间：2018-04-15 00:40

本文选题：软测量 + 动态建模　；参考：《兰州交通大学》2015年硕士论文

【摘要】：软测量建模是目前化工过程领域的热点研究方向。软测量建模方法能够有效地解决生产过程中在线分析仪表测量滞后大、价格昂贵、维护保养复杂等问题。对于具有强非线性动态系统的生产过程,动态软测量建模方法能够有效地改善以往静态软测量建模方法建模精度差、难以满足实际生产过程实时性需求的问题。目前,动态软测量建模中,神经网络、SVM等计算智能方法是主要的建模工具。回声状态网络(echo state networks,ESN)作为一种新的动态递归神经网络(RNN)已引起广泛关注。ESN的隐层结构由具有回声状态特性(Echo State Property,ESP)的状态储备池(State Reservoir,SR)构成,在网络学习训练时只需计算网络输出权值,具有很强的动态逼近能力。因此,ESN的学习算法具有计算简单有效、收敛速度快的优点。针对动态软测量建模方法,在ESN的基础上,本文研究了一类泄漏积分型回声状态网络(Leaky integrator Echo State Networks,LiESN)方法,给出了相应的离线、在线学习算法及其改进的学习算法。将本文方法应用到精炼厂化工生产过程实例中,实现了对过程变量的实时监控,实验结果表明所提出的方法具有良好的建模效果和应用潜力。主要研究内容概括如下:(1)研究动态软测量的建模方法,重点分析基于神经网络与时间序列模型相结合的动态软测量方法,运用统计学的四图分析对建模效果进行评价。(2)在ESN网络基本理论基础上,研究一类LiESN方法,包括基于全局参数约束的LiESN方法及其梯度下降学习算法;一种扩展的LiESN方法及其相应的岭回归离线学习算法和递推最小二乘(Recursive Least Squares,RLS)在线学习算法。(3)将所研究的LiESN动态软测量建模方法与非线性滑动平均(Nonlinear Moving Average,NMA)、非线性自回归(Nonlinear Autoregressive,NARX)时间序列动态模型结构相结合,应用于一类强非线性化工过程实例中,以验证所提出方法的有效性。具体包括预测脱丁烷塔底部丁烷组分的含量及计算硫回收装置中尾气的组分含量进行软测量建模。在同等条件下,为进一步衡量本文方法的有效性,通过与基本的ESN动态软测量建模方法及现有的SVM、LS-SVM、前馈神经网络等软测量建模方法进行对比,实验结果证实,所提出方法具有很高的建模精度,网络具有快速、稳定的学习速度。研究成果对有效地提高产品质量,实现对污染物的最小化排放有重要意义。
[Abstract]:Soft sensor modeling is a hot research direction in chemical process field.The soft sensor modeling method can effectively solve the problems of large measurement lag, high price and complex maintenance in the process of production.For the production process with strong nonlinear dynamic system, dynamic soft sensor modeling method can effectively improve the modeling accuracy of the previous static soft sensor modeling method, which is difficult to meet the real time requirement of the actual production process.At present, neural network and SVM are the main modeling tools in dynamic soft sensor modeling.Echo state networks (ESNs), as a new dynamic recurrent neural network (RNNs), have attracted wide attention. The hidden layer structure of Essn is composed of state reserve pool (Echo State property ESPs) with echo state characteristics (Echo State property ESPs).Network learning and training only need to calculate the network output weights, which has a strong dynamic approximation ability.Therefore, the learning algorithm of ESN has the advantages of simple calculation and fast convergence.Based on ESN, a class of leaky integrator Echo State networks based on ESN is studied in this paper. The corresponding off-line and on-line learning algorithms and their improved learning algorithms are given.The method is applied to the example of chemical production process in refineries and the real-time monitoring of process variables is realized. The experimental results show that the proposed method has good modeling effect and application potential.The main research contents are summarized as follows: (1) the modeling method of dynamic soft sensing is studied, and the dynamic soft sensing method based on the combination of neural network and time series model is analyzed.Based on the basic theory of ESN network, this paper studies a class of LiESN methods, including the LiESN method based on global parameter constraints and its gradient descent learning algorithm.An extended LiESN method and its corresponding off-line learning algorithm for ridge regression and Recursive Least Squares (RLS) online learning algorithm.Nonlinear autoregressive (NARX) time series dynamic model structure,The method is applied to a class of strongly nonlinear chemical processes to verify the effectiveness of the proposed method.It includes predicting the content of butane at the bottom of debutane tower and calculating the content of tail gas in sulfur recovery unit.Under the same conditions, in order to further evaluate the effectiveness of this method, compared with the basic ESN dynamic soft-sensor modeling method and the existing soft-sensor modeling methods such as SVML-LS-SVM, feedforward neural network, the experimental results show that,The proposed method has high modeling accuracy and fast and stable learning speed.The research results are of great significance to improve the product quality and minimize the emission of pollutants.
【学位授予单位】：兰州交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TQ018;TP274

【相似文献】