基于泄漏积分型回声状态网络的软测量动态建模方法及应用

摘要

提出一种基于泄漏积分型回声状态网络(LiESN)的软测量动态建模方法，给出 LiESN 的岭回归离线学习算法与递推最小二乘(RLS)在线学习算法。通过引入正则化系数，岭回归离线学习算法可有效地控制输出权值的幅值，改善 ESN 的预测性能。RLS 在线学习算法能适应大数据集的处理，满足过程建模实时性的需求。将基于 LiESN的软测量方法分别用于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成，实现对精炼厂相关产品质量的实时监控，并采用模型残差的四图分析对建模性能进行评价。在同等条件下，与基本的ESN网络以及支持向量机(SVM)等软测量建模方法进行了比较，结果表明，所提出的LiESN方法取得了很好的预测性能，计算精度满足工业生产的实际要求。%Soft sensor dynamic modeling based on echo state networks (ESN) with leaky-integrator neurons was proposed, in which the off-line learning algorithm using ridge regression and on-line learning algorithm using recursive least squares (RLS) were given respectively. By adding a regularization coefficient, ridge regression algorithm could control large sizes of output weight matrix and improve the properties of ESN solution. On-line learning algorithm could allow on-line processing of large data sets and attain requirement of real time for process modeling. Leaky integrator ESN (LiESN) was used to estimate the butane(C4) concentration in the bottom flow of a debutanizer column and to compute sulfur recovery unit (SRU) tail gas composition for improving product quality monitoring and control in a refinery. Simultaneously, the modeling performance was evaluated by 4-plot analysis of model residuals. Compared with existing soft sensor modeling such as ESN, support vector machines (SVM)etc, under the same condition, experimental results confirmed that LiESN could achieve better performance and the accuracy of the model could meet practical need.