基于多神经网络结构的丙烯浓度软测量建模  

Soft sensor modeling for propylene concentration based on MNN

作　　者：张斌 

机构地区：[1]中国石油化工股份公司武汉分公司,湖北武汉430082

出　　处：《计算机与应用化学》2014年第3期374-376,共3页Computers and Applied Chemistry

摘　　要：大型气体分馏装置的分馏塔丙烯浓度的软测量一般采用统计建模法进行,根据丙烯精馏塔的工艺流程和控制系统结构的分析,确定了塔顶的温差、双温差、回流量变化以及塔顶压力变化△p作为软测量模型所需的二次变量,利用气相色谱仪进行线下化验分析获得的丙烯浓度和对应采样时刻的塔顶、塔底温度、回流量、塔顶压力等数据,通过随机抽取的方法生成含训练样本集和测试样本集。采用训练样本全集和经过K-means聚类后的数据,建立了由10,15,20个隐层节点构成的单隐层多层前向网络模型;利用Levenberg-Marquardt训练算法,通过LS连接得到软测量模型,利用离线色谱分析仪的化验分析结果作为校正信号。仿真结果表明软测量模型具有较好的精确度和鲁棒性。Soft sensor for propylene concentration in the distillation tower of large-scale Gas Distillation Unit usually adopts statistical modeling. According to analysis of process flow and control system structure for propylene rectification tower, overhead temperature difference, dual temperature difference, reflux difference and overhead pressure difference Ap are determined as secondary variables required by soft sensor modeling. The propylene concentration achieved by gas chromatograph, overhead temperature, bottom temperature, reflux amount and overhead pressure can generate training sample set and testing sample set through random selection. Based on training sample set and data clustered by Kmeans, we can establish the model for multilayer neural network of single hidden layer constructed by 10, 15, 20 hidden layer neurons. Using Levenberg-Marquardt Training Algorithm, soft sensor modeling is achieved through LS connection. The laboratory analysis result obtained by off-line chromatograph is regarded as calibration signal. The simulation result demonstrates that soft sensor modeling has good accuracy and robustness.

关 键 词：丙烯浓度 软测量模型 多神经网络结构 K-MEANS聚类 

分 类 号：TQ018[化学工程] TQ03