【作者】 孙娅苹；

【导师】 曹柳林；

【作者基本信息】 北京化工大学 ， 控制科学与工程， 2010， 硕士

【摘要】 间歇反应过程在化学工业过程中占有十分重要的地位。由于间歇过程具有很强的非线性、缺少稳态操作条件、反应过程的不确定性、未知扰动、过程变量的限制和有限的在线测量信息等特性,不易建立准确的数学模型。本文研究的灰箱建模方法,就是针对这一类复杂的非线性动态系统的建模问题而进行,以制备橡胶硫化促进剂的间歇缩合反应过程作为建模对象进行模型验证。本文的主要内容包括：以机理分析和神经网络建模为基础,结合灰箱建模思想,研究了误差补偿的灰箱建模方法(Grey-box Modeling with Error Compensation, GMEC)。该方法首先利用了系统分解思想,将复杂系统分解为线性部分和非线性部分；其次,针对系统特性,在机理分析基础上,采用对角递归神经网络(DRNN)建立线性部分的系统模型；然后,结合误差补偿思想将实际系统与线性部分模型输出的差值作为非线性部分的神经网络的训练目标,同时建立线性部分与非线性部分的关联关系,从而建立系统的输入输出模型。本文提出了一种新的灰箱建模方法：基于反应基元(Fundamental Genes, FG)的建立复杂非线性系统模型的灰箱建模方法(Grey-box Modeling Based on Fundamental Genes, GMFG)。该方法首先根据先验知识及系统特性分析引入过程的初始反应基元,并以此为出发点建立结构逼近神经网络模型(Structure Approaching Neural Network, SANN),实现基元之间的关联,赋予网络节点实际的物理意义；然后通过提出的最小化预测误差,结合逐步回归分析方法(Stepwise Regression Analysis, SRA)选择最优反应基元,优化网络结构,建立起表示系统变量关系的灰箱模型。在基于反应基元的灰箱建模方法的基础之上进行改进,提出了改进的基于反应基元的智能灰箱建模方法(Improved Grey-box Modeling Based on Fundamental Genes, IGMFG)。该智能灰箱建模方法首先构建候选反应基元池,然后通过粒子群优化算法优化选择最优反应基元来建立系统模型。在建模过程中,保留GMFG的优点的同时,对GMFG建模方法的初始化反应基元和优化选择最优基元这两大方面进行了改进。以实验室制备橡胶硫化促进剂的间歇缩合反应过程作为建模对象,对上述三种灰箱建模的仿真结果,从建模精度、“白箱化”程度,以及优缺点等方面进行分析和比较,并与黑箱建模的并联神经网络建模结果进行比较,表明了这三种灰箱建模方法的可行性和高效性,同时也证明了引入反应基元的建模思想对化工过程或对象的“白箱化”建模的创新性和重要性。最后的部分是对灰箱建模中还需深入研究的问题进行总结和展望。更多还原

【Abstract】 Batch reactor plays an important role in the chemical industry. Batch reactor modeling is the prerequisite and basis of batch reactor design, optimization and process control. However, it is hard to accurately establish batch reactor model because of the factors, such as strong nonlinear behaviors, uncertainty, unknown disturbances and the limited online measurement information. In this paper, grey-box modeling approaches have been brought forward to solve the above-mentioned problems. And detailed processes of modeling of 3 grey-box methods were described in modeling batch reactor of producing accelerant for sulfuring rubber. The content is arranged as follows:Firstly, a modeling method, called Grey-box Modeling with Error Compensation (GMEC) had been put forward. The characteristics and structures of it are described in detailed in the paper. Idea of decomposition of system had been used to decompose the complicated system into the linear and non-linear part. Diagonal Recurrent Neural Network modeling (DRNN) was used to develop model of the linear part. Then, the output error between the actual system and the linear model was regarded as the training target of the non-linear model. The test result proves that GMEC is effective.Secondly, a modeling approach, named Grey-box Modeling Based on Fundamental Genes (GMFG) had been brought forward. Fundamental Genes had been introduced to describe the system behavior with insight information and prior-knowledge. A Structure Approaching Neural Network (SANN) was established, and Fundamental Genes (FGs) were regarded as the network nodes. Then, with Step-wise Regression Analysis (SRA), the Optimal FG (OFG) was selected to optimize the model structure and establish the relationship between system variables. The test result fully demonstrates that GMFG is effective.Thirdly, based on the GMFG, Improved Grey-box Modeling Based on Fundamental Genes (IGMFG) was developed to improve the deficiency in GMFG. It is mainly carried out two major improvements. The test result fully proves that IGMFG is effective.Finally, the summary and perspectives of the 3 grey-box modeling approaches are addressed.更多还原