Neural Network Based Hammerstein Model Identification And Its Application

Hammerstein model is a typical block-oriented nonlinear model, which is composedof a nonlinear static function and a linear dynamic subsystem in series. Many practicalsystems, such as solid oxide fuel cell, continuous stirred tank reactor, and distillation tower,can be described by the Hammerstein model. Identification of Hammerstein model is aproblem, where an effective nonlinear static description and a simple identificationalgorithm is the key to the problem. Although the traditional neural networks, such asmultilayer perceptrons and radial basis function neural network, have strong capability ofnonlinear approximation, these networks contain many parameters and complex structure,which increases the complexity of the identification algorithm inevitably. In this thesis, theidentification and application of Hammerstein model based on two kinds of neuralnetwork with simple structure, i.e., functional link artificial neural network (FLANN) andextreme learning machine (ELM) are studied, and the main work includes:Firstly, we put forward a two-step identification method for Hammerstein modelbased on FLANN. In our method, the nonlinear part is described by the FLANN, and aspecial input signal is utilized to separate the nonlinear part from the linear part. Then, theleast square method and the gradient descent method are adopted to estimate theparameters of the linear part and FLANN respectively. The experimental results show theeffectiveness of the proposed method.Secondly, the identification algorithm of FLANN-Hammerstein model proposed onlast chapter may fall into the local optimal sometimes. To this situation, we regard the sumof square error as objective function, and proposed an improved particle swarm algorithmbased on Lévy variation to estimate the parameters in FLANN-Hammerstein model. Thismethod can not only identify parameters of the linear part and nonlinear partsimultaneously, but also avoid falling into local optimal point and improve theidentification precision.Thirdly, we propose a nonlinear system identification method based onELM-Hammerstein model. In order to determine the structure of ELM-Hammerstein model, the Lipschitz quotient is used to ascertain the model structure from theinput-output data. Then, the generalized ELM algorithm is given to estimate the modelparameters. The effectiveness of proposed method is proven with three experiments.Finally, after the analysis of the structure and dynamic behavior of solid oxide fuelcell (SOFC) system, the proposed ELM-Hammerstein model is used to establish itscorresponding nonlinear model.