Study On Nonlinear System Identification Based On Recurrent Neural Network

The identification of nonlinear dynamic system is always the difficulty and the focus in the fields of control research.There are many defects of Traditional feedforward neural networks in dealing with nonlinear dynamic system identification.And the recurrent neural network can use its own internal system of state-feedback to reflect the non-linear characteristics,so it is more suitable for nonlinear system identification.There are many different types of recurrent neural networks, and this paper has only described five common recurrent neural networks:Hopfield networks, Jordan network, Elman network, DRNN network and ESN network.In this paper,Elman, DRNN, ESN network models have used in the actual nonlinear system identification.The research of learning algorithm has always been the core of the neural network research.The learning algorithm of the recurrent neural network commonly use BPTT algorithm, RTRL algorithm, which are based on gradient descent.Therefore the commonly algorithms can not prevent many defects to the error gradient.After the study of three kinds of intelligent optimization algorithm,this paper has proposed two kinds of improved optimization algorithms and has compared the optimization results of three benchmark functions to find the best performance algorithm in the five kinds of optimization algorithms.The IACA algorithm has the best performance,and the algorithm has used to optimize the neural network structure and parameters.Taking the factory in Lanzhou,the dynamic chemical system petrochemical ethylene, propylene distillation column for the identification objects,First,It is advantages and disadvantages of the experimental study of the Elman,DRNN network for nonlinear system identification.There are many shortcomings of the traditional recurrent network training algorithm.So it has a better recognition performance for using the IACA algorithm to design the network structure and parameters.In addition,this paper has detailed studied a new type of recurrent neural networks which is ESN network,and has used it for the identification problem of ethylene, propylene distillation.Also this paper has studied the three important parameters (reserve pool size N, the spectral radiusγ,connection density (?)) of ESN network on the Network Identification properties.It is easily lead to pathological problems of ESN network regular learning algorithm,so this paper has used the IACA algorithm for optimizing the output weights of ESN network, and it has obtained the relative best results of identification.Through a large number of simulation experiments,the results show that recurrent neural networks are more suitable for nonlinear system identification than feedforward neural networks.Using the IACA algorithm to design the network structure and parameters,it can Effectively improve the network efficiency,improve the training speed of network,also it can solve the local minimum problem of traditional gradient descent and pathological problem of ESN Network.Finally,the distillation product quality real-time prediction program,which used the ESN network based on IACA algorithm,has proposed for the study identification of ethylene, propylene distillation.In the final of this paper,it summarizes the work of this article and proposes some future research directions and priorities.