Study On Nonlinear System Identification Of DRNN Based On Differential Evolution

System identification is an important part of control theory research, and nonlinear system identification is always the focus and difficulty in this field. As the artificial neural networks have been proposed, it opens up a new way for the identification of complex nonlinear system. In many kinds of neural networks, Back-Propagation (BP) neural network is the most widely used. As a multilayer feed-forward neural network, BP neural network has many problems and disadvantages. Diagonal recurrent neural network (DRNN) is a dynamic network with feed-back parts. By storing its internal status to reflect the dynamic characteristics, DRNN is more suitable to for the identification of nonlinear dynamic system. The learning algorithm is the core issues in the system identification. Traditionally, the BP algorithm which is always called gradient descent algorithm can not meet our needs of identification accuracy and convergence rate. In order to solve this problem, the improvement and investigation of various algorithms are carried out, and their achievements are proved to be effective. So as to further enhance the identification accuracy and convergence rate, this thesis makes an in-depth exploration and investigation of DRNN nonlinear system identification based on differential evolution (DE).In the first part of the thesis, the present situations of the system identification with neural networks and differential evolution are summarized and analyzed. The problems need to be solved are also pointed out, besides the ideas and significance of the investigation are given. The basic model and system identification principle of DRNN are introduced. BP algorithm, improved BP algorithm and genetic algorithm (GA) are respectively used as the learning algorithm in the identification of two typical nonlinear systems. Through the matlab simulation, the performances of those three algorithms are compared. The experiment results show that the nonlinear system identification of GA is much better than the other two algorithms.After that, in order to further enhance the identification accuracy, a new learning algorithm-DE algorithm is introduced and used as the learning algorithm in DRNN. Three different strategies of DE are discussed and analyzed. According to the merits of those three strategies, a hybrid mutation strategy of DE is proposed. Through the Benchmark functions testing, the effiency of hybrid strategy of DE is proved. Through the matlab simulation, the identification accuracy and convergence rate of DE is better than GA.At last, Memetic technology is introduced for overcoming the disadvantages of DE which uses global search strategy. The Simplex local search method is added in DE to get an improved DE algorithm which is called DE-Simplex. Comparing with the Benchmark test function results of DE, DE-Simplex has significantly improved. The effectiveness of DE-Simplex is verified. Then, DE-Simplex is applied to the identification of nonlinear system. The comparison results show that the identification accuracy and convergence rate are further improved.