Nonlinear System Identification Using Improved Biogeography-Based Optimization Trained DRNN

The nonlinear identification problem is always one of the hotspots and difficulty in control field. Recurrent neural network is more suitable for nonlinear dynamic system identification because of its unique feedback characteristics. As a kind of recurrent neural network, diagonal recurrent neural network (DRNN) is easy to implement, but also adapt to time-variation. BP algorithm, the most widely used algorithm for training DRNN neural network, contains the limitations of low identification accuracy and slow convergence. Therefore, some scholars propose new algorithms based on this method. In order to further enhance the identification accuracy and convergence rate, improved biogeography-based optimization based method is utilized to the identification of DRNN nonlinear system in this thesis.In this thesis, a new learning algorithm-biogeography-based optimization (BBO) algorithm is introduced to train the DRNN. To verify the feasibility of BBO algorithm for training recurrent neural network, BP algorithm, GA algorithm, BBO algorithm are respectively applied to train the DRNN to identify nonlinear system. Experimental results show that the identification accuracy and the identification error of BBO algorithm are better than that of BP algorithm and GA algorithm. Therefore, the non-linear system identification of DRNN network based on BBO algorithm is feasible.As BBO algorithm has the limitations of weak exploitation ability, and is easy to fall into premature and local optimum, DE algorithm and BBO algorithm are combined to train the DRNN network initial weights to generate a BBO-DE algorithm. Then the proposed BBO-DE algorithm is used to optimize the network. The analysis and comparison of BBO algorithm, DE algorithm and BBO-DE algorithm show that the non-linear system identification of DRNN based on BBO-DE algorithm is more effective.In order to further enhance the search ability of BBO-DE algorithms, a non-uniform mutation operator based BBO-DE (nDEBBO) algorithm is designed. Compared with the BBO-DE algorithm and the DE-Simplex algorithm proposed in literature20, the proposed algorithm is feasible and efficient.Finally, to avoid the influence on network weight case by gradient, a R-nDEBBO algorithm which combines RPROP algorithm with nDEBBO algorithm is designed and then adopted to train DRNN. Compared with nDEBBO algorithm, the R-nDEBBO algorithm is more effective, and the training time is shortened after the introduction of RPROP algorithm. Therefore, the designed R-nDEBBO algorithm in this thesis is advanced.