Identification of nonlinear system based on improved neural network sequential learning algorithm

Abstract

Neural network is one of the most effective tools in nonlinear control system design. In practical online application, sequential learning algorithms are preferred over batch learning algorithms because they do not require retraining whenever a new data is trained. However, the existing sequential learning algorithms only utilize the current instant estimation of the nonlinear system for constructing the network structure. Therefore they do not characterize temporal variability well. To overcome this problem, the multi-step-ahead predictor of the nonlinear system is introduced to the growing and pruning network for constructing network structure. Furthermore, a sliding window model is used to prevent the network from fitting the noise if there is noise in the input data. And in order to reduce the computation load, the winner neuron strategy is utilized to update the parameters of the neural network using extended Kalman filter. Experimental results show that the proposed algorithm can obtain more compact network along with smaller errors in mean square sense than other typical sequential learning algorithms.