A Method to Model Nonlinear Systems by Neural Networks

Many processes in reality exhibit nonlinear characteristics and in most of cases they cannot be treated satisfactorily using linearized approach in a large operating range. In this paper, an approximate approach is introduced to overcome the inaccuracy and inconsistency between the linearized model and the real process, due to linear representation of the nonlinear system, such as using Taylor series expansion by treating the nonlinear system as a linear uncertain system, that consists of a linear part, and an uncertain part. A neural network with Gaussian radial basis function in the hidden layer is employed to approximate the uncertain system. The approach can incorporate prior knowledge in its framework and provide a more transparent insight than the neural "black box" approach. The simulation results reveal that the proposed modeling approach to nonlinear systems is effective.