Development of a Model-Based Dynamic Recurrent Neural Network for Modeling Nonlinear Systems

In this study we develop the theory lying behind a model-based dynamic recurrent neural network (MBDRNN) that has been previously used to improve the linearized models of nonlinear systems. The initial structure of the MBDRNN is based on the linearized system model. Afterwards, the MBDRNN is trained to represent the system´s nonlinearities by adapting the weights of its nodes´ activation functions using Back-Propagation. The MBDRNN is applied with analytical detail to an arbitrarily chosen Single-Input/Single-Output (SISO) second order nonlinear system, and comparisons are made between the linearized and MBDRNN models, showing that the MBDRRN effectively improved the linearized model