Modeling of nonlinear dynamic systems via discrete-time recurrent neural networks and variational training algorithm

This paper proposes a discrete-time recurrent neural network architecture and parameter adaptation algorithm for modeling of nonlinear dynamic systems. The learning algorithm is based on variational calculus and operates off-line. A neural network based current transformer nonlinear model is presented as a demonstration of the proposed architecture and learning algorithm. It is designed for power engineering needs in power systems and is suited for real-time applications in digital relay protections.