Application of Neural Networks Methodology to Power System Modelling and Control

Artificial feedforward networks are studied as nonlinear function approximators used to identify forward and inverse mappings of discrete time dynamic systems. They are found to provide significant advantages over other modelling techniques such as polynomial approximations, especially if the extrapolation beyond the region covered by the learning data is involved. We apply the neural network methodology to a simple second order approximation of a single-machine infinite-bus power system controlled by means of modifying the reactance of the line. Accurate off-line identification of forward and inverse dynamics of the system is performed by means of single hidden layer neural networks, and both models are then used in a direct inverse control configuration. The controller simulations show very good quality of transients for severe short-circuit fault.