Neural networks for constrained transient stability flows

A weighted neural network (WNN) and a weightless neural network (WLNN) were compared for output accuracy dependent on the number of training data and distribution. If the number of training inputs is limited, having an appropriate distribution is important. Sobol's method was used to generate a quasi-random sequence of training inputs, providing good coverage over a specified range. These Sobol sequences (Sob) were employed to select the training patterns for WNN and WLNN designed to determine the limiting power flows over critical lines under transient stability conditions of a 4-machine 11 bus and a 10-machine 39 bus New England system with variations in load level and fault location. The results indicate that the constrained flows to maintain given transient stability margins in operation can be efficiently estimated to be better than 5% using both WNN and WLNN, but WLNN is recommended for its ease and speed of training.