Centralized and decentralized training strategies of artificial neural networks for transient stability assessment

Back-propagation (BP) and radial-basis function (RBF) neural networks were trained for estimating critical clearing times (CCT) of faulted power systems. Various strategies for training the neural nets like input data from the entire system, data only local to the event, data collected from various input features of the generators, etc were investigated. Of the two nets tested, radial-basis function network was, generally, found to be more suited in terms of speed of computation and accuracy of prediction. It was observed that the nets could be trained to estimate CCT resonably accurately with a smaller number of features, and also with data from critical generators in the vicinity of the faults.