Adaptive protection strategies for detecting power system out-of-step conditions using neural networks

This paper presents new strategies for adaptive out-of-step (OS) protection of synchronous generators based on neural networks. The neural networks architecture adopted, as well as the selection of input features for training the neural networks, is described. A feed forward model of the neural network based on the stochastic backpropagation training algorithm is used to predict the OS condition. Two adaptive OS protection strategies are suggested. The first approach depends firstly on detecting the case of the system through case detection neural networks by some prefault local measurements at the machine to be protected, and then calculating the new OS condition through an adaptive routine. The second approach is based on creating a large neural network to be trained using different outage cases of the power system. The capabilities of the developed adaptive OS prediction algorithms are tested through computer simulation for a typical case study. The results demonstrate the adaptability of the proposed strategies.