Fault classification using Kohonen feature mapping

Applications of neural networks to power system fault diagnosis have provided positive results and shown advantages in process speed over conventional approaches. This paper describes the application of a Kohonen neural network to fault detection and classification using the fundamental components of currents and voltages. The Electromagnetic Transients Program is used to obtain fault patterns for the training and testing of neural networks. Accurate classifications are obtained for all types of possible short circuit faults on test systems representing high voltage transmission lines. Short training time makes the Kohonen network suitable for on-line power system fault diagnosis. The method introduced in the paper can be easily extended to any size power system since the only information required for the NN to function are those that are recorded at substation fault recorders. With fast NN hardware now becoming available, on-line implementation is only a question of economics.