The application of artificial neural networks to transformer protection

A new method of using Artificial Neural Networks to identify the magnetizing inrush currents that may occur in transformers is developed in this thesis. It is based on the fact that the magnetizing inrush current has large harmonic components. A feed-forward neural network (FFNN) has been trained using the back-propagation algorithm, to discriminate between transformer magnetizing inrush and fault currents. The proposed ANN-based inrush detector uses magnitudes of fundamental and up to the fifth harmonic components as the inputs and provides inrush or no-inrush indication to the differential relay. Some important issues such as the neural network's and the simulated sample power network's structures; simulation, selection and preprocess of the sample data are discussed. The A NN is trained and tested by using simulated data from the EMTP program. The trained network was verified using field test data from a laboratory transformer. The simulation and field test results are included in this thesis and indicate that the ANN-based inrush detector is fairly efficient with good performance and reliability. The work reported here is a description of an experimental demonstration that a feedforward neural network could be used as an alternative method to correctly discriminate between magnetizing inrush and internal fault currents in power transformers. The network and its training process were adapted to the goal of implementing the algorithm in a digital differential protective relay.Applied Science, Faculty ofElectrical and Computer Engineering, Department ofGraduat