Comparison of PD classification capabilities for transformer failure and typical noise models with neural network applications

This paper presents three kinds of Neural Networks (NNs) for classifying Partial Discharges (PDs) of failure models which are extracted from internal insulation configurations of power transformers and typical noise in substations. The test results show that three networks can fairly classify the designed models. The performance of Back-Propagation (BP), Learning vector Quantization (LVQ) and Fuzzy ARTMAP networks is evaluated. The classification accuracies obtained from the three networks are compared with each other. In addition to the classification accuracy, the neural networks are analyzed for their generalization capability and stability of the results. Best results (accuracy and convergence time) are obtained with the Fuzzy ARTMAP network. Classification rate for the designed models is 100% at voltage level 3. Simulation results serve to illustrate the properties of various networks we used as well as the stability with respect to various critical parameters.