Metodologia para a classificação de descargas parciais utilizando redes neurais artificias.

Abstract

Insulation failure in power systems plant may be caused by manufacturing defects or by electrical, mechanical, thermal and chemical process which occur during operation. These processes create defects, including voids in solids, which locally reduces the dielectric insulation strength. Electrical discharge, may occur in these regions of reduced dielectric strength without discharge occurring in the regions of non-degraded insulation. A category of this electrical discharges is labeled Partial Discharges. The Partial Discharges measurement has long been used to evaluate insulation system design and as a quality assurance test for High Voltage apparatus prior to installation and during the equipment operational time. Partial Discharge is characterized by high frequency current pulses originating in gas ionization processes when damaged insulation is submitted to high values of electric field. The aim of this thesis is to present a comprehensive methodology for emulation, processing, and automatic classification of several types of partial discharge signals. For the generation of partial discharges in laboratory it was developed a hermetic cell that allowed the generation of well defined partial discharges signals. It was verified during the measurements that the partial discharges signal were susceptible to noise existing in measurement area. To mitigate this problem, it was used the Wavelet Transform. Several partial discharges signals obtained in laboratory were evaluated and was verified a significat noise level reduction after the filtering process. Among the evaluated wavelets families, the Daubechies have presented the best performance, besides a lower computational processing time in relation to the other wavelet families. In the sequence, feature extraction was performed using statistical parameters calculation for each insulation defect configuration. The obtained feature data was used as input in a variety of Artificial Neural Network (ANN) topologies for an automatic identification and interactive determination of the most suitable ANN topology (i.e. number of artificial neurons and hidden layers) for this purpose. Three topologies of ANN were implemented, the first was based in one hidden layes, the second in two hidden layes and the last in tree hidden layers. The performance of classification was satisfatory, mainly the topology based on two hidden layers. It was obtained a global recognition rate of aproximatedely 96%, presenting a better result in comparison with the other topologies. The performance of the ANN topologies was evaluated using the mean squared error and confusion plots.