NEURAL NETWORKS APPLICATION TO FAULT DETECTION IN ELECTRICAL SUBSTATIONS

This paper proposes an application of neural networks to fault detection in electrical substations, particularly to the Parada Angelica Electrical Substation, part of the AMPLA Energy System provider in Rio de Janeiro, Brazil. For research purposes, that substation was modeled in a bay oriented fashion instead of component oriented. Moreover, the modeling process assumed a substation division in five sectors or set of bays comprising components and protection equipments. These five sectors are: 11 feed bays, 2 capacitor bank bays, 2 general/secundary bays, 2 line bays and 2 backward bays. Electrical power engineer experts mapped 291 faults into 134 alarms. The employed neural networks, also bay oriented, were trained using the Levenberg-Marquardt method, and the AMPLA experts validated training patterns, for each bay. The test patterns were directly obtained from the SCADA (Supervisory Control And Data Acquisition) digital system signal, suitably decoded were supplied by AMPLA engineers. The resulting maximum percentage error obtained by the fault detection neural networks was within 1.5 % which indicates the success of the used neural networks to the fault detection problem. It should be stressed that the human experts should be the only ones responsible for the decision task and for returning the substation safely into normal operation after a fault occurrence. The role of the neural networks fault detectors are to support the decision making task done by the experts.