Analysis of training techniques of ANN for classification of insulators in electrical power systems

• Published in: IET generation, transmission & distribution transmission & distribution, 2020-04, Vol.14 (8), p.1591-1597
• Main Authors: Frizzo Stefenon, Stéfano, Waldrigues Branco, Nathielle, Nied, Ademir, Wildgrube Bertol, Douglas, Cristian Finardi, Erlon, Sartori, Andreza, Henrique Meyer, Luiz, Bartnik Grebogi, Rafael
• Format: Article
• Language: English
• Subjects: artificial intelligence techniques; artificial neural network models; classification problem; defect condition classification; electrical power systems; faulty insulating devices; field inspections; gradient descent method; inspection; insulators; learning (artificial intelligence); Levenberg–Marquardt method; neural nets; optimisation; optimisation methods; outdoor insulating systems; pattern classification; power engineering computing; power system reliability; quasi‐Newton method; Research Article; specific inspection equipment; training capacity; ultrasound detector
• ISSN: 1751-8687; 1751-8695; 1751-8695
• DOI: 10.1049/iet-gtd.2019.1579

Identifying defects in electrical power systems during field inspections is a difficult task, since faults are generally not visible and may be intermittent. To find possible adverse conditions, specific inspection equipment is used. The ultrasound detector is the equipment normally used to inspect outdoor insulating systems; however, using it demands operator experience. To improve the defect condition classification, artificial intelligence techniques are applied to assist the operator in the decision task and thereby facilitate the identification of faulty insulating devices in the grid. The training of artificial neural network (ANN) models is an important step in solving the classification problem. This study aims to evaluate the training capacity in terms of the performance of different optimisation methods for the calculation of the mean square error after convergence. Traditional methods such as Gradient Descent and its variations will be presented, as well as methods that employ high computational effort such as quasi-Newton and Levenberg–Marquardt. In order to base these concepts, a review will be presented on the use of these algorithms and on the problem of classification of insulators in distribution networks. The results show that there is a considerable performance difference between the calculation methods.