Affiliations: [a] Department of Electrical Engineering, Regional University of Blumenau (FURB), R. São Paulo 3250 (Itoupava Seca), Blumenau SC, Brazil
| [b] Department of Information Systems and Computing, Regional University of Blumenau (FURB), R. Antônio da Veiga, 140 (Itoupava Seca), Blumenau SC, Brazil
| [c] Department of Electrical Engineering, Santa Catarina State University (UDESC), R. Paulo Malschitzki 200 (Zona Industrial Norte), Joinville SC, Brazil
| [d] Fondazione Bruno Kessler, Istituto per la Ricerca Scientifica e Tecnologica, Povo, Via Sommarive 18, 38123 Povo, Trento TN, Italy
| [e] Computer Science and Artificial Intelligence, University of Udine, Via delle Scienze 206, 33100 Udine UD, Italy
Correspondence:
[*]
Corresponding author. Alessandro Medeiros, Department of Electrical Engineering, Regional University of Blumenau (FURB), R. São Paulo 3250 (Itoupava Seca), Blumenau SC, Brazil. E-mail: [email protected].
Abstract: Contamination in insulators results in an increase in surface conductivity. With higher surface conductivity, insulators are more vulnerable to discharges that can damage them, thus reducing the reliability of the electrical system. One of the indications that the insulator is losing its insulating properties is its increase in leakage current. By varying the leakage current over time, it is possible to determine whether the insulator will develop an irreversible failure. In this way, by predicting the increase in leakage current, it is possible to carry out maintenance to avoid system failures. For forecasting time series, there are many models that have been studied and the definition of which model is suitable for evaluation depends on the characteristics of the data associated with the analysis. Thus, this work aims to identify the most suitable model to predict the increase in leakage current in relation to the time the insulator is outdoors, exposed to environmental variations using the same database to compare the methods. In this paper, the models based on linear regression, support vector regression (SVR), multilayer Perceptron (MLP), deep neural network (DNN), and recurrent neural network (RNN) will be analyzed comparatively. The best accuracy results for prediction were found using the RNN models, resulting in an accuracy of up to 97.25%.
Keywords: Failure prediction, time series forecasting, artificial neural network, insulators
