Diagnóstico inteligente de faltas em transformadores baseado na análise de gás dissolvido em óleo

Abstract

The National Interconnected System (SIN) for the production and transmission of elec trical energy Sileiro is a large hydro-thermal-wind system, with a predominance of power plants hydroelectric plants, whose representation results from the combination of differ ent generation systems, by a robust network of transmission lines and numerous substa tions surrounding the network basic energy from the voltage class of 230kV. The power transformer tency presents itself as a connecting link between generation and transmis sion, playing a essential in electrical power systems, whose early fault detection is crucial for such systems, due to the high maintenance cost and the impact of defects in these equipment. In this context, several methods, both intelligent as well as conventional, for detecting flaws based on the analysis of dissolved gases in insulating oil (DGA) have been developed and standardized. This work presents a DGA database composed of real sam ples collected from transformers around the world over 20 years of operation, in addition to data from consolidated literature. To the more than 2000 samples allow the design of thermal and electrical fault classifiers in transformers by machine learning (ML). The study details the exploration data and evaluates classifiers such as Logistic Regression (RL), Vector Machine Support System (SVM), Artificial Neural Networks (ANN), K Nearest Neighbors (KNN), in addition to the conventional methods of Duval’s triangle, Rogers relations, Key Gas, Doernenburg and IEC 60599. The results indicate that a hy brid architecture, composed by classifiers KNN, ANN and the conventional Duval triangle method, has better results than the individual use of the methods tested in this work. Where the classification of the test samples, highlighted the performance of the hybrid architecture in 98% in the diagnosis of incipient failures in transformers.