Optimal Selection of Input Variables by BPSO for Diagnosis of Incipient Failures in Power Transformers (by DGA)

Abstract

Power transformer immersed in oil is a valuable asset in the operation of the electrical system, therefore, it is of interest to the operating companies to keep the power transformers in perfect operating conditions. Early diagnosis of a fault condition in the power transformer is a fairly addressed research topic, however, inappropriate use and the limited number of data do not allow formulating a robust methodology for a real implementation in the electrical system. This document presents an optimal selection of input variables in diagnosis of power transformer failures by DGA, the sample of inputs is generated from the gas contents (hydrogen, methane, acetylene, ethane and ethylene) and the selection of optimal inputs (VE-BPSO) is extracted with Binary Particle Swarm Optimization (BPSO) in the nearest neighbor classification (Conventional K-NN Classifier). In the validation process for 63 independent data in both Conventional K-NN Classifier and Artificial Neural Network (ANN) the performances for VE-BPSO are superior to the conventional approach (IEC 60599 standard inputs). Therefore, the input variables with the best characterization (clustering) in diagnosis of faults in TP is VE-BPSO, which is the main contribution of this paper.