International Conference on Computer and Knowledge Engineering

Home / 12th International Conference on Computer and Knowledge Engineering

Intelligent Interpretation of Frequency Response Signatures to Diagnose Radial Deformation in Transformer Windings Using Artificial Neural Network

Authors :

Reza Behkam¹ Hossein Karami² Mehdi Salay Naderi³ Gevork B. Gharehpetian⁴

1- Department of Electrical, Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran 2- High Voltage Studies Research Department Niroo Research Institute 3- Department of Electrical, Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran 4- Department of Electrical, Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran

Keywords :

Transformer،frequency response analysis (FRA)،radial deformation (RD)،artificial neural network (ANN)،intelligent interpretation

Abstract :

Transformers are vital elements of a power system network in which continuous service is of great importance, and high reliability of the entire network depends on health condition of the transformers. Transformer windings are susceptible to mechanical tensions as a result of poor operation or transit. Radial deformation (RD) as a mechanical winding defect exerts disruptive influences on the performance of the transformer. In the field of transformer monitoring, frequency response analysis (FRA) has established itself as a reliable diagnostic tool. Nonetheless, complexity and open questions surround the decipherment of FRA results because reliable interpretation code is unavailable. This study presents an artificial intelligence-based code for interpreting frequency response traces. In this study, RD faults are practically applied to the windings of a 1600 kVA distribution transformer operating at 20 kV. Practical measurements are taken of FRA traces, and then feature vectors are extracted using adequate and sensitive numerical indexes, including cross-correlation factor (CCF), normalized root mean square deviation (NRMSD), Lin's concordance coefficient (LCC), and fitting percentage (FP). All four parts of frequency responses which are magnitude, angle, real, and imaginary parts are investigated. In addition, an artificial neural network (ANN) regarded as an intelligent classifier employing extracted features to distinguish locations of RD defects. In order to assess the performance of the proposed intelligent network, K-fold cross validation technique is utilized. The most appropriate numerical index beside the most effective part of the frequency response are introduced.