TRANSFER FUNCTION ESTIMATION AND AI APPLICATION FOR TRANSFORMER FRA INTERPRETATION

Abstract

Winding fault takes up a considerable proportion among all the fault types of transformer. Frequency response measurement of transformer has been developed as an effective tool to detect the mechanical integrity of windings. In Frequency Response Analysis (FRA) technique, the diagnosis measurement should be compared with the reference measurement, and winding displacement/deformation may be suggested by the occurrence of difference. However, no IEEE or IEC standard has been published regarding the interpretation of frequency responses. Also, a large amount of frequency responses have been accumulated by the utilities over many years. Utilities may or may not know transformer design information, such as the winding construction types. Different winding construction types are susceptible to different modes of mechanical deformations, and the same asset management method can be applied to transformers with same winding construction types. Two methods are proposed to derive the mathematical expression for the FRA trace, in the format of a transfer function. Using this mathematical expression data can be generated in the same format, no matter how or by which FRA measurement device the original FRA trace is produced; and the same data format is of necessity for applying numerical indices and AI techniques to interpret FRA. Feature Extraction Method divides the frequency range into several regions, and complex poles and zeros are extracted to form a feature transfer function, the difference between which and the measured data is then corrected by real poles and zeros and the constant. It is validated by 48 frequency responses of eight 400/275/13 kV autotransformers. Extreme Points Identification Method detects extreme points and iterates to optimise the transfer function’s parameters. The Feature Extraction Method is good at describing the subtle feature of frequency responses, whilst the Extreme Points Identification Method ensures the simplistic expression to be identified, which is physically achievable by filter design principle. Both methods have been successfully applied for the diagnosis of faulty transformer winding. Two methods are proposed to identify the winding construction types using frequency responses. The supervised machine learning method, Support Vector Machines (SVM), is utilised to build an identification model, using FRA traces of transformers with known winding type. After testing and validating, the SVM model is then applied to FRA traces with unknown winding type information. A set of data from the UK’s National Grid FRA database, was used to demonstrate and verify the SVM model. The proposed method can successfully identify winding types including multiple layer, plain disc, interleaved disc and single helical windings. The unsupervised machine learning method, Hierarchical Clustering, is utilised to cluster frequency responses according to the similarity and dissimilarity. Once the frequency responses are in the same cluster, it is in default to think the windings should share the same winding construction type. It was applied to UK’s National Grid FRA database so the frequency responses of transformers with unknown winding type can be identified by being clustered into a group together with a frequency response with known winding type.

Date of Award	1 Aug 2020
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Zhongdong Wang (Supervisor) & Peter Crossley (Supervisor)