Abstract:
Transformer Frequency Response Analysis (FRA) is a powerful diagnostic tool for detecting and identifying winding deformation, short-circuits recognition, and other abnormalities in power transformer active part. Accurate interpretation of FRA results is crucial for timely maintenance and the prevention of catastrophic failures in transformers. However, conventional interpretation methods often rely on expert knowledge, which can be subjective and limited. This study aims to develop a machine learning-based approach for interpreting FRA data that enhances the accuracy and efficiency of detecting winding issues in power transformers.
In this thesis, various machine learning algorithms are investigated for their suitability in interpreting FRA data. A comprehensive dataset, comprising experimental data from multiple transformers with different configurations, are assembled and used to detect and recognize a fault severity in transformer winding and estimating fault contact resistance and its strictness. The dataset is used to train, validate, and test the performance of the machine learning models, which included Decision Trees, Random Forests, K-Nearest Neighbors (KNN), and Support Vector Regression (SVR).
The results demonstrated that the machine learning-based approach significantly will improve the interpretation of FRA data. In particular, the Random Forest and KNN models exhibited high accuracy and low error rates in predicting winding short-circuits severity. Furthermore, a user-friendly Graphical User Interface (GUI) is developed to facilitate the deployment and interpretation of the trained models, making the approach more accessible for industry professionals. The outcome of this thesis will support and improve any missed or incorrect measurement in transformer FRA data and is able to advise the level of short-circuit fault to operator or utility manager.