Exploring the Usability of Adaptive Weighting Factors in Transformer Health Diagnostic Practices

This paper examines the potential of using adaptive weighting factors in the Health Index (HI) method for transformer health diagnostics. Typically, HI methods rely on weighting factors that are set by experts. While effective, this approach often static and may not always incorporate the latest insights or data. This study aims to enhance the HI method's accuracy and reliability by integrating adaptive, data-driven weighting factors. To achieve this, the usability of deep neural network algorithms to derive these weighting factors from actual transformer oil data is investigated. The dataset includes 500 high-voltage transformer oil samples, each with five oil testing features and known oil quality conditions. This data is used to develop and evaluate the approach. The dataset is analyzed at different scales (5, 25, 50, 100, 250, and 500 randomized samples) to ensure a comprehensive and repeatable evaluation. The main procedure involves feeding the dataset into a neural network, which then extracts the optimal weighting factors. These factors are subsequently applied in the conventional HI formula to assess their accuracy in classifying transformer oil quality. The study focuses on three key areas: firstly, the influence of the dataset size on the precision of the adaptive weighting factors, secondly, the effect of varying the neural network's structure, and third, the choice of optimizers. The aim of this study is to shift from a predominantly expertbased approach to a more data-driven methodology for reliable transformer health diagnostics framework.