Power transformer health index (HI) prediction using Kolmogorov-Arnold Network

Power transformers are critical assets in electrical power systems, and ensuring their reliability and longevity is paramount to maintaining the stability of power grids. One of the most significant threats to transformer health is internal degradation, which can manifest through electrical and thermal stresses. Dissolved Gas Analysis (DGA) is a widely accepted method for diagnosing transformer health by detecting specific gases dissolved in the transformer oil. These gases, including hydrogen, methane, ethylene, and others, are by-products of insulation breakdown and oil degradation, which occur due to faults like partial discharge, arcing, and overheating. While traditional methods for interpreting DGA data, such as the Rogers Ratio, Duval Triangle, and Key Gas Method, have been valuable, they often rely on fixed thresholds or expert judgment, which can limit their accuracy and scalability. Recent advancements in machine learning (ML) offer a promising alternative by enabling data-driven approaches that can identify complex patterns and relationships within DGA features. Machine learning models can learn from historical data, providing more accurate and timely predictions of transformer health and fault conditions. This paper presents an ML approach for power transformer health assessment, employing the Kolmogorov- Arnold Network(KAN) model. The study leverages DGA data to train and evaluate the models, aiming to enhance accuracy in fault detection. KAN is used to capture the non-linear relationship and complexity within the DGA data, while Multilayer Perceptron (MLP) serves as a baseline for comparison. The simulation results show the KAN model with 3 layers (and a single univariate layer), with an RMSE of 0.0011 andR2 of 1.000 in independently reserved test data. This result significantly outperforms the MLP and demonstrates a superior ability to handle nonlinearities and reduce prediction error.