Application of Artificial Neural Network Algorithm to DGA Diagnosis for Transformer Fault Classification

Abstract

One of the most crucial parts of the electrical grid is the power transformer. Due to the high cost of a power loss, extending the lifespan of transformers is essential. Condition monitoring has emerged as the standard method for determining a transformer's overall health, which helps to lessen the likelihood of unexpected failure and guarantees the occurrence of unscheduled outages. Numerous diagnostic approaches have been developed to achieve this aim. Dissolved Gas Analysis (DGA) is one of the procedures that has gained favor for checking the transformer's condition. For the simple reason that it gives a true picture of the transformer's current condition and the nature of any problems developing inside it. The correct analytical method allows for the removal and identification of dissolved gases. A total of seven gases, including H₂, CH₄, C₂H₆, C₂H₄, C₂H2, CO, and CO₂, are identified as being crucial. It's possible that rule-based expert systems won't be able to learn from new/novel defect data, and traditional interpretation techniques may not converge on the best answer in most situations. In this research, we use DGA data as inputs to a Multi-Layer Perceptron (MLP) network in an effort to identify problems with power transformers. In order to detect these early problems, we propose a three-tiered MLP network. To train the MLP, we use three different training algorithms and assess their diagnostic efficacy: Gradient Descent (GD), Scale Conjugate Gradient (SCG), and Levenberg-Marquardt (LM). The accuracy of the SCG training algorithm is rapidly improving. When compared to more traditional techniques of diagnosis, the suggested MLP network shows promising results.