Fault Diagnosis in Smart Transformers Using Machine Learning Techniques

Abstract

Smart transformers bring major change to present-day power systems, allowing for better voltage control, easy monitoring and improved connection with the grid. Because they are used in smart substations and for renewables, these transformers have sensors and communication technology built in so that they can be controlled from a distance and diagnosed while they are operating. The typical ways to detect faults like dissolved gas analysis (DGA), thermography and manual inspection are mostly used once problems occur, are expensive and have limited capabilities for ongoing and prompt diagnostics. This study therefore suggests a strong and flexible machine learning-based method for automatically detecting faults in smart transformers. It includes bringing in data from many sensors, processing the signals with the discrete wavelet transform and cutting down on the amount of data to be processed with principal component analysis. These supervised learning models—Support Vector Machines (SVM), Random Forest (RF) and Convolutional Neural Networks (CNNs)—are all trained with data collected from simulations and real sensors. They are assessed for accuracy, precision, recall, F1-score and AUC. The CNN-based approach is better than classic classifiers, reaching an accuracy rate of over 96% in identifying faults while having very little trouble with both false positives and brief interruptions in data. CNN gets such high accuracy since it can learn useful features by itself from the temporal patterns of the signals. The structure was built to fit into edge computing, letting it address real-time applications with low resources. Integrating advanced signal processing and deep learning helps find faults at an early stage in smart transformers, cutting maintenance expenses, lowering downtime and strengthening the overall resistance of upgraded power distribution systems.