A Smart Cloud-integrated LSTM Framework for Transformer Fault Prediction Using Sequential DGA Data

Abstract

Power transformers are vital components of electrical networks, where unexpected failures can lead to widespread outages and significant economic losses. Dissolved Gas Analysis (DGA) is a well-established technique for fault diagnosis; however, traditional ratio-based methods such as IEC, Rogers, and Duval, along with classical machine learning, are limited to instantaneous fault detection and overlook the temporal evolution of gases, thus lacking predictive capability. This study proposes an intelligent deep-learning framework that integrates DGA interpretation with sequential modeling and cloud-based monitoring. A real-world, time-series field database was developed from Iraqi transformers, covering eight key gases, including CO2 and H2O, in addition to operational variables (temperature, pressure, and load). A Long Short-Term Memory (LSTM) model was employed to capture gas dynamics and predict transitions from normal to fault conditions, achieving 97% mean accuracy and outperforming CNN and traditional methods. The integrated MATLAB–ThingSpeak–MongoDB Atlas system enables real-time monitoring (cloud) and early warning, providing an effective tool for predictive maintenance and enhancing transformer reliability.