Power Transformer Fault Prediction Using Dissolved Gas Analysis and Neural Networks

In this work, we present a neural network-based study capable of predicting faults in oil-insulated power transformers through the analysis of dissolved gases. The advantage of this study lies in using data already collected by electric power companies, which gather it to comply with international or regional standards; however, they sometimes act only after the equipment is already in a faulty condition. Therefore, the challenge in this work was data regularization, as collections typically occur at long intervals of 6 to 12 months. Furthermore, samples are often irregular, as data collection depends on factors such as weather and the availability of maintenance teams. As a result of this work, Multilayer Perceptron (MLP), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM) were used to predict failures with advanced forecasts ranging from 1 to 6 months, achieving accuracies of 97.5% and 85%, respectively. Thus, these models prove to be important tools for maintenance planning, enabling adequate predictability for organizing equipment shutdowns without the need for high investments in installing tools to capture this information online and adapting substations to send data to control rooms or other analysis centers.