Diagnóstico de fallas en transformadores de potencia usando modelos de redes Neuronales Perceptrón Multicapa

Power transformers are among the most expensive and essential elements in the functioning of electrical networks. Dissolved gas analysis (DGA) enables the detection of dissolved gas concentrations in insulating oil, associated with thermal faults, partial discharges, and both high- and low-energy discharges. These gas correlations enable the assessment of the transformer’s condition. In recent years, the use of machine learning techniques, such as artificial neural networks, has seen an increase in the prediction, diagnosis, and management of faults, displaying high performance in spotting anomalies and aiding decision-making in the field of maintenance. The present study developed a multi-label Multilayer Perceptron (MLP) model, which considers transformer fault diagnoses calculated using the Dornenburg, Rogers, Duval triangle, and gas procedures recommended by IEC & IEEE standards. This enables the utilization of several fault diagnosis for each transformer via each approach. The model underwent validation by k-Fold cross-validation, achieving an exact match rate of 90.95%, which pertains to instances where the model aligned with all labels specified by each diagnostic procedure. The ROC curve, with an area under the curve of 99%, and the Accuracy-Completeness curve, with an average accuracy of 99.3%, were generated graphically.