Precise transformer fault diagnosis via random forest model enhanced by synthetic minority over-sampling technique

Power transformers are considered one of the power system's most critical and expensive assets. In this regard, it is vital to assess the fault within the power transformer considering numerous operational aspects. In the literature, dissolved gas analysis (DGA) is the routine in-service test for power transformers and one of the most important tests to ensure sufficient system reliability. Specifically, this test can detect dissolved gases in transformer oil which are then interpreted to detect the fault type of the transformer. Previous studies reported that the graphical Duval pentagon is one of the most accurate and consistent DGA interpretation techniques. However, it still has limitations on the complexity of the implementation in large amounts of data. To cover these issues, this study mitigates the limitation and complexity of implementing the graphical Duval Pentagon Method (DPM) in large amounts of data. To reach this goal, we develop a precise machine-learning-based fault identification model by employing the Random Forest algorithm with Synthetic minority over-sampling technique (SMOTE) preprocessing. The proposed Random Forest models with SMOTE perform satisfactorily in diagnosing faults for the evaluation dataset, with a total accuracy of 96.2% for DPM1 and 96.5% for DPM2. The proposed models were also compared to other machine learning algorithms, performing better both in classification accuracy and consistency due to uncertainty.