Comparative Analysis of Neural Networks and Random Forests for Transformer Fault Diagnosis Based on Dissolved Gas Analysis (DGA)

This paper investigates the application of deep learning techniques, Random Forest and Neural Network, for classifying transformer faults based on Dissolved Gas Analysis (DGA). A real-world data set has been used, with fault labels assigned according to the IEC 61198:1993 standard, including fault classes. In order to ensure statistical reliability, each model has been evaluated over ten runs, and performance metrics have been averaged. The Random Forest model has achieved a mean accuracy of 80.87% (±9.87%), outperforming the Neural Network, which has reached a mean accuracy of 68.70% (±12.51%). Confusion matrices and class-wise metrics have showed that Random Forest has classified the majority of Normal and T3 cases correctly, while Neural Network had higher misclassification rates, especially between Normal and D1. Overall, Random Forest has exhibited greater robustness and interpretability, making it a more suitable choice for practical fault diagnostics in power transformers.