Power transformers are vital in ensuring the reliability of electrical power systems, necessitating accurate fault classification for their efficient operation. This research evaluates a novel Transformer Deep Learning model architecture for fault classification using dissolved gas analysis (DGA) data, leveraging feature engineering and an over-sampling technique to address high-dimensionality and class imbalance challenges. The model demonstrated substantial accuracy improvements across datasets of varying sizes and preprocessing stages, particularly with SMOTE-enhanced data. These findings underscore the effectiveness of Transformer deep learning architectures in advancing the state-of-the-art in fault classification for power transformer systems.