A CatBoost-Based Framework for Data-Driven Fault Diagnosis in Power Transformers Using Dissolved Gas Analysis

• A CatBoost-based DGA framework achieves 97.6% fault classification accuracy. • 1680-dimensional time-series data reduced to 34 physics-informed features (98% reduction). • Strong class-wise performance: 98.9% (Normal), 94.7% (PD), 93.9% (LED), 89.0% (LTO). • Hydrogen and engineered gas ratios dominate feature importance. • Lightweight, interpretable model enables real-time transformer monitoring. Reliable transformer operation is essential for reducing outages and maintaining power system stability, yet existing diagnostic techniques still face significant limitations. Although Dissolved Gas Analysis is widely adopted for condition monitoring, traditional ratio-based approaches such as the Duval Triangle and Key Gas Method often become unreliable when gas signatures overlap or when faults are at an incipient stage, leading to ambiguous interpretations. Machine learning methods have improved classification performance, but many existing models remain sensitive to noise, data imbalance, and overfitting, which restricts their robustness in real world applications. To overcome these challenges, this study proposes a CatBoost based diagnostic framework that integrates statistical and energy related features of H₂, CO, C₂H₂, and C₂H₄ with carefully engineered gas ratios to capture complex inter gas relationships linked to discharge and thermal faults. With optimized hyperparameters, the model achieved 97.6 percent overall accuracy, along with strong class specific results across Normal, Partial Discharge, Low Energy Discharge, and Low Temperature Overheating conditions. Feature importance analysis highlights the dominant contribution of hydrogen and ratio-based features, while stable and rapid convergence demonstrates computational efficiency. The proposed approach offers a robust, interpretable, and practically deployable solution for accurate transformer fault detection.