A Hybrid CNN-LSTM Approach for Transformer Remaining Useful Life Prediction Using Dissolved Gas Analysis

Power transformers are critical assets in electrical grids, and accurately predicting their remaining useful life (RUL) is essential for predictive maintenance strategies. This paper presents a novel hybrid CNN-LSTM approach for Dissolved Gas Analysis (DGA) data that prioritizes computational efficiency while maintaining high prediction accuracy. The approach combines CNN for extracting spatial patterns from multi-dimensional gas concentrations and LSTM for modeling temporal dependencies, integrated with Cox proportional hazards regression for probabilistic survival predictions. Using 30 years of DGA data, the proposed model achieved a C-index of 0.822, comparable to LSTM-only model while reducing training time by 89.3%. This 9.3× faster training speed makes the model highly suitable for industrial deployment where rapid model updates are essential. Optimization techniques including CBAM and mixed-precision training further enhanced efficiency.