Power Transformer Health Index and Life Span Prediction Using Supervised Machine Learning: A Hybrid Autoencoder Enhanced Gradient Boosting Approach

Abstract

The health index (HI) assessment serves as a vital indicator for ensuring reliability and evaluating the operational state of power transformers. However, condition monitoring systems often operate under constraints of computational efficiency and latency, necessitating compact yet informative feature representations. This study proposes a hybrid model that jointly optimizes the extraction of deep features through autoencoders using a gradient boosting regressor (GBR) to estimate transformer HI and lifespan, resource-efficiently, without accuracy loss. The proposed autoencoder-GBR model is subsequently benchmarked against traditional feature reduction techniques and a full uncompressed feature set, together with the conventional as well as deep learning-based regression algorithms. The comparative analysis demonstrates that the recommended model has been able to perform better with an average R2 performance of 0.993 and root mean squared error (RMSE) of 1.466 with dimensionality reduction of features by 35.7%, which benefits both predictive ability and computational efficiency. The proposed model has less inference latency and has an average model size with insignificant peak memory. It also shows strong robustness with acceptable predictive performance under synthetic noise and missing feature scenarios, hence ensuring runtime feasibility and reliability for a real-world monitoring environment. The same autoencoder-GBR model is employed for lifespan prediction, which achieves 0.9946~R2 and 1.3022 RMSE, further confirming its robustness. Overall, these findings demonstrate that the nonlinear feature compression not only boosts prediction accuracy but also ensures scalability, making the proposed framework highly suitable for real-time and field-deployable transformer health monitoring systems.