Predicting Transformer Hotspot Temperature: A Comparative Study of FNN and LSTM-PCA

Power transformers are a key element of the power system, and their availability is essential for maintaining it reliable and stable. High transformer loading leads to elevated operating temperatures, which in turn accelerate insulation deterioration and ultimately shorten the expected service life of these critical components. This paper aims to accurately predict the hotspot temperature, serving as a key indicator of transformer health and lifespan. Accurate prediction of this parameter enables adaptive load management and helps prevent premature aging. While previous studies have employed both physics-based and machine learning models to estimate hotspot temperatures, few have examined performance under limited data conditions. This paper addresses this gap by presenting a comparative study of a long short-term memory (LSTM) model with principal component analysis (PCA) and a feedforward neural network (FNN), both using the same input data. Their performance is benchmarked against the IEEE thermal model across two cooling types (with and without forced air). Results from a 22-month dataset covering 17 oil-immersed transformers indicate that both data-driven models outperform the IEEE model under forced air cooling conditions. However, only the FNN performs better than the IEEE model for the oil natural air natural cooling type, while the LSTM-PCA struggles to achieve this. These findings highlight the practical advantages of data-driven models, particularly FNNs for real-world transformer monitoring and asset management.