An online adaptive evolution method for traction transformer temperature estimation model based on multi-scale periodic updates

• A two-stage progressive update architecture. • A short-term dynamic calibration method based on transfer learning. • A long-term knowledge consistency maintenance method based on knowledge distillation. • A balance between adaptability and stability. Accurate estimation of traction transformer winding temperature is critical for assessing operational status and ensuring the safety and reliability of urban rail system. However, conventional offline models, trained on historical data, exhibit limited adaptability to evolving system conditions, leading to accumulated estimation errors and diminished long-term usability. Therefore, we propose an online adaptive evolution framework for transformer temperature estimation, incorporating a multi-scale periodic update mechanism. Specifically, a two-stage progressive update strategy captures temperature variations across multiple temporal scales, maintaining accuracy under varying operational conditions. On one hand, a transfer learning-based feature calibration method is proposed, partially freezing model parameters. It enables rapid adaptation to short-term load fluctuations. On the other hand, to preserve long-term knowledge integrity, a self-distillation strategy is introduced. Through self-distillation of historical model, thermodynamic consistency preservation with adaptability to scheduling changes and equipment aging can be balanced. Experimental results demonstrate that the proposed method achieves a computational accuracy exceeding 98%, while improving computational efficiency by 66.89% compared to full model retraining. Overall, the approach significantly enhances the long-term usability and adaptability of transformer temperature models, providing strong support for deployment in urban rail transit systems.