A Surface Thermal Sensing Framework for Internal Winding Temperature Estimation in Oil-Immersed Converter Transformers

Accurate monitoring of internal winding temperature is essential for assessing the thermal state and operational reliability of oil-immersed transformers. However, direct deployment of distributed temperature sensors inside transformer windings is difficult because of insulation constraints, structural complexity, and potential reliability risks. To address this problem, this paper proposes a non-invasive internal winding temperature estimation method based on surface temperature sensing and a hybrid deep learning model. In the proposed framework, external surface temperature measurements are used as sensor inputs to infer the internal transient thermal state of the transformer. First, an extreme gradient boosting (XGBoost) model is employed to evaluate the contribution of different surface temperature measurement points and select the sensing locations that are most strongly correlated with internal winding temperature variations. Then, the selected surface temperature time-series data are used to train a Long Short-Term Memory (LSTM) network, which captures the temporal evolution of the transformer temperature field under different operating conditions. The proposed method is verified through both numerical simulation and experimental testing on a scaled single-phase oil-immersed converter transformer model (D-800/35) developed in this study. The results show that the proposed XGBoost-LSTM model can estimate internal winding temperature with an error of less than 1.5 K. Compared with direct internal sensing, the proposed method provides a non-invasive and sensor-efficient solution for internal temperature monitoring. The results demonstrate its potential for real-time thermal state estimation, condition monitoring, and fault diagnosis of oil-immersed converter transformers.