Physics-Guided Top-Oil Temperature Prediction for Operational Thermal-State Assessment of Oil-Immersed Power Transformers

Top-oil temperature is a key operational variable for oil-immersed power transformers, governing loading-guide calculations, hot-spot estimation, and insulation-aging risk screening. Under field operation, however, ONAN transformers exhibit equipment-specific thermal inertia, varying ambient boundaries, and changing cooling efficiency, which constrains both fixed-parameter thermal models and unconstrained data-driven predictors. This paper proposes a physics-guided framework, termed TIR-MSTLA, for short-term top-oil temperature prediction and transformer thermal-state assessment. A first-order heat-balance recursion builds a loading-guide-consistent physical reference from the load factor, ambient temperature, historical top-oil temperature, and nameplate parameters. Adaptive thermal parameters, residual learning, multi-scale time-delay attention, and dynamic bias correction then compensate for field deviations while preserving thermal-inertia consistency. The method is validated on one-year field datasets from 110 kV and 220 kV ONAN transformers. Relative to the fixed-parameter IEEE/IEC standard thermal model and three data-driven baselines, it reduces the test MAE by 77.0% and 53.1% on the two datasets and achieves MAE/RMSE values of 0.1450/0.3577 ◦C and 0.1776/0.4862 ◦C, with all continuous-segment absolute errors below 2 ◦C. The resulting error band provides a field-validated top-oil input for transformer condition monitoring, loading-margin assessment, and insulation-aging risk screening.