Online-monitoring-oriented prediction of transformer oil temperature using VMD-CEEMDAN and hybrid temporal learning

Abstract

Transformer oil temperature is an important online-monitoring indicator reflecting the thermal state, insulation stress, and operational safety margin of power transformers. Accurate short-term prediction of oil temperature can provide actionable support for condition monitoring, early warning, and operation and maintenance decision-making under fluctuating load and environmental conditions. However, oil-temperature series usually exhibit strong nonstationarity, multiscale coupling, and thermal-inertia-induced lag, which make direct end-to-end prediction on the raw sequence prone to unstable generalization during operating-condition transitions. To address this issue, this study develops a decomposition-enhanced hybrid temporal learning framework for short-term transformer oil-temperature prediction. The raw oil-temperature sequence is first decomposed by VMD and CEEMDAN into time-aligned multichannel components to reduce scale aliasing and improve the separability of operational fluctuations and disturbance-related patterns. These components are then modeled by a hybrid Transformer–BiGRU architecture to jointly capture cross-time-step dependency and local temporal evolution. Experiments on the ETTh2 transformer temperature dataset show that the proposed method outperforms representative baseline models, achieving RMSE = 0.0219, MAE = 0.0160, R ²  = 0.9977, MAPE = 4.87 %, sMAPE = 4.73 %, and WQE = 4.835. Compared with the BiGRU benchmark, the proposed model reduces RMSE, MAE, MAPE, sMAPE, and WQE by 24.33 %, 26.34 %, 19.90 %, 19.97 %, and 20.07 %, respectively. Compared with the strongest conventional baseline, Ridge, it further reduces RMSE, MAE, MAPE, sMAPE, and WQE by 5.19 %, 6.43 %, 3.37 %, 3.47 %, and 3.65 %, respectively. In addition to improved numerical accuracy, the model demonstrates better tracking of peak–valley variations and a more concentrated residual distribution, indicating stronger robustness under nonstationary operating conditions. These results suggest that the proposed framework is a promising data-driven tool for transformer condition monitoring and early warning, and may support more reliable thermal-state assessment in practical power-system operation.