Accurate prediction of the remaining useful life (RUL) of power transformers is critical for ensuring the safety, reliability, and intelligent operation of modern power systems. However, transformer operating data are typically nonlinear, nonstationary, and multi-source coupled, posing significant challenges for conventional models in feature extraction and temporal modeling. To overcome these limitations, this study proposes a hybrid predictive framework that integrates signal decomposition, intelligent optimization, and deep learning—the SSA-CEEMDAN-Transformer-BiGRU model. First, the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm is employed to perform multi-scale decomposition of the transformer oil temperature series, effectively isolating noise, periodic fluctuations, and long-term degradation trends. Second, the Sparrow Search Algorithm (SSA) is utilized to conduct global adaptive optimization of key hyperparameters in the Transformer-BiGRU network, thereby improving convergence speed and generalization capability. Finally, the Transformer module captures global temporal dependencies through its multi-head self-attention mechanism, while the BiGRU network characterizes local dynamic variations via a bidirectional gated structure. Experimental results on the ETTh2 dataset demonstrate that the proposed model substantially outperforms traditional statistical and deep learning approaches in both prediction accuracy and stability, achieving an R² of 0.9721 and an MSE of 0.031 on the test set. Ablation and feature-importance analyses further confirm the critical contribution of the CEEMDAN and SSA modules to overall performance enhancement. The findings indicate that the proposed methodology provides a high-precision, interpretable, and practically deployable solution for intelligent condition monitoring and lifetime management of power transformers.
Liu et al. (Tue,) studied this question.