Accurate prediction of voltage degradation in Proton Exchange Membrane Fuel Cells (PEMFCs) is essential for improving operational reliability and enabling predictive maintenance. In this paper, a hybrid framework integrating Variational Mode Decomposition (VMD) and Transformer network is proposed for short-term voltage degradation prediction. Initially, original voltage signals are denoised using wavelet thresholding and subsequently decomposed through VMD to isolate frequency-specific degradation components. Each mode is independently modeled using a Transformer network to capture temporal dependencies, and the final voltage trajectory is obtained by summing the predicted components. The proposed method is evaluated on PEMFC degradation data and outperforms both LSTM-only and Transformer-only models, achieving an RMSE of 0.0032 and a MAPE of 0.06%. The model exhibits strong capability in isolating and learning low-frequency degradation patterns that primarily govern PEMFC voltage decay. These results underscore the effectiveness of combining frequency-domain decomposition with deep sequence modeling, as VMD enhances signal interpretability by separating critical degradation-related components while suppressing noise and irrelevant features. The proposed framework offers a promising solution for fault diagnosis and remaining useful life prediction in PEMFCs.