Abstract: The proton exchange membrane fuel cell (PEMFC), as a core component of new energy technologies, requires accurate performance degradation estimation for practical applications. While conventional data-driven approaches predominantly utilize nonlinear algorithms to forecast PEMFC system performance degradation and refine predictive accuracy through algorithmic architecture optimization, they frequently overlook critical challenges such as the multi-time-scale aging behaviors of individual components and the complex voltage recovery dynamics inherent in fuel cell operation. This study proposes a dual data-driven approach for fuel cell voltage prediction by extracting performance degradation signatures from voltage data. The methodology first decomposes and reconstructs raw voltage data into multiple aging time-scale sequences using a time-frequency domain decomposition algorithm (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise, CEEMDAN). Subsequently, distinct data-driven methods are implemented to separately predict the global degradation trend and local recovery features within the decomposed sequences. Experimental validation on dynamic PEMFC datasets demonstrates that the proposed method achieves 37.2%-43.0% prediction accuracy improvement compared with standalone approaches.