To improve the prediction accuracy of performance degradation in proton exchange membrane fuel cells (PEMFCs)

particularly the nonlinear degradation of stack voltage

and toovercome the limitations of traditional neural networks in capturing long-term dependencies

this paper proposes a prediction method that optimizes TimesNet based on Bayesian optimization (BO). The method utilizes frequency domain analysis to extract low-frequency features from voltage signals and reshapes them into two-dimensional tensors

thereby achieving high-precision predictions of stack voltage degradation trends. Additionally

Bayesian optimization automatically tunes the hyperparameters of TimesNet

enhancing the model’s ability to capture long-term dependencies. Experimental results show that the proposed method reduces prediction errors by 72.33%~80.71% under dynamic cycling conditions of fuel cells. The method also demonstrates strong robustness and adaptability across various operating conditions

which verifies its excellent generalizability and effectiveness.