Interval prediction of fuel cell degradation based on voltage signal frequency characteristics with TimesNet-GPR under dynamic conditions

Proton exchange membrane fuel cells (PEMFCs) are crucial modern sustainable energy generation devices. The accurate assessment of their state of health (SOH) and the forecast of their remaining useful life (RUL) are critical for their practical deployment. Current mainstream methods typically use time-domain voltage decay as the health indicator (HI) and rely on recurrent neural networks. However, PEMFC voltage decay results from multiple factors, including internal component degradation, changes in operating conditions, and environmental impacts. Low-frequency domain analysis can effectively detect degradation in the proton exchange membrane and gas diffusion layer, leading to more accurate SOH estimation for fuel cells. This study reshapes time-domain voltage signals into frequency factors in a 2D space based on frequency domain features to more accurately reflect the aging characteristics of PEMFCs. We propose a TimesNet-GPR method to accurately quantify the uncertainty in degradation prediction, demonstrating good adaptability with different lengths of training data and various dynamic conditions. This method uses TimesNet for point estimation prediction, overcoming the limitations of neural networks in capturing long-term dependencies, improving point estimation accuracy by 39.18%–70.14% on dynamic cycling condition datasets. In order to evaluate uncertainty during point estimation and provide more accurate confidence interval predictions, Gaussian Process Regression (GPR), is utilized.