Traditional time–frequency domain methods face critical limitations in predicting voltage degradation of proton exchange membrane fuel cells (PEMFCs). Time-domain models struggle to robustly separate long-term degradation-related low-frequency trends from contaminated voltage signals under highly dynamic and non-stationary conditions, while conventional frequency-domain analysis loses essential time-localized information during feature extraction. Both approaches exhibit significantly degraded prediction performance under limited data conditions. To overcome these challenges, this paper proposes a time–frequency fusion algorithm that integrates TimesNet with long short-term memory (LSTM), effectively combining 2D frequency-domain representations with 1D temporal memory to enhance voltage degradation prediction under dynamic conditions. Based on the capability of TimesNet-LSTM to extract low-frequency voltage features, a transfer learning technique grounded in low-frequency consensus knowledge (LCK-TL) is further developed. By selectively transferring low-frequency voltage features that robustly reflect aging patterns, LCK-TL considerably reduces distribution discrepancy between source and target domains, achieving joint optimization of predictive modeling and transfer mechanisms. Leveraging the inherently low computational cost of transfer learning, LCK-TL enables rapid multi-step predictions while maintaining accuracy, providing effective guidance for cross-device and cross-condition fuel cell health management. • 2D frequency analysis integrated with 1D temporal memory for PEMFC prognostics. • Co-optimization framework integrating prognostics and transfer learning. • Domain discrepancy reduction via low-frequency voltage characteristics. • Online cross-domain knowledge transfer for accurate PEMFC degradation prediction.
Wu et al. (Sun,) studied this question.