Integrated Temporal-frequency Feature Fusion Algorithm Research for Fuel Cell Performance Degradation Prediction

The proton exchange membrane fuel cell (PEMFC), as a critical energy conversion device for new energy vehicles and stationary power stations, exhibits complex degradation mechanisms characterized by pronounced nonstationarity. Conventional time-domain analyses inadequately quantify dynamic fluctuation patterns, while frequency-domain approaches often neglect temporal dependencies, leading to fragmented representations of degradation modes when applied individually. To address these limitations, this study proposes a time-frequency fusion-based TimesNet-long short-term memory (LSTM) model. Leveraging TimesNet's multi-period temporal convolutional network, the model captures periodic frequency-domain features from degradation signals, while the LSTM architecture deciphers long-term temporal dependencies through its gated mechanisms. A cross-weighted time-frequency feature integration framework is further designed to synergistically optimize degradation characterization. Experimental results demonstrate that the proposed method effectively fuses multi-dimensional degradation information from both domains, achieving 66.9% and 40.9% improvements in prediction accuracy, respectively, compared to single-domain approaches. Multi-dimensional analyses further validate the superiority of the framework. This work establishes a high-precision cross-domain analytical paradigm for PEMFC performance evaluation.