基于Transformer及其变体的商用质子交换膜燃料电池系统电压衰减预测

Abstract Transformer and its variants show significant potential for predicting proton exchange membrane fuel cell performance degradation, enabling accurate capture of degradation patterns to inform control strategies and extend lifespan. However, despite advancements, their applicability to commercial high-power fuel cells remains unclear, as existing researches focus primarily on small-scale laboratory stacks. Addressing this gap, this study investigates a 60 kW commercial fuel cell system under two 1000-hour aging test modes with different hydrogen supply conditions (ambient vs. low temperature). A characteristic current-based data extraction method was employed for the raw data associated with each mode. Three representative characteristic currents were selected based on the current distribution of dynamic load cycles for data extraction. Following the preprocessing, aging datasets for three characteristic currents were obtained. Single cell voltage was selected as the aging feature parameter to construct Transformer and four variants (Informer, Half-Transformer, Half-Informer, and Autoformer) for degradation prediction. Comparative analysis revealed Autoformer’s superior aging voltage prediction accuracy. Its robustness was further validated under multi-step prediction, training set missing, and multivariate input scenarios, maintaining high accuracy across diverse conditions. The deviation of absolute prediction errors at the 80 % and 90 % cumulative distribution levels remained below 10 mV. These results demonstrate Autoformer’s strong potential for integration into fuel cell control systems, offering promising applications in health management to enhance practical value.