Remaining useful life prediction of vehicle-level fuel cells based on self-attention gated recurrent unit modeling

This paper introduces a modified relative voltage loss rate as an enhanced degradation indicator to characterize the primary operating points of eight onboard fuel cell city buses. The modified relative voltage loss rate is derived using a sliding window strategy combined with locally estimated scatterplot smoothing, effectively incorporating voltage decay acceleration under primary operating current conditions. Based on this refined degradation metric, a deep sequence model incorporating a self-attention gated recurrent unit is established to predict the remaining useful life of vehicle-level fuel cells. The self-attention mechanism enhances feature extraction, while the gated recurrent unit model captures temporal dependencies, forming a comprehensive time-series prediction framework. Comparative analysis demonstrates that the deep sequence model outperforms various conventional neural network architectures, including the baseline gated recurrent unit model and the self-attention long short-term memory model. The proposed deep sequence model achieves a mean absolute percentage error below 3.79% for all eight vehicle-level fuel cell systems, highlighting its robustness and generalizability.