Investigation of fuel cell stack performance degradation based on 1000 h durability experiments and long short-term memory prediction frameworks under dynamic load conditions

• 1000 h durability tests of a 5-kW stack were conducted under dynamic load conditions. • LSTM and Attention-LSTM models were developed with 16 key parameter dimensions. • Voltage degradation speeds range from 25 to 60 μV h -1 among all current modes. • LSTM model shows superior prediction capabilities in tracking the voltage dynamics. • Attention-LSTM model performs better when the advanced forecast time reaches 200 h. Investigating the proton exchange membrane fuel cell (PEMFC) stack performance degradation phenomena is of vital importance for product development. In the study, the 1000 h durability experiment of a 5-kW fuel cell stack was performed under dynamic cyclic test conditions, and the test data containing 16 key parameters was utilized to develop the performance prediction framework based on long short-term memory (LSTM) model and LSTM model incorporating attention mechanism (Attention-LSTM). Data preprocessing and postprocessing for eight current modes as well as incremental learning approach were also presented. Experimental results show that the voltage degradation ratio is about 2.0% at the total dynamic cyclic duration of 500 h and approximately 4.8% at 1000 h. The degradation ratio at higher stack operating currents is found larger than that of lower operating currents. The calculated voltage degradation speeds among all current modes fall within the range of 25∼60 μV h -1 . When it comes to model prediction performances, both LSTM and Attention-LSTM models could effectively capture the voltage variations under current rising and dropping conditions. The LSTM model exhibits superior transient prediction capabilities near current change moments while the Attention-LSTM model demonstrates smaller prediction deviations at relatively stable conditions. When the advanced forecast time reaches or exceeds 200 h, the Attention-LSTM model predictions agree better with the bench test data, and it maintains consistent prediction accuracy across different current modes. The study contributes to fuel cell stack durability performance analysis and degradation prediction.