Automotive fuel cell performance degradation prediction using Multi-Agent Cooperative Advantage Actor-Critic model

The performance degradation of automotive proton exchange membrane cell (PEMFC) has long been a bottleneck hindering its commercial applications. Predicting fuel cell voltage degradation is vital, as it provides practical guidance for fuel cell health management to extend its lifetime. While the Reinforcement Learning (RL) model, Advantage Actor-Critic (A2C), has shown promise in predicting degradation, its high learning costs and poor prediction stability limit its application. In this work, a fuel cell mechanism model, which extracts degradation features according to different operating conditions, is integrated into a novel Multi-Agent Cooperative Advantage Actor-Critic (MAC-A2C) framework. The extracted aging parameters under different load currents are assigned to individual agents whose respective actor networks are updated asynchronously to learn the particular degradation trends. The model is trained and validated using a dataset from a self-designed 80 kW PEMFC engine operating on a city bus for 3566 h. Results indicate that our method effectively tracks the voltage degradation trend, achieves high prediction accuracy with a mean average percentage error of 0.38 %, reduces learning costs by 87.5 %, and significantly improves prediction stability. The proposed method can offer potential solutions for online health management, thereby extending the lifespan of fuel cell vehicles.