Multi-agent reinforcement learning based coordinated control of PEMFC gas supply system

This paper proposes a multi-agent reinforcement learning algorithm for the coordinated gas supply control of PEMFC to improve efficiency and extend service life by regulating differential pressure of anode and cathode, oxygen excess ratio and anode inlet humidity. The coordinated gas supply model of PMEFC is firstly established to provide the environment used to interact with agents, and then the control frame considering air supply, hydrogen supply and anode circulation gas supply is designed to realize coordinated control. In the structure of multi-agent reinforcement learning, each agent has a pair of critic networks and actor networks for action value estimation and action selection respectively. By designing appropriate reward function and giving reward feedback, the agents update their actor and critic networks with the method of centralized training and decentralized execution. The control algorithm is finally evaluated in the matlab&simulink environment, and compared with the traditional PID. The results show that the proposed algorithm is superior to traditional PID in steady-state error and dynamic response.