Large-scale multi-agent deep reinforcement learning-based coordination strategy for energy optimization and control of proton exchange membrane fuel cell

Abstract

To maximize the net output power and ensure the safe operation of proton exchange membrane fuel cells (PEMFCs), a coordination strategy is proposed in this paper to realize better optimization and control of a cell’s oxygen excess ratio (OER). There are two agents in this strategy. An optimization agent is employed to obtain an ideal OER in different states by considering both the maximum net output power and the power control performance and safety of the PEMFC. A control agent effectively controls the OER of the PEMFC by determining the ideal OER. In addition, an evolutionary curriculum imitation large-scale multiagent deep reinforcement learning algorithm (ECILS-MADDPG) is introduced for this framework, which draws from the pedagogical technique of curriculum learning and imitation learning to improve the robustness of the coordination strategy. Proven by experiments, the proposed coordination strategy can obtain the maximum net power while maintaining OER control performance and the safety constraints of the PEMFC. Moreover, the proposed method increases the net power by up to 0.70% compared with an equivalent net power minimization strategy, and the integrated absolute error (IAE) in the OER is reduced by up to 99%.