Deep Meta-Reinforcement Learning for the Flow-Pressure Coordination of a Vehicle PEMFC System Considering Multivariable Constraints

In the proton exchange membrane fuel cell (PEMFC), both air flow and cathode pressure affect the output performance and safety. To improve the PEMFC performance, this work presents a data-driven flow-pressure coordinated control (DD-FPCC) method, where the air compressor controller and backpressure valve controller are equated as two agents. A centralized training method enables the two agents to cooperate with each other in offline training. To facilitate this method, this work combines meta-reinforcement learning with multi-agent reinforcement learning and proposes an effective exploration multi-agent deep meta deterministic policy gradient (EEMA-DMDPG) algorithm, which includes two learners. The base learner is responsible for completing the learning of the underlying task. The meta-learner identifies the most suitable exploration noise parameters for the base learner, which improves the DD-FPCC performance. A simulation validation on a 75-kw PEMFC shows that the DD-FPCC improves the control performance in terms of the output voltage, oxygen excess rate (OER) and air pressure and prevents constraint violations.