Multi-agent deep meta-reinforcement learning-based active fault tolerant gas supply management system for proton exchange membrane fuel cells

The output performance of proton exchange membrane fuel cells (PEMFCs) is highly susceptible to the influence of various parameters, which in turn are affected by PEMFC failures. A data-driven active fault-tolerant control (DD-AFTC) method is proposed to achieve the stable control of a PEMFC in the event of a fault. In addition, this proposed method combines meta-reinforcement learning with multiagent reinforcement learning to offer a distributed multiagent deep meta-deterministic policy gradient (DMA-DMDPG) algorithm, which provides the agents with independent multitask cooperative learning capabilities, thus ensuring excellent robustness. This algorithm consists of a meta-learner and a base learner. The base learner equates the hydrogen controller and the oxygen controller as independent decision-making agents. At the same time, the meta-learner is responsible for identifying PEMFC faults and selecting the appropriate joint policy according to each specific PEMFC failure. Experimental validation for a 75 kW PEMFC illustrates that DD-AFTC can improve control performance in terms of output voltage and oxygen excess rate (OER) under fault-induced conditions.

Automated summary

This paper proposes a data-driven active fault-tolerant control method for stable control of proton exchange membrane fuel cells (PEMFCs) in the event of a fault. It combines meta-reinforcement learning with multiagent reinforcement learning to provide independent multitask cooperative learning capabilities, ensuring excellent robustness.