Active fault-tolerant coordination energy management for a proton exchange membrane fuel cell using curriculum-based multiagent deep meta-reinforcement learning

Abstract

This paper addresses the challenge of active fault-tolerant coordination control (AFTCC) for proton exchange membrane fuel cells (PEMFCs), which are complex nonlinear systems with multiple inputs and outputs. Conventional fault-tolerant control methods cannot properly coordinate multiple operating variables and prevent constraint violations in PEMFCs. Our proposed AFTCC method seeks to stabilize the output performance of four operating variables and avoid PEMFC operating constraint violations during failure scenarios. Our method is supported by a curriculum-based multiagent deep meta-deterministic policy gradient (CMA-DMDPG) algorithm, which integrates meta-reinforcement learning, multiagent reinforcement learning and curriculum learning to achieve multitask collaboration of multiple agents, thereby enhancing PEMFC robustness. The algorithm consists of a meta-learner and a base learner. The base learner regards the hydrogen controller, oxygen controller, pump controller and radiator controller as four independent agents and thus achieves a cooperative control policy. The meta-learner detects PEMFC faults and selects an appropriate cooperative control policy. The performance of AFTCC under various stochastic and fault conditions is evaluated using a 75 kW PEMFC model. The results showed that the performance of AFTCC surpassed 11 other fault-tolerant control methods in terms of output voltage, oxygen excess ratio, and stack temperature, and avoided constraint violations.