Large-scale multi-agent reinforcement learning-based method for coordinated output voltage control of solid oxide fuel cell

Abstract

For a solid oxide fuel cell (SOFC) power system containing a hydrogen fuel reformer and a DC–DC converter, it is necessary to coordinate the controllers of the two devices in order to maintain effective power tracking control and also prevent constraint violations of fuel utilization. A data-driven SOFC output voltage coordinated control method is proposed for maintaining a stable fuel utilization whilst satisfying load demand requirements in this paper. To that end, a Pygmalion effect-based multi-agent double delay deep deterministic policy gradient algorithm (PEB-MA4DPG) is presented in this work. This algorithm is a combination of a comprehensive exploration, imitation learning and curriculum learning policy, which altogether constitute a coordinated strategy of high robustness. By employing the controllers for the fuel reformer and DC–DC converter as the two agents, the proposed algorithm generates an optimal coordinated policy via centralized training and distributed implementation. The experimental verify the effectiveness of the proposed method.