A Policy optimization-based Deep Reinforcement Learning method for data-driven output voltage control of grid connected solid oxide fuel cell considering operation constraints

Abstract

Solid oxide fuel cells (SOFCs) have many applications in microgrids, but they often face the challenge of maintaining the fuel utilization within a safe range, which affects their lifespan. To address this issue, we propose a Data-driven Output voltage control (DDD-OVC) approach that treats the SOFC output voltage controller as an intelligent agent. The agent can design a reward function that incorporates the fuel utilization constraint and learn the optimal control policy through offline training. The objective is to optimize both the SOFC output voltage control performance and lifetime. Furthermore, we develop a policy optimization-based Deep Reinforcement Learning (PO-DRL) algorithm that adopts the idea of proximal policy optimization to enhance the learning speed and convergence of the agent, as well as the policy stability and quality of DDD-OVC. We validate the effectiveness of our method on a 5kW SOFC model.

Keywords

• Solid oxide fuel cell
• Policy optimization-based Deep Reinforcement Learning
• Data-driven output voltage control
• 5kw SOFC model
• Proximal policy optimization