MAPIRL: A Hyperbolic Tangent-Enforced Physical-Informed RL for Multi-IESs Optimal Dispatch

In optimal dispatch (OD) of multiple integrated energy systems (MIES), purely data-driven reinforcement learning (RL) methods often encounter challenges such as transient data boundaries, robustness, and interpretability. For this problem, this paper proposes a multi-agent physics-informed reinforcement learning (MAPIRL) method for MIES optimal dispatch. MAPIRL analytically integrates safety constraints using hyperbolic tangent functions, implementing a physics-informed learning process that enforce these constraints within the actor networks. The well-trained MAPIRL can achieve highly generalized real-time decision making. The MAPIRL method is compared with none-physics-based classical RL in a MIES market. The results demonstrate that MAPIRL not only facilitates highly safe and reliable dispatch decisions but also surpasses other methods in convergence efficiency. Additionally, embedding physics knowledge enhances the interpretability for the intelligent OD process.