Multi-agent energy scheduling in renewable-rich grids using a cooperative reinforcement learning game framework

The rapid integration of renewable and distributed energy resources has increased uncertainty and operational complexity in modern smart grids, challenging traditional energy scheduling methods that rely on static or deterministic assumptions. To address these challenges, this paper proposes a risk-aware and fair multi-agent energy scheduling framework for renewable-rich power systems by integrating reinforcement learning with cooperative game theory. In the proposed framework, generators, energy storage systems, prosumers, and electric vehicles are modeled as autonomous agents capable of forming dynamic coalitions to enhance renewable energy utilization and reduce system operating costs. Uncertainties in renewable generation and demand are explicitly captured through a risk-adjusted coalition value function based on entropic value-at-risk. Coalition coordination and energy scheduling policies are learned using Q-learning and a multi-agent proximal policy optimization algorithm under a centralized training and decentralized execution paradigm, enabling stable, adaptive, and fairness-aware decision making. A decentralized settlement mechanism is employed to ensure transparent and equitable payoff allocation among coalition members. Simulation studies conducted on the IEEE 30, 57, and 118-bus test systems demonstrate that the proposed approach achieves operating cost reductions of 5.55%, 6.09%, and 5.90%, respectively, along with renewable energy utilization improvements of 10.84%, 8.37%, and 12.91%, compared to non-cooperative scheduling strategies. The results highlight the effectiveness of the proposed cooperative reinforcement learning framework as a scalable and sustainable solution for intelligent energy management in next-generation green power systems.