A Low Carbon Collaborative Scheduling Model for Hybrid New Energy Power Systems Based on MARL

This paper proposes a multi-agent reinforcement learning (MARL) framework for low-carbon collaborative scheduling in hybrid energy systems. The model incorporates role-specific policy training, carbon-aware reward shaping, and sparse inter-agent communication to coordinate distributed energy resources (DERs), including storage, generation, and flexible loads. It is evaluated on three diverse datasets — empirical, synthetic, and microgrid — to assess generalizability across different energy environments. Results show the framework effectively reduces operational costs and carbon emissions while maintaining load balance and respecting system constraints. An ablation study further validates the contribution of key components such as entropy regularization, modular policy design, and consistency enforcement, each of which improves convergence, coordination, and robustness. The architecture supports scalability as agent numbers grow and remains resilient under uncertain forecasts and partial observability. Overall, the proposed MARL approach demonstrates strong potential for real-time, decentralized energy management in smart grids focused on sustainability and adaptability.