Enhancing economic efficiency and operational stability of high-penetration renewable distribution networks: A multi-timescale coordinated optimization method leveraging multi-agent graph reinforcement learning

The increasing integration of distributed generation into high-penetration renewable distribution networks challenges conventional single-dimensional optimization strategies. These strategies struggle to balance multi-objective requirements, such as reducing network losses, ensuring voltage stability, and accommodating renewable energy. Existing approaches often neglect critical interactions among network topology, reactive power compensation, and active power regulation, leading to compromised performance under high DG penetration. This paper proposes a multi-agent graph reinforcement learning framework for multi-timescale coordinated rolling optimization of topology, reactive power, and active power. Our method integrates graph convolutional networks to model grid-structured data and double deep Q-networks to address complex decision-making. By hierarchically coordinating slow- and fast-timescale resources through rolling optimization, it adaptively manages stochastic fluctuations in DG outputs and loads. Case studies demonstrate enhanced cross-timescale coordination, achieving higher renewable utilization, stabilized voltage profiles, improved operational flexibility, and robust uncertainty handling capabilities. Furthermore, the framework exhibits excellent scalability and computational efficiency, enabling effective optimization even for large-scale distribution networks, which is critical for practical deployments in complex high-renewable scenarios. The framework offers a scalable solution for intelligent grid management in high-renewable scenarios.