Multi-agent reinforcement learning for global virtual power plant collaborative scheduling…

The integration of high-penetration renewable energy sources (RES) into global power systems necessitates advanced scheduling strategies to ensure supply-demand balance. Virtual Power Plants (VPPs) serve as critical aggregators for distributed resources; however, coordinating VPPs across multiple regions is hindered by the curse of dimensionality, partial observability, and stochastic volatility. Conventional centralized optimization lacks scalability for real-time applications, while single-agent approaches fail to effectively address complex collaborative dynamics. To overcome these limitations, this paper proposes a collaborative scheduling framework based on Multi-Agent Reinforcement Learning (MARL). We model the global system as a multi-regional environment where heterogeneous agents operate under a Centralized Training with Decentralized Execution (CTDE) architecture. A composite reward function is designed to balance economic efficiency with RES absorption, utilizing an attention-based mechanism to exploit time-zone complementarity. Simulation results demonstrate that the proposed method significantly outperforms baseline strategies. Specifically, it achieves a global RES accommodation rate of 94.2% and maintains a minimal tie-line violation rate of 0.8%, compared to only 76.5% accommodation with rule-based heuristics. Furthermore, the approach exhibits superior robustness in extreme-volatility scenarios where standard methods degrade. This study validates the efficacy of distributed intelligence in solving large-scale energy dispatch problems, offering a scalable and privacy-preserving pathway for managing the Global Energy Interconnection.