Research on Distributed Optimal Scheduling of Virtual Power Plants Based on Multi-Agent Reinforcement Learning

Virtual Power Plants (VPPs), as a key technology for integrating Distributed Energy Resources (DERs) and demand-side resources, face challenges in scheduling, such as multi-agent coordination, uncertainty, and complex constraints. Traditional centralized optimization methods struggle to adapt to distributed scenarios, while existing reinforcement learning studies often focus on single-agent systems, neglecting the cooperation and competition among multiple agents [1]. The multi-energy virtual power plant optimization scheduling framework based on the PPO algorithm proposed by Sun Dongchuan et al. [2] achieves multi-energy coordination through a hierarchical attention mechanism, providing important reference for this research. This study creatively proposes a distributed optimization scheduling framework based on Multi-Agent Reinforcement Learning (MAPL). By integrating game theory with Deep Reinforcement Learning (DRL), a multi-agent collaborative strategy is developed to address scheduling problems in complex scenarios. A hierarchical attention mechanism and a distributed reward allocation mechanism are designed to achieve multi-energy coordination optimization for electricity, heat, and storage [3]. The results show that, compared to traditional centralized optimization, the proposed method reduces operational costs by 15.2% and increases the renewable energy consumption rate by 11.5%.