An Optimization Method for Grid Look-Ahead Dispatch Based on Adaptive Multi-Agent Reinforcement Learning

This paper presents an adaptive look-ahead scheduling optimization method for power grids, combining adaptive multi-agent reinforcement learning (MARL) and adaptive K-means clustering. As the integration of renewable energy sources such as wind and solar power increases, their variability and uncertainty pose significant challenges to traditional scheduling methods, especially regarding the scalability and generalization of single-agent reinforcement learning approaches. To address this, the proposed method uses adaptive K-means clustering to classify large-scale input scenarios and assigns them to corresponding agents for parallel training, effectively addressing the generalization limitations of single-agent methods. Additionally, a global shared experience pool is introduced, enabling agents to share experiences, accelerate learning, and improve convergence. A case study based on the IEEE 118-bus system demonstrates that the proposed method outperforms single-agent approaches in terms of scheduling efficiency, renewable energy integration, and computational time.