Frequency Coordination Control of Multi-Area Power Systems Based on Multi-Agent Deep Reinforcement Learning

As renewable energy sources like wind and solar increasingly penetrate power systems, frequency fluctuations have become more pronounced, creating a more complex and variable environment for balancing power supply and demand. This paper proposes a data-driven frequency control method based on a multi-agent reinforcement learning algorithm. First, the dynamic responses of both thermal power units and the system are evaluated, and a two-area frequency response model is developed. Secondly, the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm based on the framework of “centralized training, decentralized execution” is adopted. During centralized training, agents share experiences and collaboratively learn to optimize control strategies for coordinated operation across areas. In decentralized execution, each agent autonomously implements the learned strategies using local state data to generate control commands for its units. Additionally, multi-agent controller are designed based on the developed model. Finally, the simulation results verify that the controller based on MADDPG can quickly reduce frequency deviations and tie-line power deviations, and the algorithm demonstrates a fast convergence rate during the training process.