Scholarlink

This paper proposed a novel distributed secondary control strategy for islanded alternating current (AC) microgrids to compensate for frequency and voltage deviations caused by primary droop control. The strategy adopts an “offline training–online deployment” reinforcement learning framework to enhance power quality, in which two independent twin delayed deep deterministic policy gradient (TD3) agents are trained separately for frequency and voltage regulation. Based on local observations and information from neighboring nodes, the agents generate control actions in real time to dynamically adjust the droop control. To improve learning efficiency and control precision, a normalized reward function is designed, incorporating deviation magnitude, node consistency, and action penalty. The proposed control system operates without centralized coordination and relies solely on local communication to achieve global cooperation. Simulation results on an islanded microgrid with four inverter-based distributed generators (DGs) demonstrate that the TD3-based controllers significantly reduce frequency and voltage overshoot, shorten regulation time, and decrease steady-state error compared to traditional proportional-integral (PI) controllers. These results verify the superior dynamic response and robustness of the proposed method.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Reinforcement Learning Based Secondary Control Strategy for Islanded Microgrids

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yuntao Shi,
  • Haifeng Guo,
  • Xiang Yin,
  • Keqing Gao,
  • Shufeng Zhang

摘要

This paper proposed a novel distributed secondary control strategy for islanded alternating current (AC) microgrids to compensate for frequency and voltage deviations caused by primary droop control. The strategy adopts an “offline training–online deployment” reinforcement learning framework to enhance power quality, in which two independent twin delayed deep deterministic policy gradient (TD3) agents are trained separately for frequency and voltage regulation. Based on local observations and information from neighboring nodes, the agents generate control actions in real time to dynamically adjust the droop control. To improve learning efficiency and control precision, a normalized reward function is designed, incorporating deviation magnitude, node consistency, and action penalty. The proposed control system operates without centralized coordination and relies solely on local communication to achieve global cooperation. Simulation results on an islanded microgrid with four inverter-based distributed generators (DGs) demonstrate that the TD3-based controllers significantly reduce frequency and voltage overshoot, shorten regulation time, and decrease steady-state error compared to traditional proportional-integral (PI) controllers. These results verify the superior dynamic response and robustness of the proposed method.