Scholarlink

<p>The operation of a microgrid (MG) system with multiple nodes not only needs to solve the optimization problem of economic dispatch but also has to consider the optimization goals of the safe and environmentally friendly operation. Therefore, the purpose of each node with renewable resources is to collaborate to achieve the optimization of multiple objectives. In this paper, the authors shall design a deep reinforcement learning (DRL) algorithm to perform the multi-objective optimization which can handle continuous action space and determine the specific output power of each device. Unlike the existing algorithms that learn policies with holistic reward signals, the proposed algorithm decomposes the reward into multiple parts and trains multiple critic networks for sub-objectives to get the Pareto optimal solutions. The proposal of single-actor multi-critic architecture not only can avoid task-specific local optimal policies but also does not need to set weight values. The effectiveness of the algorithm is verified by case studies on a modified IEEE-30 bus system and a modified IEEE-118 bus system. After training, the DRL agent can adapt to the high uncertainty of the photovoltaics and exploit the capacity of battery energy storage stations safely, which is more practical in a real system.</p>

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

Eco-Friendly Multi-Objective Dispatching Strategy of Microgrid System Based on Deep Reinforcement Learning with SAMC Architecture

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Xiaowen Wang,
  • Shuai Liu,
  • Qianwen Xu,
  • Xinquan Shao

摘要

The operation of a microgrid (MG) system with multiple nodes not only needs to solve the optimization problem of economic dispatch but also has to consider the optimization goals of the safe and environmentally friendly operation. Therefore, the purpose of each node with renewable resources is to collaborate to achieve the optimization of multiple objectives. In this paper, the authors shall design a deep reinforcement learning (DRL) algorithm to perform the multi-objective optimization which can handle continuous action space and determine the specific output power of each device. Unlike the existing algorithms that learn policies with holistic reward signals, the proposed algorithm decomposes the reward into multiple parts and trains multiple critic networks for sub-objectives to get the Pareto optimal solutions. The proposal of single-actor multi-critic architecture not only can avoid task-specific local optimal policies but also does not need to set weight values. The effectiveness of the algorithm is verified by case studies on a modified IEEE-30 bus system and a modified IEEE-118 bus system. After training, the DRL agent can adapt to the high uncertainty of the photovoltaics and exploit the capacity of battery energy storage stations safely, which is more practical in a real system.