引用本文: | 刘林鹏,朱建全,陈嘉俊,叶汉芳.基于柔性策略-评价网络的微电网源储协同优化调度策略[J].电力自动化设备,2022,42(1): |
| LIU Linpeng,ZHU Jianquan,CHEN Jiajun,YE Hanfang.Cooperative optimal scheduling strategy of source and storage in microgrid based on soft actor-critic[J].Electric Power Automation Equipment,2022,42(1): |
|
摘要: |
近年来,微电网中的可再生能源与储能占比不断增大,给其优化调度带来了新的挑战。针对微电网源储协同调度问题中非凸非线性约束带来的求解困难,利用深度强化学习算法构建基于数据的策略函数,通过不断地与环境进行交互学习寻找最优策略,避免了对原非凸非线性问题的直接求解。考虑到训练过程中策略函数可能不满足安全约束,进一步提出了一种利用部分模型信息的微电网源储协同优化调度安全策略学习方法,得到了满足网络安全约束的优化策略。此外,针对强化学习的智能体在训练过程中与环境的交互耗时较长的问题,采用神经网络对环境进行建模以提高学习效率。 |
关键词: 微电网 可再生能源 储能 柔性策略-评价网络 强化学习 深度学习 安全约束 |
DOI:10.16081/j.epae.202110036 |
分类号: |
基金项目:国家自然科学基金资助项目(51977081);电力系统国家重点实验室资助课题(SKLD20M15);广东省自然科学基金资助项目(2019A1515010877);广东省普通高校青年创新类人才项目(2019GKQNCX040) |
|
Cooperative optimal scheduling strategy of source and storage in microgrid based on soft actor-critic |
LIU Linpeng, ZHU Jianquan, CHEN Jiajun, YE Hanfang
|
School of Electric Power, South China University of Technology, Guangzhou 510640, China
|
Abstract: |
In recent years, the proportion of renewable energy and energy storage in microgrid is increasing, which brings new challenges to its optimal scheduling. Aiming at the difficulty in solving the cooperative optimal scheduling problem of source and storage in microgrid due to the non-convex nonlinear constraints, the deep reinforcement learning algorithm is used to construct the data-based strategy function, and the optimal strategy is found out through continuous interactive learning with the environment, so that avoiding the direct solution of the original non-convex nonlinear problem. Considering the strategy function may not meet the security constraints in the training process, furthermore, a learning method of cooperative optimal scheduling secure strategy of source and storage in microgrid based on partial model information is proposed, and the optimal strategy meeting the network security constraints is obtained. In addition, aiming at the problem of long time-consuming due to the interaction between agents and environment in the training process for reinforcement learning, the neural network is used to model the environment, so as to improve the learning efficiency. |
Key words: microgrid renewable energy energy storage soft actor-critic reinforcement learning deep learning security constraint |