Semimonthly

ISSN 1000-1026

CN 32-1180/TP

+Advanced Search 中文版
Self-adaptive Uncertainty Economic Dispatch Based on Deep Reinforcement Learning
Author:
Affiliation:

School of Electrical Engineering and Automation, Wuhan University, Wuhan 430072, China

Fund Project:

This work is supported by National Key R&D Program of China (No. 2016YFB0900100), Excellent Youth Foundation of Hubei Province of China (No. 2018CFA080) and National Natural Science Foundation of China (No. 51707136).

  • Article
  • |
  • Figures
  • |
  • Metrics
  • |
  • Reference
  • |
  • Related
  • |
  • Cited by
  • |
  • Materials
    Abstract:

    When the large-scale intermittent generation sources including wind and photovoltaic generation connects to the power system, in order to cope with their uncertainty, the economic dispatch model for power system needs to be built based on the uncertainty modeling. The accuracy of modeling will directly affect the accuracy of the dispatch results. However, when considering the complex uncertainty of both load and intermittent generation sources such as wind and photovoltaic generation, it is particularly difficult to accurately model the overall uncertainty of the system. In view of this problem, the deep deterministic policy gradient (DDPG) algorithm in the deep reinforcement learning is introduced. The work of uncertainty modeling is avoided. Instead, the uncertainty is adapted by the DDPG algorithm relied on the mechanism of interacting with the environment and improving the strategy based on feedbacks. In order to guarantee the applicability of the algorithm, the generalization method for the DDPG model is proposed. Aiming at the stability problem of the algorithm, two mechanisms are designed, including perception-learning ratio adjustment and experience replay improvement. The result of example shows that the dynamic economic dispatch problem of power systems in any scenario can be solved by the proposed method based on adapting the system uncertainty.

    表 1 调度结果经济性对比Table 1 Economic comparison of dispatch results
    图1 DDPG算法流程Fig.1 Flow chart of DDPG algorithm
    图1 DDPG算法流程Fig.1 Flow chart of DDPG algorithm
    图2 训练过程中的总回报平均值Fig.2 Average total reward during training process
    图3 典型日调度结果Fig.3 Dispatching results for a typical day
    图4 系统备用对不确定性的平衡Fig.4 Power uncertainty balanced by system reserve
    图5 功率缺额概率分布与系统预留备用情况Fig.5 Situation of power shortage probability distribution and system reserve
    图6 机制设计对DDPG算法收敛性的影响Fig.6 Effect of mechanism design on convergence of DDPG algorithm
    图 Critic网络及其目标网络结构Fig. Structure of critic network and its target network
    图 Actor网络及其目标网络结构Fig. Structure of actor network and its target network
    图 Critic网络及其目标网络结构Fig. Structure of critic network and its target network
    图 过拟合时第一隐层网络参数Fig. The parameters of the first hidden layer when overfitted
    图 网络收敛情况(损失函数)Fig. Network convergence (Loss function)
    图 过拟合时第一隐层网络参数Fig. The parameters of the first hidden layer when overfitted
    图 网络收敛情况(损失函数)Fig. Network convergence (Loss function)
    图 损失函数曲线发散Fig. Loss function curve diverges
    图 学习停滞Fig. The stagnation of learning
    图 训练时间随机组规模的变化Fig. The training time varies with the scale of power units
    Reference
    Related
    Cited by
Get Citation

PENG Liuyang,SUN Yuanzhang,XU Jian,et al.Self-adaptive Uncertainty Economic Dispatch Based on Deep Reinforcement Learning[J].Automation of Electric Power Systems,2020,44(9):33-42.DOI:10.7500/AEPS20190706003

Copy
Share
Article Metrics
  • Abstract:
  • PDF:
  • HTML:
  • Cited by:
History
  • Received:July 06,2019
  • Revised:December 20,2019
  • Adopted:
  • Online: May 10,2020
  • Published: