Model-augmented safe reinforcement learning for Volt-VAR control in power distribution networks. (1st May 2022)

• Record Type: Journal Article
• Authors: Gao, Yuanqi; Yu, Nanpeng
• Is Part Of: Applied energy. Volume 313(2022)
• Publication Date: 2022-05-01
• Subjects: Volt-VAR control -- Data-driven -- Deep reinforcement learning -- Pathwise derivative -- Safe exploration; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2022.118762
• Languages: English
• ISSNs: 0306-2619

Abstract

Abstract: Volt-VAR control (VVC) is a critical tool to manage voltage profiles and reactive power flow in power distribution networks by setting voltage regulating and reactive power compensation device status. To facilitate the adoption of VVC, many physical model-based and data-driven algorithms have been proposed. However, most of the physical model-based methods rely on distribution network parameters, whereas the data-driven algorithms lack safety guarantees. In this paper, we propose a data-driven safe reinforcement learning (RL) algorithm for the VVC problem. We introduce three innovations to improve the learning efficiency and the safety. First, we train the RL agent using a learned environment model to improve the sample efficiency. Second, a safety layer is added to the policy neural network to enhance operational constraint satisfactions for both initial exploration phase and convergence phase. Finally, to improve the algorithm's performance when learning from limited data, we propose a novel mutual information regularization neural network for the safety layer. Simulation results on IEEE distribution test feeders show that the proposed algorithm improves constraint satisfactions compared to existing data-driven RL methods. With a modest amount of historical data, it is able to approximately maintain constraint satisfactions during the entire course of training. Asymptotically, it also yields similar level of performance of an ideal physical model-based benchmark. One possible limitation is that the proposed framework assumes a time-invariant distribution network topology and zero load transfer from other circuits. This is also an opportunity for future research. Highlights: A model-augmented RL algorithm for VVC is proposed to boost sample efficiency. A quadratic programming-based policy network is proposed to enhance the safety. A mutual information regularizer is designed to improve the constraint layer accuracy.