A parallel multi-scenario learning method for near-real-time power dispatch optimization. (1st July 2020)

• Record Type: Journal Article
• Authors: Guan, Jinyu; Tang, Hao; Wang, Ke; Yao, Jianguo; Yang, Shengchun
• Is Part Of: Energy. Volume 202(2020)
• Publication Date: 2020-07-01
• Subjects: Power dispatch -- Deep reinforcement learning -- Near-real-time -- Policy network -- Parallel environments; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2020.117708
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: Power dispatch problems become more complex when the weight of uncertain renewable resources in the power system gradually increases in recent years. To make use of renewable energy, such as wind energy, more adequately, wisely and intelligently, higher requirements are placed on the level of inter-region power dispatch coordination. In the context, solving the problem of power dispatch on a large scale in near-real-time (5 min in this paper) becomes more important. In this paper, the power dispatch was treated as a sequential decision-making problem and Deep Reinforcement Learning (DRL) with continuous control was introduced to offer a smarter solution. In this way, we designed a novel interactive learning environment based on the economic power dispatch model for the DRL algorithm and we proposed two feasible implementations to handle the different application scenarios. As a result, DRL with a continuous control method has a great performance in our proposed implementations. Moreover, we found that dispatching data richness has a significant influence on the generalization of the learned policy. Graphical abstract: Image 1 Highlights: A novel learning method for ED within 5-min considering the injection of large-scale wind power is presented. This paper designed an interactive parallel learning framework based on multiple source-lord scenarios. It improves learning efficiency and brings effective utilization of historical data. Data richness has a significant influence on the generalization of the learned policy.