Mode-decomposition memory reinforcement network strategy for smart generation control in multi-area power systems containing renewable energy. (1st February 2022)

• Record Type: Journal Article
• Authors: Yin, Linfei; Wu, Yunzhi
• Is Part Of: Applied energy. Volume 307(2022)
• Publication Date: 2022-02-01
• Subjects: Smart generation control -- Empirical mode decomposition -- Reinforcement learning algorithm -- Long short-term memory networks -- Proportional-integral-derivative; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2021.118266
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: Disturbances and uncertainty of load and renewable energy are considered. A multi-area and multi-output smart generation control framework is built. A joint strategy "Divide and Conquer" is improved for frequency regulation. The controller breaks the boundary between classical and intelligent algorithms. Lower frequency deviation and generation cost are achieved simultaneously. Abstract: The large-scale application of renewable energy can promote the global goal of carbon neutrality. However, the stochastic nature of wind and solar energy aggravates the active power imbalance and increases the frequency deviation. These obstacles hinder the load frequency control with the traditional proportional-integral-derivative as the primary approach for automatic generation control. Inspired by the "Divide and Conquer" strategy, a mode-decomposition memory reinforcement network strategy is proposed to reduce the impact of random fluctuations and uncertainties on power systems. The proposed strategy combines the traditional methods and intelligent algorithms for smart generation control. The proposed strategy includes empirical mode decomposition, proportional-integral-derivative, long short-term memory networks, and reinforcement learning algorithms. Firstly, the historical data that has been decomposed by the empirical mode decomposition is utilized to train long short-term memory networks. Then, the trained long short-term memory networks decompose and reorganize the frequency deviation into the high-frequency and low-frequency signals in real-time. Finally, reinforcement learning and proportional-integral-derivative respectively optimize the generation commands by the high-frequency and low-frequency signals to mitigate frequency deviation. Two cases results prove that the mode-decomposition memory reinforcement network has a higher control effect and lower generation cost than the other four strategies. Significantly, the frequency deviation and generation cost are respectively reduced by at least 9.77% and 4.39% in the four-area power system.