Construction of offline predictive controller for wind farm based on CNN–GRNN. (October 2022)

• Record Type: Journal Article
• Authors: Wang, Yu; Wei, Shanbi; Yang, Wei; Chai, Yi; Li, Penghua
• Is Part Of: Control engineering practice. Volume 127(2022)
• Publication Date: 2022-10
• Subjects: Offline predictive control -- Dynamic wake model -- Offshore wind farm -- CNN–GRNN; Automatic control -- Periodicals; 629.89
• DOI: 10.1016/j.conengprac.2022.105290
• Languages: English
• ISSNs: 0967-0661

Abstract

Abstract: In order to reduce Levelized Cost of Energy(LCOE), the wind farm development trend includes mainly two points: (1) rotor diameter of wind turbines will be larger, (2) density of offshore facilities will be higher. Consequently, the wake interference effect will be more severe, and predictive control has become one of the main studies for reducing wake disturbance. However, most of these strategies are usually computationally expensive and have limitations for large-scale wind farms since the optimization cost tends to increase exponentially with prediction horizons and number of wind turbines, which is unfavorable for real-time control. Considering control effectiveness and online computational efficiency, an offline predictive controller using convolutional neural network-general regression neural network (CNN–GRNN) is proposed for wind farm power control considering fatigue load. A CNN–GRNN hybrid network is trained to carry out the mapping between system states and optimal control settings calculated by particle swarm optimization (PSO) algorithm under various wind directions and wind speeds, and CNN–GRNN acts as an offline optimizer to provide ideal control laws with negligible online computation workload. The power at Horns Rev 1 can be improved up to 5.6% in comparative experiments with lookup table (LUT) strategy while keeping the fatigue load within the preset limitation, and the extensive simulation results under various wind conditions indicate that this proposed strategy can provide a more desirable performance for wind farm power control with fatigue constraint while retaining the advantage of negligible online computational burden. Highlights: Build a dynamic wake model involving delay time of wake and control settings. Power control problem considering the fatigue load is carried out. Replace optimization algorithm with a neural network in receding horizon. CNN–GRNN serves as an offline optimizer to provide desirable control laws.