Model Predictive Control for Wind Farm Power Tracking With Deep Learning-Based Reduced Order Modeling

Dynamic power control of wind farms (WFs) is necessary to provide automatic generation control (AGC) services for the power system. However, cooperative WF control for AGC remains a great challenge because of the nonlinear and high-dimensional nature of the wake flow dynamics. To address this challenge, this article proposes a model predictive control (MPC) framework with deep learning-based reduced-order modeling (ROM). Two novel neural network architectures are designed, which successfully formulate a WF ROM capturing the dominant wake steering dynamics in a computationally efficient manner. Compared to physical models, the data-driven ROM reduces the number of model states by orders of magnitude. Then, a novel WF AGC framework embedding the derived WF ROM is proposed. Thrust coefficient and yaw steering are both employed to optimize WF power tracking performance. Compared to prior WF AGC controllers, the dynamic yaw actuation is first optimized for AGC considering the wake steering dynamics. Case studies validate the effectiveness of the deep learning-based WF ROM at capturing the wake traveling dynamics. The WF controllers were stress-tested under time-varying inflow directions. The proposed MPC can react to different wind directions and generates higher-quality control performance than existing alternatives with extended trackable AGC range and better dynamic power tracking performance.