Multi-Objective Wind Farm Controller Optimization with Combined Thrust and Yaw Control using Koopman Model Predictive Control

Kurzfassung

A novel approach to wind farm control, leveraging the Koopman framework for non-linear system identification, is shown to be an efficient way to derive a pareto-optimal farm control algorithm. The objective in both wind turbine and farm control is to minimize the levelized cost of energy (LCoE) via control settings that optimize a combination of power, mechanical loads and actuator criteria subject to actuator constraints, [1,2]. Power criteria are the minimization of the difference to grid power refences and the minimization of the power standard deviation. Load criteria ensure the minimization of mechanical damage equivalent loads on tower and blades. Including actuator criteria, i.e. penalizing actuator outputs and rates, ensures that the wear on the actuators is minimized. There are three actuators available per turbine: With the turbine yaw, the angle between turbine rotor and wind direction is set, typically setting the turbine rotor perpendicular to the main wind direction. For thrust control, the generator torque is used to maximize the power yield at below rated wind and the blade pitch for keeping the power constant at above rated wind and to ensure structural integrity. Two approaches to collaborative farm control are axial induction control (AIC) and wake redirection control (WRC): In AIC, the generator torque and blade pitch angles are varied from individual optimal setting to change the axial induction or thrust coefficient. In WRC, the wake of upstream wind turbines is redirected away from downstream wind turbines, often via yawing the upstream turbine. Both approaches have shown to yield promising results. However, there are relatively few investigations into combining thrust and yaw control for optimal farm yield. The main contributions of the presented work are: The WFSim environment [3] is leveraged to demonstrate the power gain possible with WRC. Physically motivated Koopman lifting functions and dynamic mode decomposition are then shown to be well suited for modelling the dynamics between the two control inputs, yaw and thrust, and the overall farm output. This is an extension of the Koopman models for AIC presented in [4]. The resulting linear model is then shown to provide a good basis for model predictive control using combined AIC and WRC with the goal to minimize the difference between grid reference power while minimizing actuator energy as well as keeping the yaw angle as small as possible, as larger yaw misalignment, while resulting in overall power increase, also increase the mechanical loads on the turbine.