Conceptual Design of a Reinforcement Learning Agent for Active Wake Control in Wind Farms

Active wake control for wind farms offers the potential to increase the total power production and reduce mechanical stress on turbine components by manipulating the propagation of wake formations through the wind farm. However, the determination of suitable set points for the wake control methods is challenging. Varying wind conditions and a large number of turbines mean that the impact of a large number of possible set point combinations must be continuously evaluated to ensure optimal performance. Reinforcement learning for active wake control has received increasing scientific attention in recent years because, as a type of machine learning model, it learns from its interaction with the wind farm, e.g. in an atmospheric simulation model, and generalizes an optimal policy that maps the current wind conditions to wake control set points. No explicit model building of the complex wake interactions is necessary. In this paper, a concept is presented that uses a function approximation agent to determine optimal set points for wake redirection considering the constraints of modern commercial wind farms. The relevant properties of wind farms are assessed and the agent structure and learning process is crafted to match the available measurement data as well as the wind field dynamics that complicate the evaluation of beneficial control actions in large wind farms. In the future, the resulting concept will be implemented and tested using the Smart Windpark Laboratory and a dedicated wake model.