Time-Accurate Wind Farm Control in Dynamic Flow Conditions using Deep Reinforcement Learning

Wake steering is a form of Wind Farm Control (WFC) in which upstream turbines are yawed in order to redirect their wakes away from neighbouring downstream turbines. The reduction in power of the upstream turbines can be more than compensated for by the increase in power from the downstream turbines, as they see a faster and less turbulent inflow; over a wind farm, this can give a net increase in farm power. Calculating the required yaw set-points for WFC over a range of wind conditions and for multiple turbines can be computationally intensive for traditional optimisation methods. Reinforcement Learning (RL) is a form of machine learning where agents learn beneficial behaviour patterns through trial-and-error to achieve long-term goals. It has become a topic of interest for WFC, primarily through using time-averaged (steady-state) wind farm models to train the RL agents. In this work, RL agents are trained using a dynamic flow solver, with incoming wind from a wide range of directions. These RL agents are able to yaw the turbines towards set-points that increase the farm power, despite initially losing power while travelling. Using a realistic wind time history, their performance was compared to similar RL agents trained with a steady-state model. Both training approaches gave agents that can successfully control the wind farm to gain 2-3% additional power. However, the results suggest that it is not necessary to expend the additional wall-clock or computational time to train RL agents using dynamic flow over steady-state.