A graph-based multi-agent approach for Wind Farm Flow Control using reinforcement learning

As wind turbines generate power they extract energy from the flow causing wakes, which are areas of slower-moving and more turbulent flow behind the turbines. These wakes can occlude downstream turbines leading to power reductions. Wind Farm Flow Control optimises the settings of all turbines within a farm to achieve farm-level goals such as maximising the total power generated. Wake steering is a popular Wind Farm Flow Control technique where upstream turbines are yawed to misalign their rotors with the incoming wind to deflect their wakes away from their downstream neighbours. Here, a novel method is developed that combines a Reinforcement Learning algorithm with a Graph Neural Network to give a turbine-level agent optimised for wake steering. After training on a nine-turbine farm under a range of wind directions, the agent implemented successful wake steering on this farm to gain 10% farm power over four wind directions. Using a Graph Neural Network meant the agent could be applied to different layouts, and gained 12% farm power over greedy control on a 46-turbine farm. An agent trained with this method against farm-specific agents from a contemporary work achieved comparable gains in farm power across three layouts with significantly reduced overall training time.