Wind farm optimization using RL and large-scale simulation

In collaboration with Microsoft and Vestas, we developed wind farm controllers that boost annual energy production by 1-2%. The controller achieves this by adjusting the yaw of upstream wind turbines in the farm to minimize wake losses at downstream turbines. The training of such deep reinforcement learning-based controllers for wind farms required large-scale and time-consuming computational fluid dynamics simulations. During the training of the controllers we used up to 15,000 CPU cores in parallel, using our platform DeepSim. DeepSim is minds.ai's end-to-end machine learning platform that uses Ray for distributed computing, model serving, and data processing. In this talk, we will discuss the project and how the controllers were efficiently trained in the cloud.