A reinforcement-learning approach for individual pitch control

Individual pitch control has shown great capability of alleviating the oscillating loads experienced by wind turbine blades due to wind shear, atmospheric turbulence, yaw misalignment, or wake impingement. This work presents a novel controller structure that relies on the separation of low-level control tasks and high-level ones. It is based on a neural network that modulates basic periodic pitch angle signals. This neural network is trained with reinforcement learning, a trial and error way of acquiring skills, in a low-fidelity environment exempt from turbulence. The trained controller is further deployed in large eddy simulations to assess its performances in turbulent and waked flows. Results show that the method enables the neural network to learn how to reduce fatigue loads and to exploit that knowledge to complex turbulent flows. When compared to a state-of-the-art individual pitch controller, the one introduced here presents similar load alleviation capacities at reasonable turbulence intensity levels, while displaying very smooth pitching commands by nature.