Implementation and benchmarking of model based and model free controllers for wind turbine

Wind turbines are sophisticated machines designed to harness the raw power of the wind and convert it into electrical energy. Developing a controller for wind turbines is a challenging task due to their complex nature. Their operation is indeed influenced by turbulent external winds and characterized by nonlinear dynamics dictated by aerodynamics. The state of the art in wind turbine control employs classical controllers, but recently the community started to explore AI solutions to improve their performances. This thesis focuses on the development and benchmarking of such control techniques. It articulates on both numerical and experimental phases. On the numerical side, a custom-built Python environment is developed to simulate the behavior of different controllers. The model utilizes a simplified, nonlinear first-order representation of wind turbines. The project begins by establishing baseline control techniques, including the widely-used Proportional Integral (PI) controller and the K-omega squared technique. These techniques are validated against the existing controllers, specifically the ROSCO package developed by NREL, demonstrating a high level of agreement. Model-based methodologies such as Model Predictive Control (MPC) and data-driven approaches like Genetic Programming (GP), Bayesian Optimization (BO), and Reinforcement Learning (RL) are explored. Comparisons are conducted to evaluate their performance. The experimental campaign is conducted in the Low Speed Wind Tunnel L-2B at the von Karman Institute. The wind tunnel features a test section area of 0.35 cm x 0.35 cm, with a maximum incoming flow velocity of 35 m/s. Efforts are dedicated to implementing the reading of rotational speed from an encoder with an Arduino master-slave architecture, which interacts with a Python code responsible for wind turbine control. Experimental testing involves the PI controller and the optimization of the PI controller’s weights using Bayesian Optimization (BO). Lastly, a specific experimental setup is designed to simulate real-world conditions in wind turbine farms, positioning one wind turbine downstream of another. This configuration enables the validation of the controller’s performance in mitigating wake interference and optimizing overall wind farm performance. In conclusion, based on the promising outcomes achieved with data-driven controllers in the numerical simulations, further experimental testing of these controllers would be of great interest for future research.