Wind power plays a vital role in the global effort towards net zero. Control systems are at the core of wind turbine and farm operations and have essential influences on the power capture efficiency and maintenance cost. Substantial efforts have been made to investigate the control methods of wind turbines and farms. However, there still exists a research gap due to the complex and nonlinear systems of modern wind turbines and farms. Particularly, reinforcement learning (RL) is a cutting-edge machine learning technique that can handle high-complexity and nonlinear control problems. Therefore, this thesis aims to investigate RL-based wind turbine and farm control technologies to increase power generation, decrease loads, and improve control performance in the presence of uncertainties and faults.

Firstly, at the wind turbine level, the wind turbine torque and pitch control are studied. A novel RL controller combining deep neural network (DNN) and model predictive control (MPC) is designed to achieve power generation maximization and maintenance. Secondly, a novel RL-based control scheme, incorporating individual pitch control (IPC) and collective pitch control (CPC), is proposed for floating offshore wind turbines (FOWTs) to balance load mitigation and power regulation. Thirdly, considering wind turbine actuator and sensor faults, an RL-based faulttolerant control (FTC) strategy – incremental model-based heuristic dynamic programming (IHDP) is developed subject to unknown system models. Finally, at the wind farm level, a model-free deep RL (DRL) method is investigated for maximizing the total power generation of the whole wind farm. A novel double-network-based deep deterministic policy gradient (DN-DDPG) approach is designed to generate control policies for thrust coefficients and yaw angles simultaneously and separately. Numerical simulations based on a high-fidelity wind turbine simulator – FAST (Fatigue, Aerodynamics, Structures, and Turbulence) and a dynamic wind farm simulator are conducted to demonstrate the effectiveness of these RL-based control methods.