Reinforcement learning-based pitch control for wind turbines using double deep Q-networks

This paper presents a novel reinforcement learning-based framework for pitch control in wind turbines, addressing the challenges posed by high nonlinearity and uncertainties in wind dynamics. Traditional pitch control methods, such as proportional-integral controllers, often struggle with adaptability and load mitigation under fluctuating wind conditions. To enhance control efficiency, the proposed framework employs a two-stage approach: policy transfer and policy refinement. In the policy transfer stage, a proportional-integral controller is used to estimate an initial control policy. This policy is then refined using the double deep Q-Network algorithm. Furthermore, a novel reward function is introduced as an improved version of a recent formulation in the literature, ensuring better alignment with practical operational scenarios. Simulations demonstrate that the proposed controller outperforms the industry-standard ROSCO controller, achieving a 15.87% reduction in power fluctuations, while preserving comparable structural load levels, with slight reductions in some cases. These results highlight the potential of reinforcement learning-based control to enhance wind turbine power regulation without compromising structural safety in terms of loads and vibrations. • A novel two-stage RL-based methodology: policy transfer and policy refinement. • OpenFAST is used to train the pitch controller under a realistic WT environment. • A reward function is proposed to account for a variety of operational scenarios. • The RL-based controller outperforms the ROSCO across various wind profiles.