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

Area of Science:

• Renewable Energy Systems
• Control Engineering
• Artificial Intelligence

Background:

• Traditional wind turbine pitch controllers face limitations in adapting to nonlinear wind dynamics and mitigating fluctuating loads.
• Proportional-integral (PI) controllers exhibit challenges in maintaining optimal performance under variable wind conditions.

Purpose of the Study:

• To develop and evaluate a novel reinforcement learning (RL)-based framework for enhanced wind turbine pitch control.
• To improve power regulation efficiency and structural load mitigation in wind turbines operating under uncertain wind conditions.

Main Methods:

• A two-stage RL approach involving policy transfer and refinement using the double deep Q-Network (DDQN) algorithm.
• Development of a novel reward function tailored for practical operational scenarios, refining an initial PI controller-based policy.

Main Results:

• The proposed RL controller demonstrated superior performance compared to the industry-standard ROSCO controller.
• Achieved a 15.87% reduction in power fluctuations.
• Maintained comparable or slightly reduced structural load levels, ensuring operational safety.

Conclusions:

• Reinforcement learning offers a powerful approach to enhance wind turbine power regulation.
• The developed RL framework effectively improves control efficiency and maintains structural integrity without compromising safety.