Reinforcement learning increases wind farm power production by enabling closed-loop collaborative control

Area of Science:

• Fluid dynamics
• Renewable energy systems
• Artificial intelligence

Background:

• Traditional wind farm control optimizes turbines individually, limiting overall energy production.
• Coordinated wake steering and dynamic control show potential but are hindered by low-fidelity simulations.
• High-fidelity simulations are needed to capture turbulent fluctuations for effective dynamic control.

Purpose of the Study:

• To develop and evaluate a reinforcement learning controller for dynamic, coordinated wind farm control.
• To leverage high-fidelity turbulence-resolving simulations for training the controller.
• To demonstrate improved wind farm power maximization through real-time response to atmospheric conditions.

Main Methods:

• Utilized high-fidelity, turbulence-resolving simulations for controller training.
• Implemented a reinforcement learning approach for dynamic, closed-loop control.
• Compared the RL controller against static optimal yaw and Bayesian optimization methods in a three-turbine test case.

Main Results:

• The reinforcement learning controller achieved a 4.30% increase in wind farm power output.
• This performance nearly doubled the gain from static optimal yaw control (2.19%).
• It also surpassed global wind direction-based dynamic control (2.67%) in a three-turbine test case.

Conclusions:

• Reinforcement learning effectively utilizes high-fidelity simulation data for dynamic, flow-responsive wind farm control.
• This approach significantly enhances wind farm power output compared to existing methods.
• The findings support accelerated renewable energy deployment towards net-zero targets.