FALCON- FArm Level CONtrol for wind turbines using multi-agent deep reinforcement learning

Turbines in a wind farm dynamically influence each other through wakes. Therefore trade-offs exist between energy output of upstream turbines and the health of downstream turbines. Using both model-based predictive control (MPC) and machine learning techniques, existing works have explored the energy-fatigue trade-off either in a single turbine or only with few turbines due to issues of scalability and complexity. To address this gap, this paper proposes a multi-agent deep reinforcement learning-based coordinated control for wind farms, called FALCON. FALCON addresses the multi-objective optimization problem of maximizing energy while minimizing fatigue damage by jointly controlling pitch and yaw of all turbines. FALCON achieves scale by using multiple reinforcement learning agents; capturing the global state-space efficiently using an auto-encoder; and pruning the action-space using domain knowledge. FALCON is evaluated through a real-world wind-farm case study with 21 turbines; and performs better than the default baseline PID controller and a learning-based distributed control.