Load constrained wind farm flow control through multi-objective multi-agent reinforcement learning

A team led by Teodor Åstrand and Marcus Binder Nilsen at the Technical University of Denmark (DTU) has introduced a multi-agent reinforcement learning (MARL) framework for load-constrained wind farm flow control, published on arXiv (2604.22795) and accepted for the Torque 2026 conference. The framework addresses a key challenge in wake steering: while adjusting turbine yaw can increase total farm power, it often increases structural loads on downstream turbines, risking fatigue damage. The researchers integrated an Independent Soft Actor-Critic (I-SAC) architecture with a data-driven surrogate model that provides real-time estimates of Damage Equivalent Loads (DELs) based on local inflow sector averages. This allows each turbine agent to optimize power production while respecting specific load increase thresholds (10%, 20%, or 30%) relative to a baseline controller. The system was implemented in the WindGym environment using the DYNAMIKS flow solver with a Dynamic Wake Meandering (DWM) model to capture non-stationary wake physics.

Results demonstrate that the MARL agents successfully learn collaborative policies that prioritize power gain while actively retreating from high-DEL control strategies. By incorporating load estimates into a shaped reward function, the agents can balance energy output against structural integrity without requiring centralized coordination. This approach offers a practical path for wind farm operators to increase energy yield without exceeding turbine design limits, potentially extending equipment lifespan while boosting revenue. The framework's ability to handle dynamic wake conditions and provide real-time load feedback makes it a promising tool for next-generation wind farm control systems, though real-world validation and computational efficiency remain areas for further development.