Wind turbines: by directing wakes and noise, a reinforcement learning agent meets the 40 dB limit by losing only 2% production

Noise Steering controls both the orientation and the power of each turbine; where conventional bridging costs 17% production, it loses only 2% on average.

To meet noise thresholds near homes, wind farms typically reduce the power of their turbines, regardless of the wake interactions or direction in which the sound is emitted. Researchers propose Noise Steering, a reinforcement learning strategy that jointly adjusts the yaw angle and power control of each wind turbine.

The agent, trained at a lower cost by pairing a simplified acoustic model with the FLORIS simulator, learns to divert wakes and noise from the protected area while redistribute production between the machines. On a test fleet of six turbines, it holds the stress of 40 dB over the entire range of turbulent winds, with an average deficit of 2% compared to a stress-free operation (up to 7.7% per strong wind). The industrial reference method, reduced mode operation, loses an average of 17%.

Transferred to the more faithful FAST.Farm simulator, politics retains most of its performance, suggesting a better compromise between renewable energy production and local acceptability.