Researchers at the University of Warwick and the Technical University of Munich have used artificial intelligence to coordinate three scale-model wind turbines as wind direction changed. Their offline reinforcement-learning controller increased the turbines' combined output by about 10% over a standard control strategy in a wind tunnel, according to a preprint submitted on September 11.
A turbine extracts energy from the wind and leaves slower, more turbulent air behind it. This wake can reduce the output of turbines farther downwind. Most turbines point directly into the incoming wind to maximize their own production. Wake steering turns an upstream turbine slightly away from the wind so that its wake misses machines behind it, trading some output at the first turbine for a larger gain across the farm.
The new controller learned from an 80-gigabyte archive of wind-tunnel measurements collected under several earlier control policies. This is the “offline” part of offline reinforcement learning: the system trains on fixed records without testing exploratory actions on operating equipment. Its inputs included the measured wind direction and the yaw angle of each turbine. Its output was a set of new yaw commands. The researchers also penalized abrupt changes between similar commands because turbine yaw mechanisms respond slowly.
The test used three turbines with 1.1-metre rotors on a 13-metre turntable at the Politecnico di Milano wind tunnel. The machines were spaced five rotor diameters apart. A five-day wind-direction record from northern Germany was compressed into nine profiles lasting a total of 90 minutes, with direction varying by 15 degrees on either side of the turbine row. Wind speed remained at 5.35 metres per second and turbulence intensity at 6%.
Across the full test, three versions of the offline controller each produced about 10% more combined power than the baseline that kept every turbine facing the wind. The strongest version matched a lookup-table controller built from a calibrated wake model and performed about as well as an online reinforcement-learning controller. It used roughly 5% of the online controller's training cost. The first turbine sometimes generated less power, while the second and third gained enough to raise total output.
The result comes from three small turbines under a fixed wind speed and turbulence level. It does not measure annual production, structural loads or performance at a commercial wind farm. The authors plan full-scale testing. For perspective, the US National Renewable Energy Laboratory estimates that wake steering can raise annual energy production at existing wind facilities by about 1% to 2%, so the 10% wind-tunnel gain should be read as a controlled test result.
Sources
- Yuhan Su and colleagues, arXiv, September 11, 2026. Wind-tunnel setup, offline reinforcement-learning method, training data, comparisons and limitations.
- US National Renewable Energy Laboratory. Wake-steering explanation and estimated annual energy gain at existing facilities.
- Lead image: Graham Robson, Barmoor Wind Farm, used without alteration under the CC BY-SA 2.0 licence. The photograph is illustrative and does not show the experiment.
