Physics checks exposed overoptimistic wind-farm settings.
A turbine slows the wind behind it. Turning it slightly can leave more wind for its neighbours. Simulation lets us test when that tradeoff helps the whole farm.
Simulate a wind farm under changing wind and turbine settings. Train a model to predict power, use it to choose how the turbines should turn, then check those choices in the physics solver.
Edited demonstration briefThe result
Model-selected settings helped in some winds and hurt in others.
The model-selected settings increased total power in some simulated wind conditions, but reduced it in others. Re-running them in the physics solver exposed where the model overstated the benefit.
Steady wind-farm simulations. Benefits vary with the conditions; no annual energy gain or operating-farm control has been established.
Recorded simulation
Solver checks reveal the power delivered by the proposed settings.
Choose a recorded wind condition. Compare the model's prediction with the power computed by the physics solver.
Recorded outputs, replayed in your browser. No new simulation is running.
The best setting for a turbine can cost the farm.
The world we created
OpenFOAM models airflow through a virtual wind farm. Changing the wind and the direction each turbine faces changes the power available across the farm. Each case represents a steady wind condition.
What this enables
A learned model predicts turbine power from the conditions and settings. An optimizer uses those predictions to choose how the turbines should turn. The selected settings are then checked in fresh physics simulations.
Wind and turbine angles changed power across the farm.
Vary wind conditions and the direction the turbines face. Record the power produced by each turbine and the farm.
An optimizer chose settings using learned power predictions.
Train a model on the simulated cases. An optimizer searches its predictions for settings that could improve farm output.
Fresh solver runs exposed both gains and losses.
Compare the selected settings with turbines facing the wind under the same held-out conditions. Keep both the gains and the losses.
The simulation supplies training examples and a reference for checking the proposed settings. Those checks reveal where an accurate predictor still misleads an optimizer.
How the model was tested
Physical model. OpenFOAM simpleFoam solves steady incompressible RANS with a k-omega SST turbulence model and actuator disks. Turbine yaw changes the disk direction. A neural-network ensemble predicts power; a numerical optimizer selects yaw settings. FLORIS provides a separate engineering-model comparison.
Training experience. Structured parameter sweeps and Latin-hypercube samples vary inflow speed, turbulence, direction and turbine yaw. Separate cases are reserved for prediction tests and solver checks of the selected settings. Recorded outputs include turbine power, thrust and yaw.
Evaluation. The surrogate is evaluated on held-out CFD cases. Its optimized settings are compared with wind-facing baselines in fresh OpenFOAM runs. Predicted benefits are optimistic, and some selected settings reduce total power. Published farm measurements provide a limited reference comparison with mismatched turbines and layout.
- The simulation is steady RANS, with no resolved time-varying vortices. Playback of different cases is not a transient flow simulation.
- Every sweep case reached a fixed iteration budget. The declared residual stopping criterion did not trigger; a completed run is not proof that this criterion was met.
- The CFD campaign covers a limited wind-direction sector and does not model turbulence building up inside the farm. Its control checks do not establish annual energy gains.
- Agreement with digitised measurements from a different farm is limited by layout and turbine mismatches. There was no deployment of the controller at an operating wind farm.
Original research artifacts
The original files retain the recorded measurements and model details.