the turbines behind get cheated

Problem statement

In large wind plants, upstream turbines extract momentum from the wind and create wakes — regions of reduced velocity and increased turbulence — that propagate downstream for 5–15 rotor diameters. Deep within a wind plant, turbines operating in the accumulated wakes of multiple upstream rows can experience 10–20% lower power production and significantly higher fatigue loads compared to front-row turbines. No existing model accurately predicts the combined effects of turbine-generated wakes interacting with atmospheric boundary layer dynamics (stability, shear, veer) across an entire wind plant, especially during the unstable atmospheric conditions most relevant to energy production.

Why this matters

Wind energy currently supplies ~10% of global electricity and is projected to reach 30–50% by 2050. As wind plants grow larger (offshore plants now exceed 1 GW and span hundreds of km²), wake-related losses compound nonlinearly. A 1% improvement in wake-loss prediction accuracy across the global fleet would be worth approximately $1–2 billion annually in avoided underperformance. More importantly, unreliable energy production forecasts increase balancing costs for grid operators and reduce investor confidence in wind plant financial projections. Wake effects also extend beyond individual plants: plant-to-plant wakes from large offshore clusters in the North Sea have been measured at 50+ km downstream.

What’s been tried and why it hasn’t worked

Engineering wake models (Jensen, Bastankhah-Porté-Agel) use simplified momentum-deficit profiles and can compute quickly but systematically underpredict wake losses in large arrays and cannot capture wake meandering (lateral oscillation). Mesoscale weather models (WRF) capture atmospheric dynamics but treat wind plants as parameterized drag sources without resolving individual turbine interactions. Large Eddy Simulation (LES) can resolve both wake dynamics and atmospheric turbulence but is computationally prohibitive for design optimization — a single wind plant simulation at relevant resolution requires ~10⁶ CPU-hours. Field measurement campaigns (SCADA data from operating plants) provide ground truth but are site-specific, commercially sensitive, and cannot separate wake effects from atmospheric variability without controlled experiments that are impossible at full scale.

What would unlock progress

Convergence of three currently separate research communities is needed: atmospheric science (boundary layer dynamics and mesoscale modeling), wind engineering (turbine aerodynamics and control), and data science (machine learning from SCADA data). Specifically, ML-augmented wake models that learn correction factors from high-fidelity LES and field data could achieve engineering-model speed with LES-level accuracy. Open-access SCADA datasets from large wind plants (currently restricted by commercial sensitivity) would enable community-wide model validation.

Entry points for student teams

A student team could implement and compare 3–4 engineering wake models (Jensen, Bastankhah, Gaussian-curl hybrid, Floris framework) against published SCADA data from open-access wind plants (e.g., Horns Rev, Lillgrund) and quantify prediction errors as a function of atmospheric stability class. Alternatively, teams could train ML surrogate models on LES wake simulation data and test whether these surrogates generalize across wind plant configurations. The NREL FLORIS framework (open source) provides a ready-made computational environment. Relevant disciplines: fluid mechanics, atmospheric science, data science, renewable energy engineering.