Wind turbines do not operate in isolation. Every turbine in a commercial wind farm is embedded in the wake field of upstream rotors — a region of reduced wind speed, elevated turbulence intensity, and altered velocity profile that persists for 5 to 15 rotor diameters downstream depending on atmospheric stability and surface roughness. The energy lost to wake interactions at the farm level is not a rounding error: inter-turbine wakes reduce aggregate farm output by 8–15% compared to free-stream estimates across typical commercial layouts.
For a 200 MW wind installation, that means the production forecast could be structurally wrong by 16–30 MW under certain wind direction and atmospheric stability conditions — before accounting for any NWP input uncertainty. Ignoring wake sector effects in an energy forecast is not missing a refinement. It is accepting a known systematic bias.
The physics of wake deficit and turbulence intensity
When a turbine extracts kinetic energy from the freestream, it creates a velocity deficit in its downwind shadow. In the near wake (2–4 rotor diameters), the turbulence generated at the blade tips and hub creates complex mixing. In the far wake (6+ rotor diameters), the deficit profile broadens and the intensity decreases, but the velocity recovery is incomplete at typical turbine spacings of 7–10 rotor diameters in the prevailing wind direction.
Turbulence intensity in the wake matters beyond production loss. Elevated TI increases fatigue loading on downwind turbine blades and actuates more frequent pitch and yaw corrections, which temporarily reduces output below the static power curve value. This effect is intermittent and varies with atmospheric boundary layer stability — under stable nighttime conditions, wake recovery is slower and TI effects persist further downstream than in the convective mixing of a daytime unstable boundary layer.
Grid-scale NWP models running at 3 km resolution (NAM-3km) or coarser (GFS at 0.25°, ECMWF at 0.1°) cannot represent individual turbine wakes. Their output is effectively a freestream wind speed at the farm centroid. Any asset-level wind energy forecast that ingests NWP output without a wake correction model is implicitly assuming that all turbines in the farm operate at freestream conditions — a systematic overestimation that is worst in the prevailing wind direction sectors.
Sector-based wake correction at the asset level
The practical approach for operational wind forecasting is not full computational fluid dynamics (CFD) — CFD is appropriate for farm layout design, not real-time dispatch. The operational approach is a sector-based statistical correction: for each wind direction sector (typically 15–30° bins), apply a pre-computed wake loss factor derived from turbine layout geometry and validated against historical SCADA actuals.
The correction takes the form of a wind direction-conditional power output adjustment. For a given asset, you characterize the upstream turbine rows for each directional sector, estimate the aggregate wake loss at the array level using an analytical wake model (Jensen, Bastankhah-Porté-Agel, or equivalent) tuned to the site geometry, and then apply that correction as a multiplicative adjustment to the NWP-derived power curve output.
Consider a 150 MW wind farm on the high plains of eastern Colorado — a site we can characterize with reasonable accuracy from the Colorado Wind Energy Atlas. Prevailing winds at 250–270° (westerly) drive turbine rows aligned roughly perpendicular to the prevalent wind vector, meaning the dominant wake direction cuts across multiple turbine rows. In this sector, an uncorrected forecast may estimate 130 MW of output while an array-aware forecast, accounting for the three-row internal wake cascade, yields a production-matched 112 MW. On a day when the dispatch operator committed reserves against the 130 MW estimate, that 18 MW miss in a high-wind period forces RTM corrections.
Stability-conditional wake loss factors
A static wake loss factor per sector is a significant improvement over no wake model, but it still misses atmospheric stability effects that cause the same wind direction to produce different wake depths across days. Under neutral stability conditions (moderate wind, overcast sky, strong surface mixing), wakes recover quickly and TI-driven production loss is moderate. Under stable stratification (calm wind, clear sky, nocturnal boundary layer), wake recovery is suppressed — the Jensen wake length scale can extend 50% further than under neutral conditions for the same wind speed.
Adding a stability-conditional dimension to the wake correction — using NWP-derived Monin-Obukhov length, sensible heat flux, or proxy variables like downwelling longwave radiation and surface temperature difference — captures the night/day asymmetry in aggregate farm performance that a direction-only correction misses. In practice, this means maintaining separate wake loss lookup tables indexed by both direction sector and stability regime, with the stability category derived from NWP output at each forecast interval.
We're not suggesting that full LES (large eddy simulation) is needed for every production forecast. The marginal accuracy gain from LES over a well-tuned analytical wake model with stability conditioning does not justify the computational overhead in an operational 6-hourly update cycle. The goal is a fast, physically grounded correction that captures 80% of the systematic bias without requiring turbine-level CFD at each model run.
What SCADA actuals reveal about wake model performance
The best validation signal for a wake correction model is directional analysis of SCADA actuals against uncorrected power curve estimates. When you bin the historical forecast error by wind direction sector, wake-affected sectors show consistent positive bias in the uncorrected model — the model systematically over-predicts output. Sectors aligned with the prevailing wind show the largest positive bias. Off-angle sectors where no upstream turbines exist may show near-zero systematic bias from wake effects (though they may show other biases from terrain channeling or shear).
This directional error signature is diagnostic: if you see bias that is random across directions, it is not a wake model problem — it may be a power curve fit issue or a turbine availability signal mixed into the actuals. If the bias is strongly directional and correlated with the upstream turbine layout geometry, the wake correction model is the right place to intervene.
Running this analysis on 90 days of SCADA actuals per site provides enough directional sample coverage for a reasonably well-characterized wind rose. For sites with strong directional anisotropy — where 60% of energy comes from a 45° arc — you need longer history to get statistical confidence in the off-angle sectors, but those sectors contribute less to aggregate error anyway.
Impact on aggregate portfolio forecast accuracy
At the individual asset level, adding wake sector correction typically reduces the systematic positive bias in prevailing wind sectors by 60–80%. At the portfolio level — where a balancing authority or asset manager might be aggregating 10–20 individual wind installations — the wake correction matters differently: systematic per-site biases that all trend in the same direction on high-wind days can aggregate into a portfolio-level miss that is larger than any individual site error.
A 5 MW systematic over-forecast on each of 12 wind farms in the same prevailing wind corridor is a 60 MW portfolio miss on high-wind days — exactly when the reserve commitment based on that forecast is most consequential because thermal units are backed down. Wake-corrected per-site forecasts reduce that correlated bias, and the portfolio-level benefit exceeds the per-site improvement ratio.
The work is in the data collection upfront — turbine layout coordinates, rotor diameter, hub height, historical SCADA actuals with sufficient directional sampling — and in the ongoing retraining cycle that keeps the wake correction factors calibrated as turbines age, performance degrades unevenly across the array, and the actual wind resource distribution shifts relative to historical patterns.