Initial results from a field campaign of wake steering applied at a commercial wind farm – Part 1

A wind turbine is deliberately pointed slightly away from the wind. This is not due to a malfunction, nor is it because the grid is asking for less power. An engineer has made this choice. The turbine is voluntarily sacrificing its own output, and the question is whether that sacrifice pays off for the farm as a whole. This is the entire premise of wake steering, and a team from the National Renewable Energy Laboratory and NextEra Energy spent a summer at a real commercial wind farm to find out whether it actually works. The problem they were solving is a structural issue in wind energy. Wind farms are not simply collections of isolated machines; turbines interact aerodynamically, and those interactions cost energy. When one turbine is positioned upstream of another, it leaves behind a turbulent, low-velocity wake. The downstream turbine enters this disturbed air, experiences weaker inflow, and produces less power. In a closely spaced array, this occurs constantly, and the cumulative losses are significant. Wake steering proposes a solution: yaw the upstream turbine off the wind so its wake deflects sideways, away from the downstream machine. The downstream turbine then encounters cleaner air. The upstream turbine pays a small penalty for pointing off-axis, but the idea is that the downstream gain outweighs that cost. Simulations and wind tunnel experiments had indicated this could work. Fleming and colleagues wanted to know if it held up in the field. They selected a cluster of five turbines at a commercial wind farm: three central machines labeled T2, T3, and T4, with T1 and T5 on the flanks acting as uncontrolled references. T3 is positioned in the near wake of T4, which made this pair the focus. The Phase 1 campaign ran from May 4 through July 11 of 2018, during which summer winds predominantly came from the south-southeast, so the team concentrated on that approach direction. To measure atmospheric conditions, they deployed a Leosphere Windcube version two profiling lidar and two Vaisala Triton sodar wind profilers, with range gates every 20 meters from 40 up to 200 meters of altitude. The southern sodar became the primary inflow sensor for the experiment. Conducting this at an operating commercial farm introduced practical friction immediately. The southern sodar only provided data once every ten minutes, while turbine telemetry was received at one hertz. The team downsampled turbine data to one-minute bins and filled in the gaps using a zero-order hold, which means holding the last known value forward until the next measurement arrived. The southern terrain was also complex, with escarpments extending south of T4, influencing both atmospheric behavior and wake evolution. These were not controlled laboratory conditions. The controller itself was built on FLORIS, the FLOw Redirection and Induction in Steady state model. This computational wake model was used by the team to calculate the optimal yaw offset for T4 under different wind speed and direction combinations. Those optimized offsets were incorporated into a static lookup table. At runtime, the controller read filtered wind speed and direction from the southern sodar, pulled the appropriate offset from the table, filtered that offset again through a 30-second low-pass filter, and injected the result into T4's yaw system by modifying the nacelle vane signal. Offsets were capped at 20 degrees for wind speeds of 12 meters per second or less, based on load safety constraints from earlier work by Damiani and colleagues. To isolate the controller's effect from natural wind variability, the team ran a toggling protocol: one hour of wake steering on and one hour off, alternating throughout the three-month campaign. This cadence provided comparable amounts of data in each state and ensured the two datasets encountered similar distributions of wind speeds and atmospheric conditions, making the comparison clearer. To analyze the data, Fleming and colleagues constructed an energy ratio metric. They computed a synthetic reference power based on the sodar's rotor equivalent wind speed, which is a weighted average of sodar measurements across the rotor swept area, and then used the FLORIS power coefficient lookup to convert that to a reference power. They sorted every one-minute sample into two-degree wind direction bins, weighted each sample to correct for any mismatch in wind speed composition between the on and off periods, and computed the ratio of actual turbine power to the reference. Bootstrapping, with at least a thousand resampling iterations, provided ninety-five percent confidence intervals on each bin. The headline numbers are clear. Over the ten-degree sector from 140 to 150 degrees, the downstream turbine T3 gained roughly 14.6 percent in energy when T4 was steered. That is a meaningful increase. However, T4, by pointing off the wind, gave some of that gain back. When combining T3's gain with T4's loss and looking at the pair together, the net increase is about 4.1 percent over that same sector. These field measurements aligned well with FLORIS predictions—in fact, the observed net gain for the pair was somewhat larger than the model's estimate. The four percent number deserves some attention. It might sound modest, but it represents a real, measured net gain at a commercial-scale installation. Furthermore, T3 was deep in T4's wake—the energy ratio at its lowest dropped below 40 percent of the reference, meaning T3 was producing less than half of what it would in clean air. Steering significantly improved that situation. However, the complications were real and instructive. Atmospheric stability turned out to be a significant factor. Fleming and colleagues classified conditions using the Obukhov length, which measures how thermally stratified the atmosphere is, following the Wharton and Lundquist classification system. Stable conditions, characterized by low turbulence, are the most favorable for wake steering because stable atmospheres preserve wakes more coherently, making deflection more effective. Unfortunately, only about one-third of the Phase 1 data fell into the stable category. The summer campaign biased toward higher turbulence, unstable conditions, which worked against them. Lower wind speeds tended to be more unstable, while higher wind speeds were more stable. As a result, the observed gains are a blend of conditions, with the cleaner, more favorable stable regime data being the minority. Controller performance posed another challenge. The combined filtering in the system, which included thirty-second low-pass filters on both the inflow signal and the offset command, in addition to the inherent dynamics of the turbine's yaw controller, introduced substantial lag. Fleming and colleagues provide a concrete example: a wind direction shift that should trigger a 20-degree offset at minute two doesn't result in the turbine reaching that offset until around minute five. This causes three minutes of suboptimal steering every time the wind changes direction significantly. Additionally, there was a general tendency toward undershoot, with the turbine consistently falling short of its commanded offset. The FLORIS model used for controller design also failed to incorporate terrain modeling and a well-tuned near-wake model, despite T3 being squarely in T4's near wake and the southern terrain being complex. These gaps contributed to uncertainty in the quantitative comparison. Fleming and colleagues emphasize that these are lessons learned, not just failures. The Phase 1 trial produced a clear positive outcome—wake steering worked, the downstream gain was real, and it matched the model's general prediction—while also generating a specific diagnostic list of what needs improvement. The north campaign planned for Phase 2 targets a simpler site geometry, aims to use higher-frequency inflow sensing, including lidar, and calls for a more dynamic controller design that can respond faster and reduce the lag and undershoot that limited Phase 1's performance. Improved near-wake and terrain modeling are also on the agenda. What Phase 1 established is this: deliberately misaligning an upstream turbine at a commercial wind farm, using a model-based lookup table and a relatively simple control architecture, generated a measurable 14.6 percent energy increase at the downstream turbine and a net 4.1 percent gain for the pair. Both numbers were consistent with FLORIS predictions. This represents the first rigorous field validation of the concept at commercial scale. The counterintuitive act of pointing a turbine away from the wind, it turns out, yields dividends. The question moving forward is how much more can be extracted when the controller is faster, the model is sharper, and the conditions are better understood. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.