Field test of wake steering at an offshore wind farm

It's dawn on a wind farm off the coast of China. One turbine — let's call it C1 — is spinning steadily, feeding four megawatts into the grid. Fifty meters behind it, another turbine sits in its wake: slower air, more turbulence, and less power. The downstream machine is being robbed by the one in front of it, and there's nothing in its controller that can do anything about that. Now, here's the bet that Fleming and colleagues from the National Renewable Energy Laboratory and Envision Energy decided to make: what if you deliberately pointed C1 slightly off the wind — not to maximize its own output, but to deflect that slow-air shadow away from the turbines downstream? That is wake steering. This paper describes the first time it was tested on a real, operating, commercial offshore wind farm. Wind turbines in a farm are not independent generators. Every spinning rotor sheds a turbulent, slower-moving wake that travels downwind and degrades the output of whatever turbine is in its path. At the Longyuan Rudong Chaojiandai offshore wind farm in Jiangsu, China, Fleming and colleagues selected a subset of 25 Envision EN136 four-megawatt turbines. Turbine C1 was the controlled, upstream machine. Behind it sat three downstream targets: D1 at seven rotor diameters away, D2 at eight and a half, and D3 at fourteen and a third. A fourth turbine, R1, sat outside the wake geometry entirely and served as a reference baseline. The three pairings exercise three different spacings — and the physics of wake recovery means each spacing tells a different story. Prior modeling by Gebraad and colleagues using FLORIS — FLOw Redirection and Induction in Steady State — had predicted a thirteen percent total power increase for a six-turbine case with wake steering. Fleming's own earlier work suggested a four-and-a-half percent net improvement for a two-turbine setup. Those are model results. The Rudong campaign was designed to find out whether the real world agrees. The strategy for the field test was built on two models working in tandem. SOWFA — the Simulator for Wind Farm Applications — is a high-fidelity computational fluid dynamics tool that uses an actuator-line coupling to simulate how a rotor interacts with the full atmospheric boundary layer. FLORIS, on the other hand, is a steady-state engineering model built for speed, not fidelity — the kind of tool you can actually run in a control loop. The approach was to use SOWFA to generate detailed wake data, then tune FLORIS to match those results, and finally use FLORIS to produce the actual yaw offsets deployed in the field. Tuning FLORIS required adjusting four parameters: the yaw-power exponent — let's call it pP — which governs how quickly an upstream turbine's own power drops as it yaws away from the wind; the wake expansion rate; the wake deflection recovery rate; and an initial wake deflection term. The optimized values for the Envision turbine came out as pP equal to 1.43, notably lower than the 1.88 used for the NREL five-megawatt reference turbine and the roughly 2.0 reported from wind-tunnel experiments by Medici. A lower pP means the upstream turbine loses less power per degree of yaw misalignment — which matters enormously for whether the steering trade-off is worth making at all. That tuned FLORIS then produced a lookup table: for each wind direction, a prescribed yaw offset for C1. Envision engineers modified C1's controller to implement those offsets in real time. Practical constraints applied — offsets were limited to positive, counterclockwise yaw only, capped at 25 degrees, and disabled entirely in sustained winds above ten meters per second for safety. The controller was open loop: it issued offsets based on wind direction but never observed whether the downstream wake was actually moving. The field campaign ran in two sequential phases, each lasting about four months. From April to August 2016, the farm ran normally — for baseline data collection. From August to December 2016, the offset controller was active on C1. The supervisory control and data acquisition system was the primary measurement source. To isolate the effect of steering from day-to-day wind variability, the team used a per-day power curve fitting procedure, fitting a scalar gain to a nominal power curve for each turbine each day, and then summarizing the distributions across the campaign using the twenty-fifth, fiftieth, and seventy-fifth percentiles. R1, the unaffected reference turbine, anchored the comparison. Fleming and colleagues are candid about the design's limitations. Because the two phases ran sequentially rather than alternating, seasonal and atmospheric shifts could blur the comparison. They estimate that roughly 30 to 40 separate days of testing per controller condition would be a reasonable target to resolve differences reliably — and they had fewer than that. The implemented offsets also varied widely from moment to moment — the average offset tracked the lookup table, but instantaneous values scattered — which means the real controller operated somewhere between baseline and the theoretical optimum. Now, what did the data actually show? The clearest result came from the closest pairing: C1 and D1 at seven rotor diameters. In the principal wake direction, D1's normalized power rose from 0.59 in baseline operation to 0.76 with wake steering active — a 29 percent increase. That is a real, directionally specific gain. FLORIS had predicted a larger peak gain of roughly 40 percent for that direction, so the observed improvement fell short of the model's optimistic estimate, but the direction of the effect and its approximate magnitude were consistent with the prediction. The upstream turbine C1 showed no strong overall reduction in power. That result tracks directly from the low pP value: with an exponent of 1.41 fitted from field data — almost identical to the FLORIS-tuned value of 1.43 — the cosine-based power loss relationship means yawing by a moderate angle costs the upstream turbine relatively little. The field-derived pP and the model-tuned pP agreed to within two hundredths. That is a meaningful form of model validation: the one parameter that most governs whether wake steering is worth doing at all was independently confirmed in the field. At the mid-range spacing — C1 and D2 at eight and a half rotor diameters — the results were more ambiguous. Some improvement appeared in the main wake direction around 50 degrees, but the twenty-fifth to seventy-fifth percentile bands overlapped enough that the paper describes the result cautiously. At the longest spacing, C1 and D3 at fourteen and a third rotor diameters, no measurable gain appeared. That outcome makes physical sense: at that distance, the wake has recovered substantially, so the baseline power loss is small and there is less to recover. Fleming and colleagues are explicit about what the models got right and where uncertainty remains. SOWFA's prediction that the Envision turbine would lose power more slowly with yaw than prior literature suggested was confirmed by the field. The directional structure of the gains — strongest at seven rotor diameters, fading with distance — matched the qualitative prediction from FLORIS. But the quantitative agreement was imperfect, and the authors attribute part of the gap to the controller itself. By smoothing the sharp yaw offsets prescribed by FLORIS, the real controller delivered a blurred version of the optimal strategy, trimming the peaks. Three gaps stand out as the clearest next steps. First, turbine loads. Sustained yaw misalignment changes the structural loading on both the upstream and downstream turbines, but load measurements were not available in this campaign. Zalkind and Pao, Schulz and colleagues, and a separate instrumented field test by Fleming's group are all working on this problem. Second, closed-loop control. The Rudong campaign ran open loop — the controller never looked at the actual wake. Moving to a system that tracks the wake in real time using lidar or estimation algorithms is an active research area, with large multiyear programs already underway. Third, longer and better-designed campaigns. Alternating controller conditions rather than running two sequential phases, and accumulating the 30 to 40 test days per condition the authors recommend, would sharpen what can be concluded. The Rudong result is a proof of concept, not a deployment specification. But it is the kind of proof that matters: a real farm, a real controller, real supervisory control and data acquisition data, and a downstream turbine that gained 29 percent in its worst-case wake direction. The models built to design that strategy held up where they mattered most. What the offshore wind industry does with that result depends on whether the load question, the closed-loop question, and the data quantity question can be answered at scale — and those are the experiments that come next. 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.