Wake steering is a form of wind farm control in which intentional yaw misalignments are applied to upstream wind turbines to change their wakes to benefit downstream turbines . Research has demonstrated that by implementing well-designed control strategies, it is possible to increase the combined power production of a given set of wind turbines.

Engineering, or control-oriented, models of wakes and wake steering are critical to the design of successful control strategies. In order to be used in the optimizations that design controllers, or perhaps better in the online control system itself, it is necessary that the models are computationally efficient. They are therefore typically analytical or statistical in nature.

The FLOw Redirection and Induction in Steady State (FLORIS) software framework is one such engineering tool, and it includes several engineering models of wake and wake steering, as well as the tools used in the design and analysis of wind farm control strategies. The original FLORIS model, detailed in , combined a multizone adaptation of the Jensen/Park model with a model of wake steering as deflection, as described in .

The current underlying wake model used in FLORIS-based design and analysis is the Gaussian wake model described in and and the model of deflection provided in . However, this model is modified by an analytic approximation of the curl model to adjust the deflection of nonsteered wakes by interaction with steered wakes (secondary steering), as well as an effect called “yaw-added recovery”, which adjusts the velocity recovery of steered wakes. This model is called the Gauss–curl hybrid (GCH) and is presented in . Since its incorporation into FLORIS, it is the standard tool used in design, optimization, and analysis of wind farm control, because of its ability to predict the second-order effects of secondary steering and yaw-added recovery, in addition to wake deflection.

In order to achieve optimal results, it is necessary to validate the predictions of engineering models like FLORIS. Original validation experiments of FLORIS were compared to the large-eddy Simulator fOr Wind Farm Applications (SOWFA) . Wake steering is compared to SOWFA in , , , and . Validation of the engineering models with respect to higher-fidelity simulations is still ongoing.

There is also growing literature on wind tunnel investigations used to assess engineering models of wake steering. These include two-turbine , three-turbine , and five-turbine studies. These experiments allow for controlled and detailed experimentation and examination of wake steering.

There are further experiments of wake steering completed at full scale. In one experiment, a 1.5 MW wind turbine had a rear-facing lidar installed scanning the wake while different fixed yaw offsets were applied in . The results were then compared against several models within FLORIS in .

Finally, there are tests of wake steering at commercial-scale sites. These include tests made using fixed offsets over a fixed sector of wind directions (in both cases 40∘ sectors are used) and offsets applied using an optimal schedule based on wind directions . In this article, we review the results of a new wake-steering experiment. The campaign is different from previous campaigns conducted by the National Renewable Energy Laboratory (NREL) in that fixed offsets (as opposed to a lookup table of optimal offsets per wind direction) are applied for all directions where the wind turbine's wake impacts another turbine. This sort of test has several useful attributes. First, by not varying the offset amount, wake-steering effects can be observed at full scale isolated from issues of tracking a changing offset amount. In , for example, it is posited that the main gap in achieved versus modeled wake-steering results is attributable to this tracking/controller problem. This experiment offers a test of that supposition. A second advantage is that because the same offset is applied regardless of whether the wake would be steered away from a downstream wind turbine (normal wake steering) or toward a downstream wind turbine (what might be called “wrong-way steering”), we can examine the asymmetry of wake steering. This provides an important test of the yaw-added recovery component of the GCH model of wake steering. Finally, model predictions are compared to measured results to assess agreement.

We conducted the experiment on an 11-turbine wind farm illustrated in Fig. . All turbines at the test wind farm are the same model and dimensions. Exact details are withheld, but the rated power is close to 2 MW, the rotor diameter is in the range 80 to 90 m, and the hub height is between 60 and 70 m.

In the test, we applied offsets to the yaw controller of a single wind turbine. This turbine, T5, is referred to as the controlled turbine and is indicated in Fig. . A second turbine, T4, is designated as the reference and is in the freestream for the wind directions of interest. Finally, turbines T6 and T10 are the downstream test turbines, and we assessed the impact of various offsets on those turbines. In addition to the wind turbines, the site includes two ground-based lidars. One of the ground-based lidars was located approximately 2.5 rotor diameters south of T4. The second was located approximately 2.5 rotor diameters north of T5. Finally, prior to the experiment, we equipped several of the turbines (T4, T5, T6, T3) with nacelle-mounted lidars.

Data from the turbines are reported in 10 min averages, and there was no possibility to inspect higher-frequency data. This is different than the 1 min averages used in , for example. While 10 min average data are conventional for most supervisory control and data acquisition (SCADA) analysis in wind energy, we believe it can be a challenge for wind farm control analysis, which is highly sensitive to wind direction.

A first step in preparing for the campaign was constructing the FLORIS model of the test site. An example simulation from this model is shown in Fig. . Details of the FLORIS wake model used are provided in the Appendix.

With the first initial comparisons of the FLORIS model to the SCADA data collected before the start of the campaign, we observed that the absolute wind directions (those giving the compass direction of the wind and not the vane-measured relative angle to the wind turbine nacelle) were approximately +9∘ offset from FLORIS' expectations. We define a positive yaw offset as one in which the turbine is rotated counterclockwise of the incoming wind direction (as illustrated in Fig. ).

Comparison with the two ground-based lidars revealed a similar discrepancy. Because four of the wind turbines had nacelle-mounted lidars installed, we were further able to assess that the relative yaw misalignment (i.e., relative error between the turbine yaw angle and measured wind direction) of each of those turbines appeared to be close to 7∘. We then assigned the residual error over 7∘ to the nacelle position encoders. Figure  illustrates these direction signals and the sign and naming conventions used in this article. For the seven other turbines, we assumed that the error could be similarly divided between the vane and nacelle position error. These results are summarized in Table .

With these findings, we corrected the raw nacelle position and wind vane data using the offset values in Table . The assumed offsets imply that the normal condition of the wind farm is for all the turbines to have an approximately 7∘ offset. This update is incorporated into the FLORIS model of the wind farm by correcting the raw values. Figure  shows an illustration of the finished FLORIS model of the site assuming these offset values.

Note that the offsets are positive angles, which in our convention implies a counterclockwise rotation, which is the direction we use for wake steering. Accounting for this issue when resimulating the data in FLORIS should mean the SCADA data and model can be directly compared. However, it is important to note that this default offset changes the nature of the test from a comparison of a yawed to an aligned wind turbine to a comparison of a smaller and larger offset.

Following the discovery of the systemic small positive offsets, we decided to divide the static test into three phases. In all phases, a target offset would be sent to T5, the controlled wind turbine. In the first phase, we applied an offset of -7∘ (a clockwise rotation) to confirm the lidar-measured offset by seeking to observe an increase in power of the controlled wind turbine when the offset is applied. Then, we applied a +10 and +20∘ (counterclockwise) offset in phases 2 and 3.

The first phase -7∘ offset target test was run from 18 February through 29 March 2020. The second phase test targeting +10∘ offset ran from 7 April through 23 June 2020. Finally, the +20∘ offset period ran from 25 June through 2 November 2020. Note that for this site, which is located in the Northern Hemisphere, the winds are most prevalent in the winter and least in the summer, so to some extent these periods are a bit unlucky in producing annual wind speeds that are lower than average.

In all phases, the controller is toggled on and off hourly to generate a data set with an equal amount of baseline (no applied offset) versus controlled (offset applied cases), with a similar distribution of observed wind speeds and directions.

In Fig. , the yaw offsets of the controlled turbine (T5) are shown. The offsets measured both by the wind turbine's vane (black) and then computed by comparing its yaw position to the wind direction measured by the reference T4 (red) are shown for the three periods. The top row represents the baseline conditions when no offset is applied, and so the wind turbine is at +7∘ measured either by itself or by the reference. In the controlled cases, there are a few details to note. First, the offset is limited to wind speeds of 14 m s-1 and below. Second, there is a certain amount of undershoot in the self- and reference-measured offsets. The average offsets are summarized in Table . The gap between self- and reference-measured offset grows with the size of the target offset and most likely reflects a need to apply corrections to the vane signal when operating in yaw (although it could also be partially attributable to noise issues). Note that when simulating the data in FLORIS, we used the reference-measured offsets.

In this study, we first cleaned and filtered the collected data. The data are limited to directions of interest, and wind speeds are limited to 14 m s-1 and below. The data are limited this way because as shown in Fig. , above 14 m s-1 the offset is turning off. Next, we removed any data that are flagged as curtailed or not fully operational for any of the relevant wind turbines. The data were then further filtered to remove power production anomalies using the filtering toolkit in NREL's OpenOA software . Specifically, the power production data for each wind turbine were divided into 50 kW bins. Within each power bin, samples for which the nacelle-measured wind speed is more than 1 SD (standard deviation) from the median wind speed are treated as outliers and removed. Finally, the wind vane and nacelle measurements for all the wind turbines were corrected according to Table .

Next, we resimulated all the data points using FLORIS. The wind speed and direction are provided by the reference turbine (T4) (and not the lidars to maximize data availability, noting that the turbine's measurements are calibrated against the lidars).

This article presents results from a wake-steering study using fixed yaw offset targets, instead of a lookup table of optimal yaw offsets as a function of wind speed and direction. The results allowed us to investigate model predictions in conditions that would normally be avoided, such as intentional wrong-way steering. Comparing the model and measured results shows that the gains from right-way steering were in line with expectations, whereas the losses from wrong-way steering were less. The lower-than-expected losses from wrong-way steering are particularly promising because the GCH model of wake steering to which the results are compared already predicts a decrease in losses compared to the standard Gaussian wake model as a consequence of the yaw-added recovery effect.

We believe this article adds to the growing literature that validates wake steering at a range of scales and fidelities. Comparisons between model predictions and SCADA results were generally positive within the limits of the data uncertainty. The results show that engineering wake-steering models for small numbers of wind turbines and when wind-direction-based offset strategies and associated control implementation difficulties are removed are reasonably accurate.

For wind farm control to be realized in practice in industry, validation campaigns which increase confidence in the engineering models used in the design of wind farm controllers are critical . We believe this campaign contributes importantly to that need. However, the campaign, like previous campaigns, is of only a limited number of turbines. The performance of wake steering in arrays of turbines larger than 10 would validate additional important effects, such as blockage, secondary steering, wind farm boundary layers, and deep-array effects, all of which may pose problems that do not manifest themselves at the scale of a few turbines. Future research campaigns should address these issues through implementations on larger numbers of turbines.

Finally, design and analysis of controllers for achieving the maximum possible performance, given practical issues such as modeling error, wind direction variability, and yaw activity limitations, as well as opportunities to match control strategies to specific atmospheric conditions, are recommended for future research.