This is a well-written paper that discusses the design of a combined model and data-driven wake steering controller for a large wind plant and evaluates the ability of the underlying wake model to predict power production at unwaked and waked turbines. The paper clearly and thoroughly explains methods and design choices for determining turbine Cp and Ct curves required for the engineering wake model (EWM) as well as estimating required wind inputs to the EWM (ambient wind speed, wind direction and turbulence intensity). A detailed explanation of the design and training steps of the data-driven output corrector stage of the wake model is provided as well. Some of the main contributions of this paper to the wind farm control literature include 1) showing the design of an EWM-based wind farm controller and the input estimation steps for large wind farm applications, 2) showing how the EWM can be augmented with a data-driven model, and 3) showing how the power prediction performance of the resulting model can be incrementally improved through several modifications to the controller (e.g., calibrating absolute wind direction, including wind speed heterogeneity, and adding the output corrector stage).  

The main area where the paper can be improved is showing how well the wake models predict the change in power or energy from wake steering. The results focus on the ability of the different model variations to predict the power ratios of different sets of turbines relative to unwaked turbines, which is useful for general wake model validation. But since the goal of the models is to optimize yaw offsets for wake steering to increase power capture, the ability of the models to predict the change in power with wake steering compared to normal operation should be investigated in more detail. For example, Figs. 12 and 13 show the power ratios as a function of wind direction for two waked turbines with and without wake steering, but the expected change in power from wake steering is probably small for these turbines and it is hard to determine how accurately the models capture this change from the separate plots. Plots showing the measured and predicted difference between the power ratios for wake steering and normal operation cases would be more effective at showing how well the model captures the change in energy from wake steering. Further, similar power ratio plots for upstream turbines (where we'd expect a power loss from wake steering), combined upstream and downstream turbines, and larger clusters of turbines (e.g., turbines 9-12 and 19-24) could help validate the models in a wider variety of scenarios capturing both power losses and gains at different turbines from wake steering. Aggregate metrics such as the total power or energy gain over a range of wind directions could be helpful too. 

Additional specific comments are provided below.

1. Pg. 1, ln. 18: Another reference to consider citing for the magnitude of predicted wake losses is Lee and Fields (2021): https://doi.org/10.5194/wes-6-311-2021, which shows typical losses between 5% and 20%.
2. Pg. 2, ln. 29: "non-waked inflow": Or more accurately, the downstream turbine will experience less wake overlap or a reduction in wake velocity deficits.
3. Pg. 3, ln. 59: "the timescales on which wind characteristics change.": Can you elaborate on this? For example, the wind characteristics change too quickly for the extremum seeking controller to converge?
4. Pg. 3, ln. 70: "helps the system account for spatial variation in wind characteristics.": It is hard to tell what helps the system account for spatial variation in wind characteristics. Can you clarify what part of the discussion above you are referring to?
5. Pg. 5, ln. 100: "The farm layout presents a challenge for wake steering control": Another challenge worth mentioning here is that because of the long distance between turbines in the predominant wind direction, wake losses are expected to be relatively low (if that is the case), so there is probably not much room for wake steering to increase power.
6. Section 2.4.1: When estimating the Cp and Ct curves using nacelle wind speed measurements, do you account for potential biases (that are also potentially turbine-specific) between the nacelle wind speed measurements and the true freestream wind speed (i.e., by determining and then applying a nacelle transfer function)? Can you discuss how these biases could affect your estimation of the Cp and Ct parameters? 
7. Eq. 3: The power loss is more accurately modeled in FLORIS by scaling the effective wind speed experienced by the turbine by cosine(gamma)^(p/3). This way power is inherently kept at rated power above the rated wind speed when the turbine yaws.
8. Pg. 9, ln. 187 "These can be excluded by filtering any turbine wind speeds…": Could you clarify whether you apply this filtering in your wind speed estimator?
9. Pg. 10, ln. 230: "we also have access to the wind direction reported by GNSS": A GNSS measures the nacelle orientation, but how do you determine the wind direction from the GNSS?
10. Pg. 11, ln. 237: "The wind direction obtained through this approach, however, may differ from that needed in the model…": If you are using GNSS wind direction measurements, which should already provide accurate nacelle orientation measurements relative to true north, why would further calibration be needed? Also, is the "manual" calibration of SCADA signals mentioned in the previous paragraph the same process as described in the rest of this section?
11. Pg. 11, ln. 239 and pg. 12, ln. 271: "waking directions": How is "waking direction" different from "wind direction" as used here? If there is no difference, consider using "wind direction" to avoid confusion.
12. Equation 9: What temporal averaging period do you use when calculating the power ratios?
13. Pg. 12, ln. 255: "we filter for yaw errors below 5 degrees": Do you require all of the reference turbines to have yaw errors less than 5 degrees as well, or only the target turbine?
14. Section 2.5: To confirm, how exactly is the combined EWM/output corrector model intended to be used for optimization? Are the input features measured at the current 1-minute time step k used to find the optimal yaw offsets to be implemented during 1-minute period k+1?
15. Pg. 16, ln. 345: "we have mTk possible features": What does "T" represent?
16. Pg. 17: Why aren't any EWM inputs used as features in the models? Were they not helpful in improving estimation accuracy?
17. Pg. 17, ln. 389: "1 minute rolling average of SCADA power" is listed twice.
18. Pg. 17, ln. 392: "but with the target turbine's wind speed and power measurements omitted". Are the 1-minute lagged power measurements at the target turbine omitted as well, or are they still used as features?
19. Pg. 18, ln. 405: "was reduced by a factor of 8 from the IEC value": Why did you choose to make the waked sectors so narrow? This seems too narrow to capture the true width of the wakes during higher turbulence intensities.
20. Pg. 18, ln. 416: "Any column that was missing more than 60% of the data…": Can you clarify what columns and rows correspond to in the feature matrix? Does this mean that features are removed if they are missing for more than 60% of the turbines, and then turbines are removed if any of their features are missing?
21. Pg. 18, ln. 416: Additionally, can you discuss how the output predictor model deals with missing input features during the training and prediction stages? As stated earlier, a maximum of 404 features can be used to predict waked turbines' power. It seems likely that some of this data would be missing a significant amount of the time. Can the model still be used to predict power if it uses a different combination of available input features than were used for training? And can the training stage be performed using different combinations of available input features for each 1-minute sample?
22. Pg. 18, ln. 430: "the model should accurately capture the mean power loss due to waking as a function of wind direction, wind speed, and TI.": In addition to wind direction, wind speed, and TI, the model should be able to predict power as a function of yaw misalignment. How is this validated? 
23. Pg. 22, ln. 489: "The correlation coefficient of the power ratio is also strongly affected by modeling errors across the entire plant because incorrect power predictions at the reference turbines will affect the power ratio predicted at all turbines": Wouldn't some errors in power prediction tend to cancel out across the farm when using the power ratio metric? For example if power is underpredicted at all turbines in the farm (both unwaked and waked), then the errors would tend to cancel out when calculating the power ratio.
24. Table 2: Why does the correlation coefficient for turbines 8-13 decrease between models 1 and 2? The only change in the model appears to be adding a 5-degree offset to the wind direction, and these turbines are unwaked anyway.
25. Pg. 27, ln. 550: "While the reason for the time lag is unclear…" and pg. 29, ln. 571: "Model 4 has an apparent 1 minute time lag…": The model features include the target turbine's 1-minute rolling average of SCADA power lagged by one minute for freestream turbines (and maybe for waked turbines? (see comment 18)). It seems likely that this lagged power feature is one of the best predictors of power during the next 1-minute period, so the model heavily relies on this measurement, explaining the time lag in the prediction time series. Can you discuss this possibility in the paper? Additionally, for waked turbines, the current 1-minute averaged power of freestream turbines are used as features. Since the wind speeds at the freestream turbines could tend to travel downstream at the mean wind speed and interact with the waked turbines after some time delay, could this behavior also contribute to the visible prediction time lag?
26. Fig. 18: Please discuss this plot further. Is model 4 only trained with yaw angle magnitudes > ~8 degrees, so the model reverts to the EWM model for smaller yaw angles? Why is the predicted power gain so much higher for model 4 than model 3? The gain above 1 appears to be an order of magnitude higher with model 4 than model 3.

see attachment