This paper presents a comparison of engineering wake models (available in FLORIS) to the mesoscale WRF simulator for comparing farm-to-farm wakes in the North Sea. Overall, the paper is well-written and its scope is well-defined. The results may help to explain the strengths and weaknesses of different engineering wake models, although some clarity is needed to understand the modeling procedure. I have also left some open questions that the authors may consider addressing (but these should not be treated as necessary).

I will add a disclaimer here that I am a developer of the FLORIS engineering wake modeling package that is used throughout this work.

*** Comments on modeling procedure.

- The validation procedure for WRF seems fairly limited. As I understood it, there are a handful of locations used for validation, but even the highest point is still less than half of expected turbine hub heights as it is only for a few locations. Moreover, the WRF simulations used for calibration contained no wind farms, since none were built yet; but this does beg the question of how useful such validation is when the rest of the study is comparing wakes. The authors do, to some extent, point out the WRF simulations are not to be treated as "truth"; however, I think the limitations of validation process should be articulated more clearly.

- Similarly, I feel the statement that WRF should not be treated as "truth" should appear more strongly earlier in the paper. Although the authors do state that WRF is not the truth, my impression is that this message often gets lost in comparisons (in this work and others), which may mean we select parametrizations for engineering models based on faulty criteria. 

- Finally, along this vane, the statement that "advanced" features of FLORIS (heterogeneity, shear(?), veer) appears only in the conclusions (lines 428--430), and I think should have been stated in section 2.4. There, homogeneity is mentioned, but not shear and veer, and since some level of shear is commonly the default configuration for FLORIS (and appears in the example inputs of FLORIS v3.4), this should be stated clearly. 

- The authors ran FLORIS for a grid of wind speeds and wind directions to generate a look-up table of results, which were then mapped to the year-long case. This procedure is reasonable, but I did not quite understand what wind conditions are used to sample the look-up table (line 182). The authors say "the time series extracted from the no WF WRF run", but I assume the wind conditions varied across the domain---how did the authors chose a scalar wind speed and direction to sample the gridded FLORIS results? Were these selected from a single point in the simulation domain, or some sort of average? Lines 381 and 382 suggest (if I'm reading them correctly) that background flow heterogeneity may be partially the cause of differences, so the choice of representative background flow condition is important here.

- Fig. 10. (c) and (d) seems to imply that the FLORIS results depend on the stability class. However, the authors state that turbulence intensity is fixed at 6% throughout for FLORIS simulations. Is there some other dependency on stability that is used, or is the variation in the FLORIS result due to other correlated effects (different wind speed profiles, for example)?

*** Open questions (again, these do not need to be directly addressed in the article, but came to mind as I was reading the text and could perhaps be used to clarify conclusions).

- To what extent do the authors feel that retuning certain parameters (e.g. the A of the TurboPark model) could have corrected differences between WRF and FLORIS, not only for the single AEP prediction but also for the time-step to time-step comparison? Is it reasonable to assume that engineering wake models will need to be tuned for specific scenarios? This is touched on briefly in lines 93--96 but not really returned to. 

- The mechanism for further downstream farms might be experiencing "negative" wake losses wasn't clearly explained to me. I understand that the authors are not attempting to investigate this phenomenon closely; however, I didn't really understand the candidate explanation in lines 327--329. I have thought about this (even ignoring higher turbulence) as being explained by a change (decrease) in thrust at the front row turbines (due to the presence of an upstream farm causing lower velocities at the front row turbines), which causes a shallower wake and therefore (in certain cases) more power at the downstream turbines. Is that equivalent to what you are saying?

This article provides a comparison of WRF vs engineering wake models for comparing internal and external wake losses in the North Sea, focusing in particular on the impact from a new cluster, Princess Elisabeth. As other reviewers have provided feedback on the engineering model comparison, I focus here only on the WRF analysis. Overall, it is an interesting study and many insights are provided. However, I am concerned about some of the results which I find difficult to understand. Most importantly, the large amount of negative wake losses presented throughout the plots are difficult to understand. Although flow acceleration around wind farms is indeed a phenomenon, the magnitude and breadth of these negative wake losses is concerning. I think there may be an issue in the calculation that should be investigated. 

I also feel the paper could use a tighter focus. There is a significant section devoted to WRF performance at different sites and atmospheric conditions, and another section devoted to comparison against engineering models, but only at the Belgian-Dutch cluster. In my opinion, this a natural split and could motivate two papers, each of which targeted at answering a specific problem.

I would recommend revisiting the analysis and consider splitting the paper before resubmission.

Line 21: small point, but it is convention to express ranges from lowest to smallest number

Line 23: Unclear what “qualitatively” refers to here. Aren’t spatial distributions of wakes or waked wind speeds still fundamentally quantitative? Perhaps a better term could be used here.

Line 45: “limiting their application to specific atmospheric conditions”. Not true anymore. Companies like Whiffle are running GPU-powered LES models that can run full year simulations at costs often deemed reasonable by stakeholders. Today, computational limits are less prohibitive.

Line 46: Why can engineering wake models account for more atmospheric conditions but high fidelity microscale models can’t? I feel like the opposite is more true. Engineering models have tuning often ignore atmospheric conditions completely, apart from maybe turbulence intensity. I think some clarity on what high fidelity models can do vs engineering models would be useful here.

Line 51: Range is more like 500m-3km. Not sure who is confidently using data at 5km resolution. Also, the size of the turbine isn’t the relevant comparison here; it’s the spacing of the turbines, distance between wind farms, etc. that matters

Line 53: Some brief description of how wind farms are parameterized in mesoscale models would be useful here, with emphasis on the default Fitch parameterization.

Line 54: Have different wind farm parameterizations been used, or different parameters to the WRF model? Important distinction here.

Line 58: Not sure I agree here. Lately, running 1 year + simulations is the norm and is certainly sufficient to compute annual impacts, especially if a typical meteorological year is chosen. Again, cost is becoming less and less of an issue, and it could be useful to make that point, i.e., we used to run shorter periods due to computational limits, but are now getting longer.

Line 75: I think a better distinction between mesoscale and engineering models is needed. Fundamentally, mesoscale models make no assumptions about wake behavior. How wakes propagate, interact, dissipate etc. is determined entirely by prevailing atmospheric conditions simulated in the model. Engineering models, on the other hand, fundamentally make explicit assumptions about wake behaviour. It’s very important to make this clear.

Line 99: I’d be specific about language. Mesoscale models are not wake models. They parameterize wind farms directly in the simulations, and we can deduce wakes from those simulations. They aren’t wake models though, unlike the engineering models.

Line 130: (placeholder) why are we comparing against obs given PE doesn’t exist yet? Why not use deployed lidars in the region?

Line 135: What value was used for the TKE generation parameter?

Line 180: where does the wind rose come from? What wind data are you using to inform the engineering wake models, and how does this data compare to the WRF simulations?

Table 5:  The varying degrees of bias are concerning, as is the 1.22 m/s bias at SW. What QC was performed on the measurement data, and over what period were the data analyzed? Some info on data quality, recovery rate, etc. would be helpful. Also, in the end, validation against measurements under 40m nearly as important as validation at hub height, again pointing at the importance of using lidar data instead

Figure 3: Struggling a bit to interpret here. I don’t see any real impact apart from the main downwind cluster. Also why does site B have losses on the north end but nowhere else, and which also doesn’t align with the wind speed deficit map? Further, the wind speed deficit map is spotty and difficult to explain (why the northern deficit?). This is likely due to numerical anomalies between the different simulations, and I recommend limiting impacts to 0.05 m/s or so, or some value that eliminates these artifacts. Finally, why limit wind deficits to max of 0.2? We’re missing a lot of structure in that big blob in the middle.

Table 6: Do you discuss the negative values anywhere in the paper?

Line 274: The trend should be very statistically significant. We should see considerably higher wake losses in stable conditions. And how are you quantifying statistical significance? There should be a much stronger signal here. Part of the problem may be the Obukhov length, which is surface specific. Suggest using bulk Richardson number, or vertical temperature gradient, as alternates to compare.

Figure 4: Why are there external wake losses in sectors other than SW, and what do you make of the negative losses?

Line 283: Not necessarily related to thrust. At high winds, you’re well above rated capacity, and even with wakes you can stay above rated, hence no losses. It’s more a question of wind speed magnitude. At low wind speeds, your denominator is low, hence percentage based losses are high. It’d be useful to see a side by side of percentage and power/energy based losses

Figure 6: I don’t understand these negative wake losses in multiple wind directions. May I ask how you’re calculating the average? Are you calculating a percentage loss for each timestep, then averaging those percentages by sector? Or are you (and more correctly) averaging/summing power for each timestep for each sector, and only calculating percentage loss based on those final averages? The two approaches will yield very different results and the latter is more defensible.

Table 7: I don’t understand how KF and GS can have negative external wake losses arising from the Prinses Amalia deployment. They’re too far away. I think there is an issue with your calculations.