This study introduces an innovative bat protection strategy for the operation of onshore wind farms that limits the negative effect on annual energy production. This strategy relies on bat detection mechanisms and shuts down the turbines when bat activity is detected, i.e. dynamic, rather than pre-defined “blankets” that depend on atmospheric conditions. The authors also analyze the effect of these strategies in terms of structural loading, demonstrating that the dynamic strategies do not increase structural loads while guaranteeing lower losses in energy production.

The topic is well aligned with the current interests of the scientific community and the results presented here are promising. I include here some comments to improve the quality of the manuscript.

Comments:

1. The manuscript is quite long. I suggest shortening a few parts, such as the detailed description and comparison of the machine learning models. Othwerwise, some parts could be moved to the Appendix, such as Fig.5 and Table 4.
2. The tuning procedure described in Section 2.1 is not very clear. Figure 2 facilitates comprehension, but it could be better supported by the text. For instance, mention the division into the two types of coefficients more clearly, the cost function, and the total number of coefficients that are tuned. If the length increases significantly, the section could be moved to the Appendix.
3. In Section 4.7 (line 496), it is stated that bat activity occurs more frequently at lower wind speeds. However, I am not sure that this information can be easily obtained from Figure 9 (left) since the distribution is very similar to the typical Weibull distribution of wind speed occurrence. In other words, if the distribution of bat activity were the same as that of the wind speed occurrence, it would imply that bat activity does not depend on the wind speed. I suggest plotting these data without the bias of the wind speed occurrence to better convey your message.
4. Previous studies highlight the need to consider multiple sectors instead of only one rotor-averaged inflow estimate to predict the structural loads using surrogate models. This is mentioned in Section 4. However, the impact of this assumption could be tested with tools available in the literature, e.g. surrogate models that take sector-averaged inflow as input instead of rotor-averaged inflow. I recognize that this may be beyond the scope of the study but it would strengthen the results of this work.

Minor comments:

1. In the Introduction (lines 45-50), the statement on the loads of the transients refers to floating wind farms. I suggest finding a reference for onshore farms or justifying why this also applies here.
2. In Figure 2, the orange dots on the flow field are difficult to see. I suggest changing the color.
3. In Figure 6, is it possible to add the DELs obtained from the aeroelastic solver? Keep it as possible suggestion.
4. I recommend using the word ‘blanket’ also in the description of the methodology (Section 2.8) for consistency with the rest of the manuscript.
5. In Figure 12, I suggest including the acronym (e.g. TBTOR) for each sublot to be consistent with the text, for instance, below each subtitle.
6. In Figure 13, the plot on the right is difficult to read. The x-axis labels are a bit messy and the bar widths are very small. I suggest increasing the bin size of the TI.
7. For consistency with the other metrics, I suggest presenting the results of Table 6 as a histogram and including the information related to the inter-annual variability in Table 7 instead.

The manuscript by Pettas et al. combines multiple different fields of wind turbine and wind farm research to research the effects of bat protection strategies on wind farm performance using a specific case study. The research, including its assumptions and shortcomings, is well executed and well documented. The paper has significant scientific value, and I recommend it to be published in WES. I do however have five larger and a couple of minor comments that I recommend be addressed before publication.

Major comments:

1. The manuscript is very long. Part of this can be explained by the fact that the manuscript covers many different areas (describing the specific case study, the engineering wind farm model, the turbine load surrogate model, the bat protection strategy, and the effects in terms of both power and loads). However, the same information could in my opinion still be provided in a significantly shorter manuscript. The authors use a very verbose writing style, and many different paragraphs provide little to no new information with respect to provided figures and/or previous paragraphs. In the minor comments below I will make several suggestions for shortening the manuscript, but generally using a more concise writing style would be my number 1 recommendation.
2. At the end of section 2, on line 542, the authors say: “the dynamic cases are assumed to correspond to a 100% bat protection level”. I would argue that this assumption is in no way realistic in this case study. The wind farm uses only one bat sensor on the very edge of the farm, so it is highly unlikely that every bat activity within the entire wind farm is captured by the sensor. From Figs 9 and 10, it follows that bats have relatively predictable behavior, which is obviously why static protection algorithms are usually implemented. As a result, bat activity resulting in false-negative measurements by the bat sensor are likely still captured by these static algorithms, providing additional protection to bats that the dynamic algorithm does not. This effect is not mentioned anywhere in the manuscript, not here and only partly in Section 4. I understand that it is hard (perhaps impossible) to quantify this effect using the available dataset, but it should at the very least be mentioned more explicitly as a limitation of the dynamic protection algorithms.
3. The figures in the manuscript do not match the quality of the manuscript itself. Figures use acronyms not explained anywhere in the surrounding text, some of the captions/textboxes in the figures are too small to read, and some of the data are too small to reasonably interpret. I suggest cleaning this up in the final manuscript. See minor comments below for some suggestions.
4. Line 331-333: “ Instead, wake ... of wakes”. Is it possible to quantify how well this method captures loads induced by wakes? If you have wind direction and load SCADA signals, you should also be able to see whether a turbine is waked. So, you could make a figure such as Figure 5 comparing waked and unwaked turbines. Such a figure would in my opinion be more valuable than the current Figure 5, as it is well-known that waked turbines see substantially increased loads and this situation therefore contributes significantly to the overall fatigue damage. If the error of your model in waked situations is much higher than in unwaked situations, as I would expect, that is another unmentioned limitation of this study.
5. Section 3: What I am missing in this section is an analysis of how well the different strategies really save bats. For example, how much of the measured bat activity in year 2 is missed by implementing the Static90, Static99, and Static100 strategies? Do we still achieve the promised percentage of bat protection? If not, how can this be improved (by using rolling average temperatures perhaps?). What if it were reversed and year 2 would have been the benchmark year?

Minor comments:

Line 35: “In practice … wind farm”. A source would be appreciated to support this statement

Figure 1: FLORIS, DELs, KPI are all undefined at this point in the text. The graphs are too small to make out and add no value.

Figure 2: The cost function formulas add absolutely no value/clarity if they are not elaborated on/explained in the paper. I suggest removing the formulas from the figure.

Table 1: All the parameters listed in this table are not defined anywhere in the paper and therefore do not support the text in any way. If the authors insist on providing these values, I suggest moving them to the appendix (possibly together with the section describing the FLORIS tuning, as this is not the main contribution of the paper).

Line 203: The case study wind farm and its turbines are mentioned here but are not defined yet. This caused confusion for me as I was reading it. I suggest moving section 2.6 up to before the current section 2.2.

Table 2: Seed-avg points are not defined or mentioned anywhere else. I’m therefore not sure I understand what you mean by this.

Section 2.3: This section is VERY long. It describes and compares two different surrogate models, while this is as far as I see it not one of the main contribution of this paper, as both methods have been used and described in literature before. Given the length of the paper, I suggest considering to remove or to significantly trim this section.

Line 262: “Li et al. (2018)” should be “(Li et al, 2018)”.

Line 270: How is the model accuracy measured? Do you use higher frequency SCADA data? This might be mentioned earlier in the text but it’s not completely clear to me.

Figure 6: This figure looks very sloppy. The font size is too small, the use of “20k” is not very scientific, the legend uses inconsistent naming (“start”, “shut”), and it is unclear to me why these specific channels are plotted while the third dimension channels and the tower-top channels are omitted.

Line 339: “response”, not “resposnse”

Section 2.4: This section is again quite long, while in my opinion providing limited new information with respect to literature. Consider trimming to only describe the most relevant findings.

Section 2.5: See above. This section could potentially be omitted completely or merged into one of the previous sections in my opinion.

Line 466-469: “Missing timestamps … temporal coverage”. How reliable is this method? Would it not be better to simply omit the data that is unavailable, especially if it is only 0.4% of the dataset?

Line 498-499: “The seasonal …during May”. I assume this can be explained (at least partly) by the daily mean temperatures, i.e., year 1 having a warmer spring than year 2? Did you investigate that, and consider using that as a potential indicator of when to initiate any of the bat protection protocols?

Figure 9: This figure would be more useful in my opinion (at least the lefthand part) if it were shown relative to occurrence of each wind speed bin. Now it is hard to compare this figure with the lefthand side of figure 8 to truly assess when bats are most active. Furthermore, there’s a space missing between “Right:” and “bat”.

Line 500-510: this whole section seems repetitive w.r.t. earlier parts of the manuscript.

Figure 10: I suggest plotting sunrise/sunset as a line instead of as dots, as it develops continuously throughout the year. I would also move the legend box, so it doesn’t unnecessarily overlap with part of the plotted points, and cut off the figure at 0 and 24, as there is no 25^th hour in a day. Similarly, I would suggest using 6-hour intervals for the y-label markers, which makes much more sense for time of the day.

Line 555: Why do you use a hashtag (year #1) here and throughout the manuscript instead of just “year 1” or even “2022” (or whatever year it was).

Figure 11: I would be very interested to learn how the baseline downtime and start/stop events compare to the actual measured values. Please consider adding this to the manuscript.

Line 559 and after: you use decimal points to describe time, but you have a maximum accuracy of 10/60 minutes = 0.166 hour > 0.1. Please round up to full hours or use integer 10-minute intervals.

Section 3.2: Again very long, and a lot of the text adds no information that isn’t provided by the figures.

Figure 12: It would make a lot more sense to me if all the same elements (tower-base, blade-root, etc) are grouped together in the same row or column. Furthermore, because the Static100 case results in so much higher load variations compared to all other cases, the y-axis is stretched too much for all other signals, and subsequently, the value of the boxplot is diminished. Consider removing the Static100 case from this figure or plotting it separately.

Line 599: period missing after Static100.

Line 611-612: “This is illustrated in Figure 13”: In my opinion, this also follows logically from Table 5. I therefore see limited value in including Figure 13.

Figure 13: If you do keep this figure, define acronym WSP. Furthermore, the rhs of this figure is unreadable right now with the number of bins shown. Either increase the size of this figure or increase the bin size so the number of bars is reduced.

Line 624-633: This paragraph in my opinion summarizes the true relevance of this section, combined with Figure 12. A lot of what is previously written in this section can honestly be removed.