This manuscript presents a data-driven approach based on ANN for estimating the thrust load and then fatigue damage on jacket foundation, by using SCADA measurements and tower top acceleration. I have some comments that are needed to be addressed to improve the presentations:

-The methodology section should improve a lot. Section 2.2.1 discusses Tier one which is about the estimation and validation of thrust load. Then Section 2.2.2 is focused on feature selection so that it reviews different techniques but does not clearly answer the question that which technique is applied to this work. The reader expects to realize what you have applied in the methodology section. Please provide references for the techniques that you have mentioned there. Maybe it is better to name the Section methodology, then creating two subsections Input data and the Proposed Estimation Algorithm. Please improve the structure to make it easier for readers to follow the work.

-Please explain why ADAMAX is used for optimizing ANN training. What are other alternatives and what is making ADAMAX different in this context?

-Is it possible to provide a confidence interval for the estimation error by applying different techniques like what is used in the following publication:

Online condition monitoring of floating wind turbines drivetrain by means of digital twin, FK Moghadam, AR Nejad - Mechanical Systems and Signal Processing, 2022

-Please list the general specification of the under investigation turbines, including the rated power and speed.

-Could you give a comment on Figures 6 and 13 results, and explain in the text how the error is correlated to wind speed?

-The highest error in DEL does not correspond to the highest error in thrust load estimation. Could you add a comment on that? Is it possible to do a sensitivity analysis to see which parameter contributes the most to the DEL error?

-Could you move the schematic diagram presented in Fig. 15 related to the farm-wide DEL to the methodology section?

-Line 125 “To avoid excessive drifts in this transformation a lower frequency bound of 0.1Hz is used.”, please elaborate more or possibly provide a reference.

-Line 152 “In the current contribution we focus primarily on the DEL estimation in the FA direction, as this is considered most relevant for the current jacket foundations.” Could you provide a reference?

-Line 170 “The thrust load can be obtained from measurements by low-pass filtering the bending moment timeseries (with an upper frequency bound of 0.2 Hz).” Why is the cut-off frequency chosen to be 0.2 Hz?

-Site-to-side in figure 2 should change to side-to-side.

-Multiple grammatical errors. e.g. “a intermediate” in Line 27 that should be corrected.

Summary:

The article proposes a two-tier data-driven model to predict the 10-min fore-aft damage equivalent fatigue loads on the jacket foundation of offshore wind turbines. Data utilized in this research correspond to SCADA, estimated thrust load and a “low quality, i.e. 1Hz sampled” nacelle-based accelerations signals. The motivation behind this research is that strain gauge instrumentation for direct measurements of fatigue loads on OWT jackets is not economical, and a substitute surrogate data-driven model should be developed.

The topic is very relevant. The paper is well written.

Comments:

At the top of Figure 4, I read the expression “Measured thrust load”. I would not agree that thrust is strictly measured, it is rather estimated from the strain measurements.

At the top of Figure 4, I read that you provide a validation set for the thrust load. This indicates that the thrust has been validated independently somehow. Might you please elaborate on this point? Were validation performed via an aeroelastic simulation? In order to get a good estimate of thrust, the strain gauges need to be calibrated for a know load level. How was this done, in order for one to trust that the Thrust estimates are correct?

On page 9: I suggest that authors explain the basis by which they selected the features of the time series.

Section 2.2.2: another powerful feature selection approach is https://pypi.org/project/BorutaShap/

Page 12 line 289-290: “carefully selected as to be representative of all operating conditions”. What are the criteria for such a selection? Turbine operating in partial load, full load and transition? Turbine operating at max Cp and rated power? Operating conditions covering all possible combination of pitch-tipSpeedRatio? Please explain what are all the representative operating conditions.

Page 12 line 292: why do you correct the thrust for the air density? Are you calculating the thrust coefficient?

Page 13 line 299: ANN as used in this article cannot extrapolate, i.e. they cannot make correct predictions when the input are not within the range of the training set. How do the authors ensure that the validation data (3 months of data outside of the training period) fall within the range of the training set?

Page 13 line 300: the cross-validation applied to a different turbine. Has the cross-validation set been chosen in such a way to reflect the conditions that occurred in the original training set? i.e. you cannot cross-validate on a set where the other turbine is known to be in a waked condition, right?

Table 1: you want to consider adding this to the mix of methods: https://pypi.org/project/BorutaShap/

Page 15: please note that random forest feature importance can be derived either based on mean decrease in impurity or based on feature permutation. Both methods could give slightly different results. Mind you that the results could also be affected by dependent input features. Please make sure you do not bias your feature selection based on the above. Kendall’s Tau takes such dependence into account, and probably why it gives the lowest MAE in figure 7 at the cost of high number of features

Section 3.2.2:  Table 2: once 1-second SCADA data is made available, by default then the 10min-SCADA can be calculated. Did you consider a scenario where both 10min and 1Hz SCADA are made available to the model?

Section 3.2.2:  it is perfectly acceptable to propose several scenarios of various data sources, and check their effect on the model predictive error. However, wouldn’t a more principled approach involving sensitivity analysis and sensors selection optimization with value of information be more adequate? Please discuss.

Page 20, line 472: please elaborate on the 80-20 train-test split. Do you respect the temporal evolution of the data or do you perform randomized split?

Page 21, line 489: “…for a different turbine”. Please specify where this other turbine located with respect to the reference turbine where the model was trained.

Page 22, line 505-508: this might the case, but the question is whether the training set of the reference turbine included any data corresponding to wake? Indeed not all wakes/partial wakes/multiple wakes are created equal because of dependence on atmospheric stability, turbulence and shear. It is worth discussing this issue. The logical consequence of this is that your training set for the reference turbine should include a much larger amount of data in order to take into account the various wake effects in order to generalize to another turbine…

Page 22, line 505-508: 8-18 m/s according to the reference turbine where the model was trained or according to the different (second) turbine? If the second turbine were wake affected, then its wind speed should account for the velocity deficit.

Page 22, line 505-508:   Avendaño-Valencia, L. D., Abdallah, I., and Chatzi, E.

Figure 16: what is mean by mean_DEL? DEL is a short term measure of fatigue conditional on wind speed. Is the mean calculated by weighting according to the pdf of the wind speed? Please elaborate how the mean is computed in this case.

Figure 18: same comment as above.

Figure 16: DEL across various wind turbines in the farm will be highly influenced by the mean wind speed (at each wind turbine), turbulence and shear. Maybe you would want to plot Figure 16 for various wind speed bins (e.g. 6-9m/s, 10-14m/s and >15m/s). You will notice that the DEL across wake-free and wake affected wind turbines will change quite a bit.

General comment:

There are multiple grammatical and orthographic mistakes, and a proof-read is necessary.

It is not mandatory, but you might want to consider comparing the performance of your method to other state of the art methods dealing with a similar subject matter.

Methodological Suggestion:

A more direct approach avoiding the two-tier approach proposed in this article might be to use variational auto-encoders neural networks such as proposed here:

https://onlinelibrary.wiley.com/doi/full/10.1002/we.2621

https://link.springer.com/chapter/10.1007%2F978-3-030-12075-7_21

This manuscript describes use of an ANN for estimating the fatigue damage rates due to fore-aft bending on the jacket foundation of offshore wind turbines. This allows for the prediction of farm-wide fatigue assessment without necessitating condition monitoring installations on all turbines. The motivation is thus to create a data-driven model that estimates fatigue loads as a more economic alternative to farm-wide condition monitoring. Metrics used for inputs to the ANN is calculated from SCADA data, 1Hz nacelle acceleration and estimated thrust loads.

The method is useful in that it estimates fatigue on turbines with no structural health monitoring installations. This can ultimately be used to inform life-extension decisions. This is a valuable contribution to the problem identified in the manuscript.

The work is interesting and well presented.

General comments:

There are some typing and/or grammatical errors that should be revised. For example, ‘within the an offshore wind farm’ in Line 114-115.

One of the unique aspects of this work is stated to be the use of ANN estimated thrust load as an input for the fatigue estimator (Line 72). It would then be sensible to compare the effect of including or excluding the thrust load on the accuracy of the fatigue estimation. Only one such comparison is possible in Section 3.2.2 between scenario D and scenario F. The cross-validation for these scenarios however show that the inclusion of the thrust load increases the MAE which may result due to overfitting. This is briefly mentioned in Line 441-444, but for the farm wide evaluation (Section 4.3) the thrust load is included even though it has been shown to result in overfitting. It is recommended that the authors evaluate and discuss whether the use of estimated thrust load is beneficial to a model that can be used for farm-wide estimates or whether scenario D would not better address the goals of this work.

In addition, the impact of including/excluding thrust load is said to be studied in future work (Line 581). If the inclusion of the thrust load is indeed deemed to be unique to this work, it falls within the scope of this paper and should be addressed thorougly.

Comments:

Line 118: Rotational speed is also referred to as RPM which is confusing as the unit in which rotation is measured is cpm and not RPM. Recommended use of only rotational speed.

Line 119: Unit of ambient temperature is given as ° which does not distinguish between °C and °F.

Line 368: This is the first time that reference is made to the X direction in the text. At that stage it was unclear what the X direction is. It was found that Figure 2 also refers to X and Y, but it might be clearer if the text mentions that X and Y is the measurement directions of the accelerometers (perhaps in Line 121-122). Possibly also show in a figure like Figure 16 how these directions are defined.

Line 388: Y vibration is said to be unnecessary. Is this due to yaw angles close to 0° or 180° for most data used during feature selection? If the yaw angle is close to 90° or 270°, the FA direction would be the same as the Y direction. Would this case not result in the Y SHM measurements to be better correlated to the fatigue? Perhaps the exclusion of Y measurements would not generalize for all conditions of the studied turbines or generalize well for all wind farms.

Line 424: The final scenario is that of the final instrumentation setup with only the selected features used as inputs for the ANN. It is recommended that an additional scenario is included in this section which uses the same instrumentation but no feature selection. This would be beneficial to see the impact of the feature selection on the errors for validation and cross-validation. This would then show the increased performance due to feature selection which is referred to in Line 229.

Line 451: The use of the word ‘clearly’ seems to contradict the initial understanding of Figure 10. This is evidenced by the need of the author to explain the lack of generalization in Line 452. If one were to choose a case that is generally ‘best’ based on Figure 10, it seems that scenario G performs better. Please explain how these results are used to determine the superiority of scenario H.

Line 471: Is the 8000 datapoints each selected randomly or is a period of 8000 consecutive datapoints selected randomly?