See pdf

This is an interesting, well-written, concise paper that presents field validation of neural network-based wind shear and veer observers using blade root bending moment measurements. The study builds on previous work by some of the authors that discussed a neural network-based observer for vertical shear and yaw misalignment. The current study is a valuable contribution to the literature because it presents the first field validation of a veer observer known to the authors and validates the shear and veer observers using reference wind measurements across the full height of the rotor using a ground-based vertically profiling lidar.

Although I don’t have any major concerns about the paper, there are many places where additional details or clarification should be provided, as explained in the comments below.

Specific comments:

1. Section 2.1: A few more details about the neural network-based shear and veer observers would be helpful for readers, especially non-experts. In particular, after equation 1, can you explain what "p" represents? It is explained a couple paragraphs later, but it would be better to mention the variable as soon as it is shown in Eq. 1. Some discussion or references about "sigmoid activation functions" and the "backpropagation" training method should be provided for those unfamiliar with neural networks. Lastly, how is the number of hidden neurons M hyperparameter chosen? Is this optimized as part of the observer identification (could also be discussed in Section 3.1)?
2. Ln. 91: "30 sec moving average of the rotor-effective wind speed": I suggest adding "estimated" rotor-effective wind speed.
3. Figure 3: I suggest labeling the lidar measurement heights in the figure as belonging to the "lidar", or mentioning this in the caption, to make it clear what the heights refer to. 
4. Section 2.3: Why don’t you use the more common power law shear exponent definition of vertical wind shear in the estimator instead of the linear shear? This would likely be a more useful term for many applications of the observer. I understand that the blade load harmonics are more easily connected to linear shear across the rotor, but since a neural network is used, it seems the NN could relatively easily be trained to estimate the equivalent power law shear as well.
5. Ln. 149: Fig. 4a lists 0.122 as the MAE, but the text states 1.22 m/s. Is there a typo in the text?
6. Section 3.1: Please clarify what data sample period is used for the observer identification. In Section 3.2, it appears that the observers use 1 Hz input data. Is this the same frequency used for training the observer? Do you expect the trained observers can be applied to data with a range of sampling frequencies, or should they only be used with data of the same frequency as the training data?
7. Ln. 162: "of which about 67%... were used for training": Can you explain how the training data points were selected? Are they randomly distributed throughout the data set, or are the first 67% of the data used for training and the last 33% used for validation? 
8. Fig. 4 caption: Please mention that 10-minute averages are shown in this figure.
9. Ln. 172: "using the aerodynamic torque obtained from the dynamic torque-balance equation, and on measured power, pitch and rotor speed from the SCADA data stream." The dynamic torque-balance equation generally requires the measured generator torque and generator acceleration. Should these measurements be added to the list here?
10. Ln. 194: "The quality of these results seems to be more than capable of supporting applications such as the one described in Sucameli et al.": I don’t think this is incorrect necessarily, but it's kind of an arbitrary statement. Can you explain this a little more? What kind of accuracy/quality are you looking for in the estimates for them to be considered capable?
11. Ln. 201: "Interestingly, both the observed shear and veer appear to be rather insensitive to TI and wind speed, only the shear error exhibiting a growing trend for low wind speeds.": This statement doesn't quite seem correct. At most higher wind speeds (>= 8 m/s), there is clearly higher error for higher turbulence levels in Fig. 9. Further, both shear and veer errors appear to exhibit a growing trend at low wind speeds.
12. Ln. 217: "the ability to follow rapid changes of the duration of the order of tens of minutes": Should this be tens of seconds? Depending on the application, it doesn’t seem correct to consider tens of minutes as rapid. Or can you discuss this further?   
13. Ln. 221: "there is a good robustness with respect to TI, which has only minor effects on the quality of the results.": As mentioned in comment 11, there is significantly larger error for higher TI at some higher wind speeds.
14. Ln. 223: "some degradation appearing only for veer at the lowest wind speeds.": As also mentioned in comment 11, based on Fig. 9 it seems this is the case for shear as well.
15. Conclusion section: Can you comment on whether you expect the trained observer can be applied to other turbines of the same model? Or would the training procedure need to be applied to each individual turbine?
16. Ln. 242: How is the blade flap angle beta defined? Further, what is the sign convention of the azimuthal blade position and cross-flow velocity? A figure would be helpful here.
17. Eq. A2: Can you provide a reference for this equation? Also, what is "u_p"? Should this be "u_n"? Lastly, it would be good to clearly define rho as the air density as well.
18. Ln. 250: "Inserting the second expression of Eq. (A2) into Eq. (A1)…": It isn’t clear where this would be inserted in Eq. A1. Please state where lift appears in Eq. A1. Or do you mean inserting the second expression of A1a into A2?
19. Ln. 251: "shear leaves a mark on the 1P harmonics of blade loads, and veer on their 2P harmonics". Please describe why the veer observer also uses the 1P harmonics as inputs, if veer appears to only depend on 2P harmonics as shown here.