Please, see supplement

The topic of the paper is very relevant, and the approach followed—anchored in two large datasets of experimental data collected from two operating wind turbines and exploring many alternative models—is quite valuable.

The objectives of the paper are clearly stated and pertinent. As the authors state at the end of the paper, some of the conclusions might depend on the site and turbine model, but still, the in-depth analysis and the results obtained for two significantly different offshore models represent an important contribution to the current state of the art.

The following points could be improved:

• Section 3.1 – In the strain data processing, it is relevant to mention that before converting strains to stresses, it is crucial to remove the effects of temperature (this is not mentioned in the paper, but there is probably a temperature sensor next to each strain gauge) and any potential strain drift over time (these are more critical in electrical strain gauges; the paper does not specify whether the strain gauges are electrical or fiber-optic sensors).
• Figure 3 – In the strain pre-processing, it would be preferable to convert measured stresses to bending moments (this needs to be explained, since with the use of six measuring points, a fitting procedure should be devised). These could then be oriented in the compass direction or in the FA/SS direction, and from these, the stresses at any point of the cross-section could be obtained. The naming “stresses in FA and SS direction” is misleading—the stresses under analysis are vertical!
• Equation (1) – In design codes, the fatigue of steel elements is calculated using a bi-linear S-N curve with m equal to 3 and 5. This should be commented on.
• Section 3.2 – The following sentence is unclear: “Invalid operational states refer to intervals lacking valid SCADA-derived statistics, typically involving transient events such as rotor start-up or shutdown.” Transient events are not considered in fatigue accumulation? Please clarify this point.
• Section 5, Table 3 – Some candidates for selected variables present strong correlations. Please clarify how this correlation may have influenced the selection of the features.
• Section 5.1 – The most standard variables used for fatigue estimation with SCADA data are wind velocity and turbulence. It would be useful to compare a model just based on these variables with all the others that have been tested.