The article presents a method of developing LSTM models for predicting blade response using multi-stage modelling and multi-task learning with dimensionality reduction techniques. Overall, the article is well-written, presenting a novel and relevant contribution in data driven methods. There are several minor aspects of the article which could be modified to improve its quality

1. Article can be made more concise
   • Consider shortening section 2 – not necessary to explain how TurbSim works, possible to reorganise by removing section 2 entirely and including the important parts as part of section 6
   • Sufficient to simply state that IEA15MW reference turbine is used – Figure 2 and Table 1 not necessary
     
     
     
     
2. In Section 1 or 5 – should include a justification on why LSTM (or RNNs in general) is used in this work, as opposed to conventional Neural Networks.
3. Table 3 – Optimized value for Fully connected layer is at upper limit – why is range not expanded to ensure it is actually optimal?
4. Table 5 – Between RMSE and signals shown in Figures 14 to 17, it is clear that LSTM model is accurate. However, RMSE may not necessarily be the best metric to measure model performance, as it does not take into account the variation of the signal (e.g. oop deflection has much larger magnitudes than ip deflection). Consider the use other metrics such as Variance Accounted For or Confidence Index.
5. Figure 10 – Unclear on what this figure represents – some additional explanation in caption may be helpful
6. Example LSTM predictions (Figures 9,12,14 to 17) show results at TI=11.54 – however, range from Table 2 shows that max TI should be 7.629.
7. Example LSTM predictions (Figures 9,12,14 to 17) only show U = 13.65m/s upwards and TI=6.19 to 11.54 – may be interesting to see predictions at lower speeds or higher TI (if possible)

The article presents a method to a) a simple dynamic model of wind turbine b) unsteady wind order reduction and c) developing a LSTM model to predict the blade deformations from the wind time series reduced space. Although this work is interesting, and clearly, the authors put a lot of effort into it, it needs more work to be publishable. I commented on the article extensively and in detail, which you can find in the supplement. These are my general comments:

• The structure of the paper can be improved. It doesn't necessarily follow the chronological order. I think there needs to be a clearer separation between methodology and results. For example, Fig 4 can be in section 8.1.2, and the explanation about DCT needs to move the part where you explain DCT.
• The article is well-written, but it could be more concise. Some sentences are long and written in passive voice, making them hard to follow.
• The quality of the figures can be improved. Figures with time series data are small and not legible. The flowcharts can have a better graphical quality.
• It is necessary to show how the LSTM results match OpenFAST and how much data is lost when reconstructing your wind time series from the reduced space (PCA or DCT).
• The DCT part needs to be explained better. What do you do with your time when you are going through DCT? Section 8.1.2 makes this a bit clearer, but it needs to be explained earlier.