Unscented Kalman Filter-Aided Long Short-Term Memory Approach for Wind Nowcasting
Round 1
Reviewer 1 Report
Please see attached file for comments and suggestions to authors.
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Attached.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The proposed manuscript has presented a hybrid LSTM based model for wind nowcasting considering its applications in the aviation industry. The authors have introduced the problem statement interestingly and the results presented in the case study are interesting, however, the case study and overall manuscript are not suitable for publication in their current state. The following comments must be considered while revising the manuscript: 1. The motivation behind selecting only LSTM and its hybrid model is not very clear. It is obvious that the performance of LSTM must be better than MLP because of the basic differences in the structure of these methods. Besides, GP regression usually fails to achieve higher accuracy in the stochastic time series. Therefore, it is advised to consider other (e.g., statistical (ARIMA, PSF), hybrid (EEMD, VMD), etc) methods to compare the performance of the proposed hybrid method. 2. Statistical details and characteristics of the used time series are not provided in the manuscript. 3. Comparison of the proposed model with others on the basis of RMSE and R-squared values can not be sufficient. It is advised to consider other error metrics since all of these metrics are helpful in understanding the performance of prediction. The authors may refer to this article for further details: https://www.mdpi.com/1996-1073/13/10/2578 4. The presented case study is for the nowcasting of the wind speed and the process of nowcasting is predicting almost in real-time. In such applications, it is essential to keep track of the delays in the prediction process. Therefore, it is very crucial to compare the time and space complexities of the models. It is advised to compare the complexities of the models used in the study. This comparison also is helpful in understanding at what cost the prediction accuracy achieved with the proposed hybrid model. 5. The performance evaluation based on a single time instance can not be considered valid proof to decide the best model for the given time series. Therefore, it is advised to perform the comparison at different sections of the time series. The authors must consider the cross-validation approach for better comparison evaluation. 6. The resolution of the figures is poor to understand the prediction results of the models. It is advised to enhance them for better visibility. 7. The domain of wind nowcasting has now progressed significantly. So, for a new research contribution, it is essential to describe the state-of-the-art progress in this domain. It is advised to extend the literature review by reviewing recently published review articles in this domain.Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
The topic “Unscented Kalman Filter-Aided Long Short Term Memory Approach for Wind
Nowcasting” is modern and important for readers/scientists.
General comments:
Please add the list of Abbreviations and in the text you can use abbreviations.
Conclusion – this point is too short in my opinion – add 5 – 10 sentences
Detailed comments:
Figure 1 – the source of picture?
Line 128 “Additionally, we decomposed the time-series data into a training phase (i.e., from January to October) and a validation phase (i.e., from November to December) for the holdout validation.” What about testing data set for on final quality measure checking after tuning hiperparameters of proposed models? What is the proportion of training and validation set-write in paper.
Line 130 “the inputs become x1:N-1 and their corresponding outputs are x2:N.” Why you did not tested more sets of input data for example N-2, N-3? Only one back is very simply model.
Figure 4 – in the output layer activation function is nonlinear or linear – I see nonlinear.
Table 1– add information about range of tested hyperparameters
Table 2 – the same situation as above, and dropout – in which layer?, number of nodes in each layer you should add
RMSE error-write equation
Table 3 and Table 4 – LSTM method has different result
Table 4 the results are unbelievably strange!!!!. I have been forecasting for 20 years. It is not possible to reduce the error from 22 RMSE to 3 RMSE - that's almost 800 percent. There must be a mistake!!!. There is no possibility that the hybrid method is so much better than the very good LSTM model.
Page 15: " 32. Lee, K.; Choi, Y.; Johnson, E.N. Kernel Embedding-Based State Estimation for Colored Noise Systems. IEEE/AIAA 36th Digital
Avionics Systems Conference (DASC), 2017, pp. 1–8. doi:10.1109/dasc.2017.8102036.” – 2017 should be bolded, also number 27 in References
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
I appreciate the authors response. Unfortunately, I still have serious concerns concerning the revised version.
1. The authors’ contributions still seem minimal. The authors acknowledge that very similar methods have been applied to very similar problems. In my estimation, the authors contributions are not substantial enough to warrant publication.
2. The authors still do not provide a comparison between their method and state-of-the-art methods. The only methods that the authors compare their method to are the simulated sensor measurements and the LSTM network that does not receive any sensor measurements. There are many existing methods that could be applied to this problem (for example Hur 2021), and the authors need to demonstrate that their method is better than the existing methods.
For these reasons, I still do not recommend publication of this paper.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The authors have improved the article well, especially in the results section. Here are some minor changes that can be made to improve the article.
- Line 78 - Line 89: Authors are encouraged to move this section to Section 3 "3. Data Pre-processing".
- The third section "3. Data Pre-processing" can be given as the initial section in Section 4 "4. Proposed Methodology", and is better called material and method.
- Line 33 and 34. “Wind forecasting has recently been recognized as one of the most challenging tasks in the aviation industry. “Add a reference such as, https://doi.org/10.1016/j.enconman.2021.114002
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The updates in the manuscript and response to the reviewers' comments are satisfactory.
Author Response
Thank you so much for taking the time to review our manuscript. We would like to express our appreciation to you for your invaluable feedback and comments.
Reviewer 4 Report
I read the answer and I am satisfied with the replies and improvements in the article
Author Response
Thank you so much for taking the time to review our manuscript. We would like to express our appreciation to you for your invaluable feedback and comments.
