Attention-Based Distributed Deep Learning Model for Air Quality Forecasting

Round 1

Reviewer 1 Report

The authors have addressed an attention-based convolutional Bi-LSTM autoencoder model for air quality prediction. Some concerns are listed below:

1) The differences and similarities between this work and reference [15] should be clearly stated.

2) Some parts of this work are similar to the following references; while not mentioned:

• Wang, Z. Yan, J. Lu, G. Zhang, and T. Li, "Deep multi-task learning for air quality prediction," in International Conference on Neural Information Processing, 2018: Springer, pp. 93-103.
• Xiao, M. Yang, H. Fan, G. Fan, and M. A. Al-Qaness, "An improved deep learning model for predicting daily PM2. 5 concentration," Scientific Reports, vol. 10, no. 1, pp. 1-11, 2020.

3) Literature review is not complete. Many related works such as the following have not been reviewed:

• Dairi, F. Harrou, S. Khadraoui, and Y. Sun, "Integrated multiple directed attention-based deep learning for improved air pollution forecasting," IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1-15, 2021.
• Heydari, M. M. Nezhad, D. A. Garcia, F. Keynia, and L. De Santoli, "Air pollution forecasting application based on deep learning model and optimization algorithm," Clean Technologies and Environmental Policy, pp. 1-15, 2021.
• -S. Chang, H.-T. Chiao, S. Abimannan, Y.-P. Huang, Y.-T. Tsai, and K.-M. Lin, "An LSTM-based aggregated model for air pollution forecasting," Atmospheric Pollution Research, vol. 11, no. 8, pp. 1451-1463, 2020.
• Arsov et al., "Multi-Horizon Air Pollution Forecasting with Deep Neural Networks," Sensors, vol. 21, no. 4, p. 1235, 2021.
• Guo, G. Liu, and C.-H. Chen, "Air Pollution Concentration Forecast Method Based on the Deep Ensemble Neural Network," Wireless Communications and Mobile Computing, vol. 2020, 2020.
• …

4) “When the YOLOv5 model detects the object, the model cannot define the same objects over the different frames.”. What is the reason for this?

5) What is the reason for the relatively sharp changes in the rate error versus the number of nodes in Figure 17?

6) No reference is provided for any of the equations; while most of them need reference.

7) Figures 3, 11 and 20 lack visual quality.

8) Some writing mistakes:

• Page 2: two fundamental drawbacks: (1) A lack of precision. (2) Time and energy → two fundamental drawbacks: (1) a lack of precision; (2) time and energy
• Page 3: As we can see → As can be seen
• Page 6: “In the presented equations, …”, A new paragraph should not be started. Similarly after Eqs. (5) and (6), and so on.
• Please improve the whole text in terms of writing.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper proposes the application of an attention-based convolutional Bi-LSTM autoencoder model. Evidently, the proposed deep learning model has been trained by using a distributed framework called data parallelism to forecast the intensity of particle pollution. The topic is important. However, some problems must be handled.

• The abstract should be more concise with main findings and contribution.
• The references are limited. Some important works on improved LSTM should be analyzed briefly to the references more comprehensive (https://doi.org/10.3390/app112411820). There is the similar study on the LSTM with attention (https://doi.org/10.1016/j.energy.2021.121756). Please analyze the difference to highlight the contribution.
• The conclusion part: The main findings should be further highlighted. A discussion of the advantage and underlying drawbacks of the proposed method is welcome in the conclusion part.
• Generally speaking, how to set the parameters of the proposed hybrid LSTM for better performance?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Some previous comments are not well addressed (the following):

1) Comments 4 and 5, although answered, should also appear in the manuscript.

2) Some figures still do not have quality.

3) Some formulas still need reference.

4) There are still some writing issues.

5) The references mentioned in the previous round, even if they are different works, should at least be used as relevant references.

As a piece of advice to the authors, highlighting the revised version in its current form is not legible at all. Suppose you are the reviewer and you are going to check it. This way of putting the old and new text together, and the continuous underlining of the sentences really makes the job difficult to review. Please make the revision version more friendly.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Some of my comments were handled. However, some problems must be further handled.

1. It is very hard to read this manuscript with lots of marks. It is better to provide a revision with colored text.
2. The way to set the parameters of the proposed hybrid LSTM is not general. The technology contribution is weak.
3. The experiments must be strengthened. The authors should compare their work with more popular approaches (LSTM with attention) from recent papers (2019-) in important journals in this field.
4. The literature review is a little far from being accepted. In Section 1, it is better to show the existing works with the used technologies and results in a table.
5. Section 3.3: Try to further analyze the results from the theory aspect.
6. The writing can be improved thoroughly with the help of a native English speaker.
7. Some figures (9, 19, etc.) are ugly.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 3

Reviewer 2 Report

No further comments