You are currently viewing a new version of our website. To view the old version click .
MDPIMDPI
Search all articles
by
  • Fei Yan1,* and
  • Huawei Wang2

Reviewer 1: Pampa Sadhukhan Reviewer 2: Gilberto Lechuga Reviewer 3: Anonymous Reviewer 4: Chao Sun

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study proposes a hybrid architecture that integrates state-space models with a dynamic graph attention to predict air traffic flow under several disruptive weather conditions. The results from experiments on the real-world dataset in this study report better performance of the proposed model in prediction accuracy and robustness, compared to other approaches. The study also points out the limitations of the proposed model and discusses its significant applicability beyond air traffic flow prediction. The article is written well. However, several issues related to its technical content and presentation exist, based on which the following suggestions to improve the article further are recommended.                                                                                 

  1. The definitions of parameters n and k used in equation 2 are required in the text following their appearance in subsection 3.2.
  2. Defining all mathematical notations/parameters used herein in a separate table at the beginning of Section 3 would make the proposed method more comprehensible to the readers.
  3. Before introducing graph attention networks in subsection 3.3, the operational principles of GNNs and their relevance to the air traffic systems should be outlined to point out the scientific significance of applying such networks to the latter.  
  4. The reference to Figure 1 is missing in the text. Moreover, neither the operation of the individual steps of the internal flow diagram of the proposed method, shown in Figure 1, is described, nor does the textual description of the proposed method match the given flow diagram. Furthermore, at the beginning of Section 4, a brief introduction to the core components of the proposed architecture should be provided.
  5. Usually, the original dataset is divided into 80:20 to prepare the training and test sets for the ML models. Also, to evaluate the generalizability of the model, testing with unseen data is needed. However, in subsection 5.5, the authors reported that 20% of episodes were used for model training while the remaining were used for testing to evaluate the model’s generalizability. The above statement should be justified with proper reasoning.
  6. Minors:
    1. The abbreviated terms like NOAA and ADS-B should be expanded on their first use.
    2. Definition of parameters N and K used in equations 18, 19 & 20 is needed after their appearance.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

 

  1. Briefly describe what you mean by nonlinear dynamics of basic temporal patterns.

 

  1. Briefly explain what you mean by the computational complexity generated by window sizes and very long sequences.

 

  1. Explain the meaning of the letters n and k in Equation (5).

 

  1. In equations (10) and (11), what is the discretization magnitude of t?

 

  1. Clearly explain the meaning of Equation (15), which is very confusing.

 

 

  1. It is suggested to use the format of an algorithm to describe the steps shown at the beginning of section 4.4.

 

  1. What is the role of the lambda values ​​in equation (16)? Is it a convex combination?

 

  1. Explain why you prefer to use the LeakyReLU function instead of the standard ReLU function in equation (6). Provide more information about the meaning of the parameter a in the same equation.

 

  1. Is it possible to apply the dynamic graph approach to the analysis of vehicle routing systems in general?

 

  1. In equation (7), is node i the origin node and j the destination node? Provide some examples of the forms this function can take.

 

 

  1. In equation (16), specify what you mean by a sequence of length L. Is it time that you are measuring?

 

  1. Is equation (8) an original proposal by the authors?

 

  1. Regarding Fig. (1), improve its quality and clarity and relate it to the text of the document.

 

  1. Expand your explanation regarding the computational complexity of length L, O(NL). Can any relationship be established between O(NL) and O(Nk)?

 

  1. List the acronyms used in your document and explain their function and units of measurement.

 

  1. (4) is confusing and unrelated to the document. Expand on its meaning and describe its relationship to the document's narrative.

 

  1. Expand on your explanation in paragraph 6.4 about O(N) complexity.

 

  1. Would it be possible to include one or more computational runs showing the results of your proposal?

 

  1. Some improvements to the style and spelling of the document should be addressed.

 

  1. Authors should standardize the way they make the corresponding citations according to the journal's guidelines.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper presents a significant advancement in air traffic flow prediction by effectively addressing the dual challenges of long-range temporal modeling and adaptive spatial relationship learning under weather disruptions. The hybrid design of State-DynAttn demonstrates superior performance across various weather scenarios and prediction horizons. With some refinements to address the identified areas for improvement, the proposed architecture has the potential to make a substantial impact in the field of air traffic flow prediction.

 

In the Introduction section, the critical importance of air traffic flow prediction in ensuring aviation safety and efficiency should be more emphasized, which not only helps in optimizing flight schedules and reducing delays but also plays a pivotal role in preventing potential air traffic congestion and ensuring the safety of flight operations, especially under adverse weather conditions.

 

As for the dataset description and data processing, the introduction to the datasets is somewhat fragmented. It is recommended to consolidate it into two subsections: "Dataset Description" and "Data Preprocessing." Detailed explanations of the data sources, record types, and coverage ranges for both the ATWID and OpenSky Network Traffic Dataset should be provided separately. Subsequently, systematically elaborate on preprocessing steps such as temporal alignment, representation of active flight routes, and extraction of weather features to enhance logical coherence.

 

In formula derivation section, the titles of Figures 1 to 4 merely describe the content of the charts. It is suggested to further refine these titles to encapsulate the core conclusions of the charts, such as "Figure 1: Internal Workflow and Module Interactions of the State-DynAttn Model" and "Figure 2: Comparison of Air Traffic Flow Prediction on Network Edges under Normal and Disrupted Conditions (Blue: Normal; Red: Disrupted)" to enable readers to quickly grasp the value of the charts.

 

In Experiment and Results section, Table 1 should explain more, "As shown in Table 1, the MAE value of State-DynAttn in 6-hour traffic flow prediction is lower than that of other models, indicating higher prediction accuracy." The analysis of results could delve deeper, such as by discussing, in conjunction with Figure 3, the adaptability of the model's dynamic attention mechanism to weather changes during peak disruption periods.

 

The formatting of references in the paper should be unified, which requires careful verification and correction.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This paper proposes a hybrid architecture that combines the state space model (SSM) with the dynamic graph attention mechanism for robust air traffic flow prediction under weather interference. However, some problems need to be addressed in the following:

  1. The SSM branch assumes that the state evolves smoothly and responds slowly to sudden events, which may lead to prediction lag. An event-triggered mechanism can be introduced. For example, when a sudden change in weather or a sharp drop in traffic is detected, the weight of the attention branch can be temporarily increased.
  2. Top-k pruning and weather thresholds may accidentally remove long-range or low-frequency but crucial spatial dependencies, which will affect the experimental results. Causal discovery methods (such as the PC algorithm) can be combined to pre-identify key paths, thus avoiding blind pruning and accidentally removing crucial spatial dependencies.
  3. The model did not observe an obvious process of performance degradation under extreme weather conditions (such as volcanic ash) in the training data. It is recommended to introduce adversarial training or data augmentation (such as synthesizing extreme weather scenarios) to improve the generalization ability and enhance the reasoning ability for unknown weather patterns.
  4. It is recommended to improve the quality and clarity of the pictures in the paper. For example, in Figure 3, the legend blocks an important curve. It is suggested to adjust the position of the legend to avoid affecting the reading.
  5. The experiments in the paper were only conducted on historical data and did not address practical challenges such as real-time data streams, latency, and system integration. It is recommended to conduct real-time simulation tests to verify the feasibility of the method.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Well the required adjustments.

Reviewer 4 Report

Comments and Suggestions for Authors

The authors have already answered the inquiries and carefully revised themanuscript based on the reviewers' comments. In my opinion, the revised manuscript should bepublished after checking the template.