This document is downloaded from DR‑NTU (https://dr.ntu.edu.

sg)
Nanyang Technological University, Singapore.

Data‑driven network modeling for airspace

optimization through airway re‑structurization
and traffic flow prediction

Ma, Chunyao

2022

Ma, C. (2022). Data‑driven network modeling for airspace optimization through airway
re‑structurization and traffic flow prediction. Doctoral thesis, Nanyang Technological
University, Singapore. https://hdl.handle.net/10356/164436

https://hdl.handle.net/10356/164436

https://doi.org/10.32657/10356/164436

This work is licensed under a Creative Commons Attribution‑NonCommercial 4.0

International License (CC BY‑NC 4.0).

Downloaded on 18 Apr 2025 01:16:21 SGT

 DATA-DRIVEN NETWORK MODELING FOR

AIRSPACE OPTIMIZATION

THROUGH AIRWAY RE-STRUCTURIZATION

AND TRAFFIC FLOW PREDICTION

MA CHUNYAO
SCHOOL OF MECHANICAL AND AEROSPACE ENGINEERING
2022
DATA-DRIVEN NETWORK MODELING FOR

AIRSPACE OPTIMIZATION

THROUGH AIRWAY RE-STRUCTURIZATION

AND TRAFFIC FLOW PREDICTION

MA CHUNYAO

School of Mechanical and Aerospace Engineering

A thesis submitted to the Nanyang Technological University

in partial fulfilment of the requirement for the degree of

Doctor of Philosophy

2022

II
 Statement of Originality

I hereby certify that the work embodied in this thesis is the result of original
research, is free of plagiarised materials, and has not been submitted for a higher
degree to any other University or Institution.

III
 Supervisor Declaration Statement

I have reviewed the content and presentation style of this thesis and declare it is free
of plagiarism and of sufficient grammatical clarity to be examined. To the best of my
knowledge, the research and writing are those of the candidate except as acknowledged
in the Author Attribution Statement. I confirm that the investigations were conducted
in accord with the ethics policies and integrity standards of Nanyang Technological
University and that the research data are presented honestly and without prejudice.

IV
 Authorship Attribution Statement

This thesis contains material from six papers published in the following peer-
reviewed journals / from papers accepted at conferences in which I am listed as an
author.
Chapter 3 is published as: 1) Chunyao Ma, Sameer Alam, Qing Cai, and Daniel
Delahaye, “Critical Links Detection in Spatial-Temporal Airway Networks Using Com-
plex Network Theories,” IEEE Access, vol. 10, pp. 27925-27944, 2022.
2) Chunyao Ma, Qing Cai, Sameer Alam, and Vu N. Duong. “A Percolation Theory
Based Approach for Identification of Bottleneck Links in an Airway Network,” In 9th
International Conference on Research in Air Transportation, Online, 15 September,
2020.
The contributions of the co-authors are as follows:

• Professor Sameer Alam provided the initial project direction and edited the
manuscript drafts.

• I conducted the literature survey, developed the research methodology, carried

out the experiment, analyzed research results, and prepared the manuscript
drafts.

• Dr Qing Cai assisted me in understanding the network theories which are the
basis for the methodology development in the study.

• Professor Daniel Delahaye and Dr Qing Cai revised the manuscript drafts.

• Professor Sameer Alam provided weekly feedback to the research included in the
paper.

Chapter 4 is published as: 1) Chunyao Ma, Qing Cai, Sameer Alam, Banavar
Sridhar, and Vu N. Duong, “Airway network management using Braess’s Paradox,”
Transportation Research Part C: Emerging Technologies, vol. 105, pp. 565-579, 2019.

V
 2) Qing Cai, Chunyao Ma, Sameer Alam, Vu N. Duong, and Banavar Sridhar.
“Airway Network Flow Management using Braess‘s Paradox,” In 13th USA/Europe
Air Traffic Management Research and Development Seminar, Vienna, Austria, 17-21
June, 2019.
The contributions of the co-authors are as follows:

• Professor Sameer Alam provided the initial project direction and edited the
manuscript drafts.

• I conducted the literature survey and prepared the manuscript drafts.

• Dr Qing Cai and I designed the methodology and conducted the case study based
on the proposed methodology.

• Professor Vu N. Duong revised the manuscript drafts.

• Dr Banavar Sridhar provided suggestions on the research methodology and re-

vised the manuscript drafts.

• Professor Sameer Alam provided weekly feedback to the research included in the
paper.

Chapter 5 is published as: Chunyao Ma, Qing Cai, Sameer Alam, and Vu N.
Duong, “Airspace capacity overload identification using collision risk patterns,” in
2020 International Conference on Artificial Intelligence and Data Analytics for Air
Transportation (AIDA-AT), Singapore, IEEE, 3-4 February, 2020, pp. 1–9.

• Professor Sameer Alam provided the initial project direction and edited the
manuscript drafts.

• I conducted the literature survey, developed the research methodology, carried

out the experiment, analyzed research results, and prepared the manuscript
drafts.

• Professor Vu N. Duong revised the manuscript drafts.

VI
 • Dr Qing Cai provided suggestions on the research methodology and revised the
manuscript drafts.

• Professor Sameer Alam provided weekly feedback to the research included in the
paper.

Chapter 6 is published as: Chunyao Ma, Sameer Alam, Qing Cai, and Daniel
Delahaye, “Sector Entry Flow Prediction Based on Graph Convolutional Networks,”
In International Conference on Research in Air Transportation, Tampa, FL, USA,
19-23 June, 2022.

• Professor Sameer Alam and I developed the initial research problem.

• I conducted the literature survey, developed the research methodology, carried

out the experiment, analyzed research results, and prepared the manuscript
drafts.

• Professor Sameer Alam and Dr Qing Cai revised the manuscript drafts.

• Professor Daniel Delahaye provided suggestions on the research methodology

and revised the manuscript drafts.

• Professor Sameer Alam provided weekly feedback to the research progress in-
cluded in the paper.

VII
 Acknowledgement

First and foremost, I would like to express my sincere gratitude to my supervisor,

Associate Professor Sameer Alam, from the School of Mechanical and Aerospace En-
gineering (MAE), Nanyang Technological University (NTU). In my four-year Ph.D.
research journey in air traffic management, he has shown me the path to conducting
innovative research and built my interest and confidence in this research area. His
assiduous research ethics has constantly reminded me of doing meaningful works that
can make a difference in this domain and even in people’s lifestyles. His meticulosity
in the details has inspired me to refine my research methodologies with maximum
effort, write research papers with heartful consideration of the readers, and apply the
“always try my best” altitude to every tiny thing in my work and life. Moreover, his
dedication to helping people under difficult circumstances and his sincerity in tend-
ing to the well-being of his teammates nurture me to become a person with love and
care. Furthermore, his willingness to face challenges has motivated me to be brave
and strong under stress and predicaments. During the four years of my Ph.D., I am
fortunate to have Professor Alam as my supervisor. He has helped me become the
person I would always like to be, both as a researcher and a human.
I would also like to convey my sincerest appreciation for my Ph.D. mentor Dr. Cai
who is a research fellow at the Air Traffic Management Research Institute (ATMRI),
School of Mechanical and Aerospace Engineering (MAE), Nanyang Technological Uni-
versity (NTU). His continuous guidance and invaluable suggestions were essential in
developing my raw research ideas. He has been working closely with me to improve
my coding, reading, writing, and scientific thinking skills. He is not just an excellent
mentor but also my good friend who I can always talk when I am facing depression
and confusion. He is a dedicated person to research whom you can always find sitting
in front of the laptop screen. His work passion and positive attitude inspire me to

VIII
heartfully love and enjoy research. I value the research skills that I have learned from
him and the friendship we have built over the years. I wish him happiness in his whole
life, which he truly deserves.

I want to express my heartfelt thanks to my friend Dr. Goh Sim Kuan, who is
now a faculty in Xiamen University Malaysia. He is a very kind fellow who always
talks softly and says yes to friends’ requests. During the COVID 19 period, when it
was difficult for us to visit families and friends overseas, we formed a small hang-out
team with Dr. Cai to spend our spare time visiting interesting places and eating nice
food together in Singapore. His friendship made Singapore a warm city and helped
me overcome the depression of being far from families. I wish him a bright career
in research and a happy life with his family. Now that the world has recovered from
COVID19 and travel restrictions are being lifted, I hope we will be able to reunite in
Singapore soon!

I would also like to thank my friend Dr. Tran Ngoc Phu, who started a new research
career in the U.S. a few months ago. He offered me a lot of help when I struggled
with package installations in Python. Many times when I felt lost on this journey, his
wise words brought me energy and direction. He is a talented researcher and a caring
friend. I hope he will enjoy life with his family in the U.S.

Finally, I would like to thank my friends in ATMRI, Mr. Yash Guleria, Ms. Yixi
Zeng, Ms. Lim Zhijun, Mr. Ang Hao Jie, Dr. Hasnain Ali, Dr Duc-Thinh Pham,
Mr. Tan Tranh Nam, and Mr. Nguyen Duy Ahn. They have been integral parts of
this journey. Their participation has made this journey filled with joy and cheers. I
hope they will enjoy research and complete their Ph.D. program soon. I value their
friendship deeply, and I feel very lucky to have them as my colleagues and my dear
friends.

This research is supported by the National Research Foundation, Singapore, and

the Civil Aviation Authority of Singapore, under the Aviation Transformation Pro-

IX
gramme. Any opinions, findings and conclusions or recommendations expressed in
this material are those of the author(s) and do not reflect the views of National Re-
search Foundation, Singapore and the Civil Aviation Authority of Singapore.

X
Contents

Chapter 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Air Traffic Growth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Research Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2.1 Airspace Structure Optimization . . . . . . . . . . . . . . . . . 7

1.2.2 Airspace Capacity Overload Identification . . . . . . . . . . . . 9

1.2.3 Air Traffic Flow Prediction . . . . . . . . . . . . . . . . . . . . . 10

1.3 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.4 Data-Driven Methodology . . . . . . . . . . . . . . . . . . . . . . . . . 11

1.4.1 Network Science in Airway Network Optimization . . . . . . . . 12

1.4.2 Machine Learning in Airspace Optimization . . . . . . . . . . . 14

1.5 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

Chapter 2 Related Literature & Research Questions . . . . . . . . . 18

2.1 Airspace Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.1.1 Airspace Sectorisation . . . . . . . . . . . . . . . . . . . . . . . 18

2.1.2 Airway Network Management . . . . . . . . . . . . . . . . . . . 21

2.1.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.1.4 Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.2 Airspace Capacity Overload Identification . . . . . . . . . . . . . . . . 25

2.2.1 Quantitative Methods: Airspace Capacity Estimation . . . . . . 25

2.2.1.1 Controller Workload Based Analytical Models . . . . . 25

2.2.1.2 Machine Learning Based Models . . . . . . . . . . . . 26

2.2.2 Qualitative Methods: Other Capacity Overload Indicators . . . 27

2.2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.2.4 Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.3 Air Traffic Flow Prediction . . . . . . . . . . . . . . . . . . . . . . . . . 29

2.3.1 Trajectory-based Flow Prediction Methods . . . . . . . . . . . . 30

XI
 2.3.2 Aggregate Flow Prediction Methods . . . . . . . . . . . . . . . . 30

2.3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.3.4 Gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.4 Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

Chapter 3 Airway Network Critical Link Identification . . . . . . . 36

3.1 Related Backgrounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3.1.1 Spatial-Temporal Networks . . . . . . . . . . . . . . . . . . . . 37

3.1.2 Network Centrality . . . . . . . . . . . . . . . . . . . . . . . . . 38

3.1.3 Percolation Theory . . . . . . . . . . . . . . . . . . . . . . . . . 38

3.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3.3 Research Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3.3.1 Methodology Overview . . . . . . . . . . . . . . . . . . . . . . . 40

3.3.2 Network Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 41

3.3.3 Critical Link Detection . . . . . . . . . . . . . . . . . . . . . . . 43

3.3.3.1 Centrality Method . . . . . . . . . . . . . . . . . . . . 43

3.3.3.2 Percolation Based Method . . . . . . . . . . . . . . . . 44

3.3.4 Critical Link Determination . . . . . . . . . . . . . . . . . . . . 46

3.4 Experimental Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

3.4.1 Network Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 47

3.4.2 Critical Link Detection . . . . . . . . . . . . . . . . . . . . . . . 48

3.4.2.1 Variations of the Number of Critical Links . . . . . . . 50

3.4.2.2 Spatial Distribution of the Detected Critical Links . . 53

3.4.3 Critical Link Determination . . . . . . . . . . . . . . . . . . . . 55

3.4.3.1 Number of the Determined Critical Links . . . . . . . 55

3.4.3.2 Temporal Distribution of the Determined Critical Links 55

3.4.3.3 Spatial Distribution of Determined Critical Links . . . 55

3.4.4 Validation of the Identified Critical Links . . . . . . . . . . . . . 57

XII
 3.4.4.1 Air Traffic Volume Perspective . . . . . . . . . . . . . 58

3.4.4.2 Airspace Design Perspective . . . . . . . . . . . . . . . 58

3.4.4.3 Operational Perspective . . . . . . . . . . . . . . . . . 59

3.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

Chapter 4 Airway Network Structure Management . . . . . . . . . 71

4.1 Braess’s Paradox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

4.2 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

4.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

4.3.1 Methodology Overview . . . . . . . . . . . . . . . . . . . . . . . 78

4.3.2 Trajectory Registration . . . . . . . . . . . . . . . . . . . . . . . 79

4.3.3 Cost-Function Formulation . . . . . . . . . . . . . . . . . . . . . 81

4.3.4 Origin-Destination (OD) Pair Extraction . . . . . . . . . . . . . 82

4.3.5 BP Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

4.3.5.1 Minimization of Objective Functions . . . . . . . . . . 83

4.3.5.2 BP Links Identification . . . . . . . . . . . . . . . . . 86

4.4 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

4.4.1 Cost Function Modeling . . . . . . . . . . . . . . . . . . . . . . 87

4.4.2 BP Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

4.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

Chapter 5 Airspace Capacity Overload Identification . . . . . . . . 98

5.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

5.1.1 Methodology Overview . . . . . . . . . . . . . . . . . . . . . . . 100

5.1.2 Collision Risk Modeling . . . . . . . . . . . . . . . . . . . . . . 101

5.1.3 Collision Risk Heatmap Generation . . . . . . . . . . . . . . . . 104

5.1.4 Collision Risk Pattern Recognition . . . . . . . . . . . . . . . . 106

5.2 Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

5.2.1 Collision Risk Computation . . . . . . . . . . . . . . . . . . . . 109

XIII
 5.2.2 Heatmap Generation . . . . . . . . . . . . . . . . . . . . . . . . 109
5.2.3 Collision Risk Pattern Recognition . . . . . . . . . . . . . . . . 110
5.2.4 Cluster Parameter Setting for Experiment . . . . . . . . . . . . 110
5.2.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
5.3 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Chapter 6 Short-term Air Traffic Flow Prediction . . . . . . . . . . 115
6.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.2.1 Methodology Overview . . . . . . . . . . . . . . . . . . . . . . . 117
6.2.2 Data Denoising . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.2.3 Spatial-temporal Graph Construction . . . . . . . . . . . . . . . 121
6.2.3.1 Nodes Extraction . . . . . . . . . . . . . . . . . . . . . 121
6.2.3.2 Edges Generation . . . . . . . . . . . . . . . . . . . . . 122
6.2.3.3 Graph Construction . . . . . . . . . . . . . . . . . . . 123
6.2.4 Sector Entry Flow Prediction . . . . . . . . . . . . . . . . . . . 123
6.3 Experimental Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.3.1 Graph Construction . . . . . . . . . . . . . . . . . . . . . . . . 125
6.3.2 Sector Entry Flow Prediction . . . . . . . . . . . . . . . . . . . 126
6.3.3 Result Analysis and Comparison . . . . . . . . . . . . . . . . . 127
6.4 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Chapter 7 Conclusions and Future Work . . . . . . . . . . . . . . . . 132
7.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

XIV
List of Figures

1 Research simulation indicates that the number of conflicts in the airspace

increases exponentially with the number of aircraft in the airspace [1]. 1

2 Overview of the future long-term European traffic (IFR movements) [2]. 2

3 Annual growth in global air traffic passenger demand between 2006 and
2022 [3]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

4 Gate to Gate phases of a typical flight, with the cruise phase dominating
the typical flight time [4]. . . . . . . . . . . . . . . . . . . . . . . . . . 4

5 A visual illustration of the research motivation. . . . . . . . . . . . . . 5

6 Airway network structure in Singapore FIR. . . . . . . . . . . . . . . . 8

7 Concept diagram of the data-driven research framework in this thesis. . 13

8 Overview of the thesis structure. . . . . . . . . . . . . . . . . . . . . . . 16

9 Phylogenetic tree of the investigated research problems in this thesis. . 34

10 Concept diagram of the proposed method for critical link detection in

airway networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

11 An example of the graph percolation process on the SEAN. . . . . . . . 45

12 The network structure (a) and a weighted graph representation (b) of

the SEAN. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

13 Temporal distributions of the number of critical links detected using

the network percolation theory. . . . . . . . . . . . . . . . . . . . . . . 50

14 Spatial distributions of the detected critical links using the betweenness

centrality and percolation theory. . . . . . . . . . . . . . . . . . . . . . 52

15 The distribution of the number of final critical links determined in the

SEAN. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

16 Critical links identified over time under different time slot lengths. . . . 56

17 Spatial distributions of the critical links in the SEAN. . . . . . . . . . . 57

18 Geographical positions of the identified critical links. . . . . . . . . . . 60

XV
19 Eastbound flight routes in the SEAN passing through the identified
critical links. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

20 Westbound flight routes in the SEAN passing through the identified

critical links. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

21 Visualization and simulation of air traffic data for SMEs’ analysis. . . . 66

22 Traffic scenario examples under which “MABAL – VISAT” or “RAXIM

– VMR” is identified as a critical link. . . . . . . . . . . . . . . . . . . . 67

23 A graphical network example presented in [5] to illustrate the BP phe-

nomenon. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

24 The layered en-route structure airspace network from ADS-B data used
in the study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

25 Concept diagram of the proposed method for BP detection in a given

airway network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

26 Registration of trajectories onto airways. . . . . . . . . . . . . . . . . . 81

27 An example of the data-driven cost function formulation for one link of

an airway network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

28 Airway structure of SAN. . . . . . . . . . . . . . . . . . . . . . . . . . 88

29 Cost functions modeled from the traffic data for two representative links
in SAN. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

30 Assessment of the data-driven based cost function modeling using MSE,

RMSE, and R2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

31 Air traffic demand over 180 days in SAN at flight level 330. . . . . . . . 90

32 Comparisons of the total travel time costs occurred in SAN under UE,
SE, and BE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

33 The optimal flows on each link in SAN with respect to UE, SE, and BE
on day 175. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

34 BP detection results on day 175 for SAN. . . . . . . . . . . . . . . . . . 93

XVI
35 Conceptual diagram of the proposed method for airspace capacity over-
load identification. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
36 The heatmap generated in an airspace sector with air traffic data during
20 minutes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
37 Determination of the number of clusters: average SSE and minimum
SSE are plotted versus the number of clusters. . . . . . . . . . . . . . . 110
38 The twelve recognized collision risk patterns (p1-p12). . . . . . . . . . . 111
39 Collision risk versus aircraft count under different patterns. . . . . . . . 111
40 Collision risk versus aircraft count under low, medium, and high colli-
sion risk patterns. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
41 Visualization of one-day flight trajectories in French airspace. . . . . . . 116
42 Conceptual diagram of the proposed method for short-term sector entry
flow prediction based on graph convolutional networks. . . . . . . . . . 118
43 Illustration of the proposed ADS-B data denoising approach. . . . . . . 120
44 An example of nodes extraction from three flight trajectories. . . . . . 121
45 The constructed graph for French airspace. . . . . . . . . . . . . . . . . 126
46 Flow prediction result on node “LFBBP2-LFBBP1-LFBBP3” from 00:00
to 23:59 on 30th December 2019 (UTC). . . . . . . . . . . . . . . . . . 127
47 Prediction performance (MAE, MSE and RMSE) of the proposed GCN-
based method and the LSTM model versus the prediction window. . . . 128
48 Cross-correlation between the true values and the lagged (shifted) copies
of predicted values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

XVII
List of Tables

1 The ratio of flights (r) on each of the five high-frequency links being
detected as critical links under different time slot lengths. . . . . . . . 58
2 Characteristic information of the identified critical links. . . . . . . . . 62
3 Statistical results for BP detection in SAN at flight level 330. . . . . . . 94
4 Performance comparison between the proposed GCN-based method and
the LSTM model using three metrics: MAE, MSE and RMSE. . . . . . 128

XVIII
Acronyms

ADS-B Automatic Dependent Surveillance-Broadcast. 12

ANSP Air Navigation Service Provider. 1

AOC Airlines Operations Center. 76

ASE Altimetry System Error. 101

ASTGCN Attention-based Spatial-Temporal Graph Convolutional Network. 116

ATC Air Traffic Control. 6

ATCC Air Traffic Control Center. 4

ATCO Air Traffic Controller. 3

ATFM Air Traffic Flow Management. 4

ATM Air Traffic Management. 15

ATS Air Traffic Service. 37

BE Braess’s Equilibrium. 77

BP Braess’s Paradox. 71

CNN Convolutional Neural Network. 31

CPA Closest Point of Approach. 102

FIR Flight Information Region. 7

FRA Free Route Airspace. 5

FTE Flight Technical Error. 101

XIX
GCN Graph Convolutional Network. 128

IFR Instrumental Flight Rule. 2

LDSM Linear Dynamic System Model. 31

LOP Links of Paradox. 79

LSTM Long Short-Term Memory. 15

MAP Monitor Alert Parameter. 26

OD Origin-Destination. XIII, 82

SAN Singapore Airway Network. 78

SE System Equilibrium. 75

SME Subject Matter Expert. 60

TLS Target Level of Safety. 102

UE User Equilibrium. 75

XX
Abstract

The increasing air traffic demand is leading to airspace congestion which not
only elicits traffic delays causing substantial economic losses but may also exceed
the airspace capacity causing potential safety and inefficiency issues. Accommodating
the future air traffic demand becomes a major challenge confronted by air navigation
service providers (ANSPs). This problem is further compounded due to air traffic’s
constantly evolving nature, owing to weather disturbances, frequent flight rerouting,
Special Use Airspace (SUA) management, and tactical Air Traffic Flow Management
(ATFM). However, according to a consensus in the aviation community, the pro-
jected increase in traffic demand can be accommodated through improved and effec-
tive airspace management, which urges research to optimize the airspace structure and
flow for better air traffic management.

In the literature, research efforts on airspace optimization mainly focus on sector

structure optimization and air traffic flow optimization. Sector structure optimiza-
tion usually leads to a redesign of the entire airspace structure, which is difficult to
implement due to operational constraints, while the effectiveness of air traffic flow
optimization is limited by the constantly changing nature of air traffic due to airspace
capacity constraints, weather, and varying traffic demand. Therefore, it is essential
to develop innovative approaches that can effectively manage the airspace by mak-
ing minimal changes to the underlying network structure and can provide timely and
accurate predictions about traffic flow conditions for effective flow management. In
view of this, this thesis aims to seek answers to the following research questions which
can contribute to airspace optimization: 1) how to identify the structural bottlenecks
in a given airspace such that minimal changes can be made to the airspace network
structure to improve the traffic flow? 2) how to exploit the structural and flow fea-
tures of an airspace to make effective predictions of air traffic capacity and flow? To
this end, this thesis concentrates on three specific research topics: airspace structure
optimization, airspace capacity overload identification, and air traffic flow prediction.

Note that most traditional methods in airspace optimization are based on ana-
lytical models derived from empirical analysis or traffic simulations, which may not
accurately represent the complex, dynamic, and interdependent nature of real-world
air traffic. With the widespread availability of ADS-B air traffic surveillance data,
data-driven approaches offer a new paradigm for solving the problem of airspace op-
timization. Compared to analytical models, data-driven approaches can be more ef-
fective in representing the complex relationships among the system variables without
explicit knowledge of the physical behavior of the air traffic system. Therefore, this
thesis aims to develop data-driven airspace optimization approaches based on real-
world air traffic data.

In airspace structure optimization, this thesis proposes to improve airway network

performance based on complex network theories. Based on the network percolation
and centrality theories, this thesis proposes a complex network approach for critical
link detection in airway networks to identify airway links that are vital for the struc-
tural integrity and performance of the networks in accordance with time-varying traffic
situations. Moreover, inspired by the Braess’s Paradox (BP) phenomenon, a scenario
where an alteration to a traffic network to improve traffic conditions actually has the
reverse effect, this thesis provides a “counter-intuitive” perspective towards airspace
optimization by removing airways/links from a given airway network to improve the
air traffic flow efficiency. In terms of airspace capacity overload identification, this
thesis proposes to use airspace collision risk patterns as an indicator of capacity over-
load. With air traffic data and airspace configurations, the collision risk distributions
inside an airspace are organized into different patterns, such as normal patterns and
overload patterns, based on the density and intensity of the collision risk. The capacity
overload status can be identified by recognizing the best-fit pattern of the collision risk

XXII
distribution, which can indicate the anticipated saturation in the airspace. As for air
traffic flow prediction, this thesis proposes to predict short-term en-route traffic flows
based on graph convolutional networks. A graph representation of air traffic structure
is built from real-world air traffic data, and the dynamic spatial-temporal features of
air traffic flow are explored from a graph perspective. Such prediction can help plan
and execute tactical measures for better air traffic management.
Feedback from subject-matter experts demonstrates that the complex network
theory-based method for critical link identification proposed in this thesis can dynam-
ically identify links, both structurally and operationally, critical under time-evolving
air traffic scenarios. Research results presented in the thesis show that optimizing
the airway network structure by removing some links can save up to 3.8% of the
travel time for a one-day traffic sample at a given flight level. Moreover, results of
airspace capacity overload identification demonstrate that distinct collision risk pat-
terns exist for the normal state, transition phase, and capacity overload state. The
proposed method can identify the capacity situation of a given traffic situation by
recognizing its best fit pattern. The results of air traffic flow prediction show that the
proposed graph convolutional network-based method outperforms the well-established
Long Short-Term Memory (LSTM) model. It exhibits a better capability to predict
rapid changes in traffic flow and has a relatively smaller decrease in prediction accu-
racy as the prediction window increases. Through airspace optimization, i.e., airway
network optimization to improve air traffic flow efficiency and capacity overload identi-
fication/air traffic flow prediction to advise ATFM to manage anticipated capacity/air
traffic flow circumstances, this thesis paves the way to a better accommodation of the
increasing air traffic.

XXIII
 and the scaling of the relationships between its input and gated
output parameters. For gas-models, the shape of the model is by rep
dependent on the modeling of the number of combinations of displa
two aircraft, i.e. the first component of gas models, see (1). The 1) Eff
scaling, on the other hand, is dependent on the computation Before
Chapter 1. Introduction
of the conflict probability between aircraft, i.e, the second model
component
1.1. of gas models, see (1), since conflict probability,
Air Traffic Growth results
as defined in this paper, is independent of traffic density. Fig. 1
Present day, airspace is reaching its operational capacity as air transportation is
The model adjustment derivations described in section III of the
have showed
growing thatoperationally
physically and all traffic[6,scenario properties
7]. The increasing only
air traffic affect
demand the
is leading
condit
computation
to of congestion,
significant airspace the conflict which probability
will cause flightbetween
delays and aircraft for
huge economic speed
UA, i.e., the scaling of the model. As such, the number of for UA
losses to Air Navigation Service Providers (ANSPs), airlines, and customers [8–10].
combinations of aircraft, or equally the shape of gas models, is distrib
Itnot affected
is reported thatbythetraffic scenario
annual total cost of distributions.
air transportationFor delaysthis
was reason,
over $30
ground
the analysis
billion in the UnitedthatStatesfollows for each
[11, 12]. Moreover, of the demand,
the increasing four simulation
as well as the result
experiments will focus on whether the analytical models can affect
congestion in airspace, may compromise air traffic controller’s workload capability,
correctly compute the conflict probability between aircraft, or probab
equally,
which thetoscaling
may lead of the
potential safety number issues
and inefficiency of combinations
in the airspace [13,of14].twoAs
ground
aircraft.
shown in Fig.Moreover, if the analytical
1, fast-time simulation models
experiments reveal can’t
that the predict
number the
of conflicts
scaling correctly, then the analysis will consider if the model 2) Eff
in the airspace can increase exponentially with the number of aircraft in the airspace,
adjustments derived in section III can be used to improve The ac
accuracy
leading to ATC forworkload
non-ideal traffic
and safety scenarios.
issues [1]. in Tab
it sho
assum
distrib
the ac
distrib
trend
speed
horizo
distrib
justme
high m
speed
distrib
C. He
Figure 1: Research simulation indicates that the number of conflicts in the airspace increases expo- The
Fig. 9. Simulation results (scatter points) and analytical model prediction (solid
nentially with the number of aircraft in the airspace [1].
lytical
line) for the ideal traffic scenario picture
Currently, one of the major challenges confronted by ANSPs is accommodating
the foreseen growth in air traffic demand. This problem is further compounded by

1
 53% more than 2017 (Figure 1). That is 1.9% average annual growth per year over the 2017-2040 period,
a rather slower growth rate than before 2008. Indeed, over the 20 years before the economic crisis, the
number of IFR movements in Europe doubled from 5 million IFR movements in 1988 to 10 million in
2008. The deceleration in growth over the next 20 years is explained by slower rates of economic growth,
increasing fuel prices and increasing congestion at airports.

Thetraffic’s
air recent return to trafficand
complex growth after 10 years
dynamic of stagnation
nature, such as hasthe
beendynamically
vigorous, and there are newer
changing airspace
growth drivers—long-haul, low-cost new aircraft types, middle-class growth in China, changes in
propensity to
capacity, thefly—which
complex aredecision-making
under-represented inmechanism
our forecast models
of airbecause
trafficofflow
their short histories.
management, and
For this reason we recommend that, in addition to the Regulation and Growth (most-likely) scenario,
particular
the attention isof
vulnerability paid
airtotraffic
the Global Growthto
systems scenario, which forecasts
convective weather84% [15].growth in flights to
19.5 million.
20
Global Growth

15 Regulation & Growth

Fragmenting World
Flights in Europe (Million)

10

5%

Annual Growth
0 0%
Annual Growth
-5%

Actual Forecast
1980 1985 1990 1995 2000 2005 2010 2015 2020 2025 2030 2035 2040

Figure 1 Figure 2: ofOverview

/ Overview the future of the future
long-term traffic long-term European
(IFR movements). The ratetraffic (IFR
of growth movements)
in traffic is [2].
expected to slow down over the next twenty years, compared to the pre-2008 period.

According to ICAO’s forecasts, global passenger traffic will double by 2035 with
/ / / // /////// /////// / // / // / // / //// /// /// /// /// /// /////////////////////////////
an annual growth rate of 4.3% [16]. In 2018 alone, the air transportation system
served nearly 8 billion passengers and operated over 90 million flights worldwide.
According to the Airports Council International (ACI) World Airport Traffic Report
2019, passenger numbers were estimated to have reached 8.8 billion in 2019, growing
by 6.4% compared to 2018 [17]. According to the “European Aviation in 2040” report
published by EUROCONTROL, European traffic (Instrumental Flight Rule (IFR)
movements) has passed 10 million flights per year since 2016. By 2040, air traffic is
expected to grow beyond 16 million flights per year and even close to 20 million in the
context of an increasingly globalized world, as shown in Fig. 2.

Although passenger demand during the outbreak of the COVID-19 pandemic al-
most came to a standstill, with travel restrictions being lifted gradually, air traffic
demand is now on its way to ramping up [18]. As shown in Fig. 3, in 2021, global air-

2
 Figure 3: Annual growth in global air traffic passenger demand between 2006 and 2022 [3].

traffic-passengers had an annual growth of 18 percent and was estimated to increase

by 51 percent in 2022. In 2021, overall traveler numbers were 47% of 2019 levels, which
is expected to improve to 83% in 2022, 94% in 2023, 103% in 2024, and 111% in 2025
[19]. This brief pause for the air transportation industry could also be an opportunity
to develop and adopt new approaches to accommodate future air traffic growth.

However, this air traffic growth is unsustainable, as evidenced by increasing de-

lays, flight cancellations, airspace congestion, taxiway incidents, Air Traffic Controller
(ATCO) workload, and environmental concerns [20, 21]. The continued growth in
air traffic, the reduced robustness to external conditions such as weather, and the
increased safety and environmental concerns have brought new challenges to air navi-
gation service providers, airlines, airports, and other stakeholders.

1.2. Research Motivation

According to a consensus in the aviation community, the projected increase in air

traffic demand can be accommodated through airspace optimization, which adapts

3
the airspace structure to the dynamically evolving air traffic flow and expedites the
orderly flow of air traffic through effective air traffic flow management [22].

Figure 4: Gate to Gate phases of a typical flight, with the cruise phase dominating the typical flight
time [4].

En-route airspace offers the most flexibility and opportunity for optimization. As
shown in Fig. 4, this is the airspace outside terminal areas, where the cruise phase, i.e.,
part of the flight from the end of the take-off and initial climb phase to the commence-
ment of the approach and landing phase, of the flight take place and within which
various types of air traffic services are provided [23]. En-route airspace is highly com-
plex with a number of actors (such as ANSPs, Air Traffic Control Centers (ATCCs),
airline operators, and national/international regulatory authorities) interacting with
one another, conducting different activities in the short, medium, or long term. Sub-
ject to the constraints posed by these actors, en-route airspace handles a huge number
of flights with consideration of many factors, including airspace structure, Air Traffic
Flow Management (ATFM), flight separation, airspace capacity, etc.

Compared with the terminal maneuvering area (TMA), en-route airspace conges-
tion becomes more prominent in busy airspaces around the world, such as Europe
and Southeast Asia [24]. Additionally, the departure and arrival phases of flights
are highly regularized by Standard Instrument Departure (SID) [25] and Standard
Instrument Arrival (STAR) [26], which gives en-route airspace a better potential to
be improved. Innovative research has started to be conducted to improve the en-

4
 Inadequate capacity Congestion Delay Safety concerns

Increasing demand Airspace optimization

Improvement in airspace structure Improvement in flow management

Airway network optimization Capacity overload identification Air traffic flow prediction

Improve flow efficiency Help ATFM facilitate effective utilization of the airspace

Figure 5: A visual illustration of the research motivation. The increasing air traffic demand is leading
to a series of adverse effects, including inadequate capacity, air traffic congestion, flight delays, and
safety concerns, which require the optimization of the airspace to adapt to the increasing demand.
As the promising options for en-route airspace optimization, improvement in the airspace structure
and traffic flow management have been well recognized and shown promising results. The airspace
structure, such as the airway network structure, influences the air traffic flow structure and controller
workloads. Optimizing the airway network structure concerning the evolving air traffic demand may
improve the air traffic flow efficiency and the airspace’s capability in air traffic accommodation. By
predicting the air traffic flow in view of identifying airspace capacity saturation, ATFM strategies
can be applied more efficiently to make the most effective utilization of the airspace. Therefore,
airway network optimization, airspace capacity overload identification, and air traffic flow prediction
are three major steps to be explored to optimize the airspace for better air traffic accommodation.

route airspace capability in handling the increased traffic demand. Research programs
such as Single European Sky ATM Research (SESAR) [27] and Next Generation Air
Transportation System (NextGen) [28, 29] have commenced research on optimizing
the en-route airspace for the adaption of aviation to future air traffic demand [30],
ranging from route network design [31], Free Route Airspace (FRA) implementation
[32] and trajectory optimization [33], to air traffic flow management [34].

As the promising options for en-route airspace optimization, optimizing the airspace
structure and improving the ATFM have been well recognized and shown promising

5
results over the decades [35, 36]. Firstly, from a spatial view of airspace optimization,
the airspace structure, including sectorisation schemes and airway network structure,
influences the air traffic flow structure and the controller workload, which are the
primary limiting factors of flight efficiency and airspace capacity [37, 38]. Optimiz-
ing the airspace structure to adapt to the evolving air traffic demand can improve
flight efficiency and balance the controller workload in the airspace, contributing to
the airspace’s capability to accommodate the increasing traffic demand [39].

Secondly, from a temporal view of airspace optimization, improvement in ATFM

can facilitate activities related to optimal traffic organization when demand exceeds or
is expected to exceed the available capacity of the Air Traffic Control (ATC) system.
For ATFM to develop effective flow control strategies in advance, the provision of the
anticipated saturation in the airspace capacity and the forthcoming air traffic flow is a
vital prerequisite. With predicting the air traffic flow in view of the capacity saturation,
flow control strategies, such as flight re-routing and speed adjustment, can be applied in
a more efficient manner to make the most effective utilization of the airspace [40, 41].
Therefore, air traffic flow prediction and airspace overload identification are crucial
enablers of the safe, orderly, efficient flow of air traffic.

In view of the above analysis, this thesis aims to contribute to airspace optimization
from three research perspectives: airspace structure optimization, air traffic flow pre-
diction, and airspace capacity overload identification. The overall research diagram of
this thesis is depicted in Fig 5. The increasing demand is leading to inadequate capac-
ity, air traffic congestion, flight delay, and safety concerns. Addressing the increasing
demand as well as its adverse effects requires the optimization of the airspace, which
can be achieved through improvement in the airspace structure (e.g., airway network
optimization to improve the traffic flow efficiency) or improvement in ATFM (capacity
overload identification and air traffic flow prediction to help ATFM facilitate effective
utilization of the airspace). In the following paragraphs, detailed research background

6
of airspace optimization will be introduced from each of the three research perspec-
tives.

1.2.1. Airspace Structure Optimization

In en-route airspace, air traffic has long been artificially concentrated on airways
with intermediate waypoints [42], i.e., airway networks. Fig. 6 shows the airway net-
work structure in Singapore Flight Information Region (FIR) as an example. The solid
lines depict the sector boundaries, the triangles denote the waypoints, and the dash
lines represent the airways. Research has proved that the structure of airway networks
can influence the airspace robustness, efficiency, and capacity [43–45]. Present airway
networks have evolved, over a period of time, without any scientific basis. As new
waypoints and airways are added to airway networks to accommodate the increasing
traffic in an as-needed manner, the structures of airway networks are becoming more
and more complicated [46–48]. Such networks may also lead to increased ATCO work-
load and reduced flight efficiency in the airspace [49, 50]. Therefore, airway network
design and optimization is an important step to improving the airspace structure and
mitigating the increasing traffic pressure.

Note that traffic congestion usually emerges on small parts of the airway network
and propagates to the proximity [51], which essentially inhibits traffic flows in the
airspace and exacerbates traffic congestion. Therefore, one way to improve airway
network performance is to identify the critical links in the airway network, limiting
the network traffic flow efficiency structurally and operationally. Accordingly, flow
management measures can be taken in advance to improve the air traffic flow and
prevent the potential failure of such links to ensure the smooth flow of air traffic
through the air traffic network.

Besides recognizing the critical links in an airway network to facilitate flow manage-
ment to address the inefficiencies in airspace, optimizing the airway network structure
to improve the network flow efficiency is another way to better air traffic accommo-

7
Figure 6: Singapore Flight Information Region and its airway network structure, including waypoints
and airways. The solid lines depict the sector boundaries, the triangles denote the waypoints, and
the dash lines represent the airways.

dation. Airway network optimization can result in a completely new design or partial
modification of the airway network. Considering the variety of stakeholders (Military,
ANSPs, Airports, Airlines) and the series of airspace constraints (navigation infras-
tructure, special use airspace), the complete airway network redesign to obtain the
theoretical optimal airway structure for a given airspace/FIR is difficult and costly to
be implemented in real air traffic operation [52]. Moreover, due to the highly dynamic
structural features of the air traffic flow, various optimal airway network structures
may exist under different traffic scenarios [53]. Therefore, it is difficult to determine a
static global optimal airway network structure to accommodate the evolving air traf-
fic flow. Instead, it may be of great significance to develop a scientific and flexible
airway network management method to improve the airway network performance in
accordance with the varying air traffic flow structure by making minimal changes to
the airway network structure.

8
1.2.2. Airspace Capacity Overload Identification

Optimization of airway network structure offers an opportunity to break through

an airspace’s structural limitations and improve the network air traffic flow from a
spatial point of view. Note that the primary cause of air traffic congestion is the
limited airspace capacity being overloaded by the increasing air traffic demand, which
can induce excessive controller workload and jeopardize flight safety [54]. From a tem-
poral perspective, knowledge of the airspace capacity overload status can facilitate
flow management measures in addressing anticipated saturation in airspace capacity
and limiting air traffic to a level that does not overload the system or the controllers
excessively. When the capacity of an airspace is overloaded or foreseen to be, flow
management measures can be triggered, such as coordinating with adjacent sectors
that have an immediate impact on the airspace capacity, managing the airspace struc-
ture to balance the controller workload, and altering traffic movement to avoid flight
conflicts [55]. En-route delays and conflicts can be controlled, and the flight cost can
be reduced consequently [56]. Thus, identification of the airspace capacity overload is
an important enabler in maintaining a safe and smooth flow of air traffic.

In most air traffic control systems, the airspace capacity is quantified as the maxi-
mum number of flights that may enter an airspace per hour averaged over a sustainable
period of time and is used to manage a safe, orderly, and efficient traffic flow [57]. Such
measurement of the airspace capacity depends on many factors, e.g., the airway net-
work structure, navigation accuracy, convective weather, airspace restrictions, and
controller workload [13]. Given the uncertainties in the variety of influencing factors,
the airspace capacity can be highly dynamic. Therefore, the controller’s judgment
about the airspace capacity overload can not be simply summarized by the number
of flights, although they are correlated. Therefore, to ensure that the flights provided
with air traffic control service do not overload the airspace capacity under the pre-
vailing circumstances, more innovative metrics and methods need to be investigated

9
to advise ATFM about the airspace capacity overload status. More specifically, to
identify the phase transitions of airspace capacity, i.e., the traffic demand is under
airspace capacity, during the transition phase, or overloading the airspace capacity.

1.2.3. Air Traffic Flow Prediction

In addition to the identification of airspace capacity overload, improvement in air

traffic flow prediction system can help forecast air traffic demand and accommodate
the future air traffic flow by ATFM.

Usually, ATFM is carried out in three phases: strategic planning, pre-tactical plan-
ning, and tactical operations [13]. Tactical ATFM operations consist of: 1) executing
the agreed tactical measures in order to provide a reduced and even flow of traffic where
demand would otherwise have exceeded capacity; 2) monitoring the evolution of the
air traffic situation to ensure that the ATFM measures applied are having the desired
effect; 3) taking or initiating remedial actions when long delays are reported, including
flight re-routing and flight level allocation [58], in order to utilize the available ATC
capacity to the maximum extent.

Accurate tactical air traffic flow prediction is essential in advising air traffic con-
trollers about the forthcoming traffic flow, which can support the decision-making and
evaluation of control strategies of tactical ATFM [59]. When the traffic demand is
foreseen to exceed the capacity of a particular airspace, the responsible ATFM unit
can inform the flights planned to fly in the affected area as soon as practicable of the
delays expected or the restrictions that will be applied. The flight routes and flight
levels can be allocated more reasonably and the available ATC capacity can be uti-
lized maximally. In this way, tactical ATFM operations can be planned and executed
in advance in a more efficient manner to accommodate the increasing air traffic de-
mand. Moreover, to reduce air traffic congestion and flight delays in the presence of
disruptions, such as convective weather and airport conditions, innovative modeling
and prediction methods for air traffic flow are effective supporting tools for ATFM to

10
ensure that air traffic moves smoothly and efficiently in the airspace [60].

1.3. Research Objectives

To handle the increasing air traffic demand, this thesis proposes to optimize the
airspace from three perspectives: airway network optimization, airspace capacity over-
load identification and tactical air traffic flow prediction. The specific research objec-
tives of this thesis are as follows:

1 To investigate effective methods for airway network optimization through crit-

ical link identification and to improve flight efficiency through airway network
structure management.

2 To develop airspace capacity overload identification methods using airway struc-

ture and traffic flow features. Such methods may help identify the airspace
capacity threshold transitions.

3 To develop effective air traffic flow prediction models by leveraging upon net-
work flow features. Such models may assist with efficient execution of ATFM
measures.

1.4. Data-Driven Methodology

En-route airspace is a complex environment, involving many actors from different

domains, operating at different spatial and temporal horizons, such as airlines, ANSPs,
and air traffic controllers [61, 62]. Activities in the airspace are subjected to uncer-
tainties, such as weather and airspace restrictions. Given the variety of influencing
actors and factors, modeling and optimizing the airspace based on physical laws or
expert experiences is a challenging and demanding task.

Data-driven approaches provide a flexible alternative to the physical laws of traffic

flow, which helps to see and plan what can be done to improve the airspace performance
with minor money and time costs [63, 64]. By processing and exploiting the air

11
traffic data, insights can be derived to improve the performance of an airspace, such
as supporting the decision-making of ATCOs and helping use past information to
forecast the future [65]. This gives air traffic data a key role in better assessing and
understanding the airspace performance, enhancing the system’s predictive power, and
creating new decision support tools to assist ATFM from strategic to tactical levels.

With Automatic Dependent Surveillance-Broadcast (ADS-B), open-source flight

tracking data has become increasingly accessible, offering an opportunity to analyze
flight movement in the airspace on a global scale. Therefore, aiming to optimize the
airspace for better air traffic flow accommodation, this thesis explores air traffic data to
investigate novel methods and concepts that can contribute to airspace optimization.
Fig. 7 depicts the concept diagram of the data-driven framework adopted in this thesis.
With the assistance of big data based techniques and theories, such as machine learning
and network science, data-driven approaches expedite the move of air transportation
from a system that relies on experience, intuition, and “golden rules” toward a system
that is more intelligent, adaptive, and proactive.

1.4.1. Network Science in Airway Network Optimization

Complex network theory-based analysis is one of the most promising approaches

to data-driven airway network optimization [66]. The study of complex networks is an
active area of scientific research primarily inspired by the empirical study of real-world
networks, such as computer networks and social networks [67]. The research content
mainly includes the geometrical properties of the networks, the formation mechanism
of the networks, the stability of the network architecture, and the dynamic mechanism
of the network evolution [68–70].

En-route airspace, which comprises a number of airways and intermediate way-

points, can be seen as a network of links between waypoints. The horizontal projection
of such an airway network can be seen as a planar network, with the nodes being the
waypoints and links being the airways. Airways in the airway network can be seen

12
Figure 7: Concept diagram of the data-driven research framework in this thesis. Open-source flight
tracking data offers an opportunity to optimize the airspace using data-driven methodologies, such as
machine learning and network science, expediting the move of air transportation from an experience,
intuition, and “golden rules” based system toward a more intelligent, adaptive, and proactive system.
An airway network can be modeled as a weighted network with links being the edges, waypoints
being the nodes, and edge weights describing the relation between nodes, such as travel efficiency,
connectivity, and edge centrality. Modeling and analyzing such weighted networks from real-world
data based on network science help understand the function and optimization of the airway network
in a way of practical significance. Machine learning-based approaches have shown particular success
in traffic pattern recognition and prediction without facing the challenge of modeling the complex
physical laws of air traffic. It has the potential to discover new insights from the big air traffic data
for airspace capacity overload identification and air traffic flow prediction.

as a sequence of links starting and ending at fixed nodes(waypoints). In recent years,

many researchers have begun to use network theory to investigate airport networks.
The topological structure of the world airport network was studied in [71] as a complex
network. It was found that the worldwide airport network is a small-world network
with power-law decaying degree and betweenness centrality distributions. Du’s team
analyzed the Chinese air transportation network using each province as a network node
[72]. It also concluded that the Chinese air transportation network is a small-world
network.

13
 The analysis of airway networks in en-route airspace is still a novelty in the litera-
ture. Network science theories and concepts have significant potential to contribute to
airway network-related research [73]. Rather than focusing on the simple node-to-node
interactions, a network science approach examines the interactions, relationships, and
inter-connectedness between hosts within a network [74]. In [75], Du et al. identified
the vital edges in the Chinese airway network, from a pure network perspective, by
examining the robustness of the new network after removing some edges via memetic
optimization. In [76], two strategies, adaptively modifying and permanently modifying
the network’s topology, were analyzed to increase the resilience of air traffic networks
using a network graph theory approach.

In the current literature, most network theory-based research on air transportation

and airway networks was conducted from a network topological perspective. However,
the operations of airway networks are influenced by many regulations and constraints.
Results obtained by pure topological analysis may be distinctly different from the
desired solutions in real situations. Note that in a weighted network, each link has
a connection strength value. The weights and their distribution significantly impact
the nature and function of the network, which can help understand the nature and
function of the complex systems more deeply [77]. Therefore, as shown in Fig. 7, it
is essential to incorporate real-world air traffic data into the weighted network-based
approaches to optimize the airway network in the way of practical significance.

1.4.2. Machine Learning in Airspace Optimization

With the increasing availability of air traffic data and the improvement in compu-
tational power, machine learning has been widely applied and shown to be particularly
successful in traffic pattern recognition and prediction [64, 78, 79]. In recent years,
machine learning techniques are becoming popular in air traffic management-related
research. For example, in [80], clustering methods are used for airspace collision risk
hot-spot identification. Deep learning has been widely applied in conflict resolution

14
[81], traffic flow prediction [82], flight delay prediction [83], etc. Long Short-Term
Memory (LSTM) networks are shown to be effective in time series forecasting [84].
Machine learning is also acting as an important enabler of the digitization of air traffic
management, which is a very relevant topic today. The theme for the Civil Air Navi-
gation Services Organisation’s (CANSO) 2018 Global Air Traffic Management (ATM)
Summit, one of the most important ATM events globally, was “Air Traffic Management
in the Age of Digitisation and Data” [85].

As shown in Fig. 7, given the sound performance of machine learning in pattern

recognition and traffic prediction, without having to face the challenge of modeling
the complex physical laws of air traffic, machine learning-based approaches have the
potential to discover new insights and mechanisms from the big air traffic data for
airspace capacity overload identification and air traffic flow prediction.

However, there remain many challenges in utilizing machine learning technology

for airspace optimization. These include the availability of big data required to train
the machine learning model, the heavy noise in the data set, and variations in data
sources. Careful feature engineering and data processing are required when employing
machine learning techniques.

1.5. Thesis Outline

The outline of this thesis is described below, while the structure of this thesis and
the relations between each chapter are depicted in Fig. 8.

Chapter 1 is the introduction to this thesis. Firstly, the background of this thesis,
i.e., air traffic growth, is introduced. Secondly, to accommodate the air traffic growth,
the research motivation and objectives for airspace optimization are illustrated from
three perspectives: airspace structure optimization, airspace capacity overload iden-
tification, and air traffic flow prediction. The end of this chapter briefly describes
the data-driven methodology in this thesis, including network science and machine
learning.

15
7/14/22, 12:00 PM thesis_outline.drawio.svg

Thesis Outline

Research Motivation Research Spatial Approaches to Temporal Approaches to Conclusions &

& Objectives Questions (RQ) Airspace Optimization Airspace Optimization Future Work

Chapter 1 Chapter 2 Chapter 3/RQ1 Chapter 4/RQ1 Chapter 5/RQ2 Chapter 6/RQ3 Chapter 7

Airway Critical Airway Network Capacity

Related Air Traffic Flow Conclusions &
Introduction Links Structure Overload
Literature Prediction Future Work
Identification Management Identification

Figure 8: Overview of the thesis structure.

Chapter 2 introduces the related literature of this research, including airspace

structure optimization, airspace capacity overload identification, and air traffic flow
prediction. Based on summarizing the state-of-the-art methodologies and the research
gaps in the literature, three research questions on airspace optimization, respectively
focusing on airway network management, airspace capacity overload identification,
and air traffic flow prediction, are illustrated at the end of this chapter.

Chapters 3 and 4 present the detailed data-driven research methodology, experi-

mental study, preliminary results, and the results discussion of airway network opti-
file:///C:/Users/Administrator/Downloads/thesis_outline.drawio.svg
mization. Chapter 3 focuses on the airway network’s critical link identification through
complex network theories, i.e., network centrality and percolation theory. Chapter 4
focuses on managing the airway network structure through minimal changes based on
the inspiration from Braess’s Paradox.

Chapter 5 presents the detailed research methodology, experimental study, prelim-

inary results, and discussion of airspace capacity overload identification. This chapter
adopts aircraft mid-air collision risk patterns as an indicator of airspace capacity over-
load and identifies different phase transitions of airspace capacity through pattern
recognition.

Chapter 6 presents the detailed research methodology, experimental study, prelim-

inary results, and the discussion of air traffic flow prediction. This chapter models

16
air traffic flows as a weighted spatial-temporal graph using Word2vec neural networks
and adopts graph convolutional networks to predict air traffic flow.
Chapter 7 discusses the research outcomes of this thesis and makes recommen-
dations for future work. Firstly, the research findings concerning each of the three
research questions on airway network management, airspace capacity overload identi-
fication, and air traffic flow prediction are discussed, respectively. Secondly, the poten-
tial contributions of the research findings to the future air traffic system/procedures
are discussed, including sectorless/flight-centric ATC service, virtualization of ATM,
integration of airspace entrants, and digitization of the aviation system.

17
Chapter 2. Related Literature & Research Questions

In the literature, many efforts have been put into improving airspace performance
through airspace structure optimization, airspace capacity overload identification, and
air traffic flow prediction. To develop innovative approaches to airspace optimization,
this chapter studies the related literature regarding airspace design, airspace capacity
overload identification, and air traffic flow prediction, respectively. At the end of this
section, based on analyzing the current literature and identifying the research gaps,
the research questions of this thesis are illustrated.

2.1. Airspace Design

Effective airspace design will enhance the ability of ANSPs and airspace users
to accomplish conflict management and increase ATM system safety, efficiency, and
capacity [86]. In the literature, lots of research has been done on airspace design to
manage air traffic congestion induced by the increasing traffic demand. The main
approaches are airspace sectorisation and airway network management [87]. This
section will introduce the related literature regarding airspace sectorisation schemes
and airway network optimization.

2.1.1. Airspace Sectorisation

One frequently-used air traffic accommodation strategy is increasing the airspace

capacity by improving the sectorisation schemes [88]. An airspace sector is the ba-
sic controlling unit in the air transportation system. They were originally designed
following predefined rules, including historical or geographic considerations [89]. How-
ever, along with the rapidly increasing air transportation over the past decade and
the changing traffic flow structures, several problems have arisen, such as the highly
variable traffic patterns and unbalanced workload across different sectors [90], making
the current sector design difficult to accommodate the evolving air traffic flows [91].

18
Therefore, airspace sectorisation schemes need to be adjusted to provide maximum
efficiency and flexibility in response to the changing traffic conditions.

The original idea of airspace sectorisation was realized by the “Merge and Di-
vide” operation [92]. However, the flexibility of this approach is limited by the fixed
boundaries of these sectors across different periods [93, 94]. Therefore, the concept of
dynamic airspace configuration was proposed [95], which allowed both boundaries and
the number of sectors to change according to the evolving traffic situations.

At present, there are mainly three categories of methods in dynamic sectorisation

[93, 96]: geometric computation based methods, clustering based methods, and graph
based methods. In the geometric computation based methods, Voronoi diagrams were
combined with genetic algorithms to find the optimal design and number of sectors in
an airspace [97–99]. Tang et al. [90] proposed a multi-objective optimization method
to achieve dynamic sectorisation and compared the effects of several different sector
partition models, such as bisection and KD-tree [100]. Wong et al. created airspace
sectors with a Voronoi diagram. They proposed a multi-objective formulation to obtain
a range of solutions with varying trade-offs among different categories of controller
workload [101]. They further extended this work by changing the boundaries of the
sectors concerning future traffic distributions in a rolling horizon framework [102].

In the clustering based methods, Brinton formed sectors directly by clustering the
trajectory points [103]. Yousefi and Donohue separated the airspace into three layers
based on the altitude ranges [104]. Each layer was divided into hexagonal cells based
on evaluating the controllers’ workload. Then, clustering techniques were applied to
the hexagonal cells to form different sectors. Drew improved the sectorisation results
in [104] by utilizing boundary-smoothing strategies to eliminate the jagged boundary
segments [105]. Targeting future flight-centric operations, Gerdes et al. came up
with the idea of automatic sectorisation, which clustered traffic flows based on fuzzy
partition and used evolutionary algorithms to find the optimal sectorisation scheme

19
[91].

The graph based method relies on a weighted undirected graph representation of the
airspace, which adopted a set of sub-graphs to represent different sectors [106]. Li et al.
represented the air route network by a weighted graph [107]. Then, the sectorisation
problem was modeled as a graph cut problem, which was solved by iterative spectral
bisection. Chen and Zhang proposed a spectral clustering-based method to cluster
vertices in the graph [108]. The spectral clustering approach was further refined by
mixing general weighted graph cuts algorithms to achieve better performance in terms
of controller workload balancing. Sergeeva et al. created 3D airspace configurations
using graph partitioning techniques, and the airspace configurations were optimized
using a genetic algorithm to minimize the controller workload [39].

In [109], Zelinski delivered an overall comparison of the various sectorisation meth-

ods. The results demonstrated that different approaches would induce significantly
different sector shapes. Geometric computation-based methods are straightforward,
whereas they are optimally inferior due to the limited methods to cut the airspace.
While the second approach, i.e., the clustering based method, may have the best
performance in workload balancing, it has difficulty dealing with other sectorisation
objectives or constraints. Undirected graph cuts-based methods show great poten-
tial for multi-objective and multi-constraint optimization. However, these methods
are facing two critical challenges. The first is the validation of graph cutting method
results, which must guarantee compliance with all problem constraints. The second
is the difficulty in delineating the specific sector boundaries from the generated sub-
graphs. Additionally, dynamic sectorisation is considered operationally infeasible by
many researchers since it takes time for traffic controllers to adapt to the new shape
of the sector. Further, airspace sectorisation almost reaches the structural limits of
the system and can raise safety and efficiency issues.

20
2.1.2. Airway Network Management

In an airspace, air traffic has long been artificially concentrated on airways with
intermediate waypoints [42], i.e., airway networks, which could result in congestion as
well as other airspace inefficiencies [43]. Note that en-route congestion usually emerges
locally in an airway network. The local congestion on some links can propagate to
their vicinity [110], which then essentially impedes traffic flows in the airspace and
exacerbates the traffic congestion. In the presence of the limited airspace capacity and
the saturated airway network, it is of great significance to identify such critical links
in an airway network to assist air traffic managers with better traffic flow planning
and decision making.

Generally speaking, a critical link in a complex network is a link whose failure will
significantly affect the network’s performance in terms of structural integrity, function-
ality, etc [111]. Critical link detection in complex networks is not new and has long
been explored [112–114]. One of the most popular methods for critical link detection
in complex networks is based on network centrality metrics [115]. A centrality met-
ric provides a straightforward way of calculating how central a network’s component
is. The second widely adopted method is based on network vulnerability analysis in
the presence of link failures [116]. The underlying idea is that the critical links of a
network should be the links whose failures will decrease its robustness in the face of
perturbations [117, 118]. In recent years, many researchers have applied network the-
ories to detect critical links in transportation networks by investigating the networks’
performance, such as the overall travel cost [119], the total demand losses [120], the
decreased network capacity [121], etc.

Although various methods have been proposed for critical link detection in trans-
portation networks, existing approaches mainly have three drawbacks. Firstly, the
model-driven methods [122] neglect the merits of real-world traffic data, which may
limit its implementation in the real-world context [123]. Secondly, many methods only

21
consider a single metric, such as network centrality [124, 125], traffic capacity [121],
or network robustness [126, 127], to quantify the criticality of a network’s link [113].
Multiple criteria should be taken into account for critical link detection for real ap-
plication purposes to avoid uncertainty and unreliability [128]. Thirdly, as indicated
in [129] that the majority of existing methods only deal with static networks, while
real-world transportation networks are time evolving. Meanwhile, most if not all of
the existing studies are oriented to urban traffic networks [130, 131], rendering their
direct applications to the three-dimensional air traffic networks infeasible.

Apart from the critical link identification, adapting the airway network structure
to the increasing traffic flow is essential for better air traffic accommodation. Present
airway networks have evolved, over a period of time, without any scientific basis [132].
As new waypoints and airways are added to airway networks to accommodate the
increasing traffic in an as-needed manner, the structures of airway networks are be-
coming more and more complicated [46–48]. Such networks may also lead to increased
ATCO workload [49, 50]. Research has proved that the structure of airway networks
indeed influences air traffic congestion [43–45]. In [4], the authors proposed a dynamic
rewiring strategy that can locally reshape the structures of airway networks. They
discovered that partially changing the structures of airway networks can alleviate air
traffic congestion. Note that applying changes to airway networks to form a better dis-
tribution of traffic flow may reduce the congestion and provide more flexibility to flight
schedules and routes. Thus, it is necessary to investigate novel methods to restructure
airway networks so as to improve flight efficiency in the airspace.

In line with this idea, researchers have made significant efforts toward designing
airspace with optimal airway structures [38, 133, 134]. In [135], a multi-objective Par-
ticle Swarm Optimization (PSO) algorithm was introduced to optimize the crossing
waypoints’ location of an airway network with the objective of maximizing flight effi-
ciency as well as the airspace capacity simultaneously. However, this method does not

22
consider actual location restrictions and operating limits of the waypoints. In [136],
an airway network optimization model was developed to minimize the total opera-
tional cost, with airspace restriction and air route network capacity being the major
constraints. The disadvantage of this work is that it targets the fragmented airspace in-
stead of the general airspace. In [137], a multi-objective genetic algorithm was adopted
to optimize the airway network based on geographic information to reduce the number
of airways, travelers’ travel time, and the aggregation of passengers. In [138], an opti-
mization framework based on an Artificial Potential Field (APF) model was proposed
to optimize the airway network structure considering the attractive forces produced by
the optimal routes and repulsive forces generated by obstacles (Prohibited, Restricted
and Dangerous areas). In [139], constrained by airspace availability and route cross-
restrictions, the crossing waypoint locations in airway networks were optimized based
on a flight conflict coefficient using the Genetic Algorithm (GA). The three models
above mainly focus on two network topologies, i.e., point-to-point or/and hub-and-
spoke. The two topologies are difficult to capture the complex en-route air traffic flow
and airway network structure that can be influenced by the airspace infrastructure,
the navigation aids, or operational constraints.

Essentially, at present, most of the solution algorithms for the airway network
optimization model, such as GA, PSO algorithm [140], and Differential Evolution
(DE) algorithm, are both complex and easy to converge to the local optimal because
of limited exploration ability in the search space. The stability of the results is also
a challenge. More importantly, these efforts often end up with complete new airway
structures for a given airspace. The challenge is that, due to operational constraints,
the complete redesign of airway networks, so as to obtain the global optimal airway
structures (for a given airspace/FIR), is impractical.

Some researchers proposed new concepts for the operation of airways, such as
defining new airways dedicated to the most important traffic flows [141, 142] and 4D

23
airways design [143–146] that can reduce congestion by planning the space and time
of flights. 4D airways can reduce the effects of air traffic congestion problems through
flight control. However, its effectiveness is limited by the capacity of the airspace.
Moreover, its good performance is subject to good control of uncertainties, such as
weather and traffic demand.

2.1.3. Summary

Increasing the airspace capacity by improving the sectorisation schemes is one of

the most frequently-used air traffic congestion management strategies for airspace opti-
mization [88]. Dynamic airspace sectorisation is an approach to restructuring airspace
to achieve capacity-demand balance by managing air traffic controllers’ workload to
ensure an orderly flow of traffic [90, 95]. Airway network design and optimization is
another way to contribute to airspace optimization. Some researchers focus on iden-
tifying the critical links in the airway network with regard to the local presence of
traffic congestion on some airway links. Meanwhile, most research pays attention to
making structural improvements in the airway network, which usually end up with a
completely new airway network structure. Some researchers introduced the concept
of 4D airway design that can reduce congestion by doing spatial and temporal airway
planning for flights.

2.1.4. Gaps

Dynamic airspace sectorisation has reached the structural limits of the system and
may induce safety and inefficiency problems. The performance of critical link detection
methods in the literature is limited by the adoption of a single criticality evaluation
metric, the assumption of static network typologies, or the failure of considerations
to the real-world data. Complete redesign of airway networks is a challenging task
that has to consider many competing factors, such as the needs of civil and mili-
tary airspace users, the traffic structure, the ATM capacity, a proper sectorisation

24
scheme, and the connectivity to the terminal area [89]. Given a variety of stakeholders
(Military, ANSPs, Airports, Airlines, etc.), a complete redesign of airway networks is
impractical. 4D airway design mitigates air traffic congestion by flight control. How-
ever, its effectiveness is limited by the capacity of the airspace. Moreover, good results
are subject to good control of uncertainties such as weather and traffic demand. Thus,
it is necessary to develop new airway network design and optimization methods so
that, by making minimal changes to the structure of an airway network, the airspace
efficiency can be improved.

2.2. Airspace Capacity Overload Identification

Knowledge of the airspace capacity overload status, i.e., the airspace is operating
under capacity, at a transition phase, or over capacity, can be helpful for controllers
to plan and develop flow management strategies in advance to address the anticipated
traffic saturation in the airspace. In past decades, multiple methods were developed
concerning airspace capacity overload identification, including quantitative methods,
i.e., airspace capacity estimation based methods, and qualitative methods, i.e., other
capacity overload indicators based methods.

2.2.1. Quantitative Methods: Airspace Capacity Estimation

One way to identify airspace capacity overload is to estimate the airspace capacity,
i.e., the maximum number of aircraft which can be accepted over a given period of
time within the airspace, and identify whether the traffic demand will exceed the
estimated capacity value. In the literature, there are mainly two types of airspace
capacity estimation models: controller workload based analytical models and machine
learning based models.

2.2.1.1. Controller Workload Based Analytical Models. Analytical description of airspace

capacity is to develop a static mathematical formula to compute the airspace capacity
using a set of complexity factors [147]. A large set of these methods are based on

25
regression models of controller workload generated by complexity factors for different
air traffic situations. In the Monitor Alert Parameter (MAP) model, the mean sector
transit time (the average time required for aircraft to fly through the sector) is used
as the sole complexity factor for determining the controller workload and the sector
capacity [148]. While the full workload model [147] accounts for all task types (tran-
sit, recurring, conflict, background) constituting controller workload. With a preset
workload intensity threshold, the sector capacity can be described as a function of
the threshold and the complexity factors contributing to the workload intensity. The
assessment of the workload intensity induced by complexity factors has mainly been
done by asking the subjective personal opinion of controllers [149], by observing con-
trollers at work and measuring the time they spend on various actions [150] or by
measuring their physiological parameters [151]. Finally, regression techniques, which
can be simple linear [57, 152] or logistic regression [153] as well as more developed
methods, such as neural networks [154–156], are used to find the function that best
fits the relationship between the sector capacity and the complexity factors. Once
this function is defined, it can be used to estimated the sector capacity under tar-
geted traffic situations subject to the workload intensity threshold. These analytical
models can result in precise enough estimations in a certain airspace. However, the
biggest disadvantage of these analytical models derives from their overly elaborated
way of describing complexity factors. The resulting mathematical formula to calculate
the capacity is static, making it hard to capture the intrinsic mechanism behind the
complexity factors as well as the evolution of the sector.

2.2.1.2. Machine Learning Based Models. The models introduced so far all contex-
tualize air traffic as a set of single aircraft. It is, however, also possible to model
air traffic as a structure of air traffic flows. In [157], a Self-Organizing Map (SOM)
based model was adopted to represent the correlation between the characteristics of
air traffic flows and traffic patterns. Under each traffic pattern, the authors used the

26
average flight distance inside the sector as an indicator to identify a sector capacity,
which was realized by observing a phase transition in the average flight distance when
the number of aircraft increases. Then for a given traffic situation, the capacity was
determined through pattern recognition. This work sheds light on dynamic capacity
estimation using data-driven methods instead of analytical models based on controller
workload. The shortcoming of this method comes from the adoption of SOM to build
traffic patterns. Since SOM is an unsupervised learning method, the result is hard
to be explained and validated objectively [158]. Only three traffic patterns have been
extracted from the dataset using SOM, which limits the precision and diversity of
capacity estimation and prediction. Moreover, the validation of the usage of average
flight distance inside a sector as a capacity indicator is not manifested. In [159], the
airspace capacity was estimated using a spatial-temporal graph convolutional network
model considering three input features, including the estimated airspace capacity, the
air traffic demand, and convective weather. However, this supervised learning model
requires an estimated capacity value as the model input, which largely reduces its
practical significance due to the unavailability of foreknowledge about airspace capac-
ity.

2.2.2. Qualitative Methods: Other Capacity Overload Indicators

In the above research, aircraft count is the only measurement of airspace capacity
which can vary with traffic complexity and other factors. However, the representation
of airspace capacity using aircraft count has obvious limitations. The capacity of an
airspace is not simply determined by the number of aircraft handled by the controller
in scope. It is also influenced by numerous other factors like the disposition of aircraft
in the airspace [160], relative aircraft velocity, aircraft heading, aircraft types, airspace
structure, weather, and availability/quality of ATC equipments [161]. Sometimes the
controller in charge of an operational sector can accept more aircraft than the actual
capacity, while on some occasions will refuse some flights even though the capacity has

27
not been reached [162]. Thus, the controller’s decision about the capacity overload can
not be summarized by how many aircraft can be maintained in the airspace, though
they are correlated.

In recent years, researchers have started to use new capacity measurements which
identify the capacity overload by determining the system limit. One example is the
“sector operation”, i.e., the merge and split of sectors, which reflects whether the
traffic situation in a sector can be safely handled by controllers [160, 163]. Another
example is the “Domino Effect Parameter”, i.e., whether the conflicts in a sector can be
solved without causing conflict chain reactions [164]. These representations are more
comprehensive and integrated to reflect the capacity overload of an airspace. However,
neither of the two works can identify phase transitions in the airspace capacity. For
instance, the two models cannot identify the phase transition of air traffic operating
from under capacity to over capacity. Moreover, the “Domino Effect Parameter” is
calculated through fast time simulations which can be affected by the conflict resolution
algorithm adopted and may not apply to real traffic situations. In [165], the airspace
capacity overload was estimated using an augmented cell transmission model based
on the airspace structure with constraints on the separation requirements and flight
speed. The main disadvantage of this method is the lack of consideration of operational
constraints and traffic flow features, which limits its application under real air traffic
scenarios.

2.2.3. Summary

Knowledge of the airspace capacity overload can benefit flight planning and ATC
operations to avoid airspace saturation and air traffic congestion. One way to identify
airspace capacity overload is the quantitative approach, i.e., estimating airspace ca-
pacity through controller workload based analytical models or air traffic pattern based
machine learning models. The estimated capacity, represented as aircraft count, can
be compared with the traffic demand to identify whether the airspace capacity is over-

28
loaded. In recent years, new indicators, i.e., qualitative approaches for system limit
determination, have been introduced to identify airspace capacity overload, such as
“sector operations” (the merge and split of sectors, which reflects whether the traffic
situation in a sector can be safely handled by controllers), and “Domino Effect Param-
eter” (whether the conflicts in a sector can be solved without causing conflict chain
reactions).

2.2.4. Gaps

The most significant disadvantage of analytical capacity models is their overly

elaborated description of the workload factors. The resulting mathematical formula
is static, making it hard to capture the changes and evolution of the airspace. Using
traffic flow patterns to identify capacity is still at an early stage. Only a few patterns
and a few capacity values can be developed, which limits the precision and diversity
of the airspace capacity analysis results. In the above research, the implementation of
aircraft count as the measurement of airspace capacity can not summarise the decision
of the ATCOs about capacity overload, though they are correlated. The “Domino
Effect Parameter” is calculated through fast time simulations which can be affected by
the conflict resolution algorithm adopted and may not apply to real traffic situations.
Evaluation of capacity overload through neither sector operations (merge and split) nor
the “Domino Effect Parameter” can reflect the phase transitions in airspace capacity.
Therefore, it is necessary to seek new metrics and methods which can leverage the
airspace structural features and air traffic flow patterns to identify airspace capacity
phase transitions.

2.3. Air Traffic Flow Prediction

In the presence of the increasing traffic flow and limited capacity, an accurate short-
term air traffic flow prediction, i.e., 0-2 hour look-ahead time [82], in the airspace can
be vital for ATCOs managing the forthcoming traffic flow situation and evaluating

29
control strategies [59]. In this way, tactical ATFM measures can be planned and
executed in advance in a more efficient manner to maximize the use of the existing
airspace capacity. In the current literature, there are mainly two types of short-term air
traffic flow modeling and prediction methods: trajectory-based methods and aggregate
methods.

2.3.1. Trajectory-based Flow Prediction Methods

Trajectory-based air traffic flow prediction is based on propagating the flight tra-
jectories over time and using the predicted trajectories to count the future number
of flights in the airspace [166]. However, research shows that the prediction error
of the trajectory-based method increases exponentially as the forecast time horizon
increases beyond 20 minutes, which may be caused by its sensitivity to various influ-
encing factors, such as weather conditions, airspace restrictions, and human factors
[167]. Furthermore, the dimension of the trajectory-based method depends on the
number of flights under consideration, which requires substantial computational costs
making such methods untenable in the real-world context.

2.3.2. Aggregate Flow Prediction Methods

Efficient ATFM demands prediction of the overall traffic flow situations in the
specified airspace instead of the temporal trajectory of individual aircraft. Thus, the
aggregate air traffic flow models focus on predicting the overall distributions of the air
traffic flow in the airspace [168]. Since flights in the airspace are spatially aggregated,
the computational cost can be reduced significantly compared to the trajectory pre-
diction based approaches. Moreover, without considering the individual behaviors of
flights, the aggregated prediction methods are less sensitive to the uncertainty factors
related to individual flights. Therefore, a longer forecast time horizon with smaller
prediction errors can be achieved.

Recently, the aggregated approach has been widely discussed in the literature,

30
which mainly predicts air traffic flow based on probabilistic analysis or machine learn-
ing techniques. Based on the flight transition probabilities between adjacent air traffic
control centers, a stochastic framework with Linear Dynamic System Model (LDSM)
was developed by Sridhar et al [60]. The traffic flow is predicted as the flight count in
each control center during a short term, e.g., a few hours, using the flights transition
probability between adjacent centers extracted from historical data, the number of
departure/arrival flights in the control center, and the departure uncertainties. The
major shortcoming of the stochastic framework is that it adopts Poisson distributions
to model the departures at each time step, ignoring the fact that the departure traffic
flow varies significantly during days, weeks, and seasons under different traffic con-
ditions. Gilbo et al. further explored the stochastic LDSM to predict the air traffic
flow at the sector levels using a linear regression model [169]. However, their research
mainly focused on improving the flow prediction results of the Enhanced Traffic Man-
agement System (ETMS) without a deeper investigation of the prediction mechanism
of air traffic flow prediction. Probabilistic models of flight or airspace were also stud-
ied to illustrate stochastic features, such as flight speed and traffic density, of the air
traffic and further used to predict the traffic flow [170]. However, the performances
of probabilistic methods are affected by handcraft models or analyses, which makes it
difficult to fully depict the real traffic situation and capture the sharp changes in the
highly dynamic traffic flow.

In recent years, machine learning algorithms have started to be implemented in

aggregated air traffic flow prediction [64]. In [82], the authors discretized the airspace
into small cubes and predicted the flight count in each cube. The prediction was
achieved by a Convolutional Neural Network (CNN) and recurrent neural network to
model the spatial and temporal correlations of air traffic flow. The main shortcoming
of this model is that the disintegration of airspace into cubes ignores the structural
features of the airspace, such as airway network structure and sectorisation schemes.

31
This may limit the practical implementation of the prediction results in ATFM. In
[171], the authors adopted a Phase-Space Reconstruction Extreme Learning Machine
(PSR-ELM) model to learn and predict the time series of traffic demand passing given
waypoints in the airspace. In [172], by exploring ADS-B data, the authors adopted
LSTM models and Support Vector Machine (SVM) to predict the air traffic demand
between two airports. The limitation of such time-series models is that they fail
to consider the interactions between the time-series flows at different waypoints or
airports, which may restrain the prediction accuracy of the models [173].

2.3.3. Summary

Short-term air traffic flow prediction can be used in tactical ATFM to facilitate
traffic demand and mitigate congestion. One way to predict air traffic flow is based on
trajectory prediction of flights and counting the number of anticipated aircraft in the
airspace. Another way is to consider flights in the airspace as aggregated flows and
use analytical models, probabilistic models, or machine learning models to describe
and predict air traffic flow.

2.3.4. Gaps

Trajectory-based flow prediction error increases sharply when the forecast time
horizon exceeds 20 minutes. Moreover, the dimension of the model depends on the
number of aircraft under consideration, which is computationally expensive. In prac-
tice, it is not easy to make good online air traffic control strategies based on the
trajectory-based model due to the short forecast horizon and the expensive compu-
tational cost. The analytical aggregated model in the literature, i.e., LDSM, ignores
the fact that the traffic flow varies significantly during days, weeks, and seasons under
different traffic conditions. The probabilistic method adopts handcrafted probabilis-
tic models, which fail to fully depict the real traffic situations and capture the sharp
changes in traffic flows. Neural network models in the literature for time series pre-

32
diction fail to consider the spatial dependencies of the air traffic and may restrain
the prediction accuracy of the models. The main shortcoming of the CNN-based ap-
proach is that the prediction is made for a set of disintegrated cubes in the airspace.
It ignores the structural characteristics of the air traffic flow in the airspace, such
as airway network structure and sectorisation schemes. Such prediction outcome has
limited significance for ATFM, whose operations essentially rely on the airspace struc-
ture. Therefore it is necessary to investigate new short-term air traffic flow prediction
methods considering the highly dynamic, structural, and interrelated nature of air
traffic flows.

2.4. Research Questions

To fill the gaps in the literature and serve better the purpose of airspace optimiza-
tion through airspace structure optimization, airspace capacity threshold identifica-
tion, and air traffic flow prediction, this thesis aims to seek solutions to the following
research questions:

Research Question 1:
How to identify the structural critical links and flow bottlenecks in a given
airway network and make minimal changes to the network structure, in order
to improve the air traffic flow efficiency without compromising its structural
integrity?

Research Question 2:

How to identify airspace capacity phase transitions by leveraging airspace

structural and air traffic flow features?

Research Question 3:
How to improve short-term air traffic flow prediction by leveraging and learning
the spatial-temporal dependencies of air traffic flow in a given airspace?
33
Figure 9: Phylogenetic tree of the investigated research problems in this thesis.

34
 Fig. 9 depicts a phylogenetic tree of the investigated research problems in this
thesis. The topics presented in black letters show the research viewpoints this thesis
focuses on, while the topics in gray letters show other efforts in the literature to address
the airspace optimization problem. In the following chapters, the detailed research
methodology of this thesis will be illustrated regarding airway network management,
capacity overload identification, and air traffic flow prediction.

35
Chapter 3. Airway Network Critical Link Identification

The research in this chapter has been published in the journal paper: Chunyao Ma, Sameer Alam,

Qing Cai, and Daniel Delahaye, “Critical Links Detection in Spatial-Temporal Airway Networks Using

Complex Network Theories,” IEEE Access, vol. 10, pp. 27925-27944, 2022.

Knowledge of critical links in an airway network is important for ATCOs in the

presence of limited airspace capacity and the saturated airway network. Generally
speaking, a critical link in a complex network is a link whose failure will significantly
affect the network’s performance in terms of structural integrity and functionality [111].
When an airway network is of concern, the critical link is regarded as the airway link
(connecting two waypoints), which acts as a pivot link and whose failure may decrease
the network’s structural integrity concerning the given traffic scenarios. Therefore,
this chapter focuses on identifying the critical links in an airway network, which may
assist with better air traffic flow management while addressing the increasing traffic
demand.

Existing studies for critical link detection primarily deal with static networks, which
do not fit the time-evolving nature of airway networks. Dynamically detecting critical
links from airway networks over time using real-world data could be more appealing.
Moreover, in the literature, researchers mainly label the critical links in networks us-
ing a single metric. Adopting an assembled metric could provide a more reliable way
for critical link detection. In view of this, this chapter proposes to detect the critical
links spatially and temporally with respect to different traffic situations. Furthermore,
instead of using a single metric, two metrics in network theory, i.e., network central-
ity [115, 174] and percolation theory [175, 176], are adopted to identify the critical
links in the spatial-temporal networks. More specifically, a weighted spatial-temporal
airway network is constructed first based on given flight track data. Then, edge be-
tweenness centrality and network percolation theory are introduced to detect critical

36
links in each snapshot of the spatial-temporal network. Finally, comparisons among
the detection results using the two metrics are made to determine the final critical
links for the snapshot of given traffic scenarios. With the proposed method, a case
study is carried out on the Southeast Asia airway network derived from one-month
flight track data for the calendar year 2018 to demonstrate the efficacy of the proposed
method. The studied network covers the Air Traffic Service (ATS) routes in Singapore
FIR and the transiting links between Singapore FIR and its neighboring airspace. For
the purpose of simplicity, this network is termed as Southeast Asia Airway Network
(SEAN) throughout this thesis.

3.1. Related Backgrounds

In this section, the related backgrounds, i.e., the related network theories, of the
proposed methodology for critical link identification will be introduced, including
spatial-temporal networks, network centrality, and percolation theory.

3.1.1. Spatial-Temporal Networks

A complex network is usually depicted by a graph composed of a set of node/vertices

and links/edges. Mathematically, a graph G is denoted by G = {V, E} with V and E
respectively being the sets of nodes and links. Usually, n = |V | and m = |E| are used
to denote the number of nodes and edges in G.

In reality, the nodes of complex networks like airway networks carry geographical
coordination information. Such networks are generally called spatial networks. Note
that some complex networks are time-evolving, i.e., their structures change over time.
Such networks are generally called temporal networks. Mathematically, a temporal
network G can be denoted by a network sequence, i.e., G = {Gt0 , ..., Gti , ...} with Gti
being the snapshot at time point ti or for a certain time slice. A complex network
carrying both spatial and temporal information is normally modeled as a spatial-
temporal network [177].

37
3.1.2. Network Centrality

For a given complex network, one may wish to know which nodes or edges are
more important than others concerning the network structure. The network centrality
metric provides an outlet for that purpose. In the literature, many centrality metrics
are available [115, 174]. There are mainly two types of centrality metrics, namely node
centrality and edge centrality, while the latter is generally the extension of the former
ones.

In this study, the betweenness centrality metric (CEB ) is adopted to detect critical
links for a given airway network. The betweenness centrality has been tested to have a
large impact on a network’s robustness [115]. For a given network G, the betweenness
centrality for an edge e ∈ E is calculated as

2 X pij (e)
CEB (e) = (1)
n(n − 1) i̸=j Pij

in which pij (e) is the number of shortest paths between nodes i and j running through
edge e, and Pij is the total number of shortest paths between nodes i and j.

3.1.3. Percolation Theory

Complex networks, in reality, suffer from various perturbations. Consequently, the

components of a network may break down, and potential risk is likely to happen.
To better design the structure of a network to make it robust to perturbations, it
is pertinent to analyse the dynamics of a complex network subject to perturbations.
Percolation theory has proven as an effective instrument for analyzing the capability
of a complex network in the face of perturbations [175, 176].

Suppose that 1 − p fraction of network components are disconnected to the rest

of the focal network due to external/internal perturbations. The disconnection of
those failed network components can fragment the focal network into pieces amongst
which there exists the Largest Connected Component (LCC) [178]. The LCC of a

38
network is an essential indicator for capturing the network’s capability in response
to perturbations. When p = 0, the LCC of the network disappears, simulating the
scenario that the focal network is entirely down due to perturbations. For p = 1,
it corresponds to the situation that the network is not suffering from perturbations
and the LCC keeps its original state. When p increases from 0 to 1, the size of the
LCC changes with p. When p reaches a particular value, the size of the LCC shows
a notable change, such as a sharp decline or becoming extremely small or even zero.
Such a value of p is generally termed as the percolation threshold denoted by pc .

3.2. Problem Description

This study aims to identify the critical links in a given airway network from the
network theory perspective. Specifically, critical links are defined as the airway links
that meet the following two requirements:

1) the links act as the traffic pivots through which shortest paths frequently pass;

2) the links act as the bridges whose failures will significantly decrease the network’s
structural integrity in the face of perturbations.

Note that for different traffic situations, the critical links may vary. Therefore,
directly detecting critical links in a static and unweighted airway network is not of
practical usage. Thus, this chapter proposes to detect the critical links spatially and
temporally for different traffic situations. Moreover, instead of using a single metric,
two metrics in network theory, i.e., network centrality and percolation theory, are
adopted and combined to identify the critical links in the spatial-temporal networks.
Centrality is a widely applied metric for quantifying the importance of a network’s
components (nodes and links). Large centrality values can distinguish the pivots links
through which shortest flight paths frequently pass, while the percolation theory is an
effective method for measuring the structural integrity of a given network concerning
network component failures.

Fig. 10 presents a graphical illustration of the research problem. In the network

39
 Final critical links determination
Air Traffic Data

Critical links detection

Network modelling t
Network percolation Overlapping

Edge betweenness centrality calculation

Weighted
networks
Critical links

Figure 10: A graphical illustration to the proposed research problem of detecting critical links in a
given airway network by making use of complex network theories.

modeling step, for a given airway network, the flight track data are processed to
construct weighted spatial-temporal airway networks G = {Gt0 , ..., Gti , ...} with Gti
being the weighted network snapshot built for a certain time slice ti . Then, centrality
metric and network percolation theory are adopted to identify critical links from each
network snapshot Gti . Finally, in the final critical link determination step, the critical
links obtained by percolation theory and edge centrality are compared and integrated
to determine the final critical links.

3.3. Research Methodology

3.3.1. Methodology Overview

The proposed method for critical link detection contains three key steps: spatial-
temporal network modeling, critical link detection based on network theories, and final
critical link determination, as shown in Fig. 10.

In the first step, i.e., spatial-temporal network modeling, given the air traffic
data during different periods, the weighted spatial-temporal airway networks are con-
structed based on the flight fixes and flight paths information to describe the changing
traffic situations on the airway network during different time snapshots. The second
step, critical link detection, works on the constructed spatial-temporal airway net-
works. This step will leverage two network metrics, i.e., edge centrality metric and

40
percolation theory, to detect critical links from each network snapshot of the spatial-
temporal networks. For each network snapshot, the critical link detection step is likely
to yield different critical links from the two metrics. The third step, the critical link
determination step, compares and analyzes all the detected critical links and merges
them to filter out the final critical links for the studied airway network. In the follow-
ing three subsections, i.e., Section 3.3.2, 3.3.3 and 3.3.4, the detailed descriptions of
each of the three steps are illustrated respectively.

3.3.2. Network Modeling

The purpose of network modeling is to construct the weighted spatial-temporal

airway networks for the critical link detection step. Network modeling consists of two
steps: extracting the airway network configuration and determining the weight of each
link from the flight track data. In this study, the normalized average ground speed on
each link is used as the link weight. The main reason is that the buffer time for airborne
congestion as well as the risk of delay propagation through the entire airway network
can be reduced if flights can transit faster through their routes [179]. Moreover, when
an airway link is faced with congestion or other disruptions, major reactive measures
in operational air traffic flow management are flight vectoring and speed adjustment
[13]. These measures will consequently influence the speed of flights transiting through
the affected links. The travel speed on a link is normalized using the daily maximum
speed. Therefore, the lower weight of a link indicates that the current traveling speed
on the link is more degenerated than the best daily traffic situation.

The flight track data required for network modeling includes the flight paths of
each flight, i.e., the flight fixes of the flight trajectory and the time when the flight is
reported to be at these fixes. From the flight path information, the airway network
configuration can be constructed by setting the flight fixes as the nodes and determin-
ing the connections between nodes, i.e., links, from the path of each flight. If there are
flights whose paths pass the link between two nodes, the two nodes will be considered

41
connected. Consequently, a link will connect the two nodes on the resulting airway
network. In this manner, the airway network will be constructed entirely from the
flight track data.

With the time information of flights reaching the fixes, the average speed sa,j of a
flight fa on the link ej of its path can be computed by averaging the length Lj of ej
over the flight duration Tja on ej :

Lj
sa,j = (2)
Tja

Note that the critical links are identified for the k th time slice Tk (from t0k to t1k ),
therefore the required weight wj of link ej is the normalized average flight speed of
all flights passing ej during Tk , instead of simply normalizing the average flight speed
of all flights passing ej . This means that for flight fa , only the part la,j
k
on ej , that
has been flown during Tk , will be considered, From the available traffic data, it is not
able to obtain the exact time line of flight fa when it is on ej during Tk . Therefore, to
reduce the bias, it is assumed that the flight fa is flying on ej with the constant speed

en and exit time tex of fa , the flight duration Ta,j of

sa,j . Then, given the entry time ta,j a,j k

fa on ej during Tk can be determined:

k
Ta,j = min{ta,j 1 a,j 0
ex , tk } − max{ten , tk } (3)

Then, la,j
k
can be estimated as Lj weighted by the proportion of Ta,j
k
to Tk :

k
Lj × Ta,j
k
la,j = (4)
Tk

The average speed skj on link ej during Tk can be calculated by averaging the sum

42
of flight distances of all the flights on ej over the according sum of flight duration:

PF k
la,j
skj = PFa=1 k
(5)
a=1 Ta,j

where F is the total number of flights passing link ej during Tk .

Finally, the weight wj on link ej during the given time slice Tk is determined by
normalizing the average speed skj with the daily maximum speed sjmax on link ej :

skj
wj = (6)
sjmax

In this way, the weight on link ej is determined, which will be a number between
0 and 1.

3.3.3. Critical Link Detection

To measure how critical an edge of a network is, two methods are adopted from
the network theory point of view. The first method adopts the edge centrality metric
while the second method is based upon network percolation theory.

3.3.3.1. Centrality Method. Centrality is a class of straightforward metrics for quan-

tifying the importance of a network’s components. For a given airway network, Eq.
1 is used to calculate the centrality for each link. Note that there exist a couple of
edge centrality metrics in the literature. In this work, only the betweenness centrality
metric is adopted. The main reason is that the betweenness centrality is involved with
the shortest paths in a network which is more appealing to airway networks.

The edges are ranked based on their betweenness centralities. Edges with the
highest centralities are regarded as the critical links. Specifically, the highest K links
of the betweenness centrality are taken as the critical links. In the experiments in this
chapter, K is set as K = 10. Note that K cannot be too large as it would not be
possible for an airway network to have too many critical links. Also, K cannot be too

43
small as the detection results need to compare in the subsequent analysis with what
is detected by using the percolation theory.

3.3.3.2. Percolation Based Method. Percolation theory has been widely applied to
investigate the structural properties of diverse complex networks, including trans-
portation networks. For example, the authors introduced percolation theory to detect
critical links in urban traffic networks [111]. Percolation theory uses statistical physics
principles and graph theory to analyse changes in the structure of a complex network
subject to perturbations. The percolation threshold pc signifies the transition point of
a given network, thus, shedding light on probing its critical sub-structures.

1) Critical Threshold Identification

Note that for a given network, the lower weight of a link shows that the current
traveling speed on the link is more degenerated than the best daily flight speed on
this link. Therefore, these lower-weight links can be regarded as failed links that can
potentially slow down the flights and induce congestion. By incrementally closing
these low-weight links, one can observe the links whose closure leads to a transition of
the airway network from the phase of connected to the phase of disconnected.

Note that an airway link Ej is characterized by the weight wj . Therefore, for a

given threshold of q ∈ [0, 1], the link Ej can be classified into two categories: functional
when wj ≥ q and dysfunctional when wj < q. This can be represented as:



 1,
 wj ≥ q
Ej = (7)

 0,
 wj < q

As the value of q increases, more low-weight links are closed, which makes the
network sparser. Note that the weight on a link refers to the normalized average
flight speed on that link. It indicates that as q increases, links with low flight speeds
are closed, and links with higher flight speeds remain active. In this way, a functional

44
 (a) (b)

(c) (d)
Figure 11: Percolation process on the SEAN. (a) The relative sizes of the LCC1 and the LCC2 are
shown as a function of q. The sudden decrease of the size of the LCC1 and the increase of the size
of the LCC2 indicate that, at the percolation threshold qc , the LCC1 is broken into several isolated
clusters because of the closure of some links. Therefore, the network transits from a largely connected
phase to a disconnected phase at qc . The LCC1 of SEAN (b) at and (c) after the critical threshold
qc with respect to the percolation process. (d) The critical links with respect to the critical threshold
qc . Each node number represents a unique node in the network.

airway network can be constructed for a given q value according to the traffic dynamics
of the original airway network.

As q increases, the original network will be disintegrated into several isolated clus-
ters because of the closure of some low-speed links. Therefore, the size of the largest
cluster decreases, and the second-largest cluster reaches a maximum at the percola-
tion threshold qc , which is the transition to the disconnected phase from the connected
phase of an airway network. Fig. 11a demonstrates the percolation process. The y-
axis of Fig. 11a refers to the fraction of the size of the largest connected component

45
(LCC1) or the second-largest connected component (LCC2), which is a value between
0 and 1.

As an indicator of the robustness characteristics of network connectivity [180],

the percolation threshold qc in this percolation process quantifies the organization
efficiency of real air traffic. Flights can travel to most nodes in the airway network
(the largest connected component of the airway network) with normalized speeds below
qc , while flights will be trapped in small isolated clusters when they fly with normalized
speeds above qc . Hence, qc measures the maximum normalized speed with which flights
can travel over a large part of the airway network, reflecting the global efficiency of
air traffic from a network perspective.

2) Critical Links Identification

At the critical threshold qc , the LCC1 of an airway network exists, and the LCC2 is
relatively small in size. As reflected in Fig. 11a, when more links are further removed
from the LCC1, the LCC1 will break down into small pieces. As a consequence, the
size of the LCC2 grows. Fig. 11b, Fig. 11c, and Fig. 11d further take SEAN as an
example to demonstrate this process visually. Fig. 11b displays the structure of the
LCC1 in the weighted SEAN at the critical threshold qc with respect to the percolation
theory. In Fig. 11c, the LCC1 fragments into pieces after qc due to the removal of the
critical links. Fig. 11d shows the critical links. Therefore, the identification of critical
links using the percolation theory is formulated as follows:

Ec = {e ∈ E | we < qc + δ} (8)

in which δ is the interval length for the variation of variable q, q ∈ [0, 1].

3.3.4. Critical Link Determination

For each snapshot Gti of the constructed spatial-temporal networks, there are two
sets of detected critical links, one yielded by the centrality metric and the other by the

46
percolation theory. Note that the edge betweenness centrality helps identify links that
act as the traffic pivots through which the shortest paths frequently pass, while the
percolation theory identifies links that act as the bridges whose failures will decrease
the network’s structural integrity significantly. Links identified by both percolation
theory and betweenness centrality can meet the requirements of critical links defined
in this thesis, i.e., pivot links contributing to shortest flight paths and bridge links
contributing to structural integrity. Therefore, the overlapped links of the two link
sets are determined as the final critical links Ecti of Gti .

The critical links Ecti are identified for time slice ti . Note that the critical links will
evolve over time following different traffic situations. This critical link determination
process will dynamically provide critical links for different periods. By observing the
changes of critical links, temporal distribution of the critical links can be obtained. By
further merging Ecti across the time horizon, a holistic view of the spatial distribution
of the critical links for the airway network G can be achieved.

3.4. Experimental Study

3.4.1. Network Modeling

The above section describes the proposed method for critical link detection in an
airway network with given flight track data. To check the efficacy of the proposed
method, this section carries out a case study on the SEAN. The SEAN sits on the
juncture of six neighboring FIRs and covers south China sea airspace, where most of
the air traffic from China converges towards South-East Asia. Despite its small size,
it has features of various airspace, e.g., radar, procedural, and oceanic. Moreover, the
SEAN has complex traffic structures comprising the confluence of en-route air traffic
from neighboring airspace, climbing traffic from the terminals, and descending traffic
to the terminals. The complexity of the traffic structure and the high traffic demand
are likely to induce air traffic congestion during peak hours, making it significant to
detect critical links in such an airway network. The network structure of SEAN is

47
shown by blue lines in Fig. 12a. The black dashes represent the sector boundaries of
Singapore FIR.

The SEAN shown in Fig. 12a consists of 118 nodes and 174 links. In the exper-
iments, one-month (1st December 2018 to 31st December 2018) en-route flight track
data provided by the Civil Aviation Authority of Singapore is used. December is the
peak season for air transportation due to the increasing travel demand during the
holiday. During this period, the high traffic demand provides the advantage and pos-
sibility to reveal the critical links in airway networks which significantly affect the
network’s performance in terms of structural integrity and functionality. The tested
one-month data records the information of 44215 flights, including flight fixes and the
time passing those fixes.

Based on the one-month flight track data, the weighted spatial-temporal airway
networks can be constructed. Note that the distributions of the critical links may
change over time. Therefore, the spatial-temporal airway networks are constructed
over time for different time slots. Fig. 12b displays a snapshot of the weighted spatial-
temporal airway networks for SEAN during a 30-minute time slot.

3.4.2. Critical Link Detection

To study the evolution of critical links overtime under different traffic situations,
spatial-temporal networks are constructed for different time slots characterizing the
evolution of traffic situations. Moreover, the length of the time slots chosen may influ-
ence the critical link detection outcome. Therefore, in the experiments, the weighted
spatial-temporal airway networks are constructed with different time slot length.

In this study, the time slot lengths are set respectively to be 15 minutes, 30 minutes,
45 minutes, and 60 minutes, resulting in four sets of spatial-temporal networks. The
reason for adopting the four slot lengths between 15 to 60 minutes is that “15-minute”
and “60-minute” (1 hour) are the commonly used horizons for evaluation of controllers’
workload and air traffic planning purposes [181–184]. For example, in the MAP model,

48
 ELSAS
LAXOR

TEGID
AKMON
BIDAG
TIDAR IPRIX ESPOB MELAS LAGOT
UTHAI DUDIS
ODONO TODAM
UPRON ASISU
ENREP
DOLOX OLKIT
LATIL
TEPUS DAMOG MABLI TERIX GULIB
PULIP TAXULDUBSA
ELGOR LUSMO
KIMAT VPKEGOLO RIZAN
SAROX LIPRO OMLIV
VMR BOBOB KAMIN
AROSO OLSAM AGOBA OKTOG
AKOMA OBLOT
HOSBA TOMAN NIMIX ATETI
ASUNA ANIPU
TAROS REMES ARUPA
REPOV ATPOM OBDOS
KADAR BAVUS
ANITO KIKOR
PARDI SURGA

(a)

(b)

Figure 12: (a) The network structure of SEAN with spatial information. The black dashes represent
the sector boundaries of Singapore FIR, and the green nodes represent the waypoints whose names
are shown in red letters. (b) The weighted SEAN with the weights being the normalized average
flight speed on the links.

the airspace capacity is computed on a 15-minute basis [185].For the one-month en-
route flight track data from 1st December 2018 to 31st December 2018, there will
be 2976, 1488, 992, 744 weighted networks constructed over time when the time slot
lengths are set as 15 minutes, 30 minutes, 45 minutes, and 60 minutes, respectively.
Then the critical links will be detected from each set of the networks using both
percolation theory and edge betweenness centrality metric.

49
#Critical links
10

0
Dec 01 Dec 04 Dec 07 Dec 10 Dec 13 Dec 16 Dec 19 Dec 22 Dec 25 Dec 28 Dec 31
Time 2018

(a) 15 minutes
#Critical links

10

0
Dec 01 Dec 04 Dec 07 Dec 10 Dec 13 Dec 16 Dec 19 Dec 22 Dec 25 Dec 28 Dec 31
Time 2018

(b) 30 minutes
20
#Critical links

10

0
Dec 01 Dec 04 Dec 07 Dec 10 Dec 13 Dec 16 Dec 19 Dec 22 Dec 25 Dec 28 Dec 31
Time 2018

(c) 45 minutes
20
#Critical links

10

0
Dec 01 Dec 04 Dec 07 Dec 10 Dec 13 Dec 16 Dec 19 Dec 22 Dec 25 Dec 28 Dec 31
Time 2018

(d) 60 minutes
Figure 13: The number of critical links detected from 1st December 2018 to 31st December 2018
using the percolation theory applied to the spatial-temporal SEAN with different time slot lengths.
The average number of critical links detected for each snapshot under 15 minutes, 30 minutes, 45
minutes, and 60 minutes are 1.9543, 2.5673, 2.8697, and 3.1116 respectively.

3.4.2.1. Variations of the Number of Critical Links. Fig. 13 visualizes the temporal
distribution of the number of critical links detected by using network percolation
theory. In the network percolation process, the interval δ for the variation of variable
q, q ∈ [0, 1], is set to be δ = 0.001. The curves in Fig. 13 show that the number
of detected critical links varies over time. However, the maximum number does not
exceed 18, while the average number of critical links detected for each snapshot under
15 minutes, 30 minutes, 45 minutes, and 60 minutes are 1.9543, 2.5673, 2.8697 and
3.1116, respectively.

It can be observed from Fig. 13 that as the length of the time slot for constructing
the temporal networks increases from 15 minutes to 30 minutes, 45 minutes, and 60
minutes, there is also a slight increase in the number of the critical links identified.
The possible reason could be that the fluctuations in the traffic situation on some links

50
and the influence of a single link over the network percolation process are neutralized
with the increase of the time slot for network construction. More specifically, critical
link detection for a 15-minute slot identifies the critical links for a short period. When
we look at a longer period, such as 60 minutes, different critical links can exist during
different short sub-periods of the long period. Therefore, the criticality of the varying
critical links during different short periods will be neutralized. All of the varying
critical links can appear critical for this long period. Thus, the number of critical links
can increase with the length of the time slot.

Moreover, it can be observed from Fig. 13, especially from Fig. 13(a), that the
number of identified critical links seems to change cyclically daily. The troughs in
the curve usually appear from 18:00 UTC to 23:00 UTC, especially around 21:00
UTC. From 23:00 UTC to the following 18:00 UTC, some peaks in the identified
critical links show up. The reason for such a phenomenon is likely to be the hub
airport nature of Singapore Changi airport. This oscillatory behavior of the number
of identified critical links is due to the hub nature of Singapore Changi Airport. An
overview of the hourly level of activity of Singapore Changi Airport reveals distinct
patterns related to its connectivity. Changi airport, acting as a hub (the airport as an
intermediary location), has several noticeable surges of activity during the day. Surges
at hub airports are often characterized by several inbound flights arriving within a time
frame, and about 2 hours later, a surge of outbound departures can be observed. For
most hub airports, there is a peak of activity around 7 AM and another peak around
7 PM, which mostly corresponds to short-haul flights and preferences for passengers
to depart in the morning and return in the evening.

As explained in Section 3.3.3.1, for the centrality metric, the 10 (K = 10) links
with the highest centrality values are chosen to be the critical links of each weighted
network. The main reason for choosing K = 10 in this experiment is that the number
of the critical links detected using the percolation theory is usually less than 10 (as

51
Figure 14: Critical links detected using the betweenness centrality metric (left column) and perco-
lation theory (right column) when applied to the spatial-temporal networks formed under different
time slot lengths (i.e., 15 minutes, 30 minutes, 45 minutes, and 60 minutes). The link thickness is
proportional to the frequency of being identified as a critical link.

52
shown in Fig. 13). Note that K cannot be too large as critical airway network links
are in the minority. Meanwhile, K cannot be too small as the detection results need
to be compared with what is detected using the percolation theory. As a consequence,
K = 10 is a suitable value for this experiment. For different airway networks, the
suitable choices of K could vary. The parameter K needs to be tuned based on the
merits of the situation.

3.4.2.2. Spatial Distribution of the Detected Critical Links. For different network snap-
shots, the detected critical links could be unique. Therefore the spatial distributions
of the detected links are further compared. Specifically, for each critical link detec-
tion metric, i.e., network percolation and edge betweenness, the critical link detection
results in each weighted network are integrated by counting the frequencies of each
link being detected as critical in all time slots. Fig. 14 presents the overall spatial
distribution of the critical links in the SEAN detected using the betweenness central-
ity metric and percolation theory, respectively. In Fig. 14, the critical links detected
by the betweenness centrality are marked in purple, while the ones detected by the
percolation theory are marked in red. The thickness of each critical link in the figure
is proportional to its frequency of being identified as critical.

Fig. 14 shows that the frequencies of the critical links detected using the centrality
metric are relatively higher than that of percolation theory. The reason is that each
network snapshot has ten critical links when the centrality metric is of concern. While
using percolation theory, the average number of critical links detected for each snapshot
is 2 to 3. Whichever method is used, Fig. 14 demonstrates that only a small portion
of the detected critical links have relatively high frequencies.

Fig. 14 presents the detected critical links that vary in number and frequency
under different time slot lengths for constructing the corresponding spatial-temporal
weighted networks. It can be observed that, as the time slot length of the network
snapshots increases, the frequency of the detected critical links decreases (thickness of

53
the red/purple links reduces). The main reason is that fewer network snapshots will
be constructed based on the one-month traffic data when the slot length increases.
A short time slot captures the air traffic within a short window, thus providing a
microscopic view to investigate the network dynamics. A long time slot gauges the
air traffic over a long period, providing a macroscopic view of the network dynamics.
When constructing the temporal networks, there is no need to fix the time slot. As
a result, a decision-maker can choose a proper time granularity concerning a specific
task and purpose.

Figure 15: The distribution of the number of final critical links determined from 1st December 2018
to 31st December 2018 in the spatial-temporal SEAN modeled with different time slot lengths, i.e.,
15 minutes, 30 minutes, 45 minutes, and 60 minutes. The X-axis represents the number of critical
links identified, while the Y-axis shows the corresponding frequency.

The above experiments mainly demonstrate the overall comparisons between the
spatial distributions of the critical links detected by the edge betweenness centrality
and the percolation theory. In what follows, the critical link determination results are
presented.

54
3.4.3. Critical Link Determination

3.4.3.1. Number of the Determined Critical Links. For each time slice, the critical
links of the corresponding network snapshot are determined as the overlapped links of
the two critical link sets detected using the two network metrics. The distribution of
the number of the final critical links identified for the network snapshots is shown in
Fig. 15.
By comparing Fig. 13 and Fig. 15, it can be observed that the number of critical
links determined for each network snapshot reaches a reasonable level (between 1 to
5) after overlapping the critical links determined using the network percolation metric
and the edge betweenness centrality. Fig. 15 also shows that multiple critical links
exist under some time slots, while most of the time, only one critical link is identified.

3.4.3.2. Temporal Distribution of the Determined Critical Links. The proposed net-
work approach can detect the critical links in a given airway network both spatially
and temporally. Fig. 16 demonstrates the temporal distribution of the determined
critical links in the SEAN. In Fig. 16, the X-axis represents the link ID of the 174
links in the SEAN, while the Y-axis represents the time. More specifically, the X-axis
view of Fig. 16 shows the exact links out of the 174 links that have been detected as
critical for a given time horizon, while the Y-axis view of Fig. 16 presents the criti-
cality evolution (critical or non-critical) of a specific link over time. The black block
(ID, time) illustrates that the link with the corresponding ID on the X-axis has been
identified as a critical link at the corresponding time on the Y-axis.
It can be observed from Fig. 16 that there exist four links that are frequently
identified as critical. These links are (identified by entry-exit waypoints) as follows:
“MABAL – VISAT”, “RAXIM – VMR”, “KILOT – OTLON”, and “KILOT – LIPRO”
(marked with red arrows and their corresponding names).

3.4.3.3. Spatial Distribution of Determined Critical Links. Fig. 17 visualizes the spa-
tial distributions of the final determined critical links in the SEAN. The red links

55
 (a) 15 minutes

(b) 30 minutes

(c) 45 minutes

(d) 60 minutes

Figure 16: Critical links identified over time under different time slot lengths (i.e., 15 minutes (a),
30 minutes (b), 45 minutes (c), and 60 minutes (d)) in modeling the spatial-temporal networks. The
X-axis, ranging from 1 to 174, represents the 174 links in the SEAN, while the Y-axis represents
the period from 1st December 2018 to 31st December 2018. When a link (whose coordinate is x
on the X-axis) is identified as critical during a time slice (whose coordinate is y on the Y-axis), the
corresponding area of the coordinate (x, y) will be marked as a black block.

denote the determined critical links, and the thickness of the links is proportional to
their frequencies.

It can be seen from Fig. 17 that the majority of the critical links are located
in the sector highlighted in purple. This sector is the most heavily loaded sector
in the airspace covered by SEAN, within which traffic complexity and density are
distinctly higher than in other sectors [186]. Also, it can be observed that as the
time slot length of network snapshots increases, the number of detected critical links
together with their frequencies decreases (thickness of the red links reduces). Two
reasons are attributable to this phenomenon. First, fewer network snapshots will be
constructed based on the one-month traffic data when the time slot length becomes

56
 (a) 15 minutes (b) 30 minutes

(c) 45 minutes (d) 60 minutes

Figure 17: Spatial distributions of the critical links in the SEAN. The red links are the determined
critical links. The link thickness is proportional to the frequency of being identified as a critical link.

larger. Therefore, the frequencies decrease as there are fewer network observations.
Second, each spatial-temporal network snapshot is constructed based on the traffic
for a given period. If the time slot length for constructing the weighted networks is
long, the weights on the airway links do not distinguish from each other, resulting in
homogeneous network observations. Consequently, both the centrality metric and the
percolation theory will not be effective on a homogeneous network as each link in the
network acts importantly the same as others do.

3.4.4. Validation of the Identified Critical Links

The above experiments have demonstrated the experimental study on the critical
link detection in the SEAN using the suggested network theories. In this section, the
criticality of the detected critical links is validated from three perspectives, i.e., air

57
traffic volume perspective, airspace design perspective, and operational perspective.

3.4.4.1. Air Traffic Volume Perspective. It can be observed from Fig. 15 that the
maximum number of the simultaneously detected critical links in the SEAN is 5. Note
that there were a total number of 44215 flights passing through the SEAN in December
2018. The ratio of flights passing each of the five critical links, appearing with the
highest frequency under different time slot lengths, over the total number of flights is
presented in Table 1.

Table 1: The ratio of flights (r) on each of the five high-frequency links being detected as critical
links under different time slot lengths.

15minutes 30minutes 45 minutes 60 minutes

Link Name r (%) Link Name r (%) Link Name r (%) Link Name r (%)
MABAL – VISAT 10.68 MABAL – VISAT 10.68 KILOT – OTLON 5.68 RAXIM – VMR 5.69

KILOT – OTLON 5.68 KILOT – OTLON 5.68 AKOMA – VMR 13.37 AKOMA – VMR 13.37

LUSMO – OPULA 10.69 RAXIM – VMR 5.69 MABAL – VISAT 10.68 KILOT – OTLON 5.68

LUSMO – TERIX 9.90 KILOT – LIPRO 5.68 RAXIM – VMR 5.69 MABAL – VISAT 10.68

KILOT – LIPRO 5.68 LUSMO – OPULA 10.69 KILOT – LIPRO 5.68 KILOT – LIPRO 5.68

Note that the four links marked in bold in Table 1 are the commonly identified
critical links under different time slot lengths. It can be observed from Table 1 that the
four links have high relative usages by flight as the ratios of being transited through
by flights are large. If anyone of the four links is blocked due to weather or airspace
restriction, the corresponding portion of flights, as recorded in Table 1, will be directly
affected. For example, if the link “MABAL – VISAT” is blocked, 10.68% of the total
flights will be affected, and air traffic control operations, such as flight re-routing,
speed control, and vectoring, would be required.

3.4.4.2. Airspace Design Perspective. Jet routes are equipped with ground-based nav-
igation beacons such as VOR/DME stations. A VOR/DME beacon emits radio signals
to provide surveillance information (range and bearing) for flights to navigate through
the sky [187, 188]. VOR/DME stations serve as important navigational aids connecting

58
all significant traffic flow sources [189]. Fig. 18 presents a snapshot of the aeronautical
chart of SEAN in which the four high-frequency critical links are annotated.

It can be seen from Fig. 18 that the three links “MABAL – VISAT” (on airway
M758), “KILOT – LIPRO” and “KILOT – OTLON” (on airway M761) lie on airways
(radio signals) radiated from the PEKAN VOR/DME, which pilots and controllers
will primarily choose for easy use of navigation. Moreover, links “KILOT – LIPRO”
and “KILOT – OTLON” are on airway M761 between the outbound radio of two
VOR/DMEs (PEKAN VOR/DME and KUCHING VOR/DME). This airway serves
the heavy traffic between Kuala Lumpur (in west Malaysia) and Kuching (in East
Malaysia) [190]. Link “MABAL – VISAT” is on the airway from PEKAN VOR/DME
to waypoint “LUSMO”, the metering point for flights entering sector 5 in Singapore
airspace and flying to Indonesia, Philippines, East Malaysia, and Japan. Waypoints
“MABAL”, “VISAT”, “OTLON”, “KILOT”, and “LIPRO” are the crossing points for
airways radiated from the MERSING and PEKAN VOR/DME, which serve major
air traffic flows from airports in Singapore and Malaysia. Link “RAXIM – VMR”
caters to the heavy traffic flow between Singapore and China. Waypoint “VMR” is
at the MERSING VOR/DME, a crucial metering point for flights from/to Singapore
airports. Additionally, apart from the four links, the rest of the seven links listed in
Table 1 (“LUSMO – OPULA”, “LUSMO – TERIX”, and “AKOMA – VMR”), are all
located on airways radiated from VOR/DMEs.

3.4.4.3. Operational Perspective. This section presents the validation of the criticality
of the identified links from an operational perspective, considering the feedback from
operational experts. The detection results are validated from an operational perspec-
tive, including i) spatial view based on airspace traffic flow structures and ii) temporal
view based on real traffic scenarios.

i) Spatial view based on airspace and air traffic flow structures

Table 2 summarizes the characteristic information of the four critical links “MABAL

59
Figure 18: Geographical positions of critical links 1 - “MABAL – VISAT”, 2 - “KILOT – OTLON”, 3 -
“KILOT – LIPRO” and 4 - “RAXIM – VMR”. The picture is excerpted from the website of SkyVector
[191].

– VISAT”, “KILOT – OTLON”, “KILOT – LIPRO” and “RAXIM – VMR”. Subject

Matter Experts (SMEs) use this information to analyse the spatial criticality of the
identified critical links. In the following paragraphs, the analysis of each of the four
links will be illustrated one by one.

1) Spatial criticality of link “MABAL – VISAT”: Airway link “MABAL

– VISAT” is located on the ATS route M758, which is a bidirectional airway. It
accommodates high traffic volumes as the ratio of flights in the SEAN transiting
through “MABAL – VISAT” is 10.68%. Given the short length of the link, which is 41
nautical miles (nm), “MABAL – VISAT” possesses a high traffic density considering
the peak number of aircraft transiting through the link is 6 ~ 7 per 15 minutes derived
from the traffic data.

Table 2 lists the Entry – Exit waypoints of flights in the SEAN transiting through
the link “MABAL – VISAT”. The right-heading arrow indicates that the flights are
transiting in the direction presented in the table header, i.e., “MABAL – VISAT”,
and vice versa. The corresponding flight paths, connecting these Entry – Exit pairs
of flights transiting in the direction of “VISAT – MABAL”, are highlighted by green
dashes in Fig. 19, while paths in the direction of “MABAL – VISAT” are highlighted

60
by green dashes in Fig. 20. It can be seen from Fig. 19 that flights flying outbound
of Singapore FIR and taking the three major ATS routes, namely, L625, N884, and
M758, will transit through the link in “VISAT – MABAL” direction. The ATS route of
M758 facilitates the smooth flow of heavy air traffic between Peninsular Malaysia and
East Malaysia, Brunei [192]. ATS route M758 handles approximately 742 movements
a week in 2016 [192]. The unidirectional ATS routes N884 and L625 cater to the main
traffic flow flying eastbound to the Philippines, far east (Japan), USA, etc [190]. Fig.
20 shows that inbound flights, from Peninsular Malaysia, Brunei, Philippines and far
east via ATS routes M767 (unidirectional), M758 (bidirectional), merge to route M758
and transit through link “MABAL – VISAT” to the west. From this point of view, the
critical link “MABAL – VISAT” serves as a pivot to spread the eastbound flights in
the SEAN to the north-east world and caters to westbound flights entering the SEAN
from the north-east world.

The criticality of link “MABAL – VISAT” not only depends on its high traffic
density and presence on the major ATS route M758, but also on its position in the
airspace. Link “MABAL – VISAT” is located in the busiest sector in the airspace
covered by the SEAN, with the highest traffic load and complexity. The trunk route
M758 intersects with the major ATS routes M771 and N892, which cater for flights
to and from the north (China, Vietnam, and Thailand) [186, 192], at the waypoints
“VISAT” and “MABAL” respectively. Therefore, managing the confluence of the ATS
route M771 and N892 against the South China Sea air traffic flow on ATS routes
M758 is a demanding task and puts pressure on the link “MABAL – VISAT” due to
the complexity exacerbated by the high density of crossing air traffic [193].

2) Spatial criticality of link “LIPRO – KILOT” and “KILOT – OTLON”:

Airway links “LIPRO – KILOT” and “KILOT – OTLON” are on ATS route M761
(bidirectional airway). The two links handle 5.68% of flights in the SEAN. Both of
them have a short length, 32nm and 21nm respectively, and handle 4 ~ 5 flights per 15

61
Table 2: Characteristic information of critical links “MABAL – VISAT”, “KILOT – OTLON”, “KILOT
– LIPRO”, and “RAXIM – VMR”. The right-heading arrow “→” represents that the information is
for flights transiting in the same direction as presented in the table header, e.g., from “MABAL” to
“VISAT”, and vice versa. “Entry – Exit in SEAN” denotes the entry waypoint and exit waypoint in
the SEAN of flights transiting through the corresponding critical link, which are visualized with the
corresponding flight path in Figs. 19 and 20.

MABAL – VISAT LIPRO – KILOT KILOT – OTLON RAXIM – VMR

ATS route M758 M761 M761 M771

Direction bidirectional bidirectional bidirectional unidirectional
Length (nm) 41 32 21 47
Flight ratio 10.68% 5.68% 5.68% 5.69%
Peak #aircraft/15mins 6~7 4~5 4~5 4~5
Mean speed (knots)→ 474.4 472.8 485.9 ———————–
Speed range (knots)→ 246.2 ~ 615.6 322.3 ~644.6 315.4 ~ 630.7 ———————–
Mean speed (knots)← 446.6 445.8 451.4 442.3
Speed range (knots)← 273.6 ~ 615.6 322.3 ~ 644.6 315.3 ~ 630.7 282.9 ~ 565.7
Entry–Exit in SEAN→ OLKIT – SAROX, AGOBA – SAROX AGOBA – SAROX
GULIB – SAROX KAMIN – SAROX KAMIN – SAROX
OLKIT – IDSEL AGOBA – OTLON AGOBA – OTLON ———————–
TEGID – SAROX KAMIN – OTLON KAMIN – OTLON
TEGID – IDSEL
GULIB – IDSEL
Entry–Exit in SEAN← KIMAT – OLKIT KIMAT – AGOBA KIMAT – AGOBA VMR – DUDIS
KIMAT – AKMON KIMAT – KAMIN KIMAT – KAMIN PARDI – DUDIS
KIMAT – LAXOR OTLON – AGOBA OTLON – AGOBA VTK – DUDIS
IDSEL – OLKIT OTLON – KAMIN OTLON – KAMIN PU – DUDIS
IDSEL – LAXOR OTLON – ARUPA OTLON – ARUPA VMR – AKMON
IDSEL – AKMON
KIMAT – TERIX
VMR – AKMON
KIMAT – SAROX
KIMAT – LUSMO

62
minutes. The traffic density on the two links reaches a relatively high level, considering
the short lengths of the two links, leading to a short space of time for reaction.

“LIPRO – KILOT” and “KILOT – OTLON” locate on M761, which is a trunk route
for air traffic in the east – west direction. Table 2 presents the Entry – Exit waypoints
in the SEAN of flights transiting through links “LIPRO – KILOT” and “KILOT –
OTLON”. The right-heading arrow indicates that the flights are transiting in the
westbound direction of “LIPRO – KILOT” and “KILOT – OTLON”, and vice versa.
The corresponding flight paths connecting the listed Entry – Exit pairs are highlighted
by red dashes in Fig. 19 (for eastbound flights) and Fig. 20 (for westbound flights). It
can be observed from Fig. 19 that the eastbound flights transiting through “OTLON
– KILOT – LIPRO” on ATS route M761 spread to ATS routes M761, M646, and
N875, which accommodate the major flows of air traffic between east Malaysia and
Peninsular Malaysia, Brunei, Philippine, and Indonesia [192]. Similarly, Fig. 20 shows
that the westbound flights enter Singapore FIR through ATS routes M646 and N875.
Flights then merge to ATS route M761 and fly to the west through links “LIPRO –
KILOT – OTLON”.

Similar to link “MABAL – VISAT”, besides their high traffic density and crucial
position, links “LIPRO – KILOT” and “KILOT – OTLON” are located within the
highly utilized sector (the sector highlighted in purple in Fig. 17). Additionally, ATS
route M761 crosses ATS routes M771 and N892, which handle major traffic flows to and
from the north SEAN [186, 192], at waypoints “OTLON” and “KILOT” respectively.
Traffic flow on ATS route M761 crosses the north-eastbound traffic flow on ATS route
N884 at waypoint “LIPRO”. The high density of crossing traffic and the short space of
time for reaction due to the short lengths of the two links have increased the pressure
on handling the high volume of air traffic on links “LIPRO – KILOT” and “KILOT –
OTLON”, making the two links critical in the SEAN.

3) Spatial criticality of link “RAXIM – VMR”: Link “RAXIM – VMR” is

63
Figure 19: Routes of flights in the SEAN transiting through airway links “VISAT → MABAL” (green
dashes), “OTLON → KILOT → LIPRO” (red dashes) and “VMR → RAXIM” (purple dashes).

on the unidirectional ATS route M771, accommodating outbound traffic flows heading
north. Therefore, only the south-to-north flights in the SEAN will transit through it,
and there will be no flights on “RAXIM – VMR” in the opposite direction, i.e., north
to south. It is a highly utilized link through which around 5.69% flights in the SEAN
will transit. The length of “RAXIM – VMR” is 47nm, and the peak number of flights
on “RAXIM – VMR” is normally 4 ~ 5 per 15 minutes.

Link “RAXIM – VMR” has an important position in the SEAN. By referring to

Fig. 18, it can be seen that waypoint “VMR” is located at the MERSING VOR/DME,
which is a crucial beacon point for aircraft flying in/out of Singapore. The majority
of northbound traffic flows in the SEAN will transit through “RAXIM – VMR”. The
Entry – Exit waypoints of flights transiting through the link “RAXIM – VMR” are
shown in Table 2. The corresponding flight paths connecting the Entry – Exit pairs
are highlighted by purple dashes in Fig. 19. Combining Fig. 19 and Table 2, it can
be observed that three major traffic flows will converge at “VMR” and head to the

64
Figure 20: Routes of flights in the SEAN transiting through airway links “MABAL → VISAT” (green
dashes), “LIPRO → KILOT → OTLON” (red dashes).

north via the link “RAXIM – VMR” on route M771: (1) northbound flights taking off
from Singapore, transiting through waypoints “PU”, “VTK” and heading to “VMR”,
(2) northbound traffic flow from Johor Bahru (in Malaysia), (3) northbound traffic
flow from Jakarta FIR via ATS routes G579 (one of the busiest international routes
in this region) whose destination is beyond Singapore. The pressure of handling the
confluence of traffic at “VMR” and accommodating the northbound flights to transit
onto ATS route M771 adds to the criticality of “RAXIM – VMR”.

The convergence of different types of traffic also contributes to the criticality of

link “RAXIM – VMR”. Departure flights from Singapore and Johor Bahru, which
are still in their phase of climbing to the cruise level, will need to step climb on the
link “RAXIM – VMR” due to en-route traffic from Jakarta. Moreover, traffic on the
link “RAXIM – VMR” enters the busiest sector in Singapore FIR at “RAXIM” and
immediately crosses the west-east direction traffic on ATS route M761, which leads to
a potential area of conflict at “RAXIM” and affects the smoothness of traffic flow on

65
“RAXIM – VMR”.
In summary, all four links accommodate a high volume of air traffic flow in the
SEAN. Traffic flows on ATS routes where the four links are located constitute a sig-
nificant part of traffic flows within the SEAN [192]. The four links serve a high traffic
density due to their short lengths and the heavy traffic demands. The four links are
inside or connecting to the heavily utilized sector, i.e., the sector highlighted in Fig.
17, within which trunk routes in different directions intersect with each other. The
four links lie on the trunk ATS routes, and their waypoints are the intersection points
of several trunk links. Considering the above factors, emergencies, such as system
failure, will pose significant inefficiency challenges on these links, and the controllers
must anticipate and solve such issues immediately.

Figure 21: Visualization and simulation of air traffic data for SMEs’ analysis. The critical links
identified under the corresponding traffic scenario are marked as red segments. Two flight levels
out of eleven are presented in this figure. The critical links identified evolve in accordance with the
changes in air traffic.

ii) Temporal view based on real-time traffic scenarios

The above analyses manifest the criticality of the critical links from a spatial view.
To manifest the efficacy of the critical link identification results from a temporal per-
spective, the real-time flight movements are visualized with flight information (callsign,
aircraft type, speed, flight level), as shown in Fig. 21. Two layers of flight traffic are

66
shown in this figure as an example. The blue dots represent flights in the airspace.
The corresponding critical links identified under these traffic scenarios are marked red
and change dynamically over time.
This visualization will allow the dynamic observation of traffic situations and val-
idation of the real-time critical link identification results at different periods. This
section shows two examples of air traffic scenarios and the corresponding critical links
identified.

(a) MABAL – VISAT (b) RAXIM – VMR

Figure 22: Traffic scenario examples under which “MABAL – VISAT” or “RAXIM – VMR” is identified
as a critical link. The critical links are marked in red. Each blue dot represents one flight in the
SEAN.

As shown in Fig. 22a, at UTC time 1st December 2018 03:12:00, “MABAL –
VISAT” (marked in red) is identified as a critical link. This time is 11:12:00 local time
in Singapore. As shown in the figure, a high volume of flights transit through the
link “MABAL – VISAT” on ATS route M758, from both eastbound and westbound.
Meanwhile, many north-east direction flights fly along ATS route M771 and cross the
traffic flow on M758 at waypoint “VISAT”. This situation leads to a potential area of
conflict at “VISAT”. On the rest of the network, either the traffic density is not high,
or the pressure of handling crossing traffic is low, making link “MABAL – VISAT”
critical at this period.
As shown in Fig. 22b, “RAXIM – VMR” is identified as a critical link (marked

67
in red) at UTC 1st December 2018 05:41:00. UTC 05:41 is between 13:00 and 14:00
local time in Singapore. At this time, more east-west bound flights are transiting
through ATS routes M758 and M761, which can also be observed from Fig. 22b.
Meanwhile, many flights transiting from Singapore to MERSING (“VMR”) are flying
in a north-north-east direction via ATS route M771. Traffic flow on M771 enters the
sector filled with heavy traffic at waypoint “RAXIM” and immediately crosses the east-
west direction traffic flow on M761. This situation puts potential conflict pressure on
“RAXIM”. Moreover, departure flights on “RAXIM – VMR” will need to step climb
to the cruise level due to the crossing traffic on M761, adding to traffic complexity on
the link “RAXIM – VMR”. The above facts make “RAXIM – VMR” a critical link at
this moment compared to other links.

The above analyses from the perspective of traffic volumes, aeronautical charts,
and the operational perspective manifest that the proposed method is effective in
identifying critical airway links and can dynamically identify critical links over time
in accordance with changing traffic conditions.

3.5. Chapter Summary

Note that identifying critical links in an airway network can assist with air traffic
flow management, flight scheduling, and resource allocation. This chapter proposed
complex network models to dynamically detect critical links in a given airway network.
In order to quantify how critical a link of an airway network is, two metrics were
introduced, i.e., edge betweenness centrality (identify links acting as the traffic pivots
through which shortest paths frequently pass) and percolation theory (identify links
acting as the bridges whose failures will decrease the network’s structural integrity
significantly). As the critical links of an airway network can vary over time, spatial-
temporal airway networks were constructed based on flight track data. Then the two
network metrics were individually applied to each network snapshot for critical link
detection, and their results were spatially intersected to determine the final critical

68
links.

The proposed methodology is generic in the sense that it can be applied to any air
traffic network given the sufficient air traffic data. However, critical link detection in
some air traffic networks might be challenging. For example, in European airspace, air
traffic has significant vectoring, while in Chinese airspace, air traffic usually adheres
to flight plans in en-route airspace. The proposed method was applied on the SEAN
with one-month flight track data. The detection results showed that the critical links
in the SEAN vary over time. Most of the critical links were concentrated in the sector
that witnessed heavy transition traffic in the airspace covered by the SEAN. Some of
the critical links appeared with a high frequency and amongst which the four airway
links “MABAL – VISAT”, “KILOT – OTLON”, “KILOT – LIPRO”, and “RAXIM –
VMR” distinguished themselves from the rest.

Furthermore, the four critical links locate at airways connecting two navigation aids
(VOR/DME) or one navigation aid (VOR/DME) and an important metering point.
Observations from aeronautical charts showed that waypoints “MABAL”, “VISAT”,
“OTLON”, “KILOT”, “LIPRO” and “RAXIM” are the crossing points of airways on the
outbound radial of three navigation aids (PEKAN VOR/DME, KUCHING VOR/DME,
and MERSING VOR/DME), while the waypoint “VMR” is a metering fix. These ob-
servations manifest that the detected critical links based on the proposed method
have operational significance. Further operational analysis by controllers validates the
operational criticality of the detected critical links from both spatial and temporal
views.

It is expected that the proposed method, which is based on complex network theory,
can help identify dynamic airway links that are operationally critical as identified by
SMEs. Moreover, considering the temporal nature of the proposed method, with
good traffic flow prediction tools in the future, this method can be adopted to predict
critical links in airway networks, which can facilitate better resource allocation in the

69
airspace and assist controllers in real-time air traffic management. Once a link is
identified as critical for a given period, ATFM measures can be applied in advance
to prevent the potential failure of the critical link to reduce its impact on the flow of
air traffic. At the strategic planning and pre-tactical planning stages of ATFM, by
examining the forthcoming demand and assessing the traffic pressure on the critical
links (such as the aforementioned four critical links), measures can be taken to balance
the traffic pressure and operational efficiency on the critical links, such as arranging
with the ANSPs to provide adequate capacity on the critical links at the required
time, optimizing air traffic flows to reduce the traffic pressure on the critical links,
scheduling or rescheduling flights to avoid the critical links, and deciding the need for
tactical ATFM measures on the critical links. For tactical ATFM operations, traffic
re-routing and flight level allocation can be applied according to the changing traffic
situation and the corresponding critical links to ensure the smooth flow of air traffic
through the air traffic network.

70
Chapter 4. Airway Network Structure Management

The research in this chapter has been published in the journal paper: Chunyao Ma, Qing Cai, Sameer

Alam, Banavar Sridhar, and Vu N. Duong, “Airway network management using Braess’s Paradox,”

Transportation Research Part C: Emerging Technologies, vol. 105, pp. 565-579, 2019.

Apart from identifying the critical links, applying changes to airway networks may
provide more flexibility to flight schedules and improve the flight efficiency in the
airspace. Thus, it is necessary to investigate novel methods for restructuring airway
networks for better air traffic flow accommodation.

Most existing airway network optimization methods end up with complete new net-
work designs whose implementation can be impractical due to operational constraints
and a variety of stakeholders. Note that decreased flight duration can reduce the buffer
time for airborne congestion as well as the risk of delay propagation through the entire
airway network [179]. Therefore, this chapter aims to investigate whether it is possible
to make minimal changes to the structure of the airway network in order to reduce
the flight duration and increase flight efficiency. More precisely, can the airspace per-
formance be improved by removing some of the airways, i.e., network links, from the
air traffic network? The answer may be yes for some networks.

This counterintuitive phenomenon was observed in road transportation networks

in 1969. When a new road was constructed in Stuttgart, the travel time of users was
worsened and only improved after the road was closed. In 1990, when 42nd street
was closed in New York City, the travel time of users was actually reduced instead
of the anticipated traffic congestion. These paradoxical phenomena are named as the
Braess’s Paradox (BP) [194] [195]. Under the future air traffic management paradigm
where aircraft operators can make the user preferred route (UPR) choices, BP could
be a common phenomenon in the network. However, to the best of my knowledge,
this phenomenon has not been explored or investigated in air transportation networks.

71
One possible reason might be the centralized air traffic management and the lack of
methods for large air traffic data processing.

To detect the existence of BP in airway networks and identify the links leading
to BP, this chapter first quantifies the performance of an airway network as the total
duration spent by all the aircraft on the given airway network. Then a generic method
is proposed for BP detection for a given airway network. The proposed method consists
of three key steps. The first step mainly involves pre-processing air traffic data and
modeling the airway network. The second step is about BP detection, which identifies
the potential links that may cause BP. The third step is BP verification which aims
to validate whether the links suggested by the second step are indeed the links that
cause BP.

4.1. Braess’s Paradox

BP provides an explanation for the situation where an alteration to a traffic network

to improve traffic condition actually has the reverse effect and impedes traffic through
it. BP was named after Braess, a mathematician who articulated the phenomena
[194, 195].

The simple network shown in Fig. 23 is used to demonstrate the occurrence of

BP. As illustrated in Fig. 23, many network users travel from the origin ‘A’ to the
destination ‘D’. Assume that each user chooses an A − D path independently and
selfishly, so as to minimize their own time cost. For simplicity, it is assumed that the
cost of a road link for drivers is equal to the the travel time.

The network shown in Fig. 23 consists of 5 links: L1 : A → C, L2 : B → D, L3 :

A → B, L4 : C → D, L5 : B → C. The flow on link Li is denoted as ϕi . There are
three paths from the origin A to the destination D: P1 : A → C → D, P2 : A → B →
D, P3 : A → B → C → D. The flow choosing path Pi is denoted as fi . Considering
the flow on each link equals the sum of flows on paths passing through the link, the
flows on the five links in Fig. 23 can be represented as:

72
 B

L  I   L  I

L  I  

A
L  I L  I  

C
Figure 23: A graphical network example presented in [5] to illustrate the BP phenomenon.

ϕ1 = f1 , ϕ2 = f2 , ϕ3 = f2 + f3 , ϕ4 = f1 + f3 , ϕ5 = f3 (9)

Each link Li of the network is characterized by its cost function, which describes
the time spent by the users traveling through the link. As shown in Fig. 23, the time
cost t of each user on each link is defined as:

t1 = 10ϕ1 , t2 = 10ϕ2 , t3 = ϕ3 + 50, t4 = ϕ4 + 50, t5 = ϕ5 + 10 (10)

which indicates that the travel time spent by each user on L1 and L2 is ten times of
the traffic flow volume on the links. The travel time spent by each user on L3 and L4
is the traffic flow volume on the link plus a constant cost 50. The travel time for each
user on L5 is its flow volume plus a constant cost 10.

When a user chooses path Pi , his travel cost from the origin ‘A’ to the destination
‘D’ will be the sum of his travel cost on the links composing Pi . Therefore, the travel

73
cost on each path can be formulated as:

T1 = t1 + t4 , T2 = t2 + t3 , T3 = t3 + t4 + t5 (11)

Assume that each user on this network attempts to find a path on which the travel
cost is minimum for his own. Thus, when all users find their own optimal paths, the
system reaches an equilibrium. In other words, the travel cost is equal on all paths in
the network, which can be represented as:

T1 = T2 = T3 (12)

Therefore, the total travel time TEq

+
of all users on the network under equilibrium
is:
3
X 31Φ2 + 1010Φ
+
TEq = Ti = (13)
i=1
13

Φ is the total traffic flow in the network, which can be formulated as: Φ = f1 + f2 + f3 .

When link L5 is removed from the network, there remain two paths from the origin
A to the destination D: P1 : A → C → D, P2 : A → B → D. The total travel time
TEq for users can be determined in the same way of calculating TEq
+
before the removal
of L5 . Accordingly, TEq under equilibrium can be represented as:

2
X 11Φ2 + 100Φ
TEq = Ti = (14)
i=1
2

Assume that the traffic demand, i.e., the total amount of traffic flow, in the network
is 6, namely: Φ = 6. In the equilibrium status of the network shown in Fig. 23, each
of the three paths, P1 : A → C → D, P2 : A → B → D, P3 : A → B → C → D,
has a flow of 2 with the corresponding cost for each user being 92. However, if the
link L5 : B → C in Fig. 23 is removed, in the ensuing equilibrium status, half
of the flow are attached to path P1 : A → C → D and the other half to path

74
P2 : A → B → D, resulting a cost of 83 for each user, which is better than that of
keeping link L5 : B → C. This phenomenon indicates that removing links can improve
the performance of the equilibrium flow of a selfish routing network.

BP happens when the total travel time on the network without link L5 : B → C
is less than that on the network with link L5 : B → C, that is:

+
TEq ≤ TEq (15)

By solving inequation 15, the interval of demand for BP to happen can be deter-
mined as: 0 < Φ < 80
9
. This indicates that, with a set of cost functions for each link
of a traffic network, BP will appear when the traffic demand is within a certain range.
In this simple network, the mechanism of BP is shown by simple calculations. When
it comes to large and complex networks, BP also widely exists, however, in a much
more intractable manner.

Since its discovery in 1968 [194], BP has generated significant research in the ground
traffic networks. BP has been studied concisely in its original network form [196–198]
and in a more general context. It is also proved that BP is likely to occur in a natural
random network model [199]. More precisely, for a given appropriate total flow, it is
shown that in almost all networks, there is a set of links whose removal improves the
travel time at the equilibrium status of traffic [199].

In a traffic network consisting of links associated with non-decreasing flow functions

known as cost functions, ‘users’ unilaterally choose the least cost (or shortest) path
on the network. The resulting flow is known as User Equilibrium (UE), also known
as Wardrop Principle 1 [200]. However, there is a better flow pattern if the users
cooperate with each other to come out with a flow that is best for all the users in
the system. ‘Best for all users’ means that the total cost incurred to all the users is
minimum. Such a flow is termed as System Equilibrium (SE), also known as Wardrop
Principle 2 [200]. When a new link is added to a network, it is supposed to reduce

75
the total cost in the system. However, the users may pursue a new UE under the
new network structure, and the total costs incurred to all users under the new UE
might become larger than the total costs of the previous UE on the original network.
BP is rooted in the behavior of the users: users choose their own least-cost routes
without any regard for how their choices may affect others. If the users decided to
collaborate to constitute SE, there would be no place for BP. In air transportation, this
is analogous to the way airlines do their flight planning exercises. For a given flight,
the Airlines Operations Center (AOC) chooses the best possible routes accounting for
aircraft performance and flight profile, subject to safety and operational constraints.
AOC carries out this exercise using its own business models without any consideration
for other airlines. ATCOs largely adhere to the flight plan, though they do make
tactical changes to ensure flight safety in conflict situations or bad weather conditions.

Previous investigations on BP were commonly based on the classical, symmetric

four-link network introduced by Braess. There are few studies on BP in large-scale
transportation networks with real-world transportation demand. Researchers prove
that BP detection is highly intractable, especially in real-world networks, and no
efficient method has been introduced. A heuristic methodology based on the Genetic
Algorithm has been studied to detect BP [201, 202], in which links that might cause
BP is identified by simply testing their closure one by one and then an algorithm
is adopted to run over these links to find a combination whose closure improves the
traffic cost. This method is effective for detecting BP in small-scale networks and is
therefore not applicable to large, complex real-world networks. In a different context,
a path-based formulation has been developed [203] to detect BP in a stable dynamic
network. The major limitation of this path-based (not link-based) formulation is that
its computational time cost increases exponentially if the network size grows.

Note that most studies on BP detection mainly identify the links/paths that cause
BP by enumerating all possible removals to see if the removals indeed improve the

76
travel costs on the focal transportation network, with their critical drawback being
the time-consuming implementations. In order to assist in better identification of the
links/paths that cause BP to a traffic network, the authors in [204, 205] presented
a variant to the BP model which aims to explore a solution that will make some
users better-off, while no user is worse-off compared with the solution to UE. To keep
consistency with UE and SE, here it is named as Braess’s Equilibrium (BE). As BP
detection in real-world networks is NP-hard [206], the BE model has proven to be
conducive to the detection of BP in traffic networks. Inspired by [204, 205], in this
research, the BE model is adopted to assist in identifying the potential airways/links
that cause BP.

4.2. Problem Description

An airspace, comprising several airways and intermediate waypoints, can be seen

as a network of links between waypoints. The horizontal projection of such an airway
network can be seen as a planar network with the nodes being the waypoints and links
the airways. An airway in an airway network is defined as a sequence of segments
starting and ending at fixed waypoints [133]. To keep en-route vertical separation, so
as to assure en-route flight safety, an airspace network often has a layered structure,
as shown in Fig. 24.

Fig. 24 illustrates the vertical distribution of flight trajectories on the layered

airway network. It can be seen from Fig. 24 that an airway network is practically in
a multi-layer structure. Flights have to follow certain airways in the airway network
at given flight levels.

As an airspace has its capacity limit, when air traffic demand reaches the capacity
threshold of the airspace, air traffic congestion is likely to occur [207]. In order to
mitigate congestion, one possible way is to add new airways to the airway network to
dispatch the congested traffic on alternate routes. Note that the design of an airway
network has to take into account many factors such as procedural, technical, and

77
 35000
34750
34500

Altitude (ft)
34250
34000
33750
33500
33250
33000
15
10 )
100 5 (deg
105 0 itude
Longit 110 115 t
ude (d
eg) 120 −5 La

Figure 24: The layered en-route structure airspace network from ADS-B data used in the study.

geographical constraints. An airway network can possess a very complicated structure.

Consequently, it is difficult to mitigate congestion by adding extra airways/links to an
airway network.
This chapter suggests a counter-intuitive method to optimize the airway network
with the inspiration drawn from BP. Specifically, for a given airway network at a given
flight level, this chapter aims to identify the airways/links whose removal will reduce
the total traveling time spent by flights without compromising the traffic demand.

4.3. Methodology

4.3.1. Methodology Overview

The proposed methodology encompasses three key components: trajectory regis-

tration, cost function formulation, and BP detection. To better describe how the three
components interact with each other, a visual representation of the proposed method
is presented in Fig. 25.
Fig. 25 takes Singapore Airway Network (referred to as SAN hereafter) as an
example to illustrate the proposed method. The Trajectory Registration step mainly
aims to map the aircraft trajectory data points, in ADS-B data, to the SAN and

78
 Figure 25: Concept diagram of the proposed method for BP detection in a given airway network.

extract the air traffic demand (number of aircraft) information from the ADS-B data
for later calculation.

The outcome of the Trajectory Registration step is the relationship between the
flow (number of flights) and the flight duration (total flight time) on each link of the
SAN. In the flow-duration network shown in Fig. 25, the size of circles on both ends of
a link represents the significance of flight duration on that link. The larger the circles
are, the longer the flight duration is. The Cost Function Formulation step develops
the cost function for each link of the SAN.

From the first two steps of the proposed method, the cost functions and total
traffic demand information can be obtained. Then the final step, i.e., BP Detection, is
performed through non-linear programming to identify the Links of Paradox (LOP) in
the airway network that cause BP. In what follows, how each step works is delineated
in detail.

4.3.2. Trajectory Registration

BP detection requires the flow-duration function on each link of the airway network.
With regard to this, the aircraft trajectory data points are registered, from ADS-B
data for a given period of time, on the SAN. The trajectory registration algorithm is
presented in Algorithm 1. The notations are as follows:

Ti : (T1i , T2i , T3i , ..., Tki , ..., Tni i ) is the i-th trajectory with Tki being the k-th waypoint

79
Algorithm 1 Trajectory registration algorithm
Input: ADS-B Data

1. Find the best matched airway segments for trajectory Ti : By traversing all air-
ways, find an airway Aj contains a sequence of segments S = (Ajk , ..., Ajm , ..., Ajh )
which has the minimum distance Dmin to Ti .
2. If Dmin > D0 , Ti will be regarded as noise and discarded.
3. If Dmin ≤ D0 , find the best matched segments T s = (Tli , ..., Tpi , ..., Tqi ) on Ti for
As, so that T s has the minimum distance to As.
4. One unit flow will be added to each segment Asy = (Ajy , Ajy+1 ), y = k, ..., m, ..., h,
in As.
5. Each segment T sx = (Txi , Tx+1
i
), x = l, ..., p, ...q − 1, in T s will be registered to
the airway segment (Am , Am+1 ) in As that has the minimum distance to T sx .
j j

6. The rest part of the trajectory Ti , (T1i , ..., Tl ) and (Tqi , ..., Tni i ), will be regarded
as new trajectories and being registered from step 1.

Output: Flow-flight duration relation on each link of the airway network.

on Ti . Ti consists of ni − 1 consecutive links. When randomly choosing two way points

Tki and Tgi forming the segment ST ikg , there will be n2i different trajectory segments.

Aj : (Aj1 , Aj2 , Aj3 , ..., Ajh , ..., Ajmj ) is the j-th airway with Ajh being the h-th waypoint
on Aj . Aj consists of mj −1 consecutive links. When randomly choosing two way points
Ajh and Ajl forming the segment SAjhl , there will be m2j different airway segments.

The distance between segments STki and SAjhl is defined as the sum of the distances
between the starting point and ending points of each segment: D = (xik − xjhl )2 + (yki −
j 2
yhl ) + (xik+1 − xjh(l+1) )2 + (yki − yh(l+1)
j
)2 . The bigger the difference between the two
segments, the larger the value of the distance.

As illustrated in Fig. 26, from a micro perspective, for each pair of matched
airway segment and trajectory segment, links composing the trajectory segment will
be registered onto links composing the airway segment. Time spent on each trajectory
link will be added to its matched airway link duration. From a macro perspective, the
entire trajectory Ti is assigned onto the airway segment with the minimum D between
them, with the constraint D ≤ D0 . When the minimum D is larger than D0 , this
trajectory is discarded because it may be corrupted by noise. D0 is the upper bound

80
Figure 26: Registration of trajectories onto airways. From a macro view, part 1 of trajectory T1 will
be registered to airway A1 , while part 2 of trajectory T1 will be registered to airway A2 . One unit
flow will be added to A1 and A2 . From a micro view, the time duration on trajectory links ST11 ,
ST21 , ST31 will be added to airway link SA212 , while the time duration on trajectory links ST41 , ST51 ,
ST61 will be added to airway link SA223 .

of distance. When a distance is larger than D0 , it suggests a similarity between the

trajectory and any airway segment barely exists. The trajectory segment is matched
onto the corresponding segment of the assigned airway. Suppose a part of a trajectory
has the minimum distance to an airway. In that case, this part will be registered
onto the corresponding airway. The rest of the trajectory will be regarded as a new
trajectory and undergo the registration process again. This step is repeated until
every trajectory part is matched onto an airway segment or discarded as noise. When
a trajectory segment is registered onto an airway segment, a unit flow will be added
to the links comprising the airway segment.

4.3.3. Cost-Function Formulation

After mapping all the trajectories’ segments onto their best-fitted airway links, the
flow-duration data on each link can be obtained for a given period. The flow on a
link equals the number of trajectories matched onto it, while the flight duration on a
link is the time flights spend on the matched trajectory segments. Fig. 27 gives an
example of the flow-duration relation on a link of an airway network.

81
 30 30

25 25

Cost (min.)
20
Cost (min.)
20

15 15

10 10

5 5

0 0
0 2 4 6 8 10 0 2 4 6 8 10
Flow Flow

(a) (b)

Figure 27: An example of the data-driven cost function formulation for one link of an airway network.
(a) shows the original flow-duration relation obtained from the Trajectory Registration step. (b) shows
the linear cost function for the focal link.

Upon getting the relationship between flow and duration, to determine the overall
trend of the relationship without loss of accuracy, data points under the same flow are
fused by taking the average of them, whereby the red points shown in Fig. 27(a) are
converted into red points in Fig. 27(b). Afterward, the least mean square method is
used to determine the coefficients of the linear cost function as displayed in Fig. 27(b).

Note that in the literature, when investigating BP on road traffic networks, re-
searchers have utilized linear cost functions [208–210], polynomial cost functions [211],
and general monotonous non-decreasing cost functions [212]. Results showed that cost
function regression using a linear, polynomial, and cubic function has similar perfor-
mance. Therefore, in this research the linear function, L = c0 + c1 x, is adopted in
accordance with the principle of parsimony [213].

4.3.4. Origin-Destination (OD) Pair Extraction

Section 4.1 shows that, apart from the cost functions for all the links of an airway
network, knowledge of the OD pair and traffic demand information in the airway
network are also required to detect BP.

From the ADS-B data, entry-exit points for each flight in the SAN can be extracted.

82
After mapping each flight trajectory to the SAN, the waypoints nearest to the entry
and exit points can be determined. The two waypoints are labeled as an OD pair.
From all the trajectories during a given period, the set of OD pairs on the focal airway
network can be extracted.

For each OD pair, the traffic demand, i.e., the number of aircraft traveled across
the OD pair, can be obtained from the ADS-B data. The total traffic demand on the
focal airway network can be determined by adding up the demand on each OD pair.
The next step illustrates the BP detection process for a given airway network with the
cost functions, OD pairs, and traffic demand information at hand.

4.3.5. BP Detection

The BP detection process involves two steps. The first step is to minimize the
objective functions with respect to the UE, SE, and BE models so as to obtain the
corresponding optimal solutions. The second step compares the corresponding optimal
solutions yielded by the first step and identifies the links that cause BP. The details
for each step are presented in the following paragraphs.

4.3.5.1. Minimization of Objective Functions. In order to detect BP from a traffic

network, one needs to determine the optimal traffic flow distribution on the focal net-
work based on a given principle like UE. In the literature, the UE, SE, and BE models
have been formulated as mathematical optimization problems [204, 205]. Therefore,
the optimization problem in this chapter is introduced by presenting the following
pertaining notations and definitions first.

Given an airway network denoted by an undirected network G = (V, E) with V

being the set of waypoints (nodes) and E the set of links connecting the waypoints.
Let the nodes of the network be indexed as i = 1, ..., n and links be index as eij
if the link connects nodes i and j. Let A be the adjacency matrix of G, with its
entry aij = 1 if there is a link eij between nodes i and j, otherwise, aij = 0. Let

83
EE = {EE1 , ..., EEt , ..., EET } be the set of the T OD pairs on G. The t-th OD
pair is EEt = {Ot , Dt } with Ot and Dt respectively being the origin and destination
waypoints. Let D = {d1 , ..., dt , ..., dT } be the demand vector for the T OD pairs on
G, i.e., dt is the demand on the t-th OD pair EEt . Assume that for the t-th OD pair
there are maximum Rt accessible paths respectively denoted by Pt1 , Pt2 , ..., Ptr , ..., PtRt .
The flow on path Ptr is represented by ftr . The set of available paths in the airway
network G is denoted by P = {P1 , P2 , ..., PT }. A link e ∈ p when the link is traversed
by the path p. Define a function δ(tr,ij) , with δ(tr,ij) = 1 if link eij ∈ Ptr , otherwise,
δ(tr,ij) = 0. Let L be the latency matrix of G with its entry lij = c0ij + c1ij xij being the
cost function for link eij . xij is the amount of flow on link eij .

With the above notations and definitions, the optimization problem with respect
to the UE model for an airway network is formulated as follows:

n X
X n Z xij
UE : min FU E (X) = aij lij (t)dt
X={xij }
i,j=1,...,n i=1 j=1 0

s.t. ftr ≥ 0
PRt
r=1 ftr = dt (16)
PT
t=1 dt = D
PT PRt
t=1 r=1 δ(tr, ij)ftr = xij

lij (xij ) = c0ij + c1ij xij

In the UE optimization problem presented above, ftr denotes the flow on the path
Ptr . Constraint ftr ≥ 0 assures that the traffic flow on every path of an airway net-
work is non-negative. Constraint R r=1 ftr = dt indicates that the sum of flows on
P t

all paths between the t-th OD pair should satisfy its demand requirement. Con-
straint Tt=1 dt = D means that the sum of demands on all OD pairs should equal
P

the demand on the studied airway network under a given spatial-temporal context.
Constraint r=1 δ(tr, ij)ftr = xij converts flows on paths into flows on links
PT PRt
t=1

since the objective function is involved with links flows. The last constraint presents

84
the linear cost functions for each link in an airway network. By minimizing FU E , the
optimal flow distributions can be obtained, which are the outcome of users’ selfish
routing.

n X
X n
SE : min FSE (X) = aij xij lij (xij )
X={xij }
i,j=1,...,n i=1 j=1

s.t. ftr ≥ 0
PRt
r=1 ftr = dt (17)
PT
t=1 dt = D
PT PRt
t=1 r=1 δ(tr, ij)ftr = xij

lij (xij ) = c0ij + c1ij xij

Analogous to the UE optimization problem, the SE optimization problem is for-

mulated in Eq. 17, which shares the same constraints as Eq. 16. The SE model is the
outcome of users’ collaborative routing. As a result, the minimization of FSE generates
the optimal solutions for the air traffic flow assignment in the airway network. The BE
optimization problem is similar to the SE optimization, but with one more constraint
added to ensure that no users’ travel cost is increased compared to the cost under UE.

n X
X n
BE : min FBE (X) = aij xij lij (xij )
X={xij }
i,j=1,...,n i=1 j=1

s.t. ftr ≥ 0
PR t
r=1 ftr = dt
PT (18)
t=1 dt = D
PT PRt
t=1 r=1 δ(tr, ij)ftr = xij

lij (xij ) = c0ij + c1ij xij

u∗tr ≥ utr

In the constraint u∗tr ≥ utr , the variables u∗tr and utr denote the time costs on path
Ptr obtained by optimizing FU E and FBE respectively. Constraint u∗tr ≥ utr requires

85
that the optimal solution to FBE makes some users’ travel cost better-off but no users’
cost worse-off compared to the UE model.

By minimizing the three objective functions, the corresponding flow distribution

matrix X can be obtained. The entry xij represents the optimal flow on link eij in the
studied airway network.

4.3.5.2. BP Links Identification. Optimization of the three objective functions provide

three flow distribution matrices: XU E , XSE , and XBE . This section determines the
potential links that may cause BP by comparing the entries of XU E , XSE and XBE .
Specifically, if the following conditions are satisfied:

xSE BE
ij − xij →0
(19)
xUijE ≥ δxBE
ij

link eij will be regarded as a potential link that causes BP.

ij and xij
The first condition requires the value of xSE BE
being close to each other.
XU E is the outcome of users’ selfish routing which is in a way similar to the Nash
solution in the Game Theory [204, 205]. XSE is the global optimal solution. XBE can
be regarded as a local optimal solution compared to XSE . In the second condition, the

ij . According
parameter δ determines the difference between the values of xUijE and xBE
to the BP introduction presented in Section 4.1, if the difference between xUijE and
xBE
ij on link eij is large, it suggests that the flow on link eij could have detrimental
influences on the flight cost. Thus, link eij will be a potential link that causes BP.
Therefore, this study assumes that if the difference between xUijE and xBE
ij exceeds one

order of magnitude, the corresponding link eij will be regarded as a potential link
leading to BP in the network. As a consequence, δ is set as δ = 10 in the case study
section of this chapter.

Through on the above steps, one can determine all the potential links that may
cause BP. Afterward, by removing those links from the network and re-optimizing

86
 ′
FU E on the new network, a new flow distribution matrix XU E can be yielded. If the
following condition is satisfied:

′
FSE (XU E ) < FSE (XU E ) (20)

it implies that after removing those links from the airway network, if the total travel
cost on the network is reduced, the links indeed cause BP to the focal airway network.

4.4. Case Study

The above section describes the proposed method for BP detection for a given
airway network with flight trajectory data. To verify the efficacy of the proposed
method, this section carries out a case study on SAN using six-month ADS-B data
from 1st June 2017 to 30th December 2017.

4.4.1. Cost Function Modeling

First, the network structure of SAN is constructed from the ADS-B data. Fig. 28
displays the airway network structure of SAN, which consists of 198 nodes and 272
links.

The trajectory data at flight level 330 are selected in this study. Through trajec-
tory registration, the flow-duration relations for all links of SAN can be extracted.
Specifically, trajectory registration is implemented at each airway in SAN during each
period of a day. The registration results over the six months are averaged and used
to approximate the cost functions on the airway links. As shown in the top panel of
Fig. 29a, on some days, 30 or more flights could transit through a particular airway.
Note that the cost function of a link is a function of air traffic flow depicting the total
duration of flights on the link. As long as the flow, e.g., 2-3 aircraft, on a link is
known, one can obtain the flight duration that the aircraft will spend on an airway
through the cost functions. In this way, flow assignment on airways can be achieved for
BP detection. The cost function matrix L is approximated using the linear function

87
 NOBEN
TOKON
REGOS
AGSAM VAVUM
DAMVO MIMUXALDAS
BENSA PQ LOSON
TUNPO LATHA
ELSAS MAPNO
ANHOA LAXOR PR
TONIKXONAN DAGAG
DAGAK
MOXON TEGID
BIBAN CONSON
ADBOP VINIK
SUWAN
TIKAL
DUGON SAMOG AKMON NODIN
SUMLA
BIDAG OSANU
BITOD LAGOT
IPRIX
IGARI
TIDAR DUDIS MELAS
ODONO TODAM
HTY GOLUD
VKB UPRON ASISU
PADLO
IKUMI
IKUKO
GUGIT VIDIPKINABALU
VPG DOGOGUKIBA
LABUAN
ANIRU
GUNBO ENREPDOLOX OLKIT BOMPOLEDAM BRUNEI
KERTEHNOPAT AGUPO
SAKMA BUTAX
DAMOG MALBI
MABLI TERIX GULIB MIRI
KADMO
ANSOM VELPI
VEPLI
TAXUL SUSAR
DUBSA DARMU UDEROMAMOK
PULIP VKN UGPEK
BUVAL MUMSO LUSMO
MATSU
ADNUT
PIBOS DOVOL
VISAT
URIGO
IDSEL
PEKAN
EGOLO ELGOR
MABAL ONAPO SARVO
KIMAT
ISTAN MANIM
OBDAB
PALDIOTLON
RAXIM
KILOT LEBIN TAMIBBINTULU
VKL LATUK PULAU
UKASA
SAROX NIVAM LIPRO
KIBOL
LENDA BUNTO ALEMO
SAPAT
GEMAS
ENKOL PEKLA
VMR VERIN
BOBOB KAMIN
BATAR REDUK
GUPTAAROSO SABIP PILAX
VSI
SIBU
MASBO
OGAKO
MITOS MESOG OLSAM
ESPITAGOBA
TOPOR
ARAMA
REKOP OBLOT OKADA
SUKRIPU
VTKHOSBA TOMANOBGETNIMIX KUCHING
PAPSA SJ
SAMKO
BOBAG
ASUNA ANIPU
TAROS REMES TPG
IDMAS
PEKANBARU ATPOM OBDOS ARUPA
REPOV
ANITO KADARBAVUS
PONTIANAK
PARDI KIKOR
SURGA OMEGA
YUANA
WIDIA
LAWIB
RUSMA
APARI
PALEMBANG

GUVIL

Figure 28: Airway structure of SAN.

800 150
input input
predicted 100 predicted
Cost (min.)

600
interval interval
Cost (min.)

50
400
0
200
-50
0
-100

-200 -150
0 10 20 30 1 2 3 4
Flow Flow
40 3

20 2
residual error

residual error

0 1

-20 0

-40 -1

-60 -2
0 10 20 30 1 2 3 4
Flow Flow
(a) (b)
Figure 29: Cost functions modeled from the traffic data for two representative links in SAN. (a) cost
function for link 88. (b) cost function for link 201.

88
y = c0 + c1 x. The entry lij of L includes two elements, c0ij and c1ij , which are used to
calculate the time cost on link eij of SAN.
Fig. 29 displays the cost functions on two representative links, i.e., link 88 and link
201, of SAN. It can be seen from Fig. 29(a) that link 88 had been frequently traveled
by flights, while link 201 was relatively less traveled. In the calculations of the matrix
L, three statistical metrics, MSE (Mean Square Error), RMSE (Root Mean Square
Error), and R2 (R-square), are adopted to evaluate the cost functions. The statistical
results for the 204 links of SAN are shown in Fig. 30.

500
MSE

0
20 40 60 80 100 120 140 160 180 200
Link ID

20
RMSE

10

0
20 40 60 80 100 120 140 160 180 200
Link ID
1
R2

0.5

0
20 40 60 80 100 120 140 160 180 200
Link ID

Figure 30: Assessment of the data-driven based cost function modeling using statistical metrics. MSE
is the mean square error, and RMSE is the root mean square error. R2 is the R-square statistical
metric.

Normally, when the value of R2 is larger, the approximated linear cost function will
have higher confidence. It can be seen from Fig. 30 that the values of MSE and RMSE
are large only for a few links, while for the majority of them, the values are quite small.
This observation supports the feasibility of adopting linear representations of the cost
functions. The values of R2 for the cost function regression on all the links are large
except for the function on link 201. The cost function on link 201 is shown in Fig.

89
29(b). It can be seen that although the value of R2 is small, the cost function is still of
high confidence. The reason is that the R2 metric makes some exceptions, i.e., when
data samples approximately follow a distribution parallel to the X-axis, the value of
R2 will be close to zero.

4.4.2. BP Detection

Before BP detection, the demand and OD pair information need to be determined.

Fig. 31 displays the air traffic demand over 180 days in SAN at the flight level 330.

160
Demand Distribution

140

120
Demand (#Aircrafts)

100

80

60

40

20
20 40 60 80 100 120 140 160 180
Day

Figure 31: Air traffic demand over 180 days in SAN at flight level 330.

With the cost functions as well as the daily demand on the OD pairs in SAN,
FU E (X), FSE (X) and FBE (X) can be minimized to identify the three optimal solutions,
i.e., XU E , XSE , and XBE . Each of the optimal solutions represents the optimal flow
distribution with respect to the corresponding equilibrium model with the given travel
demand. By substituting X into FSE (X), one can obtain the total travel cost of flights
in the airway network.

Fig. 32 shows the difference between the total costs under UE and SE, as well as

90
 70
UE-SE
UE-BE
60

50

Residual Cost (min.)

40

30

20

10

-10
20 40 60 80 100 120 140 160 180
Day

Figure 32: Comparisons of the total travel time costs occurred in SAN under UE, SE, and BE over
the studied period of 180 days. The red curve represents the residual costs of the total costs with
respect to UE and SE, while the blue curve denotes the residual costs of the total costs with respect
to UE and BE.

the difference under UE and BE. Note that BP could occur if the difference between
the total costs is significant, as a small difference between the total costs indicates that
the Nash solution XU E is close to the global optimal solution XSE or to the Pareto
solution XBE . It can be observed from Fig. 32 that, for some days the differences are
significant, while for some days the differences are small (less than 10 minutes).

Fig. 32 only presents a holistic view of the total costs with respect to the three
optimization models. Identification of the links that cause BP requires comparisons
between the optimal solutions XU E , XSE , and XBE . As each of the optimal solutions
is a matrix, to facilitate the comparisons, they are converted into three vectors re-
spectively: x1 = {x11 , x12 , , ..., x1204 }, x2 = {x21 , x22 , , ..., x2204 }, and x3 = {x31 , x32 , , ..., x3204 },
with xi being the flow on the i-th link of SAN.

Without loss of generality, day 175 is taken as an example to show the BP detection
result. Fig. 33 shows the three optimal flows x1 , x2 , and x3 with respect to UE, SE

91
and BE on day 175. Note that for x1 , x2 , and x3 , x1i = x2i = x3i exists for some i-s. To
keep the figure clear, flows on such links are not shown Fig. 33.

Statistics for link flow (Day 175)

35
UE Flow
SE Flow
30 BE Flow

25 10
10
Flow (#Aircrafts)

5
5
20
0 0
113 114 115 162 163 164
15

10

0
0 20 40 60 80 100 120 140 160 180
Link ID

Figure 33: The optimal flows on each link in SAN with respect to UE, SE, and BE on day 175.

In Fig. 33, the length of each colored bar signifies the flow on the corresponding
link. It can be seen from Fig. 33 that, on most of the links, the flows under UE,
SE, and BE are very close to each other. Only the flow on links 114 and 163 show
x1114 = x1163 = 0.1656 and x2114 = x3114 = x2163 = x3163 = 0, which satisfies the condition
presented in Eq. 19. Therefore, links 114 and 163 are regarded as the potential links
that can cause BP.
The waypoints associated with links 114 and 163 are shown in Fig. 34. To verify
whether links 114 and 163 indeed cause BP, the two links are removed from the SAN.
Then FU E (X) is re-optimized and the total cost on the network is recalculated. The
corresponding results are shown in Fig. 34. It can be observed from Fig. 34 that
before removing links 114 and 163, the total time duration on the network was 8661.15
minutes. Removing those two links saves the travel time on the network by 3.8%. Thus,
links 114 and 163 are the links that cause BP on day 175. This improvement in the

92
 Link ID: 163
LIPRO -- KILOT
Link Flow: 0.1656

Link ID: 114

UGPEK -- ENREP
Link Flow: 0.1656

Figure 34: BP detection results on day 175 for SAN. Links 114 and 163 are detected as the links that
cause BP. The original flight duration on the network was 8661.15 minutes. After removing the two
links from the network, the flight duration is reduced to 8328.64 minutes.

93
Table 3: Statistical results for BP detection in SAN at flight level 330. T C1 and T C2 represent
the total costs obtained by minimizing FU E (X) before and after removing the BP links from SAN.
SC = T C1 − T C2 is the saved cost.

Day Demand BP Link Link Flow T C1 (Min.) T C2 (Min.) SC (Min.) SC/T C1 (%)
1 49 67 2.887e-02 2021.312 2014.060 7.251 0.359
8 63 134 1.804e-01 3503.380 3501.663 1.717 0.049
12 60 121 8.956e-02 2765.423 2764.178 1.245 0.045
27 2.160e-02
17 58 3536.038 3516.810 19.228 0.544
71 1.445e-01
28 82 144 1.003e-01 5170.283 5161.725 8.558 0.166
30 30 158 1.891e-01 1131.530 1123.548 7.982 0.705
38 84 78 1.270e-01 7108.310 7107.388 0.922 0.013
44 92 70 8.820e-02 9642.960 9590.713 52.247 0.542
45 94 156 4.856e-02 8848.947 8846.660 2.286 0.026
48 92 45 8.307e-02 7120.268 7101.940 18.328 0.257
54 88 28 2.964e-02 6265.240 6239.498 25.743 0.411
64 115 45 1.009e-01 13823.800 13814.417 9.384 0.068
66 102 78 2.061e-02 9999.972 9999.807 0.166 0.002
67 99 58 8.068e-02 11940.628 11936.217 4.412 0.037
71 87 134 5.167e-02 7670.403 7668.630 1.773 0.023
72 107 74 5.331e-02 11172.695 11169.220 3.475 0.031
78 111 21 8.447e-02 14093.982 14091.357 2.625 0.019
91 119 114 2.279e-02 10629.088 10397.777 231.312 2.18
93 106 186 4.128e-02 10711.080 10709.265 1.815 0.017
165 2.475e-02
96 73 6838.770 6816.277 22.494 0.329
177 1.335e-01
107 85 71 5.762e-02 8335.593 8290.270 45.324 0.544
112 116 122 1.747e-01 10854.922 10853.098 1.823 0.017
115 130 101 1.700e-01 10421.997 10400.513 21.484 0.206
58 1.516e-01
126 112 9354.567 9320.438 34.128 0.365
169 1.516e-01
129 114 74 8.605e-02 10751.427 10725.627 25.800 0.240
143 110 177 7.539e-02 8440.668 8434.132 6.538 0.077
146 122 177 1.074e-01 8585.915 8566.662 19.254 0.224
148 112 122 8.963e-02 5675.065 5665.760 9.305 0.164
150 109 186 3.376e-02 9150.042 9149.582 0.459 0.005
165 122 45 3.111e-02 11706.637 11701.168 5.469 0.047
168 114 28 8.776e-02 9694.607 9661.287 33.320 0.344
114 1.655e-01
175 118 8661.152 8328.643 332.509 3.839
163 1.655e-01

94
travel time is only for one day and one flight level of the given airway network. For all
flight levels and all OD pairs in the airway network, the improvement can be highly
significant.

To verify the validity of the proposed method and the primary result, the proposed
method is applied to the entire dataset of 180 days to identify BP links. The statistical
results are reported in Table 3. The days in bold are the top five days that have reduced
flow duration after removing the identified BP links. It can be clearly seen from Table
3 that BP does occur in SAN, and after removing the BP links, the flow duration on
the airway network is indeed reduced.

4.5. Chapter Summary

This chapter investigated the occurrence of the BP phenomenon in airway networks

to optimize the structure of airway networks. To achieve this goal, this chapter first
developed a generic method to assist in detecting links that can cause BP in any given
airway network. Then a case study was carried out over South-East Asia airspace
covering Singapore Airway Network (SAN) using six months of ADS-B data from 1st
June 2017 to 30th December 2017. The proposed method involves two key components,
i.e., cost function calculations and BP modeling. The cost functions for the SAN
links are derived from real-world data instead of assumptions. Experimental results
demonstrate that the BP phenomenon does occur in airway networks. In the case
study, it is found that after removing BP links, for one-day air traffic, the total travel
time was reduced from 8661.15 minutes to 8328.64 minutes, saving 332.5 minutes.
This amount of travel time saving in SAN is 3.8%, and it is just for one flight level.

Although only the case study on the SAN was carried out, it is positive that the
findings of BP in airway networks could be a prevalent phenomenon. SAN is one of
the most congested and dense airspaces in the South-East Asia region (e.g., Singapore-
Kuala Lumpur is the world’s busiest route) as it sits on the juncture of six neighboring
FIRs and covers the South China Sea airspace, where most of the air traffic from China

95
converges towards South-East Asia. Further, despite its small size, SAN has features of
a variety of airspace, e.g., radar, procedural, RVSM, and oceanic. In future work, the
proposed method can be extended to cover larger airspace, specifically which applies
advanced ATM procedures such as Direct-To, Lateral Offset, and Required Navigation
Performance (RNP), which may alter the air traffic flow as well as the structure of a
given airway network.

Note that present airway networks have evolved, over a period of time, without
any scientific basis. As and when major traffic flows were identified, new airways were
established and added to the airway network. It was an incremental approach, which,
though considered traffic demand and capacity, the airspace structure complexity and
connectivity, and the balance of ATCO workloads, led to a highly rigid and complex
airspace design rendering it impractical to change [161]. These pre-defined airway-
waypoint networks constitute one of the major elements used today to channel air
traffic, mainly for the en-route part of the flight. As air traffic has been concentrated on
airway networks, redesigning such networks that have not been systematically designed
may improve traffic flow and help manage air traffic congestion. Complete redesign
of airway networks is a challenging task that has to consider many competing factors,
such as the needs of both civil and military airspace users, the traffic structure, the
ATM capacity, a proper sectorisation scheme, and the connectivity to terminal areas
[89]. Given a variety of stakeholders (Military, ANSPs, Airports, Airlines, etc.), a
complete redesign of airway networks is impractical. The findings of this work suggest
that the “removals” of some airways could reduce the total travel time for the flights
flying through the airway network at a given flight level without cutting down the air
traffic demand. Therefore, this work’s idea sheds new light on the optimal structure
design of airway networks for better air traffic services.

The BP phenomenon, as illustrated in Fig. 23, indicates that the removal of some
extra links in an airway network may improve the traffic efficiency. Note that extra

96
links in an airway network may be there for a variety of reasons, such as Special
Use Airspace (SUA) in that region. They might be active at some point or due to a
Letter of Agreement (LOA) with the neighboring airspace for special events (such as
airshows). Removing them may lead to a reduction in user flexibility at such times.
Nevertheless, the proposed method can help identify such links, and the identification
may help redesign such links more efficiently.
With regard to the BP detection method, one may argue based on Fig. 23 that an
extra link in a traffic network could dispatch traffic flow, and the merge of traffic flow
could cause congestion in the network. Therefore, one would suggest the detection
of such links by locating all the possible links as shown in Fig. 23 in a given airway
network. However, one thing that should be pointed out is that the traffic flow in the
airspace may not necessarily form the network shown in Fig. 23 as the traffic flow is
dynamically changing over time. Nevertheless, this factor can be considered to gain
more insights into the methodology development in further work.
Although the proposed method is able to identify possible links that cause BP, it
cannot identify the conditions under which BP happens in an airway network. In the
future, efforts can be put to mathematically exploit the condition of the occurrence
of BP in airway networks as well as identify the critical demand for BP occurrence.
Although airspace optimization is arduous and requires interdisciplinary efforts, this
study sheds light on the possible way to improve air traffic flow.

97
Chapter 5. Airspace Capacity Overload Identification

The research in this chapter has been published in the conference paper: Chunyao Ma, Qing Cai,

Sameer Alam, and Vu N. Duong, “Airspace capacity overload identification using collision risk pat-

terns,” in 2020 International Conference on Artificial Intelligence and Data Analytics for Air Trans-

portation (AIDA-AT), Singapore. IEEE, 3-4 Feb, 2020, pp. 1–9.

The previous two chapters optimize the airspace from a spatial view, i.e., airway
network optimization. This chapter will investigate the optimization of airspace from a
temporal perspective: identification of the airspace capacity overload, which may assist
ATCOs in addressing anticipated saturation in airspace capacity. The most popular
capacity overload identification method is the controller workload-based analytical
model, which is difficult to capture the changes and evolution of the air traffic patterns
and airspace structure. Other capacity overload indicators in the literature, such as
airspace merge/split operation and “Domino Effect Parameter”, are not able to identify
different phase transitions in the airspace capacity. Therefore, this chapter focuses on
new metrics and methods to identify the phase transitions in airspace capacity, which
leverages the airspace structural features and air traffic flow patterns.

Current air traffic control operations are highly structured to help ATCOs manage
traffic flow within the constraints imposed by safety requirements [214]. Air Traffic
Control procedures are designed to facilitate air traffic’s safe and efficient flow through
the airspace. Under IFR, responsibilities for safety and efficiency are shared between
the pilots and ATCOs. In general, the pilots follow an IFR flight plan (filed before
the flight) that defines the horizontal and vertical components of the flight route.
The pilot is responsible for flying along the route within certain limits of accuracy.
These limits assure that each aircraft remains in its assigned airspace and is proce-
durally separated from other aircraft flying different routes. ATCOs are responsible
for assuring that aircraft under their supervision remain separated to avoid the risk

98
of collision. Therefore, managing air traffic is a safety-critical operation that requires
close interactions between human ATCOs, spacing and separation procedures, and
technical systems. The trade-off between safety and capacity, inherit to ATM, implies
that without a simultaneous validation of safety, validation of capacity is meaningless
[215]. Therefore, this chapter proposes to use the airspace safety indicator: aircraft
mid-air collision risk, to measure the airspace capacity overload, i.e., identify different
phase transitions in airspace capacity.
When two aircraft cross each other, there will be a risk of collision at this crossing
point. The distribution of the crossing points along with the risk value associated
characterize the collision risk inside an airspace [80]. To identify airspace capacity
overload using collision risk, this chapter first computes the collision risk distributions
inside an airspace through collision risk modeling [80]. Then a set of heatmaps are
generated to represent the collision risk distributions, while the heats on the heatmap
are characterized by the density of crossing points in this area as well as the collision
risk values. Eventually, the heatmaps are clustered into a set of collision risk patterns
using image processing techniques from which different phase transitions in airspace
capacity can be further identified. The proposed method has been applied to one-
month flight data of a sector in Singapore airspace, and three states have been clearly
observed: normal state, transition state, and overload state. Upon identifying the
three collision risk phases, for a given traffic situation, one can identify the airspace
capacity overload status by recognizing the best-fit pattern of the collision risk distri-
bution. If the recognized pattern is the overload state, it suggests that the airspace
has reached the capacity limit, and the collision risk will increase dramatically if more
flights are accepted.

99
5.1. Methodology

5.1.1. Methodology Overview

Fig. 35 shows the concept diagram of the proposed method. It can be seen from
Fig. 35 that the proposed approach encompasses three key steps: collision risk mod-
eling, collision risk heatmap generation, and collision risk pattern recognition.

Figure 35: Conceptual diagram of the proposed method for airspace capacity overload identification.

Given the air traffic data and the airspace configuration, a collision risk model is
introduced to identify the distribution of crossing points as well as their associated
collision risk values. The outcome of the collision risk modeling step, i.e., collision risk
distribution, is then converted into heatmaps in the heatmap generation step based
on the collision risk densities and intensities. In the third step, the pattern recogni-
tion technique, specifically the image clustering technique, is applied to the generated
heatmaps to uncover the latent collision patterns. The recognized patterns are further
categorized into three classes, viz, low, medium, and high collision risk, which corre-
spond to the normal, transition, and overload state of the capacity. Eventually, for
a new traffic scenario, one can identify the corresponding airspace capacity overload
state by recognizing the best fit pattern of the generated heatmap. In what follows,
how each step works is delineated in detail.

100
5.1.2. Collision Risk Modeling

Collision Risk is defined by ICAO as “the expected number of mid-air aircraft acci-
dents in a prescribed volume of airspace for a specific number of flight hours due to loss
of planned separation” [216]. Vertical collision risk accounts for two primary factors:
the likelihood of the loss of vertical separation and the exposure to the loss of verti-
cal separation. This exposure occurs when two aircraft at adjacent flight levels cross
each other. Depending on traffic geometry, the crossing of aircraft may be broadly
subdivided into same direction traffic (zero intersection angle), opposite direction traf-
fic (180º intersection angle), and crossing traffic (remaining intersection angles) [217].
Slightly different models exist for the different traffic geometries. Flight data are the
essential data source in estimating the exposure to the risk due to the loss of vertical
separation. The likelihood of the loss of vertical separation due to typical aircraft
height deviations depends on the probability distributions of Altimetry System Error
(ASE) and Flight Technical Error (FTE).

1) Data Preparation

Before computing collision risk, it is necessary to prepare the required air traffic
data, including flight data and airspace configuration data. The flight data comprises
aircraft type, departure/destination airport, entry/exit points, entry/exit level, en-
try/exit time, and flying duration in the airspace. The required airspace configuration
data is the intermediate waypoints of the airways in the target airspace [218]. With
the information of intermediate waypoints, the route of a given flight can be best
represented instead of a straight segment connecting the entry/exit points.

2) Counting Crossings

As mentioned previously, two aircraft are exposed to the risk of vertical collision
when flying at adjacent flight levels and crossing each other. In order to determine the
crossing between flights, the nominal projection method proposed in [219] is adopted
to estimate the positions of the aircraft based on their current status. By doing so,

101
for each pair of flights, one can identify the Closest Point of Approach (CPA) between
them based on their predictable routes. If the CPA is within 1000 feet vertically and
5 nm horizontally, the the crossing is counted.

With the crossing points being prepared, they can be fed forward to the collision
risk calculation. Further, for those potential crossings, the CPA positions (latitude,
longitude, altitude) together with the time taken by the two aircraft flying to the CPA
are recorded for later use in generating collision risk heatmaps.

3) Collision Risk Computation

The collision risk calculation model in [80, 220] is adopted to compute the collision
risk of the crossing points. Collision risk can be acquired using the air traffic data
prepared in Section 5.1.2 while taking into consideration of multiple factors: the nav-
igation accuracy, lateral/vertical deviations, airway network structures, tracks of two
crossing aircraft, aircraft number and size, and the flight duration within the airspace.
The collision risk is required to be no more than the Target Level of Safety (TLS) of
the fatal accident per flight hour to ensure flight safety in the airspace.

The vertical collision risk of two aircraft flying at adjacent flight levels along the
same route can be described as:

 
|ẏ| λxy |ż|
Naz = 2Pz (Sz )Py (0)nz (equiv) 1 + + (21)
2V λz 2V

where Naz represents the estimated number of fatal aircraft accidents per flight hour
 
|ẏ| |ż|
due to the loss of vertical separation. The factor 1 + 2V + λλxy
z 2V
is a function of
relative aircraft speed and aircraft dimension. V is the average ground speed of the
aircraft. |ẏ| is the average relative crossing speed between two aircraft at adjacent flight
levels. |ż| is the average relative vertical crossing speed between two aircraft that have
lost the vertical separation of Sz feet. A typical aircraft can be characterized by two
parameters when computing the technical vertical risk, namely, the average diameter
(λxy ) and the average height (λz ) of a standing cylinder representing the aircraft.

102
 Pz (Sz ) is the vertical overlap probability for aircraft at adjacent flight levels with
the vertical separation being 1000ft (Sz ). This suggests that two aircraft still have
the potential of vertical overlap while properly flying at adjacent flight levels. Py (0)
represents the lateral overlap probability of aircraft following the same route. It can be
calculated through the time ratio α of an aircraft flying in the airspace with satellite
navigation (GNSS) versus radio navigation (VOR/DME):

!
2
α2
Py (0) = 2λy (1−α)
√
2σV OR/DM E π
+q 2
2α(1−α)
2
√ + √
2σGN SS π
(22)
σV OR/DM E
+σGN SS 2π

where σV OR/DM E and σGN SS stand for the standard deviation for VOR/DME navi-
gation and GNSS navigation respectively. α denotes the proportion of flights flying
with satellite navigation. λy denotes the average wingspan of an aircraft. Vrel (θ) is the
average relative horizontal speed between aircraft at adjacent flight levels following
two intersecting routes with an intersection angle θ. Vrel (θ) can be formulated as:

(23)
p
Vrel (θ) = V 2(1 − cosθ)

 
|ẏ| λxy |ż|
Pz (Sz ), Py (0) and 1 + 2V
+ λz 2V
are calculated for the entire airspace, so they
remain constant for all crossing points. The subscript z is used to label aircraft flying
at adjacent flight levels.

The only remaining term is nz (equiv), which is calculated taking into consideration
of the contribution of all crossing points. One crossing point will contribute 1 count to
nz (opp), nz (same) or nz (θ), which denote the frequencies of opposite direction crossing,
same direction crossing, and angled crossing respectively. nz (equiv) is formulated as:

103
  
|ẏ| λ |ż|
+ xy
|∆V | λz |∆V |
nz (equiv) = nz (opp) + nz (same)  |ẏ| λ |ż|
+
1+ + xy
2V λz 2V
n
π
 
λ |ż|
(24)
X
1 2 xy

|ẏ| λ |ż|
 nz (θi ) 1 + Vrel (θi )2λz
Py (0) 1+ + xy
2V λz 2V i=1
n
X
= nz (opp) + Anz (same) + Bi nz (θi )
i=1

where

   
|ẏ| λxy |ż| |ẏ| λxy |ż|
A= |∆V |
+ λz |∆V |
/ 1+ 2V
+ λz 2V
1 
(25)
Bi = 
|ẏ| λ |ż| π λ |ż|
xy

Py (0) 1+ + xy 1+ V 2 (θ )2λ
2V λz 2V rel i z

|∆V | represents the relative longitudinal speed between two aircraft. nz (θi ) denotes
frequencies of crossings associated with angle θi , i = 1, ..., n. By rearranging the terms,
Eq. 24 can be reorganized as a sum of the collision risk contribution of all types of
individual crossing points. The reorganized nz (equiv) can be written as:

nz (opp) nz (same) nz (θ)

X X X
nz (equiv) = 1+A 1+ Bk (26)
i=1 j=1 k=1
 
|ẏ| λxy |ż|
Representing the kinematics factor 1 + 2V
+ λz 2V
as C, it can be seen from Eqs.
26 and 21 that the collision risk at a same direction, opposite direction, and angled
direction crossing point is 2Pz (Sz )Py (0)AC, 2Pz (Sz )Py (0)C, and 2Pz (Sz )Py (0)BC,
respectively.

5.1.3. Collision Risk Heatmap Generation

With the required flight data prepared for a particular airspace during a certain
time, together with the equations in Section 5.1.2, it will be able to discover when
and where the crossing of two aircraft occurs, what type the crossing is, what the
collision risk at the crossing is, and the distribution of the crossing points. In other
words, the distribution of crossing points, as well as their associated collision risk,

104
can be determined for air traffic during any time slot of interest. This crossing point
distribution and the collision risk values can be delineated by a heatmap. The heat
value of a point on the heatmap reflects the density of the crossing points in the nearby
area and their associated collision risk values.

Assume that during a time slot, there are a number of n crossing points. For each
crossing point ci , i = 1, 2, 3, ..., n, its position is denoted by (xi , yi ) with xi and yi
representing the longitude and latitude respectively. The collision risk value of ci is
denoted by wi . In a heatmap, the heat influence of a crossing point ci on its proximity
shall be stronger as the distance reduces and the collision risk value at the crossing
point increases, which can be described using a two-dimensional Gaussian function
centered at (xi , yi ) and weighted by its collision risk value wi :

(x−xi )2 +(y−yi )2
f (x, y) = wi e− σ2 (27)

in which σ defines the range of the influence of the crossing point at (xi , yi ). Therefore
the overall heat value at a point (x, y) in the heatmap can be described as the sum of
influences from all crossing points in the area:

n
X (x−xi )2 +(y−yi )2
H(x, y) = wi e − σ2 (28)
i=1

By computing the heat values of every point in the area and reflecting the heat
values to different colors, a heatmap can be generated to represent the distribution
and risk values of the crossing points in a 2D plane (latitude-longitude). Fig. 36 shows
an example of a heatmap generated for a sector in the Singapore airspace based on
collision risks calculated during 20 minutes. The area in “red” is subject to a high
density of crossing points and high collision risk values, while the area in “blue” is
subject to low density and low values. There are no crossing points in the “white”
area; thus, the collision risk is zero.

105
 Figure 36: The heatmap generated in an airspace sector with air traffic data during 20 minutes.

5.1.4. Collision Risk Pattern Recognition

The above section presents how to generate collision risk heatmaps for traffic data
during given time periods. These heatmaps reflect the collision risk distribution inside
an airspace, which has the potential to indicate the airspace capacity overload. This
section aims to recognize collision risk patterns from the collision risk heatmaps. To
do so, image processing techniques that involve feature selection and clustering are
adopted.

1) Image Feature Selection

Each heatmap is handled as an image with a fixed size of m×n. For a heatmap, the
most important features are the variation of colors at different positions. Therefore,
the RGB features of each heatmap are extracted, which can be further reshaped into
a row vector with m × n × 3 elements.

2) Image Clustering

Clustering is to organize a set of observations, such as data items or feature vectors,

into classified patterns [221]. Clustering has been widely used in many research fields.
However, it still remains a difficult problem since it requires proper assumptions,
research context, and expert knowledge. In air traffic management domain, clustering
is also a very popular technique which has been widely applied on pattern analysis of
air traffic [222, 223], such as identifying spatial traffic patterns from flight track data
for monitoring, anomaly detection, and airspace redesign purposes [224, 225]. Gariel
et al. developed a framework for airspace monitoring using density-based clustering

106
to learn typical patterns of operation and assess the conformance of flight trajectories
[224]. Similarly, Rehm and Enriquez respectively relied on hierarchical and spectral
clustering to identify nominal and abnormal spatial traffic patterns to/from a specific
airport [225, 226] . In [227], clustering methods were applied to analyze route patterns
to reduce the number of paths between two airports.

In this research, the well-known k-means++ clustering method developed in [228]

is applied to discover the collision risk patterns from the heatmaps. The k-means++
algorithm has improved the traditional k-means algorithm [229] by spreading out the
k initial cluster centers: the first cluster center is chosen uniformly at random from
the data points being clustered, after which each subsequent cluster center is chosen
from the remaining data points with a probability proportional to its squared distance
to the closest existing cluster center.

Although k-means++ is simple, quick and satisfactory when clustering 2D data,

its main drawback is the requirement for the prescription of k, i,e., the number of
clusters. To determine the best value of k, researchers [230] suggested a ‘knee’ point
method which runs the clustering method with respect to different k values ranging
from kmin to kmax , and the Sum-Square-Error (SSE), the sum of squared distances of
all samples to the corresponding cluster centers, is utilized as the validation index to
evaluate the clustering results. When one visualizes the values of SSE with respect to
k, the optimal number of clusters kbest is determined as the knee point on the curve.

Assume that there are nh heatmaps. Let V = {v1 , v2 , v3 , ..., vnh } be the set of
representative vectors of heatmaps and C = {c1 , c2 , c3 , ..., ckc } be the set of kc non-
overlapping clusters. The clustering algorithm searches for the optimal number of
∗
clusters kc∗ and the optimal partition P kc = {p1 , p2 , p3 , ..., pkc∗ } of the vectors. The
clustering algorithm is developed as follows:

Let kmin and kmax represent the minimum and maximum number of clusters, and
r represent the number of repetition of the clustering process. To determine the

107
Algorithm 2 Heatmap clustering algorithm
Input: nh Heatmaps

1. Convert nh heatmaps into nh representative vectors V = {v1 , v2 , v3 , ..., vnh };

2. for kc = kmin to kmax :
for i = 1 to r:
Run k-means++ with kc clusters for the nh vectors;
Compute qki c , i.e., SSE;
end
qkc = min{qk1c , ....qkrc };
Record the corresponding partition P kc ;
end
Plot qkc against kc and identify the “Knee” point;
∗
Output: Optimal kc∗ and the corresponding partition P kc .

optimal number of clusters kc∗ , for each kc ∈ [kmin , kmax ], k-means++ algorithm is run
repeatedly for r times and the ‘best’ validity index qkc , i.e., minimum SSE, as well as
the corresponding partition P kc = {p1 , p2 , p3 , ..., pkc } is recorded. qk is then plotted
against k, and the k at the ‘knee’ point of the plot is determined as the optimal number
∗
of clusters kc∗ . Correspondingly, the optimal partition will be P kc = {p1 , p2 , p3 , ..., pkc∗ }.
The clustering algorithm is summarized in Algorithm 2.

3) Collision Risk Pattern Identification

Applying the clustering algorithm to the generated heatmaps will form kc∗ clusters
of heatmaps. Each cluster suggests one collision risk pattern, and each pattern is
related to a particular phase of the airspace capacity: normal state, transition state,
or overload state. Therefore, for a given traffic scenario, by identifying its best-fit
collision risk pattern, one will be able to identify whether the airspace is operating
below capacity, going through a transition phase, or with capacity overloaded.

5.2. Experiment

Section 5.1 describes the proposed method for capacity overload identification for a
given airspace with air traffic data. To verify the efficacy of the proposed method, this
section carries out an experimental study on an en-route sector in Singapore airspace

108
using one-month air traffic data. Note that most of the time, the sector is operating
below capacity, and it is difficult to collect enough high workload data. Therefore, to
theoretically identify the capacity overload, the one-month flight data are intensified
three times as the original data.

5.2.1. Collision Risk Computation

First, the airspace data for the target sector is determined, i.e., the entire set
of airways and corresponding intermediate waypoints. Then the one-month en-route
flight data between flight level 290 and flight level 410 in this sector are selected.
For each flight, the entry/exit time, entry/exit waypoint, entry/exit level, depar-
ture/destination airport, and aircraft type are identified. The airspace data, together
with the flight data, are fed into a collision risk simulator with the built-in collision
risk computation model. Upon the completion of the simulation, the crossing points
associated with their time of occurrence, the positions of occurrence, and collision risk
values can be obtained for further processing.

5.2.2. Heatmap Generation

With the collision risk distribution obtained in the last step, the heatmaps are
generated at a granularity of per minute. Since the average flight time in the target
sector is around 20 minutes, a flight flying across the sector can possibly be affected
by the collision risk that occurred in the last 20 minutes. Therefore, the collision risk
heatmap generated at each minute considers the collision risks within a 20-minute time
window. In other words, each heatmap represents the collision risk that occurred in
the last 20 minutes. For one-month flight data, 16342 heatmaps are generated. Note
that the proposed method is a generic method that can be applied to any airspace
and any air traffic data. To show the experimental results more clearly and for the
convenience of data processing, 2000 heatmaps are randomly sampled from the original
set for collision risk pattern recognition.

109
5.2.3. Collision Risk Pattern Recognition

To prepare the 2000 heatmaps for clustering, image processing is first applied.
Each heatmap is meshed into 365 × 278 pixels, and each pixel is described by its RGB
features. Thus each heatmap is represented by a 365 × 278 × 3 3D matrix which is
further reshaped into a vector containing 365 × 278 × 3 elements. The vectors can
then be fed to the clustering algorithm in Section 5.1.4 to organize the heatmaps into
different clusters, i.e., different collision risk patterns.

10 6
2.6
min SSE
avg SSE
2.4

2.2

2 k=12
SSE

1.8

1.6

1.4

1.2
0 5 10 15 20 25
k
Figure 37: Determination of the number of clusters: average SSE and minimum SSE are plotted
versus the number of clusters.

5.2.4. Cluster Parameter Setting for Experiment

To determine the appropriate number of clusters, k-means++ clustering algorithm

is run 10 times for different k that ranges from 3 (kmin ) to 22 (kmax ). The average
SSE and minimum SSE of in the 10 runs for each k are recorded and the results are
shown in Fig. 37. It can be seen from Fig. 37 that the decrease in average SSE and
minimum SSE have distinctly slowed down at k = 12, which is regarded as the “knee”

110
(a) p4 (b) p1 (c) p6 (d) p12

(e) p5 (f) p7 (g) p10 (h) p8

(i) p3 (j) p11 (k) p2 (l) p9

Figure 38: The twelve recognized collision risk patterns (p1-p12).

-13 best k=12

10
5 p1
p2
4.5 p3
p4
4 p5
p6
3.5 p7
p8
Collision Risk

3 p9
p10
p11
2.5 p12
AvgCR
2

1.5

0.5

0
0 10 20 30 40 50 60
Aircraft Count
Figure 39: Collision risk versus aircraft count under different patterns.

111
point. Therefore the number of clusters is set as 12.

5.2.5. Results

Organizing the heatmaps into twelve clutsers, the resulted twelve patterns are
shown in Fig. 38. Each cluster is regarded as a collision risk pattern. It can be
observed from Fig. 38 that, in patterns p4, p1, p6, and p12, the “heat” is low and
distributed in a sparse manner. While in patterns p5, p7, p10, and p8, the “heat”
starts to increase and turns into a higher density. In patterns p3, p11, p2, and p9, the
“heat” becomes very high and dense.

Fig. 39 shows the collision risk versus aircraft count under different patterns.
Each data point in Fig. 39 represents one collision risk heatmap generated for a 20-
minute time slice. Its vertical coordinate, i.e., the collision risk value, is the total
collision risk of all crossing points that have appeared during the 20 minutes, while
the horizontal coordinate, i.e., the aircraft count, is the number of flights that entered
the sector during the 20 minutes. The red curve delineates the average collision risk
of the data points in accordance with the aircraft count. It can be seen that as more
aircraft enter the sector, the collision risk increases quadratically, although showing
some fluctuations which are likely caused by different traffic complexities and weather
conditions. This also manifests the efficacy of collision risk as an indicator of airspace
capacity overload. It can also be observed that the collision risk values under the
“low-heat” patterns, i.e., pattern p4, p1, p6, and p12, as shown in Fig. 40a, is low,
while 94.71% of them are below 0.8 × 10−13 and slightly increase (almost insensitive)
with aircraft count. As for the “medium-heat” patterns, p5, p7, p10, and p8, as shown
in Fig. 40b, 74.01% of the collision risk values fall between 0.8 × 10−13 to 2 × 10−13
and increase with a higher speed than under “low-heat” patterns. This percentage is
relatively low because the “medium-heat” patterns are the transition phases from low
collision risk to high collision risk. Therefore the range of this pattern covers part of
the low collision risk area and part of the high collision risk area. Under the “high-

112
heat” patterns p3, p11, p2, and p9, as shown in Fig. 40c, 90.52% of the collision risk
values are above 2 × 10−13 and increase quadratically with the aircraft count.

10-13 10-13 10-13

4 4 5
p4 p5 p3
p1 3.5 p7 p11
3.5
p6 p10 p2
p12 p8 4 p9
3 3 AvgCR AvgCR
AvgCR

Collision Risk

Collision Risk
Collision Risk

2.5 2.5
3
2 2

1.5 1.5 2

1 1
1
0.5 0.5

0 0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60 0 10 20 30 40 50 60
Aircraft Count Aircraft Count Aircraft Count

(a) p1, p4, p6 and p12 (b) p5, p7, p8 and p10 (c) p2, p3, p9 and p11

Figure 40: Collision risk versus aircraft count under low, medium, and high collision risk patterns.
Pattern p1, p4, p6, and p12 refer to the low collision risk area in the left panel. Pattern p5, p7, p8,
and p10 refer to the medium collision risk area in the middle panel. Pattern p2, p3, p9, and p11 refer
to the high collision risk area in the right panel.

In fact, this indicates that under “low-heat” patterns, the sector is operating nor-
mally and has enough workload/safety margin to accommodate more aircraft without
causing an evident increase in collision risk. Under “medium-heat” patterns, airspace
safety is still at an acceptable level, but a distinct increase in collision risk can be
observed. While under “high-heat” patterns, the collision risk increase dramatically,
and it is becoming impossible to handle more aircraft. Therefore, the three sets of
patterns can be regarded as the normal state, transition state, and capacity overload
state, respectively. Once a new traffic scenario is given, by generating a heatmap and
matching it to the best-fit pattern, one can identify whether this sector is overloaded.

5.3. Chapter Summary

In this chapter, to identify the transition phases in airspace capacity, a data-driven

method was proposed based on clustering techniques. This chapter first developed a
generic method to recognize different collision risk patterns in a given airspace with
flight data. The proposed method comprises three components: collision risk compu-
tation, collision risk heatmap generation, and collision risk pattern recognition. Then

113
an experimental study was carried out over an en-route sector in Singapore airspace
using one-month ADS-B flight data with the proposed method. The results show that
there are twelve patterns of collision risk in the sector, which can be further organized
into three classes: normal state, transition state, and capacity overload state. Then
for a given traffic scenario, the sector capacity phase can be identified by matching the
collision pattern to its best-fit pattern. The results highlight the efficacy of collision
risk as an effective indicator of airspace capacity phase transitions. By identifying the
phase transitions in airspace capacity, ATCOs can be better aware of the traffic situ-
ations in the airspace and address anticipated airspace saturation or safety concerns
for better air traffic accommodation.

114
Chapter 6. Short-term Air Traffic Flow Prediction

The research in this chapter has been published in the conference paper: Chunyao Ma, Sameer Alam,

Qing Cai, and Daniel Delahaye, “Sector Entry Flow Prediction Based on Graph Convolutional Net-

works.” In International Conference on Research in Air Transportation, Tampa, FL, USA, 19-23

June, 2022.

Besides identifying the airspace capacity overload, in the presence of limited ca-
pacity, an accurate short-term air traffic flow prediction, i.e., 0-2 hour look ahead time
[82], can be vital for ATCOs managing the forthcoming traffic flow situation and eval-
uating control strategies [59]. For instance, ATCOs can execute the agreed tactical
measures, such as traffic re-routing and flight level allocation, in advance, to provide
a smooth and efficient flow of traffic where demand is foreseen to exceed capacity or
disruptions such as convective weather appears, to utilize the available capacity to the
maximum extent [60].

Note that most of the current literature in air traffic flow prediction lacks consid-
eration of the highly dynamic, structural, and interrelated nature of air traffic flows
in the airspace, such as considering the air traffic flow at different positions in the
airspace as different time-series isolated from each other. This chapter proposes to
predict airspace sector entry flows based on a graph-represented structure of air traf-
fic, which intrinsically considers the dynamic spatial-temporal evolution of air traffic
from a graph perspective. First, the air traffic flow prediction problem is formulated
as a three-stage sector entry flow prediction, which considers three stages of an entry
flow to an airspace sector: the upstream sector, the entry sector, and the downstream
sector. Each sector entry flow is then represented by a node in the graph. Further,
the interactions among the entry flow in the airspace are described by the weighted
edges between nodes, which are modeled by a Word2vec neural network [231] using
the sequences of traffic flow transitions among the nodes. With the weighted graph

115
constructed and the temporal flows on the nodes extracted from the flight trajecto-
ries, an Attention-based Spatial-Temporal Graph Convolutional Network (ASTGCN)
module is adopted to capture spatial-temporal features of recent, daily-periodic, and
weekly-periodic air traffic flows in the airspace. The outputs from the ASTGCN mod-
ule based on the recent, daily, and weekly historical data are fused to predict the traffic
flow on the nodes (sector entry).

6.1. Problem Description

An airspace sector’s capability in handling air traffic is not only determined by the
number of flights that can be served but also influenced by the distribution of the flights
inside the sector [223]. Thus, besides predicting the number of future flights entering
a sector, it is also essential for air traffic flow predictions to provide the distribution
of the upcoming traffic flow of the sector, such as the number of flights crossing the
entry/exit points of the sector, the number of flights transiting through the air routes
in the sector, or the number of flight passing certain waypoints in the sector.

Figure 41: Visualization of one-day flight trajectories in French airspace.

In recent years, the gradual implementation of FRA has given flights freedom to
plan their routes between an entry point and an exit point without referring to the ATS
routes [232]. Fig. 41 shows the one-day flight trajectories (red lines) in French airspace

116
(sector boundaries shown as black lines), where the FRA has been implemented in
nearly 50% of the airspace above flight level 195 [233]. It can be observed from Fig.
41 that most flight trajectories follow straight paths instead of specific air routes in the
airspace. The entry/exit points of flights into a sector are no longer fixed at certain
positions at the sector boundary. Under this context, it is impractical to predict air
traffic flow on airways, fixed entry/exit points on sector boundaries, or waypoints.

Note that in the FRA, flights will remain subject to control from the sector ATCOs.
It can be noticed from Fig. 41 that most of the flows entering a sector are aggregated
by their upstream and downstream sectors. Therefore, predicting the number of flights
in the entry flow along with downstream and upstream information, i.e., where the
flow is from and where it is going to, can provide ATCOs an understanding of the
emerging flow in a sector to develop flow control strategies.

Given the above analysis, this chapter aims to predict the number of flights in the
sector entry flows. An entry flow of a sector is defined as the flow entering the sector
in which the flights have the same upstream and downstream sectors. The prediction
outcome will be the number of flights in each sector entry flow during a look-ahead
time. For instance, flow “S1S2S3” represents one entry flow to Sector “S1”, which
is from Sector “S2” and going to Sector “S3”. This chapter proposes to predict the
number of flights in flow “S1S2S3” during a given look-ahead time.

6.2. Methodology

6.2.1. Methodology Overview

The proposed method to predict the sector entry flows comprises seven steps.
The first step denoises the raw ADS-B data. The second to fourth steps organize
the denoised ADS-B data and airspace structure data into a spatial-temporal graph.
The last three steps build the machine learning framework to learn from the spatial-
temporal traffic information and make predictions about future traffic flow.

The flight information used in the chapter is extracted from ADS-B data. It faces

117
 Airspace nodes with temporal flow
Structure
edges
... node
sequence Word2vec
𝑆3 nodes vectors
𝑆3 𝑆4 𝑆…
𝑆1 𝑆2 𝑆4 0.205 − 1.048 1.398
𝑆4
𝑆1 𝑆4 𝑆1 𝑆3
𝑆4 𝑆1 𝑆3 0.648 − 0.327 0.195
𝑆1 𝑆2 𝑆4
𝑆3 𝑆4 𝑆… 1.023 − 0.793 0.785
𝑆2 ... ...

time-series on graph
prediction
ASTGCN 𝑡𝑖𝑚𝑒
recent Module
FC 𝑡 + ∆𝑡
Fusion daily
𝑡

ST block

⋯
4D trajectories weekly GCN+ConV
2
S/T Attention
Denoise 1

ST block
GCN+ConV
ADS-B recent
S/T Attention
daily
Feature weekly

Figure 42: Conceptual diagram of the proposed method for short-term sector entry flow prediction
based on graph convolutional networks.

plenty of noises in measuring barometric altitudes, which is susceptible to variations in

atmospheric conditions [234]. Therefore the first step is to remove the noises from the
barometric altitude data. From the denoised ADS-B data, 4D flight trajectories are
obtained, i.e., time-series data with geographic flight information. Therefore, in the
second step, by referring to the airspace structure data, i.e., sector boundaries, one can
map each 4D trajectory into a sequence of nodes at sector boundaries and determine
the corresponding times when the flight passes the nodes. In the third step, the node
sequences are fed to a Word2vec model, which extracts the notion of relatedness across
nodes and encodes each node into a numerical vector. By calculating the similarity
between the vectors, the relatedness between nodes can be quantified, i.e., weights on
the edges connecting the node pairs. In the fourth step, a spatial-temporal graph is
built to describe the flow of air traffic in the airspace by combining the flight time
series on each node and the weights on the edges.

118
 Based on the spatial-temporal graph established in the previous four steps, the last
three steps adopt a graph-based neural network to achieve the prediction of the sector
entry flow. Step five organizes the spatial-temporal graph obtained in step 4 into a
time series sliding over different time slices, from which the weekly, daily, and recent
flow features can be extracted. Step six adopts an ASTGCN module to capture the
dynamic spatial and temporal patterns from the weekly, daily and recent data. In the
seventh step, the outputs from the ASTGCN module of the three features are fused
to obtain the air traffic flow prediction for the future time slices.

The following subsections illustrate details about data denoising, spatial-temporal

graph construction, and the graph-based neural network to learn and predict air traffic
flow.

6.2.2. Data Denoising

ADS-B data for aircraft surveillance can give highly accurate aircraft position and
velocity information, providing a useful source for analytical solutions to effective and
efficient airspace usage [235]. However, ADS-B data still includes some noise in flights’
geographic information, especially barometric altitude [236]. The black line in Fig. 43
shows a flight trajectory extracted from the ADS-B data. Many spikes are observed
in different phases of the flight, especially in take-off, cruise, and landing. This section
will describe removing these spikes from the flight trajectory.

To remove the spikes from the cruise phase data, a median filter [237] is adopted.
The median filter moves through the trajectory data one by one and replaces each
value with the median value of its neighboring data. The green line in Fig. 43 shows
the denoised flight trajectory, in which the spikes in the cruise phase are smoothed
after applying the median filter.

However, the noise in the altitude data cannot be wholly flatten by a median filter,
as a spike can still be observed at the initial part of the green curve. The cause of the
altitude data busts during the take-off/landing phase may be the failure to change the

119
necessary QNH [238] setting during the take-off/landing phase to the correct altimeter
setting [239]. Thus, during the take-off/landing phase, the outlier data constantly show
up over a relatively long time window, compared to the cruise phase during which
the outlier data appear individually and are easy to be removed through a median
filter. Given the constant occurrence of inaccurate barometric altitude data during
take-off/landing, it is better to remove the trajectory data during these phases, i.e.,
below a certain flight level threshold, to avoid influencing the reliability of the research
results. Therefore, the Robust Locally Weighted Scatterplot Smoothing (RLOWESS)
[240], which uses locally weighted linear regression to smooth data, is applied onto
the filtered data (green line) to fit a smooth curve from which the overall trend of the
trajectory can be identified. The blue line in Fig. 43 shows the smoothed curve after
applying RLOWESS. By removing the corresponding data in the green line below a
certain flight level (above which the flight altitude measurement is stable) indicated
by the blue line, the final denoised flight trajectory can be obtained, as shown by the
red curve in Fig. 43.

Figure 43: Illustration of the proposed ADS-B data denoising approach.

120
6.2.3. Spatial-temporal Graph Construction

This section organizes the denoised flight trajectories and the airspace structure
information into a spatial-temporal graph representation G = (V, E), where V is
the set of nodes in the graph, and E is the set of edges indicating the connectivity
between the nodes. This section consists of nodes extraction, edges generation, and
graph construction.

6.2.3.1. Nodes Extraction. From the ADS-B data, 4D flight trajectories, containing
the geographic information of flights over time, can be obtained. By intersecting the
4D flight trajectories with sector boundaries, it can be determined where and when
a flight enters into each sector. An entry flow of a sector can be described as a node
defined by the entry sector, the upstream sector, and the downstream sector. More
specifically, when a flight fk enters sector Si from sector Si−1 then goes to sector Si+1 ,
it belongs to the entry flow of sector Si described by the node Si Si−1 Si+1 . ...

𝑆
𝑆3 𝑆1 𝑆…
...
...

𝑆1 𝑆2 𝑆3
𝑆3 𝑆3 𝑆4 𝑆…
𝑆3 𝑆1 𝑆…
𝑆4
𝑆4 𝑆1 𝑆3 𝑓1
𝑆1 𝑆2 𝑆3
𝑆1
𝑓2
𝑆1 𝑆2 𝑆4

𝑆2
𝑓1
𝑓2 𝑓3

Figure 44: An example of nodes extraction from three flight trajectories.

Fig. 44 presents an example of nodes extraction from three flight trajectories.

Flight f1 and f2 enter sector S1 from sector S2 , then go to sector S3 . Therefore,
according to the definition of entry flow in the Problem Description section, the two

121
flights belong to the entry flow of sector S1 represented by the node S1 S2 S3 . Similarly,
upon entering sector S3 , the two flights are classified to the entry flow of S3 represented
by the node S3 S1 S... , while S... represents the subsequent sector after exiting S3 . In
the same way, the nodes passed by flight f3 are S1 S2 S4 , S4 S1 S3 , and S3 S4 S... , when
entering sector S1 , S4 , and S3 , respectively. By identifying the nodes from each flight
trajectory in the ADS-B data, all the nodes in the traffic graph can be determined.

6.2.3.2. Edges Generation. After determining the nodes, the next step is to construct
the edges between the nodes. In this study, if two nodes are passed by a flight consec-
utively, they will be considered connected, and there will be an edge between them.
Most current research in graph convolutional networks for traffic flow prediction either
adopts a binary graph [241] or set the edge weights according to the geometric distance
between the nodes [242]. However, these simple representations of the weights may fail
to capture many complex features in the graph, e.g., some edges are frequently trav-
eled while some are seldom used, which may limit the ability of graph convolutional
networks to learn from the features of neighboring nodes [243]. Therefore, instead of
adopting a simple representation of the weights, this study proposes to learn the edge
weights from the traffic data, i.e., the sequence of nodes generated from each flight
trajectory.

Word2vec is a technique for natural language processing published in 2013 [231].

It uses a neural network to model word associations by training from a text corpus. In
this process, Word2Vec can build a vocabulary from the training text corpus and learn
the vector representations of each word. The cosine similarity between the vectors can
indicate semantic similarity between the words represented by those vectors [244].

Making an analogy between a node sequence and a text sentence with the nodes
being the words in the sentence, the set of node sequences generated from the flight
data can be regarded as the text corpus. Thus, by feeding the node sequences to
a Word2vec model, the vector representation for each node can be obtained. Sub-

122
sequently, the edge weights between nodes can be described by the cosine similarity
between the corresponding vectors. In this way, the weighted edges between the nodes
in the graph are determined.

6.2.3.3. Graph Construction. In the previous two steps, nodes and edges for a graph
have been determined. Each node, based on the time of each flight passing it, can be
associated with a time-varying traffic flow described by a sequence of flights on the tem-
poral dimension. More specifically, if flights f1 , f2 , ..., fk pass the node Si Si−1 Si+1 at an
increasing sequence of time t1 , t2 , ..., tk , the temporal flow on node Si Si−1 Si+1 can be
represented by a sequence {(f1 , t1 ), (f2 , t2 ), ..., (fk , tk )}. The nodes with dynamically
changing flow and the weighted edges between nodes constitute the spatial-temporal
graph structure.

6.2.4. Sector Entry Flow Prediction

Given the spatial-temporal graph constructed in the previous step, this section
predicts the flows, i.e., the number of flights, passing the nodes in the graph during a
future time slice.

With the temporal flight sequence recorded on each node in the graph G, a traffic
flow time series for each node over different time slices can be determined. xit ∈ R is
used to denote the traffic flow of node i at time t. Xt = (x1t , x2t , ..., xN
t ) ∈ R denotes
T N

the traffic flow values of all nodes at time t. N is the number of nodes in the graph.
χ = (X1 , X2 , ..., Xτ )T ∈ RN ×τ denotes the traffic flow values of all nodes over τ time
slices. Therefore, the air traffic flow prediction problem can be specified as: given χ,
traffic flow values at all the nodes on the graph over past τ time slices, predict future
traffic flow sequences Y = (y1 , y2 , ..., yN )T ∈ RN ×Tp of all the nodes on the whole
traffic graph over the next Tp time slices, where yi = (xiτ +1 , xiτ +2 , ..., xiτ +Tp ) ∈ RTp
denotes the future traffic flow of node i from τ + 1 to τ + Tp .

As shown in Fig. 42, the proposed method incorporate three time series features,

123
i.e., recent flow series, daily flow series, and weekly flow series, to predict the future
traffic flow. The three features are designed to respectively model the recent, daily-
periodic and weekly-periodic dependencies of the historical data. Assume the sampling
frequency is q times per hour, the current time is t0 , and the prediction window
size is Tp . Three time series segments during the last Th hours, Td days and Tw
weeks are extracted as the input of the recent, daily-period and weekly-period feature
respectively. Details about the three time series segments are illustrated below.

The recent segment is a segment of historical data adjacent to the predicting

period, which is designed to capture the influence of the recent traffic evolution
on the future traffic flows. This recent flow segment can be specified as: χh =
(Xt0 −Th ×q+1 , Xt0 −Th ×q+2 , ..., Xt0 ) ∈ RN ×Sh , while Sh = q × Th .

The daily-periodic segment is a time series of traffic flow over the past few days
during the same time period as the predicting period, which is designed to capture
the daily repeated patterns of air traffic flow. The daily segment can be specified as:
χd = (Xt0 −Td ×24q+1 , ..., Xt0 −Td ×24q+Tp , Xt0 −(Td −1)×24q+1 , ...,
Xt0 −(Td −1)×24q+Tp , ..., Xt0 −24q+1 , ..., Xt0 −24q+Tp ) ∈ RN ×Sd , while Sd = Tp × Td .

The weekly-periodic segment is composed of the flow series on last few weeks, which
have the same week attributes and time intervals as the forecasting period. It is de-
signed to capture the weekly repeated patterns of air traffic flow. The weekly segment
can be specified as: χw = (Xt0 −7Tw ×24q+1 , ..., Xt0 −7Tw ×24q+Tp , Xt0 −7(Tw −1)×24q+1 ,
..., Xt0 −7(Tw −1)×24q+Tp , ..., Xt0 −7×24q+1 , ..., Xt0 −7×24q+Tp )
∈ RN ×Sw , while Sw = Tp × Tw .

Each of the three time series feature will be fed to the ASTGCN module, shown
in Fig. 42, respectively to model the three temporal properties of air traffic flows.
The ASTGCN module consists of a number of spatial-temporal blocks as well as a
fully connected layer in the end to keep the output dimension the same as the fore-
casting target. A spatial-temporal block mainly consists of two mechanisms. First

124
is the spatial-temporal attention mechanism which is employed to capture the dy-
namic spatial-temporal correlations in the traffic data. The second mechanism is the
spatial-temporal convolution which uses graph convolutions and common standard
convolutions to respectively capture the spatial dependencies from the neighborhood
and the temporal dependencies from the time series. The spatial-temporal attention
combined with the spatial-temporal convolution forms a spatial-temporal block. Mul-
tiple spatial-temporal blocks can be further stacked to capture larger range of dynamic
spatial-temporal correlations.
Eventually, the outputs from the ASTGCN module based on each of the three
features are weighted fused to generate the final prediction results of traffic flow on
the nodes, depending on the significance of the influence of the three features. The
final prediction result after the fusion can be described as:

Ŷ = Wh ◦ Ŷh + Wd ◦ Ŷd + Ww ◦ Ŷw (29)

where ◦ represents the Hadamard product. Wh , Wd and Ww are the learned parameters
to quantify the influence of the three features on the forecasting target.

6.3. Experimental Study

With the proposed method, this section carries out an experimental study on the
French airspace using one-month ADS-B data from 1st December 2019 to 31st Decem-
ber 2019. A number of 158856 flights in French airspace are included in the study. The
prediction target is the hourly sector entry flows. The look ahead time investigated
in this experiment ranges from 1 hour to 2 hours. The following subsections presents
the result of the experimental study.

6.3.1. Graph Construction

By intersecting the flight trajectories with the sector boundaries in the French
airspace, 6028 nodes have been extracted for the traffic graph to represent different

125
entry flows of the airspace sectors. The edges in the graph are determined from the
sequence of nodes on a flight trajectory. The sequences of nodes are further fed to the
Word2vec neural network to model the edge weights. Fig. 45 shows the constructed
graph for the French airspace. Each red node in the graph denotes a sector entry flow,
and the node’s size is proportional to the size of the traffic flow passing through it.
Each blue line represents an edge in the graph, while the line thickness is proportional
to the edge weight.

Figure 45: The constructed graph for French airspace. Each red node denotes a sector entry flow and
each blue line denotes an edge in the graph.

6.3.2. Sector Entry Flow Prediction

The last three hours, six days, and one week flow features, i.e., Th = 3, Td = 6
days, and Tw = 1, of all nodes in the graph are extracted from the time series data and
fed to the ASTGCN module to predict the next one hour to two hours traffic flows
on the nodes. The ASTGCN module in this study adopts a widely used structure
that stacks two spatial-temporal blocks. The mean square error (MSE) between the
predicted and true values is used to compute the loss function of the model. The
sampling frequency of the hourly flow on the nodes is set as twelve times per hour,
i.e., q = 12. The batch size is five during the training phase, and the learning rate is

126
0.0003. The training, validation, and test data consist of 60%, 20%, and 20% of the
whole data set, respectively. Concretely, 18 days of data are used for model training,
the following six days for model validation, and the last six days for testing the model.

Figure 46: Flow prediction result on node “LFBBP2-LFBBP1-LFBBP3” from 00:00 to 23:59 on 30th
December 2019 (UTC). The upper panel plots the 0-1 hour prediction results, while the lower panel
plots the 1-2 hours prediction results. The red solid shows the true flow values, the green dash shows
the predicted values using the proposed GCN-based method, and the black dash shows the predicted
value using LSTM.

6.3.3. Result Analysis and Comparison

The node flow prediction results of the proposed method are compared with the
state of art time series prediction model LSTM. Given there are over 6000 nodes in
the graph, here a busy node in French airspace, node “LFBBP2-LFBBP1-LFBBP3”, is
used as an example to present the prediction result on the test dataset. According the
node definition, node “LFBBP2-LFBBP1-LFBBP3” represents an entry flow of sector
“LFBBP2”, which is from sector “LFBBP1” and heading to sector “LFBBP3”.

Fig. 46 presents the traffic flow prediction result from 00:00 to 23:59 on 30th
December 2019. The upper panel plots the prediction results for the next 1 hour

127
Table 4: Performance comparison between the proposed GCN-based method and the LSTM model
using three metrics: MAE, MSE and RMSE.

Look ahead Metric GCN LSTM

MAE 0.864 0.887
0-1 hr MSE 1.587 1.984
RMSE 1.260 1.409
MAE 1.402 2.016
0.5-1.5 hr MSE 3.384 7.956
RMSE 1.839 2.821
MAE 1.784 2.535
1-2 hr MSE 5.792 12.477
RMSE 2.406 3.532

Figure 47: Prediction performance (MAE, MSE and RMSE) of the proposed GCN-based method
and the LSTM model versus the prediction window.

(0-1 hour), while the lower panel plots the prediction for the next 1-2 hours. The
red solid shows the true number of flights passing node “LFBBP2-LFBBP1-LFBBP3”,
the green dash shows the predicted value using the proposed Graph Convolutional
Network (GCN) based method, and the black dash shows the predicted value using
LSTM. It can be observed from Fig. 46 that both the proposed method and the LSTM
model give forecasts that are close to the actual flow value in the coming 1 hour, while
the proposed method can better capture the sharp increases in traffic flow. When the
prediction window increases to 2 hours, the performance of the two models decreases.

128
The performance of the LSTM model reduces sharply while the proposed method can
still capture the traffic flow trends.

Table 4 shows the average performance comparison between the proposed method
and the LSTM model, including three metrics: Mean Absolute Error (MAE), Mean
Squared Error (MSE), and Root Mean Squared Error(RMSE). It can be seen from Ta-
ble 4 that the proposed method achieves better performance in terms of all evaluation
metrics. Fig. 47 shows the changes in the prediction performance of the two models as
the prediction window increases. Generally, the prediction challenge becomes greater
when the prediction window is longer. Consequently, the prediction errors will in-
crease. As can be seen from Fig. 47, LSTM can achieve good results in the short-term
prediction. However, with the increase in the prediction window size, the prediction
accuracy drops dramatically.

Figure 48: Cross-correlation between the true values and the lagged (shifted) copies of predicted
values. The horizontal axis shows the value of the time lag. The vertical axis shows the normalized
value of the cross-correlation.

The reason for the sharp decrease in the performance of the LSTM model may be
that the LSTM model learns the correlations in the flow times series on the target
node instead of the causal relationship between flows. Therefore its predictive power
reduces sharply when trying to estimate the target at a later stage. As shown in

129
Fig. 48, the predicted values using LSTM are highly correlated to the latest true
value available in the input data. When the prediction window is 1 to 2 hours, one
can observe a one-hour lag between the truth and the prediction of LSTM, while a
zero lag of the proposed method. The proposed method considers the spatial-temporal
correlations which are more important in the long-term prediction. It has the potential
to learn the causal relationship between flows from a graph perspective and produce
good prediction results even when the prediction window is larger.

6.4. Chapter Summary

This chapter proposed a method for short-term sector entry flow prediction based
on graph convolutional networks. Each entry flow of a sector was defined as a three-
stage flow based on the entry, upstream, and downstream sectors. The prediction
of each sector entry flow in the airspace was modeled as the prediction of temporal
features on each node in a graph. A Word2vec neural network is adopted to model
the interactions among the sector entry flows in the airspace, i.e., the weighted edges
between nodes, based on the transition of air traffic in the graph. The weighted graph
and the temporal flow series (hourly, daily, and weekly flows) on the nodes were input
to an ASTGCN module to capture spatial-temporal dependencies from historical data
and produce the final flow prediction results. An experimental study was carried out
using the proposed method in the 164 sectors of French airspace using the flight data
from 1st December 2019, to 31st December 2019, including 158856 flights. Results
showed that, compared with the LSTM model, the proposed method could better
capture sharp increases in traffic flow. And as the prediction window increased, the
proposed method had shown a slower drop in prediction accuracy, while the accuracy
of the LSTM model decreased drastically.

This flow prediction method can provide the number of flights in a sector entry flow
that comes from a particular upstream sector and goes to a certain downstream sector
during a given future time slice. The prediction result can help ATCOs better aware

130
of the distribution of anticipated air traffic flows in the sector and develop control
strategies, such as traffic re-routing and flight level allocation, in advance so that the
air traffic flows can be handled in a more organized manner.
The prediction of air traffic flow in this chapter only incorporates the feature of
flight number in each entry flow. Various other factors can also influence airspace
traffic situations, such as convective weather, airspace congestion, and operational re-
strictions. In the future, this traffic flow prediction method can be improved by taking
more influencing factors into consideration to deliver a better prediction accuracy.

131
Chapter 7. Conclusions and Future Work

7.1. Conclusions

This thesis investigated how to optimize the airspace to cater for the increasing
air traffic demand by modelling air traffic/airspace features as complex networks and
spatial-temporal graph representations for effective learning and prediction using data-
driven models.

Research Question 1:

How to identify the structural critical links and flow bottlenecks in a given
airway network and make minimal changes to the network structure, in
order to improve the air traffic flow efficiency without compromising its
structural integrity?

Findings:

This thesis proposed to improve airway network performance based on complex

network theories. Taking advantage of the network percolation and centrality
theories, this thesis proposed a complex network approach for critical link
detection in airway networks to dynamically identify airway links that are
vital for the network’s structural integrity and performance concerning time-
varying traffic situations. Moreover, inspired by the Braess’s Paradox (BP)
phenomenon, a scenario where an alteration to a traffic network to improve
traffic conditions actually has the reverse effect, this thesis provided a “counter-
intuitive” perspective towards airspace optimization by closing airways/links
in a given airway network to improve the air traffic flow efficiency. Through
the closure of airway links causing Braess’s Paradox, the one-day flight travel
duration in the airway network could be reduced by up to 3.8% at one flight
level.

132
 Note that air traffic congestion usually emerges locally on some airway links, which
impedes the smooth flow of air traffic through the airspace. Therefore, in airspace
structure optimization, this thesis proposed a network percolation and centrality the-
ory based critical link identification method, which identifies the links that are vital
for the structural integrity and performance of the networks concerning time-varying
traffic situations. A case study on the SEAN using one-month ADS-B data showed
that the proposed method, which is based on complex network theory, can help identify
dynamic airway links that are operationally critical, as identified by SMEs.

Apart from critical link identification, this thesis proposed an airway network struc-
ture management method to reduce flights’ travel duration in the airspace through
minimal modification of airway links. Inspired by Braess’s Paradox (BP), a situa-
tion where an alteration to a traffic network to improve traffic conditions actually has
the reverse effect, this thesis investigated a “counter-intuitive” approach to improve
air traffic efficiency by removing airways/links from a given airway network. An ex-
perimental study on SAN using six-month ADS-B data demonstrated that the BP
phenomenon does occur on airway networks. Through the closure of links causing BP,
for one day of traffic and one flight level, the total travel time can be reduced by up to
3.8%. Through critical link identification and minimal airway network structure mod-
ification, this thesis provides a spatial approach to airspace optimization to improve
air traffic flow.

From a temporal view for airspace optimization, this thesis proposed to make
airspace capacity overload identification and air traffic flow prediction to provide AT-
COs with knowledge about the airspace conditions so that the anticipated capacity
saturation and air traffic flow can be addressed more efficiently.

133
 Research Question 2:

How to identify airspace capacity phase transitions by leveraging structural

and air traffic flow features?

Findings:

This thesis proposed to use the airspace collision risk pattern as an indicator of
capacity phase transitions, which leverages the air traffic flow and the airspace
structural features. With air traffic data and airspace configurations, the colli-
sion risk distributions inside an airspace were organized into different patterns
based on the density and intensity of the collision risk to represent different
phase transitions in airspace capacity. An experimental study in Singapore
airspace demonstrated that distinct collision risk patterns exist for the normal
state, transition phase, and capacity overload state. The proposed method
identified the capacity phase transitions of given traffic situations by recogniz-
ing their best-fit collision risk patterns.

In terms of capacity overload identification, the mid-air collision risk pattern for a
given airspace was used as the indicator of airspace capacity overload. The collision
risk in an airspace was modeled from the air traffic data and airspace configurations,
which was further represented by heatmaps and clustered into different patterns based
on the collision risk density and intensity. For new traffic data during a given period,
the capacity overload status can be identified by recognizing the best-fit pattern of
the collision risk distribution. An experimental study in Singapore airspace identified
three phase transitions in airspace capacity: normal state (collision risk was kept at
a low level as traffic demand increased), transition phase (collision risk started to
gradually increase with the traffic demand), and capacity overload state (collision
risk increased sharply with the traffic demand). Compared to the approaches in the

134
literature, which either estimate the capacity using an analytical formula or identify
the capacity overload by observing the system limitation, the proposed method is able
to identify different phase transitions of the airspace capacity and notify ATCOs of
the potential airspace saturation.

Research Question 3:

How to improve short-term air traffic flow prediction by leveraging and learning
the spatial-temporal dependencies of air traffic flow in a given airspace?

Findings:

This thesis proposed to predict short-term en-route traffic flows based on a

graph convolutional network (GCN). A graph representation of air traffic struc-
ture was built from the real-world air traffic data based on learning word asso-
ciations from a large corpus of text, and the dynamic spatial-temporal depen-
dencies of air traffic flow were explored from a graph perspective. The results
of air traffic flow prediction showed that, compared to the well-established
Long Short-Term Memory (LSTM) model, the proposed method had better
capability to predict rapid changes in traffic flow with a smaller decrease in
prediction accuracy as the prediction window increased.

As for air traffic flow prediction, this thesis proposed to make short-term en-route
air traffic flow predictions based on graph convolutional networks. The air traffic
flow prediction problem was formulated as a three-stage sector entry flow prediction,
considering the upstream sector, the entry sector, and the downstream sector of an
entry flow to an airspace sector. A graph representation of the air traffic flow was
modeled from a Word2vec neural network based on real-world traffic data and the
airspace structure. Attention-based Spatial-Temporal Graph Convolutional Network

135
(ASTGCN) modules were adopted to capture spatial-temporal air traffic flow fea-
tures in the graph and make predictions about future traffic flow. An experimental
study in the French airspace demonstrated that, compared with the LSTM model,
the proposed method could better capture sharp increases in traffic flow and showed
less cross-correlation between the predicted values and the past time series data. As
the prediction window increased, the proposed method showed a slower drop in pre-
diction accuracy, while the accuracy of the LSTM model decreased drastically. Such
prediction may help ATFM plan and execute tactical measures for better air traffic
accommodation.

7.2. Future Work

The proposed approaches not only contribute to the airspace safety by advising the
potential overloads in the airspace but also improve the network traffic flow perfor-
mance in the airspace by optimizing the air traffic efficiency from a spatial-temporal
network perspective. Airway network management provides more flexibility for the
airspace structure to adapt to the varying traffic flow. Moreover, predicting the air traf-
fic flow and the airspace capacity overload improves the predictability of the airspace,
which can help ATCOs better utilize the airspace resources and contribute to a cost-
efficient air traffic management system.

Through airway network optimization, airspace capacity overload identification,

and air traffic flow prediction, this thesis paves the way to a better accommodation of
the increasing air traffic. However, the contribution of this thesis does not stop here.
The proposed approaches have the potential to act as an enabler of the future aviation
paradigm: sectorless/flight-centric ATC service, virtualization of ATM, integration of
airspace entrants, and digitization of the aviation system.

Sectorless/flight-centric ATC service: Facing the growth of air traffic demand

and the difficulties in subdividing heavily loaded sectors, many research teams, such
as SESAR and German Aerospace Center DLR, have been examining and testing

136
the concept of sector-less ATM/flight-centric ATC, which relies on a trajectory-based
individual control and monitoring of flights. This concept views the airspace as a whole
instead of the current practice of dividing the airspace into small sectors. Controllers
are responsible for a number of assigned aircraft from the entry into the sectorless
airspace until the exit. In this way, the number of ATCOs, as well as the assignment
of controller workload in an airspace, can be dynamically and flexibly optimized based
on the air traffic flow in scope since there is no longer a fixed number of geographical
sectors to operate.

Although the sectorless concept may have many benefits, its implementation has
been rather slow. One challenge is the coordination between controllers. In the tra-
ditional sector-based airspace, the coordination is primarily between different sectors,
while in a sectorless ATM environment, coordination between controllers is primarily
used to avoid potential conflicts in the air traffic flows. The highly dynamic nature of
air traffic brings uncertainties into the flight routes, traffic flow structure, and airspace
capacity, adversely affecting the effective, efficient, and safe coordination between con-
trollers for conflict resolution. Note that controllers are responsible for the entire
trajectory in the sectorless airspace. Accurate prediction of air traffic flows in the
airspace can provide airspace users network-wide flow information to plan and coordi-
nate the flight trajectories in advance so that complex situations can be avoided where
coordination is needed for conflict situations involving large numbers of aircraft. Iden-
tification of capacity overload in the airspace based on safety indicators can inform
the airspace users to avoid the potential saturated areas on the flight trajectories to
minimize flight conflicts and safety concerns.

Modification or redesign of the airway route structure to adapt to this new air
traffic management paradigm is another challenge. Similar to free flight operations,
in sectorless airspace, flights are no longer required to be handed over at certain
waypoints, and user-preferred trajectories can be planned. Given that flights can fly

137
their preferred trajectories to meet their business goals without being constrained by
airspace sector configurations, the complex structure of the current airway network
can lead to inefficiencies in air traffic flow and flights coordination, which requires the
gradual adaption of airway network, e.g., partial changes at some moments in time,
to the sectorless operations. The BP based airway network management method
modifies the airway network structure, i.e., closure & opening of some airway links
based on the traffic flow structure, under the user-preferred route paradigm to avoid
adverse interactions between the flight route choices of different airspace users. This
sheds light on improving the network flow efficiency through minimal airway network
changes in the future sectorless airspace.

Virtualization of ATM: Note that the traditional Air traffic control service
is established in geographically partitioned sectors, organized and managed by the
national ANSP in accordance with expected traffic demand and airspace availability.
This fragmentation of service provision may generate imbalanced controller workload,
cost-inefficiencies, and rigidity that make it impossible to adequately address the air
traffic growth of today and tomorrow. Moreover, the fragmented airspace sectors may
impede the transition of air traffic management to the future flow-driven paradigm,
which requires seamless control of flights across different geographical areas.

The virtualization of the ATM system, such as the virtual center and virtual tower,
provides an operating environment for different ATS units across different ANSPs to
appear as a single unit. It is an essential enabler of delivering ATC services to one
or more FIRs, irrespective of physical infrastructure or geographical location. The
virtualization of ATM overcomes the geographical limitations of ATS units and allows
each flight to be optimized as a whole rather than segmented portions of trajectory
in different sectors, which may be an opportunity for the collaborative planning and
decision-making of airspace users, flight operations centers, air traffic controllers, and
ATFM.

138
 This new development of the ATM network relies on improved data sharing among
different stakeholders and the support of common automation interfaces/tools. This
thesis provides airspace optimization strategies for airway network management, ca-
pacity overload identification, and air traffic flow prediction. Under the context of the
virtualized ATM, the proposed approaches can be further integrated into a mutual
decision support tool for the delivery of higher network performance of air traffic ac-
cording to the spatial-temporally varying air traffic density and complexity. With the
shared network-wide air traffic data in the virtualized ATM system, the air traffic flow
and airspace capacity phase transitions can be predicted based on what critical airway
links can be identified, and the airway network structure can be modified through BP
detection. This integrated decision support tool may enable consolidated, harmonized,
and cost-efficient ATM solutions to the adaption of ATFM operations, air traffic con-
trol strategies, and airspace structures to better demand and capacity management.
The proposed approaches have the potential to increase the integrity, predictability,
and coherence of the ATM system, bringing it one step closer to the flexible, dynamic,
network-centric optimization and management of air traffic in the future flight-centric
and virtualized ATM context.

Integration of airspace entrants: With the exponential growth in the use of

new forms of air transportation, such as remotely piloted aircraft systems (RPAS),
supersonic flights, and very high-altitude vehicles (above flight level 600), more pro-
found changes in the core ATM functionality are needed to support not only tradi-
tional airspace users but also very low/high-altitude airspace users. In tandem with
the more efficient air transportation infrastructure and services, the traditional ATM
stakeholders have started to roll out the interconnected ATM network. New entrants,
e.g., Unmanned Aircraft System (UAS) operators and very high-altitude operators,
are added into different airspace classes of the traditional aviation environment.

Facilitating the safe and effective integration of the aviation system calls for de-

139
veloping innovative traffic management solutions to enable greater levels of autonomy
and connectivity in both manned and unmanned aviation. The involvement of short-
term spatial-temporal air traffic flow prediction in managing heterogeneous airspace
users is essential for providing network situational awareness of different forms of air
traffic flow. By providing the anticipated air traffic flow from different airspace users,
accurate prediction of air traffic flow can enable the management of traffic flows in the
network in a more collaborative and dynamic manner with the increased involvement
of various airspace users. The transparency, robustness, and efficiency of complicated,
dynamic, and interconnected airspace operations can be improved consequently. It
may become possible for fleets of unmanned vehicles to perform a wide range of aerial
tasks safely and efficiently alongside manned flights. Thus, air traffic flow predic-
tion is an essential enabler of the seamless and safe management of different airspace
operators in the future compounded aviation ecosystem.

Digitization of the ATM system: Rapid advancement in enabling technolo-

gies, e.g., computational power, AI, and cloud database, with the development of
new concepts, e.g., flight-centric ATM, virtualization of ATM, and new forms of air
transportation, accelerate the digital transformation of the underlying aviation infras-
tructure system. The digital transformation leverages modern technologies through
a data-rich and cyber-secure connected digital ecosystem, which allows ground and
airborne resources (on the ) to be connected and optimized across the network, irre-
spective of geographical positions, classes of airspace, manned or unmanned aircraft
types. Compared with the current infrastructure blocks following geographically frag-
mented airspace, isolated technical and operational solutions, and limited interoper-
ability, the digitized aviation system will become more modular, connected, efficient,
and intelligent. More extensive air transportation services can be supported, and a
better experience can be delivered to all stakeholders.

The digitization of air traffic management can mainly be achieved by improved

140
data availability, enhanced data connectivity, and advanced data analytics for full
optimization of the network flows. The data-driven airspace optimization approaches
presented in this thesis are based on the exploration and generalization of ADS-B data
in the air transportation network and modeling traffic flow as graph representation
with neural networks to model the spatial-temporal flow dependencies. This enables
the development of the predicted traffic situation purely from available air traffic data,
which helps find the bottlenecks in the airway network and improves the network traffic
efficiency. Investigation of network theory with data-driven models may help identify
the critical links and assist in restructuring the airway network to accommodate the
demands under emerging traffic patterns. The findings may help bring forward a
digitized ATM environment to enable a smarter and safer ATM that constantly learns
from the environment.

141
References

[1] E. Sunil, Þ. Olafure, J. Ellerbroek, and J. Hoekstra, “Analyzing the effect

of traffic scenario properties on conflict count models,” in Proceedings of the
Conference: 2018 International Conference for Research on Air Transportation,
Barcelona, Spain, 26-29 June, 2018.

[2] EUROCONTROL, “Challenges of growth - annex 1 - flight fore-

cast to 2040,” https://www.eurocontrol.int/sites/default/files/2019-07/
challenges-of-growth-2018-annex1_0.pdf, accessed on 03 July, 2022.

[3] Statista, “Annual growth in global air traffic passenger demand

from 2006 to 2022,” https://www.statista.com/statistics/193533/
growth-of-global-air-traffic-passenger-demand/, accessed on 03 July, 2022.

[4] B. Monechi, V. D. Servedio, and V. Loreto, “Congestion transition in air traffic

networks,” PloS one, vol. 10, no. 5, p. e0125546, 2015.

[5] S. Çolak, A. Lima, and M. C. González, “Understanding congested travel in

urban areas,” Nature Communications, vol. 7, p. 10793, 2016.

[6] F. Netjasov, “Framework for airspace planning and design based on conflict risk
assessment: Part 1: Conflict risk assessment model for airspace strategic plan-
ning,” Transportation Research Part C: Emerging Technologies, vol. 24, pp.
190–212, 2012.

[7] H. A. Blom, J. Krystul, and G. Bakker, “A particle system for safety verification
of free flight in air traffic,” in Proceedings of the 45th IEEE Conference on
Decision and Control, San Diego, CA, USA. IEEE, 13-15 December, 2006, pp.
1574–1579.

[8] H. Balakrishnan, “Control and optimization algorithms for air transportation

systems,” Annual Reviews in Control, vol. 41, pp. 39–46, 2016.

142
 [9] “European air traffic control delays loom over summer air travel,” https://
www.iata.org/pressroom/pr/Pages/2018-07-18-01.aspx, accessed on January 28,
2019.

[10] W.-B. Du, M.-Y. Zhang, Y. Zhang, X.-B. Cao, and J. Zhang, “Delay causality
network in air transport systems,” Transportation Research Part E: Logistics
and Transportation Review, vol. 118, pp. 466–476, 2018.

[11] M. Ball, C. Barnhart, M. Dresner, M. Hansen, K. Neels, A. Odoni, E. Peterson,

L. Sherry, A. Trani, B. Zou et al., “Total delay impact study,” in NEXTOR
Research Symposium, Washington DC. http://www. nextor. org, 2010.

[12] FAA, “Cost of delay estimates,” https://www.faa.gov/data_research/aviation_

data_statistics/media/cost_delay_estimates.pdf, accessed on 17 March, 2019.

[13] ICAO, ICAO PANS-ATM Doc.4444, ATM/501, 15th Edition, Procedures for
Air Navigation Services. ICAO, 2010.

[14] H. Hind, A. El Omri, N. Abghour, K. Moussaid, and M. Rida, “Dynamic airspace

configuration: Review and open research issues,” in 2018 4th International
Conference on Logistics Operations Management (GOL), Le Havre, France.
IEEE, 10-12 April, 2018, pp. 1–7.

[15] V. Bulusu and R. Sengupta, “Improving productivity of low-altitude airspace:

Speed vs direction control,” IEEE Transactions on Intelligent Transportation
Systems, vol. 22, no. 9, pp. 5904–5913, 2021.

[16] ICAO, “ICAO long-term traffic forecasts,” INTERNATIONAL CIVIL AVIA-

TION ORGANIZATION, Tech. Rep., April, 2018. [Online]. Available: https://
www.icao.int/sustainability/Documents/LTF_Charts-Results_2018edition.pdf

143
[17] ACI, “Annual world airport traffic report, 2019 edition,” Airports Council
International, Tech. Rep., 2019. [Online]. Available: https://store.aci.aero/
product/annual-world-airport-traffic-report-2019/

[18] S. Gudmundsson, M. Cattaneo, and R. Redondi, “Forecasting temporal world

recovery in air transport markets in the presence of large economic shocks: The
case of covid-19,” Journal of Air Transport Management, vol. 91, p. 102007,
2021.

[19] IATA, “Air passenger numbers to recover in 2024,” https://www.iata.org/en/

pressroom/2022-releases/2022-03-01-01/, accessed on 17 March, 2019.

[20] M. Zhang, S. Chen, L. Sun, W. Du, and X. Cao, “Characterizing flight delay
profiles with a tensor factorization framework,” Engineering, vol. 7, no. 4, pp.
465–472, 2021.

[21] L. Hao, M. Hansen, Y. Zhang, and J. Post, “New york, new york: Two ways
of estimating the delay impact of new york airports,” Transportation Research
Part E: Logistics and Transportation Review, vol. 70, pp. 245–260, 2014.

[22] P. K. Menon, G. D. Sweriduk, and K. D. Bilimoria, “New approach for model-

ing, analysis, and control of air traffic flow,” Journal of Guidance, Control, and
Dynamics, vol. 27, no. 5, pp. 737–744, 2004.

[23] “Eurocontrol, sass-c glossary,” https://ext.eurocontrol.int/lexicon/index.php/

En-route_airspace, accessed on 10 October, 2019.

[24] Y. Lin, L. Li, P. Ren, Y. Wang, and W. Szeto, “From aircraft tracking data to
network delay model: A data-driven approach considering en-route congestion,”
Transportation Research Part C: Emerging Technologies, vol. 131, p. 103329,
2021.

144
[25] J. Chevalier, D. Delahaye, M. Sbihi, and P. Marechal, “Departure and arrival
routes optimization near large airports,” Aerospace, vol. 6, no. 7, p. 80, 2019.

[26] M. Liang, D. Delahaye, and P. Marechal, “Conflict-free arrival and departure

trajectory planning for parallel runway with advanced point-merge system,”
Transportation Research Part C: Emerging Technologies, vol. 95, pp. 207–227,
2018.

[27] S. Consortium et al., “The roadmap for sustainable air traffic management-
European ATM master plan,” 2012.

[28] S. Corver and G. Grote, “Uncertainty management in enroute air traffic control:
a field study exploring controller strategies and requirements for automation,”
Cognition, Technology & Work, vol. 18, no. 3, pp. 541–565, 2016.

[29] B. I. Team, “NextGen airspace optimization study,” Puget Sound Regional Coun-
cil, Information Center, Seattle, Washington, Tech. Rep., 2016.

[30] J. Rosenow and H. Fricke, “Impact of multi-criteria optimized trajectories on Eu-

ropean airline efficiency, safety and airspace demand,” Journal of Air Transport
Management, vol. 78, pp. 133–143, 2019.

[31] B. Monechi, V. Servedio, and V. Loreto, “An air traffic control model based local
optimization over the airways network,” Proceedings of the SESAR Innovation
Days EUROCONTROL, Madrid, Spain, 25-27 November, 2014.

[32] J. Rosenow, D. Strunck, and H. Fricke, “Free route airspaces in functional air
space blocks,” in 8th Sesar Innovation Days, Salzburg, Austria, 3-7 December,
2018.

[33] J. Rosenow, M. Lindner, and J. Scheiderer, “Advanced flight planning and the
benefit of in-flight aircraft trajectory optimization,” Sustainability, vol. 13, no. 3,
p. 1383, 2021.

145
[34] B. Hoffman and A. Bayen, “Concepts and models for NEXTGEN air traf-
fic flow management,” in NEXTOR Research Symposium. FAA Headquarters,
Washington DC, 6 March, 2008.

[35] T. Treimuth, “Dynamic optimization of airspace sector grouping,” Ph.D. disser-

tation, INPT, 2018.

[36] V. Ho-Huu, S. Hartjes, H. G. Visser, and R. Curran, “An optimization frame-

work for route design and allocation of aircraft to multiple departure routes,”
Transportation Research Part D: Transport and Environment, vol. 76, pp. 273–
288, 2019.

[37] T. Mihetec, S. Steiner, and D. Odić, “Utilization of flexible airspace structure

in flight efficiency optimization,” Promet-Traffic&Transportation, vol. 25, no. 2,
pp. 109–118, 2013.

[38] E. Sunil, J. Hoekstra, J. Ellerbroek, F. Bussink, A. Vidosavljevic, D. Delahaye,

and R. Aalmoes, “The influence of traffic structure on airspace capacity,” in
ICRAT2016-7th International Conference on Research in Air Transportation,
Philadelphia, USA, 20-24 June, 2016.

[39] M. Sergeeva, D. Delahaye, C. Mancel, and A. Vidosavljevic, “Dynamic airspace

configuration by genetic algorithm,” Journal of Traffic and Transportation
Engineering (English edition), vol. 4, no. 3, pp. 300–314, 2017.

[40] J. Nosedal, M. A. Piera, A. O. Solis, and C. Ferrer, “An optimization model to

fit airspace demand considering a spatio-temporal analysis of airspace capacity,”
Transportation Research Part C: Emerging Technologies, vol. 61, pp. 11–28,
2015.

[41] J. Nosedal, M. A. Piera, S. Ruiz, and A. Nosedal, “An efficient algorithm for

146
 smoothing airspace congestion by fine-tuning take-off times,” Transportation
Research Part C: Emerging Technologies, vol. 44, pp. 171–184, 2014.

[42] N. Doble, R. Hoffman, P. Lee, J. Mercer, B. Gore, N. Smith, and K. Lee, “Cur-
rent airspace configuration practices and their implications for future airspace
concepts,” in 8th Aviation Technology Integration Operations Seminar, AIAA,
Anchorage, Alaska, 14-19 September, 2008, pp. 8936–8948.

[43] L. L. Cruciol, A. C. de Arruda Jr, L. Weigang, L. Li, and A. M. Crespo, “Reward

functions for learning to control in air traffic flow management,” Transportation
Research Part C: Emerging Technologies, vol. 35, pp. 141–155, 2013.

[44] A. Cardillo, M. Zanin, J. Gómez-Gardenes, M. Romance, A. J. G. del Amo, and

S. Boccaletti, “Modeling the multi-layer nature of the European air transport
network: Resilience and passengers re-scheduling under random failures,” The
European Physical Journal Special Topics, vol. 215, no. 1, pp. 23–33, 2013.

[45] L. Lacasa, M. Cea, and M. Zanin, “Jamming transition in air transportation net-
works,” Physica A: Statistical Mechanics and its Applications, vol. 388, no. 18,
pp. 3948–3954, 2009.

[46] D. Gillen and W. G. Morrison, “Regulation, competition and network evolution

in aviation,” Journal of Air Transport Management, vol. 11, no. 3, pp. 161–174,
2005.

[47] G. Burghouwt, J. Hakfoort, and J. R. van Eck, “The spatial configuration of

airline networks in Europe,” Journal of Air Transport Management, vol. 9, no. 5,
pp. 309–323, 2003.

[48] K. Cai, J. Zhang, C. Zhou, X. Cao, and K. Tang, “Using computational intelli-
gence for large scale air route networks design,” Applied Soft Computing, vol. 12,
no. 9, pp. 2790–2800, 2012.

147
[49] A. Majumdar and W. Ochieng, “Factors affecting air traffic controller work-
load: Multivariate analysis based on simulation modeling of controller work-
load,” Transportation Research Record: Journal of the Transportation Research
Board, vol. 1788, no. 8, pp. 58–69, 2002.

[50] P. Kopardekar, J. Rhodes, A. Schwartz, S. Magyarits, and B. Willems, “Relation-

ship of maximum manageable air traffic control complexity and sector capacity,”
in 26th International Congress of the Aeronautical Sciences (ICAS), Anchorage,
Alaska, 14–19 September, 2008, pp. 15–19.

[51] A. Churchill, D. Lovell, and M. Ball, “Flight delay propagation impact on strate-
gic air traffic flow management,” Transportation Research Record: Journal of
the Transportation Research Board, vol. 2177, no. 13, pp. 105–113, 2010.

[52] M. Zhang, B. Liang, S. Wang, M. Perc, W. Du, and X. Cao, “Analysis of flight
conflicts in the Chinese air route network,” Chaos, Solitons & Fractals, vol. 112,
pp. 97–102, 2018.

[53] L. B. Real, I. Contreras, J.-F. Cordeau, R. S. de Camargo, and G. de Miranda,

“Multimodal hub network design with flexible routes,” Transportation Research
Part E: Logistics and Transportation Review, vol. 146, p. 102188, 2021.

[54] W. Tian, X. Dai, and M. Hu, “Systemic congestion propagation in the airspace
network,” Mathematical Problems in Engineering, vol. 2018, 2018.

[55] J.-P. B. Clarke, S. Solak, L. Ren, and A. E. Vela, “Determining stochas-

tic airspace capacity for air traffic flow management,” Transportation Science,
vol. 47, no. 4, pp. 542–559, 2013.

[56] C. Ma, Q. Cai, S. Alam, and V. N. Duong, “Airspace capacity overload identifica-
tion using collision risk patterns,” in 2020 International Conference on Artificial

148
 Intelligence and Data Analytics for Air Transportation (AIDA-AT), Singapore.
IEEE, 3-4 February, 2020, pp. 1–9.

[57] J. D. Welch, J. Y. Cho, N. K. Underhill, and R. A. DeLaura, “Sector work-

load model for benefits analysis and convective weather capacity prediction,” in
Tenth USA/Europe Air Traffic Management Research and Development Seminar
(ATM2013), Chicago, IL, USA, 10-13 June, 2013.

[58] W. Ding, Y. Zhang, and M. Hansen, “Downstream impact of flight rerouting,”

Transportation Research Part C: Emerging Technologies, vol. 88, pp. 176–186,
2018.

[59] D. Chen, M. Hu, Y. Ma, and J. Yin, “A network-based dynamic air traffic
flow model for short-term en route traffic prediction,” Journal of Advanced
Transportation, vol. 50, no. 8, pp. 2174–2192, 2016.

[60] B. Sridhar, T. Soni, K. Sheth, and G. Chatterji, “Aggregate flow model for air-
traffic management,” Journal of Guidance, Control, and Dynamics, vol. 29, no. 4,
pp. 992–997, 2006.

[61] E. Aydoğan and C. Çetek, “Point merge concept for en route air traffic flow
management,” Journal of Aircraft, vol. 55, no. 6, pp. 2203–2216, 2018.

[62] J. Carreras-Maide, O. Lordan, and J. M. Sallan, “Cost savings from trajectory

deviations in the European airspace,” Journal of Air Transport Management,
vol. 88, p. 101887, 2020.

[63] J. Zhang, F.-Y. Wang, K. Wang, W.-H. Lin, X. Xu, and C. Chen, “Data-driven
intelligent transportation systems: A survey,” IEEE Transactions on Intelligent
Transportation Systems, vol. 12, no. 4, pp. 1624–1639, 2011.

149
[64] N. G. Polson and V. O. Sokolov, “Deep learning for short-term traffic flow pre-
diction,” Transportation Research Part C: Emerging Technologies, vol. 79, pp.
1–17, 2017.

[65] A. M. Avila and I. Mezić, “Data-driven analysis and forecasting of highway traffic
dynamics,” Nature Communications, vol. 11, no. 1, pp. 1–16, 2020.

[66] Q. Zhou, W. Yang, and J. Zhu, “Mapping a multilayer air transport network
with the integration of airway, route, and flight network,” Journal of Applied
Mathematics, vol. 2019, 2019.

[67] S. Kim, M. E. Lewis, and C. C. White, “Optimal vehicle routing with real-time
traffic information,” IEEE Transactions on Intelligent Transportation Systems,
vol. 6, no. 2, pp. 178–188, 2005.

[68] D. J. Watts and S. H. Strogatz, “Collective dynamics of ‘small-world’networks,”

Nature, vol. 393, no. 6684, p. 440, 1998.

[69] A.-L. Barabási and R. Albert, “Emergence of scaling in random networks,”

Science, vol. 286, no. 5439, pp. 509–512, 1999.

[70] C. Molinero, R. Murcio, and E. Arcaute, “The angular nature of road networks,”
Scientific Reports, vol. 7, no. 1, pp. 1–11, 2017.

[71] R. Guimera and L. A. N. Amaral, “Modeling the world-wide airport network,”

The European Physical Journal B, vol. 38, no. 2, pp. 381–385, 2004.

[72] W.-B. Du, B.-Y. Liang, C. Hong, and O. Lordan, “Analysis of the Chinese
provincial air transportation network,” Physica A: Statistical Mechanics and its
Applications, vol. 465, pp. 579–586, 2017.

[73] A. Tizghadam and A. Leon-Garcia, “Robust network planning in nonuniform

150
 traffic scenarios,” Computer Communications, vol. 34, no. 12, pp. 1436–1449,
2011.

[74] Y. T. Adam, “Constructing a network science toolchain for analyzing network

traffic,” Ph.D. dissertation, Carnegie Mellon University Pittsburgh, PA, USA,
2018.

[75] W. Du, B. Liang, G. Yan, O. Lordan, and X. Cao, “Identifying vital edges in Chi-
nese air route network via memetic algorithm,” Chinese Journal of Aeronautics,
vol. 30, no. 1, pp. 330–336, 2017.

[76] S. Dunn and S. M. Wilkinson, “Increasing the resilience of air traffic networks us-
ing a network graph theory approach,” Transportation Research Part E: Logistics
and Transportation Review, vol. 90, pp. 39–50, 2016.

[77] Z. Tian, L. Jia, H. Dong, F. Su, and Z. Zhang, “Analysis of urban road traffic
network based on complex network,” Procedia Engineering, vol. 137, pp. 537–
546, 2016.

[78] B. D. Ripley, Pattern recognition and neural networks. Cambridge university

press, 2007.

[79] Y. Lv, Y. Duan, W. Kang, Z. Li, and F.-Y. Wang, “Traffic flow prediction
with big data: a deep learning approach,” IEEE Transactions on Intelligent
Transportation Systems, vol. 16, no. 2, pp. 865–873, 2014.

[80] M.-H. Nguyen and S. Alam, “Airspace collision risk hot-spot identification using
clustering models,” IEEE Transactions on Intelligent Transportation Systems,
vol. 19, no. 1, pp. 48–57, 2017.

[81] Z. Wang, H. Li, J. Wang, and F. Shen, “Deep reinforcement learning based
conflict detection and resolution in air traffic control,” IET Intelligent Transport
Systems, vol. 13, no. 6, pp. 1041–1047, 2019.

151
[82] Y. Lin, J.-w. Zhang, and H. Liu, “Deep learning based short-term air traffic
flow prediction considering temporal–spatial correlation,” Aerospace Science and
Technology, vol. 93, p. 105113, 2019.

[83] B. Yu, Z. Guo, S. Asian, H. Wang, and G. Chen, “Flight delay prediction for
commercial air transport: A deep learning approach,” Transportation Research
Part E: Logistics and Transportation Review, vol. 125, pp. 203–221, 2019.

[84] B. Yang, S. Sun, J. Li, X. Lin, and Y. Tian, “Traffic flow prediction using LSTM
with feature enhancement,” Neurocomputing, vol. 332, pp. 320–327, 2019.

[85] T. Kistan, A. Gardi, and R. Sabatini, “Machine learning and cognitive er-
gonomics in air traffic management: Recent developments and considerations
for certification,” Aerospace, vol. 5, no. 4, p. 103, 2018.

[86] ICAO, “Doc 9854: Global air traffic management operational concept, interna-
tional civil aviation organization,” 2005.

[87] A. S. A. Alochukwu and M. O. Olusanya, “On modeling and optimisation of air

traffic flow management problem with en-route capacities,” Ph.D. dissertation,
University of KwaZulu-Natal, South Africa, 2016.

[88] M. Schultz, I. Gerdes, T. Standfuß, and A. Temme, “Future airspace design

by dynamic sectorization,” in ENRI International Workshop on ATM/CNS,
Nakano, Japan. Springer, 14-16 November, 2017, pp. 19–34.

[89] EUROCONTROL, “Eurocontrol manual for airspace planning, 2003.”

[90] J. Tang, S. Alam, C. Lokan, and H. A. Abbass, “A multi-objective approach

for dynamic airspace sectorization using agent based and geometric models,”
Transportation Research Part C: Emerging Technologies, vol. 21, no. 1, pp. 89–
121, 2012.

152
[91] I. Gerdes, A. Temme, and M. Schultz, “Dynamic airspace sectorisation for flight-
centric operations,” Transportation Research Part C: Emerging Technologies,
vol. 95, pp. 460–480, 2018.

[92] C. S. Y. Wong, S. Sundaram, and N. Sundararajan, “Cdas: A cognitive decision-

making architecture for dynamic airspace sectorization for efficient operations,”
IEEE Transactions on Intelligent Transportation Systems, vol. 20, no. 5, pp.
1659–1668, 2018.

[93] X. Zou, P. Cheng, B. An, and J. Song, “Sectorization and configuration transition
in airspace design,” Mathematical Problems in Engineering, vol. 2016, 2016.

[94] D. Gianazza and N. Durand, “Ant colony systems for optimizing sequences of
airspace partitions,” in 2020 International Conference on Artificial Intelligence
and Data Analytics for Air Transportation (AIDA-AT). IEEE, 2020, pp. 1–10.

[95] P. Kopardekar, K. Bilimoria, and B. Sridhar, “Initial concepts for dynamic

airspace configuration,” in 7th AIAA ATIO Conf, 2nd CEIAT Int’l Conf on
Innov and Integr in Aero Sciences, 17th LTA Systems Tech Conf; followed by
2nd TEOS Forum, Belfast, Northern Ireland, 18-20 September, 2007, p. 7763.

[96] G. Mohammed and M. El Bekkaye, “Fuzzy dynamic airspace sectorization prob-

lem,” in Machine Intelligence and Data Analytics for Sustainable Future Smart
Cities. Springer, 2021, pp. 229–250.

[97] S.-C. Han and M. Zhang, “The optimization method of the sector partition based
on metamorphic voronoi polygon,” Chinese Journal of Aeronautics, vol. 17, no. 1,
pp. 7–12, 2004.

[98] M. Xue, “Airspace sector redesign based on voronoi diagrams,” Journal of

Aerospace Computing, Information, and Communication, vol. 6, no. 12, pp.
624–634, 2009.

153
 [99] ——, “Three-dimensional sector design with optimal number of sectors,” Journal
of guidance, control, and dynamics, vol. 35, no. 2, pp. 609–618, 2012.

[100] K. Zhou, Q. Hou, R. Wang, and B. Guo, “Real-time kd-tree construction on

graphics hardware,” ACM Transactions on Graphics (TOG), vol. 27, no. 5, pp.
1–11, 2008.

[101] C. W. S. Yin, T. Venugopalan, and S. Suresh, “A multi-objective approach for

3D airspace sectorization: A study on singapore regional airspace,” in 2016
IEEE Symposium Series on Computational Intelligence (SSCI), Athens, Greece.
IEEE, 6-9 December, 2016, pp. 1–8.

[102] C. Wong, S. Suresh, and N. Sundararajan, “A rolling horizon optimization

approach for dynamic airspace sectorization,” IFAC Journal of Systems and
Control, vol. 11, p. 100076, 2020.

[103] C. R. Brinton and S. Pledgie, “Airspace partitioning using flight clustering and
computational geometry,” in 2008 IEEE/AIAA 27th Digital Avionics Systems
Conference, St. Paul, MN, USA. IEEE, 26-30 October, 2008, pp. 3–B.

[104] A. Yousefi and G. Donohue, “Temporal and spatial distribution of airspace

complexity for air traffic controller workload-based sectorization,” in AIAA
4th Aviation Technology, Integration and Operations (ATIO) Forum, Chicago,
Illinois, USA, 20-22 September, 2004, p. 6455.

[105] M. Drew, “Analysis of an optimal sector design method,” in 2008 IEEE/AIAA

27th Digital Avionics Systems Conference, St. Paul, MN, USA. IEEE, 26-30
October, 2008, pp. 3–B.

[106] P. Jägare, P. Flener, and J. Pearson, “Airspace sectorisation using constraint-

based local search,” in Tenth USA/Europe Air Traffic Management Research

154
 and Development Seminar (ATM2013), Chicago, IL, USA. Federal Aviation
Administration, 10-13 June, 2013.

[107] J. Li, T. Wang, M. Savai, and I. Hwang, “Graph-based algorithm for dynamic
airspace configuration,” Journal of Guidance, Control, and Dynamics, vol. 33,
no. 4, pp. 1082–1094, 2010.

[108] Y. Chen and D. Zhang, “Dynamic airspace configuration method based on a

weighted graph model,” Chinese Journal of Aeronautics, vol. 27, no. 4, pp. 903–
912, 2014.

[109] S. Zelinski, “A comparison of algorithm generated sectorizations,” Air Traffic

Control Quarterly, vol. 18, no. 3, pp. 279–301, 2010.

[110] A. Wellens and M. M. Mota, “Modeling approach for managing the demand in
congested airport networks: The case of mexico city airport,” in 2017 Winter
Simulation Conference (WSC), Las Vegas, NV, USA. IEEE, 3-6 December,
2017, pp. 4535–4537.

[111] D. Li, B. Fu, Y. Wang, G. Lu, Y. Berezin, H. E. Stanley, and S. Havlin, “Perco-
lation transition in dynamical traffic network with evolving critical bottlenecks,”
Proceedings of the National Academy of Sciences, vol. 112, no. 3, pp. 669–672,
2015.

[112] A. Cook, H. A. Blom, F. Lillo, R. N. Mantegna, S. Micciche, D. Rivas,

R. Vázquez, and M. Zanin, “Applying complexity science to air traffic man-
agement,” Journal of Air Transport Management, vol. 42, pp. 149–158, 2015.

[113] E. Yu, D. Chen, and J. Zhao, “Identifying critical edges in complex networks,”
Scientific Reports, vol. 8, no. 1, pp. 1–8, 2018.

155
[114] L. Liu, B. Yin, S. Zhang, X. Cao, and Y. Cheng, “Deep learning meets wireless
network optimization: Identify critical links,” IEEE Transactions on Network
Science and Engineering, vol. 7, no. 1, pp. 167–180, 2018.

[115] Q. Cai, M. Pratama, and S. Alam, “Interdependency and vulnerability of mul-

tipartite networks under target node attacks,” Complexity, vol. 2019, Art. No.
2680972, 2019.

[116] D. Z. Wang, H. Liu, W. Szeto, and A. H. Chow, “Identification of critical com-

bination of vulnerable links in transportation networks–a global optimisation
approach,” Transportmetrica A: Transport Science, vol. 12, no. 4, pp. 346–365,
2016.

[117] X. Xu, A. Chen, and C. Yang, “An optimization approach for deriving upper
and lower bounds of transportation network vulnerability under simultaneous
disruptions of multiple links,” Transportation Research, vol. 94, no. 9, pp. 338–
353, 2018.

[118] E. L. de Oliveira, L. da Silva Portugal, and W. P. Junior, “Indicators of reliability

and vulnerability: Similarities and differences in ranking links of a complex road
system,” Transportation Research Part A: Policy and Practice, vol. 88, pp. 195–
208, 2016.

[119] P. Gauthier, A. Furno, and N.-E. El Faouzi, “Road network resilience: how
to identify critical links subject to day-to-day disruptions,” Transportation
Research Record, vol. 2672, no. 1, pp. 54–65, 2018.

[120] R. Vodák, M. Bíl, T. Svoboda, Z. Křivánková, J. Kubeček, T. Rebok, and

P. Hliněnỳ, “A deterministic approach for rapid identification of the critical links
in networks,” PloS One, vol. 14, no. 7, p. e0219658, 2019.

156
[121] M. Du, X. Jiang, and L. Cheng, “Alternative network robustness measure using
system-wide transportation capacity for identifying critical links in road net-
works,” Advances in Mechanical Engineering, vol. 9, no. 4, pp. 1–12, 2017.

[122] M. Guettiche and H. Kheddouci, “Critical links detection in stochastic networks:

application to the transport networks,” International Journal of Intelligent
Computing and Cybernetics, vol. 12, no. 1, pp. 42–69, 2019.

[123] P. Ren and L. Li, “Characterizing air traffic networks via large-scale aircraft
tracking data: A comparison between china and the US networks,” Journal of
Air Transport Management, vol. 67, pp. 181–196, 2018.

[124] T. Bröhl and K. Lehnertz, “Centrality-based identification of important edges in

complex networks,” Chaos: An Interdisciplinary Journal of Nonlinear Science,
vol. 29, no. 3, p. 033115, 2019.

[125] I. Kivimäki, B. Lebichot, J. Saramäki, and M. Saerens, “Two betweenness cen-

trality measures based on randomized shortest paths,” Scientific Reports, vol. 6,
no. 1, pp. 1–15, 2016.

[126] Y. Shen, N. P. Nguyen, Y. Xuan, and M. T. Thai, “On the discovery of critical
links and nodes for assessing network vulnerability,” IEEE/ACM Transactions
on Networking, vol. 21, no. 3, pp. 963–973, 2012.

[127] E. L. de Oliveira, L. da Silva Portugal, and W. P. Junior, “Determining critical

links in a road network: vulnerability and congestion indicators,” Procedia-Social
and Behavioral Sciences, vol. 162, pp. 158–167, 2014.

[128] C. Yang, A. Chen, X. Xu, and S. Wong, “Sensitivity-based uncertainty anal-

ysis of a combined travel demand model,” Transportation Research Part B:
Methodological, vol. 57, pp. 225–244, 2013.

157
[129] H. Feng, F. Bai, and Y. Xu, “Identification of critical roads in urban transporta-
tion network based on GPS trajectory data,” Physica A: Statistical Mechanics
and its Applications, vol. 535, p. 122337, 2019.

[130] X. Yang, L. Liu, Y. Li, and R. He, “Identifying critical links in urban traffic
networks: a partial network scan algorithm,” Kybernetes, vol. 45, no. 6, pp.
915–930, 2016.

[131] Y. Zhou and J. Wang, “Critical link analysis for urban transportation systems,”
IEEE Transactions on Intelligent Transportation Systems, vol. 19, no. 2, pp.
402–415, 2017.

[132] L. Dai, B. Derudder, and X. Liu, “The evolving structure of the southeast asian
air transport network through the lens of complex networks, 1979–2012,” Journal
of Transport Geography, vol. 68, pp. 67–77, 2018.

[133] N. Durand, D. Gianazza, J.-B. Gotteland, and J.-M. Alliot, Metaheuristics for
Air Traffic Management, Chapter 2. John Wiley & Sons, 2015.

[134] S. Alam, S. Chaimatanan, D. Delahaye, and E. Féron, “A distributed air traffic

flow management model for European functional airspace blocks,” in 12th ATM
R&D Seminar, 2017.

[135] C.-l. Chen and W.-b. Du, “A multi-objective crossing waypoints location opti-
mization in air route network,” in 2011 3rd International Workshop on Intelligent
Systems and Applications, Wuhan, China. IEEE, 28-29 May, 2011, pp. 1–4.

[136] W. Shijin, C. Xi, L. Haiyun, L. Qingyun, H. Xu, and W. Yanjun, “Air route net-
work optimization in fragmented airspace based on cellular automata,” Chinese
Journal of Aeronautics, vol. 30, no. 3, pp. 1184–1195, 2017.

158
[137] M. Borhani, “A multicriteria optimization for flight route networks in large-scale
airlines using intelligent spatial information,” International Journal of Interactive
Multimedia & Artificial Intelligence, vol. 6, no. 1, 2020.

[138] B. Pang, W. Dai, X. Hu, F. Dai, and K. H. Low, “Multiple air route crossing
waypoints optimization via artificial potential field method,” Chinese Journal of
Aeronautics, vol. 34, no. 4, pp. 279–292, 2021.

[139] Y. Li, Y. Zhang, and B. Feng, “A design for crossing waypoints location
in air route network based on genetic algorithm,” in 2020 IEEE 9th Joint
International Information Technology and Artificial Intelligence Conference
(ITAIC), Chongqing, China, vol. 9. IEEE, 11-13 December, 2020, pp. 827–
832.

[140] K.-q. Cai, J. Zhang, and C. Zhou, “Optimization of the crossing waypoints in
air route network,” in 29th Digital Avionics Systems Conference, Salt Lake City,
UT, USA. IEEE, 3-7 October, 2010, pp. 2–E.

[141] H. Hering, “Air traffic freeway system for Europe,” Eurocontrol Experimental
Centre, EEC Note, no. 20/05, 2005.

[142] D. Delahaye, S. Puechmorel, S. Alam, and E. Féron, “Trajectory mathematical

distance applied to airspace major flows extraction,” in EIWAC 2017 The 5th
ENRI International Workshop on ATM/CNS, 2017.

[143] X. Tang, Y. Zhang, P. Chen, B. Li, and S. Han, “Strategic deconfliction of

4d trajectory and perturbation analysis for air traffic control and automation
system,” Discrete Dynamics in Nature and Society, vol. 2016, 2016.

[144] N. Barnier and C. Allignol, “4D-trajectory deconfliction through departure

time adjustment,” in 8th USA/Europe Air Traffic Management Research and
Development Seminar, Napa, USA, 29 June-2 July, 2009.

159
[145] S. Ruiz, M. A. Piera, J. Nosedal, and A. Ranieri, “Strategic de-confliction in the
presence of a large number of 4D trajectories using a causal modeling approach,”
Transportation Research Part C: Emerging Technologies, vol. 39, pp. 129–147,
2014.

[146] A. Vitali, M. Battipede, and A. Lerro, “Multi-objective and multi-phase 4D tra-

jectory optimization for climate mitigation-oriented flight planning,” Aerospace,
vol. 8, no. 12, p. 395, 2021.

[147] J. Welch, “En route sector capacity model final report,” ATC-426, 2015.

[148] B. Marr and K. Lindsay, “Controller workload-based calculation of monitor alert

parameters for en route sectors,” in 15th AIAA Aviation Technology, Integration,
and Operations Conference, Dallas, TX, 22-26 June, 2015, pp. 3178–3188.

[149] P. Kopardekar and S. Magyarits, “Dynamic density: measuring and predict-

ing sector complexity [atc],” in Proceedings. The 21st Digital Avionics Systems
Conference,Irvine, CA, USA, vol. 1. IEEE, 27-31 October, 2002.

[150] S. Inoue, K. Furuta, K. Nakata, T. Kanno, H. Aoyama, and M. Brown, “Cogni-

tive process modelling of controllers in en route air traffic control,” Ergonomics,
vol. 55, no. 4, pp. 450–464, 2012.

[151] P. Averty, S. Athenes, C. Collet, and A. Dittmar, “Evaluating a new index of

mental workload in real ATC situation using psychophysiological measures,” in
Proceedings. The 21st Digital Avionics Systems Conference,Irvine, CA, USA,
vol. 2. IEEE, 27-31 October, 2002.

[152] S. K. Parambath, “Capacity of ATC sectors in chennai upper area control cen-
tre,” in 2020 IEEE International Conference on Electronics, Computing and
Communication Technologies (CONECCT), Bangalore, India. IEEE, 2-4 July,
2020, pp. 1–7.

160
[153] A. J. Masalonis, M. B. Callaham, and C. R. Wanke, “Dynamic density and
complexity metrics for realtime traffic flow management,” in Proceedings of the
5th USA/Europe Air Traffic Management R & D Seminar, Budapest, Hungary,
23-27 June, 2003, p. 139.

[154] D. Gianazza and K. Guittet, “Selection and evaluation of air traffic complexity
metrics,” in 2006 ieee/aiaa 25TH Digital Avionics Systems Conference, Portland,
OR, USA. IEEE, 15-19 October, 2006, pp. 1–12.

[155] K. Zhang, Y. Liu, J. Wang, H. Song, and D. Liu, “Tree-based airspace capacity
estimation,” in 2020 Integrated Communications Navigation and Surveillance
Conference (ICNS), Herndon, VA, USA. IEEE, 8-10 September, 2020, pp.
5C1–1.

[156] M. Majeed and R. Su, “Neural partial differentiation-based estimation of termi-

nal airspace sector capacity,” SAE International Journal of Aerospace, vol. 14,
no. 2, p. 203, 2021.

[157] L. Song, C. Wanke, and D. Greenbaum, “Predicting sector capacity for TFM,”
in 7th USA/Europe air traffic management research and development seminar
(ATM2007), Barcelona, Spain, vol. 15, 02 - 05 July, 2007.

[158] T. Kohonen, “Essentials of the self-organizing map,” Neural Networks, vol. 37,
pp. 52–65, 2013.

[159] J. Chen, K. Cai, W. Li, S. Tang, and J. Fang, “An airspace capacity estima-
tion model based on spatio-temporal graph convolutional networks considering
weather impact,” in 2021 IEEE/AIAA 40th Digital Avionics Systems Conference
(DASC), San Antonio, TX, USA. IEEE, 3-7 October, 2021, pp. 1–7.

[160] B. Számel, I. Mudra, and G. Szabó, “Applying airspace capacity estimation

161
 models to the airspace of Hungary,” Periodica Polytechnica Transportation
Engineering, vol. 43, no. 3, pp. 120–128, 2015.

[161] International Civil Aviation Organization, “Air traffic services planning man.
first edition 1984.”

[162] N. Schefers, M. A. Piera, J. J. Ramos, and J. Nosedal, “Causal analysis of airline

trajectory preferences to improve airspace capacity,” Procedia Computer Science,
vol. 104, pp. 321–328, 2017.

[163] D. Gianazza, “Learning air traffic controller workload from past sector oper-
ations,” in ATM Seminar, 12th USA/Europe Air Traffic Management R&D
Seminar, Seattle, United States, 26-30 June, 2017.

[164] E. Sunil, J. Smisek, M. M. van Paassen, and M. Mulder, “Validation of a tuning

method for haptic shared control using neuromuscular system analysis,” in 2014
IEEE International Conference on Systems, Man, and Cybernetics (SMC), San
Diego, CA, USA. IEEE, 5-8 October, 2014, pp. 1499–1504.

[165] W. Gao and M. Liang, “Airspace capacity assessment with augmented cell trans-
mission model,” Preprints, 2021.

[166] E. C. Fernández, J. M. Cordero, G. Vouros, N. Pelekis, T. Kravaris, H. Georgiou,

G. Fuchs, N. Andrienko, G. Andrienko, E. Casado et al., “DART: a machine-
learning approach to trajectory prediction and demand-capacity balancing,” in
SESAR Innovation Days, Belgrade, Serbia, 28–30 November, 2017.

[167] I. Lymperopoulos, J. Lygeros, and A. Lecchini, “Model based aircraft trajec-

tory prediction during takeoff,” in AIAA Guidance, Navigation, and Control
Conference and Exhibit, Keystone, Colorado, 21-24 August, 2006.

162
[168] P. Wei, Y. Cao, and D. Sun, “Total unimodularity and decomposition method
for large-scale air traffic cell transmission model,” Transportation Research Part
B: Methodological, vol. 53, pp. 1–16, 2013.

[169] E. Gilbo and S. Smith, “A new model to improve aggregate air traffic demand pre-
dictions,” in AIAA Guidance, Navigation and Control Conference and Exhibit,
Hilton Head, South Carolina, 20-23 August, 2007, p. 6450.

[170] D. Chen, M. Hu, H. Zhang, J. Yin, and K. Han, “A network based dynamic
air traffic flow model for en route airspace system traffic flow optimization,”
Transportation Research Part E: Logistics and Transportation Review, vol. 106,
pp. 1–19, 2017.

[171] Z. Zhang, A. Zhang, C. Sun, S. Xiang, J. Guan, and X. Huang, “Research on air
traffic flow forecast based on ELM non-iterative algorithm,” Mobile Networks
and Applications, vol. 26, no. 1, pp. 425–439, 2021.

[172] G. Gui, Z. Zhou, J. Wang, F. Liu, and J. Sun, “Machine learning aided air
traffic flow analysis based on aviation big data,” IEEE Transactions on Vehicular
Technology, vol. 69, no. 5, pp. 4817–4826, 2020.

[173] R. Dalmau-Codina, B. Genestier, C. Anoraud, P. Choroba, and D. Smith, “A

machine learning approach to predict the evolution of air traffic flow management
delay,” in 2021 ATM Seminar virtual event, 20-24 September, 2021.

[174] T. Bröhl and K. Lehnertz, “Centrality-based identification of important edges in

complex networks,” Chaos: An Interdisciplinary Journal of Nonlinear Science,
vol. 29, no. 3, p. 033115, 2019.

[175] F. Morone and H. A. Makse, “Influence maximization in complex networks

through optimal percolation,” Nature, vol. 524, no. 7563, pp. 65–68, 2015.

163
[176] S. Boccaletti, J. Almendral, S. Guan, I. Leyva, Z. Liu, I. Sendiña-Nadal,
Z. Wang, and Y. Zou, “Explosive transitions in complex networks’ structure
and dynamics: Percolation and synchronization,” Physics Reports, vol. 660, pp.
1–94, 2016.

[177] L. Liu, Z. Qiu, G. Li, Q. Wang, W. Ouyang, and L. Lin, “Contextualized

spatial–temporal network for taxi origin-destination demand prediction,” IEEE
Transactions on Intelligent Transportation Systems, vol. 20, no. 10, pp. 3875–
3887, 2019.

[178] Q. Cai, M. Pratama, S. Alam, C. Ma, and J. Liu, “Breakup of directed multipar-
tite networks,” IEEE Transactions on Network Science and Engineering, vol. 7,
no. 3, pp. 947–960, 2019.

[179] P. Baumgarten, R. Malina, and A. Lange, “The impact of hubbing concentration

on flight delays within airline networks: An empirical analysis of the US domestic
market,” Transportation Research Part E: Logistics and Transportation Review,
vol. 66, pp. 103–114, 2014.

[180] M. Barthélemy, “Spatial networks,” Physics Reports, vol. 499, no. 1-3, pp. 1–101,
2011.

[181] T. Lehouillier, F. Soumis, J. Omer, and C. Allignol, “Measuring the interac-

tions between air traffic control and flow management using a simulation-based
framework,” Computers & Industrial Engineering, vol. 99, pp. 269–279, 2016.

[182] C. Verlhac and S. Manchon, “Optimization of opening schemes,” in Proceedings

of the fourth USA/Europe Air Traffic Management R&D Seminar, Santa, USA,
4-7 December, 2001, pp. 8–49.

[183] K. Pien, K. Han, W. Shang, A. Majumdar, and W. Ochieng, “Robustness analy-

164
 sis of the European air traffic network,” Transportmetrica A: Transport Science,
vol. 11, no. 9, pp. 772–792, 2015.

[184] S. Starita, A. K. Strauss, X. Fei, R. Jovanović, N. Ivanov, G. Pavlović, and

F. Fichert, “Air traffic control capacity planning under demand and capacity
provision uncertainty,” Transportation Science, vol. 54, no. 4, pp. 882–896, 2020.

[185] K. Hanson, J. Gulding, and A. Afshar, “Analyzing the operational ca-

pacity effects of the monitor alert paramater (MAP),” in 2016 Integrated
Communications Navigation and Surveillance (ICNS), Herndon, VA, USA.
IEEE, 19-21 April, 2016, pp. 6C1–1.

[186] ICAO, “Report of the fourth meeting of the southeast Asia route re-
view task force,” https://www.icao.int/APAC/Meetings/2010/sea_rr_tf4/
SEA-RRTF4rpt.pdf, 2010, accessed on 03 July, 2022.

[187] J. Milan, “A practical capacity model of an air route network,” Transportation

Planning and Technology, vol. 11, no. 2, pp. 87–103, 1986.

[188] X. Sun, S. Wandelt, and F. Linke, “On the topology of air navigation route
systems,” in Proceedings of the Institution of Civil Engineers-Transport, vol.
170, no. 1. Thomas Telford Ltd, 2017, pp. 46–59.

[189] A. Winick and D. Brandewie, “VOR/DME system improvements,” Proceedings

of the IEEE, vol. 58, no. 3, pp. 430–437, 1970.

[190] “Aeronautical information publication & amendments,” https://www.caas.gov.

sg/docs/default-source/pdf/aip-singapore---31-dec-20.pdf, 2020, accessed on 03
July, 2022.

[191] SkyVector, “SkyVector aeronautical charts,” https://skyvector.com, 2021, ac-

cessed on 30 March, 2021.

165
[192] ICAO, “The fourth meeting of traffic flow review group: Harmonisa-
tion of effort to restructure ATS routes over the South China Sea,”
https://www.icao.int/APAC/Meetings/2016%20SCSTFRG4/SCSTFRG4%
20WP06%20Harmonization%20of%20Effort%20to%20Restructure%20ATS%
20Routes%20Over%20the%20South%20China%20Sea-Singapore.pdf, 2016,
accessed on 03 July, 2022.

[193] ——, “The fourth meeting of South China Sea traffic flow review group:
Harmonization of parallel routes structure with existing ATS routes
M751 within Kuala Lumpur FIR and M644 within Bangkok FIR,”
https://www.icao.int/APAC/Meetings/2016%20SCSTFRG4/SCSTFRG4%
20WP13%20Harmonization%20of%20Parallel%20Routes%20Structure%
20with%20Existing%20ATS%20Routes%20M751%20within%20%20KUALA%
20LUMPUR%20FIR%20and%20M644.pdf, 2016, accessed on 03 July, 2022.

[194] D. Braess, “Über ein paradoxon aus der verkehrsplanung,”

Unternehmensforschung, vol. 12, no. 1, pp. 258–268, 1968.

[195] D. Braess, A. Nagurney, and T. Wakolbinger, “On a paradox of traffic planning,”

Transportation Science, vol. 39, no. 4, pp. 446–450, 2005.

[196] E. I. Pas and S. L. Principio, “Braess’ paradox: Some new insights,”

Transportation Research Part B: Methodological, vol. 31, no. 3, pp. 265–276,
1997.

[197] M. Frank, “The Braess paradox,” Mathematical Programming, vol. 20, no. 1, pp.
283–302, 1981.

[198] C. M. Penchina, “Braess paradox: Maximum penalty in a minimal critical net-

work,” Transportation Research Part A: Policy and Practice, vol. 31, no. 5, pp.
379–388, 1997.

166
[199] G. Valiant and T. Roughgarden, “Braess’s paradox in large random graphs,”
Random Structures & Algorithms, vol. 37, no. 4, pp. 495–515, 2010.

[200] J. G. Wardrop, “Some theoretical aspects of road traffic research,” Proceedings

of the Institution of Civil Engineers, vol. 1, pp. 325–362, 1952.

[201] S. A. Bagloee, A. Ceder, M. Tavana, and C. Bozic, “A heuristic methodology

to tackle the Braess paradox detecting problem tailored for real road networks,”
Transportmetrica A: Transport Science, vol. 10, no. 5, pp. 437–456, 2014.

[202] L. Sun, L. Liu, Z. Xu, Y. Jie, D. Wei, and P. Wang, “Locating inefficient links in
a large-scale transportation network,” Physica A: Statistical Mechanics and its
Applications, vol. 419, pp. 537–545, 2015.

[203] K. Park, “Detecting Braess paradox based on stable dynamics in general con-
gested transportation networks,” Networks and Spatial Economics, vol. 11, no. 2,
pp. 207–232, 2011.

[204] J. N. Hagstrom and R. A. Abrams, “Characterizing Braess’s paradox for traffic

networks,” in 2001 IEEE Intelligent Transportation Systems. Proceedings (Cat.
No.01TH8585), Oakland, CA, USA, 25-29 August, 2001, pp. 836–841.

[205] R. A. Abrams and J. N. Hagstrom, Improving Traffic Flows at No Cost. Boston,

MA: Springer US, 2006, pp. 1–22.

[206] T. Roughgarden, “On the severity of Braess’s paradox: designing networks for
selfish users is hard,” Journal of Computer and System Sciences, vol. 72, no. 5,
pp. 922–953, 2006.

[207] K.-Q. Cai, J. Zhang, M.-M. Xiao, K. Tang, and W.-B. Du, “Simultaneous opti-
mization of airspace congestion and flight delay in air traffic network flow man-
agement,” IEEE Transactions on Intelligent Transportation Systems, vol. 18,
no. 11, pp. 3072–3082, 2017.

167
[208] R. Steinberg and W. I. Zangwill, “The prevalence of Braess’ paradox,”
Transportation Science, vol. 17, no. 3, pp. 301–318, 1983.

[209] T. Roughgarden, Selfish routing and the price of anarchy. MIT press
Cambridge, 2005. [Online]. Available: http://www.cs.yale.edu/homes/lans/
readings/routing/roughgarden-jobtalk2-2001.pdf

[210] M.-C. Hwang and H.-J. Cho, “The classical Braess paradox problem revisited: a
generalized inverse method on non-unique path flow cases,” Networks and Spatial
Economics, vol. 16, no. 2, pp. 605–622, 2016.

[211] P. N. Brown and J. R. Marden, “Studies on robust social influence mechanisms:

Incentives for efficient network routing in uncertain settings,” IEEE Control
Systems Magazine, vol. 37, no. 1, pp. 98–115, 2017.

[212] A. Nagurney, “The negation of the Braess paradox as demand increases: The wis-
dom of crowds in transportation networks,” EPL (Europhysics Letters), vol. 91,
no. 4, pp. 48 002–48 006, 2010.

[213] E. Sober, “The principle of parsimony,” British Journal for the Philosophy of
Science, vol. 32, no. 2, pp. 145–156, 1981.

[214] I. Annex, “2–rules of the air,” International Civil Aviation Organization, 2005.

[215] G. Bakker and H. A. Blom, “Air traffic collision risk modelling,” in Proceedings
of 32nd IEEE Conference on Decision and Control, San Antonio, TX, USA.
IEEE, 15-17 December, 1993, pp. 1464–1469.

[216] ICAO, “Manual on implementation of a 300 m (1000 ft) vertical separation min-
imum between FL 290 and FL 410 inclusive,” ICAO-Doc-9574, Montreal, QC,
Canada, 2001.

168
[217] G. Moek, “Post implementation collision risk assessment for RVSM in the Africa
Indian ocean region,” Nationaal Lucht- en Ruimtevaartlaboratorium, National
Aerospace Laboratory NLR, Tech. Rep., 2005.

[218] F. Symon, S. Alam, M. M. Hossain, and H. Blom, “Airspace network character-

ization for effect of intermediate waypoints on collision risk assessment,” in 6th
International Conference on Research in Air Transportation, Istanbul, Turkey,
26-30 May, 2014, pp. 1–8.

[219] J. K. Kuchar and L. C. Yang, “A review of conflict detection and resolution

modeling methods,” IEEE Transactions on Intelligent Transportation Systems,
vol. 1, no. 4, pp. 179–189, 2000.

[220] ICAO, “Unified framework for collision risk modelling in support of the manual
on airspace planning methodology with further applications, circ 319-an/181
ed,” International Civil Avaition Organization, Montreal, Canada, 2008.

[221] B. Mirkin, Clustering: a data recovery approach. Chapman and Hall/CRC,

2016.

[222] M. C. R. Murça and R. J. Hansman, “Identification, characterization, and pre-

diction of traffic flow patterns in multi-airport systems,” IEEE Transactions on
Intelligent Transportation Systems, vol. 20, no. 5, pp. 1683–1696, 2018.

[223] M. Prandini, L. Piroddi, S. Puechmorel, and S. L. Brázdilová, “Toward air traffic

complexity assessment in new generation air traffic management systems,” IEEE
Transactions on Intelligent Transportation Systems, vol. 12, no. 3, pp. 809–818,
2011.

[224] M. Gariel, A. N. Srivastava, and E. Feron, “Trajectory clustering and an appli-

cation to airspace monitoring,” IEEE Transactions on Intelligent Transportation
Systems, vol. 12, no. 4, pp. 1511–1524, 2011.

169
[225] M. Enriquez, “Identifying temporally persistent flows in the terminal airspace
via spectral clustering,” in Tenth USA/Europe Air Traffic Management Research
and Development Seminar (ATM2013), Chicago, IL, USA, 10-13 June, 2013.

[226] F. Rehm, “Clustering of flight tracks,” in AIAA Infotech@ Aerospace 2010, 2010,
p. 3412.

[227] J. S. DeArmon, C. P. Taylor, T. Masek, and C. R. Wanke, “Air route clus-

tering for a queuing network model of the national airspace system,” in 14th
AIAA Aviation Technology, Integration, and Operations Conference, Atlanta,
GA, USA, 16-20 June, 2014, p. 2161.

[228] D. Arthur and S. Vassilvitskii, “k-means++: The advantages of careful seeding,”

Stanford, Tech. Rep., 2006.

[229] J. MacQueen, “Classification and analysis of multivariate observations,” in

Proceedings of the 5th Berkley Symposium on Mathematical Statistics and
Probability, Oakland, CA, USA, 1967, pp. 281–297.

[230] M. Halkidi, Y. Batistakis, and M. Vazirgiannis, “On clustering validation tech-

niques,” Journal of Intelligent Information Systems, vol. 17, no. 2-3, pp. 107–145,
2001.

[231] K. W. Church, “Word2Vec,” Natural Language Engineering, vol. 23, no. 1, pp.
155–162, 2017.

[232] S. Samolej, G. Dec, D. Rzonca, A. Majka, and T. Rogalski, “Regular graph-

based free route flight planning approach,” Aircraft Engineering and Aerospace
Technology, vol. 93, no. 9, pp. 1488–1501, 2021.

[233] “DSNA France introduces free route airspace above 6000m,”

https://centreforaviation.com/news/dsna-france-introduces-free-route-airspace-
above-6000m-1106609, accessed on 3 February, 2022.

170
[234] B. S. Ali and N. A. Taib, “A study on geometric and barometric altitude data
in automatic dependent surveillance broadcast (ADS-B) messages,” Journal of
Navigation, vol. 72, no. 5, p. 1140–1158, 2019.

[235] C. Ma, Q. Cai, S. Alam, B. Sridhar, and V. N. Duong, “Airway network man-
agement using Braess’s paradox,” Transportation Research Part C: Emerging
Technologies, vol. 105, pp. 565–579, 2019.

[236] S. C. Mohleji and G. Wang, “Modeling ADS-B position and velocity errors for
airborne merging and spacing in interval management application,” The MITRE
Corporation, Tech. Rep., 2010.

[237] V. Amanipour and S. Ghaemmaghami, “Median filtering forensics in compressed

video,” IEEE Signal Processing Letters, vol. 26, no. 2, pp. 287–291, 2018.

[238] D. Martinez, S. Belkoura, S. Cristobal, F. Herrema, and P. Wachter, “A

boosted tree framework for runway occupancy and exit prediction,” in 8th Sesar
Innovation Days, Salzburg, Austria, 3-7 December, 2018.

[239] J. García-Heras Carretero, “Analysis of the geometric altimetry to support air-

craft optimal trajectories within the future 4D trajectory management,” Ph.D.
dissertation, Aeronauticos, 2014.

[240] K. M. Johnson, R. M. McKay, J. Etourneau, F. J. Jiménez-Espejo, A. Albot,

C. R. Riesselman, N. A. Bertler, H. J. Horgan, X. Crosta, J. Bendle et al.,
“Sensitivity of holocene east Antarctic productivity to subdecadal variability set
by sea ice,” Nature Geoscience, vol. 14, no. 10, pp. 762–768, 2021.

[241] J. Zhu, Q. Wang, C. Tao, H. Deng, L. Zhao, and H. Li, “Ast-gcn: Attribute-
augmented spatiotemporal graph convolutional network for traffic forecasting,”
IEEE Access, vol. 9, pp. 35 973–35 983, 2021.

171
[242] S. Guo, Y. Lin, N. Feng, C. Song, and H. Wan, “Attention based spatial-temporal
graph convolutional networks for traffic flow forecasting,” in Proceedings of the
AAAI Conference on Artificial Intelligence, Hawaii, USA, vol. 33, no. 01, 27
January-1 February, 2019, pp. 922–929.

[243] J. Xie, Q. Miao, R. Liu, W. Xin, L. Tang, S. Zhong, and X. Gao, “Attention
adjacency matrix based graph convolutional networks for skeleton-based action
recognition,” Neurocomputing, vol. 440, pp. 230–239, 2021.

[244] L. Ma and Y. Zhang, “Using Word2Vec to process big text data,” in 2015 IEEE
International Conference on Big Data (Big Data), Santa Clara, CA, USA. IEEE,
29 October-1 November, 2015, pp. 2895–2897.

172