Adaptative air traffic network: statistical regularities in traffic management

Transcription

1 ELSA Empirically grounded agent based models for the future ATM scenario Adaptative air traffic network: statistical regularities in traffic management 11 th ATM USA/Europe ATM R&D seminar, June 2015 G. Gurtner, C. Bongiorno, S. Miccichè, F. Lillo & R. Mantegna, S. Pozzi

2 Presentation of ELSA Part of the SESAR WP-E (long-term research) EmpiricaLly grounded agent based models for the future ATM scenario Deep Blue Gerald Gurtner Simone Pozzi Università di Palermo Christian Bongiorno Salvatore Miccichè Rosario Mantegna Scuola Normale Superiore di Pisa Fabrizio Lillo G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

3 Introduction to ELSA SESAR and NextGen try to shift from a top-down organization to a bottom-up approach i.e. flight structure emerges out of the single trajectories, The way the airspace will be used will deeply change. This changes require the understanding of how all the elements will interact together. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

4 Introduction to ELSA SESAR and NextGen try to shift from a top-down organization to a bottom-up approach i.e. flight structure emerges out of the single trajectories, The way the airspace will be used will deeply change. This changes require the understanding of how all the elements will interact together. System Large system: up to 30,000 flights a day, 500 airports, 900 sectors, 6,000 navigation points, different networks, different scales, everything is intricate, Big number of agents competing for different goals, Various types of interactions between them. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

5 First goal of ELSA Presentation of ELSA In the current scenario, finding the hidden statistical patterns which will/should be reproduced or taken into account to go to the new scenario. In particular: Monitor differences between the planned use of the network and its actual use, as determined by either controllers or pilots, Employ methods that can capture system-level regularities, but that can provide entry points to drill down and identify the local causes creating such patterns. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

6 First goal of ELSA Presentation of ELSA In the current scenario, finding the hidden statistical patterns which will/should be reproduced or taken into account to go to the new scenario. In particular: Monitor differences between the planned use of the network and its actual use, as determined by either controllers or pilots, Employ methods that can capture system-level regularities, but that can provide entry points to drill down and identify the local causes creating such patterns. Second goal of ELSA Build an Agent-Based Model (ABM) integrating many actors at different levels in order to test new scenarios in Air Traffic Management (ATM), taking into account the previous statistical facts. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

7 First goal of ELSA Presentation of ELSA In the current scenario, finding the hidden statistical patterns which will/should be reproduced or taken into account to go to the new scenario. In particular: Monitor differences between the planned use of the network and its actual use, as determined by either controllers or pilots, Employ methods that can capture system-level regularities, but that can provide entry points to drill down and identify the local causes creating such patterns. Second goal of ELSA Build an Agent-Based Model (ABM) integrating many actors at different levels in order to test new scenarios in Air Traffic Management (ATM), taking into account the previous statistical facts. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

8 Table of Contents 1 Introduction 2 Comparing planned and actual trajectories The navpoint network Comparison of metrics among countries Advanced metrics and correlations Overexpression of deviations 3 Discussion Other null models Third dimension Other metrics Spatial extension of the correlations 4 Conclusions G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

9 Table of Contents 1 Introduction 2 Comparing planned and actual trajectories The navpoint network Comparison of metrics among countries Advanced metrics and correlations Overexpression of deviations 3 Discussion Other null models Third dimension Other metrics Spatial extension of the correlations 4 Conclusions G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

10 Organization of the airspace The airspace in organized in a hierarchical way by different pieces. Main important pieces are: National Airspace, Sectors, 3d volumes controlled directly by two controllers, Navigation points (navpoints), defined via a couple of latitude/longitude values (no height!) Flight plan submission A flight plan is a sequence of navpoints with cruise altitude and hours of departure and arrival. A first flight plan is filed around two hours before the departure. The flight plan is updated just before departure, recorded in m1 files. The flight is tracked during its course, the trajectory is recorded in m3 files. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

11 Network theory approach Central Idea Study the relationships between different entities to understand how they shape a system. In particular, air traffic-based networks are built this way: Consider a type of entity used by the flight during its trajectory (airport, sector, navpoint); define all instances of this type as nodes; put a link between two nodes if a flight is going from one entity to the other in the considered time window, (optional) put a weight on each link proportional to the number of flights crossing the link in the considered time window. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

12 Network theory approach Central Idea Study the relationships between different entities to understand how they shape a system. In particular, air traffic-based networks are built this way: Consider a type of entity used by the flight during its trajectory (airport, sector, navpoint); define all instances of this type as nodes; put a link between two nodes if a flight is going from one entity to the other in the considered time window, (optional) put a weight on each link proportional to the number of flights crossing the link in the considered time window. The network is a creation of the traffic, which was based on a prior organization of the airspace, which in turn is based on past traffic. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

13 Navpoint network We build networks of navpoints, allowing to study the finest resolution of the trajectories. With planned and actual trajectories, one could build two types of networks: the planned network, heavily dependant on how the airspace has been previously organized, the realized network, including the actions of the controllers. The structure of the networks and the differences between them are the consequences of the choices from regulators and controllers. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

14 Navpoint network We build networks of navpoints, allowing to study the finest resolution of the trajectories. With planned and actual trajectories, one could build two types of networks: the planned network, heavily dependant on how the airspace has been previously organized, the realized network, including the actions of the controllers. The structure of the networks and the differences between them are the consequences of the choices from regulators and controllers. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

15 Navpoint network We build networks of navpoints, allowing to study the finest resolution of the trajectories. With planned and actual trajectories, one could build two types of networks: the planned network, heavily dependant on how the airspace has been previously organized, the realized network, including the actions of the controllers. The structure of the networks and the differences between them are the consequences of the choices from regulators and controllers. Here we use one month of data in different national airspaces to build the networks. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

16 Metrics Traffic-related metrics Number of flights in a given area, en-route delay, length of trajectory... Standard network metrics Degree: number of links of a node. It can be viewed as a very local complexity of the traffic for controllers. Strength: sum of the weights of the links pointing to a node. This is proportional to the traffic passing through the navpoint. Delta network metrics Measuring a difference between planned and real trajectories. Fork: fraction of flights rerouted at a given point,... G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

17 Navpoint network French airspace, one month of data based on the en-route part of planned trajectories. Remark: this is a two dimensional network, all flights levels are projected. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

18 Network metrics Country Nodes Deg. Max Deg. Str. Max Str. Eff. Italy (LI) France (LF) UK (EG) Spain (LE) Germany (ED) Belgium (EB) Different countries have different strategies/geographical constraints. Some concentrate the traffic in a few nodes (e.g. Belgium or Germany) whereas others spread it on a higher number of nodes hence decreasing the local traffic (strength) (e.g. UK) Some keep the local complexity (the degree) at a low level (Belgium, France), whereas others allow high local complexity (U.K.). G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

19 Non-conventional and traffic metrics Country l r l p /l p En-route delay En-route delay fork LI LF EG LE ED EB Different countries have different strategies during the tactical phase. Some keep the horizontal trajectories very stable, hence the punctuality is high (France) whereas other countries disturb the trajectories (Spain), Some countries have less deviations per navpoints but longer deviations (Spain) whereas other have higher rate of deviations (France). G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

20 Changing network with the traffic Hourly-averaged metrics for Germany, days and nights over one month Structure of the network changes with traffic. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

21 The Fork metric G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

22 The Fork metric G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

23 The Fork metric G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

24 The Fork metric G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

25 The Fork metric G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

26 The Fork metric The value of this non-standard metric displays a daily patterns correlation between Fork and the traffic. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

27 Reaction of controllers to the traffic Hourly-averaged metrics for France, days and nights over one month Negative correlation between the traffic and the number of reroutings stabilization of the horizontal trajectories when the traffic is high. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

28 Reaction of controllers in different countries Country Pearson corr. coeff. LI LF EG LE ED EB Different magnitude in different countries. Two distinct operational situations lead to the same apparent outcome: Rerouting coming from conflict, Directs asked by pilots to controllers. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

29 Reroutings during nights and days Distributions of fork during nights and days are distinct. During the night, controllers are using less nodes, with a high overall value of deviations, including the maximum, but no special nodes. On the contrary, during the day, controllers use more navpoints and spread the deviations on more nodes, but they keep a few nodes which bear many deviations with respect to their traffic. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

30 What is a statistical null model? Imagine you are looking at people in the street and you count blond people. You know that in your country there are 20% of blond people in the population. You see 10 persons every hour. 8am: over 10 persons, 2 were blond: things are pretty normal. 9am: over 10 persons, 3 were blond: things are still normal, it is just a fluctuation. 10am: over 10 persons, 1 was blond: things are still normal, it is just a fluctuation. 11am: over 10 persons, 5 were blond: this is strange, but not impossible. 12am: over 10 persons, 2 were blond: things are pretty normal... 5pm: over 10 persons, 10 were blond: this is not normal, this is not happening by chance. There MUST be another reason: you reject the fact (the null hypothesis) that this is happening by chance. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

31 Statistical null model Need of a statistical model: The local value of the fork can be the result of fluctuations or consistent use (or absent of use) by the controllers for reroutings. We use a statistical model in order to test if some navpoints are likely to be the sources of reroutings not by mere chance, but because another process is at play (most likely, the controllers). We make the hypothesis that each flight passing through a navpoint has the same probability of being rerouted. Hence, the probability of the value of the fork (fraction of rerouted flights passing through the node) is computable analytically. Probability to have the values computed with the data can be assessed and those sufficiently unlikely (under 1%) are called overexpressed. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

32 Daily evolution of overexpressed nodes Non-zero number of overexpressed navpoint: non random use of the navpoints. Daily pattern, more overexpressed navpoints during the day: possible stronger procedures during the day. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

33 Stability of the special points Some nodes are consistently used as source of rerouting by the controllers. Different nodes seem to be used during the night and during the day. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

34 Map of most stable points G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

35 Table of Contents 1 Introduction 2 Comparing planned and actual trajectories The navpoint network Comparison of metrics among countries Advanced metrics and correlations Overexpression of deviations 3 Discussion Other null models Third dimension Other metrics Spatial extension of the correlations 4 Conclusions G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

36 The ELSA Air Traffic Simulator ELSA also developed a simulator modelling the Strategic and Tactical Phases. Different modules Airspace building: the user can build its own airspace: sector, navpoints, etc. Strategic phase: the air companies submit different flight plan to the network manager, which can reject them based on the occupation of the airspace. Tactical Phase: a (super)-controller is solving the conflicts with different actions: reroutings, directs, changes of altitudes, also checking for sector capacity constraints. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

37 The ELSA Air Traffic Simulator ELSA also developed a simulator modelling the Strategic and Tactical Phases. Different modules Airspace building: the user can build its own airspace: sector, navpoints, etc. Strategic phase: the air companies submit different flight plan to the network manager, which can reject them based on the occupation of the airspace. Tactical Phase: a (super)-controller is solving the conflicts with different actions: reroutings, directs, changes of altitudes, also checking for sector capacity constraints. The result is a realistic simulation of the airspace where the user can change the structure of the airspace and the behaviors of the agents to test different scenarios. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

38 Using the Simulator for null models The ABM will be used to produce alternate null models for statistical tests and refine the analysis on the special points. In particular: By removing some constraints on the system: capacity constraint, conflict resolution, etc. By changing the behaviors of actors (directs, etc), By changing the flows between entries/exits. This will allow to make the difference between what is due to the network, what is due to the flows and what is due to the controllers themselves. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

39 Using the Simulator for null models The ABM will be used to produce alternate null models for statistical tests and refine the analysis on the special points. In particular: By removing some constraints on the system: capacity constraint, conflict resolution, etc. By changing the behaviors of actors (directs, etc), By changing the flows between entries/exits. This will allow to make the difference between what is due to the network, what is due to the flows and what is due to the controllers themselves. Advertisement The Simulator is now is pre-release and can be accessed on demand. It will be released as open source at the end of July and can be used as scenario generator in other studies. It is modular, documented with some tutorials and is designed to be used and enhanced by the academic community. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

40 Deviations in three dimensions In theory, the same kind of work can be done vertically: compute the unexpected changes of altitudes with respect to the flight plan. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

41 Deviations in three dimensions In theory, the same kind of work can be done vertically: compute the unexpected changes of altitudes with respect to the flight plan. Problem: the vertical trajectory is not reliable because it has been reconstucted by the CFMU. the data is not reliable enough. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

42 Deviations in three dimensions In theory, the same kind of work can be done vertically: compute the unexpected changes of altitudes with respect to the flight plan. Problem: the vertical trajectory is not reliable because it has been reconstucted by the CFMU. the data is not reliable enough. Possible solution: infer normal trajectories for scheduled flights from different days/weeks/months and then compute statistical outliers to see deviations. this will only work if the trajectories are stable enough in the vertical dimension. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

43 What about other metrics? Do we have the right metrics? Frac: fraction of flights which should have passed by the point and have not. Fork: fraction of flights for which a deviation begins after this point; this point is the last common point in M1 and M3 for this piece of trajectory. Antifork: contrary of the previous one, points which ends a deviation. First common point between M1 and M3 after a deviation. AfterFork: fraction of flights which had a fork at the previous point. Alt: absolute difference of altitude at this point between the planned and actual one. Dev: amount of horizontal area generated by the point, divided by the distance to the next point. Vert dev: amount of vertical area generated by the point, divided by the distance to the next point. Delay: amount of en-route delay generated by the point. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

44 Natural spatial extension of control? G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

45 Conclusions Different countries display different types of network which reflect the structure of the airspace as well as the different traffic conditions, As a consequence of these differences and of the different cultures among controllers, the deviations from the planned flight plans are also different in different countries. The controllers react to increased traffic by stabilizing the horizontal trajectories. Different countries have different policies in this regard too. Overexpression of some advance metrics among points allows to spot them geographically and to study them more in-depth via experts feedback. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

46 Changing network with the traffic random effect? not a random effect. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

47 The ELSA Air Traffic Simulator G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34

48 Network theory approach This approach can be applied at several levels for the air traffic. G. Gurtner et al. (ELSA) Adaptative air traffic network ATM seminar, June / 34