6th International Conference in Air Transport 27th-30th May 2014. Istanbul Technical University Providing insight into how to apply Data Science in aviation: Metrics and Representations Samuel Cristóbal
Origin - Destination graph Airports as nodes LEMD Two airports are connected by an edge if there is at least one flight connecting them LEMD EGKK Edges are usually weighted by either distance (geographical or time) or number of flights.
Origin - Destination graph Gives you information about the network topology and structure. Detection of subnetworks of flights structures (e.g. point-to-point or hub-and-spoke) Airports and routes can be classified by similar features (clustering) Connectivity and centrality
Origin - Destination graph The main drawback is that the time information (flight schedules) is neglected completely!! Passenger connections can not be represented properly, neither aircraft rotations.
Airport time-line graph
Airport time-line graph LEMD EGKK Multiple flights between city pairs
Airport time-line graph EGKK EGKK time-line LEMD LEMD time-line
Airport time-line graph Airport nodes EGKK EGKK time-line LEMD LEMD time-line
Airport time-line graph EGKK EGKK time-line LEMD LEMD time-line Ground arcs
Airport time-line graph Closing ground arcs EGKK EGKK time-line LEMD LEMD time-line
Airport time-line graph EGKK ground arc LEMD ground arc LFPG ground arc
Airport time-line graph Path-related Complex Network metrics (e.g. betweenness, closeness) now correctly capture passenger flows in the network. EGKK ground arc LEMD ground arc LFPG ground arc
Airport time-line graph Path-related Complex Network metrics (e.g. betweenness, closeness) now correctly capture passenger flows in the network. The Origin-Destination graph can be easily recovered by projecting over the airport s ground arc. EGKK ground arc LEMD ground arc LFPG ground arc
Airport time-line graph Delays impact on the time-line graph topology!!!
Airport time-line graph The alterations in the Airport time-line graph introduced by delays are invariant under the projections over the ground arcs!!
Airport time-line graph schedule
Airport time-line graph schedule actual operations
Airport time-line graph? schedule There are many possible outcomes!!!
Airport time-line graph Probability Density (but we ll get to that later)
Airport time-line graph Delays have an impact on the graph topology!!! This representation is especially suitable for analysing network dynamics and flows (aircrafts, passengers, crews, etc.) However, by using this representation it would be tedious (although possible) to explore reactionary delays and propagation.
Delay maps Delay maps were developed in the framework of the ongoing FP7 Resilience 2050 project as an alternative metric to measure the resilience* of the ATM system in Europe. *the ability to recover under abnormal conditions.
Delay maps E A B D C Let us consider all flights departing from a given airport A. Each flight has a departure and arrival delay (or zero if on-time).
Delay maps (departure airport) Each dot represents a flight that departed from A and arrived either at B, C, D or E E A B D C (Colour indicates arrival airport)
Delay maps (departure airport) Each dot represents a flight that departed from A and arrived either at B, C, D or E E A B D C e.g. IJ1042 from A to C flew on 12-3-2011 suffered a departure delay 47 and arrival delay 14
Delay maps (departure airport) E A B D C Let us now focus on a single route
Delay maps (single route) Each dot represents a flight that departed from A and arrived at B A B
Delay maps (data cleanse) Remove Each artefacts dot represents to obtain a a less biased flight that data Each departed set dot increasing represents from A fit and a reliability arrived flight (next that any departed step) other aport from A and arrived at B A B e.g. AB1042 on 12-3-2011
Delay Map (data fitting) The dependence between arrival and departure delay is modelled using lineal regression Linear fit A B Goodness of fit is controlled by the Pearson s correlation coefficient, errors by the SSE and RSME values.
Delay Map (delay rate) We define the Delay Rate between a pair of airports as the first coefficient (slope in %) of the linear fit. 0.9694 A -3.1% B This means that, under normal conditions, flights between A and B usually absorb 3.1% of the departure delay.
Delay Map (delay rate) A -3.1% B A positive RDR implies delay amplification whilst a negative value implies delay absorption
Delay Map (delay rate) E -0.6% A +1.1% D -3.1% -1.1% C B The whole network of delay amplification/absorption can be created using the same technique over all routes, but a very important source of delay is missing: turnaround delay.
Delay Map (turnaround delay) D E A C B Let us consider all aircraft rotations at an airport A. Each rotation would have an associated arrival! delay of the previous leg and a departure delay! of the next leg (or zero if on-time).
Delay Map (turnaround delay) Each dot represents an aircraft rotation in A A
Delay Map (turnaround delay) Each dot represents an aircraft rotation in A A e.g. aircraft A-405 on 21-4-2011 arrived with 32 delay and departed with 17 delay
Delay Map (turnaround delay) Similarly to route analysis, after a data cleanse a linear regression is performed to find a linear fit A
Delay Map (turnaround delay) We define the Turnaround Delay Rate as the first coefficient (slope in %) of the linear fit. 0.9113-8.9% A This means that, under normal conditions, up to 8.9% of the arrival delay can be absorbed throughout aircraft rotation
Delay Map (turnaround delay) -8.9% A A positive RTDR implies delay amplification at airport A whilst a negative value implies delay absorption
Delay Map -12% E -8.9% -5% +1% -0.6% D A +1.1% -3% -1.1% C B 6% The process repeats for each airport and it is combined with the route analysis producing a graph picture of (reference) delay rates
Delay Map
Delay Map Aircraft and crew rotations: Recovery/neutral/critical paths Critical paths should be avoided Graph covering by recovery/neutral paths or cycles. Network analysis Performance degradation/recovery (Resilience) ATM stress test design
Reactionary delay propagation trees Reactionary delay propagation tress were developed as part of the SESAR WP-E Long Term Research project Passenger Oriented Enhanced Metrics (POEM) addressing the following questions:! Certain key changes in ATM performance can only be observed through passenger-centric and cost-centric metrics. Current flight-centered metrics are not fully sufficient to capture passenger experience. SESAR Outstanding Project 2014
the POEM s simulator Passenger-Oriented Enhanced Metrics! Simulation engine for European Air Transport, explicitly including: Air traffic between the 199 EU busiest airports, published schedule times for 98% of flights considered major flows in and out of Europe, individual aircraft occupations and configuration individual passenger itineraries;! and stochastically modelled: En-route delays (weather, regulations), Taxi times, Airport turnaround processes.
the POEM s simulator Passenger-Oriented Enhanced Metrics! It also incorporates: Airlines passenger cost UoW s models (hard and soft cost) Non-passenger simplified cost models (crew, fuel, maintenance). Passenger reaccommodation, airport waiting list and overnight. In accordance with: Regulation (EC) No 261/2004 of the European Parliament and of the Council of 11 February 2004. Common rules on compensation and assistance to passengers in the event of denied boarding and of cancellation or long delay of flights,
the POEM s simulator Passenger-Oriented Enhanced Metrics
the POEM s simulator Passenger-Oriented Enhanced Metrics
The POEM s metrics Passenger-Oriented Enhanced Metrics A"total"of"32"core"stochas/c"metrics," including:" " Aircra8"centered"metrics:"" off:bock"delay,"" aircra8"in:block"delay,"" reac/onary"delay,"" primary"delay,"" Boarding"delay,"" at:gate"delay,"" departure"queue,"" taxi:out"devia/on,"" en:route"devia/on,""" holding,"" taxi:in"devia/on,"" Pax"centered"metrics:" load"factors," extra"/me"before"boarding," extra"journey"/me," extra"/me"spent"flying," extra"/me"spent"on"ground," missed"connec/ons" successfully"reaccommodated," pax"overnight," aborted"trips,"" extra"flights"taken," Cost"centered"metrics:" Pax"hard"cost" Pax"so8"cost" Value"of"/me" Non:pax"cost"
Reactionary delay propagation trees Nodes are flights
Reactionary delay propagation trees Nodes are flights Two flights are connected by an edge if the former induces any reactionary delay to the latter. either by aircraft rotation or late gate arrival passengers
Reactionary delay propagation trees Nodes are flights Two flights are connected by an edge if the former induces any reactionary delay to the latter. either by aircraft rotation or late gate arrival passengers Edges are weighted by the amount of delay propagated.
Reactionary delay propagation trees
Reactionary delay propagation trees Many trees apear in actual operations
Reactionary delay propagation trees Probability Density
Reactionary delay propagation trees Number of affected flights (tree sizes)
Reactionary delay propagation trees Time until reactionary delay is absorbed (longest branch)
Reactionary delay propagation trees Reactionary delay Reactionary and arrival delays Arrival delay
Reactionary delay propagation trees Total delay induced Aircraft and passengers Only aircraft rotations Only passenger connections
Reactionary delay propagation trees Delay propagated to other airports
Reactionary delay propagation trees Detect key flights propagating delay! Airport back-propagated delay! Reactionary delay due to passengers/aircraft! Airport classification: sources-sinks, amplifiers-attenuators! Role of different airports on passenger delay vs. flight delay
Thank you for your attention!! For more information please refer to: http://resilience2050.innaxis.org/ (Delay Maps) (POEM s AT model, delay propagation trees) http://complexworld.eu/wiki/poem Samuel Cristóbal sc@innaxis.org