MIT
ICAT
The Airside Departure Process:
Observations & Queuing Models
Prof. John-Paul Clarke
International Center for Air Transportation
Massachusetts Institute of Technology
MIT
ICAT
Overview
¾ Observations at BOS
¾ Queuing Model of Taxi Out Process with Aircraft
Passing
¾ Observations at EWR
¾ Queuing Model of Departure Flow with Downstream
Constraints
MIT
ICAT
Interactive Queuing System
-
N
MIT
ICAT
Uncertainty in Pushback Time
"Scheduled Departure Tim e" to "Ready for Pus h or Taxi"
80
Frequency
70
60
50
Mean = 14 min (absolute)
Std. Dev = 17 min 22 sec
40
30
20
10
0
0:00 0:08 0:16 0:24 0:32 0:40 0:48 0:56 1:04 1:12 1:20 1:28 1:36
Tim e (hr:m in)
MIT
ICAT
Passing During Taxi-Out
6
60%
92%
5
40%
4
Avg. Swap Amplitude
50%
30%
20%
3
2
1
10%
0
0%
-8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8
Histogram of swap magnitude
at Logan airport
ATL
DFW
ORD
S FO
BOS
Comparison of 5 U.S. airports
MIT
ICAT
Key Observations
¾ Airport is an interactive queuing system.
¾ Runway is the key flow constraint.
¾ Gates, Ramps and Taxiways are secondary flow
constraints.
¾ Downstream constraints manifest at different points in
the system such as at the runways and at the gates.
¾ High uncertainty in pushback time.
¾ Significant passing during taxi-out.
¾ ATC workload can be a flow constraint.
MIT
ICAT
Overview
¾ Observations at BOS
¾ Queuing Model of Taxi Out Process with Aircraft
Passing
¾ Observations at EWR
¾ Queuing Model of Departure Flow with Downstream
Constraints
MIT
ICAT
Queuing Model of Taxi Out Process
with Aircraft Passing
¾ Motivation:
Ö Current queuing models for taxi out time are based on the
number of aircraft taxiing out when aircraft being considered
is pushed back from the gate.
Ö Field observations and interviews at BOS indicate that there is
significant passing that occurs during taxi out and that the
taxi time is most strongly correlated with configuration, airline
(surrogate for terminal) and queue at runway.
¾ Modeling Approach:
Ö Divide data based on configuration and airline
Ö Quantify passing behavior as P(Q|N)
Ö Use passing behavior to map N to Q and then to T
T(N) = ∑ [T(Q) *P(Q | N)]
Q
Taxi-Out Time T
MIT
ICAT
Taxi-Out Time vs. N Aircraft@Pushback
MIT
ICAT
Passing Behavior
Aircraft that pushed back
before tout and took off after
toff (the NP aircraft that were
passed by the reference
aircraft)
Number of
departure aircraft
(N) present on the
airport surface at
pushback time tout
Queue size (Q)
(Aircraft that
took off
between tout
and toff )
Reference
aircraft
tout
Aircraft that pushed
back before tout and
took off between tout
and toff
Aircraft that pushed
back after tout and took
off between tout and toff
(the NP aircraft that
passed the reference
aircraft)
Pushback time
(ACARS Out
time)
T, taxi-out time
of the reference
aircraft
toff
Takeoff time
(ACARS Off
time)
Taxi-Out Time T
MIT
ICAT
Taxi-Out Time vs. Takeoff Queue Size
Taxi-Out Time vs. Takeoff Queue Size
(given Configuration and Airline)
Taxi-Out Time T
MIT
ICAT
T(Q) for BOS configuration 27/22L-22R/22L and American Airlines
MIT
ICAT
P(Queue Size Q | N Aircraft@Pushback)
1
0.9
0.8
N=1
0.7
N=2
P(Q|N)
P( Q| N )
0.6
N=3
0.5
N=4
N=5
0.4
0.3
0.2
0.1
0
0
1
2
3
4
Q
5
6
7
8
9
10
Q
Distributions P(Q | N) for BOS configuration 4R/4L-9/4R/4L and US airlines
Taxi-Out Time vs. N Aircraft@Pushback
(with Passing Behavior Included)
T(N)
MIT
ICAT
N
T(N) for BOS configuration 4R/4L-9/4R/4L and Continental Airlines
MIT
ICAT
Model Performance
Running Average
Mean absolute difference between
actual and predicted taxi
% predicted within 5 minutes of
actual taxi
Group
Actual - Running Avg
Actual - Queuing Model
Running Avg - Queueing Model
t Value Prob > t
-10.66 <.0001
-0.848 0.3966
-14.21 <.0001
Queuing Model
5.69 minutes
4.56 minutes
53.74%
65.63%
95% Confidence Interval for Difference of Means
Lower Limit
Upper Limit
-1.46
-1.00
-0.30
0.12
-1.30
-1.00
MIT
ICAT
Overview
¾ Observations at BOS
¾ Queuing Model of Taxi Out Process with Aircraft
Passing
¾ Observations at EWR
¾ Queuing Model of Departure Flow with Downstream
Constraints
MIT
MIT
ICAT
ICAT
Cause of Weather Delays: EWR
¾ Sensitivity to weather
11-29
Ö Runway limitations
Ö Limited gate space
¾ Frequency of adverse
weather
¾ The schedule operated at the
airport
¾ New York airspace
congestion
4R-22L Map of
4L-22R
Newark Airport
(EWR)
MIT
ICAT
EWR System Study
¾ Site visit Thursday, June 29, 2000
Ö Observations at NY TRACON, EWR Tower, CO Ramp Tower
Ö “Worst summer ever”
Ö Just before July 4th weekend
Ö Large delays the previous day
Ö Severe Weather Avoidance Program
(SWAP) active, 1pm – 10:30pm EDT
Ö Data collected focuses on departures
MIT
ICAT
June 29, 2000 – National Weather
National Doppler Radar Map
EWR
9:30:00 EDT (13:30:00 Zulu)
Source: Data Transmission Network
MIT
ICAT
EWR Fix-Destination Mapping
GAYEL
COATE
TEB
ELIOT
LGA
PARKE
JFK
LANNA
BIGGY EWR
SEA
ZNY
ZDC
PDX
MSP
DEN
SFO
MCI
LAS
SAN
HNL
PHX
DFW
IAH
WHITE
BOS
MKE
ORD
SLC
LAX
GREKI
MERIT
ELM EWR
PIT
Europe
CLE
BWI
IAD DCA
IND
RIC ORF
GSO
CLT
BNA
CHS
ATL
MSY
MCO
TPA
PBI
MIA
Central and S America
According to preferred routings
MIT
ICAT
Impact on Delays
Running plot of average departure delay of aircraft waiting to depart from EWR
Min Departure Delay
140
120
100
80
60
40
20
0
11:00 12:00
13:00 14:00
15:00 16:00 17:00
18:00 19:00 20:00
21:00 22:00
23:00
0:00
Time (EDT)
Running plot of number of aircraft waiting to depart from EWR
70
No. of Aircraft
60
50
40
30
20
10
0
11:00
12:00 13:00
14:00 15:00
16:00 17:00
18:00 19:00
20:00 21:00
22:00 23:00
0:00
Time (EDT)
Source: Flight Strips from EWR Tower
MIT
ICAT
Impact on Delays
Thursday 6/29/00
Eastern Daylight Time: 7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23
11:30
3:11
4:45
5:22
6:33
8:31
North Restrictions
West Restrictions
7:01
South Restrictions
Restrictions Applied to Fix
Min Departure Delay
6000
Fix closed
Running plot of total departure delay of aircraft waiting to depart EWR
5000
4000
3000
2000
1000
0
9:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 18:00 19:00 20:00 21:00 22:00 23:00 0:00
Time (EDT)
Source: Flight Strips from EWR Tower
MIT
ICAT
Pre-Sequencing
Running plot of average departure delay of aircraft waiting to depart EWR
140
120
Continental domestic flights
100
All other domestic flights
Departure
80
Delay
[Minutes] 60
40
20
0
12:00
13:00
14:00
15:00
16:00
17:00
18:00
19:00
20:00
21:00
22:00
Time (EDT)
High CO arrival delays
Effect of pre-sequencing
Source: Flight Strips from EWR Tower
MIT
ICAT
Key Observations
¾ Non-local restrictions (e.g. arrival restrictions at
destination) and time-windowing (e.g. DSP or EDCT)
has less airport-wide impact than flow-restrictions (e.g.
MIT metering) at local fixes.
Ö Large impact per flight but only a few flights are affected.
¾ Controllers can impose minor MIT or MINIT restrictions
over departure fix (e.g. stretching 3-mile takeoff
minimum to 5 MIT) in TRACON airspace without
impacting throughput.
Ö This flexibility is limited by workload involved in necessary
sequencing, and traffic distribution among fixes.
MIT
ICAT
Key Observations
¾ Restrictions caused by coupling of local weather,
downstream weather and traffic demand.
Ö Some fixes impacted by local weather but unrestricted
because of light traffic.
Ö Some fixes not impacted by local weather but restricted
because of downstream weather.
¾ Opportunities for reroutes not utilized because of interfacility coordination required.
Ö Gap in weather to south not utilized for west reroutes because
of coordination required with Washington Center & Cleveland
Center.
¾ Airline planning (e.g. pre-sequencing) can reduce
delays.
MIT
ICAT
Overview
¾ Observations at BOS
¾ Queuing Model of Taxi Out Process with Aircraft
Passing
¾ Observations at EWR
¾ Queuing Model of Departure Flow with Downstream
Constraints
MIT
ICAT
Queuing Model of Departure Flow with
Downstream Constraints
¾ Motivation:
Ö Current queuing models for airport surface traffic focus on
runways as major bottleneck.
Ö Field observations and interviews at BOS and EWR indicate
that downstream flow constraints as a major source of delay
and that the closure of departure fixes are the most severe
class of flow constraints.
¾ Modeling Approach:
Ö Use observations at BOS and EWR to extend queuing model
to include the effects of fix closures.
Ö Test extended model via Monte Carlo simulation with EWR
operations data collected on 2000-06-29.
MIT
ICAT
Extending the Queuing Model (I)
¾ Many aircraft were pushed
back on-time, only to be held
in parking areas or “penalty
boxes” during fix closures.
Ö Parking areas did not impede
normal surface traffic.
Ö Parking capacity was not a
limiting factor.
¾ Runway 11/29 was used for
less than 4% of flights.
Ö Queuing model only requires
a single runway queue.
MIT
ICAT
Extending the Queuing Model (II)
¾ Right side of figure:
Ö Original queuing model
¾ Left side of figure:
Ö Fix-closure effects
¾ When fix closes, A/C move to
“buffers”, and stop making
progress toward takeoff.
¾ When fix opens, A/C start
making progress again.
¾ A/C which leave runway
queue must re-enter at end of
the queue (i.e. they lose their
spot in line).
MIT
ICAT
Calibration and Testing
¾ Operations data:
Ö Taken from CATER system and TRACON logs.
Ö Data set for 11:00 to 19:00 EDT on July 29, 2000.
¾ Stochastic runway service time:
Ö Estimated from 09:00 to 11:00 period (very light restrictions).
Ö Triangular density (mode of 1 minute, range of 0 to 2 minutes).
¾ Stochastic unimpeded taxi-out times:
Ö Estimated for major passenger carriers from 1998 ASQP data.
Ö Same distribution (9.75±2.75 min.) assumed for other carriers.
¾ Testing Procedure:
Ö Run Monte Carlo simulation of traffic over 11:00 to 19:00.
Ö Compute average results over 40 Monte Carlo runs.
MIT
ICAT
Monte Carlo Results:
Early Period (11:00 to 15:00 EDT)
¾
Simulation accurately tracks observed number of departing A/C.
¾
88% of flights (107 out of 122) have taxi-time error‡ of 10min or less.
Of the 15 outliers, 8 used LANNA fix, with 40MIT from 13:05 to 15:50.
¾
In contrast, original queuing model gives errors of up to 3 hours for
flights hit by fix closures, and wider distribution of errors for other flights.
‡: Def’n of taxi-time “error”: Difference between actual taxi-time and average simulated taxi-time.
MIT
ICAT
Monte Carlo Results:
Late Period (15:00 to 19:00 EDT)
¾
Much larger number of restrictions during this period.
Simulation tracks poorly until 17:00 when it starts to recover.
¾
Flights through ELIOT and PARKE account for most of the extreme
errors. Both fixes closed from 16:00 to 16:30 and again from 18:10 to
19:00, but had only 20MIT during remaining period; hence large delays
must be due to some other cause.
MIT
ICAT
Research Questions
¾ Question: After a fix closure is lifted, what is the expected
response-time for fix throughput to reach capacity? Can this
response-time be minimized by buffering A/C in different
parts of the system?
¾ Question: At many airports, departure fixes are clustered,
e.g. the West fixes at EWR. Are there workable re-routing
procedures which could switch fixes (and thus avoid
closures) in response to local weather?
¾ Question: What is the utility of lead-time information on
upcoming changes to fix restrictions?
¾ Question: What level of MIT/MINIT spacing effectively closes
a fix?
MIT
ICAT
Further Extensions to the Model
¾ Severe spacing restrictions can
combine with high traffic levels,
overloading the flexibility to meet
such restrictions via re-sequencing
before takeoff or vectoring after
takeoff. To correctly model effects
of spacing restrictions, need to
capture sequencing logic of
controllers.
¾ Airspace buffering for departures is
not currently observed. However,
may be worth extending model in
that direction, to evaluate potential
benefits (setting aside workload and
procedural issues for the moment).