The National Airspace System as a Cyber-Physical SystemHamsa Balakrishnan
Massachusetts Institute of Technology
Joint work with: Diana Michalek, Harshad Khadilkar, Hanbong Lee,
Ioannis Simaiakis and Varun Ramanujam
NSF−UC Berkeley−MIT ActionWebs Summer Meeting
July 23, 2010
Research objectives
Increase efficiency of operations Multi-objective control techniques for balancing tradeoffs Decrease fuel burn and environmental impact Increase operational robustness in the presence of
weather More flexible and dynamic trajectories More decentralized decision-making
The use of onboard sensing in ActionWebs, combined with the development of hybrid systems models of aircraft trajectories can help us achieve these objectives
Motivation In 2007, domestic air traffic delays [JEC, US Senate]
Had a $41 billion impact on the US economy Cost airlines an additional $1.6 billion in fuel costs Consumed an additional 740 million gallons of jet fuel Released an additional 7.1 million tons of CO2 into the atmosphere
In 2007, aircraft in the U.S. spent over 63 million minutes taxiing in to their gates, and over 150 million minutes taxiing out for departure [FAA]
Taxiing aircraft burn fuel, and contribute to surface emissions of CO2, hydrocarbons, NOx, SOx and particulate matter An estimated 6 million tons of CO2, 45,000 tons of CO, 8,000 tons of NOx
and 4,000 tons of hydrocarbons are emitted annually by aircraft taxiing out
In Europe, aircraft are estimated to spend 10-30% of their time taxiing A short/medium range A320 expends as much as 5-10% of its fuel on
the ground [Airbus]
Boston Logan airport, 1830hrs, 01/06/2010
Efficient and equitable arrival/departure runway scheduling
Given a set of flights with estimated arrival times at the airport, the aircraft need to be sequenced into the landing (takeoff) order, and the landing (takeoff) times need to be determined Minimum separation between operations for wake avoidance
(wt. class dependent) (Safety) Currently FCFS; resequencing could
increase throughput (Efficiency)
Separation H L/B757 S
H 96 157 196
L/757 60 69 131
S 60 69 82
Trailing a/c
Lead
ing
a/c
Separation in secondsH SH S196 s 60 s 196 s
452 s
S H60 s 96 sS 60 s H 216 s
*[Dear 1976, Psaraftis 1980, Garcia 1990, Volckers 1990, Neuman and Erzberger 1991, Venkatakrishnan, Barnett and Odoni 1993, Trivizas 1998, Beasley 2000, Anagnostakis 2001, Carr 2004]
Fairness: Constrained Position Shifting*
FCFS… …
… …k = 2
6 7 8 9 10
Runway scheduling under Constrained Position Shifting (CPS)
[Balakrishnan and Chandran AIAA GNC 2006; Operations Research (2010, to appear)]
Key result: We have proved that this is not true Basic idea: We propose a way to represent solution space as a network
whose size is linear in the number of aircraft
[USA/Europe ATM R&D Seminar 2007; Amer. Control Conf. 2007, Amer. Control Conf. 2008, Proceedings of the IEEE (2008)]
Well-studied problem, using various (mostly heuristic) approaches Challenge: Scheduling under CPS has been conjectured to have
exponential computational complexity in the number of aircraft [Carr 2004]
Various interesting extensions can be solved in (pseudo-)polynomial time as shortest-path problems on variations of this network Multiple objectives: Average delay, max. delay, sum of delay costs, fuel costs,
operating costs, robustness to uncertainties, etc. Evaluation of tradeoffs between multiple objectives
CPS network
Size of network is linear in the number of aircraft and exponential in k (≤3) Easy to incorporate arrival time windows and precedence constraints Precedence constraints make problem easier to solve
n = 6, k = 1
ts
3-4-53-4-64-6-54-5-65-4-63-5-63-6-5
1-2-31-2-41-3-21-3-42-1-32-1-42-3-4
1-21-32-12-3
12
2-3-42-3-52-3-62-4-52-4-62-5-42-5-63-4-53-4-6
3-5-64-3-54-3-64-5-6
3-5-4
1-2-31-2-41-2-51-3-41-3-51-4-31-4-52-3-42-3-5
2-4-53-2-43-2-53-4-5
2-4-3
Stage 1 Stage 2 Stage 3 Stage 4 Stage 5 Stage 6
Length = (2k+1)Length = min(Stage#,2k+1)
# of arcs
[Balakrishnan and Chandran, AIAA GNC 2006; Operations Research (2010, to appear)]
Tradeoffs between throughput and delay costs
Randomly generated sequence, 40% Heavy, 40% Large, 20% SmallDelay costs of Heavy and Large are 9x (Delay cost of Small aircraft)
This gain in throughput comes at a very high cost
[Lee and Balakrishnan, Proceedings of the IEEE, 2008]
throughput gain
increase in cost
Is speeding up necessarily a good idea?
Accelerating from the nominal speed requires more fuel
However, speeding up the first aircraft in a sequence may decrease the delay incurred by subsequent aircraft Concept of Time advance
[Neuman and Erzberger 1991, Lee and Balakrishnan ACC 2008]
0
200
400
600
800
1000
1200
0 1 2 3 4 5Time Advance (min)
Extra fuel cost rel. to
ETAs ($)
FCFS
1-CPS
2-CPS
3-CPS
Evaluating the benefits of Time Advance
Total estimated fuel cost vs. allowable time advance
[Lee and Balakrishnan, Proceedings of the IEEE, 2008]
Modeling the taxi-out process
Queuing network model of the taxi-out process
Model inputs (ASPM database) Pushback schedules Gate assignments Runway configurations Weather conditions
Desired outputs Taxi-out time for each flight Level of congestion on airport surface Loading on departure queues
Estimating taxi-out time Taxi-out time expressed as
t = tunimpeded + ttaxiway + tdep.queue
Unimpeded taxi-out time (tunimpeded) dominates when the number of aircraft on the surface is low
Departure queue wait time (tdep.queue) dominates when the number of departures on the surface is high
Taxiway congestion term (ttaxiway) is important under medium traffic conditions
Parameter estimation: Unimpeded taxi-out time Departure throughput saturation Runway service process
[Simaiakis and Balakrishnan, AIAA GNC, 2009]
Improvement through including taxiway congestion term Assume ttaxiway = aR(t) R(t) is the number of aircraft on ramps and taxiways Choose a for best parameter fit
Without taxiway congestion term
With taxiway congestion term
[BOS; 4L,4R|4L,4R,9; VMC]
[Simaiakis and Balakrishnan, AIAA GNC, 2009]
Model validation: departure throughput
Model validated on 2008 data
Without taxiway congestion term
With taxiway congestion term
[BOS; 4L,4R|4L,4R,9; VMC]
[Simaiakis and Balakrishnan, AIAA GNC, 2009]
Taxi times as a function of congestion
High (N ≥ 17)
Medium (9 ≤N ≤16)
Low (N ≤ 8)
[Simaiakis and Balakrishnan, AIAA GNC, 2009]
Prediction of taxi-out timesm
inm
in
[Simaiakis and Balakrishnan, AIAA GNC, 2009]
One potential strategy: “N-Control” Conceptually simple: Limit the buildup of queues on the airport
surface by controlling the pushback times of aircraft If number of aircraft in movement area > Nctrl
Add any departing aircraft that requests clearance to a virtual departure queue, unless there is an aircraft waiting to use the gate, in which case, clear departure for pushback
If number of aircraft in movement area ≤ Nctrl Clear aircraft in virtual departure queue for pushback in FCFS order,
unless there is a flight waiting to use the gate of an aircraft in the virtual departure queue, in which case, clear departure for pushback
Model can be used to evaluate surface traffic management strategies
Pushback requests
Pushback clearances
Departure queue
Departure throughput
(Virtual)pushback queue for N-control
Ramp and taxiway interactions
Runway
Evaluating delay-emissions tradeoffs
Effect of stopping and starting while taxiing
Potential fuel burn impact from stopping on the surface
No significant impact
Effect of stopping and starting while taxiing
Impact depends on pilot actions
No significant change in thrust setting
thrust setting (%)
Using CFDR data to estimate impact of different taxi profiles
ICAO emissions databank assumes that aircraft taxi at a constant thrust setting of 7%
Using CFDR data (from Swiss Air) corresponding to taxi profiles of various aircraft, we Developed a regression model for fuel burn, that considers the
baseline fuel burn and the impact of stop-start events Stop-start impact: Estimate of the form
“The extra fuel burn from a start-stop event is equivalent to x additional minutes of taxi time”
Developed a (linear) regression model between fuel burn and thrust settings
Conducted above analysis for 9 aircraft types
Fuel burn / √Temp = Baseline fuel burn rate*(taxi time) + (Stop-start impact)*(# of stop-start events)
CFDR estimates vs. ICAO fuel burn rates
9.9%8.8%
1.6%
7.4%
4.0%
6.3%
27.4%
4.1%
5.9%
[Khadilkar, Balakrishnan & Reynolds, working paper, 2010]
Labels indicate thrust settings; ICAO assumes 7% constant thrust during taxi.
Results of CFDR analysis
Taxi time is by far the dominating factor in determining taxi fuel burn
Impact of stops and turns appears to be due to the increase in taxi time, rather than due to a significant change in the fuel burn rate The majority of fuel consumed during stops appeared to be due to
the additional time, rather than the application of breakaway power; the exact estimates varied (depending on aircraft type) from 3%-15% of the average fuel burn during a stop corresponding to the breakaway
In the data set considered (mostly from European airports), each stop appeared to add on average 2 min to the taxi time, while each turn added on average 20 seconds
[Khadilkar, Balakrishnan & Reynolds, working paper, 2010]
Comparison of CFDR and surface surveillance data
Hybrid estimation to filter surface tracks Compare time spent in different taxi modes
ASDE-X data is only from BOS; CFDR is from mostly European and a few US airports
Predictive model for routing
Can we build a model of the airport surface that allows us to predict taxi times and recommend taxi routes?
Mean error: -32 sec, stdev: 141 secAt first node in path
Mean error: -4sec, stdev: 121 secAt second node in path
Take-offs from Node 7
Predictive model for routing
Predictive model for routing
Mitigating the impact of weather on air traffic operations
Objectives: More flexible and dynamic trajectories Increase efficiency of operations Increase operational robustness in the presence of weather
Routing flexibility increases capacity E.g., can we find a route within B km of original route that does not
pass through weather
Forecast Actual
B
[Michalek and Balakrishnan, ATM R&D Seminar 2009]
Dynamic fix relocation: 60-min ahead
Pilots typically assumed to avoid Level 3+ weather
Dynamic fix relocation: 60-min ahead
Pilots typically assumed to avoid Level 3+ weather
Dynamic fix relocation: 60-min ahead
Pilots typically assumed to avoid Level 3+ weather
Dynamic fix relocation: 60-min ahead
Pilots typically assumed to avoid Level 3+ weather
Dynamic fix relocation: 60-min ahead
Pilots typically assumed to avoid Level 3+ weather
Dynamic fix relocation: 60-min ahead
Pilots typically assumed to avoid Level 3+ weather
Dynamic fix relocation: 60-min ahead
Pilots typically assumed to avoid Level 3+ weather
Dynamic fix relocation: 60-min ahead
Pilots typically assumed to avoid Level 3+ weather
Resectorization and rerouting/relocating fix
Forecast Observed
WHI
NZCADI
T
LAGR
ANGE BRAVS
SINCA
DOOLY
GEETK
ROME
WHI
NZCADI
T
LAGR
ANGE BRAVS
SINCA
DOOLY
GEETK
ROME
New Sector boundaryOld sector boundaryNew FixOld Fix
Legend
Multi-objective optimization: Maximize probability of fixes being open; minimize movement of sector boundary; minimize deviation of route from procedural route; try to keep all sectors open
Pilots typically assumed to avoid Level 3+ weather
Summary statistics of resectorization (fix/route relocation)
Note that these metrics do not reflect the probability with which fix was predicted to be open.
77%71%85%72%95%88%86%85%
0%0%7%7%4%5%5%3%
[Michalek and Balakrishnan, CDC 2010]
Summary statistics of resectorization (fix/route relocation)
[Michalek and Balakrishnan, CDC 2010]
Pilot behavior
Of course, pilots do occasionally penetrate weather today Can provide additional flexibility Can we identify the factors that determine their behavior?
0 1 2 3 4 5 60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
VIL Level
Prob
abili
ty o
f pilo
t pen
etra
tion
0 1 2 3 4 5 60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
VIL Level
Prob
abili
ty o
f pilo
t pen
etra
tion
“3 Weather days” “Really bad weather day”
Conclusions
Air transportation system presents many important problems that require the development of CPS methodologies
Solutions have the potential to increase system efficiency (reduce delays), robustness and energy efficiency (reduce fuel burn), and decrease environmental impact
Key objectives should include More dynamic, flexible trajectories (and system structure) Multi-objective control and balancing of tradeoffs More decentralized decision-making