M I T I n t e r n a t i o n a l C e n t e r f o r A i r T r a n s p o r t a t i o n M I T I n t e r n a t i o n a l C e n t e r f o r A i r T r a n s p o r t a t i o n Potential Approaches to Increase Milestone Prediction Accuracy in NAS Movements Georg Theis, Francis Carr, Professor JP Clarke Presented at NEXTOR/CDM Review Meeting, September 25, 2003
Agenda CDM Mission Milestones for NAS Movements Benefits of Improved Prediction Methods Outlook: Using Real-Time Ground Handling Milestones during Turnaround to Improve the Accuracy of Pushback Time Prediction Future Research 2
CDM Mission Collaborative Decision Making provides Airline Operations Centers and the FAA with real-time access to National Airspace System status information including weather, equipment, and delays (FAA website: http://ffp1.faa.gov/tools/tools_cdm.asp) NAS Status Information Dissemination (NASSI) (CDM Summary Report, 01/08/03) GDP Projected Demand and Capacity Departure Delays Planned and Actual Pushback Times Airport Acceptance Rates Airports Configurations Integrated Arrival and Departure Management 3
Milestones for NAS movements where are we today? ground event flight AA 123 ground event taxi-out flight time taxi-in off blocks on blocks takeoff touchdown Published Schedule Agents should predict when they are ready to hand the aircraft over to the next agent ( handoff ): Airline Tower, TRACON Airline Tower, TRACON Estimates Ready for Pushback Ready for Takeoff Ready for Touchdown Ready for onblocks Models e.g. Idris, Clarke et al. Airlines (e.g. Lido, proprietary) How do estimates benefit downstream processes? How can we come up with the missing estimates? 4
Research Questions To what extent does an improved taxi-out time prediction increase the accuracy of touchdown prediction? Outlook: How can pushback prediction be improved based on real-time ground handling data? 5
Available Data Flights LH422 FRABOS for August 2002 Actual taxi-out times Pre-calculated flight times at day of operations and actual flight times ( Pre-calculated flight times are calculated two hours before departure based on planned departure/approach path, winds, etc.) Taxi Out FRABOS LH422 01AUG- 31AUG Precalculated and Actual Flight Times LH 422 9 8 7 6 5 4 3 2 1 0 6 Frequency 0 6 12 18 24 30 More Bin Frequency 8:09 7:55 7:40 7:26 7:12 6:57 6:43 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 hh:mm day of operation, Aug 2002 PCFT Actual Flight Times
Two Steps Part A: Prediction of touchdown time based on the available data Part B: Estimation of average wait times and runway idle time based on touchdown predictions 7
Part A: Methodology Predicting touchdown based on different levels of taxi-out time prediction accuracies Base case: o taxi-out prediction: actual taxi-out time deviation from taxi-out mean o flight time prediction: actual flight time deviation from PCFT Test case: o improved taxi-out prediction: 50% of actual taxi-out time deviation from taxi-out mean o flight time prediction: actual flight time deviation from PCFT Convolve taxi out deviation distribution and flight time deviation distribution Compare standard deviations of touchdown time 8
Part A: Methodology Taxi-out Time Prediction Deviation (Act. Taxi Mean Taxi) 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0-15 -10-5 0 5 10 15 Flight Time Prediction Deviation (Act. FT PCFT) 0.2 0.18 0.16 0.14 9 Base Case + 0.12 0.1 0.08 0.06 0.25 0.04 0.02 0 0.2 0 1 2 3 4 5 6 7 8 9 10 11 12 0.15 0.1 0.05 Test Case 0-15 -10-5 0 5 10 15
Part A: Calculations No significant correlation between taxi-out time prediction or flight time prediction 0:12 0:11 0:10 0:08 0:07 0:05 0:04 0:02 0:01 0:00-10 -5 0 5 10 15 Dev. Actual Taxi Out from Median of Taxi Out 10 Dev. Actual FT from PCFT
Part A: Convolution Results Deviation from touchdown prediction Base Case Test Case 0.08 0.1 0.07 0.09 0.06 0.08 0.05 0.07 0.06 0.04 0.05 0.03 0.04 0.02 0.03 0.01 0-15 -10-5 0 5 10 15 20 25 0.02 0.01 0-15 -10-5 0 5 10 15 20 25 σ= 6.03 min σ= 4.05 min 11
Discussion Part A In this example, an improvement of taxi-out time prediction by 50% leads to an improvement of touchdown prediction by approx. 33% More accurate taxi-out time prediction has a considerable effect on the accuracy of touchdown predictions 12
Part B: How does improved touchdown prediction change wait times and runway idle time? Analyze average wait time and average runway idle time for different touchdown prediction accuracies at constant expected arrival headways Assumptions: Runway only used for landings Service Time 1.5 min, i.e. capacity of 40 aircraft/hour Expected headway of landing aircraft 1.5 min. with different standard deviations for base and test case 16 hour operation First come first serve Simulation: 25 Runs 13
Part B: Methodology E(Touchdown)= [Median Taxi-Out + Mean (Deviation Median vs. Actual Taxi Out)] + [PCFT + Mean (Deviation PCFT vs. Actual Flight Time) -> set E(Touchdown Deviation)=0 Standard Deviation as calculated in Part A Wait time per aircraft: Departure Time Service Time Arrival Time Runway Idle Time: Time when no aircraft is serviced (because no aircraft is available in queue) 14
Pushback Origin A Pushback Origin B Pushback Origin D Aircraft have an expected arrival every 1.5 minutes Base Case Destination Airport Touchdown Test Case Destination Airport Pushback Origin C Touchdown Pushback Touchdown Origin A Touchdown Pushback Origin B Touchdown Pushback Origin C Touchdown Pushback Origin D Touchdown 0 1.5 3 4.5 t/min Touchdown Test Case: µ=0, σ=4.05 0 1.5 3 4.5 t/min Base Case: µ=0, σ=6.03 15
Results Wait times can be reduced by approximately one minute by using improved taxi-out time prediction Idle times can be reduced by 34% 6.03 (base case) 4.05 (Test Case) Headway Idle Time Aver. Wait Idle Time Aver. Wait 1.5 1.45% 5.72 min 0.95% 4.68 min 16
Changing the headway gives consistent results 6.03 (base case) 4.05 (Test Case) 3.13 (perfect taxi) Headway Idle Time Wait Time Idle Time Wait Time Idle Time Wait Time 1.4 0.87% 33.68 0.48% 33.66499 0.51% 32.88 1.45 0.93% 18.44 0.46% 18.33 0.48% 17.65 1.5 1.45% 5.72 0.95% 4.68 0.87% 4.16 1.55 4.48% 3.9 4.10% 2.74 3.86% 2.44 1.6 7.29% 2.7 6.80% 2.26 6.83% 1.94 17
Results Managed Arrival Reservoir 100 10 1 E (Headway)= service time=1.5 min Aver. wait time per a/c in min. σ=6.03 σ=4.05 σ=3.13 0.00% 2.00% 4.00% 6.00% 8.00% 10.00% Runway Idle Time Share Test Case Base Case Perfect Taxi Out Pred. 18
Implications More accurate taxi-out time can improve the prediction accuracy of touchdown Wait times in holding patterns can be reduced and runway utilization rates improved Reduction in holding patterns: At $3000/block hour, reduction of $ Mio 11.7 p.a. o 40aircraft/hour*1min*16 hours=10.7 delay hours/day o 10.7 delay hours/day*365 days*$3000=$ Mio 11.7 p.a. Passenger delay costs: At a value of 30$ per hour and 100 pax/aircraft, reduction of another $11.7 Mio p.a. o 10.7 delay hours/day*30*100=$32100/day o $32100*365 days= $11.7 Mio p.a. 19
Limitations/Next Steps Limitations Small sample Assumption was made that pushback at origin can be moved arbitrarily to assure an arrival at a specific expected time Assumption that pushback time setting has no effect on taxi/flight time or en-route constraints Headway is fixed Possible Next Steps Verify results with larger sample and more simulation runs Allow for dual use runway and different aircraft sizes Allow changes in headway over time 20
Outlook: How can ground handling data be used to improve the accuracy of pushback time prediction? Improved Prediction only helps after Pushback Controllers need to know how many aircraft are about to push How can we already predict the downstream touchdown during turnaround? -> We need an accurate pushback time prediction: Ongoing Research of Francis Carr/Georg Theis/Professors Clarke, Feron Pushback Prediction 21
Ground Handling Data to our knowledge currently only available to Lufthansa Airlines ALLEGRO measures crew on position, beginning and end of all main turnaround processes Deboarding Passenger Buses Core processes Standard processes Not Ready Message Exceptional processes Crew Processes Gate Processes Catering Cleaning Fueling Boarding Passenger Buses Deicing Un-/Loading Aircraft Positioning Aircraft Acceptance Towing Pushback For all processes, beginning and end are reported real-time For every involved vehicle, the system gets the messages at aircraft, begin service, end service 22
Used Data ALLEGRO measures crew on position, beginning and end of all main turnaround processes Deboarding Passenger Buses Core processes Standard processes Not Ready Message Exceptional processes Crew Processes Gate Processes Catering Cleaning Fueling Boarding Passenger Buses Deicing Un-/Loading Aircraft Positioning Aircraft Acceptance Towing Pushback Source (translated): Theis, G. (2002) Telematik Anwendungen im Luftverkehr, Internationales Verkehrswesen, 54(5), 225-228. See also: http://www.fraport.com/online/general/en/download/presentation_220802.pdf, 12-13 23
Methodology Filtered out exogenous influences GDP Technical Failure Weather With these causes being eliminated, ready for pushback = actual pushback from a ground handling perspective Different methods On block based : Sched.TDep= max (Sched.TDep, (on block + Minimum Ground Time)) Age based: use elapsed ground time to improve updates Status based: update pushback prediction after end of major processes 24
Preliminary Results Status based predictions improve the quality of pushback estimates However, uncertainty in airline turn process has significant lower bound Subsequent processes catch up for late previous processes, i.e. basing the prediction on planned process times reduces the estimate s quality Decision Support Tools for ground-air handoffs will have to take into account intrinsic stochasticity of pushback estimates Pushback Prediction 25
Future Research Check what data sources are available to use for Taxi Out Time Model (Idris, Clarke, et al.) Pushback Time Prediction Model Make Models Responsive to Dynamic Stream of Data Updates Estimate Potential Benefits 26