An Econometric Analysis of US Airline Flight Delays with of-day Effects

Transcription

1 An Economeric Analysis of US Airline Fligh Delays wih Time-of of-day Effecs Chieh-Yu Hsiao Mark Hansen Naional Cener of Excellence in Aviaion Operaions Research (NEXTOR) Deparmen of Civil and Environmenal Engineering Universiy of California a Berkeley 01/20/06 1

2 Ouline Inroducion Model Facors Daa Esimaion Resuls Delay Change Decomposiion Conclusions 2

3 Inroducion: Arrival Delays & Scheduled Arrivals Arrival Delay Scheduled Arrivals 3 Arrival Delay (Minues per Fligh) Scheduled Arrivals (1000 per Day) Time Period

4 Goals Develop saisical model o forecas delay in he NAS Cerain advanages over simulaion models Esimaed over large number of days Reduced daa needs Insananeous calculaion Fuure year predicion based on applicaion o a full year of days raher han a few sample days Compare resuls wih simulaion 4

5 Objecives Model daily average arrival delay Assess impacs of Queuing Volume Weaher (sorm and erminal condiions) Seasonal Effecs Invesigae ime of day effecs for queuing delay 5

6 Model AvDelay() = f ( Queuing(), SFlighs(), GAFlighs(), MFlighs(), IFR(), Windspeed(), THX (), Evens(), Season(), Period()) + ν 6

7 Queuing Calculae deerminisic queuing delay on a daily basis for each of 32 DOT airpors Based on Quarer-hr demand profile (based on schedule and adjused for cancelled flighs) Quarer-hr AAR 7

8 Queuing: Procedure Airpor level Consruc Cumulaive Arrival Demand Curve Consruc Cumulaive Arrival Coun Curve Calculae Delay as Area Beween Demand and Coun Curves Sysem level Summed airpor delays o ge oal delay Divided by oal arrivals o ge average delay Cumulaive Arrival Demand or Coun :00 Demand Curve 7:15 7:30 7:45 8:00 Coun Curve 8:15 8:30 Time a Beginning of Period 8:45 9:00 9:15 8

9 En Roue Weaher (Sorms) Based on Surface Summary of he Day from NOAA (Naional Oceanographic and Amospheric Adminisraion) Daily summary from ~1500 US weaher saions Each saion repors binary (yes/no) hundersorm variable Used proporion reporing hundersorms as sorm meric Consruc area-specific hundersorm merics on la-long quandrangles Include each meric as separae explanaory variable 9

10

11 Terminal Weaher Proporion of he day under IFR Airpor IFR proporion: proporion of ime under IFR Sysem IFR proporion: weigh airpor IFR proporion by # of operaions Higher IFR proporion, expeced higher delay due o lower AAR Wind speed Airpor wind speed: avg. wind speed of he day Sysem wind speed: weigh airpor wind speed by # of operaions Higher wind speed, expeced higher delay 11

12 Oher Effecs Volume Scheduled arrivals GA operaions Monhly fixed effecs Based on monh Effecs of winds and oher weaher facors Yearly fixed effecs Year-o-year rends no oherwise accouned for May reflec FAA performance among oher facors UA srike effec 9/11 Effec 12

13 Esimaion Daa Daily daa: Jan June 2005 Excludes 12/20-25, 9/11-30/01, and sandard/dayligh savings ransiion days Esimaions OLS model: Heeroskedasiciy and auocorrelaion GARCH model Model form y = β + v ε X ' v = ε AR1 v 1 = h e h e = ARCH GARCH1 h ~ IN(0,1) 2 ARCH1 ε HET1 Q( ) 13

14 Caegory Queuing Volume Terminal Weaher En Roue Weaher Even Esimaion Resuls Variable (parameer for caegory GARCH) Baseline Sandard Esimae Error Time-of-Day Sandard Esimae Error Ineracion Sandard Esimae Error Full Sandard Esimae Error Inercep Queuing Queuing (00:00--08:00) Queuing (08:00--12:00) Queuing (12:00--16:00) Queuing (16:00--24:00) Queuing Scheduled arrivals (000) GA operaions (000) IFR raio Wind speed Sch*IFR raio Sch*Wind speed Sorms in region Sorms in region Sorms in region Sorms in region Sorms in region Sorms in region Miliary Operaions (000) Sch*Sorms in region Sch*Sorms in region Sch*Sorms in region Sch*Sorms in region Sch*Sorms in region Sch*Sorms in region Sch*Miliary Operaions (000) UA srike dummy nd half 2001 afer 9/11dummy R-Squared Noe: "Scheduled arrivals" is abbreviaed as "Sch."; Time dummies and GARCH variables are used bu no shown here. 14

15 Key Findings Time of day queuing effecs Vary srongly by ime-of-day Greaes impac in morning (delay propagaion) Concave relaionship Volume Scheduled arrivals: 1000 addiional scheduled flighs increases average delay by minues on exremely good weaher days GA operaions: 1000 addiional GA operaions in increases average delay by 0.1 minues Miliary operaions: srong negaive relaionship wih delay (!) 15

16 Key Findings (con) Terminal weaher effecs Depend on volume Have impac even when conrolling for queuing En roue weaher effecs Depend on volume Wide geographic variaion Oher effecs 2000 UA job acion had pronounced effec Pos 9/11 effec no significan (capured by oher variables) 16

17

18 Conclusions Saisical models help us explain and predic delay changes Use of deerminisic queuing increases fideliy of hese models and allows ime-of-day effecs o be capured En roue convecive weaher effecs can be capured in fairly simple ways Subsanial unexplained variaion remains 18