Modeling Flight Delays and Cancellations at the National, Regional and Airport Levels in the United States Banavar Sridhar Yao Wang Alex Klein Richard Jehlen National Airspace System Performance Workshop Asilomar, CA April 13-16, 2009
Modeling Flight Delays and Cancellations at the National, Regional and Airport Levels in the
United States
Banavar SridharYao WangAlex Klein
Richard Jehlen
National Airspace System Performance WorkshopAsilomar, CA
April 13-16, 2009
2
Motivation
• Weather is the major cause of delay in the National Airspace System (NAS)
• Four possible scenarios
• Relate delay, cancellations and other NAS performance metrics to the weather conditions to improve Traffic Flow Management
PoorWeatherForecast
AccurateWeatherForecast
Poor operationalresponse
X X
Properoperationalresponse
X Bestoutcome
Results
• Developed flight delay and cancellation models at the national, regional and airport levels
• Expected number of aircraftimpacted by weather goodproxy for delay
• Different models for summer and winter• All metrics can be estimated to same level of accuracy• FAA (ASPM and OPSNET) databases are complementary• Neural Network models perform slightly better
3
4
Outline
• Objectives • Databases• Airspace Performance Metrics• Modeling/Estimation of Metrics
– Regression Models– Neural Network Models
• Results• Conclusions
5
Objectives
• Develop NAS performance metric models based on FAA operational traffic databases– Different metrics– Impact of databases – Approach
• Linear regression models• Neural networks models
6
Databases
• FAA Operations Network (OPSNET)– Data available from 1990– Daily values– 45 airports– Total national delay
• Aviation System Performance Metrics (ASPM)– Data available from 2000– Every 15 minutes– 75 airports– Total OAG-based and flight-plan based arrival delays,
EDCT hold minutes, airborne delay, flight cancellations• Paper uses data from 2005-2008
7
NAS Performance Metrics
8
Weather Impacted Traffic Index (WITI)
Severe weatherAircraft positions
Weather Impacted Traffic Index (WITI)
• Grid-based WITI• National Weather Index (NWX)
– En-route WITI (E-WITI), representing convective weather impact on major flows between city pairs
– Terminal WITI (T-WITI), representing weather impact on major airports
– Airport Queuing Delay (Q-Delay), representing surface and terminal-airspace weather impact on major airports in a non-linear fashion
9
10
• Number of aircraft affected by weather ( )• Number of aircraft affected by weather in each
Center ( )• Performance metric ( ) • Models
– Linear Regression (LR)
– Multiple Linear Regression (MLR)
– Neural Networks
– Dynamic Models
Modeling/Estimation of Metrics
11
Performance of Regression Models
OPSNET Total Delay Flight Cancellations
Estimate
Error
12
Performance of Regression Models
• Regression models perform a good job of accounting for the impact of weather on delays and flight cancellations
• For systems with demand-capacity imbalance, growth in delay is non-linear
Atlanta Airport
Nonlinear Models
Inputs PerformanceMetric(Target)
Performance of National Model
14• Neural Network models perform slightly better
Performance of different WITI definitions
15
• Models using NWX perform slightly better• Difference not significant while using MLR or NN
Seasonal performance of national delay models
16
• Higher correlation during summer• Lower correlation in winter may be due to higher number ofcancellations on days with heavy snow, very low ceilings/visibility
Airport delay models using Regression analysis
17
• 34 major airports on the OEP-list
• Good delay estimates for ORD, ATL,..• Delay at ten airports in Eastern U.Snot influenced by NWX in the neighboring Centers
Behavior of airport models
18
Welcome Joe User Today is Thursday, August 14, 2008
Print | Help | Log Out | Customize
APF Dashboard For Wednesday, August 13, 2008
OPS – Delayed & On-Time Gate Arrivals Last 8 Days Holding Minutes Last 8 Days
OPSNET Delays Last 8 DaysWITI Last 8 Days
0
1000
2000
3000
4000
5000
6000
EnRt Conv Local Conv Wind Snow
IMC Queue Del Unfav Rwy Config WITI
Efficiency Last 8 Days
88
89
90
91
92
93
94
95
96
97
98
Terminal_AER System_AER
TMI Minutes Last 8 Days
0
1000
2000
3000
4000
5000
6000
AFP_Minutes GDP_Minutes GS_Minutes
Advisories Last 8 Days
0
20
40
60
80
100
120
140
160
NAS Reroute NAS GDP NAS GS NAS Other
19
Diverts, MIT, Predictability Last 8 Days
Default Customer View
NAS Service Area ARTCC Airport
Customer View ATO View
Days 1 3 8 30
0
10000
20000
30000
40000
50000
On-Time Ops Del'd Ops
0
2000
4000
6000
8000
10000
12000
< 100nm Actual < Rate < 100nm Actual >= Rate > 100nm
0
5
10
15
20
25
30
35
40
0
10000
20000
30000
40000
50000
60000
70000
80000
90000
NAS_MIT_MinuteMiles x 1000Total Diversions_countAverage Predictability_Range
0
200
400
600
800
1000
1200
1400
1600
1800
2000
Weather Vol Term Vol Center Rwy TaxiEquip Other Non-OEP 35
20
Concluding Remarks
• Estimation/Modeling of performance metrics resulting from the use the two databases are comparable
• For all metrics, neural networks produce higher correlation and reduced errors than regression methods
• Different methods of reducing neural network complexity produce similar results