Taxi-out Time Prediction

for Departures at Charlotte Airport

Using Machine Learning Techniques

Hanbong Lee

3rd Joint Workshop

for KAIA/KARI – NASA ATM Research Collaboration

NASA Ames Research Center

October 24-26, 2016

https://ntrs.nasa.gov/search.jsp?R=20170000660 2020-05-20T07:34:54+00:00Z

Outline

• Introduction: Aircraft taxi time prediction

• Charlotte Douglas International Airport (CLT)

• Taxi-out time data analysis

• Taxi time prediction using machine learning techniques

• Prediction performance evaluation

• Ongoing work for ATD-2

– Linear regression model with live data at CLT

2/24

Motivation

• Taxi-out time for departing aircraft

– Ground movement time from pushback to takeoff

– Depend on taxi route and surface congestion

• Aircraft taxi time prediction

– Increase takeoff time predictability

– Improve efficiency in airport surface operations

– Help controllers find better takeoff sequences to maximize runway throughput

• However, accurate prediction is difficult.

– Uncertainties in airport operations

– Operational complexity

3/24

Previous Research

• Queuing models for taxi-out time estimation

• Machine learning based approaches

– Linear regression models, Neural network model, Reinforcement learning algorithms, etc.

– Independently applied to limited data at several airports

• Taxi time prediction using machine learning methods and fast-time simulation (Lee, 2015)

– Used human-in-the-loop simulation data for CLT

– Possibly over-trained with limited datasets

4/24

Objectives

• Analyze actual taxi time data at Charlotte airport (CLT)

– Identify unique operational characteristics of CLT

– Determine key factors affecting taxi times

• Develop precise taxi time prediction modules

– Based on taxi-out time data analysis

– Using machine learning techniques

• Evaluate taxi time prediction performance

– Using actual surface surveillance data at CLT

– Comparison of prediction methods

• Apply the taxi time prediction module to live data and incorporate it with a tactical scheduler for ATD-2 project

5/24

Charlotte International Airport (CLT)

6/24

Hardstand

A

CB

D

E

18R

23

18L

18C

5

36L36C

36R

1

2

3

4 56

7

12

11

10

8

Ramp Area

Taxi-Out Time Data Analysis

• Taxi-out time data

– Used actual flight data at CLT in 2014

– Analyzed 246,083 departures after data filtering

• Taxi-out times categorized by

– Terminal concourse

– Spot

– Runway

– Departure fix

– Aircraft weight class

– Month

7/24

A6.1%

B16.7%

C19.8%

D6.1%

E44.6%

Unknown

6.7%

0

5

10

15

20

25

30

35

A B C D E Unknown

−−− 2014 AverageStandard Deviation

8/24

Taxi Time by Terminal

Average taxi time seems insensitive to terminal concourse, except for concourse D used by international flights.

Departure distribution by terminal concourse

Average taxi-out time (in minutes) by terminal concourse

9/24

Taxi Time by Spot

S13.2%

S212.1%

S30.8%

S47.7%

S51.1%

S66.0%

S720.6%

S812.8%S10

4.5%

S110.5%

S1211.2%

Unknown

19.3%

0

5

10

15

20

25

30

35

40

45

Spots S10, S11 and S12 are assigned to flights from concourse D/E to runway 18L, leading to short taxi time.

Departure distribution by spot

Average taxi-out time (in minutes) by spot

−−− 2014 Average

close to hardstand close to runway 18L

10/24

Taxi Time by Runway

50.4%

18C25.5%

18L30.7%

18R0.0%

230.6%

36C23.7%

36L0.0%

36R19.0%

Unknown

0.0%

0

5

10

15

20

25

30

35

40

45−−− 2014 Average

Taxi distance from terminal to runway affects taxi-out time directly.

Departure distribution by runway

Average taxi-out time (in minutes) by runway

close to terminal

South flow traffic

North flow traffic

far from terminal

11/24

Taxi Time by Departure Fix

MERIL20.0%

NALEY19.6%

BUCKL13.4%

BNA9.5%

BGRED9.1%

ZAVER4.2%

LILLS4.1%

TAY3.6%

GANTS3.4%

VXV2.8%

GIPPR1.9%

Others7.6%

Unknow

n

1.0%

0

5

10

15

20

25

30

35−−− 2014 Average

Taxi times of top 3 fixes for miles-in-trail (MIT) constrained departures are similar to the whole year average.

Departure distribution by departure fix

Average taxi-out time (in minutes) by departure fix

top 3 fixes for MIT constraints use short taxi routes

12/24

Taxi Time by Weight Class

Heavy1.6%

Large93.4%

Small+2.5%

Small0.0%

Unknown

2.5%

0

5

10

15

20

25

30

35

Heavy Large Small Plus Small Unknown

−−− 2014 Average

Heavy aircraft have relatively longer taxi times, whereas small aircraft have shorter taxi times.

Departure distribution by weight class

Average taxi-out time (in minutes) by weight class

13/24

Taxi Time by Month

Jan8.2% Feb

7.0%Mar8.7%

Apr8.3%

May8.6%Jun

8.5%Jul

8.9%

Aug8.7%

Sep8.0%

Oct8.5%

Nov8.1%

Dec8.5%

0

5

10

15

20

25

30

35

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

−−− 2014 Average

Average taxi times are insensitive to month, meaning no seasonal effect on taxi-out time.

Departure distribution by month

Average taxi-out time (in minutes) by month

• Separate data analysis using live data on 9/16-23/2016

• Average ramp taxi time as a function of congestion level in ramp area

Taxi Time by Congestion Level

400

450

500

550

600

650

700

750

800

850

0 1-5 6-10 11-15 16-20Ave

rag

e r

am

p t

ax

i ti

me (

se

co

nd

s)

# departures in the ramp taxiing to the same runway

Runway 36R

Runway 36C

14/24

Taxi Time Prediction Methods

• Machine learning techniques tested

– Linear Regression (LR)

– Support Vector Machines (SVM)

– k-Nearest Neighbors (kNN)

– Random Forest (RF)

– Neural Networks (NN)

• Dead Reckoning (DR) method

– Baseline for comparison

– Based on unimpeded taxi times, defined as 10th percentile of taxi times having the same gate, spot, and runway

15/24

16/24

Features

• Terminal concourse and Gate

• Spot

• Runway

• Departure fix

• Weight class and Aircraft model

• Taxi distance

• Unimpeded taxi time

• Scheduled pushback time of day

• Number of departures and arrivals on the surface

17/24

Training and Test Datasets

Traffic flow

Weather Dataset Dates Data size Avg. Taxi time (min)

Std. Dev. (min)

Southflow traffic

Good weather

Training 6/1, 6/2, 6/4, 6/7, 6/15

3,361 17.11 6.65

Test 8/15 689 17.78 6.59

Rain Training 6/11, 6/12, 6/25, 7/9, 8/11

3,280 17.98 6.99

Test 8/12 644 17.68 6.51

Northflow traffic

Good weather

Training 6/6, 6/20, 8/25 2,134 19.32 6.13

Test 8/26 684 19.36 6.09

Rain Training 7/21, 8/1, 8/23 1,944 18.83 6.25

Test 8/24 621 19.31 6.32

• Two runway configurations: south flow and north flow

• Two weather conditions: good weather and heavy rain

18/24

Prediction Results – South Flow

Machine learning algorithms show better performance than Dead Reckoning (DR) method. Linear Regression (LR) and Random Forest (RF) are the best.

South-flow traffic, good weather South-flow traffic, heavy rain

Taxi Time Difference (Actual – Predicted) (in minutes) Taxi Time Difference (Actual – Predicted) (in minutes)

19/24

Prediction Results – North Flow

Linear Regression (LR) and Random Forest (RF) are still the best prediction methods for both traffic flow.

Taxi Time Difference (Actual – Predicted) (in minutes) Taxi Time Difference (Actual – Predicted) (in minutes)

North-flow traffic, good weather

North-flow traffic, heavy rain

Conclusions

• Analyzed the whole year taxi time data at CLT

– Found several factors affecting taxi-out time

– No seasonal effect on taxi time

• Applied various machine learning techniques to actual flight data at CLT for taxi-out time prediction

– Machine learning methods were better than Dead Reckoning method based on unimpeded taxi time.

– Linear Regression and Random Forest methods showed the best prediction performance.

– Considered various operational factors, but still needs to be improved.

20/24

21/24

Ongoing Work for ATD-2

• Apply a linear regression model to live data– Focus on ramp taxi time prediction

• Update taxi speed decision trees used in Tactical Scheduler– Current taxi speed decision trees based on historical flight

data and taxi route data• Two decision trees for estimating taxi-out times of

departures and taxi-in times of arrivals

• Taxi speed values both in AMA and Ramp in knots

• Branches by runway, spot, ramp area, and weight class

– Need to account for congestion on the surface• Count the number of aircraft moving on the surface when a

departure is ready to push back

• Formula

– xf: variables for flight f– yf: predicted ramp taxi time of flight f– Constant and Coefficients determined by training dataset

• Variables– Ramp taxi distance (from gate to spot)– Binary variables

• Ramp area, spot, runway, weight class, and EDCT– Scheduled off-block time– Congestion factors

• Number of departures in ramp area (by runway and ramp area)• Number of arrivals in ramp area (by ramp area)

– Departures in the previous 15 minutes• Number of flights going to the same runway, and their mean taxi time• Number of flights going to the same fix, and their mean taxi time

Linear Regression Model

y f =Const + Coeffi × xif

i=1

n

å

22/24

0

20

40

60

80

100

120

140

0 2 4 6 8 10 12 14 16 18 20 22 24

Nu

mb

er

of

de

par

ture

s

Taxi-out time (minutes)

Actual Taxi Time Distribution

0

20

40

60

80

100

120

140

-15-13-11 -9 -7 -5 -3 -1 1 3 5 7 9 11 13

Nu

mb

er

of

de

par

ture

s

(Predicted) – (Actual taxi time) (minutes)

Taxi Time Difference Distribution

• Live data from CLT

– North-flow traffic both in training dataset (9/16-22/2016) and test dataset (9/23/2016)

• Prediction accuracy

• Departures within ±5-min error window: 714 (89.8%)

• Departures within ±3-min error window: 549 (69.1%)

Linear Regression Result

(Pred.) - (Actual):Average: 0.37Std.Dev: 3.23Minimum: -14.45Maximum: 8.42Median : 1.06

Total flights: 795Average: 9.38Std.Dev: 3.62Minimum: 2.12Maximum: 23.57Median : 8.73

23/24

• AAL1832 from CLT to SAT (A319)– Taxi route: B8 S13 36C

• Default ramp distance from gate to spot: 370.5m

– Number of departures taxiing on surface: 6• Two aircraft from each Concourse B, C, and E to runway 36C

• Linear Regression model

– Predicted ramp taxi time: TaxiTLR = 0.2735*370.5 + 166.2 + 28.6 + 189.6 + 74.2

+ 9.9*2 + (-1.3) *2 + 4.6*2

= 586.3 seconds

• Actual ramp taxi time: 573 seconds (Difference: 13.3 seconds)

– Predicted taxi speed in ramp area: 370.5/(586.3 – 260) = 2.2 knots

Linear Regression Example

Variable Ramp Distance

B_EAST Spot 13 Runway 36C

Weight Class D

Dep# B to 36C

Dep# C to 36C

Dep# E to 36C

Coefficient

0.2735 166.2 28.6 189.6 74.2 9.9 -1.3 4.6

24/24