neural-geo-temporal approach to travel demand modelling
TRANSCRIPT
ANALYSIS OF THE EVOLUTION OF TRAVEL DEMAND IN ANALYSIS OF THE EVOLUTION OF TRAVEL DEMAND IN URBAN AREAS: A NEURAL-GEO-TEMPORAL MODELLING URBAN AREAS: A NEURAL-GEO-TEMPORAL MODELLING
APPROACHAPPROACH
Andre Dantas, University of Canterbury, Christchurch, New Zealand.Andre Dantas, University of Canterbury, Christchurch, New Zealand.Yaeko Yamashita, University of Brasilia, Brasilia, Brazil.Yaeko Yamashita, University of Brasilia, Brasilia, Brazil.Marcus Vinicius Lamar, Federal University of Parana, Curitiba, BrazilMarcus Vinicius Lamar, Federal University of Parana, Curitiba, Brazil
Department of Civil EngineeringDepartment of Civil EngineeringMaster in Transportation EngineeringMaster in Transportation Engineering
•IntroductionIntroduction•Neural Networks and GISNeural Networks and GIS•Theoretical conception of the NGTMTheoretical conception of the NGTM•Case StudyCase Study•ConclusionsConclusions
Outline of the presentationOutline of the presentation
Complex commuting patterns all over the
city.
t=1
t=2
t=n
Central displacements on foot;
Travel Demand
Long-motorized travel from suburbs to CBD;
Introduction - Urban changes and travel demandIntroduction - Urban changes and travel demand
Agglomerations in urban centers and
subcenters
Concentration of activities
Provision of variousservices and facilities
Attraction of high amount of daily commuting
TrafficCongestion
CarAccidents
Pollution
NO
ParkingSpace
CONSEQUENCES
LIFEQUALITY
Introduction - Urban changes and travel demandIntroduction - Urban changes and travel demand
Introduction - travel demand modellingIntroduction - travel demand modelling
Considerable amount of research effortsConsiderable amount of research efforts
Land useLand use ..........
Application of economic theoryApplication of economic theory
Urban development modelsUrban development models
To explain the configuration and evolution of urban structuresTo explain the configuration and evolution of urban structures
Integrated Land use-ransport modelsIntegrated Land use-ransport models
Incorporate the most important spatial processes of development Incorporate the most important spatial processes of development in conjunction with travel demand forecastingin conjunction with travel demand forecasting
Travel demandTravel demand
Introduction - travel demand modellingIntroduction - travel demand modelling
Travel demand models have been criticizedTravel demand models have been criticized
BUT...BUT...
Massive and costly data Massive and costly data requirementsrequirements
for application to real problems.for application to real problems.Non-incorporation of temporal dynamic and Non-incorporation of temporal dynamic and
realistic dimensions of urban realityrealistic dimensions of urban reality
Still based on the traditional four-Still based on the traditional four-step approachstep approach
Introduction - travel demand modellingIntroduction - travel demand modelling
DynamicDynamic modellingmodelling
Land useLand useTransport system Transport system
interactionsinteractions
Non-linearNon-linear modellingmodelling
Travel demand models - RESEARCH PERSPECTIVESTravel demand models - RESEARCH PERSPECTIVES
Introduction - travel demand modellingIntroduction - travel demand modelling
Exploring new modelling techniques!Exploring new modelling techniques!
Employing available Employing available technology!!technology!!
Using Geo-spatial data!!! Using Geo-spatial data!!!
How do we reach this new modelling frontier?How do we reach this new modelling frontier?BUT...BUT...
Introduction - travel demand modellingIntroduction - travel demand modelling
Is it technically feasible???Is it technically feasible???
BUT...BUT...
What’s available???What’s available???
Can WE really think in a different way?Can WE really think in a different way?
Neural
Geo
Temporal
Model (NGTM)
Introduction - NGTMIntroduction - NGTM
GIS
Neural Networks (NN)
Consider spatial-temporal evolution of urban areas
To incorporate temporal interactions between the transportation system and
land use patterns
Overcome limitations
of linear modelling
Neural
Geo
Temporal
Model (NGTM)
Introduction - NGTMIntroduction - NGTM
NN AND GISNN AND GIS
•What is GIS?A special type of Information System;
Database, personnel and Technology in a systemic and interactive manner;
Manipulation, storage, visualization of georeferenced data; and
SPATIAL ANALYSIS providing new and useful information for decision-making/planning activities.
NN AND GISNN AND GIS
SPATIALANALYSIS
Entry
Manipulate
Present
Storage
ModellingWhat if…?
ROUTINGBest way to...?
TRENDWhat has changed?
CONDITIONHow is it…?
LOCATIONWhat is at…?
NN AND GISNN AND GIS
•What is a NN?A non-linear extension of conventional spatial statistical models;
Model reached without a priori assumptions;
To compute a function that expresses the correlation between independent and dependent variables;
An “analogy” of human brain processing;
NN AND GISNN AND GIS
wk1
wk2
wkj
x1
x2
xj
ukInput Signals
yk (uk )
wkpxp
SummingFunction(adder)
LinkWeights
ActivationFunction
Output
)exp(11)(
kk u
u
The Back-Propagation Algorithm
i
ii wOywE
2)(
21)(
)()()1( )()(
)( tww
wEtw lijl
ij
lij
Cost (error) Function
produced outputby the net, dependingon the weight matrix
training valueoutput
Learning Rule
learn rate used = 0.01 puls term used = 0.9
1
0
2)(1 M
iii Oy
ME
1
0
1 P
EP
E
Pyx ik ,...,1,,
Training Set ofInput-Output pairs
Average Root Mean Square (AveRMS)
NN AND GISNN AND GIS
NN
Training Data set
Test Data set
Training Process
Modelling function
Pre–ProcessingData Set
NN AND GISNN AND GIS
at the ith zone
TransportationSystem
SpatialLocation
Land Use patterns
Population
Interactions(UI)
Trip Generation
(TG)f
NGTM - THEORETICAL CONCEPTIONNGTM - THEORETICAL CONCEPTION
Local Spatial Interactions
Surrounding Spatial Interactions
Global Spatial Interactions
TransportationSystem
LandUse
Zone i Zone i Zone i
NGTM - THEORETICAL CONCEPTIONNGTM - THEORETICAL CONCEPTION
t=1t=2
t=n
Local Spatial Interactions
Surrounding Spatial Interactions
Global Spatial Interactions
NGTM - THEORETICAL CONCEPTIONNGTM - THEORETICAL CONCEPTION
UIi(1)
TGi(1)
UIi(2)
TGi(2)
UIi(z)
TGi(3) TGi
(z) TGi(z+1)
NGTM - RECURSIVE MODELLINGNGTM - RECURSIVE MODELLING
zi
zi
zi
zi
Zi
POSLLUTSUI ,,,
Input Vector zone ith, time z
Transportation system
Land Use
Spatial location
Population
NGTM - REPRESENTING SPATIAL-TEMPORAL INTERACTIONSNGTM - REPRESENTING SPATIAL-TEMPORAL INTERACTIONS
INDEPENDENT VARIABLES
1ZiTG
NGTM - REPRESENTING SPATIAL-TEMPORAL INTERACTIONSNGTM - REPRESENTING SPATIAL-TEMPORAL INTERACTIONS
DEPENDENT VARIABLES
Trip production - number of trips FROM zone i
Trip attraction - number of trips TO zone i
.....
NGTM - REPRESENTING SPATIAL-TEMPORAL INTERACTIONSNGTM - REPRESENTING SPATIAL-TEMPORAL INTERACTIONS
zi
zii TGIUX ,
1z
ii TGY
InputLayer
...
TGiz+1
...HiddenLayer
OutputLayer
UIiz TGi
z
z-1
ContextLayer
...-1 -1
NGTM - APPLICATION OF AN ELMAN NNNGTM - APPLICATION OF AN ELMAN NN
•326,35 Km2
•2,2 million people (1998)
CASE STUDY - LOCATION AND CHARACTERISTICS OF NAGOYA CASE STUDY - LOCATION AND CHARACTERISTICS OF NAGOYA CITYCITY
CASE STUDY - LOCATION AND CHARACTERISTICS OF NAGOYA CASE STUDY - LOCATION AND CHARACTERISTICS OF NAGOYA CITYCITY
4
21
Data collection & gathering
t = 0t = 1
t = n
NN
Data updatet = n+1
3
GIS
CASE STUDY - INTEGRATION NN-GISCASE STUDY - INTEGRATION NN-GIS
GIS
1 2
3
I
V
Georeferencing
Visualization & analysis
SpatialPatterns
Database
II
Queries
III
ntntntnt TripsZ,.....,Y,Xf
Modelling FunctionIVSpatial
analysis
VIII
VI VII
CASE STUDY - INTEGRATION NN-GISCASE STUDY - INTEGRATION NN-GIS
Road Transportation system Railway Transportation system
CASE STUDY - MAIN FEATURES OF THE GIS DATABASECASE STUDY - MAIN FEATURES OF THE GIS DATABASE
Large scale commercial land use
Commerce 1991
Commerce 1981
Commerce 1971
CASE STUDY - MAIN FEATURES OF THE GIS DATABASECASE STUDY - MAIN FEATURES OF THE GIS DATABASE
Features (Spatial data) Attributes (Non-spatial data)
Point Line Area AT2 …… ATvAT1
TZStations
Bus StopsBus Lines
Railways
NH rampNHRoads
Land Use
•248 Traffic Zones (TZ)•Data from 3 different sources
CASE STUDY - MAIN FEATURES OF THE GIS DATABASECASE STUDY - MAIN FEATURES OF THE GIS DATABASE
Introduction - travel demand modellingIntroduction - travel demand modelling
BUT...BUT...3 years to obtain/construct the 3 years to obtain/construct the
DATABASEDATABASE
And people say...And people say...““You were lucky!!!”You were lucky!!!”
CASE STUDY - TEMPORAL EVOLUTIONCASE STUDY - TEMPORAL EVOLUTION
Pattern No. Pattern Description No. Cases No.Cases%1 PO 71=PO 81 PO 81>PO 91 57 22.98%
13 PO 71=PO 81 PO 91>PO 81 40 16.13%
11 PO 71=PO 81 PO 81>PO 91 PT 71<PT 81 PT 81=PT 91 12 4.84%
3 PO 71=PO 81 PO 81>PO 91 PT 71=PT 81 PT 81<PT 91 10 4.03%
21 PO 71=PO 81 PO 81>PO 91 RL 71<RL 81 RL 81=RL 91 8 3.23%
14 PO 71=PO 81 PO 91>PO 81 PT 71<PO 81 PT 81=PT 91 7 2.82%
41 PO 71=PO 81 PO 91>PO 81 RL 71<RL 81 RL 81=RL 91 7 2.82%
6 PO 71=PO 81 PO 81>PO 91 PT 71=PT 81 PT 81<PT 91 RL 71=RL 81 RL 81<RL 91 7 2.82%
9 PO 71=PO 81 PO 81>PO 91 PT 71<PT 81 PT 81<PT 91 7 2.82%
22 PO 71=PO 81 PO 81>PO 91 RL 71=RL 81 RL 81<RL 91 7 2.82%others 86 34.68%
61 patterns - TS/LU changes61 patterns - TS/LU changes
CASE STUDY - TEMPORAL EVOLUTIONCASE STUDY - TEMPORAL EVOLUTION
CASE STUDY - TEMPORAL EVOLUTIONCASE STUDY - TEMPORAL EVOLUTION
Pattern No. Pattern Description No.Cases No.Cases%3 TG 71<TG 81 TG 81>TG 91 155 62.50%
1 TG 71>TG 81 TG 81>TG 91 59 23.79%
4 TG 71<TG 81 TG 81<TG 91 21 8.47%
2 TG 71>TG 81 TG 81<TG 91 13 5.24%
CASE STUDY - TEMPORAL EVOLUTIONCASE STUDY - TEMPORAL EVOLUTION
NN
2
3
Training Data set
Test Data set
BTraining Process
C
Modelling function D
Pre–ProcessingData Set
A
CASE STUDY - NN PROCESSINGCASE STUDY - NN PROCESSING
zi
zi
zi
zi
zi
POSLLUTSUI ,,,
zi
zi
zi
zi
NHRTPTTS ,,
zi
zi
zi RLPLCLz
iLU ,,
ii HDSDziSL ,
ZiGT Zonal Trip
ends
trips
(all modes)/hour
CASE STUDY - FORMATION OF INPUT AND OUTPUT VECTORSCASE STUDY - FORMATION OF INPUT AND OUTPUT VECTORS
NN
2
3
Training Data set
Test Data set
BTraining Process
C
Modelling function D
Pre–ProcessingData Set
A
CASE STUDY - NN PROCESSINGCASE STUDY - NN PROCESSING
Data set
496 vectors
Training / Testing
random selection
Testing vectors (124)
Training vectors (372)
Input Data normalization 1
minmaxmin ][][][][8.01.0][
UIUIUIUIIU tzi
zi
Output Data 1
minmaxmin8.01.0
AAAAA zi
zi
CASE STUDY - FORMATION OF INPUT AND OUTPUT VECTORSCASE STUDY - FORMATION OF INPUT AND OUTPUT VECTORS
NN
2
3
Training Data set
Test Data set
BTraining Process
C
Modelling function D
Pre–ProcessingData Set
A
CASE STUDY - NN PROCESSINGCASE STUDY - NN PROCESSING
TestingTraining
CASE STUDY - FORMATION OF INPUT AND OUTPUT VECTORSCASE STUDY - FORMATION OF INPUT AND OUTPUT VECTORS
NN
2
3
Training Data set
Test Data set
BTraining Process
C
Modelling function D
Pre–ProcessingData Set
A
CASE STUDY - NN PROCESSINGCASE STUDY - NN PROCESSING
UIi(1971)
Ai(1971)
UIi(1981)
Ai(1981) Ai
(1991)
CASE STUDY - NGTM TRAININGCASE STUDY - NGTM TRAINING
NN parameters: NN parameters: •learning rate=0.01learning rate=0.01•sigmoid activation sigmoid activation
functionfunction•MSE=0.000229MSE=0.000229
•36677 iterations36677 iterations•20 hours of 20 hours of
processing (Pentium processing (Pentium 200 MHz)200 MHz)
NN
2
3
Training Data set
Test Data set
BTraining Process
C
Modelling function D
Pre–ProcessingData Set
A
CASE STUDY - NN PROCESSINGCASE STUDY - NN PROCESSING
CASE STUDY - NGTM TESTINGCASE STUDY - NGTM TESTING
Average Average error=87.11error=87.11
Standard Standard deviation of deviation of
the the errors=128.errors=128.
3535
Average Average relative relative
error=23.8%error=23.8%
R2=0.93
Introduction - travel demand modellingIntroduction - travel demand modelling
Why???Why???
0
0.5
1
1.5
2
2.5
3
3.5
4
zones
Trip
s (x1
000)
-60
-40
-20
0
20
40
60
80
100
120
Rela
tive
Erro
r(%
)
GVi91 Yi91 Erro Relativo (%)TGi91 Relative Error (%)
CASE STUDY - NGTM TESTINGCASE STUDY - NGTM TESTING
CASE STUDY - NGTM TESTINGCASE STUDY - NGTM TESTING
AA9191 =3373; =3373; YY91 91
=3170 =3170 AA8181>>AA7171
AA8181>>AA9191
TSTS9191>>TSTS7171
Commercial LU upCommercial LU up
AA9191 =113; =113; YY91 91 = = 232232
21% PO increase21% PO increaseAA8181==126126
AA7171==5252
UIi(1991)
Ai(2001)
UIi(1971)
Ai(1971)
UIi(1981)
Ai(1981) Ai
(1991)
CASE STUDY - NGTM FORECASTINGCASE STUDY - NGTM FORECASTING
[-160;-50[ [-50;-25[ [-25;0[
[0;50[ [50;100[ [100;200[ 200
Relative variation 1991-2001 (%)•Maximum positive
variation =441%
•Maximum negative variation =-160%
•Average variation =44.81%
CASE STUDY - NGTM FORECASTINGCASE STUDY - NGTM FORECASTING
[0;500[ [500;1000[ [1000;2000[
[2000;3000[ [3000;5000[ 5000
Zonal trip ends 2001
•Max. =7831
•Min. =0
•Average=574.12
•Variation % (2001-1991)= 1.45
CASE STUDY - NGTM FORECASTINGCASE STUDY - NGTM FORECASTING
Achievements in the case study
NGTM’s contribution
Modelling function capable of computing a very complex reality with time variation;
Remarble barriers on constructing the GIS database for the multi-year period.
Future improvements
Incorporation of temporal dimension into travel demand modelling;
Non-linear approach based on NN; and
Incorporation of geo-spatial data expressing urban interactions using GIS.
Classified output & use of extensive temporal database
CONCLUSIONCONCLUSION
Andre Dantas, University of Canterbury, Christchurch, Andre Dantas, University of Canterbury, Christchurch, New Zealand.New Zealand.
SCHOOL OF ENGINEERINGSCHOOL OF ENGINEERINGCIVIL ENGINEERINGCIVIL ENGINEERING4F4FROOM 406ROOM [email protected]@CANTERBURY.AC.NZEXTENSION 6238EXTENSION 6238
Department of Civil EngineeringDepartment of Civil EngineeringMaster in Transportation EngineeringMaster in Transportation Engineering
THANK YOU!!
OBRIGADO!!