Managing Transport Modelling
ROLE OF TRANSPORT MODELLING
How to use Transport Models to assist
decision making; assumptions and constraints
Luis Willumsen
BETTER FORECASTS
We use models to improve decision making
How to use transport models to assist decision
making.
• Better Models -> Better Forecasts
• Better Forecasts -> Better Decisions
• Short Term Decisions -> Better Accuracy
• Long Term Decisions -> Better basis for decision
making
2
Why are models required?
• We cannot do experiments, in most cases
• Transport investments takes a long time so we
must envisage future conditions that do not yet
exist
• We can test alternative solutions in this modelled
future
3
How do we use models?
Formulation of the problem
Model specification
Data collation and
collection
Build, calibrate and
validate analytical model
Generate solution packages Forecast underlying
to test (planning) variables
Test solutions in the model
Sensitivity and risk tests
Evaluate solutions and
recommendations
Implement solutions
4
MODELS AND REALITY
The nature of Analytical Models
• Models are a simplification of reality based on some useful
theoretical assumptions and sufficient data to estimate
them
• We can deliver forecasts based on our models provided the
theoretical assumptions remain reasonable and data about
the future is reliable
• These conditions are never 100% met, so we need to
interpret the results from our models
• Even perfect models will not be able to deliver totally
accurate forecasts; some uncertainties cannot be eliminated
5
Accuracy and precision
• Sometimes pursuing precision may lead to inaccuracy
6
Accuracy and precision
• Sometimes pursuing precision may lead to inaccuracy
• It is better to be approximately right than precisely wrong
ONE SHOT
is not
enough!
7
8
Where models can help
MODEL OBJECTIVES
Assessing options
• Scheme testing
• Policy testing
Assessing impacts
• Traffic and congestion
• Economic (user, non-user benefits)
• Financial (revenue)
• Environmental (CO2, NOX emissions), Safety (accidents)
• Performance
• Equity: winners and losers
Forecasting usage (patronage, traffic levels etc.)
9
MODEL SPECIFICATION
How to specify a Model
Need to understand and define the problem
Have an idea of the possible solutions
What are the main behavioural responses to these
solutions
Resources Available: Time and Budgets, Data, Software,
Skills
The costs of “getting it wrong”
Important to identify what the model should be able to
do
And what the model will NOT be able to do
Interpretation and post modelling adjustments are fair
10
MODEL DESIGN STEPS
Model design steps
Scope of the model
Area to be covered
Behavioural responses to be included
Government standards (WebTag, etc.)
Data available
Data to be collected
Calibration/validation to be conducted of various stages
Sensitivity Tests to be carried out
The resources spent in modelling and analysis should be
related to the cost of a wrong decision
11
Area to be covered and level of detail
AREA TO BE COVERED
Directly related to the scope of
the model
Also related to the
understanding travel patterns
Type of study, strategic or
tactical
How much detail is needed for
route choice and other
responses
Source: ptv
12
Responses to include
RESPONSES
Networks and
Activity System
services
• Typical 4/5 stage model
response. Change in:
• Number of trips
• Destination Trip Generation/Attraction,
• Mode used trip frequency choice
• Time of travel
• Route used Destination choice
• Generally
• Small area models with
Mode choice
network issues –route
choice and assignment Time of departure
choice
only
• PT models – Mode Choice Route choice and
assignment
very important
• Strategic Models – All
responses Flows by link and mode
Travel costs, especially
time
13
Data availability
DATA AVAILABILITY
Data is the key in any model development task
Each response model requires different data; always look for
what is already available
Previous studies, Census data, National Database, etc.
If no other study specific data is available then essential to
carry out surveys, most commonly:
Traffic Counts, Bus, Rail, Metro occupancy surveys/boarding-alighting
counts, Intercept surveys: Road Side Interviews and PT user OD
surveys, Travel time and service quality surveys
A rich source of travel user behaviour is a detailed Household
Interviews (HHI) or Household Travel Surveys (HTS) including a
Travel Diary; generally a 1% to 2% sample.
The future is BRIGHT with new sources of mobility data
New sensors and probes:
• Radar, loops
• CCTV
• ANPR
• SmartCards
• Transponders/Tags
• Bluetooth
• WiFi & Others
• GPS
• Mobile phones
14
15
ESTIMATION, CALIBRATION AND VALIDATION OF MODEL
Calibration, Estimation and Validation
Three distinct aspects of model design:
• Calibration of a model is finding the best values for its
parameters
• Model estimation reflects the fact that the functional form
is not decided before hand in some cases
• Model Validation follows post estimation and tests the
model against data not used in calibration; backcasting is
ideal but seldom possible
Means different things for different response models
Demand model elements: total demand, Trip Length Distribution, mode
shares: compared against observed totals
Assignment model: Traffic flows, Boarding Alighting, Patronage
DATA ERRORS
Sources of error: Data and ……. Models
• Base Year Data errors: Poor and small sample, limited data
for calibration/validation;
• Better methods and QA would reduce these errors
• Future Data errors in the input variables are unavoidable;
typically
• GDP, Population, employment Growth
• Key inputs like fuel prices, fares of competing modes
• Parameter changes: willingness to pay, Values of Time; fixed
tastes assumptions
• And also future Scenarios (not entirely under control)
• New Developments that do/do not materialise
• Competing facilities that DO materialise
• New disruptive technologies
16
NZ GDP
New Zealand GDP forecasts
17
SPECIFICATION ERRORS
Sources of error: Model specification
• Bad modelling
• Poor sensitivity to price, too aggregate Values of Time
• Poor models of delay, crowding
• Ignoring relevant behavioural responses
• Including irrelevant behavioural responses
• Overconfidence in the model and its implicit assumptions
(that are never 100% right!):
• Users have perfect information and can process it efficiently
• Users respond instantly to changes
• Etc….
• Over specifying the model: unrealistic precision
18
Consider a modelling effort with 30 years planning horizon
Base year data is obtained the usual way (2-3 % HH survey
plus some intercept, travel time and service surveys plus
counts);
Model is calibrated/estimated for year 0;
…and then applied for future periods, say every 10 years;
Using projected planning data (GDP, population,
employment, etc. allocated over time and space);
..and some assumptions about changes in the technologies,
networks, competitors, economic regeneration, etc.
19
ERRORS IN FORECASTS
Sources of errors in forecasts
The model. It is simplified, ignores some behavioural
responses, network is not perfect, assumes constant
parameters, etc. It is likely to be even more uncertain as time
progresses
Base year data: small sample, on certain days, blended with
distribution/destination choice models to produce semi-
synthetic matrices: declining influence
Future year data: uncertainty grows over time horizon; gets
more uncertain the more granular (disaggregated) it is:
increasing influence
Scenario uncertainty: the elements we do not influence and
cannot fully predict: competitors in a PPP, success of land
use planning and control, policy changes, disruptive
technologies: increases over time
20
FORECASTING ERRORS
Sources of errors in forecasts
No onal sources of uncertainty in forecas ng
45%
40%
35%
30%
No onal Uncertainty Index
Scenario Uncertainty
25%
20%
15% Future data
Base Year
10% Data
5% Uncertainty due to model quality
0%
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Years a er forecast
21
BETTER DATA AND MODEL
With improved Base Year Data and Model
No onal sources of uncertainty a er improvements in data and model
45%
40%
35%
30%
No onal Uncertainty
25% Scenario uncertainty
20%
15%
Future data uncertainty
10%
Be er model
5%
0%
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Years a er original forecast
Be er model Be er data Fut. data a er Scenario a er Original model Data Future data Scenario
22
IMPACT OF DATA REQUIREMENTS
Model with more disaggregated data requirements
No onal sources of uncertainty a er improved model requiring more data
50%
45%
40%
35%
No onal Uncertainty
Scenario uncertainty
30%
25%
20%
Greater future data requirements
15%
10%
5% Be er model
0%
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Years a er original forecast
Be er model Be er data Fut. data a er Scenario a er Original model Data Future data Scenario
23
MODELS AND EQUILIBRIUM
A model is always a simplification of reality
Some simplifications make our life easier, others are essential
Take Route Choice and Assignment
We model it as a steady state phenomenon: all vehicles clear
the network during the modelled (peak) period.
The software finds a solution with an iterative process until
“equilibrium” is reached:
Wardrop’s Equilibrium:
Under equilibrium conditions traffic arranges itself in congested
networks in such a way that no individual trip maker can reduce
his path costs by switching routes
Certain conditions with respect of the relationship between
flow and travel time must be met to achieve this equilibrium
24
Issues
• If equilibrium of routes and costs is not reached the flows
(and costs) depend on when do we stop the iterative process
• This may affect our choice of scheme
• However, reality is never in “equilibrium”
• Noise in traffic (empty taxis, delivery vehicles, cruisers, lost)
• Noise in the network: stalled vehicles, rain, poor light, roadworks
• Day-to-day variability in the trip matrices
• Moreover, the better we represent delays at junctions, the
more difficult (and sometimes impossible) it is to ensure
“equilibrium”.
25
EQUILIBRIUM
Equilibrium Activity System
Networks and
services
There is a risk of oscillations
in the complete model for
future runs:
1. A lot of congestion, shift Trip Generation/Attraction,
to PT trip frequency choice
2. Less congestion, back to
car
Destination choice
3. A lot of congestion, shift
to PT.
4. Etc….
Mode choice
Time of departure
choice
Route choice and
assignment
Travel costs, especially
Flows by link and mode
time
26
Ten messages to take with you
10 MESSAGES
6. Do results make sense? Check
1. You may outsource the Model model results by other means (e.g.
but not the Planning/Decision back of envelope calculations)
Making
7. The model will be audited: good
2. Engage in Model Specification, documentation & transparency
Develop/retain capacity to essential; better if peer-reviewed
interpret results, protect and throughout
transfer acquired knowledge
8. Model results will be used in a
3. In specifying the model focus wider context and by other
on What Matters; do not aim disciplines; avoid intellectual
for the academic ideal imperialism
4. Decide how will you deal with 9. Is the model WebTAG compliant?
the inevitable uncertainties If not, does it matter and how
5. What assumptions were made does it impact its reliability?
in the model build? What 10. Has the model been validated
assumptions are made about through alternative techniques to
the future in the application? WebTAG, e.g. benchmarking?
27
ROUND UP
Round up
a. Models can be very useful tools to support decision making
b. Resources devoted to analysis should be aligned to the cost
of poor decisions
c. Government guidance often drive model specification
d. Models and their application, however validated, are not
error free
e. There are significant risks in over-specifying a model: delays,
spurious accuracy and excessive faith
f. But they can be very useful to compare solutions on a
common basis and eventually under alternative scenarios
g. The interpretation of model results requires understanding
what they include and what they leave out
h. …and judgment based on experience
28
THANK YOU
29
