CIVL2700/9700

Transport Systems
Topic 4: Travel Demand Forecasting

Andres Fielbaum

1
 What did we learn next week?
➢ Discrete Choice: How to model when people need to to choose among a discrete set of
alternatives?

➢ Particularly important in transport: Mode choice!

➢ But mode choice does not come alone… We first decide whether and where to travel.

➢ This week: Travel demand models.

CIVL2700: Transport Systems - Week 4 2

 Relevance to Real-world problems
➢ What are the transport implications of building new infrastructure (e.g. WestConnex)?

➢ How congested would Sydney roads be in 2029? What about in 2034?

➢ Which areas of Sydney need more transport investment?

➢ How wide should a new tunnel be if it has a planned lifespan of 50 years?

CIVL2700: Transport Systems - Week 4 3

Agenda
• Travel Demand Forecasting
• Traveller Decision Modelling
• Traveller Decisions and the Chicken and Egg Problem
• General 4-Step Framework
• Trip Generation and Attraction
• Trip Distribution
• Mode/Destination Choice
• Route Choice
• Summary

CIVL2700: Transport Systems - Week 4 4

Transport Demand Modelling
Previously on CIVL2700: (week 1)

CIVL2700: Transport Systems - Week 4 5

 Introduction
• Travel demand changes over time
o Economical forces
o Major modifications of infrastructure
• Two interrelated elements must be considered in forecasting travel demand:
o Overall regional traffic growth/decline (long-term)
▪ Primarily influenced by economic and social factors
▪ Regions with increased business activity will also lead to increases in shopping
centers and residential communities
▪ Social characteristics of a particular population influence attitudes toward
transportation modes, or even trip-making in general

o Traffic diversion (short-term)

CIVL2700: Transport Systems - Week 4 6

 Introduction
• Travel demand changes over time
o Economical forces
o Major modifications of infrastructure
• Two interrelated elements must be considered in forecasting travel demand:
o Overall regional traffic growth/decline (long-term)

o Traffic diversion (short-term)

▪ With new facilities are built or modification of existing facilities, traffic will divert from
more congested routes to now less congested routes and may also change departure
time.
▪ Significant changes to any part of the network results in changes to that part as well as
many other parts of the network.
▪ Thus, the transportation network must be viewed and analyzed as a system.
CIVL2700: Transport Systems - Week 4 7
 Introduction
• Travel demand forecasting is a complex problem, as it requires
o Accurate regional economic forecasts
o Accurate forecasts of roadway users’ social and behavioral attitudes regarding trip-
oriented decisions

• Both are very difficult to do

• Forecasts of travel demand volumes should be based upon some understanding of

traveler decisions (behavioral)
o Quantity
o Spatial distribution (i.e. route selected)
o Temporal distribution (i.e. time of day)

CIVL2700: Transport Systems - Week 4 8

 Traveller Decisions
Travellers can be viewed as making four distinct but interrelated decisions:
1. Temporal (time) decisions
2. Destination decisions
3. Modal decisions
4. Spatial (route) decisions
• Temporal decisions include the decision whether to travel and when to travel
• Destination decisions concern the selection of a specific destination (e.g. shopping center)
• Modal decision relates to how the trip will be made (e.g. auto, bus, bicycle)
• Spatial decision concerns which route will be used to travel between origin and chosen
destination
CIVL2700: Transport Systems - Week 4 9
 Traveller Decisions - The Chicken and Egg
• A factor such as travel time (which is itself a
function of several factors) will influence a
traveller’s decision to use a particular mode
and route.

• The traveller’s decision (when aggregated

with all other travellers) will in turn affect a
factor (e.g. traffic volume) that is used to
base a travel decision on.

• This interdependence adds even more

complexity to the traffic forecasting problem.

CIVL2700: Transport Systems - Week 4 10

Traveller Decisions - The Chicken and Egg

Land Use

Transportation Transportation
infrastructure Demand

CIVL2700: Transport Systems - Week 4 11

 Traveller Decisions - The Chicken and Egg
• Last week we said that a User’s Mode choice depends on the travel time of the
different modes, but doesn’t this travel time depend on route choice?

• But if we try to solve route choice first, doesn’t that depend on us knowing the
number of people that are planning to drive?

• This is the chicken and egg problem

• How do we solve this?

- Ideally, trying to find the equilibrium
- In practice, this might be unfeasible (and do cities reach an equilibrium?). So
frequently, doing lots of iterations and analyse the evolution.

• In this course we will typically only do the first iteration 12

Framework (3-step)
• Complex problem, hence some simplifications must be made to make the problem
tractable.
• Viewing traveller decisions as a sequence of three distinct decisions:
1. Trip generation
2. Mode/Destination Choice (last week).
3. Route Choice

Before we said 4 steps.

What is the difference?

CIVL2700: Transport Systems - Week 4 13

 Trip Generation
• Objective: to develop an expression that
predicts exactly when a trip is to be
made.

• This is indeed an extremely difficult and

complex task

• Reasons: wide variety of trip types and

activities undertaken by a traveller in a
sample day

CIVL2700: Transport Systems - Week 4 14

 Trip Generation Simplifications
• Aggregation of Decision-Making Units: Grouping relatively homogenous units
together—in this case, travellers living in the same household.
These travellers share social and economic characteristics, and their trip-making
activities are often intertwined.
• Segmentation of Trips by Type (Purpose): Different types of trips are more likely to be
taken during certain times of day. And some trip types are more likely to be made
multiple times during the day than other trip types. To account for this, three trip types
are used:
• work trips
• shopping trips
• social/recreational trips
• Temporal Aggregation: Trying to predict the overall number of trips taken during some
time period (e.g., peak hour), rather than trying to predict when an individual traveller
will make a trip. (We will extend this at the end of this week).

CIVL2700: Transport Systems - Week 4 15

Example of Segmentation of Trips by Type

CIVL2700: Transport Systems - Week 4 16

 Trip Generation/Attraction Models
• Trip generation models typically assume a linear form and a period (e.g. morning peak)
• The number of vehicle-based trips is generally a function of various socio-economic
(e.g. age, income) characteristics

𝑇𝑖 = 𝑏0 + 𝑏1 𝑧1𝑖 + 𝑏2 𝑧2𝑖 + ⋯ + 𝑏𝑘 𝑧𝑘𝑖

where:
𝑇𝑖 = number of vehicle-based trips of a given type (e.g. work or shopping) in some specified time period
made by household i
𝑧𝑘𝑖 = characteristic k 𝑏𝑘 = coefficient estimated from traveler survey data and corresponding to
characteristic k

• The coefficients (b’s) are typically estimated by the method of least squares regression
using data collected from traveler surveys.
CIVL2700: Transport Systems - Week 4 17
 Trip Generation - Attributes
𝑇𝑖 = 𝑏0 + 𝑏1 𝑧1𝑖 + 𝑏2 𝑧2𝑖 + ⋯ + 𝑏𝑘 𝑧𝑘𝑖

Which households’ attributes are relevant?

• Number of members (and their age!)

• Income
• Characteristics of the neighborhood
o Amenities nearby.
o Connectivity by public transport/car.
• Do they have a car?
• Do they have drivers’ license?
• Do they have other vehicles (bike, moto, etc)?

CIVL2700: Transport Systems - Week 4 18

 Trip Attraction Models
Think-Pair-Share: How would a similar model look when modelling trips attraction?

Answer: Roughly the same, but now we aggregate by zones the characteristics are those
defining the attractibility of the zone:

• m^2 commerce
• Number of parking spaces
• Number of jobs
• Number of students
• …

In this class, we model destination and mode jointly, using Discrete Choice models (last
week!).
CIVL2700: Transport Systems - Week 4 19
 Example 1
A simple linear regression model is estimated for shopping-trip generation (𝑇𝑖 ) during a
shopping-trip peak hour.

𝑇𝑖 = 0.12 + 0.09 (household size) + 0.011 (household income in thousands of dollars) –

0.15 (number of retail shops in hundreds)

A particular household has 6 members and an annual income of $50,000. They currently
live in a neighborhood with 450 retails shops but are moving to a new home in a
neighborhood with 150 retails shops. Calculate the predicted number of vehicle-based
peak-hour shopping trips the household makes before and after the move.

CIVL2700: Transport Systems - Week 4 20

 Example 2 - Attributes
A neighbourhood has 205 retail shops and 700 households that can be categorized into
four types, with each type having characteristics as follows:

There are 100 type 1, 200 type 2, 350 type 3, and 50 type 4 households. Assuming that
shopping, social/recreational, and work vehicle-based trips all peak at the same time
determine the total number of peak-hour trips (work, shopping, social/recreational)
using the below generation models.
CIVL2700: Transport Systems - Week 4 21
 Example 2 – Trip Generation Models
A neighbourhood has 205 retail shops and 700 households that can be categorized into
four types, with each type having characteristics as follows:

There are 100 type 1, 200 type 2, 350 type 3, and 50 type 4 households.

Number of peak-hour vehicle-based shopping trips per household

= 0.12 + 0.09 (household size) + 0.011 (household income in thousands of dollars) – 0.15
(number of retail shops in hundreds)

Number of peak-hour vehicle-based social/recreational trips per household

= 0.04 + 0.018(household size) + 0.009(annual household income in thousands of dollars)
+ 0.16(number of nonworking household members)
CIVL2700: Transport Systems - Week 4 22
 Example 3 – Model Accuracy
Consider the simple linear regression model for shopping-trip generation (𝑇𝑖 ) that was
used in example 1: 𝑇𝑖 = 0.12 + 0.09 (household size) + 0.011 (household income in
thousands of dollars) – 0.15 (number of retail shops in hundreds)

Suppose the suburb has 300 retail shops and 2000 households, and that the average
household has 4 members and an annual income of $80,000:
a) Calculate the predicted number of shopping trips
b) If the actual measured number of shopping trips is 1500, is this model accurate?
c) Suppose there was another model that was just Ti = 0.75 Which model is more
accurate?
d) What if the number of households changed to 3000 and the actual number of trips
changed to 2500?

CIVL2700: Transport Systems - Week 4 23

Mode & Destination Choice – Just like last week!
• Depending on the Trip Type, Combining Mode & Destination makes more or less sense:
o Work (Mode Choice Only)
𝑒 𝜇𝑈𝑖𝑛
o Shopping (Mode & Destination) 𝑃 𝑖 𝐶𝑛 =
o Social/recreational (Mode & Destination) σ𝑗∈𝐶𝑛 𝑒 𝜇𝑈𝑗𝑛

• Mode & Destination traveller decision making:

o Microeconomic Concept of Utility Maximization (Like Last Week)
o A traveller will select the combination of mode and destination that provides the
most utility

• It can be challenging to create an expression for utility based on both mode and
destination alternatives

CIVL2700: Transport Systems - Week 4 24

 Mode & Destination Choice Example
Consider a residential area and two shopping centres that are possible destinations. From
7:00 to 8:00 P.M. on Friday night, 900 vehicle-based shopping trips leave the residential
area for the two shopping centres. A joint shopping-trip mode-destination choice logit
model (choice of either auto or bus) is estimated, giving the following coefficients:

Constant

0.12 0.12
m2)

If shopping centre 2 has 40,000 m2 of commercial floor space and shopping centre 1 has
25,000 m2, determine the number of Friday night shopping trips by destination and
mode.
CIVL2700: Transport Systems - Week 4 25
 Think – pair - share
What novel sources of data can give more information about how people is travelling?

How do these data relate to the methods we discussed this week?

CIVL2700: Transport Systems - Week 4 26

 Origin-Destination matrix
We completed the first two steps. What do we have?
R: Given a mode M and period t, we know the number of people going from each
origin to each destination. This information is typically presented in an OD-Matrix

SOURCE: DR. JEAN-PAUL RODRIGUE, DEPT. OF GLOBAL STUDIES & GEOGRAPHY , HOFSTRA
UNIVERSITY, NEW YORK, USA.
CIVL2700: Transport Systems - Week 4 27
 Origin-Destination matrix
Formally, and OD matrix is a square matrix, where rows and columns represent the
potential origins and destinations. The element (i,j) of the matrix represents the
number of trips occurring from i to j (for the given mode and period).

There are other alternatives to estimate an OD matrix: direct observation, the

four-steps model (see assignment 1), historical data, etc. A virtue of the 3-steps
model is that regardless of how you obtained the OD Matrix, the next step is the
same.

And what is the next step?

CIVL2700: Transport Systems - Week 4 28

 Third step: Route choice
The third step is route choice:

• If the chosen mode is car, we go back to week 2 and the Wardrop Equilibrium!

• If the chosen mode is public transport, this is frequently modelled using logit, but
some sophisticated techniques might be needed because the alternatives are not
independent.

• If the chosen mode is walking or biking, often assuming the shortest path is good
enough.

CIVL2700: Transport Systems - Week 4 29

 How does this model work In Real Life?
State-of-the-practice of travel demand forecasting
1
• Studied geographic region is segmented into nearly
homogeneous areas based on similarities in land use,
socioeconomic conditions, etc: traffic analysis zones (TAZs)
• TAZs are used to determine the origins and destinations of trips
• The choice of the number of TAZs is a trade-off between
accuracy and ease of implementation
• A single point is usually chosen within the TAZ as the assumed
origin/destination point of all TAZ trips: Centroid
2
• The traffic network includes links.
• Links are connected by nodes (e.g. intersections)
• Link’s performance function (to satisfy user equilibrium route
choice)
• The performance function often used in practice (originally
developed by the U.S. Bureau of Public Roads) is:
CIVL2700: Transport Systems - Week 4 30
 How does this model work IRL?
State-of-the-practice of travel demand forecasting
1
• traffic analysis zones (TAZs)
2
• Link’s performance function (to satisfy user equilibrium), BPR
function is:

CIVL2700: Transport Systems - Week 4 31

 How does this model work IRL?
State-of-the-practice of travel demand forecasting
1
• traffic analysis zones (TAZs) EXOGENOUS
2
• Link’s performance function (to satisfy user equilibrium), BPR
function. EXOGENOUS
3
• Trip generation models and mode/destination choice models
are then used to determine the number of vehicles traveling
between all TAZs during a specified time period (usually the
peak hour).
• The resulting vehicle trips are used to create an origin-
destination matrix (O-D) that gives the total number of vehicle
trips going between each TAZ during the analysis period.
4
• With the vehicle O-D trip matrix, traffic flows on each link are
determined, usually by assuming that user-equilibrium holds.
CIVL2700: Transport Systems - Week 4 32
 Summary
1. Trip generation Alternative
2. Mode/Destination Choice 1. Trip generation and attraction
3. Route Choice 2. Trip Distribution -> How?
3. Mode Choice
4. Route Choice

CIVL2700: Transport Systems - Week 4 33

Extension: Time departure choice
Let us now study a related trip-generation problem where we will combine different
techniques we have learned before.

Consider one unit of passengers (e.g. 1000 pax) that want to travel by car from O to
D. They all start working at D at Q (e.g. Q=8 AM). If a person departs at time t, the
time need to go from O to D is given by

𝑐 𝑡 = 𝐴 + 𝐵𝑥𝑡

Where 𝑥𝑡 is the number of people departing at time t. We want to model and

predict at which time users will choose to depart.

CIVL2700: Transport Systems - Week 4 34

Extension: Time departure choice
𝑐 𝑡 = 𝐴 + 𝐵𝑥𝑡
To do this, we will use a deterministic utility model. Assume that the utility of a person is given
by

𝑈 𝑡 = −𝑐 𝑡 − 𝛼𝐸𝑡 − 𝛽𝐿𝑡

Where 𝐸𝑡 is how early the person arrived, and 𝐿𝑡 how late (note that they cannot be positive at
the same time).

1) Which do you expect to be greater, 𝛼 or 𝛽?

R: 𝛽. Actually, we need 𝛼 < 1 (i.e., one minute early in the office is better than one minute in
the car

2) What would be the equivalent to the Wardrop principles in this case?

R: All the t’s that have passengers should have the same utility, and greater than in the rest.35
Extension: Time departure choice
𝑐 𝑡 = 𝐴 + 𝐵𝑥𝑡 , 𝑈 𝑡 = −𝑐 𝑡 − 𝛼𝐸𝑡 − 𝛽𝐿𝑡
The utility of someone arriving late:

𝑈 𝑡 = −𝐴 − 𝐵𝑥𝑡 − 𝛼 𝑄 − 𝑡 + 𝐴 + 𝐵𝑥𝑡 = 𝛼𝑡 − 𝐴 − 𝛼𝑄 + 𝛼𝐴 + 𝑥𝑡 (𝛼𝐵 − 𝐵)

This is a constant value. If we denote it by K, we obtain:

𝛼𝑡 − 𝐴 − 𝛼𝑄 + 𝛼𝐴
𝑥𝑡 = −𝐾
𝐵 − 𝛼𝐵

Note that 𝑥𝑡 increases linearly with t, up to the time where the passenger arrives
just in time (with maximum congestion).

36
Extension: Time departure choice
𝑐 𝑡 = 𝐴 + 𝐵𝑥𝑡 , 𝑈 𝑡 = −𝑐 𝑡 − 𝛼𝐸𝑡 − 𝛽𝐿𝑡
The utility of someone arriving late:

𝑈 𝑡 = −𝐴 − 𝐵𝑥𝑡 − 𝛽 𝑡 + 𝐴 + 𝐵𝑥𝑡 − 𝑄
= −𝛽𝑡 − 𝐴 − 𝛽𝐴 + 𝛽𝑄 − 𝑥𝑡 (𝐵 + 𝛽𝐵)
So:

−𝛽𝑡 − 𝐴 − 𝛽𝐴 + 𝛽𝑄
𝑥𝑡 = −𝐾
𝐵 + 𝛽𝐵

Now 𝑥𝑡 decreases linearly with t, from the time where the passenger arrives just
in time (with maximum congestion).
37
Extension: Time departure choice
As a result, the graph showing how
many people chooses each possible
departure time is triangular.
Comments:

• This is because everything is linear.

• How do we calculate the min and
max departure times?
• Can you plot the resulting arrival
times?

38