CIVL2700 · Transport Systems
Travel Demand Forecasting
Travel demand forecasting is the Week-4 backbone of CIVL2700 Transport Systems at the University of Sydney: the trip-based framework that turns land use and travel behaviour into the vehicle flows on every link. It chains trip generation (how many trips, by regression or category analysis), mode/destination choice (a multinomial-logit utility model with the IIA property), and route choice (user equilibrium on BPR links) — the classic four-step model collapsed into this unit's three-step model. Because link travel times feed back into the earlier choices, the sequence is solved as a system that equilibrates.
What this chapter covers
- 01The four-step model (generation, distribution, modal split, assignment) vs the course's three-step model
- 02Trip-based vs activity-based demand models; travellers' temporal, destination, modal and spatial decisions
- 03Trip generation by linear regression: T = b0 + b1·Z1 + … + bk·Zk, fitted by least squares
- 04Trip purposes (work, shopping, social/recreational) and scaling a per-household rate to a zone
- 05Category (cross-classification) analysis: empirical cell rates × households summed over cells
- 06Comparing generation models by absolute / percentage error against observed counts
- 07Mode/destination choice: multinomial logit P(i) = e^Vi / ∑ e^Vj, with linear-in-attributes utility
- 08The IIA property: P(i)/P(j) = e^(Vi − Vj) is independent of other alternatives
- 09Route choice: user equilibrium and the BPR volume-delay function t = tf[1 + a(x/c)^β]
- 10System equilibration: iterating the steps until utility travel times match assignment travel times
Multinomial-logit mode choice, plus an IIA add-a-mode test
- +1Utilities. V(car) = 1.8 − 0.25(9) − 0.02(20) = 1.8 − 2.25 − 0.40 = −0.85. V(train) = −0.25(4) − 0.015(35) = −1.00 − 0.525 = −1.525. Time and cost enter negatively, so higher time/cost lowers utility.
- +1(a) Shares and demand. e^(−0.85) = 0.4274, e^(−1.525) = 0.2176; sum = 0.6450. P(car) = 0.4274/0.6450 = 0.663, P(train) = 0.337. Demand = 0.663 × 4,000 = 2,650 car and 1,350 train (shares sum to 1).
- +1(b) Add the bus with V(bus) = V(train) = −1.525. New sum = 0.4274 + 0.2176 + 0.2176 = 0.8626. P(car) = 0.4274/0.8626 = 0.495; P(train) = P(bus) = 0.2176/0.8626 = 0.252. Demand = 1,982 car, 1,009 train, 1,009 bus.
- +1(c) Odds ratio. Before: P(car)/P(train) = 0.663/0.337 = 1.96. After: 0.495/0.252 = 1.96 — unchanged, equal to e^(−0.85 − (−1.525)) = e^0.675. The car:train odds depend only on those two utilities: this is the Independence of Irrelevant Alternatives (IIA) property.
Key terms
- Four-step model
- The classic trip-based demand model (1950s): trip generation → trip distribution → modal split → route assignment. This unit uses a three-step version that merges distribution and modal split into one joint mode/destination choice step.
- Trip generation
- Step 1: predicting the number of trips (by purpose — work, shopping, social/recreational) a household or zone produces in a period, usually the peak hour, as a function of socio-economic and land-use characteristics.
- Trip-generation regression
- A linear model T_i = b0 + b1·Z1i + … + bk·Zki, where Z are household/zone attributes and the b coefficients are fitted by least-squares regression on household-survey data. Each Z must enter in its coefficient's units (e.g. income in $'000s, retail in hundreds).
- Category (cross-classification) analysis
- A trip-generation method that sorts households into cells defined by category variables (e.g. income stratum by number of workers) and applies the empirical mean trip rate observed in each cell; the zone total sums cell rate times households per cell — no coefficients are fitted.
- Multinomial logit (MNL)
- The mode/destination choice model P(i) = e^Vi / ∑_j e^Vj over the choice set, with utility V linear in attributes (time, cost, floor space). Predicted demand on an alternative equals its probability times the total generated trips.
- IIA (Independence of Irrelevant Alternatives)
- A property of the logit: the odds between two alternatives, P(i)/P(j) = e^(Vi − Vj), depend only on those two utilities and are unchanged by adding or removing other options. It makes the model tractable but fails for close substitutes (the red-bus/blue-bus problem).
- User equilibrium (UE)
- Wardrop's first principle used in route choice: every used route between an origin-destination pair carries equal, minimal travel time, so no traveller can cut their time by switching route. Link times use the BPR volume-delay function t = tf[1 + a(x/c)^β].
- System equilibration
- The feedback loop that closes the model: the travel times used in the mode/destination utilities are the outputs of route assignment, so the steps are iterated until the times used and the times produced converge. That fixed point is the forecast.
Travel Demand Forecasting FAQ
What is the difference between the four-step and three-step travel demand models?
Both are trip-based. The four-step model keeps trip distribution (which destination zone) and modal split (which mode) as two separate stages, giving generation → distribution → modal split → assignment. CIVL2700's three-step model estimates distribution and modal split together as one joint mode/destination choice, giving generation → mode/destination choice → route choice. The modern alternative to both is the activity-based model, which simulates a person's whole daily activity schedule instead of independent trips.
Why does adding a new mode change the shares but not the odds ratio in a logit model?
Because logit probabilities are P(i) = e^Vi / ∑ e^Vj, the ratio for any two alternatives is P(i)/P(j) = e^(Vi − Vj) — it contains only those two utilities. Adding an alternative enlarges the denominator, so every absolute share drops, but the pairwise ratio is untouched. That is the Independence of Irrelevant Alternatives (IIA) property, and it is also the model's limitation for options that are close substitutes.
Can AI help me with travel demand forecasting in CIVL2700?
Yes — a study assistant like Sia can explain the four-step and three-step framework step by step, check that you have entered each regression variable in the right units, walk through a multinomial-logit share and demand calculation, and test your understanding of IIA and user equilibrium with practice questions. It supports your own working and revision, explaining each step; it never hands over answers, completes your In-class Test or exam for you, or guarantees a grade — always confirm methods and assessment details against Canvas and your lecturer.
Exam move
Carry the three steps in order and know what each produces: generation gives trip counts, mode/destination gives an O-D matrix by mode, assignment gives link flows. For generation, write every substitution with its units (income in $'000s, retail in hundreds) and read the sign of each coefficient before you trust the answer — more local retail lowers outbound shopping trips, and remember to scale the per-household rate by the number of households. For mode/destination, always exponentiate-sum-divide rather than quoting utilities as probabilities, keep time and cost negative, and be ready to demonstrate IIA by adding a mode of equal utility and showing the odds ratio holds. For route choice, state user equilibrium and the BPR form, and note that one pass is not the answer — the model iterates until the utility travel times match the assignment travel times. Practise the tutorial-style 'which model is more accurate?' questions by comparing each model's error against the observed count at every scenario given. Week-4 content is examined in the supervised In-class Test 1 (Weeks 1-4) and in the final exam; confirm all dates and whether the exam is open- or closed-book on Canvas.
Working through Travel Demand Forecasting in CIVL2700? Sia is AskSia’s AI Engineering tutor — ask any CIVL2700 Travel Demand Forecasting question and get a clear, step-by-step explanation grounded in how CIVL2700 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
