Travel Demand Forecasting

Forecast traffic by asking where trips start, where they go, and how they get there.

Before this Lesson

A developer wants to build 400 homes on the far side of the river, and the city has to decide now whether the two-lane bridge into downtown can take the extra traffic. There's nothing to count. The homes don't exist yet, and neither do the people who would drive out of them.

So the traffic gets built on paper instead, one question at a time, starting with the homes and ending with cars on the bridge.

Four questions, in order

A forecast splits the region into zones, such as neighborhoods, and asks four questions about the trips between them. Each question needs the answer to the one before it.

  1. Trip generation: how many trips? This step counts the trips that start or end in each zone.
  2. Trip distribution: where do they go? It fills in the trips between every pair of zones. The gravity model does this.
  3. Mode choice: how do they get there? It splits each zone pair's trips among car, bus, rail, and so on. The logit model does this.
  4. Traffic assignment: which route? It loads the trips onto paths through the road and transit networks and totals the trips on every link. This is where the bridge finally gets its number.
1Trip generation How many trips? trip rates, regression 2Trip distribution Where do they go? gravity model 3Mode choice By what mode? logit model 4Traffic assignment Which route? loading trips onto routes trips produced and attracted by each zone trips between each pair of zones zone-to-zone trips by mode volume on every road and transit link
Each step answers one question and hands its output to the next

The order isn't arbitrary. You can't pick a mode until you know where the trip is going, because the bus might run downtown but not to the office park. You can't pick a route until you know whether it's a car trip or a bus trip. The Handbook doesn't describe the four steps, but it prints the tools for the middle two, the gravity and logit models, under Traffic Flow Relationships. For assignment, the exam asks only what the step does.

Checkpoint
1 of 2

A method takes the trips leaving each zone and splits them among destination zones, sending more to zones with many jobs and fewer to zones that are far away. Which step does it perform?

A.
Trip generation
B.
Mode choice
C.
Traffic assignment
D.
Trip distribution

Trip generation: counting trips that don't exist yet

Trips come from activities. People leave home to work, shop, and go to school, so a zone's homes, jobs, and stores tell you how many trips it will see. Planners label a trip's two ends by role: homes produce trips, and workplaces, stores, and schools attract them. For a trip with one end at home, the production is always the home end, so the drive home from work still counts as produced by the home zone. That's a convention, and a useful one: a zone's productions follow its households, and its attractions follow its jobs and floor space.

The simplest forecast is a trip rate, in trips per dwelling unit for housing or trips per 1,000 ft² of floor area for offices and stores. The subdivision's 400 homes at 9.5 daily trips per home make 3,800 trips a day. For floor area, divide the square footage by 1,000 before multiplying, because that's the unit the rate counts in.

The other common form is a regression equation fitted to travel surveys, which predicts trips from variables such as households, persons, vehicles, or employees. Before plugging in, read what the equation predicts: trips per household or trips for the whole zone. A per-household answer still has to be multiplied by the number of households. The Handbook prints no trip rates or trip-generation equations, so a problem always supplies them.

Checkpoint
1 of 2

A development has 180 apartments, at 6.6 daily trips per dwelling unit, and a 24,000 ft² retail center, at 37.8 daily trips per 1,000 ft². How many daily trips does it generate?

trips/day

Trip distribution: the gravity model

Now the trips need destinations. Models distribute each trip purpose separately, so take the 1,050 work trips among the subdivision's 3,800 daily trips. Where do they go?

Two things pull a trip toward a destination: how much is there, and how close it is. A huge office park draws workers, but fewer of them if it's across the county. That's the reason for the name. Newton's pull grows with mass and fades with distance, and this one grows with a zone's attractions and fades with travel time.

The Handbook's Gravity Model gives each destination a weight and hands out the zone's trips in proportion to it:

Tij=Pi[AjFijKij∑jAjFijKij]T_{ij} = P_i \left[ \frac{A_j F_{ij} K_{ij}}{\sum_j A_j F_{ij} K_{ij}} \right]

Here TijT_{ij} is the trips produced in zone ii and attracted to zone jj, PiP_i is the total produced in zone ii, and AjA_j is the trips attracted to zone jj. The friction factor FijF_{ij} measures how easy the trip is. The Handbook defines it only as an inverse function of travel time, so it falls as the trip gets longer. Its exact form is calibrated from local data, so a problem states it, often as F=1/t2F = 1/t^2 or as a list of values. KijK_{ij} is a socioeconomic adjustment for one zone pair, and it's 1 unless a problem gives it.

Look at the bracket. The top is one destination's weight and the bottom is the sum of all of them, so the bracket is that destination's share, and the shares add to 1. Whatever the weights, the trips from zone ii add back up to PiP_i, which gives you a free check on the arithmetic.

t = 5 min t = 10 min t = 20 min Zone 1 (subdivision) P1 = 1,050 Zone 4 (hospital) A4 = 1,500 Zone 2 (downtown) A2 = 3,000 Zone 3 (office park) A3 = 6,000
The subdivision's work trips and three places they could go
Each destination circle's area is proportional to the zone's attractions, and its distance from zone 1 is proportional to the travel time.
Worked Example Distributing the subdivision's work trips
1 of 3

Zone 1, the subdivision, produces 1,050 work trips a day. Downtown (zone 2) attracts 3,000 trips at a travel time of 10 min, the office park (zone 3) attracts 6,000 at 20 min, and the hospital (zone 4) attracts 1,500 at 5 min. With F=1/t2F = 1/t^2 and K=1K = 1, how many trips go to each zone?

The fixed total has a consequence worth slowing down for. Every destination's trips come out of the same PiP_i, so the destinations compete for them, and a change to any one weight changes the total that every share is divided by.

Checkpoint
1 of 2

Back to the subdivision's 1,050 work trips. A new tower doubles downtown's attractions to 6,000, and nothing else changes. Which of these are true?

A.
Trips to downtown double, to 600
B.
Trips to downtown rise, but by less than double
C.
Trips to the office park and the hospital fall
D.
Zone 1 produces more trips in total

Mode choice: the logit model

Of the work trips headed downtown, how many go by car and how many by bus? Each traveler weighs time against money, and the logit model turns that weighing into shares.

It starts with a utility for each mode, a score of how attractive it is. The Handbook's Logit Models section builds it as a sum over the mode's attributes:

Ux=∑iaiXiU_x = \sum_i a_i X_i

Each XiX_i is an attribute value, such as the travel time in minutes or the cost in dollars, and aia_i is its coefficient. Time and cost are things travelers would rather not spend, so the Handbook notes that the coefficients are negative. That makes most utilities negative too, which looks odd until you read each one as minus a penalty. The less negative mode is the more attractive one, so a utility of −1 beats −2.

To turn utilities into shares, the model gives each mode a weight of eUe^{U} and hands out the trips in proportion. Why eUe^{U} and not just UU? Try it with utilities of −1 and −2. As weights, they give the better mode 1/3 of the trips and the worse mode 2/3, which is backward. The exponential is positive for any utility, and it shrinks as the utility gets more negative, so the better mode always carries more weight.

It does something subtler too. The ratio of two weights is eU1/eU2=eU1−U2e^{U_1}/e^{U_2} = e^{U_1 - U_2}, so only the gap between utilities matters, never their level. Each point of gap multiplies the weight ratio by e≈2.72e \approx 2.72, wherever the two modes sit. The Handbook's probability of choosing mode xx out of nn modes is:

P(x)=eUx∑k=1neUkP(x) = \frac{e^{U_x}}{\sum_{k=1}^{n} e^{U_k}}

The Handbook writes the sum with xx as its index, but either way it runs over every mode, the one you're asking about included. This is the gravity model's machinery again: a fixed total, a weight for each option, and shares that add to 1. Only the weights are different.

−3 −2 −1 0 1 2 3 0 0.25 0.50 0.75 1.00 0.27 0.50 0.73 bus less attractive bus more attractive Ubus − Ucar P(bus)
With two modes, the share depends only on the utility gap
Equal utilities split the trips evenly. A gap of 1 in the bus's favor makes its weight ee times the car's, for a 73% share, and a gap of 1 against it leaves 27%.
Worked Example Car or bus downtown
1 of 3

Of the subdivision's work trips, 300 go downtown. The utility is U=−0.06t−0.30cU = -0.06t - 0.30c, with tt the travel time in minutes and cc the cost in dollars. The car takes 15 min and costs $5.00 with parking. The bus takes 40 min door to door and costs $2.00. How many of the 300 trips go by car?

When one mode changes

Suppose the bus gets faster. Its utility rises and its weight grows, and nothing else changes in the numerators: every other mode's weight is exactly what it was. They're just divided by a bigger total. So every other mode's share shrinks by the same factor. A new mode works the same way, arriving as one more weight in the denominator.

Checkpoint
1 of 2

A city's trips split 60% car, 25% bus, and 15% rail. A bus upgrade raises the bus's utility, and its share climbs to 40%. The car's and rail's utilities don't change. What is the car's new share?

%

Working backward to a target share

Planners often ask the question in reverse: how good would upgraded transit have to be to win 45% of the trips? Lump every other mode's weight into one number, SS, the sum of their eUe^U. Then the target mode's share is P=w/(w+S)P = w/(w + S), where w=eUw = e^U is its own weight. Solve for ww, and undo the exponential with ln⁡\ln:

w=P S1−PU=ln⁡ww = \frac{P\,S}{1 - P} \qquad U = \ln w

The utility coefficients then turn that utility into the minutes or dollars the upgrade has to deliver.

Worked Example A bus that wins 45%
1 of 2

On the subdivision's commute, Ucar=−2.40U_{\text{car}} = -2.40 and Ubus=−3.00U_{\text{bus}} = -3.00, which gives the bus 35% of the trips. What bus utility would capture 45%, and how much faster would the bus trip have to be? The time coefficient is −0.06 per minute.

Recap

  • The four-step model asks how many trips (generation), where they go (distribution), by what mode (mode choice), and by which route (assignment), and each step feeds the next.
  • Trip generation multiplies a rate per dwelling unit or per 1,000 ft² by the size of each land use, or evaluates a regression. Homes produce trips, and jobs and stores attract them.
  • The gravity model shares a zone's productions among destinations in proportion to AjFijKijA_j F_{ij} K_{ij}. The friction factor falls with travel time, and the trips always add up to PiP_i.
  • The logit model shares trips among modes in proportion to eUe^{U}. Utilities are usually negative, the less negative mode is more attractive, and improving one mode shrinks every other mode's share by the same factor.
  • For a target share, lump the other weights into SS, then w=PS/(1−P)w = PS/(1 - P) and U=ln⁡wU = \ln w.