Activity Based Model
SANDAG is in the early stages of transitioning from an enhanced four-step transportation model to an activity-based model (ABM). An ABM simulates individual and household transportation decisions that compose their daily travel itinerary.
Next –Generation Transportation Modeling: The Activity Based Model
SANDAG is in the early stages of transitioning from an enhanced four-step transportation model to an activity-based model (ABM). An ABM simulates individual and household transportation decisions that compose their daily travel itinerary. People travel outside their home for activities such as work, school, shopping, healthcare, and recreation (Figure 1), and the ABM attempts to predict whether, where, when, and how this travel occurs.
The SANDAG ABM strives to be as behaviorally realistic as possible and is based on empirical data collected by SANDAG, Caltrans, and the federal government. The model development has been regularly peer-reviewed by the ABM Advisory Committee, a panel of national experts in the travel demand forecasting field. The SANDAG ABM is innovative, and it builds off of a proven model design that has been tested in practice in the San Francisco, Atlanta, and New York regions. This document briefly describes how the SANDAG ABM works and highlights some of the model’s key assumptions.
Figure 1: Travel is often a necessary part of engaging in desired activities
How Will an ABM Enhance SANDAG's Planning Analysis?
Activity-based models are at the forefront of travel demand modeling technology. These models allow for a more nuanced analysis of complex policies and projects. The powerful analytic capabilities of an ABM are particularly helpful in evaluating Transportation Demand Management (TDM) policies, social equity, carpooling, transit access, parking conditions, tolling and pricing. Because an ABM tracks the characteristics of each person, the model can be used to analyze the travel patterns of a wide range of socioeconomic groups. For example, a household with many members may be more likely to carpool, own multiple vehicles, and share shopping responsibilities.
Figure 2: A very wide range of policies can be analyzed with an ABM
The ABM will allow for more in-depth examination of transportation pricing policies such as tolls and congestion charges. The model accounts for the distinct ways that different groups of people are likely to respond. For example, consider the construction of a new managed lane facility. Some groups (e.g. high-income individuals heading to work or a parent on the way to pick up a child from daycare), may not be discouraged much by the toll and may even be attracted to the road because of the time savings that result. Other groups (e.g.some retirees, students), for whom time may be more flexible, are likely to seek out a route that avoids the toll. A rush-hour-only toll might incentivize certain commuters to leave home a bit earlier to avoid the charge.
As mentioned, the ABM accounts for the demographic and socioeconomic characteristics of travelers. This leads to a number of analytic advantages, including the ability to predict the impact of transit fare changes that apply to specific subgroups of the population, such as low-income and student populations. In fact, the model can be used to study the impacts of transportation policies and projects on a very wide range of demographic slices of the population, a capability that is very important for environmental justice analysis and other equity analysis.
The SANDAG ABM captures detailed traveler responses to high-occupancy vehicle (HOV) facilities and other carpool policies. A significant portion of carpool trips involve members of the same household. SANDAG’s ABM accounts for household decisions about when to share a ride, and the model considers factors like the number of people in a household, whether there are kids, and whether household members’ schedules align in a way that would make carpooling feasible.
Parking is a major aspect of travel by car, and the SANDAG ABM is sensitive to a variety of parking policies. It can predict how travelers respond to parking price changes, an increase in parking supply, and other policies that impact the cost and availability of parking. The ABM can also predict where people park in parking-constrained areas such as downtown San Diego. Depending on the price of parking, the availability of parking, and traveler characteristics, drivers will decide whether to park at their ultimate destination or at a nearby lot.
The future of the San Diego region may involve more flexible work schedules and increasing substitution of communication technology for travel. The SANDAG ABM is suitable for analyzing these emerging trends and the policies that target them. The SANDAG ABM specifically accounts for the worker decision of whether to stay at home or commute to work on a given day. It can also predict how people’s non-work travel patterns change if they work from home. For example, a person may work at non-traditional hours and go shopping in the middle of the day if they are able to telecommute.
The ABM is a forecasting tool that is versatile and realistically responsive to transportation policy changes. The model has been tailored specifically to meet SANDAG’s analytical needs, considering current and future plans, and also taking into account unique sources of travel demand that exist in the San Diego region.
A Methodologically-Advanced Approach to Forecasting Travel Demand
The SANDAG ABM includes a number of methodological strengths. It predicts the travel decisions of San Diego residents at a detailed level, taking into account the way people schedule their day, their behavioral patterns, and the need to cooperate with other household members. When simulating a person’s travel patterns, the ABM takes into consideration a multitude of personal and household attributes like age, income, and gender. The model’s fine temporal and spatial resolution ensures that it is able to capture subtle aspects of travel behavior.
Figure 3 maps out one day’s worth of travel for one typical San Diego resident named Ben, a married father of two. At 7:30a.m., he rolls out of his home driveway with a daughter in the passenger seat. After dropping her off at school at 7:40a.m., he drives to work downtown, pulling into a parking lot a few blocks from his office at 8:20a.m. Ben then walks from his parking spot to the office building. The rest of his workday is spent at the office, with the exception of a quick lunch-time stroll to a nearby restaurant. Ben leaves work at 5:30p.m., walks to his car, and then drives to a pharmacy to fill a prescription, arriving back home at 6:30p.m. Later that evening, Ben and his wife drive to a furniture store to browse couches. They depart together at 7:45p.m. and are back home at nine.
Figure 3: An example of one person’s simulated travel over the course of one day
The description of Ben’s travels presented here is a narrated version of what the ABM could have simulated for a person similar to him. The model’s output is a predicted travel itinerary for each person. Notice that there is a spatial and temporal consistency to Ben’s travel patterns. The ABM ensures that people move from one place to the next in a plausible manner, taking into account when they need to be at important places such as work or school. The 30 minute time intervals used in ABM help ensure temporal consistency among travel activities. Also notice that Ben’s schedule is coordinated with the schedule of other people in his household. He carpools with his daughter in the morning and goes shopping with his wife in the evening. The ABM makes sure that travel involving multiple members of a household fits in each person’s schedule. Finally, notice that Ben’s choice of travel modes is consistent. He goes to work in a car and returns home in a car. For short trips near work, he travels by foot. The ABM ensures that the sequence of modes used by each person is logical. For example, if Ben had parked at a trolley stop to take the trolley to work, the model would always predict that he picks up his car at the end of the day. Consistency of model output is an important aspect of the ABM.
When people travel outside the home to participate in activities, they sometimes stop by other places en route to their ultimate destination. For example, Ben stops by the pharmacy on the way home. In the language of travel demand analysis, this sort of behavior is called ‘trip-chaining’. Trips (segments of travel with a given origin and destination) are strung together to reflect the way people often travel. A chain of trips that starts and ends at home is called a ‘tour’. Ben undertakes two tours- one where the primary destination is his workplace and another where the primary destination is the furniture store.
The socio-demographic characteristics of Ben and his family have a significant influence on his travel behavior. For example, Ben’s occupation helps to determine where he will work. Ben’s status as a parent means that he will often need to escort kids and run errands related to his parental responsibilities. His income influences how willing he is to pay for highway tolls. When simulating a person’s travel patterns, the ABM takes into consideration these attributes and a multitude of other personal and household attributes. Because the ABM tracks the characteristics of each person, the model can be used to analyze the travel behavior of almost any group of people.
In the description of Ben’s travel patterns, very specific times and places are mentioned. One gets a very detailed view of his daily itinerary. The ABM makes predictions at a similar level of detail. In the model, time is represented as thirty-minute intervals, and space is divided into more than 22,000 zones, which is substantially more than the number of zones used by most similarly-sized regions. This fine temporal and spatial resolution is able to capture subtle aspects of travel behavior. The ABM can predict how near or far from his office Ben will park, and it can predict whether he leaves earlier from home if a peak-period toll is instituted. Spatial detail also means that the model offers reliable forecasts of walking and biking because these forms of travel involve short distances and are influenced by neighborhood characteristics.
The Steps in the SANDAG ABM
Because the personal travel decisions of over 3 million San Diego residents are complex, the ABM is comprised of numerous, interacting components (“steps”). The first step is to build a representative population that looks like the real San Diego. Once a representative population is created, the model predicts long-term and medium-term decisions such as a choice of work location and a household’s choice of number of cars to own. Next, each person’s day is scheduled, taking into account the priority of various activities. Once all journeys to and from home have been scheduled, the model predicts specific travel details such as mode and the number of stops to make. The final step of the ABM is traffic assignment.
Figure 4: The sequence of ABM components
The steps will be discussed below, in the order that each step is taken within the SANDAG ABM framework (see Figure 4).
Step 1: Build a representative population that looks like the real San Diego
The first step is to create a ‘synthetic’ population of San Diego County. A synthetic population is a table that has a record for every individual and household, with the individual’s and the household’s characteristics. For example, if there are 40,000 18-year-old males in the region, there would be 40,000 records in the table for males age 18, with each record also having other characteristics such as school enrollment and labor force participation status. Taken as a whole, this synthetic population represents the decision-makers whose travel choices the ABM will simulate in later steps. For each simulation year, a full population is synthesized to match the forecasted socio-economic and housing characteristics of each part of the region at the zonal level. These forecasts, a key ABM input, come from the land use model (PECAS). Synthesis works by replicating a sample of Census records (each containing complete household and individual characteristics) and placing them around the region in such a way that the forecasted characteristics of each zone are matched.
Step 2: Assign a work location to workers and a school location to students
The second step predicts where each individual will go to work or school, if applicable. The model simulates each worker’s choice of work location, taking into account many factors, including ease-of-travel and the number of employees by occupation type in each location. The model also simulates each student’s choice of school, taking into account factors that include the distance from home to school, school enrollment, and district boundaries. The results from this step affect later travel choices significantly because of the prominent role that workplace and school usually play in the itinerary of workers and students.
Step 3: Determine certain mobility characteristics of individuals and households
This step predicts the number of automobiles each household owns, whether each household owns a toll transponder, and whether worker parking costs are employer-reimbursed. The model assigns each household zero cars, one car, two cars, or ‘three or more’ cars, taking into account a number of criteria, including household size, income, and how easy it is to reach destinations from the household’s place of residence. This step sets certain mobility characteristics that influence how people travel.
Step 4: Schedule the day
The fourth step begins by predicting a ‘daily activity pattern’ for each individual. A daily activity pattern is a theme that overarches an individual’s schedule. A ‘mandatory’ pattern means that an individual travels to work and/or school, and then schedules other activities around work/school. An ‘at-home’ pattern means that an individual’s daily schedule involves no travel in the region. A ‘non-mandatory’ pattern means that an individual’s daily schedule involves traveling, but only to destinations other than work or school. An individual’s daily pattern-type is influenced by the pattern-type of other household members. For example, if a child stays home from school, a working parent might be more likely to stay home from work as well.
Once an individual’s daily activity pattern has been selected, the model schedules the tours that he or she will take. Recall that a tour is a journey that begins and ends at home, and it can include stops at other destinations on the way to or from the primary destination. The ABM deals with three main categories of tours: mandatory tours, joint tours, and non-mandatory tours. Mandatory tours have work or school as the primary destination. Joint tours involve out-of-home activities that multiple members of a household partake in together. Non-mandatory tours involve purposes other than work or school that an individual undertakes independent of other members of his or her household. The model schedules each tour type by predicting how many tours of that type there are, who will participate in the tour, where the main destination is, and when to depart and arrive (see Figure 5).
Figure 5: For each tour type, scheduling involves predicting a number of details
(Note to reader: Mode choice is not included in this diagram because it comes in Step 6.)
For individuals assigned a ‘mandatory’ activity pattern, the ABM first assigns the number of work tours and/or school tours they will make. After the number of these mandatory tours has been determined, the model selects the time of departure from and arrival back home for each tour.
After mandatory tours have been scheduled, the time remaining for other tours is calculated. Remaining intervals of time are called “residual time windows”, and other tours can only be scheduled in these open slots (see Figure 6 for an example).
Figure 6: After each tour is scheduled, remaining tours can only be placed in open time windows
In time remaining after mandatory tours are scheduled, the model determines the number of joint tours to be made for each household. Joint tours can only be scheduled in the time windows that overlap between individuals after mandatory activities are accounted for. After the number and purpose of these joint tours has been determined, the model decides which household members will participate in each joint tour and whether the joint tour must involve a combination of children and adults. The model then chooses a specific destination for the tour and the specific times when tour participants will depart from and arrive back home together.
Next, non-mandatory tours are scheduled. For each household, the model decides what other tours need to be made for the purpose of household ‘maintenance’ activities such as shopping. These tours are assigned to specific household members to carry out individually. For the person who is assigned each maintenance tour, the model selects a specific destination and schedules the tour to take place in a time window that mandatory tours and joint tours have left open. Finally, in what time remains, the model decides whether each individual will take non-mandatory ‘discretionary’ tours. These low-priority tours involve activities related to recreation, eating out, and social functions. Discretionary tours can only take place in time windows that remain after all other tours have been scheduled. The model chooses a specific destination and departure/arrival combination for each discretionary tour a person makes.
Step 5: Make tour and trip-level decisions
The ABM then selects more detailed characteristics of each tour for every traveler. This step fills in travel details after the major aspects of the day have been scheduled. Tour characteristics that need to be determined include: primary mode of the tour, how many times to stop, where to stop, and when to depart from each stop to continue the tour. Available modes are shown in Figure 7. After tour characteristics are set, the model determines the mode of each trip (conditional upon tour mode). Recall that trips are segments of tours that have a given origin and destination. If the trip mode involves an automobile and the destination is a parking-constrained area, then the model chooses a parking location for the traveler at the trip destination.
Figure 7: Tour and Trip Modes
Step 6: Aggregating and assigning trips
The previous step provided travel details for each person down to the trip level. In this final step, the model sums all trips taken by individuals in San Diego County along with trips generated by other models that represent special categories of travel within the region that are not covered by the ABM. The model then arranges aggregated trip data in a zone-to-zone matrix and assigns trips to the transportation network in the same fashion as in the four-step model (see four-step documentation for more information). Output is used to calculate measures of performance and impact (see Figure 8).
Figure 8: Summing up trips, assigning them a path, then measuring system performance/impacts
Model integration is an important aspect of the SANDAG system of regional models. The ABM needs to be integrated with PECAS and various travel models. The entire model system will be run iteratively.
The ABM predicts the travel patterns of San Diego County residents, but it does not predict travel generated by visitors, commercial vehicles, freight transportation, and special events such as professional football games. A number of special models (commercial vehicle model, truck model, special events model, air passenger model, external trip model, visitor model, cross-border model) account for these other sources of transportation demand. After network assignment, software such as EMFAC calculates the resulting emissions.
SANDAG is in the process of developing a new land use model, PECAS. Integration between PECAS and the ABM occurs on a number of levels (see Figure 9). PECAS provides the socio-demographic and housing forecasts that underlie ABM’s population synthesis step. ABM creates a population that matches forecasted zonal attributes from the land use model (gender, age, race, children, household size, workers, income, housing type, and group quarters type). Additionally, each synthesized household and individual chooses travel destinations and workplace locations based in part on the PECAS-forecasted distribution of employment in the region. For example, a healthcare worker is more likely to commute to a zone with a large number of healthcare jobs. A shopper is more likely to visit a zone with extensive retail employment.
Going in the other direction, ABM calculates measures of accessibility that PECAS utilizes to determine how attractive locations are for the purpose of economic activity and real estate development. ABM and PECAS exchange output to reflect the interdependence of transportation and land use patterns.
Figure 9: Transportation-land use interactions in the SANDAG model system
Validation and Calibration
If necessary, adjustments to the model estimates will be made so that output better fits with observed data (this process is called calibration). The ABM will be calibrated to match various survey data. To gauge the quality of ABM results, model output will be compared with independent sources of data (this process is called validation). For example, predicted automobile travel can be compared with traffic counts, and travel by workers can be compared with worker flow data from the Census. The population synthesis step of the ABM is validated by comparing characteristics of the synthetic population with Census data.
Running an ABM requires not only extensive computing power but also a software architecture that allows for model improvement and growth. The SANDAG ABM is implemented in Java and takes full advantage of object-oriented programming principles. Adherence to these principles ensures that the model’s software components will be flexible, interchangeable, and easy to maintain and enhance. The ABM software implementation supports model evolution and facilitates cooperation in model development.
The SANDAG ABM utilizes a variety of data as inputs. The most important data source is household travel survey data. The latest household travel survey conducted for SANDAG was the 2006 Household Travel Behavior Survey (TBS06). TBS06 surveyed 3,651 households in San Diego County. The survey asked all household members to record all trips for a specified 24-hour weekday period using a specially designed travel log.
Additional data needed for the mode choice components of the ABM come from a transit on-board survey. The most recent SANDAG survey of this kind is the 2009 Transit On-Board Survey (OBS09). OBS09 collected data on transit trip purpose, origin and destination address, access and egress mode to and from transit stops, the on/off stop for surveyed transit routes, number of transit routes used, and demographic information. The total number of OBS09 survey records is 42,854.
Population synthesis requires two types of data: 1) individual household and person Census records from San Diego County, and 2) aggregate data pertaining to the socio-demographic characteristics of each zone in the region. The first type of data is available from the Public Use Micro-data Sample (PUMS), a representative sample of complete household and person records that is released with the Census and American Communities Survey. The second type of data is from the Census for the base-year and from land use forecasts for future years.
The data sources mentioned above, along with other necessary sources of data, are listed in Table 1. Modeling parking location choice and employer-reimbursement of parking cost depends on parking survey data collected from 2010 into early 2011, as well as a parking supply inventory. The transponder ownership submodel requires data on transponder users. Data needed for model validation and calibration include traffic counts, transit-boarding data, and Census Transportation Planning Package (CTPP) data.
Table 1: ABM Input Data