Travel Demand Modeling Process & Data

Travel Demand Modeling Process

Transportation planners and engineers conduct travel demand modeling to predict future traffic patterns, congestion, and transportation infrastructure needs. A typical travel demand modeling process, which involves both the development and application of a travel demand forecasting model (TDFM), includes the following steps:

• Specification

In this step, the model’s mathematical form is specified (e.g. regression, cross classification, logit, lookup table) and the variables of interest are identified.

• Estimation

In model estimation, one or more mathematical procedures are used to statistically determine the likely values of the model parameters and coefficients. For example, when estimating the likely coefficient values for a logit model, the method of maximum likelihood estimation (MLE) is generally used. Empirically estimated models rely upon individual behavior data, which is usually derived from surveys (e.g. Census, household travel surveys, air passenger surveys). Most estimation work is done with software packages such as SAS, SPSS, R, Stata, Alogit, LIMDEP/NLOGIT, Biogeme or LARCH.

• Implementation

Once a model is estimated, it needs to be implemented so that it can be applied. Most travel models are implemented and applied using computer software. The TPB travel models make use of both proprietary (e.g. Bentley Systems Cube) and open-source (e.g. ActivitySim, TRANSIMS ModeChoice, Python, R) software packages.

• Calibration/Validation

Model calibration and validation generally occur in an iterative fashion. In model calibration, model parameters or constants are adjusted so that model outputs match observed data at an aggregate level. Subsequently, the model is validated in a "base year" against observed data to make sure that it adequately and reasonably represents reality. Based on the performance of the model in model validation, additional adjustments may be made to the model parameters or constants in model calibration until the model accurately replicates observed patterns and behavior. Ideally, the model is validated to a different set of observed data than was used for model estimation or model calibration. A "future year" validation can also be performed. Although there is no observed data for a future year, one can make sure that the model forecasts in a sensitivity testing setting are reasonable and consistent with expectations. All TPB travel models are validated against observed data.

• Application

In the final step of the process, models are applied, generally using computer software, so that they may be used for developing forecasts that support various transportation planning processes, such as Long-Range Transportation Plan/ Metropolitan Transportation Plan (MTP) update, planning studies, and traffic impact analysis.

The above process is usually iterative, and each step can feed back to the previous step.

Travel Demand Forecasting Methodologies

COG/TPB’s Gen2 and Gen3 Travel Models represent two mainstream paradigms of travel demand modeling that are used in practice. The Gen2 Travel Model is a four-step, trip-based travel demand model (FSM) that simulates trips from each zone, based on household characteristics, and aggregates trips for the purposes of trip distribution and mode choice, while the Gen3 Travel Model is a tour-based/activity-based travel model (ABM) that micro-simulates travel at the person level and associates travel-related decisions with more specific household- and person-level data such as household income, ages of people in the household, and person type. The main difference between trip-based and activity-based models is the level of detail and granularity in modeling travel behavior. Trip-based models focus on individual trips, while activity-based models represent travel as both trips and tours, and consider the entire daily activity pattern of individuals and households. A tour is a series of associated/chained trips. For example, a home-based work tour starts at home, goes to work, and returns home, and could also include other intervening trips, like a trip to the store on the way to or from work.

An FSM contains four major steps:

• Step 1: Trip generation - How many trips are generated?

Step one of the process is to determine the number of daily trips that take place in the region. This procedure is called trip generation, and it estimates the number of "trip ends" produced in and/or attracted to each transportation analysis zone (TAZ) in the region. Each trip is made of two "trip ends," one at the production end of the trip and one at the attraction end of the trip.

• Step 2: Trip distribution - Where do the trips go?

In step two, trip distribution, the motorized person "trip ends" developed in trip generation are linked geographically into complete trips, from an origin/production zone to a destination/attraction zone. For example, work trip ends "produced" in a zone in Gaithersburg, Maryland are matched with the work trip ends "attracted" to other zones throughout the region. When trip ends are linked to create complete trips, the linking may occur within the same zone, or between adjacent zones, or between zones that are some distance apart, such as a zone in Gaithersburg and a zone in downtown Washington, D.C. The same process is used to connect all of the trip ends produced in or attracted to various zones in the region into complete trips.

• Step 3: Mode choice - What travel mode is used for each trip?

The third step of the modeling process is known as mode choice, which is used to predict the likely travel modes of all travelers in the region. Both Gen2 and Gen3 models consider three auto modes (drive alone, shared ride 2-person, and shared ride 3+person) and four transit modes (commuter rail, all bus, all Metrorail, and combined bus/Metrorail) in mode choice. The Gen3 Model also considers additional motorized and non-motorized travel modes (e.g. ride hailing, school bus, walking and biking) in mode choice.

• Step 4: Trip assignment - What is the route of each trip?

The final step in the forecasting of travel behavior is to determine the routes travelers choose to reach their destinations. This step is known as trip assignment. Only motorized person trips are assigned, which includes both trips made by automobile/car and trips made by public transit.

The workflow of a typical four-step model is illustrated below.

An ABM simulates the inter-connected travel-related choices made by individual households/persons throughout a typical day. These choices include long-term choices, daily scheduling choices, tour-level choices, and daily trip-level choices.

• Long-term choices include choices like a person’s work or school location (if applicable) and how many vehicles a household will own.
• Daily scheduling choices include choices like whether to make mandatory tours, non-mandatory tours, or stay home during the day, as well as the number of work and school tours during the day.
• Tour-level choices include choices like tour mode, tour timing, and number of stops on a tour.
• Finally, trip-level choices determine specific travel modes, times, locations, and purposes for each segment of each tour.

All these choices are linked together in the model system. The flowchart below (from the Gen3 Model user’s guide) illustrates an example ABM implementation of this choice model system. The Gen3 Model uses ActivitySim, an open-source platform for implementing ABMs.

Modeling Data & Sources

The models that make up the regional travel demand model are estimated, calibrated, and validated using observed data, supplemented by professional judgment when necessary. COG/TPB’s Gen2 Travel Model was estimated and calibrated to 2007 conditions, while the Gen3 Model was estimated and calibrated using 2018 observed data. The exogenous demand components of both models were developed based on a series of exogenous trip surveys conducted between 1994 and 2005. Here are some of the major sources of observed data used for the development of the Gen2 and Gen3 travel models:

• Census data, such as
  • Census Transportation Planning Products (CTPP)
  • American Community Survey (ACS)
• Regional household travel surveys, such as
  • 2007/2008 COG/TPB Household Travel Survey (HTS)
  • 2017/2018 COG/TPB Regional Travel Survey (RTS)
• Transit on-board surveys, such as
  • Metrorail Passenger Surveys (2008 & 2016)
  • 2008 Regional Bus Survey
  • On-Board Survey of Maryland Transit Administration (MTA) Riders, which include users of MARC train service (2008 & 2016)
  • Virginia Railway Express (VRE) Passenger Survey (2005 & 2019)
• 2005 COG/TPB Commercial Vehicle Survey (in addition to non-freight commercial vehicles, this survey also included some truck counts)
• 1994 COG/TPB Auto External Survey
• 1996 COG/TPB Truck Internal Survey
• 1996 COG/TPB Truck External Survey
• State and local government traffic counts (various years)

Model Inputs & Outputs

In the broadest sense, the TPB's regional travel demand forecasting models consist of three elements:

1. Input data
2. A series of models (mathematical procedures and representations)
3. Output data (“results”)

The three basic inputs to TPB's current regional travel demand model are:

• Forecasts of future population, households, and employment throughout the region.
  • For TPB’s trip-based Gen2 Model, TAZ-level population, household, and employment forecasts are directly used as land use inputs to the model.
  • For the activity-based Gen3 Model, population and household forecasts are used as marginal controls to generate a representative population of the region, which is used as an input to the model; TAZ-level employment forecasts, along with forecasts of other land use activities, such as golf course and park acres, are used as size-terms which are inputs to model steps like workplace location model and tour destination choice.
• Information about future transportation networks -- changes that are planned, or potential changes to be tested -- that would improve today's transportation system.
• Policy assumptions, such as how transit fares and other travel costs will change over time.

Population, household, and employment forecasts are regularly updated through COG’s Cooperative Forecasting Program, reflecting the best judgments of local officials regarding the location of future housing, commercial and industrial development within the region. The forecasts are developed by the Cooperative Forecasting and Data Subcommittee (CFDS), reviewed by the Planning Directors Technical Advisory Committee (PDTAC), and approved by the COG Board of Directors.

The main outputs from TPB's trip-based Gen2 Travel Model are:

• From trip distribution or mode choice
  • Estimated origin-destination (O-D) movements, represented by zone-to-zone person trip tables.
  • Although the trip distribution model is calibrated to the observed trip-length frequency distributions at the regional level and the mode choice model is calibrated by geographic segment, neither zone-level nor jurisdiction-level O-D information is validated. Thus, staff do not recommend its use.
• From the mode choice model
  • Estimated mode splits for zone-to-zone trips
  • Output data is summarized to the jurisdictions and the region
• From traffic assignment
  • Estimated vehicle trips and volumes on road segments/“road links” (though not validated at this level)
  • Estimated traffic speeds on road segments (though the speed data is not validated)
  • Output data is validated and summarized to screenlines, jurisdictions, and the region
• From transit assignment
  • Estimated person trips/volumes on transit links (though not validated at this level) for AM Peak and Off-Peak periods
  • All transit trips are assigned to the transit network, but we currently validate only total boardings by transit sub-mode and Metrorail boardings by station group
• From the MOVES mobile emissions model and the travel-speed post-processor model
  • Estimated air pollution (mobile emissions) from cars and trucks for both criteria air pollutants (e.g. NOx and VOC) and greenhouse gas (GHG) emissions.
  • Estimated information summarized at the regional and jurisdiction levels.

The demand-side output from TPB’s activity-based Gen3 Travel Model is substantially different than that from the Gen2 Model. Specifically, it features disaggregated travel behavior data that is generated from ActivitySim for individual households, persons, zones, vehicles, trips and tours through a series of Monte Carlo simulations.

• The final household table contains all the original synthetic population household attributes (e.g. household size, household income) as well as the ones created in ActivitySim (e.g. vehicle ownership, joint tour frequency).
• The final person table contains all the original synthetic population person attributes (e.g. age, sex, person type) as well as the ones created in ActivitySim (e.g. workplace location, school location, transit pass availability, telecommute frequency).
• The final zone table contains all the original land use fields (e.g. households, population, employment) as well as the ones created in ActivitySim (e.g. parking cost, terminal time).
• The final vehicle table contains all the information on the vehicles simulated in ActivitySim (e.g. vehicle type, operating cost, range, MPG, whether it is an Autonomous Vehicle).
• The final tour table contains all the information on the tours created in ActivitySim (e.g. tour type, origin, destination, mode, stop frequency, departure time, arrival time, duration).
• The final trip table contains all the information on the trips created in ActivitySim (e.g. purpose, origin, destination, mode, departure time, destination logsum, mode choice logsum). Note that the individual trips contained in the final trip table can be aggregated to O-D trip matrices at the TAZ or jurisdictional levels. The logsum, which is equal to the log of the denominator of a multinomial logit model, is a measure of composite utility or expected maximum utility of a choice set.

By contrast, the supply-side output from the Gen3 Travel Model is very similar to that from the Gen2 Model since both models use aggregate assignment procedures. The Gen3 Model uses the same highway assignment procedure as the Gen2 Model. The transit assignment processes are slightly different, as the Gen3 Model switches from Cube TRNBUILD to Cube Public Transport (PT) for transit modeling, and simulates transit trips in O/D format for four time periods (as opposed to P/A format and two time periods in the Gen2 Model).

Note: The trip-based Gen2 Travel Model produces many fine-grained outputs, such as link-level outputs (e.g., the number of vehicles traveling on each link in the AM peak period) and zone-interchange-level outputs (e.g., the number of bus person-trips traveling from TAZ X to TAZ Y). The Gen3 Model produces completely disaggregated outputs, such as individual tours and trips. However, neither model has been calibrated or validated to these fine-grained levels, so it is not recommended that one directly uses these fine-grained outputs from the travel model. A general rule is that, before using or reporting any model outputs, they should be summarized or aggregated to the same, or a higher, level as was used in model calibration or model validation. For example, although the model produces estimated link-level traffic volumes, this information should be aggregated to the screenline level, jurisdiction level, or regional level, before it is used or reported. Please see the latest user’s guides for the Gen2 and Gen3 travel models for more information.