Activity-Based Travel Demand Models: A Primer (2014)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

131 SHRP 2 C10A JACKSONVILLE, FLORIDA, AND BURLINGTON, VERMONT Objective The primary objective of the SHRP 2 C10A project was to make operational a regional-scale dynamic integrated model and to demonstrate the model’s performance through validation tests and policy analyses. The model system was designed to cap- ture changes in demand, such as time-of-day choice and peak-spreading, destination, and mode and route choice, in response to capacity and operational improvements such as signal coordination, freeway management, and variable tolls. An additional goal was to develop a model system that could be transferred to other regions, as well to incorporate fi ndings from other SHRP 2 efforts. The model system was imple- mented in two regions: Jacksonville, Florida, and Burlington, Vermont. In both re- gions, implementation of the model system was primarily performed by a consultant team, with the Jacksonville MPO and the Florida DOT providing data and support (Resource Systems Group et al. 2014). The information presented in this example was gathered from project reports and interviews with project team members. Model System Design and Components The SHRP 2 C10A model system comprises two primary components. They are DaySim and the TRANSIMS Router and Microsimulator. DaySim is an activity-based travel demand forecast model that predicts household and person travel choices at a parcel level on a minute-by-minute basis. The TRANSIMS Router and Microsimulator is dynamic network assignment and network simulation software that track vehicles on a second-by-second basis. DaySim simulates 24-hour itineraries for individuals with spatial resolution as fi ne as individual parcels and temporal resolution as fi ne as single minutes, so it can generate outputs at the level of resolution required as input to dynamic traffi c simulation. The TRANSIMS network microsimulation process assigns a sequence of trips or tours for individual household persons between specifi c 5 CASE EXAMPLES

132 Part 2: ISSUES IN ADOPTING INTEGRATED DYNAMIC MODELS SYSTEMS activity locations to paths on a second-by-second basis for a full travel day. The net- work includes detailed information regarding the operational characteristics of the transportation facilities that may vary by time of day and by vehicle or traveler type such as the number of lanes; the lane use restrictions; traffic controls, signal timing, and phasing plans; turning restrictions; and tolls and parking fees (Resource Systems Group et al. 2014). The C10A model system was implemented for two regions. The Jacksonville model includes four counties in northern Florida with a population of approximately 1.2 million people, while the Burlington model includes one county in Vermont with a population of approximately 150,000 people. The model system employs multiple spatial resolutions—the base spatial data describing employment and households are at the level of individual parcels, while the network performance indicators are avail- able at either the TAZ level or the activity location (AL) level. The model system also employs multiple temporal resolutions. On the demand model side, the core time- of-day models within DaySim operate at temporal resolutions as fine as 10 minutes and are subsequently disaggregated to individual minutes. On the supply model side, TRANSIMS follows vehicles on a second-by-second basis, measures of link-level net- work performance are typically collected using 5-minute intervals, and measures of O-D network performance (or skims) are also generated using temporal resolutions as fine as 10 minutes, consistent with the demand side time-of-day models. The DaySim model incorporates significant typological detail, including basic persontypes, as well as identifying trip-specific values of time reflective of travel purpose, traveler income, and mode. This value-of-time information is incorporated into the TRANSIMS net- work assignment process using approximately 50 value-of-time classes. TRANSIMS also generates network skims by value-of-time class for input into DaySim. The model system is configured to run a fixed number of assignment iterations and system itera- tions and is designed to achieve sufficient levels of convergence as necessary to gener- ate meaningful performance metrics for planning purpose (Resource Systems Group et al. 2014). Lessons Learned Developing the inputs to the DaySim activity-based demand model components was relatively straightforward, though significant cleaning was required. Transferring the DaySim activity-based demand component from Sacramento to Jacksonville radically reduced the amount of time required to implement the activity-based demand model component of the model system. In contrast, developing detailed and usable networks for microsimulation required a significant level of effort, although this effort was miti- gated by using TRANSIMS tools to perform network development tasks and the avail- ability of spatially detailed network data. Correcting topological errors; resolving attri- bute discontinuities; coding intersection controls; and iteratively evaluating, adjusting, and testing the networks by running simulations is time-consuming. In addition, there are numerous challenges when developing future-year or alternative network scenarios (Resource Systems Group et al. 2014).

133 Chapter 5: CASE EXAMPLES As noted in the SHRP 2 C10A final report, “Configuring DaySim to generate temporally, spatially, and behaviorally detailed travel demand information for use in TRANSIMS was straightforward, as was configuring TRANSIMS to generate the skims for input to DaySim. More sophisticated methods of providing TRANSIMS- based impedances to DaySim, such as implementing efficient multistage sampling of destinations (and corresponding impedances) at strategic points in the DaySim looping process or integrating DaySim and TRANSIMS so that DaySim can call TRANSIMS to extract the required measures quickly, could potentially be implemented.” (Resource Systems Group et al. 2014). The new model system is more sensitive to a wider range of policies than a tradi- tional travel demand model system, and this sensitivity is further enhanced by the detailed representation of temporal dimension. Extensive testing of the model system was necessary to determine the number of network assignment and model system iter- ations required to ensure that differences between alternative scenario model results were attributable to these policy and investments and not obscured by noise in the model system. Extracting, managing, and interpreting these results was not difficult; however, the level of effort required to effectively test different types of improvements varied widely, from as little as an hour to as more than a week. It is safe to say that a higher degree of knowledge and patience is required when interacting with the new integrated model system than is required when using a traditional trip-based model system (Resource Systems Group et al. 2014). SHRP 2 C10B SACRAMENTO, CALIFORNIA Objective As with the SHRP 2 C10A project, the primary objective of the SHRP 2 C10B project was to make operational a regional-scale dynamic integrated model and to demon- strate the model’s performance through validation tests and policy analyses. However, there are two notable distinctions between the scopes of the C10B and C10A projects. First, the size of the Sacramento, California, region used in the C10B project is approx- imately twice as big as the Jacksonville region used in the C10A project, and more than 10 times as large as the Burlington region used in the C10A project. Second, and more significantly, the C10B project included a dynamic transit demand network assignment model in addition to a dynamic roadway network assignment model. Even though many dynamic roadway network assignment models represent interactions between transit vehicles and private vehicles, there are very few models that provide the capa- bility to assign transit demand and represent the effect of this demand on network per- formance. Although the development of the integrated model system was primarily led by a consultant team, SACOG staff were actively involved in the C10B effort, such as performing the model sensitivity test runs (T. Rossi, personal communication, Oct. 17, 2013). The information presented in this example was gathered primarily from inter- views with project team members and from project reports.

134 Part 2: ISSUES IN ADOPTING INTEGRATED DYNAMIC MODELS SYSTEMS Model System Design and Components The SHRP 2 C10B model system comprises three primary components: DaySim, Dynus-T, and FAST-TrIPs. DaySim is a travel demand forecast model that predicts household and person travel choices at a parcel level on a minute-by-minute basis. Dynus-T is the dynamic roadway traffic assignment tool, which tracks vehicles on the network on a second-by-second basis, and FAST-TrIPs is the dynamic transit demand assignment tool, which tracks transit travelers on a second-by-second basis ( Cambridge Systematics, Inc. et al. 2014). DaySim simulates 24-hour itineraries for individuals with spatial resolution as fine as individual parcels and temporal reso- lution as fine as single minutes, so it can generate outputs at the level of resolution required as input to dynamic traffic simulation. Dynus-T assigns a sequence of trips or tours for individual household persons between specific activity locations to paths on a second-by-second basis for a full travel day and incorporates significant capabili- ties to adjust the input demand in order to generate more realistic results. Dynus-T operates at a mesoscopic scale, which is different from the microscopic simulations of the TRANSIMS, TransModeler, and Dynameq software used in the other integrated model development efforts described in this document, although Dynus-T shares some similarities with microscopic car-following-based models (Cambridge Systematics, Inc. et al. 2014). FAST-TrIPs is a transit assignment tool that is designed to accurately represent transit operations, to capture the operational dynamics of transit vehicles, to provide both schedule-based and frequency-based transit traveler assignment, and to generate skims for feedback to the activity-based travel demand model (Cambridge Systematics, Inc. et al. 2014). The C10B model system was implemented in the Sacramento, California, region, which includes approximately 2.3 million people. Like the C10A project, the C10B model system employs multiple spatial resolutions—the base spatial data describing employment and households are at the level of individual parcels, while the network performance indicators are available at the TAZ level. The model system also employs multiple temporal resolutions. On the demand model side, the core time-of-day models within DaySim operate at temporal resolutions of 30 minutes, which are sub- sequently disaggregated to individual minutes. On the supply model side, Dynus-T and FAST-TrIPs follow vehicles on a second-by-second basis, and ultimately skims of O-D network performance are generated using a temporal resolution of 30 minutes, consis- tent with the demand side time-of-day models. The model system is configured to run a fixed number of assignment iterations and system iterations and is designed to achieve sufficient levels of convergence as necessary to generate meaningful performance met- rics for planning purposes (T. Rossi, personal communication, Oct. 17, 2013). Lessons Learned The most important lesson learned from this effort is that it demonstrates the feasi- bility of implementing a regional-scale integrated activity-based model and dynamic traffic and transit assignment models. The project clearly illustrated the potential ben- efits of a more continuous representation of time such as the ability to generate more

135 Chapter 5: CASE EXAMPLES detailed network performance skims, as well as the ability to more precisely character- ize the location, extent, and duration of congestion (T. Rossi, personal communica- tion, Oct. 17, 2013). However, implementing the model system was a significant undertaking. Exten- sive efforts were required to develop and calibrate the roadway and transit assign- ment models, and additional efforts are likely required to achieve a level of confidence required to support project evaluations. Applying the model was also complicated by the relatively long model system run times, and by the fact that the integrated model system requires a relatively high level of modeler involvement to execute a complete integrated run. Interpretation of model results also proved to be challenging, and fur- ther work is required in order to ensure that the network assignment models and the overall model system are reasonably well converged before being suitable to sup- port policy and investment analyses. Stochasticity in the model results appears to be an issue that will require further investigation. Finally, the project team felt that the iterative development and expansion of modeled area may not be the most effective method for getting to a full regional model implementation (B. Griesenbeck, personal communication, Oct. 17, 2013). SAN FRANCISCO COUNTY TRANSPORTATION AUTHORITY’S “DTA ANYWAY” Objective The goal of the San Francisco County Transportation Authority’s “DTA Anyway” project was to develop an application-ready tool that the SFCTA could use to evalu- ate projects throughout the city. SFCTA staff were particularly interested in under- standing the effects of congestion pricing on transit performance and traffic diversion, representing operational strategies, and producing realistic traffic flows in which the forecast demand does not exceed assumed capacities. This citywide dynamic network model built on an earlier dynamic traffic (or network) assignment (DTA) model built for the northwest quadrant of the city that had been successfully applied to analyze construction phasing for a major roadway project. Beyond being able to use the model to support SFCTA’s planning activities, SFCTA staff also sought to assist future DTA deployment efforts by building a toolkit in Python programming language that pro- vides capabilities such as network data exchange and conversion procedures and reporting capabilities. Additionally, SFCTA staff tried to fully document the process and assumptions used in implementing the citywide DTA model and to reveal DTA performance in the context of a congested grid network in which there is significant interaction with transit vehicles and demand (Parsons Brinckerhoff and San Francisco County Transportation Authority 2012). Model System Design and Components The SFCTA DTA Anyway model system comprises two primary components: the SF-CHAMP activity-based model system and the Dynameq network simulation model. SF-CHAMP is an activity-based travel demand forecast model that predicts household and person travel choices. Dynameq is microscopic traffic simulation model

136 Part 2: ISSUES IN ADOPTING INTEGRATED DYNAMIC MODELS SYSTEMS with sufficient detail to consider lane-level modeling. The SF-CHAMP activity-based model generates demand for the entire San Francisco Bay Area region, while the DTA Anyway model implementation covers San Francisco County only. In order to bridge these different spatial extents, a subarea extraction process using a static network assignment model is used in which vehicle flows into and out of San Francisco are sum- marized. This method can be effectively applied in San Francisco because of the unique geographical features that define the city. SF-CHAMP forecasts DAPs for all regional residents using a spatial resolution of very small TAZs within San Francisco and a temporal resolution of multihour time periods. SF-CHAMP is sensitive to the impact of travel times and costs on time of day, mode, destination, and activity generation, and generates lists of person-level tours and trips. However, these tours and trips are not used directly in the model. Rather the trips lists are aggregated to matrices of flows by time of day and mode, these flows are assigned to SF-CHAMP’s static model networks, and subarea matrices are derived from this assignment. The Dynameq traffic microsimulation model assigns discrete vehicle trips to a detailed San Francisco network. This network includes information on the actual sig- nal and timing plans for all traffic signals in the city (there are more than 1,100) as well as the locations of other intersection controls, such as more than 3,000 stop-sign locations. The network also includes detailed information regarding the operational characteristics of the transportation facilities that may vary by time of day and by vehicle type or traveler type such as the number of lanes, lane use restrictions, traffic controls and signal timing and phasing plans, turning restrictions, tolls, and parking fees. The network simulation also includes transit vehicles, which are an important segment of the vehicle fleet operating in transit-rich San Francisco. A key challenge in integrating the SF-CHAMP activity-based demand model and the Dynameq traffic microsimulation model is the different temporal resolutions used by these two tools. It is necessary to temporally disaggregate the broad time-period demand produced by the activity-based model down to the finer time slices used by Dynameq. In the San Francisco Dynameq network model, changes in network performance by time of day that are used to build paths are represented using 7.5-minute intervals, although the simulation of vehicle interactions uses a significantly finer temporal resolution. The network simulation is performed only for the 3-hour p.m. peak period, although a one-hour warm-up period, and a one-hour cool-down period are also simulated. During this time period, approximately 450,000 vehicle trips are assigned. It should be noted that because of the limited temporal and spatial extents of the traffic micro- simulation model, it is not feasible to generate skims for feeding back input to the activity-based model. Thus, the integration of the model is one way. Lessons Learned SFCTA staff and their consultant team members learned a number of meaningful les- sons from the development of the dynamic traffic model for the city and the inte- gration of activity-based demand into this model. From a practical perspective, the automated procedures for aligning the Dynameq and SF-CHAMP data assumptions

137 Chapter 5: CASE EXAMPLES proved to be invaluable when applying the model to evaluate project alternatives. An additional data-related conclusion was that it is much better to use actual data rather than synthesized data, to the greatest extent possible. Actual data can include observed signal timing information when building networks, observed traffic flow properties when calibrating traffic flow parameters, and traffic counts and speeds when validating network model results. This effort also proved that it is not necessary to use matrix-estimation techniques to create input demand; this finding is significant because the use of matrix estimation in the context of future or alternative scenarios is problematic. A limitation of the current integration scheme that may be addressed in future model development phases is the lack of temporal information when extracting subarea demand. This effort also revealed to SFCTA staff a number of traffic microsimulation model sensitivities. For example, the traffic simulation model proved to be very sensitive to small changes in input assumptions. Staff described how a one-foot increase in the effective vehicle length caused systemwide network performance issues. Similarly, a bottleneck at a single intersection could also cause the entire network simulation to crash. Regarding using the model to evaluate alternative scenarios, staff discovered the following two points: The dynamic network model generally predicts more traf- fic diversion than static assignment techniques as a result of the sensitivity to actual capacity constraints, and stochasticity can be an issue when comparing scenarios, necessitating higher levels of convergence in order to draw meaningful conclusions. Finally, SFCTA staff advocated that sensitivity testing is an essential part of the model calibration and validation process (Parsons Brinckerhoff and San Francisco County Transportation Authority 2012). Next steps for model development include development of a disaggregate dynamic transit assignment model and better representation of parking behavior within San Francisco. Other key development tasks include the development of a full 24-hour simulation and the associated development of dynamic network-model-based skims for the entire day (Parsons Brinckerhoff and San Francisco County Transportation Authority 2012). MARICOPA ASSOCIATION OF GOVERNMENTS INNER LOOP TRAFFIC MODEL Objective The Maricopa Association of Governments developed the Inner Loop Traffic Model to support the Central Phoenix Transportation Framework Study. The purpose of this study is to identify the transportation strategies and investment needs for the central portion of the Phoenix region. The Phoenix core freeway system is still relatively new, but there are a significant number of chokepoints. Rather than only consider capac- ity expansion investments, the region wanted to have a tool that provided sensitivity to operational strategies in order to be able to more fully understand the interactions between the region’s highway system and regional arterials, and to strategically iden- tify how arterials can accommodate projected travel demand. In addition, significant

138 Part 2: ISSUES IN ADOPTING INTEGRATED DYNAMIC MODELS SYSTEMS investments in transit are being made by the region, and there is a tremendous focus around planning and developing high-capacity transit corridors. In order to have sen- sitivity to these strategies, it was determined that a more detailed network model than found in the trip-based demand model would be required. The development of the Inner Loop Traffic Model is considered the first part of a multiphase effort to develop regional simulation capabilities. This initial effort demonstrated the proof of the con- cept that it is feasible to develop and calibrate a regional-scale traffic simulation model (R. Hazlett, personal communication, Oct. 3, 2013). The information presented in this example was gathered from interviews with project team members. Model System Design and Components The focus of the Inner Loop Traffic Model development effort was to establish a large-scale network simulation model rather than to fully integrate the regional travel demand and network simulation components. The traffic model is only loosely linked with traditional trip-based travel demand model, and there is not yet an integrated demand–supply model runstream. Travel demand from the region’s trip-based model was used to seed the traffic simulation calibration effort, but the trip-based model demand was refined by applying a dynamic O-D matrix-estimation process that uses 15-minute counts. As a result, the present version of the model is more oriented toward shorter-term operations and engineering analyses rather than long-term future demand analyses. The project team is developing a process for creating detailed future-year simulation demand by pivoting off of the future demand model outputs. At present, network performance indicators, such as travel time and cost skims, are not being fed back to the trip-based demand model. The spatial extent of the simulation model is approximately 530 square miles. The model maintains a consistent geographic resolution with the trip-based model, including approximately 800 internal zones and 90 external or interface locations, but there are significantly more network loading locations than in the trip-based model. The original model design called for the simulation and calibration of two 3-hour peak periods, but ultimately the model included a broader temporal extent. This inclusion of the broader temporal extent was necessary to calibrate the 3-hour period that pre- cedes each of the peak time periods in order to ensure reasonable peak period network performance. This approach was especially necessary for the p.m. peak period. In the a.m. peak approximately 900,000 trips are simulated, while in the p.m. peak approxi- mately 1.2 million trips are simulated (D. Morgan, personal communication, Oct. 3, 2013). The project team was able to implement a microscopic model for the entire mod- eled area that incorporates the accurate representation of signals, meters, and bus routes and schedules. Use of a microscopic scale model provides better sensitivity to operational phenomena like traffic across lanes, weaving, merging, signals, and bottle- necks. The dynamic user equilibrium seeking solution method implemented in the model does not rely on the traditional method of successive averages as many dynamic network models do. Rather, at every iteration, every driver’s paths are informed by the latest network performance information. This approach is similar to the methodology

139 Chapter 5: CASE EXAMPLES used in the SHRP 2 C10A project. The model uses a temporal resolution of 15 min- utes for network pathbuilding, and a temporal resolution as fine as 0.1 seconds for the simulation step size (D. Morgan, personal communication, Oct. 3, 2013). The model development work was performed almost exclusively by consultants, although MAG staff received some training, and MAG has dedicated a staff person to ongoing model management. Agency staff are also now performing testing of the model to ensure that the tool is incorporating realistic assumptions and is producing reasonable results. The consultant selected to implement the model is a developer of one of the major traffic simulation software packages, and this familiarity with the software was one of the primary factors in selecting this consultant. Lessons Learned The most important lesson learned from this effort is that it is possible to build a regional-scale microscopic traffic simulation. Microscopic models can provide better sensitivity to operational phenomena such as traffic across lanes, weaving, merging, signals, and bottlenecks than mesoscopic models. Microscopic models have longer run times, but given operational considerations of interest to MAG, this trade-off was acceptable. Learning how to harness current hardware and software in order to achieve better run times was a key learning component of the project. An obvious lesson learned—but one that still bears repeating—is that regional simulation models require lots of good data and that it is preferable to use observed data rather than synthesized data. However, there are limitations on the availability of actual data; it is unavoidable that some assumptions and synthesis of data are necessary in establishing the model. In addition, calibration of the model system is challenging. Stochasticity, or random variation, was a particular focus of the team in develop- ment of the model. Stochasticity is intrinsic to the simulation model, as well as intrinsic to the real world, and this effort revealed there is significant investigation to be done to understand how the models can be run and how the results can be applied, interpreted, and communicated to member agencies and to the public. Two primary next steps are envisioned for this model. First, the regional net- work simulation is to be expanded to cover the entire region, rather than just the core 500 square miles. This work is currently ongoing. Second, MAG has also been developing a regional activity-based model system. Although the initial version of this activity-based model system incorporates a traditional static network assignment component, it is anticipated that at some point the activity-base demand model and regional traffic microsimulation model will be linked (R. Hazlett, personal communi- cation, Oct. 3, 2013). These case examples confirm the potential for integrated dynamic models to give more comprehensive and more detailed information to decision makers and to pro- vide sensitivity to a wider set of policy and investment alternatives. The C10A and C10B projects demonstrated these enhanced capabilities through a set of diverse policy sensitivity tests, while the SFCTA’s DTA Anyway project illustrated how an activity- based and dynamic network model model system could be used to inform real project choices. The SFCTA as well as the MAG efforts are also indicators of future directions

140 Part 2: ISSUES IN ADOPTING INTEGRATED DYNAMIC MODELS SYSTEMS in travel demand forecasting practice, as both agencies intend to move toward more fully integrated dynamic model systems with future model development efforts. How- ever, these case examples also illustrate issues (e.g., integration strategies and com- putational resource requirements) with developing an integrated model. In addition, implementing an integrated dynamic network model necessitates addressing the issues independently associated with implementing an activity-based model and a regional- scale dynamic network model. The following sections in Chapter 6 consider some of the critical issues faced when implementing integrated dynamic model systems.