Advanced Practices in Travel Forecasting (2010)

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.

The goal of this chapter is to provide the reader with enough background knowledge to understand the key aspects of advanced models, providing a foundation for the remainder of the report. As noted, the report will focus primarily on activity- based travel demand models and their integration with land use models and dynamic network models. Freight models are also considered, as they are reaching the same levels of sophistica- tion and are based on much of the same data, principles, and mindsets. Most of the models described in this report have been applied at the urban and metropolitan levels; however, nothing precludes them from being applied at the statewide level as well, and several examples of such applications are presented. A separate discussion of statewide models and issues unique to their scale are provided to contrast them with the more familiar urban models. Regardless of the specific applica- tion, many of the institutional issues associated with imple- menting advanced models and lessons learned are the same. Many of the models described can easily be labeled as a “demand model” or “supply model.” Indeed, practitioners often describe themselves as falling into one camp or the other. Many developers of activity-based travel models place themselves in the first camp, whereas network modelers identify with the latter. Users of traffic simulation models have traditionally worked outside the orbit of both groups. One of the biggest promises of the advanced modeling paradigms discussed in this report is the fusion of these heretofore sep- arate areas of practice. For example, several researchers are exploring the integration of activity-based travel demand and dynamic traffic assignment models. Much has already been learned from land use modelers who for some time have tightly coupled the demand supply sides of their model. Although the discussion in this chapter follows the literature and prac- tice to date—which stops just short of merging these two camps—it is clear that the convergence of these two streams of research and practice is close at hand. TOUR- AND ACTIVITY-BASED TRAVEL MODELS Over the past decade, several activity-based models have been developed in the United States, and applied successfully in practice, and more are in the process of being developed. The locations of these models are shown in Figure 1. The first generation of those used extensively in practice includes models in New York, San Francisco, and Columbus (Ohio), with more recent applications in Oregon, Sacramento, and 8 Lake Tahoe (Nevada). Activity-based models are currently being actively developed in locations including Atlanta, Denver, Ohio (statewide), Portland, San Diego, the San Fran- cisco Bay Area, and Seattle. With this breadth of applica- tions, activity-based models are making a major advance- ment with the greatest penetration into practice. The general characteristics of such models are described in this section and compared with traditional trip-based models. The goal of this section is not to provide a comprehensive guidebook for these models, but to describe them in just enough detail so that the reader is provided with some context for the remainder of this report. Recognizing that each model developed so far is different, the focus remains on the com- monalities among these models, rather than the differences between them. When comparing activity-based and trip- based models a representative “good practice” model is described for each, rather than the model implemented in a specific location. Trips, Tours, and Activities Traditional models use a trip as the basic unit of analysis, hence the moniker “trip-based models.” A trip is defined as a unit of travel connecting two locations. Figure 2 shows an example of a day containing five trips. The trips connect four locations: home, work, a store, and a park. At each of these locations some activity or set of activities occurs, which might include working, eating a meal, shopping, and recreation. In a trip-based model, each trip is modeled as an indi- vidual unit with no knowledge of its context beyond its endpoints. Therefore, in the example, trip 5 has no knowl- edge that it is the reverse of trip 4, nor does trip 3 know that it is related to trip 2. In this example, trips are categorized by what happens at either endpoint into three purposes: home-based work (HBW), home-based other (HBO), and non-home-based (NHB). Home-based trips are trips with either end at home, whereas NHB trips have neither end at home. This distinction is important in a trip-based model, because if the trip has one end at home then the model can take into consideration the demographic characteristics of households in that home zone. Conversely, for NHB trips, the model cannot con- sider any demographic characteristics because doing so would CHAPTER TWO GENERAL DESCRIPTION OF ADVANCED PRACTICES

9In Figure 3 the same five trips are grouped into two tours, where a tour is a sequence of trips that starts and ends at home. This is the type of grouping that occurs in a tour-based model, where the tour becomes the primary unit of analysis. For each tour, a primary purpose is assigned based on the most impor- tant activity that happens on that tour. In this case there is one work tour and one other tour. Each tour is assigned a primary destination (where the primary activity occurs) and a primary mode. Tour-based models ultimately consider decisions made at the trip level, but the choices for trips are constrained to be consistent with the tour of which it is a part. For Tour 2 in this example, both trips would be forced to occur between the same two zones in opposite directions. In Tour 1, the second trip must start from the work location, and the stop location on the return commute would be chosen based on how far it deviates from the work to home path. With the grouping into tours, tour-based models overcome much of the knowledge loss associated with trip-based models because the trips in the model have some knowledge of their context. Furthermore, because information can be traced back to the home, demographic characteristics of the traveler can be considered for all trips, not just home-based trips. An activity-based model goes further in recognizing travel as a derived demand. That is, the demand for travel is derived FIGURE 1 Activity-based models in the United States. FIGURE 2 Example of trips. require tracing beyond the trip’s endpoint and therefore beyond the scope of knowledge for that trip. The second major distinction within the home-based trip purposes is between work trips and other trips. HBW trips are of particular importance owing to the significance of home-to- work commutes to travel during the congested peak periods and to transit markets. Note that even though trips 2 and 3 add up to a work com- mute with a stop on the way home, neither is classified as a HBW trip. Therefore, when these two trips are modeled there is a loss of knowledge in the model—knowledge that the two trips are related, that the stop is likely to occur somewhere between home and work, and that the time-of-day and mode preferences associated with work commutes are relevant. Trip-based models work best in situations where there is ample direct “there and back” travel between home and a single location. The loss of knowledge in a trip-based model is higher in situations where there is a lot of trip chaining— the linking together of more than two trips while away from home—and therefore a high share of NHB trips. In the United States, the share of NHB trips has been increasing in recent decades as auto ownership and the number of multi-worker families has increased and as suburban land use patterns have become more dominant. FIGURE 3 Example of tours.

from the desire to participate in activities, rather than the desire to travel for the joy of being in the car. Figure 4 shows the same travel day, but lists the activities in which the person is engaged. The person works at work, shops at the store, and recreates at the park. Further, several activities are listed as occurring in the individual’s home. The individual begins her day by sleeping, then eating breakfast. After returning home from work, the individual eats before going to the park to recreate, then returns home to sleep. Activity-based models recognize that a person is motivated to work (coincidentally another derived demand, associated with the desire to get paid) not to make HBW trips, and that the person is motivated to shop, not to make HBO trips. Therefore, the activity-based models are capable of modeling the tradeoff between participating in a shopping activity as a stop on the work tour versus making an additional tour for the sole purpose of shopping. That tradeoff can be a function of the desirabil- ity of shopping locations near the home zone—the data show that persons living in highly accessible locations are more likely to make additional tours, whereas persons in inacces- sible locations will seek chain trips into more complex tours. If an activity-based model were to consider the details of in-home activities, it could explicitly consider the tradeoff between shopping online at home and traveling to a store to shop. Although there has been research on this topic, the models implemented in practice do not go to this level of detail. The literature is not in agreement as to a precise definition of what constitutes a tour-based model versus an activity-based model. Some would argue that a model is not truly activity- based unless it involves a list of activities and whether they occur in-home or out-of-home. Although there is some intuitive appeal to such a definition, such a distinction is a minor point compared with the large differences with traditional trip-based models. Therefore, in colloquial use, the two terms are often used interchangeably. That practice is continued in this report. Demand Simulation Many of the advanced models described in this report depart from traditional deterministic approaches familiar to practi- tioners of trip-based models. Closed form mathematical 10 equations have been used extensively in trip-based models. An obvious but important property of such deterministic models is that they can be exactly solved obtaining an invariant solu- tion. That is, such a model solved repeatedly will always obtain the exact same solution. Simulation models, on the other hand, are often used when analytical solutions cannot capture the salient characteristics of the system under study. They typi- cally include stochastic (random) effects or variables, such that repeated trials give rise to different outcomes. The degree of difference depends on how much random variation is intro- duced and at what level the difference is measured. Although most advanced models are constructed using richer behavioral detail and interactions, moving from a deterministic to sto- chastic framework entails an equally large change in the analytical mindset. Travel models, both traditional and advanced, are generally built on a core set of probabilistic choice models. A mode choice model, for example, predicts the probability that a user will choose each available mode given the level-of-service for each mode and relevant attributes of the traveler or trip. Tra- ditionally, such models are applied using a fractional proba- bility approach where the probability of each choosing each mode is multiplied by the total number of trips to get the total number of trips on each mode. If the probability of choosing auto is 0.75, the probability of walking is 0.125, and the prob- ability of taking transit is 0.125, and there are 10 total trips, then the model would predict 7.5 auto trips, 1.25 walk trips and 1.25 transit trips. An alternative way of applying the same model is to simu- late the choice made by each trip using a Monte Carlo approach. In this case, a random number between 0 and 1 would be drawn for each of the 10 trips. If the random number is in the range 0 to 0.75 then the trip is assigned a choice of auto, if it is in the range 0.75 to 0.875 it is assigned a choice of walk, and if it is in the range 0.875 to 1, it is assigned a choice of transit. In this example, there might be seven trips choosing auto, one choosing walk, and two choosing transit. Over a large sample the simulation will produce a result equivalent to the frac- tional probability model, although there will certainly be some variation owing to the simulation. In the travel forecasting practice, this approach is referred to as demand simulation, microsimulation, or pseudo-random sample enumeration. This method usually feeds into a standard user equilibrium highway assignment and should not be confused with traffic micro- simulation, where individual vehicles are simulated traversing highways. Both approaches are based on the same core probabilistic models; therefore, they would produce the same outcome in the aggregate. However, there are several practical reasons for demand simulation (Vovsha et al. 2002). The first main moti- vating factor is that there are significant computational and data storage benefits for large problem sizes. Second, explicitly tracking the choices of individual agents allows for the down- stream choices to be constrained to be logically consistent withFIGURE 4 Example of activities.

11 the choices that have already been made. Put together, these allow for the formulation and implementation of more complex model systems. These differences are discussed as each method is described in further detail. Note that it is possible to apply a trip-based model using demand simulation, and it is possible to apply an activity-based model using fractional probabilities. This is not commonly done in current practice, however, with the existing trip-based models usually using a fractional probability approach and the existing activity-based models all using a microsimula- tion approach. This is primarily because demand simulation makes it easier to manage the added complexity in activity- based models. Consider a traditional fractional probabilities model with 3,000 traffic analysis zones (TAZs), as outlined in Figure 5. The model begins with a list of households in each TAZ in each of three auto ownership levels. Trip generation rates are applied to each set of households for three trip purposes, result- ing in 9 vectors of trips. When those trips are distributed to 3,000 possible destinations, there are 9 matrices. When a mode choice model is applied, those 9 matrices are divided into 27. Thus, for this simple model the software must store and process 27 matrices (3 auto levels × 3 purposes × 3 modes) with 9 mil- lion (3,000 TAZs × 3,000 TAZs) elements each, for a total of 243 million elements. Each element is a fractional number, many of which are very small. If each is stored as a 4-byte floating point number, this amounts to about 900 megabytes (MB) of data, a manageable amount for today’s hardware. In the fractional probabilities approach, the problem size increases quickly as the complexity of the model increases. Consider instead a model with 5 purposes and 10 modes, which is not uncommon. Here the model would require 150 matrices (3 auto levels × 5 purposes × 10 modes) and 5 gigabytes (GB) of data. If the model were also segmented by 3 income levels, the problem size would triple to 450 matrices and 15 GB of data. Segmenting fully by 4 times-of-day would quadruple the problem size, for a total of 1,800 matrices and 60 GB of data. Of course, this does not account for other terms such as age, gender, and household composition that may be useful descrip- tive variables in a travel model. As the complexity increases, the application becomes both computationally intensive and cumbersome to manage. Next, consider the same model implemented using micro- simulation. Here the software is relieved of the need to store every possible combination of outcomes for every zone pair and instead only the chosen outcomes are stored. Rather than storing vectors or matrices of data at the zonal level, the model stores a table of data with one record for each trip and a column for each field of interest. In the example shown in Figure 6, if there are 10 million trips and 8 columns of data, the model would need to process and store 80 million ele- ments. If each is stored as a 4-byte-long integer, this would amount to about 300 MB of data. Adding modes or purposes would not increase the amount of data that needs to be stored. Adding income level and time of day would each require an additional column, increasing the problem size to 100 million elements and 380 MB of data. It becomes apparent that as the complexity of the problem increases microsimulation allows for the necessary data to be efficiently stored. The true advantage to demand simulation is that it allows modelers to develop more sophisticated core probabilistic models without devoting undue resources to managing the complexity. It is less of a burden to include additional variables, such as household composition, age, or additional purposes and modes that may be useful to the model formulation. Further, because the outcomes from the upstream models are stored as discrete values, it is easy to make downstream models depen- dent on those choices. In the previous section it was noted that in a traditional model the scope of knowledge for a trip is its two endpoints, meaning that the model cannot consider the context of the trip within a more complex chain. Tour- based models can overcome this by extending the scope to the entire tour. Demand simulation allows the scope of knowl- edge to extend even further, because the decisions are tracked explicitly. For example, time-of-day models can know when other tours have already been scheduled, and avoid overlap- ping. Also, the model can know what other members of the same household are doing, allowing for joint-travel and intra- household interactions to be modeled. The bottom line is that with less information loss the models can be more powerful. Typical Model Flow To illustrate the major components of activity-based models an example model structure is described here and contrasted with the structure of an example trip-based model. The activity- based model is assumed to be implemented in a microsimula- tion framework and the trip-based model uses the traditional fractional probabilities approach. In both cases a representative 3 Auto Ownership Levels 3000 Zone 3 ModesX XX 3 Purposes FIGURE 5 Example of traditional model application.

“good practice” model is described, rather than a specific model system in existence. The individual components are tabulated in a way to illustrate the parallels and differences between the systems. The basic structure of a good practice trip-based model is summarized in Table 2. Often such models are referred to as 4-step models, referring to the steps of trip generation, trip distribution, mode choice, and assignment. It is common, however, for such models to have more than four steps when components such as time of day, household submodels, and auto ownership are included. The left (stub) column of Table 2 indicates the main stage of the model, the center column lists the individual outcomes associated with those stages, and the right column shows how the data are stored. The main inputs to the model system include highway net- works, transit networks, and a list of households and employ- ment by TAZ. The networks are used to create level-of-service skim matrices representing the travel time and cost between each TAZ pair, and the household and employment data are stored as a list of the total in each TAZ. Often, trip-based models include a set of household sub- models that are used to disaggregate household characteris- 12 tics from averages to groups. This might include a submodel to convert from the average household size in each zone to the number of households in each zone of size 1, 2, 3, or 4+. Similarly, a submodel commonly converts from the average household income in a zone to the number of low, medium, and high income households. At this point, the model would be storing for each zone the number of households (usually a fractional number) in each size and income group. Next, the model would forecast any long-term choices, which would typically only include auto ownership. Auto ownership is a function of the household size, income, and usually some accessibility terms at the zonal level. Auto own- ership is included because it is such an important predictor of mode choice. Trip generation rates are applied to these vectors of house- holds by category, resulting in the number of trips generated in and attracted to each TAZ. The trip lists for each purpose here would be segmented by auto ownership and income for home- based trips, with no segmentation of non-home-based trips. Next the core trip-level models are applied. The first is destination, where trips are typically distributed using a gravity model. Mode choice is then applied using a logit choice model. HH ID Person # Trip # Autos Purpose ATAZPTAZ Mode … Work 1557 Other 1557 Auto Other 1557 Auto Work 152 School 152 1826 973 976 48 14 Other 152 179 Other 978 647 Other 978 395 Other 978 1792 Auto 1 1 1 2 2 2 3 3 3 3 1 1 1 1 2 2 1 2 3 4 1 2 3 1 1 2 1 1 1 1 1 1 1 2 2 2 2 2 2 2 Other 978 857 Auto Transit Walk Auto Auto Auto Auto FIGURE 6 Example of microsimulation model application. Model Stage Data and Outcomes Data Representation Inputs Highway network Transit network Households and employment by TAZ Household Submodels Number of households by income group and size Long-term Auto ownership Generation Number of trips by purpose Lists of total by TAZ Trip Level Destination Mode Time of day and peak spreading Matrices by TAZ Assignment Auto volumes on each link Transit volumes on each link Auto and transit travel times Loaded networks TABLE 2 STRUCTURE OF GOOD PRACTICE TRIP-BASED MODELS

13 Finally, the trip tables are factored by time of day, usually using fixed factors. Through each of these three steps, trips are stored in zone-to-zone matrices. The final step is to assign the trip tables to the networks, resulting in highway volumes on each link, transit volumes on each line, and congested travel times. Table 3 illustrates the structure of a good practice activity- based model. Again, the left column shows the main stages, the middle shows individual outcomes, and the right shows the format in which the data are stored. The inputs to the tour-based model are identical to those of a trip-based model, including the highway and transit networks and a list of households and employment by TAZ. The first step of the model is to create a synthetic popula- tion of households in the region. This can be done by select- ing representative households from the Public Use Microdata Sample (PUMS) to correspond to the aggregate characteris- tics of households in each zone. For example, if the zonal data show that TAZ 1 contains two middle-income, two-person households, then two households meeting those criteria will be randomly selected from the PUMS and associated with that TAZ 1. When this is done, all the other uncontrolled attributes of the household are included as well. Therefore, the first household might include two working adults age 45 and 50 with their associated occupations, whereas the second household might include a single mother and her teenage son. Additional dimensions can be controlled, such as number of workers, number of children, or age of householder, but if they are uncontrolled, the synthetic population will simply match what is found in the PUMS. The end result is a table with one record for each household and person in the region, with attributes representative of the regions’ actual population. Population synthesis is somewhat analogous to household submodels converting aggregate household data to something more disaggregate. Although both involve disaggregation, the difference is that population synthesis results in a fully disaggregate population that will be used with microsimula- tion methods, whereas household submodels still maintain fractional probabilities. As with the traditional model structure, long-term models are applied next. This includes auto ownership, as was done before, but adds a model of usual workplace location. The usual workplace location model predicts for each worker the zone in which their workplace is located. It is a bit different from a HBW trip distribution model in that the usual work- place location model predicts the work location even if the person does not go to work on the travel day, and even if the person chains trips such that there are no HBW trips. Thus, the result is directly comparable to the U.S. Census journey- to-work data. Workplace is included as a long-term decision based on the concept that it does not tend to change from day to day. Next, a set of generation models is applied to predict the number of activities by purpose, how those activities are formed into tours, and any joint travel. These models encom- pass trip generation, but provide significantly more information because they include the entire day’s pattern of activity and travel results. It is within this generation stage that the biggest structural differences between activity-based model systems exist. In the San Francisco model and its derivatives each per- son is given a choice of a daily activity pattern, and thus each person chooses the full package of what he/she will do during the day. In the Columbus model and its derivatives, tours are scheduled one at a time, with each building on the available time windows, and with consideration of intra-household joint travel. Although the workings differ the end result is the same—the set of activities and their composition into tours for each person. Coming out of this model there is now a list of tours in addition to a list of households and a list of persons. Data representationData and outcomesModel stage Inputs Highway network Transit network Households and employment by TAZ Lists of totals by TAZ Population Synthesis List of representative households with associated income, size, and other attributes Long-term Usual workplace location Auto ownership Generation Number of activities by purpose Formation of activities into tours Joint travel Tour Level Destination Time of day Mode Trip Level Stop location Time of day Mode List of each household, person, Assignment Auto volumes on each link Transit volumes on each link Auto and transit travel times Matrices by TAZ Loaded networks tour, or trip TABLE 3 STRUCTURE OF A GOOD PRACTICE ACTIVITY-BASED MODEL

The tour-level models are then applied to this list of tours. A primary destination is chosen for each tour. A hierarchy of activity importance is defined, with work or school at the top and serving as the primary destination if it is included in the tour. If the tour is a work tour, the primary destination is pre- determined by the result of the usual workplace location model. For non-work tours, the destination is chosen. A time-of-day model is applied to predict the departure and return time for each tour. The tour mode choice model predicts the primary mode of each tour based on the roundtrip level-of-service for each mode between the home location and the primary desti- nation. These three models are parallel to the trip-level models in a trip-based model, but applied to tours instead of trips. The list of tours is converted into a list of trips, and for each trip the trip-level models are applied. An important detail is that the trip-level models are constrained to be consistent with the tour-level models. For destination choice, the location of any intermediate stops is chosen taking into consideration the deviation from the path between home and the primary destination. For time of day, the timing of individual trips must be consistent with the timing of the tour as a whole and, in mode choice, the trip modes must be consistent with the tour modes. In mode choice, for example, if a traveler takes transit to work, he/she cannot drive home from work. The end result here is a full list of trips with destination, time-of- day, and mode details. This final list of trips is aggregated into zone-to-zone trip tables, which are assigned to the highway and transit networks. The assignment steps are the same as in a trip-based model. Models with Coordinated Travel The initial efforts at activity-based models in San Francisco and New York treated each individual as completely inde- pendent, with no knowledge of what the other members of their households do. In reality, however, it is very common for household members to jointly participate in travel and to coor- dinate their travel. The Columbus model made an important advancement in that it explicitly accounts for these intra- household interactions. Although the individual-traveler mod- els can include variables such as household size, they cannot capture the behavior in a consistent way across household members. Modeling a person without the context of a house- hold is like modeling a trip without the context of the tour. Accounting for joint travel is important, because members of the same household account for a significant majority of carpools, and understanding joint travel improves the ability of the model to forecast travelers’ willingness or lack thereof to carpool outside the home, and the effectiveness of high- occupancy vehicle lanes and regional carpool promotion poli- cies. The difference between the two families of models is largely in the structure of the generation step seen in Table 3. 14 It is worth noting, however, that regardless of the approach used in generation, either model structure is a major advance over trip-based models. Research Models Several advanced activity-based models have been developed in the academic world. These include CEMDAP, developed at the University of Texas; FAMOS, developed at the Uni- versity of South Florida; TASHA; developed at the Univer- sity of Toronto; and ALBATROSS, developed at Eindhoven University of Technology in the Netherlands. Each of these models is an activity-based travel demand model that offers something of interest beyond what has been implemented in practice. CEMDAP, for example, simulates activities and travel in a continuous time domain. It has been implemented in the Dallas–Fort Worth region, and is being deployed by the Southern California Association of Governments. FAMOS uses the notion of time-space prisms to constrain the available travel and activities. TASHA is a 24-h activity and travel sim- ulation. ALBATROSS uses a decision tree method to imple- ment a series of if-then rules for making decisions. A detailed review of each of these models is beyond the scope of this document. From the standpoint of advanced practice, how- ever, there is room for some of the lessons learned from these academic models to be incorporated into practice. DYNAMIC NETWORK MODELS Static traffic assignment models route fixed specifications of demand through a network representation of the transporta- tion system of the modeled area. In the past the demand has typically been specified as daily flows, or the same divided into three or more periods of the day. The resulting link flows and travel times are important performance measures and model outputs, as well as serving as inputs to other compo- nents in the modeling chain. The topic of static user equilib- rium assignment has been extensively covered in the litera- ture, although much of it focuses on theoretical issues such as problem formulation and increasingly more efficient solution methods. With the possible exception of on-going vendor efforts to parallelize such methods it appears that from a pragmatic standpoint there have not been widespread break- throughs in static assignment methods over the past decade. The static methods are often criticized on several grounds. Traffic signals are not represented in the models, such that their major influence on system performance and delay must be captured through link capacity functions. Static models can- not represent queue formation and dissipation. Moreover, sta- tic assignment routes all of the demand through the network, regardless of whether there is enough capacity to accommo- date it. As a consequence, it is not uncommon to see large numbers of links in forecast years with a volume-to-capacity ratio in excess of 1.0. Although mathematically simple to

15 handle, such flows cannot be accommodated on the road- way, leading to severe congestion, long delays, and possible cancellation or postponement of travel. Even when splitting demand into several periods of the day it is difficult to capture time-varying traffic properties or control strategies and behav- ioral response to them. As a result, it is difficult to adequately model Intelligent Transportation System (ITS) strategies or other operational scenarios. The alternative is to relax the assumptions of invariant demand, abstract control representation, and effects of con- gestion. Traffic simulation models have been in use almost as long as travel forecasting models (Gerlough 1964; Brown and Scott 1970), with current models being capable of mod- eling large urban systems. However, such models are com- monly referred to in practice and in the literature as traffic simulation models. The same convention is adopted here to retain consistency and avoid confusion.] However, few attempts to model large urban areas have been completed, and none by practitioners seeking to use it as an adjunct to or replacement for traffic assignment. However, two some- what related technologies—TRANSIMS and dynamic traffic assignment—have emerged as viable alternatives. The term “dynamic network model” will be used where any of these three approaches can be described or used interchangeably, and by their specific terms otherwise. The expanding roles played by dynamic network models in transportation planning were illuminated in a recent workshop on Integrated Corridor Systems Management Plans held in Irvine, California. Co-sponsored by the California DOT (Caltrans) and TRB, the workshop focused on best prac- tices, which typically involve the analysis of urban freeway corridors 20 to 40 miles in length. The size of these corridors makes them regional in significance, yet their analyses require operational details typically obtained using only microsimu- lation models. The policies examined often include relatively low-cost improvements such as ramp metering, improved incidence response times, spot widening at chokepoints, and arterial signal timing optimization at or near exit and entry ramps. Case studies from Minneapolis and Monterey were presented. Although many studies still solely make use of traffic simulation models a small but growing interest in using dynamic traffic simulation was evident, especially for larger study areas. However, it was equally clear that such thinking is just now occurring, and opportunities for using newer approaches will become more common in the next decade. TRANSIMS Dissatisfaction with the four-step modeling paradigm led David Albright, then director of research at the New Mexico DOT, to approach experts in the field of large-scale simula- tion at Los Alamos National Laboratory (LANL) in 1991. His challenge to them was to start from the requirements of the newly enacted ISTEA and Clean Air Act Amendments and to design, from first principles and without recourse to practices in use, a systems dynamics approach to modeling urban transportation systems. The resulting white paper, long since lost, formed the basis for a proposal to the DOE that eventually funded the TRANSIMS initiative. The resulting broad design is shown in Figure 7. The Los Alamos team made significant strides in the devel- opment of the population synthesizer (Beckman et al. 1996) and began work on the route planner and traffic microsimula- tor. This was in contrast to the work of others, such as Kitamura et al. (1996) and other early researchers in activity-based modeling, who focused on the demand side of the equation (depicted as the activity generator in Figure 7) and largely ignored the supply side of the model. Regrettably, the two camps never collaborated to their strengths, largely as a result of the differences in opinion over how the overall framework would be structured or developed. By early 1995, the LANL team implemented a cellular automata microsimulator and the distributed computer infra- structure (hardware and software) necessary to implement it. Tests on small prototype networks revealed promising emergent traffic flow properties (Nagel et al. 1997). The first TRANSIMS case study was carried out in partnership with the North Central Texas Council of Governments in 1996–1998. Because the activity generator had not been attempted, static demand from the Council’s regional travel model was sliced into small time intervals, from which trips and tour were syn- thesized. The entire Dallas–Fort Worth region was included in the model at a coarse level, with North Dallas being modeled FIGURE 7 TRANSIMS architecture.

in great detail. The case study sought to prove the concept was viable, and the test was widely viewed as successful (Fed- eral Highway Administration 1999). A second case study was launched in Portland, Oregon, in the spring of 2001. The goal was much more ambitious, includ- ing plans to model the entire region using the router and microsimulator and to commence work on the demand side (activity generator) of the model. Achieving the latter would result in operational modules for each part of the system. Extensive network development and testing was undertaken, with two network configurations compared. The develop- ment team had long thought that an “all roads” representation of the region was necessary, whereas others believed that the network used for traffic assignment from Portland Metro’s regional model would prove adequate. By late 2004, tests of both approaches led to the conclusion that the latter could effectively represent network conditions, although continued refinement of the network and testing of transit networks continued afterwards. In 2005, the TRANSIMS team at LANL disbanded, having completed the work on the initial framework funded by inter- nal sources, DOE funds, and limited U.S.DOT support. The software was licensed as open source software a year later. During the same time AECOM undertook an extensive over- haul of several parts of the system to include porting the sys- tem to Microsoft Windows™ and improving the graphical user interface (GUI) and other user interaction components. Several more tests of the emerging system, carried out by aca- demics in several locations, continued to add to the knowledge base about the model. The open source implementation of TRANSIMS is the largest and perhaps the most successful such undertaking to date in transportation planning. FHWA decided on the desired outcomes, to include at the outset their long-term strategy for TRANSIMS and gradual transition to a user- supported community. The team made the decision to use an existing widely used open source license rather than writing one themselves. They also adopted common distribu- tion strategies, making significant use of Internet technologies such as web pages, wiki pages, and accessible version control systems. The result is a vibrant, online, user community that is supporting several initiatives and projects. The web- site for the current version of the software can be found at http://transims-opensource.net. FHWA established an early deployment program in 2000 and continued to seek case study locations for the system. In contrast to the Dallas–Fort Worth and Portland case studies, most of the current deployments are for smaller studies. A listing of current TRANSIMS research and applications are included in Appendix C. FHWA is also sponsoring research into the integration of TRANSIMS with activity-based travel demand models in Sacramento and Columbus. This work was beginning as this report was finalized. 16 Dynamic Traffic Assignment Dynamic traffic assignment (DTA) models occupy the middle ground between macroscopic traffic assignment and microscale traffic simulation models. They employ the familiar network and demand specifications of traffic simulation models, but operate at a much finer level of temporal detail. Because they typically employ link-based simulation models they produce more robust estimates of link travel times and costs. Research in DTA techniques began before the TRANSIMS project got underway, but was largely limited to academics working on its formulation and theoretical aspects. DTA overcomes the limitations of static models noted earlier, although at the cost of increased data requirements and computational burden. Moreover, software platforms capable of solving the DTA problem for large urban systems are just now becoming avail- able. Unlike the activity-based travel and land use models, all known operational DTA packages capable of handling large networks are proprietary or have closed code. DTA models can generally be classified by how they model intersection delay. Analytical DTA models treat it in the same manner as static equilibrium assignment models, with no explicit representation of signals. Link capacity functions, often similar or identical to those used in static assignment, are used to calculate link travel times. Analytical models have been widely used in research and for real-time control system applications, but not in practice. Simulation- based DTA models include explicit representation of traffic control devices. Such models require detailed signal param- eters to include phasing, cycle length, and offsets for each signal in the network. Delay is calculated for each approach, with vehicles moving from one link to the next only if avail- able downstream capacity is available. The underlying traf- fic model is often different, but at the network level they behave in a similar fashion. Demand is specified in the form of origin–destination matri- ces for short time intervals, typically 15 min each. Trips are randomly loaded onto the network during each time interval. As with traffic microsimulation models, adequate downstream capacity must be present to load the trips onto the network. The shortest paths through time and space are found for each origin–destination pair, and flows loaded to these paths. A gen- eralized flowchart of the process is shown in Figure 8. As with static models, the process shown in the figure is iteratively solved until a stable solution is reached. The memory and com- puting requirements of DTA, however, are orders of magnitude larger than for static assignment, reducing the number of iter- ations and paths that can be kept in memory. Instead of a single time period, as with static assignment, DTA models must store data for each time interval as well. A 3-h static assignment would involve only one time interval. A DTA model of the same period, however, would require 12 inter- vals (assuming 15 min each). These are all in addition to the memory requirements imposed by the number of user classes and zones.

17 Although its use in planning studies was perhaps always intended (Peeta and Ziliaskopoulos 2001), most of the early investigations focused on freeway control and ITS applica- tions (Van Aerde and Yagar 1988; Mahmassani et al. 1993). Only a few large-scale applications of the model in tandem with regional travel demand models have been attempted. Dynameq has been successfully applied to a large subarea of Calgary and to analyses of the Rue Notre-Dame in Montréal. Although user group presentations of both applications have been made, and reported very encouraging results, the work is currently unpublished and inaccessible except through contact with the developers. The largest known DTA application to date is described by Hicks (2008). The network from the Atlanta Regional Commission’s (ARC’s) travel model formed the starting point for the DTA network. Intersections were coded, cen- troid connectors were re-defined, and network coding errors were corrected. A signal synthesizer derived locally optimal timing parameters for more than 2,200 signalized intersec- tions in the network. Trip matrices from the ARC model were divided into 15-min intervals for the specification of demand. Approximately 40 runs of the model were required to diagnose coding and software errors. Unfortunately, the execution time for the model was approximately one week per run. The resulting model eventually validated very well to observed conditions; however, the length of time required to render it operational and the run time required prevented it from being used in studies as originally intended. Subse- quent work by the developer has resulted in substantial reductions in run time, but this remains a significant issue that must be overcome before such models can be more widely used. A number of cities are currently testing DTA models, but are not far enough along in their work to share even prelimi- nary results. At least a dozen such cases are known to be in varying stages of planning or execution, suggesting that the use of DTA models in planning applications is about to expand dramatically. However, in addition to the issue of long run times, a number of other issues must be addressed before such models are likely to be widely adopted. • The integration of DTA and travel demand models has only been attempted on an ad hoc basis, although the topic has received considerable research interest (Boyce 1986; Lin et al. 2008). Operational models formally incorpo- rating feedback between the two modeling realms will be attempted as part of the SHRP 2 C10 project which will run from 2010 to 2012. Parallel work in two cities—Sacramento and Jacksonville (Florida)—will be pursued as part of that project. • Transit has not been credibly tackled in DTA models, and considerable work will likely be required before it catches up with the existing capabilities for handling auto trips. It might be possible to combine static transit assignment methods with DTA models of auto and truck flows on an interim basis, feeding link travel times from the latter to the former. • Traffic signal timings have a significant effect on net- work performance. However, most of the research on DTA models has been on node-abstract analytical solu- tions. Practical and scalable methods for developing signal timing inputs to regional DTA models have yet to emerge despite considerable evidence of its influence on capacity and operations (Berg and Do 1981; Boyce et al. 1989; Rakha and Van Aerde 1996). • Criteria for the validation of such models have not been widely accepted. The paucity of traffic counts in most urban areas, and especially at 15-, 30-, or 60-min inter- vals, is a significant barrier to definitive assessment of these models. It can be noted that these shortcomings apply equally to TRANSIMS and traffic simulation models, although the path to overcoming them may vary by platform. FIGURE 8 Typical DTA solution methodology.

Traffic Simulation Models Traffic simulation models have been used for several decades to conduct detailed analyses of roadway designs and opera- tional plans. Individual vehicles traverse detailed networks in very short time steps (typically 0.5 to 5 s intervals) in these models, and they explicitly model driver behavior such as lane changing and car following. Initially developed about the same time as the first four-step models, they likewise were initially executed on mainframe computers. However, whereas four-step models went on to become institutionalized in the urban planning process under strong federal leadership, traffic simulation models progressed at a slower rate. Over the past several decades traffic simulation models have caught up with travel demand models in terms of sophistication, quality of software, experience in practice, and to some extent, suit- ability for large-scale planning studies. They excel at visual- ization. Once in the domain of different specialists (traffic engineers and transportation modelers) that seldom collabo- rated, the two streams are converging as they are being used more in combination. As with the other types of dynamic network models dis- cussed, traffic simulation models typically use the familiar concepts of roadway networks and trip matrices. The former are typically far more detailed in geometric representation and lane configurations, and the latter are not only spatially more detailed but also more detailed in temporal respects as well. Such models also require explicit coding of traffic detec- tors and control systems to include traffic signal timing plans. Unlike DTA models, some traffic simulation packages can also optimize signal timings, reducing the amount of input data required to deploy them. Owing to the amount of data required to develop them and their heavy computational requirements traffic simulation models have traditionally been restricted to small area studies, often encompassing no more than a few dozen traffic signals and the detailed land use patterns and networks accompany- ing them. Other simulation models were developed specifi- cally for the study of freeway corridors. In recent years the advent of GIS, remote sensing capabilities, online traffic data, and signal timing optimization programs has enabled these models to be used for successively larger study areas. GUIs have significantly eased the coding and checking of input data, and some packages have interfaces to macroscopic traffic assignment models. Aside from TRANSIMS there have been few attempts to replace traffic assignment models in urban areas with traffic simulation models. Rickert and Wagner (1996) built a model of the German Autobahn network, and Rakha et al. (1998) described the application of the INTEGRATION model in Salt Lake City. Both were prototypical applications that did not lead to their use by public agencies, although further work was anticipated in both instances that would have resulted in calibrated and useful models. There are no known on-going 18 attempts in North America to microsimulate traffic flows for entire urban networks outside of the TRANSIMS program. At the corridor and subarea level, traffic simulation mod- els are much more widely used in urban areas, with varying degrees of integration with travel demand models. In many instances, traffic count data and origin–destination matrices are used to develop demand estimates for the traffic simula- tion model. The former are usually directly transferable if in small enough units of time (15-, 30-, or 60-min intervals). The latter are generally too coarse for direct use in such mod- els. Intermediate steps to divide the data into the finer tem- poral intervals needed by the model, often based on observed diurnal patterns, are required. Most TAZs in travel demand models are too large to serve analogous roles in traffic simula- tion models, requiring an additional step to divide the origin– destination flows into sinks and sources within zones. Getting traffic dynamics right in traffic simulation models entails coding traffic entering and leaving the network where they do in real life, which is often small feeder streets, parking lots, etc. These are known as sinks and sources. Several, and per- haps dozens, of them might be coded in the same area as a single TAZ. Such allocations are typically based on counts, field observation, or aerial or satellite photography. The tour- and activity-based models described earlier can mesh with traffic simulation models in several ways. Most tours and their constituent trips are assigned a discrete starting time, or grouped into finer intervals than possible with trip- based models, obviating the need to slice peak periods into the finer intervals required by traffic simulation models. Because synthetic population generators create individual households it is possible to geocode them as point locations, although fur- ther research is required to ensure realistic outcomes. Agen- cies with land use models will find it easier to get such micro- positioning correct. However, these synthetic households will only rarely match the characteristics of a real household at that location. The aggregate characteristics of the population at the level at which the synthesis was constrained (typically census tract or public use microdata area) will match, but indi- vidual observations within them will not necessarily do so. The manner in which such models can work together can be illustrated through several examples. One is the afore- mentioned implementation in Atlanta. However, the major focus was on the traffic simulation models, which were used to evaluate a wide variety of operational and design issues relating to I-285, a 64-mile circumferential freeway. Traffic simulation models were the only ones sensitive enough to cap- ture the effects of some of the contemplated changes. How- ever, it was also recognized that a number of factors well beyond the freeway affected the demand, such as broad demo- graphic shifts and changing travel behavior, particularly with respect to growing congestion in the region. After consultation with experts in the field, it was decided to construct a three-tier model to study the I-285 freeway, as

19 shown in Figure 9. Demand data and networks from the regional model were linked to a DTA model of the entire region, with emphasis (in detail as well as calibration) on the study area. As noted earlier, a matrix variegator was devel- oped to divide the peak period regional trip matrices into the 15-min intervals required by the model, using observed peak- ing characteristics and travel survey data (Simons 2006). Synthetic matrix estimation could have been used to enforce consistency between the demand data and traffic counts, but was not required. Some regional strategies, particularly those related to ITS, were evaluated at the DTA level. Finally, demand data from the DTA model was mapped to the traffic simulation model, the level at which most scenarios were tested. Performance measures were compiled at that level as well. Because the corridor was already congested and most signals operated on pre-timed plans during the peak periods the existing timings and parameters were used in the traffic simulation (VISSIM) models, although provision for using the Synchro model for timing optimization was built in. Opportunities for feedback to the DTA model handled cases where excess demand was supplied to the traffic simulation models. Microsimulation was also used in tandem with the activity- based travel model maintained by San Francisco County Transportation Authority (SFCTA) to study alternatives for the reconstruction and seismic retrofit of the major roadway providing access from the city to the Golden Gate Bridge. As with the I-285 study, demand matrices from the regional model were adapted for use in the traffic simulation model (Paramics) used in the study. Both spatial and temporal disaggregation was required; however, as in Atlanta, this process obtained reasonable travel patterns without recourse to further pre- processing (e.g., synthetic matrix estimation). Once developed and validated the model was used to study a number of sce- narios, to include sensitivity testing of electronic toll collection impacts on the bridge toll plaza, various roadway design alter- natives, incident simulation, and construction staging. None of these scenarios could be addressed using the static traffic assignment model used with the SFCTA model. Moreover, the measures of effectiveness—which included estimates of delay at each intersection, incidence and severity of bottle- necks, and other operational metrics—are below the level of resolution of regional models. The animated display of traffic flows, now a standard feature in traffic simulation packages, proved invaluable for conveying model outcomes to the public. AIR QUALITY MODELING The output of traffic assignment in regional travel demand models is routinely transformed into the inputs required for emissions models used for air quality conformity analyses. In the recent past the MOBILE6 package (Vehicle Emissions Testing Software) developed by the EPA has been used for such analyses, with California using a variant known as EMFAC (EMission FACtors). Both models produce estimates of tailpipe emissions by vehicle type of various pollutants to include nitrous oxide, carbon monoxide and dioxide, partic- ulate matter, and other toxins. The programs have evolved over time to incorporate changes in fuel composition, typical driving behavior, and improvements in the test data used to develop the various exhaust and evaporative emission rates under varying driving cycles. Additional toxins have been added as well. These models can be applied at varying levels of geography, allowing them to be used for national-, regional-, and project-level analyses. The processing of travel model outputs required to use these models includes an exoge- nous estimate of average vehicle speeds based on link flows and capacity and analysis of the detailed peaking characteris- tics of travel within the study area, such that estimates of vehi- cle miles of travel by vehicle class and speed can be generated from the assignment results. Aside from continued improvements to the models them- selves there has been little change in how they are used in the transportation planning process. Their inputs are specified in exactly the same manner whether a sketch planning aggre- gate travel model or highly detailed dynamic network model is used to generate them. In the latter case, the preservation FIGURE 9 Multi-tiered modeling approach for the I-285 study.

of detailed microdata about the vehicle’s drive cycle and the underlying travel plans giving rise to it can be mined, allow- ing for higher granularity in the inputs. That is, a larger num- ber of traveler–vehicle categories can be used to permit more detailed levels of analyses. Insomuch as a dynamic network model provides more accurate depictions of microscale driving dynamics, the resulting estimates of travel time and delay will obviate the current requirement for post- processing static assignment results to obtain more realis- tic travel speeds. The EPA has recently developed the MOVES (Motor Vehicle Emissions Simulator) model to replace the MOBILE family of models. Information about the MOVES model, as well as the software, can be found at http://www.epa.gov/ otaq/models/moves. As with its predecessors, it is built from a vast amount of vehicle testing data. However, MOVES is more than simply an update of the previous models. The underlying software has been largely rewritten, and now includes a GUI. MOVES is capable of modeling emissions from the national level down to that of individual projects. Although not yet formally approved for air quality confor- mity analyses at the time of this writing, the software is expected to replace MOBILE6 in the near future. In some respects it remains a work in progress, with some issues relat- ing to compatibility between the various beta versions that have been made available. However, it represents a major advancement over previous emissions models, especially in how it estimates greenhouse gas emissions. Similar efforts are underway to update the EMFAC model in California. MOVES2010 and EMFAC2009 are the current versions of these packages. The MOVES model will continue as a standalone program, requiring data in its unique format. The situation has been improved somewhat by its ability to read required inputs from relational databases, which will ease the translation from net- work models. One logical progression is the convergence of emissions and dynamic network models, where the former becomes a standard metric calculated by the latter. Such could be achieved by tightly coupling the two packages so that they share a common database, exchange data dynamically using shared libraries or application program interfaces, or are fully integrated pieces of software. The TRANSIMS emissions estimator is an example of such integration. Although it can- not be used as a substitute for MOBILE or MOVES for con- formity analyses, it can be used seamlessly to assess the envi- ronmental impacts of the scenarios tested with TRANSIMS. Whether other dynamic network models provide comparable functionality or tight linkages to the MOVES model remains to be seen. Similarly, tools or procedures for incorporating emissions into pricing analyses, which implies a feedback from emissions to travel demand model, have yet to emerge. Separate tools designed to evaluate carbon emissions have begun to appear, in part because the MOBILE package was not thought to deal with them satisfactorily. The GreenSTEP 20 model developed by Gregor (2009) is perhaps one of the most developed to date. It goes beyond being an emissions calcu- lator by instituting a framework for the codification of the effects expected from and rapid testing of a large number of emission reduction strategies. In Oregon it is designed for application at the statewide level, although nothing in its design precludes it from being used for metropolitan areas. As with MOVES, it contains a variety of user-defined inputs for vehicle fleet composition and their changes over time, fuel characteristics, and prices. However, it also generates a syn- thetic population and assigns travel characteristics to it, as well as capturing some aspects of land use. The resulting esti- mates of greenhouse gas emissions and fuel consumption can be used to screen a number of strategies, alone or in combi- nation. The structure of the model is shown in Figure 10. GreenSTEP and similar tools complement, rather than substitute for, more detailed urban and statewide travel demand models and emissions calculators such as MOVES. Its strength lies in its statewide focus, straightforward inter- face, ability to test a large number of scenarios or combina- tions thereof, and ability to define travel characteristics and other inputs that are consistent with other models used by the agency. The need for a tool like GreenSTEP may be reduced over time as more of its features are incorporated into main- stream travel models; however, in the near term it fills a niche not currently filled by other tools. LAND USE MODELS Simulating land use improves the outcome of the transporta- tion model. By explicitly simulating the land use and trans- port interactions, observed behavior of traveling, household relocation, job change, shopping location choice, etc., may be modeled more realistically. It also creates a logical con- sistency between land use and transportation forecasts, and the performance measures derived from them. The transportation and land use systems closely interact, as illustrated in Figure 11, by the land use/transport feedback cycle (Wegener 2004, p. 130). Starting at the bottom of the cycle (Land Use), the locations of population and employ- ment determine the origins and destinations of most trips sim- ulated by travel models (Activities). The simulation of the Transport System allows calculating Accessibilities, which describe how accessible one zone is to all other zones. This accessibility shapes land use. Both households and businesses search for locations that are—among other location factors— readily accessible. This closes the land use transport feedback cycle. Traditionally, changes in the distribution of households and employment are given exogenously based on a consen- sus forecast. Simulating land use allows for generating logi- cally consistent forecasts that are independent of stakeholder interests, which commonly affect consensus forecasts. Further-

21 by Forrester (1969) was a milestone in explicitly simulating businesses, dwellings, and households. The Lowry Model (Lowry 1964) acquired even more popularity. Its relatively simple model structure allowed for the development of many applications, some of which are still in use or undergoing even further development today. A large variety of different land use model approaches cur- rently is in operation. Comprehensive overviews on existing land use models may be found in Kain (1987), Wegener (1994, 1998, 2004), Wegener and Fuerst (1999, p. 42 ff), the U.S. Environmental Protection Agency (2000, p. 27 ff), Kanaroglou and Scott (2002), Timmermans (2003), or Hunt et al. (2005). Although the motivation of many land use models are land use-related analyses, most of these models are integrated with travel models. Therefore, the majority of these mod- els support analyzing how land use models may improve travel models. Design Principles of Land Use Models At the detailed level, most land use models work differently; however, many land use models are designed based on a sim- ilar rationale as shown in Figure 12. A common design prin- ciple is the distinction of three major players on the land use FIGURE 10 GreenSTEP model structure. FIGURE 11 Land use–transport feedback cycle (Source: Wegener 2004). more, consensus forecasts commonly forecast land use at the municipal and in some cases at urban district levels. A land use model, in contrast, generated land use forecasts at any geographical level of interest. Ever since personal computers became available for aca- demic research, land use simulation models have been devel- oped. The pioneering work of Herbert and Stevens (1960) and Harris (1966) was fundamental in exploring how com- puter models may be applied for urban analysis. Though aspatial in design, the theory of urban interaction developed

market: population, employment, and developers. Whereas population and employment are using dwellings or floor- space to locate, developers build additional dwellings or floorspace based on demand and available developable land. Accessibilities influence location decisions of population, employment, and developers. The updated locations of pop- ulation and employment are then fed back into the travel model. Commonly, the simulation of households is done in two steps. Demographic changes are aspatial changes, such as ageing, marriage, birth of a child, divorce, death, change of educational level, receiving a driver’s license, etc. House- hold moves are the spatially explicit relocation of house- holds. Many demographic changes trigger household moves. For example, a daughter who leaves the parental household needs to find a dwelling to establish her own household. The same distinction is commonly made for the simulation of employment. Firmography, a contraction of firm and demography, is the study of the structure and evolution of businesses. Such changes are simulated in the firmographic module including nonspatial events such as business establishment, growth, decline, or closure. Business relocation is the move of the entire firm or of a part of the firm. Again, some firmographic events, namely business establishment and growth, may trigger a business move. Commonly, nonspatial and spatial events are distinguished because they are simulated differ- ently. Often, aspatial events are steered by an economic model simulating employment and population growth or decline. Another distinction is that land use policies may be tested that aim at influencing location behavior; however, policies are seldom tested that influence demographic changes. Hence, spatial modules are designed to be sensitive to a large vari- ety of policy changes, whereas nonspatial modules simulate events that happen to population or employment. 22 Although the basic design structure is similar for most land use models, there are at least three fundamental design features handled differently in several land use models: 1. Behavioral or structure-explaining approach, 2. Bid-rent or discrete choice approach, and 3. Aggregate or microscopic simulation. Behavioral approaches aim at simulating the explicit behavior (such as marriage, birth, or relocation), whereas structure-explaining approaches attempt to simulate the out- come (such as population distribution) directly without deal- ing with the motivation that led population to be distributed in a certain way. Certainly, this distinction is vague and many models are somewhere between these two approaches. The model design in the example in Figure 12 shows a behavioral approach, as the behavior that leads to a certain distribution of population and employment is simulated explicitly. A common example for a structure-explaining model is a Cel- lular Automata that simulates the state of a single raster cell based on the state of the surrounding raster cells. Raster cells are equally sized quadratic or hexagonal zones that cover the entire study area. Being equal in size the zonal system simpli- fies simulating spatial interaction between neighboring cells. Cellular Automata models do not explain the change of a raster cell, but rather simulate the outcome. Structure-explaining models tend to be less sensitive to policy scenarios because behavior is not represented in the model. However, Cellular Automata allow for building a land use model even if few data are available. As a consequence, many Cellular Automata models are implemented in developing countries where data availability is limited. A classic distinction in land use models is the bid-rent approach and the discrete-choice approach. The bid-rent theory FIGURE 12 Common design of urban land use models.

23 was first developed by Alonso (1960). According to this theory, every actor on the land use market is making bids for a piece of land, and the bidder with the highest offer gets the land. Because of transportation costs, everybody is willing to bid more for a location in the city center than for a location on the outskirts. Because most office firms value transportation costs more highly than most households, the office employment makes the higher bid in the city center, whereas the house- hold bids higher in the suburbs. This explains why office buildings are located downtown while most residential areas are in the suburbs. The discrete-choice theory was devel- oped by Domencich and McFadden (1975). Frequently, logit models are used to implement the discrete-choice approach. Households, firms, and developers make choices among a finite set of alternatives. The utility of every choice is used to select one alternative; the higher the util- ity of a given alternative, the greater the probability this alternative will be selected. Not everyone chooses the per- fect solution, but some deviation from the optimum distri- bution is implemented. An advantage of the bid-rent approach is that prices are sim- ulated endogenously in the bidding process. A well-calibrated model generates realistic prices that equal the highest bid made for every location. To reach the equilibrium price, the model needs to iterate a few times until prices are found and no one is willing to make a higher bid for any location. The bid-rent approach assumes market transparency and users who maxi- mize their profit. The discrete-choice approach requires an additional land-price model, as prices are not updated auto- matically. Limited information is introduced explicitly in the discrete-choice approach by logit models: owing to a lack of time and money to analyze all alternatives as well as the result of personal preferences, habits, and prejudices, some users make seemingly nonoptimal choices. Overall, actors in the discrete-choice approach aim to satisfy their needs and not to maximize their profits. Martinez (1992) has shown that the two approaches lead to similar model results. As a rule of thumb, bid-rent approaches work best in markets that are highly competitive and transparent. Discrete-choice approaches work better in markets that react with some time lag and in which users have to make decisions at a certain level of uncertainty. The third characteristic analyzed in this context is the dis- tinction between aggregate and microsimulation land use models. Aggregate models cluster actors into certain groups, such as households by household type or firms by industry type. All actors in each group are assumed to have homoge- nous preferences. With a smaller number of groups, aggregate models store data efficiently and tend to have shorter run times. If more detail is added to the model, the handling of many groups may become cumbersome, and a disaggregated approach may become more appropriate (see the earlier dis- cussion of activity-based models). Ever since Orcutt (1960) introduced microsimulation, both land use and transporta- tion models have been developed that simulate every actor individually. The great advantage of microsimulation is the explicit simulation of the interaction between individuals. Hägerstrand (1967) proved in his theory of spatial diffusion how innovations are spread from a single actor to other actors who live in spatial proximity. Individuals who received the innovation become a sender themselves, further spreading this innovation at the microscopic level. Nobel Prize laureate Schelling (1978, p. 147 ff) showed with the self-forming neighborhood model how microscopically simulated house- holds choose more segregated locations than the aggregate segregation desire would suggest. As discussed in the earlier section on activity-based models, microsimulation models allows for storing complex data sets more efficiently. Often, microscopic approaches are easier to communicate, as describing the behavior of single actors is less abstract than describing the homogenous behavior of groups. Because microscopic models simulate individual inter- action explicitly, model results tend to be more coherent with urban theory. However, model developments obsessed with adding ever more detail do not lead to the best models. By adding detail, model run times may suffer, and in some cases the complexity of the model may exceed time and budget allo- cated to the model development. Microsimulation models require a random number generator to simulate choices. With different random numbers in each model run, the results in every run are slightly different owing to the sto- chastic variation. This difference is insignificant if a very large number of actors are simulated (such as a location choice of 1 million households). Stochastic variation makes model output invalid if the output is analyzed at a detailed level (such as location behavior of a hundred households of household type 1 in neighborhood A). If microsimulation is applied, analyses of model results may only be done at a fairly aggregated level. Examples of Land Use Models in Practice After a wave of urban model developments in the 1960s, increasing deregulation and a shift from the synoptic plan- ning paradigm to incremental planning decreased the inter- est in urban models outside academia. Whereas the synoptic planning paradigm sought to take into account many plan- ning aspects simultaneously and hence favored comprehen- sive modeling, an incremental planning approach aims at addressing single issues one at a time, which requires less understanding of the big picture. In the 1990s, the interest in urban models as a planning support tool was revived as a result of federal regulations by the EPA and a general disappointment with the success of market-driven incre- mental planning approaches. Ever since, several models have been developed that are applied in practice. To provide an overview of the most common land use models, seven have been selected that have been applied to more than one study

area and as such been in practice outside of academic research projects. The models discussed here are, in alphabetical order, DELTA, LUSDR, MEPLAN, MUSSA, PECAS, TRANUS, and UrbanSim. Though LUSDR has been applied so far to only one study area, it has been included here as an operational model with multifaceted policy applications. The DELTA model has been developed by Simmonds (1999, 2001; Simmonds and Feldman 2007). The acronym DELTA was derived from the five major sub-modules: Devel- opment, Employment status and commuting, Location and property market, Transition and growth, and Area quality. This aggregate model simulating economic growth and land use changes may be applied at the regional or urban level. At the urban level, DELTA simulates developers, household location, demographic changes, auto ownership, and employ- ment location. At the regional level, DELTA adds modules that simulate long-distance migration and an economic model. All location decisions are based on the discrete-choice approach using logit models. DELTA has been applied to several cities and regions in the United Kingdom, as well as the Auckland region in New Zealand. The Land Use Scenario DevelopeR (LUSDR) was designed and implemented by Gregor (2007) for the Rogue Valley MPO in Oregon. A guiding principle was to build a policy-sensitive land use model that limits data requirements to a minimum. LUSDR creates a synthetic population with households and employment establishments. An iterative process allocates households and firms to development types. Subsequently, development types are allocated to zones based on their land availability, plan compatibility, prices, and accessibility using Monte Carlo sampling. The bid-rent approach is used to adjust floorspace supply based on the demand by households and firms. The model was developed using the R statistical pro- gramming language. Currently, an application of the LUSDR model is under development for the Salem–Keizer MPO in Oregon. Another Oregon land use model, MetroScope, has been used by Portland Metro for almost two decades (Conder and Lawton 2002). Originally developed as a land consump- tion model, it includes both residential and nonresidential supply and demand components that include elaborate and sophisticated econometric models. Pivoting off of externally specified amounts of land supply and zoning the model esti- mates land consumption by tenure, type, and location. In recent years it has been integrated with Portland Metro’s regional travel model. MEPLAN was developed by Echenique et al. (1969, 1990). This aggregate land use model initially was based on the Lowry Model for distinguishing basic (exporting) and non- basic (supplying the local market) employment. MEPLAN is integrated with an aggregate transportation model, and both land use and transportation models iterate until equilibrium is reached. Land use is simulated as economic activities by households that live in housing units and employment that is located on floorspace. An economic input–output model 24 simulates the flows of goods and the required labor to feed the economy. MEPLAN has been applied to many regions worldwide. MUSSA (Modelo de Uso de Suelo de SAntiago) was developed by Martinez at the University of Chile in Santiago (Martinez 1996; Martinez and Donoso 2007). MUSSA con- tains a microeconomic approach to simulate demand and supply of real estate. By developing new floorspace, equi- librium between demand and supply is reached. A logit model is used to simulate bids of users that are constrained by the available budget. Developers add real estate based on expected rents and construction costs, while land use regulations are used as a constraint. The model has been developed with a GUI that allows running the model and visualizing the output. In cooperation with Citilabs, Cube Land, which integrates MUSSA with a transportation soft- ware, was released recently. PECAS (Production, Exchange and Consumption Alloca- tion System) was developed at the University of Calgary by Hunt and Abraham (2003). A land developer module simulates the behavior of real estate developers. An aggregate system simulates the exchange of goods, services, and labor. Prices are defined in an equilibrium process. Flows from production to consumption are allocated by nested logit models using prices and transport disutilities as impedance. A large num- ber of PECAS applications are under development in the United States, including those for the state of Oregon, state of California, Montgomery MPO in Alabama, and the Balti- more Metropolitan Council in Maryland. PECAS is one of the few land use models that has been applied both at the urban and the statewide levels. TRANUS (Transporte y Uso del Suelo) was been devel- oped) in Venezuela by de la Barra and colleagues (de la Barra and Rickaby 1982; Barra et al. 1984, 1989). As an integrated land use transport model, TRANUS simulates location of activities, the real-estate sector, and a multimodal transporta- tion system. Based on the Lowry Model and similar to the MEPLAN model, TRANUS distinguishes basic and non- basic employment. Change of employment in the basic sector is allocated first, and non-basic employment is treated as induced demand. An equilibrium approach iterates between changes in demand and supply to simulate land rents. TRANUS has been applied to cities and regions in America, Europe, and Asia. UrbanSim was developed at the University of Washing- ton by a team led by Waddell (Waddell 2002; Waddell et al. 2003). This microscopic model simulates households, employ- ees, developers, and real estate prices. Location decisions are simulated based on multinomial logit models. To select a location, a uniform distribution is used to randomly sample a set of nine alternatives in addition to the site with the high- est utility. The final location is selected from these ten alternatives. Land values are updated by hedonic regression.

25 Hedonic models are common in real estate analyses and esti- mate how much the individual characteristics of the land con- tribute to its value. Recently, the spatial resolution was increased from raster cells to parcels. UrbanSim applications are under development in many urban regions worldwide. Today, PECAS, UrbanSim, and TRANUS are common land use models in the United States. As different as the design concepts of the three models are, all allow for expanding pure transportation models to integrated land use/transportation models. The two Latin American and the two English mod- els described previously also have been shown to success- fully integrate land use and transportation simulations. Furthermore, several academic land use models have proven to be operational and are promising to provide useful tools for land use analyses (Wegener 2004). There is no one model that fits all purposes; the best selection for an agency must be based on its requirements, capabilities, and resources, with particular emphasis on the scale and type of land use ques- tions that will be studied. If time and funding permit, a custom- made model can even allow for tailoring a land use model to the specific local needs. FREIGHT AND COMMERCIAL MOVEMENT MODELS The state of practice in urban freight modeling remains far behind that of person-travel modeling, especially with respect to advanced modeling concepts. Although there has been two decades of intensive research and development of activity- based models, there is virtually no comparable activity in freight or commercial travel modeling. Historically, this has been the result of a lack of emphasis on freight, owing to its predominately private-sector nature and the relatively low percentage of trucks on most urban roadways. Data are diffi- cult and expensive to collect compared with person travel, and the underlying behavior is more complex. Multiple decision makers, some with conflicting goals, influence the transporta- tion choices made in the distribution of freight. As a conse- quence, little has been accomplished in this field over the same period of time that other advanced models have flour- ished. However, there are encouraging signs of increased progress and momentum in this area, ranging from improve- ments in existing techniques to emerging advanced models. Trip-Based Urban Truck Models In some respects having an explicit freight or commercial vehicle model at all might be considered an advanced mod- eling practice, as most urban areas have relatively simplistic representations of commercial vehicle flows. In many cases, they amount to little more than growth factoring of long-ago observed or imputed truck trip matrices. In other cases, a par- tially or completely synthetic four-step sequential modeling process is employed: 1. Trip generation (typically carried out for specific truck classes rather than by trip purpose) 2. Trip distribution 3. Time-of-day factoring 4. Traffic assignment. Mode choice is not modeled explicitly, because the mode (truck type) is implicit in trip generation. Moreover, virtually all commercial flows within urban areas are by truck, offer- ing little opportunity for mode choice. Freight mode choice is typically a function of the existence of carrier contracts, price differentials between competing carriers and modes, reliability concerns, and other factors not represented in typi- cal urban transportation models or networks (Donnelly 2007). However, there is increasing evidence that pricing strategies may also influence mode choice (as well as other decisions) within urban areas (Zamparini and Reggiani 2007). Traffic assignment is typically carried out in conjunction with person-travel flows using a multi-class equilibrium assign- ment. Uncommon only a decade ago, multi-class assignments now appear to be the norm in most instances; not only do they permit the concurrent assignment of different truck classes, but they also permit partitioning of person-travel demand by vehicle occupancy or toll use. A number of such models have been successfully imple- mented. An important NCHRP synthesis of freight modeling was prepared by Kuzmyak (2008). It contains a comprehensive review of recent models including case studies of practice- leading models in Ohio, Oregon, Los Angeles, and Calgary. Profiles of other urban freight models in an additional eight cities are included. Most reported using surveys to build the traditional four-step model described earlier. Most are freight models, although two model only trucks and some attempt to incorporate flows through trans-shipment points. The work in Los Angeles is noteworthy in that they have recently invested in the development of a new heavy-duty truck model and linkages with air quality models. As with some of the others, it also includes explicit handling of trans- shipment centers. A recently completed revision of the Quick Response Freight Manual (Beagan et al. 2007) is another important tool for modeling urban freight. As its name implies, it does not include nonfreight commercial movements. However, it does represent a tremendous improvement in content and organi- zation over the previous 1996 version, making it a valuable resource that will lower the barriers for agencies lacking the resources to complete surveys and model development activities of their own. A second valuable resource is a col- lection of sketch planning and aggregate modeling techniques used to account for all commercial vehicles (both freight and non-freight) in urban areas developed by Cambridge System- atics et al. (2004). As with most other such models practition- ers are familiar with them. Their work is significant not nec- essarily from a methodological standpoint, although it does nicely bring together several traditional methods in a cohesive framework, but rather because it is almost singular in its

ability to address the full range of commercial vehicle travel occurring in urban areas. Synthetic Matrix Estimation Models Synthetic matrix estimation (SME) techniques have been employed by some researchers and practitioners to help over- come the paucity of spatially indexed behavioral data. Despite some differences in solution method, all of these models attempt to adjust an estimated, obsolete, or partially observed trip matrix to match observed traffic counts. Earlier methods typically employed maximum likelihood estimates of maxi- mum entropy to arrive at a solution. Such models often suffered from unexpectedly large differences in outcomes owing to small changes in inputs (Van Aerde et al. 2003), as well as their inability to reconcile inconsistent or erroneous traffic counts (Yang and Zhou 1998; Hazelton 2003). A substantial amount of literature exists on this topic, although almost all relates to person-trip modeling and estimation of origin– destination flows for traffic control systems. Munuzuri et al. (2004) developed an SME model for truck movements in Seville that included five different retail markets and one for home deliveries. The demand was consolidated into a single seed matrix and adjusted using a gradient descent method developed by Spiess (1987, 1990). More recent formulations have permitted multiple sources of data with reliability estimates attached to each, the ability to handle multiple classes of vehicles, and the use of linear pro- gramming techniques to reduce untoward responses to small changes in input (Logie and Hynd 1990; List and Turnquist 1994; List et al. 2001). Synthetic models are relatively easy to construct and have straightforward data requirements. However, they are not suitable for many types of analyses, owing largely to their lack of behavioral basis. Ríos et al. (2002) noted that the link counts themselves have the greatest impact on model accu- racy, which is hardly surprising in that the models use them as the constraint against which to work. However, the result- ing process is more geared toward replicating observed flows than explaining why they are there in the first place. These techniques are appropriate for evaluating network responses to changes in supply or operation, but cannot be used to address many of the issues facing policymakers and analysts in transportation planning agencies. An interesting variant on the approaches described here is one developed by Tardif (2003) in the Canadian province of Ontario. Using a database of approximately 78,000 roadside interviews at 240 locations, he built a database of truck trip records that included origin, destination, vehicle type, com- modity, and weight. Using truck counts at the survey locations as targets he employed sample enumeration to characterize total daily demand in Ontario, which could then be summa- rized or segmented as needed for the analysis at hand. Although sample enumeration has been proposed for activity-based 26 person-travel models (Kitamura et al. 1996; Shiftan and Suhrbier 2002), this application in freight modeling is unique. However, it suffers from the same limitations described earlier for other synthetic estimation methods. Criticisms of Current Practice Although arguably appropriate for modeling person-travel, the four-step sequential modeling process is inappropriate for analyzing freight flows. The motivation for and charac- teristics of person-travel are well informed by an extensive body of survey research and can be efficiently represented by relatively few market segments, homogeneous household and travel characteristics, and similar travel budgets. Most person-travel is characterized by roundtrips from home to principal destination, and back again. Stops are sometimes made along the way for secondary purposes, which the traveler does to simultaneously increase their utility while minimiz- ing travel cost. As noted earlier, recent advances in person- travel modeling have focused on the explicit representation of person tours or activity chains to better represent observed travel behavior. Unfortunately, freight does not emanate from or move according to the same principles as person flows. Freight flows are “the economy in motion,” the trade between pro- ducers and consumers that underpins modern economies. The factors driving the economy are more diverse and com- plex than those motivating personal travel, involve multiple entities (such as producers, carriers, distributors, regulators, and consumers), and are optimized to reduce the cost and uncertainty associated with their conveyance. Trucks are far less likely to make roundtrips serving only a single customer per day, because the lower productivity compared with trip chaining would be prohibitive for most firms. Indeed, within urban areas truck tours with several pickups and destinations comprise a significant share of observed truck flows. More- over, the widespread adoption of just-in-time and supply chain logistics has increased the use of distribution centers and trans-shipment terminals. A recent study of freight move- ments within and between Ontario, Canada, found that more than half of all truck trips involve such facilities (Donnelly et al. 2002; Tardif 2003). Network assignment processes that simply route freight between each origin and destination miss these important dynamics of freight. Not surprisingly, the resulting models do not accurately replicate observed conditions (Taylor and Button 1999; Wigan and Southworth 2005). Even if the origin–destination patterns were correct, the practice of routing each origin–destination interchange separately will still result in flow patterns that do not match observed conditions. Slavin (1979) developed a model that accounted for truck tours in Boston that is still elegant by today’s stan- dards, and found that it significantly improved the accuracy

27 of the model. Russo and Carteni (2005) formulated a tour- based urban freight distribution model as a series of nested logit models, proceeding from distribution strategy through first-stop choice to subsequent stop choices. The demand is specified exogenously. The model was successfully applied in Italy. Holguín-Veras and Thorson (2003) implemented a tour-based model of empty commercial vehicles in Guatemala City that was linked to previous trips in the tour. Their contri- bution is significant in that it is one of only a few that addresses empty vehicle movements, although such movements account for between 20% and 30% of urban truck trips (Holguín-Veras and Thorson 2003; Raothanachonkun et al. 2007). Several others have also proposed tour-based models (Oppenheim 1993; Boerkamps et al. 2000); however, there is no evidence they were implemented. Tour-Based Microsimulation Models The desire to incorporate the important unique characteristics of urban commercial travel, such as trip chaining (tours), increasing use of distribution centers, and optimization of rout- ing, has spurred the development of models that share many of the characteristics noted earlier for activity-based and land use models. Hunt and Stefan (2007) describe the development of tour- based microsimulation for Calgary. In 2000, they used a com- modity flow survey of 3,454 business establishments in Calgary to build a tour-based commercial vehicle model. It focused on all trips made using commercial vehicles, of which freight movements constituted only one-third. The remaining trips were made for service delivery, business travel, etc. The resulting model is an adaptation of the person-tour-based mod- eling approaches described earlier, with some elements unique to modeling freight. They attempted to model all commercial travel in the region, which encompasses nonfreight movements as well. The structure of the model is shown in Figure 13. Three classes of trucks are modeled based on relationships derived from their earlier commodity flow survey (Hunt et al. 2006). The model uses a nested set of logit models at the level of individual tours, which are generated as a function of land use rather than economic activity. Their contribution is unique in that the tours are not defined or optimized beforehand; rather, a decision is made at each stop whether to continue on to another destination or return to the origin. The probability of making another stop is calculated in part by the angle formed by the truck’s current location, its origin, and the location of the next stop chosen from a list of all available stops. Stops significantly out-of-direction are rejected in favor of those that move the tour back toward the origin. The resulting tours are sub-optimal from a routing standpoint. The model has been calibrated to the targets defined in the data, and early validation work appears to show that it matches observed commercial vehicle counts well. The model is being used in the regional modeling system in the city of Calgary. Donnelly (2007) describes the development of a tour-based freight model used as part of Oregon’s statewide model. It attempts to overcome the lack of holistic data on freight flows and characteristics by fusing a wide array of disparate and heterogeneous data using a microsimulation approach where different actions are modeled using data most appropriate for that decision. The overall model structure is shown in Figure 14. The model transforms production–consumption flows modeled by the first generation PECAS model embed- ded in the Oregon statewide model measured in annual dol- lar terms into daily flows by tonnage, commodity, and mode of transport. The resulting flows are expressed in weekly origin–destination matrices. For any given origin–destination flow the model calculates the probability of the goods flowing through a distribution center or transportation terminal. If so, the origin–destination flow is split into two, with opportunities for a different mode (generally smaller trucks) for the local portion of the overall movement. The rates were obtained from Canadian surveys, as comparable data are not available in the United States. Once this trans-shipment is accounted for, sampling from observed distributions of shipment sizes, carrier type, and vehi- cle type are used to transform the weekly flows into discrete daily shipments. These are assigned to specific vehicles, whose itineraries are optimized (if two or more stops are required) using a traveling salesman algorithm. The resulting flows are assigned to a multimodal network along with auto flows using a multiclass traffic assignment. Both models would appear to be portable. The Calgary model is being tested as part of the statewide modeling sys- tem in Ohio. Results from these and other research efforts are FIGURE 13 Structure of the Calgary commercial vehicle model.

expected to dramatically change the way commercial vehi- cles are modeled in urban areas. Statewide Models Many of the models and approaches already described have been at the metropolitan level. In most cases these models are not limited by scale, and could be used at the regional or statewide level as well. However, doing so might be imprac- tical owing to the data requirements and lack of need for data at such a high level of detail. Thus, statewide models often sacrifice spatial detail to gain wider coverage. With improve- ments in GIS and computer technology such limitations are in some places disappearing, prompting increasingly more ambitious statewide models. At the present time, approximately two-thirds of all U.S. states are known to have a statewide model of some type (Horowitz and Farmer n.d.). Most use a sequential modeling paradigm, either based on travel surveys completed across the state or models borrowed from elsewhere. Michigan, Ohio, and Oregon have all made substantial investments in the collection of short- and long-distance travel surveys, with enough observations by different divisions of geography to model the unique characteristics of each area. Ohio and Oregon are unique in that their statewide models employ all of the advanced modeling approaches listed previ- ously except for dynamic network modeling. The two models are conceptually similar, incorporating several components that are each fairly sophisticated models in their own right. 1. A macroeconomic model provides statewide forecasts of growth by economic sector and aggregate demographic changes. 2. A synthetic population generator allocates households to TAZs, and updates the population in response to changes in the macroeconomic forecast. 28 3. A production allocation process is used to allocate employment to zones, balancing floorspace consump- tion with demand by sector. 4. An activity-based person-travel model is used to model both short- and long-distance travel. 5. A commercial vehicle model also estimates short- and long-distance travel. 6. A traffic assignment model allocates the person and commercial trips to least cost paths on a multimodal network. The second and third components correspond to the land use models described earlier, whereas the final three focus on the transportation side of the system. Collectively they comprise an integrated land use–transportation modeling system. In both cases, the production allocation model decides where to obtain its workers. Thus, the linkage between workplace and residence is defined, obviating the need for a destination choice model for NHW trips in the transport models. Indeed, tours involving work locations are anchored to these loca- tions, such that intermediate stops are influenced by the work- place choice made before the activity-based component of the model even begins. The land use and transport components are integrated in other ways as well. Transportation costs and disutilities are directly used by the choice models in the land use components. This permits accessibility to various modes of transporta- tion to be considered in the location choice decisions. The relationships between producing and consuming industries, expressed as input–output make and use coefficients, are also used by the commercial vehicle models to define the linkage between industries and the commodities they pro- duce and consume. These linkages across different models, as well as the behavioral assumptions included within each component, complicate considerably the task of calibrating and validat- FIGURE 14 Structure of the Oregon commercial transport module.

29 ing such models. Many of the techniques commonly used in travel modeling, such as automatic calibration of destination choice models, certainly cannot be used in this situation. Anomalies in the modeled trip length frequency distributions for the home-to-work portion of the tours, for example, must be resolved in the production allocation model, which in turn affects the calibration of other parts of the model. Therefore, effective strategies and appropriate targets for calibrating such models are only now being learned. Both models operate at two levels of geography. A coarser zone system (500 to 800 zones) is used for the produc- tion allocation process, whereas a finer level of resolution (3,500 to 4,500 zones) is used for the transportation mod- eling components. The Ohio model also includes a focus- ing utility that will allow for a more detailed analysis of specific corridors or subareas. The Ohio model is nearing completion, although the Oregon model has been more widely tested to date. In Maryland, a somewhat different approach has been used in the recent development of a statewide model. It was designed as a multi-level model, allowing different types of travel choices to be represented at the most appropriate level. The model levels are shown in Figure 15, and include the fol- lowing modules at each level: • The regional level covers North America with 132 zones, with more detail near Maryland and less farther away. Economic forecasts are generated at this level. Sample enumeration from the long-distance element of the 2000–2001 National Household Travel Survey was used to model long-distance person-travel by residents (those who reside within statewide-level zones) and visitor trips. FHWA Freight Analysis Framework 2 was used to define internal–external and through freight trips. • The statewide level covers 1,607 zones in Maryland and parts of the surrounding states and the District of Columbia. A more detailed network is used at this level as well. Short distance person and freight trips are mod- eled at this level using traditional sequential modeling approaches. An important difference is that internal– external and through trips are not modeled at all—they are defined by the models at the regional level. Traffic assignment is also carried out at this level. • The urban level provides data from the metropolitan models. No statewide modeling is done at this level, but information from the MPO models is retrieved from this level, and comparisons of the statewide model out- comes are made to comparable MPO outputs. In many respects the travel models used at the statewide level are simple extensions of traditional sequential models and therefore not noteworthy in a review of advanced mod- els. The long-distance person and visitors models, how- ever, are microsimulation models that directly mine National Household Travel Survey data, which are unique. The first gen- eration of this model was recently completed, with a primary emphasis on being able to accurately portray multimodal travel in the Baltimore–Washington corridor. INTEGRATION ISSUES Whenever a model consists of two or more modules, integra- tion of different modules demands attention. Integration in this context simply means the blending or creation of links between otherwise separate models or modeling platforms. The goal is generally to create a more holistic model that per- forms better than the sum of the parts. However, integration can occur in several different ways within the context of modeling. A common example is passing travel time data from a travel model for calculating accessibilities and dis- utilities in a land use model. A particular challenge has to be addressed when models work at different levels of resolution. For instance, if a local travel model is designed as a DTA while a regional model is a traditional aggregate transport model, regional aggregate flows that enter the local study need to be translated into single vehicles to be added to the DTA, while local flows leaving the study area need to be aggregated to the vehicle representation at the regional level. The aggregation of local flows into regional flows commonly is a simple addition. The disaggregation of regional flows into local flows, however, requires some procedure that rea- sonably splits aggregate vehicles into logical vehicle classes and the time-of-day of the local DTA model. Where no data exist, acceptable assumptions must be made. This integration may have two dimensions (see Figure 16). First, models of otherwise comparable phenomena may work at different geographical levels, such as integrating a travel FIGURE 15 Multi-level modeling architecture of the Maryland statewide model.

model for a metropolitan area with a long-distance travel model for internal–external trips (geographical integration). At a minimum, the output of the two models needs to be com- bined, and often output from a model at one geographic layer directly influences the model behavior at another geographic layer. Second, a model could consist of several modules with different modeling tasks for the same geography, such as a transportation model and a land use model covering the same study area (component integration). The two modules are likely to improve by exchanging information. Each level is discussed separately in the following sections. Geographic Integration Models at different geographies allow for simulating the same task (such as a person trip) with different approaches catered to each level. As a benefit, the models may be designed differently, and the spatial resolution of different modules may differ to fit each model’s purpose. Although a gravity model may work well to distribute person-trips at the local level, this model becomes difficult to calibrate in a way that it works both for short- and long-distance trips. Thus, the same task of a person-trip may be simulated with different methods at the local and the regional levels. The spatial res- olution may be finer at the local level and much coarser at the regional level. For a trip that stays within the study area, the detailed locations of origin and destination are of interest. For a trip that leaves the study area for a destination 100 miles away, the precise location of the destination most likely is irrelevant. A geographic distinction in different model layers may be less relevant for urban models, but offers value to models that cover regional study areas. A common example is a statewide model that feeds into a metropolitan model and vice versa. If trips are simulated at two geographic layers, special attention has to be given to minimize inconsistencies at the border between the two layers. Figure 17 shows an example of the borderline between two layers. The local travel model on the left side has small zones, such as TAZs, whereas the regional model on the right side could have counties as spatial representation. The trip length distribu- tion shown with the gamma function and the resulting cir- cle on the zone system would capture a larger number of zone centroids of the local geographical level, but would miss most zone centroids at the regional geographical level. If such a model system is implemented, it is important that this border effect be addressed. One way to handle this issue 30 FIGURE 16 Example of two dimensions of model integration. FIGURE 17 Reconciliation for geographic resolution. is to simulate internal–internal trips by the local travel model, and to handle internal–external trips by the regional model. Component Integration It is common to build several different models that work at the same geographical level. The list could include a person- travel demand model, a truck model, a land use model, and an emissions model, as well as others. Every model is likely to benefit from (if not require) an integration with some or all other models. For example, the person-travel model may require the location of population and employment from the land use model and may provide travel times to the land use model and traffic volumes to the emissions model. The land use model may require travel times from the person and truck travel models and noise emissions (as a location factor) from the emissions model. The land use model may provide the location of population and employment to all other models (compare also land use/transportation feedback cycle in the earlier section on land use models). Such integration may become fairly complex and requires a close communication between the developers of the different models. Both geographic and component integration have been implemented in many transportation models. The vast major- ity of transportation models have some same-level integra- tion implemented, most commonly consisting of a personal transport, a freight transport, and/or a land use model. Geo- graphic integration is not as widespread, although multi- level models that require geographic integration are becom-

31 ing more common in the transportation modeling world. Examples for geographic and component integration in practice are the statewide models of Ohio and Oregon. In terms of geographic integration, both integrate long- distance personal travel models with a statewide person- travel model and long-distance goods flows with statewide freight models. In terms of component integration, both Ohio and Oregon integrate personal travel, freight travel, and land use models. Levels of Model Integration There are different levels of how closely models are integrated technically (see Figure 18 for three examples). The most com- mon is Integration Level 1. Every model runs independently. After a model has started, it reads the output data of several other models, does its own simulation, writes new output data, and is closed. After one model has finished another model can start. Building one single piece of modular software (Integra- tion Level 2) that contains all modules may be advantageous. Having all modules in one piece of software saves run time because a large amount of data can be kept in working mem- ory. Keeping data in memory saves the time one module needs to write data plus the time another module needs to read these data. The integration into one piece of software that is likely to improve the run time requires, on the other hand, a very close interaction between the developers of all modules. The third level of integration runs all modules at the same time. This close integration is only feasible in a microsimulation. Events of each module are run in random order, such as a per- son makes a trip to work, another household moves, a truck delivers groceries, a child is born, a person goes to the cin- ema, etc. This very close integration resembles how events happen in reality. So far, however, this level of integration rarely has been achieved in applied models. A somewhat different approach to multi-level integration is evident in several of the dynamic network models. A few commercial packages, allow for different levels of analyses on different links or in different subareas within the same simulation. Freeways might be modeled using a micro- simulation approach, whereas arterials are modeled using a mesoscopic or macroscopic formulation. Alternatively, micro- simulation might be used within a subarea of interest, while distant parts of the study area unlikely to be affected by the scenarios tested are modeled using a macroscopic approach. Such approaches can dramatically reduce the amount of data and computer run times required to execute the model, while retaining the flexibility to expand the resolution and fidelity of the model without losing the investment already made in data, interfaces with demand models, and proficiency with the software. RISK AND UNCERTAINTY In addition to the models themselves, ample room exists in the industry for advanced applications of both traditional and advanced models. Historically, the practice of travel forecast- ing has focused on obtaining a single “right” answer that both modelers and decision makers are comfortable with. How- ever, we are operating in a world with many uncertainties, as evidenced by recent radical swings in both fuel prices and the economy. Uncertainties in these dimensions are compounded by what may happen with major policy decisions such as greenhouse gas policy. With such challenges it may be unre- alistic to expect any model estimated from and calibrated to past behavior to correctly predict the future 30 years out. In a world of such uncertainty, it is still incumbent on the modeler to provide useful information to decisions makers. To the extent advanced models can overcome some of these uncertainties they could certainly be implemented. However, in some cases, the model itself may matter less than how it is used. Specifically, an explicit acknowledgment of uncertainty and a strong focus on scenario testing may be beneficial. Ulti- mately, the strongest project is not necessarily the one that performs best under the base conditions, but may be one that is robust across a range of scenarios. Some of the recent focus in FTA New Starts projects on scenario testing provides a good example of the form that such analysis may take. Projects are analyzed not just for base conditions, but also a series of “what if” analyses. What if the central business grows half as much as projected? What if the rail travel times are slower than anticipated? What if congestion is lower than expected? These are the sort of tests that can reveal much about the robustness of a project; they are well within our existing capabilities to perform with either traditional or advanced models, requiring only the additional time and effort. FIGURE 18 Three degrees of integration.