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.
70 chapter seven Case examples Five case examples were selected to demonstrate typical implementations of statewide and megaÂ regional models. However, standards for model application do not exist, as every implementation is unique and geared toward the needs of an individual state. Therefore, the case examples attempt to proÂ vide an overview of the breadth of statewide and megaregional modeling. These case examples include: â¢ The Chesapeake Bay Megaregion Model, which is the only operational megaregional travel demand model in the United States. Even though it is not actively used now, it provides an intriguing example of largeÂscale modeling. â¢ The Arkansas Statewide Model employs a more traditional tripÂbased model that is an analog of the urban travel forecasting process. Being a smaller state, resources are more limited than in some larger states. Arkansas also provides an example of state of practice for integrating information from the FAF and external economic forecasts. â¢ The California Statewide Model is of interest for several reasons. The issues encountered while trying to integrate several quite different MPO models are worth describing. California also has notably ambitious carbon reduction targets that have affected model design. They have recently implemented findings from a household travel survey of approximately 44,000 households, making interesting discoveries. â¢ The Florida Statewide Model must deal with a geographically and socially diversity state and its forecasting needs. Florida has not only implemented a statewide model, but also urban modeling standards in conjunction with it, and interfaces between the two. Florida has also invested heavily in freight modeling, operating one of the most complex statewide freight models in the country. â¢ The Oregon Statewide Model is arguably the most complex and ambitious statewide model in use, making it well suited to represent the state of the art in statewide modeling. Oregon operates an integrated economic/land use/transport/environment impact model designed to address issues most states do not address. The model has been under development a decade, making it one of the more mature statewide models as well. This case study also provides an interesting illustration of the rationale and experience of investing in such a heavyweight model. These case examples do not attempt to present the modeling system and supporting data in detail, but rather selected interesting archetypal aspects of every model are described. Readers interested in more details of these models are referred to the corresponding model documentations and user guides. This chapter aims to draw attention to the unique and exemplary details of each model, and document the breadth of approaches found in statewide and megaregional modeling. 7.1 Chesapeake Bay megaregion model With the beginning of the 21st century, a revived interest in megaregions surfaced. Several research projects on the extent and relevance of megaregions were conducted, sponsored primarily by FHWA. Most notable is the book published by Ross (2009) on the delineation of megaregions across the United States. Driven by her research, additional studies were conducted to explore the feasibility of modeling megaregional transportation systems. In 2010, FHWA commissioned a project to develop the first megaregional model in the United States to a team of universities and consultants led by the University of Maryland (Ducca et al. 2013; Moeckel et al. 2015). One goal of this project was the development of a framework for a megaregion model that was tested in an implementation for the Chesapeake Bay megaregion around Washington, D.C.
71 The Chesapeake Bay Megaregion Model (CBM) for the megaregion around Washington, D.C. is the only operational megaregional transportation model in the United States. Arguably, there are other megaregional or quasiÂmegaregional models implemented elsewhere, such as Gunn (1994) for the Netherlands and Wegener (2008) for the European Union. However, no other complete travel demand model at the megaregional scale has been implemented in the United States. Zhang et al. (2013) built a quasiÂmegaregional model for the southern U.S. states, which was built to model traffic flows from evacuating the region if a hurricane is approaching. Thereupon, Zhangâs model focuses on the assignÂ ment step, leaving out the travel demand side. The CBM remains the only operational megaregional model in the United States to date that includes all steps from travel demand to the assignment. 7.1.1 geography The CBM spans the area from the southern border of Pennsylvania through Maryland and eastern Virginia to Norfolk and Virginia Beach. It includes all or part of five states: Pennsylvania, Maryland, Virginia, Delaware, and West Virginia, as well as the District of Columbia. The CBM is defined by its primary environmental resource, the Chesapeake Bay. Figure 48 provides a map of the CBM along with major surface transportation infrastructure. Given the geographic extent of a megaregion, a single scale model is insufficient to capture relevant activities, travel behavior, and their impacts. This is true for many statewide models as well, particularly if trips with external origins and destinations are modeled explicitly. Therefore, a twoÂlayer approach was chosen to distinguish a megaregional layer represented in more detail and a national layer capturÂ ing relevant activities and flows outside of the main study area. Given the interactions between different FIGURE 48 The Chesapeake Bay Megaregion.
72 megaregions nationally, and to some respect even globally, the twoÂlayer approach facilitates to represent the study area with sufficient detail, yet acknowledges that megaregions cannot be treated as monolithic islands. 7.1.2 model overview The megaregion framework implemented for the CBM is shown in Figure 49. A series of economic models predicts the growth and decline of different economic sectors based on assumptions of changes in the global economy. The economic models cover both the megaregional layer and the national layer to account for a global economy that affects growth and decline in the megaregion. The land use model simulates changes in population and employment subject to this forecast. The locations of population and employment are used in the travel demand model to simulate both person travel and freight flows. Accessibilities are fed back into the land use model to influence land use changes in the following simulation period, allowing land use to evolve over time. Finally, both traffic flows and the output of the land use model are used in a series of indicator models to analyze the model results. 7.1.3 economic module To address the larger economic interest of a potential megaregion authority, the analysis framework emphasizes links with the economy, both nationally and locally. The megaregion model chain is driven by a national economic forecast component that predicts economic activity for large regions covering North America. The national Computable General Equilibrium (CGE) economic forecastÂ ing model built by the INFORUM group at the University of Maryland is applied (McCarthy 1991). The model employs interÂindustryâmacroeconomic general equilibrium models to examine past employment trends and to forecast future employment across 65 sectors of the economy. The priÂ mary model, LIFT (LongÂterm InterÂindustry Forecasting Tool), uses econometric equations to preÂ dict final demand and output at the national level, based on interÂindustry inputÂoutput relationships and valueÂadded behavioral equations. The second component STEMS (State Employment ModelÂ ing System) allocates the national forecast to states, considering the industryâs mix of basic and nonÂbasic employment and personal income. Basic employment is driven by LIFT national industry FIGURE 49 Concept of the CBM megaregion model.
73 trends, while nonÂbasic employment is driven by stateÂspecific personal income forecasts. The land use model further allocates these state control totals to model zones. Figure 50 shows countyÂlevel economic flows modeled for the CBM area. To account for local transport impacts on the regional economy beyond the national economic drivers that have little knowledge of local conditions, an economic postÂprocessor was developed. The postÂprocessor combines knowledge of countyÂtoÂcounty goods movement by industry and its sensitivity to travel impedance (generalized cost including time, distance, and travel cost per mile) to determine the impact of the local transport on each industryâs supply chain relationships (upÂ and downstream). 7.1.4 land Use module The land use module works at two different geographies: a national geography with 126 zones coverÂ ing North America and a megaregional geography with 2,075 zones. Economic flows generated by the economic model (âEconomic Modelsâ) serve as control totals for the regional level: â¢ National Model: Historic employment data from the U.S. Bureau of Labor Statistics (BLS) Quarterly Census of Employment and Wages are disaggregated to counties. Employment data at the county level are used to calculate the share of employment in each county in the baseline FIGURE 50 County-level economic flows in the Chesapeake Megaregion.
74 years 2000, 2005, and 2009. The ratio of employment for each county is then extrapolated for 5Âyear periods beginning in 2010 and ending in 2030 using an exponential smoothing method. Employment data from 101 BLS subÂsectors for the same period are then reconciled with 65 STEMS industries to match the BLS subÂsectors. This process simply employs a straight average of the two employment ratios. The purpose of this method is to allocate future STEMS projections based on the historic difference in BLS data and STEMS projections. â¢ Megaregional Model: The local model generates socioÂeconomic data at the resolution of megaÂ regional modeling zones. The initial allocations are made based on transportation costs and the employment distribution of basic employment, or exportÂoriented employment. The allocation of households and employment is based on the Lowry model (Lowry 1964). The model is repeated in 5Âyear increments until the horizon year 2030 is reached. The location of basic employment provides inertia in the allocation of nonÂbasic employment. Obtaining data for a land use model at the scale of a megaregion is a challenging task, particularly if several states are involved. In contrast, a Lowry model requires less data but may provide reasonable land use sensiÂ tivities within computationally acceptable runtimes. Because of its simplicity and broad simplificaÂ tions, the Lowry model has become rare at the urban level. At the megaregional level, however, such a model may provide the right balance between model sensitivities and comprehensiveness. 7.1.5 Travel demand model Similar to the land use model, the travel demand model works at two geographic layers. Although the local travel demand model is closer to a traditional statewide model, the national models are built from nationwide surveys for both personal and truck longÂdistance travel. Travel demand of the two layers is combined in the traffic assignment process that assigns shortÂ and longÂdistance traffic flows concurrently in a multiÂclass assignment. The megaregional travel demand model is built as a fiveÂstep aggregate model that includes trip generation, destination choice, mode split, timeÂofÂday split, and assignment. The original model conÂ cept was borrowed from the Baltimore Metropolitan Council model (BMC 2007), refined and applied for shortÂdistance trips of 50 miles or less for the entire CBM study area. Trip generation rates were derived from a 2007 Household Travel Survey for the Baltimore and Washington, D.C., metropolitan areas, as well as travel times, mode split, and timeÂofÂday split. Thereby, the model is calibrated to the core area of the CBM study area. A survey for the entire study area was not available. This model includes a threeÂstep local truck model based on the QRFM (Beagan et al. 2007), which was applied for local trips under 50 miles within the CBM study area. The comparison with VMT estimates revealed that the QRFM method, which is based on a 1992 truck trip survey from Phoenix, generated too many truck trips. Therefore, parameters were scaled down to resemble local VMT truck estimates and truck count data. At the national layer, the National Estimate of LongÂDistance Travel (NELDT) has been impleÂ mented to simulate longÂdistance person trips greater than 50 miles (Moeckel & Donnelly 2011). The longÂdistance element of the 2002 NHTS, the most recent survey that covered explicitly longÂdistance trips in the United States, provided data input. To expand the survey to cover all longÂdistance trips in the United States, air travel data from the Bureau of Transportation Statistics (BTS), which covers 10% of all ticketed air travel passengers, was used as a control total to scale NHTS records to observed flight volumes. Auto trips were generated using the same scaling factor and added to the multiÂclass assignment with local traffic. For freight longÂdistance travel, commodity flows from the Freight Analysis Framework 3 (FAF3) were used for truck trips of 50 miles or more. Though the 123 FAF zones reflect centers of economic activity, the resolution is too coarse to model truck trips on a network with much more detail. Therefore, flows between FAF regions were first disaggregated to flows between counties and then further disÂ aggregated to flows between model zones. To disaggregate commodity flows, make/use coefficients and employment by type were used to allocate flows to the most likely producers and consumers of every commodity.
75 Commodity flows between zones are converted into trucks trips using average payload factors proÂ vided by FAF. These factors describe how many tons of a given commodity are carried by a truck on average. To account for empty truck trips, an emptyÂtruck trips rate was added globally to all truck trips. 7.1.6 environmental and Fiscal impacts A few indicator models were added to estimate the environmental and fiscal impacts. The results of the indicator models are not fed back to the other model components, but were used to evaluate scenarios: â¢ Gaseous Emissions: This model captures traffic emissions using the EPA Motor Vehicle Emission Simulator (MOVES) model (EPA 2016). The MOVES model uses VMT and linkÂlevel volumes and speed data output of the travel model to estimate GHG and other mobile emissions. â¢ Water Quality: This model captures the impact of alternative policies on water quality. A nutrientÂ loading model estimated flows draining into the Chesapeake Bay. It forecasts the annual loads of nitrogen, phosphorus, and sediments. The model uses detailed land cover changes from a rasterÂ based land use model to identify changes in nutrient runoff resulting from land cover changes. â¢ Infrastructure Costs: This model estimates state and local governmentsâ costs to provide public infrastructure in support of new development (e.g., roads, sewer, water). Established relationÂ ships between the current development and the provision of infrastructure were applied to project future improvements needed to satisfy additional settlement activity. Data included resiÂ dential development classified by housing type, existing water and road infrastructure, property value trends and tax rates, among others. 7.1.7 lessons learned from the CBm model The design of the CBM model is similar to a statewide model. A comparable twoÂlayer geography is represented in many statewide models to distinguish longÂdistance and shortÂdistance travel. Special attention was given to additional model components, such as economic, land use as well as environÂ ment and fiscal impact models, though many other statewide models use such âaddÂonsâ as well. A significant challenge for the CBM was a study area that included parts of five different states and the District of Columbia. Although many statewide models cover parts of neighboring states at a higher resolution, it is rare that data from as many states are included. Creating a network from different sources, socioÂ economic data from different agencies, and count data in different formats created a challenge for the development of the CBM. As such, the model was operational but only implemented as a prototype. Detailed model validation has never been conducted outside of Maryland. As with many statewide models, the CBM coped with two levels of model integration (Figure 51). On the one hand, vertical integration covers the interaction between different geographies, here the FIGURE 51 Horizontal and vertical integration in a megaregional model.
76 North America level with the megaregional level. On the other hand, horizontal integration captures the interaction between different modeling domains, such as transport models with land use models. A core finding was that the level of integration depends on (1) the frequency of information exchange, (2) the direction (oneÂway or biÂdirectional) and (3) software implementation constraints. Tighter inteÂ gration turned out to not always be better. In some cases, loose coupling of models may be sufficient and much easier to implement. The work resulted in an extensive report (Ducca et al. 2013) and a paper published (Moeckel et al. 2015), but was never used as a true policy advice tool. Being funded by FHWAâs Exploratory Advanced Research Program, it was not expected to be applied in practice. However, the work has not inspired the development of other megaregional models yet. Most likely, this is the result of the lack of megaregional institutions across the United States that might use largeÂscale models to advise decision making. Therefore, the CBM model remains a proof of concept that may assist other megaregions in the future. 7.2 arkansas sTaTewide model The Arkansas State Highway and Transportation Department (AHTD) completed a Phase II update of its statewide model in 2015. All parts of the model and underlying data were addressed during the update. The resulting modeling system represents an excellent example of best practices in statewide modeling, incorporating current thinking about how such models might evolve in the future. These qualities are shared by other models reviewed in this chapter. What makes Arkansas unique, however, is that they could leverage prior work and experience to do so within a modest budget, and their creative approach to overcoming a lack of local travel behavior data. A multiÂscale modeling approach was employed to develop a model that provided a regional and national context for Arkansas, while still providing a high enough level of detail for accurate forecastÂ ing of policy and project scenarios within the state. The state is covered by 5,849 traffic analysis zones, designed to include urban model zones as already defined, and to be compatible with and nest within Census block and block group boundaries otherwise. Care was also taken to respect natural boundaries that impede flow, as well as not crossing major roadways and highways. The rest of North America is covered by 144 zones formed by aggregations of Economic Areas defined by the Bureau of Economic Analysis (U.S. Department of Commerce) with the United States, and states and provinces within Mexico and Canada. As with many other statewide models, urban model networks within the state are fused with the FHWA National Highway Planning Network to represent the roadway system. A rail network maintained by Oak Ridge National Laboratory was used to represent the freight rail system. The agency model is described next, followed by a discussion of its innovative approach to data development and lessons they have learned. Their model documentation, spread across eight reports, is available upon request from AHTD. 7.2.1 long-distance Freight models LongÂdistance truck and rail flow forecasts are developed using a fourÂstep sequential modeling proÂ cess, based upon commodity flow forecasts. The model is based upon a 2008 Global Insight Transearch flow database. Unlike many states, that map the flows to truckload equivalents, the data are used to build a commodity flow model. This provides a degree of policy sensitivity that static forecasts, such as the Transearch and FAF static forecasts, do not. In this case, changes in employment can be used as a proxy for changes in gross state product, growth in certain industries relative to others, and the like. Linear regression models are used to predict annual tons shipped by all modes of transportation as a function of employment within industries that produce such goods, for 15 commodity groups. The production model was estimated at the county level within Arkansas, and state level outside of it, for 13 of those groups (reported flows in the Transearch data were used for mining and coal flows). The resulting r2 statistics ranged between 0.46 and 0.84, providing reasonable fits. Attraction models were likewise estimated using linear regression, using inputÂoutput tables to understand better the linkage
77 between commodities and consumption. Special generators were developed for counties whose actual tonnage (from the Transearch data) deviated significantly from the regression estimates. A doubly constrained gravity model is used for trip distribution, using friction factors based on the distances coded in the Transearch data. The latter are often used instead of mathematical functions to replicate the multiple peaks in observed trip distance distributions. An incremental mode choice model, based upon existing mode shares by commodity and distance ranges, was developed using the Transearch data. The rail tonnage was assigned directly, with truck tonnage converted into daily truckload equivalents using payload and annualization factors. The latter were combined with auto flows in a multiÂclass traffic assignment. 7.2.2 person and other Truck models Person local and external travel, as well as nonÂfreight commercial vehicles, are modeled using the widely accepted tripÂbased modeling paradigm, with a few interesting extensions. A schematic diaÂ gram of the modeling system is shown in Figure 52. Nine person and four truck trip purposes are modeled, to include two of each devoted to external trips (i.e., those crossing the state border). Two of the purposes represent infrequent longÂdistance travel, which includes trips longer than 100 miles. Like an increasing number of states, person trip generation rates are differentiated by area types (four in this case, as a function of population and employment density) in addition to household size and income groups. Attraction rates were calculated using borrowed data, stratified by the same four area types and income groups, as well as total households and employment by broad categories. Special FIGURE 52 Schematic diagram of Arkansas Statewide Model.
78 generator attractions were calculated for unique land use types, such as universities, hospitals, military bases, and airports. Doubly constrained gravity models are used for trip distribution, using distance as the measure of impedance for infrequent longÂdistance trips and travel time for all other trip purposes. Friction facÂ tors were fitted to smoothed trip length frequency distributions for each trip purpose, and iteratively calibrated to survey data. A nested logit mode choice model that handles both local and intercity travel was developed, based heavily upon FTA guidance and typical practices. Several of the model coefficients were asserted, based upon FTA advice, and then calibrated to local conditions. CalibraÂ tion targets were generated from 2009 NHTS data for auto trips, while data from two Arkansas transit agencies were used for transit trips. The daily person trips by mode were allocated to four periods of the day, using diurnal factors by trip purpose and direction. Finally, the person flows were converted into vehicle trips through the application of auto occupancy rates, where appropriate. External trips are those with an origin, and possibly destination as well, outside of Arkansas. This corresponds to the familiar externalâinternal (EI) and externalâexternal (EE, or through) trip purÂ poses. Because both local and longÂdistance trips by Arkansas residents are modeled there is no need to include internalâexternal (IE) trips, for they are already accounted for within the modeling steps described previously. Traffic and vehicle classification counts at the state border are used to constrain the flows to observed or forecasted levels. The external trips go through the same diurnal factoring as internal trips before being combined with them for traffic assignment. The person travel models include truck flows that are not represented in the Transearch data described previously. These include trucks making local trips, as well as nonÂfreight trucks traveling longer disÂ tances. They are modeled in a similar manner as persons, using data borrowed from Idaho. Regression models by truck type were developed for trip generation, which are assumed to also represent trip attractions. Trip distribution models such as those developed for person trips were calibrated for three categories of trucks. Diurnal factors were borrowed from Idaho and compared with those found in the literature and other statewide models. When combined with the longÂdistance truck trips described in the previous section these flows represent the entire population of trucks operating within Arkansas. 7.2.3 Traffic assignment The Arkansas model includes both highway and freight rail assignment procedures. The former is carried out for four time periods and includes feedback loops to obtain more accurate travel times, as shown in Figure 52. Several user classes are jointly assigned. Trip matrices are segmented by 18 combinations of trip purpose, auto occupancy, and income groups to define the user classes. This is more detail than thought to be necessary in most statewide models. A multiÂclass static user equilibrium assignment process is run separately for four periods of the day, with each converging through feedback separately. The widely used Bureau of Public Roads (BPR) volumeÂdelay function is used to update link travel times during assignment. The model coefficients were updated in the Arkansas model, drawÂ ing upon experience in other states and the literature. Ten variants are used to capture differences in perceptions and values within each user class. The function coefficients are also segmented not only by functional class, but by whether the downstream node is a signal or stop sign controlled, or not an intersection. Considerable effort apparently went into derivation and testing of these coefficients, as well as sensitivity testing of toll costs. A simpler assignment process was also developed, to be used to reduce the computational burden and data handling associated with a large number of user classes. Three user classes (passenger, comÂ mercial trucks, and heavy trucks) are used, although doing so precludes analyses of toll and highÂ occupancy lanes. The modeling system also includes the capability to assign freight rail tonnage or annual train equivalents on a simplified rail network. The originÂdestination flows are assigned to the shortest distance path, rather than attempting to capture the complexity of rail routing decisions that include factors such as interlining cost, track ownership, and cost minimization. The goal is to visualÂ ize the spatial dimensions of rail traffic, not its exact routing, for which their method is quite adequate.
79 7.2.4 innovative data development The developers of the Arkansas model had to overcome several data shortcomings, with the lack of household and truck travel survey data being the most significant. It is common in such cases to borÂ row or assert models, or synthesize them using data from different sources, and to calibrate the model to what local data are available. Like other states, the developers turned to the NHTS for this task. However, there are too few observations within Arkansas to build a statistically stable and reliable model from them. However, both Tennessee and Texas had purchased additional samples during the 2009 NHTS cycle, which proved invaluable for the development of the Arkansas model. Summaries from both states were compared during each step of the person travel model. The Texas data were used in each case, for their much larger sample size (22,255 households) was an order of magnitude larger than collected in Tennessee. However, the differences between them, as well as summaries of the small samples from other nearby states and a 2003 household survey from Central Arkansas, defined the range of such values. This helped identify which patterns and relationships were the most stable versus those where Arkansas would likely vary from its neighbors. Similar data were borrowed for other parts of the model, to include: â¢ Workplace surveys from Arkansas were similarly not available. The team obtained permission to use data from four Texas workplace surveys to develop trip attraction rates. The resulting rates compared well with other models and the literature. â¢ Typical mode choice model structures, parameters, and coefficients were drawn from FTA New Starts guidance, as well as advice from senior modelers in the agency. The collection of onboard travel survey data, coding and testing of transit networks, and formal estimation of parameter estimates is timeÂconsuming and of questionable value in settings such as these. Synthesizing these data saved the developers considerable time, permitting them to focus more time on caliÂ bration and sensitivity testing, both of which led to better modeling outcomes. â¢ Data from a 2008 commercial vehicle survey conducted for the Community Planning Association of Southwest Idaho (COMPASS), the MPO for the Boise area, were used to develop the truck characteristics in the person travel model (e.g., those trucks not included in the Transearch data, to include local and nonÂfreight truck movements). These data proved invaluable for the developÂ ment of that part of the model. â¢ Experiences with traffic assignment models, to include volumeÂdelay function parameters, were heavily informed by models implemented in several other states by the developers. The effective use of these data reduced both the cost and time required to develop the Arkansas model. Judging from the reported validation statistics, these borrowed data enabled the model to perform as accurately as reported in many other states. Thus, Arkansas has a fully functional model, when they might still be in development stages had they invested in a largeÂscale household survey across the state. In a sense, the availability of these data has made statewide modeling affordable for Arkansas and therefore accessible when it might not otherwise have been. 7.2.5 lessons learned from arkansas Arkansas is a relative newcomer to statewide modeling, having embarked upon its first version in 2010, and recently updating it. It achieved this through leveraging work done elsewhere, as well as learning from their experience. Some of the more important and interesting successes include: â¢ The NHTS is a very important resource for statewide modelers, for many states lack the resources required to conduct statewide travel surveys. However, little has been done to compare how estimates for one state stack up against its neighbors, or whether data from several states can be pooled to gain acceptable sample sizes. Several developers have already done so, but few comÂ pare the results in each step of the model like Arkansas has done. This underscores the importance of funding addÂon surveys (i.e., paying for the collection of additional surveys) in future versions of the NHTS for those states unable to conduct such surveys on their own, as well suggesting the value of several adjacent states pooling funds to do so.
80 â¢ Considerable work went into the development of a graphical user interface and standardized reporting, which make the system and its outputs more accessible to both modelers and their clients. Having a model that is easier to use enables AHTD staff to spend more time evaluating results, rather than setting up and running the model in an ad hoc manner. â¢ The focus was on the usability of the model throughout its development. This not only drove the initial design, but remained in focus when evaluating validation outcomes and deciding upon changes and possible model enhancements. This is best illustrated in their adoption of FTA guidance and advice in the construction of their mode choice model. Rather than developing a model and testing its utility, the developers started with a recommended approach and worked backward to determine how best to calibrate and test it. This approach saved considerable time and effort for a part of the modeling system that is often as costly as the rest combined. The agencyâs current model appears capable of meeting its shortÂterm needs, and represents a sound investment. How the model will evolve over time to address new and more complex issues not well addressed using tripÂbased modeling systems remains an open question. There is no doubt, however, that Arkansas is on a path that will enable them to do so in a highly costÂeffective manner. 7.3 CaliFornia sTaTewide models On paper, California has the most expensive statewide model in existence, by a wide margin. Although it is difficult to pinpoint the full cost of statewide modeling, the costs shown in Figure 37 suggest that their costs far exceed those of other states. A closer look reveals that their program involves far more than just statewide travel modeling, including forwardÂthinking and ambitious investment in data collection that eclipses those of almost all other states. Moreover, there are two statewide models of California, which unto itself is an interesting story. Their modeling systems and what they can do with them, how they have partnered with others in their development, and how they might complement one another are described in this section. Extensive and current documentation about both modeling systems is available online. These include descriptions of the models and their development, as well as reports about supporting data, results of model testing, and forecasts developed using them. In the case of its HSR model the findings of an independent peer review panel are also published on the Authorityâs website. This might not appear remarkable today, until one attempts to find similar documentation for most other statewide models online. Viewed in that light, California is commendably progressive in making extensive documentation of its statewide travel data and models publicly accessible. Each modeling system is described in the following sections, with a focus upon the unique and noteworthy aspects of each. 7.3.1 The California statewide Travel demand model Caltrans has developed a comprehensive multimodal statewide model of all personal and commercial travel by residents of the state. The model system is used by Caltrans for a variety of analyses. The system was designed to meet several analytical needs: â¢ Provide insights into longÂdistance travel between and through multiple regions of the state. â¢ Model behavior for all California residents and firms, expressed as multimodal tours and trips. â¢ Provide data on external travel for urban and metropolitan travel forecasting models. â¢ Evaluate changes in GHG emissions, as required by several state laws. â¢ Provide aggregate travel statistics that can be used to resolve differences between adjacent urban and regional models. â¢ Provide a travel demand model for agencies that do not maintain a model of their own. Descriptions of the modeling system and data, as well as how they have partnered with other agencies, are presented in the next sections.
81 Partnerships Caltrans has partnered with several agencies to fund and direct its statewide data and modeling proÂ gram. The firstÂgeneration model was funded through university research, as is work on a new freight modeling framework. Its recent statewide travel survey, described here, was designed to collect data needed for both transportation and environmental models. The list of collaborators is impressive and unmatched by any other state except Oregon: â¢ California Air Resources Board, â¢ California Strategic Growth Council, â¢ California Department of Public Health, â¢ California Department of Housing and Community Development, â¢ California Energy Commission, â¢ California HighÂSpeed Rail Authority, â¢ Metropolitan planning organizations, and â¢ Regional transportation planning agencies. Not all partners are involved at the same level; however, the statewide modeling program in CaliÂ fornia has greatly benefited, and had more resources at their disposal as a result. Granted, California has passed legislation requiring such collaboration to achieve sustainable planning and GHG reducÂ tion targets. Specific multiÂagency mandates and deadlines are required in several laws, to include: â¢ California Global Warming Solutions Act (AB 32, passed in 2006); â¢ Sustainable Communities and Climate Protection Act (SB 375, passed in 2008); and â¢ California Homes and Job Act (SB 391, passed in 2013). Regardless of the motivations, the level of interÂagency cooperation and pooling of resources to solve large problems is helping to drive the agenda for statewide modeling, both directly and indirectly. Data Development Many agencies rely on the NHTS or data borrowed from other areas to build their statewide models, as revealed in Table 2. California, like a few other states, has opted instead to collect the data it believes are required for robust models. Some have no parallel in other places, and have enabled Caltrans to develop the advanced model that it currently uses. Two of the most important recent data collection activities included: â¢ The 2010â2012 California Household Travel Survey (CHTS) collected data from households in all California counties and three in Nevada (Nustats 2013). A traditional 24Âhour travel diary was used to record trips by all modes of transportation for all respondents, and in some cases, three types of GPS data collection [wearable, inÂvehicle, and connected to the onÂboard diagnostics (OBD) port]. Data were collected from the GPS units for periods of 3 or 7 days, depending upon the device used. Households were also asked to report longÂdistance travel over the past 8 weeks using a separate retrospective survey. Complete data from a total of 42,431 households were collected, including 5,717 from which GPS data were obtained. An additional dataset of partially completed surveys for 20,651 households was provided, which some of the survey sponÂ sors considered still useful. â¢ The 2014 California Vehicle Inventory and Use Survey (CalÂVIUS) is designed to fill the gap left by the cancellation of the federal survey a decade ago. Like its national predecessor, the survey is designed to collect data on the physical and operational characteristics for a sample of all commercial vehicles. The survey, being conducted by the Institute of Transportation Studies at the University of CaliforniaâIrvine, is being conducted in two parts. The first is a webÂbased survey of fleet managers, used to collect information about the vehicle and its annual operation. A driver survey is completed from a mobile phone app over a 3Âday period. These data will be invaluable for developing and validating the commercial vehicle models. The survey is currently in progress.
82 Overview of the Modeling System The second version of the California Statewide Travel Demand Model (CSTDM) was delivered in 2014 (Cambridge Systematics 2014). It is a multimodal, tourÂbased travel modeling system with five major components, as shown in Figure 53. All the models employ a microsimulation framework, in whole or in part. The ShortÂDistance Commercial Vehicle Model (SDCVM), for example, uses aggregate models to generate tours, and microsimulation to add their attributes. The system simuÂ lates travel by all California residents and firms for a typical weekday in the fall or spring. Databases have been developed to support forecasts for 2015, 2020, 2035, 2040, and 2050. In a nutshell, the five components include: â¢ The ShortÂDistance Personal Travel Model (SDPTM) consists of a series of logit models that generates tours and trips by five periods of the day. Starting from a synthetic population, longÂ term choices are modeled first, to include work and school location and driver license status. Daily tours are then generated, as well as their primary destination and mode of travel. Secondary destinations are then chosen, followed by trip mode choice. The models are applied somewhat differently for each tour purpose. â¢ The LongÂDistance Personal Travel Model (LDPTM) is based upon the approach used in the statewide HSR model, which begins with a tour formation model. It is run in parallel with the SDPTM. Tour attributes are then generated, with models of party formation (size), tour properÂ ties (duration and time of day), destination choice, and mode choice. The latter includes main, access, and egress choice models, based upon an earlier version of the HSR longÂdistance mode choice model. An interesting aspect is how both models interact, as shown in Figure 54. â¢ The SDCVM is an adaptation of a tourÂbased microsimulation of commercial vehicle (CV) flows, based on a model originally developed in Calgary (Hunt and Stefan 2007), and based on parameter values from its implementations in Alberta and establishment survey of California businesses. The model structure is illustrated in Figure 55, which simulates flows from six categories of establishments for three types of trucks. â¢ The LongÂDistance Commercial Vehicle Model (LDCVM) is based upon truck originÂ destination flows derived from a ProductionÂExchangeÂConsumption Allocation System (PECAS) model implemented in the firstÂgeneration statewide model. Growth factors are used to factor the trip matrices for future years, based upon growth in zoneÂlevel demographics. Although the original flows were developed using PECAS, the current model does not depend upon it for current or future forecasts. â¢ The Extracted Timing Models (ETMs) handle flows to and from 51 external stations in the model, including three marine ports. They are organized into six districts across the state. Existing counts at the external stations, and extrapolations into the future, are used to generate external trips. A logitÂbased destination choice model is used, and trips are allocated to time periods using observed patterns. The resulting modeling system includes 5,474 traffic analysis zones, which nest within California counties and 524 land use zones used by the PECAS model (Hunt and Abraham 2005). The latter FIGURE 53 Structure of the California statewide travel demand model.
83 was an integral part of the first generation statewide model, but is not an active part of the current version. The network includes more than 325,000 nodes, including its sparse representation of the national Interstate and U.S. highway system. One innovative feature is the use of a simplified representation of local transit service, using a technique described by Circella et al. (2014). Fixed guideway and intercity transit lines are explicitly represented in the model, but the level of service times and costs are synthesized based upon roadway speeds, land use variables, and assumptions about transit service levels. The same methodology is used in the Oregon Statewide Model to obtain reasonable approximations of local transit accessibility without the huge effort required to code every local bus line within the state. Next Steps Caltrans plans to complete model updates on a 5Âyear basis, incorporating data that has become availÂ able or been updated since the last update. Caltrans is sponsoring the development of a new multiÂ modal freight model that will eventually replace the LDCVMs and SDCVMs in the current system. Both steering committees and peer review panels will be used to provide advice about the longÂterm evolution of the statewide model and its supporting data programs. FIGURE 54 Integration of the California short- and long- distance person travel models. FIGURE 55 Structure of the Calgary commercial vehicle model.
84 7.3.2 The California high-speed rail ridership and revenue model The California HighÂSpeed Rail Authority maintains a separate ridership forecasting model used to support business and system planning, as well as corridor studies and analysis of alternative alignments and phasing of implementation. It is a complete statewide model, and uses some of the same data used to develop the Caltrans model. It incorporates a more sophisticated mode choice model better suited for evaluating HSR alternatives, but otherwise covers the same travel markets as the Caltrans model, and at comparable levels of spatial, temporal, and behavioral resolution. It also includes an explicit risk and uncertainty assessment process. This is a necessity for HSR forecasting, but unfortunately unique among the statewide models reviewed for this report. The Authority oversees the ridership forecasting work, with assistance from its Rail Delivery PartÂ ner (RDP) and an independent peer review panel. However, all model development and application work has been undertaken by consultants, under contract to the RDP. The associated cost of developÂ ing and using this model is not published, but is thought to be comparable to the investment made in the Caltrans model and data. Moreover, scores of forecasts have been produced with it. Because these were undertaken by the consultant, those costs likely account for a significant portion of the total cost incurred by the Authority. Several aspects of the California HSR model and how it is used are noteworthy. A summary of their modeling system, as well as data and institutional aspects of its development and deployment, are described in the following sections. Open and Transparent Analytics The California HSR model is unique in several aspects, both in comparison to other HSR forecasting efforts and most other statewide models. Most HSR forecasts, both domestically and abroad, have been undertaken by consultants who use proprietary models and data to develop them. The overall approach is described in their reports, as well as characterization of major market trends driving the trajectory of the model. Narrative, tabular, and graphical summaries of the forecasts are also proÂ vided, along with their interpretation. However, the forecasts are the delivered product, not the data or models used to develop them. This precludes reuse of the data and models, as well as detailed review and testing. California breaks from that tradition, having developed the data and platform used for its analyses. This gives it the ability to test new alternatives quickly, as well as generating new forecasts to reflect changing alignment, service levels, fares, etc. It should be transferable to other regions, although no known attempt has been made to do so. The model was originally developed under contract to the Metropolitan Transportation CommisÂ sion in the Bay Area. It was based, in part, upon results from stated preference surveys carried out in 2005 within the state and older statewide travel survey data. It was subsequently extended to support statewide analyses for the Authority, and updated with new data and models. The current version of the system, known as the Business Plan ModelÂV3 (BPMÂV3) incorporated new survey data from the CHTS and a 2013â14 SP survey performed for the Authority, a completely revised mode choice model, and improvements to almost all the other modules. The model is extensively documented (Cambridge Systematics 2016a), as are the data used to develop and apply it. Documentation about the model structure, key parameter values and their deriÂ vation, and tests of the effect of major assumptions are available on the Authorityâs website. Data Development An intercept survey of stated and revealed intercity mode preferences of California residents was undertaken in 2013â2014. Data about respondent attitudes toward, and experience with, HSR elseÂ where were also collected. The data, and how they compared with results obtained in an earlier survey completed in 2005, were used to develop a new mode choice model (Cambridge Systematics et al.
85 2015). The SP portion consisted of six experiments presented to each respondent, consisting of several options for mode of transport and variations on important characteristics of each. A total of 4,314 respondents provided complete and usable data. About 40% were from conventional rail riders, while the remainder was split almost evenly between air and auto travelers. The results of the CHTS, and particularly the longÂdistance element, were also used in model estiÂ mation and calibration. The daily diary portion of the survey was large enough to accurately portray the incidence of longÂdistance travel on the survey day, something smaller surveys struggle to capture. While the number of such observations was too small to build robust models, the CHTS travel diaries did prove useful for validating the incidence of longÂdistance travel. Information about the latter was also gleaned from an additional 8Âweek recall survey of longÂdistance travel. This optional part of the CHTS provided data on 32,641 longÂdistance trips by California residents, which was used for some parts of model development and validation of other parts. The aggregate results were also adjusted based on 2011 data from a Harris Panel survey to understand gaps in the data, particularly with respect to uncertainty about how CHTS respondents reported repetitive trips. Bi-level Modeling Framework The modeling system employs the biÂlevel structure shown in Figure 56. The longÂdistance model includes trips within the state longer than 50 miles, stratified by four trip purposes. They include business, commuting, recreational, and all other trips, a common scheme used in longÂdistance travel models. The trip frequency and destination choice models were estimated using the CHTS longÂ distance survey data and RP portion of the 2005 RP/SP survey. The mode choice model is a combined model of main, access, and egress mode choice, estimated from the weighted RP/SP surveys alone. Several nesting structures were tested, and the final model was calibrated using the expanded CHTS, rail boardings, and air travel flow data. The shortÂdistance model includes person trip tables by mode and trip purpose from the regional travel models used by the Southern California Association of Governments (SCAG) and the MetroÂ politan Transportation Commission (MTC). They are the MPOs for the Los Angeles Basin and San Francisco Bay Areas, respectively. Trip tables for their base and forecast years are used to represent shortÂdistance travel within the BPMÂV3 framework, and are translated into the format required by the shortÂdistance mode choice model. The mode choice model is an adaptation of the 1996 Baycast model developed for MTC, which has been calibrated to reproduce base year transit ridership within each metropolitan area. More recent models in both regions are further apart in structure, precluding their direct use within the BPMÂV3 framework. The resulting system provides consistent forecasts of shortÂdistance mode choice within both regions. The station spacing outside of those regions is too far apart to enable shortÂdistance HSR trips, obviating the need for shortÂdistance models within the rest of the state. FIGURE 56 Structure of the California HSR Business Plan Model-V3.
86 The BPMÂV3 system and its predecessors have been used to generate ridership and revenue estimates for initial and detailed system planning and in support of corridor studies and evaluation of candidate initial operating segments. This has included the environmental studies required at all levels of analyses for the program, and use for stationÂlevel impact analyses. The latter has required postÂprocessing of the model outputs, for the model was not designed to support detailed analyses at a fine level of geography. The SCAG and MTC models are wellÂsuited for that level of analyses, but only for stationÂlevel analyses within their modeled area. The BPMÂV3 and its predecessors have also been used to generate forecasts for the Authorityâs biennial business plan. The process used to develop the forecasts and their interpretation is described by Cambridge Systematics (2016b). Independent Peer Review The early forecasts developed using the first version of the model were criticized with respect to their derivation and accuracy (Brownstone et al. 2010). The Authority formed an independent peer review panel of prominent experts in travel demand forecasting to review the entire modeling system, not just the controversial parts, and advise the Authority on its fitness for use. The panel conducted a thorough review over a period of several months, and made several recommendations that guided the development of the second and third versions of the modeling system. The group was renamed the Ridership Technical Advisory Committee in 2014, as their role changed from the independent review of the original model to providing advice to the Authority on longÂterm goals for their forecasting work, review of interim products, and continued review of technical aspects of the modeling system. Their work and findings are reported in a dozen reports published on the Authorityâs website. 7.3.3 lessons learned from California Several regulatory and policy issues seemingly unique to California have influenced how their stateÂ wide models have evolved. However, climate change, economic interaction, sustainability, and new technologies are issues that other states are beginning to grapple with as well. Some of the key lessons learned from California include: â¢ The importance of, and value gained from, data collection cannot be overstated. California has invested heavily in person and freight travel data collection, and is committed to continuing these programs in the future. This has enabled them to gather information about their unique local markets, overcome longÂstanding gaps in knowledge about intercity modeling, learn how California residents might react to new modes of travel and changes in existing ones, and focus the survey upon information required for their advanced models. â¢ Caltrans was not afraid to significantly overhaul the statewide model, enabling them to refocus their efforts on aspects that proved important during use of their first model, and streamline other parts of it. Doing so was costly, but provided them with improved model performance and capabilities. â¢ Their investment in rigorous peer review enabled the HighÂSpeed Rail Authority to quickly address controversy about the model and forecasts generated using it. Such a group not only brings new ideas to the table and asks hard questions, but they also bolster the credibility of the agency. This is important in cases where new modeling approaches are being used, or when major decisions are being made based on the forecasts. â¢ There is more that is similar than unique about both California models, enabling them to share data and lessons learned. The ability to compare one with the other, especially in model calibraÂ tion and validation, is particularly valuable. Very few states have reinvented their statewide modeling programs in the way that California has. Their commitment to data collection is especially laudable, for it has enabled them to break past the barrier of knowing little about patterns unique to California, and their interactions with external markets. These progressive ideas will hopefully define a new standard in statewide modeling in North America.
87 7.4 Florida sTaTewide model The Florida Statewide Model is a particularly interesting case example for this report for two reasons. First, the consistency between MPO models and the statewide model in terms of input data and model design is remarkable. Second, the Florida DOT staff appears to be doing an excellent job in explaining the value and limitations of statewide modelling to decisionÂmakers and others. This might be part of the reason why, per the survey, Florida has retained the largest number of staff members working on statewide modeling among all state DOTs in the U.S. (even though Florida is only the fourth bigÂ gest state by population). The combination of continuous exchange with potential model users and a strong champion promoting statewide modeling appears to have established the statewide model as an integral part of transportation planning in Florida. Floridaâs statewide modeling efforts started in 1988 with software written in Fortran and TRANPLAN, making Florida not the first but one of the early adopters in statewide modeling. Like other states, Floridaâs intention was to model travel demand outside and between MPO areas, where much of Floridaâs population growth has been occurring over the past two decades. Highway corridors are of particular interest, as well as longÂdistance freight flows. In 2000, the model was transferred to Citilabsâ CUBE transportation modeling package. 7.4.1 Coordination Between statewide and mpo models In Florida, statewide modeling and MPO modeling are closely coordinated. The Florida Standard Urban Transportation Model Structure (FSUTMS) has been developed to serve both statewide and urban modeling needs. The FSUTMS provides the same model structure for statewide and urban models, including model scripts, software, zone systems, networks, socioÂdemographic data, and concepts for model application. These standards were developed by a soÂcalled Modeling Task Force (MTF), consisting of representatives from MPOs, Florida DOT districts, Floridaâs Turnpike Enterprise, transit agencies, usersâ groups, and FHWA. In addition, five technical committees were formed under the MTF: Data, Freight, GIS, Model Advancement, and Transit Committees. These comÂ mittees meet regularly in person or by means of the web and serve as the technical backbone of MTFâs modeling standards. These committees make recommendations on data purchases, model developÂ ment, and model applications. Other transportation professionals throughout the state, including academics and consultants, participate in discussions and technical committee activities without voting rights. Florida DOT (2016) maintains a website for all model users in the state to provide updates on model development, data, and joint activities. Finally, FSUTMS also provides training workshops free of charge to transportation modelers. Training topics covered in the past include CUBE scripting, express lane modeling, data for calibration and validation, destination choice, and others. The consistent model and data structures for statewide and urban models allow close cooperaÂ tion between agencies involved in transportation modeling in Florida. Data are stored in a hierarchiÂ cal system where the statewide model serves as a data resource for all MPO models. A model update implemented in the statewide model can easily be transferred to MPO models, and vice versa. This close integration provides consistency across models in Florida rarely seen elsewhere. Expenditures for model development and data can be shared among agencies, making investments more efficient than in agencies that work in parallel. On the flip side, the statewide transportation model needs to carry considerable detail that is relevant for MPO models. For example, the statewide model includes all urban model TAZs, adding up to 8,518 zones for Florida. With an additional 594 zones in Georgia and Alabama and 185 zones for the rest of the world, the Florida Statewide Model is one of the most detailed models in the country in terms of spatial resolution. Despite the detailed spatial representation, the stateÂ wide model can complete a model run overnight, including person and freight travel. Florida DOT carefully reviews model improvements on runtime and attempts to balance increased runtimes with added value.
88 7.4.2 model structure The person transportation model is a traditional tripÂbased, 3Âstep travel demand model. Floridaâs TurnÂ pike Enterprise has developed a separate model that validated well and has been used to provide input data for the development of the statewide model. Travel demand from the turnpike model and traffic counts were used in a Synthetic Matrix Estimation (SME), sometimes called OriginâDestination Matrix Estimation (ODME), to provide countyÂspecific trip generation rates. Although those rates were someÂ what artificially created based on traffic count data, this approach allowed for distinguishing significant differences in travel behavior between coastal and inland regions. Special generators for airports, ports, and Walt Disney World have been added to trip generation. Trip purposes include homeÂbased work, homeÂbased shop, homeÂbased recreational or social, homeÂbased other, nonÂhomeÂbased, external/ internal, and truck/taxi. Truck/taxi accounts for shortÂdistance trips (<50 miles) of commercial vehicles and is based on a QRFM approach (compare âTravel Demand Modelâ). Trip distribution follows a traditional gravity model approach, taking into account travel times and tolls. It has been kÂfactored using the previously mentioned turnpike model as well as NHTS data. Trip distribution was adjusted at the zonal level to closely match observed travel. Mode choice is limited to an auto occupancy model at this point. The trip generation rates used in the statewide model only cover auto trips, making mode choice superfluous. When trip rates are transferred to MPO models that include transit, they need to be increased accordingly. Florida DOT is currently working on a statewide mode choice model that is expected to include the modes auto (by occupancy), bus, rail and air. External trips are generated at 59 external stations along the state border with Georgia and Alabama. These include internalÂtoÂexternal (IÂE) and externalÂtoÂinternal (EÂI) trips. ExternalÂtoÂexternal (EÂE) trips were considered, though given the peninsular topology of Florida, EÂE trips barely exist and have been excluded. External trips are split into âlocalâ external trips of less than 50 miles and âlongÂ distanceâ external trips that are between 50 and 200 miles. ShortÂ and longÂdistance external trips are calibrated separately using different gravity models. The freight model called FreightSim is the newest addition to the Florida Statewide Model, comÂ pleted in spring 2016. Although the person travel demand model was implemented by AECOM, the freight model was added by RSG. This model was originally developed for Chicago (Outwater et al. 2013) with FHWA funding and then transferred to Florida. The model starts with synthesizing firms by 20 industry categories and seven size categories for the entire U.S. Based on inputÂoutput tables, commodities by 43 SCTG categories are produced and consumed by each synthetic firm depending on their industry category. Next, supplier firms are selected, that is, producers of a given commodity are linked with consumers of that commodity. The supply chain of goods flows may travel through distribution centers and warehouses; such alternative âpathsâ are evaluated, and the preferred path is chosen with a multinomial logit model. Shipment size and shipping frequency are selected next using a discrete choice model. Finally, a mode choice model using a multinomial logit formulation is used to select modes and mode combinations, including road, rail, air, and waterway. The criteria for selecting a mode include travel time, costs, characteristics of the shipment (e.g., perishable, expedited, or containerized), characteristics of the distribution channel (i.e., whether a shipment is routed through a warehouse or distribution center), and whether the shipment includes an intermodal transfer (such as truck to rail to truck). On the truck side, the model further distinguishes heavy and medium trucks. In contrast to the person travel demand model, the freight model is cutting edge and one of the most advanced statewide freight models applied in the country. Florida DOT makes a concerted effort to balance the level of innovation with benefits for the task at hand. According to respondents from the Florida DOT, consultants sometimes try to push for more complex models, while the DOT attempts to carefully assess the âreturn of investment.â They also reported that in some cases simpler models fulfil their needs. The fact that Florida DOT decided to move their freight model to one of the most complicated models on the market underlines the relevance of freight in their statewide transportation planning. At this point, there are no plans to upgrade the person travel demand model by the same degree.
89 The network covers the entire country, with more detail in Florida and a decreasing degree of detail outside of Florida. Network resolution and zone system resolution decrease in detail similarly and gradually the further one travels away from Florida. The Florida Statewide Model does not distinguish timeÂofÂday periods but generates daily traffic. The assignment algorithm is a static user equilibrium approach. Traffic counts were available on 14,726 links (8% of all links in the network). After careful review of count data and elimination of implausible counts (such as counts that changed substantially from one year to the next), the model validated remarkably well. With a RMSE between counts and modeled volumes of 44.4% and a corÂ relation of R2 = 0.81, the model validates better than most other statewide models. In part, this may be the result of the kÂfactoring of trip generation and trip distribution, and the simplification of not having to model mode choice. 7.4.3 model applications The Florida Statewide Model was designed with common scenarios in mind. Model applications focus on statewide analyses of traffic flows, corridor analyses, and analyses of areas outside of MPO model study areas. Typical scenarios tested with the Florida statewide transportation model include: â¢ Operation analyses, â¢ LevelÂofÂservice analyses across the state, â¢ Performance measures for future years, â¢ Roadway projects, â¢ Support of the statewide LRTP, â¢ Freight analyses, particularly in relationship to port activities, â¢ Evaluation of reversibleÂlane counter flows versus allowing hardÂshoulder flows, and â¢ Corridor analyses. The current top priority is the analysis of reversible lanes and hardÂshoulder flows. Freight is expected to play a much larger role in the near future, though given that the advanced freight model was completed only recently, possible applications have not yet been defined. One common application the statewide model is not prepared to handle in its current state relates to managed lanes with variable tolls throughout the day. As the statewide model covers daily volumes only, timeÂofÂday dependent tolls cannot be modeled. At this point, such scenarios are handled by the turnpike model that has a finer temporal resolution. The Florida DOT is currently exploring the posÂ sibility of using a DTA model for selected corridors to model managed lanes with travel demand data from the statewide model. 7.4.4 Future development The present base year is 2010 and the future year is set to 2040. Florida DOT is currently updatÂ ing these years to 2015 and 2045. There are plans to replace the current stick network by a HomeÂ land Emergency Response Exchange (HERE) network to provide enhanced spatial resolution and improved visualization of traffic patterns. Florida DOT is considering a destination choice model that would replace the gravity model in trip distribution, as well as a mode choice model that would add travelers by bus, rail, and air. Particular attention is given to tourist travel. At this point, external travel is represented rather simply through count volumes at external stations and special generators at airports. In the future, Florida DOT envisions generating synthetic travelers that visit Florida and follow selected activiÂ ties. Like the activityÂbased modeling concept, tourists shall be modeled in much more detail to reflect their relevance for state revenues. The tourist model shall be sensitive to, among other factors, increases in energy prices.
90 To further improve freight modeling, a tourÂbased truck model is envisioned as well. The FreightÂ Sim model system was originally developed with a tourÂbased truck model for Chicago. In Florida, however, it was decided to implement the freight demand generation first, gain experience with the model system, and consider tourÂbased truck assignment at a later point. Special effort is given to vehicle flow data collection. Being frustrated with the level of detail proÂ vided by commercial vendors of passively collected data, Florida DOT has started collecting its own data in house. More than 1,000 Bluetooth sensors are used across the state to monitor vehicle flows. Volumes and speed can be gathered from these data collection points. A few portable Bluetooth senÂ sors can be located near gates that are predominately used by trucks, such as driveways near ports. Sensing the same vehicles at other screenlines throughout the state allows Florida DOT to understand truck flows specifically. Over the last 10 years, Florida DOT has invested approximately $1,000,000 per year into their statewide model. This includes expenditures for data and consultant fees, but excludes salaries for DOT staff. Although this is roughly twice as much as spent by the average state DOT in the United States, the expenditure appears rather moderate given the size of Florida and the extra effort the DOT provides through supporting modeling for all MPOs in the state. 7.4.5 lessons learned from Florida Florida DOT makes a concerted effort to explain the benefit of statewide modeling to model users and decision makers. Its survey respondents stated that they are at risk of being in constant âmodelÂ update mode,â overlooking valuable applications using the existing model. Florida DOT attempts to be highly responsive to model requests; in cases where the model is not suited to fulfil the request, staff at the agency attempts to provide other data that may help answer the question, even if the originally requested information cannot be provided by the statewide model. At the same time, the modeling team at Florida DOT attempts to market its efforts proactively to decision makers and staff members in other agencies. Explaining the relevance of statewide modeling helps to keep the program alive and relevant. Specifically, the Florida Statewide Model is actively used to provide input to: â¢ The stateâs highway maintenance program, â¢ The LRTP, â¢ MPO transportation modeling efforts, and â¢ The development of port models. In addition, the agency keeps staff that is working exclusively on training in model applications. To fit into tight schedules of busy employees, the team developed 15Âminute training sessions that cover selected topics, such as model convergence, managed lanes, or external travel. The Model Task Force described in the introduction to this case example turned out to be an excellent vehicle for gathering everyone on the same page regarding transportation modeling. Regular meetings ensure that people are informed and have a forum to contribute to model development and applicaÂ tion. Annual technical meetings and annual application conferences have been instrumental in bringing together experts and advertise the potential of transportation modeling. These meetings have also helped to some degree to reign in unreasonable expectations of what models would be able to accomplish. The biggest challenge for Florida DOT remains the turnover of staff. Too many times has the agency invested into new employees who, once they were wellÂtrained, decided to move into industry. This ongoing shift in personnel requires continuous training of new hires and consumes a substantial amount of resources. 7.5 oregon sTaTewide model The Oregon Statewide Model has been under development since 1998, under the auspices of Oregonâs Transportation and Land Use Modeling Improvement Program (TLUMIP). The modelÂ ing system seeks to fully integrate economic, land use, and multimodal transportation models
91 within a single framework, with linkages to various evaluation and impact models. It motivated a revival in land useÂtransportation modeling in the U.S. and spawned the development of models used elsewhere. The second generation statewide integrated model (SWIM2) is possibly the most ambitious statewide model currently in operation in North America, and the first land useÂtransport model operated at the statewide level. 7.5.1 The oregon modeling Context Oregon has pushed the envelope in land use planning over the past two decades, particularly in the Portland region. An effort was made at the outset of TLUMIP to identify the likely constituents of the model and the issues they were facing. These include legislative staff, decision makers within the Oregon DOT, and MPOs. At that time, the interactions between land use and transportation were of high priority and heavily influenced the design of the modeling system. A little over a decade later the policy focus had shifted considerably. Applications of the firstÂ generation model focused more on the interactions between the economy, jobs, and transportation than land use. Strategies for reducing GHG emissions had become an important topic, as did pricÂ ing, least cost planning (then a priority within the department), and the effects of the 2008â2010 economic downturn. At this writing, most of the previous issues remain, but several new ones have arisen. The ability to understand dynamic pricing is still a priority. Perhaps even more pressing is the need for informaÂ tion about the likely effects of connected and automated vehicles, which appear closer to reality than just a few years ago. Related to that are mobility services such as Uber and Lyft, and many others would like to enter this space. These will change auto ownership patterns, as well as car and rideÂ sharing possibilities. Taken together, these requirements point to the need for a highly flexible and scalable modeling framework that can quickly adapt to new needs. Oregon has invested in strategic visioning models (chapter five, âStrategic Visioning Modelsâ) and agile development [Statewide Integrated Model (SWIM)] to help meet these needs, and continues to seek new approaches that will enable them to do so. 7.5.2 data development The first phase of TLUMIP demonstrated the proof of concept, proving that largeÂscale land useÂ transport modeling was feasible and capable of meeting agency needs. A statewide model was impleÂ mented within the TRANUS framework (De la Barra 2005), and the initial development of an urban model in EugeneâSpringfield evolved into the initial version of UrbanSim (Waddell 2002). A conÂ siderable amount of effort was devoted to assembling the data required to build both models. The goal in both cases was to create the initial models using secondary data sources, both to reduce their cost and implement them quickly. Oregon did not have a prior statewide model, necessitating the development of zone systems and networks from scratch, as well as data unique to land useÂtransport models. The latter included land pricing data, parcel and tract level databases, and economic inputÂ output flows and coefficients. Surveys were conducted at truck weigh stations to better understand the relationship between tonnage by commodity and truckload equivalents. The required socioÂeconomic and land use data are unusually detailed for a statewide model. They include: â¢ 18 household types, stratified by household size and income, â¢ 52 employment categories, â¢ 14 occupation categories, â¢ six residential floorspace types, â¢ 14 nonÂresidential floorspace types, and â¢ 41 SCTG commodity classifications. A statewide travel survey of 14,000 households was carried out in 1998â1999. It was designed to support the development of small urban and MPO travel forecasting models, as well as the statewide
92 travel models. The Oregon Household Activity Survey (OHAS) of around 18,000 households was subsequently conducted in 2009â2011. Both surveys included households across the state, with about 6,500 of them within the PortlandâVancouver metropolitan area as part of the OHAS. Oregon has otherwise relied upon secondary public or government data sources, or has purchased thirdÂparty data, to support their modeling needs. Like other states, they have borrowed data from other places, includÂ ing longÂdistance travel survey data from Ohioâs 2001â2003 statewide household travel survey. As with many other states, Oregon uses the FAF data to depict commodity flows entering or leaving the state. These data are fused with Carload Waybill Sample data to better understand rail traffic patterns. 7.5.3 statewide integrated model (swim) Work began a decade ago on the current statewide model, after completing several applications with the firstÂgeneration prototypes. It was decided to focus on a single framework, rather than separate models at the urban and statewide scales. A modular approach was adopted, where each component would interact with others through data files and messaging. This was thought to facilitate the develÂ opment of each module by different teams, as well as to test different approaches for each, without disrupting the work of others. The resulting modeling platform is a hybrid, in that some models are aggregate equilibrium models, while others are microsimulated. The modeled area covers the state of Oregon and a halo of roughly 50 miles covering counties in the neighboring states of Washington, Idaho, Nevada, and California. A multiÂscale system is used to represent geography at scales approÂ priate for each part of the modeling system: â¢ 2,950 alpha zones cover the model area, and are comparable to traffic analysis zones in most statewide models. The transport models are applied at this level, to include traffic assignment. â¢ 518 beta zones are aggregations of alpha zones and used by the economic allocation and land use modules for computational efficiency reasons. â¢ The rest of the world is represented in six world markets, of which four are shown in Figure 57. The fifth world market is the rest of the world, and the sixth is local markets in neighborÂ ing states, which are used to represent shortÂdistance flows that cross the SWIM study area boundary. FIGURE 57 Representation of external domestic and Canadian markets.
93 A flow chart of the SWIM2 system in shown in Figure 58. The economic, land use, and transportaÂ tion models are tightly integrated. Travel costs and disutilities from the traffic assignment model are fed back to the land use modules in the subsequent simulation period, and land use changes feed back into the economic model. The economic and land use modules run in annual time steps through the simulation, while the transport models are typically run every third year to reduce model run time. A complete model run, covering 30 years in annual steps, takes approximately 8 hours per year using a cluster of six quadÂcore computers, making extensive use of process threading and distributed execuÂ tion. The modules currently implemented include: â¢ The New Economics and Demographics (NED) module determines modelÂwide production activity levels, employment, and imports and exports based on official Oregon state forecasts. This model is an aggregate formulation and is capable of producing different economic futures. â¢ The Synthetic Population Generator (SPG) module samples household and person demographic attributes (SPG1), and subsequently assigns the household to an alpha zone (SPG2). â¢ The Aggregate Land Development (ALD) module allocates modelÂwide land development decisions to alpha zones considering floorspace prices and vacancy rates. â¢ The Activity Allocations (AA) module generates production and consumption flows of land (floorspace), labor, and commodities at the beta zone level. It includes the commodity (goods, services, floorspace, and labor) quantities and prices in all exchange zones to clear markets, including the location of businesses and households by beta zone. â¢ The Person Travel (PT) module generates activityÂbased person trips for each person in the synÂ thetic population during a typical weekday and assigns a workplace by alpha zone. It includes both shortÂ and longÂdistance travel flows, which are combined with commercial vehicle flows in a trip list for traffic assignment. â¢ The Commercial Transport (CT) module maps interÂregional commodity flows into daily truckÂ load equivalents, both within the modeled area and between it and external markets. It also generates internal (to the modeled area) truck tours using a microsimulation framework, to include trips made through distribution centers and intermodal drayage. The roadway and transit network assignments are carried out using the thirdÂparty VISUM packÂ age for four periods of the day. Still under development is an economic feedback (EF) module, which is a simplified dynamic feedback that adjusts the NED modules fixed modelÂwide economic forecast, considering the statewide composite location utilities by industry from the AA module. FIGURE 58 Oregon statewide model flowchart.
94 The SWIM2 system was calibrated and implemented in a staged manner. The preparation and testÂ ing of the software code, preparation of validation and calibration data and targets, and estimation and calibration of initial parameters were completed first. The crossÂsectional calibration and validation of the full model system (i.e., modules working together) were then carried out. Finally, the system was run through time, subjecting it to different policies and stressors (e.g., the introduction of tolling, Interstate highway bridge collapse). A lack of data precluded backcasting. Thus, reasonability checks and stress testing were carried out as part of the final model acceptance testing. The full documentaÂ tion of the statewide model is published in Donnelly and Macfarlane (2016). The design of the activity allocation (AA) module changed over the course of developing the SWIM2 platform. It started out as a microsimulation model, but slowly moved toward an aggregate equilibrium formulation that was inspired, in part, by the MEPLAN model (Echenique et al. 2007). It later evolved into the PECAS, development of which was undertaken in several places in addiÂ tion to Oregon (Hunt and Abraham 2005). As such, it is fair to describe TLUMIP as an incubator of innovative ideas to model land use changes, having already supported the initial development of UrbanSim. The SWIM2 development started with traditional multiÂyear contracting, with a comprehensive model design at the outset. This was the typical approach to largeÂscale model development in the 1990s, when TLUMIP was launched. Later experiences, and failures in some cases, led to the adopÂ tion of the agile development practices becoming more commonly used today. It is a concept from software development that proposes to start with, âthe simplest thing that can possibly work.â The components are then incrementally developed based upon user and client feedback, continual review of requirements, and performance testing. Oregon has benefited from using this approach, despite adopting it while SWIM2 was in development. The questions asked by ODOT stakeholders changed while the program was in motion, and the team shifted development and implementation priorities to accommodate them. This would have required contract amendments and considerable delay had the work been traditionally structured, with deliverables and tasks defined up front. 7.5.4 major applications and implementation The firstÂ and secondÂgeneration models have been used in a variety of studies, but arguably none as important as the ODOTâs Bridge Limitations Study. A design flaw was discovered in 2003 that increased the risk of premature cracking in several dozen bridges across the state. The potential repair costs substantially exceeded the agencyâs budget. However, not addressing them would have resulted in weight restrictions or closures. ODOTâs modeling staff proposed to the departmentâs management that the problem be couched as a community connectivity and economic impact analysis, rather than engineering design issue. It was thought that the legislature would be more receptive if the problem was defined in terms they more quickly understood the implications of. The statewide model was used to test a variety of prioritization packages and effects of bridge limitations. A $4.3 billion program was approved, in part because of the analyses presented to policy makers. Moreover, it resulted in prioritizing repairs for communities most affected by loss of connectivity and jobs, rather than the initial assumption that Interstate highway bridges be addressed first. The model has also been used in a freight bottleneck study, and is currently being used as part of the departmentâs ârough roadsâ assessment. The latter focuses on the implications of inadequate maintenance of existing roadway infrastructure across the state. It has the potential to be as important to the department as the bridge limitations study and is equally well suited for analyses of economic and social impacts. ODOT has invested $10.9 million in consultant fees over the course of the program. However, less than a third of this amount has been devoted to the development of the firstÂ and secondÂgeneration models themselves. A larger amount has been spent on data development, to include travel surveys, acquisition of thirdÂparty data (including other government agencies), and a commodity flow surÂ vey jointly undertaken with Portland Metro and the Port of Portland. Several applications of the model were carried out, both for model acceptance testing and the major studies described earlier. In
95 addition, two ODOT staff members have been involved for at least half of their time in model design, development, and implementation over the entire period that TLUMIP has been active. A final key to the success of the SWIM2 system was an active international peer review panel. The panel has frequently met since the beginning of TLUMIP to oversee model development, enviÂ sion future revisions and extensions, and to present the latest research conducted at universities and in industry. 7.5.5 lessons learned from oregon Most states have neither the need nor resources to develop as sophisticated a modeling system as Oregon. However, many of the important lessons learned apply to states with less ambitious goals. Some of the key findings from Oregon include: â¢ The questions that policy makers are facing are changing faster than models used to inform them. The key requirements changed not just once, but twice over the lifespan of TLUMIP, suggesting the need for a more flexible and quickly adaptive modeling framework than typically found in practice. â¢ Along the same lines, the analytical requirements dictated what Oregon wanted to accomplish, rather than simply desiring to have âbetter modelsâ or pursuing the state of the art. Their conÂ tinual reÂexamination of priorities of decision makers has required shifting focus several times, but has kept their work relevant to their clients. Oregon would not likely have invested in such a sophisticated modeling system without such requirements. The reminder that applications should dictate design rather than the other way around is a lesson worth repeatedly learning, and has been key to sustained support for their program. â¢ The agile development process is a far better framework for the development of major new platforms than the traditional âbig design up front,â followed by several years of data and model development. Requirements change over time, new data and methods become available, and experience with interim products often dictates small but important changes in the modeling system that better address the needs of statewide modelers and their clients. â¢ The ability to model freight, its economic linkages, and to visualize truck flow patterns is as important, and perhaps more so than modeling person travel flows. Most of the major applicaÂ tions of the model to date have involved freight, either wholly or in part. The work in Oregon remains a work in progress, and perhaps will always be so. The constantly changing policy environment, the level of questions posed by decision makers, their zeal for pushÂ ing the envelope in land useâtransport modeling, and upper management support ensures that they will remain at the forefront of statewide modeling. The investment ODOT have made in its staff is particularly evident, giving it the ability to do many things that many agencies cannot.