Transcription

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day SimulatorSoundCastActivity-Based Travel Forecasting Model for PSRCFeaturing DAYSIM—the Person Day Activity and Travel SimulatorModel System DesignSeptember 25, 2014Prepared forPuget Sound Regional CouncilPrepared byJohn L. Bowman, Ph. D.Transportation Systems and Decision Sciences5 Beals Street Apt. 3, Brookline, MA 02446 USA 1-617-232-3478 John L [email protected]://JBowman.netMARK BRADLEYBRADLEY RESEARCH & CONSULTING524 Arroyo Ave., Santa Barbara, CA 93109, USA. 1-805-564-3908 mark [email protected] Sound Regional CouncilResource Systems GroupCambridge Systematicspage 1

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day SimulatorTable of ContentsTable of Contents . 2Introduction and Model System Overview. 3Model system . 3Daysim . 4Model variables . 5Population Synthesizer . 6Base year synthetic population . 7Forecast year synthetic population . 8Long Term Choice Simulator and DaySim . 8Long term choice models . 11Day level models . 11Tour level models . 11Trip/stop level models . 12Accessibility variables . 14Recent Updates to Daysim . 17The parcel land use data . 17Changes to support the treatment of policy-based pricing . 19The use of path type choice models for all modes . 21Changes to specific DaySIM choice models . 23Supplemental Trip Modeling: External, Trucks, Special Generation, and Group Quarters . 25External Modeling . 25Special Generators . 27Group Quarters . 27Truck Model . 29Truck Model Parameters . 31Network Assignment and Skimming. 35Skim and Trip Data Exchange . 35Skims . 36Daysim Trips to EMME . 39Appendix 1—DaySim Model Features . 41Appendix 2—DaySim Variables . 44Appendix 3 : DaySIM Software and Other Detailed Improvements . 49Traveler- & tour-specific model coefficients . 50Park and ride path type and lot choice model . 51Treatment of transit pricing . 51Pay to park at workplace model - estimation . 52DaySim Software and Hardware . 52page 2

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day SimulatorIntroduction and Model System OverviewModel systemSoundCast is a travel demand model system built for the Puget Sound Region, as shown inFigure 1. The model was built to depict diverse human travel behavior and include travelsensitivity to land use and the built environment. SoundCast outputs transportation networkmeasures such as highway volumes in one hour periods in a future year or number of boardingson a transit line. It also outputs measures related to people like average distance to work byhome county or the number of transit trips different types of people will take.The three main components of SoundCast are: person trip demand in the Daysim activity-based model external, special generation, truck, and group quarters aggregate modeling assignment and skimming in EMMEDaySim is a modeling approach and software platform to simulate resident daily travel andactivities on a typical weekday for the residents of a metropolitan region or state.In essence, DaySim replaces the trip generation, trip distribution and mode choice steps of a 4step model, while representing more aspects of travel behavior (auto ownership, trip chaining,time of day scheduling, detailed market segmentation, etc.)Daysim integrates with EMME by generating resident trip matrices for assignment and uses thenetwork skims from assignment for the next global iteration of DaySim.The major inputs to SoundCast are transportation networks and modeled household andemployment data from UrbanSim. In Daysim, The Population Synthesizer (PopSyn) creates asynthetic population, comprised of Census PUMS households, that is consistent with regionalresidential, employment and school enrollment forecasts. Long-term choices (work location,school location and auto ownership) are simulated for all members of the population. ThePerson Day Activity and Travel Simulator (DaySim) creates a one-day activity and travelschedule for each person in the population, including a list of their tours and the trips on eachtour.The trips predicted by DaySim are aggregated into EMME trip matrices and combined withpredicted trips for special generators, external trips and commercial traffic into time- and modespecific trip matrices. The EMME network traffic assignment models load the trips onto thenetwork. Traffic assignment is iteratively equilibrated with the Long Term Choice Simulator,DaySim and the other demand models. The parcel level land use inputs come from UrbanSim.page 3

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day SimulatorFigure 1: New PSRC Regional Travel Forecasting Model SystemDaysimThe following section describes the design features of PopSyn, the long term choices andDaySim. These include a description of each model component, definitions of the variablesincluded in the simulated output, details about accessibility variables employed to help integratethe model system, and the sampling procedure used in the destination choice models. The submodels in the system are:1. Work Location2. School Location3. Pay to Park at Work4. Transit Pass Ownership5. Auto Ownership6. Individual Person Day Pattern7. Exact Number of Tourspage 4

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day Simulator8. Work Tour Destination9. Other Tour Destination10. Work-based subtour Generation11. Work Tour Mode12. Work Tour Time13. School Tour Mode14. School Tour Time15. Escort Tour Mode16. Escort Tour Time17. Other Tour Mode18. Other Tour Time19. Work-Based Subtour Mode20. Work-Based Subtour Time21. Intermediate Stop Generation22. Intermediate Stop Location23. Trip Mode24. Trip TimeModel variablesTable 1 lists the variables that will be produced by the Daysim models. The variables are at fivedifferent levels: household, person, person day, tour and trip. The table also lists the range ofvalues that will be used for each output variable. Table 1 contains only the most elementalvariables. More output variables can be computed in combination with the network and/or zonaldata, such as the VMT traveled by a person.page 5

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day SimulatorTable 1—Elemental variables produced by PopSyn and DaySimLevelVARIABLE RNOGENDAGEWORKERSTUDENTHRSWORKWPCLSUPARCELPerson DayTourTOURNOPDTYPEVariable Descriptionhousehold ID number# persons in HH# vehicles in HHtotal household incomehousehold residence parcelperson ID numberGenderAgeemployment statusstudent status# hours worked per weekusual work location parcelusual school location parceltour ID number(in simulation order)primary destination purposetypeRange of Values0-100-4 0-98 employed, not employedUniversity student, grade school student, nonstudent1-work5-shopping2-school6-mealOPCLTour origin location parcelDPCLPrimary destination loc. parcelMMODEtour main mode(may be an aggregated set ofthe 9 modes)1 –walk 2 –bike3 – sov 4 –hov2 5 –hov3 6 –walk-transit7 –park and rideTrip tour halfTrip ID within tour half(outward from primary dest)Trip origin purpose typeTrip destination purpose typeTrip origin parcelTrip destination parcelTrip origin arrival timeTrip origin departure timeTrip destination arrival timeTrip destination departure timeTrip mode1st, ec4-per.bus8-homeHome parcel for home-based toursWork tour destination location for work-based tours8-school bussee tour primary destination purposesee tour primary destination purpose30-minute time periods30 10-minute time periods30 10-minute time periods30 10-minute time periodssee tour main modePopulation SynthesizerThis model/procedure produces a list of household and person records from the PUMSmicrodata. Each household is defined in terms of income and household size, plus the age,gender, employment status and student status of all household members. Using CTPP and STFtables in the base year, appropriate numbers of each type of household are allocated to eachTAZ. In forecast years, these numbers are adjusted according to demographic forecasts from theland use model and any additional sources. Parcel level inputs on residential land use are used tofurther allocate households to parcels.SoundCast uses the population synthesizer (PopSyn) also used by Atlanta Regional Commission.Figure 2 provides a schematic of PopSyn, showing key inputs and outputs for the base year anda forecast year, and the procedures are described in the next two subsections.page 6

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day SimulatorBase year synthetic populationBy far the best available detailed information about households comes from the US census.Therefore, the model system is set up to use a census year (2000) as the base year for modelforecasts, and PopSyn is designed to extensively use census data to create the base year SynPop.Census SF1, SF3 and CTPP tables provide rich information about the distribution of variousimportant household characteristics within each census block [SF1] or block group [SF3, CTPP].Many of these tables are multidimensional; that is, the table provides information about the jointdistribution of two or more important variables. PopSyn is set up so that it can synthesize a baseyear population that matches any number of desired multidimensional SF1, SF3 and CTPPdistributions at the TAZ level of detail.The distribution of households is synthesized through an iterative proportional fitting (IPF)procedure called ‘Balancer’ that is like a traditional Fratar procedure for balancing trip ends,except the ‘cells’ of the joint distribution are defined by household characteristics and the controlvalues can apply to any designated subset of cells. For the base year, Balancer’s ‘seed’distribution is the joint distribution observed in the census 5% Public Use Micro Sample(PUMS). The PUMS distribution is used because each PUMS household has enough dataavailable to assign it precisely to one household demographic category (HHCat) defined jointlyby several different variables. This allows us to define HHCats to take advantage of the SF1,SF3 and CTPP tables, and still have a reliable seed distribution. Since PUMS data is stripped ofdetailed geographic information, the seed distribution for each TAZ is the distribution of thePUMA to which it belongs.Figure 2: Basic inputs, processes and outputs of population synthesizer (PopSyn)SF3 tablesCTPP tablesSF1 tables(Provide controltotals for Balancer)PUMS 5% sampleLand Use Forecasts(provides seed forbase year Balancer,plus HH for Drawer)(Control totals for Balancer)PECAS (or PLACE3S)by TAZ:--HH by income--# jobs--floorspace by housing typeFORECAST YEARBASE YEARBalancerBalancer(Estimate joint distributionby iteratively fitting PUMA’sPUMS seed matrix tocensus table control totals)Drawer(draw PUMS HH)Base YearSynthetic population--1 record per HH--1 record perpersonBase year jointdistribution(provides seed forforecast year Balancer)Regionwide:--Pop under age 18--Pop age 65 (Estimate joint distributionby iteratively fitting baseyear joint distribution toSACOG forecast controltotals)Drawer(draw PUMS HH)Forecast yearjoint distributionForecast YearSynthetic population--1 record per HH--1 record perpersonOnce Balancer determines the distribution of households by HHCat within TAZ, then the secondmajor step in PopSyn—HHDrawer—creates the SynPop by drawing, for each TAZ, the correctnumber of households of each HHCat from the PUMS households with matching HHCat andpage 7

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day SimulatorPUMA. Then, parcel level inputs on residential land use are used to further allocate householdsto parcels. Since the number of households determined by Balancer is fractional, HHDrawer ispreceded by a procedure that ‘integerizes’ the IPF results, preserving the distribution as much aspossible. Also, since the number of households within a particular HHCat for a given PUMAmay be small, Drawer is set up to draw from similar PUMAs if the same household wouldotherwise be drawn more than a prescribed number of times. PUMA similarity and themaximum number of times that a household may be drawn is specified in the control file.In summary, PopSyn creates the base year SynPop in two steps called Balancer and HHDrawer.Balancer is an iterative proportional fitting procedure that estimates the base year distribution ofhouseholds by household demographic category (HHCat) for each TAZ. HHDrawer is asampling procedure that populates each TAZ by drawing the correct number of households ofeach HHCat from census PUMS data. For the base year, PopSyn matches exactly the targetsdetermined by census SF1, SF3 and CTPP tables at the TAZ level, while preserving to the extentpossible the full multi-dimensional distribution observed in PUMS at the PUMA level.Forecast year synthetic populationPopSyn uses the same two steps, Balancer and HHDrawer, to synthesize the population for aforecast year, but it uses regional forecasts from (PLACE3S or PECAS) as input instead ofcensus data. Balancer creates a forecast population distribution that matches the following PSRCforecasts: (a) households by income category in each TAZ, (b) number of jobs held by employedpersons living in each TAZ, (c) floorspace by housing type in each TAZ, (d) number of personsaged 65 and older in the region, and (e) number of persons aged 0-17 in the region. Like thebase year, PopSyn’s forecast inputs come from input parameters in its control file, so it would bepossible, without software programming, to fairly quickly and inexpensively adjust PopSyn tomatch other regional forecasts.Since the available forecast year information can be quite limited, and the distribution ofhousehold and personal characteristics change gradually over time, Balancer is set up to preservethe base year distribution as much as possible while matching the above-described forecastcontrol totals. That is, Balancer uses the base year distribution created by PopSyn as its seeddistribution for the forecast year. However, since the distribution at the TAZ level of geographymay not be very stable over time, Balancer’s seed distribution for each TAZ is a blend of theTAZ, census tract and PUMA base year distributions. The exact blend for each TAZ depends onthe sizes of the TAZ and its tract, and is determined by easily changed parameters in the controlfile; the bigger the TAZ, the more heavily it weighs in the blend.Long Term Choice Simulator and DaySimFigure 3 presents the DaySim model hierarchy, embedded within the program looping structurein which the models will run. Program loops are bounded by lines starting with ‘Begin’ and‘End’, and indentation indicates embedded sub-loops. The models themselves are numbered.For each household, the long term choice models (1.2-1.4) run first. Then, a loop runs for eachperson, in which their day pattern (models 2.1-2.2) is simulated. Within that loop, each tour ofthe pattern is simulated in turn (models 3.1-3.4), and each stop is simulated within each tourpage 8

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day Simulator(models 4.1-4.4). Work-based tours are modeled as tours, but at the same level of priority asstops on the way to and from work.The next subsections describe each of the model types. Additional details about each model canbe found in tabular form in Appendix 1, including the model type, output variables, andimportant variables that it uses. Appendix 2 provides a detailed list of variables produced by theDaySim models, including for each a reference to the model that produces it.page 9

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day SimulatorFigure 3—DaySim models (numbered) within the program looping structureBegin{Read run controls, model coefficients, TAZ data, LOS matrices,population controls, and Parcel data into memory}{Draw a synthetic household sample if specified}{Pre-calculate destination sampling probabilities}{Pre-calculate (or read in) TAZ aggregate accessibility arrays}{Open other input and output files}{Main loop on households}{Loop on persons in HH}{Apply model 1.1 Work Location for workers}{Apply model 1.2 School Location for students}{Apply model 1.1 Work Location for students}{End loop on persons in HH}{Apply model 1.3 Household Auto Availability }{Loop on all persons within HH}{Apply model 2.1 Activity Pattern (0/1 tours and 0/1 stops)and model 2.2 Exact Number of Tours for 7 purposes}{Count total home-based tours and assign purposes}{Initialize tour and stop counters and time window for the person-day before looping on tours}{If there are tours, loop on home-based tours within person in tour priority sequence,with tour priority determined by purpose and person type}{Increment number of home-based tours simulated for tour purpose (including current)}{Apply model 3.1 Tour destination}{If work tour, apply model 3.2 Number and purpose of work-based subtours}{Loop on predicted work-based sub tours and insert then tour array after current tour}{Apply model 3.3 Tour mode}{Apply model 3.4 Tour primary destination arrival and departure times}{Loop on tour halves (before and after primary activity)}{Apply model 4.1Half tour stop frequency and purpose}{Loop on trips within home-based half tour (in reverse temporal order for 1st tour half)}{Increment number of stops simulated for stop purpose (including current)}{Apply model 4.2 Intermediate stop location}{Apply model 4.3 Trip mode}{Apply model 4.4 Intermediate stop departure time}{Update the remaining time window}{End loop on trips within half tour}{End loop on tour halves}{End loop on tours within person}{Write output records for person-day and all tours and trips}{End loop on persons within household}{End loop on Households}{Close files}{Create usual work location flow validation statistics}End.page 10

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day SimulatorLong term choice modelsWork location (1.2) and School location (1.3)These are essentially destination choice models, but they determine the longer term choice ofusual work and school locations (parcel within TAZ). These, along with residence location, tendto structure a person’s spatial activity patterns. The choice is primarily a function of travelaccessibility across all modes and land use characteristics in and surrounding each possible TAZand parcel. Key segmentation variables include income for workers and age group for students.In the model sequence, work location conditions the school location for most workers, but foruniversity and young driving age students, school location conditions work location.Auto availability (1.4)This model is applied at the household level, and determines the number of vehicles available tothe household drivers. Key variables are the numbers of working adults, non-working adults,students of driving age, children below driving age, income, auto and non-auto accessibilities towork and school locations, and more general pedestrian, transit and auto accessibility to retailand service locations.Day level modelsDay activity pattern (2.1-2.2)This model is a variation on the Bowman and Ben-Akiva approach, jointly predicting the numberof home-based tours a person undertakes during a day for seven purposes, and the occurrence ofadditional stops during the day for the same seven purposes. The seven purposes are work,school, escort, personal business, shopping, meal and social/recreational. The pattern choice is afunction of many types of household and person characteristics, as well as land use andaccessibility at the residence and, if relevant, the usual work location. The main pattern model(2.1) predicts the occurrence of tours (0 or 1 ) and extra stops (0 or 1 ) for each purpose, and asimpler conditional model (2.2) predicts the exact number of tours for each purpose.Tour level modelsWithin each tour, three main models are used, to first simulate the tour’s destination, then thebeginning and ending period of the tour’s primary activity, and finally the main mode used forthe tour. For work tours, the number of work-based subtours is also modeled, after destinationchoice, and before timing and travel mode.Destination choice (3.1)Similar to the work and school location models, these models determine the primary destinationTAZ and parcel for home-based tours and work-based subtours. For the primary tourdestination, the logsum from the mode choice model across all modes is used as the main level ofservice variable.page 11

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day SimulatorThe universal choice set of destinations is very large, including all parcels within themetropolitan area. In any given situation, some of the parcels will be infeasible, either becausethe location cannot be reached in the available time, or because the desired activity cannot beaccomplished there. Also, for the sake of computational feasibility, the huge size of the choiceset makes it necessary to sample alternatives when applying the destination choice models. Asampling procedure has been designed to deal with both of these issues. The availablealternatives are sampled in a way that allows the probability of being drawn into the sample to becalculated for each drawn alternative. Statistical procedures are then used during modelestimation and application to allow the sample to represent the entire set of available alternativeswithout biasing the results.The chosen sampling procedure is called two-stage importance sampling with replacement. Inthe first stage, a TAZ is drawn with a known probability approximately equal to its chance ofcontaining the chosen destination. Then, a parcel is drawn within that TAZ with a knownprobability approximately equal to its chance of being the chosen parcel within the TAZ. Thetwo main criteria used in the design of the procedure are statistical soundness and computationalefficiency. A later technical memo on the location choice models will document theseprocedures in detail.Number and purpose of work-based tours (3.2)For this model, the work tour destination is known, so variables measuring the number andaccessibility of activity opportunities near the work site are expected to influence the number ofwork-based tours.Tour main mode (3.3)The tour mode choice model determines the main mode for each tour (a small percentage oftours are multi-modal), with the alternatives being drive to transit, walk to transit, school bus, carshared ride 3 , car shared ride 2, car drive alone, bike and walk.Primary activity periods (3.4)The dependent variables of this choice model are a pair of 30 minute time periods representingthe times that the person arrives at and departs from the tour primary activity location. Ittherefore provides an approximation of both time-of-day and activity duration. Since entiretours, including stop outcomes are modeled one at a time, first for work and school tours andthen for other tours, the periods away from home for each tour become unavailable forsubsequently modeled tours. The time period of a work-based subtour is constrained to bewithin the time period of its parent tour.Trip/stop level modelsAlthough the presence of extra (intermediate) stops in the day pattern is determined in the patternmodel, the exact number of stops for each purpose is a result of the stop level models. Withineach tour, the stops are modeled one-by-one, first for stops before the tour destination, and thenfor stops after the tour destination. This is an iterative model structure, very similar to the oneused in Model 3.2 for the number and purpose of work-based subtours.page 12

SoundCast: PSRC Activity-Based Travel Forecasting ModelFeaturing DAYSIM—the Person Day SimulatorStops before the tour destination are modeled in reverse temporal sequence. First the possibleparticipation in a stop is modeled simultaneously with the stop’s purpose (4.1). If the stopoccurs, then its location (4.2), and then its trip mode (4.3), and finally the 10-minute time periodof the arrival at the tour destination (4.4) are modeled. These results also determine the timeperiod in which the trip from the stop location begins, since the trip mode and travel level ofservice are known. If a stop occurs, then the possible participation and purpose of a prior stopare modeled, along with details of location, trip mode and timing. This continues, constructingthe trip chain from the tour primary destination to the tour origin in reverse chronologicalsequence until the model predicts no more stops (at which point, the “final” trip between the“last” stop and the tour origin is modeled). The reason for modeling in reverse chronologicalsequence for the first half tour is the hypothesis that people aim to arrive at the primarydestination at a particular time, and adjust their tour departure time so as to enable completion ofthe desired intermediate stops. After the trip chain for the first half-tour is modeled, the tripchain for the second half-tour back to the tour origin is similarly modeled, but this time in regularchronological order.Number and purpose of intermediate stops (4.1)Throughout the construction of the trip chains, the making of intermediate stops by purpose isaccounted for, so that as stop purposes called for by the pattern model are accomplished, thelikelihood of additional stops decreases.Intermediate stop location (4.2)For intermediate stop locations, the main mode used for the tour is already known, so the choiceis primarily a tradeoff between the additional deviation and impedance of making another stopby that mode versus the accessibility to additional land use opportunities in alternative zones andparcels.As with tour destinations, a sampling procedure is required for the stop location models, and aprocedure has been designed that employs importance sampling with replacement. The exactprocedure is different, however, because the sampling problem

12. Work Tour Time 13. School Tour Mode 14. School Tour Time 15. Escort Tour Mode 16. Escort Tour Time 17. Other Tour Mode 18. Other Tour Time 19. Work-Based Subtour Mode 20. Work-Based Subtour Time 21. Intermediate Stop Generation 22. Intermediate