APPLICATION OF A COMBINED TRAVEL DEMAND AND MICROSIMULATION MODEL FOR A SMALL CITY

Transcription

1 APPLICATION OF A COMBINED TRAVEL DEMAND AND MICROSIMULATION MODEL FOR A SMALL CITY Danel Morgan Transportaton Engneer Calper Corporaton 1172 Beacon Street, Newton, MA Phone: (617) Fax: (617) danel@calper.com Rck Mayberry Transportaton Engneer Calforna Department of Transportaton 1656 Unon Street, Eureka, CA Phone: (707) Fax: (707) rck_mayberry@dot.ca.gov Ths paper descrbes the development and applcaton of a mult-modal mcrosmulaton model for the Greater Eureka Area (GEA), a small communty of 25,000 n Northern Calforna. The travel demand model for the GEA was used to develop estmates of the traffc demand for base and forecast years. A rgorous data collecton and calbraton effort was made to calbrate the smulaton model for the base year. The smulaton of pedestran actvty and bus routes s ncluded n the model. The applcaton s unque n ts methods and n ts ultmate objectve. Frst, the mcrosmulaton model was developed on a geographc nformaton system platform shared wth the travel demand model, allowng the fuson of geographc nformaton and the applcaton of geographc analyss methods to assst n the refnement of peak perod trp tables for smulaton. The travel demand model was used to develop ntal estmates of the traffc demand. Addtonal analyss was performed to develop a dynamc temporal profle n the demand. Smulaton-based dynamc traffc assgnment methods were used to calbrate route choces n the model. Second, the mcrosmulaton model was desgned not for a specfc and fnte project n the common tradton of plannng and engneerng practce but for the purpose of becomng a lvng model to be adopted and mantaned by local authortes for use n all manner of plannng and traffc mpact studes, both bg and small, throughout the cty. Thus, the mcrosmulaton model wll serve as a natural corollary and complement to the travel demand model. A varety of alternatves, ncludng the addtonal lanes on a key corrdor and traffc sgnal optmzaton, are analyzed to demonstrate the effectveness of mcrosmulaton for mprovng the estmates of project mpacts n the plannng context. In addton to servng as an llustratve case study for the applcaton of mcrosmulaton n small and medum-szed communtes, ths paper demonstrates the advantages of GIS for makng the development of a mcrosmulaton model for small and medum-szed communtes feasble and cost-effectve. Lessons learned and gudance for smlar applcatons elsewhere are provded. Submttal s for a presentaton and wrtten paper. Descrpton of paper: A case study demonstratng the development, calbraton, and applcaton of a mcrosmulaton model n a small communty, ncludng travel demand forecastng, trp matrx estmaton, and dynamc traffc assgnment technques. Key words: traffc smulaton, mcrosmulaton, travel demand model, dynamc traffc assgnment Student status: Nether of the authors s a student

2 2 APPLICATION OF A COMBINED TRAVEL DEMAND AND MICROSIMULATION MODEL FOR A SMALL CITY INTRODUCTION Wth the rsng profle of mesoscopc and mcroscopc traffc smulaton models and hgher fdelty, notably dynamc, traffc assgnment methods n the plannng communty, there s lttle n the way of gudance or standards of practce when t comes to estmatng, calbratng, or applyng these models. Even as the experts struggle to defne dynamc traffc assgnment (DTA), there s lttle experence for travel demand modelers to draw upon. Ths paper descrbes a methodology and case study for a successful applcaton of a combned travel demand and mcrosmulaton model for a small cty. More specfcally, ths paper wll focus on the mcroscopc smulaton-based DTA methods that were used n the estmaton and calbraton of the model. The mcrosmulaton model spans the whole of Eureka, Calforna, whch represents the core of the Greater Eureka Area (GEA) Travel Model, the travel demand model that covers all of Humboldt County. The purpose of the project s to develop a traffc mcrosmulaton model that extends and complments regonal plannng actvtes dependent, for the most part, on the Greater Eureka Area Travel Model, whch, lke most travel demand models, lacks the detal and fdelty to be used to evaluate projects or to estmate the operatonal mpacts of growth, changes n land use, or roadway and traffc sgnal mprovements. To a lesser extent, modelng actvtes nvolve other traffc smulaton and traffc sgnal tmng analyss software, but none that bear any formal relatonshp wth the travel demand model or wth one another. Gven that the Humboldt County Assocaton of Governments (HCAOG) has nvested over a perod of years, and wll contnue to nvest, the energes and resources of ts member governments to develop and mantan the travel demand model, a mcrosmulaton model that bulds on and complements the travel demand model was also a motvatng factor. The desgn of the mcrosmulaton model s such that t wll rely on the travel demand model to produce base estmates of the traffc demand for peak perods, but use feld data to mprove those estmates by way of ndustry-standard calbraton and valdaton technques. The geographc scope of the mcrosmulaton model s far greater than that of a typcal traffc smulaton project, ncludng every street n a relatvely dense grd network coverng more than 16 square mles. To contend wth the trp table estmaton and route choce challenges that are pvotal to the successful calbraton and valdaton of such a model, local geographc nformaton, ncludng nformaton about the predomnant land uses n each traffc analyss zone, s used. Once estmates of the trp tables for the AM and PM peak perods are produced usng the travel demand model, feld data, and geographc nformaton, the most formdable techncal barrer to smulatng areas as wde as even a small cty s the queston of route choce. The smulaton-based DTA model that s the subject of ths paper s the prncple tool used to decde the routes that drvers wll take n the GEA mcrosmulaton model. The DTA model s based on prncples that are famlar to planners, such as User Equlbrum (UE), but uses route choce methods to determne the

3 3 paths to whch vehcles are assgned and mcroscopc traffc smulaton models (e.g., car followng, lane changng) to determne lnk performance (.e., loaded travel tmes and delays). PROBLEM STATEMENT The methodology n ths project s desgned to answer two questons, the answers to whch are nextrcably lnked and nterdependent: (1) What are the volumes of vehcles travellng between orgn and destnaton zones n the network? (2) What are the lkely routes drvers take between those orgn and destnaton zones? The challenges n answerng these questons effectvely and accurately stem mostly from lmtatons n exstng methodologes and n the data that s typcally used to answer them. To truly understand the trp pattern and route choces of drvers n a regon, t s mperatve to drectly observe the orgns, destnatons, and routes. Ths can be acheved wth lcense plate surveys, for example. Other nnovatve methods have been used that track or match the denttes of vehcles observed at dfferent locatons at dfferent tmes, such as vdeo recorded from arplanes crculatng above a ste. TRAFFIC DATA Traffc data used to estmate trp tables are commonly lmted to what can be nexpensvely collected or has already been collected, namely traffc counts. Ths dependence almost exclusvely on counts has a number of drawbacks: (1) Counts reveal nether the orgns and destnatons of vehcles nor ther routes. (2) Poor coverage of the study area may leave lnks on key routes between orgn-destnaton pars countless, whch degrades the qualty of the orgn-destnaton matrx estmaton (ODME) soluton. (3) Analysts are tempted to combne counts from dfferent days, or even years, to ncrease coverage, or to average counts together to reflect an average day. These problems are sources of error and uncertanty n the trp table soluton, and tend to be, at worst, poorly understood or overlooked and, at the very least, underapprecated. ODME METHODS Furthermore, ODME methods are mperfect. ODME methods use tradtonal, typcally statc (.e., one tme perod) traffc assgnments to load trps from a matrx onto a network. Loaded flows are compared wth counted volumes n order to calculate an adjustment to the matrx that, when loaded agan, wll mprove the match between assgned flows and counts. Ths procedure contnues teratvely untl the match between the flows loaded from the estmated matrx and the counts cannot be mproved further. Two shortcomngs wth these methods are, lke the shortcomngs n the traffc count data descrbed above, also not well or wdely understood:

4 4 (1) The soluton s heavly nfluenced by the matrx used n the ntal loadng (.e., the seed matrx). (2) Volumes n a cell n the matrx are adjusted based on the flows and counts on lnks on the used path(s) between the correspondng orgn-destnaton (OD) par; crude heurstcs must be used to estmate volumes between OD pars between whch no counts are avalable on the path(s) between them. (3) No unque soluton can be proven to exst, meanng that any number of estmated matrces mght match the counts equally well when assgned to the network. In other words, a good match wth the counts does not n and of tself prove a good estmate of the trp pattern. The seed matrx that s of such crtcal mportance to the qualty of the ODME soluton s usually produced by a subarea analyss n a travel demand model. Thus, a poorly calbrated travel demand model can be, alongsde the traffc count data, yet another source of error n the ODME soluton. In summary, effectve use of state-of-the-practce ODME methods requres a thoughtful consderaton of these lmtatons. Experence applyng the methods and the know-how to dentfy defects n the soluton are almost a precondton for success. ROUTE CHOICE CONSISTENCY Lastly, once a matrx s estmated, assumng that t can be trusted to be representatve of the real pattern of travel n the study area, traffc smulaton models use route choce models to determne the paths vehcles choose to take. Routes determned by tradtonal statc UE assgnments n travel demand models, a temptng source for the routes to be used a mcrosmulaton model, are nether unque nor relable, and are n all lkelhood probably nfeasble. More recent developments n traffc assgnment methods n the lterature (Slavn et al., 2009; Calper Corporaton, 2010) mght one day change ths fact. However, there s not yet any proof or reason to beleve that the routes reported by a statc UE assgnment of any knd wll yeld a sensble set of paths for mcrosmulaton. The more useful route choce models n the mcrosmulaton context are behavoral, meanng that drvers choose a route that mnmzes some perceved total travel tme or generalzed cost or that acheves some other objectve. Each drver makes hs or her decson out of self nterest, ndependently of the decsons of others. The loadng mechansm (e.g., user equlbrum) used n the ODME methods are thus nherently nconsstent wth the route choce loadng. In other words, even f the match between the loaded flows and counted volumes resultng from the ODME s very good, there s no guarantee that the route choce model wll produce the same loadng. The match between the smulated flows and the counts mght possbly be worse than that acheved by the ODME. Only a smulaton-based ODME, usng a route choce loadng consstent wth the smulaton, can overcome ths nherent nconsstency. The methodology presented n ths paper recognzes these shortcomngs and proposes alternatve methods to help mprove the estmaton and calbraton.

5 5 GENERAL METHODOLOGY The methodology used to estmate and calbrate a dynamc trp table for mcrosmulaton of the GEA s summarzed n Fgure 1. Analyss of Land Use Data Constrant Matrx Traffc Counts Yes Fgure 1. A flow chart llustratng the trp table and route choce estmaton methodology SUBAREA ANALYSIS The frst step n the methodology s to perform a subarea analyss n the travel demand model. The core of the GEA study area s approxmatly 3.5 mles by 3.5 mles, spans the cty lmts, and covers parts of surroundng Humboldt County. Stretches of Route 101 several mles beyond that 3.5 x 3.5- mle core are ncluded n the subarea. The subarea s performed n the most recently calbrated base year 2005.

6 6 The subarea analyss s a tradtonal statc UE traffc assgnment for the AM and PM peak hours. It s run to a relatve gap of 10-6, beyond whch only very mnute changes n the flow vector occur over contnued teratons, where the relatve gap s computed as: Gap = fk tk I k K I I d t d mn, t mn, where: Gap = Relatve gap; I = Set of all O-D pars ; K = Set of paths used by trps travelng between O-D par ; f k = Number of trps takng path k (.e. path flow); t = Travel tme on path k; k d = Demand departng; t mn, = Travel tme on shortest path between O-D par ; The output of ths procedure s the matrx that s used as the seed for the ODME. ORIGIN-DESTINATION MATRIX (TRIP TABLE) ESTIMATION Traffc count data was collected throughout the cty n the Sprng of 2009, the base year for the mcrosmulaton model calbraton. So that all counts could be observed durng the same hours of the same days for purposes of consstency, cameras were used to record ntersectons so that turnng movement volumes could be counted durng back-offce data reducton. Elsewhere, manual turnng movement counts and pneumatc tubes were used to collect traffc counts. These count data were processed and compled nto a geographc count database n TransModeler, the mcroscopc traffc smulaton platform used n the project. The mnmum requrements to perform the ODME are a seed matrx and traffc counts. However, t was found that, due to some of the ssues wth ODME methods descrbed earler, the ODME solutons based on these nputs alone are nadequate. Eureka has a farly dense, mostly grd layout, whch allows for numerous feasble paths between a sgnfcant majorty of orgn-destnaton pars. Furthermore, the traffc analyss zones are farly small, resultng n a rather dense geographc dstrbuton of centrods (.e, around 50 per square mle). The dense zonal structure and grd street network present stff challenges for the ODME effort. For nstance, the ODME soluton may acheve excellent results n terms of matchng counts by producng unreasonable volumes of trps travelng short dstances between agrcultural and low-densty resdental areas. Thus, to steer the ODME toward a more probable soluton, geographc analyss was used to produce a matrx of constrants that would lmt the volume of trps that can be produced n the ODME between zones a very short dstance a part and havng predomnantly agrcultural, low-densty resdental and other land uses not expected to be sgnfcant producers or attractors of trps.

7 7 Usng the seed matrx from the subarea analyss, the traffc counts, and the constrant matrx, an ODME soluton was produced for the AM and PM peak perods that matched counts wth a root mean square error of less than 10% and that satsfed general a pror expectatons about trp pattern n the cty. SIMULATION-BASED DYNAMIC TRAFFIC ASSIGNMENT The network loadng (.e., the lnk flows resultng from the trp assgnment) n the ODME method used n ths applcaton s based on a tradtonal statc (.e., sngle tme perod) UE traffc assgnment. Thus, the ODME loadng does not account for the capacty and delay effects of traffc sgnals, stop sgns, and other causes of nterrupted flow that are prevalent n Eureka, a cty whose street network s predomnantly, f not almost entrely, made up of urban surface arterals and local street. The ODME loadng also allows volumes to exceed capacty. The smulaton-based DTA model, on the other hand, s senstve to all of the aforementoned effects and does not permt volumes to exceed capacty. In fact, the loadng resultng from the smulaton-based DTA ought to be dfferent from, and wll be nconsstent wth, that resultng from the ODME, the bass on whch the matrx of trps s estmated. Ths means that the step-wse approach that begns wth statc ODME and s followed by DTA, though t represents the best tool set the state of the practce has to offer, has defcences that should be recognzed up front. That sad, addtonal steps, whch wll be descrbed later, can be taken to rectfy some dscrepances between the two technques. In order for reasonable route choces to be smulated, congested (.e., loaded ) travel tmes on whch the route choces are based must be estmated. Ths s the prmary functon of the smulaton-based DTA method n TransModeler. The smulaton s run to completon for the entre tme perod teratvely, wth the method of successve averages appled to output travel tmes every teraton. The route choces of each run are thus a functon of the travel tmes smulated and averaged over pror runs. In Eureka, a 15-mnute temporal profle n the demand was estmated based on 15-mnute count data. Thus, dynamc, 15-mnute travel tmes were used. Through the smulaton-based DTA, those dynamc travel tmes (and the dynamc route choces) are expected to stablze. The assgnment runs untl t has converged to a target relatve gap measure for UE or untl a maxmum number of teratons s reached. The relatve gap calculaton s smlar to that of the subarea analyss, but a gap s computed both for the entre smulaton perod as well as for each tme nterval τ as follows: Gap τ where: = f τ τ k k I k K I τ τ d tmn, I t d t τ τ mn, τ Gap = Relatve gap n tme nterval τ; I = Set of all O-D pars ; K = Set of paths used by trps travelng between O-D par ; f = Number of trps takng path k (.e. path flow) n tme nterval τ; τ k

8 8 τ t k = Travel tme on path k n tme nterval τ; τ d = Demand departng n tme nterval τ; τ t mn, = Travel tme on shortest path between O-D par n tme nterval τ; Very good are acheved usng the DTA to estmate the route choces of trps generated from the estmated trp tables. Routes observed vsually between OD pars and passng through crtcal lnks all satsfy basc a pror expectatons about the feasble routes drvers take. Unreasonable routes, such as those along corrdors nterrupted by stop sgns every block, are effectvely fltered out of the set of route choces. Furthermore, multple route choce alternatves are frequently used between many OD pars. The ODME and DTA together produce very good results n Eureka gven the complexty of the route choce problem n the cty s grd street network and the mperfectons nherent n ther step-wse applcton. However, as expected, the qualty of the goodness of ft wth the counts, the ultmate target of the calbraton effort, degrades as the smulated (DTA) loadng dverges from the ODME loadng. In Eureka, the root mean square error n the volumes ncreased from under 10% as determned by the ODME loadng to as much as 35% (see Fgures 2 and 3). However, for any gven applcaton, ths wll vary dependng on the scale and complexty of the model. Ths necesstates the fnal step of the methodology, whch s to systematcally refne the trp table based on gaps that are observed n the model between smulated volumes and counts. TRIP TABLE REFINEMENT Refnement of the trp table s the last step n the methodology, and generally requres a number of teratons before a desrable degree of match between the smulated and observed volumes s reached. Usng a crtcal lnk analyss based on the path flows resultng from the smulated route choces, the domnant OD pars producng trps passng lnks where the match s poor are dentfed. The crtcal OD pars n the trp table are thus dentfed by the software and scaled by some approprate factor provded by the user to mprove the match. If the changes are modest, another smulaton-based dynamc assgnment can be avoded. Rather, a sngle smulaton can be run to determne the mprovement n agreement wth the counts. Ths can be repeated as many tmes as s necessary to mprove the trp table and to reduce the error to acceptable calbraton targets. Ths process effectvely mrrors the automated ODME routne of alternatng matrx adjustments and assgnments, but s subjectvely prortzed and drven by the modeler. The manual refnement process has two mportant merts. Frst, the model volumes are a reflecton of the route choces made n the smulaton model, so there are no ssues of consstency. Second, the process nvolves and ndeed requres the nteracton of the user, brngng nto the process human judgement, ntuton, and knowledge of the study area, all thngs the automated statc ODME method lacks. However, the adjustment of the matrx should be done wth thoughtful consderaton of the effects scalng all the OD pars dentfed va the crtcal lnk wll have on other lnks n the network. Judgment must be used n order to leverage both the systematc and objectve nature of the computerzed ODME process and the njecton of local knowledge about the general traffc pattern n the study area that s a vrtue of the manual refnement.

9 9 The refnement of the trp table resulted n a substantal mprovement n all goodness-of-ft measures that were used to compare the smulated lnk volumes wth ground counts. The goodness-of-ft measures that were consdered ncluded the relatve (.e., percent) root mean square error (RMSE), absolute error, relatve error, and the GEH statstc. The RMSE s a poor measure for goodness-of-ft n the wde area smulaton context because t can exaggerate large relatve errors on lnks wth low counts errors that are typcally more acceptable to the practtoner than the same relatve error on a lnk wth hgh counted volumes, artfacts for whch measures lke the GEH statstc are desgned to account. However, the relatve RMSE gves a helpful sngle-value pcture of the goodness-of-ft, at least when used to compare wth other solutons. As shown n Fgure 2, the relatve RMSE deterorated from 9% to 33% from the ODME to the DTA, then mproved to 17% after the trp table refnement. Fgure 3 shows the same for the PM case, where the relatve RMSE ncreased from 10% to 35% after the DTA, then mproved to 18% after the trp table refnement. Fgure 2. Relatve RMSE n the AM perod counts at varous stages of the calbraton process Fgure 3. Relatve RMSE n the PM perod counts at varous stages of the calbraton process

10 APPLICATION OF THE SIMULATION-BASED DTA Morgan, Mayberry 10 The smulaton-based DTA requres no more data preparaton or model development tme than that requred to develop a model for smulaton. The DTA analyss s clearly more meanngful when used wth a dynamc trp table, whch s not the standard n the present state of smulaton practce, though t s the opnon of the authors that t ought to be, and that the mpacts of tme-varyng demand on network performance, both smulated and real, are grossly underestmated or overlooked. To apply the DTA, however, demands modern computng resources, by whch we mean desktop computers wth 4 GB of RAM or more and multple cores, or processors, operatng at clock speeds of 2.00 GHz or more. Computers wth these specfcatons are readly commercally avalable at reasonable prces. Modern computng power s needed because the DTA must smulate the full study perod many tmes over for a gven scenaro or tme perod. Each executon of the smulaton produces new reported travel tmes and turnng movement delays that wll be used to reevaluate route choces pror to the next smulaton. For large models such as the GEA that have many route choces, ths can mean runnng the smulaton dozens of tmes such that all potentally competng route choce alternatves between all OD can be explored, and such that the best paths (.e., those that mnmze travel costs) can be allowed to dstngush themselves from neffcent ones (.e., those nvolvng a lot of turns or sufferng excessve turnng delays). Havng no blueprnt from the research or practce, nor any practcal justfcaton, for choosng a partcular convergence crteron or maxmum number of teratons, many experments were tred. As many as 150 teratons of the smulaton were executed n a sngle applcaton of the DTA. To run many smulatons quckly, TransModeler takes full advantage of all the computer s processors. On a computer wth 8 GB of RAM and a 2.0 GHz Intel Core 7 processor, whch has four cores hyperthreaded, thus makng eght threads avalable for smulatng vehcles, TransModeler s able to smulate the base year PM scenaro n the GEA model more than 40 tmes faster than real tme. To smulate four hours of traffc between 2:00 and 6:00 PM takes a mere 5.7 mnutes on average. Addtonal processng s performed between teratons to average the travel tmes or path flows and to compute the relatve gap for the convergence threshold, but ths means that large numbers of DTA teratons can be executed n well under 24 hours. Larger models wth greater and longerlastng congeston than that found n the GEA wll requre longer runnng tmes, but are entrely wthn the means of today s md-to-hgh-performance desktops. The applcaton of smulaton-based DTA n the GEA model held a number of lessons for practce. Perhaps most nterestngly, because of the stochastc nature of route choce and mcroscopc traffc smulaton, and the nose generated n these models, the UE relatve gap alone s not a suffcent stoppng crteron. The relatve gap alone masks the effects of subtle shfts n route choce that can have mportant mplcatons for calbraton and valdaton locally (.e., at an ntersecton or along a corrdor). For these reasons, the GEA model s allowed to run for many more teratons (.e., on the order of ) beyond the pont at whch the relatve gap can vsbly be seen to level off. As shown n Fgure 4, ths occurs as early as 30 teratons nto the DTA.

11 11 Fgure 4. Relatve RMSE n the PM perod counts at varous stages of the calbraton process Addtonal convergence nformaton, such as the maxmum flow change on any lnk n the network, are beng consdered for future applcatons. SUMMARY Ths paper gves a methodology for estmatng and calbratng a mcrosmulaton model for an entre cty usng a combnaton of travel demand methods, geographc analyss, orgn-destnaton matrx estmaton, and dynamc traffc smulaton. The Eureka case study demonstrates that such an approach can be successful n overcomng some of the more dffcult challenges n the applcaton of mcroscopc traffc smulaton to wde areas, partcularly n the plannng context. Trp table estmaton and calbraton methods n the state of the practce can be sgnfcantly mproved by explotng geographc nformaton and analyss and by usng emergng dynamc traffc assgnment methods.

12 12 REFERENCES Slavn, H., Brandon, J., Rabnowcz, A., and Sundaram, S. Applcaton of accelerated user equlbrum traffc assgnments to regonal plannng models, Presented at the 12 th TRB Natonal Transportaton Plannng Applcatons Conference, May 17-21, 2009, Houston, TX Calper Corporaton (2010) What TransCAD Users Should Know about New Statc Traffc Assgnment Methods, Newton, MA Calper Mappng Software