Forecasting Passenger Congestion in Rail Networks

Transcription

1 Forecasting Passenger Congestion in Rail Networks Craig McPherson an Naomi Langon Sinclair Knight Merz, Melbourne, Australia ABSTRACT Just as traffic congestion constrains the efficient movement of people an goos on the roa system, passenger congestion on the rail network affects the on-time running of trains an operation of stations. In Melbourne, population growth, fare policy changes an fuel price instability have contribute to unpreceente growth in public transport patronage in recent years, leaing to significant crowing issues in parts of the rail network. This paper escribes a new approach to the forecasting of passenger congestion on rail networks using simulation to preict crowing on trains an in stations. The simulation technique moels the choices of iniviual passengers as they select their eparture time, train service, interchange points an walking paths through stations. The position of each passenger is then mappe in time an space to etermine when an where crowing may occur. The moel has been successfully applie in Melbourne to test changes to the train timetable an to estimate when congestion mitigation measures may be require in central city stations. It is hope that this approach will help planners make better use of existing infrastructure before costly capital works are require. 1 INTRODUCTION Just as traffic congestion constrains the efficient movement of people an goos on the roa system, passenger congestion on the rail network affects the on-time running of trains an operation of stations. In Melbourne, train patronage has grown by about 40% over the last three years, leaing to significant crowing issues in parts of the rail network (Heral Sun 2009). To evelop solutions to overcrowing on the rail network, the Victorian Department of Transport (DoT) commissione a stuy in 2007 to evaluate the capacity of Melbourne s inner city stations an train services. The City Loop an Inner Core (CLIC) stuy s objectives inclue the following: etermine how train loas might be affecte by alternative train timetables, incluing ifferent combinations of through an irect services; forecast peestrian throughput at each inner-city station an ientify capacity constraints; forecast the number of passengers transferring between rail services, taking into account the ease of transfer (e.g. cross-platform versus subways/briges); hanle unconstraine or capacity-constraine rolling stock loaings; incorporate peak spreaing, that is, the re-timing of trips in response to crowing; calculate the proportion of staning an sitting passengers; an prouce etaile passenger travel time outputs for economic evaluation purposes. To aress these objectives, it was clear that some form of moel was require to forecast passenger emans an to test the effects of new rail services an infrastructure. However, the stuy managers realise that a traitional four-step strategic transport moelling approach was not going to be sufficiently etaile to answer the questions pose by the stuy. A new, more etaile moel was neee. This paper escribes the new rail network moel that was evelope for the CLIC stuy. Working han-in-han with a strategic moel, the ClicSim moel not only forecasts crowing levels on iniviual trains but also simulates peestrian movements through stations. Unlike other train loaing moels, the new moel uses a behavioural choice component to represent passengers ecision-making processes, an integrates this with an agent-base simulation to moel ynamic congestion levels on trains an in stations. 32 n Australasian Transport Research Forum Page 1

2 Section 2 of this paper reviews some of the common techniques for forecasting train loas an station congestion, an which approaches was aopte by the ClicSim moel. Section 3 escribes the structure of the moel an Sections 4 an 5 emonstrate the practical outcomes an conclusions from the moel s evelopment. 2 PASSENGER CONGESTION FORECASTING TECHNIQUES This section reviews some of the common forecasting techniques use for train loa estimation an peestrian flow moelling. 2.1 Train Loas Averaging Techniques At the simplest level, estimates of train loas can be etermine by taking the total number of passengers using a rail line at a given location an iviing by the number of trains available uring the time perio. Passenger volumes can be etermine irectly from passenger counts (boarings an alightings) or erive from a strategic moel. Usually a location is chosen in the network where train loas are known to be at a maximum (e.g. immeiately outsie major city stations or interchange points). While this technique gives an estimate of average train loas, it oes not take into account the relative attractiveness of express an stopping services, nor oes it allow for variability in train frequencies 1. To better represent these effects, a more sophisticate passenger ecision-making moel, such as the Rooftop Moel, may be more appropriate The Rooftop Moel The Rooftop Moel was originally evelope by British Rail in the 1970s to estimate the relative passenger loas on fast an slow train services (Tyler an Hassar 1973, Shilton 1982). Since then, it has been applie in Australia to preict train loas for the Victorian Regional Fast Rail project (McPherson an Ashley 2004). The Rooftop Moel takes into account each passenger s esire eparture time an allocates passengers to available trains by minimising the train travel time an scheule elay (i.e. the ifference between the passenger s esire an actual times of eparture). By selecting suitable penalty factors for the scheule elay, the moel is able to ifferentiate the loas on express an stopping services more effectively than the simple averaging methos escribe earlier. While the Rooftop Moel is quite effective in corrior stuies (McPherson an Ashley 2004), it is ifficult to apply in complex urban rail networks with frequent trains, branching an looping lines, an passengers interchanging between services. To allow for these complexities, a process that consiers a broaer representation of train choice attributes is neee Choice Moels Passenger train choice moels are a further generalisation of the averaging an Rooftop Moel allocation techniques escribe above. In essence, choice moels etermine a set of feasible train journey alternatives for each passenger an etermine the probability of each 1 Assuming that passengers enter a station at a constant rate uring a particular time perio, one coul assume that boaring volumes will be proportional to the time between successive trains (all other conitions being equal). 32 n Australasian Transport Research Forum Page 2

3 alternative being chosen. Each alternative has a certain utility (or attractiveness) which is typically base on attributes such as travel time, number of interchanges, waiting time an scheule elay. Using common iscrete choice moel formulations such as logit or probit moels (see for example, Train 2003), the proportion of passengers choosing each alternative can be estimate. A choice-base moelling framework provies a flexible way of representing the explanatory variables that govern passenger train choices. One of the main ifficulties in applying these techniques is the nee to enumerate all of the feasible train paths for each passenger. In a large urban rail network such as Melbourne or Syney, the enumeration of alternatives can rapily lea to combinatorial overloa if there are frequent services an multiple interchange points. The ClicSim moel solves this problem by using a culling process to avoi combinatorial overloa. The metho is base on a branch-an-boun approach propose by Frierich, Hofsäß an Wekeck (2001). Section 3 of this paper escribes the approach in more etail. 2.2 Station Capacity To ientify capacity constraints an congestion hot spots in a railway station, planners will often use some form of peestrian moel. Peestrian moels simulate the progression of passengers through platforms, ticket barriers, walkways, stairs an escalators. Harney (2002) provies an overview of various techniques use to moel peestrian flows. Most of these moelling techniques fall into two broa categories: flow propagation moels an simulation moels Flow Propagation Moels Flow propagation moels consier the flow rates of peestrians through each part of the network using relationships between crow ensity an travel times. Densities an flows are calculate in iscrete area units (or blocks ). The moels also take into account average walking spees an constraints such as ticket barriers an escalators. Such moels calculate peestrian levels of service, travel times an the ynamic propagation of queues through the network. One of the popular flow propagation moels use since the 1990s is PEDROUTE (Buckman an Leather 1994). PEDROUTE, evelope by Halcrow in association with Lonon Unergroun an the British Airport Authority, enables planners to conuct rapi capacity assessments of stations an airports. Although not as complex as simulation moels, flow propagation moels are relatively quick to set up an simple to use, an are still favoure for high-level capacity stuies Peestrian Simulation Moels In the last ecae, a new class of peestrian moels has gaine popularity in research an inustry. Agent-base simulation moels simulate the goals, behaviours an movement of iniviual peestrians (or agents), allowing a much more etaile representation of crow behaviour an congestion effects. Many of the popular commercial moelling software applications (e.g. Legion, SimWalk, Paramics an VISSIM 2 ) use this approach. 2 See n Australasian Transport Research Forum Page 3

4 3 THE CLICSIM PASSENGER SIMULATION MODEL 3.1 Overview The ClicSim moel uses a choice-base train allocation technique in conjunction with a custom-built agent-base peestrian moel. This integrate approach provies a great eal of flexibility for planners: the complete journey of each passenger on the rail system is moelle both the train component an station (walking) component; accurate transfer penalties (that take into account the walking time between platforms) are automatically calculate; the peaks in peestrian congestion cause by trains arriving simultaneously at a station can be simulate; crowing on trains will cause some retiming of passenger journeys, simulating the peak spreaing effect; an outputs from the moel are available at the iniviual passenger level, allowing planners to iagnose why certain trains are overloae. 3.2 Inputs an Outputs The main inputs to the moel are: forecast passenger flow patterns expresse in a station-to-station origin-estination (OD) matrix; the train timetable; a specification of the rail network; an a specification of the peestrian networks within each station (i.e. platforms, walkways, escalators an exits). In the Melbourne implementation, the passenger OD matrix was erive from an extensive passenger interview survey. Future passenger flows were then generate by applying growth factors to the observe matrix. Growth factors woul typically be calculate by a strategic transport moel or similar forecasting moels. The rail network an internal station peestrian networks are represente hierarchically within an XML file. The XML format allows moellers a great eal of flexibility in specifying the networks, transparent version control an auiting, an the ability to interface with spreasheets or other moels. An example of the XML coing is shown in Error! Reference source not foun.. The main outputs from the moel are: boarings an alightings by train, station, platform an time of ay; train loas by train, line section an time of ay; numbers of passenger transfers by station, trains, platforms an time of ay; peestrian congestion measures (level of service an area occupancy) by station, location within the station, an time of ay; passenger travel times (comprising walking time, in-vehicle time an transfer time); an animate outputs showing the movement of trains an ynamic peestrian ensities within stations. All outputs can be isaggregate to the level of the iniviual peestrian if require, proviing a powerful way of interrogating the moel. 32 n Australasian Transport Research Forum Page 4

5 <stations> <station noeid="1" name="flagstaff" coe="fgs" transferpenalty="16" > <walknoes> <walknoe noeid="3" name="unpai concourse 1" area="60" traveltime="00:00:09" coorinatex="800" coorinatey="400" /> </walknoes> <exits> <exit noeid="1" name="la Trobe Street Exit" entryproportion="0.17" exitproportion="0.3" area="75" coorinatex="600" coorinatey="200"/> </exits> <platforms> <platform name="1"> <waitingarea noeid="6" boarproportion="0.25" alightproportion="0.25" /> </platform> </platforms> <walklinks> <walklink startnoe="3" ennoe="4" exittype="queue" servicepoints="2" queueservicetime="00:02.4" biirectional="false"/> </walklinks> Figure 1: XML snippets showing the structure of the network input file Train timetable Train choice moule Train paths Transfer costs Station exit proportions Station path moule Peestrian paths Crowing Path generator Passenger OD matrix Passenger time profiles Simulator Accumulator Boarings an alightings Train loas Transfers Peestrian occupancies Passenger travel times Figure 2: Structure of the ClicSim Moel 32 n Australasian Transport Research Forum Page 5

6 3.3 Moel Structure The main components of the moel are shown in Figure 2 an escribe below Train Choice Moule The train choice moule uses a branch-an-boun public transport assignment technique base on the work of Frierich et al (2001). The moule was implemente in the following way for each cell in the passenger origin-estination (OD) matrix: 1) Split eman into time intervals the passenger volume between each pair of stations is apportione to a set of 15-minute intervals uring the peak perio. The proportion of travel allocate to each interval is base on the observe arrival an eparture profiles at each en of the trip. 2) Enumerate all possible train combinations all train combinations between the chosen origin an estination are etermine. These inclue all possible eparture times an interchange combinations. Because the number of combinations woul be prohibitively large for a typical trip (literally tens of thousans of paths), limitations are place on the total number of interchanges mae by each passenger an the stations at which interchanges are permitte. The set of train paths is represente in a connection tree as escribe by Frierich et al (2001). 3) Cull infeasible combinations the connection tree is prune to remove any infeasible combinations. Infeasible combinations inclue those where: - passengers alight, then immeiately reboar, the same train; - for a given origin-estination pair, a train eparts earlier an arrives later than the preferre train 3 (in this case, the train eparting earlier will be culle from the selection set); - interchange times are shorter than the minimum calculate walking time between platforms; or - transfer times are greater than a preefine maximum allowable transfer time (in the Melbourne moel, this was nominally set at one hour to allow for infrequent country trains). 4) Calculate the perceive travel cost for each combination. The perceive cost C (in minutes) is given by: C IVT c transfer t access t egress F crow F time where: - IVT is the in-vehicle time (in minutes); - c transfer is the total transfer cost (see below); - t access is the walking time between the origin station entrance an boaring platform; - t egress is the walking time between the alighting platform an estination station exit; - F crow is a crowing penalty factor (see below); an 3 In practice, a small travel time tolerance is use so that if a train is one or two minutes earlier, it woul still be consiere. This allows for variability in train running times an passengers perceptions of travel times. 32 n Australasian Transport Research Forum Page 6

7 - F time is a eparture/arrival time penalty factor (see below). The total transfer cost is the sum of iniviual transfer costs at each station s in the set of transfer stations S on the passenger s journey. The total transfer cost is given by: c transfer s S a s t s b s where: - t s is the transfer time (in minutes) at station s; - a s is a station-specific waiting time coefficient; an - b s is a station-specific constant travel time penalty. The crowing penalty factor F crow represents the relative attractiveness of seate, staning an crush conitions on a train. The moel calculate the value of F crow from the piecewise linear relationship epicte in Figure 3. Figure 3: Crowing penalty factor calculation In this figure: - F stan is the staning penalty factor; - F crush is the crush loaing penalty factor; an - the train loa is the eparture loa moelle at the boaring point on the passenger s journey. In the Melbourne application, the crush loaing penalty factor was set to an arbitrarily high value (10,000) to prevent passengers from travelling on overloae trains. The time penalty factor F time is calculate as shown in Figure n Australasian Transport Research Forum Page 7

8 Figure 4: Time penalty factor calculation In this figure: - the preferre interval is the current time interval being moelle (see step 2); - F early is the penalty multiplier applie to arrival or eparture times actual before the start of the passenger s preferre time interval preferre start F time F early preferrestart actual actual preferrestart - F late is the penalty multiplier applie to arrival or eparture times actual after the en of the passenger s preferre time interval preferre en F time F late actual preferreen actual preferreen In the Melbourne moel, 15-minute time intervals were aopte. 5) Calculate the probability of using each train path a multinomial logit moel is use to allocate passengers to each train path from each station: p t exp( exp( i S Ct ) C ) i where: - p t is the proportion of passengers allocate to train path t; - S is the set of all feasible train combinations between the origin an estination; - C i is the perceive cost of travel on train path i (see step 4); an - is a moel sensitivity parameter. In the Melbourne moel, the moel sensitivity parameter was etermine through sensitivity tests. The final value of =0.3 gave a realistic istribution of passengers to express an stopping trains. 32 n Australasian Transport Research Forum Page 8

9 The apportionment of passengers is carrie out for each origin-estination pair an each time interval Station Path Moule The station path moule etermines the walking paths of peestrians between platforms, an also between platforms an station exits. Platform-to-platform paths are use to calculate realistic transfer times an penalties between platforms (e.g. a cross-platform transfer will have a lower cost than a transfer using a subway or over-brige). The transfer costs are use by the train choice moule in the etermination of train path probabilities. Platform-to-platform paths are also use by the simulation moule to simulate peestrians transferring between platforms. Platform-to-exit paths are use to etermine the routes that peestrians use for platform access an egress. The simulator uses these paths to etermine the minute-byminute locations of peestrians within each station. The proportions of peestrians using each route are etermine from surveys of passenger volumes at each station exit Path Generator Moule The path generator moule compiles the train paths an peestrian paths into a composite path for each peestrian. A separate path is generate for each passenger trip in the originestination matrix The Simulator The simulator is responsible for tracking the positions of iniviual passengers in the train an peestrian networks. The simulator upates the position of each passenger uring each time step of the simulation (typically at half-secon intervals). The progress of each passenger through the network epens on train arrival an eparture times, average walking spees an queues within stations The Accumulator The accumulator takes the time an space ata output from the simulator an etermines the total numbers of passengers in each part of the network at any given instant. Train loas calculate in the accumulation step are fe back to the train choice moule so that crowing can be taken into account in passengers choice of trains Crowing Increments The simulation moel oes not know which trains are crowe until after it has assigne passengers to trains. Therefore, the moel uses an incremental feeback process whereby it assigns a small number of passengers at a time from the OD matrix an progressively monitors the level of crowing on each train as each passenger path is calculate. The incremental crowing metho generally works well, but will still occasionally allow trains to be fille beyon capacity 4. Train overcrowing may also occur in the extreme case where all trains eparting from a particular station are alreay crowe. In this case, passengers have no choice but to travel on an alreay-crowe train. Moel results showing this type of crowing behaviour usually inicate an insufficient number of trains in the timetable. 4 In the Melbourne moel, this was ue to the size of the increment chosen. A more sophisticate algorithm that selectively assigns passengers in smaller increments coul avoi this problem. 32 n Australasian Transport Research Forum Page 9

10 3.4 Moel Valiation The ClicSim moel was valiate for Melbourne conitions by comparing moelle passenger loas on trains with observe loas from surveys. Figure 5 an Figure 6 show the moelle an observe numbers of passengers on incoming trains at Richmon Station in the morning peak. Richmon Station, locate just outsie the Melbourne CBD, was chosen because eastern suburban trains are usually most heavily loae at this point. Figure 5: Moelle an observe AM peak line volumes at Richmon for Burnley Group lines, Melbourne. Observe ata: 2007 Loa Stanars Survey, Victorian Department of Transport. The Belgrave, Lilyale an Glen Waverley lines have a larger catchment area an higher proportion of express trains than the Alamein an Ringwoo lines, an therefore attract more passengers. Figure 5 shows that the moel allocates passengers to each line to within 10% of the observe total. At the iniviual train loa level (Figure 6), the moel reprouces the general pattern of express an stopping train loas quite well, though the error on iniviual train services is higher than at the total line level. Given that the moel assumes that all trains are running on time (which oes not always happen in practice), this was consiere to be an acceptable result. Figure 7 shows the moelle peestrian ensities uring the busiest half hour at Fliners Street, Melbourne s busiest station. This plot was checke by station operations staff an confirme to accurately represent the locations where congestion is most apparent. 32 n Australasian Transport Research Forum Page 10

11 Figure 6: Moelle an observe train loas at Richmon for Burnley Group lines, Melbourne. Observe ata: 2007 Loa Stanars Survey, Victorian Department of Transport. Figure 7: Peestrian ensities at Fliners Street Station, Melbourne (AM peak busiest half hour, 2008). Warm colours (res an oranges) represent the most congeste locations. 32 n Australasian Transport Research Forum Page 11

12 4 MODEL APPLICATION The ClicSim moel has been use to simulate passenger movements across the entire Melbourne metropolitan rail network. In the AM peak, over 200,000 passengers, 1,800 trains an 200 stations were simulate. The internal station networks were moelle for the ten inner-city stations uner consieration in the CLIC stuy. At other stations, a simplifie set of platforms an walking links were use. Using a half-secon simulation time step, the train choice calculations (i.e. passenger assignment) took about two hours to process an the spatial-temporal simulation of passenger movements about five minutes. Figure 8 shows the user interface of the ClicSim software. The simulation is isplaye as an interactive animation showing the progression of trains an peestrian movements. The animation uses ifferent colours to represent the loaing levels on trains, allowing moellers to reaily ientify the times an locations in the network where train loas become critical. Figure 8: ClicSim user interface showing animation of trains (top pane) an internal station networks for two stations (lower panes). The software prouces a comprehensive set of spreasheet reports that provie moellers with a complete coverage of boarings an alightings, train loas, travel times, transfer behaviour, congestion levels an peestrian flows. The following figures provie examples of the types of outputs that can be prouce by the moel. 32 n Australasian Transport Research Forum Page 12

13 From Platform Forecasting Passenger Congestion in Rail Networks Figure 9: Comparison of train loas with two alternative timetables. The aition of more trains causes a reuction in loas, although these are mostly early in the peak To Platform Figure 10: Volumes of transferring passengers between platforms at Richmon Station. The vertical axis represents the origin platform an the horizontal axis represents the estination platform. Platforms 7 an 8 represent a cross-platform interchange, whereas Platforms 1 to 8 o not. This type of iagram can be use to iagnose inefficient train presentation patterns at multi-platform stations. 32 n Australasian Transport Research Forum Page 13

14 Figure 11: Peestrian congestion plot showing instantaneous levels of service on a walkway at Melbourne Central Station. Walkway levels of service are efine in Fruin (1971). 5 CONCLUSION This paper has escribe a rail network moel that uses an agent-base simulation technique to forecast passenger crowing levels on trains an peestrian movements through stations. The moel uses an integrate approach to moelling passenger paths; the train an walk components are both simulate by the moel, allowing the position of each passenger to be simulate as they move across the rail system. The use of a iscrete choice structure with a branch-an-boun approach enables the moel to hanle complex urban rail networks an interchange arrangements. The moel has been successfully applie to the Melbourne rail network an shown to reprouce observe train loas an station congestion quite well. It provies a powerful tool for iagnosing crowing issues an assessing the effectiveness of new train timetables an rail infrastructure. 32 n Australasian Transport Research Forum Page 14

15 Acknowlegements The authors wish to thank the staff from the Public Transport Division at the Victorian Department of Infrastructure for supporting the evelopment of the moel. References Buckman LT an JA Leather (1994), Moelling station congestion the PEDROUTE way, Traffic Engineering an Control, 35(6), pp Frierich, M, I Hofsäß an S Wekeck (2001), Timetable-base transit assignment using branch & boun, Transportation Research Recor 1752, pp Fruin JJ (1971), Peestrian Planning an Design, Elevator Worl, Mobile, Alabama. Harney, D (2002), Peestrian moelling: current methos an future irections, Roa & Transport Research, 11(4), pp Heral Sun (2009), Public transport patronage on the rise ami commuter fury, Heral Sun, March 6, McPherson, CD an DJ Ashley (2004), Estimating Passenger Deman for Fast Rail Services With the Rooftop Moel, 27th Australasian Transport Research Forum, Aelaie. Shilton, D (1982), Moelling the eman for high spee train services, Journal of the Operational Research Society 33, pp Sinclair Knight Merz (2008), City Loop an Inner Core Patronage Distribution an Station Operations - Moel Verification Report, Final report for Victorian Department of Transport, July Train, KE (2003), Discrete Choice Methos With Simulation, Cambrige University Press. Tyler, J an Hassar, R (1973), Gravity/elasticity moels for the planning of the inter-urban rail passenger business, Proceeings of Seminar G, PTRC Summer Annual Meeting, University of Sussex. 32 n Australasian Transport Research Forum Page 15