Abstract
Assigning passenger flows on a metro network plays an important role in passenger flow analysis that is the foundation of metro operation. Traditional transit assignment models are becoming increasingly complex and inefficient. These models may even not be valid in case of sudden changes in the timetable or disruptions in the metro system. We propose a methodology for assigning passenger flows on a metro network based on automatic fare collection (AFC) data and realized timetable. We find that the routes connecting a given origin and destination (O-D) pair are related to their observed travel times (OTTs) especially their pure travel times (PTTs) abstracted from AFC data combined with the realized timetable. A novel clustering algorithm is used to cluster trips between a given O-D pair based on PTTs/OTTs and complete the assignment. An initial application to categorical O-D pairs on the Shanghai metro system, which is one of the largest systems in the world, shows that the proposed methodology works well. Accompanying the initial application, an interesting approach is also provided for determining the theoretical maximum accuracy of the new assignment model.
1. Introduction
As an efficient transport system, the metro system is now the mainstay of urban passenger transport in many megacities, especially in highly populated areas [1]. Passenger flow is the foundation of making and coordinating operation plans for a metro system, while assigning passenger flows on the metro network plays an important role in analyzing passenger flows. A number of studies [2–4] have developed passenger flow assignment models. However, these models are becoming increasingly complex because of many diverse parameter types. In the case of sudden changes in the timetable or disruptions in the metro system, these models may not be valid.
Different from private cars, a metro system is operated according to the timetable, which is an important constraint for a passenger’s travel. New technologies are widely introduced into metro systems, resulting in improvements in passenger flow assignment. For example, the automatic fare collection (AFC) system has become the main method for collecting metro fares in many cities in the world. This system records the origin and destination stations of a trip and their corresponding timestamps. The transaction data obtained through these AFC systems contain a vast amount of archived information on how passengers use a metro system. Up to date, however, there are limited studies on AFC data or how to assign passenger flows efficiently by combining these data with the timetable.
This paper mainly focuses on how to efficiently model the passenger flow assignment problem for a metro network with AFC data and timetable.
1.1. Timetable and AFC Transaction Data
1.1.1. Timetable Information
The metro timetable contains the set of all train trips with arrival and departure times per station and per train number. Figure 1 is an example of the timetable for a metro line in the Shanghai metro system. Since the metro system is operated based on the timetable, a passenger’s travel time between the origin and destination stations is subject to not only the chosen route but also the timetable. For our study in this paper, the realized timetable is used.
1.1.2. AFC Transaction Data
The assignment addressed in this paper obviously requires AFC transaction data. The ID number of a smart card holder is recorded each time the holder passes the entry or exit gates, and the corresponding transaction record indicates an unlinked trip. These smart card transaction records provide information on ID numbers, the date, departure station, passage time at an entry gate, arrival station, and passage time at an exit gate. The entry and exit times are recorded in the exact number of seconds, based on which observed travel times can be obtained. Example AFC data are shown in Table 1. Our initial analysis (Figure 2) of the observed travel times indicates that the routes connecting a given O-D pair are related to their observed travel times, although there is also travel time uncertainty at the route level.
(a) A single-route trip
(b) A multiroute trip
1.2. Problem Description
The observed travel time is relevant to the passenger travel process. Figure 3 shows a typical travel process of a metro passenger. It consists of entry walking, waiting for train, traveling in-vehicle, transfer walking, waiting for transfer trains if necessary, and exit walking. Correspondingly, the observed travel time (OTT) of a passenger includes entry walking time (ENT), waiting time on platforms (WT), in-vehicle time (IVT), exit walking time (EWT), transfer walking time (IWT), and another waiting time (WT) if there is a transfer. Moreover, we define that CIT is the check-in time at the origin station recorded by AFC data, COT is the check-out time at destination station, BOT is the actual time point that the passenger boards on the train, and GOT is the actual time point that the passenger gets off from the train. Obviously, both BOT and GOT are related to the timetable. Thus, the interval between CIT and BOT is the sum of ENT and WT, and the interval between COT and GOT is EWT. Based on the abovementioned, the pure travel time (PTT), which is relevant to the timetable and an important notion that we defined in this paper, can be calculated from the interval between BOT and GOT.
As mentioned in Section 1.1, OTTs derived from AFC data are relevant to the route choices, and there may be a wide variation of OTTs for a given O-D pair, especially in a large scale network. It can be explained by the random CITs and the resulting random origin WTs. In extreme cases, the origin WT can affect a passenger’s OTT to such a great extent that there is no determinate relationship between the OTT and possible routes. For example, if the OTTs between two routes vary only by 3 minutes while the interval between services is 9 minutes, it is difficult to assign an OTT from AFC data to one of the routes as on average 4.5 minutes of WTs result from the random CITs (Figure 4).
Fortunately, it seems more promising to model the relationship between the possible routes for a given O-D pair and the corresponding PTTs which delete ENTs, EWTs, and origin WTs from OTTs (Figure 5). Therefore, the problem being studied in this paper can be stated as follows: how to abstract these OTTs/PTTs based on AFC data and then complete the passenger flow assignment with them?
The objective of this paper is to propose a methodology to assign passenger flows on a metro network mainly based on travel times (OTTs/PTTs) abstracted from AFC data. To achieve this goal, the following approach is used:(1)We propose a transit assignment model using revealed information including AFC data and realized timetable of metro systems rather than a priori knowledge.(2)We introduce a novel clustering approach to conduct the assignment. It is only based on the distance between data points and can detect nonspherical clusters and automatically the correct number of clusters.(3)We find that PTT is better than OTT when being used for clustering. It can reduce the variation of travel times for O-D pairs to a great extent.(4)We also provide an approach, accompanying the initial application to categorical O-D pairs on the Shanghai metro network, for determining the theoretical maximum accuracy of our proposed assignment model.
The remainder of this paper is organized as follows. In Section 2, relevant literature is reviewed. Section 3 presents the methodology in detail, including the PTTs abstracting and clustering approaches. Section 4 is an initial application of the proposed approach to the Shanghai metro network, demonstrating how to assign passenger flows for categorical O-D pairs and determine the theoretical maximum accuracy of the proposed assignment model. Finally, Section 5 summarizes the findings of the paper and proposes areas for further investigation.
2. Literature Review
Passenger flow is required to make and coordinate operational plans for a metro system. Conventionally, models to solve passenger flow assignment problems can be classified according to whether Wardrop’s principle is followed [5]. One model is the nonequilibrium assignment, and the other is the equilibrium assignment model [6]. Moreover, it is assumed that the process of passengers’ choice has some random characteristics because of imperfect knowledge of travel time, individual differences, measurement errors, and so on [5–7]. Therefore, confronted with today’s metro systems, the result from passengers’ route choices can be described more appropriately by the stochastic user-equilibrium (SUE) with time and space constraints, which is also proved by some simulation experiments [5, 8] and full-scale case tests [4]. Up to date, those models to solve a SUE problem are becoming increasingly complex due to the many diverse parameter types. Thorough reviews were presented in some of the literature [2, 3, 9, 10].
In recent years, automatically collected fare data such as smart card data have been used by transit service providers to analyze passenger demand and system performance. These data have been used for O-D matrices estimation [11, 12], demand analysis [13, 14], travel behavior analysis [15], operational management, public transit planning [16–18], and so on. And according to Pelletier et al. [19], the studies on the use of smart card data can be grouped into three categories: strategic (long-term planning), tactical (service adjustments and network development), and operational (ridership statistics and performance indicators). However, there are fewer studies dealing with AFC data of metro systems. Some important publications include work of Chan in 2007 [20], Kusakabe et al. in 2010 [21], Xu et al. in 2011 [22], Sun and Xu in 2012 [23], Zhou and Xu [24], and Zhu et al. in 2014 [25].
Chan [20] developed two applications based on Oyster card data in the London Underground: one of these estimated an O-D flow matrix, while the other constructed rail service reliability metrics. This is the first attempt at measuring service delivery quality using elapsed travel time. Based on Asakura’s efforts, Kusakabe et al. [21] develop a methodology for estimating which train would be boarded by each smart card holder using long-term transaction data. Their approach is based on the assumption that smart card data that could not be identified to the possible train choices would be assigned with equal probability. Xu et al. [22] try to estimate metro passengers’ route choice behaviors using smart card data and proposes a new model for passenger flow assignment based on an AFC system environment. However, the requirement of detailed calibration restricts further development of the proposed model. Sun and Xu [23] use smart card data to analyze travel time reliability and passenger route choice behavior in metro networks. However, the problem of calibrating the vast number of parameters in behavior functions such as arrival/departure distributions still exists. Zhou and Xu [24] develop a passenger flow assignment model and designed an algorithm for urban rail transit based on entry and exit time constraints from AFC data. However, because their algorithm checks each path using train diagram to determine the path validity, it needs more storage space and floating-point calculation than other normal algorithms, leading to a higher space complexity and time complexity. Zhu et al. [25] present a method for calibrating metro assignment models using AFC data. Their calibration approach uses a genetic algorithm-based framework with nonparametric statistical techniques, and consequently calculation cost is still a problem for the transit assignment model to be calibrated.
To the best of our knowledge, these existing studies on transit assignment models with AFC data are either too simple or too computationally costly and should be improved. This paper focuses on how to precisely and efficiently assign the real passenger flows on a metro network using AFC data and timetable.
3. Methodology
3.1. OTT and PTT Abstracting Approach
Since CITs and COTs are recorded in the AFC data, it is convenient to obtain OTTs. This section focuses on abstracting PTTs from AFC data. We first give some basic definitions on the train timetable. The train timetable illustrates the relationship between space and time of train operation. The main information it contains are trains’ arrival and departure times at each station. Denote the set of metro lines as and as the set of stations in line ; means station in line . Then the arrival time and departure time of th train at station can be described as . Therefore, its train path is defined as the collection . For each line, each station, and each train, the train timetable can be represented by . Moreover, an AFC data record can be described as OD(ID, CIT, COT, enter_st_no, exit_st_no), where enter_st_no and represent enter station ID and exit station ID, respectively.
3.1.1. Determination of BOT
Let be its CIT from AFC data, and let be its enter_st_no. Searching every train which runs through the station in order, the train that stops at station at time can be determined by locating such that
The passengers can choose th train to arrive at their destination or transfer station. The search process is illustrated in Figure 6. Therefore, the ATT can be set as
3.1.2. Determination of GOT
Similarly, let be its COT from AFC data, and let be its exit_st_no, as shown in Figure 7. Then search every train which stops at the station in reverse order. Passengers getting off the th train can be obtained from the condition
Passengers will check out from station once they get off trains, it is simpler for containing no waiting time comparing passengers’ entry. Thus, GOT is equal to . It should be noted that there is a minimum interval threshold between CIT and BOT as well as GOT and COT, because walking or waiting will also need time.
Therefore, smart card data can be trimmed as follows: From OD(ID, CIT, COT, enter_st_no, exit_st_no) to OD(ID, BOT, GOT, enter_st_no, exit_st_no)
3.1.3. Determination of PTT
After the AFC data record is trimmed from OD(ID, CIT, COT, enter_st_no, exit_st_no) to OD(ID, BOT, GOT, enter_st_no, exit_st_no), PTT can be expressed as
3.2. A Novel Clustering Approach
Since the AFC transaction data can be used to estimate passengers’ route choices, it is possible to use it for passenger flow assignment. To achieve this, we have applied cluster analysis techniques. Unlike the existing assignment model, the cluster analysis technique in this paper clusters trips between a given O-D pair based on PTTs/OTTs derived from the AFC data. It then assumes that similar PTTs/OTTs are linked to the same route. Cluster centers for a given O-D pair are considered the expected travel times (ETTs) of the feasible routes, and PTT/OTT is assigned to the corresponding cluster center.
Several clustering strategies have been proposed, including the -means method [26], the -medoids method [27], distribution-based algorithms [28], density-based algorithms [29], and the mean-shift method. However, a novel clustering approach was recently proposed by Laio and Rodriguez [30]. We have used this method for the following reasons.(1)The -means and -medoids methods cannot detect nonspherical clusters, because a data point is always assigned to the nearest center. The OTTs for a given O-D pair consist of nonspherical clusters.(2)Distribution-based algorithms attempt to reproduce the observed data points using a mix of predefined probability distribution functions. The accuracy of such methods depends on how well the trial probability represents the data.(3)Density-based algorithms choose an appropriate threshold which may be nontrivial, though clusters with an arbitrary shape can be easily detected by approaches based on the local density of data points.(4)The mean-shift method only works for data defined by a set of coordinates and is computationally costly, although it does allow for nonspherical clusters and does not require a nontrivial threshold.(5)The clustering approach proposed by Laio and Rodriguez [30] is superior, because it is only based on the distance between data points, it can detect nonspherical clusters, and it automatically determines the appropriate number of clusters.
The adopted clustering approach is based on the idea that cluster centers are characterized by a higher density than their neighbors and by a relatively large distance from points with higher densities. For each data point , we compute two quantities: its local density and its distance from points of higher density. Both these quantities depend only on the distances between data points, which are assumed to satisfy the triangular inequality. The local density of data point is defined aswhere if and otherwise. is a cutoff distance, and is the number of points that are closer than to point . The algorithm is only sensitive to the relative magnitudes of values for different points. This implies that, for large data sets, the results of the analysis are robust with respect to the choice of .
is determined by computing the minimum distance between point and any other point with higher density. That is,
For the point with the highest density, we conventionally take . Note that is much larger than the typical nearest neighbor distance only for points that are local or global maxima in the density. Thus, cluster centers are recognized as points for which the value of is anomalously large (as shown in Figure 8).
(a) Point distribution. Data points are ranked in order of decreasing density
(b) Decision graph for the data in (a). Different colors correspond to different clusters
After the cluster centers have been found, each remaining point is assigned to the same cluster as its nearest neighbor of higher density. The cluster assignment is performed in a single step, in contrast with other clustering algorithms where an objective function is optimized iteratively.
4. Initial Application and Analysis to Categorical O-D Pairs on the Shanghai Metro Network
4.1. Passenger Flow Assignment for the O-D Pairs with Single Route
Although the proposed approach aims to assign passenger flows to the routes between a given O-D pair, those O-D pairs with single route should be identified first of all. There are two types of O-D pairs with a single route:(1)O-D pairs with a unique physical route on the network.(2)O-D pairs that have only one feasible route when we consider the travel cost threshold, although there is more than one physical route on the network.
In both of the abovementioned cases, all the passengers for the O-D pair are assigned to only one route. And the procedure is similar to All or Nothing Assignment Model.
Taking the Shanghai metro as an example, the feasible route set for a given O-D pair is generated using a two-step route generation method. First, the th-shortest path algorithm is applied and a universal route set is generated based on the physical topology of the metro network. Second, the universal set is filtered by judging the rationality of alternative routes based on the difference in the travel costs of the alternative and shortest route. This narrows the feasible route set.
The initial statistics of the Shanghai metro network demonstrates that there is a large percentage of O-D pairs with a single route (35.98% in terms of O-D pairs and 60.15% in terms of trips).
4.2. Passenger Flow Assignment for the O-D Pairs with Multiroutes
4.2.1. Estimating Passenger Route Choices with the Clustering Technique and Determinate PTTs
Except those O-D pairs with single route, there are a large number of O-D pairs with multiple routes, for which passenger route choices can be estimated using the determinate PTTs and proposed clustering technique.
Consider an example O-D pair with two feasible routes on the Shanghai metro network. The distribution of OTTs and the corresponding probability density function are shown in Figure 9(a). Using the abstracting approach proposed in Section 3.1, OTTs can be further fined to PTTs shown in Figure 9(b). We computed two quantities for each point of PTTs in this example data: its local density and its distance from points with higher densities (), with the corresponding decision graph being shown in Figure 9(c). We can see that two points (blue and red) have large values and a sizeable density. These two points correspond to cluster centers, which represent the expected PTTs of two routes between the O-D pair. After determining the two centers, each point is assigned to a cluster, which is used to calculate route choice probabilities for the O-D pair (Figure 9(d)).
The test O-D pairs discussed in this section are those with determinate PTTs for which the passenger route choices can be estimated accurately to a great extent and consequently a precise passenger flow assignment result can be obtained. Taking the Shanghai metro network as an example, our initial calculations and analyses for all of the O-D pairs on the network showed that there are 42611 O-D pairs (35.39% in terms of O-D pairs, 22.22% in terms of trips) falling into this category of O-D pairs.
However, there are also other categories of O-D pairs that may not be suitable for the estimation of passenger route choices using PTTs. In these cases, a passenger’s travel behavior is so complex that it is difficult to determine the passenger’s PTT. For example, if both the upstream and downstream are feasible directions for the origin station of an O-D pair to the destination (Figure 10(a)), or the origin station of an O-D pair is a transfer station (Figure 10(b)), we cannot judge which train a passenger boards on in reality and consequently the corresponding PTT is not determinate. The following section discusses how to estimate these categories of O-D pairs.
(a) Both of upstream and downstream are feasible directions for the origin to the destination
(b) The origin of an O-D pair is a transfer station
4.2.2. Estimating Passenger Route Choices with the Clustering Technique and Indeterminate PTTs
For the category of O-D pairs with multiple routes and indeterminate PTTs, we use the clustering technique and OTTs to estimate passenger route choices based on which the passenger flow assignment is completed. Our initial calculations and analyses for all the O-D pairs on the Shanghai network show that there are 34472 O-D pairs (28.63% in terms of O-D pairs, 17.63% in terms of trips) falling into these categories of O-D pairs.
Of course, the result from this assignment for the abovementioned O-D pairs may not be accurate due to a possible wide range variation of OTTs. However, among these categories of O-D pairs, there is still a kind of O-D pairs for which the corresponding assignment result can be precise to a great extent. It is because the expected travel times of routes for a given O-D pair falling into this kind of O-D pairs are obviously different from each other, and consequently the corresponding OTTs can be clustered into the routes accurately. For the Shanghai metro network, the corresponding percentage is 5.06% in terms of O-D pairs, as well as 7.14% in terms of trips.
Moreover, there are some O-D pairs for which we cannot give accurate route choice estimations. Such O-D pairs include those with similar expected travel times for its different connecting routes and with small flows from several to several dozen passengers. In the case of these O-D pairs, the route choices of passengers are stochastic to a great extent. For the Shanghai metro network, the corresponding percentage is 19.46% in terms of O-D pairs, as well as 3.87% in terms of trips.
5. Discussions and Conclusions
5.1. Extended Discussions to the Proposed Approach
From the analysis in the previous sections, the proposed approach in this paper can efficiently estimate metro passenger route choices using a novel clustering technique and processed AFC data (PTTs/OTTs) and consequently provide appropriate passenger flow assignments on a metro network. Furthermore, the approach implies the potential of measuring its minimum and maximum accuracy; the minimum and maximum accuracy can be approached in practice by classifying all the O-D pairs into several categories. Taking the Shanghai metro network as an example, as shown in Table 2, we can measure the minimum and maximum accuracy of the approach as follows.(1)O-D pairs with single route: the passenger flow assignment using the proposed approach is accurate for this category of O-D pairs because there is only one feasible route between a given O-D pair and a passenger’s route choice is unique. For the Shanghai metro network, the corresponding percentage is 35.98% in terms of O-D pairs, as well as 60.15% in terms of trips. There is a large percentage of O-D pairs for which the estimated route choices are always correct, regardless of the assignment model. This is an interesting characteristic of a metro network compared with an urban road network.(2)O-D pairs with multiple routes and determinate PTTs: the passenger flow assignment using our approach is accurate for this category of O-D pairs because the variation of travel times for a route is narrowed to a smaller range by using PTTs instead of OTTs. For the Shanghai metro network, the corresponding percentage is 35.39% in terms of O-D pairs, as well as 22.22% in terms of trips.(3)O-D pairs with indeterminate PTTs but obviously different expected travel times for routes: for the route choices between an O-D pair in this category whose route expected travel times are obviously different from each other, the proposed approach can also give an accurate assignment. For the Shanghai metro network, the corresponding percentage is 5.06% in terms of O-D pairs, as well as 7.14% in terms of trips.(4)Some special O-D pairs: we cannot give accurate passenger flow assignment for them. In the case of these O-D pairs, the route choices of passengers are stochastic to a great extent. For the Shanghai metro network, the corresponding percentage is 19.46% in terms of O-D pairs, as well as 3.87% in terms of trips.(5)Others: except the above categories of O-D pairs, the remainder is those O-D pairs for which the proposed approach cannot guarantee giving an accurate assignment but may have the potential of approaching the actual route choices in theory.
In summary, based on the above discussions for different categories of O-D pairs, the minimum and maximum accuracy of the proposed approach with the clustering technique and AFC data can be measured in practice. Taking the Shanghai metro network as an example, the proposed approach is accurate for 94.10% of trips, cannot be accurate for 5.28% of trips, and may be accurate for 0.62% of trips. And the total accuracy range is 75.43%~79.54% in terms of O-D pairs with 89.51%~96.13% in terms of trips.
5.2. Concluding Remarks
A metro system is operated based on the timetable. Developments in the application of ADC systems such as AFC systems have made the collection of detailed passenger trip data in a metro network possible. In this paper, we aim to propose an efficient approach to assign passenger flows on a metro network combing AFC data and timetable. The advantages of the proposed approach include the following:(1)A posteriori transit assignment model, which uses information including AFC data and timetable of metro systems rather than a priori knowledge, was proposed.(2)A novel clustering approach was introduced to conduct the assignment. It is only based on the distance between data points and can detect nonspherical clusters and automatically the correct number of clusters.(3)It was found that PTT is better than OTT when being used for clustering, because it can reduce the variation of travel times for O-D pairs to a great extent.(4)Accompanying the initial application to categorical O-D pairs on the Shanghai metro network, an interesting approach was also provided for determining the theoretical maximum accuracy of our proposed assignment model.
However, some additional issues still need to be addressed. For example, several unusual phenomena during peak periods such as “failing to board on” should be accounted for in the assignment process, and the computational efficiency of the approach should be further improved considering the massive amounts of AFC data and timetable data. All the above mentioned is the prospective work in the future.
Overall, this study provides a promising approach that can efficiently assign passenger flows on a metro network not only in the common case but also in the case of sudden changes in the timetable or disruptions in the metro system.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
The study is financially supported by the National Natural Science Foundation of China (71271153) and Program for Young Excellent Talents in Tongji University (2014KJ015).