Transportation Research Part C: Feng Shi, Shuo Zhao, Zhao Zhou, Pu Wang, Michael G.H. Bell

Transportation Research Part C 74 (2017) 51–80
Contents lists available at ScienceDirect
Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc
Optimizing train operational plan in an urban rail corridor

based on the maximum headway function
Feng Shi a,⇑, Shuo Zhao a, Zhao Zhou b, Pu Wang a, Michael G.H. Bell c
a
School of Traffic and Transportation Engineering, Central South University, Changsha, Hunan 410075, China
b
Hisense Transtech Company Limited, Hisense Group, Qingdao, Shandong 266071, China
c
Institute of Transport and Logistics Studies, University of Sydney, NSW 2006, Australia
a r t i c l e i n f o a b s t r a c t
Article history: The train operational plan (TOP) plays a crucial role in the efficient and effective operation
Received 24 November 2015 of an urban rail system. We optimize the train operational plan in a special network layout,
Received in revised form 9 November 2016 an urban rail corridor with one terminal yard, by decomposing it into two sub-problems,
Accepted 9 November 2016
i.e., the train departure profile optimization and the rolling stock circulation optimization.
Available online 18 November 2016
The first sub-problem synthetically optimizes frequency setting, timetabling and the roll-
ing stock circulation at the terminal without a yard. The maximum headway function is
Keywords:
generated to ensure the service of the train operational plan without considering travel
Urban rail corridor
Train operational plan
demand, then we present a model to minimize the number of train trips, and design a
Maximum headway function heuristic algorithm to maximize the train headway. On the basis of a given timetable,
Time-varying demand the rolling stock circulation optimization only involves the terminal with a yard. We pro-
Optimization algorithm pose a model to minimize the number of trains and yard–station runs, and an algorithm to
find the optimal assignment of train-trip pair connections is designed. The computational
complexities of the two algorithms are both linear. Finally, a real case study shows that the
train operational plan developed by our approach enables a better match of train headway
and travel demand, and reduces the operational cost while satisfying the requirement of
the level of service.
Ó 2016 Elsevier Ltd. All rights reserved.
1. Introduction
With the rapid growth of urban populations and automobile ownership, most large cities are suffering from severe traffic
congestion. The improvement of urban rail transit has become increasingly important as a means of mitigating traffic prob-
lems in these dense population areas. Urban rail transit offers large passenger capacity and high operational speed as well as
lower energy consumption and reduced pollution. These safe, fast, comfortable, energy-efficient urban rail systems are cru-
cial for the economic, environmental, and social success of metropolitan regions (Wang et al., 2015). In recent years, oper-
ational planning and improved organization of urban rail transit systems have attracted wider attention from the field of
transportation optimization (Ceder, 1984, 1986, 2002, 2007; Guihaire and Hao, 2008).
Public transit planning generally involves five basic steps: (1) designing routes; (2) frequency setting; (3) timetabling; (4)
rolling stock circulation or vehicle assignment; and (5) crew scheduling (Guihaire and Hao, 2008; Cadarso and Marín, 2014).
Designing a model that encompasses all five planning steps at once is rather cumbersome and complex. In most cases, the
⇑ Corresponding author.
E-mail address: shifeng@csu.edu.cn (F. Shi).
http://dx.doi.org/10.1016/j.trc.2016.11.007
0968-090X/Ó 2016 Elsevier Ltd. All rights reserved.
52 F. Shi et al. / Transportation Research Part C 74 (2017) 51–80
model for each of the five planning steps is developed sequentially. The output of one step provides the input for the next
step (Ceder, 2002). In particular, the train operational plan (TOP) normally consists of three of the procedures described
above: frequency setting, timetabling, and rolling stock circulation. Each plays a crucial role in the efficient and effective
operation of an urban rail transit system.
Frequency setting and timetabling are usually studied jointly. In general, a schedule that incorporates constant headways
can reduce the total waiting time if the pattern of passenger arrival times at stations follows particular probability distribu-
tions, such as uniform distribution or Poisson distribution (Niu and Zhou, 2013). In keeping with this approach, some
researchers have designed multi-phase timetables to implement regular vehicle departure intervals for each time period.
Ceder (2009) proposed a modeling framework and developed corresponding algorithms for planning vehicle departure
times. The proposed timetable used even headways or even average loads during the same time period as well as smooth
transitions between time periods. In contrast, a number of researchers have focused on periodic timetable design.
Liebchen (2008) designed a train timetable using a periodic event-scheduling approach based on a well-established graph
model. The author generated the periodic schedules for the Berlin subway system. Odijk (1996) proposed a constraint gen-
eration algorithm to develop a model with periodic time window constraints. This model used periodic timetables to plan
arrival and departure times.
Given that travel demand has large variations (Vuchic, 2005), schedules with fixed headways usually cannot satisfy the
fluctuating demands well. Improved results are obtained when transit operators understand the dynamics of passenger
demands and respond by designing demand-sensitive timetables to make better use of limited service capacities (Ceder,
2007; Niu and Zhou, 2013). Accordingly, a variety of timetabling methods have been proposed that take into consideration
time-varying demand and non-periodic headways. Ceder (2001) developed a scheduling model to replace constant headway
by making transit vehicles even-loaded. Jamili and Pourseyed-Aghaee (2015) developed a formulation of finding optimum
stop-skip patterns and used a robust approach to obtain headway distributions which fit traffic in different weekdays and
holidays. Assis and Milani (2004) presented a method for computing optimal train schedules in metro lines using a
linear-programming-based model with predictive control formulation. This train traffic model was comprised of dynamic
equations that described the evolution of train headways and train passenger loads. This model also considered the time
variation of passenger demand and all relevant safety and operational constraints. Albrecht (2009) pointed out that the eco-
nomical and attractive operation of suburban railways can be realized only by recognizing the need for flexible headways,
adaptation of the network, and capacity of the lines. Albrecht presented a two-level approach to planning and timetabling for
a suburban railway. Niu and Zhou (2013) optimized train schedules for an urban rail transit line by including consideration
of time-dependent, origin-destination passenger demands under heavily congested conditions with limited train fleet avail-
ability. Furthermore, Niu et al. (2015) investigated a method to minimize the total passenger waiting time at stations by
computing and adjusting train timetables for a rail corridor with given time-varying O-D demand and skip-stop patterns.
Sun et al. (2014) presented three models for designing demand-sensitive timetables, in which the influence of train capacity
was considered. The authors showed the advantages of a dynamic timetable built with capacity constraints. Li and Lo (2014)
considered time-varying passenger travel demand and proposed a dynamic optimization framework for train scheduling.
Jiang et al. (2016) presented a model based on time-driven microscopic simulation to evaluate train timetable with urban
rail transit big data.
In the area of railroad operation, train timetables serve as an essential data input to locomotive and crew scheduling,
directly impacting the utilization of scarce resources such as engines, cars, and crews (Zhou and Zhong, 2007). In general,
train scheduling must consider a number of operational and safety requirements as well as limited resources. Recently,
researchers have investigated synchronized timetables, conflict prevention, and resource utilization. Wu et al. (2015) and
Xiong et al. (2015) investigated the issues for optimizing synchronized timetables in urban rail transit systems to minimize
passengers’ waiting time, delay cost, and transfer cost. Wong et al. (2008) focused on coordinated timetables that enabled
smooth interchanges with minimal delays and waiting times for passengers. Corman et al. (2010) proposed a bi-objective
model to minimize train delays and missed connections in order to provide a set of feasible non-dominated schedules.
Carey and Crawford (2007) designed a series of heuristics to find and resolve train conflicts under various operational con-
straints and objectives. Dündar and Sahin (2013) developed a decision support tool for re-scheduling problems designed to
resolve inter-train conflicts. Zhou and Zhong (2005, 2007) proposed a generalized resource-constrained project scheduling
formulation that considered segment and station headway capacities as limited resources. Their model addressed the train
scheduling problem for planning applications based on the combined objective of minimizing both expected waiting times
and total travel times of trains.
In order to utilize existing resources most effectively and avoid conflicts between trains, rolling stock circulation must be
considered; otherwise, the timetable cannot be applied in practice. Some researchers have considered an integrated opti-
mization of timetable and rolling stock circulation. Guihaire and Hao (2010) proposed an integrated timetable and vehicle
assignment optimization model for a bus transit system in which they optimized the vehicle scheduling plan by using the
method proposed by Freling et al. (2001). They embedded these procedures in an iterated local search algorithm capable of
finding feasible solutions. Petersen et al. (2012) proposed an integer programming problem to optimize the timetable and
deal with the vehicle scheduling problem simultaneously by applying a large neighborhood search metaheuristic. Ibarra-
Rojas et al. (2014) studied the trade-off between the level of service and operating costs, including the problems of timeta-
bling and vehicle scheduling. They presented two integer linear programming models for the two problems and then com-
bined them in a bi-objective integrated model. It is notable, however, that the consideration of demand in the above studies
F. Shi et al. / Transportation Research Part C 74 (2017) 51–80 53
was pre-defined or had a uniform rate flow. To include time-varying demand, Niu and Zhou (2013) studied the transit-
scheduling problem assuming heavily congested conditions with a limited train-unit fleet in which the rolling stock circu-
lation was simple, i.e., each train arriving at the terminal station turned around after the pre-defined dwell time.
Given the difficulty to incorporate time-varying demand into optimization models, recent studies generally use an urban
rail corridor (URC) as the system layout (Assis and Milani, 2004; Niu and Zhou, 2013; Sun et al., 2014). In this paper, we also
focus on a URC and optimize TOP with time-varying demand. The studied URC has one terminal yard, a configuration that is
widely used in many cities. This layout is common for newly operated rail transit systems and rail systems in small cities.
Specific examples include Line 5 of the Munich Metro, Line 7 of the Beijing Metro, and Line 2 of the Changsha Metro. In this
kind of network layout, the finite parking capacity at the terminal station without a yard would result in difficulty for parking
additional arriving trains or supplying additional departing trains when the train headway changes frequently.
Under a predetermined level of service, we aim to minimize the major components of operation costs, i.e., the number of
train trips, the number of trains used, and the number of runs between the yard and terminal station (yard-station runs). The
current paper extends the research on TOP in the following aspects:
(1) We optimize the number of train trips by maximizing train headways while satisfying travel demand, and develop a
precise formulation (the maximum headway function) to describe the change of train headway. Using this function,
we can optimize the timetable without considering travel demand, and the level of service of TOP is also satisfied.
(2) In the special network layout, the limited parking capacity at the terminal station without a yard is considered, due to
its significance to influence the quality of train schedules in a busy corridor (Qi et al., 2016). We develop an integrated
optimization for frequency setting, timetabling and the rolling stock circulation at the terminal without a yard.
(3) We design an optimization algorithm with linear computational complexity to solve the optimal solution of the rolling
stock circulation at the terminal with the yard, which minimizes the number of trains used and the number of yard-
station runs.
This paper is organized as follows. The overall problem statement and underlying assumptions for the optimization of the
TOP in a URC with one terminal yard are given in Section 2. By analyzing the travel demand of railway stations and the con-
straints of the level of service, we define maximum headway function (MHF) and an approximate generation method for
MHF in Section 3. We propose a model for optimizing the train departure profile (integrated optimization for frequency set-
ting, timetabling and the rolling stock circulation at the terminal without a yard) in Section 4 and a model for optimizing the
rolling stock circulation (only for the rolling stock circulation at the terminal with a yard) in Section 6. For the model in Sec-
tion 4, we design a heuristic algorithm with linear computational complexity in Section 5; for the model in Sections 6 and 7
shows the optimization algorithm has linear computational complexity. A numerical example is shown in Section 8 to val-
idate the proposed models and evaluate the efficiency of the designed algorithms. Conclusions follow in Section 9.
In this paper, a train trip represents a transportation task that a train runs from one terminal to the other terminal of the
line, with determined stopping stations, as well as departure, arrival and stopping times. The train headway is the interval
between the departure times of two consecutive train trips (with the same direction) from a station.
2. Problem statement
An optimal TOP must be designed to guarantee the level of service and to minimize the operational cost for the operating
company.
For convenience, travel demand is described by section demand rather than traditional origin–destination (O–D) demand.
A section can be viewed as a connection from one station to its next station. The section demand equals to the demand of
passengers passing the starting station plus the demand of passengers departing from the starting station to the section. It
can be viewed as the cumulative passenger flow of all O–D demands in the section along their chosen paths. We can obtain
the section demand by traditional investigation and statistics (Ceder, 1984, 1986), or by transit equilibrium assignment com-
bined with the analysis of passengers’ path choice behaviors (Poon et al., 2004). The section demand varies with time, and it
can be viewed as a continuous function of time.
In this section, we first introduce the concept of the level of service in Section 2.1, and then we describe the TOP optimiza-
tion for a special layout of an urban rail line, i.e., a URC with one terminal yard in Section 2.2.
In the following, Tables 1 and 2 show the input parameters and variables in this paper.
2.1. The level of service
The level of service, an important feature of TOP optimization, evaluates the quality of service. In this paper, we use three
constraints as the requirements of the level of service:
(1) The constraint for the departure times of the first train trip and the last train trip
Table 1
Input parameters.
Name Definition
smin ; smax Minimum/maximum train headway
C Transport capacity of a train
a Ideal occupancy rate
sA ; sB Two terminal stations in an URC
D; U Down/up direction
s1 ; s2 ; . . . ; sN Stations along direction D
N Number of stations
S Set of stations
sðsi ; sj Þ The travel time between two stations si ; sj 2 S
th
Q ðk;kþ1Þ ðtÞ; Q ðkþ1;kÞ ðtÞ The section demand of the k section at time t in the down direction and up direction, 1 6 k 6 N 1
½T 1 ; T 2 Planning operational period of one day
L ; tL
tD U
The latest acceptable departure times of the first train trips in the two directions
E ; tE
tD U
The earliest acceptable departure times of the last train trips in the two directions
pA The single yard
C A ; CB Parking capacities of two terminal stations
s A ; sB Turning-around times of two terminal stations
s Minimum time unit of the timetable
Table 2
Variables.
Name Definition
F D ðxÞ; F U ðyÞ Maximum headway functions in the two directions
x; y Time variables in the two directions, x; y 2 ½T 1 ; T 2
xi The departure time of train trip i (the ith train trip) in the down direction at the terminal sA
yj The departure time of train trip j (the jth train trip) in the up direction at the terminal sB
X; Y Vector of departure times of the train trips in the two directions with n elements, i.e., ðx1 ; x2 ; . . . ; xn Þ; ðy1 ; y2 ; . . . ; yn Þ
th
xsi k The departure time of the i down-direction train trip from station sk 2 S n fsB g
rsk ðtÞ The detained passenger flow at station sk after the train trip which starts at t from the terminal leaves station sk
ssi k The critical time when the waiting demand at station sk 2 S n fsB g for the i þ 1th down-direction train trip reaches the ideal train
load
tsi k The proper departure time of the i þ 1th down-direction train trip from the station si
th
W hl The waiting demands for the i þ h train trip in the deduction process, l ¼ 1; 2; . . .
t1 The departure time of the first train trip from terminal sA
th
ftk1 ; t k2 ; . . . ; t kNðkÞ g The k sequence of data points of time
th
NðkÞ Number of data points of time in the k sequence
K Number of sequences
th
tki The ith data point of time in the k sequence
Nb Number of final data points of time
n Number of train trips in the down direction or up direction
Z; Z 1 ; Z 2 Variables of the optimization objects
i ; j Minimum number of train trips in the two directions solved by the main loop of Algorithm 1
I; J Set of the down- or up-direction train trips
M Rolling stock circulation at the terminal sA
Mb Rolling stock circulation at the terminal sA or the yard pA
jMj; j Mj b Numbers of elements in set M and M b
Mþ ðjÞ; M ðiÞ Equivalent expressions of M
M; M Two maximum cardinality assignments of J to I
I0 ; J 0 Sets of train trips related to M; M
r Number of connections in the maximum cardinality assignment
th
jnrþk ; ik The k element of J0 and I0 , 1 6 k 6 r
uðtÞ Number of train trips which end at terminal sA no later than t and the corresponding M þ – 0
v ðtÞ Number of train trips which start from terminal sA no later than t and the corresponding M – 0
wðtÞ Number of trains parking at terminal sA at time t
w The simplification of wðtÞ
Me Maximum cardinality assignment of the set fjnrþk jMþ ðjnrþk Þ ¼ 0; 1 6 k 6 rg to the set fih jM ðih Þ ¼ 0; 1 6 h 6 rg
The first train trip in the down/up direction should start no later than the predetermined latest acceptable departure time
to ensure that all passengers waiting for the first train trip can get on board. Similarly, the last train trip in the down/up
direction should start no earlier than the predetermined earliest acceptable departure time.
(2) The constraint for train headway

The train headway between two consecutive train trips should be set between the minimum train headway smin and the
maximum train headway smax . Parameter smin is determined by the capability of the operation control system and safety
issues. It reflects the maximum service frequency that can be provided by the transit system in a given time period. Param-
eter smax is the maximum time accepted by passengers.
(3) The constraint for train load
We denote C as the transport capacity of a train, and að0 < a < 1Þ as the ideal occupancy rate. When the train load
exceeds C, passengers who fail to board the overloaded train generate detained passenger flow. The waiting demand at a
station is defined as the number of passengers waiting for next train trip at the station, which includes the detained passen-
ger flow left behind the last train trip and the cumulative section demand before the next train trip leaves the station. We
classify the whole operation period into off-peak hours and peak hours as follows:
i. Off-peak hours
The waiting demand at every station can be satisfied without causing train congestion, i.e., the train load does not exceed
aC.
ii. Peak hours
The waiting demand at some intermediate stations cannot be satisfied without causing train congestion, i.e., the train
load is larger than aC, even reaches C.
2.2. TOP optimization in a URC with one terminal yard
We denote sA and sB as the two terminal stations in an URC, D as the down direction (from sA to sB ) and U as the up direc-
tion (from sB to sA ). The stations along direction D are denoted as a sequence of s1 ; s2 ; . . . ; sN , and s1 ; sN are exactly sA ; sB . The
set of stations is denoted by S. The travel time between two stations si ; sj 2 S, including running time and stopping time, is
denoted by sðsi ; sj Þ. The sections along direction D and U are denoted by ðk; k þ 1Þ and ðk þ 1; kÞ; k ¼ 1; 2; . . . ; N 1. The sec-
th th
tion demand of the k section in the down direction and the section demand of the k section in the up direction are denoted
by Q ðk;kþ1Þ ðtÞ; Q ðkþ1;kÞ ðtÞ; k ¼ 1; 2; . . . ; N 1; t 2 ½T 1 ; T 2 , where ½T 1 ; T 2 is the planning operational period of one day, T 1 is the
beginning time and T 2 is the ending time. We denote t DL ; t UL 2 ½T 1 ; T 2 as the latest acceptable departure times of the first train
trips in the two directions, and t DE ; t UE 2 ½T 1 ; T 2 as the earliest acceptable departure times of the last train trips in the two
directions.
For TOP optimization in a URC with one terminal yard, the following assumptions are adopted:
(1) The URC has a single yard pA , which is adjacent to terminal sA , and its parking capacity is large enough for trains.
(2) The parking capacities of the two terminal stations are finite. We denote C A ; C B as the parking capacity of terminal
sA ; sB respectively (parking capacity is the number of parking lines, and each parking line can park only one train),
and they satisfy C A ; C B P 1. When terminal sA is fully occupied, some trains arriving at station sA can drive into yard
pA . When the parking capacity at terminal sB is reached, some trains have to leave the terminal by starting a new train
trip and thus release space for new trains’ entry.
(3) Down-direction train trips move from sA to sB , while up-direction train trips move from sB to sA .
(4) Trains can turn around at a terminal to start a new train trip in the opposite direction. The time taken by this process
should be no less than the turning-around time. We denote s A ; sB as the turning-around time at the terminal sA ; sB ,
which should satisfy sB 6 smin since trains at sB can depart again only if they have turned around.
(5) All train trips stop and take on/drop off passengers at every intermediate station in the URC.
(6) All trains are provided by the yard pA to station sA , and drive into the yard after finishing one day’s operation.
th
(7) All trains at sB follow the first-come-first-serve principle, that is, the train which executes the i down-direction train
th
trip would connect the i up-direction train trip at sB .
Assumptions (1) and (2) are illustrated in Fig. 1. When a train departs from station sA , it will get to station sB , then turn
back to station sA or stay at station sB . When a train gets to station sA , it can turn back to station sB , stay at station sA or move
to yard pA .
Given the time-varying demand and the layout of an urban rail line with one terminal yard, the optimization of TOP com-
prises two sub-problems, i.e., the train departure profile optimization and the rolling stock circulation optimization.
Considering time-varying demand, the departure times and frequencies should fit the fluctuation of demand. During peak
hours, the departure frequencies should be high, and a number of trains would arrive at or depart from the terminal station,
which requires that the terminal station should have the relevant ability to park or supply trains. However, at the terminal
without a yard, the finite parking capacity (generally 2 trains) may not satisfy the requirements for parking or supplying
trains, resulting in the imbalance between the departure and arrival of train trips in the two directions. Hence, in order
Drive into the yard

UP
Turn around
Turn around
Drive out of the yard

DOWN
Fig. 1. Operation process of an urban rail corridor with one terminal yard.
to ensure a balanced circulation of trains at the terminal without a yard, the departure times of train trips in the two direc-
tions and the rolling stock circulation at the terminal without a yard must be optimized simultaneously.
In addition, since all train trips stop at every intermediate station, only the departure times of train trips in the down/up
direction need to be determined. This is because the arrival time and departure time of a train at every intermediate station
can be simply determined by the travel time.
With time-varying demand and the assumptions described above, the train departure profile optimization satisfies the
constraints of the level of service, synthetically optimizes frequency setting, timetabling and the rolling stock circulation
at the terminal without a yard, and minimizes the number of train trips.
The rolling stock circulation optimization at the terminal with a yard is based on a given timetable and the above assump-
tions. Generally, it means the process that a train departs from a yard, completes a series of train trips, and drives back to the
yard after finishing one day’s operation. When a train finishes a train trip, it arrives at the terminal, and executes another
train trip. Therefore, rolling stock circulation can be regarded as the connection of train-trip pairs, which means the connec-
tion relationship between an ending train trip and a starting train trip. In the following, we use the connection of train-trip
pairs to describe the framework of rolling stock circulation. Rolling stock circulation involves two optimization objectives:
the first is to minimize the number of trains, and the second is to minimize the number of yard–station runs.
For the three optimization objectives, the number of train trips is the main objective which can minimize the running cost
and potentially reduce the number of trains in operation. The number of trains represents the fixed cost in the operation, and
more trains would need more energy and crew. The number of yard-station runs is an additional cost, which is related to
train routes at terminal stations and labor cost, so it is the least important. In the optimization process, the plans are com-
pared considering the different importance of the three objectives.
The optimization process of the TOP in an urban rail corridor with one terminal yard is illustrated in Table 3.
Table 3
The optimization process of the TOP in an urban rail corridor with one terminal yard.
Optimization stage Construction of MHFs Train departure profile optimization Rolling stock circulation optimization
Modeling objectives Train headways Number of train trips Number of trains and yard-station runs
Input parameters Q ðk;kþ1Þ ðtÞ; Q ðkþ1;kÞ ðtÞ F D ðxÞ; F U ðyÞ n; X; Y
smin ; smax ; s; sðsi ; sj Þ L ; t L ; t E ; tE
tD U D U
C A ; s A ; sðsA ; sB Þ
C; a; N; ½T 1 ; T 2 C B ; sB ; sðsA ; sB Þ
Output variables F D ðxÞ; F U ðyÞ n; X; Y b
M; M
3. Train headway and the maximum headway function
In the train departure profile optimization, we want to maximize the train headway under the constraints of the level of
service. Considering the time-varying demand and the level of service, we construct a maximum headway function (which is
called MHF in the following sections). If the train headway does not exceed MHF, the service of the TOP for time-varying
demand can reach the predetermined level of service; with the upper bound constraint of MHF, the maximized train head-
way can further heuristically minimize the number of train trips.
We denote MHF in the down direction and the up direction by F D ðxÞ; F U ðyÞ; x; y 2 ½T 1 ; T 2 . The function F D ðxÞ can be
described as follows: if a train trip in the down direction starts from the original terminal at time x, the maximum train head-
way for the next down-direction train trip satisfying the level of service is F D ðxÞ. The function F U ðyÞ is similarly defined for
up-direction train trips.
In this section, we first describe how to set the train headway with the constraints of the level of service and travel
demand. Then we calculate the discrete deduction formulas for maximizing train headway. Finally, an approximate con-
struction of the MHF is proposed.
3.1. Train headway setting
The number of train trips in the operation is the major part of operational costs. More train trips result in more energy
consumption and higher labor cost. Minimization of the number of train trips is the main objective of the TOP optimization.
Obviously, in an operational period, if we set a larger train headway, the number of train trips will be smaller.
If we decrease the train headway, the section demand cumulated during the train headway decreases, which further
results in fewer waiting demands and a better chance to guarantee that the train load does not exceed the ideal volume
aC. However, sometimes even when the train headway is set to smin , the waiting demand still exceeds aC, and the corre-
sponding period is viewed as peak hours. The parameters, smax and a, can be modified to ensure the predefined level of ser-
vice. When the waiting demand at one station reaches aC, a train trip should start to satisfy the constraint for train load;
however, if the waiting demand at every intermediate station does not reach aC when the cumulative time reaches smax ,
a train trip should start in order to satisfy the constraint for train headway. In the following, we show a method for maxi-
mizing the train headway while satisfying the constraints of the level of service.
(1) During off-peak hours
We increase the train headway until the waiting demand at one station is equal to aC or the train headway is equal to
smax . At this point, the train headway reaches its maximum value and the level of service is guaranteed. Obviously, all sta-
tions have no detained passenger flow in this case.
(2) During peak hours
We set smin as the train headway and make sure that the train load at any station does not exceed C. Detained passenger
flow may generate at some stations in this case.
3.2. Deduction formulas for maximizing the train headway
If the ith down-direction train trip starts from terminal sA at time xi , then the departure time of this train trip at the station
sk 2 S n fsB g will be xi k ¼ xi þ sðsA ; sk Þ, and the departure time of the train trip can be replaced by xsi 1 . We denote the detained
s
passenger flow at the station sk by rsk ðxi Þ. In the following, we determine the deduction formulas for the MHF F D ðxi Þ and the
corresponding detained passenger flow rsk ðxiþ1 Þ in the cases of off-peak hours and peak hours.
(1) During off-peak hours
R xsi k þsmin
In this case, if the train headway is set to smin , the inequality s Q ðk;kþ1Þ ðtÞdt þ rsk ðxi Þ 6 aC is satisfied for any station
xi k
R xi þsmin
sk
sk 2 S n fsB g. In this inequality, s Q ðk;kþ1Þ ðtÞdt is the cumulative section demand of the section ðk; k þ 1Þ.
xi k
For any station sk 2 S n fsB g, if there exists a time ssi k P xsi k þ smin and the equation
Z s sk
i
s
Q ðk;kþ1Þ ðtÞdt þ rsk ðxi Þ ¼ aC ð1Þ
xi k
is satisfied, then we consider setting si k xi k as the new train headway and set t i k ¼ si k . If there is no such time satisfying
s s s s
ssi k P xsi k þ smin or Eq. (1) is not satisfied, then we consider setting smax as the new train headway and set tsi k ¼ 1. Conse-
quently, we obtain the deduction formulas of the MHF and the corresponding detained passenger flow as follows:
F D ðxi Þ ¼ minfsmax ; minft i k xi k jsk 2 S n fsB ggg

s s
ð2Þ
rsk ðxiþ1 Þ ¼ 0; sk 2 S n fsB g ð3Þ

(2) During peak hours
R xsi k þsmin
In this case, there exists at least one station sk 2 S n fsB g satisfying s Q ðk;kþ1Þ ðtÞdt þ rsk ðxi Þ > aC. We set smin as the
xi k
train headway. The deduction formulas of the MHF and the corresponding detained passenger flow are expressed as follows:
F D ðxi Þ ¼ smin ð4Þ

( Z s )
xi k þsmin
rsk ðxiþ1 Þ ¼ max 0; s
Q ðk;kþ1Þ ðtÞdt þ rsk ðxi Þ C ; sk 2 S n fsB g ð5Þ
xi k
Station
Waiting Demand
0
Time
Station
Waiting Demand
0
Time
Off-peak hours Peak hours Peak hours
Fig. 2. Illustration of a small case of the deduction of maximizing train headway.
We propose a small example, in which the urban rail line has 3 stations, to demonstrate how to maximize train headway
(Fig. 2). Assuming that the departure time of a down-direction train trip i is determined, i.e., xi ¼ xsi 1 , and
rs1 ðxi Þ ¼ 0; rs2 ðxi Þ ¼ 0, the waiting demands cumulated during a certain period for the down-direction train trip i þ h at
R xs1 þsmin
s1 ; s2 can be denoted as W hl (for example, W 11 ¼ si1 Q ð1;2Þ ðtÞdt þ rs1 ðxi Þ).
xi
Next, we determine the departure time of the down-direction train trip i þ 1. Because W 11 < aC; W 13 < aC represents the
train trip is in off-peak hours, we can solve tsi 1 ; tsi 2 when the waiting demands at the stations reaching aC, and obtain
F D ðxi Þ; rs1 ðxiþ1 Þ ¼ 0 and rs2 ðxiþ1 Þ ¼ 0 using formulas (1)–(3). The departure times of the down-direction train trip i þ 1 from
s1 ; s2 are consequently determined: xsiþ1
1
¼ xsi 1 þ F D ðxi Þ; xsiþ1
2
¼ xsi 2 þ F D ðxi Þ. For the down-direction train trip i þ 2, because
W 23 > C represents that the trip is in peak hours, then we can obtain F D ðxiþ1 Þ; rs1 ðxiþ2 Þ ¼ 0 and rs2 ðxiþ2 Þ ¼ W 23 C using for-
mulas (4) and (5). Values of xsiþ2
1
; xsiþ2
2
can be determined correspondingly. For the down-direction train trip i þ 3, because
aC < W 31 ; W 32 < C represents that the trip is in peak hours, then we can obtain F D ðxiþ2 Þ; rs1 ðxiþ3 Þ ¼ 0 and rs2 ðxiþ3 Þ ¼ 0 using
formulas (4) and (5). Values of xsiþ3

1
; xsiþ3
2
can be determined correspondingly. The departure times of the subsequent train trips
and the train headways between them can be determined according to the approach described above.
Generally, we select a departure time for the first train trip x1 ¼ t 1 and use formulas (1)–(5) to calculate F D ðx1 Þ and rsk ðx2 Þ.
R t þsðs ;s Þ
The initial condition can be determined as follows: if for any sk 2 S n fsB g, it satisfies T 11 A k Q ðk;kþ1Þ ðtÞdt 6 aC, then rsk ðt 1 Þ is
calculated using formula (3); otherwise, rsk ðt1 Þ is calculated using formula (5).
3.3. Approximate construction of the maximum headway function
According to the proposed deduction formulas, if we set the departure time t 1 to T 1 , we can obtain a discrete form of MHF,
i.e. F D ðxÞ; x 2 ft1 ; t2 ; . . . ; t N g, where t iþ1 ¼ ti þ F D ðti Þ and the inequality t N 6 T 2 < tN þ F D ðtN Þ is satisfied.
For convenience of mathematical modeling, we adjust the discrete form F D ðxÞ; x 2 ft1 ; t2 ; . . . ; t N g to the continuous form of
D
F ðxÞ; x 2 ½T 1 ; T 2 through the methods of curve fitting and interpolation. However, one problem we meet is that the set of
data points of time in the sequence of ft 1 ; t 2 ; . . . ; t N g is sparse. Thus, we increase the density of data points of time and gen-
th
erate several similar sequences with the same deduction method as follows (the k sequence is denoted by ft k1 ; t k2 ; . . . ; tkNðkÞ g,
th
and NðkÞ is the number of data points in the k sequence).
As shown in Fig. 3, for k ¼ 1; 2; . . . ; K, we set the first train trip’s departure tim e tk1 ¼ T 1 þ ðk 1Þs (s is equal to the min-
imum time unit of the timetable, e.g. s ¼ 0:5 min), and rsi ðt k1 Þ can be determined by the process in Section 3.2. We set rsi ðtk1 Þ
as the initial value and calculate t k2 ; t k3 ; . . . ; tkNðkÞ successively. Then we obtain K sequences of F D ðxÞ; x 2 ft k1 ; t k2 ; . . . ; tkNðkÞ g, where
t kiþ1 ¼ tki þ F D ðt ki Þ and tkNðkÞ 6 T 2 < t kNðkÞ þ F D ðt kNðkÞ Þ are satisfied. Merging the obtained K sequences and deleting overlapped
data points of time, we can obtain a new sequence with enough data points of time, which satisfy
b is the number of data points.
t 11 < t 21 < < tK1 < t12 ¼ t11 þ F D ðt 11 Þ. We denote this new sequence by ft 1 ; t 2 ; . . . ; t b g, where N
N
Sequence Number
Fig. 3. Illustration of construction of sequences of data points.
After we increase the number of data points, the interval between each two neighboring data points is no smaller than
s ¼ 0:5 min, so the total number of time data points is no larger than 19 60 0:5 ¼ 2280 (the operation period is generally
from 5:00 to 24:00). In addition, if the intervals between the following data points exceed 0.5 min, we can selectively add
some time data points, and ensure N b is no larger than 3000 (3000 is selected as the limit). Parameter K can be adaptively
modified according to the density of data points. The change of the granularity of data points would affect the final results,
that is, less granularity would result in higher precision.
Finally, we construct the continuous form of MHF using the Lagrange linear interpolation method.
x ti D x ti1 D
F D ðxÞ ¼ F ðt i1 Þ þ F ðt i Þ; x 2 ½t i1 ; t i ; i ¼ 2; 3; . . . ; t b ð6Þ
ti1 t i ti ti1 N
4. Train departure profile optimization model
The train departure profile optimization model needs to synthetically optimize frequency setting, timetabling and the
rolling stock circulation at the terminal without a yard. In the optimization of timetabling, the departure times are restrained
by the latest/earliest acceptable departure times, and the train headways are restrained by the minimum train headway smin
as well as the MHFs, i.e., F D ðxÞ and F U ðyÞ. With the upper bound constraint of MHF, the timetable for time-varying demand
can satisfy the predetermined level of service and the maximized train headway can further heuristically minimize the num-
ber of train trips. To optimize the rolling stock circulation at the terminal sB , the number of parking trains cannot exceed the
parking capacity of the terminal without a yard. In addition, up-direction train trips cannot start before the down-direction
th
train trips connect with them, so there should be enough time between the departure time of the i up-direction train trip
th
and the arrival time of the i down-direction train trip to ensure effective connections between the train trips in two
directions.
The train departure profile optimization model can be described by the number of train trips, and the departure times of
down-direction and up-direction train trips. According to the requirements of the level of service, the MHFs and assumptions
(1)–(7), the following constraints should be satisfied:
The constraint for the latest acceptable departure time of the first train trip
The departure times of the first train trips in the two directions should not be later than the latest acceptable departure
times.
x1 6 tDL ð7Þ
y1 6 t UL ð8Þ
The constraint for the earliest acceptable departure time of the last train trip
The departure times of the last train trips in the two directions should not be earlier than the earliest acceptable depar-
ture times.
xn P tDE ð9Þ
yn P t UE ð10Þ
The constraint for the minimal train headway
The time interval between two consecutive starting train trips in the same direction should be no smaller than the min-
imal train headway.
xiþ1 xi P smin ; i ¼ 1; 2; . . . ; n 1 ð11Þ
yjþ1 yj P smin ; j ¼ 1; 2; . . . ; n 1 ð12Þ
The constraint for the MHF

The time interval between two consecutive starting train trips in the same direction should be no larger than the value
determined by the MHF.
xiþ1 xi 6 F D ðxi Þ; i ¼ 1; 2; . . . ; n 1 ð13Þ
yjþ1 yj 6 F U ðyj Þ; j ¼ 1; 2; . . . ; n 1 ð14Þ
The constraint for turning-around time at the terminal sB without a yard

th th
According to assumption (7), the j up-direction train trip is connected by the j down-direction train trip at the terminal
sB . Thus, the departure time yj of the up-direction train trip j should be no earlier than the arrival time xj þ sðsA ; sB Þ of the
down-direction train trip j plus the turning-around time sB .
yj P xj þ sðsA ; sB Þ þ sB ; j ¼ 1; 2; . . . ; n ð15Þ
The constraint for the parking capacity at the terminal sB without a yard
Considering the situation that the down-direction train trips i C B ; i C B þ 1; . . . ; i 1 all park at terminal sB , it is equiv-
alent to the inequality xi1 þ sðsA ; sB Þ < yiC B . This means that after the down-direction train trip i 1 ends at sB , the up-
direction train trip j ¼ i C B has not started from sB , thus the trains executing the down-direction train trips
i C B ; i C B þ 1; . . . ; i 1 are parking at terminal sB , and terminal sB reaches capacity saturation. Before the time when
the down-direction train trip i ends at sB , the departure time yiC B of the up-direction train trip j ¼ i C B should be no later
than the arrival time xi þ sðsA ; sB Þ of the down-direction train trip i. In addition, when xi1 þ sðsA ; sB Þ P yiC B , the formula
xi þ sðsA ; sB Þ > xi1 þ sðsA ; sB Þ P yiC B is always satisfied.
xi þ sðsA ; sB Þ P yiCB ; i ¼ C B þ 1; C B þ 2; . . . ; n ð16Þ

We can determine the rolling stock circulation at the terminal sB based on the first-come-first-serve principle, as long as
constraints (15) and (16) are satisfied.
For simplicity, we do not give the constraint for all trains driving into the yard after ending one day’s operation.
Since the constraints of time-varying demand have been guaranteed by constraints (13) and (14), to minimize the num-
ber of train trips, we propose the mathematical model (M1) for optimizing the train departure profile as follows:
min Z ¼ 2n
s:t: constraints ð7Þ—ð16Þ
5. A heuristic algorithm for train departure profile optimization
In the train departure profile optimization model (M1), if we only solve the departure time in a single direction, we can
easily obtain the solution for each train trip one by one, either the X ¼ ðx1 ; x2 ; . . . ; xn Þ of down-direction train trips or the
Y ¼ ðy1 ; y2 ; . . . ; yn Þ of up-direction train trips; constraints (11)–(14) can also be easily met. Yet, constraints (15) and (16)
require X and Y to be coordinated, and the solution should be considered systematically. Additionally, the first variables
x1 ; y1 only refer to the constraints for the latest acceptable departure time of the first train trip, i.e. constraints (7)–(8);
the last variables xn ; yn only refer to the constraints for the earliest acceptable departure time of the last train trip, i.e. con-
straints (9)–(10). In order to minimize the number of train trips and maximize the train headway, we try to heuristically
maximize the departure time of every train trip xi ; yj in model (M1).
Combining constraints (7)–(8), we first solve x1 ; y1 . Based on the initial values x1 ; y1 ; i ¼ 2; j ¼ 2, we iteratively solve the
departure time vectors X; Y.
In every iteration, we selectively determine xi ; yj for the new train trips based on solved x1 ; x2 ; . . . ; xi1 ; y1 ; y2 ; . . . ; yj1 while
satisfying constraints (11)–(16). The word ‘‘selectively” means that sometimes we determine only xi or yj , while sometimes xi
and yj are determined simultaneously. After that, the indexes i; j of the order of train trips correspondingly add one.
The core content in each iteration is the coordination of X and Y, i.e., the order determination of xi ; yj . Note that the critical
factors in the decision-making process are the constraints (15) and (16) for turning-around time and parking capacity at ter-
minal sB . Therefore, we characterize the determining order according to the time relations of relevant trains at the terminal
sB .
When the values of x1 ; x2 ; . . . ; xi1 ; y1 ; y2 ; . . . ; yj1 are determined, it implies that: the down-direction train trip i 1 has
ended at terminal sB , the up-direction train trip j 1 has started from terminal sB , and the number of parking trains is equal
to ði 1Þ ðj 1Þ ¼ i j, which satisfies the constraint 0 6 i j 6 C B . Thus we discuss the determining order of xi ; yj accord-
ing to the following cases.
(1) i j ¼ 0
There is no parking train for a new up-direction train trip. Hence, the down-direction train trip i needs to be assigned
priority, i.e., we determine the value of xi . If yj ¼ xi þ sðsA ; sB Þ þ sB in constraint (15), we determine the values of xi ; yj
simultaneously.
(2) i j ¼ C B
The parking capacity of terminal sB is reached and no arriving train can enter the station. Hence, the up-direction train trip
j needs to be assigned priority, i.e., we determine the value of yj . If xi þ sðsA ; sB Þ ¼ yj in constraint (16), we determine the val-
ues of xi ; yj simultaneously.
(3) 0 < i j < C B
In this case, we compare the arrival time xi þ sðsA ; sB Þ of the down-direction train trip i and the departure time yj of the up-
direction train trip j; and select the train trip with the minimum value as the current prior train trip.
When xi ; yj both satisfy constraints (9)–(10), we stop the iteration process and obtain the minimal value n for the number
of train trips. Finally, we assign the trains at terminal sB to return to sA as quickly as possible, and obtain all values of the
departure time vectors X; Y.
The algorithm is described as follows, where xi and y j represent the latest departure times of down-direction train trip i
and up-direction train trip j to be determined.
Algorithm 1: Departure time adjustment algorithm.

Input the latest acceptable departure times of the first train trips t D L ; t L , the earliest acceptable departure times of the
U
last train trips tE ; tE , the total travel time sðsA ; sB Þ, the minimum train headway smin , the turning-around time
D U
sB ð6 smin Þ, the MHFs F D ðxÞ; F U ðyÞ, and the parking capacity C B ðP 1Þ of terminal sB without a yard;
Output the departure time vectors X; Y and the number of train trips 2n.
Begin
minft D L ; t L sðsA ; sB Þ s g;
U B
x1
y1 x1 þ sðsA ; sB Þ þ sB ;
i 2; j 2;
while xi1 < tD E or yj1 < t E do
U

xi D
xi1 þ F ðxi1 Þ;
j
y yj1 þ F U ðyj1 Þ;
if i j ¼ 0 and y j 6 xi þ sðsA ; sB Þ þ sB then
fxi yj sðsA ; sB Þ sB ; yj
j ; i
y i þ 1; j j þ 1g;
if i j ¼ 0 and y j > xi þ sðsA ; sB Þ þ sB then
fxi
xi ; i i þ 1g;
if 0 < i j < C B and y j < xi þ sðsA ; sB Þ then
fyj j ; j
y j þ 1g;
if 0 < i j < C B and y j > xi þ sðsA ; sB Þ then
fxi
xi ; i i þ 1g;
if 0 < i j < C B and y j ¼ xi þ sðsA ; sB Þ then
fxi
xi ; yj j ; i
y i þ 1; j j þ 1g;
if i j ¼ C B and y j < xi þ sðsA ; sB Þ then
fyj j ; j
y j þ 1g;
if i j ¼ C B and y j P xi þ sðsA ; sB Þ then
fxi
xi ; yj xi þ sðsA ; sB Þ; i
i þ 1; j j þ 1g;
End while

for k ¼ i 1; i 2; . . . ; 2 do if xk1 < t D E 6 xk then i k, break;
for k ¼ j 1; j 2; . . . ; 2 do if yk1 < tE 6 yk then j
U
k, break;
n maxfi ; j g;
for j ¼ j þ 1; j þ 2; . . . ; n do yj maxftU E ; yj1 þ smin ; xj þ sðsA ; sB Þ þ s g;
B
End
We can obtain the rolling stock circulation at terminal sB by Algorithm 1, that is, for j ¼ 1; 2; . . . ; n, the down-direction train
trip j is connected with the up-direction train trip j.
Proposition 1. The solutions of Algorithm 1 are the feasible solutions for the optimization model of train departure profile (M1),
and the computational complexity of Algorithm 1 is OðnÞ.
The proof is not shown here due to the limitation of space. In order to explain our algorithm, we propose a small numer-
ical example in Appendix A.
6. Rolling stock circulation optimization model
The other part of the TOP is the rolling stock circulation at the terminal sA with yard pA . Before optimizing the rolling stock
circulation, the schedule of each train trip is determined by Algorithm 1, and the problem in this section is how to connect
the train-trip pairs at terminal sA .
We use the concept of assignment to describe the rolling stock circulation, and denote I; J as the set of the down-direction
train trips and up-direction train trips. An assignment of J to I can be viewed as a set M J I, in which
I ¼ f1; 2; . . . ; ng; J ¼ f1; 2; . . . ; ng: If the connection ðj; iÞ 2 M, then it satisfies:
0 0 00 00 0 00 0 00 0 00 0 00
(1) if ðj ; i Þ; ðj ; i Þ 2 M; j ; j 2 J; i ; i 2 I, then j – j , i – i ;
(2) yj þ sðsA ; sB Þ þ s A 6 xi .
We denote jMj as the number of connections in the set M, i.e., jMj ¼ jfðj; iÞ 2 Mgj.
The rolling stock circulation at the terminal sA with yard pA can be described by assignments M b and M, in which
b b
M M J I. M represents the rolling stock circulation at the terminal sA or the yard pA ; M represents the rolling stock cir-
culation at the terminal sA .
If ðj; iÞ 2 M, then the train that finishes the up-direction train trip j will execute the down-direction train trip i at terminal
b n M, then the train that finishes the up-direction train trip j, drives into the yard p , and executes the down-
sA . If ðj; iÞ 2 M A
b the train which executes the down-direction train
direction train trip i at the yard p . If for any i 2 I, no j 2 J makes ðj; iÞ 2 M,
A
b the train that finishes the up-direction train trip j will
trip i drives from the yard pA . If for any j 2 J, no i 2 I makes ðj; iÞ 2 M,
drive into the yard pA .
Given the departure time X; Y solved by the mathematical model (M1) (see Section 4), the constraints for the train-trip
pair connection and parking capacity are considered for the optimization of the rolling stock circulation at the terminal with
a yard.
The constraint for the train-trip pair connection at the terminal with a yard
Up-direction train trips can be connected to down-direction train trips at terminal sA or yard pA . If the up-direction train
trip j is connected to the down-direction train trip i, the arrival time yj þ sðsA ; sB Þ of the up-direction train trip j plus the
turning-around time s A should be no later than the departure time xi of the down-direction train trip i.
yj þ sðsA ; sB Þ þ s A 6 xi ; b
ðj; iÞ 2 M ð17Þ
According to the definition of assignment, constraint (17) is also a condition that a connection should satisfy.
The constraint for the parking capacity at the terminal with a yard
At any time t, the number of trains parking at terminal sA cannot exceed the parking capacity C A . In fact, we only need to
guarantee that the constraint is satisfied at time t when each up-direction train trip ends at sA . The arrival times of all up-
direction train trips at sA can be expressed as t ¼ y1 þ sðsA ; sB Þ; y2 þ sðsA ; sB Þ; . . . ; yn þ sðsA ; sB Þ. The situation where at time t
the up-direction train trip j parks at terminal sA indicates that there exists a down-direction train trip i which makes
ðj; iÞ 2 M and yj þ sðsA ; sB Þ 6 t 6 xi . The number of trains parking at time t can be denoted as
jfðj; iÞjyj þ sðsA ; sB Þ 6 t 6 xi ; ðj; iÞ 2 Mgj. Hence, the constraint for the parking capacity at the terminal with a yard can be
expressed as follows:
jfðj; iÞjyj þ sðsA ; sB Þ 6 t 6 xi ; ðj; iÞ 2 Mgj 6 C A ; t ¼ y1 þ sðsA ; sB Þ; y2 þ sðsA ; sB Þ; . . . ; yn þ sðsA ; sB Þ ð18Þ
The sequential optimization objectives of rolling stock circulation are the number of trains and the number of yard–sta-
tion runs, which are described as follows.
The first objective: minimization of the number of trains, Z 1 .
The number of trains is equal to the number of total down-direction train trips n minus the number of connected down-
direction train trips. A ‘‘connected down-direction train trip” means that a down-direction train trip is executed by a train
b is equal to the
that has finished an up-direction train trip, either at the terminal or at the yard. It is easy to validate that j Mj
number of reused up-direction train trips. Thus
b
Z 1 ¼ n j Mj ð19Þ
The second objective: minimization of the number of yard–station runs, Z 2 .
The number of yard–station runs is equal to the number of total train trips 2n minus the number of total train trips in the
assignment M, i.e. 2jMj.
Z 2 ¼ 2ðn jMjÞ ð20Þ
In summary, we construct a multi-objective optimization model (M2) of the rolling stock circulation at the terminal with
a yard as follows:
b
min Z 1 ¼ n j Mj
min Z 2 ¼ 2ðn jMjÞ
s:t: constraints ð17Þ; ð18Þ
7. An algorithm for rolling stock circulation optimization
For simplicity, we introduce a function M þ ðjÞ; j 2 J to explain an assignment M of J to I: for any j 2 J, if there exists i 2 I
which makes ðj; iÞ 2 M, we set M þ ðjÞ ¼ i; otherwise, we set M þ ðjÞ ¼ 0. Similarly, we introduce a function M ðiÞ; i 2 I to explain
an assignment M of J to I: for any i 2 I, if there exists j 2 J which makes ðj; iÞ 2 M, we set M ðiÞ ¼ j; otherwise, we set
M ðiÞ ¼ 0. The functions M þ ; M are both equivalent expressions of M. If any expression is changed, the others will be chan-
ged accordingly. For instance, if M ¼ £, then Mþ ðjÞ ¼ 0; M ðiÞ ¼ 0; j; i ¼ 1; 2; . . . ; n.
Minimizing the first objective of model (M2) has the same effect with maximizing j Mj. b Hence, Mb is the maximum cardi-
nality assignment (a maximum cardinality assignment is an assignment with the maximum number of elements). The max-
imum cardinality assignment can be obtained by Hungarian Method, but we design another algorithm to obtain the
maximum cardinality assignment, minimize the second objective and improve the efficiency of the algorithm.
First, we solve two maximum cardinality assignments of J to I, i.e., M and M.
(1) Solve the assignment M
We set M þ ð0Þ ¼ 0, and do the following process in order of j ¼ 1; 2; . . . ; n. If

fi 2 Iji > maxfM ðj Þj0 6 j < jg; yj þ sðsA ; sB Þ þ s A 6 xi g – £, then we set M þ ðjÞ ¼ minfi 2 Iji > maxfM þ ðj Þj0 6 j < jg; yj þ
þ 0 0 0 0
sðsA ; sB Þ þ s A 6 xi g; otherwise, we set Mþ ðjÞ ¼ 0. This solving method is named the ordering method, and the corresponding
assignment is named the ordering assignment.
(2) Solve the assignment M
We set M ðn þ 1Þ ¼ 1, and do the following process in order of i ¼ n; n 1; . . . ; 1. If

fj 2 Jjj < minfM ði Þji < i 6 n þ 1g; yj þ sðsA ; sB Þ þ s 6 xi g – £,
0 0 0 0
A
then we set M ðiÞ ¼ maxfj 2 Jjj < minfM ði Þji < i 6
n þ 1g; yj þ sðsA ; sB Þ þ s A 6 xi g; otherwise, we set M ðiÞ ¼ 0. This solving method is named the reverse ordering method,
and the corresponding assignment is named the reverse ordering assignment.
Obviously, the ordering assignment and the reverse ordering assignment are both maximum cardinality assignments.
Then we adopt the two assignments M and M to construct the solutions of model (M2), i.e., M; b M.
We denote r as the number of connections in the maximum cardinality assignment of J to I, namely, r ¼ jMj ¼ jMj. We
define two sets as follows:
J 0 ¼ fjnrþk ¼ M ðn r þ kÞj1 6 k 6 rg ð21Þ
I0 ¼ fik ¼ M þ ðkÞj1 6 k 6 rg ð22Þ
We obtain the solutions of model (M2) M b and M by solving the assignment of J to I0 .

0
For a given assignment M, at any time t, the number of up-direction train trips which end at terminal sA no later than t and
satisfy Mþ ðjnrþk Þ – 0 is denoted as uðtÞ. Obviously,
uðtÞ ¼ jfjnrþk 2 J 0 jyjnrþk þ sðsA ; sB Þ 6 t; Mþ ðjnrþk Þ – 0; 1 6 k 6 rgj ð23Þ
The number of down-direction train trips which start from terminal sA no later than t and satisfy M ðih Þ – 0 is denoted as
v ðtÞ. Obviously,
v ðtÞ ¼ jfih 2 I0 jxi h

6 t; M ðih Þ – 0; 1 6 h 6 rgj ð24Þ
Thus, at time t, the number of trains parking at terminal sA is denoted as wðtÞ ¼ uðtÞ v ðtÞ. Then we can obtain the equiv-
alent form of constraint (18):
wðtÞ 6 C A ; t ¼ yjnrþ1 þ sðsA ; sB Þ; yjnrþ2 þ sðsA ; sB Þ; . . . ; yjn þ sðsA ; sB Þ ð25Þ
In turn, we can also use constraint (25) to solve the assignment M. We use the concept of the current time t, i.e., the arrival
times of up-direction train trips and the departure times of down-direction train trips, to iteratively determine the corre-
sponding connections of assignment of J 0 to I0 in the time order; that is, for any t ¼ yjnrþk þ sðsA ; sB Þ or
t ¼ xih ; 1 6 k 6 r; 1 6 h 6 r, we determine the value of M þ ðjnrþk Þ or M ðih Þ. In order to conveniently describe the algorithm,
for any current time t, we denote wðtÞ by w.
When the current time t ¼ yjnrþk þ sðsA ; sB Þ, if it satisfies w ¼ C A before t, then set Mþ ðjnrþk Þ ¼ 0 to guarantee constraint
(25); otherwise, set Mþ ðjnrþk Þ ¼ 1, which indicates that the connecting object of the up-direction train trip jnrþk is to be
determined, and add 1 to w at the moment. Among all the up-direction train trips with M þ ðjnrþk Þ ¼ 1, the one with the
minimum value of k is first determined. We denote the corresponding k by m.
When the current time t ¼ xih , if w ¼ 0 before t, or yjnrþm and xih cannot satisfy constraint (17), then set M ðih Þ ¼ 0; other-
wise, add ðjnrþm ; ih Þ to M, and decrease w by 1 at the moment.
Consequently, the designed algorithm is as follows:
Algorithm 2: Rolling stock circulation optimization algorithm.

Input the departure time vectors X; Y, the total travel time sðsA ; sB Þ, the turning-around time s A , and the parking
capacity C A ;
Output the rolling stock circulations M; M, b the number of trains Z 1 , and the number of yard–station runs Z 2 .
Begin
Solve the maximum cardinality assignment M of J to I using the ordering method, i.e., M ¼ fðk; ik Þj1 6 k 6 rg;
Solve the maximum cardinality assignment M of J to I using the reverse ordering method, i.e.,
M ¼ fðjnrþk ; n r þ kÞj1 6 k 6 rg;
M £; yjnþ1 1;
w 0; k 1; h 1; m 1;
While h 6 r do
if yjnrþk þ sðsA ; sB Þ < xih ðnote : t ¼ yjnrþk þ sðsA ; sB ÞÞ then
if w ¼ C A then M þ ðjnrþk Þ 0
else M þ ðjnrþk Þ 1; w w þ 1;
k k þ 1;
else ðnote : t ¼ xih 6 yjnrþk þ sðsA ; sB ÞÞ
while M þ ðjnrþm Þ – 1 and m < k do m m þ 1;
if w ¼ 0 or yjnrþm þ sðsA ; sB Þ þ s A > xih then M ðih Þ 0
else ðnote : w > 0 and yjnrþm þ sðsA ; sB Þ þ s A 6 xih Þ
M M [ fðjnrþm ; ih Þg; w w 1;
h h þ 1;
End while
e of the set fj þ
Solve the maximum cardinality assignment M nrþk jM ðjnrþk Þ ¼ 0; 1 6 k 6 rg to the set

fih jM ðih Þ ¼ 0; 1 6 h 6 rg by the ordering method;
Mb e [ M;
M
Z1 b Z2
n j Mj; 2ðn jMjÞ;
End
The statements, M þ ðjnrþk Þ 0 and M ðih Þ 0, are added to make the algorithm more clear, although the values have been
set by M £.
Proposition 2. The assignments M; M b solved by Algorithm 2 are the optimal solutions for the optimization model of rolling stock
circulation (M2), and the computational complexity of Algorithm 2 is OðnÞ.
The proof is not shown here due to the limitation of space. In order to explain our algorithm, we propose a small numer-
ical example in Appendix B.
8. Numerical example
8.1. Description
In this section, we take a line of the urban rail transit in Shenzhen, China as a numerical example. This line has 30 stations,
and the dwell times of train trips in each station are: s5 ; s7 ; s8 ; s19 —s28 : 30s; s1 ; s29 ; s30 : 35 s; s9 ; s14 ; s15 : 45 s; other stations:
40 s. The section running times are shown in Table 4. The travel time of section ðk; k þ 1Þ or ðk þ 1; kÞ comprises the section
running time, as well as the dwell time at skþ1 or sk . The total travel time is 70 min. The nominal capacity of a train is
C ¼ 1200 persons per train, the ideal occupancy rate is a ¼ 0:75, and the minimum and maximum train headways are
smin ¼ 2:5 min and smax ¼ 15 min respectively. The turning-around time is s A ¼ sB ¼ 2 min, and the parking capacity is
C A ¼ C B ¼ 2 trains. The operational period is 05:00–24:00. The latest acceptable departure times of the first train trips
t DL and t UL are both 06:20; the earliest acceptable departure times of the last train trips tDE and tUE are both 22:30.
Table 4
Section running time (unit: seconds).
Section Time Section Time Section Time Section Time Section Time
(1, 2) 84 (2, 3) 92 (3, 4) 72 (4, 5) 123 (5, 6) 76
(6, 7) 132 (7, 8) 84 (8, 9) 76 (9, 10) 113 (10, 11) 87
(11, 12) 106 (12, 13) 124 (13, 14) 96 (14, 15) 119 (15, 16) 150
(16, 17) 90 (17, 18) 91 (18, 19) 97 (19, 20) 224 (20, 21) 110
(21, 22) 136 (22, 23) 109 (23, 24) 118 (24, 25) 106 (25, 26) 107
(26, 27) 118 (27, 28) 106 (28, 29) 100 (29, 30) 119
Table 5
Section demand per day (unit: persons).
Section Demand in down direction Section Demand in up direction

Total Max Average Total Max Average
(1, 2) 13,440 61 23.6 (2, 1) 12,080 58 21.2
(2, 3) 23,190 97 40.7 (3, 2) 22,076 100 38.7
(3, 4) 24,649 113 43.2 (4, 3) 23,616 101 41.4
(4, 5) 29,969 148 52.6 (5, 4) 28,532 154 50.1
(5, 6) 36,060 159 63.3 (6, 5) 35,667 150 62.6
(6, 7) 37,509 169 65.8 (7, 6) 36,438 142 63.9
(7, 8) 42,611 190 74.8 (8, 7) 39,722 167 69.7
(8, 9) 43,473 200 76.3 (9, 8) 40,148 174 70.4
(9, 10) 44,642 211 78.3 (10, 9) 41,162 168 72.2
(10, 11) 84,210 450 147.7 (11, 10) 95,270 523 167.1
(11, 12) 84,758 446 148.7 (12, 11) 93,221 509 163.5
(12, 13) 80,337 445 140.9 (13, 12) 88,186 496 154.7
(13, 14) 77,200 340 135.4 (14, 13) 84,834 487 148.8
(14, 15) 75,833 391 133.0 (15, 14) 83,486 482 146.5
(15, 16) 73,162 377 128.4 (16, 15) 81,138 434 142.3
(16, 17) 72,305 322 126.9 (17, 16) 72,305 394 144.8
(17, 18) 66,259 286 116.2 (18, 17) 66,259 333 132.4
(18, 19) 57,170 231 100.3 (19, 18) 57,170 254 115.9
(19, 20) 48,658 166 85.4 (20, 19) 48,658 186 97.9
(20, 21) 46,896 172 82.3 (21, 20) 46,896 167 93.8
(21, 22) 44,800 150 78.6 (22, 21) 44,800 156 88.8
(22, 23) 41,693 142 73.1 (23, 22) 41,693 141 81.1
(23, 24) 40,849 139 71.7 (24, 23) 40,849 124 77.7
(24, 25) 38,744 125 68.0 (25, 24) 38,744 119 73.8
(25, 26) 35,991 121 63.1 (26, 25) 35,991 108 67.8
(26, 27) 32,706 109 57.4 (27, 26) 32,706 89 61.2
(27, 28) 29,688 89 52.1 (28, 27) 29,688 78 56.2
(28, 29) 24,445 74 42.9 (29, 28) 24,445 68 47.4
(29, 30) 18,564 58 32.6 (30, 29) 18,564 55 37.4
Fig. 4. The density distribution of section demand (unit: time/hour, flow/person).

The total volumes, maximum values and average values of section demand in each section per day are listed in Table 5.
Fig. 4 shows the density distribution curves of section demand. For each diagram in Fig. 4, the solid line shows the results of
the down direction, and the dashed line shows the results of the up direction. Fig. 4 indicates that morning peak mainly
affects the demand in the up direction, while evening peak mainly affects the demand in the down direction.
8.2. The optimized results
According to the approach proposed in Section 3, we calculate the MHFs F D ðxÞ; F U ðyÞ in Fig. 5, where the solid line indi-
cates F D ðxÞ and the dashed line indicates F U ðyÞ. Fig. 5 shows that in the morning peak the train headway is small and the
frequency is high in the up direction, whereas in the evening peak the train headway is small and the frequency is high
in the down direction.
Using a computer with an InterCore 2.90 GHz CPU and 1.96 GB RAM, we code our algorithms on Matlab (R2010b). The
computer takes 2.23 s CPU time to complete the optimization of the TOP in the numerical example. Using Algorithms 1
and 2, we obtain the optimized TOP in Fig. 6. The first down- and up-direction train trips start at 05:08:00 and 06:20:00
respectively, and the last down- and up-direction train trips start at 22:36:05 and 23:48:05 respectively. In Fig. 6, the hor-
izontal axis indicates the operation hours, whereas the vertical axis indicates the spatial positions of terminal stations and
the yard. The oblique lines between two terminal stations, called operation lines, represent train trips in two directions,
which can be distinguished by the running direction with time. The labels of some train trips are marked in the figure, where
the letter ‘‘D” and ‘‘U” represent the two directions respectively, and the short dotted lines indicate the corresponding depar-
ture time and arrival time. The duration of the gap between each two neighboring operation lines in the same direction rep-
resents the train headway. The change of frequencies (unit: train/hour) in the two directions can be illustrated by the two
color bars, where a darker color represents a higher frequency, and some extreme values are marked beside the bars. The
rolling stock circulations at two terminal stations are denoted by broken lines and triangles at corresponding positions, in
which broken lines indicate the connections at the station, and the heads of the triangles indicate the directions of yard–sta-
tion runs, i.e. driving into or out of the yard. A solid triangle represents the train driving out of the yard for the first time or
th
driving into the yard for the last time; a hollow triangle represents that the train turned around in the yard, i.e., the i hollow
th
triangle representing driving out of the yard and the i hollow triangle representing driving into the yard generate a con-
nection in the yard.
In order to show the change of train headways in Fig. 6 clearly, the train headways between each two train trips in the
optimized TOP are calculated according to the departure times in the TOP in Fig. 7, where each discrete mark represents the
subsequent train headway between the corresponding train trip and the next one.
Fig. 5. Illustration of MHFs.

Fig. 6. Illustration of the optimized TOP.

Fig. 7. Illustration of comparison between the optimized train headways and the MHFs.
Comparing the optimized train headways and the MHFs in the two directions, we find evidences to demonstrate the coor-
dination of the departure times in the two directions. The optimized down-direction train headways are equal to the down-
direction MHF in the late periods (after the departure time of the down-direction train trip 81, i.e., 14:50:34). The optimized
train headways between down-direction train trips 1–80 can be viewed as the optimized train headways between up-
direction train trips 1–80 moving backward 72 min (total travel time plus turning-around time). This is because in the morn-
ing peak, in order to satisfy the huge travel demand in the up direction, the train headways should be determined by F U ðyÞ,
and down-direction train trips must satisfy the requirement to connect up-direction train trips. In contrast, the optimized
up-direction train headways are equal to the up-direction MHF in the early periods (before the departure time of the up-
direction train trip 88, i.e., 17:19:34). The optimized train headways between up-direction train trips 90–137 can be viewed
as the optimized train headways between down-direction train trips 92–139 moving forward 70 min (total travel time). This
is because in the evening peak, in order to satisfy the huge travel demand in the down direction, the train headways should
be determined by F D ðxÞ, and each up-direction train trip should start from sB when a new down-direction train trip enters the
terminal in order to satisfy the parking capacity (C B ¼ 2 trains).
In addition, there are some special situations:
(1) The train headway between up-direction train trips 89 and 90 is 7.88 min (the first circle in Fig. 7), which is unrelated
to MHFs. It can be seen from Fig. 7 that y89 y88 ¼ F U ðy88 Þ; y91 y90 ¼ F D ðx92 Þ, and F D ðx91 Þ < y90 y89 < F U ðy89 Þ, so the train
headway between up-direction train trips 89 and 90 is a transition point from F U ðyÞ to F D ðxÞ, and the departure time of up-
direction train trip 90 is equal to the arrival time of down-direction train trip 92 at sB .
On one hand, according to the analysis above, the reason for y91 y90 ¼ F D ðx92 Þ is that the number of parking trains at sB
has reached the parking capacity before the down-direction train trip 92 arrives at sB . It means that the up-direction train
trip 90 should start no later than the arrival time of the down-direction train trip 92 to satisfy the parking capacity, i.e.,
y90 6 x92 þ sðsA ; sB Þ < y89 þ F U ðy89 Þ. Hence, the train headway between up-direction train trips 89 and 90 is smaller than
F U ðy89 Þ. On the other hand, according to the principle that each step in Algorithm 1 maximizes the train headway, the
up-direction 90 would start no earlier than the arrival time of the down-direction train trip 92, i.e., y90 P x92 þ sðsA ; sB Þ,
and the inequality y89 < x91 þ sðsA ; sB Þ is satisfied. Hence, the train headway between up-direction train trips 89 and 90 is
larger than F D ðx91 Þ.
(2) The up-direction train trips, which start after the earliest acceptable departure time of the last train trip, drive back
into the yard as quickly as possible. The up-direction train trip 138 is the first to start after t UE . After it starts from sB , the two
parking trains will depart with the minimum train headway, i.e., y139 y138 ¼ y140 y139 ¼ 2:5 min (see the two points in the
second circle in Fig. 7). The up-direction train trips 141–147 are connected by down-direction train trips 141–147 which
immediately turn around after arriving at sB . However, the train headway between up-direction train trips 140 and 141 is
related to the departure time of up-direction train trip 140, which is another special train headway (the third circle in
Fig. 7). Then the subsequent train headways satisfy yiþ1 yi ¼ F D ðxi Þ; i ¼ 141; 142; . . . ; 146. The TOP described in Fig. 6
requires 31 trains, 294 train trips and 92 yard–station runs. Among the 92 yard-station runs, 31 yard-station runs represent
the train trips driving out of the yard for the first time, 31 yard-station runs represent the train trips driving into the yard for
the last time, and 30 yard-station runs represent the 15 pairs of train trips complete connections at yard pA , which are caused
by the limited parking capacity of terminal sA and the imbalance between the departure and arrival of train trips.
Since the rolling stock circulation at terminal sB is based on the first-come-first-serve principle, the down-direction train
trip i connects the up-direction train trip i, which can be described by the trip pair i. Thus, in order to show the rolling stock
circulation of the TOP, we only need to list the connections between trip pairs in Table 6, which represent the rolling stock
circulation of each train unit at terminal sA and yard pA . In Table 6, the symbol ‘‘–” shows the connection at terminal sA , and
the symbol ‘‘!” shows the connection at yard pA .
Table 6
Rolling stock circulation of TOP.
Train unit Trip-pair number Train unit Trip-pair number

1 1 30 48 65 82 101 17 17 41 58 76 92 123 142
2 2 32 50 67 83 103 18 18 ! 105
3 3 33 ! 115 137 19 19 ! 107
4 4 34 51 69 85 108 134 20 20 42 59 77 93 124
5 5 35 52 70 86 111 21 21 ! 109
6 6 36 53 71 87 114 22 22 43 60 78 95 126 144
7 7 ! 68 84 106 133 23 23 ! 110 135
8 8 37 54 72 88 116 24 24 44 61 79 96 127 145
9 9 ! 94 125 143 25 25 ! 112
10 10 38 55 73 89 119 26 26 45 62 ! 117 138
11 11 ! 97 128 146 27 27 46 63 80 98 129
12 12 39 56 74 90 120 140 28 28 ! 113 136
13 13 ! 100 130 147 29 29 47 64 81 99
14 14 ! 102 131 30 31 49 66 ! 118 139
15 15 40 57 75 91 122 141 31 121
16 16 ! 104 132 – –
We use the 23th train unit as an example. The 23th train unit first executes the down-direction train trip 23, and connects
the up-direction train trip 23 at sB ; when arriving at sA , it drives into the yard, and executes the tasks for the down-direction
train trip 110 and the up-direction train trip 110; when it arrives at sA , it parks at the parking line, and continues to execute
the down-direction train trip 135 and the up-direction train trip 135; after that, it finishes its tasks and drives into the yard
for the last time. The above train trips are illustrated in Fig. 6.
8.3. Comparison between the real-world plan and the optimized plan
In real-world operation of the studied line, its first down-direction train trips start at the times shown in Fig. 6. The sub-
sequent train headways and train trips are as follows: 5 train trips with a train headway of 7 min, 35 train trips with a train
headway of 3 min, 98 train trips with a train headway of 6 min, 46 train trips with a train headway of 4 min, and 16 train
trips with a train headway of 9 min. All down-direction train trips turn around immediately when ending at terminal sB .
In order to demonstrate the advantages of the optimized TOP, we make some comparisons between the real-world plan
and the optimized plan on the operation cost and the level of service.
(1) Analysis of operation cost
The statistical indexes in the real-world TOP are 400 train trips, 43 trains and 110 yard–station runs. Compared with the
real-world plan, the optimized TOP has an obvious advantage, because it needs fewer train trips (106 train trips less), trains
(12 trains less) and yard-station runs (18 yard-station runs less), which can greatly reduce the operation costs.
(2) Analysis of the level of service

(a) The constraint for the departure times of the first and last train trip
The departure times of the first train trips of the two TOPs are the same, which are both no later than the latest acceptable
departure times in two directions. The last train trips in two directions of the real-world plan start at 22:37:00 and 23:49:00,
while the last train trips in two directions of the optimized plan start at 22:36:05 and 23:48:05. Thus, they start no earlier
than the earliest acceptable departure times in two directions, and this constraint is satisfied.
Fig. 8. Comparisons between the maximum occupancy rates of train trips in two directions in the real-world plan and the optimized plan.
(b) The constraint for train headway
In terms of the train headway setting, the train headways in the real-world plan and in the optimized plan are between
the minimum train headway smin and the maximum train headway smax . Thus, the real-world plan and the optimized plan
both satisfy the constraint of train headway.
(c) The constraint for train load
Each train trip serves the waiting demand at every station. We calculate the maximum occupancy rate and average occu-
pancy rate of each train trip in the real-world plan and in the optimized plan respectively (Figs. 8 and 9). The former can
reflect the fluctuation of section demand in two directions, while the latter can show the general capacity usage of a train
trip.
As Figs. 8 and 9 show, in general both the maximum occupancy rates and the average occupancy rates in the optimized
plan are higher than those in the real-world plan. In some periods, the maximum occupancy rates and the average occupancy
rates in the optimized plan are the same with those in the real-world plan.
In summary, the optimized TOP can not only ensure the level of service, but also reduce the operation cost. The perfor-
mance is mainly attributed to the flexible and frequent variations of train headways in the optimized TOP, which fit the tra-
vel demand pattern well. The operation efficiency is improved and the level of service is also guaranteed.
8.4. Sensitivity analysis
In addition to the optimized results above, we also make some sensitivity studies.
8.4.1. The value of s in the construction of MHFs

The data points of MHFs are composed by a number of sequences, and s is the interval between two neighboring
sequences. A smaller s would result in more sequences, as well as a higher precision of MHFs and optimized results. Table 7
shows the optimized results of TOP with different values of s. The change of the value of s has no influence on the optimized
results, so it is reasonable to select the minimum time unit of the timetable as the interval, i.e., s ¼ 0:5 min.
8.4.2. The ideal occupancy rate a
(1) The influence of a on the optimized results

Fig. 9. Comparisons between the average occupancy rates of train trips in two directions in the real-world plan and the optimized plan.
The ideal occupancy rate a is an important index of the level of service. Table 8 shows the optimized results of TOP with
different ideal occupancy rates. As Table 8 shows, as the ideal occupancy rate a increases, the number of train trips and the
number of trains both decrease, which is consistent with our intuition. As the ideal occupancy rate a increases, the number of
yard-station runs first increases then decreases. The change of the number of yard-station runs is basically consistent with
our intuition when a P 0:3. However, as a decreases from 0.3, the number of yard-station runs decreases, showing that the
good combination between the arrival times of up-direction train trips at sA and the departure times of down-direction train
trips from sA achieves a high connection efficiency at sA . In addition, the change of the number of trains and yard-station runs
Table 7
The final results of TOP with different intervals between two neighbor sequences.
s (min) Number of data points in every minute between t11 and t12 Number of train trips Number of trains Number of yard-station runs
0.5 2 294 31 92
0.25 4 294 31 92
0.167 6 294 31 92
0.125 8 294 31 92
0.1 10 294 31 92
Table 8
The optimized results of TOP with different ideal occupancy rate a.
Ideal occupancy rate a Number of train trips Number of trains Number of yard-station runs
0.1 836 59 118
0.2 826 59 122
0.25 748 59 152
0.3 662 59 162
0.4 532 54 156
0.5 438 46 136
0.6 366 39 118
0.7 314 34 102
0.75 294 31 92
0.8 276 30 88
0.9 246 27 76
1.0 222 24 70
only reflects a trend, and there are some small perturbations. Since the number of train trips is the main objective of opti-
mization, and keeps decreasing with the increase of a, showing that the operation costs reduces as a increases.
(2) The influence of a on the average waiting time
The defined level of service does not include the average waiting time, but the average waiting time of passengers at each
station can be calculated by section demands and the optimized timetable, and then we can obtain the influence of a on the
average waiting time.
We take station s11 as an example, and the variations of the average waiting times in the two directions from the begin-
ning time to the current time are illustrated in Fig. 10(a). In the up direction, the average waiting time first increases dra-
matically due to the few cumulative demand before the first train trip arrives at station s11, and a small fluctuations are
observed when the first train trip is about to arrive. After the first train trip leaves station s11, the average waiting time
decreases rapidly, and then keeps fluctuating slightly. For the down direction, since the first train trip leaves station s11 rel-
atively early, there is no large fluctuation of the average waiting time.
The relationship between the daily average waiting time and a is shown in Fig. 10(b). It can be seen that the daily average
waiting time increases monotonically when the ideal occupancy rate becomes larger, while a relatively slow growth is
observed when a is small.
8.5. The variation of detained passenger flow
In order to study the correctness of the proposed approach when the detained passenger flow is generated, we propose an
example by enlarging the section demand of the former example by a factor of two. In order to ensure all passengers waiting
for the first train trip can get on board, we adjust the latest acceptable departure times of the first train trips t DL and tUL to
06 : 05. We obtained the MHFs and the optimized train headways of train trips in two directions, and illustrate them in
Fig. 11.
Comparing the obtained MHFs and the optimized train headways in Fig. 11, we find that they have the similar features
with Fig. 7. The difference between Figs. 7 and 11 is that in the morning peak and evening peak, the train headways are
2.5 min, so detained passenger flow may generate during these two periods.
Fig. 10. Illustration of variation of average waiting time at s11.

Fig. 11. Illustration of comparison between the optimized train headways and the MHFs with the consideration of large demand.
Next, we analyze the detained passenger flow in two aspects: (1) in the construction of MHF, the detained passenger flow
at every station may generate at the discrete time points, so we can draw the curves of detained passenger flow at every
station; (2) we can simulate the transit process of the optimized plan and obtain the detained passenger flow at every station
Fig. 12. Illustration of fluctuation of detained passenger flow at s11 in the construction of MHF and in the optimized plan.
generated when each train trip leaves. Thus, we can verify the effect of MHF by comparing the curves of detained passenger
flow and the simulated detained passenger flows.
Due to the large fluctuation of section demand in section (11, 10), we take s11 as an example, and illustrate the fluctuation
of detained passenger flow in the construction of F U ðyÞ; y 2 ½T 1 ; T 2 . It shows the up-direction detained passenger flow at s11
generated during 08:16:35–09:06:09 when constructing the F U ðyÞ; y 2 ½T 1 ; T 2 . With the optimized timetable, we simulate
the process in which passengers get on train trips at s11, and obtain the number of passengers left behind the overloaded
trains. In the optimized plan, the up-direction detained passenger flow at s11 is generated during 08:17:24–09:04:54
(the period between the departure times of the up-direction train trip 22 and 41 from s11). Fig. 12 shows that the detained
passenger flow when each train trip leaves the station can perfectly fit the curve of the detained passenger flow generated in
the construction of MHF, which can be verified at every station. It helps to demonstrate the correctness of the construction of
MHFs.
9. Conclusions
Focusing on a URC with one terminal yard, this paper studies the TOP optimization problem based on the time-varying
demand and a predetermined level of service, which is decomposed into two sub-problems, i.e., the train departure profile
optimization and the rolling stock circulation optimization. The first sub-problem systematically optimizes the departure
times of train trips, and the rolling stock circulation at the terminal without a yard; a heuristic algorithm with linear com-
putational complexity is designed to minimize the number of train trips. For the second sub-problem, a model is proposed to
optimize the rolling stock circulation at the terminal with a yard, and the optimal solution is solved by an optimal algorithm
with linear computational complexity, which sequentially minimizes the number of trains, and the number of yard–station
runs.
Considering both time-varying demand and the level of service, we construct the maximum headway function (MHF). If
the train headway does not exceed the corresponding value of the MHF, the service of the TOP could reach the predeter-
mined level. We present a framework for optimizing the TOP in a URC using the concept of the MHF, which helps to obtain
the maximized train headways to optimize the timetable without considering time-varying travel demand. Thus, the MHF
plays an important role in the train departure profile optimization.
The method for rolling stock circulation optimization at the terminal with a yard can be applied to the rolling stock cir-
culation at any terminal station with a yard.
In conclusion, based on MHF, the flexible and frequent variations of train headways fit the time-varying travel demand
well, so the optimized TOP can not only provide the predetermined level of service, but also reduce the operation cost.
Acknowledgments
This study was supported by the National Natural Science Foundation of China (Grant Nos. 71171200, U1334207,
61473320) and the Fundamental Research Funds for the Central Universities of Central South University (Grant No. 2016z-
zts051). The authors thank Dr. Xuesong Zhou for the valuable constructive comments on this paper.
Appendix A. A small numerical example of Algorithm 1
The relative parameters are as follows:
sðsA ; sB Þ ¼ 58 min; sB ¼ 2 min; C B ¼ 2 trains; tDL ¼ 06 : 00; tUL ¼ 07 : 00 and

6 min; x 2 ½06 : 00; 06 : 16 18 min; y 2 ½06 : 00; 07 : 15
F D ðxÞ ¼ ; F U ðyÞ ¼
16 min; x 2 ð06 : 16; 08 : 00 8 min; y 2 ð07 : 15; 08 : 00
Every step demonstrated in the following determines the departure time xi or yj of a down- or up-direction train trip or
both, as well as the latest departure times j of the next train trips to be determined in the two directions. We illustrate
xi and y
each step in Fig. 13, in which a solid line corresponds to a train trip with the determined departure time xi or yj ; a dashed line
corresponds to a train trip with the latest departure time j to be determined.
xi or y
At the beginning of Algorithm 1, we obtain the departure times of the first train trips in the two directions, i.e.,
x1 ¼ 06 : 00; y1 ¼ 07 : 00. The sequence numbers of train trips to be determined in the two directions are i ¼ 2; j ¼ 2. The lat-
est departure times to be determined are x2 ¼ x1 þ F D ðx1 ¼ 06 : 00Þ ¼ 06 : 06; y 2 ¼ y1 þ F U ðy1 ¼ 07 : 00Þ ¼ 07 : 18. Fig. 13(a)
illustrates train trips starting at x1 ; 2 .
x2 ; y1 ; y
2 ¼ 07 : 18 > 07 : 06 ¼
As i j ¼ 0; y x2 þ sðsA ; sB Þ þ sB , according to the second if statement in Algorithm 1, we can obtain
the departure time of the current down-direction train trip x2 ¼ x2 ¼ 06 : 06. Then i ¼ 3; j ¼ 2, and the latest departure times
to be determined are 2 ¼ 07 : 18 respectively. Fig. 13(b) illustrates train trips starting at x1 ; x2 ;
x3 ¼ 06 : 12; y 2 .
x3 ; y1 ; y
6:00 7:00 8:00 9:00 6:00 7:00 8:00 9:00

SA SA
SB SB
6:00 7:00 8:00 9:00 6:00 7:00 8:00 9:00
(a) train trips starting at (b) train trips starting at
6:00 7:00 8:00 9:00 6:00 7:00 8:00 9:00
SA SA
SB SB
6:00 7:00 8:00 9:00 6:00 7:00 8:00 9:00
(c) train trips starting at (d) train trips starting at
6:00 7:00 8:00 9:00 6:00 7:00 8:00 9:00

SA SA
SB SB
6:00 7:00 8:00 9:00 6:00 7:00 8:00 9:00
(e) train trips starting at (f) train trips starting at
6:00 7:00 8:00 9:00 6:00 7:00 8:00 9:00

SA SA
SB SB
6:00 7:00 8:00 9:00 6:00 7:00 8:00 9:00
(g) train trips starting at (h) train trips starting at
Fig. 13. Illustration of execution of each step in Algorithm 1.
As 0 < i j < 2; y 2 ¼ 07 : 18 > 07 : 10 ¼ x3 þ sðsA ; sB Þ, according to the fourth if statement in Algorithm 1, we can obtain
the departure time of the current down-direction train trip x3 ¼ x3 ¼ 06 : 12. Then i ¼ 4; j ¼ 2, and the latest departure times
to be determined are x4 ¼ 06 : 18; y2 ¼ 07 : 18 respectively. Fig. 13(c) illustrates train trips starting at x1 ; . . . ; x3 ; 2 .
x4 ; y1 ; y
As i j ¼ 2; y 2 ¼ 07 : 18 P 07 : 16 ¼ x4 þ sðsA ; sB Þ, according to the seventh if statement in Algorithm 1, we can obtain the
departure times of the current train trips in the two directions: x4 ¼ x4 þ sðsA ; sB Þ ¼ 07 : 16 < y
x4 ¼ 06 : 18; y2 ¼ 2 . Then
i ¼ 5; j ¼ 3, and the latest departure times to be determined are

x5 ¼ x4 þ F D ðx4 ¼ 06 : 18Þ ¼ 06 : 34; y 3 ¼ y2 þ F U ðy2 ¼ 07 : 16Þ ¼ 07 : 24 respectively. Fig. 13(d) illustrates train trips starting
at x1 ; . . . ; x4 ; 3 . In this situation, if the departure time of the up-direction train trip j ¼ 2 is y2 ¼ y
x5 ; y1 ; y2 ; y 2 ¼ 07 : 18, then
three trains corresponding to down-direction train trips i ¼ 2; 3; 4 would park at sB during the interval ½07 : 16; 07 : 18Þ,
which exceeds the parking capacity of sB , i.e., C B ¼ 2trains. Thus, the departure time of the up-direction train trip j ¼ 2 needs
to be determined by the arrival time of the down-direction train trip i ¼ 4, i.e., y2 ¼ x4 þ sðsA ; sB Þ ¼ 07 : 16, to ensure that the
number of parking trains does not exceed the parking capacity, which indicates constraint (16).
As i j ¼ 2; y 3 ¼ 07 : 24 < 07 : 32 ¼ x5 þ sðsA ; sB Þ, according to the sixth if statement in Algorithm 1, we can obtain the
departure time of the current up-direction train trip y3 ¼ y 3 ¼ 07 : 24. Then i ¼ 5; j ¼ 4, and the latest departure times to
be determined are x5 ¼ 06 : 34; y4 ¼ 07 : 32 respectively. Fig. 13(e) illustrates train trips starting at x1 ; . . . ; x4 ; 4 .
x5 ; y1 ; . . . ; y3 ; y
As 0 < i j < 2; y 4 ¼ 07 : 32 ¼ x5 þ sðsA ; sB Þ, according to the fifth if statement in Algorithm 1, we can obtain the departure
times of the current train trips in the two directions: x5 ¼ x5 ¼ 06 : 34; y4 ¼ y 4 ¼ 07 : 32. Then i ¼ 6; j ¼ 5, and the latest
departure times to be determined are x6 ¼ 06 : 50; y 5 ¼ 07 : 40 respectively. Fig. 13(f) illustrates train trips starting at
x1 ; . . . ; x5 ; 5 .
x6 ; y1 ; . . . ; y4 ; y
As 0 < i j < 2; y 5 ¼ 07 : 40 < 07 : 48 ¼ x6 þ sðsA ; sB Þ, according to the third if statement in Algorithm 1, we can obtain
the departure time of the current up-direction train trip y5 ¼ y 5 ¼ 07 : 40. Then i ¼ 6; j ¼ 6, and the latest departure times
to be determined are x6 ¼ 06 : 50; y6 ¼ 07 : 48 respectively. Fig. 13(g) illustrates train trips starting at
x1 ; . . . ; x5 ; 6 .
x6 ; y1 ; . . . ; y5 ; y
As i j ¼ 0; y 6 ¼ 07 : 48 < 07 : 50 ¼ x6 þ sðsA ; sB Þ þ sB , according to the first if statement in Algorithm 1, we can obtain the
departure times of the current train trips in the two directions: x6 ¼ y 6 sðsA ; sB Þ sB ¼ 06 : 48 < 6 ¼ 07 : 48. Then
x6 ; y6 ¼ y
i ¼ 7; j ¼ 7, and the latest departure times to be determined are x7 ¼ 07 : 04; y 7 ¼ 07 : 56 respectively. Fig. 13(h) illustrates
train trips starting at x1 ; . . . ; x6 ; y1 ; . . . ; y6 . In this situation, if the departure time of the down-direction train trip i ¼ 6 is
x6 ¼ x6 , then the earliest departure time of the up-direction train trip j ¼ 6 is x6 þ sðsA ; sB Þ þ sB ¼ 07 : 50. Since the latest

departure time of the up-direction train trip j ¼ 6 is y6 ¼ 07 : 48 < 07 : 50, x6 needs to be determined by y 6 , i.e.,
x6 ¼ y 6 sðsA ; sB Þ sB , so as to ensure the up-direction train trip j ¼ 6 can start from sB at x6 þ sðsA ; sB Þ þ sB ¼ y 6 ¼ 07 : 48,
which indicates constraint (15).
In addition, we use broken lines to represent the connections between the down-direction train trip i and the up-direction
train trip i ði ¼ 1; 2; . . . ; 6Þ in the rolling stock circulation at terminal sB , which are illustrated in Fig. 13(h). It can be seen from
Fig. 13(h) that the turning-around time sB ¼ 2 min is the lower bound of the time of each connection, which satisfies con-
straint (15); the number of parking trains does not exceed the parking capacity C B ¼ 2 trains at any time, which satisfies con-
straint (16).
Appendix B. A small numerical example of Algorithm 2
Assuming that the timetable and the rolling stock circulation at terminal sB have been solved by Algorithm 1, Fig. 13 illus-
trates the situation, and the relevant parameters are as follows: s A ¼ 2 min;C A ¼ 2 trains, the operation period of the exam-
ple is [6:00, 9:00], and the number of train trips in each direction is n ¼ 14. The indexes of the whole train trips are marked in
Fig. 14 (the time unit in Fig. 14 is set to be 2 min).
In the following, we demonstrate how to solve the rolling stock circulation at terminal sA using Algorithm 2. It can be seen
that the determination of connections focuses on the period [7:00, 8:00].
According to Algorithm 2, we first solve an ordering assignment M ¼ fð1; 9Þ; ð2; 10Þ; ð3; 11Þ; ð4; 12Þ; ð5; 13Þ; ð6; 14Þg, which
is illustrated in Fig. 15(a), and a reverse ordering assignment M ¼ fð7; 14Þ; ð6; 13Þ; ð5; 12Þ; ð4; 11Þ; ð3; 10Þ; ð2; 9Þg, which is
illustrated in Fig. 15(b). Then it implies r ¼ jMj ¼ jMj ¼ 6.
We can obtain ih ; 1 6 h 6 6, i.e., i1 ¼ 9; i2 ¼ 10; i3 ¼ 11; i4 ¼ 12; i5 ¼ 13; i6 ¼ 14 from M, and jnrþk ; 1 6 k 6 6, i.e.,
j9 ¼ 2; j10 ¼ 3; j11 ¼ 4; j12 ¼ 5; j13 ¼ 6; j14 ¼ 7 from M.
The initial values of the main loop of Algorithm 2 can be set as follows: M £; yj15 1; w 0; k 1; h 1; m 1.
We describe each execution of the main loop in order of the value of h ¼ 1; 2; . . . ; 6. With the departure times and arrival
times of train trips at terminal sA in Fig. 14, the statements and conclusions of each execution of the main loop are illustrated
as follows:
48
6:00 7:00 08 16 20 24 28 32 36 40 47 54 8:00 9:00
SA
10
11
12
13
14
1
11
10
12
13
14
6
1
SB
6:00 7:00 8:00 9:00
Fig. 14. Illustration of train timetable and rolling stock circulation at terminal sB .
7:00 8:00 7:00 8:00
10
11
12
13
14
10
11
12
13
14
9
6
10
10
6
8
7
1
7
2
9
1
7:00 8:00 7:00 8:00
(a) an ordering assignment (b) a reverse ordering assignment
Fig. 15. Illustration of an ordering assignment and a reverse ordering assignment.
pA
08 16 28 32 48
6:00 7:00 20 24 36 40 47 54 8: 00 9: 00
SA
10
11
12
13
14
1
11
10
12
13
14
6
8
1
SB 9
6: 00 7: 00 8: 00 9: 00
Fig. 16. Illustration of the optimized rolling stock circulation.
(1) h ¼ 1 6 6 ¼ r
① j9 ¼ 2; i1 ¼ 9: y2 þ sðsA ; sB Þ < x9 ð7 : 08 < 7 : 20Þ
Since w < C A , M þ ðj9 ¼ 2Þ 1; w w þ 1 ¼ 1; k k þ 1 ¼ 2.
② j10 ¼ 3; i1 ¼ 9: y3 þ sðsA ; sB Þ < x9 ð7 : 16 < 7 : 20Þ
③ j11 ¼ 4; i1 ¼ 9: x9 6 y4 þ sðsA ; sB Þð7 : 20 6 7 : 24Þ
Since Mþ ðj9 ¼ 2Þ ¼ 1, m remains 1.
Since w > 0 and yj9 þ sðsA ; sB Þ þ s A 6 xi1 ,
M fðj9 ; i1 Þ ¼ ð2; 9Þg; w w 1 ¼ 1; h h þ 1 ¼ 2:
(2) h ¼ 2 6 6 ¼ r
① j11 ¼ 4; i2 ¼ 10: y4 þ sðsA ; sB Þ < x10 ð7 : 24 < 7 : 28Þ
② j12 ¼ 5; i2 ¼ 10: x10 6 y5 þ sðsA ; sB Þð7 : 28 < 7 : 32Þ
Since M þ ðj9 ¼ 2Þ – 1 and m ¼ 1 < k ¼ 4, m m þ 1 ¼ 2.
M fð2; 9Þ; ðj10 ; i2 Þ ¼ ð3; 10Þg; w w 1 ¼ 1; h h þ 1 ¼ 3:
(3) h ¼ 3 6 6 ¼ r
① j12 ¼ 5; i3 ¼ 11: y5 þ sðsA ; sB Þ < x11 ð7 : 32 < 7 : 36Þ
② j13 ¼ 6; i3 ¼ 11: x11 6 y6 þ sðsA ; sB Þð7 : 36 < 7 : 40Þ
Since Mþ ðj10 ¼ 3Þ – 1 and m ¼ 2 < k ¼ 5, m m þ 1 ¼ 3.
M fð2; 9Þ; ð3; 10Þ; ðj11 ; i3 Þ ¼ ð4; 11Þg; w w 1 ¼ 1; h h þ 1 ¼ 4:
(4) h ¼ 4 6 6 ¼ r
① j13 ¼ 6; i4 ¼ 12: y6 þ sðsA ; sB Þ < x12 ð7 : 40 < 7 : 48Þ
② j14 ¼ 7; i4 ¼ 12: y7 þ sðsA ; sB Þ < x12 ð7 : 47 < 7 : 48Þ

Since w ¼ 2 ¼ C A , M þ ðj14 ¼ 7Þ 0; k k þ 1 ¼ 7.
③ j15 ¼ 8; i4 ¼ 12: x12 6 y8 þ sðsA ; sB Þð7 : 48 6 1Þ
M fð2; 9Þ; ð3; 10Þ; ð4; 11Þ; ðj12 ; i4 Þ ¼ ð5; 12Þg;
w w 1 ¼ 1; h h þ 1 ¼ 5:
(4) h ¼ 5 6 6 ¼ r
① j15 ¼ 8; i5 ¼ 13: x13 6 y8 þ sðsA ; sB Þð7 : 54 6 1Þ
M fð2; 9Þ; ð3; 10Þ; ð4; 11Þ; ð5; 12Þ; ðj13 ; i5 Þ ¼ ð6; 13Þg;
w w 1 ¼ 0; h h þ 1 ¼ 6:
(6) h ¼ 6 ¼ r
① j15 ¼ 8; i6 ¼ 14: x14 6 y8 þ sðsA ; sB Þð8 : 00 6 1Þ
Since Mþ ðj13 ¼ 6Þ – 1 and m ¼ 5 < k ¼ 7, m m þ 1 ¼ 6.
Since w ¼ 0, M ði6 ¼ 14Þ 0; h h þ 1 ¼ 7.
(7) h ¼ 7 > 6 ¼ r
End the main loop.
So far, we obtain M ¼ fð2; 9Þ; ð3; 10Þ; ð4; 11Þ; ð5; 12Þ; ð6; 13Þg.
Since M þ ðj14 ¼ 7Þ ¼ 0 and M ði6 ¼ 14Þ ¼ 0, M e ¼ fðj ; i6 Þg ¼ fð7; 14Þg.
14
Then M b e [ M ¼ fð2; 9Þ; ð3; 10Þ; ð4; 11Þ; ð5; 12Þ; ð6; 13Þ; ð7; 14Þg.
M
Finally, Z 1 b ¼ 14 6 ¼ 8 and Z 2
n j Mj 2ðn jMjÞ ¼ 2ð14 5Þ ¼ 18.
b which is illustrated in Fig. 16. We use broken lines to represent the con-
The rolling stock circulation at terminal sA is ðM; MÞ,
b
nections in M. The connection ð7; 14Þ in M M represents the train runs between the terminal sA and the yard p , and the A
hollow arrowheads represent the connection direction. The other train trips are irrelevant to M, b of which the down-direction
train trips start from the yard for the first time, and the up-direction train trips return to the yard for the last time. The solid
arrowheads represent the running direction of trains. The TOP needs 8 trains and 18 yard–station runs.
In contrast, it can be seen from Fig. 15 that both M and M have 6 connections at terminal sA and the yard pA , of which 2
pairs of train trips need to connect at the yard pA , i.e., ð3; 11Þ; ð6; 14Þ in M and ð4; 11Þ; ð7; 14Þ in M. Thus, the rolling stock cir-
culation solved by the ordering method or the reverse ordering method needs 8 trains and 20 yard-station runs. Therefore,
the TOP solved by Algorithm 2 needs fewer yard-station runs, which reflects the optimization of Algorithm 2.
References
Albrecht, T., 2009. Automated timetable design for demand-oriented service on suburban railways. Publ. Transp. 1 (1), 5–20.
Assis, W.O., Milani, B.E.A., 2004. Generation of optimal schedules for metro lines using model predictive control. Automatica 40 (8), 1397–1404.
Cadarso, L., Marín, Á., 2014. Improving robustness of rolling stock circulations in rapid transit networks. Comput. Oper. Res. 51, 146–159.
Carey, M., Crawford, I., 2007. Scheduling trains on a network of busy complex stations. Transp. Res. Part B: Method. 41 (2), 159–178.
Ceder, A., 1984. Bus frequency determination using passenger count data. Transp. Res. Part A: Policy Pract. 18 (5), 439–453.
Ceder, A., 1986. Methods for creating bus timetables. Transp. Res. Part A: Policy Pract. 21 (1), 59–83.
Ceder, A., 2001. Bus timetables with even passenger loads as opposed to even headways. Transp. Res. Rec.: J. Transp. Res. Board 1760, 3–9.
Ceder, A., 2002. Urban transit scheduling: framework, review and examples. J. Urban Plan. Develop. 128 (4), 225–244.
Ceder, A., 2007. Public Transit Planning and Operation: Theory, Modeling and Practice. Butterworth-Heinemann, Oxford.
Ceder, A., 2009. Public-transport automated timetables using even headway and even passenger load concepts. In: The 32nd Australasian Transport
Research Forum, Auckland, 29 Semptember–1 October 2009, pp. 1–17.
Corman, F., D’Ariano, A., Pacciarelli, D., Pranzo, M., 2010. Bi-objective conflict detection and resolution in railway traffic management. Trans. Res. Part C:
Emerg. Technol. 20 (1), 79–94.
Dündar, S., Sahin, I., 2013. Train re-scheduling with genetic algorithms and artificial neural networks for single-track railways. Trans. Res. Part C: Emerg.
Technol. 27 (2), 1–15.
Freling, R., Wagelmans, A.P.M., Paixao, J.M.P., 2001. Models and algorithms for single depot vehicle scheduling. Transp. Sci. 35 (2), 165–180.
Guihaire, V., Hao, J.-K., 2008. Transit network design and scheduling: a global review. Transp. Res. Part A: Policy Pract. 42 (10), 1251–1273.
Guihaire, V., Hao, J.-K., 2010. Transit network timetabling and vehicle assignment for regulating authorities. Comput. Ind. Eng. 59 (1), 16–23.
Ibarra-Rojas, O.J., Giesen, R., Rios-Solis, Y.A., 2014. An integrated approach for timetabling and vehicle scheduling problems to analyze the trade-off between
level of service and operating costs of transit networks. Transp. Res. Part B: Method. 70, 35–46.
Jamili, A., Pourseyed-Aghaee, M., 2015. Robust stop-skipping patterns in urban rail operations under traffic alteration situation. Trans. Res. Part C: Emerg.
Technol. 61, 63–74.
Jiang, Z., Hsu, C., Zhang, D., Zou, X., 2016. Evaluating rail transit timetable using big passengers’ data. J. Comput. Syst. Sci. 82, 144–155.
Liebchen, C., 2008. The first optimized railway timetable in practice. Transp. Sci. 42 (4), 420–435.
Li, X., Lo, H.K., 2014. Energy minimization in dynamic train scheduling and control for metro rail operations. Transp. Res. Part B: Method. 70, 269–284.
Niu, H., Zhou, X., 2013. Optimizing urban rail timetable under time-dependent demand and oversaturated conditions. Trans. Res. Part C: Emerg. Technol. 36,
212–230.
Niu, H., Zhou, X., Gao, R., 2015. Train scheduling for minimizing passenger waiting time with time-dependent demand and skip-stop patterns: nonlinear
integer programming models with linear constraints. Transp. Res. Part B: Method. 76, 117–135.
Odijk, M.A., 1996. A constraint generation algorithm for the construction of periodic railway timetables. Transp. Res. Part B: Method. 30 (6), 455–464.
Petersen, H.L., Larsen, A., Madsen, O.B.G., Petersen, B., Ropke, S., 2012. The simultaneous vehicle scheduling and passenger service problem. Transp. Sci. 47
(4), 603–616.
Poon, M.H., Tong, C.O., Wong, S.C., 2004. A dynamic schedule-based model for congested transit network. Transp. Res. Part B 38, 343–368.
Qi, J., Yang, L., Gao, Y., Li, S., Gao, Z., 2016. Integrated multi-track station layout design and train scheduling models on railway corridors. Trans. Res. Part C:
Emerg. Technol. 69, 91–119.
Sun, L., Jin, J.G., Lee, D., Axhausen, K.W., Erath, A., 2014. Demand-driven timetable design for metro services. Trans. Res. Part C: Emerg. Technol. 46, 284–299.
Vuchic, V., 2005. Urban Transit: Operations, Planning and Economics. Wiley, Hoboken.
Wang, Y., Tang, T., Ning, B., Boom, T., Schutter, B.D., 2015. Passenger-demands-oriented train scheduling for an urban rail transit network. Trans. Res. Part C:
Wong, R.C.W., Yuen, T.W.Y., Fung, K.W., Leung, J.M.Y., 2008. Optimizing timetable synchronization for rail mass transit. Transp. Sci. 42 (1), 57–69.
Wu, J., Liu, M., Sun, H., Li, T., Gao, Z., Wang, D.Z., 2015. Equity-based timetable synchronization optimization in urban subway network. Trans. Res. Part C:
Xiong, J., He, Z., Guan, W., Ran, B., 2015. Optimal timetable development for community shuttle network with metro stations. Trans. Res. Part C: Emerg.
Technol. 60, 540–565.
Zhou, X., Zhong, M., 2005. Train scheduling for high-speed passenger railroad planning applications. Eur. J. Oper. Res. 167 (3), 752–771.
Zhou, X., Zhong, M., 2007. Single-track train timetabling with guaranteed optimality: branch and bound algorithms with enhanced lower bounds. Transp.
Res. Part B: Method. 41 (3), 320–341.

Transportation Research Part C: Feng Shi, Shuo Zhao, Zhao Zhou, Pu Wang, Michael G.H. Bell

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Transportation Research Part C: Feng Shi, Shuo Zhao, Zhao Zhou, Pu Wang, Michael G.H. Bell

Uploaded by

Copyright:

Available Formats

Transportation Research Part C 74 (2017) 51–80

Contents lists available at ScienceDirect

Transportation Research Part C

Optimizing train operational plan in an urban rail corridor

2.1. The level of service

(2) The constraint for train headway

(3) The constraint for train load

ii. Peak hours

2.2. TOP optimization in a URC with one terminal yard

Drive into the yard

Drive out of the yard

3. Train headway and the maximum headway function

3.1. Train headway setting

(1) During off-peak hours

(2) During peak hours

3.2. Deduction formulas for maximizing the train headway

(1) During off-peak hours

F D ðxi Þ ¼ minfsmax ; minft i k xi k jsk 2 S n fsB ggg

rsk ðxiþ1 Þ ¼ 0; sk 2 S n fsB g ð3Þ

F D ðxi Þ ¼ smin ð4Þ

Off-peak hours Peak hours Peak hours

Fig. 2. Illustration of a small case of the deduction of maximizing train headway.

formulas (4) and (5). Values of xsiþ3

3.3. Approximate construction of the maximum headway function

Fig. 3. Illustration of construction of sequences of data points.

4. Train departure profile optimization model

xiþ1 xi P smin ; i ¼ 1; 2; . . . ; n 1 ð11Þ

yjþ1 yj P smin ; j ¼ 1; 2; . . . ; n 1 ð12Þ

The constraint for the MHF

xiþ1 xi 6 F D ðxi Þ; i ¼ 1; 2; . . . ; n 1 ð13Þ

yjþ1 yj 6 F U ðyj Þ; j ¼ 1; 2; . . . ; n 1 ð14Þ

The constraint for turning-around time at the terminal sB without a yard

xi þ sðsA ; sB Þ P yiCB ; i ¼ C B þ 1; C B þ 2; . . . ; n ð16Þ

5. A heuristic algorithm for train departure profile optimization

(3) 0 < i j < C B

Algorithm 1: Departure time adjustment algorithm.

6. Rolling stock circulation optimization model

jfðj; iÞjyj þ sðsA ; sB Þ 6 t 6 xi ; ðj; iÞ 2 Mgj 6 C A ; t ¼ y1 þ sðsA ; sB Þ; y2 þ sðsA ; sB Þ; . . . ; yn þ sðsA ; sB Þ ð18Þ

7. An algorithm for rolling stock circulation optimization

First, we solve two maximum cardinality assignments of J to I, i.e., M and M.

(1) Solve the assignment M

We set M þ ð0Þ ¼ 0, and do the following process in order of j ¼ 1; 2; . . . ; n. If

(2) Solve the assignment M

We set M ðn þ 1Þ ¼ 1, and do the following process in order of i ¼ n; n 1; . . . ; 1. If

J 0 ¼ fjnrþk ¼ M ðn r þ kÞj1 6 k 6 rg ð21Þ

I0 ¼ fik ¼ M þ ðkÞj1 6 k 6 rg ð22Þ

We obtain the solutions of model (M2) M b and M by solving the assignment of J to I0 .

uðtÞ ¼ jfjnrþk 2 J 0 jyjnrþk þ sðsA ; sB Þ 6 t; Mþ ðjnrþk Þ – 0; 1 6 k 6 rgj ð23Þ

v ðtÞ ¼ jfih 2 I0 jxi h

wðtÞ 6 C A ; t ¼ yjnrþ1 þ sðsA ; sB Þ; yjnrþ2 þ sðsA ; sB Þ; . . . ; yjn þ sðsA ; sB Þ ð25Þ

Algorithm 2: Rolling stock circulation optimization algorithm.

Section Demand in down direction Section Demand in up direction

Fig. 4. The density distribution of section demand (unit: time/hour, flow/person).

8.2. The optimized results

Fig. 5. Illustration of MHFs.

Fig. 6. Illustration of the optimized TOP.

Train unit Trip-pair number Train unit Trip-pair number

(1) Analysis of operation cost

(2) Analysis of the level of service

(b) The constraint for train headway

(c) The constraint for train load

8.4. Sensitivity analysis