Transportation Research Part E: Yihui Wang, Tao Tang, Bin Ning, Lingyun Meng

Transportation Research Part E 105 (2017) 83–104
Contents lists available at ScienceDirect
Transportation Research Part E

journal homepage: www.elsevier.com/locate/tre
Integrated optimization of regular train schedule and train

circulation plan for urban rail transit lines
Yihui Wang, Tao Tang, Bin Ning, Lingyun Meng ⇑
State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing 100044, PR China
a r t i c l e i n f o a b s t r a c t
Article history: This paper presents an integrated model to optimize the train schedule and circulation plan
Received 2 September 2016 simultaneously based on a given service pattern generated by the demand analysis and line
Received in revised form 29 April 2017 planning. The operation of train services, the turnaround operations, the entering/exiting
Accepted 1 June 2017
depot operation, and the number of available trains are involved in the model. The pro-
posed integrated and extended integrated optimization problems are transformed into
mixed integer linear programming (MILP) problems, which can be efficiently solved by
Keywords:
the CPLEX solver. Numerical examples based on the Beijing Yizhuang line are implemented
Urban rail transit
Regular train schedule
to demonstrate the performance of the proposed models and solution approach.
Train circulation plan Ó 2017 Elsevier Ltd. All rights reserved.
Integrated optimization
MILP
1. Introduction
Urban rail transit systems play an important role in public transportation since it combines high transport capacity and
high efficiency. With the increasing of passenger demands, the frequency of train operations is becoming very high, espe-
cially in large cities like Beijing, Shanghai, Tokyo, New York, and Paris, where the headway between trains is often less than
10 min and even close to 2 min for some lines. Hence, the planning process for the urban rail transit systems is becoming
more and more significant for reducing the operation costs of rail operators and for guaranteeing passenger satisfaction.
The planning process traditionally consists of five sequential processes (Bussieck et al., 1997): demand analysis, line plan-
ning, train scheduling, train (or rolling stock) circulation planning, and crew scheduling. In particular, the line planning deci-
des the service pattern, which involves the type of line services (e.g., full-length and short-turning services), the frequencies
or the headways between train services, train compositions, etc. The train scheduling process then generates a feasible train
schedule that satisfies the requirements specified in the service pattern. Based on the feasible train schedule, a train circu-
lation plan is then scheduled. Finally, the crew schedule is generated based on the train circulation plan. If one of the process
cannot generate feasible results, then the previous process or even the whole sequential processes should be calculated
again. This paper only focuses on the generation of the train schedule and train circulation plan based on a given service pat-
tern specified by the demand analysis and line planning for an urban rail transit line.
The train scheduling and circulation planning are handled separately in general due to the complexity of the integrated
problem. However, if we solve the train scheduling problem without considering the train circulation, then there could be no
feasible train circulation plan for a feasible train schedule, especially when the number of available trains is limited. In the
train circulation planning phase, the train schedules need to be adjusted in general to satisfy the constraints presented by the
⇑ Corresponding author.
E-mail addresses: yihui.wang@bjtu.edu.cn (Y. Wang), ttang@bjtu.edu.cn (T. Tang), bning@bjtu.edu.cn (B. Ning), lym@bjtu.edu.cn (L. Meng).
http://dx.doi.org/10.1016/j.tre.2017.06.001
1366-5545/Ó 2017 Elsevier Ltd. All rights reserved.
84 Y. Wang et al. / Transportation Research Part E 105 (2017) 83–104
train circulation, where the departure and arrival times of the train services should be changed and some of the train services
even need to be cancelled if there are not enough trains. Hence, we propose the integration of the train scheduling and cir-
culation planning to obtain the train schedules and circulation plans simultaneously.
1.1. Literature review
Train scheduling for regional or national rail transit systems has been studied by many researchers (Szpigel, 1972;
Petersen et al., 1986; Kraay et al., 1991; Higgins et al., 1996; Cordeau et al., 1998; Ghoseiri et al., 2004; D’Ariano et al.,
2007b,a; Corman et al., 2012), where trains could overtake or cross each other at the sidings and crossings. In particular,
the integration of the train scheduling and train routing/platforming has been considered for regional and national rail tran-
sit systems in Samà et al. (2017, 2016), Pellegrini et al. (2014), Dewilde et al. (2013), where mixed integer linear program-
ming models are formulated to tackle these problems. Moreover, the integration of train scheduling and delay management
has been researched in Dollevoet et al. (2014), Corman et al. (2014, 2016) to obtain a better trade-off between the passenger
travel time minimization and train delay minimization. The train circulation planning problem is also called train assign-
ment problem or rolling stock circulation problem. Many studies, e.g., Alfieri et al. (2006), Fioole et al. (2006), Peeters and
Kroon (2008), Cacchiani et al. (2010, 2013) and Giacco et al. (2014), have tackled the train circulation planning, where
the train units are coupled or decoupled based on the passenger demand and are scheduled through the whole rail network.
In this paper, we focus on urban rail transit systems, where the overtaking and crossing of trains are normally not allowed
during the operations. Furthermore, the trains or electrical multiple units (EMUs) are employed only in the operation of a
specific line and the composition of the trains or EMUs does not change in general.
There are two types of train schedules that are concentrated by the researchers. The first type is regular (or periodic) train
schedules, which have different fixed headways for peak hours and off-peak hours. The regular train schedules are often
applied in the current practice of urban rail transit systems. For example, every three minutes there is a train entering a sta-
tion in the peak hours and every eight minutes there is a train entering a station in the off-peak hours. Nachtigall and Voget
(1996) applied genetic algorithm to optimize regular train schedules by minimizing the passenger waiting times. A heuristic-
based evolutionary approach is proposed in Kwan and Chang (2005) to optimize the frequency (or headway) between trains
to reduce the operation costs and the passenger dissatisfaction. Ceder (2009) proposed a methodology framework to deter-
mine the departure times of train services with even headway and to provide smooth transitions between different time
periods. The train scheduling problem is formulated as a periodic event-scheduling problem based on a graph model in
Liebchen (2006), which is then solved using integer programming methods. The regular train schedules generated by the
approach proposed by Liebchen have been applied in Berlin subway systems (Liebchen, 2008). In addition, a demand-
oriented timetable design is proposed in Albrecht (2009), where the optimal train frequency and the capacity of trains
are first determined and then the schedule of trains are optimized. Moreover, a regular train schedule is optimized together
with the train speed profiles to saving energy by maximize the utilization of regenerative energy in Su et al. (2013) and Li and
Lo (2014). The second type is irregular (or non-periodic) train schedules, which have attracted more and more attentions
from researchers. Cury et al. (1980) presented a hierarchical methodology to generate nonperiodic schedules for metro lines
based on a model of the train movements and of the passenger behavior. Based on the model in Cury et al. (1980) and Assis
and Milani (2004) proposed a model predictive control algorithm to optimize the train schedule, which can effectively gen-
erate train schedules for the whole day. Niu and Zhou (2013) applied genetic algorithms to optimize train schedules for an
heavily congested urban rail transit line. A bi-level approach is proposed in Wang et al. (2014) to obtain the optimal train
schedule for an urban rail transit line with consideration of time-varying passenger demand and stop-skipping. Sun et al.
(2014) proposed three models to design demand-driven train schedules to fully capture the heterogeneity of passenger arri-
val time by minimizing total passenger waiting time. Canca et al. (2014) considered variable demand within a long time per-
iod in the train scheduling model, where train capacity is considered and useful measures of timetable quality were
presented. Barrena et al. (2014) proposed three exact linear formulations and a branch-and-cut algorithm to design train
schedules with dynamic demand. Furthermore, the passenger-demand-oriented train scheduling for an urban rail transit
network is considered in Wang et al. (2015), where the train schedules of two lines are optimized simultaneously to satisfy
the passenger demand and the transfer between different lines are included in the model formulation. Since the regular train
schedules are still widely used in the current practice of the urban rail transit systems, we focus on the regular train sched-
ules in this paper.
Both the train scheduling and train circulation planning have significant impacts on the quality of train operations, thus
collaboratively optimizing these two stages has great potential benefits. Liebchen and Möhring (2007) proposed to integrate
the train circulation into the periodic train scheduling to minimize the number of trains required for the operation. A
discrete-time model with granularity of minutes is proposed by Cadarso et al. (2012, 2013) to integrate the train scheduling
and rolling stock planning for suburban rapid transit networks, where the maximum and minimum frequencies for train ser-
vices are predefined by the line planning process. In addition, Chang et al. (2015) presented an integrated optimization
model for train scheduling and circulation planning, where they assume that there exists an expected train schedule and
the difference between the expected and optimized train schedule are minimized. Furthermore, an integration of line plan-
ning, timetabling, vehicle scheduling is proposed by Michaelis and Schöbel (2009) for bus transit systems, where the route of
each bus is first determined, then the lines and periodic timetables are calculated.
Y. Wang et al. / Transportation Research Part E 105 (2017) 83–104 85
1.2. The proposed integrated approach
The turnaround operations and the entering/exiting depot operation of trains are usually the bottlenecks for urban rail
transit lines and the constraints proposed by the train circulation are critical for the train scheduling. To deal with this issue,
an integrated model that involves train scheduling and train circulation planning is needed to consider the turnaround oper-
ations, the entering/exiting depot operations, and the availability of trains in the depots. Based on the achievements and gaps
in the literature, the main contributions of this paper are summarized as follows:
This study makes a step forward to optimize the train schedule and train circulation plan simultaneously from a given
service pattern generated by the demand analysis and the line planning. More specifically, the service pattern specifies
the partition of the operation period (e.g., peak hours and off-peak hours), the headways for the time periods, and the
number of train services for each time period.
Two mixed integer nonlinear programming models are proposed and then transformed into mixed integer linear pro-
gramming (MILP) models to deliver the train schedule and circulation plan at the same time. In particular, the number
of train services in each time period is not predefined but is also optimized to achieve the pre-specified headway in
the extended integrated model.
In our models, three conflicting objectives (i.e., headway deviations between the train schedule and the service pattern,
the number of entering/exiting depot operations, and the headway variations between consecutive trains) are considered
and the trade-off between them is explored.
The reminder of this paper is organized as follows. The integrated model for the train scheduling and train circulation
planning is formulated in Section 2, where the service pattern, the operation of train services, and the circulation of trains
are described. Section 3 formulates the objective function and the constraints for the integrated optimization problem for the
regular train schedule and circulation plan. By introducing extra binary variables to denote the relationship between train
services and time intervals in the service pattern, Section 4 presents an extended integrated optimization problem, where
the headway deviation between the train schedule and the service pattern are calculated based on the time intervals. In Sec-
tion 5, the integrated optimization problem and the extended integrated optimization problem are transformed into MILP
formulations by applying the transformation properties. In Section 6, the proposed integrated model is applied to a small
case with one headway transition and two large-scale cases for weekdays based on the data of the Beijing Yizhuang line.
Finally, the conclusions and future works are presented in Section 7.
2. Mathematical formulation for the integrated optimization problem
The mathematical model is presented in this section to collaboratively optimize the regular train schedule and circulation
plan. First, the notations and assumptions are introduced. The service pattern generated by the demand analysis and line
planning is then introduced. Finally, the details for the operation of train services and the circulation of trains are formulated.
2.1. Assumptions and notations
The urban rail transit line considered in this paper is a two-track line as given in Fig. 1, where the tracks for the up direc-
tion and down direction1 are almost parallel with each other and the operation of trains in these two directions does not inter-
act with each other in normal operations. Furthermore, the trains in the urban rail transit line do not overtake or meet with each
other in general due to the limited infrastructure settings. As illustrated in Fig. 1, the total number of stations in the urban rail
transit line is denoted by J. Let Sstation denote the set of stations, then we have Sstation ¼ f1; 2; 3; . . . ; Jg. In particular, we only con-
sider one depot for this line and the depot is connected with station 1. In this paper, a train service is defined as follows: a train
arrives at station 1 (whether the train just come out from the depot or just finished a previous train service and turned around at
station 1), goes from station 1 to station J (i.e., running in the up direction), turns around at station J, then goes from station J to
station 1 (i.e., running in the down direction). The turnaround operation of trains at station 1 and the entering/exiting depot
operation of trains are not included in the definition of train services but are considered in the train circulation plan.
2.1.1. Assumptions
For the integrated optimization of the regular train schedule and circulation plan, the operation of trains inside the depot
is out of the scope of this paper since the operation of trains inside the depot is normally scheduled when the train schedule
and the train circulation plan are fixed (Flamini and Pacciarelli, 2008). Furthermore, only one type of train services, i.e., full-
length train services, is considered and all train services stop at all the stations, so the short-turning train services, the dead-
heading train services, the stop-skip train services etc. are not included in this paper. In addition, we make the following
assumptions:
1
Note that the up and down direction in this paper are used to denote the two different operation directions of trains in urban rail transit line. In urban rail
practice, the up and down directions could be used to represent the operation directions toward the north (or west) and south (or east), or vice versa.
Fig. 1. The layout and train services of an urban rail transit line.
A 1. A platform of a station can only accommodate one train at a time and no overtaking can occur at any point in the urban
rail transit line.
A 2. The running times and dwell times at stations are constants but station dependent for all train services.
Assumption 1 generally holds for most urban transit systems, which are usually operated in first-in first-out order from
station 1 to J. With the adoption of automatic train control systems, the operation of trains and the door opening/closing
process are controlled automatically. So the fixed running times and dwell times in Assumption 2 is generally holds for
all the advanced urban rail transit systems for the train services. Moreover, the running times and dwell times are predefined
in general before the train scheduling process in the current practice for most urban rail transit systems.
2.1.2. Notations
Table 1 lists all the parameters and subscripts used in the model formulation. Since the integrated train scheduling and
circulation planning model is built on the basis of train services, departure times, arrival times, running times, and dwell
times are important elements for the train scheduling model. Furthermore, the status of trains that serves train services
is important for the train circulation plan, where the status involves whether the train that serves a train service directly
comes from the depot or not, whether the train that serves a train services enters the depot or not immediately after the
train service finishes, and whether the train serves other train services or not. Moreover, the headway deviation for the
extended integrated model depends on the binary variables that denote the relationship between train services and time
intervals in the service pattern. So the decision variables of the integrated train scheduling and utilization planning model
are given in Table 2.
2.2. Service pattern obtained by the demand analysis and line planning
In urban rail transit systems, the headway between train services is determined by the passenger demand X, the train
capacity C train , and the predefined load factor j, where the train capacity and the predefined load factor are constant param-
eters since the composition of trains does not change during the daily operation. More specifically, the predefined load factor
is introduced to limit the crowdedness of onboard passengers and to guarantee their satisfaction, which is chosen less than 1
in general. For non-peak hours, a possible value for j could be 0.75. This means if the capacity of a train is 1000 passengers,
then the maximum number of passengers on board each train is 750 when calculating the service pattern. For peak hours, j
could be taken as 0.9 or even 1. Because the passenger demand X is a time-varying quantity, the train services should have
different headways at different time periods to cope with the passenger demand. In this paper, we define the results of the
Table 1
Parameters and subscripts for the model formulation.
Notations Definition
½t start ; tend operation period
i; i0 index of train services
j; j0 index of stations
J total number of stations
Itotal total number of train services
Sstation set of stations f1; 2; 3; . . . ; Jg
Sservice set of train services f1; 2; 3; . . . ; Itotal g
½t k1 ; tk k-th time interval for the service pattern
Ik number of train services for time interval ½t k1 ; tk
Hk operation headway for time interval ½t k1 ; tk
up up operation direction
dn down operation direction
rup
i;j running time from station j and the next station in the up direction
rdn
i;j running time from station j and the next station in the down direction
sdn
i;j dwell time at station j in the down direction for train service i
sup
i;j
dwell time at station j in the up direction for train service i
r turn
J;min minimal turnaround time at station J
r turn
J;max maximal turnaround time at station J
rturn
1;min minimal turnaround time at station 1
rturn
1;max maximal turnaround time at station 1
hmin minimal headway between two consecutive train services
hmax maximal headway between two consecutive train services
N train number of available trains for the operation
M a sufficiently large number
a small positive number
tenter time for the entering depot operation of trains
texit time for the exiting depot operation of trains
Table 2
Variables for the model formulation.
Notations Definition
dn
di;j departure time of train service i at station j in the down direction
up
di;j departure time of train service i at station j in the up direction
adn
i;j arrival time of train service i at station j in the down direction
aup
i;j arrival time of train service i at station j in the up direction
ai 0–1 binary variable, if the train for train service i does not come from the depot, ai ¼ 1; otherwise, ai ¼ 0
bi 0–1 binary variable, if the train for train service i does not enter the depot, bi ¼ 1; otherwise, bi ¼ 0
ci;i0 0–1 binary variable, if train services i connects with train service i0 with i < i0 ; ci;i0 ¼ 1; otherwise, ci;i0 ¼ 0
km;i 0–1 binary variable, if the train serving service m has entered the depot and could be used to serve service i; km;i ¼ 1; otherwise, km;i ¼ 0
up
yk;i 0–1 binary variable, if di;1 6 t k ; yk;i ¼ 1; otherwise, yk;i ¼ 0
demand analysis and the line planning as service pattern, which specifies the partition of the operation period (e.g., peak
hours and off-peak hours), the headway and the number of services for each time interval. Note that with Assumption 2
the running times and dwell times are the same for all the train services, so the departure headway and the arrival headway
between consecutive trains are the same at stations. Here, we adopt the departure headway, i.e., the headway calculated by
the departure times.
In this paper, the operation period for the urban rail transit line is denoted as ½t start ; t end . In the service pattern, the oper-
ation period is split into multiple time intervals, which can be presented by ½t0 ; t1 ; ½t1 ; t2 ; . . . ; ½t k ; t kþ1 ; . . . ; ½tK1 ; t K with
t 0 ¼ t start and tK ¼ tend . The number of train services for these time intervals are denoted as I1 ; I2 ; . . . ; Ik ; Ikþ1 ; . . . ; IK , which
are integers. Then the passenger demand that can be handled for time interval ½tk1 ; tk is
~ k ¼ C train j Ik ; 8k 2 f1; 2; . . . ; Kg;
X ð1Þ
~ k should be larger than the actual passenger demand X for time interval ½t k1 ; t k . The corresponding headways are
where X
denoted by H1 ; H2 ; . . . ; Hk ; Hkþ1 ; . . . ; HK , which can be calculated by
t k tk1
Hk ¼ ; 8k 2 f1; 2; . . . ; Kg: ð2Þ
Ik
In addition, these headways should satisfy the constraints as follows
hmin 6 Hk 6 hmax ; 8k 2 f1; 2; . . . ; Kg; ð3Þ
where hmin is the minimal headway between two consecutive train services determined by the infrastructure, rolling stocks,
signaling systems, etc. and hmax is the maximal headway between two consecutive train services introduced to provide a cer-
tain level of services to passengers. Hence, the total number Itotal of train services can be calculated by
X
Itotal ¼ Ik : ð4Þ
k2f1;2;...;Kg
2.3. Operation of train services at stations
Based on the layout of the urban rail transit line shown in Fig. 1, a train service starts from station 1, runs to station J,
turns around at station J, and goes back to station 1. We denote the departure time and arrival time of train service i at sta-
up
tion j as di;j and aup
i;j for the up direction with i 2 Sservice and j 2 Sstation . For the up direction, we have
up
i;j þ si;j ; 8i 2 Sservice ; 8j 2 Sstation ;
di;j ¼ aup up
up ð5Þ
i;jþ1 ¼ di;j þ r i;j ; 8i 2 Sservice ; 8j 2 Sstation =fJg;
aup up
where sup up
i;j is the dwell time for train service i at station j in the up direction and r i;j is the running time between station j to
the next station in the up direction, i.e., station j þ 1. Similarly, the departure and arrival times for the down direction can be
dn
denoted as di;j and adn
i;j , which can be calculated by sdn dn dn
i;j and r i;j in a similar way. In particular, r i;j is the running time from
station j to station j 1. Note that the upper bounds and lower bounds for the running times and dwell times for trains
are not needed due to Assumption 2.
Remark. When the running times and dwell times (i.e., rup up
i;j and si;j ) are considered as variables, if we choose the departure
up up
and arrival times (i.e., di;j and ai;j ) at stations as independent variables, then the running times and dwell times can be
eliminated and represented by the departure and arrival times according to (5). On the contrary, if we choose the running
times and dwell times as independent variables, then the departure and arrival times at stations can be eliminated.
Train services need to turnaround at station J to change from the up direction to the down direction. The turnaround oper-
ation at station J can be described by
up
i;J ¼ di;J þ r i;J ; 8i 2 Sservice ;
adn ð6Þ
turn
where rturn
i;J is the turnaround time for train service i at station J. The turnaround time rturn
i;J is a variable, which should satisfy
rturn
J;min 6 rturn
i;J 6 J;max ;
rturn ð7Þ
where rturn
J;min is the minimal turnaround time at station J determined by track layout, train characteristics, signaling system,
train preparation, simple cleaning, etc., and rturn
J;max is the maximal turnaround time at station J specified by the rail operator. In
this paper, we consider the case that there is only one train could turn around at a time at the turnaround station J. So the
departure and arrival times of two consecutive train services should satisfy the constraints given as follows
up
i;J > 0; 8i 2 Sservice =fItotal g;
diþ1;J adn ð8Þ
which means that train service i þ 1 can start turnaround operation in the up direction only when service i has arrived at the
platform of station J in the down direction. The constraints for the turnaround operations at station 1 will be discussed in
Section 2.4.
The headway between two consecutive train services should be larger than the minimal headway and smaller than the
maximal headway. The train services are increasing indexed according to the departure time at station 1 in the up direction
in this paper. Furthermore, there is no overtaking at anywhere on the urban rail transit line based on Assumption 1, so the
order of train services holds the same for all stations. The headway constraints can be formulated as
up up
hmin 6 diþ1;j di;j 6 hmax ; 8i 2 Sservice =fItotal g; 8j 2 Sstation ;
dn dn
ð9Þ
hmin 6 diþ1;j di;j 6 hmax ; 8i 2 Sservice =fItotal g; 8j 2 Sstation :
According to Assumption 2, the running times and dwell times are the same for all train services. So we only need to consider
the headway constraints at station 1 for the up direction and the headway constraints at station J for the down direction. The
other headway constraints given in (9) can be eliminated.
2.4. Circulation of trains
Based on the layout of the urban rail transit line shown in Fig. 1, a train service could be served by a train that comes out
the depot directly or by a train that just finishes a previous train service and turns around at station 1. Binary variables ai
with i 2 Sservice are defined as follows to describe whether the train for train service i comes out from the depot or not

0 the train comes out from the depot;
ai ¼ ð10Þ
1 the train does not come out from the depot:
In addition, after the train has served train service i, the train could go back to the depot or turnaround at station 1 to serve
another train service later on. Binary variables bi with i 2 Sservice are introduced to denote whether the train for service i goes
back to the depot or not after the service

0 the train goes back to the depot;
bi ¼ ð11Þ
1 the train does not go back to the depot:
If the train does not go back to the depot after serving train services i, i.e., bi ¼ 1, this train will run another train service,
0 0
which is denoted by i with i > i. We then have ai0 ¼ 1. Binary variables ci;i0 are introduced to denote whether they are served
by the same train
(
0
0 train services i and i are not served by the same train;
ci;i0 ¼ 0 ð12Þ
1 train services i and i are served by the same train:
Note that the train here is not corresponding to a physical train, but represents that a physical train comes out from the
depot, runs on the urban rail transit line, and goes back to the depot. So a physical train could corresponds to several trains
here since the physical train could go out and back to the depot several times during the operation period.
Based on the definition of ai ; bi , and ci;i0 , the following two constraints should be satisfied
X
ai0 ¼ ci;i0 ; ð13Þ
i2Sservice and i<i0
X
bi ¼ ci;i0 ; ð14Þ
i0 2Sservice and i0 >i
0 0
where (13) means the train serving service i comes out from the depot or not depends on whether service i is served by the
same train with the previous services or not, and (14) means that the train serving service i goes back to the deport or not
depends on whether there exists a later train service that is served by the same train for service i. Furthermore, the following
constraints should be satisfied
ai0 þ bi 6 2 þ Mð1 ci;i0 Þ;
ð15Þ
ai0 þ bi P 2 Mð1 ci;i0 Þ;
0
where M is a sufficiently large number. (15) means that if train services i and i are served with the same train, i.e., ci;i0 ¼ 1,
0
then this train will not go back to depot after serving service i and the train for service i does not come from the depot, i.e.,
bi ¼ 1 and ai0 ¼ 1. In case of ci;i0 ¼ 0, (15) will be satisfied automatically.
0 0
Now we consider the turnaround operation for trains at station 1. For train services i and i with i < i , the arrival and
departure times at station 1 should satisfy
dn
aup
i0 ;1
di;1 P r turn
1;min Mð1 ci;i0 Þ;
dn
ð16Þ
aup
i0 ;1
di;1 6 r turn
1;max þ Mð1 ci;i0 Þ;
1;min and r 1;max are the minimal and maximal turnaround time at station 1, respectively. When ci;i ¼ 0, i.e., train ser-
where r turn turn 0
0
vices i and i are not served with the same train, then the constraints (16) will be satisfied automatically. When ci;i0 ¼ 1, i.e.,
0
train services i and i are served with the same train, then the turnaround time of this train should be between rturn turn
1;min and r 1;max
as given in (16). Furthermore, the departure and arrival times of consecutive train services should satisfy the turnaround
constraints at station 1
dn
X
diþ1;1 ci;i0 aup
i0 ;1
> 0; 8i 2 Sservice =fItotal g: ð17Þ
0 0
i 2Sservice and i >i
which involves the following two situations:
when the train serving train service i does not go back to the depot and ci;i0 is equal to 1, then this train also serving train
0
service i . So the train for train service i þ 1 can only depart from the platform in the down direction when the train for
0
train service i has arrived at the platform in the up direction, i.e.,
dn
diþ1;1 aup
i0 ;1
> 0; 8i 2 Sservice =fItotal g and ci;i0 ¼ 1:
0 0
when the train serving train service i goes back to the depot, i.e., ci;i0 ¼ 0 for all i 2 Sservice and i > i, then constraint (17)
dn
will be satisfied automatically since the departure time diþ1;1 is larger than 0.
In practical operation of an urban rail transit line, the number of trains available for the daily operation is limited and is
denoted by N train . In this paper we only consider one depot, so all the trains for the operation are in the depot before the daily
operation and all of them also need to come back to the depot after the operation. When ai ¼ 0, i.e., the train for train service
i comes out from the depot, we need to check whether there is still a train inside the depot for the departure of train service i.
The total number of exiting operations for the depot until the departure of train service i at station 1 in the up direction can
be calculated as
X
i
ð1 am Þ:
m¼1
Furthermore, the total number of entering operations for the depot until the departure of train service i depends on the
up dn
departure time di;1 of train service i and the departure time dm;1 for m ¼ 1; 2; . . . ; i 1. If bm ¼ 0 and
dn up
dm;1þ t enter 6 t exit for m ¼ 1; 2; . . . ; i 1, then the train for train service m has entered the depot and could be used to
di;1
serve train service i. Note that t enter and texit are the entering time from station 1 to the depot and the exiting time from
the depot to station 1, respectively. We introduce binary-valued variables km;i and define it as
dn up
½km;i ¼ 1 () ½dm;1 di;1 þ t enter þ t exit 6 0; ð18Þ
then the total number of entering operations for the depot until the departure of train services i can then be calculated by
X
i1
ð1 bm Þkm;i :
m¼1
For each train service, the following constraint should be satisfied

X
i X
i1
ð1 am Þ ð1 bm Þkm;i 6 Ntrain ; ð19Þ
m¼1 m¼1
which means that the difference between the total number of exiting operations and the total number of entering operations
should be less than the number of available trains. Note that there is only one depot for the urban rail transit line and all the
available trains should be parked in this depot; so the capacity of the depot can also be represented by the number of avail-
able trains N train .
3. Integrated optimization problem for regular train schedule and circulation
3.1. Objective function
The performance of the train schedule and circulation plan depends on many factors, such as the consistency between the
service pattern and the train schedule, the number of trains required, the entering/exiting depot operations of trains, and the
travel time of passengers. Hence, the integrated problem formulated in this paper is a multi-objective optimization problem,
the objectives of which are explained as follows.
In the service pattern given in Section 2.2, there is a predefined headway for each time interval of the operation period
based on the passenger demand. One objective of the integrated optimization problem for regular train schedule and circu-
lation plan is to minimize the headway deviations between the train schedule and the service pattern, which can be formu-
lated as
Pk
X
K X I
‘¼1 ‘
up up
f obj;1 ¼ ðdiþ1;1 di;1 Þ Hk ; ð20Þ
k¼1 Pk1
i¼ I‘ þ1
‘¼1
where the operation period is split into K time intervals and the headway for the k-th interval is Hk . Note that the headway
between the two train services for the transition of two time periods is considered the same as the headway of the first inter-
val here. When compared with the departure disparity considered in Chang et al. (2015), minimizing the headway differ-
ences between the train schedule and the service pattern could eliminate the optimization process from the service
pattern to the expected train schedule.
For the integrated optimization model, the total number of train services is determined in the service pattern. So the qual-
ity of the train schedule and the circulation plan is determined by the number of trains required and the number of entering/
exiting depot operation of trains. In order to minimize the number of required trains and depot operations, we need to max-
imize the number of train services that served by the same train, i.e.,
0
X
f obj;2 ¼ ci;i0 ;
i;i 2Sservice ;i<i0
0
which is equivalent to minimize

X
f obj;2 ¼ ci;i0 : ð21Þ
i;i0 2Sservice ;i<i0
As stated in Cury et al. (1980), the minimum passenger waiting time is achieved when the headway variations between
consecutive train services are minimized. Therefore, we also include the headway variations between consecutive train
services in the performance criterion of the integrated optimization problem. The headway variation is defined as the
absolute value of the difference between the current headway and the mean of the neighboring headways, which can be
formulated as
Pminðiþi2 ;IÞ
X up up
up up m¼maxðii1 ;2Þ ðdm;j dm1;j Þ
f obj;3 ¼ ðdi;j di1;j Þ ; ð22Þ
i2Sts =f1g;j2Ssta
minði þ i2 ; IÞ maxði i1 ; 2Þ
where i1 and i2 is the number of neighboring train services before and after train service i. In particular, (22) can be simplified
and rewritten as

X ðdminðiþi2 ;IÞ;j dmaxðii1 ;2Þ1;j Þ
up up
up up
f obj;3 ¼ ðdi;j di1;j Þ : ð23Þ
i2Sts =f1g;j2Ssta
For the resulting multi-objective optimization problem, the linear weighted method can be applied to deal with these
three objectives. The objective function for the integrated problem can be written as
f obj;1 f obj;2 f obj;3
f obj ¼ w1 þ w2 þ w3 ; ð24Þ
f obj;1;nom f obj;2;nom f obj;3;nom
where w1 ; w2 , and w3 are the weights to denote the importance of these three objectives, f obj;1;nom ; f obj;2;nom , and f obj;3;nom are
the nominal values for the headway deviation from the service pattern, the number of two train services that are served by
the same train, and the headway variation penalty. Since the objective values of f obj;1 ; f obj;2 , and f obj;3 have different orders of
magnitudes, the normalization factors f obj;1;nom ; f obj;2;nom , and f obj;3;nom are introduced to make these objective values in the
same order of magnitudes and then to be comparable with each other. The normalization factors can be estimated based
on the feasible trains schedules and circulation plans. For this objective function, if w1 is set as a larger value than w2
and w3 , it means that the rail operator would like to pay more attention to the headway deviation between the train schedule
and the service pattern. On the contrary, if w1 is set as a smaller value than w2 and w3 , then the headway deviation will be
emphasized less. Similarly, the value of w2 and w3 could be tuned according to the preferences of the rail operator.
3.2. Constraints
The constraints of the integrated optimization problem of regular train schedule and circulation plan include the
departure and arrival constraints (5), turnaround constraints (6)–(8) at station J, headway constraints (9), train circulation
constraints (13)–(15), turnaround constraints (16) and (17) at station 1, and limited number of available trains (18) and
(19). In addition, we have
up
t start 6 di;j 6 t end ; t start 6 aup
i;j 6 t end ;
dn
ð25Þ
t start 6 di;j 6 t end ; t start 6 adn
i;j 6 t end ;
for all i 2 Sservice and j 2 Sstation . Furthermore, the departure times of the first train service and the last train service at station 1
should be equal to the predefined departure times, i.e.,
up up ;
d1;1 ¼ d
up up ;
dItotal ;1 ¼ d ð26Þ
1;1 Itotal ;1
up and d
where d up
1;1 Itotal ;1 are the predefined departure times.
4. Extended integrated optimization problem for regular train schedule and circulation
The objective function f obj;1 defined in (20) is calculated based on the number of train services Ik in time period ½t k1 ; t k for
k 2 f1; 2; . . . ; Kg as illustrated in Fig. 2. In order to achieve different headways for the time periods, more trains should be put
into operation or some trains should be removed from the operation. Here, we call the process of adding/removing trains as
the transition process of headways. Since adding/removing trains needed to be coordinated with the turnaround operation of
other trains, the transition process of headways cannot be done immediately but requires a certain time. Therefore, the pre-
defined number of train services in each time period may be larger or smaller than the actual number of train services
required, which could result a time delay to the transition process of headways as shown in Fig. 2 if we fix the number of
train services for each time period.
Fig. 2. The transition process between different headways and the delay caused by the integrated optimization model.
In this section, we extend the integrated optimization problem for regular train schedule and circulation plan proposed in
Section 3 by not specifying the number of train services for each time period but only considering the headways in the ser-
vice pattern. Then the number of train services in each time period is decided by the optimization process in order to achieve
the pre-specified headways. The total number of train services in the regular train schedule could be Itotal given in the service
pattern or an integer value close to Itotal . Note that the operation cost increases with the total number of train services. Fur-
thermore, the increase of the total number of train services may require more trains for the operation, which will also
increase the operation cost. Here, for the simplicity of the notations, we also use Itotal to denote the total number of train
services in the extended integration optimization problem. As stated in Section 2.2, the operation period ½t start ; tend is divided
into K time intervals, denoted by ½tk1 ; tk with k 2 f1; 2; . . . ; Kg. Here, binary variables yk;i with k 2 f1; 2; . . . ; Kg and i 2 Sservice
are introduced to represent whether train service i is in time interval ½tk1 ; tk or not. The definition of binary variable yk;i is
given as follows
up
½yk;i ¼ 1 () ½di;1 6 t k ; 8k 2 f1; 2; . . . ; Kg; 8i 2 Sservice ; ð27Þ
up
which means if the departure time di;1 of train service i at station 1 in the up direction is less than and equal to t k , then yk;i is
equal to 1; otherwise, yk;i is equal to 0. As can be observed from Fig. 3, binary variables y‘;i with ‘ P k are equal to 1 and
binary variables y‘;i with ‘ < k are equal to 0 for time period ½tk1 ; tk . Based on these binary variables, the expected headway

h between train services i and i 1 can be expressed by
i1;i
i1;i ¼ y H1 þ ð1 y Þy H2 þ ð1 y Þy H3 þ . . . þ ð1 y
h 1;i 1;i 2;i 2;i 3;i K1;i ÞyK;i HK ; 8i 2 Sservice =f1g; ð28Þ
i1;i depends on the values of the binary variables y for k 2 f1; 2; 3; . . . ; Kg and these binary variables are determined
where h k;i
up
by the departure time di;1 of train service i at station 1 in the up direction.
In the extended integrated optimization problem, instead of (20) the headway deviation between the train schedule and
the service pattern are formulated as follows
Itotal
X
E up up :
f obj;1 ¼ ðdi;1 di1;1 Þ hi1;i ð29Þ
i¼2
In addition, the objective function for the extended optimization problem can be written as
E
E f obj;1 f obj;2 f obj;3
f obj ¼ w1 E
þ w2 þ w3 : ð30Þ
The constraints for the extended integrated optimization problem are the same as those defined in Section 3.2.
5. Solution approach – MILP
The integrated optimization problem formulated in Section 3 and the extended integrated optimization problem in Sec-
tion 4 can be transformed into mixed integer linear programming (MILP) problems. To start with this MILP transformation,
we introduce three properties according to Bemporad and Morari (1999):
Fig. 3. The definition of binary variables for train services and time intervals.
P.1 Consider the statement ~f ð~

xÞ 6 0 , where ~f : Rn ! R is affine, ~
x 2 v with v Rn and Let
~f max ¼ max~f ð~xÞ; ~f ~ xÞ:
min ¼ minf ð~ ð31Þ
~x2v ~x2v
If we introduce the logical variable d 2 f0; 1g, then the following equivalence holds:
(
~f ð~xÞ 6 ~f max ð1 dÞ;
½~f ð~xÞ 6 0 $ ½d ¼ 1 is true iff ð32Þ
~f ð~xÞ P þ ð~f Þd;
min
where e is a small positive number (typically the machine precision) that is introduced to transform a strict equality into a
non-strict inequality, which fits the MLD and MILP frameworks (Bemporad and Morari, 1999).
P.2 The product of two logical variables d1 d2 can be replaced by an auxiliary logical variable d3 ¼ d1 d2 , i.e.
½d3 ¼ 1 $ ½d1 ¼ 1 ^ ½d2 ¼ 1, which is equivalent to
8
< d1 þ d3 6 0;
>
d2 þ d3 6 0; ð33Þ
>
:
d1 þ d2 d3 6 1:
P.3 The product of a logical variable d 2 f0; 1g and a real-valued variable ~f , i.e., d~f ð~
xÞ, can be replaced by the auxiliary real-
valued variable ~z ¼ d~f ð~
xÞ, which satisfies ½d ¼ 0 ) ½~z ¼ 0 and ½d ¼ 1 ) ½~z ¼ ~f ð~xÞ. Then ~z ¼ d~f ð~
xÞ is equivalent to
8
>
> ~z 6 ~f max d;
>
>
< ~z P ~f d;
min
ð34Þ
> ~z 6 ~f ð~xÞ ~f min ð1 dÞ;
>
>
>
:
~z P ~f ð~xÞ ~f max ð1 dÞ:
For the integrated optimization problem formulated in Section 3, the nonlinear constraints (17)–(19) should also be trans-
formed into linear constraints to obtain the MILP formulation. In addition, the objective function involves absolute functions
and nonlinear terms, which need to be transformed into linear functions with linear constraints. The details of the transfor-
mation are given as follows. For the nonlinear constraint (17), the nonlinear term ci;i0 aup
i0 ;1
can be replaced by a auxiliary real-
valued variable zi;i0 . According to transformation property P.3, zi;i0 ¼ ci;i0 aup
i0 ;1
is equivalent to
8
> zi;i0 6 t end ci;i0 ;
>
>
>
< zi;i0 P tstart ci;i0 ;
ð35Þ
>
> zi;i0 6 aup
i0 ;1
t start ð1 ci;i0 Þ;
>
>
: z 0 P aup t ð1 c 0 Þ:
i;i i0 ;1 end i;i
where tstart and t end are the minimal and maximal values for aup
i0 ;1
according to constraint (25). Then constraint (17) can be
rewritten as
dn
X
diþ1;1 zi;i0 > 0; 8i 2 Sservice =fItotal g: ð36Þ
i 2Sservice and i0 >i
0
Furthermore, the definition of km;i is given in constraint (18), which is equivalent to the following linear constraints according
to transformation property P.1
8
e km;i Þ;
< ddn dup þ tenter þ texit 6 Mð1
m;1 i;1
ð37Þ
: d d þ tenter þ texit P þ ðm
dn up
~ Þkm;i ;
m;1 i;1
e and m
where M ~ are the maximal and minimal values for dm;1 di;1 þ t enter þ texit . The maximal and minimal values of dm;1 and
dn up dn
up
di;1
are t end and t start . In addition, the maximal and minimal values of tenter and t exit can be obtained by the operation practice,
e and m
which are denoted by t enter; max ; t enter; min ; t exit; max , and t exit; min . Then M ~ can be chosen as
(
e ¼ t enter; max þ texit; max ;
M
ð38Þ
~ ¼ tstart tend þ t enter; min þ t exit; min :
m
In addition, constraint (19) contains the product of two binary-valued variables, i.e., bm km;i , which can be replaced by an aux-
iliary binary-valued variable /m;i . Then /m;i ¼ bm km;i is equivalent to the following constraints according to transformation
property P.2
8
< bm þ /m;i 6 0;
>
km;i þ /m;i 6 0; ð39Þ
>
:
bm þ km;i /m;i 6 1:
With this replacement, nonlinear constraint (19) can be rewritten as the following linear constraint
X
i X
i1
ð1 am Þ ðkm;i /m;i Þ 6 Ntrain : ð40Þ
m¼1 m¼1
For the objective function of the integrated optimization problem formulated in Section 3, the objective functions given in
(20) and (23) are described by the absolute functions. New auxiliary real-valued variables ni;k and gi;j are introduced to han-
dle these two absolute functions, where are specified by

up up
ni;k P ðdiþ1;1 di;1 Þ Hk ; ð41Þ
and

ðdminðiþi2 ;IÞ;j dmaxðii1 ;2Þ1;j Þ
gi;j P ðdi;j di1;j Þ : ð42Þ
The constraints (41) and (42) are equivalent to the following constraints
up up
ni;k P ðdiþ1;1 di;1 Þ Hk ;
up up ð43Þ
ni;k P ðdiþ1;1 di;1 Þ þ Hk ;
and
ðdminðiþi2 ;IÞ;j dmaxðii1 ;2Þ1;j Þ
gi;j P ðdi;j di1;j Þ ;
ð44Þ
ðd dmaxðii1 ;2Þ1;j Þ
gi;j P ðdi;j di1;j Þ þ minðiþi2 ;IÞ;j :
Since w1 and w3 in (24) are positive, so minimizing (24) is equivalent to minimize
~f f obj;2 ~f
obj;1 obj;3
f obj ¼ w1 þ w2 þ w3 ; ð45Þ
subject to constraints (43) and (44), where

Pk
I
X
K X‘¼1 ‘ X
~f ni;k and ~f obj;3 ¼
obj;1 ¼ gi;j :
k¼1 Pk1 i2Sts =f1g;j2Ssta
i¼ I‘ þ1
‘¼1
Indeed, it is easy to verify that when we minimize the objective function (45) subject to (43) and (44), the optimal value of

up up ðdminðiþi ;IÞ;j dmaxðii ;2Þ1;j Þ
ni;k and gi;j will be equal to j ðdiþ1;1 di;1 Þ Hk j and ðdi;j di1;j Þ minðiþi 2 1
2 ;IÞmaxðii1 ;2Þ
, respectively, so the original objective
function (24) will also be minimized.
For the extended integrated optimization problem formulated in Section 4, the constraints are the same as these of the
integrated optimization problem. However, the objective function of the extended integrated optimization involves binary
variable yk;i for k 2 f1; 2; . . . ; Kg and i 2 f1; 2; . . . ; Itotal g. The definition of yk;i given in (27) can be transformed into linear con-
straints based on transformation property P.1 similarly as the definition of km;i . In addition, the product of binary variables
yk1;i yk;i in (28) can be replaced by newly introduced binary-valued variables similar as the product bm km;i by applying trans-
formation property P.2. The absolute functions in the objective function (29) of the extended integrated problem can be dealt
in the same way as these in the integrated optimization problem.
After the transformations given as before, the integrated optimization problem and the extended integrated optimization
problem formulated in Section 3 and 4 can be transformed into MILP problems, which can be solved by several existing com-
mercial and free solvers, such as CPLEX, Xpress-MP, GLPK (see e.g. Linderoth and Ralphs, 2005; Atamturk and Savelsbergh,
2005).
6. Case study
In order to demonstrate the performance of the integrated optimization model and the extended integrated model for
regular train schedules and circulation plans, we use the data of the Beijing Yizhuang line as a case study. The resulting MILP
problems are solved by the CPLEX implemented through the cplex interface function of the Matlab OPTI toolbox. All the
experiments are performed on a 3.6 GHz Intel Core i7-3520 M CPU running with a 64-bit windows operating system and
16G RAM.
The Beijing Yizhuang line was put into operation on December 30, 2010 and the daily ridership now is around 200,000.
This line has 14 stations and the length is 23.23 km. The layout of this line is shown in Fig. 4. The operation direction from
Yizhuang to Songjiazhuang is defined as the up direction and the operation direction from Songjiazhuang to Yizhuang is
defined as the down direction. There is one depot which connects with Yizhuang station. The running times and dwell times
used in this study are the same as these for the current train schedule in practice, where the running and dwell times are
determined by the current train control systems and the practical experiences. The turnaround time at station Songjiazhuang
and Yizhuang should be larger than 120 s and smaller than 720 s according to the practical requirements. Furthermore, the
minimal and maximal headway between train services are set as 150 s and 660 s, respectively.
Three cases will be considered here:
Case A: train schedule and circulation plan for two time intervals.
Case B: train schedule and circulation plan for a weekday in the current practice.
Case C: train schedule and circulation plan for a weekday with much smaller headways.
6.1. Case A: Two time intervals
6.1.1. Set-up for Case A

In Case A, the operation period considered is from 5:20 to 10:30, where the first and last train services depart from Yiz-
huang station at 5:20 and 9:00, respectively. The service pattern involves two time intervals as shown in Table 3, where the
service headway between 5:20 and 7:00 is 600 s, the headway between 7:00 to 9:00 is 390 s, and the total number of train
services is 30. Note that the start time and end time for the service pattern are the departure times of the first and last train
services at station 1, i.e., Yizhuang station, respectively. Since the running times and dwell times are the same for all train
services, the headway between train services is the same for the up direction and the down direction, respectively. However,
because the turnaround time at station J, i.e., Songjiazhuang station, varies between the minimal and maximal turnaround
time, the headways between train services for the up direction and the down direction could be slightly different from each
other. In order to choose appropriate values for the normalization factors f obj;1;nom ; f obj;2;nom , and f obj;3;nom , we first solve the
Fig. 4. The layout of the Beijing Yizhuang line.
Table 3
The service pattern of the Beijing Yizhuang line for Case A.
Time interval Headway [s] Number of train services

5:20–7:00 600 11
7:00–9:00 390 19
integrated optimization problem as single objective optimization problems by setting the normalization factors as one and
setting one of the weights (i.e., w1 ; w2 , and w3 ) as one and the others as zero. The minimal values of f obj;1 ; f obj;2 , and f obj;3 can
then be obtained and are equal to 420, 19, and 91.375. So in this case study the normalization factors f obj;1;nom ; f obj;2;nom , and
E
f obj;3;nom are taken as 420, 19, and 91.375, respectively. Moreover, the nominal value of f obj;1;nom in (30) is also taken as 420.
With the introduction of these normalization factors, the three components of the objective function are more or less in the
same order of magnitudes, the weights w1 ; w2 and w3 can be specified to indicate the relative importances of the compo-
nents. In this case study, the weights w1 ; w2 and w3 are all taken the value 1, which means these three objectives have equiv-
alent importance. In addition, the effect of different weights on the objective function is also evaluated in Case A.
6.1.2. Results for Case A

In Case A, the headway between train services changes from 600 s to 390 s and the headway transition occurs at 7:00. The
performance comparison of the integrated model and the extended integrated model is given in Table 4, where the number
of train services for the integrated model is 30 according to the service pattern given in Table 3 and the number of train ser-
vices for the extended model is taken the values between 28 and 31. As can be observed from Table 3, the computation times
for both the integrated model and the extended integrated model are less than 1 s when using the CPLEX solver for the MILP
problems. In particular, the number of variables is about 650 and the number of constraints is around 2300 for Case A after
the elimination of dependent variables and redundant constraints. When compared with the performance of the integrated
model with 30 services, the extended integrated models with 29 and 30 services have better performance, where the objec-
tive function values are smaller. This is because the feasible set of the extended integrated model is larger than that of the
integrated model due to the flexible number of train services in each time period of the extended integrated model. However,
the performance of the extended models with 28 and 31 services is worse than that of the integrated model with 30 services.
In particular, 28 train services are not enough to reach headway 390 s; so the headway deviations is much larger. Further-
more, 31 train services are too much to reach the specified headway and they result in smaller headways and then bigger
headway deviations.
The headway between train services obtained by the integrated model and the extended integrated model is illustrated in
Fig. 5. The headway transition process for the integrated model with 30 service starts before 7:00 and the headway reaches
390 s at 7:09. In addition, the extended models with 29, 30, and 31 services also start before 7:00 and the headways of these
three models reach 390 s at 7:29, 7:09, and 7:00, respectively. In particular, for the extended model with 31 services, since
there is one more services when compared with the service pattern given in Table 3, some of the headways between 7:00
Table 4
Performance comparison of the integrated optimization model and the extended integrated model for Case A.
Model Number of Computation Objective Headway Headway Number of Number of Relative

services time [s] value deviations [s] variations [s] connected services required trains gap
Integrated 30 0.42 1.552 420 137 18 12 0.00%
28 0.34 1.741 570 112 16 12 0.00%
Extended 29 0.52 1.415 308 144 17 12 0.00%
Integrated 30 0.64 1.338 330 137 18 12 0.00%
31 0.64 2.035 600 142 18 13 0.00%
Fig. 5. The headway of the train schedule with two time intervals for Case A obtained by the integrated model and the extended integrated model.
and 9:00 are less than the predefined headway 390 s and the number of required trains 13 instead of 12. The headway tran-
sition process for the extended integrated model with 28 services starts after 7:00 and the headway of the extended model
with 28 services reaches 390 s at 7:55. Note that the headway that reaches 390 s after 7:00 could result in passenger crowd-
edness on board trains and at platforms because the service pattern is calculated based on the passenger demand.
The train schedule and circulation plan obtained by the integrated model with 30 services are shown in Figs. 6 and 7,
where 9 trains come from the depot directly to serve services 1 to 9 and these 9 trains are sufficient for the headway
600 s in [5:20,7:00]. In order to satisfy the headway 390 s in [7:00, 9:00], three trains come out from the depot at 7:10,
7:29, and 7:49 to serve service 13, 16, and 19, respectively. These trains could turn around at station 1 to service other train
services or go back to the depot. As can be observed from Fig. 7, most of the trains serve 2 or 3 services before going back to
the depot and only the train for service 19 serves only one service and then goes back to the depot. Based on the train sched-
ule and the train circulation plan, the train operation inside the depot could then be scheduled.
In order to evaluate the influence of different weights in the objective function on the optimal solutions and computation
times, we set two weights as 1 and set the other weight as 0.1, 0.5, 1, 10, 50, and 100 for the extended optimization model
with 30 train services. The solutions we obtained for all these weight setting are integer optimal, i.e., the relative gaps are
equal to zero. The detailed computation results are given in Table 5, where the headway deviations, headway variations,
number of connected services and computation times for different weight settings are listed. As can be observed from
Table 5, the computational time of the CPLEX solver finding the optimal solution is influenced by the values of these three
Fig. 6. The train schedule with two time intervals obtained by the integrated model for Case A.
Fig. 7. The train circulation plan with 30 train services obtained by the integrated model for Case A.
Table 5
Optimal objectives and computation times with different weights.
w1 w2 w3 Computation time [s] Headway deviations [s] Headway variations [s] Number of connected trains
1 1 1 0.56 330 137 18
0.1 1 1 0.95 1033 78 18
0.5 1 1 0.56 518 107 18
10 1 1 0.53 210 206 18
50 1 1 0.83 210 206 18
100 1 1 0.53 210 206 18
1 0.1 1 0.67 330 137 18
1 0.5 1 0.55 330 137 18
1 10 1 1.00 330 137 18
1 50 1 1.81 1393 115 19
1 100 1 4.17 2643 562 20
1 1 0.1 0.52 210 206 18
1 1 0.5 0.72 210 206 18
1 1 10 0.89 1033 78 18
1 1 50 0.73 1461 65 17
1 1 100 0.84 1461 65 17
coefficients. The computation time is more than 1 s for the two cases with setting (1, 50, 1) and (1, 100, 1). The minimal value
for the headway deviations is 210 s for the extended optimization with 30 train services. So when the headway deviations is
larger than 210 s, this value decreases with the increasing of w1 . However, when this optimum is reached, the optimal solu-
tion does not change with the increasing of w1 . When w2 is assigned with a larger value, more emphasis is put on the con-
nection of train services. So with the increasing of w2 , the number of train connections grows from 18 to 20. In particular, the
maximum number of train connections is 20, i.e., the minimum value of f obj;2 is 20, for the extended model with 30 train
services. However, with the increasing of the number of train connections, the headway deviations and headway variations
also increase. Similarly, when w3 increases, the value of f obj;3 decreases but the values of f obj;1 and f obj;2 increase.
Remark. Note that there is only one depot for the considered urban rail transit line and the depot is connected with
Yizhuang station. So in the beginning of the operation period, trains should come out the depot and start the operation from
Yizhuang station. Moreover, trains should finish the operation at Yizhuang station and go back to the depot at the end of the
operation period. Therefore, the train schedule given in Fig. 5 provides only one direction train services in the beginning and
at the end of the operation period. This also holds for the train schedule obtained by Case B because the Beijing Yizhuang line
does not operate for the 24 h a day, but starts its service at 5:20 and finishes its service at 23:20.
6.2. Case B: Weekdays in the current practice
6.2.1. Set-up for Case B

For Case B, we consider the generation of a daily regular train schedule and circulation plan. The daily operation period of
the Beijing Yizhuang line is from 5:20 to 23:30. In particular, the first train service departs from Yizhuang station at 5:20 and
the last train service departs from Yizhuang station at 22:05. These two departure times cannot be changed in general to
provide a consistent service to passengers; so these requirements are considered as hard constraints for the first train service
and the last train service in the integrated model. Based on the passenger demand analysis of the Beijing Yizhuang line in the
weekdays, the service pattern of weekdays used in the current practice is illustrated in Table 6, where the operation period
between the first train service and the last train service is split into 6 time intervals and the total number of train services is
122. The normalization factors f obj;1;nom ; f obj;2;nom , and f obj;3;nom for Case B are chosen in a similar way as these in Case A and are
E
taken as 260, 106, and 1143, respectively. In addition, the nominal value of f obj;1;nom in (30) is also taken as 260. Similar as
Case A, the weights w1 ; w2 , and w3 are all taken the value 1.
6.2.2. Results for Case B

As stated before, Table 6 gives the service pattern for Case B, which involves 122 train services, 6 time periods, and 5
headway transitions. The headways for the morning and evening peak hours are 390 s and 350 s, respectively. The perfor-
mance of the train schedule and circulation plan obtained by the integrated model and the extended integrated model is
compared with that of the train schedule and circulation plan used in practice. The performance comparisons are illustrated
in Table 7. In this case study, the number of train services for the integrated model is 122 and the number of train services for
the extended integrated model is taken the value from 119 to 123 as given in Table 7. The headways between train services
for the practical train schedule and the train schedules with 122 train services obtained by the integrated model and the
extended integrated model are given in Fig. 8. In addition, the train circulation plan of the practical train schedule and
the circulation plan with 122 train services obtained by the extended integrated model are given in Figs. 9 and 10, respec-
Table 6
The service pattern of weekdays for the Beijing Yizhuang line for Case B.

5:20–5:30 535 2
5:30–9:00 390 33
9:00–16:00 660 39
16:00–19:00 350 30
19:00–20:00 540 7
20:00–22:05 660 11
Table 7
Performance comparison of the integrated optimization model and the extended integrated model for Case B.

services time [s] value deviations [s] variations [s] connected services required trains gap (%)
Practical 122 – 8.560 2086 1714 102 13 –
Integrated 122 15.98 1.381 260 1557 104 14 0.00
119 88.08 2.305 555 1284 101 13 0.00
Extended 120 100.15 1.167 205 1533 102 13 0.00
Integrated 121 186.31 0.740 75 1627 103 14 0.00
122 106.11 0.921 155 1493 104 14 0.00
123 149.95 2.208 545 1260 105 14 0.00
Fig. 8. The headway for the practical train schedule, and the train schedules with 122 train services obtained by the integrated model and the extended
integrated model.
tively. Furthermore, the headways of the train schedules obtained by the extended integrated model with 119–123 train ser-
vices are shown in Fig. 11.
Note that the current train schedule and circulation plan used in practice are not generated automatically by algorithms
but are designed by the experienced planners. When we use the objective function defined in Section 3.1 to evaluate the
practical train schedule and circulation plan, the calculated objective function value is 8.056 as given in Table 7. The sum
of headway deviations is 2086 s and the sum of headway variations is 1714 s. Moreover, the headway between train services
in the practical train schedule is demonstrated by the blue line with circle in Fig. 8, where the headways of the practical train
schedule are shifted between 290 s and 410 s around 17:00. This is caused by the exiting depot operation of trains, i.e., four
trains come out the depot to serve services 79, 80, 82, and 85 around 17:00 as shown in Fig. 9. Furthermore, the number of
required trains is 13 and the entering/exiting depot operations for this train schedule is 20 as can be observed in Fig. 9.
The objective function values of the integrated model and extended integrated model are smaller when compared with
that of the practical train schedule. Moreover, the smallest objective function value is obtained by solving the extended inte-
grated model with 121 train services similar as Case A. Note that for the results given in Table 7, the constraints (18) and (19)
for the number of available trains are not taken into account, but it is assume that there are sufficient trains in the depot for
Fig. 9. The train circulation plan with 122 train services used in the current practice of the Beijing Yizhuang line.
Fig. 10. The train circulation plan obtained by the extended integrated model with 122 train services for the Beijing Yizhuang line.
Fig. 11. The headway of the train schedules obtained by the extended integrated model with 119–123 train services for Case B.
the train schedule. The number of required trains for the extended integrated models with 119 and 120 services is 13, which
is equal to that of the practical train circulation plan. For the other extended integrated models and the integrated model
with 122 services, the number of required trains is 14. This is because the headways for the extended model with 119
and 120 services keep at 350 s shorter when compared with the other cases. In particular, the headways for the extended
models with 119 and 120 services reach 350 s at 17:22 and 17:12 and stay at 350 s until 18:33 and 18:51, respectively. How-
ever, the headways for the extended models with 121, 122, and 123 keep at 350 s much longer. More specifically, as can be
observed from Fig. 8, the headways of the train schedules obtained by the integrated model and extended model with 122
train services reach 350 s at 16:24 and 16:00, while the headway of the practical train schedule reaches 350 s at 17:13. This
also explains why one more train is needed for the integrated model and the extended integrated model with 122 train ser-
vices. In addition, due to the newly introduced variables and constraints for the objective function, the computation time for
the extended integrated model is longer than that of the integrated model. In particular, the number of variables and the
number of constraints for the extended optimization model are about 4300 and 14500, while these numbers for the inte-
grated optimization model are around 3000 and 11000.
When comparing the train circulation plans with 122 train services in Figs. 9 and 10, the number of trains required for the
morning peak hours is 12 for both circulation plans. The number of trains for the afternoon peak hours is also 13 for the prac-
tical circulation plan, but the number of trains required in the circulation plan with 122 services obtained by the extended
integrated model is 14. This is because the headway of the train schedule obtained by extended optimization model achieves
350 s at 16:00 in the afternoon peak hours, but the the headway of the train schedule reaches 350 s at 17:13 as can be
observed from Fig. 8. Moreover, the lasting time of this minimum headway in the optimized train schedule by the extended
model is much longer when compared with that of the practical train schedule. So the number of trains required by the opti-
mized train schedule is larger than that required by the practical train schedule as illustrated by Figs. 9 and 10.
As stated before, the constraints for the number of available trains are not considered for the results given in Table 7.
Here, we take the number of available trains as 13 as required in the practical train schedule for the integrated model with
122 services and for the extended integrated model with 121 and 122 services. The performance comparison is given in
Table 8, where the computation times are around 1500 s and are much longer than these given in Table 7. This is because
the number of binary variables and the number of constraints increase dramatically with the introduction of the available
train constraints (18) and (19). More specifically, the number of variables and the number of constraints are about 16000 and
52000 when including the available train constraints. Furthermore, due to the limited number of available trains, the objec-
tive function values in Table 8 are also higher than these given in Table 7. In particular, the extended integrated model with
121 train services achieves the best performance when the number of available trains is 13.
6.3. Case C: Weekdays with smaller headways
6.3.1. Set-up for Case C

For Case C, we consider the generation of a daily regular train schedule and circulation plan with much smaller headways.
The daily operation period of the Beijing Yizhuang line is still from 5:20 to 23:30. The departure times of the first and the last
train services at Yizhuang station are 5:20 and 22:05, respectively. With the increasing of the passenger demand, a possible
service pattern of weekdays could be taken the values demonstrated in Table 9, where the operation period between the first
train service and the last train service is split into 6 time intervals. The headway for the peak hours is 180 s, the headway for
the off-peak hours is 360, and the headway after 20:00 is 540 s. The total number of train services for workdays increases to
216, which is much bigger than 122 train services in Case B.
Table 8
Performance comparison of the integrated model and the extended integrated model for Case B with 13 available trains.

Integrated 122 1852 2.256 508 1477 105 13 0.00
Extended 121 1263 1.203 194 1644 104 13 0.00
Integrated 122 1935 2.148 443 1629 104 13 0.00
Table 9
The service pattern of weekdays for the Beijing Yizhuang line for Case C.

5:20–6:30 360 12
6:30–9:00 180 50
9:00–16:00 360 70
16:00–19:00 180 60
19:00–20:00 360 10
20:00–22:05 540 14
According to the analysis of the results of Case A and Case B, the extended integrated model can achieve better perfor-
mance than the integrated model. So we calculate the train schedules and circulation plans for Case C using the extended
E
integrated model. The normalization factors f obj;1;nom ; f obj;2;nom , and f obj;3;nom for Case C are chosen in a similar way as these
in Case A by using the extended integrated model. The values of these normalization factors are taken as 4170, 192, and
468, respectively. Similar as Case B, the weights w1 ; w2 , and w3 are all taken the value 1.
6.3.2. Results for Case C

The number of variables in the model is about 15000 and the number of constraints is around 50000. We set the max-
imum computation time as 12 h, i.e., 43200 s, and the relative gap of the obtained solution is 0.73%. We then set the relative
gap as 1%, 2%, and 3% as the stop criterion, respectively. The corresponding computation times are 1367 s, 1049 s, and 819 s.
The performance comparison of the extended integrated model with different gaps are given in Table 10. As can be observed
from Table 10, the performance of train schedule and circulation plan obtained with relative gap 0.99% is quite close to these
with relative gap 0.73%, but the computation is much shorter. In addition, the performance of the train schedules and cir-
culation plans with relative gap 1.98% and 2.86% is slightly worse than these with relative gap 0.99%, but is still acceptable
if the computation time is limited.
Fig. 12 illustrates the headways between train services for these train schedules, which are generally consistent with the
service pattern given in Table 9. Because the headway variations are considered in the objective function, the transition
between different headways is not sharp but changes smoothly. In addition, the headways of the train schedule with a higher
relative gap fit the given service pattern better because the number of connected services is smaller, which means that there
will be more entering/exiting depot operations. The train circulation plan obtained by the extended integrated model with
relative gap 0.99% is shown in Fig. 13. The number of entering/exiting depot operations in Fig. 13 is 38 and the number of
required trains is 25.
6.4. Discussions
It is concluded that for the given case study the train schedules and circulation plans obtained by the proposed integrated
optimization model and the extended integrated optimization model can achieve better performance than the practical train
schedule and circulation plan. Moreover, when comparing the headway of the train schedules obtained by the integrated
optimization model and the extended optimization model, the headway transition process of the integrated optimization
model starts and finishes later than that of the extended integrated optimization model. In addition, the computation time
Table 10
Performance comparison of the extended integrated model for Case C with different relative gaps.

216 43200 1.331 4170 591 179 25 0.73
Extended 216 1367 1.359 4170 642 178 25 0.99
Integrated 216 1049 1.403 4170 593 166 27 1.98
216 819 1.507 4170 602 167 27 2.86
Fig. 12. The headway for the train schedules with 216 train services obtained by the extended integrated model with different relative gaps.
Fig. 13. The train circulation plan obtained by the extended integrated model with 216 train services for the Beijing Yizhuang line (relative gap: 0.99%).
for a weekday train schedule and circulation plan of the Beijing Yizhuang line is less than 30 min, which is much more effi-
cient than the experienced designer. Note that for the experienced designer, it takes 3–4 days to obtain the train schedule
and circulation plan.
7. Conclusions and future works
In this paper, we have proposed a collaborative optimization model to integrated the train scheduling and circulation
planning together based on a given service pattern for urban rail transit lines. The departures and arrivals of trains, the turn-
around operations, the entering/exiting depot operations, etc. are considered in the integrated optimization model. Further-
more, the headway deviations between the train schedule, the headway variations between consecutive train services, and
the number of required trains are included in the objective function and are minimized by applying a linear weighted
method. As an extension of the proposed integrated model, binary variables are introduced in the extended model to
describe the relationship between train services and time intervals. The resulting integrated optimization problems are
reformulated into mixed integer linear programming problems. In order to illustrate the effectiveness and efficiency of
the proposed models, two cases were implemented in this paper. The first small case demonstrated the applications of
the presented models. Then, a real case study was performed on the Beijing Yizhuang line with the practical operation data.
The computational results showed that the CPLEX solver can efficiently solve the medium-scale problem and the perfor-
mance of the optimized train schedules and circulation plans is comparable with or even better than that of the practical
train schedule and circulation plan designed by experienced planners.
In this paper, the train schedule and circulation plan is collaboratively optimized based on the service pattern given by the
demand analysis and the line planning. Generating the train schedules and circulation plans directly according to the pas-
senger demand will be one topic for our future work. In addition, we focus on the comparison between the proposed opti-
mization approach and the practical train schedule obtained by the dispatchers in this paper. Furthermore, different types of
train services, e.g., short-turning train services and skip-stop train services, are introduced in the operation of urban rail tran-
sit lines to enhance the passenger satisfaction and to reduce the operation costs. In future work, we will consider the inte-
grated train scheduling and circulation planning for urban rail transit lines that have multiple depots and different types of
train services. There we need to consider the capacity of each depot and the balance issue of trains between multiple depots.
Furthermore, the maintenance constraints which limit the running distances that trains can serve should be included in the
train scheduling and circulation planning. Moreover, we will compare the formulated models with multiple depots and/or
multiple lines with the other integrated models in the existing works, such as Cadarso et al. (2012) and Chang et al. (2015).
Acknowledgment
Research supported by the National Natural Science Foundation of China (No. 61503020, 71571012), the State Key Lab-
oratory of Rail Traffic Control and Safety (Contract No. RCS2016ZJ003), the Fundamental Research Funds for the Central
Universities (No. 2017JBM076), the Beijing Municipal Science and Technology Commission (Contract No.
Z161100001016006), and the Beijing Municipal Commission of Education (Contract No. I15H100010).
References
Albrecht, T., 2009. Automated timetable design for demand-oriented service on suburban railways. Public Transport 1, 5–20.
Alfieri, A., Groot, R., Kroon, L., Schrijver, A., 2006. Efficient circulation of railway rolling stock. Transport. Sci. 40, 378–391.
Assis, W., Milani, B., 2004. Generation of optimal schedules for metro lines using model predictive control. Automatica 40, 1397–1404.
Atamturk, A., Savelsbergh, M., 2005. Integer-programming software systems. Ann. Oper. Res. 140, 67–124.
Barrena, E., Cana, D., Coelho, L., Laporte, G., 2014. Exact formulations and algorithm for the train scheduling problem with dynamic demand. Comput. Oper.
Res. 44, 66–74.
Bemporad, A., Morari, M., 1999. Control of systems integrating logic, dynamics, and constraints. Automatica 35, 407–427.
Bussieck, M., Winter, T., Zimmermann, U., 1997. Discrete optimization in public rail transport. Math. Program. 79, 415–444.
Cacchiani, V., Caprara, A., Toth, P., 2010. Solving a real-world train-unit assignment problem. Math. Program. 124, 207–231.
Cacchiani, V., Caprara, A., Toth, P., 2013. A lagrangian heuristic for a train-unit assignment problem. Discrete Appl. Math. 161, 1707–1718. 9th Cologne/
Twente Workshop on Graphs and Combinatorial Optimization (CTW 2010).
Cadarso, L., Marin, A., Maroti, G., 2012. Integration of timetable planning and rolling stock in rapid transit networks. Ann. Oper. Res. 199, 113–135.
Cadarso, L., Marin, A., Maroti, G., 2013. Recovery of disruptions in rapid transit networks. Transport. Res. Part E: Logist. Transport. Rev. 53, 15–33.
Canca, D., Barrena, E., Algaba, E., Zarzo, A., 2014. Design and analysis of demand-adapted railway timetables. J. Adv. Transport. 48, 119–137.
Ceder, A., 2009. Public-transport automated timetables using even headway and even passenger load concepts. In: Proceedings of the 32nd Australasian
Transport Research Forum, Auckland, New Zealand, pp. 1–17.
Chang, S., Jong, J., Lai, Y., 2015. Integrated optimization method for the train scheduling and utilization planning problems in mass rapid transit systems. In:
Proceedings of the 6th International Conference on Railway Operations Modelling and Analysis, Tokyo, Japan.
Cordeau, J., Toth, P., Vigo, D., 1998. A survey of optimization models for train routing and scheduling. Transport. Sci. 32, 380–420.
Corman, F., D’Ariano, A., Hansen, I.A., 2014. Evaluating disturbance robustness of railway schedules. J. Intell. Transport. Syst. 18, 106–120.
Corman, F., D’Ariano, A., Marra, A.D., Pacciarelli, D., Samà, M., 2016. Integrating train scheduling and delay management in real-time railway traffic control.
Transport. Res. Part E: Logist. Transport. Rev. http://dx.doi.org/10.1016/j.tre.2016.04.007.
Corman, F., D’Ariano, A., Pacciarelli, D., Pranzo, M., 2012. Optimal inter-area coordination of train rescheduling decisions. Transport. Res. Part E: Logist.
Transport. Rev. 48, 71–88. Select Papers from the 19th International Symposium on Transportation and Traffic Theory.
Cury, J., Gomide, F., Mendes, M.J., 1980. A methodology for generation of optimal schedules for an underground railway systems. IEEE Trans. Autom. Control
25, 217–222.
D’Ariano, A., Pacciarelli, D., Pranzo, M., 2007a. A branch and bound algorithm for scheduling trains in a railway network. Eur. J. Oper. Res. 183, 643–657.
D’Ariano, A., Pranzoand, M., Hansen, I., 2007b. Conflict resolution and train speed coordination for solving real-time timetable perturbations. IEEE Trans.
Intell. Transport. Syst. 8, 208–222.
Dewilde, T., Sels, P., Cattrysse, D., Vansteenwegen, P., 2013. Robust railway station planning: an interaction between routing, timetabling and platforming. J.
Rail Transport Plan. Manage. 3, 68–77. Robust Rescheduling and Capacity Use.
Dollevoet, T., Corman, F., D’Ariano, A., Huisman, D., 2014. An iterative optimization framework for delay management and train scheduling. Flexible Services
Manuf. J. 26, 490–515.
Fioole, P.J., Kroon, L., Marti, G., Schrijver, A., 2006. A rolling stock circulation model for combining and splitting of passenger trains. Eur. J. Oper. Res. 174,
1281–1297.
Flamini, M., Pacciarelli, D., 2008. Real time management of a metro rail terminus. Eur. J. Oper. Res. 189, 746–761.
Ghoseiri, K., Szidarovszky, F., Asgharpour, M., 2004. A multi-objective train scheduling model and solution. Transport. Res. Part B: Methodol. 38, 927–952.
Giacco, G.L., D’Ariano, A., Pacciarelli, D., 2014. Rolling stock rostering optimization under maintenance constraints. J. Intell. Transport. Syst. 18, 95–105.
Higgins, A., Kozan, E., Ferreira, L., 1996. Optimal scheduling of trains on a single line track. Transport. Res. Part B: Methodol. 30, 147–161.
Kraay, D., Harker, P., Chen, B., 1991. Optimal pacing of trains in freight railroads: model formulation and solution. Oper. Res. 39, 82–99.
Kwan, C., Chang, C., 2005. Application of evolutionary algorithm on a transportation scheduling problem – the mass rapid transit. In: Proceedings of the IEEE
Congress on Evolutionary Computation, Edinburgh, UK, pp. 987–994.
Li, X., Lo, H., 2014. An energy-efficient scheduling and speed control approach for metro rail operations. Transport. Res. Part B: Methodol. 64, 73–89.
Liebchen, C., 2006. Periodic Timetable Optimization in Public Transport. Ph.D. thesis, Technical University of Berlin, Berlin, Germany.
Liebchen, C., 2008. The first optimized railway timetable in practice. Transport. Sci. 42, 420–435.
Liebchen, C., Möhring, R.H., 2007. The Modeling Power of the Periodic Event Scheduling Problem: Railway Timetables — and Beyond. Springer, Berlin,
Heidelberg, pp. 3–40.
Linderoth, J., Ralphs, T., 2005. Noncommercial software for mixed-integer linear programming. In: Karlof, J. (Ed.), Integer Programming: Theory and Practice,
Operations Research Series. CRC Press, Boca Raton, FL, USA, pp. 253–303.
Michaelis, M., Schöbel, A., 2009. Integrating line planning, timetabling, and vehicle scheduling: a customer-oriented heuristic. Public Transport 1, 211–232.
Nachtigall, K., Voget, S., 1996. A genetic algorithm approach to periodic railway synchronization. Comput. Oper. Res. 23, 453–463.
Niu, H., Zhou, X., 2013. Optimizing urban rail timetable under time-dependent demand and oversaturated conditions. Transport. Res. Part C: Emerging
Technol. 36, 212–230.
Peeters, M., Kroon, L., 2008. Circulation of railway rolling stock: a branch-and-price approach. Comput. Oper. Res. 35, 538–556. Part Special Issue: Location
Modeling Dedicated to the memory of Charles S. ReVelle.
Pellegrini, P., Marlire, G., Rodriguez, J., 2014. Optimal train routing and scheduling for managing traffic perturbations in complex junctions. Transport. Res.
Part B: Methodol. 59, 58–80.
Petersen, E., Taylor, A., Martland, C., 1986. An introduction to computer-assisted train dispatch. J. Adv. Transport. 20, 63–72.
Samà, M., D’Ariano, A., Corman, F., Pacciarelli, D., 2017. A variable neighbourhood search for fast train scheduling and routing during disturbed railway
traffic situations. Comput. Oper. Res. 78, 480–499.
Samà, M., Pellegrini, P., D’Ariano, A., Rodriguez, J., Pacciarelli, D., 2016. Ant colony optimization for the real-time train routing selection problem. Transport.
Res. Part B: Methodol. 85, 89–108.
Su, S., Li, X., Tang, T., Gao, Z., 2013. A subway train timetable optimization approach based on energy-efficient operation strategy. IEEE Trans. Intell.
Transport. Syst. 14, 883–893.
Sun, L., Jin, J., Lee, D., Axhausen, K., Erath, A., 2014. Demand-driven timetable design for metro services. Transport. Res. Part C: Emerging Technol. 46, 284–
299.
Szpigel, B., 1972. Optimal train scheduling on a single line railway. In: Proceedings of the International Conference on Operational Research, Amsterdam,
The Netherlands, pp. 344–351.
Wang, Y., De Schutter, B., van den Boom, T., Ning, B., Tang, T., 2014. Efficient bi-level approach for urban rail transit operation with stop-skipping. IEEE Trans.
Intell. Transport. Syst. 15, 2658–2670.
Wang, Y., Tang, T., Ning, B., van den Boom, T., De Schutter, B., 2015. Passenger-demands-oriented train scheduling for an urban rail transit network.
Transport. Res. Part C 60, 1–23.

Transportation Research Part E: Yihui Wang, Tao Tang, Bin Ning, Lingyun Meng

Uploaded by

Document Information

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 E: Yihui Wang, Tao Tang, Bin Ning, Lingyun Meng

Uploaded by

Copyright:

Available Formats

Transportation Research Part E 105 (2017) 83–104

Contents lists available at ScienceDirect

Transportation Research Part E

Integrated optimization of regular train schedule and train

1.1. Literature review

1.2. The proposed integrated approach

2. Mathematical formulation for the integrated optimization problem

2.1. Assumptions and notations

2.3. Operation of train services at stations

2.4. Circulation of trains

which involves the following two situations:

For each train service, the following constraint should be satisfied

3. Integrated optimization problem for regular train schedule and circulation

3.1. Objective function

which is equivalent to minimize

5. Solution approach – MILP

P.1 Consider the statement ~f ð~

subject to constraints (43) and (44), where

6.1. Case A: Two time intervals

6.1.1. Set-up for Case A

Fig. 4. The layout of the Beijing Yizhuang line.

Time interval Headway [s] Number of train services

6.1.2. Results for Case A

Model Number of Computation Objective Headway Headway Number of Number of Relative

6.2. Case B: Weekdays in the current practice

6.2.1. Set-up for Case B

6.2.2. Results for Case B

Time interval Headway [s] Number of train services

Model Number of Computation Objective Headway Headway Number of Number of Relative

6.3. Case C: Weekdays with smaller headways

6.3.1. Set-up for Case C

Model Number of Computation Objective Headway Headway Number of Number of Relative

Time interval Headway [s] Number of train services

6.3.2. Results for Case C

Model Number of Computation Objective Headway Headway Number of Number of Relative

7. Conclusions and future works

You might also like