IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO.

2, APRIL 2014 673

Optimization of Multitrain Operations

in a Subway System
Shuai Su, Tao Tang, Xiang Li, and Ziyou Gao

Abstract—Energy efficiency is paid more and more attention should minimize the energy cost. The driving strategy during
in railway systems for reducing the cost of operation companies interstations is another important factor of energy-efficient op-
and emissions to the environment. In subway systems, the opti- eration. With a given trip time, there are many feasible driving
mizations on timetable and driving strategy are two important
and closely dependent parts of energy-efficient operations. The strategies that consume different amounts of energy.
former regulates the fleet size and the trip time at interstations, The timetable and driving strategy are two closely dependent
and the latter determines the control sequences of traction and parts of energy-efficient operations. On one hand, the cycle time
braking force during the trip. Most conventional research opti- that is influenced by the fleet size and interstations’ trip time
mized the timetable and the driving strategy separately such that that is scheduled by timetable are important constraints for the
global optimality cannot be achieved. In this paper, we analyze
the hierarchy of energy-efficient train operation and then propose optimal train-control model. On the other hand, the amount of
an integrated algorithm to generate the globally optimal opera- energy consumption determined by the optimal driving strategy
tion schedule, which can get better energy-saving performance. is the foundation of optimizing the timetable. Hence, to obtain
Within the criteria of meeting the passenger demand, the inte- the global optimality for the whole system, one cannot neglect
grated energy-efficient algorithm can simultaneously obtain the the relationship between the timetable and driving strategy.
optimal timetable and driving strategy for trains, which realizes
the combination of the high-level transportation management and Our previous work [35] has presented an approach of inte-
the low-level train operation control. The simulation results based grating the train control and distribution of the trip time, but
on the Beijing Yizhuang Subway Line illustrate that the integrated the methodology is used to optimize the operation for a single
algorithm can achieve a 24.0% energy reduction for one day, on train. In other words, our previous work did not consider the
average. In addition, the computation time is within 2 s, which is layer of headway control (see Fig. 1). In addition, the previous
short enough to be applied for real-time control system.
approach is based on the assumption that the traction force, the
Index Terms—Energy conservation, integrated energy-efficient braking force, and the running resistance are constant. This pa-
(IEE) algorithm, subway system, timetable, train operation. per extends the methodology to solve the energy-efficient train-
I. I NTRODUCTION control problem when considering the variable traction force,
braking force, and running resistance. Based on the hierarchy

S UBWAY systems, as an important part of urban trans-

port, are tasked with taking passengers to their destination
safely, quickly, and conveniently. A timetable, which deter-
of energy-efficient train operation (see Fig. 1), an integrated
energy-efficient (IEE) optimization model is formulated to re-
alize the optimal control of multitrains from the system point of
mines the fleet size, the departure time, and the arrival time of view. An IEE algorithm is proposed to optimize the timetable
trains at stations, is the foundation of train operation and has a and driving strategy, which can achieve global optimization of
great influence on energy consumption. The larger the fleet size train operation on energy saving.
is, the greater the energy consumption will be. The departure The rest of this paper is organized as follows. In Section II,
and arrival times, which determine the trip time distribution we review some important literature on timetable planning and
among different interstations, also influence the energy con- energy-efficient driving strategy. In Section III, we analyze
sumption of the whole trip [7], [35]. Therefore, within the cri- the three layers of train operation and formulate an IEE train
teria of meeting the passenger demand, an excellent timetable operation model. In Section IV, we present the algorithms for
calculating the energy-efficient driving strategy, distributing the
Manuscript received February 16, 2013; revised June 23, 2013 and August 8, cycle time, and generating the train timetable. This is integrated
2013; accepted October 1, 2013. Date of publication October 30, 2013; date to optimize the train operation. In Section V, a case study is
of current version March 28, 2014. This work was supported in part by the
National Natural Science Foundation of China under Grants 71101007 and
presented based on the infrastructure and operation data from
71371027, by the National Basic Research Program of China under Grant the Beijing Yizhuang Subway Line in China, which illustrates
2012CB725401, by the Specialized Research Fund for the Doctoral Program that the proposed approach can provide good performance with
of Higher Education of China under Grant 20110009120036, by the State Key
Laboratory of Rail Traffic Control and Safety under Grant RCS2012ZT002, and
energy saving. In Section VI, sensitivity analysis on the as-
by the Fundamental Research Funds under Grant 2013YJS019 from Beijing sumptions is presented to illustrate that the proposed approach
Jiaotong University. The Associate Editor for this paper was B. Chen. is meaningful and practical.
S. Su, T. Tang, and Z. Gao are with the State Key Laboratory of Rail Traffic
Control and Safety, Beijing Jiaotong University, Beijing 100044, China.
X. Li is with the School of Economics and Management, Beijing University
of Chemical Technology, Beijing 100029, China (e-mail: lixiang@bjtu.edu.cn). II. L ITERATURE R EVIEW
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. To achieve good utilization of capacity, subway service must
Digital Object Identifier 10.1109/TITS.2013.2285737 be tailored to cover the variable passenger demand as much

1524-9050 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
674 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

Fig. 1. Three layers for energy-efficient train operation.

as possible. The hourly cyclic timetable, for which the trains to be reduced by maximizing the use of regenerative-braking
running on the same line will follow each other with a fixed energy [11].
cycle time [13], is a good solution to the problem. The periodic Much research also focuses on how to obtain the energy-
event-scheduling problem (PESP) model introduced by Serafini efficient driving strategy between successive stations. Dating
and Ukovich [14] has been widely applied to solve the cyclic back to the 1960s, Milroy [2] proposed an optimal control
railway timetable problem (CRTP) [8], [19], [23]. The PESP model based on the assumption that the train runs on a flat track
considered the problem as a set of periodically recurring events with a constant gradient. By using the Pontryagin maximum
under periodic time window constraints. Voorhoeve [8] used the principle, the optimal driving strategy is proved to consist of
PESP model to solve the CRTP for reducing the operational maximum acceleration, cruising, coasting, and maximum brak-
cost. Nachtigall [23] used the PESP constraints to formulate the ing. The Pontryagin maximum principle was used as a typical
cyclic behavior of the railway timetables and introduced a cycle method and laid the foundation for solving the optimal train
periodicity formulation model to minimize the passengers’ control problem. For taking the theory into practice, variable
waiting time. To deal with variable trip time in an existing cyclic gradients, variable speed limits, and traction efficiency were
railway timetable, Kroon and Peeters [18] extended the PESP gradually taken into account [5], [6], [27], [37]. For example,
model, in which small deviations from the fixed trip time are Khmelnitsky [5] chose the kinetic energy as the control variable
allowable for getting a larger, useful solution space. To obtain and considered variable gradients, variable traction efficiency,
more feasible solutions to PESP, Gabrio et al. introduced a and arbitrary speed limits in the optimal train-control model.
periodic service intention framework for the partial periodic Then, he gave an analytic solution to solve the optimal train
timetables in order to deal with changing demand [21]. In control sequences. Liu and Golovitcher [6] also analyzed the
addition, Liebchen and Peeters [22] modeled the CRTP by a optimal driving strategy by using the Pontryagin maximum
constraint graph and a periodic tension and introduced a more principle and gave the possible switching of optimal controls.
general concept of integral cycle bases. This characterized the Then, an analytical solution was proposed to find the switching
integral cycle bases using the determinant of cycle bases. Other points of control sequences while considering variable gradi-
recent developments in this field mainly concentrated on the ents and speed limits. In addition, Howlett et al. [31] provided
design of robust cyclic timetables to cope with stochastic delays an analytical method for the problem with more than one steep
[3], [4], [17], [20], [36]. For example, Jamili et al. [36] used a slope, and a local optimization principle was used to solve
hybrid metaheuristic algorithm to deal with scheduling different the energy-efficient driving strategy for each part of the route.
types of trains in a single railway track for minimizing train de- Li et al. [32] studied the optimal driving strategy with an energy
lays. Shafia et al. [3] formulated a mixed integer programming constraint since the minimum energy consumption uniquely
model for developing robust train-timetabling on a single-track corresponds to the given trip time. In addition to analytical and
railway line. Then, a branch-and-bound algorithm, along with a numerical methods, evolutionary algorithms such as a genetic
new heuristic beam search algorithm, is presented to solve the algorithm [1], [26] and an ant colony algorithm [33] were
model with a short computation time. Theories about PESP and applied to solve the optimal train-control problem. Lin and
corresponding algorithms were put into practice with coordi- Sheu [15] proposed an adaptive optimal control algorithm to
nation among trains through cooperative timetables [11], [12]. regulate the train operation with less energy consumption. The
For example, the movement of trains in subway systems can be method can be used to optimize the train regulator by adjusting
synchronized to allow for energy consumption from substations the running time and dwell time.
SU et al.: OPTIMIZATION OF MULTITRAIN OPERATIONS IN A SUBWAY SYSTEM 675

Furthermore, some research has studied the integration be- α utilization of the maximum train capacity;
tween the train-control and timetable design. For example, f service frequency (in trains per hour);
Albreche et al. [24] designed a dynamic programming approach v(s) train speed (in meters per second);
to solve the optimal trip time for a single train and then V speed limit (in meters per second);
presented a solution to the optimal train-control problem based v0j initial speed for the jth section (in meters per second);
on the Pontryagin maximum principle. Ghoseiri [25] optimized vtj final speed for the jth section (in meters per second);
the timetable based on the assumption that trains have constant s train position (in meters);
speeds, which did not have a precise estimation of the energy Li length of section i (in meters);
consumption for interstations. Su et al. [35] proposed an ap- Tp pure trip time (in seconds);
proach to find the optimal driving strategy of the entire route Td total dwell time during the trip (in seconds);
by distributing the reserve time to different sections, which can Tt turn-back time at terminals (in seconds);
optimize the train-control sequences, as well as the trip time E energy consumption per unit mass (in kilowatthours);
for interstations. However, the traction force, braking force, F (v)maximum traction force per unit mass (in meters per
and the resistance in this paper are assumed as constants, and square second);
the proposed approach can only be used for optimizing the B(v) maximum braking force per unit mass (in meters per
operation of a single train. square second);
r(v) running resistance per unit mass (in meters per square
III. M ODEL F ORMULATION second);
g(s) gradient resistance per unit mass (in meters per square
A. Assumptions and Symbols
second);
For a better understanding of this paper, the assumptions, m number of interstations.
parameters, and decision variables are first introduced.
• Trains in a subway system are controlled by the signal
system, which will permit the train operation at the stated B. Energy-Efficient Operation Model
headway. As shown in Fig. 1, the highest layer of the train operation is
• A subway system uses the hourly timetable that is sched- the headway control, which is followed by the middle layer of
uled for different time periods since the passenger demand distribution of the cycle time. In addition, the energy-efficient
presents periodical change at different hours based on the driving strategy is the lowest layer. The timetable in subway
operational experience. systems is designed according to the passenger demand, which
• The advanced automatic train operation (ATO) system is the foundation of train operation. It usually determines the
has been widely used in the subway signaling system, fleet size, headway between trains, and the trip time for each
which can track the target speed precisely and fulfill the interstation. The headway in the timetable determines the ca-
automatic train control instead of drivers. Generally, the pacity of the system
ATO system uses the same target speed for a given section
during a fixed time period. Hence, all trains are assumed 3600
C =c×f ×α=c× ×α (1)
to have the same driving strategy for a given section in a h
given time period.
in which the headway can be further formulated as
• The turn-back times are assumed as constants at terminals
since the ATO system is capable of completing automatic T
turn-back. h= . (2)
n
• The passenger demand can be obtained from the operation
control center through the smart card system. For minimizing the energy consumption W (in kilowatthours)
• The approach proposed in this paper focuses on timetable in a fixed time period Q (in seconds), we formulate the follow-
planning. Hence, the dwell times are assumed as constants ing optimization criterion:
at stations based on the operational experience, although
Q
the practical dwell times are not uniform due to unpre- min W = nE(T ). (3)
dicted number of passengers alighting and boarding. T
Decision Variables: For a given subway system, n and T should follow the fleet size
Tj trip time at the jth interstation (in seconds); and the total trip time constraints
T cycle time (in seconds);
n number of service trains; n ≤ nmax , Tmin ≤ T ≤ Tmax . (4)
h headway (in seconds);
kt (s) relative traction force; The energy consumption E(T ) during a cycle time is the sum
kb (s) relative braking force; of the energy consumption at all interstations, i.e.,
Parameters: 
m
c train capacity (in persons/train); E(T ) = Ej (Tj ) (5)
C line capacity (in persons); j=1
676 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

where the interstation trip time should satisfy the following trip
time constraints:


m
T = Td + Tt + Tj , Tjmin ≤ Tj ≤ Tjmax , j = 1, 2, · · · , m.
j=1
(6)
Furthermore, Ej (Tj ) is the energy consumption during the
traction process, which is determined by the driving strategy

Tj
Ej (Tj ) = kt (t)v(t)F (t)dt. (7) Fig. 2. Integrated approach for optimizing the train operation.
0
In this paper, an integrated model is proposed to optimize the
The train movement can be described by energy-efficient timetable and the optimal train control as a
whole
dv(t) ⎧
= kt F (t) − kb B(t) − g(s) − r(v) (8) ⎪ 
m Tj
dt ⎪
⎪ Q
⎪
⎪ min W = n kt (t)v(t)F (t)dt
⎪
⎪
T
j=1 0
⎪
⎪
with the boundary conditions ⎪
⎪ s.t. C = c × 3600 h × α, h = n , n ≤ nmax
T
⎪
⎪
⎪
⎪ m
⎪
⎨ T = Tj , Tjmin ≤ Tj ≤ Tjmax
v(0) = v(Tj ) = 0 (9) j=1
(13)
⎪
⎪
dv(t)
= k F (t) − k B(t) − g(s) − r(v)
⎪
⎪ dt t b
and the constraints on the trip distance, the train speed, ⎪
⎪ Tj
⎪
⎪
the relative traction force, and the relative braking force, as ⎪
⎪ L = v(t)dt
⎪
⎪
i
shown in ⎪
⎪
0
⎪
⎪ v(0) = v(Tj ) = 0
⎩
kt ∈ [0, 1], kb ∈ [0, 1], v ≤ V .
Tj
Li = v(t)dt, v ≤ V , kt ∈ [0, 1], kb ∈ [0, 1]. (10)
0
IV. IEE A LGORITHM
In this section, we present the algorithms for each layer of
Within the criteria of meeting the passenger demand, timetable the energy-efficient train operation and integrate them to solve
optimization aims to minimize the total traction energy con- the globally optimal solution, which is called IEE algorithm.
sumption W of all trains for a period of time Q, which is As shown in Fig. 2, we propose an integrated approach
presented as the following timetable optimization model: to optimize the train operation in this section. The optimal
⎧ train control (low layer), the distribution of the cycle time
⎪ 
m
⎪
⎪ min W = Q (middle layer), and the headway optimization (high layer) are
⎪ Tn Ej (Tj )
⎪
⎪ formulated as two closed loops. When the middle layer is
⎪
⎪
j=1
⎪
⎨ s.t. C =c× 3600
h ×α optimized, the low layer will act as the input. These two layers
h = Tn (11) form the first loop of energy-efficient operation for single train.
⎪
⎪ 
⎪
⎪
m Similarly, the two layers regarded as a whole is considered as
⎪
⎪ T = Tj
⎪
⎪ the input when the high layer of headway control is designed. A
⎪
⎩
j=1
close relationship among the three layers fulfills the integrated
Tjmin ≤ Tj ≤ Tjmax .
framework of the train operation.
With the given timetable, the optimal train-control model fo-
A. Optimal Train-Control Algorithm
cuses on traveling the interstation with the minimum energy
consumption There has been much research concentrating on how to get
the optimal driving strategy such that the energy consumption
⎧ Tj
⎪ is minimized [5], [6], [27], [30], [31]. The authors described the
⎪
⎪ min Ej (Tj ) = kt (t)v(t)F (t)dt
⎪
⎪ problem with an optimal train-control model, which has been
⎪
⎪ 0
⎪
⎪ dv(t) presented in (12).
⎨ s.t. dt = kt F (v) − kb B(v) − g(x) − r(v)
By applying the Pontryagin maximum principle, the optimal
Tj (12)
⎪
⎪ control problem is transformed to maximize the following
⎪
⎪
Li = v(t)dt
⎪
⎪ 0 Hamiltonian function with respect to variables kt and kb :
⎪
⎪
⎪
⎩ v(0) = v(Tj ) = 0, v ≤ V
kt ∈ [0, 1], kb ∈ [0, 1]. H = (p1 − v)kt F − p1 kb B − p1 (r + g) + p2 v (14)
SU et al.: OPTIMIZATION OF MULTITRAIN OPERATIONS IN A SUBWAY SYSTEM 677

TABLE I constraint. In conclusion, the optimal speed sequences can be

O PTIMAL D RIVING S TRATEGY
solved with the following algorithm.

Algorithm 4.1: The algorithm for obtaining the optimal

speed sequences for a given section is described as follows.

Step 1. Initialize initial speed v0 , final speed vt , and energy

consumption E;
Step 2. Divide the section into l parts; the speed sequences
{vi , 0 ≤ i ≤ l} are corresponding to the train posi-
tion {si , 0 ≤ i ≤ l};
Step 3. Set i = 0, and vi = v0 ;
Step 4. Set i = i + 1. Calculate

a = F (vi−1 ) − r(vi−1 ) − g(si−1 )

vi2 = vi−1
2
+ 2a(si − si−1 )
E = E − F (vi−1 )(si − si−1 ); (15)

Step 5. If vi < V̄ (si ) and E > 0, then go to step 4;

Step 6. If E > 0, then let vi+1 = vi and
Fig. 3. Switching of the optimal control sequences.
E = E − (r(vi ) + g(si )) (si+1 − si ); (16)

Step 7. If E > 0, set i = i + 1 and go to step 5;

Step 8. Set k = i and pk = vk ;
Step 9. If k ≤ l, generate

p2k+1 = p2k − 2(sk+1 − sk ) (r(pk ) + g(sk )) . (17)

Set k = k + 1 and repeat step 8;

Step 10. Let j = l and qj = vt ;
Step 11. If j > i, then calculate
2
qj−1 = qj2 + 2(sj − sj−1 ) (B(qj ) + r(qj ) + g(sj )) . (18)

Set j = j − 1 and repeat step 10;

Step 12. Let i = i + 1. Set
Fig. 4. E−T function.
vi = min{qi , pi }. (19)
from which we can obtain that for a given section whose
gradient and speed limit are constants, the optimal driving strat- If i ≤ l, repeat step 11;
egy consists of maximum acceleration (MA), cruising (CR), Step 12. Return the optimal speed sequences {vi , 0 ≤ i ≤ l}
coasting (CO), and maximum braking (MB) [7], [30], [35] (see and the trip time of this section t with the following
Table I). Furthermore, switching of the phase sequences can be equation:
formulated by Fig. 3 (see literature [6]). 
l
si+1 − si
In general, the energy consumption will decrease as the trip t= . (20)
time increases. As shown in Fig. 4, the minimum energy con- i=1
vi
sumption is uniquely determined by the trip time and vice versa.
Hence, we formulate an algorithm to calculate the optimal
speed sequences with the energy constraint instead of using
the time constraint. As shown in Fig. 5, the speed sequences
B. Distribution Algorithm of Cycle Time
in the MA phase are first calculated with the given energy
consumption. If the speed reaches the speed limit during the Considering the variable speed limits and gradients, we first
acceleration process, the rest energy will be used for generating partition the route into “k” section with a constant speed limit
the speed sequences of the CR phase. For the rest journey, and gradient (see Fig. 6). According to the conclusion of the
a coasting speed sequence pk from the end speed of the CR optimal train-control model, the optimal control sequences in
phase and a braking speed sequence qj from the end of the MB each section consist of MA, CO, MB, and CR.
phase are calculated; then, the minimum speed value of the two Second, we find a primary solution for this optimal con-
sequences is obtained as the optimal speed with given energy trol problem. The train applies the MA at first to achieve a
678 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

Fig. 5. Sketch map for Algorithm 4.1.

Fig. 7. Distribution of reserve time.

Algorithm 4.2: The distribution algorithm of the cycle time

is designed as follows.

Step 1. Initialize the cycle time T , and calculate the pure trip
time Tp as

Fig. 6. Interstation and sections. Tp = T − Td − Tt . (21)

Step 2. Divide the route into “k” sections such that each
high speed (for example, trains usually try to accelerate to section has constant gradient and speed limit;
50 km/h in the subway system), which is necessary in practical Step 3. Initialize a primary solution and record the speed
operation and regarded as a fixed part of the optimal control sequences for each position;
sequences. Then, the train will coast for a while until the Step 4. Calculate the primary trip time Tx according to (20);
speed of the train reaches the braking curve. Finally, the train Step 5. Add an energy unit ΔE to the ith section (0 < i ≤ k)
will apply the maximum available braking force to the station and recalculate the trip time Tx with Algorithm 4.1;
or the speed limit. We assume that the trip time and energy then, the time reduction is obtained with
consumption calculated by the above primary solution are Tx
ΔTi = Tx − Tx ; (22)
and Ex , respectively. It is easy to find that Ex is the minimum
energy consumption with the trip time Tx . Step 6. If ΔTj = max{ΔTi }, distribute ΔE the jth section;
The practical energy consumption will be more than that of Step 7. Return the new speed sequences and trip time Tx ;
the primary solution, and the trip time should be shortened to Step 8. Set Tx = Tx . If Tx < Tp , go to step 5;
meet the operation demand. Hence, we design an algorithm Step 9. Return the final speed sequences and the trip time at
to get the optimal control sequences for interstations (or the each interstation after distribution of the cycle time.
entire route) based on the primary solution by increasing the
energy consumption of each section until the trip time Tx is
exactly shortened to the total trip time of the interstation Tj
(or cycle time of the entire route T ). At each time, we try to
C. Headway Control
allocate an energy unit ΔE for one section, and then the speed
sequences are regenerated for all the following sections with In practice, the train-control system is traditionally designed
Algorithm 4.1. The low layer of optimal train control is inte- by the signaling engineers, whereas the timetable is decided
grated into the middle layer, from which we can recalculate the by the operation company according to the operational require-
trip time Tx and the time reduction ΔT . The same amount of ment and the obtained data from the signaling system. For ex-
energy added to different sections may obtain different time ample, a signaling engineer can calculate the minimum trip time
reductions, which depends on the ratio ΔE/ΔT [7], [35]. for an interstation as 104 s. The length of the route is 1400 m,
Hence, we distribute the energy unit to obtain the maximum and the maximum trip time for this interstation is 126 s if
time reduction at each time. Taking Fig. 7 as an example, the lowest average traveling speed of 40 km/h (11.1 m/s) is
we can get more time reduction in Section I since the ratio offered to the passengers. Then, the operation company may
ΔE/ΔT1 is smaller than ΔE/ΔT2 . Repeat the distribution schedule the trip time of this interstation as 115 s for adding
process until the trip time Tj (or T ) is delivered, and we will get some time supplementation to create a robust timetable instead
the optimal speed sequences and the trip time for all sections. of considerations to save energy. Theoretically, the timetable
SU et al.: OPTIMIZATION OF MULTITRAIN OPERATIONS IN A SUBWAY SYSTEM 679

design is a complex work and has been proved to be NP-hard

[10], [14]. The variables will be numerous, and the computation
time of the algorithm will be much longer when the timetable
and train control are integrated. This cannot fulfill the critical
requirement of the advanced signaling system in subways since
the real-time control for multitrains is used to ensure the system
safety. In conclusion, the timetable and train control are not
efficiently combined to reduce the energy consumption of train
operation. In addition, the numerical IEE algorithm proposed
in this section has a short computation time and can be used for
both off-time optimization and real-time control.
Optimizing the headway of the subway system must be
within the criteria of meeting the passenger demand. Generally,
passenger demand varies with different time periods and line
capacity should also be changed to cope with the variable
passenger demand. For a fixed time period, the loading factor
and train capacity are constant, and headway determines the
line capacity of subway systems. Furthermore, headway is
closely related to the cycle time and fleet size according to
(1) and (2). Hence, three methods are concluded to increase
(decrease) the line capacity as follows:
• increasing (decreasing) the fleet size;
• reducing (extending) the cycle time of the train;
• combination of the last two methods.
For example, when the passenger demand increases, dis-
patchers will shorten the headway by increasing the fleet size
and reducing the cycle time, which aims to provide adequate
transporting capacity, reduce the waiting time of the passengers
and avoiding congestion in stations. When the passenger de-
Fig. 8. Flowchart of the IEE algorithm.
mand is smaller, dispatchers will shrink the fleet size or extend
the cycle time appropriately to reduce the operation cost. In
addition, the subway system must offer a certain minimum Algorithm 4.3: The IEE algorithm for optimizing the train
frequency of service (maximum acceptable headway) required operation is described as follows.
from a level-of-service point of view [9].
Fleet size and the cycle times have a great influence on
the energy consumption of subway systems. On one hand, a Step 1. Based on the passenger demand and loading factor,
larger fleet size consumes more energy consumption during headway h is calculated according to (1);
a fixed time period. On the other hand, the cycle time and Step 2. Initialize the fleet size n and calculate the cycle time
its distribution are also important factors on energy saving. T according to (2);
Hence, as a precondition of meeting the passenger demand, the Step 3. With parameters n and T , we calculate the energy
operation plan of the fleet size and cycle time should be energy consumption in a fixed time period by Algorithm 4.2;
efficient. Furthermore, the fleet size and the cycle time are Step 4. Take n = n + 1 and repeat the steps 2 and 3 until
closely dependent through the headway constraint. Separately n = nmax ;
optimizing the fleet size or the cycle time (the first two methods) Step 5. Make a comparison on the energy consumption of
cannot achieve the globally optimal solution. Hence, we adopt different schedules and output the energy-efficient
the combination method in our model to optimize the headway fleet size nk ;
for train operation. First, the passenger demand is obtained Step 6. Based on T k and nk , we calculate the optimal trip
from the operation control center. Then, based on the passenger time and driving strategy for each interstation by
demand and the loading factor, we calculate line capacity and Algorithm 4.2.
the necessary headway. To achieve this headway, we may have
many schedules, depending on the fleet size and cycle time.
The flowchart of the IEE algorithm is shown in Fig. 8.
By utilizing the available number of service trains, we calcu-
late the energy consumption for each schedule according to
V. C ASE S TUDY
Algorithms 4.1 and 4.2. The middle layer of distribution of the
cycle time and the low layer of optimal control are integrated to The Beijing Yizhuang Subway Line, which started operation
the high-layer optimization, among which we choose the most in 2010, is an important subway line for linking the downtown
energy-efficient one as our final operation plan. of Beijing and the Yizhuang Economic Development Zone.
680 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

Fig. 9. Yizhuang Subway Line in Beijing.

TABLE II
O PERATION S ITUATION OF THE Y IZHUANG S UBWAY L INE

Fig. 10. Hourly variations of passenger demand on the Yizhuang Line and its
scheduling periods [16]. TABLE III
T IMETABLE OF THE Y IZHUANG S UBWAY L INE
It covers a length of 22.73 km and consists of 14 stations,
including Songjiazhuang, Xiaocun, Xiaohongmen, Jiugong,
Yizhuangqiao, Wenhuayuan, Wanyuan, Rongjing, Rongchang,
Tongjinan, Jinghai, Ciqunan, Ciqu, and Yizhuang (see Fig. 9).
Seventeen trains are assigned for operation each day, and each
train is composed of six car units. The maximum speed and
the average speed are 80 and 40 km/h, respectively. Generally,
the cyclic period of the timetable is 77 min, which contains
55.2-min running time, 12.8-min dwell time, and 9-min termi-
nal time (including the turn-back and self-checking time). In
addition, the total daily passenger demand ranges from 100 000
to 200 000, and the hourly passenger demand can reach 18 000
in peak hours (see the red columns in Fig. 10). For dealing with
the variations of the passenger demand, an hourly timetable
is used in the Yizhuang Subway Line. For example, 11 or 12
s1 and sj . The corresponding relative traction force is recorded
trains will serve in lines during peak hours, and 7 trains are in
as {kt (i)}, and the maximum traction force can be investigated
operation during off-peak hours (see Table II), whereas the trip
in [34]. Hence, the energy consumption for this interval can be
time at interstations does not change (see Table III).
calculated as
The current signaling system can record the operation data for
all trains, such as the train speed, position, and relative traction 
j−1
force, from which we can evaluate the energy consumption of Ep = kt (i) × F × (si+1 − si ). (23)
each train. For example, the train accelerates between positions i=1
SU et al.: OPTIMIZATION OF MULTITRAIN OPERATIONS IN A SUBWAY SYSTEM 681

TABLE IV performance on energy efficiency and is a good development of

O PTIMAL S CHEDULES FOR D IFFERENT T IME P ERIODS
the previous work.
Specifically, we take the time period between 7:00 A . M . and
8:00 A . M ., for example, to show where the energy savings come
from. In Table VI, the first two columns denote the practical en-
ergy consumption for each interstation. By applying the energy-
efficient train-control algorithm (the previous work), the driving
strategies are optimized for each interstation and the energy
is saved by 12.9% for 1 h (see the third and fourth columns
in Table VI). In addition, the result with only energy-efficient
control does not change the fleet size and the cycle time.
TABLE V Hence, the energy consumption for a single train is calculated as
C OMPARISON OF E NERGY C ONSUMPTION W ITH P REVIOUS W ORK 365.74 kW/h during a cycle time. The integrated approach
calculates the optimal fleet size as 12 with respect to 11 in
practice. It increases the number of service trains but enables a
longer cycle time such that the energy consumption for a single
train is only 251.94 kW/h. The energy savings for a single
train is more than 100 kW/h. Thus, the energy savings of 11
service trains can reach more than 1000 kW/h, which not only
compensates the energy of the extra service train but also leaves
much reserve energy savings. Consequently, the total energy
consumption of the whole system is efficiently reduced.
The optimal trip times for each interstation at different time
periods are shown in Table VII, and the optimal driving strategy
generated by the IEE algorithm is shown in Fig. 11.
From the simulation results, we can also find that the signif-
icant energy savings comes from longer cycle time and larger
fleet size. First, the longer trip time may bring a loss of travel
time to passengers. To deal with this, the constraints on the
cycle time [see (4)] can actually be variable with time when the
algorithm is applied in practice. For example, Tmax is larger
in peak hours than in off-peak hours such that the passengers’
By applying the proposed algorithm, we solve the optimal waiting time can be efficiently reduced for maintaining the
schedule for each time period (see Table IV), in which the quality of service. In addition, the additional trains efficiently
fleet size, the distribution of the cycle time, and the driving reduce the energy consumption of the system while this also
strategy are all optimized. (Note that the subway system must increases other operation costs (such as the maintenance cost).
offer a certain minimum frequency of service during off-peak Hence, whether the proposed approach can cut down the total
hours. Hence, for off-peak hours, we do not change the fleet operational cost of the system need a further improvement in
size and only optimize the trip time and driving strategy for the future work.
interstations.) According to the practical energy consumption Given the passenger demand, the computation time of gen-
Ep and the energy consumption Eo calculated by the IEE erating the IEE timetable is about 1.89 s on average. The
algorithm, we obtain energy reduction R as computation time is 0.15 s if the algorithm is applied to solve
the optimal driving strategy for a single train. Therefore, the
R = (Ep − Eo )/Ep × 100%. (24) algorithm is quick enough to be applied to the real-time control
system.
As shown in Table V, the proposed algorithm can reduce the
energy consumption by over 25.4% during peak hours and
VI. S ENSITIVITY A NALYSIS ON A SSUMPTIONS
15.9% during off-peak hours, and the energy reduction can
reach 24.0% for the whole day, which validates the good In this section, we provide some sensitivity analysis on the
energy-efficient performance of the IEE algorithm. In addition, assumptions that the turn-back time and the utilization of the
our previous work [35] has designed an integrated algorithm to maximum capacity are constant in this paper to illustrate that
optimize the operation for a single train in which the driving the assumptions are reasonable and the optimizing result is
strategy and the distribution of the cycle time (the two lower meaningful in practice.
layers) is calculated and fleet size is not considered. Based on Based on practical experience, the minimum automatic turn-
(3) and (24), we can calculate the energy consumption Eo of back time for a train is about 200 s. Hence, we analyze the
the whole system with the previous work for different time influence of the turn-back time on the energy consumption
periods and the corresponding energy reduction (see Table V). when the turn-back time ranges from 200 to 500 s (see the
It is obvious that the proposed integrated approach has a better left graph in Fig. 12), and the energy consumption calculated
682 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

TABLE VI
C OMPARISON OF E NERGY C ONSUMPTION A MONG THE C URRENT O PERATION , E NERGY-E FFICIENT R ESULTS
W ITH O NLY O PTIMAL C ONTROL , AND O PTIMAL R ESULTS F ROM THE I NTEGRATED A PPROACH

TABLE VII
O PTIMAL T IMETABLE

Fig. 11. Optimal driving strategy for the whole line from 9:00 to 10:00. Fig. 12. Sensitivity analysis on the turn-back time.
SU et al.: OPTIMIZATION OF MULTITRAIN OPERATIONS IN A SUBWAY SYSTEM 683

R EFERENCES
[1] R. Chevrier, “Multi-objective evolutionary algorithm for speed tuning op-
timization with energy saving in railway management,” in Proc. RailRome
Conf., 2010.
[2] I. P. Milroy, Aspects of Automatic Train Control. Leicestershire, U.K.:
Loughborough University, 1990.
[3] M. A. Shafia, M. P. Aghaee, S. J. Sadjadi, and A. Jamili, “Robust train
timetabling problem: Mathematical model and branch and bound algo-
rithm,” IEEE Trans. Intell. Transp. Syst., vol. 13, no. 1, pp. 307–317,
Mar. 2012.
[4] S. G. Gao, H. R. Dong, Y. Chen, B. Ning, G. R. Chen, and X. X. Yang,
“Approximation-based robust adaptive automatic train control: An approach
for actuator saturation,” IEEE Trans. Intell. Transp. Syst., to be published.
[5] E. Khmelnitsky, “On an optimal control problem of train operation,” IEEE
Trans. Autom. Control, vol. 45, no. 7, pp. 1257–1266, Jul. 2000.
[6] R. Liu and I. Golovitcher, “Energy-efficient operation of rail vehicles,”
Transport. Res. A, vol. 37, no. 10, pp. 917–932, Dec. 2003.
[7] A. Thomas, Railway Timetable and Traffic. Hamburg, Germany: Eurail-
press, 2008.
[8] M. Voorhoeve, “Rail scheduling with discrete sets,” Eindhoven University
of Technology, Eindhoven, The Netherlands, Tech. Rep., 1993.
[9] R. Vukan, Urban Transit: Operations, Planning and Economics.
Fig. 13. Sensitivity analysis on the utilization of the maximum capacity. Hoboken, NJ, USA: Wiley, 2004.
[10] L. Peeters, “Cyclic railway timetable optimization,” Ph.D. dissertation,
Erasmus University Rotterdam, Rotterdam, The Netherlands, 2003.
from the proposed model keeps constant at 2043 kW/h until [11] M. Pena-Alcaraz, A. Fernandez, A. P. Cucala, A. Ramos, and
R. R. Pecharroman, “Optimal underground timetable design based on
the turn-back time is over 475 s where the systematic energy power flow for maximizing the use of regenerative-braking energy,” Proc.
consumption gradually rises. Therefore, we conclude that the Inst. Mech. Eng., Part F, J. Rail Rapid Transit, vol. 12, no. 2, pp. 374–383,
turn-back time contributes a little to the energy consumption 2011.
[12] L. F. Li, H. Zhang, X. F. Wang, W. Lu, and Z. P. Mu, “Urban transit
with the available turn-back time. The practical turn-back time coordination using an artificial transportation system,” IEEE Trans. Intell.
will not be over 300 s and the assumption on the constant turn- Transp. Syst., vol. 12, no. 2, pp. 374–383, Jun. 2011.
back time is reasonable. [13] M. Wardman, J. Shires, W. Lythgoe, and J. Tyler, “Consumer benefits and
demand impacts of regular train timetables,” Int. J. Transp. Manag., vol. 2,
The influence of the utilization of the maximum capacity no. 1, pp. 39–49, 2004.
on the energy consumption is analyzed in Fig. 13 (see the [14] P. Serafini and W. Ukovich, “A mathematical model for periodic schedul-
right graph), which implies that the utilization of the maxi- ing problems,” SIAM. J. Discrete. Math., vol. 2, no. 4, pp. 550–581,
Nov. 1989.
mum capacity has a great influence on the systematic energy [15] W. S. Lin and J. W. Sheu, “Optimization of train regulation and energy
consumption. Specifically, the energy consumption is about usage of metro lines using an adaptive-optimal-control algorithm,” IEEE
2400 kW/h when the value of α is 0.7, as compared with Trans. Autom. Sci. Eng., vol. 8, no. 4, pp. 855–864, Oct. 2011.
[16] C. Gao and X. Xiao, “Passenger demand for Yizhuang subway line,” Tech.
1900 kW/h when the value of α is 0.9. Therefore, it is more Rep., 2012.
important to have an accurate estimation on α based on the [17] L. Kroon, G. Maróti, M. R. Helmrich, M. Vromans, and R. Dekker,
historical operational records, and this will have a significant “Stochastic improvement of cyclic railway timetables,” Transp. Res. B,
vol. 42, no. 6, pp. 553–570, Jul. 2008.
help to the operators to make an energy-efficient timetable. [18] L. Kroon and L. Peeters, “A variable trip time model for cyclic railway
However, in practical, the utilization of the maximum capacity timetabling,” Transp. Sci., vol. 37, no. 2, pp. 198–212, May 2003.
during a fixed time period will not vary a lot. For example, the [19] M. A. Odijk, “A constraint generation algorithm for the construction of
periodic railway timetables,” Transp. Res. B, vol. 30, no. 6, pp. 455–464,
experiential value of α is 0.75 for a given time interval, and then Dec. 1996.
the practical value ranging from 0.72 to 0.77 will not influence [20] M. A. Odijk, H. E. Romeijn, and M. H. Van, “Generation of classes of
the energy consumption of the system according to Fig. 13. robust periodic railway timetables,” Comput. Oper. Res., vol. 33, no. 8,
pp. 2283–2299, Aug. 2006.
Hence, the proposed approach that assumes α as constant also [21] C. Gabrio, L. Marco, S. Kaspar, W. Stefan, and F. Martin, “The periodic
has a great significance for optimizing the train operation. service intention as a conceptual framework for generating timetables
with partial periodicity,” Transp. Plan. Technol., vol. 34, no. 4, pp. 329–
339, 2011.
[22] C. Liebchen and L. Peeters, “Integral cycle bases for cyclic timetabling,”
VII. C ONCLUSION Discrete Optim., vol. 6, no. 1, pp. 98–109, Feb. 2009.
[23] K. Nachtigall, “A branch and cut approach for periodic network program-
This paper has formulated an IEE train operation model to ming,” Hildesheimer Informatik-Berichte, Univ. Hildesheim, Hildesheim,
optimize the timetable and driving strategy. This is an important Germany, Techn. Rep. 29, 1994.
improvement on energy-efficient train operation since it first [24] T. Albreche and S. Oettich, “A new integrated approach to dynamic sched-
ule synchronization and energy-saving train control,” Comput. Railways
integrates the fleet size and cycle time decision, distribution of VIII, pp. 847–856, 2002.
cycle time, and driving strategy optimization. An IEE algorithm [25] K. Ghoseiri, F. Szidarovszky, and M. J. Asgharpour, “A multi-objective
is designed to find the globally optimal schedule. The case train scheduling model and solution,” Transp. Res. B, Methodological,
vol. 38, no. 10, pp. 927–952, Dec. 2004.
study based on the operation data of Beijing Yizhuang Subway [26] C. S. Chang and S. S. Sim, “Optimising train movements through coast
Line shows that the proposed approach can reduce the energy control using genetic algorithms,” IEE Proc.-Elect. Pow. Appl., vol. 144,
consumption by 24.0% on average for the whole day. In addi- no. 1, pp. 65–73, Jan. 1997.
[27] J. Cheng and P. G. Howlett, “A note on the calculation of optimal strate-
tion, the computation time of the proposed algorithm is short gies for the minimization of fuel consumption in the control of trains,”
enough to apply the algorithm to the real-time control system. IEEE Trans. Autom. Contr., vol. 38, no. 11, pp. 1730–1734, Nov. 1993.
684 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

[28] P. G. Howlett, “Optimal strategies for the control of a train,” Automatica, Xiang Li received the B.S. degree from Jilin Uni-
vol. 32, no. 4, pp. 519–532, Apr. 1996. versity, Changchun, China, in 2004 and the Ph.D.
[29] P. G. Howlett, “The optimal control of a train,” Ann. Oper. Res., vol. 98, degree in operations research and cybernetics from
pp. 65–87, Dec. 2000. Tsinghua University, Beijing, China, in 2008.
[30] P. G. Howlett and P. J. Pudney, “Energy-efficient train control,” in Adv. He is a Professor with Beijing University of
Ind. Control. New York, NY, USA: Springer-Verlag, 1995. Chemical Technology, Beijing, China. He has au-
[31] P. G. Howlett, P. J. Pudney, and X. Vu, “Local energy minimization thored or coauthored more than 40 articles on in-
in optimal train control,” Automatica, vol. 45, no. 11, pp. 2692–2698, ternational journals, including Information Sciences,
Nov. 2009. European Journal of Operational Research, IEEE
[32] X. Li, C. F. Chien, L. Li, Z. Y. Gao, and L. X. Yang, “Energy-constraint T RANSACTIONS ON F UZZY S YSTEMS, and IEEE
operation strategy for high-speed railway,” Int. J. Innov. Comput., Inf. T RANSACTIONS ON I NTELLIGENT T RANSPORTA -
Control, vol. 8, no. 10(A), pp. 6569–6583, Oct. 2012. TION S YSTEMS.
[33] B. R. Ke, M. C. Chen, and C. L. Lin, “Block-layout design using max–min Dr. Li served as the Executive Editor-in-Chief of Journal of Uncertainty
ant system for saving energy on mass rapid transit systems,” IEEE Trans. Analysis and Applications and as an Editorial Board Member of International
Intell. Transp. Syst., vol. 10, no. 2, pp. 226–235, Jun. 2009. Journal of General Systems, International Journal of Advanced Intelligence
[34] M. Kondo, “Traction and braking performance of Beijing subway line,” Paradigms, and Journal of Applied Mathematics and Physics.
Dongyang Manufacturer of Motor, Incheon, Korea, Tech. Rep., 2009.
[35] S. Su, X. Li, T. Tang, and Z. Y. Gao, “A subway train timetable optimiza-
tion approach based on energy-efficient operation strategy,” IEEE Trans.
Intell. Transp. Syst., vol. 14, no. 2, pp. 883–893, Jun. 2013.
[36] A. Jamili, M. A. Shafia, S. J. Sadjadi, and R. Tavakkoli-Moghaddam,
“Solving a periodic single-track train timetabling problem by an efficient
hybrid algorithm,” Eng. Appl. Artif. Intel., vol. 25, no. 4, pp. 793–800, Ziyou Gao received the Ph.D. degree in operations
Jun. 2012. research and control theory from Chinese Academy
[37] I. Golovitcher, “Control algorithms for automatic operation of rail vehi- of Sciences, Beijing, China, in 1994.
cles,” Automat. Rem. Contr., vol. 11, pp. 118–126, 1986. From October 1993 to September 1995, he was
a Postdoctoral Researcher majoring in transporta-
tion management with the Transportation Simulating
Shuai Su received the B.S. degree in 2010 from Center, Beijing Jiaotong University, where he was
Beijing Jiaotong University, Beijing, China, where previously a Professor, then later a Doctor Advisor
he is currently working toward the Ph.D. degree in and currently the Director of the State Key Labora-
the State Key Laboratory of Rail Traffic Control and tory of Rail Traffic Control and Safety.
Safety. Dr. Gao is also the Vice-Chairman of the China
His research interests include energy-efficient op- Transportation Systems Engineering Society, an Advanced Member of the
eration and control in railway systems. China Academy of Railway Sciences, a Coeditor of Traffic and Transportation
Systems Engineer and Information, and a member of the editorial board of
Transportmetrica. He was also elected as a Foreign Member of the Russian
Academy of Natural Sciences in 2003.

Tao Tang received the Ph.D. degree from the

Chinese Academy of Sciences, Beijing, China,
in 1991.
He is the Academic Pacesetter with the National
Key Subject Traffic Information Engineering and
Control and the Director of the State Key Laboratory
of Rail Traffic Control and Safety, Beijing Jiaotong
University, Beijing. He is also a Specialist with the
National Development and Reform Commission and
the Beijing Urban Traffic Construction Committee.
His research interests include both high-speed and
urban railway train control systems, as well as intelligent control theory.