Transportation Research Part B 64 (2014) 7389

Contents lists available at ScienceDirect

Transportation Research Part B

journal homepage: www.elsevier.com/locate/trb

An energy-efcient scheduling and speed control approach

for metro rail operations
Xiang Li a,, Hong K. Lo b
a
b

School of Economics and Management, Beijing University of Chemical Technology, Beijing 100029, China
Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology, Hong Kong

a r t i c l e

i n f o

Article history:
Received 13 August 2013
Received in revised form 25 March 2014
Accepted 25 March 2014

Keywords:
Metro rail system
Energy-efcient operation
Timetable
Speed prole
Net energy consumption

a b s t r a c t
Due to increasing environmental concerns and energy prices, what is very important but
has not been given due consideration is the energy efciency of metro rail systems. Train
energy-efcient operation consists of timetable optimization and speed control. The former synchronizes the accelerating and braking actions of trains to maximize the utilization
of regenerative energy, and the latter controls the train driving strategy to minimize the
tractive energy consumption under the timetable constraints. To achieve a better performance on the net energy consumption, i.e., the difference between the tractive energy consumption and the utilization of regenerative energy, this paper formulates an integrated
energy-efcient operation model to jointly optimize the timetable and speed prole. We
design a genetic algorithm to solve the model and present some numerical experiments
based on the actual operation data of Beijing Metro Yizhuang Line of China. It is shown that
a larger headway leads to smaller energy saving rate, and the maximum energy saving rate
achieved is around 25% when we use the minimum allowable headway of 90 s. In addition,
compared with the two-step approach optimizing the timetable and speed prole separately, the integrated approach can reduce the net energy consumption around 20%.
2014 Elsevier Ltd. All rights reserved.

1. Introduction
The increased use of metro rail systems is often proposed as a solution to problems caused by urban trafc congestion.
Traditional deployment analysis typically covers rail line alignment, scheduling and capacity. Due to increasing environmental concerns and energy prices, what is very important but has not been given due consideration is the energy efciency of
metro rail systems. The emergent question is how the scheduling and speed control of metro rail system can be optimized
such that regenerative energy from a braking train can be used by nearby accelerating trains. Studies indicate that maximizing the use of such regenerative energy can lead to substantial improvements in the energy-efciency of the system.
Train energy-efcient operation is one of the most effective energy saving technologies, and generally consists of two levels. On the rst level, the trafc management develops a timetable, includes information on the number of service trains,
travel time at sections and dwell time at stations, which aims to improve the utilization of regenerative energy by synchronizing the operations of accelerating and braking trains (Yang et al., 2013). On the second level, the automatic train operation
system or the driver assistance system calculates the speed prole with minimum tractive energy consumption under the
Corresponding author. Address: Beijing University of Chemical Technology, School of Economics and Management, 15 BeiSanHuan East Road, Beijing
100029, China. Tel.: +86 15810979849.
E-mail addresses: lixiang@mail.buct.edu.cn (X. Li), cehklo@ust.hk (H.K. Lo).
http://dx.doi.org/10.1016/j.trb.2014.03.006
0191-2615/ 2014 Elsevier Ltd. All rights reserved.

74

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

timetable constraints (Su et al., 2013), and guides the drivers to track the speed prole under the supervision of a trafc management center.
Past studies typically consider these two levels separately. However, since timetable and speed prole are closely related and both of them have a direct inuence on the tractive energy consumption as well as the regenerative energy utilization, this two-step optimization method is defective from the view of global optimality on energy conservation.
Therefore, it is necessary to take into account the speed control and tractive energy consumption on the timetable optimization level. Recently, there have been some papers on minimizing the net energy consumption, i.e., the difference between the tractive energy consumption and the utilization of regenerative energy (Bocharnikov et al., 2010; Dominguez
et al., 2012; Rodrigo et al., 2013), but limitations are always observed. The complexity of the rail trafc management
system, related to the dynamic characteristic of the train synchronization problem, can be regarded as the main reason
(Tuyttens et al., 2013).
In order to achieve a better performance on energy saving, this paper proposes an integrated energy-efcient operation
model to jointly optimize the timetable and speed prole with minimum net energy consumption. We design a genetic algorithm and provide a real-life Beijing Metro Yizhuang Line simulation study to demonstrate the efciency of our approach.
Numerical experiments show that a larger headway leads to smaller energy saving rate, and the maximum energy saving
rate achieved is around 25% when we use the minimum allowable headway of 90 s. In addition, compared with the two-step
optimization method, our approach can reduce the net energy consumption around 20%.
The rest of this paper is organized as follows. In Section 2, we review some recent literature on train energy-efcient operation. In Section 3, we quantitatively analyze the tractive energy consumption, utilization of regenerative energy, and net
energy consumption, and formulate an energy-efcient scheduling and speed control model to minimize the net energy consumption. In Section 4, we design a genetic algorithm to obtain a satisfactory solution. In Section 5, we present some numerical experiments based on the actual operation data of Beijing Metro Yizhuang Line of China. At the end of this paper, a brief
summary is given.
2. Literature review
Train operation strategy, consisting of the timetable and speed prole, has a great inuence on the amount of energy consumption. As one of the most efcient energy saving technologies, train energy-efcient operation employing the mathematical optimization method to seek the operation strategy with better energy performance, has been widely studied and well
applied in metro rail system.
Train energy-efcient operation is a very complicated and difcult optimization problem because not only the timetable should be well dened for synchronizing the accelerating and braking actions of trains in the same substation, but also
the speed prole should be controlled to reduce the tractive energy consumption under the speed limits, travel time and
distance constraints (Tuyttens et al., 2013). Therefore, the traditional studies usually divide it into a timetable optimization problem and a speed prole optimization problem. The former adjusts the number of service trains, travel time at
sections and dwell time at stations to maximize the synchronization time between accelerating trains and braking trains
such that the regenerative energy from braking trains can be effectively utilized by the accelerating trains, while the latter
controls the accelerating, cruising, coasting and braking time and speed such that the tractive energy consumption can be
minimized.
Many metro systems have a cyclic timetable, which requires trains to follow each other with a xed headway, and take
the same dwell time at stations and travel time at sections. Voorhoeve (1993) rst considered a Periodic Event Scheduling
Problem (PESP) based model for the cyclic timetable problem, which was followed by many researchers, e.g., Kroon et al.
(2008) and Cordonea and Redaellib (2011).
Regenerative braking is an energy recovery mechanism which can convert the kinetic energy into electricity during the
braking phase by using the electric motor as an electric generator. Although regenerative braking can recover about 40% of
the tractive energy, little work observed in the literature studies its efcient utilization method. Ramos et al. (2007) proposed
a timetable optimization model to maximize the overlapping time between accelerating and braking actions of trains in the
same substation. Nasri et al. (2010) studied the inuence of headway on the utilization of regenerative energy. They veried
that the utilization of regenerative energy is higher if the minimum acceptable headway is used because of the simultaneous
presence of braking and accelerating trains. Pena-Alcaraz et al. (2012) designed a mathematical optimization model to synchronize the braking trains with the accelerating trains for improving the utilization of regenerative energy. David et al.
(2012) formulated an optimization model to maximize the utilization of regenerative energy by subtly modifying dwell time
for trains at stations. A hybrid genetic/linear programming algorithm was implemented to tackle this problem. Yang et al.
(2013) proposed a train cooperative scheduling rule to synchronize the accelerating and braking actions of successive trains.
Then they formulated a cooperative scheduling model to maximize the overlapping time, and designed a genetic algorithm
to solve the optimal timetable. A real-life Beijing Metro Yizhuang Line simulation study showed that the cooperative scheduling model can signicantly improve the overlapping time around 22%. Furthermore, Li and Yang (2013) proposed a stochastic cooperative scheduling model focusing on the randomness of departure delay for trains at busy stations, which
could save energy around 8% compared with the cooperative scheduling approach. When maximizing the utilization of
regenerative energy in timetable optimization model, we should consider the possible increase on tractive energy consumption. Otherwise, the total energy efciency may be reduced if the increase on tractive energy consumption is larger than the

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

75

utilization on regenerative energy. As an extension of the above works, this paper will propose an integrated energy-efcient
operation model which considers both the tractive energy and the regenerative energy.
Literature on the speed prole optimization can date back to 1960s. Ishikawa (1968) proposed the rst optimal control
model to determine the speed prole. Asnis et al. (1985) assumed that the acceleration rate was a continuous control variable with uniform bounds, and used the Pontryagin maximum principle to nd the necessary conditions on an optimal control strategy. For seeking a more rigorous mathematical justication, Howlett (1988) proved the existence of an optimal
strategy within the framework of functional analysis theory and then produced the rst theoretical conrmation that an
optimal speed prole should use an maximum accelerationcruisingcoastingmaximum braking phase sequence (Howlett,
1990), which lays the foundation for the continuous optimal train control theory. Benjamin et al. (1989) considered the discrete optimal train control problem on a typical diesel-electric locomotive, in which only a nite number of acceleration rate
is available. The theoretical basis for the discrete optimal train control was given by the Scheduling and Control Group at the
University of South Australia. For example, Howlett et al. (1994) outlined the theoretical basis for Metromiser system and
Long-haul fuel conservation system, which provides driving advice for suburban railways and long-haul rail networks, respectively. A good survey on the optimal train control theory was given by Howlett and Pudney (1995). Howlett (2000) considered the problem of determining an optimal strategy with a generalized equation of motion, and proved that the optimal
strategy for discrete control can be used to approximate as closely as we please the optimal strategy obtained using continuous control. For taking the theory into practice, gradients, speed limits, track curvature and traction efciency were considered (Khmelnitsky, 2000), which makes the optimal control problem more difcult. Liu and Golovitcher (2003) gave an
analytical approach to solve the switching time among different phases. Howlett et al. (2009) provided another analytical
method on the consideration of steep slopes. They proved that the cruising phase must be interrupted by phases of accelerating for steep uphill sections and coasting for steep downhill sections. Albrecht et al. (2013) proved that the optimal
switching points are uniquely determined for each steep section and the global optimal strategy is also unique. Su et al.
(2013) designed an iterative algorithm to solve the energy-efcient speed prole for a single train along the whole line. A
real-life Beijing Metro Yizhuang Line simulation study was given, which illustrates that the tractive energy consumption
can be reduced around 14%.
Except these exact solution methods, there are also some studies using the evolutionary algorithms. For example, Chang and
Sim (1997) applied the genetic algorithm to generate an optimal coast control with predetermined number of coasting points.
Ke et al. (2009) designed an ant colony optimization algorithm to seek the energy-efcient speed prole between successive
stations for metro rail systems. Kim and Chien (2011) developed a simulated annealing algorithm to search for the optimal
train operation by considering track alignment, speed limit, and schedule adherence. Ding et al. (2011) formulated the train
energy-efcient operation as a two-level optimization problem and designed a genetic algorithm to search for the optimal
solution. The rst level is designed to decide the appropriate coasting point of inter-station run for trains, and the second level
arranges the train travel time of inter-station run for minimizing the tractive energy consumption. Lu et al. (2013) proposed a
distance-based train speed prole searching model under the timetable, traction equipment characteristics, speed limits and
gradients constraints, ant colony optimization algorithm, genetic algorithm and dynamic programming are respectively applied to search for the optimal speed prole. The above studies on speed prole optimization lay the foundation for this paper.
Inspired with the analysis on the optimal phase sequence and the evolutionary algorithms for nding the optimal switching
points between different phases, this paper will design a genetic algorithm to solve the optimal speed prole.
Recently, some studies are observed on speed prole optimization with minimum net energy consumption. For example,
Bocharnikov et al. (2010) presented a single train speed prole optimization model which considers both the tractive energy
consumption and the utilization of regenerative energy. Furthermore, the authors performed a multi-train simulation to estimate the benets and effects of the optimal speed prole on minimizing the net energy consumption. Dominguez et al.
(2012) took into account the storage of regenerative energy in substations and designed a genetic algorithm to solve the optimal speed prole with minimum net energy consumption. Based on the assumption that all trains must obey the predetermined timetable, Tuyttens et al. (2013) designed a genetic algorithm to determine the speed proles for all trains running
within a specied period. In order to provide a real-time trafc control strategy, the empirical driving modes are used in the
solution of optimal speed prole. Rodrigo et al. (2013) designed a Lagrangian multipliers method to solve the speed prole
optimization model with maximum utilization of regenerative energy.
A set of studies have been focused on the train energy-efcient operation, but very few of them take into account the
combination of timetable optimization and speed prole optimization. In order to achieve a better performance on energy
saving, this paper proposes an integrated energy-efcient operation model and designs a genetic algorithm to optimize the
timetable and speed prole with minimum net energy consumption. A comparison between our work and the existing literature is given by Table 1.

3. Integrated energy-efcient operation model

Net energy consumption is the difference between the tractive energy consumption and the utilization of regenerative
energy, which is the electricity absorbed by trains from the substations (see Fig. 1). This section formulates an integrated
energy-efcient operation model to determine the timetable and speed prole with minimum net energy consumption.
For better understanding of this paper, the notation system is rst introduced as follows.

76

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

Table 1
Recent publications on train energy-efcient operation and our work.
Publication

Decision variable

Objective function

Algorithm

Ramos et al. (2007)

Pena-Alcaraz et al. (2012)
David et al. (2012)
Yang et al. (2013)
Li and Yang (2013)
Howlett (2000)
Khmelnitsky (2000)
Liu and Golovitcher (2003)
Howlett et al. (2009)
Albrecht et al. (2013)
Su et al. (2013)
Chang and Sim (1997)
Ke et al. (2009)
Kim and Chien (2011)
Ding et al. (2011)
Lu et al. (2013)

Timetable
Timetable
Dwell time
Timetable
Timetable
Speed prole
Speed prole
Speed prole
Speed prole
Speed prole
Speed prole
Coasting point
Speed prole
Speed prole
Coasting point + Timetable
Speed prole

Synchronization time
Regenerative energy utilization
Regenerative energy utilization
Synchronization time
Regenerative energy utilization
Tractive energy consumption
Tractive energy consumption
Tractive energy consumption
Tractive energy consumption
Tractive energy consumption
Tractive energy consumption
Tractive energy consumption
Tractive energy consumption
Tractive energy consumption
Tractive energy consumption
Tractive energy consumption

Bocharnikov et al. (2010)

Dominguez et al. (2012)
Tuyttens et al. (2013)
Rodrigo et al. (2013)

Speed
Speed
Speed
Speed

Net
Net
Net
Net

CPLEX Solver
SBB Solver
Genetic/linear algorithm
Genetic algorithm
Genetic algorithm
Analytical method
Analytical method
Analytical method
Analytical method
Analytical method
Iterative algorithm
Genetic algorithm
Ant colony optimization
Simulated annealing
Genetic algorithm
Ant colony optimization
Genetic algorithm
Dynamic programming
Genetic algorithm
Genetic algorithm
Genetic algorithm
Lagrangian multipliers

Our paper

Timetable + Speed prole

prole
prole
prole
prole

energy
energy
energy
energy

consumption
consumption
consumption
consumption

Net energy consumption

Genetic algorithm

Electric power network

Conversion loss

Substation
Diode
Transmission loss

Overhead contact line

Regenerative energy

Tractive energy
Accelerating
train

Yes
V < MTV

No

Braking train
Resistor

Track
Fig. 1. An illustration on energy ow model.

3.1. Notation
The notation system includes the indices, parameters and decision variables.
Indices and parameters
t
time index, t 1; 2; . . . ; T
k
train index, k 1; 2; . . . ; K
i
cycle index, i 1; 2; . . . ; I
n
segment index, n 1; 2; . . . ; N
j
substation index, j 1; 2; . . . ; J
I
collection of stations, I # f1; 2; . . . ; Ng
n
w
maximum dwell time at station n
wn
minimum dwell time at station n
L
passenger demand
l
train capacity
j
train capacity utilization rate
h
headway

Conversion loss

Net energy

Transmission loss

77

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

un
pn
A
B
Rn

speed limit at segment n

starting position of segment n
the maximum tractive force per unit mass for trains
the maximum braking force per unit mass for trains
the running resistance per unit mass for trains at segment n
acceleration rate for trains at segment n during accelerating phase
deceleration rate for trains at segment n during coasting phase
deceleration rate for trains at segment n during braking phase
conversion factor from electricity to kinetic energy
conversion factor from kinetic energy to regenerative electricity
transmission loss factor of regenerative energy
turnaround time
train mass
taking value 1 if segment n is located in substation j; otherwise, taking value 0

an
bn

cn
h1
h2
h3

r
M
dn; j

Decision variables
mn
the maximum speed for trains at segment n
qn
speed for trains at position pn
akin
arriving time for train k at the nth segment of cycle i
ckin
switching time from accelerating phase to coasting phase for train k at the nth segment of cycle i
bkin
switching time from coasting phase to braking phase for train k at the nth segment of cycle i
ekin
departuring time for train k at the nth segment of cycle i
v kin
speed prole for train k at the nth segment of cycle i
C
cycle time, i.e., the period required to complete one cycle for one train
I
the number of cycles
K
the number of trains

For metro rail system, it is common that sections have piece-wise speed limits for trains arising from curvature, signaling
and track condition that make higher speeds hazardous. In this paper, we will divide the total route into several segments
such that each segment has a constant speed limit (see Fig. 2). First, we assemble the turnaround stations, stations and sections from the up direction and down direction as a cycle. Furthermore, we divide each section into several segments such
that each segment has a constant speed limit. Finally, we consider stations and turnaround stations as special segments with
speed limit zero. Assume that there are N segments in the cycle. Except the turnaround stations, each physical segment is
counted twice by distinguishing the up direction and the down direction. Therefore, N is an even number, and segment
N=2 and segment N are the turnaround stations.
Headway is determined by the passenger demand L, train capacity l and its utilization rate j. If we use u to denote the
number of service trains per hour, we have

L l  j  u:

3:1

Furthermore, it follows from h 3600=u that the headway is

3600  l  j
:
L

3:2

The train capacity l and its utilization rate j are both constant parameters. Since the passenger demand L is a time-varying
quantity, trains should have different headway at different periods.

Segment

Station

N-1

N-3

N/2-4

Section

N-2

N-4

N/2-3

N/2-2

Section

N/2+4

N/2+3

N/2-1

Station

N/2+2

Fig. 2. An illustration on stations, sections and segments.

N/2+1

N/2

Station
Turnaround

Turnaround
Station
Segment

N/2

78

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

Generally speaking, the headway changes according to the partition of peak hours and off-peak hours for busy lines, while
keeps unchanged for non-busy lines. In this paper, we assume that the headway is a constant number. For the busy lines, we
can decide the operation strategy for peak hours and off-peak hours respectively with different headway.
According to the optimal train control theory (Howlett, 2000; Khmelnitsky, 2000; Liu and Golovitcher, 2003; Howlett
et al., 2009), the train energy-efcient movement at each segment consists of maximum acceleration, cruising, coasting,
and maximum braking. However, for trains with short travel distance, such as metro trains, the energy-efcient movement
only contains accelerating, coasting and braking phases (Howlett and Pudney, 1995; Bocharnikov et al., 2010; Su et al., 2013).
Then for each segment with nonzero speed limit, we should rst draw trains to the maximum speed mn with acceleration
rate an , then coast for a while without external force, and nally brake with deceleration rate cn . In order to solve the cycle
time C, the number of trains K, the number of cycles I and the speed proles v kin t with 1 6 k 6 K; 1 6 i 6 I and 1 6 n 6 N,
we need to determine the switching time akin ; ckin ; bkin ; ekin among different phases and segments (see Fig. 3). Since the
switching time ckin and bkin can be uniquely determined by speeds mn and qn , we denote the decision variables as these
switching time and speeds

x fakin ; ekin ; qn ; mn j 1 6 k 6 K; 1 6 i 6 I; 1 6 n 6 Ng:

For each 1 6 k 6 K, we have ekin akin1 for all 1 6 i 6 I and 1 6 n < N, and ekiN aki11 for all 1 6 i < I.
Remark 3.1. Assume that all trains are identical. The periodicity of the metro rail operation is reected in the following three
aspects: (i) trains follow each other with a xed headway; (ii) each train repeats the same operation at different cycles; and
(iii) different trains comply with the same operation strategy except a headway difference.
3.2. Speed prole
Let x be a decision vector. For each 1 6 k 6 K; 1 6 i 6 I and 1 6 n 6 N, we consider the speed prole for train k at segment
n of the ith cycle. The argument breaks down into two cases according to the segment type. If segment n is a station or turnaround station, it follows from un 0 that

v kin x; t 0; 8akin 6 t 6 ekin :

3:3

Otherwise, segment n is in a section with nonzero speed limit. According to the train motion equation, we have

8
if akin 6 t < ckin ;
>
< qn an t  akin ;
v kin x; t > mn  bn t  ckin ; if ckin 6 t < bkin ;
:
qn1 cn ekin  t; if bkin 6 t 6 ekin :

3:4

The rst row denotes the accelerating prole with acceleration rate an and initial speed qn , the second row denotes the coasting prole with deceleration rate bn and initial speed mn , and the last row denotes the braking prole with deceleration rate
cn and terminal speed qn1 .
It follows from the continuity of the train speed that the accelerating prole and the coasting prole should intersect at
ckin and the coasting prole and the braking prole should intersect at bkin . Then switching time ckin and bkin should satisfy the
following linear equations

qn an ckin  akin mn  bn ckin  ckin ;

mn  bn bkin  ckin qn1 cn ekin  bkin ;

Fig. 3. An illustration on decision variables.

3:5

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

79

which implies that they can be expressed in terms of decision vector x and given parameters as follows

ckin akin mn  qn =an ;


bkin cn ekin  mn qn1  bn ckin cn  bn :

3:6

Note that an A  Rn ; bn Rn and cn B Rn , where Rn is the resistance arising from the gradient, mechanical friction and
aerodynamic friction for trains at segment n.
In practice, the maximum tractive force, the maximum braking force and the running resistance all depend on the train
speed. However, in the context of this study, it is sufcient to assume constant maximum tractive force A, maximum braking
force B and running resistance Rn for calculating the approximate energy use and thereby synchronizing the accelerating and
braking patterns to conserve the net energy consumption. As a result, the acceleration rate an and the deceleration rates bn
and cn are all constant numbers.
3.3. Net energy consumption
This section analyzes the net energy consumption for trains during the daily operation period, i.e., the difference between
the tractive energy consumption and the utilization of regenerative energy. First, we analyze the energy consumption for
accelerating trains. For each akin 6 t < ekin with 1 6 n 6 N; 1 6 i 6 I and 1 6 k 6 K, according to the mechanical power equation, the required electricity for accelerating train k at time unit t; t 1 is


fkin x; t

MAv kin x; t=h1 ; if akin 6 t < ckin ;

0;

if ckin 6 t < ekin ;

3:7

where M is the train mass, A is the maximum tractive force per unit mass, and h1 is the conversion factor from electricity to
kinetic energy. Since there is no energy consumption during the coasting and braking phases, fkin x; t takes value zero when
ckin 6 t < ekin . Taking summation on fkin x; t with t and n, the electricity used for accelerating train k at the ith cycle is

F ki x

N eX
kin 1
X
fkin x; t:

3:8

n1 takin

Furthermore, taking summation on F ki x with i and k, the tractive energy consumption for all trains at all cycles is

Fx

K X
I X
N eX
kin 1
X
fkin x; t:

3:9

k1 i1 n1 takin

It follows from the periodicity of metro rail operation that the same train on the same segment will consume the same
amount of tractive energy during different cycles. Then each train will consume the same amount of electricity during different cycles. On the other hand, since trains are assumed to be identical, different trains will consume the same amount of
electricity at the same cycle. Based on the above analysis, we have

Fx K  I  F 11 x:

3:10

Now, we analyze the utilization of regenerative energy. The following two points should be noted. First, the regenerative
energy from braking trains can only be absorbed by trains which are accelerating. Second, the regenerative energy from
braking trains can only be transmitted among trains which are located in the same substation. For each akin 6 t < ekin with
1 6 n 6 N; 1 6 i 6 I and 1 6 k 6 K, according to the mechanical power equation, the electrical energy regenerated from train
k at time unit t; t 1 is

g kin x; t

0;

if akin 6 t < bkin ;

MBv kin x; th2 1  h3 ; if bkin 6 t < ekin ;

where B is the maximum braking force per unit mass, h2 is the conversion factor from kinetic energy to regenerative electricity, and h3 is the transmission loss factor of the regenerative energy. Since there is no regenerative energy during the
accelerating and coasting phases, g kin x; t takes value zero when akin 6 t < bkin . If there are other trains accelerating in the
same substation with train k, the regenerative energy will be transmitted to these trains via the overhead contact line. Otherwise, it will be dissipated by the onboard resistor. Divide the time interval 0; T into I 1 subintervals as follows

0; T 0; C [ C; 2C [    [ IC  C; IC [ IC; T:
First, we analyze the regenerative energy utilization at subinterval 0; C. In this case, trains either operate in their rst
cycle or stop at the depot. If we use nkt to denote the operation segment for train k at time t, we have

nkt

0; if 0 6 t < ak11 ;
n; if ak1n 6 t < ek1n

for all 1 6 n 6 N. Here nkt 0 denotes that train k stops at the depot at time t. For each 1 6 t < C and 1 6 j 6 J, the regenerative energy utilization for trains in substation j at time unit t; t 1 is

80

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

G1tj x min

(
K
X

f k1nkt x; tdnkt ; j;

k1

K
X

)
g k1nkt x; tdnkt ; j ;

k1

where we assume fk10 x; t g k10 x; t 0 for all 1 6 k 6 K. Here, the Kronecker delta dnkt ; j denotes whether segment nkt is
located in substation j, the rst summation denotes the tractive energy consumption for accelerating trains at substation j,
and the second summation denotes the amount of energy regenerated from braking trains at substation j. Taking summation
on G1tj x with j and t, the utilization on regenerative energy during 0; C can be formulated as

G1 x

J
C1 X
X
G1tj x:
t0 j1

Furthermore, for each 1 6 i 6 I  1, we analyze the regenerative energy utilization at subinterval iC; iC h. In this case,
except the rst train enters its i 1th cycle, other trains are all in the ith cycle. Then we have

n1t n; if a1i1n 6 t < e1i1n ;

nkt n; if akin 6 t < ekin

for all 1 6 n 6 N and 2 6 k 6 K. The regenerative energy utilization for trains in substation j at interval t; t 1 is

Gi1tj x min

(
K
X

)
K
X
f kinkt x; tdnkt ; j f1i1n1t x; tdn1t ; j;
g kinkt x; tdnkt ; j g 1i1n1t x; tdn1t ; j

k2

k2

for each iC 6 t < iC h and 1 6 j 6 J. Since the rst train is in a different cycle, the amounts of tractive energy consumption
and regenerative energy are calculated separately. Taking summation on Gi1tj x with t and j, we get the regenerative energy utilization for trains at subinterval iC; iC h. For each subinterval iC kh; iC k 1h with 1 6 k < K, considering
the k 1th train as the rst train, it is easy to see that the regenerative energy utilization at this subinterval is the same
as at subinterval iC; iC h. Therefore, the regenerative energy utilization during iC; iC C is

Gi1 x K 

iCh1
X
tiC

J
X
Gi1tj x:
j1

Again, it follows from the periodicity of metro rail operation that the same train on the same segment will consume the same
amount of tractive energy and regenerate the same amount of energy during different cycles. Hence, we have
G2 x G3 x    GI x.
Finally, we analyze the regenerative energy utilization at subinterval IC; T. In this case, trains are either in their Ith cycle
or at the depot. For each 1 6 k 6 K, it is easy to prove that

nkt

n; if akIn 6 t < ekIn ;

3:11

0; if ekIN 6 t 6 T

for all 1 6 n 6 N. The utilization on regenerative energy during this period can be formulated as

GI1 x

J
T1 X
X

(
)
K
K
X
X
min
fkInkt x; tdnkt ; j;
g kInkt x; tdnkt ; j ;

tIC j1

k1

k1

where fkI0 x; t g kI0 x; t 0 for all 1 6 k 6 K.

Based on the above analysis, the total regenerative energy utilization for trains at operation period 0; T is

Gx G1 x I  1  G2 x GI1 x:
In conclusion, the net energy consumption, i.e., the difference between the tractive energy consumption and the regenerative energy utilization, is

Ex Fx  Gx:

3:12

3.4. Constraints
In this paper, we consider the entire operation process, i.e., the period from the time when the rst train is put into operation to the time when the last train returns back to the depot. During the operation period, all trains run the same number
of cycles. It follows from the denition of headway that each cycle is divided into K units of headway time, and after the rst
train returns back to the depot, additional K  1 units of headway time is needed for the other K  1 trains return back to the
depot. The number of service trains and cycle time should satisfy the following constraints

jeKIN  a111  Tj 6 D;
C Kh; IC K  1h eKIN  a111 ;

3:13
3:14

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

81

where D is a given positive number. According to the periodicity of metro rail operation, the switching time should satisfy
the following constraints

81 6 k 6 K  1; 1 6 i 6 I; 1 6 n 6 N;
81 6 k 6 K; 1 6 i 6 I:

ak1in  akin h;

3:15

ekiN  aki1 C;

3:16

The turnaround time needed for trains change from up/down direction to down/up direction at segments N=2 and N is
assumed to be a constant number since the operation is generally completed by the automatic train operation system,
see Fig. 4. If we use r to denote the turnaround time, we have

ekiN=2  akiN=2 r;

ekiN  akiN r;

81 6 k 6 K; 1 6 i 6 I:

3:17

If we use un to denote the speed limit for trains at the nth segment, the boundary speeds qn ; qn1 and the maximum speed
mn for trains at this segment should satisfy

0 6 qn ; mn ; qn1 6 un ;

81 6 n 6 N;

3:18

where qN1 0. Especially, since un 0 for each station n 2 I, we have qn mn qn1 0. For any segment 1 6 n 6 N with
nonzero speed limit, in order to ensure that trains can complete the transportation task safely and smoothly, the boundary
speeds qn and qn1 should satisfy the following inequalities

u2n  q2n u2n  q2n1

u2  q2n u2n  q2n1

6 pn1  pn 6 n

:
2an
2cn
2an
2bn

3:19

The left inequality ensures that trains can be controlled to arrive at position pn1 with speed less than qn1 , and the right
inequality ensures that trains can be controlled to arrive at position pn1 with speed more than qn1 . In general, Eq. (3.19)
ensures that speed variables mn ; qn and qn1 are feasible.
For each segment 1 6 n 6 N with xed boundary speeds qn and qn1 satisfying conditions (3.19), the minimum travel time
is obtained by accelerating the train to un and then coasting and braking to segment n 1 with speed qn1 , and the maximum
travel time is obtained by accelerating the train to certain speed and then coasting to segment n 1 with speed qn1 . Therefore, the travel time for trains at segment n should satisfy the following constraints

8
s


>
>
an bn q2n q2n1
q
q
>
> ekin  akin 6
 n  n1 ;

2p

p

>
n1
n
<
an bn an bn
an bn
s


>
>
u

q
u
q
cn  bn an bn 2 q2n q2n1
>
n
n
n
n1
>e  a P
>



u



2p

p

:
kin
: kin
n1
n
an
bn
cn
bn cn
an bn n an cn

3:20

The rst inequality is the maximum travel time constraint, and the second inequality is the minimum travel time constraint.
In general, Eq. (3.20) ensures that time variables akin and ekin are feasible. Finally, with feasible boundary speeds qn and qn1
satisfying conditions (3.19) and switching time akin and ekin satisfying conditions (3.20), the braking speed zn and the maximum speed mn should satisfy the travel distance constraint

m2n  q2n m2n  z2n z2n  q2n1

pn1  pn ;
2an
2bn
2cn

3:21

where the rst friction is the accelerating distance, the second fraction is the coasting distance, the last fraction is the braking
distance, and pn1  pn is the length of segment n. At the same time, they should also satisfy the travel time constraint

mn  qn

an

mn  zn zn  qn1

ekin  akin ;
bn
cn

3:22

where the rst friction is the accelerating time, the second fraction is the coasting time, the last fraction is the braking time,
and ekin  akin is the total travel time at segment n.
The dwell time for trains at stations should satisfy a time window constraint according to the passenger demand and the
operation time for opening/closing screen door. In general, for each 1 6 k 6 K; 1 6 i 6 I and n 2 I, we have

 kin :
wkin 6 ekin  akin 6 w

3:23

Upline

Downline
Fig. 4. An illustration on turnaround operation.

82

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

Finally, for each 1 6 k 6 K; 1 6 i 6 I and 1 6 n 6 N, the switching time should be integer numbers, that is,

akin ; ekin 2 f0; 1; 2; . . . ; Tg:

3:24

3.5. Optimization model

Based on the above analysis, we propose the following integrated energy-efcient operation model

min Ex;
x

s:t:

3:25

Constraints 3:133:24:

4. Genetic algorithm
Genetic algorithm is a stochastic search method for optimization problems based on the mechanics of natural selection
and natural genetics. In the past decades, it has obtained considerable success in providing satisfactory solutions to many
operations management problems for railway trafc (Yang et al., 2013; Tuyttens et al., 2013; Mazloumi et al., 2012; Dominguez et al., 2012; David et al., 2012; Bagloee and Ceder, 2011; Ding et al., 2011; Bocharnikov et al., 2010). In this section, we
design a genetic algorithm to solve the integrated energy-efcient operation model (3.25).
4.1. Representation structure
The rst problem for genetic algorithm is how to construct a one to one mapping between the solution space and the
chromosome space such that the crossover and mutation operations can be simplied. In this paper, a chromosome is dened as a N  5-dimensional matrix X, in which Xn; 1 a11n ; Xn; 2 qn ; Xn; 3 mn ; Xn; 4 qn1 and Xn; 5 e11n
for all 1 6 n 6 N.
Conversely, let X be a feasible chromosome such that XN; 5  X1; 1 is the integer times of headway h. Then the cycle
time is C XN; 5  X1; 1, the number of service trains is K C=h, and the number of cycles is

K 1; if T > KC D;
K;
if T 6 KC D;

4:1

where K bT  K  1h=C c. Furthermore, according to the periodicity, for each 1 6 k 6 K; 1 6 i 6 I and 1 6 n 6 N, we

have

akin Xn; 1 k  1h i  1C;

ekin Xn; 5 k  1h i  1C:

4:2

4.2. Initialization
For each 1 6 n 6 N, we rst randomly generate the boundary speeds Xn; 2 and Xn; 4 satisfying (3.18) and (3.19). Furthermore, we randomly generate the switching time Xn; 1 and Xn; 5 according to the following three cases:
 if segment n is a station, then Xn; 1 and Xn; 5 should satisfy the dwell time constraint (3.23);
 if segment n is a turnaround station, then Xn; 1 and Xn; 5 should satisfy constraint (3.17);
 if segment n has a nonzero speed limit, then Xn; 1 and Xn; 5 should satisfy constraint (3.20).
Finally, we calculate the maximum speed Xn; 3 satisfying Eqs. (3.21) and (3.22).
Note that X may be infeasible. Therefore, we need to update X according to the following iterative process: dene the
number of service trains as K bXN; 5  X1; 1=hc; randomly select a segment n excluding the turnaround stations such
that Xn; 5  Xn; 1  1 is a feasible operation time; set Xk; 5 Xk; 5  1 for all k P n and Xk; 1 Xk; 1  1 for all
k > n; do above procedure XN; 5  X1; 1  Kh times. After this process, we have XN; 5  X1; 1 Kh. For simplicity,
we call this process as Regulating Process.
Dene pop-size as the size of population. Based on above process, randomly generate pop-size chromosomes as the initialized population. The initialization process is summarized as follows:
Step
Step
Step
Step

1.
2.
3.
4.

Set i = 1.
Randomly generate a feasible chromosome X i .
If i < pop-size, set i i 1. Go to step 2.
Return the initialized population X i ; i 1; 2; . . . ; pop-size.

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

83

4.3. Evaluation function

Evaluation function assigns each chromosome a probability of reproduction so that its likelihood of being selected is proportional to its tness relative to the other chromosomes in the population. That is, the chromosomes with higher tness will
have more chance to produce offspring. For each chromosome X, the tness is dened as the objective value, that is, the net
energy consumption.
Without loss of generality, we assume that these pop-size chromosomes have been rearranged from good to bad according
to their tness values such that X 1 is the chromosome with the smallest net energy consumption, and X pop-size is the chromosome with the largest net energy consumption. For each a 2 0; 1, we dene the rank-based evaluation function as follows,

Ev alX i a1  ai1 ;

i 1; 2; . . . ; pop-size:

4:3

The evaluation process is summarized as follows:

Step
Step
Step
Step

1.
2.
3.
4.

Initialize a real number a 2 0; 1.

Calculate the tness values for all chromosomes X i ; i 1; 2; . . . ; pop-size.
Reorder these chromosomes from good to bad according to these tness values.
Calculate the evaluation values for all chromosomes

Ev alX i a1  ai1 ;

i 1; 2; . . . ; pop-size:

4.4. Selection process

During each successive generation, a proportion of the existing population is selected to breed a new generation. The
selection process is based on spinning the roulette wheel pop-size times, and selecting a chromosome each time. The roulette
wheel is a tness-proportional selection, where tter chromosomes are typically more likely to be selected. The selection
process is summarized as follows.
Step 1. Calculate the reproduction probability

r0 0;

ri

i
X
Ev alX j ;

i 1; 2; . . . ; pop-size:

j1

Step 2. Randomly generate a real number r from 0; r pop-size .

Step 3. Select the chromosome X i such that ri1 < r 6 ri .
Step 4. Repeat the second and third steps pop-size times and obtain pop-size chromosomes.
4.5. Crossover operation
Crossover is one of the mainly used operations for generating a second population. First, we dene a parameter pc to denote the crossover probability. Repeat the following process pop-size times: generate a random number r from 0; 1, and select the chromosome X i if r < pc . Note that not all chromosomes can be selected, but different chromosomes have the equal
chance to be selected. Without loss of generality, assume that chromosomes X 1 ; X 2 ; . . . ; X c are selected and c is an even number. Divide them into the following pairs:

X 1 ; X 2 ; X 3 ; X 4 ; . . . ; X c1 ; X c :
We illustrate the crossover operation by parents X 1 ; X 2 . Denote the children as Y 1 and Y 2 . For each 1 6 n 6 N, generate a
random number kn from the unit interval 0; 1 such that Y 1 n;2;Y 1 n; 4 kn X 1 n;2;X 1 n; 4 1  kn X 2 n;2;X 2 n; 4
and Y 2 n; 2; Y 2 n; 4 1  kn X 1 n; 2; X 1 n; 4 kn X 2 n; 2; X 2 n; 4 satisfy condition (3.19). Randomly generate
Y 1 n; 1; Y 1 n; 5 and Y 2 n; 1; Y 2 n; 5 satisfying constraints (3.20). Calculate the maximum speeds Y 1 n; 3 and Y 2 n; 3
according to Eqs. (3.21) and (3.22). Update Y 1 and Y 2 by using the Regulating Process. Then we select the best two chromosomes from the parents and children to replace the parents.
The general process for crossover operation is summarized as follows:
Step
Step
Step
Step
Step
Step
Step

1.
2.
3.
4.
5.
6.
7.

Initialize a crossover probability pc , and a pop-size-dimensional vector C.

Set i 1 and k 0.
Generate a random number r from 0; 1.
If r 6 pc , select chromosome X i as the parent, set Ck i; k k 1 and i i 1.
If i 6 pop-size, go to step 3.
Denote the selected parents by X C1 ; X C2 ; . . . ; X Ck . Set j 1.
Generate two children Y Cj and Y Cj1 from parents X Cj and X Cj1 .

84

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

Step 8. Replace the parents with the best two chromosomes from the parents and children. Set j j 2.
Step 9. If j < k, go to step 7.
4.6. Mutation operation
Mutation is another important operation for updating the chromosomes. We dene a parameter pm to denote the mutation probability, and randomly select some chromosomes as parents in a similar way to the process of selecting parents for
crossover operation.
For each selected parent X, we randomly generate an index n excluding n N and n N=2. If segment n is a station, we
 11n , and set
randomly generate a dwell time w from w11n ; w

Xk; 5 Xk; 5  Xk; 5  Xk; 1  w; if k P n;

Xk; 1 Xk; 1  Xk; 5  Xk; 1  w; if k > n:

4:4

If segment n is in a section with nonzero speed limit, we randomly generate the boundary speeds qn ; qn1 and the switching
time a11n ; e11n under the constraints (3.18)(3.20). Set Xn; 2 qn ; Xn; 4 qn1 , and

Xk; 5 Xk; 5  Xn; 5  Xn; 1  e11n  a11n ; if k P n;

Xk; 1 Xk; 1  Xn; 5  Xn; 1  e11n  a11n ; if k > n:

4:5

Calculate the maximum speed Xn; 3 satisfying Eqs. (3.21) and (3.22). Update X by using the Regulating Process.
The general process for mutation operation is summarized as follows:
Step
Step
Step
Step
Step

1.
2.
3.
4.
5.

Initialize a mutation probability pm , and set i 1.

Generate a random number r from 0; 1.
If r 6 pm , randomly generate a segment index n with n N=2 and n N. Update chromosome X i .
Set i i 1.
If i 6 pop-size, go to step 2.

4.7. General procedure

Following evaluation, selection, crossover and mutation operations, a new population of chromosomes is generated. Genetic algorithm will terminate after a given number of iterations of the above steps. We now summarize the general procedure for genetic algorithm as follows.
Step
Step
Step
Step
Step
Step
Step

1.
2.
3.
4.
5.
6.
7.

Randomly initialize pop-size chromosomes.

Calculate the net energy consumptions for all chromosomes.
Calculate the evaluation values for all chromosomes.
Select the chromosomes by spinning the roulette wheel.
Update the chromosomes by using crossover and mutation operations.
Repeat the second to fth steps max-generation times.
Return the best found chromosome and decode it to a solution.

5. Experimental studies
In order to illustrate the efciency of the integrated energy-efcient operation model on energy saving, we conduct
extensive numerical experiments with different headway. The genetic algorithm is coded in Matlab 7.1 under the running
environment: a Windows 7 platform of personal computer with processor speed 2.4 GHz and memory size 2 GB. When genetic algorithm is performed, we set pop-size 40; pc 0:8; pm 0:5, and max-generation 70.
The Yizhuang line, opened on December 30, 2010, is one of the 17 lines of Beijing rail transit network, with daily ridership
approximately 180,000. This line is 23.23 km long with 14 stations, including 6 underground and 8 on the surface. Trains
running between Songjiazhuang station and Yizhuang station are powered by 6 substations (see Fig. 5). The current timetable is shown by Table 2, and the data on speed limit is described by Table 3 (Su et al., 2013). The whole line is divided into
108 segments, including 78 segments in sections, 28 stations and 2 turnaround stations. Parameters are listed as follows:
conversion factor from electricity to kinetic energy h1 0:9, conversion factor from kinetic energy to regenerative electricity
h2 0:6, transmission loss factor of regenerative electricity h3 0:1, acceleration rate at accelerating phase a 0:8 m=s2 ,
deceleration rate at coasting phase b 0:02 m=s2 , deceleration rate at braking phase c 0:4 m=s2 , train mass
M = 287,080 kg, turnaround time r 180 s, and operation time T = 64,800 s.
The following examples are simulation studies, which compare the current timetable of Yizhuang line with the best found
timetable with respect to the theoretical energy calculations using the formulae in this paper.

85

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

Fig. 5. Substation distribution for Beijing Metro Yizhuang Line.

Table 2
The practical timetable of Beijing Metro Yizhuang Line.
Station
Arrival time
Dwell time

Songjiazhuang
0
30
Wanyuan

Arrival time
Dwell time

Arrival time
Dwell time

Arrival time
Dwell time

Arrival time
Dwell time

Xiaocun
135
30

Xiaohongmen
264
30

Jiugong
431
30

Yizhuangqiao
607
35

Wenhuayuan
800
30

Rongjing

Rongchang

Tongjinan

Jinghai

Ciqunan

1061
30

1203
30

1320
30

1482
30

1665
35

Ciqu

Yizhuang

Turnaround

Yizhuang

Ciqu

Ciqunan

1802
45

2034
40

2074
180

2254
40

2491
45

2628
35

932
30

Jinghai

Tongjinan

Rongchang

Rongjing

Wanyuan

Wenhuayuan

2816
30

2978
30

3095
30

3237
30

3366
30

3498
30

Yizhuangqiao

Jiugong

Xiaohongmen

Xiaocun

Songjiazhuang

Turnaround

3686
35

3867
30

4034
30

4163
30

4290
30

4320
180

Table 3
Speed limit of Beijing Metro Yizhuang Line.
Section

Speed limit/segment

Speed limit/segment

Speed limit/segment

SongjiazhuangXiaocun
XiaocunXiaohongmen
XiaohongmenJiugong
JiugongYizhuangqiao
YizhuangqiaoWenhuayuan
WenhuayuanWanyuan
WanyuanRongjing
RongjingRongchang
RongchangTongjinan
TongjinanJinghai
JinghaiCiqunan
CiqunanCiqu
CiquYizhuang
YizhuangCiqu
CiquCiqunan
CiqunanJinghai
JinghaiTongjinan
TongjinanRongchang
RongchangRongjing
RongjingWanyuan
WanyuanWenhuayuan
WenhuayuanYizhuangqiao
YizhuangqiaoJiugong
JiugongXiaohongmen
XiaohongmenXiaocun
XiaocunSongjiazhuang

10/0128
13/0130
13/0129
13/0128
13/0120
14/0130
14/0130
14/0130
13/0130
14/0123
14/0128
14/0130
13/0133
13/0133
14/0130
14/0128
14/0123
13/0130
14/0130
14/0130
14/0130
13/0120
13/0128
13/0129
13/0130
10/0128

20/1281147
20/1301085
20/1291897
20/1282073
20/1202143
20/1301168
20/1301087
20/1301350
20/130796
20/1231781
20/1282165
20/1301151
20/1332300
20/1332300
20/1301151
20/1282165
20/1231781
20/130796
20/1301350
20/1301087
20/1301168
20/1202143
20/1282073
20/1291897
20/1301085
20/1281147

14/11471332
13/10851286
11/18972086
13/20732265
12/21432331
13/11681354
13/10871280
13/13501544
13/796992
13/17811975
13/21652369
13/11511349
10/23002610
10/23002610
13/11511349
13/21652369
13/17811975
13/796992
13/13501544
13/10871280
13/11681354
12/21432331
13/20732265
11/18972086
13/10851286
14/11471332

Example 5.1. This example illustrates the efciency of the integrated energy-efcient operation model on energy
conservation. Take h 90 s, which is used by most of the Beijing metro rail lines. According to the practical timetable, 50
trains are put into operation, and each train runs 13 cycles with 4500 s for each cycle. The net energy consumption for all
trains at all cycles is 203,850 kW h.

86

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

Table 4
The optimal timetable of Beijing Metro Yizhuang Line.
Station
Arrival time
Dwell time

Arrival time
Dwell time

Arrival time
Dwell time

Arrival time
Dwell time

Arrival time
Dwell time

Songjiazhuang
0
27

Xiaocun
135
30

Xiaohongmen
278
26

Jiugong
492
27

Yizhuangqiao
696
32

Wenhuayuan
918
28

Wanyuan

Rongjing

Rongchang

Tongjinan

Jinghai

Ciqunan

1069
27

1210
33

1378
26

1500
27

1688
27

1904
34

Ciqu

Yizhuang

Turnaround

Yizhuang

Ciqu

Ciqunan

2051
49

2320
37

2357
180

2537
37

2770
44

2929
31

Jinghai

Tongjinan

Rongchang

Rongjing

Wanyuan

Wenhuayuan

3129
33

3300
27

3423
28

3579
28

3715
33

3863
29

Yizhuangqiao

Jiugong

Xiaohongmen

Xiaocun

Songjiazhuang

Turnaround

4073
32

4273
32

4453
27

4596
29

4742
28

4770
180

Table 5
Comparisons on energy performance among IEEO model, CS model, and EEO model.
Headway
90

120

150

180

240

300

Model

Tractive energy

Regenerative energy

Net energy

C/K/I

IEEO
CS
EEO
IEEO
CS
EEO
IEEO
CS
EEO
IEEO
CS
EEO
IEEO
CS
EEO
IEEO
CS
EEO

195090
257,830
191,020
146,670
177,900
142,980
115,180
144,320
113,510
101,120
116,700
99,996
83,264
95,721
82,886
68,109
70,838
66,536

45,520
68,100
31,310
23,180
32,250
13,000
14,200
16,090
12,180
3630
10,280
2044
2320
3314
1504
5294
5573
3690

149,570
189,730
159,720
123,490
145,650
129,980
100,980
128,230
101,330
97,485
106,420
97,950
80,943
92,407
81,383
62,815
65,265
62,846

4950/55/12
4770/53/13
4950/55/12
4920/41/12
4800/40/13
4920/41/12
4950/33/12
4800/32/13
4950/33/12
4860/27/12
4860/27/12
4860/27/12
4800/20/13
4800/20/13
4800/20/13
4800/16/13
4800/16/13
4800/16/13

Now we solve the integrated energy-efcient operation model by performing the genetic algorithm. After 70 iterations,
the best found solution (see Table 4) shows that 55 trains should be put into operation, and each train needs to run 12 cycles
with 4950 s for each cycle. Compared with the practical timetable, ve more trains are used which reduces the number of
cycles from 13 to 12 for each train. Most importantly, it lengthens the cycle time from 4500 s to 4950 s, which can
signicantly reduce the tractive energy consumption, although a minor inuence on the regenerative energy utilization may
happen arising from the reduction on synchronization of accelerating and braking actions. The net energy consumption of
the best found solution is 149,570 kW h. It is easy to calculate that the integrated energy-efcient operation model achieves
an energy conservation by (203,850149,570)/203,850 = 26.63%.
Since the total travel time increases by 450 s for each cycle, compared with the practical timetable, about 17 s is added for
each travel between two successive stations. However, it is generally acceptable since most of passengers take short travels.
Example 5.2. This example makes comparisons among energy-efcient operation model (EEO, Su et al., 2013), cooperative
scheduling model (CS, Yang et al., 2013) and integrated energy-efcient operation model (IEEO) on tractive energy consumption, regenerative energy utilization and net energy consumption. For the EEO model, we rst solve the speed prole with
minimum tractive energy consumption for a single train along the whole line. Then we extend the speed prole to all trains
at all cycles according to the periodicity of metro rail operation, and calculate the regenerative energy utilization and the net
energy consumption. For the CS model, we rst solve the timetable with maximum regenerative energy utilization. Then we
calculate the energy-efcient speed prole for trains at sections, and nally calculate the net energy consumption. In this
case, this extended CS model essentially returns the net energy consumption value of the traditional two-step optimization
approach.

87

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

Taking the headway from 90 s to 300 s, the computation results are recorded by Table 5, where the third column denotes
the tractive energy consumption, the fourth column denotes the regenerative energy utilization, the fth column denotes the
net energy consumption, and the last column records the cycle time C, the number of service trains K and the number of
cycles I. It is concluded that the EEO model can minimize the tractive energy consumption at the cost of a negative utilization
of regenerative energy, while the CS model can maximize the regenerative energy utilization at the cost of a larger tractive
energy consumption. As a trade-off between the tractive energy and the regenerative energy, the IEEO model is able to
signicantly cut down the net energy consumption. For example, when h 90 s, the IEEO model reduces the net energy
consumption by 21.17% compared with the CS model, and 6.35% compared with the EEO model.
The headway has a signicant inuence on energy consumption. Generally speaking, a larger headway leads to smaller
energy consumption. Taking the IEEO model for example, the net energy consumption decreases from 149,570 kW h to
62,185 kW h as the headway increases from 90 s to 300 s. The reason is that headway is equal to the quotient of the cycle
time and the number of service trains, i.e., h C=K, which implies that a larger headway means either a larger cycle time or a
smaller number of service trains, both of them lead to a smaller energy consumption. According to the last column of Table 5,
it is easy to see that the number of service trains decreases from 55 to 16 as the headway increases from 90 s to 300 s.
Compared with EEO model, the advantage of IEEO model on energy saving gradually disappears as the increasing of
headway. In Fig. 6, when headway is larger than 150 s, the IEEO model almost returns the same value of net energy
consumption as the EEO model. This is because that the regenerative energy utilization contributes little to the energy saving
in case of there is a large headway (see the fourth column of Table 5). Therefore, the energy-efcient efciency of IEEO model
is more signicant for busy metro rail lines with smaller headway.
x 105
2
IEEO Model
CS Model
EEO Model

Net Energy Consumption

1.8
1.6
1.4
1.2
1
0.8
0.6

100 120

140 160 180

200 220 240 260

280

300

Headway
Fig. 6. Relation between the net energy consumption and headway.

1.66

x 105

Net Energy Consumption

1.64
1.62
1.6
1.58
1.56
1.54
1.52
1.5
1.48

10

20

30

40

50

60

70

80

Iteration
Fig. 7. Convergence of the genetic algorithm.

90

100

88

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

Example 5.3. This example illustrates the convergence of the genetic algorithm. Taking h 90 s, we implement the genetic
algorithm with max-generation 100 and record the best found objective value at each generation by Fig. 7. Note that the
objective value hardly changes after the 70th iteration and converges to 149,570 kW h. Therefore, it is reasonable to take
max-generation 70 in above numerical experiments.
6. Conclusions and future research
The main contribution of this paper is to develop a mathematical optimization approach to improve the operations management of metro system on energy efciency. Quantitative analysis of tractive energy consumption, regenerative energy
utilization, and net energy consumption was given, and an energy-efcient scheduling and speed control model was formulated to minimize the net energy consumption, which may be used to help the trafc management to develop a much improved energy-efcient operation strategy. The method presented here is designed for but not limited to metro rail system. It
can be extended and applied to the scheduling and speed control for high-speed trains and other types of passenger trains.
A Beijing Metro Yizhuang Line simulation study with actual operation data was given, which shows that a larger headway
leads to smaller energy saving rate, and the maximum energy saving rate achieved is around 25% when we use the minimum
allowable headway of 90 s. Compared with the traditional two-step optimization approach, the proposed approach can reduce the net energy consumption around 20%.
One of our future research directions is to extend this integrated energy-efcient operation approach to consider the variable acceleration rate, deceleration rate, running resistance, distance-based energy transmission loss factor, and stochastic
departure delay for trains at busy stations for making more realistic calculations of energy consumption.
In addition, this paper assumes that the headway takes a constant value. However, the actual Beijing Metro Yizhuang Line
adopts a CBTC (communication based train control) system for train control, which is equivalent to a moving block train control method such that the exible safety headway is available. In future, we will extend this work to consider the exible
headway.
The generated train speed prole in the timetabling stage may be unavailable in real-time operations, since the planned
segment running time may be different from the real-time running time due to uctuations in the passenger demand operations conditions. In this case, we need to regenerate the train speed prole according to the real-time segment running
time. In future, from a practice point of view, we will study the integrated train speed control and real-time railway dispatching problem.
Finally, although the proposed approach has been proved to be efcient on numerical experiments, we need to further
conduct empirical studies to test its efciency in practical operations management of metro trains.
Acknowledgments
This work was supported by the National Natural Science Foundation of China (Nos. 71101007, 71371027), Program for
New Century Excellent Talents in University under Grant No. NCET-13-0649, the Fundamental Research Funds for the Central Universities (No. ZZ1316), the State Key Laboratory of Rail Trafc Control and Safety (No. RCS2012ZT002) and General
Research Fund #616113 of the Hong Kong Research Grants Council. The authors thank the three anonymous referees for constructive comments which greatly improve this paper.
References
Albrecht, A., Howlett, P., Pudney, P., Vu, X., 2013. Energy-efcient train control: from local convexity to global optimization and uniqueness. Automatica 49
(10), 30723078.
Asnis, I.A., Dmitruk, A.V., Osmolovskii, N.P., 1985. Solution of the problem of the energetically optimal control of the motion of a train by the maximum
principle. USSR Computational Mathematics and Mathematical Physics 25 (6), 3744.
Bagloee, S.A., Ceder, A., 2011. Transit-network design methodology for actual-size road networks. Transportation Research Part B 45 (10), 17871804.
Benjamin, B., Milroy, I., Pudney, P., 1989. Energy-efcient operation of long-haul trains. In: Proceedings of the 4th International Heavy Haul Railway
Conference, National conference publication (Institution of Engineers, Australia), Barton, ACT, Australia, pp. 369372.
Bocharnikov, Y.V., Tobias, A.M., Robe, C., 2010. Reduction of train and net energy consumption using genetic algorithms for trajectory optimisation. In:
Proceedings of IET Conference on Railway Traction Systems, IEEE Publisher, Birmingham, UK, pp. 3236.
Chang, C., Sim, S., 1997. Optimising train movements through coast control using genetic algorithms. IEE Proceedings Electric Power Applications 144 (1),
6573.
Cordonea, R., Redaellib, F., 2011. Optimizing the demand captured by a railway system with a regular timetable. Transportation Research Part B 45 (2), 430
446.
David, F., Denis, M., Francois, F., 2012. Energy optimization of metro timetables: a hybrid approach. In: Proceedings of the 18th International Conference on
Principles and Practice of Constraint Programming, Lecture Notes in Computer Science, vol. 7514, Springer Publisher, Quebec City, Canada, pp. 712.
Ding, Y., Liu, H.D., Bai, Y., Zhou, F.M., 2011. A two-level optimization model and algorithm for energy-efcient urban train operation. Journal of
Transportation Systems Engineering and Information Technology 11 (1), 96101.
Dominguez, M., Fernandez-Cardador, A., Cucala, A.P., Pecharroman, R.R., 2012. Energy savings in metropolitan railway substations through regenerative
energy recovery and optimal design of ATO speed proles. IEEE Transactions on Automation Science and Engineering 9 (3), 496504.
Howlett, P., 1988. Existence of an Optimal Strategy for the Control of a Train. School of Mathematics Report 3, University of South Australia.
Howlett, P., 1990. An optimal strategy for the control of a train. Journal of the Australian Mathematical Society Series B 31, 454471.
Howlett, P., 2000. The optimal control of a train. Annals of Operations Research 98 (14), 6587.
Howlett, P., Pudney, P., 1995. Energy-Efcient Train Control. Springer.
Howlett, P., Milroy, I., Pudney, P., 1994. Energy-efcient train control. Control Engineering Practice 2 (2), 193200.

X. Li, H.K. Lo / Transportation Research Part B 64 (2014) 7389

89

Howlett, P., Pudney, P., Vu, X., 2009. Local energy minimization in optimal train control. Automatica 45 (11), 26922698.
Ishikawa, K., 1968. Application of optimization theory for bounded state variable problems to the operation of trains. Bulletino of JSME 11 (47), 857865.
Ke, B., Chen, M., Lin, C., 2009. Block-layout design using maxmin ant system for saving energy on mass rapid transit systems. IEEE Transactions on
Intelligent Transportation Systems 10 (2), 226235.
Khmelnitsky, E., 2000. On an optimal control problem of train operation. IEEE Transactions on Automatic Control 45 (7), 12571266.
Kim, K., Chien, S., 2011. Optimal train operation for minimum energy consumption considering track alignment, speed limit, and schedule adherence.
Journal of Transportation Engineering 137 (9), 665674.
Kroon, L., Maroti, G., Helmrich, M., Vromans, M., Dekker, R., 2008. Stochastic improvement of cyclic railway timetables. Transportation Research Part B 42
(6), 553570.
Li, X., Yang, X., 2013. A stochastic timetable optimization model in subway systems. International Journal of Uncertainty, Fuzziness & Knowledge-Based
Systems 21 (Supp. 1), 115.
Liu, R., Golovitcher, I., 2003. Energy-efcient operation of rail vehicles. Transportation Research Part A 37 (10), 917932.
Lu, S.F., Hillmansen, S., Ho, T.K., Roberts, C., 2013. Single-train trajectory optimization. IEEE Transactions on Intelligent Transportation Systems 14 (2), 743
750.
Mazloumi, E., Mesbah, M., Ceder, A., Moridpour, S., Currie, G., 2012. Efcient transit schedule design of timing points: a comparison of ant colony and genetic
algorithms. Transportation Research Part B 46 (1), 217234.
Nasri, A., Fekri Moghadam, M., Mokhtari, H., 2010. Timetable optimization for maximum usage of regenerative energy of braking in electrical railway
systems. In: Proceedings of the International Symposium on Power Electronics, Electrical Drives, Automation and Motion, IEEE Publisher, Pisa, Italy, pp.
12181221.
Pena-Alcaraz, M., Fernandez, A., Cucala, A., Ramos, A., Pecharroman, R., 2012. Optimal underground timetable design based on power ow for maximizing
the use of regenerative-braking energy. Journal of Rail and Rapid Transit 226 (4), 397408.
Ramos, A., Pena, M., Fernndez-Cardador, A., Cucala, A.P., 2007. Mathematical programming approach to underground timetabling problem for maximizing
time synchronization. In: Proceedings of the International Conference on Industrial Engineering and Industrial Management, IEEE Publisher, Madrid,
Spain, pp. 8895.
Rodrigo, E., Tapia, S., Mera, J.M., Soler, M., 2013. Optimizing electric rail energy consumption using the lagrange multiplier technique. Journal of
Transportation Engineering 139 (3), 321329.
Su, S., Li, X., Tang, T., Gao, Z., 2013. A subway train timetable optimization approach based on energy-efcient operation strategy. IEEE Transactions on
Intelligent Transportation Systems 14 (2), 883893.
Tuyttens, D., Fei, H.Y., Mezmaz, M., Jalwan, J., 2013. Simulation-based genetic algorithm towards an energy-efcient railway trafc control. Mathematical
Problems in Engineering, Article ID 805410, 112.
Voorhoeve, M., 1993. Rail Scheduling with Discrete Sets. Technical Report, Eind-hoven University of Technology, Eindhoven, The Netherlands.
Yang, X., Li, X., Gao, Z., Wang, H., Tang, T., 2013. A cooperative scheduling model for timetable optimization in subway systems. IEEE Transactions on
Intelligent Transportation Systems 14 (1), 438447.