This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 1

Event-Triggered Predictive Control for Automatic

Train Regulation and Passenger Flow in
Metro Rail Systems
Xi Wang, Shukai Li , Tao Tang , Senior Member, IEEE, and Lixing Yang

Abstract— Focusing on improving the operation efficiency and way. A well-designed train schedule directly contributes to
riding comfort of metro rail lines in the peak hours, this article elevating the operation efficiency and enhancing the level of
investigates the real-time train regulation and passenger load passenger service quality [3], [4].
control problem with respect to frequent disturbances. To better
illustrate the relationship between the train timetable and the In general, a nominal timetable is designed based on a
on-board passengers, the variations of the departure time and series of commercial operation factors such as the line layout,
the passenger load are elaborated in the form of a state-space the signalling system, the fleet size, the passenger demand
model. Based on the Lyapunov stability theory, the problem model and so on. However, the reference timetable may
of minimizing an upper bound on the quadratic performance become infeasible in practice for the following reasons. On one
function is transformed to a dynamic optimization problem with
a set of linear matrix inequalities (LMIs), and a predictive control hand, a variety of disturbances may disturb the normal train
strategy is designed to guarantee the actual train schedule and travelling, and thereby cause significant deviations from the
number of in-vehicle passengers track the nominal timetable and desirable train timetable. Due to the high-frequency feature of
the expected passenger load with a given disturbance attenuation metro networks, the train delay is easily propagated from one
level. With the objective to reduce the computational workloads station to the successive ones, which leads to the instability of
and cut down the utilization of wireless transmitting resources, an
event-triggered strategy is developed to implement the proposed the metro transit systems and the accumulation of passengers
stabilizing feedback controller only when the measurement error especially in the rush hours [5], [6]. On the other hand, the
exceeds certain threshold, which has better adaptability to the accelerating urbanization and centralized economic activities
application in large-scale metro networks. Some numerical exam- bring a sharp rising of demand to the urban transportation
ples based on the Beijing Yizhuang Metrol Line are provided for system, and many metro transit stations are suffering from
illustration of the effectiveness of the proposed scheme.
the over-saturated passenger flow in peak hours. For example,
Index Terms— Automatic train regulation, passenger flow con- it is estimated that the in-vehicle passengers reach as many
trol, event-triggered condition, model predictive control (MPC). as 120% of the expected volume in some Beijing transit lines
during peak hours. The crowded passengers not only prolong
I. I NTRODUCTION the delay time, but also incur concerns to the operational
safety [7], [8]. Considering the complexity of the metro
I N RECENT years, the metro rail transit lines are experi-
encing a rapid development in satisfying the ever-growing
demand for passenger transportation. As a key issue of public
systems and the real-time nature of the problem, it makes
sense to design computational efficient train regulation and
transit operations and management, providing a satisfactory passenger flow control methodology to compensate the delays
passenger service becomes the main concern for railway and adjust the on-board passenger load such that the required
enterprises [1], [2]. To achieve the operating objective, one service efficiency and quality level are enhanced.
of the important tasks is to bring all of the passengers from
their origin to their destination in a punctual and comfortable A. Literature Review
Manuscript received December 13, 2018; revised March 2, 2020 and The passenger train timetable design problem on an urban
June 20, 2020; accepted September 22, 2020. This work was supported in transit line aims to determine the arrival and departure time
part by the National Key Research and Development Program of China under
Grant 2018YFB1600900, in part by the National Natural Science Foundation for every train at each station. It can be further divided into
of China under Grant 62073024, in part by the Research Foundation of State two subproblems: the timetable planning problem and the train
Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University rescheduling problem [9]. The timetable planning problem
under Grant RCS2020ZY001, and in part by the Fundamental Research Funds
for the Central Universities of China under Grant 2019JBM076. The Associate refers to create a static optimal train schedule based on a
Editor for this article was R. Goverde. (Corresponding author: Shukai Li.) macroscopic consideration of system resources. The schedul-
Xi Wang is with the State Key Laboratory of Rail Traffic Control and ing problem focuses on re-arranging each train’s travelling and
Safety, Beijing Jiaotong University, Beijing 100044, China, and also with
the School of Electronics and Information Engineering, Beijing Jiaotong dwelling time to reduce the delays especially when unexpected
University, Beijing 100044, China. disturbances happen, which is a typical dynamic process.
Shukai Li, Tao Tang, and Lixing Yang are with the State Key Laboratory of Different from other types of railway transportation such as
Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing 100044,
China (e-mail: [email protected]). inter-city and high-speed railways, the metro transit is charac-
Digital Object Identifier 10.1109/TITS.2020.3026755 terized by its short inter-station distance and high frequency.
1524-9050 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

These features make it is more vulnerable to external distur- its distinguished advantages in relaxing the conventional peri-
bances [10]. When delay occurs in the metro transit line, there odic implementation requirements and deceasing the unneces-
are two possible ways to achieve the rescheduling of on-line sary waste of network resources, while retaining satisfactory
trains. First, the automatic train regulation (ATR) technology closed-loop properties [25], [26]. Under the framework of
can be adopted to guarantee the practical timetable tracks the event-triggered control, the control forces remain constant
the reference train schedule [11], [12]. Second, in presence until the violation of a well-designed event-triggering condi-
of large delays, where the deviations may go beyond the tion, which is a significant novelty compared with the ordinary
limit of adjusting forces, a new nominal timetable should periodical control. As far as we know, few studies take advan-
be worked out to re-arrange the departure and arriving time tages of the event-triggered control methodology to develop
for all trains [13], [14]. From the perspective of maintaining a joint train regulation and passenger load control strategy
the service quality, it is desirable that the actual operation of under a disrupted metro network. These facts make it essential
trains is stick to the optimized schedule. Hence, this article to achieve the event-triggered scheduling of train traffic and
mainly considers the first situation and pays the attention passenger load so as to provide the system with flexibility for
on recovering the reference timetable from disturbances by recovery from disturbed situations, and meanwhile reduce the
utilizing the ATR technology. number of re-computations of the control efforts and the usage
There is a wealth of literatures on the dynamic train of communication resources.
regulation approaches to recover train delays. Reference [15]
proposed a method for the allocation of buffer time based B. The Proposed Scheme and Contributions
on the train delay distributions and passenger waiting time to
adjust the railway timetable. To compensate the train delays Motivated by the above discussions, this article first designs
and improve energy-efficiency, [16] designed a numerical a predictive controller to guarantee the nominal train timetable
algorithm to realize the optimal distribution of the buffer and desirable in-vehicle passenger load are recovered from
time among different sections. Although the buffer time can uncertain disturbances. Then, an event-triggered mechanism
be utilized to absorb the delays, it reduces the flexibility of is proposed to determine whether the sampled data should be
the train timetable, and this method becomes infeasible when transmitted and whether the control efforts should be updated.
the delay time exceeds the threshold of the reserved time. Specifically, the main contributions of this study are presented
By exploiting the global information of the train traffic, [17] as follows.
developed an integer programming model to eliminate the (1) Compared with the existing literatures on the train
train delays, and minimize the energy consumption during scheduling problem [9], [12], the constructed state-space
the rescheduling process. Above methods mainly employ the model integrates the dynamic states of the train traffic with the
dynamic programming or nonlinear programming algorithm to variations of on-board passengers, which can demonstrate the
solve the formulated optimization problem, which are usually coupling relationship between the train traffic and passenger
time-consuming. To improve the computational efficiency, the flow in the disturbed scenario.
model predictive control (MPC) approach was adopted to (2) A state feedback control law is designed to guarantee the
transform the original schedule optimization problem into a actual train schedule and the number of in-vehicle passengers
series of quadratic programming problems [18], [19], which track the nominal timetable and the expected passenger load
could be computed in polynomial time. Based on the MPC with a given disturbance attenuation γ , and meanwhile the
framework, [20] and [21] designed efficient real-time control minimization of an upper bound on the cost function repre-
algorithms to deal with the regulation problem for metro senting both service quality and control cost is ensured.
lines. Reference [22] used the fuzzy passenger arrival rate to (3) Different from the traditional MPC method, where
address the uncertainty of passenger flow, and developed a the control efforts are calculated at each decision stage, a
fuzzy predictive controller to recover the nominal timetable event-triggered condition is proposed to execute the control
from delays. For rush hours with a large amount of arriving task whenever the measurement error exceeds certain threshold
passengers, it may lead to the over-congestion of in-vehicle compared with the state norm, which reduce the amount of
passengers and thereby degrade the riding comfort. To address controller updates and avoid the unnecessary waste of network
this issue, [23] studied the train scheduling problem consid- resources.
ering the origin-destination (OD) characteristics of passenger The remainder of this article is organized as follows. In
demands. [24] designed a state feedback controller to adjust Section 2, the dynamic model of the train traffic and on-board
the running and dwell time as well as the passenger load. passengers is described. In Section 3, the joint control strategy
It is worthy noting that the majority of existing train for train regulation and passenger load adjustment is designed,
regulation schemes need to update the control gain at each and the event-triggering condition is detailed. In Section 4,
stage in the real-time application, which is a computationally numerical examples are presented to illustrate the effectiveness
demanding procedure for a large-scale metro transit network. of the proposed strategies. Finally, this article is summarized
In addition, the states of all running trains and passenger in Section 5.
flows should be collected and transmitted through the wireless
network each period, which results in the over-utilization II. P ROBLEM F ORMULATION
of the available network resources. In recent years, the In this study, we consider a metro transit line with I trains.
event-triggered control method has received much attention for According to the nominal timetable, each train is dispatched

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG et al.: EVENT-TRIGGERED PREDICTIVE CONTROL FOR AUTOMATIC TRAIN REGULATION AND PASSENGER FLOW 3

TABLE I
PARAMETERS AND VARIABLES

Fig. 1. The illustration of the metro rail transit line.

from station 1, and leaves the line at the terminal station N +1.
At each station, the train dwells certain predetermined time for
passengers boarding and alighting. The operation of a metro
rail line is illustrated in Fig. 1.
According to the operational requirements of the metro
transit and characteristics of the passenger flow, we make
the following assumptions for the operation of trains and the
variation of passengers.
(A1). This study considers a one-direction and single track
metro transit line. All of the trains are assumed to begin service
from the starting station, and the overtaking and crossing are
unallowable at any stations and sections.
(A2). Each train stops at all wayside stations to allow
passengers to embark and disembark, i.e., the stop-skipping
is not considered in this article.
(A3). The dwell time of a train is mainly affected by the
number of boarding and alighting passengers.
(A4). The total amount of boarding passengers is closely
related to the waiting time between successive trains, and the
number of alighting passengers is supposed to be proportional
to the amount of on-board passengers.
(A5). The delay disturbance is in a certain range and the
initial passenger load for trains does not exceed the train
capacity.
In order to simplify the operation model for the trains,
departure time is described by the following dynamic equation:
assumption A1 is adopted in this article, which is reasonable
and adopted in existing literatures such as [10] and [27]. i
dn+1 = dni + tni + τni , (1)
Assumption A2 and A3 generally hold for most metro transit
lines, which are normally operated in this manner [28], [29]. where dni denotes the practical departure time of train i from
Assumption A4 enables us to calculate the number of boarding station n, tni represents the practical running time of train i
and alighting passengers at each station, and thereby deter- from station n to n + 1, and τni is the actual dwell time of
mine the dwell time. Assumption A5 makes it is possible to train i at station n.
study the nominal train schedule and passenger load tracking According to assumption A3, the dwell time of a train is
problem. One can find similar assumptions in many existing mainly affected by the arrival and alighting passenger flow.
researches about train regulation and passenger flow control Here, we introduce a delay coefficient δ to express the time
problems [21], [24]. consumption of one passenger getting on or getting off the
Based on the above assumptions, the joint train operation train, and then the dwell time τn+1
i
is described as
and on-board passenger dynamic model is proposed by link- i
τn+1 i
= δ(bn+1 i
+ ln+1 ) + Mn+1 + u d in+1 + wd in+1 , (2)
ing the connections between the traffic dynamics and the
i
variation of on-board passengers. Then, the event-triggered where bn+1is the number of passengers who want to board
predictive control problem for the automatic train regulation train i at station n + 1, ln+1
i denotes the amount of passengers
and passenger flow control is formulated. The parameters and leaving train i at station n + 1, Mn+1 indicates the minimal
variables used in describing the train operation process are dwell time at station n+1, u d in+1 represents the applied control
listed in Table I. forces to tune the dwell time for train i at station n + 1, and
wd in+1 means the suffered uncertain disturbance when train i
A. The Dynamic Model for Automatic Train Regulation stops at station n + 1.
i
For each train, the real traveling time tn+1 is modeled as
In this study, we use the discrete event model to formulate
the variation of train traffic [5], and the evolution of the train i
tn+1 = Tni + u t in + wt in , (3)

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

where Tni refers to the expected travelling time of train i from

station n to station n + 1, u t in is the control input to adjust
the traveling time of train i from station n to n + 1, and wt in
represents the uncertain disturbance to the running time when
train i departs for station n+1 from station n. As the delay time
is limited in urban transit systems, it is reasonable  to assume
that bothwd in and wt in are of finite energy, i.e., ∞ n=1 (wd n ) <
i 2
∞
∞ and n=1 (wt n ) < ∞.
i 2

In actual engineering applications, one of the most conve-

nient ways to achieve the train traffic regulation is to adjust
the departure time. It can be accomplished by deploying a
train departure indicator at the platform or giving a sign on the
driver-machine interface. By substituting equations (2) and (3)
into (1), and letting ū in = u d in +u t in+1 and wni = wd in +wt in+1 , Fig. 2. The illustration of on-board passengers.
the dynamic equation associated with the departure time is
constructed as
i flow control process is repeated until the final decision step.
dn+1 = dni + δ(bn+1
i i
+ ln+1 ) + Tni + Mn+1 + ū in + wni . (4) To better understand the dynamics of on-board passengers,
In equation (4), ū in is the control strategy to magnify the the evolution of in-vehicle passengers is demonstrated in Fig.2.
running and dwell time of train i between stations n and n +1. When a train stops at the station, there are passengers
i
It is apparent that dn+1 is not only determined by the departure getting on the train and passengers getting off the train. The
time from the preceding station, but also closely related to the boarding passengers is closely related to the waiting time
passenger flow in station n + 1. When the disturbance occurs, between successive trains, and can be represented as a linear
the departure time will be delay from the nominal schedule. function of train departure time. According to assumption A4,
the actual numbers of entering passengers bn+1i and exiting
Under this case, the negative value of ū in will compensate the
i
delay time, and make the movement of a train stick to the passengers ln+1 are computed as
nominal timetable. i−1
i
bn+1 = ηn+1
i
(dn+1
i
− dn+1 ), (6)
B. The Dynamics of on-Board Passengers
i
ln+1 = μin+1 pni , (7)
In the rush hours, the train may become over-congested where ηn+1i denotes the passengers arrival rate of train i at
if the number of boarding passengers is not effectively con- station n + 1, and μin+1 means the alighting coefficient of
i
trolled. To avoid this situation, we let fn+1 denote the control on-board passengers for train i at station n + 1. In equa-
force to achieve the adjustment of the in-vehicle passenger tion (7), the number of alighting passengers is assumed to
load, and the variation of the passenger load of the train i at be proportional to the number of passengers in the train. With
the station n + 1 is derived as the wide application of real-time monitoring and smart card
i
pn+1 = pni + bn+1
i
− ln+1
i
+ f n+1
i
, (5) technologies, we can collect the passenger flow information
including the alighting number, on-board number and OD table
where pni is the actual load of passengers when train i is for a period of time. Then, based on the obtained historical
i
dispatched from station n. In equation (5), fn+1 is the control data, we can further estimate the approximated proportion of
strategy to adjust the number of passengers entering train i at the alighting passengers to the whole on-board passenger at
station n +1. Under the control strategy for the passenger flow, different stations. In existing literature, there have been studies
the actual number of passengers boarding the train is changed that make similar assumption on the alighting patterns and
i
to bn+1 + f n+1
i . During the peak hours or special holidays,
assume a fixed proportion of passengers currently on-board
i
the negative value of f n+1 is acted to reduce the passenger the train to alight at each stops [21], [30]. Considering
load in dealing with the sudden gathering of passengers. In the time-dependent characteristics of the practical passenger
practice, the proposed passenger flow control strategy could flow, it requires a more comprehensive model to precisely
be adopted by the entrance facilities outside the stations predict passengers’ arriving and alighting at each station. By
to limit the entering passengers. By using the Automatic comparison, although the assumption is restrictive in practice,
Fare Collection (AFC) technology, the number of passengers it reflects the passenger load dynamic and the correlation
that successfully enter the subway station could be exactly of train traffic and passenger load in a straightforward way.
recorded. When the number of entering passengers reaches Moreover, this linear formulation facilitates the design of the
the calculated number of allowable boarding passengers during state feedback controller and the rigorous proof of the robust
the interval between the departure instants of two successive stability for the coupled system.
trains, the entrance facilities are closed temporarily such that Integrating equations (5)–(7), the state-space model for
all of the entering passengers could get on the coming train. At in-vehicle passengers is formulated as
the next decision step, entrance gates are re-opened, and the
i−1
queueing passengers could board the next train. The passenger i
pn+1 = pni + ηn+1
i
(dn+1
i
− dn+1 ) − μin+1 pni + f n+1
i
, (8)

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG et al.: EVENT-TRIGGERED PREDICTIVE CONTROL FOR AUTOMATIC TRAIN REGULATION AND PASSENGER FLOW 5

i
which indicates that the passenger load pn+1 is influenced by In the literature, there have been studies that address the
the train departure time and the number of existing on-board train regulation problem based on this assumption [24], [30].
passengers jointly. Based on assumption A5, the focus of this In this article, we also adopt this assumption to construct
study is to design train regulation and passenger flow control the state-space model of the joint train motion and passenger
strategy to track the given nominal train schedule and passen- load model, which allows the proposed formulation in a linear
ger load in presence of external disturbances, i.e., stabilizing discrete dynamic system framework and enables us to apply
the actual train schedule and passenger load at the nominal the real-time state feedback information to design the train
ones. In practice, the nominal train schedule is set up based dynamic regulation scheme from a control-oriented perspec-
on a macroscopic consideration of system resources and the tive. Recently, a novel train dynamic model has been proposed
characteristics of passenger flow, which not only ensures the to overcome this limitation [31], [32]. In [32], the authors used
safety distance between successive trains, but also guarantees a kj to represent the instant of the kth arrival time to station j ,
the maximum number of passengers does not exceed the limit and introduced a new boolean variable b j to indicate the
of train capacity in case of high passenger arrival rate. If the number of trains being on segment between station j −1 and j .
stability of the studied regulation strategy is guaranteed, the It worthy noticing that k does not index trains, but counts the
actual train schedule and passenger load decrease gradually number of arrivals. Thus, this model allows more than one
and converge to their nominal values, i.e., the actual number train traveling between two consecutive stations, and could
of on-board passengers is within the limit of train capacity, be adopted to improve the efficiency for metro rail systems,
and the stability of passengers’ dynamics (6)-(8) is ensured. which will be an interesting topic for future work.
Then, by substituting equation (11) into equation (8), equa-
C. Formulation of the Joint Model tion (8) is equivalently expressed as
i
First, noting that the actual departure time dn+1 is affected i ηn+1
i
δηn+1
i
i−1
i
by the passenger flow control strategy f n+1 , the equation (4) pn+1 = dni − dn+1 + (1 − μin+1
1 − δηn+1
i 1 − δηn+1
i
is rewritten as
δμin+1 ηn+1
i
1
i
dn+1 = dni i
+ δ(bn+1 i
+ ln+1 + i
f n+1 )+ Tni + ) pni + i
f n+1
1 − δηn+1
i
1 − δηn+1
i
+ Mn+1 + ū in + wni . (9)
ηn+1
i

Combine equations (6) and (7), it follows from (9) that + (Tni + Mn+1 + ū in + wni ). (12)
1 − δηn+1
i
 i i−1 
i
dn+1 = dni + δ ηn+1 (dn+1
i
− dn+1 ) + μin+1 pni + f n+1
i
Let z ni = [dni , pni ]T and u in = [ū in , fn+1
i ]T , it can be obtained

+ Tni + Mn+1 + ū in + wni , (10) from equation (11) and (12) that
i−1
which can be further rewritten as i
z n+1 = Ain z ni + Ain z n+1 + Bni u in + Cni (Tni + Mn+1 + wni ),
1 δηn+1
i
i−1 (13)
i
dn+1 = d i
n − dn+1
1 − δηn+1
i 1 − δηn+1
i
where
⎡ ⎤
δμin+1 i δ i 1 δμin+1
+ pn+1 + fn+1 ⎢ ⎥
1 − δηn+1
i 1 − δηn+1
i
⎢ 1 − δηn+1
i
1 − δηn+1
i
⎥
Ain = ⎢ ⎥,
1 ⎣ ηn+1i δμ i η i
n+1 n+1 ⎦
+ (Tni + Mn+1 + ū in + wni ). (11) 1 − μn+1 +
i
1 − δηn+1
i
1 − δηn+1
i 1 − δηn+1i
⎡ ⎤ ⎡ ⎤
The proposed train traffic model (11) involves both the i th −δηn+1i 1 δ
and the i − 1th trains, where the departure time of train i at ⎢ 0⎥ ⎢ 1 − δηi ⎥
n+1 1 − δηn+1 ⎥
i
⎢ 1 − δηn+1i
⎥ ⎢
station n + 1 is not only determined by its departure time at Ain = ⎢ ⎥ , Bni = ⎢ ηi ⎥,
⎣ −ηn+1i
⎦ ⎣ n+1 1 ⎦
station n, but also by the departure time of its preceding train 0
i − 1 at station n + 1. There exists an assumption behind 1 − δηn+1
i 1 − δηn+1
i
1 − δηn+1
i

this model that there is only one train traveling between 1 ηn+1
i

two neighboring stations. In metro rail systems, the practical Cni = [ , ]T .

1 − δηn+1
i 1 − δηn+1
i
headway between trains is restricted by the turnaround time,
and the nominal headway is normally more than 150s. As In an ideal environment without any delays, the headway
the distance between two consecutive stations is short, the between two adjacent trains is fixed, and the nominal train
actual travelling time on segments is usually shorter than schedule and passenger flow satisfies the following equations:
 i 
the nominal headway, which implies this assumption holds i
Dn+1 = Dni + Tni + δ ηn+1 H + μin+1 L in + Mn+1 , (14)
for most scenarios. Taking the Beijing Yizhuang metro line
L in+1 = (1 − μin+1 )L in + ηn+1
i
H, (15)
as an example, the nominal running time between stations
varies from 60s to 100s, and the nominal headway is around where Dni means the nominal departure time of train i from
180s. Under this assumption, the safety distance between two station n to n + 1, L in represents the expected number of
neighbouring trains is ensured, which guarantees the safety of passengers on train i between station n and n + 1, H =
i−1
train operation. i
Dn+1 − Dn+1 denotes the nominal headway.

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

To handle the undesirable train delay and enhance the

riding comfort, we introduce the error state vector x ni =
[dni − Dni , pni − L in ]T , and then the error state-space model
associated with x ni is constructed by subtracting equation (13)
from (14) and (15)
i i−1
x n+1 = Ain x ni + Ain x n+1 + Bni u in + Cni wni . (16)
By generalizing the dynamic equation (16) to the metro
network, the matrix form of train regulation and passenger
flow control model is derived as follow:
x(k + 1) = Ax(k) + Bu(k) + Cw(k), (17)
where
x(k) = [x 1k−1 , x 2k−2 , . . . , x Nk−N ]T ,
Fig. 3. The diagram of state transition.
u(k) = [u k0 , u k−1 k−N+1 T
1 , . . . , u N−1 ] ,
w(k) = [w0k , w1k−1 , . . . , wk−N+1
N−1 ] ,
T
⎡ ⎤
Ak0 0 0 ··· 0 forces, which is defined as an indicator of the control cost.
⎢ Ak−1 Ak−1 0 ··· 0 ⎥ The minimization of this term can avoid the utilization of
A=⎢ ⎣ 0
1 1 ⎥
⎦ ,
··· ··· ··· 0 large control actions, which may put a high requirement on
0 ··· k−N+1
0 A N−1 k−N+1
A N−1 the traveling speed. In practice, the weighted matrices Q in and
2N×2N
⎡ k ⎤ Rni can be tuned to achieve the trade-offs among different
B0 0 0 ··· 0 regulation purposes.
⎢0 B1k−1 0 ··· 0 ⎥
B =⎢ ⎣0
⎥ , In cost function (18), the state constraints are related to the
0 ··· ··· 0 ⎦ headway between two neighbouring trains and the limit on
k−N+1
0 0 ··· 0 B N−1 2N×2N passenger load. According to assumption A5, if the stability
⎡ k ⎤ of the studied regulation strategy is guaranteed, the practical
C0 0 0 ··· 0
⎢ 0 C k−1 0 ··· 0 ⎥ passenger load and train delay will decrease gradually and con-
C =⎢ ⎣0
1 ⎥ .
0 ··· ··· 0 ⎦ verge to their nominal values. During the regulation process,
k−N+1 the actual number of on-board passengers is within the limit
0 0 ··· 0 C N−1 2N×N of train capacity, and the safety headway between successive
To better illustrate the variation of the state variable in trains is guaranteed, i.e., the state constraints are ensured by
equation (17), the transition from x(k) to x(k + 1) is shown the design of the stabilizing state feedback controller. Based on
in Fig. 3. As the uncertain disturbances always lead to the above discussion, this article concentrates on solving the
deviation of the actual traffic and passenger load state from LQ problem in form of (18) subject to the following control
the nominal state, an appropriate schedule regulation and constraints:
passenger flow control is desirable to make the error state x(k)
converges to zero. Hence, the stability of the metro traffic is −U l ≤ u l (k + j |k) ≤ Ûl , (19)
equivalently converted to the stability problem of the dynamic
system (17), which allows us to utilize the system stability where u l (k), U l and Ûl denote the lth entry of u(k), U and
theory to develop the corresponding control strategy. Û , respectively, U is the lower bound on the control input,
and Û represents the upper limit of the control input.
To better deal with the future uncertainties and real-time
D. Formulation of the Event-Triggered Control Problem
updated system parameters, the MPC framework is adopted to
To balance the operational efficiency of recovering the train compute the optimization problem (18) circularly such that the
delays from unexpected disturbances and the control cost, optimal control effort is derived at each stage. Then, by consid-
a multi-objective cost function is considered for the joint ering the dynamic traffic model (17) and the cost function (18),
dynamic equation (17) as follows: the train regulation and passenger flow adjustment strategy can
T T be obtained by solving the following optimal control problem:
J (k) = x ni (k)Q in x ni (k) + u in (k)Rni u in (k) , (18)
i n k f −k k f −k
where Q in > 0 and Rni > 0 are given constant weights. min J (k) = x T (k + j |k)Qx(k + j |k)
u(k+ j |k)
Remark 1: In the optimization criterion (18), the first term j =0 j =0
represents the summation of instant errors of the practical + u T (k + j |k)Ru(k + j |k) , (20)
timetable from the nominal schedule and the real on-board
passenger from the desired passenger load. The minimization where x(k + j |k) denotes the predicted error state of train
of this part helps to enhance the punctuality and riding traffic and passenger load at decision step k + j , u(k + j |k) is
comfort. The second term denotes the magnitude of the control the control input at stage k + j , and k f is the terminal stage.

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG et al.: EVENT-TRIGGERED PREDICTIVE CONTROL FOR AUTOMATIC TRAIN REGULATION AND PASSENGER FLOW 7

To guarantee the stability of the train traffic system with III. T HE C ONTROLLER D ESIGN
respect to uncertain disturbance, we introduce the following
In this section, the sufficient conditions are first given to
state feedback control law:
compute the optimal control gains F(k) such that stability of
u(k) = F(k)x(k), (21) the metro train traffic and the on-board passenger dynamics
is guaranteed under uncertain disturbances, and meanwhile
where F(k) is the control gain to be designed. In practical
the upper bound on the cost function is minimized. Then,
operations, at each stage k, we can obtain the practical train
the event-triggered mechanism (25) is addressed to implement
departure times and the real-time passenger flow information
the proposed controller.
using real-time monitoring technology. For the given nominal
departure time Dni and expected passenger load L in , we can
further calculate the error state vector x ni = [dni − Dni , pni − A. Stability Condition for the Predictive Controller
L in ]T , i.e., the error state information of train i at stage k.
Then, based on the error state x(k) of all trains at stage k, we Based on the Lyapunov stability theory and H∞ approach,
can calculate the value of the controller u(k) according to the the following theorem provides sufficient conditions for the
proposed state feedback control law. existence of the proposed predictive controller.
For the sake of reducing the amount of controller updates Theorem 3: Consider the error state-space model (17) for
and avoiding the unnecessary waste of network resources, the the train traffic and on-board passenger load with con-
event-triggered control scheme is adopted to address this issue. straints (19). For constant weight matrices Q > 0, R > 0,
Based on the event-triggered theory [26], the control force is and a given disturbance attenuation parameter γ > 0, if there
updated at certain discrete time instants, which is defined as exists matrices X (k) > 0, G(k) > 0, E(k) and (k) with
appropriate dimensions such that the following optimization
u(k) = F(ks )x(ks ), k ∈ [ks , ks+1 ), (22) problem is feasible
where ks is the stage when control forces are re-calculated,
min ϑ(k) (26)
i.e., the control signal remains unchangeable in the interval of E(k),G(k),X (k), (k)
two sampling stages.
subject to the following LMI constraints (27)–(30), shown at
At each stage, a event-triggering condition is checked to
the bottom of the next page where (k) = X (k) − E(k) −
decide whether or not to update the system states and control Û +U Û −U
signals. To depict the deviation of the true states from the E T (k), ϒ(k) = AE(k) + B (k), Ul = l 2 l − | l 2 l |
sampled states, the state measurement error is defined as and h l indicates the l th column of the identity matrix, then
the MPC control law F(k) = (k)E −1 (k) is obtained to
ξ(k) = x(ks ) − x(k), k ∈ [ks , ks+1 ). (23) guarantee the train traffic tracks the nominal timetable and the
Based on the above descriptions, the event-triggered predic- in-vehicle passenger follows the nominal load with the given
tive control problem for train traffic and passenger load to be H∞ performance, and the minimization of an upper limit on
addressed in this article is summarized as follows. the objective function (18) is ensured.
Definition 2: For the error metro train regulation and pas- Proof: At each decision stage k, the Lyapunov function
senger load control model (17) with input constraints (19), is constructed as follows:
(1). Determine a suitable control gain F(k) for the state
feedback control law (21) such that the closed-loop train traffic V (x(k)) = x T (k)P(k)x(k), (31)
system with w(k) = 0 is stable, and the error states satisfy where P(k) > 0 holds.
k f −k To guarantee the stability, we impose the following constraints
1
x T (k + j |k)x(k + j |k) 2
on the difference of the Lyapunov function:
j =k0
k f −k V (x k+ j +1 ) − V (x k+ j )
1  T 
≤γ w (k + j |k)w(k + j |k) ,
T 2 (24) ≤ − x k+ j +1 Qx k+ j +1 + u k+ j Ru k+ j ,
T
(32)
j =k0
where x k+ j denotes x(k + j |k), and u k+ j implies u(k + j |k).
for any nonzero w(k) ∈ L 2 [0, ∞], where k0 is the beginning By summing up (32) from j = k0 to j = k f − k, it yields
stage, and γ > 0 denotes a given disturbance attenuation level; that
(2). The minimization of the upper limit on the train traffic
performance (20) is ensured; J (k) ≤ x kT P(k)x k − x kTf +1 P(k)x k f +1 ≤ V (x k ), (33)
(3). The event-triggered strategy is activated to implement
the proposed stabilizing feedback controller according to (22), which demonstrates that V (x k ) sets the upper bound of J (k).
if the following inequality holds: Setting V (x k ) ≤ ϑ(k), the original optimization problem (20)
can be solved by minimizing ϑ(k), which is equivalently
ξ(k)2 > κx(k)2 , (25)
expressed by the constraint (28) with X (k) = ϑ(k)P −1 (k).
where κ is the parameter to be determined. Next, we detail the derivation of the stability conditions for
If the above three properties hold, it is said that the the desired H∞ robustness. By calculating the difference of
event-triggered predictive control problem is solved. V (x(k)) along the trajectory of the train traffic and on-board

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

passenger model (17) with w(k) ∈ L 2 [0, ∞], it yields disturbances, and the minimization of ϑ(k) is ensured to
that achieve the operative performance.
In addition, to facilitate the derivation of the condition
V (x k+ j +1 ) − V (x k+ j )
 T  T  associated with the input constraints, we reformulate inequal-
x k+ j (k)P(k) (k) − P(k) T (k)P(k)C U +Û Û −U Û +U
ity (19) as − l 2 l ≤ u l − l 2 l ≤ l 2 l . Considering
= ·
wk+ j C T P(k) (k) C T P(k)C Û −U Ûl −U l
  |u l − l 2 l | ≤ |u l | + | 2 |, the input constraint (19) can be
x
× k+ j , (34) ensured by letting
wk+ j
where (k) = A + B F(k). |u l | ≤ Ul , (39)
Noting that −E T (k)X −1 (k)E(k) ≤ (k) and letting Û +U Û −U
where Ul = l 2 l − | l 2 l |.
(k) = F(k)E(k), we use −E T (k)X −1 (k)E(k) to replace
Based on the Cauchy-Schwarz inequality, inequality (39) is
the (k) in (27), and pre-multiply and post-multiply both
relaxed as follows:
sides of (27) by diag{ϑ 2 (k)E −T (k), ϑ − 2 (k)I, . . . , ϑ − 2 (k)I }
1 1 1

and its transpose. Then, the following inequality is derived |u l (k + j |k)|2

1 1
from (27):
⎡ ⎤ = |(Fl (k)P − 2 (k)P 2 (k)xl (k + j |k))|2
−P(k) ∗ ∗ ∗ ∗ ∗ ≤ Fl (k)P −1 (k)FlT (k)xlT (k + j |k)P(k)xl (k + j |k)
⎢ 0 −γ 2 I ∗ ∗ ∗ ∗ ⎥
⎢ ⎥ ≤ ϑ(k)Fl (k)P −1 (k)FlT (k), (40)
⎢ (k) C −1
−P (k) ∗ ∗ ∗ ⎥
⎢ ⎥<0
⎢ I 0 0 −I ∗ ∗ ⎥ where Fl (k) represents the l th row of the control gain F(k).
⎢ ⎥
⎣ F(k) 0 0 0 −R −1 ∗ ⎦ Following similar transform techniques and pre-multiplying
I 0 0 0 0 −Q −1 and post-multiplying both sides of (29) by diag{I, E −1 (k)},
(35) condition (29) is converted to the following form:
 
which can be further equivalently rewritten as the following −G(k) ∗
≤ 0. (41)
form based on the Schur complement F T (k) −ϑ −1 (k)P(k)
 T 
(k)P(k) (k) − P(k) + I T (k)P(k)C Combining inequality (40) and (41), it is obvious that
C T P(k) (k) C T P(k)C − γ 2 I
  ϑ(k)Fl (k)P −1 (k)FlT (k) ≤ G(k), G ll (k) ≤ Ul2 , (42)
Q + F T (k)R F(k) 0
+ < 0. (36) where G ll (k) implies the l th diagonal entry of G(k).
0 0
Thus, the proposed predictive controller for urban train
By integrating inequalities (34) and (36), we have
regulation and passenger load control is robustly stable under
T 2 T
V (x k+ j +1 ) − V (x k+ j ) + x k+ j x k+ j − γ wk+ j wk+ j < 0.(37) operative constraints (19), and the minimization of the perfor-
mance criterion is guaranteed. The proof is completed.
Under the zero-initial condition, we sum up both sides of (37)
Remark 4: The results presented in Theorem 3 show that
from j = k0 to j = k f − k, and the following inequality is
the predictive controller is capable of adjusting the practical
derived:
train departure time and passenger load dynamically such that
k f −k k f −k
the train traffic and in-vehicle passengers track the nominal
T
x k+ j x k+ j −γ 2
wk+
T
j wk+ j ≤ 0, (38) timetable and the expected load under uncertain disturbances.
j =k0 j =k0 By minimizing the upper limit of the objective function, the
which means that the desirable H∞ performance is satisfied. balance between efficiency and control cost is achieved. Based
Thus, the train traffic and in-vehicle passenger load sys- on the state-feedback information x(k), the desirable control
tem (17) is stabilized to the zero point under the uncertain forces u(k) can be calculated at each decision stage based on

⎡ ⎤
(k) ∗ ∗ ∗ ∗ ∗
⎢ 0 −ϑ(k)γ 2 I ∗ ∗ ∗ ∗ ⎥
⎢ ⎥
⎢ϒ(k) ϑ(k)C −X (k) ∗ ∗ ∗ ⎥
⎢ ⎥ < 0, (27)
⎢ E(k) 0 0 −ϑ(k)I ∗ ∗ ⎥
⎢ ⎥
⎣ (k) 0 0 0 −ϑ(k)R −1 ∗ ⎦
E(k) 0 0 0 0 −ϑ(k)Q −1
 
−1 ∗
≤ 0, (28)
x(k|k) −X (k)
 
−G(k) ∗
T (k) ≤ 0, (29)
(k)
h lT G(k)h l ≤ Ul2 , (30)

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG et al.: EVENT-TRIGGERED PREDICTIVE CONTROL FOR AUTOMATIC TRAIN REGULATION AND PASSENGER FLOW 9

the control law F(k) = (k)E −1 (k). In the actual circum- trend to the nominal one, i.e., less than the train capacity
stance, the practical departure time can be exactly collected and will not have negative values under the stability results
due to the wide application of the communication based train in Theorem 3. Therefore, under condition (A5), the proposed
control (CBTC) system, and the real-time passenger load can passenger flow control strategy in Theorem 3.1 ensures the
be dynamically measured by the smart card technology. As stability of the passengers’ dynamics (6)–(8). The proof is
there are limited stations in an urban transit line, the control completed.
strategy could be solved in polynomial time by exploiting the
Matlab LMI toolbox.
Remark 5: It should be noted that the formulated optimiza- B. The Event-Triggering Condition
tion problem (26) may be infeasible under some extreme To reduce the computational burden and avoid the unnec-
conditions. For example, when a great accident occurs, most essary waste of communication resources, the design of the
trains will be severely behind the nominal schedule, which event-triggered mechanism (25) is discussed in this part.
requires large control inputs to compensate the delay. How- Theorem 8: Consider the proposed controller F(k) =
ever, the possible solutions are restricted by the operative (k)E −1 (k) for the train regulation and in-vehicle passenger
constraints (19), which leads to infeasibility of the studied control system (17) with input constraints (19). If the con-
optimization problem. In this case, the original timetable no trol force is updated according to (22) when the following
longer meets the operating requirements under the disturbed event-triggered condition is satisfied
condition, and a new schedule should be regenerated to set
arrival and departure times for the affected trains. ξ(k) > κx(k), (43)
Remark 6: For the LMI conditions (27)–(30) in Theorem 3,
constraint (28) demonstrates the relationship between the where
system stability and the initial train delay and passenger load 
deviations from the nominal value, which together with other F T (ks )B T P(ks )B F(ks )
κ= 2
LMI constraints in Theorem 3.1 provide a sufficient condition λmin ((ks ))

for thresholds on initial passenger loads and delay disturbances  T (ks )P(ks )B F(ks )2 −1
to the stability of passengers’ dynamics. If the initial passenger +4 ,
λ2min ((ks ))
loads deviation and train delay disturbances satisfy condition
(28), and other LMI constraints in Theorem 3 hold, the stabil- (ks ) = P(ks ) − T (ks )P(ks ) (ks ), λmin ((ks )) and
ity of passengers’ dynamics is guaranteed. It should be noted λmax ((ks )) represent the smallest and largest eigenvalues of
that the main results of this study provide a sufficient condition (ks ), respectively, ks is the recent sampling stage and  is
in an implicit form of LMI for the stability of the passengers’ a constant satisfying 0 <  < 1, then the stabilization of the
dynamics. To better explicitly express the correlation between studied system is guaranteed.
the stability of the passengers’ dynamics and the thresholds of Proof: In the event-triggered formulation, we introduce
initial passenger loads and delay disturbances, it is desirable the measurement error ξ(k) in form of (23) to the train traffic
that the sufficient and necessary condition is obtained, which and passenger load model with w(k) = 0, and the closed-loop
will be a challenging and meaningful task for our future system is obtained as follows:
research.
Based on the results of Theorem 3, the following propo- x(k + 1) = (ks )x(k) + B F(ks )ξ(k). (44)
sition is presented to explain the stability of the passengers’
dynamics (6)–(8) under condition (A5). Consider a quadratic Lyapunov function of the form (31).
Proposition 7: Under condition (A5), the proposed passen- Based on the results of [33], [34], the Lyapunov function
ger flow control strategy in Theorem 3 ensures the stability V (x(k)) satisfies the input-to-state stable (ISS) property if
of the passengers’ dynamics (6)–(8), i.e., the actual train load there exists K∞ functions φ1 , φ2 , φ3 and a K function ψ
will not exceed the train capacity, and will not have negative such that
values.    
Proof: According to stability results in Theorem 3, the φ1 V (x(k)) ≤ V (x(k)) ≤ φ2 V (x(k)) , (45)
   
term for the error passenger load (x p in )2 = ( pni − L in )2 is V (x(k + 1)) − V (x(k)) ≤ −φ3 x(k) + ψ ξ(k) . (46)
non-increasing with the number of the station under the pas-
senger flow control strategy, i.e., ( pn+1
i −L in+1 )2 ≤ ( pni −L in )2 . Defining φ1 (y) = λmin (P(ks ))y 2 and φ2 (y) = λmax (P(ks ))y 2 ,
Consider the condition (A5) that the initial passenger load pni it is obvious that inequality (45) holds.
does not exceed the train capacity. If pni exceeds the nominal Then, by calculating the difference of V (x(k)) with respect to
load L in ( pni − L in > 0), we have pn+1
i
− L in+1 ≤ pni − L in under the dynamic equation (44), we have
i
the stability results in Theorem 3, i.e, pn+1 ≤ pni for the same
nominal load for each train i (L n+1 = L n ), which indicates
i i V (x(k + 1)) − V (x(k))
that the actual train load has a decreasing trend to the nominal = −x T (k)(ks )x(k) + 2x T (k) T
(ks )P(ks )B F(ks )ξ(k)
T T T
one, i.e., does not exceed the train capacity and will not have + ξ (k)F (ks )B P(ks )B F(ks )ξ(k), (47)
negative values under (A5). Otherwise, if pni does not exceed
the nominal load L in ( pni − L in ≤ 0), pn+1
i will have an increase where (ks ) = P(ks ) − T (k
s )P(k s ) (ks ).

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

Comparing the formation of (46) and (47), we find that the TABLE II
property (46) can be satisfied by denoting T HE PASSENGER A RRIVAL AND A LIGHTING R ATES
AT D IFFERENT S TATIONS
λmin ((ks ))x(k)2
φ3 (x(k)) = , (48)
 2
 
ψ ξ(k) = F T (ks )B T P(ks )B F(ks )

 T (ks )P(ks )B F(ks )2
+2 ξ(k)2 . (49)
λmin ((ks ))
Note that the dynamic model (44) is ISS-stable, if the follow-
ing inequality holds for ξ(k)
   
ψ ξ(k) ≤ φ3 x(k) , (50)
with 0 <  < 1. By substituting inequality (50) into
condition (46), it follows that
 
V (x(k + 1)) − V (x(k)) ≤ ( − 1)φ3 x(k) , (51)
exceed 30s and the decreasing is less than 40s, and control
which guarantees that V (x(k)) is decreasing. force of the passenger flow is supposed to vary in the interval
Combining (48) – (50), it is easy observe that the train traffic of [-20, 0], i.e., the reduction of boarding passengers is not
and passenger load system (17) can be stabilized to the origin to exceed 20. The proposed method is applied to adjust the
under the event-triggered formulation of (44) by updating the train departure time and in-vehicle passenger load during the
control efforts to enforce condition (43). interval of 7:30 a.m. to 9:00 a.m., which is believed to be the
Thus, the stabilization of the studied system is guaranteed typical morning peak hours for working days.
if the control signal is updated according to (22) under the In example 1, the proposed train traffic model and joint reg-
event-triggered condition (43). The proof is completed. ulation strategy are validated by comparing with a no control
Remark 9: Compared to the conventional MPC based train scenario and an actual regulation strategy. Then, example 2 is
regulation and flow control methods [21], [24], where the used to verify the effectiveness of the proposed event-triggered
control force is computed repeatedly at each decision stage, the mechanism. All of the simulations are performed on a Win-
event-triggered control scheme proposed in Theorem 2 allows dows 7.0 platform with a 3.3 GHz CPU and a 8G RAM.
the control task to be executed when the well-designed con-
dition (43) is activated, which is capable of reducing the A. Example 1
computation times for the control gain, while remaining the In this example, we illustrate the performance of the pro-
desirable closed-loop performance. Furthermore, the proposed posed controller in regulating the train traffic and in-vehicle
method only requires the updated train traffic and passenger passengers under uncertain disturbances. The weight para-
load states to be transmitted only after the condition is meters for the objective function (18) are defined as Q =
triggered. Considering the bandwidth of the wireless network R = diag{0.1, . . . , 0.1}, which indicates that the system
is always limited in the CBTC system, it can effectively avoid performance and control cost are equally treated. We use
the over-utilization of the wireless communication resources. the mincx function from Matlab LMI toolbox to solve the
optimization problem in form of (26), so that the predictive
IV. N UMERICAL S IMULATIONS control law F(k) = (k)E −1 (k) is obtained at each step based
In this section, we present some examples to illustrate the on the real-time traffic and passenger flow information.
effectiveness of the proposed approach for the adjustment of For the initial stage, it is assumed that train 10, 11 and
train traffic and passenger flow. The experiments are imple- 12 are delayed at station 3, 2 and 1. The delay time of train 10,
mented based on real scenarios of Beijing Yizhuang subway 11 and 12 is given as 35s, 45s and 60s, respectively. The initial
line, which includes 14 stations and 13 sections, i.e., N = 13. error states of on-board passenger loads for train 10, 11 and
It is assumed that each train travels from Songjiazhuang station 12 are set as 35, 40 and 50, respectively. Suppose that all of
to Yizhuang station in an one way manner. The predefined the trains are affected by disturbances at the 7th stage, and the
headway is set as H = 180s, and the nominal running delay time is assigned randomly in the interval of [10s, 20s].
time between stations varies from 60s to 100s with respect The H∞ disturbance attenuation level is chosen as γ = 0.25.
to the actual train timetable. Based on the real survey data First, we suppose that no control force is applied to adjust
and the typical value for the delay coefficient in the existing the train traffic and in-vehicle passengers. According to the
literature [24], δ is given as 0.02, and the arrival and alighting joint dynamic model (17), the evolutions of delay time and
rates of passengers at different stations are listed in Table II. passenger load deviation from the nominal value for train 10,
The maximum capacity of the train for passengers Cmax = 11, and 12 are shown in Fig. 4. From Fig.4, we find that both
2500. The input constraints on the schedule regulation are the delay time and the number of passengers keep increasing
given as U l = 40s and Ûl = 30s, which means that the as the trains travel along the line. When the unexpected
increasing of running time and dwell time is not allowed to disturbances happen at the 7th stage, train 10, 11 and 12 are

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG et al.: EVENT-TRIGGERED PREDICTIVE CONTROL FOR AUTOMATIC TRAIN REGULATION AND PASSENGER FLOW 11

TABLE III
T HE C ONTROL F ORCES FOR THE T RAIN R EGULATION AND PASSENGER
L OAD C ONTROL U NDER THE P ROPOSED S TRATEGY

downward tendency, and finally converge to the expected

passenger load after a few stages, which demonstrates that the
nominal timetable and on-board passenger level are recovered,
and the stability of the transit line is therefore ensured.
Moreover, during the adjusting process, we record the applied
Fig. 4. The delay time and passenger load deviation from the nominal value
of train 10, 11 and 12 without control efforts. control efforts, which are listed in Table III. It is easy to notice
that the maximum absolute value of control forces used in
regulating the schedule is 24s, and the magnitude of control
inputs associated with the passenger flow control is 11, which
indicates that all of the inputs are within the boundary of
control constraints.
To examine the robustness of the joint control method with
respect to uncertain disturbances, we introduce the randomly
selected disturbances at stage 7. Different from the situation
in Fig. 4, where the uncertain disturbances make the delays
become even worse without an effective regulation approach,
the proposed strategy can calculate the control efforts dynam-
ically to compensate the disturbances. It is shown from Fig. 5
that the delays are reduced as the trains travel along the
stations, and are finally stabilized to the equilibrium point
after several stages, which demonstrates the robustness of
the proposed controller under a perturbed situation. It worthy
noting that the train delays not only cause the deviations of the
Fig. 5. The delay time and passenger load deviation from the nominal value
practical timetable from the nominal schedule, but also lead
of train 10, 11 and 12 under the joint control strategy. to the fluctuations of the error states of on-board passengers.
Under the state feedback control law, the actual passenger load
converges to the nominal level in presence of the uncertain
further disturbed by 10s, 14s and 18s at station 9, 8 and 7, disturbances, which highlights the importance of designing the
respectively, which causes sharp increasing of the total delay train regulation and passenger load control strategy jointly.
time. It is calculated that the delay time reaches 50s, 65s and Moreover, the obtained results are compared with the real
86s for train 10, 11 and 12 as they arrive at the terminal, where data based on a practical regulation strategy. In actual metro
exists 43%, 44% and 43% increments compared to their initial transit operation, the train regulation is implemented manually
states. Similarly, the number of on-board passengers also rises with the help of the automatic train supervision (ATS) system.
significantly after the trains leave the starting station, which There are several pre-calculated nominal movement curves
greatly degrades the riding comfort. for each inter-station, which corresponds to different running
Next, we apply the proposed train regulation and passenger and dwell times. The operator should determine to choose
load control approach to compute the adjusting strategy. Based one curve from the given movement curves set based on
on Theorem 3, the desirable control efforts are computed at the real movement situation. If a train is greatly delayed
each decision stage, and the simulation results are illustrated from the nominal schedule, the speed curve with shorter
in Fig. 5. As the dimension of F(k) is very high, it is not operation time is chosen to compensate the delay. Otherwise,
listed here to keep the paper concise. the speed curve with longer operation time is adopted to save
It is shown from Fig. 5 that the proposed joint controller is energy consumption. According to the practical operating data
able to reduce the delay time effectively, and the error states of Beijing Changping metro line, the train operation time
of departure time finally recover to the zero point after about corresponding to the nominal strategy (NS) and the adjusting
five stages, which implies that the trains are then running in strategy (AS) with shorter operation time is given in Table IV.
accordance with the nominal schedule. Meanwhile, the error Suppose that train 10 is delayed 60s at the first station,
states of the in-vehicle passenger load also show an apparent and the evolutions for the delay under the proposed and the

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

TABLE IV
T HE D IFFERENT O PERATION T IME F ROM S TATION 1 TO 6

Fig. 6. Comparison with the practical adjusting strategy.

practical adjusting strategies are plotted in Fig. 6. It is easy Fig. 7. The delay time and passenger load deviation from the nominal value
of train 10, 11 and 12 under the event-triggering mechanism.
to observe from Fig. 6 that the delay evolution curve under
the proposed scheme follows the same tendency as the real
TABLE V
curve, which validates the effectiveness of the proposed model.
T HE C ONTROL F ORCES FOR T RAIN R EGULATION AND PASSENGER L OAD
Moreover, as the practical regulation scheme is pre-designed, C ONTROL U NDER THE E VENT-T RIGGERING C ONDITION
it only compensates a fixed delay time at each station, which
leads to that the train is slightly behind the nominal schedule
at station 6 and ahead of the nominal timetable at station 7.
While, the proposed method calculates the control strategy
dynamically and ensures the train tracks the nominal schedule
at station 6, which demonstrates better flexibility.

B. Example 2 To better compare the control performance, the correspond-

In this section, we validate the effectiveness of the ing control efforts under the event-triggering framework are
event-triggering mechanism for the proposed predictive con- recorded in Table V. For Fig. 5 and Fig. 7, the control gain
troller. To facilitate the comparison, all of the simulation are both calculated according Theorem 3.1. By introducing the
conditions remain the same as Example 1. proposed evert-trigger policy, the control forces are updated
At each stage, we check the event-triggering condition based every two steps in Fig. 7, which is the main difference. Taking
on Theorem 8, where the parameter is set as  = 0.15. If the the first two steps of train 10 as an example, in Fig. 5,
condition (43) is satisfied, we first calculate the control gain the control effort used to compensate the train delay time -
F(k) according to Theorem 3, then compute the control forces 14s at the first step, which leads to the real delay time reduce
with the updated control gain and system states. Otherwise, to 46s. Then, based on the recently collected delay time 46s,
we just apply the same control efforts obtained last sampling the control input is obtained as -9s, While, in Fig. 7, as the
stage. The evolutions for the train schedule and the passenger control gain and initial state are equal to the values in Fig. 5,
load deviation from the nominal value are plotted in Fig. 7 (a) the control effort is still -14s at stage 1. Although the actual
and (b), respectively. delay time has reduced to 46s, the controller still uses the
It can be inferred from Fig. 7 (a) and (b) that both the delays obtained -14s to adjust the train movement at stage 2, where
of the trains and the number of on-board passengers show an additional 5s is compensated for the train delay in Fig. 7.
significant reductions as the number of stations increases. By This explains the reason that the train delay is recovered more
comparing the experimental results in Fig. 5, we notice that rapidly than the results in Fig. 5.
the nominal timetable is recovered in the next four stages, It should be noticed that the negative value in Fig.7(b)
which is faster than the predictive control method without indicates that there are fewer passengers than the nominal
the event-triggering strategy. When the uncertain disturbances value, and the total number of onboard passengers is still
happen at stage 7 the error states of the train timetable positive. For instance, the nominal passenger load for the 6A
and the on-board passenger load are eventually stabilized at type train of Beijing subway is 1860, and the negative value
the original, i.e., the disturbed train traffic is recovered to -3 for for train 12 means that there are 3 passengers less
the nominal service level, which demonstrates the stability than the nominal passenger load, i.e., the practical passenger
and robustness of the proposed event-triggering condition in load for train 12 is 1857. The reason for the negative value
handling the train regulation and passenger flow adjustment is due to the event-triggered mechanism. Specifically, as the
problem. event-triggered condition isn’t activated when train 12 arrives

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

WANG et al.: EVENT-TRIGGERED PREDICTIVE CONTROL FOR AUTOMATIC TRAIN REGULATION AND PASSENGER FLOW 13

at station 4, the control force remains the same as the obtained [5] V. Van Breusegem, G. Campion, and G. Bastin, “Traffic modeling and
value last stage, and decreases the same number of on-board state feedback control for metro lines,” IEEE Trans. Autom. Control,
vol. 36, no. 7, pp. 770–784, Jul. 1991.
passengers, which leads to that the actual passenger load is [6] H. Niu and X. Zhou, “Optimizing urban rail timetable under time-
less than the nominal value. Under this condition, the event- dependent demand and oversaturated conditions,” Transp. Res. C, Emerg.
triggered mechanism is automatically activated, and makes Technol., vol. 36, no. 11, pp. 212–230, 2013.
[7] F. Corman and L. Meng, “A review of online dynamic models and
the number of actual on-board passengers converges to the algorithms for railway traffic management,” IEEE Trans. Intell. Transp.
nominal value. Syst., vol. 16, no. 3, pp. 1274–1284, Jun. 2015.
In addition, the average calculation time is approximate 76s [8] X.-Y. Xu, J. Liu, H.-Y. Li, and M. Jiang, “Capacity-oriented passenger
flow control under uncertain demand: Algorithm development and
at each stage. As the nominal headway is usually 90s - 120s real-world case study,” Transp. Res. E, Log. Transp. Rev., vol. 87,
in the morning peak hours, the proposed scheme can meet the pp. 130–148, Mar. 2016.
operational requirement of the real-time application. [9] J. Yin, L. Yang, T. Tang, Z. Gao, and B. Ran, “Dynamic passenger
demand oriented metro train scheduling with energy-efficiency and wait-
ing time minimization: Mixed-integer linear programming approaches,”
V. C ONCLUSION Transp. Res. B, Methodol., vol. 97, pp. 182–213, Mar. 2017.
[10] J. Yin, D. Chen, L. Yang, T. Tang, and B. Ran, “Efficient real-time train
This article investigated the joint train traffic regulation and operation algorithms with uncertain passenger demands,” IEEE Trans.
passenger flow control problem based on the event-triggered Intell. Transp. Syst., vol. 17, no. 9, pp. 2600–2612, Sep. 2016.
control methodology. A state-space model that combined the [11] W.-S. Lin and J.-W. Sheu, “Automatic train regulation for metro lines
evolution of the train departure time with the variation of using dual heuristic dynamic programming,” Proc. Inst. Mech. Eng., F,
J. Rail Rapid Transit, vol. 224, no. 1, pp. 15–23, Jan. 2010.
the in-vehicle passenger load was established, and an optimal [12] S. Li, L. Yang, Z. Gao, and K. Li, “Robust train regulation for metro lines
control problem was formulated to balance the efficiency with stochastic passenger arrival flow,” Inf. Sci., vol. 373, pp. 287–307,
and the control cost. According to the Lyapunov stability Dec. 2016.
[13] L. Yang, X. Zhou, and Z. Gao, “Credibility-based rescheduling model in
theory and the MPC scheme, a set of LMIs was given as the a double-track railway network: A fuzzy reliable optimization approach,”
sufficient conditions for the existence of the train adjustment Omega, vol. 48, pp. 75–93, Oct. 2014.
and passenger load control strategy with respect to certain [14] L. Kang, J. Wu, H. Sun, X. Zhu, and B. Wang, “A practical model for last
train rescheduling with train delay in urban railway transit networks,”
disturbance attenuation level. To reduce the computational Omega, vol. 50, pp. 29–42, Jan. 2015.
burden and avoid over-utilization of the available network [15] P. Vansteenwegen and D. V. Oudheusden, “Developing railway timeta-
resources, this article further designed a novel event-triggered bles which guarantee a better service,” Eur. J. Oper. Res., vol. 173, no. 1,
pp. 337–350, Aug. 2006.
mechanism, where the control gain was updated and the [16] S. Su, X. Li, T. Tang, and Z. Gao, “A subway train timetable optimization
sampled data was transmitted once a triggering condition approach based on energy-efficient operation strategy,” IEEE Trans.
involving the norm of the measurement error was activated. Intell. Transp. Syst., vol. 14, no. 2, pp. 883–893, Jun. 2013.
[17] X. Yang, A. Chen, J. Wu, Z. Gao, and T. Tang, “An energy-efficient
The numerical simulations indicated that the proposed control rescheduling approach under delay perturbations for metro systems,”
method could ensure the practical timetable and number of Transportmetrica B, Transp. Dyn., pp. 1–15, Jan. 2018.
on-board passengers track the nominal train schedule and [18] B. De Schutter, T. van den Boom, and A. Hegyi, “Model predictive
control approach for recovery from delays in railway systems,” Transp.
passenger load in presence of uncertain disturbances. Res. Rec., J. Transp. Res. Board, vol. 1793, no. 1, pp. 15–20, Jan. 2002.
As a future research, it is meaningful to consider multiple [19] W. O. Assis and B. E. A. Milani, “Generation of optimal schedules for
trains running on the segments between successive stations, metro lines using model predictive control,” Automatica, vol. 40, no. 8,
and design a real-time train regulator based on the comprehen- pp. 1397–1404, Aug. 2004.
[20] A. Fernandez, A. P. Cucala, B. Vitoriano, and F. de Cuadra, “Predictive
sive train traffic and passenger dynamic model. In addition, as traffic regulation for metro loop lines based on quadratic programming,”
a more general formulation for the state constraints, the safe Proc. Inst. Mech. Eng. F, J. Rail Rapid Transit, vol. 220, no. 2,
distance constraint can be constructed as dni − dni−1 ≥ Hmin, pp. 79–89, 2006.
[21] S. Li, B. De Schutter, L. Yang, and Z. Gao, “Robust model predictive
where Hmin is the minimum allowable safe headway, and the control for train regulation in underground railway transportation,” IEEE
on-board passengers constraint is in form of pni ≤ Cmax . It Trans. Control Syst. Technol., vol. 24, no. 3, pp. 1075–1083, May 2016.
should be admitted that the stability analysis for the feedback [22] X. Wang, S. Li, S. Su, and T. Tang, “Robust fuzzy predictive control for
automatic train regulation in high-frequency metro lines,” IEEE Trans.
regulators under state inequality constraints becomes more Fuzzy Syst., vol. 27, no. 6, pp. 1295–1308, Jun. 2019.
complicated from a control perspective. Such a nonlinear [23] Y. Wang, B. Ning, T. Tang, T. J. J. van den Boom, and B. De Schutter,
programming problem is usually solved in a rolling horizon “Efficient real-time train scheduling for urban rail transit systems using
iterative convex programming,” IEEE Trans. Intell. Transp. Syst., vol. 16,
and optimization schemes to satisfy the real-time requirement, no. 6, pp. 3337–3352, Dec. 2015.
which is another future research topic. [24] S. Li, M. M. Dessouky, L. Yang, and Z. Gao, “Joint optimal train
regulation and passenger flow control strategy for high-frequency metro
lines,” Transp. Res. B, Methodol., vol. 99, pp. 113–137, May 2017.
R EFERENCES [25] P. Tabuada, “Event-triggered real-time scheduling of stabilizing control
[1] C. Mannino and A. Mascis, “Optimal real-time traffic control in metro tasks,” IEEE Trans. Autom. Control, vol. 52, no. 9, pp. 1680–1685,
stations,” Oper. Res., vol. 57, no. 4, pp. 1026–1039, Aug. 2009. Sep. 2007.
[2] J.-W. Sheu and W.-S. Lin, “Adaptive optimal control for designing [26] W. P. M. H. Heemels and M. C. F. Donkers, “Model-based periodic
automatic train regulation for metro line,” IEEE Trans. Control Syst. event-triggered control for linear systems,” Automatica, vol. 49, no. 3,
Technol., vol. 20, no. 5, pp. 1319–1327, Sep. 2012. pp. 698–711, Mar. 2013.
[3] T. Dewilde, P. Sels, D. Cattrysse, and P. Vansteenwegen, “Improving the [27] L. Yang, K. Li, and Z. Gao, “Train timetable problem on a single-line
robustness in railway station areas,” Eur. J. Oper. Res., vol. 235, no. 1, railway with fuzzy passenger demand,” IEEE Trans. Fuzzy Syst., vol. 17,
pp. 276–286, May 2014. no. 3, pp. 617–629, Jun. 2009.
[4] R. Liu, S. Li, and L. Yang, “Collaborative optimization for metro train [28] H. Wang, F. Schmid, L. Chen, C. Roberts, and T. Xu, “A topology-based
scheduling and train connections combined with passenger flow control model for railway train control systems,” IEEE Trans. Intell. Transp.
strategy,” Omega, vol. 90, Jan. 2020, Art. no. 101990. Syst., vol. 14, no. 2, pp. 819–827, Jun. 2013.

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

[29] Y. Wang, B. De Schutter, T. J. J. van den Boom, B. Ning, and Tao Tang (Senior Member, IEEE) received the
T. Tang, “Efficient bilevel approach for urban rail transit operation Ph.D. degree in engineering from the Chinese Acad-
with stop-skipping,” IEEE Trans. Intell. Transp. Syst., vol. 15, no. 6, emy of Science, Beijing, China, in 1991. He is
pp. 2658–2670, Dec. 2014. currently a Professor with the School of Electron-
[30] Z. Hou, H. Dong, S. Gao, G. Nicholson, L. Chen, and C. Roberts, ics and Information Engineering, Beijing Jiaotong
“Energy-saving metro train timetable rescheduling model considering University, Beijing, and the Director of the Rail
ATO profiles and dynamic passenger flow,” IEEE Trans. Intell. Transp. Traffic Control and Safety State Key Laboratory.
Syst., vol. 20, no. 7, pp. 2774–2785, Jul. 2019. His research interests include communication-based
[31] N. Farhi, C. Nguyen Van Phu, H. Haj-Salem, and J.-P. Lebacque, “Traffic train control, high-speed train control, and intelli-
modeling and real-time control for metro lines. Part II–The effect of gent transportation systems. He is a member of the
passenger demand on the traffic phases,” in Proc. Amer. Control Conf. Experts Group of High Technology Research and
(ACC), May 2017, pp. 3828–3833. Development Program of China (863 Program) and leads the Field of Modern
[32] N. Farhi, “Physical models and control of the train dynamics in a metro Transportation Technology Experts Group. He is also a specialist of National
line without junction,” IEEE Trans. Control Syst. Technol., vol. 27, no. 4, Development and Reform Commission and Beijing Urban Traffic Construction
pp. 1829–1837, Jul. 2019. Committee.
[33] Z.-P. Jiang and Y. Wang, “Input-to-state stability for discrete-time
nonlinear systems,” Automatica, vol. 37, no. 6, pp. 857–869, Jun. 2001.
[34] A. Eqtami, D. V. Dimarogonas, and K. J. Kyriakopoulos, “Event-
triggered control for discrete-time systems,” in Proc. Amer. Control
Conf., Jun. 2010, pp. 4719–4724.

Xi Wang received the M.S. and Ph.D. degrees in

traffic control and information engineering from
Beijing Jiaotong University, Beijing, China,
in 2006 and 2017, respectively. He is currently
an Associate Professor with the School of
Electronics and Information Engineering, Beijing
Jiaotong University. His research interests include
communication-based train control, intelligent train
operation, and model predictive control.
Lixing Yang received the B.S. and M.S. degrees
from the Department of Mathematics, Hebei Uni-
versity, Baoding, China, in 1999 and 2002, respec-
Shukai Li received the M.S. and Ph.D. degrees tively, and the Ph.D. degree from the Department
in systems engineering from Tianjin University, of Mathematical Sciences, Tsinghua University,
Tianjin, China, in 2010 and 2013, respectively. He is Beijing, China, in 2005.
currently an Associate Professor with the State Since 2005, he has been with the State Key Labo-
Key Laboratory of Rail Traffic Control and Safety, ratory of Rail Traffic Control and Safety, Beijing
Beijing Jiaotong University, China. His research Jiaotong University, Beijing, China, where he is
interests include nonlinear systems and robust con- currently a Professor. He has published more than
trol, and their applications in railway traffic. 70 articles in national conferences, international con-
ferences, and premier journals. His current research interests include uncertain
programming, intelligent systems, and applications in transportation planning
and rail traffic control systems.

Authorized licensed use limited to: Carleton University. Downloaded on November 06,2020 at 11:30:22 UTC from IEEE Xplore. Restrictions apply.