Transportation Research Part C 69 (2016) 255–275

Contents lists available at ScienceDirect

Transportation Research Part C

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

Multiple-phase train trajectory optimization with signalling

and operational constraints
Pengling Wang a,b,⇑, Rob M.P. Goverde a
a
Department of Transport and Planning, Delft University of Technology, Delft, The Netherlands
b
School of Electrical Engineering, Southwest Jiaotong University, Chengdu, China

a r t i c l e i n f o a b s t r a c t

Article history: The train trajectory optimization problem aims at finding the optimal speed profiles and
Received 5 January 2016 control regimes for a safe, punctual, comfortable, and energy-efficient train operation.
Received in revised form 29 May 2016 This paper studies the train trajectory optimization problem with consideration of general
Accepted 7 June 2016
operational constraints as well as signalling constraints. Operational constraints refer to
Available online 16 June 2016
time and speed restrictions from the actual timetable, while signalling constraints refer
to the influences of signal aspects and automatic train protection on train operation. A rail-
Keywords:
way timetable provides each train with a train path envelope, which consists of a set of
Train trajectory optimization
Railway signalling
positions on the route with a specified target time and speed point or window. The train
Pseudospectral method trajectory optimization problem is formulated as a multiple-phase optimal control model
Optimal control and solved by a pseudospectral method. This model is able to capture varying gradients
and speed limits, as well as time and speed constraints from the train path envelope.
Train trajectory calculation methods under delay and no-delay situations are discussed.
When the train follows the planned timetable, the train trajectory calculation aims at min-
imizing energy consumption, whereas in the case of delays the train trajectory is re-
calculated to track the possibly adjusted timetable with the aim of minimizing delays as
well as energy consumption. Moreover, the train operation could be affected by yellow
or red signals, which is taken into account in the train speed regulation. For this purpose,
two optimization policies are developed with either limited or full information of the train
ahead. A local signal response policy ensures that the train makes correct and quick
responses to different signalling aspects, while a global green wave policy aims at avoiding
yellow signals and thus proceed with all green signals. The method is applied in a case
study of two successive trains running on a corridor with various delays showing the ben-
efit of accurate predictive information of the leading train on energy consumption and train
delay of the following train.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction

Improving transport capacity and saving energy consumption are the most urgent challenges faced by modern railway
transportation all around the world. Optimizing train operation is one promising method, which doesn’t need extra
infrastructure, but improves rail traffic efficiency by optimizing train speed and control trajectories. One core function of
train operation optimization is train trajectory calculation, which uses optimal control theory to calculate the optimal speed

⇑ Corresponding author at: Department of Transport and Planning, Delft University of Technology, P.O. Box 5048, 2600 GA Delft, The Netherlands.
E-mail addresses: P.L.Wang@tudelft.nl (P. Wang), R.M.P.Goverde@tudelft.nl (R.M.P. Goverde).

http://dx.doi.org/10.1016/j.trc.2016.06.008
0968-090X/Ó 2016 Elsevier Ltd. All rights reserved.
256 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275

profiles and control regimes, aiming at safe, on-time and energy saving train operation. These profiles are used to generate
driving advise to support train drivers in train control.
Research on train trajectory calculation started in the 1960s. The solution methods of the train trajectory optimization
problem can be divided into three categories: indirect methods, direct methods and artificial intelligence algorithms. The
indirect approach solves the problem indirectly by converting the optimal control problem to a boundary-value problem.
The direct method finds the optimal solution by transcribing a continuous optimization problem to a nonlinear program-
ming problem (NLP). Researchers who focus on indirect methods are interested largely in solving differential equations,
while researchers who focus on direct methods are interested more in optimization techniques (Betts, 1998; Rao, 2009). Pon-
tryagin’s Maximum Principle is a typical indirect method. The optimal train control strategy following from application of
Pontryagin’s Maximum Principle to a long journey on flat track with sufficient running time supplement consists of the
sequence Maximum Power–Cruising–Coasting–Maximum Braking (Cheng and Howlett, 1992; Howlett and Pudney, 2012;
Milroy, 1980). For a train operating on a track with varying speed limits and gradients the optimal control strategy is a
sequence of these optimal regimes where the succession of regimes and their switching points also depends on the speed
limits and gradients (Howlett, 1996; Khmelnitsky, 2000; Liu and Golovitcher, 2003; Pudney and Howlett, 1994). Finding
the optimal switching points is a difficult problem except for simple cases such as a single speed limit and flat track
(Albrecht et al., in press-b; Albrecht et al., in press-c). Direct approaches transform the optimal control problem into a math-
ematical programming problem. Wang et al. (2013, 2015) and Wang and Goverde (2016) reformulate the problem as a
multiple-phase optimal control model, and solve it with Pseudospectral methods (Gong et al., 2008; Rao, 2003; Ross and
Fahroo, 2004; Ross and Karpenko, 2012). Pseudospectral methods transcribe the continuous-time optimal control problem
into a nonlinear programming problem, after which nonlinear programming solvers are adopted to directly solve the prob-
lem. In addition, a variety of heuristic and artificial intelligence algorithms have been applied to locate the energy-efficient
train trajectory, such as fuzzy predictive algorithms (Yasunobu and Miyamoto, 1985) and genetic algorithms (Kang, 2011).
The classic single-train control problem focuses on one independent train from one station to the next under a scheduled
traffic plan. Dynamic influences such as delays and signalling systems, are considered only recently. Delays or other distur-
bances cause deviations from the traffic plan, in which case the train may meet yellow or red signals, which require speed
reductions and unscheduled stops. A rescheduling process is required to produce a new timetable when the deviation is big
enough. As a result, the train trajectories also need to be adjusted accordingly. Albrecht et al. (2010) considered the influence
of signalling and automatic train protection on the train trajectory optimization. This research is based on the optimal con-
trol regimes obtained from Pontryagin’s Maximum Principle, and focuses on finding the optimal switching points to handle
the influence of the signalling system. Albrecht et al. (2011) discuss energy-efficient delay recovery strategies for trains in
opposite directions. They find a set of interaction times that allows each affected train to recover from delays as well as
to save energy consumption, but energy-efficient train trajectory calculation is not discussed. Albrecht et al. (2015a) study
the safe separation problem for two trains travelling in the same direction. To satisfy the safe separation for two following
trains. An optimal set of specified intermediate clearance times for each section is calculated, which also aims at minimizing
total energy consumption. Wang et al. (2014) consider the train trajectory planning problem under fixed and moving block
signalling systems. They transform the optimal control problem into a mixed-integer linear programming problem. The non-
linear train dynamic movement model is simplified into a linear model, which speeds up the computation process but
degrades the solutions’ accuracy.
This paper gives several contributions to the literature. First, a rescheduling process may change the train’s traffic plan,
which requires adjustment of the train speed to track the new timetable. Second, a safe separation between trains running
on the same line should be guaranteed. Third, an accurate calculation taking into account operational and signalling con-
straints is necessary since the train trajectory is designed to help train drivers in practical operation. Based on these three
points, this paper formulates the real-time traffic plan for each train as a train path envelope (TPE), which was proposed first
by Albrecht et al. (2013), see also ON-TIME (2014). Similar to the quadratic time geography theory (Ma et al., 2016; Zhou
et al., 2015), train path envelopes set bounds to feasible trajectory ranges, including time and/or speed ranges at specific
points. These time and/or speed ranges are available in real operation along a train run, within which the train can move
without running late with respect to the timetable and hindering other trains’ operations. If the timetable is changed by
a rescheduling process, a new train path envelope must be generated and sent to the train trajectory calculation module.
Based on this, new speed and control profiles are calculated.
Train separation is guaranteed by a signalling system. Generally speaking, if the train operation deviates from a conflict-
free timetable, it might meet yellow or red signals. The influence from the signalling system cannot be ignored when calcu-
lating train trajectories, and in particular the information available about the future signal aspects affects the train trajectory
calculation. Two different scenarios are proposed about the amount of signal information available to the trajectory opti-
mization. In the first scenario, we assume that only information is available about the next signal aspect. An optimization
strategy called signal response policy is developed to ensure that the train makes correct and quick responses to different sig-
nalling aspects. In the other scenario, we assume that a full prediction is available about the signal aspect timings in rear of
the train ahead. A green wave policy (Corman et al., 2009) is then used to avoid yellow signals and thus separate successive
trains. The focus of this paper is on successive trains in the same direction over the same line. The signalling system dis-
cussed is the Dutch signalling system, which is a variant of a three-aspect two-block system with additional speed indica-
tions together with a continuous ATP system. The method is however generic and any signalling system can be taken into
account. Moreover, the work assumes an advanced traffic management environment such as the ON-TIME real-time railway
 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275 257

traffic management framework (Quaglietta et al., 2016). Within such framework, real-time communication is possible for
real-time train trajectory optimization.
For accurate calculations, a nonlinear model is used for the train movement formulation with accurate varying speed lim-
its and gradients. The train trajectory optimization problem is built as a multiple-phase optimal control model, and solved
with a Pseudospectral method. The multiple-phase optimal control model and Pseudospectral method have been used before
for modeling the train trajectory optimization problem in our previous works (Wang and Goverde, 2016; Wang et al., 2015).
This paper extends it to real-time train trajectory optimization with consideration of signalling and operational constraints.
The paper is organized as follows: Section 2 introduces the basic train dynamic movement model and the train path
envelope and shows how to formulate the problem into a multiple-phase optimal control problem and solve it with a Pseu-
dospectral method. In Section 3, the train trajectory calculation with consideration of delays is developed with respect to two
optimization policies incorporating the impact of the signalling system and the online information available. Section 4 illus-
trates the approach in a case study, and finally Section 5 ends the paper with conclusions.

2. Train trajectory optimization modeling and solving method

2.1. Basic train trajectory optimization model

The movement of a railway vehicle is determined by a set of physical constraints such as the timetable, speed limits, and
other vehicle-related factors. The general equation of train motion can be written as follows (Hansen and Pachl, 2014; Wang
et al., 2015):

dv ðsÞ h1 f ðsÞ  h2 bðsÞ  Rtrain ðv Þ  Rline ðsÞ

¼ ð1Þ
ds q  m  v ðsÞ
dtðsÞ 1
¼ ; ð2Þ
ds v ðsÞ
where s the traversed path [m], v ðsÞ is the train velocity [m/s], q the rotating mass factor, m the train mass [t], f ðsÞ the trac-
tion force [kN], bðsÞ the braking force [kN], Rtrain ðv Þ the train resistance force [kN], Rline ðsÞ the line resistance force [kN], tðsÞ
the traversed time [s], and h1 ; h2 2 f0; 1g two binary parameters with h1  h2 ¼ 0. Distance is chosen as the independent vari-
able because gradients and speed limits occur as functions of distance rather than of time.
The train resistance Rtrain ðv Þ comprises rolling, bearing, dynamic and wind resistances (Hansen and Pachl, 2014), and can
be described as

Rtrain ðv Þ ¼ 0:001  q  m  g  ða þ b  v þ c  v 2 Þ; ð3Þ

where g is the acceleration of gravity, and a; b and c are constant coefficients. The line resistance Rline ðsÞ is a function of posi-
tion and consists of two components: grade resistance Rgrade and curve resistance Rcurv e ,

Rline ðsÞ ¼ Rgrade þ Rcurv e ; ð4Þ

where the grade and curve resistances are constant around position s. Train traction and braking power are limited by the
adhesion between the wheels and the rails as well as the maximum power possible to be produced by the engine, so that

0 6 f ðsÞ 6 F max ð5Þ

0 6 bðsÞ 6 Bmax ð6Þ
0 6 f ðsÞ  v ðsÞ 6 Pmax ; ð7Þ

where F max ; Bmax and P max are the upper bounds on traction force, braking force and traction power.
The train speed cannot exceed the speed limits, i.e.,

0 6 v ðsÞ 6 V max ðsÞ; ð8Þ

where V max ðsÞ is the train speed limit at position s, including static and temporary speed restrictions.
The riding comfort is usually measured by train acceleration, which should satisfy

dv ðsÞ
Amin 6 6 Amax ; ð9Þ
dtðsÞ

where Amin and Amax are the lower and upper bound of acceptable riding comfort, respectively.
For a train running between two stops, the timetable restricts the departure and arrival time, which can be formulated as

v ðs0 Þ ¼ 0; tðs0 Þ ¼ T 0 ; v ðsf Þ ¼ 0; tðsf Þ ¼ T f ; ð10Þ

where s0 and sf are the positions of the departure and arrival stations, respectively, and T 0 and T f are the scheduled departure
and arrival time assigned by the timetable.
258 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275

p1 position p2 position

T2
T1
arrival time
Hpassing

departure time T3
Hpassing

time
time

T4

(a) stops (b) passing through stations

p3 position p4 position

T9
T5
Hovertaking
T6 Hjunction
Hovertaking T7
Hjunction
time

time
T8

(c) overtaking (d) passing through junctions

:Time target point :Time window

Fig. 1. Time-distance diagrams of example target time points and time windows.

Generally, the train trajectory optimization problem is to find a series of control laws for the train traction and braking
forces that minimizes train energy consumption, i.e.,
Z sf
J¼ f ðsÞds: ð11Þ
s0

The problem can be formulated as a generic optimal control problem with the train speed and time as the state variables and
the traction and braking force as the control variables. Define the state vector x ¼ ½x1 ; x2 0 ¼ ½v ; t0 and the control vector
0
u ¼ ½u1 ; u2 0 ¼ ½f ; b . Then (1)–(11) can be written as the generic optimal control problem:
8 Rs
>
> Minimize J ¼ s0f ‘ðxðsÞ; uðsÞ; sÞds ðCost functionÞ
>
>
< _
subject to xðsÞ ¼ f ðxðsÞ; uðsÞ; sÞ ðDynamic constraintsÞ
ð12Þ
>
> g 6 gðxðsÞ; uðsÞ; sÞ 6 g max ðPath constraintsÞ
>
> min
:
eðx0 ; xf ; s0 ; sf Þ ¼ E ðBoundary constraintsÞ

where x0 ¼ xðs0 Þ and xf ¼ xðsf Þ. The cost function is minimizing the energy consumption (11), the dynamic constraints con-
sist of (1) and (2), the path constraints represent (5)–(9), and the boundary constraints are (10).

2.2. Train path envelope

Model (12) only considers the departure and arrival time constraints in (10), while more time constraints exist such as
passing-through times at junctions or non-served stations. A train path envelope is used to describe those time and speed
allowances. A train path envelope is a series of targets for each train at specific positions such as station platform stops
or signals, based on estimated earliest and latest passing times and speed limits. The targets are defined by triples of position,
time and speed information. Two kinds of targets are distinguished:

1. Target points ðp; t; v Þ: indicating that a train must reach target position p at the specified time t and speed v,
2. Target windows ðp; ½tmin ; tmax ; ½v min ; v max Þ: indicating that a train must reach target position p within a time window
½tmin ; tmax  and speed window ½v min ; v max .
 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275 259

The target windows may also be specified for a fixed time of speed, in which case the lower and upper bound of the
window are equal. With this notation, the train path envelope for a given train i can be written as a series of target points
and windows
 m
TPEi ¼ ðpi;j ; ½ti;j;min ; t i;j;max ; ½v i;j;min ; v i;j;max Þ j¼1

where m is the number of target positions along train’s route, pi;j refers to the j-th target points for train i; j 2 ½1; 2; . . . ; m, and
pi;j < pi;jþ1 . ti;j;min ; ti;j;max ; v i;j;min and v i;j;max are respectively the minimum and maximum time and speed limits for train i at
pi;j .
The target positions can be stations, junctions, signal positions, and route release points. The target speed for stop points
is defined as v ¼ 0. For other target positions speed information is optional. Only some crucial positions need speed restric-
tions to assure the speed limits or operational constraints such as minimum speeds before slopes or tunnels. The upper
bounds of the time targets in the train path envelope need to be bigger than the minimal technical running time under con-
sideration of the speed limits, maximum traction and braking power, and lower than some maximal running time. In addi-
tion, the train path envelopes of the various trains must be mutually exclusive so that the trains have no conflicts as long as
they stay within their envelopes. This means that sufficient headway times are required between the envelopes of adjacent
trains.
In the following we will discuss the time restrictions at stations and junctions. Since a wide range of situations may occur,
we only present four typical examples (Fig. 1). In these examples, we assume for simplicity local minimum headway times
independent of the train orders, although in practice the minimum headway times depend on the location and order of the
train sequence.

Stops: The real-time traffic plan for each train indicates fixed arrival and departure times at their stops (Albrecht and
Dasigi, 2014). Even slight delays might be perceived directly by customers. Therefore, the arrival and departure times
p
are modeled as target points. In Fig. 1(a), the target points of train T 1 at station p1 can be formulated as (p1 ; AT11 ; 0)
p p p
and (p1 ; DT11 ; 0), where AT11 and DT11 are respectively the scheduled arrival and departure time of train T 1 at stop position
p1 , and the target speeds are 0.
Passing through stations: Each passing event has its own passing time which is assigned by the timetable. A small time
deviation without impacting other trains and passengers is acceptable. Fig. 1(b) shows an example of three trains
T 2 ; T 3 and T 4 around a station at position p2 . Train T 2 has a scheduled stop at the station while train T 3 and T 4 are
non-stop. The passing times at the station for trains T 3 and T 4 are modeled as target time windows. The time window
is computed in such a way that no two different trains have overlapping windows and a minimum headway time is main-
tained between two successive trains. Taking train T 3 in Fig. 1(b) as an example, the time window is calculated as
p p p p p p
½DT22 þ Hpassing ; ðP T23 þ P T24  Hpassing Þ=2, where DT22 is the departure time of train T 2 at p2 ; P T23 and PT24 are the scheduled
passing through times of train T 3 and T 4 at p2 , and Hpassing is the minimum headway time.
Overtaking: Overtaking is similar with passing through a station for a non-stop train. The passing time can be modeled as
a time window respecting the given flexibility in the timetable. Consider the situation in Fig. 1(c), where train T 6 over-
takes train T 5 at a station at position p3 . The overtaking time of train T 6 should be kept within a range, otherwise the over-
taken train T 5 has to wait and is delayed. The minimum headway times for overtaking must be respected by train T 6
p p p p
resulting in the target time window ½AT35 þ Hov ertaking ; DT35  Hov ertaking , where AT35 and DT35 are the scheduled arrival and
departure time of train T 5 at p3 , and Hov ertaking is the minimum overtaking headway time.
Passing through junctions: Junctions are locations where different lines cross or merge, but none of the trains has a planned
stop. Train operations at junctions may result in knock-on delays, so that a passing time window for each train is adopted
to avoid interactions. Consider the situation shown in Fig. 1(d). The time window of the passing time of train T 7 at junc-
p p p p p p p
tion p4 can be calculated as ½ðPT47 þ P T49 þ Hjunction Þ=2; ðPT47 þ PT48  Hjunction Þ=2, where P T47 ; PT48 and PT49 are the passing
through times of train T 7 ; T 8 and T 9 indicated by the timetable, and Hjunction is the minimum headway.

Above four cases show how to transform the timetable into detailed formulation of time and speed constraints in train
path envelope. It can be adopted easily for the computations related to train speed trajectory optimization and control. An
application example is shown in Section 4. Note all the time windows above should intersect with the minimum and
maximum running time constraints. Time constraints at stations and junctions might be changed in case of delays. And more
time constraints might be present at signal positions, which is discussed in Section 3.

2.3. Multiple-phase optimal control model

It is difficult to use model (12) in Section 2.1 for solving the trajectory optimization problem, since the speed limits and
gradients change along the railway track. And not only the arrival and departure time constraints considered in (12) should
be respected by train operation, but also more time and/or speed limits from train path envelope should be considered. So
we re-formulated the problem in a multiple-phase optimal control model. The advantages of this model is accurate illustra-
tion of changing speed limits, gradients and time/speed constraints from train path envelope.
260 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275

In a multiple-phase optimal control model, the trajectory consists of a collection of phases (Rao, 2003). A phase is any
segment of the complete trajectory, where any particular phase of an optimal control problem has its own cost function,
dynamic model, path constraints, and boundary conditions. The complete trajectory is then obtained by properly linking
adjacent phases via linkage conditions. The total cost function is the sum of the cost functions within each phase. The opti-
mal trajectory is then found by minimizing the total cost functional subject to the constraints within each phase and the
linkage constraints connecting adjacent phases.
Since the speed limits and gradients change along the rail track, the complete train trajectory can be divided into several
segments by the critical points of speed limits and gradients, so that each phase has a unique speed limit and line resistance.
Consider a train trajectory that consists of R distinct phases, let the independent variable in phase r 2 ½1; . . . ; R lie in the
ðrÞ 0
xðrÞ ¼ ½x1 ; x2  ¼ ½v ðrÞ ; tðrÞ 
ðrÞ ðrÞ ðrÞ 0
interval s 2 ½s0 ; sf  and denote the state and control in phase r as and
ðrÞ ðrÞ ðrÞ 0 ðrÞ ðrÞ 0
u ¼ ½u1 ; u2  ¼ ½f ; b  . The dynamic constraints in phase r 2 ½1; . . . ; R are given as
ðrÞ
_ ðrÞ
x ðsÞ ¼ f ðx ðsÞ; uðrÞ ðsÞ; sÞ;
ðrÞ
ðr 2 ½1; . . . ; RÞ: ð13Þ
where
2 ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ
3
h1 u1 h2 u2 Rtrain ðx1 ÞR ðsÞ
line
ðrÞ 6 qmxðrÞ 7
f ðxðrÞ ðsÞ; uðrÞ ðsÞ; sÞ ¼ 4 1
5; ð14Þ
1
ðrÞ
x1

ðrÞ
and the value of Rline ðsÞ changes from phase to phase depending on the value of the gradient and curve in phase r. The path
constraints in phase r 2 ½1; . . . ; R are given as
ðrÞ
g min 6 g ðrÞ ðxðrÞ ðsÞ; uðrÞ ðsÞ; sÞ 6 g ðrÞ
max ; ðr 2 ½1; . . . ; RÞ: ð15Þ
where
2 3 2 ðrÞ 3 2 3
0 u1 F max
6 7 6 7 6 7
6 0 7 6 uðrÞ 7 6 Bmax 7
6 7 6 2 7 6 7
6 ðrÞ 7
¼6 7 ¼6 7
ðrÞ
g min 6 0 7; g ðrÞ ðxðrÞ ðsÞ; uðrÞ ðsÞ; sÞ ¼ 6 uðrÞ  x1 7 ; g ðrÞ 6 Pmax 7: ð16Þ
6 7 6 1 7 max
6 ðrÞ 7
4 0 5 6 ðrÞ 7 4 V max 5
4 x1 5
Amin ðrÞ ðrÞ
x_ 1 =x_ 2 Amax

Here, g ðrÞ includes the constraints of train traction, braking force, engine power, speed limit and riding comfort. Note that
ðrÞ ðrÞ
V ðrÞ
max is the maximum speed allowance from s0 to sf , which is a constant value in phase r, but might have different values
in different phases.
The trajectory is also partitioned into more phases by the target positions of the train path envelope. For each target posi-
tion of the train path envelope a time and speed allowance is needed to restrain train operations, which can be formulated as
the boundary constraints. For example, assume target position sp of the train path envelope is the linkage point of phase r
and r þ 1 (r 2 ½1; . . . ; R  1). If there is a target point constraint at sp , we have
" #
V sp
xðsp Þ ¼ ; ð17Þ
T sp

where V sp are T sp are the target speed and time at sp , and if the constraint is a flexible target window we have
" # " #
V sp ;min V sp ;max
6 xðsp Þ 6 ; ð18Þ
T sp ;min T sp ;max

where V sp ;min ; V sp ;max ; T sp ;min and T sp ;max are the minimum and maximum speed and time at sp , respectively. Since sp is the
boundary between two adjacent phases, constraints (17) and (18) work as the terminal boundary constraints for phase r
as well as the initial boundary constraints for phase r þ 1. In general, the boundary constraints (if any) in phase
r 2 ½1; . . . ; R are given as
ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ
emin 6 eðrÞ ðx0 ; s0 ; xf ; sf Þ 6 eðrÞ
max ; ðr 2 ½1; . . . ; RÞ: ð19Þ
ðrÞ ðrÞ
where the lower and upper bound in emin and emax are equal if the constraint is a target point.
The whole train trajectory is divided into multiple segments with critical points of speed limits or gradients and curves,
and target positions of the train path envelope. For each two consecutive phases k and k þ 1; k 2 ½1; . . . ; R  1, a set of con-
ditions is used to connect the trajectories in phase k and k þ 1. In particular, the state variables must be continuous at the
boundary between phase k and phase k þ 1. Therefore, the following linkage conditions must be satisfied for all
k 2 ½1; . . . ; R  1:
 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275 261

ðkÞ ðkþ1Þ ðkÞ ðkþ1Þ

sf  s0 ¼ 0; xðsf Þ  xðs0 Þ ¼ 0: ð20Þ
ðkÞ ðkþ1Þ
An exception occurs if the linkage point of phase k and k þ 1 is the stop point, in which case x2 ðsf Þ and x2 ðs0 Þ represent
the arrival and departure time, and thus
ðkþ1Þ ðkÞ
x2 ðs0 Þ  x2 ðsf Þ ¼ DsðkÞ ; ð21Þ
f

ðkÞ ðkþ1Þ
where DsðkÞ is the dwell time of the train at sf (or s0 ). The general linkage conditions can be formulated as
f

ðkÞ ðkÞ ðkþ1Þ ðkþ1Þ

lðxf ; sf ; x0 ; s0 Þ ¼ L; ðk 2 ½1; . . . ; R  1Þ: ð22Þ

The cost function J ðrÞ in phase r 2 ½1; . . . ; R can be the energy consumption given as
Z ðrÞ
sf
J ðrÞ ¼ ‘ðrÞ ðxðrÞ ðsÞ; uðrÞ ðsÞ; sÞds; ð23Þ
ðrÞ
s0

ðrÞ
‘ðrÞ ðxðrÞ ðsÞ; uðrÞ ðsÞ; sÞ ¼ u1 ðsÞ, although later we also include delay. The objective of the multiple-phase train trajectory opti-
mization problem is to minimize the cost function over all phases
X
R
J¼ J ðrÞ ; ð24Þ
r¼1

subject to the dynamic constraints (13), path constraints (15), boundary constraints (19), and the linkage conditions (22).

2.4. Pseudospectral method

The multiple-phase optimal control model in Section 2.3 is solved by a Pseudospectral method. Pseudospectral method
transcribes the continuous-time optimal control problem into a nonlinear programming problem. The state and control func-
tions are approximated using a set of orthogonal polynomials (Chebyshev or Lagrange polynomials), where specified collo-
cation points are used for collocation of the dynamics and a quadrature approximation of the integrated Lagrange cost term.
The Pseudospectral method has a simple structure and converges exponentially. The most well-developed Pseudospectral
methods are the Gauss Pseudospectral Method (GPM) (Benson, 2005; Huntington, 2007), the Radau Pseudospectral Method
(RPM) (Garg, 2011), and the Lobatto Pseudospectral Method (LPM) (Elnagar et al., 1995).
For the Radau Pseudospectral Method, the multiple-phase train trajectory optimization model described into the previous
ðrÞ ðrÞ
section is transcribed to a NLP as follows. The first step is to map the physical domain s 2 ½s0 ; sf  to a computational domain
rðrÞ 2 ½1; 1 by means of the affine transformation
ðrÞ ðrÞ
2s sf þ s0
rðrÞ ¼ ðrÞ ðrÞ
 ðrÞ ðrÞ
: ð25Þ
sf  s0 sf  s0

Next, let N ðrÞ be the number of Legendre–Gauss–Radau (LGR) points rjðrÞ in phase r 2 ½1; . . . ; R, with
r 2 ½1; 1Þ; j 2 ½1; . . . ; N , and r ¼ 1 and r
ðrÞ
j
ðrÞ ðrÞ
1
ðrÞ
NðrÞ
< 1. Let r ðrÞ
NðrÞ þ1
¼ 1, which is a non-collocation point. Then the state
and control of phase r 2 ½1; . . . ; R are approximated by a basis of Lagrange interpolating polynomials
ðrÞ
NX þ1
x ðr Þ  X ðr Þ ¼ X j Lj ðrðrÞ Þ;
ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ
ð26Þ
j¼1

N ðrÞ
X
uðrÞ ðrðrÞ Þ  U ðrÞ ðrðrÞ Þ ¼ Uj eL j ðrðrÞ Þ;
ðrÞ ðrÞ
ð27Þ
j¼1

where X j ¼ X ðrÞ ðrj Þ; U j ¼ U ðrÞ ðrj Þ,

ðrÞ ðrÞ ðrÞ ðrÞ
and the Lagrange polynomials Lj ðrðrÞ Þ
ðrÞ
ðj 2 ½1; . . . ; N ðrÞ þ 1Þ and e
L j ðrðrÞ Þ
ðrÞ

(j 2 ½1; . . . ; N ðrÞ ) are defined as

ðrÞ
NY N ðrÞ
Y
þ1
rðrÞ  rðrÞ e rðrÞ  rðrÞ
Lj ðrðrÞ Þ ¼ L j ðrðrÞ Þ ¼
ðrÞ ðrÞ
i
and i
: ð28Þ
i¼1;i–j rj  ri i¼1;i–j rj  ri
ðrÞ ðrÞ ðrÞ ðrÞ

     
rðrÞ ¼ 1 and Lj ri ¼ 0 for i – j, and likewise for e
L j ðrðrÞ Þ, so that X j ¼ xðrÞ rj
ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ
Note that Lj j is exact at the LGR
 
points and the additional point rðrÞ
NðrÞ þ1
used in the state approximation, and likewise U j ¼ uðrÞ rj
ðrÞ ðrÞ
is exact at the LGR
points.
262 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275

Next, the derivative of the state (13) and cost function (23) are approximated. Let xðrÞ and DðrÞ be the weights and differ-
entiation matrix in phase r 2 ½1; . . . ; R corresponding to the choice of N ðrÞ . The collocated dynamics at the N ðrÞ LGR collocation
points are expressed as
ðrÞ
NX ðrÞ ðrÞ
þ1 sf  s0
f ðX i ; U i ; ri ; s0 ; sf Þ ¼ 0; Dij ¼ L_ j ðri Þ;
ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ
Dij X j  ði 2 ½1; . . . ; NðrÞ Þ: ð29Þ
j¼1
2

The cost function in phase r is

ðrÞ ðrÞ NðrÞ
sf  s0 X
i ‘ ðX i ; U i ; ri ; s0 ; sf Þ:
xðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ
J ðrÞ ¼ ðrÞ
ð30Þ
2 i¼1

Likewise, with approximations of the path constraints (15), boundary constraints (19), and linkage conditions (22) (Rao,
2003; Garg, 2011), the multiple-phase optimal control problem can be rewritten as
8
>
> X R
>
> Minimize J ¼ J ðrÞ
>
>
>
>
>
>
r¼1
>
> ðrÞ
NX
>
>
þ1 ðrÞ
s s
ðrÞ
>
< subject to Dij X j  f 2 0 f ðX i ; U i ; ri ; s0 ; sf Þ ¼ 0
ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ

j¼1 ð31Þ
>
>
>
> g min 6 g ðrÞ ðX i ; U i ; ri ; s0 ; sf Þ 6 g max
ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ ðrÞ
>
>
>
>
>
> ðrÞ ðrÞ ðrÞ
emin 6 eðrÞ ðX 1 ; s0 ; X NðrÞ þ1 ; sf Þ 6 emax
ðrÞ ðrÞ ðrÞ
>
>
>
>
>
: ðkÞ ðkÞ ðkþ1Þ ðkþ1Þ
lðX NðkÞ þ1 ; sf ; X 1 ; s0 Þ ¼ L;

where the constraints range in r 2 ½1; . . . ; R; i 2 ½1; . . . ; N ðrÞ , and k 2 ½1; . . . ; R  1.
In this way, the continuous optimal problem is transformed into a nonlinear programming problem. The resulting
nonlinear programming problem can be solved by nonlinear optimization algorithms (Gill et al., 2002). There are several
well-developed packages that implement the Pseudospectral method, in which GPOPS is a Matlab-based open source tool
that uses the Radau Pseudospectral Method to solve the multiple-phase optimal control problem (Darby et al., 2011; Rao
et al., 2010).

3. Train trajectory optimization in case of delays

Section 2 describes the basic modeling and solving methods for the train trajectory optimization problem following
planned timetables. This section discusses the train trajectory optimization in case of disturbances.

3.1. Problem description

In the presence of disturbances a rescheduling process may produce a new conflict-free timetable by changing the refer-
ence times and speeds for particular points or track sections, train routes, or even train sequences. The train path envelope
then must also be updated accordingly with the adjusted times and route characteristics. The central issue of the train tra-
jectory calculation is to respect and follow the new train path envelope. The detailed process of train trajectory calculation in
case of disturbances is as follows: The train path envelope for each train is first determined based on the new timetable with
the method from Section 2.2. New time and speed constraints of the train path envelope as well as train parameters, track
gradients, curves, speed limits and energy saving requirements are considered by the new trajectory calculation. The method
of calculating the new trajectory is the same as the one in Sections 2.3 and 2.4.
If rescheduling is not necessary, the delayed trains may have to speed up to get back to the timetable. Furthermore, it is
also important to reduce the impact of delays on other trains. The train path envelope for each train is first checked on
whether minimum running times are satisfied or not. When delayed trains do not have enough time to reach the next target
position, a free time supplement is added. The arrival and departure time constraints at stops change into flexible target win-
dows instead of mandatory target points because there is no guarantee that the delayed train can arrive at (depart from)
stops on time. However the train is not expected to arrive (depart) earlier than its scheduled time. So the flexible time win-
dow of the arrival (departure) event can be ½ta ; ta þ ts  (½t d ; t d þ t s ), where ta and t d are the scheduled arrival and departure
time, and ts is an extra time supplement. The lower bounds make sure that the trains do not arrive (depart) early, while the
upper bounds make sure that the trains have enough running time. Dwell times can be reduced to make up delay, but the
minimum dwell time at a stop should be respected. This minimum dwell constraint is included into the multiple phase opti-
mal control model, that is
   
ðkþ1Þ ðkÞ
x2 s 0  x2 sf P DsðkÞ ;min ; ð32Þ
f
 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275 263

where phase k and k þ 1 are two sequential phases where the connecting point is a stop point. DsðkÞ ;min is the minimum dwell
f

time there. (21) is adopted for non-delayed trains, and (32) for delayed trains.
The optimization objective of delayed trains is no longer just energy-efficient driving. Another important issue is to min-
imize delays to stimulate that the train paths will get back to the scheduled ones as soon as possible. So the cost function for
each phase is now given as
  Z ðrÞ
sf
ðrÞ ðrÞ ðrÞ
J ¼ w1  x2 sf þ u1 ðsðrÞ Þds; ð33Þ
ðrÞ
s0

ðrÞ ðrÞ
where x2 ðsf Þ is the time to be optimized at position sf and w1 is a nonnegative weight factor. The first item is to reduce the
arrival time at the terminal point of each phase. In turn, it is to minimize the arrival time at stations. On the other hand, a
time window constraint ½t a ; t a þ ts  is used to make sure that the train cannot arrival at stations earlier than the scheduled
arrival time. So the objective function makes the arrival time get close to the scheduled one, that in turn reduces train delays.
In all, the first item aims at reducing delays by minimizing the running times in each phase, while the second term aims at
saving energy consumption. The weight w1 reflects the trade-off between these two objectives. Different values of w1 result
in different optimization solutions. The effect of w1 is discussed by Wang and Goverde (2016). In this paper, we use w1 ¼ 103 .
Delay changes headway between successive trains and train path conflicts might occur if the scheduled buffer time
between train paths is not big enough. In practice this means that a train will meet yellow or even red signals, and has to
decelerate to a restricted speed and possibly wait in rear of a stop signal. Signalling influences are complicated and should
be taken into account the train trajectory calculation.

3.2. Signalling influences

The response to signals depends on the specific signalling system and characteristics of the ATP system, and can be quite
different. In this paper, we consider the Dutch signalling system NS’54 with the Dutch ATP system ATB (Automatische Trein
Beïnvloeding). NS’54 operates by wayside light signals that give speed commands to drivers such that trains can always brake
before a signal at danger (Goverde et al., 2013). Normally, NS’54 is a three-aspect two-block system with clear (Green, G),
approach (Yellow, Y) and stop (Red, R) aspects. In case of a green signal aspect, the train is unaffected by the signalling sys-
tem and can proceed with its normal operational speed. A yellow signal orders to reduce speed to a restricted speed of
40 km/h and prepare to stop before a red signal. A red signal is a stop order, which implies that the train should stop before
the signal. Near stations also short blocks are applied which have a length shorter than the maximum braking distance from
the line speed by which trains can follow at a shorter headway. In that case, NS’54 indicates already one (or more) signals
before that the train has to slow down to an indicated speed that must be reached before the next signal, so that the train will
enter the short block with a lower speed associated to the short block length. This progressive speed signalling is given by a
Yellow signal plus a white numeral indicating the permitted speed at the next signal. For instance, ‘Yellow 8’ (Y8) indicates
that the train has to reduce speed to 80 km/h before the next signal.
Moreover, Dutch trains are equipped with the ATB system, which supervises braking after an approach aspect as well as
certain ceiling speeds and intervenes with an emergency brake to standstill if the driver doesn’t obey the braking or super-
vised speed limits forced upon the train. So after a speed reduction order the driver has to apply the brakes until the per-
mitted speed is reached. In the case of a yellow signal the driver has to reduce speed to a restricted speed of 40 km/h
and then drive on-sight for the remainder of the block. If a driver does not brake sufficiently after a speed reduction order
then ATB warns the driver and if the driver still does not react ATB will intervene with an emergency brake to standstill. The
train is able to re-accelerate after an improved signal aspect ahead, otherwise it has to stop in rear of the signal and wait until
a new proceed aspect (green or yellow) is given.
Signalling influences cannot be ignored during train trajectory calculation. An important part is that information about
current and future signalling state is required. The train trajectory calculation strategies in case of two different scenario
assumption about available information are discussed below. The first scenario assumes that the driver only gets informa-
tion from the signal aspect at the beginning of a block. This scenario simulates the limited information that train drivers can
receive based on current commutation systems. A signal response policy then is developed considering train trajectory opti-
mization assuming only this limited signalling information. Another assumption is that the states of the upcoming signals
along the running track can be predicted. This assumption is based on advanced train control and commutation systems.
The future states of signals are predicted by the movements of previous trains. A green wave policy is developed for this case.
The following two sections give detailed illustrations of signal response policy and green wave policy.

3.3. Signal response policy

The signal response policy first calculates an optimal speed trajectory from the current location to the next stop before the
train leaves the station. Because of limited information about the signal aspects, the optimal trajectory is calculated with the
method in Section 2 if the train is not delayed or the method in Section 3.1 if the train is delayed. The trajectory is followed
until meeting a yellow or red signal aspect. A yellow or red signal means that the train should decelerate or stop, so the train
cannot just follow the old trajectory calculated. A new train trajectory is re-calculated for the remaining journey, which
264 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275

(x, u)
x(si+1)
x(si)
x(si+2)
x(si-1)

u*i-1 u*i u*i+1

*
u i-2

ti 2
ti ti 1

i 1

si-1 si si+1 si+2 s

Fig. 2. Illustration of the signal response policy with the calculated optimal controls at each step (black dotted lines) and the actual used control (red solid
line). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

should make proper responses to signal aspects. The train follows the new trajectory until meeting another yellow or red
signal aspect, or the first green signal aspect after yellow or red. Then a proper trajectory is calculated again. Hence, the sig-
nal response policy keeps detecting the signal state when the train reaches the sight distance of a signal. Once a yellow or red
signal aspect, or the first green signal aspect after yellow or red is detected, a new trajectory is required. Fig. 2 shows an
example of re-calculating a train trajectory after meeting a yellow signal:
At sight distance before the signal, information about the signal aspect ahead is available. Then a measurement of the
train speed and time is taken. Denote by si the sight distance before the signal, and xðsi Þ the state of the train speed and time
at si . Based on the measurement, an optimal trajectory from si to si;end is calculated, where si and si;end are respectively the
start and end points of the new trajectory. si;end depends on which kind of signal aspect is detected at si . Denote
ui ðsi ; si;end Þ as the optimal control trajectory calculated at si . Suppose that ui is available after time tðsi Þ þ Dt i , where Dt i is
a computational delay and tðsi Þ is the travelled time at si . Then the new trajectory ui is applied after tðsi Þ þ Dt i . When the
train gets close to the next signal siþ1 , the signal aspect ahead is available, and a new trajectory might be calculated if the
signal at siþ1 is yellow or red, or the first green signal aspect after yellow or red. To simplify the problem, a reasonable
assumption is made that the calculation time of the new trajectory is small enough so that the system can get a new trajec-
tory before the train passes the signal, i.e., within the sight and reaction time before the signal which is usually about 12 s.
The calculation of the optimal controller depends on the signal aspect. As an example of the section between si and siþ1 . If
the signal aspect is Yellow, then the train has to reduce speed to an approach speed (v approach ¼ 40 km/h in the Dutch rail-
ways). Braking has to start immediately after passing the yellow signal because of the ATB system. In view of this, the
new trajectory after the yellow signal is calculated as follows. The first step is to calculate the decelerating curve with max-
imum braking force until the restricted speed v approach is reached (Fig. 3(a)). If speed v ðsi Þ at si is lower than v approach , skip this
step (Fig. 3(b)). The yellow signal means that the next signal is red; therefore a stop curve is calculated after the decelerating
curve. The stop curve starts with the end point of the decelerating curve si;y (Fig. 3(a)) or si (Fig. 3(b)) and ends at the next
signal siþ1;signal (we assume siþ1 is the sight distance before signal siþ1;signal ). The terminal speed at siþ1;signal is zero. A new tra-
jectory is calculated with the multiple-phase optimal control model and pseudospectral method solver. v approach is taken as
the speed limit for all phases and (33) is chosen as the objective function since braking after yellow signals wastes time. In
summary, the controller ui ðsi ; si;end Þ consists of a deceleration stage (if any) and a stop stage, and si;end ¼ siþ1;signal .
In case of a Yellow 8 aspect, the train must reduce speed to v approach ¼ 80 km/h (and similar for other speed indications). If
speed v ðsi Þ at si is higher than 80 km/h then first a decelerating curve to the indicated speed must be computed with the

0 0
u* u*
v Bmax v v Bmax

vapproach

vapproach vapproach

si si , y sred s si sred s si s

Yellow Red Yellow Red Yellow 8 Yellow

(a) (b) (c)
Fig. 3. Illustrative speed-distance diagrams to optimization strategies in face of a yellow signal aspect. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)
 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275 265

Uto Bnk
Ut
Htn Htnc Cl
Utl
Gdm

Utzl Ddr

Utlr

Oss Ht
Gdm

Nm

Cl Zbm

Ehv
Tb

Fig. 4. The infrastructure between Utrecht and ’s Hertogenbosch.

maximum braking force as shown in Fig. 3(c). If v ðsi Þ is lower than 80 km/h then this step can be skipped. The second step is
to calculate the optimal curve over the remaining sections, from the end point of the decelerating curve (or the train’s cur-
rent position if there is none) to the next signal siþ1;signal . This curve assumes that the maximum operation speed over the
remaining section may not exceed 80 km/h, that is, the speed limit of each phase and point should be lower or equal to
80 km/h. With this new speed limit the optimal train trajectory is calculated using the Pseudospectral method.
When the signal aspect at si is Red, the stop curve calculated after the yellow signal aspect at si1 is followed until stand-
still. The train keeps waiting in rear of the signal until the signal aspect improves.
If the signal at si is Green, the train can proceed if the planned train trajectory covers train operation in the section
between si and siþ1 . The green signal may also appear after yellow or red signals. In that case, a new trajectory is calculated.
si;end is set as the closest next stop point. A multiple phase train trajectory model is generated with the method in Section 2.3
and solved with the Pseudospectral method with (33) as the objective function for each phase to minimize delay and energy
consumption.

3.4. Green wave policy

The signal response policy above has a proper response to the different signal aspects. But this also means that the train
has to decelerate to a lower speed when meeting yellow signal aspects and sometimes needs to re-accelerate when the sig-
nal aspect improves, energy-efficiency and riding comfort are not guaranteed. In practice, we like to avoid frequent decel-
eration and re-acceleration, and have the train operate in a more smooth way. The green wave policy is a train operation
strategy which avoids yellow signals and has the train operate under green signals for energy-efficient operation. The green
wave policy makes the trains only encounter green signals and ultimately the railway network achieves smoother operations
and as additional benefits, reduced energy consumption and reduced risk of signals passed at danger (Caimi et al., 2012;
Corman et al., 2009).
It is based on the assumption that the states of the upcoming signals along the running track can be predicted. Signalling
states include the changes of signal aspects and corresponding changing times. Take one signal ps along the train journey as
an example, and denote the predicted time that the signal aspect changes to green as T ps ;min . A green wave means that the
train will only pass through the signal with a green aspect. So T ps ;min is the earliest possible time for the train to pass through
signal ps . That is
tðps Þ P T ps ;min ; ð34Þ

where tðps Þ is the time for the train to pass through signal ps .
Every signal along the train journey has such a time constraint in the green wave policy. Those constraints are included in
the train path envelope. In other words, the train path envelope not only consists of the time and speed constraints at time-
table points but also the time constraints at signals.
For the train trajectory calculation with green wave policy, the train journey is partitioned into multiple phases by critical
points of speed limits or gradients and curves and target positions of the TPE, including signal positions. The target windows
(34) are the boundary conditions at the phases that start or end at a signal. (33) is adopted as the cost function for each phase
since the train is delayed. The optimal model is then solved using a Pseudospectral method.
266 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275

Table 1
Characteristics of Sprinter and Intercity train.

Characteristic Value
Sprinter Intercity
Train mass [t] 220 391
Rotating mass factor 1.06 1.06
Train length [m] 138 162
Maximum traction power [kW] 1918 2157
Maximum traction force [kN] 170 214
Maximum braking rate [m/s2] 0.8 0.66

170 KN Sprinter 214 KN Intercity

Force [KN]
Force [KN]

Speed [km/h] Speed [km/h]

Fig. 5. Traction force (solid line) and train resistance (dashed line) of Sprinter and Intercity.
time [min]

Ut Utl Htn Htnc Cl Gdm Zbm Ht

Distance

Fig. 6. Timetable for the corridor from Utrecht to ’s Hertogenbosch within a basic hour pattern.

4. Case studies

This section demonstrates the approach with some case studies. The optimization model and solution algorithm are
implemented in Matlab based on GPOPS 4.1 (Rao et al., 2010). The calculations were carried out on a laptop equipped with
a 3.2 GHz Pentium R processor. The Dutch corridor between Utrecht and ’s Hertogenbosch is adopted for the case studies,
which is a 50 km long double-track line with some multiple-track parts with traffic in both directions having their own
tracks (Fig. 4). Eight stations are located along this corridor: Utrecht (Ut), Utrecht Lunetten (Utl), Houten (Htn), Houten
Castellum (Htnc), Culemborg (Cl), Geldermalsen (Gdm), Zaltbommel (Zbm) and ’s Hertogenbosch (Ht). The infrastructure
characteristics consist of an accurate description of all track sections, points, speed signs, gradients, and signals over the
entire track layout from Utrecht until ’s Hertogenbosch.
We consider two different passenger train types running on the corridor in the direction from Utrecht to ’s Hertogen-
bosch: a local train (Sprinter) and an Intercity. The characteristics of the Sprinter and Intercity are shown in Table 1, includ-
ing train mass, rotating mass factor, train length, maximum traction force and power, and maximum braking rate. Since the
 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275 267

Table 2
Basic arrival and departure times (mm:ss) of the Sprinter and Intercity from Ut to Ht (repeating each 15 min).

Station Sprinter Intercity

Arrival time Departure time Arrival time Departure time
Ut - -:- - 00:00 - -:- - 12:00
Utl 04:00 05:00 - -:- - - -:- -
Htn 08:00 09:00 - -:- - - -:- -
Htnc 11:00 12:00 - -:- - - -:- -
Cl 17:00 18:00 - -:- - - -:- -
Gdm 23:00 29:00 - -:- - - -:- -
Zbm 34:00 35:00 - -:- - - -:- -
Ht 44:00 45:00 40:00 - -:- -
Speed [km/h]
Gradient [%o]

Distance [m]

Fig. 7. Illustration of route partitioning into multiple phases.

braking rate is the only accessible data characterizing the braking behavior, we set the braking force equal to the braking rate
times train mass. The assumed traction force and train resistance curves of both trains are shown in Fig. 5.
Fig. 6 shows the timetable for the direction from Utrecht to ’s Hertogenbosch within a basic hour. This timetable is based
on the practical timetable in use in 2015. Sprinter train stops at every station while the Intercity only stops in Utrecht and ’s
Hertogenbosch. Each 15 min a pair of a Sprinter and an Intercity departs from Utrecht, while the Sprinter is overtaken by the
Intercity at Gdm. The scheduled dwell times of the Sprinter at all stations is assumed 1 min except for the dwell time in Gdm
which is 6 min. The minimum dwell time for the Sprinter train in Gdm remains 6 min while the minimum dwell times in the
other stations are all assumed 0.5 min. Table 2 shows the arrival and departure times of a pair of Sprinter and Intercity (red
lines in Fig. 6) within the basic hour pattern.

4.1. Case A: Trajectory optimization for scheduled conditions

First case study is about train trajectory optimization for scheduled conditions. We discusses the train trajectory opti-
mization based on a conflict-free traffic plan without consideration of delays or disturbances. Assume that the Sprinter
and Intercity trains follow the timetable of Table 2, the focus of Case A is to calculate the optimal speed profiles for one Sprin-
ter and one Intercity train using the method from Section 2. The steps of building up train path envelope and multiple phase
optimal control model are introduced in detail.
The first step of the optimal trajectory calculation is to transform the timetable into a train path envelope. Table 2 shows
the Sprinter has 8 stops. Its train path envelope uses a series of target points indicating the departure and arrival times and
speed at stop points. TPEsprinter = {(244, 0, 0), (4574, 240, 0), (4574, 300, 0), (8193, 480, 0), (8193, 540, 0), (9934, 660, 0),
(9934, 720, 0), (18366, 1020, 0), (18366, 1080, 0), (26554, 1380, 0), (26554, 1740, 0), (34737, 2040, 0), (34737, 2100, 0),
(48707, 2640, 0), (48707, 2700, 0)}. Each target point constraint is a triple which consists of the stop position [m], the
scheduled departure or arrival time [s], and the speed target [m/s]. Speed target equals 0 everywhere because the Sprin-
ter stops at each station. For the Intercity, there is one departure event at Ut, one overtaking event at Gdm and one arri-
val event at Ht. So the train path envelope includes two target points and one flexible target window. Assuming the
minimum overtaking headway time at Gdm is 2 min, the time window of the Intercity to pass through Gdm is calculated
as ½1500; 1620 s based on the method presented in Section 2.2. TPEintercity = {(244, 720, 0), (26,554, [1500, 1620],
[0, 140]), (48,707, 2400, 0)}.
268 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275

Speed [km/h]
Ut Utl HtnHtnc Cl Gdm Zbm Ht
Distance [km]
Force [kN]

Ut Utl HtnHtnc Cl Gdm Zbm Ht

Distance [km]

Fig. 8. Optimal speed trajectories (top) and traction and braking force curves (bottom) of the Sprinter train.
Speed [km/h]

Ut Utl Htn Htnc Cl Gdm Zbm Ht

Distance [km]
Force [kN]

Ut Utl Htn Htnc Cl Gdm Zbm Ht

Distance [km]

Fig. 9. Optimal speed trajectories (top) and traction and braking force curves (bottom) of the Intercity train.

The second step is to build the multiple-phase train trajectory optimal control model. Both train trajectories are parti-
tioned into multiple phases by the critical points of speed limits or gradients, the signals, and the target positions of the train
path envelope. For example, Fig. 7 shows the partitioning of the Sprinter’s journey from Utrecht to Houten into 22 phases.
The lower black solid lines are the gradients, the red upper solid lines are the speed limits. The low speed limits before a
speed increase are extended by the train length (cyan speed lines) which models that the train will only accelerate after
its rear has safely passed low speed limit. The dotted lines with circles indicate the stop positions: Utrecht, Utrecht Lunetten
and Houten, which are the target points of the train path envelope. The dotted lines with crosses are the signal positions, and
the other dotted lines are the critical points of speed limits or gradients. Each segment between two dotted lines is a phase in
the multiple-phase optimal control model, with a unique speed limit and gradient. The entire corridor from Ut to Ht is par-
titioned into 125 phases for the Sprinter, while the Intercity route is partitioned into 90 phases.
The optimized results of state and control trajectories for the Sprinter and Intercity trains are shown in Figs. 8 and 9,
respectively. In each figure, the upper plot shows the optimized speed profiles (solid black lines) and the static speed limits
(red horizontal lines). The lower plot shows the optimized forces, where the solid lines refer to the traction force and the
dotted lines are the braking force curves. The speed profiles show that both train speeds stay below the speed limits and
the trains stop at the planned stop target positions, which means that safety and accurate stops are ensured. The control
regimes include using maximum traction during the outbound processes, maximum braking force for the inbound processes,
cruising at maximum speeds, and coasting before braking to save energy, which matches the theoretical Maximum Power,
Cruising, Coast and Maximum Brake optimal control regimes. Moreover, the speed profiles show that the trains are able to
regulate their speed at the varying speed limits via decelerating before low speed limits and accelerating before high speed
limits, where the trains only accelerate after the entire train passed the low speed limits. From the traction and braking force
curves, we can see that the maximum traction force is used during the outbound processes. The traction force reduces
 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275 269

Table 3
The results of the train trajectory optimization for the scheduled conflict-free case.

Train Section Running time [s] Energy consumption [J] Number of phases Computation time [s]
Sprinter Ut–Utl 240 7:0923  107 13 3.80
Utl–Htn 180 8:5819  107 9 2.20
Htn–Htnc 120 3:7637  107 4 1.01
Htnc–Cl 300 2:3219  108 18 6.29
Cl–Gdm 300 1:6357  108 22 7.75
Gdm–Zbm 300 2:2234  108 21 6.75
Zbm–Ht 540 2:9830  108 38 17.25
Intercity Ut–Ht 1860 1:2023  109 90 46.60
Speed [km/h]

Ut Utl HtnHtnc Cl Gdm Zbm Ht

Distance [km]
Speed [km/h]

Ut Utl Htn Htnc Cl Gdm Zbm Ht

Distance [km]
(a) Optimal speed trajectories of the Sprinter (top) and Intercity (bottom) using the signal response policy.
Speed [km/h]

Ut Utl HtnHtnc Cl Gdm Zbm Ht

Distance [km]
Speed [km/h]

Ut Utl Htn Htnc Cl Gdm Zbm Ht

Distance [km]
(b) Optimal speed trajectories of the Sprinter (top) and Intercity (bottom) using the green wave policy.

Fig. 10. Optimal speed trajectories of the Sprinter and Intercity in case of 210 s departure delay of the Sprinter from Utrecht.

gradually once the maximum power has been reached. In the cruising regimes the traction force constantly adjusts to main-
tain the optimal cruising speed, while it becomes zero during the coasting regimes so that no energy consumption is pro-
duced. The maximum value of the traction and braking force is kept within a range since the optimization algorithm also
takes riding comfort as a constraint.
Table 3 shows the results of the train trajectory optimization for the scheduled conflict-free case. The 1st column indi-
cates the train type, the 2nd column is the running section in the corridor. The 3rd column shows the computed running
times within that section, which are equal to the scheduled ones implying that the train trajectories ensure punctuality
as well. The 4th column gives the energy consumption. The 5th column is the number of phases in the multiple-phase opti-
mal control models and the 6th column gives the computation times. These last two columns indicate that the computation
270 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275

Speed [km/h]
Ut Utl HtnHtnc Cl Gdm Zbm Ht
Distance [km]
Speed [km/h]

Ut Utl Htn Htnc Cl Gdm Zbm Ht

Distance [km]
(a) Optimal speed trajectories of the Sprinter (top) and Intercity (bottom) using the signal response policy.
Speed [km/h]

Ut Utl HtnHtnc Cl Gdm Zbm Ht

Distance [km]
Speed [km/h]

Ut Utl Htn Htnc Cl Gdm Zbm Ht

Distance [km]
(b) Optimal speed trajectories of the Sprinter (top) and Intercity (bottom) using the green wave policy.

Fig. 11. Optimal speed trajectories of the Sprinter and Intercity in case of 270 s departure delay of the Sprinter from Utrecht.

time grows with the number of phases in this case. For the total 50 km train trajectory, the computation time is limited
within 1 min.

4.2. Case B: Trajectory optimization with signalling constraints

Section 3 shows the method of train trajectory calculation under disruptions and small disturbances. Rescheduling pro-
cess might produce new timetable in case of small disruptions. New generated timetable is conflict-free, which means sig-
nalling system wouldn’t influent train operation if the train is controlled according to the train path envelope based on new
timetable. So new train trajectory can be calculated with method in Section 2 without using signal response policy or green
wave policy. However in the presence of small delays, trains need to recover from delays and get back to original timetable
by regulating their own speed. Signalling influences are obvious since the headway between successive trains might become
small because of delays. Hence the focus of this case study is small delay situations and the interactions between trains and
signals. We assume that the Sprinter with scheduled departure from Ut at :00 is delayed at this station. Both the signal
response policy and green wave policy are used to calculate the train trajectories for the delayed Sprinter. The trajectories
between every two stops are calculated separately.
For the signal response policy, a trajectory is calculated each time before the Sprinter departs from a station. If the Sprin-
ter has delay, objective (33) is chosen to minimize the energy consumption as well as delay. Meanwhile, the arrival time
constraint is changed into a time window, where the lower bound is the scheduled arrival time and the upper bound is
the scheduled arrival time plus an extra time supplement. The objective (33) makes the Sprinter reach the next station as
soon as possible, while the new arrival time constraint provides enough running time and ensures that the train cannot
arrive earlier than the scheduled time. With the two cost terms, the new trajectory is able to recover the delay and have
 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275 271

Speed [km/h]

Ut Utl HtnHtnc Cl Gdm Zbm Ht

Distance [km]
Speed [km/h]

Ut Utl Htn Htnc Cl Gdm Zbm Ht

Distance [km]
(a) Optimal speed trajectories of the Sprinter (top) and Intercity (bottom) using the signal response policy.
Speed [km/h]

Ut Utl HtnHtnc Cl Gdm Zbm Ht

Distance [km]
Speed [km/h]

Ut Utl Htn Htnc Cl Gdm Zbm Ht

Distance [km]
(b) Optimal speed trajectories of the Sprinter (top) and Intercity (bottom) using the green wave policy.

Fig. 12. Optimal speed trajectories of the Sprinter and Intercity in case of 330 s departure delay of the Sprinter from Utrecht.

the train path return to the scheduled one. The minimum dwell times are used when the Sprinter train is delayed. When the
Sprinter has recovered from the delay, objective (23) is chosen to concentrate on energy saving only.
During the train run, the signal response policy checks every signal aspect at the sight distance before the signal. If the
Sprinter meets a yellow signal aspect, the trajectory is recalculated with the method from Section 3.3. The new trajectory
may include a decelerating curve in face of a yellow signal, a stop curve to stop the train before a red signal, and a re-
acceleration curve when the next signal aspect improves.
Green wave policy calculates the train trajectory every time before the Sprinter departs from a station. The states of the
upcoming signal aspects along the track are predicted with the train trajectory of the train running in front of the delayed
Sprinter. With the predicted signalling information, additional time constraints at the signal positions are added to make the
Sprinter run under all green signals. The optimal trajectories are computed using the Pseudospectral method with (33) as the
cost function for each phase.
Delayed Sprinter may cause a train path conflict to the Intercity that overtakes the Sprinter at Gdm, which means that the
Intercity might meet yellow signals during the section between Ut to Gdm. New Intercity’s trajectory is re-calculated by the
signal response policy and the green wave policy. For the signal response policy, a trajectory between Ut to Ht is calculated at
once before the Intercity departs from Ut with (23) as the objective because there is no delay of the Intercity at Ut. During the
train run, the signal response policy checks each signal aspect at sight distance before the signals. Proper responses are made
in face of different signal aspects with the signal response policy. For the green wave policy, the signal states over the track
from Ut to Gdm are predicted with the trajectory of the delayed Sprinter. The corresponding time constraints at signals are
added to ensure safe separations between the two trains. After station Gdm the Intercity train becomes the leading train of
the delayed Sprinter. The time constraints of the signals for the Intercity in the section between Gdm and Ht depend on a
previous Sprinter train in front of the Intercity instead of the delayed one. Moreover, the Sprinter’s delay may be propagated
to the Intercity at Gdm, which means that the overtaking time of the Intercity at Gdm is changed, since the overtaking time
272 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275

Time [min]

Time [min]

Time [min]
Distance [km] Distance [km] Distance [km]
Time [min]

Time [min]
Time [min]

Distance [km] Distance [km] Distance [km]

Fig. 13. The scheduled (gray) and optimized (red) train paths of the Sprinters and Intercities from Ut to Ht during an hour, including three initially delayed
Sprinters by 210 s (left), 270 s (middle) and 330 s (right). The upper and lower plots correspond to the signal response and green wave policy, respectively.
(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Table 4
Optimization results for various Sprinter delays from Utrecht (SR: signal response policy, GW: green wave policy).

Initial delay at Ut [s] Sprinter Intercity

Energy consumption Delays at Ht [s] Energy consumption Delays at Ht [s]
[ 109 J] [ 109 J]
SR GW SR GW SR GW SR GW
210 1:2045 1:2041 0 0 1:2023 1:4951 0 0
270 1:2052 1:2052 50.81 50.81 1:5938 1:4428 25.93 0
330 1:2052 1:2052 110.92 110.92 1:6329 1:3732 85.53 46.88

window for the green wave policy depends on the arrival and departure times of the delayed Sprinter. Furthermore, the arri-
val time constraint at Ht becomes a flexible target window constraint, using the scheduled arrival time as the lower bound
and the scheduled arrival time plus an extra time supplement as the upper bound. The arrival time window is to ensure a
sufficient running time to reach Ht. With the new time constraints for signals, overtaking, and arrival events, the green wave
policy calculates the Intercity’s optimal trajectory between Ut and Gdm with objective function (33) aiming at minimizing
delays and energy consumption.
Figs. 10–12 show the train trajectories of the Sprinter and Intercity trains in case of three different initial delays of the
Sprinter at Ut: 210 s, 270 s and 330 s. The red lines are the static speed limits and the black lines are the optimized speed
profiles. The trajectories are calculated with both the signal response policy (Figs. 10(a), 11(a) and 12(a)) and the green wave
policy (Figs. 10(b), 11(b) and 12(b)). Table 4 gives the resulting energy consumptions from Ut to Ht and the delays at Ht for
the two trains, the three different initial delays, and the two policies.
In the case of 210 s delay of the Sprinter from Ut (Fig. 10), the Sprinter applies less coasting but cruises more at the max-
imum speed in the early several sections. The delay decreases gradually by the increased average speed and reduced dwell
times. Compared with the energy consumption for the Sprinter travelling from Ut to Ht in Case A, which in total is
1:1108  109 J, more energy consumption is required to recover from the delay. The delay at station Ht is zero, which means
that the Sprinter has entirely recovered from the delay and is back on schedule. The Intercity’s speed profile using the signal
response policy is the same as the one shown in Fig. 9. The Intercity does not meet yellow signals in this case, so the speed
profile remains the same as the one planned before departing from Ut. However, the speed trajectory obtained with the
green wave policy is different. The train uses less coasting and thus consumes more energy compared to the result of the
signal response policy. This is because the objective function of the green wave policy is (33), which trades off minimizing
energy consumptions as well as delays, while the objective function of the signal response policy in the first stage is (23),
 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275 273

which aims at minimizing the energy consumption without consideration of delays. Moreover, the green wave policy adds
more time constraints at signal positions, which also changes the speed profile.
In the case of 270 s initial delay of the Sprinter from Ut (Fig. 11), more power is used by the Sprinter to improve its aver-
age speed and reduce running time, but still some delay remains at Ht station. The objective (33) is used both in the signal
response policy and green wave policy. As a result of the trade-off between minimizing energy consumption and delay still
coasting regimes are included in the Sprinter’s speed profile. The Intercity gets affected by the signalling system because the
headway with the Sprinter becomes small in the section between Ut and Gdm due to the Sprinter’s delay. Fig. 11(a) shows
that the Intercity train meets a Yellow 8 signal aspect (blue dashed lines) in the signal response policy, so that the train has to
decelerate to the approach speed (80 km/h) and proceed with a speed lower than that 80 km/h until the next signal aspect
improves. After this braking, the train re-accelerates to make up for the delay caused by the unscheduled braking but still has
some delay at Ht. Instead, the green wave policy makes the train coast to a lower speed before the critical signals, as shown
in Fig. 11(b). Hence, the green wave policy avoids the yellow signals, and in doing so avoids a delay in Ht and consumes less
energy consumption, see Table 4.
Fig. 12 shows the speed trajectories in the case that the Sprinter has 330 s delay from Ut. The optimized speed trajectories
and energy consumption for the Sprinter are the same as the results for 270 s initial delay. The speed profiles show that the
Sprinter applies full power for almost the whole corridor from Ut to Ht, with some minor coasting regimes. The delay is again
not fully recovered at Ht, and is about 60 s more than in the case of 270 s delay. With the signal response policy, the intercity
meets a yellow signal followed by a green signal afterwards (blue dashed lines in Fig. 12(a)). So the train decelerates to
40 km/h and then accelerates again. Moreover, the overtaking time is affected by the Sprinter’s delay at Gdm. The results
show that the Intercity uses coasting to consume time before passing through Gdm station for a smooth overtaking. The
speed profile with the green wave policy is different. It shows that the train uses coasting to avoid a yellow signal and
for smooth overtaking as well. This saves energy consumption, but results in a small delay in Ht, but less than with the signal
response policy.
The Sprinter’s trajectories with respect to the signal response policy and the green wave policy are more or less the same.
This is because the Sprinter is not affected by yellow or red signals. During the runs between Ut and Gdm, sufficient sepa-
ration between the Sprinter and the previous Intercity makes sure that there is no path conflict between the two trains. After
Gdm, the Sprinter train runs after the next Intercity, but the 2 min overtaking headway is sufficient for the Sprinter to again
meet no yellow and red signals after Gdm. Besides, the signal response policy and green wave policy use the same objective
function in case of delays. So the optimization results with the two policies are very similar. The results show that although
the delay is absorbed gradually, it is not entirely recovered and the bigger the initial delay, the more stations get a delay.
Furthermore, the delay can also be propagated to other trains. Figs. 11 and 12 show that the Intercity is delayed because
of unplanned braking caused by yellow signals and late overtaking. The energy consumption in case of delays is larger than
the one in case of the scheduled conditions.
The trajectories of the next and previous Sprinters and Intercities have also been calculated. Fig. 13 shows the time-
distance paths for an hour horizon, where the gray lines are the scheduled time-distance paths and the red lines are the opti-
mization results obtained using the signal response policy (top) and the green wave policy (bottom). From left to right, the
subfigures show the results in case of 210 s, 270 s, and 330 s departure delay of the Sprinter from Ut. The other train runs are
not affected by the delays except for the IC train overtaking the delayed Sprinter at Gdm. The distance-time paths show that
they track their scheduled paths well. So sufficient buffer time is available to prevent further delay propagation for these
initial delays up to 330 s. The delay of the Sprinter reduces gradually in each policy and delay scenario. Only small delays
can be found for the IC train that overtakes the delayed Sprinter.

5. Conclusions

This paper presented a model and algorithm for the train trajectory optimization problem. We formulated the real-time
traffic plan as a train path envelope. The train path envelope provides a convenient formulation of time and speed allowances
at timetable points and signals, witch can be used easily in the train trajectory calculation and updated with changes of the
timetable. Train trajectories based on the up-to-date train path envelope ensure that the train can track the new timetable
after the rescheduling process. The train trajectory optimization problem was formulated as a multiple-phase optimal con-
trol model, which provides an accurate train movement model with varying speed limits and gradients along the route as
well as time and/or speed constraints from the train path envelope. A Pseudospectral method was adopted for problem solv-
ing, while case studies show that the algorithm was able to find optimal or near optimal solutions.
In addition, the train trajectory calculation in case of delays was discussed. In the presence of disturbances, a rescheduling
process may adjust the timetable. As a consequence, the train path envelope needs to be updated and the train trajectory is
re-calculated based on the new train path envelope. Alternatively, trains may need to recover from delays and get back to the
timetable by regulating their own speed. The trains may meet yellow or red signals since the headway between successive
trains might become small because of the delays. Two optimization strategies were developed – the signal response policy
and the green wave policy – which take different responses to the signalling into account. The signal response policy is based
on a limited information scenario corresponding to only the signal aspect of the next signal. The idea is to ensure that the
train makes correct and quick responses to different signalling aspects. On the other hand, the green wave policy is based on
274 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275

a full information scenario corresponding to an accurate prediction of the leading train’s dynamic behavior. Predictions of
yellow or red signals can be used in target windows to avoid early arrivals to signals and thus have a train running according
to a green wave. Results show the benefit of the predictive information of the leading train on energy consumption and train
delay.
This work has been designed for real-time train trajectory calculation to support a driver advisory system. So smooth con-
trol profiles are next target of our optimization. The work is also useful for timetable optimization (Zhou et al., 2016) or traf-
fic flow and delay propagation analysis (Su et al., 2013) in railway networks. Future work will be devoted to improve the
computation times to seconds. Train trajectory optimization is an important module in developing driver advisory system,
which requires fast computation times. Currently this work is based on an existing pseudospectral tool. More work is needed
to further improve the solution time for real-time applications.

Acknowledgements

The authors thank the Dutch Infrastructure Manager ProRail for making available the infrastructure data. This work was
partially supported by the China Scholarship Council CSC (No. 201407000015).

References

Albrecht, A., Koelewijn, J., Pudney, P., 2011. Energy-efficient recovery of delays in a rail network. In: Proceedings of 2011 Australasian Transport Research
Forum. Adelaide, Australia.
Albrecht, A.R., Howlett, P.G., Pudney, P.J., Vu, X., Zhou, P., 2015a. Energy-efficient train control: the two-train separation problem on level track. J. Rail
Transport Plan. Manage. 5 (3), 163–182.
Albrecht, A.R., Howlett, P.G., Pudney, P.J., Vu, X., Zhou, P., 2015b. The key principles of optimal train control. Part 1: Formulation of the model, strategies of
optimal type, evolutionary lines, location of optimal switching points. Transport. Res. Part B: Methodol. (in press).
Albrecht, A.R., Howlett, P.G., Pudney, P.J., Vu, X., Zhou, P., 2015c. The key principles of optimal train control. Part 2: Existence of an optimal strategy, the local
energy minimization principle, uniqueness, computational techniques. Transport. Res. Part B: Methodol. (in press).
Albrecht, T., Binder, A., Gassel, C., 2013. Applications of real-time speed control in rail-bound public transportation systems. IET Intell. Transport Syst. 7 (3),
305–314.
Albrecht, T., Dasigi, M., 2014. On-time: a framework for integrated railway network operation management. In: Transport Research Arena, Paris, France.
Albrecht, T., Gassel, C., Binder, A., van Luipen, J., 2010. Dealing with operational constraints in energy efficient driving. In: IET Conference on Railway
Traction Systems. Birmingham, UK, 13–15 April 2010.
Benson, D., 2005. A gauss pseudospectral transcription for optimal control Ph.D. thesis. Massachusetts Institute of Technology, Cambridge, USA.
Betts, J.T., 1998. Survey of numerical methods for trajectory optimization. J. Guid. Control Dyn. 21 (2), 193–207.
Caimi, G., Fuchsberger, M., Laumanns, M., Lüthi, M., 2012. A model predictive control approach for discrete-time rescheduling in complex central railway
station areas. Comput. Oper. Res. 39 (11), 2578–2593.
Cheng, J., Howlett, P., 1992. Application of critical velocities to the minimisation of fuel consumption in the control of trains. Automatica 28 (1), 165–169.
Corman, F., D’Ariano, A., Pacciarelli, D., Pranzo, M., 2009. Evaluation of green wave policy in real-time railway traffic management. Transport. Res. Part C:
Emerg. Technol. 17 (6), 607–616.
Darby, C.L., Hager, W.W., Rao, A.V., 2011. An hp-adaptive pseudospectral method for solving optimal control problems. Opt. Control Appl. Meth. 32 (4), 476–
502.
Elnagar, G., Kazemi, M., Razzaghi, M., et al, 1995. The pseudospectral legendre method for discretizing optimal control problems. IEEE Trans. Autom. Control
40 (10), 1793–1796.
Garg, D., 2011. Advances in Global Pseudospectral Methods for Optimal Control Ph.D. thesis. University of Florida, Gainesville, USA.
Gill, P.E., Murray, W., Saunders, M.A., 2002. SNOPT: an SQP algorithm for large-scale constrained optimization. SIAM J. Optim. 12 (4), 979–1006.
Gong, Q., Ross, I.M., Kang, W., Fahroo, F., 2008. Connections between the covector mapping theorem and convergence of pseudospectral methods for optimal
control. Comput. Optim. Appl. 41 (3), 307–335.
Goverde, R.M.P., Corman, F., D’Ariano, A., 2013. Railway line capacity consumption of different railway signalling systems under scheduled and disturbed
conditions. J. Rail Transport Plan. Manage. 3 (3), 78–94.
Hansen, I.A., Pachl, J., 2014. Railway Timetabling and Operations. Eurail Press.
Howlett, P., 1996. Optimal strategies for the control of a train. Automatica 32 (4), 519–532.
Howlett, P., Pudney, P., 2012. Energy-Efficient Train Control. Springer.
Huntington, G.T., 2007. Advancement and Analysis of a Gauss Pseudospectral Transcription for Optimal Control Problems Ph.D. thesis. Massachusetts
Institute of Technology, Cambridge, USA.
Kang, M.H., 2011. A GA-based algorithm for creating an energy-optimum train speed trajectory. J. Int. Council Electr. Eng. 1 (2), 123–128.
Khmelnitsky, E., 2000. On an optimal control problem of train operation. IEEE Trans. Autom. Control 45 (7), 1257–1266.
Liu, R., Golovitcher, I.M., 2003. Energy-efficient operation of rail vehicles. Transport. Res. Part A: Policy Pract. 37 (10), 917–932.
Ma, J., Li, X., Zhou, F., Park, B.B., 2016. Parsimonious shooting heuristic for trajectory control of connected automated traffic. Part II: Computational issues
and optimization. Transport. Res. Part B: Methodol.
Milroy, I.P., 1980. Aspects of Automatic Train Control Ph.D. thesis. Loughborough University, Leicestershire, UK.
ON-TIME, 2014. Best Practice, Recommendations and Standardisation. Deliverable ONT-WP01-DEL-003.
Pudney, P., Howlett, P., 1994. Optimal driving strategies for a train journey with speed limits. J. Aust. Math. Soc. Ser. B. Appl. Math. 36 (01), 38–49.
Quaglietta, E., Pellegrini, P., Goverde, R.M.P., Albrecht, T., Jaekel, B., Marlière, G., Rodriguez, J., Dollevoet, T., Ambrogio, B., Carcasole, D., 2016. The ON-TIME real-
time railway traffic management framework: a proof-of-concept using a scalable standardised data communication architecture. Transport. Res. Part C:
Emerg. Technol. 63, 23–50.
Rao, A.V., 2003. Extension of a pseudospectral legendre method to non-sequential multiple-phase optimal control problems. In: AIAA Guidance, Navigation,
and Control Conference and Exhibit. Austin, USA, pp. 11–14.
Rao, A.V., 2009. A survey of numerical methods for optimal control. Adv. Astronaut. Sci. 135 (1), 497–528.
Rao, A.V., Benson, D.A., Darby, C., Patterson, M.A., Francolin, C., Sanders, I., Huntington, G.T., 2010. Algorithm 902: GPOPS, a matlab software for solving
multiple-phase optimal control problems using the gauss pseudospectral method. ACM Trans. Math. Software 37 (2), 22.
Ross, I.M., Fahroo, F., 2004. Pseudospectral knotting methods for solving nonsmooth optimal control problems. J. Guid. Control Dyn. 27 (3), 397–405.
Ross, I.M., Karpenko, M., 2012. A review of pseudospectral optimal control: from theory to flight. Ann. Rev. Control 36 (2), 182–197.
Su, S., Li, X., Tang, T., Gao, Z., 2013. A subway train timetable optimization approach based on energy-efficient operation strategy. IEEE Trans. Intell.
Transport. Syst. 14 (2), 883–893.
Wang, P., Goverde, R.M.P., 2016. Two-train trajectory optimization with a green wave policy. Transport. Res. Record, 2546.
 P. Wang, R.M.P. Goverde / Transportation Research Part C 69 (2016) 255–275 275

Wang, P., Goverde, R.M.P., Ma, L., 2015. A multiple-phase train trajectory optimization method under real-time rail traffic management. In: 2015 IEEE 18th
International Conference on Intelligent Transportation Systems. IEEE, Las Palmas, Spain, pp. 771–776.
Wang, Y., De Schutter, B., van den Boom, T.J., Ning, B., 2013. Optimal trajectory planning for trains – a pseudospectral method and a mixed integer linear
programming approach. Transport. Res. Part C: Emerg. Technol. 29, 97–114.
Wang, Y., De Schutter, B., van den Boom, T.J., Ning, B., 2014. Optimal trajectory planning for trains under fixed and moving signaling systems using mixed
integer linear programming. Control Eng. Pract. 22, 44–56.
Yasunobu, S., Miyamoto, S., 1985. Automatic train operation system by predictive fuzzy control. Ind. Appl. Fuzzy Control 1 (18), 1–18.
Zhou, F., Li, X., Ma, J., 2015. Parsimonious shooting heuristic for trajectory control of connected automated traffic. Part I: Theoretical analysis with
generalized time geography. Available from: arXiv:1511.04810.
Zhou, L., Zhou, X., Tang, J., Tong, L., 2016. Jointly optimize high-speed train route, timetable and trajectories: a unified modeling approach using space-time-
speed grid networks. Transport. Res. Part B: Methodol.