9 9

160 mm mm 21 mm mm 160 mm

The performance of many railway networks and the quality of

Hansen • Pachl
service offered is becoming more and more critical. The main
issues to be addressed are the increasing traffic volumes and making
the best use of the available capacity, at the same time resolving train
scheduling and management problems.

This is an updated, revised and extended edition of ‘Railway Timetable

& Traffic’, published in 2008. It describes the state-of-the-art methods
of railway timetabling and optimisation, capacity estimation, train Editors: Ingo Arne Hansen · Jörn Pachl
operations analysis and modelling, simulation, rescheduling and

Railway Timetabling

Railway Timetabling & Operations

performance assessment. The intention is to stimulate their broader
application in practice and to highlight current and future research

& Operations
areas. It is directed at academics, Masters and PhD students, as well
as professionals from the railway industry. It will also be of interest to
the public authorities that tender, monitor and perhaps fund railway
service provision. The overall aim is to improve the attractiveness and
efficiency of the train services which can be offered to the public. Analysis · Modelling · Optimisation
Simulation · Performance Evaluation
The key to achieving a higher efficiency and quality of train opera-

246 mm
246 mm

tions is an awareness of the impact of availability, reliability and

robustness of the subsystems on train processes. A deeper insight
into the probability of incidents and the propagation of train delays
depends on a thorough analysis of real world railway operations
and the feedback obtainable. This leads to an optimisation of the
timetable and a network-wide improvement in traffic management
performance.

This know-how should increase the efficiency of the railway system,

making it more attractive for regular, occasional and new customers,
and ensure that the railways continue to innovate. They will then be
able to make the maximum contribution possible to the transport
needs of the future.

iti en ed
ISBN 978-3-7771-0462-1

d
ed ext vis
on de
d re
an 2nd
 13460 12420
P-CST P-CST
13460
12420 67:44
04:32 10:36 16
06:20 12:50
41
D 27:26
41
13620
20 42:50
13691
09:42
P-CST P CST
P-
P-CST
3610 13620
20
16:32
13691 12510 20:04
05:02 11:08 P CS
P-CS
P- CST
03:40 0
08:58
3610 12570 12510 14590
09:53
P-CSTT
IVU.rail
12570
21:05
21 5
17:42
17 42
06:58
P-CST
15:13
14590
09:28 18:53 12:35 21:51
THEP-CST
RIGHT RESOURCE. 14420
P-CST
14750

:22 THE14 RIGHT

14:27
27
14420
PLACE.
23:23
23 23
3
13:23
12:46
14750
19:09

CST
13835 THE RIGHT
P-CST
P
16460
1
TIME. P-CST 13815 P-CST
16460
9:32 ONE
16:01 SOFTWARE.
13835 1381520:58
2 D
18:03 5:58
03:09 09:32 23:00 03:35 10:02 16:00 19:58
20:00
13650 13805 13710
0 14790
CST P-CST
13805
www.ivu.de/rail
P-CST
14790
13650 28:30
0 137100
03:11 08:58
8:58 19:00 13:28 17:06 22:5
10:01
13865 05:01 13540 10:01
-CST P-CST
13865 13540 P-CST
18:12 21:52
03:25 09:32 03:44 09:39
46:33 12760
P-CST 2
12670 D 12760 14400
12:16 9:4
T 12:16 P-CST 18:39
1413 12670 14415 13410 14770 14400
P-CST 26:04P-CST P-CST
7:31
1413
13:30 14415 13410
15:3414770
3590 06:27 14:16 22:51
13420
04:57 10:39 16:49 20:43
11553 16798
P-CST 13710 16460
3590 P-CST P-CST 13420 P-CST
18:0716798 18 57
18:57
11553 10:39 04:46 13710 10:05 16460
07:28 15:21 23:00 05:01 10:01 16:01 23:0
6 14400 12440
1673 16799 13610 11810
P-CST P-CST P-CST
P-CST P-CST P-CST
61673 1440016799 13610 12440 11810
09:3206:21 15:34 15:21 00:20
:
23:00 06:27
05:02 09:53 13:51 17:04 20:04
1493
3710 14770 16797 13420
P-CST P-CST P-CST P-CST13876 P-CST P-CST
1171
1493 16797 13420
3710 05:57
14770
15:50 22:45
13876
04:46 10:05 19:58
1171
10:01
11513 16:49 20:43 03:07
11690 09:32 15:46
14650
P-CST P-CST
12640 13835 16797
T11513 P-CST
11690 P-CST
14650
07:21
12640 19:58 05:25
13835 11:27 19:20
16797
1433 11510 14610
08:04 17:32 P-CST 03:09 P-CST
09:32 15:50
anzeige_eurail.indd 8 15.07.2014 14:33:10

Persönliches Exemplar für [email protected]

© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Railway Timetabling
& Operations
Analysis · Modelling · Optimisation
Simulation · Performance Evaluation

Editors
Prof. Dr.-Ing. Ingo A. Hansen, Delft University of Technology
Prof. Dr.-Ing. Jörn Pachl, Technical University of Braunschweig

Authors
Dr.-Ing. Thomas Albrecht, CSC Deutschland Solutions GmbH
Dr. rer. nat. Olaf Brünger, DB Systel GmbH
Dr. Gabrio Caimi, BLS AG
Dr. Francesco Corman, Delft University of Technology
Dr. rer. nat. Elias Dahlhaus, DB Systel GmbH
Dr. Andrea D’Ariano, Roma Tre University
Dr. Rob M. P. Goverde, Delft University of Technology
Prof. Dr.-Ing. Ingo A. Hansen, Delft University of Technology
Dr. Dennis Huisman, Erasmus University Rotterdam
Dr. Pavle Kecman, Delft University of Technology
Prof. Dr. Leo Kroon, Erasmus University Rotterdam
Dr. Gábor Maróti, VU University Amsterdam
Prof. Dr.-Ing. Ullrich Martin, University of Stuttgart
Dr.-Ing. Giorgio Medeossi, University of Trieste
Prof. Dr. Chris Nash, University of Leeds
Prof. Dr.-Ing. Nils Nießen, RWTH Aachen University
Prof. Dr. Dario Pacciarelli, Roma Tre University
Prof. Dr.-Ing. Jörn Pachl, Technical University of Braunschweig
Priv.-Doz. Dr.-Ing. habil. Alfons Radtke, Technical University of Braunschweig
Dr. Andrew Smith, University of Leeds
Dr. Robert Watson, Loughborough University
Dr. Phillip Wheat, University of Leeds
Dr. Jianxin Yuan, Arcadis Nederland BV

Persönliches Exemplar für [email protected]

© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
Bibliographic information published by the Deutsche Nationalbibliothek
The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data are available on the internet at http://dnb.dnb.de

Publisher DVV Media Group GmbH | Eurailpress

Nordkanalstr. 36
20097 Hamburg, Germany
Phone +49 (0)40 237 14 - 02
Fax +49 (0)40 237 14 - 236
E-mail: [email protected]
Internet: www.eurailpress.de, www.dvvmedia.com
Project Management Dr. Bettina Guiot (responsible), Anita Dahlinger
Proofreading John Glover
Advertisements Silke Härtel (responsible), Silvia Sander
Cover Design Andreas Gothsch
Frontispiece Image Source / Getty Images
Typesetting and Printing TZ-Verlag & Print GmbH, Roßdorf, Germany

© 2008, 2014 DVV Media Group

2nd revised and extended edition 2014
ISBN 978-3-7771-0462-1

This publication is protected by copyright. It may not be exploited, in whole or in part, without
the approval of the publisher. This applies in particular to any form of reproduction, translation,
microfilming, and incorporation and processing in electronic systems.

If some authors of photographs and illustrations were not determined despite careful research,
copyrights are nevertheless protected. Please notify the publisher if applicable.
A DVV Media Group publication

Persönliches Exemplar für [email protected]

© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Contents

Contents

1 Introduction ....................................................................9

2 Timetable Design Principles ........................................13

2.1 The Purpose of Scheduling .................................................................... 13
2.2 Basic Terms of Railway Operation ......................................................... 13
2.3 Diagramming Traffic ................................................................................ 18
2.4 Scheduled Running Time........................................................................ 20
2.5 Modelling Train Paths ............................................................................. 21
2.5.1 Principles of Train Separation ..................................................................................21
2.5.2 Application of the Blocking Time Model ..................................................................23
2.5.3 Modelling Specific Signalling Systems .....................................................................27
2.5.4 Modelling Interlockings and Overlaps ......................................................................30
2.5.5 Stochastic Blocking Times ......................................................................................32
2.5.6 An Alternative Approach: The Protected Zone Model ..............................................33

2.6 Headways and Buffer Times................................................................... 34

2.7 Consumed Capacity................................................................................ 38
2.8 Scheduling Methods ............................................................................... 40
2.8.1 Manual Scheduling .................................................................................................40
2.8.2 Computer-based Scheduling ..................................................................................42

2.9 Cyclic Timetables .................................................................................... 43

3 Infrastructure Modelling ...............................................47

3.1 Introduction ............................................................................................. 47
3.2 Graph Theory and its Application........................................................... 48
3.3 Macroscopic Models .............................................................................. 52
3.4 Microscopic Models................................................................................ 53
3.5 Differences between Infrastructure Models........................................... 56
3.6 Migration between Infrastructure Models .............................................. 57
3.7 Application of Infrastructure Models...................................................... 60
3.8 Outlook .................................................................................................... 62
3
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
Contents

4 Running Time Estimation .............................................65

4.1 Introduction ............................................................................................. 65
4.1.1 Fundamentals .........................................................................................................65
4.1.2 Speed Profile ..........................................................................................................66

4.2 Infrastructure and Train Data .................................................................. 68

4.2.1 Tractive Effort FT ......................................................................................................68
4.2.2 Vehicle resistances..................................................................................................70
4.2.3 Line Resistances FRl ................................................................................................71
4.2.4 Combining Efforts and Resistances, Rotating Masses .............................................71

4.3 Moving Sections...................................................................................... 72

4.3.1 Characteristic Sections ...........................................................................................73
4.3.2 Behaviour Sections .................................................................................................73

4.4 Modelling Running Time Calculations.................................................... 74

4.4.1 Determination of Braking and Acceleration Sections ...............................................75
4.4.2 Modelling the Train as a Mass Point ........................................................................77
4.4.3 Modelling the Train as a Homogeneous Strip ..........................................................78

4.5 Solving the Differential Equations by a Difference Equation Approach .. 79

4.6 Solving the Differential Equations Analytically ...................................... 81
4.6.1 Traction Effort Depends Piecewise Parabolically on Speed ......................................82
4.6.2 Traction Depends Hyperbolically on Speed ............................................................83
4.6.3 Passing a Section ...................................................................................................84

4.7 Solving the Differential Equations by Gauss Quadrature Integral

Formulae ....................................................................................................... 86
4.8 A Probabilistic Approach ........................................................................ 87
4.9 Calculating Blocking Times .................................................................... 87

5 Energy-Efficient Railway Operation ............................91

5.1 Minimisation of Mechanical Energy Consumption ................................ 91
5.1.1 Energy-Efficient Driving in Undisturbed Operation ...................................................91
5.1.2 Energy-Efficient Driving in Disturbed Operation .......................................................97
5.1.3 Operational Requirements for Energy-Efficient Driving .............................................97

5.2 Optimisation under Practical Billing Systems...................................... 100

5.2.1 Energy Supply and Billing......................................................................................100
5.2.2 Optimisation under Train-Based Billing Systems ....................................................102
5.2.3 Optimisation under Substation-Based Billing Systems ..........................................104

4
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Contents

5.3 Measures to Introduce Energy-Efficient Driving .................................. 105

5.3.1 Driver Training .......................................................................................................105
5.3.2 Driver Support Systems ........................................................................................106
5.3.3 Automatic Train Operation.....................................................................................114
5.3.4 Effects of Energy-Efficient Operation .....................................................................114
5.4 Conclusions ........................................................................................... 114

6 Queueing ....................................................................117
6.1 Introduction and motivation ................................................................. 117
6.2 Scheduled waiting times....................................................................... 119
6.2.1 Scheduled waiting times on railway lines ...............................................................119
6.2.2 Scheduled waiting times at railway nodes .............................................................124
6.2.3 Set of tracks as multiple server queues .................................................................126
6.3 Estimating knock-on delays ................................................................. 127
6.4 Summary and future prospects ............................................................ 130

7 Timetable Stability Analysis .......................................133

7.1 Introduction ........................................................................................... 133
7.2 Stability .................................................................................................. 135
7.3 Max-Plus Algebra .................................................................................. 136
7.4 Max-Plus Modelling .............................................................................. 138
7.4.1 The Higher-Order Max-Plus Model ........................................................................138
7.4.2 The First-Order Max-Plus Model ...........................................................................140
7.5 Critical Circuit Analysis ......................................................................... 142
7.6 Recovery Time Analysis ........................................................................ 144
7.7 Delay Propagation................................................................................. 146
7.8 Stochastic systems ............................................................................... 148
7.9 Conclusions ........................................................................................... 151

8 Optimisation Models for Railway Timetabling ..........155

8.1 Introduction ........................................................................................... 155
8.2 Cyclic Timetabling................................................................................. 157
8.2.1 A mathematical formulation...................................................................................157
8.2.2 Example ...............................................................................................................159
8.2.3 Objectives .............................................................................................................161
8.2.4 Solving PESP ........................................................................................................161

5
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
Contents

8.3 Non-cyclic Timetabling ......................................................................... 162

8.3.1 Graph representation of the timetable ...................................................................162
8.3.2 Arc based timetabling model.................................................................................163
8.3.3 Path based timetabling model...............................................................................164

8.4 Robust timetabling ................................................................................ 165

8.4.1 Notation ................................................................................................................166
8.4.2 Timetabling part of the model ...............................................................................167
8.4.3 Simulation part of the model .................................................................................168
8.4.4 Results and timetable experiment in practise ........................................................169

8.5 Conclusion............................................................................................. 170

9 Bi-level scheduling in highly utilised networks .........175

9.1 Introduction ........................................................................................... 175
9.2 The bi-level approach ........................................................................... 175
9.3 Macro scheduling.................................................................................. 177
9.3.1 The Periodic Event Scheduling Problem ................................................................178
9.3.2 Flexible PESP .......................................................................................................178

9.4 Micro scheduling in condensation zones ............................................ 180

9.5 Micro scheduling in compensation zones ........................................... 184
9.6 Computational results ........................................................................... 184
9.7 Conclusion and outlook ........................................................................ 188

10 Simulation ...................................................................191
10.1 Introduction – the role of simulation..................................................... 191
10.2 Methods of simulation .......................................................................... 193
10.3 Research and development .................................................................. 195
10.4 Commercial software ............................................................................ 196
10.5 Simulation in practice ........................................................................... 199
10.6 Limitations and areas for development of microscopic simulation .... 209
10.7 Conclusions ........................................................................................... 213

11 Statistical Analysis of Train Delays and Movements 217

11.1 Introduction ........................................................................................... 217
11.2 Train delays, running times and dwell times ........................................ 218
6
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Contents

11.3 Data collection ...................................................................................... 219

11.3.1 Train Describer Data .............................................................................................219
11.3.2 GPS and Train Event Recorder data ......................................................................221

11.4 Distribution fitting of train delays ......................................................... 222

11.4.1 Selection of statistical distributions ........................................................................223
11.4.2 Common approaches to parameter estimation .....................................................224
11.4.3 Testing the goodness-of-fit....................................................................................224
11.4.4 Assessing distribution models by fine-tuning the parameters ................................226

11.5 Estimation of stochastic train running times ....................................... 229

11.5.1 Calibration using on-board collected data and performance factors ......................230
11.5.2 Calibration results ................................................................................................231

11.6 Conclusions ........................................................................................... 234

12 Train delay prediction .................................................237

12.1 Introduction ........................................................................................... 237
12.2 Monitoring ............................................................................................. 239
12.3 Traffic State Prediction and Conflict Detection ................................... 240
12.4 Data-Driven Approach to Monitoring and Traffic State Prediction ..... 241
12.4.1 Methodological Framework ...................................................................................241
12.4.2 Process Mining of Train Describer Event Data .......................................................243
12.4.3 Robust Process Time Estimates ...........................................................................243
12.4.4 Online Prediction of Train Event Times ..................................................................247
12.4.5 Model Performance ..............................................................................................248

12.5 Conclusions ........................................................................................... 251

13 Rescheduling ..............................................................253
13.1 Introduction ........................................................................................... 253
13.1.1 Background ..........................................................................................................253
13.1.2 Problem description ..............................................................................................254
13.1.3 Current dispatching practice .................................................................................256
13.1.4 Real-time railway traffic management ....................................................................256

13.2 Review of the related literature............................................................. 257

13.2.1 Centralised approaches ........................................................................................257
13.2.2 Co-ordinated approaches .....................................................................................258
13.2.3 Discussion ............................................................................................................259

7
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
Contents

13.3 A real-time train traffic management approach ................................... 260

13.3.1 Load information ...................................................................................................261
13.3.2 Disruption recovery ...............................................................................................262
13.3.3 Real-time traffic optimisation .................................................................................262
13.3.4 Train speed co-ordination .....................................................................................264

13.4 Traffic optimisation model .................................................................... 265

13.4.1 Alternative graph model ........................................................................................265
13.4.2 CDRFR problem formulation .................................................................................266

13.5 Conclusions ........................................................................................... 271

14 Performance Evaluation.............................................275
14.1 Aims of Performance Evaluation .......................................................... 275
14.2 Methods of Performance Evaluation.................................................... 276
14.2.1 General Remarks ..................................................................................................276
14.2.2 Applying Effectiveness and Efficiency as Indicators ...............................................277
14.2.3 Assessing Punctuality ...........................................................................................278
14.2.4 Use of Capacity Research.....................................................................................281
14.2.5 Performance-dependent Funding of Infrastructure ................................................287

14.3 Conclusion and Further Development ................................................. 290

15 Economic performance assessment in railways ......295

15.1 Introduction ........................................................................................... 295
15.2 Background ........................................................................................... 295
15.3 Outputs and inputs of rail systems ...................................................... 296
15.4 Productivity analysis ............................................................................. 297
15.5 Measuring technical efficiency ............................................................. 298
15.6 Measuring Cost Efficiency .................................................................... 299
15.7 An example ............................................................................................ 302
15.8 Conclusions ........................................................................................... 303

16 Conclusions ................................................................307
Glossary ................................................................................................. 311
Keywords ............................................................................................... 321
The Authors ........................................................................................... 325
8
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 1 Introduction

1 Introduction
Ingo Arne Hansen

Timetabling, capacity analysis, traffic control and performance evaluation are basic elements
of railway transport management, which are inherently linked. The function of the timetable
in practice, however, is often not closely connected with its execution. Customers, train
drivers, conductors, and dispatchers depend on specific sources of information, interpret
the timetable and perceive the quality of train service differently. Further, the feasibility of train
schedules and operations is subordinated to constraints of the infrastructure, power supply,
signalling and safety systems, as well as rolling stock. For trip planning and decision making,
railway passengers use live travel information increasingly and book via the internet rather than
purchasing tickets at railway stations.
In many railways, however, train drivers and conductors still receive only a printed daily
schedule. This will indicate the relevant departure, passing and arrival times, as well as the
times and locations for the start, change and end of their duties, and a train radio or mobile
phone for communication with the traffic control centres. Dynamic train delay information
in the driver’s cab has, so far, been introduced only by a limited number of train operators,
while actual travel information in passenger coaches and at station platforms is generally
made available. During the journey, drivers must observe the signals and cab displays,
which indicate the necessary safety information with respect to their movement authority,
permitted maximum speed, and in some modern trains the critical target distance. However,
continuously updated actual delay information for train drivers is still mostly missing. A more
precise and actual traffic information feedback to drivers and conductors would help them to
better recognise deviations from schedules, to anticipate conflicts, and to raise operational
performance.
The information made available to signallers, dispatchers and network supervisors in traffic
control centres has improved significantly. This has been due to the introduction of electronic
interlocking, associated with computer screen monitoring of the actual set-up of routes,
track occupancy and sectional clearances for each train. Further, time-distance diagrams of
the daily scheduled train paths and those actually used have become available at the newer
traffic control centres. These are equipped with modern information, telecommunications and
computer technology.
Advanced tools for timetable design, simulation and construction have been implemented in
the past decades. These enable a much higher quality of timetabling and contribute to more
reliable operations. However, the reliability and train punctuality levels achieved by railways, in
general, do not give evidence of higher performance − except from Japan. It is hypothesised
that the disappointing train punctuality performance of many networks stems only partly
from the high traffic volumes and capacity usage. The main cause is an insufficient degree of
optimisation, accuracy and robustness of the current practice of timetabling, which misses the
feedback of detailed operational performance data.
This book intends to narrow the gap between the theory and practice of railway timetabling,
capacity analysis, traffic management and performance evaluation. It does so by compiling
the current state-of-the-art in this field and indicating a number of remaining problems and
research issues. It is directed at academics, Masters and PhD students, professionals from
the railway industry, and also public authorities that tender, contract and supervise railway
services. Its aim is to enable readers to better understand the underlying theoretical models
and to become acquainted with the actual state of technology.

9
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
1 Introduction

The authors of the individual chapters are mostly members of the International Association
of Railway Operations Research (IAROR)1, founded in 2004, and were selected on the basis
of expertise in specific areas. All are involved currently in railway research projects and have
published individually in journals or international conference proceedings. The book forms a
compendium of the main methodological aspects of railway timetabling, capacity and stability
analysis, scheduling, simulation, real-time traffic management and performance evaluation,
including some examples of calculation and practical application.
This book is an updated, partly revised and extended version of the previous edition Railway
Timetable & Traffic, published in 2008. Completely new in this edition are Chapter 9 on macro-
micro conflict-free scheduling in highly utilised networks, Chapter 10 Simulation, Chapter 12
Train Delay Prediction, Chapter 13 Rescheduling, and Chapter 15 Economic Performance
Assessment.
The book is structured around the elementary steps of railway timetable design, modelling of
infrastructure and train operations, and the analysis, optimisation, simulation, and evaluation of
timetables and operations respectively. Pachl outlines in Chapter 2 the principles and variables
of timetable design, and explains the principal methods for scheduling, modelling of train
paths, train separation, specific signalling systems, as well as estimation of capacity use. In the
following Chapter 3, Radtke deals with the models for description of the infrastructure based
on graph theory, while showing the differences between node and link oriented, as well as
macroscopic and microscopic models.
Chapter 4 contains a detailed explanation by Brünger & Dahlhaus of the calculation of train
running times, depending on infrastructure and rolling stock characteristics, followed by a
discussion of standard and advanced deterministic and stochastic calculation methods.
Albrecht continues in Chapter 5 by describing the practice of energy supply and billing, and
then explains the fundamentals and an intelligent tool for energy-efficient train control and
driving, depending on the actual state of operations.
The following two Chapters focus on mathematical models for the estimation of waiting
times, train delays, and delay propagation in networks. Nießen explains in Chapter 6 the
fundamentals of queuing theory and its application, while Goverde presents in Chapter 7 a
Petri net approach by means of the so-called max-plus algebra technique for stability analysis
and delay propagation in large network timetables.
The next two Chapters are focusing on current optimisation methods for the design of large
network timetables. Kroon, Huisman & Maróti describe in Chapter 8 linear programming
approaches for optimal train scheduling in cyclic and acyclic network timetables. In Chapter 9
Caimi explains the bi-level approach applied in Switzerland for generating optimal conflict-
free schedules in highly utilised networks through macro- and micro scheduling, as well as
distinction of condensation and compensation zones.
Watson & Medeossi describe in Chapter 10 the role of simulation, discuss the key
differentiators of simulation methods and give a comprehensive overview on the development
and application of current simulation models for timetable design and train operations.
In Chapter 11, Yuan & Medeossi present suitable approaches for the detailed statistical
analysis of train delays and validation of running and dwell times based on standard track
occupation and release data, including goodness-of-fit test and estimates for the distributions
and its parameters.

1 www.iaror.org

10
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 1 Introduction

Kecman & Goverde introduce in Chapter 12 a recently developed automatic data-driven

traffic state and route conflict prediction model based on train describer, track occupation
and clearance records. Chapter 13 consists of a review of the few existing models for the
rescheduling of trains in case of service irregularities and disruptions. Pacciarelli, D’Ariano &
Corman distinguish centralised from co-ordinated approaches and describe in detail the
architecture, main modules and procedures of their real-time rescheduling model Railway
traffic Optimisation by Means of Alternative graphs (ROMA).
In Chapter 14 Martin discusses the methods and key indicators for performance evaluation
of railway operations, capacity use of networks and standard cost-benefit assessment of
infrastructure investments. Wheat, Nash & Smith describe in the following Chapter 15 the
inputs and outputs for assessing the economic performance of railway systems through
productivity analysis, measurement of technical efficiency and cost efficiency.
Finally, the Editors of this book, Hansen and Pachl, derive the main conclusions concerning
the current state-of-knowledge on railway timetabling and traffic management from the
preceding Chapters. They then summarise the challenges and main topics for future research.
Their aim is to encourage readers of the book to contribute further to a better comprehension
of railway processes and to improve the attractiveness and efficiency of train services for
railway customers.
The approaches discussed cover the main current state-of-the-art theories and analytical,
combinatorial optimisation, simulation and economic models for railway timetabling and
operations world-wide. This book is a German-Dutch-British-Italian co-production, where
the academic and professional quality of railway operations research and education, and the
application of railway technology, are amongst the best in the world. The Editors hope that
researchers and professionals from any country will enjoy the content and use it for improving
the analysis, modelling and performance evaluation of railway timetabling and operations on
their own.

11
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.2 Basic Terms of Railway Operation

2 Timetable Design Principles

Jörn Pachl

2.1 The Purpose of Scheduling

Scheduling performs the following functions:

– It co-ordinates the train paths in the planning process for optimum use of the infrastructure.
– It ensures the predictability of train traffic.
– It produces timetable data for passenger information.
– It is essential for traffic control, locomotive and rolling stock usage and crew scheduling.
In passenger operations, scheduling is essential to provide predictable travel times for
passengers. The safe operation of trains also requires the co-ordination of slots. In freight
operations, the situation is different. Due to changing demands from the shippers, establishing
long term timetables is often not possible. Some freight operators abolished timetables
completely and run their trains entirely on-demand with priority-based train dispatching. Train
crews just get a “timetable” that does not contain any times, but only the data required for safe
train driving (e. g. speeds, braking conditions). This is typical for North American railways but
also for a number of mining railways in other parts of the world. European freight operators
are also facing changing demands. In the past, this problem was solved by establishing on-
demand paths in the yearly timetable. These paths could be used to run scheduled freight
trains on a daily basis.
Today, computer-based scheduling allows the principle of pre-constructed on-demand paths
to be replaced by running an increasing number of freight trains as extra trains. However,
running “extra” does not mean running “unscheduled”. An extra train has to be scheduled
as accurately as a regular train, but on a more dynamic basis. On many railways, a freight
train operating company can order a train path for an extra train just a few hours in advance.
Compared with completely unscheduled operation, this kind of flexible scheduling will lead to
a much higher degree of predictability of traffic. On lines with a mixed traffic of freight and
passenger trains, it will also ensure the required quality of service.
On railway networks that are operated on an open access basis, scheduling is not just a
planning procedure but also the commercial interface between infrastructure operators
and train operating companies. There, the scheduled train path is the product sold by the
infrastructure operator to the train operating company concerned. Scheduling means to
coordinate the train paths ordered by competing train operating companies. Assigning a train
path to a train operating company means to sell the right to run a train on that path under
specified operating conditions.

2.2 Basic Terms of Railway Operation

For the explanation of scheduling aspects, an understanding of the basic terms of railway
operation is required. Unfortunately, railways worldwide use very different terminologies which
may cause misunderstandings in international use. There is especially a big difference between
British and North American railway terms. This textbook is based mainly on the terms used
in UIC Code 406 (UIC, 2004) which is neither pure British nor pure American. However, at
some points, significant differences between the terms of UIC Code 406 and the terms used
in practical operation of British and North American railways are mentioned. For more detailed
explanation of the fundamentals of railway operation see Pachl (2009).
13
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

Organisations
The effect of EU Directives has been to separate the operational and infrastructure sides of the
railway businesses, in the EU but also in some other countries. This does not, of itself, require
separation of ownership. It has, however, introduced the concept of a track authority with a
network management, which grants access to operators of passenger and/or freight traffic
on payment of a fee. With rules which may be determined differently in the various countries
concerned, it has to grant track access as may be demanded by qualified and licenced
operators, to the extent that this is practical. Hence, efficient timetabling has become of even
more major importance in recent years.
The term open access needs to be treated with caution, since it also requires suitable train
paths to be available. In time, that may result in consequential changes to the timetable
already in place and hence to the services of existing operators. That in turn may have an
effect on the contractual arrangements between those operators and the track authority.
Alternatively, should conditions so dictate, the request may have to be refused.
Train operators may be classified generally as Train Operating Companies (TOCs), though it
should be noted that in Britain this term does not include freight operators.

Tracks, Routes and Stations

Tracks are the routeways of a railway system. A railway network in UIC terms consists
of nodes connected by links. In UIC Code 406 the links between nodes are called lines.
Consecutive lines and nodes as a whole, between a defined source and target, are called a
route (in a second meaning, the term route is also used for a route through an interlocking or
point zone). In its most simple form, a node is just a junction where trains can transfer from
one line to another. More complex nodes are extended station areas in which overtaking,
crossing, or direction reversal of trains is possible. A terminal is an assemblage of facilities
provided at a terminus or intermediate point of a route for the purpose of shunting,
assembling, sorting, and marshalling trains. A typical freight terminal consists of track groups
for arrival and departure of trains, a shunting district with yard facilities for sorting and storing
of wagons, and connecting tracks to industrial sidings. A typical passenger terminal consists
of a group of platform tracks and some yard facilities for storing passenger vehicles. In large
terminals, there are also often servicing facilities for the rolling stock. Large terminals are often
located at the intersection or junction of different routes.
In Britain, the term track is mainly used when talking about the track itself, or the permanent
way. In a station, the term “platform” usually includes reference to the track on which the train
runs at that point. Thus a train is said to be standing at platform 2, not track 2.
In railway operation, tracks are often referred to as lines. For operational purposes, tracks
are divided into two classes. Running lines are the tracks on which trains move through the
network. Running lines must be equipped with signalling appliances required for regular train
movements. A loop is a running line that is used for passing and overtaking trains. Sidings
are lines used for assembling trains, storing vehicles and trains, loading and unloading and
similar purposes, but not for regular train movements. A track layout consisting of several
sidings is called a yard. In North America, a line is not just a track but may consist of several
parallel tracks. Running lines are called main tracks, and loops are called sidings. In a literal
translation, the term main track is also used by many railways worldwide, e. g. in Continental
Europe. Since the British use of the term line differs from UIC Code 406 and may lead to
misunderstandings, the term main track is preferred against the term running line in the
sections which follow.

14
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.2 Basic Terms of Railway Operation

In UIC Code 406, the term station is only used for the points of a network where overtaking,
crossing, or direction reversal is possible. Many railways refer to all places designated as
stopping places by name in the timetable as stations. A passenger station is that with one or
more platforms for scheduled stops of passenger trains. It is not necessarily associated with a
station in the sense of UIC Code 406.
References made to interchanges between services at stations may, according to context,
include those to and from other modes, notably bus. Again, contractual elements may have a
bearing on how problems such as late running by any party should be resolved.

Track elements
A turnout is an assembly of rails, movable points, and a frog, which effect the tangential
branching of tracks and allows trains or vehicles to run over one track or another. The term
turnout is mostly used in civil engineering. In railway operation and signalling, a turnout is
usually referred to as a pair of points (points in short), although this term in its original meaning
only applies to the part of a turnout where the points are located. On North American railways,
this part of a turnout is called a switch, and the point machine is called a switch machine. In
North American railway operation and signalling, turnouts are usually referred to as switches.
A crossing is an assembly of rails that enables two tracks to cross at grade. The inner part of
a crossing is called a “diamond”. Crossings with a large angle of intersection are constructed
rigidly. In case of a small angle of intersection, a crossing may have movable points instead
of frogs. Small angle crossings may be equipped with additional points providing a slip
connection to permit movements from one track to another. A crossing with a slip connection
at one side is called a “single slip”, and a crossing with slip connections at both sides is called
a “double slip”. On North American railways, slip crossings are called slip switches.
Derailers are trackside devices that are used to protect train movements against unattended
movement of vehicles on converging tracks. An unsafe movement is derailed before it could
join the protected route. In the protecting position, a derailing piece is raised over one rail.
On many railways, derailers must not be installed in main tracks. Instead of derailers, some
railways also use trap points which have much the same effect.

Running and Shunting Movements

Trains consist of a minimum of a locomotive on its own but more generally with coaching stock
for passengers, or freight wagons for goods. Other trains may consist of electric or diesel
multiple units, and specialist works trains for engineering purposes. The term wagon in this
book may or may not include reference to passenger vehicles.
Running movements are locomotives with or without wagons, with authority to operate on a
main track under the control of main signals. The operating conditions for a running movement
(maximum speed, braking conditions) are specified in the working timetable. On lines with
non signal-controlled operation, a running movement needs a movement authority from the
operator responsible for controlling traffic on main tracks. Every running movement displays
rear end markers (tail lights in Britain) to enable the lineside staff to check the train’s integrity.
On European railways, running movements are always scheduled movements.
Shunting movements are those for making up trains, moving vehicles from one track to
another, and similar purposes. Shunting movements are accomplished under simplified
requirements at a very low speed that allows stopping short of any vehicle or obstruction.
Shunting units may enter occupied tracks. Movements in industrial sidings are also carried
out as shunting movements. Shunting movements are not scheduled when establishing the
timetable, but the need for them and the time necessarily taken must be considered.

15
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

Train runs is not a term normally found in Britain; used here, it refers to the complete end-to-
end journey undertaken by the train(s) concerned.
In North America, the sharp distinction between running movements and shunting movements
does not exist. All movements with authority to run on a main track are called trains. For more
details see Pachl (2009) and White (2003).

Signals and Interlockings

Lineside signals indicate if a movement may enter the section of track beyond the signal.
There are different signals for running and shunting movements. A main signal indicates if a
running movement may enter a section of line. In a territory with a fixed block system (see
Section 2.5.1), lines are divided into block sections for the purpose of safe train separation. A
train must not enter a block section until that section has been cleared by the train ahead. In
fixed block territory, running movements are authorised by signal indications. Except on lines
with a low speed, a signal that authorises a running movement requires an approach aspect
at the braking distance in approach to the signal, because the stopping distance is generally
longer than the range of vision.
In a territory where the distance between signals does not exceed the braking distance very
much, the approach aspect is usually provided by the signal in rear. This is called multiple-
section signalling, since a main signal provides information for at least two following block
sections. In a territory with very long distances between main signals, distant signals are
placed at the braking distance in the approach to a main signal (Fig. 2.1). A distant signal
can only provide an approach aspect for the signal ahead but it cannot show a stop aspect.
This signalling principle is called one-section signalling, since a main signal can only provide
information for one block section.

Fig. 2.1: Block Sections limited by Main Signals

16
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.2 Basic Terms of Railway Operation

Shunting signals are used to authorise shunting movements and to protect running
movements from shunting movements. In a territory where shunting movements may
enter main tracks, a shunt aspect is also integrated into the main signal, so that shunting
movements may be authorised to pass main signals in the stop position. For shunting signals,
an approach aspect is not provided because shunting movements are carried out at a very
low speed that allows stopping short of any vehicle or obstruction.
On North American railways, there is no differentiation between main and shunting signals
because of different operating rules.
Interlocking is a description of the arrangement by which points and signals are electrically
or otherwise interconnected in a way that each movement follows the other in a proper and
safe sequence. Signalled routes for trains on main tracks are usually interlocked. Signals
that govern train movements through an interlocking may be called interlocking signals. An
interlocking signal can also limit a block section. The points and signals within the interlocking
are controlled either by a local interlocking station or from a remote control centre. Local
interlocking stations are called signalboxes or signal cabins on railways that follow British
principles, and interlocking towers in North America. The main signals between controlled
interlockings are often called intermediate block signals. On most railways, the signals that limit
an interlocking are called home signals.
There are two basic signal arrangements at an interlocking (Fig. 2.2). First, there are
interlockings without consecutive interlocking signals within the home signal limits. The home
signals provide the authority to run through the entire interlocking into the next block section.
Second, there are interlockings with consecutive interlocking signals. The section of line
within the home signal which is limited by consecutive interlocking signals is often called a
station track, which is separated from the sections of the so-called open line. Although many

Fig. 2.2: Different Signal Arrangements at Interlockings

17
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

railways do not refer to these station track sections as block sections, they have the same
effect on train separation as a block section. Different from block sections of the open line,
station tracks may be used for overtaking, crossing, and direction reversal of trains. A station
platform for scheduled stops of passenger trains is not necessarily associated with a station
track which need not have a platform. There are also stations with their associated platforms
on the open line, which do not necessarily imply any provision of pointwork or the ability to
reverse trains within the signalling system.
On North American railways, interlockings with consecutive interlocking signals are not
common. The reason is that the distinction between station tracks and tracks of the open line
does not exist in the North American rules. In modern British signalling centres, there is also
no longer any distinction between station tracks and sections of the open line. However, from
the viewpoint of signalling, there is still a difference between track sections on which trains are
protected by controlled interlocking signals and track sections on which trains are controlled
by automatic block signals.

2.3 Diagramming Traffic

On most railways, traffic diagrams are used both as the basis of all planning of railway traffic
and also as essential documents for the control of the current operation. The only exceptions
are the North American railways where traffic diagrams (there called “stringline graphs”) are
used for capacity analysis and in very early stages of operation planning, while in current
operation tabular sheets are preferred.
The amount of traffic on a line is described in form of a time-distance diagram that consists
of a time axis and a station axis. At every station, there is a line parallel to the time axis. This
location represents the point where scheduled times for that station apply. Train movements
are represented by train paths (time-distance graphs) with a train description inscribed on
them. At the intersections of train paths and station lines, the actual time is marked by minute
numbers (Fig. 2.3). Traffic diagrams may be shown with a horizontal or with a vertical station

Fig. 2.3: Principle of a Traffic Diagram (Double Track Line)

18
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.3 Diagramming Traffic

Fig. 2.4: Traffic Diagram with Horizontal and Vertical Station Axis

axis (Fig. 2.4). Information of the two types of diagrams is of the same value. However, for
different railways, it is a matter of tradition as to which is used. In addition to the traffic diagram
of a line, additional diagrams are used to describe the traffic inside stations or interlockings
where different tracks are available per direction. In such a station traffic diagram, every track is
represented by a straight line parallel to the time axis.
The track occupancies are marked by stripes which have the description, arrival, and
departure times inscribed on them (Fig. 2.5). Small bent pieces at the end of the stripes are
suggested time-distance graphs to show the direction of movement. On small layouts, station
timetables are often used instead of those more complex station traffic diagrams. On some
railways, the occupation of station tracks is integrated in the line traffic diagrams. For smaller
intermediate stations, this is sufficient. For larger terminals, a separated station traffic diagram
leads to a better overview.

Fig. 2.5: Station Traffic Diagram

19
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

2.4 Scheduled Running Time

The scheduled running time of a train consists of the following components:

– The pure running time between scheduled stops
– The dwell time at scheduled stops
– Recovery time
– Scheduled waiting time
The pure running time between scheduled stops is the shortest possible running time as a
result from a running time calculation (see Chapter 4). To enable a train to make up small
delays, recovery time must be added to the pure running time. There are two kinds of recovery
time:
– Regular recovery time
– Special recovery time
The regular recovery time is added to every train path as a percentage of the pure running
time. The typical time supplement is 3 % – 7 % on European railways and 6 % – 8 % for
passenger trains on North American railways. On some railways, the regular recovery time
is evenly spread over the train path while other railways prefer to concentrate the recovery
time at the end of the run and at large intermediate terminals (Fig. 2.6). Sometimes, at large
terminals, the recovery time is not added to the running time in the approach to the terminal
but to the dwell time within it.

Fig. 2.6: Principles of Adding Recovery Time to a Schedule

The special recovery time is used to compensate the influence of maintenance and
construction works and of sections with temporarily bad track conditions. Different from
the regular recovery time, it is not added as a percentage of the running time but as a fixed
supplement to the running time on the section concerned. Scheduled waiting time is added
for scheduling reasons, e. g. to synchronise schedules of different passenger lines at changing

20
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.5 Modelling Train Paths

points, to synchronise schedules to a cyclic timetable (see Section 2.9), and to wait for a
scheduled passing or overtaking. Scheduled waiting times are mostly added to the dwell time
of scheduled stops but sometimes also to the running time.

2.5 Modelling Train Paths

A train path describes the usage of the infrastructure for a train movement on a track and in
time. As a basic requirement, a timetable must not contain any schedule conflicts between
train paths. It is not sufficient to describe a train path just by its time-distance graph. There
must also be a model to describe the “time channel” a train movement produces around
its time-distance line. A model that describes that time channel very exactly is the so-called
blocking time model. The same principle is also well known in analytical models for capacity
management, see Pachl & White (2003), Pachl (2009), Pachl (2013).
The idea of the blocking time model was developed in the late 1950s by O. Happel, who
was at the time the chief scheduling officer in the Frankfurt headquarters of German Federal
Railways (Deutsche Bundesbahn), see Happel (1959). The term blocking time is a literal
translation of the German term “Sperrzeit” which was introduced by Happel. Independently
from Happel, a similar method was developed by G. Adler at the former Dresden College of
Transportation, see Adler (1967). Implementation of that idea in practical railway scheduling
lasted several decades until computer-based scheduling systems became available. Use
in daily operation required yet more time and computer system development. The German
infrastructure operator DB Netz introduced a computer-based scheduling system based on
the blocking time model in 1998, see Brünger (2000). Beside the German scheduling system,
the blocking time model is also used in scheduling systems of other European railways, e. g. in
Austria, Denmark, Sweden, the United Kingdom and Switzerland.

2.5.1 Principles of Train Separation

Correct application of the blocking time model requires a basic understanding of the principles
of train separation. In highway traffic, separation of vehicles is effected by the principle of
relative braking distance. If one vehicle brakes, the driver of a following vehicle will notice the
brake rear lights and start braking. Two vehicles can follow in the distance that equals the
difference of the braking distances of the vehicles plus an additional safety distance to ensure
sufficient reaction time for the driver of the second car.
In a steel wheel on steel rail system, the coefficient of adhesion is on average eight times less
than in highway traffic. As a result, the braking force that can be transmitted from a train to the
track is also eight times less than the braking force that can be transmitted from a highway
vehicle to the road surface. That leads to braking distances for railway vehicles which may
significantly exceed the viewing range of the driver. Thus, train separation by the sight of the
driver is only possible when running at a restricted speed (usually no more than 25 km/h). This
is only acceptable for shunting movements and for running movements in degraded mode
operation. For regular running movements, procedures of train separation are required that
work independently from the viewing range of the driver.
The principle used for safe train separation depends on the following criteria:
– how movement authority is transmitted from track to train,
– how the line is released for further traffic behind a train.

21
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

If movement authority is only transmitted at discrete points, e. g. at lineside signals, or by

written or verbal orders, this will necessarily lead to train separation in a fixed block distance
since each movement authority has to cover the entire section up to next point where further
authority may be received. This principle is in effect on all lines where trains are governed by
fixed lineside signals and protected by an intermittent ATP system. This works for all kinds of
intermittent ATP systems including the Level 1 of the European Train Control System ETCS.
On lines where train separation in block distance is used, the track is divided in block sections.
A block section may be occupied exclusively by only one train. In a signalled fixed block
operation, the block sections are limited by signals which provide movement authority to enter
the block section protected by the signals. To clear a signal for a train that is to enter a block
section, the following conditions must have been fulfilled:
– The train ahead must have cleared the block section.
– The train ahead must have cleared the overlap beyond the next signal (only on lines where
block overlaps are used).
– The train ahead must be protected from following train movements by a stop signal.
– The train must be protected against opposing movements.
On railways where block overlaps are not required, the control length of a signal equals the
block section (Fig. 2.7 a). Examples are mainline railways in North America and in Russia. Other
railways require a control length of a signal that is longer than the block section (Fig. 2.7 b).

Fig. 2.7: Control Length of Signals in Fixed Block Territory

22
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.5 Modelling Train Paths

The difference is called “overlap” because in that area the control length of a signal overlaps
with the control length of the next signal. Overlaps are a common safety feature on railways
with high density passenger operation. The purpose of overlaps is to protect train movements
against trains that are overrunning a stop signal by a short distance due to bad brake
handling. In North America, overlaps are typical for subways and heavy rail rapid transit but
they are not normally used on standard railways. In Europe, where passenger operation is the
backbone of the railway system, overlaps are also frequently used on standard railways. Since
a main signal can only be cleared when both the block section and the overlap behind the next
main signal are clear, a train that is approaching a stop signal will always have a clear overlap
beyond that signal.
On lines where trains are governed continuously by a cab signal system, lineside signals are
no longer needed. However, continuous transmission of movement authority is not yet a
sufficient criterion to abolish fixed block sections. In addition, the train has to release the track
not at fixed intervals but continuously. This requires a permanent train-borne checking of train
integrity. Since for traditional railway systems a sufficient solution for that problem has not yet
been found, train separation in fixed block distance is still the standard principle for safe train
spacing on most railways worldwide.
From the viewpoint of train separation, the essential benefit of a fixed block system with cab
signalling compared with a fixed block system with lineside signals is the independence of
the cab signals from the approach distance of the lineside signal system, i. e. the distance
between the signal at the entrance of the block section and the signal in rear that provides the
approach indication. This allows trains to run at higher speeds. That is why cab signalling is
today the standard system on all high speed lines, worldwide.
On most railways, cab signalling is combined with a continuous ATP system. The most
common systems are the German LZB which uses a cable loop track antenna, systems based
on coded track circuits (e. g. in the French TVM system), and systems based on digital radio
as used in ETCS Level 2. However, on some railways, e. g. in Russia and North America,
cab signalling is also used on conventional lines as a pure signalling system without brake
enforcement.
Cab signalling without fixed block sections is generally known as moving block. This term
expresses the fact that fixed block sections no longer exist or, in other words, the block length
is reduced to zero. Since the train location data is transmitted at short intervals, the time
interval for safe train location is part of the minimum headway for two successive trains. Due
to the unsolved problem of train-borne checking of train integrity, moving block systems are
not yet used in standard railway operation but on a couple of transit railways. However, with
respect to future applications, the ETCS specification contains a Level 3 which is based on
moving block.

2.5.2 Application of the Blocking Time Model

The blocking time is the total elapsed time a section of track (e. g. a block section, an
interlocked route) is allocated exclusively to a train movement and therefore blocked for other
trains. Therefore, the blocking time of a track section begins with issuing a train its movement
authority for this section (e. g. by clearing a signal). The movement authority must be issued
before the train has reached the braking distance in its approach to this section. For example,
in signalled operation the train must not yet have passed the signal that gives the approach
indication to the signal at the entrance of the section. The blocking time ends after the train
has completely left the section and all signalling appliances have been reset to normal position

23
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

so that movement authority can be issued to another train to enter the same section. Thus,
the blocking time of a track section is usually much longer than the time the train occupies
the section. In a territory with lineside signals, the blocking time of a block section for a train
without a scheduled stop at the entrance of that section consists of the following time intervals
(Fig. 2.8):
– The time for clearing the signal
– A certain time for the driver to view the clear aspect at the signal that gives the approach
indication to the signal at the entrance of the block section (this can be a block signal or a
separate distant signal)
– The approach time between the signal that provides the approach indication and the signal
at the entrance of the block section
– The running time between the block signals
– The time to clear the block section and – if required – the overlap with the full length of the
train
– The release time to “unlock” the block system

Fig. 2.8: Blocking Time of a Block Section

24
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.5 Modelling Train Paths

On lines equipped with ETCS Level 1, the trackside data point that initiates braking curve
supervision when approaching a stop signal may be located at a specified distance in the
approach to the lineside signal which provides the approach indication. The reason is that
the braking supervision curves of ETCS Level 1 are based on a lower braking deceleration
compared with the braking deceleration used in traditional intermittent ATP systems. As
a consequence, the approach time is no longer the running time between the signal that
provides the approach indication and the signal at the entrance of the block section, but
the time the train runs through the braking distance that is supervised by the ETCS Level 1
system. This results in an increased approach time. For a more detailed analysis of that effect
see Wendler (2006).
For a train without a scheduled stop at the entrance of the block section, the approach time
is obsolete. In some publications, the blocking time is called occupation time which is a
literal translation of the German term “Belegungszeit”. That term was used in older German
textbooks, e. g. by Potthoff (1980). It is no longer preferred in the German terminology since
it may lead to misunderstandings. The same is true for the term occupation time since the
operational use of a block section by a train movement is significantly longer than the time
the section is physically occupied by the train. Therefore, the term blocking time seems much
more appropriate than the term occupation time. Stallybrass suggested the term “occupation
footprint” for the blocking time diagram of a block section (Stallybrass (2005)) which is also not
much clearer.
Drawing the blocking times of all block sections a train occupies into a time-over-distance
diagram leads to the so-called “blocking time stairway“ (Fig. 2.9). The blocking time stairway
represents the operational use of a line by a train. With the blocking time stairway, it is possible
to determine the minimum headway between two trains. The blocking times directly determine
the signal headway as the minimum time interval between two successive trains, considering
only one block section. The minimal line headway is the minimum time lag between two

Fig. 2.9: Determining Minimal Headways by Blocking Time “Stairways”

25
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

trains considering not only one block section but the blocking time stairways for the entire
line between two stations. In this case the blocking time stairways of two successive trains
would touch each other without any tolerance in at least one block section (the “critical block
section”). Although Fig. 2.9 shows trains following each other in the same direction as typical
for double track lines, blocking times can also be used in single track operation to find the
critical points between the paths of opposing trains.
To calculate the minimum line headway of two trains, the train paths are virtually put one over
the other with the same departure time. This results in overlapping blocking times in the block
sections. The blocking time overlap of a block section equals the amount of time the train path
of the second train has to be postponed to eliminate the blocking time conflict in this block
section (Fig. 2.10).

Fig. 2.10: Principle of calculating the Blocking Time Overlap of a Single Block Section

When the blocking time overlaps for all block section have been calculated, the maximum
value equals the time the train path of the second train has to be postponed to eliminate all
conflicts between the blocking time stairways. Thus, this time also equals the minimum
line headway between these two trains. After having prepared the blocking time data of all
reference trains in a calculation sheet, the blocking time overlap of each section can easily be
calculated (blocking time overlap = end of blocking time of 1st train – begin of blocking time of
2nd train) to find the greatest value.
th,ij = max{tbe,1(k) – tbb,2(k)} for k = 1…nb (2.1)

th,ij minimum line headway for train j following train i

tbe,i(k) end of blocking time of train i in block section k
tbb,i(k) begin of blocking time of train i in block section k
nb number of block sections

For a detailed calculation example, see the Chapter on capacity research in Pachl (2009).
Calculating minimum line headways by static blocking times is a deterministic approach, while
it is evident that real train operations are stochastic. However, modelling train paths by static
blocking times is still common practice in railway scheduling. Current research concentrates
on the subject of how the stochastic character of railway operations can be considered in
blocking time calculations. For stochastic modelling see Section 2.5.5 and Hansen (2004).

26
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.5 Modelling Train Paths

2.5.3 Modelling Specific Signalling Systems

Lineside Signals with short Block Sections

Concerning blocking time calculation, the main difference between block signal systems is the
allocation of the approach time. The approach time of a block section always starts at the first
signal in approach to that section that may slow down a train, when the signal at the entrance
of the block section shows a stop aspect. The signal where the approach time begins may
either be a separate distant signal (in one-section signalling) or a main signal (in multiple-
section signalling). Fig. 2.8 shows an example with two-section signalling in which every main
signal also provides the approach information for the next main signal. Therefore, one signal
provides information for two block sections. If the stopping distance exceeds the block length,
the braking curve takes more than one section. This can be achieved by different signalling
principles:
– Advance approach signalling
– Double approach signalling
– Progressive speed signalling
In advance approach signalling, a signal aspect exists that provides information that the
second signal ahead is at stop. Actually, this leads to three-section signalling. Thus, the train
has two full block sections for braking. Double approach signalling is used by some railways
on lines with very short block sections. There, the regular approach aspect is already displayed
at the second signal in approach to a stop signal. At the signal located directly in rear of the
stop signal, the approach aspect is usually repeated (often with an additional marker light to
indicate a stop signal to follow in reduced braking distance).
In progressive speed signalling, a train approaching a stop signal is progressively slowed
down by speed indications. The maximum speed at which the train may pass the last signal
in approach to a stop signal will ensure a safe braking within the short block section. Although
more than three signal aspects are used (mostly four aspects but in some installations even
more), it is only a two-section signalling system because approach information is given
only for the next signal ahead. There is no advance approach aspect. In progressive speed
signalling, on lines with mixed traffic of trains of different speed classes, the signal where the
approach time begins may depend on the timetable speed of an individual train. The approach
time always starts at the signal that requires the train to slow down. For more details on the
influence of the block signal systems on blocking time, see Pachl (2005).

Cab Signalling
In a territory with cab signalling, the principle is quite similar but the approach time is now
the time the train runs through the braking distance that is signalled by the cab signal system
(Fig. 2.11).
In cab signalling territory, lineside signals are generally obsolete. However, many railways
remove only the automatic block signals but not the interlocking signals. This leads to a
reduced lineside signalling system which enables a train without working cab signals to
follow another train in the distance between the interlocking signals. Thus, the block sections
between interlocking signals are an overlay to the cab signal block sections. In the blocking
time model, a train without working cab signals blocks the section between two interlocking
signals at once but releases the track in accordance to the short cab signal block sections
(Fig. 2.12).

27
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

Fig. 2.11: Blocking Time of a Block Section in Cab Signal Territory

Fig. 2.12: Blocking Time Stairways on a Cab Signal Line with Reduced Lineside Signalling

28
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.5 Modelling Train Paths

The blocking time can also be determined for trains that are controlled by a moving block
signalling system. The running time between block signals would be replaced by very short
intervals of less than a second for train location and the running time over the sight distance
is reduced to a safe reaction time of the driver. While all other components of the blocking
time are valid in the case of a moving block signalling system, its duration is limited essentially
to the running time of the train over its own braking distance and train length. On lines with
optimised distances between block signals according to the scheduled speed, the sum
of these other components is much greater than the part of the blocking time that could
be eliminated by a moving block signalling system. Only the “steps” of the blocking time
“stairway” for a line with fixed block sections would be eliminated and the blocking time
diagram is transformed into a continuous time channel (Fig. 2.13).
In many discussions, the benefit through the introduction of moving block signalling is
often over-estimated. By comparison of the blocking time diagrams for different signalling
systems, the effect on the use of capacity of moving block signalling can be very clearly
demonstrated. Its improvement potential is much less than expected provided the fixed block
signalling system is optimised with regard to the scheduled train speed and frequency, as
the infrastructure bottlenecks (e. g. switches, platform tracks) govern the minimal headway of
routes and lines. For more discussion on the implementation of the blocking time model on
lines with moving block, see Wendler (1995, 2006).

Fig. 2.13: Blocking Time of Fixed Block compared to Moving Block

29
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

2.5.4 Modelling Interlockings and Overlaps

Before a signal that protects a route through an interlocking can be cleared, all points must
be locked in the required position. The locking of a track element must not release before the
train has completely passed over that element (route locking). In older interlocking systems,
the points within the point zone beyond a main signal have a common release point located
behind the last pair of points. Thus, all points form a track section which is locked and
released in one piece. Modern interlocking systems have sectional route release. There, each
point releases separately after the passage of a train. For safety reasons, all track elements
that may release separately are connected by so called sequence locking. This prevents the
release of a track element until the element in rear has released. For this procedure, track
sections within the point zone that do not contain any points but have separate track clear
detection, are also logically locked and released. Crossings without movable points are also
locked logically to prevent conflicting movements through the crossing route.
In interlocking areas, every track element that may release separately has its own blocking
time. When a route is set up, all track sections from the home signal that governs the route
up to the next signal are blocked at the same time. After the train has entered the route, the
blocking time of the track sections the train has cleared ends according to the passage over
the release points for route release (Fig. 2.14).

Fig. 2.14: Blocking Time Diagram for a Train Movement running through an Interlocking

30
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.5 Modelling Train Paths

The reason for block overlaps was explained in Section 2.5.1. The handling of pure block
overlaps in blocking time analysis is quite simple. The blocking time of the overlap is only
considered as part of the blocking time of the block section but it is not necessary to show
a separate blocking time graph for the overlap. This can be done because the blocking time
of the overlap influences the clearing time of the block section, but it would never lock other
movements. In Fig. 2.15 this applies to signals 11, 13, and 17.

Fig. 2.15: Blocking Time Diagram in an Interlocking Area with Overlaps

However, in station areas on railways where overlaps are used, there are often successive
interlocking signals within the same home signal limits (a situation that is also not typical on
North American standard railways). There, an interlocked route leads from the home signal

31
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

to another interlocking signal that is often called a destination or exit signal. In complex
interlockings, the overlap behind the exit signal of an interlocked route may extend into the
point zone behind that signal. In Fig. 2.15 this applies to signal 15. So, as long a route is set
up at the home signal, the points within the overlap are locked.
The handling of such interlocked overlaps in blocking time analysis is different from pure
block overlaps. During the blocking time of the overlap, other routes that lead through the
overlap are locked. The overlap will release after the train has stopped (usually effected by an
automatic time release). Thus, the blocking time of the overlap may end at a time when the
track section in approach to the signal is still blocked. That is why the blocking time graph
must contain an extra blocking time for the overlap. Such interlocked overlaps generate route
conflicts which reduce the capacity of an interlocking. Railways that use interlocked overlaps
developed different principles to minimise the probability of having routes being locked by
overlaps of other routes.
On railways where overlaps are not used, a blocking time for locked overlaps does not exist.
The blocking times of block sections are also shorter than on a line with block overlaps
because the clearing time of a block section does not contain any time for clearing the overlap
behind the next signal. The same is true for the station track section before the exit signal. For
more details concerning the modelling of specific overlap arrangements, see Pachl (2005).

2.5.5 Stochastic Blocking Times

While the implementation of the blocking time model in computer-based scheduling systems
was a big step forward compared with simplified models used in manual scheduling and less
advanced computer-based scheduling systems (see Section 2.8), it still has the weakness of
being a purely deterministic approach. Two train paths are considered conflict-free, if there are
no overlapping blocking times. However, due to stochastic impacts on train moves, conflicts
may occur that have not been detected by overlapping blocking times. On the other hand,
delays in traffic operations that were previously detected by blocking time overlaps may make
conflicts disappear. In a deterministic model, the only way to cope with stochastic deviations
from the scheduled train runs is to add buffer times and recovery times based on the
operational experience.
In Medeossi et al. (2011), the Authors describe a new approach to obtain reasonable
distributions of the parameters needed to introduce stochastic modelling of train behaviour
into blocking time models. This includes both stochastic parameters to be applied directly
in the motion equation, and stochastic variations of departure times and dwell times. In that
model, the blocking time window of a track section does no longer have fixed limits at the
beginning and the end, but is represented by a blocking probability decreasing with distance
from the train path. Instead of a fixed limit, the blocking probability fades out at the edges. In
Fig. 2.16, the blocking probabilities are represented by different shades of grey. The diagram
refers to Fig. 2.9 that shows deterministic blocking time stairways for the same example. In
Fig. 2.16, the limits of the deterministic blocking times are marked by dotted lines. In areas
where the diffuse blocking zones overlap each other, a conflict probability can be calculated
from the overlapping blocking probabilities. The minimum headway between two trains paths
is no longer a deterministic value but depends on an accepted level of conflict probability.

32
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.5 Modelling Train Paths

Fig. 2.16: Stochastic Blocking Time Stairways

2.5.6 An Alternative Approach: The Protected Zone Model

In some computer-based systems for scheduling and capacity research, train paths are
modelled by a principle different from the blocking time model as explained above. It is here
called the protected zone model. The basic idea is that behind every train, the signalling
system or operating rules always provide a certain zone to protect that train against following
trains. The protected zone can be considered as a time interval or as a track length behind
a train. Fig. 2.17 shows the protected zone as a track length. The protected zone can be
divided into two parts (Fig. 2.17 a). First, there is a part of absolute protection. This part
consists of all track sections the protected train has exclusive authority to occupy. Second,
there is a part in which other trains are forced to slow down to prevent them from running into
the part of absolute protection. That zone starts at the first restrictive signal encountered by a
following train. The protected zone reaches its greatest length when a train is going to clear
a block section (Fig. 2.17 b). After the train has cleared the block section, the protected zone
has its shortest length and starts to grow again (Fig. 2.17 c). The change from the situation
of Fig. 2.17 b to the situation of Fig. 2.17 c represents a “step” in the blocking time stairway.
Diagramming the protected zones with additional times for operating and watching the signals
along the path of a train movements lead to picture similar to a blocking time stairway. The
main difference is that the approach time is not added at the beginning but at the end of the
occupation of the block section.
While the protected zone model is in some way similar to the blocking time model, there is
one essential shortcoming. The protected zone model cannot handle cab signalling systems
in which the signalled braking distance depends on the actual speed of a train. This is the
case in all moving block systems but also in most cab signalling systems based on fixed
block sections, e. g. ETCS Level 2. In such systems, a train can only calculate its own braking

33
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

Fig. 2.17: The Protected Zone Model

distance but not the braking distance for a following train. That is why a model that calculates
the braking distance ahead of a train, as the blocking time model does, is more appropriate
than a model that considers the braking distance as part of a protected zone behind the rear
of the train. In traditional signalling systems, in which all trains are governed by lineside signals,
the protected zone model works quite well, however.

2.6 Headways and Buffer Times

Planning sufficient headway between train paths is an essential requirement of scheduling.

Headways can be considered in two different ways:
– Headways assigned to the stations that limit a section
– Headways assigned to the section between two stations
Assigning headways to stations is the more traditional principle (Fig. 2.18 a). It is still common
in scheduling. There are four types of headways, see Potthoff (1980):
– The headway between two trains that depart sequentially onto the same line (“depart-
depart” headway)
– The headway between two trains that arrive from the same line (“arrive-arrive” headway)
– The headway between the arrival of a train and the departure of an opposing train towards
the same line (“arrive-depart” headway)
– The headway between the departure of a train and the arrival of an opposing train from the
same line (“depart-arrive” headway)
The point where time applies is usually not a signal but the station line of the train diagram.
This principle leads to different headways at both sides of a section. If the minimum headways
for the different train combinations are put into a spreadsheet (e. g. as input data for
scheduling systems that have no blocking time model implemented), each section has two
sheets for the two sections limiting that section. However, the average headway at one station
equals the average headway at the station at the other end of the section.
If headways are assigned to a section between two stations, there are no different types of
headways. The headway is always the difference between the times two successive trains of
either direction enter the section (Fig. 2.18 b). In recent years, that principle has become quite

34
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.6 Headways and Buffer Times

Fig. 2.18: Principles of assigning Headways

popular for the analysis of line capacity. The advantage is that the minimum headways for all
train combinations of a section can be described by just one sheet. This advantage is also
interesting for scheduling systems that use minimum headway sheets as input data.
This principle allows calculation of the headways in two different ways. The headway can
either be calculated from the times the trains pass a specified point when entering the section
(e. g. a signal or the point where the scheduled departure time applies), or from the beginning
of the blocking time of the first block section between the two stations (Fig. 2.19). The second
principle is very valuable for capacity research. For scheduling, calculation of headways from
the passage at specified points is more appropriate.
The scheduled headway between two trains must consist of the minimum line headway plus
the required buffer time to compensate for small delays. The buffer time is the smallest slot
between the blocking time stairways of two trains (Fig. 2.20 and 2.21). On some railways,
the buffer time is called a recovery margin. However, this term must not be confused with the
recovery time. The recovery time enables a train to make up a small delay while the buffer time
prevents a small delay from being transmitted to other trains. On the other hand, the recovery
time extends the running time of a train while the buffer times reduce the number of trains that
may be scheduled.
The amount of buffer time depends on the required level of service. Usually, the buffer time
is determined depending on the kind of line headway. Most railways use the following basic
rules:
– Large buffer time when the second train has a higher priority than the first train
– Small buffer time when the first train has a higher priority than the second train
– Average buffer time when both trains have the same priority
Today, most railways still allocate buffer time in a deterministic way, i. e. to have fixed minimum
buffer times to be added to the different kinds of train combinations. Current research activities
try to develop procedures for a stochastic modelling of buffer time allocations, see Hansen (2004).

35
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

Fig. 2.19: Principles of calculating Headways

Fig. 2.20: Buffer Times on a Double Track Line

36
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.6 Headways and Buffer Times

Fig. 2.21: Buffer Times at a Passing Station

Swiss railways already use the principle that in terminal areas with a very high density of traffic, train
paths many be planned without buffer if sufficient recovery time for making up delay is added to the
running time between terminals, see Herrmann et al. (2006). An exact allocation of a buffer time to
each train path in accordance with the rules above only makes sense in an accurately scheduled
operation where the train sequence seldom changes. On lines where trains often run out of their
schedules or with a lot of extra trains, the traffic may be scheduled without buffer times for every
single train. This is typical for freight lines and for connecting lines inside large terminal areas. On
such lines, the buffer that is required for
a good quality of operation could be
better provided by the rule that a certain
number of trains must be followed by a
clear buffer path (Fig. 2.22).
Beneath the train path buffer times,
buffer times are also required at
connecting stations where scheduled
transfer connections between trains
exist and where crew or equipment
changes from one train to another. To
avoid the transmission of delay at such
points, a buffer time must be added to
the time that is required for the transfer
(Fig. 2.23). For a stochastic modelling
of the transfer time see Goverde
(2005).

Fig. 2.22: Buffer Path

37
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

Fig. 2.23: Buffer Time for Passenger Transfer at a Connecting Station

2.7 Consumed Capacity

The line capacity consumed by a timetable can be visualised very clearly by virtually moving
the blocking time stairways together as close as possible, without any buffer times but without
changing the sequence of trains (Fig. 2.24). This principle is also known as the “compression
method”. Analysis of a given timetable consists of the following steps:
– Determination of all buffer times in the traffic diagram
– Finding the critical buffer path, e.g. the path through the diagram with the minimum sum of
buffer time
– Calculating the route occupancy per time period on the critical path

Fig. 2.24: Visualisation of the Consumed Capacity

38
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.7 Consumed Capacity

In complex timetable structures, in particular on lines with a lot of single track sections,
identifying the critical buffer chain might become a complicated job that requires support by
a computer-based tool. For a comprehensive discussion of how to apply the compression
method on single track routes see Landex et al. (2007).
The route occupancy is calculated as the ratio of the sum of the minimum line headway times
divided by the total time period. The sum of the minimum line headway times equals the
number of trains multiplied by the average minimum line headway:
 = n · th / tp (2.2)

 ratio of consumed capacity

th average minimum line headway
tp total elapsed time of the considered period

To calculate the average minimum line headway, trains with similar movement characteristics
should be grouped together into reference train classes. Then, the average minimum line
headway can be calculated by the following equation:
th =  (th,ij · fij) (2.3)
th average minimum line headway
th,ij minimum line headway for train j following train i
fij relative frequency of combination: train j following train i

The relative frequency of every train combination can be read directly from that timetable.
To evaluate the capacity independently from the timetable, the relative frequency of every
combination of two reference trains can be calculated by this formula:
fij = (ni · nj) / n2 (2.4)
fij relative frequency of combination: train j following train i
ni number of train i
nj number of train j

If the traffic pattern changes over 24 hours, the relative frequencies should be calculated
separately for different periods and then consolidated into a value for the entire day.
From the experience of European railways (mostly double track with mixed traffic of passenger
and freight trains), for a good quality of operation, the ratio of consumed capacity should not
significantly exceed
– 50 % during 24 hours,
– 80 % during a peak period of 4 hours.
That means, the average buffer time between two trains should not be a lot less than the
average minimum line headway. On single track lines with large distances between stations
used as scheduled passing points and on double track lines with low speed differences (e. g.
rail rapid transit), a higher ratio of consumed capacity can be accepted. In 2004, the UIC
issued a code with more detailed recommendations to the limits of the consumed capacity,
see Table 2.1. However, these limits are mainly based on European operating experience. On
railways with very different operating conditions, e.g. in North America, these limits may differ.

39
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

Type of line Peak hour Daily period

Dedicated suburb 85 % 70 %
passenger traffic
Dedicated high 75 % 60 %
speed lines
Mixed traffic lines 75 % 60 %

Table 2.1: Recommended limits of consumed capacity in UIC Code 406 (UIC 2004)

2.8 Scheduling Methods

The scheduling process can be done in two ways:

– Manual scheduling
– Computer-based scheduling
For more than 150 years, manual scheduling was the typical scheduling method on almost
all railways. Only the development of high performance computers in the 1990s offered the
opportunity to introduce computer-based scheduling methods. While some such systems use
the blocking time model for representing train paths, there also exist simpler scheduling tools
that adopted the principles of manual scheduling. Therefore, an explanation of the train path
modelling in manual scheduling is still valuable.

2.8.1 Manual Scheduling

In manual scheduling, the train path is constructed as a polygon from station to station.
The running times are read from special tables, which consist of the running times between
all stations and the extra times required for acceleration and braking. These running times
already contain the regular recovery time. Special recovery times for maintenance must be
added manually. Since the running time calculation is rather complex, these tables are usually
prepared by a computer. Thus, manual scheduling is not conducted completely without
the support of computers (on some railways with only a few different train classes, running
time can be determined by physical checks).To determine the minimum line headways, a full
calculation of the blocking time graphs is not practical in manual scheduling. The minimum
line headways are determined in a simplified way. The blocking times of the block sections
are considered by a general supplementary time added to the running time inside the section.
On European railways, this extra time is usually 1 min (Fig. 2.25). Furthermore, a buffer time is
added, which usually lies within the range of one to three minutes.
Generally, the minimum buffer time is determined in the same manner as described in
Section 2.5. At scheduled stops, the correct allocation of the dwell time to the block
sections depends on the signal arrangement of the station. At a station with a signal both
at the entrance and the exit of the station track, the dwell time neither belongs to the block
section in approach to the station nor to the block section beyond the station (Fig. 2.26 a).
At a station with a signal only at the exit of the station, the dwell time belongs to the block
section in approach to the station (Fig. 2.26 b). The block section in approach to the station will
only release after the train has departed. At a station with a signal only at the entrance of the
station, the dwell time belongs to the block section beyond the station (Fig. 2.26 c). The block
section in approach to the station is released after the train has arrived.

40
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.8 Scheduling Methods

Fig. 2.25: Principle of Manual Scheduling

It is not good practice to plan stations with a signal only at the entrance of the station. When a
train has to wait at the signal because of an occupied block section ahead, it cannot enter the
platform track and use the waiting time for passengers to leave and board. On North American
railways, in automatic block territory without cab signals, the operating rules require a train
that has a stop at a station beyond an automatic block signal but without a signal at the exit
of the station, to proceed at restricted speed to the next signal. That makes such a signal
arrangement even more disadvantageous.

Fig. 2.26: Allocation of the Dwell Time to Block Sections

41
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

2.8.2 Computer-based Scheduling

Some computer-based scheduling systems use simplified models. These may be based either
on the same principle as described for manual scheduling, or on using pre-defined matrices of
minimum line headways associated with line sections or stations. This allows the infrastructure
model to be kept quite simple. However, in such systems, especially in complex track layouts,
some scheduling conflicts may not be detected. That is the reason why these systems would
still require quite a lot of experience and control by the user. This kind of computer-based
scheduling is very successful compared with pure manual scheduling and will reduce the
effort significantly. It is a reasonable option for infrastructure operators with a not very complex
network structure. However, in complex networks and big terminal regions, scheduling
systems based on simplified models will reach their limits.
The heart of more advanced computer-based scheduling systems is a calculation of the
blocking time stairways. Some railways use the same simulation systems for scheduling that
are used for capacity research. But with pure simulation software, it is not easy to detect

Fig. 2.27: Visualisation of a Scheduling Conflict

42
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.9 Cyclic Timetables

the exact location of train path conflicts where buffer times have to be added. To meet the
demands of the railway companies, many suppliers of simulation software have recently
produced scheduling modules as add-ons to their software. These modules can calculate
and display blocking time stairways to the simulated train paths in a way very similar to
pure scheduling systems. The blocking time stairways can be visualised on a screen. So,
scheduling conflicts are exactly detected by the overlapping of blocking times (Fig. 2.27).
This kind of a computer-based scheduling system requires a very sophisticated infrastructure
database, which not only contains the track layout with all speed limits and line restrictions
but also all signalling data that is necessary for the calculation of blocking times. The timetable
designer can solve scheduling conflicts by displacing or modifying the time-distance curves.
In the near future, there will be scheduling systems that assist the scheduler not only by
visualisation of conflicts but also by suggestions for solving conflicts based on background
knowledge.

2.9 Cyclic Timetables

A cyclic timetable is based on a traffic pattern that repeats itself every hour. Within this pattern,
trains of the same route are scheduled at fixed intervals. Cyclic timetables are very common
on European passenger lines, for commercial reasons. On single track lines thus operated,
the running time between two scheduled meeting points (passing tracks) equals half of the
fixed scheduled headway (Fig. 2.28). When two passing stations are too close together, the
running time or the dwell time has to be extended up to this amount. This is a big constraint
regarding the construction of a timetable with fixed time intervals on single track lines. Due to
the mathematics of the situation, the number of meeting points on a single track line required
for a timetable with fixed time intervals can easily be determined by the following equation:
nMP >= 2 · trt / tfix (2.5)
nMP number of meeting points
trt running time of the line
tfix fixed scheduled time interval between trains

The number of train sets needed to operate a route on fixed time intervals is calculated as
follows:
nT = tc / tfix (2.6)
nT number of train sets
tc cycle time
tfix fixed scheduled time interval between trains

The cycle time is the total time between two successive departures of the same train set at the
same station in the same direction (Fig. 2.29).
Concerning the combination of several train routes, there are three scheduling strategies used
in cyclic timetables:
– Non-symmetrical cyclic timetables
– Symmetrical cyclic timetables
– Integrated cyclic timetables

43
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

Fig. 2.28: Constraints of fixed interval Schedules on Single Track Lines

As shown in Fig. 2.28, trains from opposing directions meet twice within the fixed scheduled
time interval. Thus, for the trains of one route, the timetable is always symmetrical at the
meeting time of opposing trains. This time is also called the “symmetry time” of the timetable.
If different fixed interval scheduled train routes run over the same line, these train routes may
have different symmetry times. In such a case, the timetable is called a non-symmetrical
timetable. If all train routes have the same symmetry time, it is a symmetric timetable. For a
detailed comparisons of these two basic types of cyclic timetables with evaluation of their
advantages and disadvantages under specific operating conditions, see Liebchen (2006).
Some railways with an extensive passenger operation have connected the cyclic timetables of
different routes of the network. In such integrated cyclic timetable, trains on different routes are
scheduled in a way that, in a connecting terminal, all trains meet always at the same time to

44
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 2.9 Cyclic Timetables

Fig. 2.29: Cycle Time

enable the passengers to change between all lines (Fig. 2.30). That means, in these terminals,
the trains of connected routes must have the same symmetry time. Because of many constraints
on every line and between the different lines, the operation of such a system of integrated cyclic
timetables requires a very carefully scheduled and controlled traffic.
A comprehensive discussion of optimisation of cyclic timetables is provided in Chapter 8.

Fig. 2.30: Principle of an Integrated Cyclic Timetable

45
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
2 Timetable Design Principles

References
Adler G. (1967): Die Verkettung der Streckenbelegungen und der Belegungsgrad einer
Gesamtstrecke. Diss. HfV Dresden.
Goverde R. (2005): Punctuality of Railway Operations and Timetable Stability Analysis. TRAIL
Thesis Serie no. T2005/10, Delft.
Hansen I. (2004): Increase of capacity through optimised timetabling. Computers in Railways
IX. WIT Press, pp. 529 – 538.
Happel O. (1959): Sperrzeiten als Grundlage für die Fahrplankonstruktion. Eisenbahntech-
nische Rundschau, Vol. 8, No. 2, pp. 79 – 90.
Herrmann T., Burkolter D., Caimi G. (2006): Bewertungsgrundlagen ausgelasteter Fahrpläne.
Verkehr und Technik, Vol. 59, No. 8, pp. 307– 311.
Landex A., Kaas A. H., Jacobsen E. M., Schneider-Tilli J. (2007): The UIC capacity method
used on single track sections. Rail Hannover 2007 Proceedings (CD-Rom), International
Association of Railway Operations Research, Hannover.
Liebchen Ch. (2006): Fahrplanoptimierung im Personenverkehr – Potenzial mathematischer
Optimierung. Eisenbahntechnische Rundschau, Vol. 55, No.7/8, pp. 504 – 510.
Medeossi G., Longo G., de Fabris S. (2011): A method for using stochastic blocking times to
improve timetable planning. Journal of Rail Transport Planning & Management 1, pp. 1 – 13.
Pachl J. (2005): Modelling Specific Signalling Features in Computer-based Scheduling
Systems. Rail Delft 2005 Proceedings (CD-Rom), International Association of Railway
Operations Research, Delft.
Pachl J. (2009): Railway Operation and Control. 2nd Edition. VTD Rail Publishing, Mountlake
Terrace.
Pachl J. (2013): Systemtechnik des Schienenverkehrs – Bahnbetrieb planen, steuern und
sichern. 7th Edition. Springer Vieweg, Wiesbaden.
Potthoff G. (1980): Verkehrsströmungslehre, Band 1 – Die Zugfolge auf Strecken und in
Bahnhöfen. 3rd Edition, transpress Verlag, Berlin.
Stallybrass M. (2005): Proving and improving timetables. Railway Technical Review, Vol. 54,
No. 2, pp. 26 – 30.
UIC (2004): Code 406 – Capacity. 1st Edition.
Wendler E. (1995): Weiterentwicklung der Sperrzeitentreppe für moderne Signalsysteme.
Signal und Draht, Vol. 87, No. 7/8, pp. 268 – 273.
Wendler E. (2006): ETCS und Kapazität. Proceedings VDE Kongress 2006 Aachen, Vol. 2, pp.
369 – 374.
White T. (2003): Elements of Train Dispatching 1. VTD Rail Publishing, Mountlake Terrace.
White T. (2005): North American Experience with Timetable Free Railway Operation. Rail Delft
2005 Proceedings (CD-Rom), International Association of Railway Operations Research, Delft.

46
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 3.1 Introduction

3 Infrastructure Modelling
Alfons Radtke

3.1 Introduction

An accurate infrastructure model is the fundamental base for various railway planning tasks. In
this Chapter, the basic facts about modelling railway infrastructure, their limitations and general
hints, will be presented. The railway infrastructure dealt with in this Chapter consists typically of
the following elements:
– Civil infrastructure
– Electrical infrastructure
– Signalling infrastructure
– Telecommunications infrastructure
– Geo-reference based information
Different modelling approaches and applications will be discussed in detail. The modelling of
railway infrastructure is the basis for the tasks involved in the planning of railway operations
such as:
– Running time calculation
– Calculation of headways
– Depiction of different signalling systems
– Timetable Construction, including block occupation for conflict detection and resolution
– Planning of possessions and track work
– Planning of infrastructure restrictions
– Searching for train routes or train paths in networks
– Capacity calculation
– Railway operational simulation
In practice, there are various different sources that can be used to create infrastructure data
and later on maintain the models. To do so either manually or electronically, the source of the
infrastructure data has to be checked for:
– Data quality
– Possibilities for data maintenance
– Intellectual property status of the data
Data quality is an extremely important issue. Using incorrect or inconsistent information may
cause serious problems in the planning process later on. Thus incorrect gradient information
will in turn cause errors in running time calculations. Therefore, it is essential to establish a
permanent quality check of the infrastructure model. This could be performed by visualisation
of the data and a manual crosscheck by experienced planners, or software based testing, for
example a comparison of calculated running times with real running times, or the calculation of
typical network indicators such as the length of the network, the number of signals etc.
The second point to take into account is the maintenance of the computer based
infrastructure models. It is necessary that all changes and alterations of the real infrastructure
are replicated in the model. For this purpose, it is essential that a suitable business process
is established in the railway organisation to transfer the relevant information between various
business units. This business process needs to be maintained irrespective of departmental
reorganisations. Another question is the IT based merging of different infrastructure models.
This task requires sophisticated software programmes.

47
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
3 Infrastructure Modelling

This last point is related to the intellectual property and exchange of the raw data and the
developed computer based infrastructure models. Questions like “Who has legal access to
the raw data and to the computer based infrastructure models?” and “Is it possible to transfer
either raw data or modelled infrastructure to third parties?” must be answered. As of late,
there are two approaches. On the one hand, there is more and more published infrastructure
information available, particularly on the internet. However, in most cases this data is not
sufficient for the exact planning of railway operations (e. g. missing gradient information to
perform running time calculations and timetabling). Furthermore, the quality of the data needs
to be checked beforehand.
On the other hand, there are public (and still openly available) sources of information, such as
OpenStreetMap®. With this, it is possible to create at least the “backbone” of infrastructure
models, taking into account coordinates or information on distances. In order to develop an
interchangeable format for infrastructure data, the RailML® initiative provides xml-schemes
for infrastructure modelling, currently with the version 2.2 and with the prospect of moving on
imminently to version 3.0.
Possible raw data sources are in general based on paper or IT systems:
– Track layout plans
– Signalling plans
– Illustrative schematics
– Data on various electronic devices such as
– Databases (e. g. MySQL, Oracle)
– RailML® or XML files
– EXCEL-sheets
– Public internet web sites
– OpenStreetMap® as internet GIS based information
– Other formats

3.2 Graph Theory and its Application

It is best practice in the modelling of railway infrastructure to use structures derived from
Graph Theory. Graph models Hauptmann (2000), Radtke and Watson (2007) have the proven
advantage that they are very flexible and a powerful means to depict even the most complex
railway infrastructure in an efficient mathematical model. Furthermore, these models allow
modular and redundancy free storage and handling of the data in large computer models. These
models can be extended and modified, and their size is generally not limited. The real existing
railway infrastructure (tracks, signalling systems or attributes like gradient, radius, overlaps etc.) is
separated into pieces (links). These links are bounded by nodes, which join the links.
For a better understanding of railway infrastructure modelling, a few definitions will be
introduced:
– A node (in mathematical terms called ‘vertex’) is a representation of an arbitrary location in
a railway network.
– A link (or mathematically ‘edge’) is a connection between two nodes.
The level of infrastructure detail is defined as follows:
– Microscopic node-link-models contain the highest possible level of details on nodes and
links, depending on the purpose.
– Macroscopic node-link-models contain aggregated information on nodes and links.
– Mesoscopic node-link-models are syntheses of microscopic and macroscopic
infrastructure models.

48
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 3.2 Graph Theory and its Application

For all infrastructure models a valued graph (see Fig. 3.1)

G: = (V,E,c)

with a weight function

c(e) ≥ 0 ∀e ∈ E

and the nodes V (vertex) and the links E (edges) can be defined as:
– Directed: if two adjacent nodes are linked with at least one link, the direction is illustrated
with an arrow.
– Simple: if G does not contain parallel links or loops.
– Connected: if for any two nodes of G, links exist connecting the nodes.

Fig. 3.1: Example of link and node model

In the simple example illustrated in Fig. 3.1, four links can be seen: ({1-2}, {2-3}, {3-4} and
{2-5}). This graph is directed (see the arrow at the end of the link), contains no loops and node
{2} is incorporated in three links.
Using a graph as a mathematical model, various railway-related problems can be transformed,
to be stored and processed by computer algorithms:
– Concrete railway infrastructure
– Abstract dependencies, rules or flows
In Fig. 3.2, a simple graph model of the railway infrastructure of a small station including track
numbering can be seen. It should be noted that not all node types appear in the illustration.
There are for example no nodes for changes in speed or gradient visible.

Fig. 3.2: Example of a simple link and node model of a station

For various optimisation problems and their depiction in graphs, very efficient mathematical
algorithms and heuristics exist. For the modelling of railway infrastructure, the following two
problems are among the most significant:

49
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
3 Infrastructure Modelling

– Calculation of shortest paths in graph networks between two nodes according to different
criteria (shortest path problem). The Dijkstra algorithm could be used to solve this kind of
problem, see Dijkstra (1959).
– Calculation of maximum or minimum flows in graph-networks taking into account certain
capacity functions to the links or nodes and not depending on time (maximum and
minimum flow problems, Out-of-Kilter or Edmonds and Karp algorithm could be used here,
see Schumacher (2004), Radtke (2006).
In addition, certain railway problems (e. g. allocation of vehicles) can only be described using
more complex approaches on graphs like the multi-commodity flow, maximum concurrent
flows, dynamic flows or even a combination with heuristics algorithm like simulated annealing
or genetic algorithms, see Kettner (2005), Schumacher (2004).
The attributes of a railway infrastructure modelled in a graph are assigned to the links and
nodes. Typical attributes of nodes are:
– Geographical attributes (kilometre marking, position co-ordinates, names etc.)
– Type of node (signal, point, timing points etc.)
For the attribution of links two approaches exist:
– Link-orientated models: Each link contains all relevant information (e. g. speed, gradient,
radius, electrification, direction).
– Node-orientated models: Special node types indicate the change of attributes (e. g. speed
by a speed change node, gradient by a gradient change node, etc.). In general, the links
do not contain any information in node-orientated models. Each node has positional co-
ordinate. The length of an edge can be calculated by the difference of the position of the
start and the end nodes.
Both approaches have advantages and disadvantages. The redundant assignment of all
railway infrastructure attributes to each single link in a link-orientated model can be seen as a
disadvantage, wasting storage capacity and implying problems in handling changes to those
attributes. If a user needs to change a certain attribute in a defined area, all links of that area
have to be considered. In some cases, links have to be separated and a new node created in
order to obtain a modified assignment of attributes.
However, modern infrastructure editors provide efficient support with easy to use functions
to overcome the data-handling problem. Furthermore, the amount of data to be stored
in the RAM of a computer is no longer critical. In theory, the entire network of the German
Railway with approximately 64,000 km of track would result in 64,000,000 links of one-metre
length. In practice, one-metre links are the exception, the entire German microscopic railway
infrastructure network with approximate 850,000 links and 830,000 nodes can be held in
2 GB RAM on a standard PC.
The advantage of a node-orientated attribution model is certainly the possibility for redundancy
free data storage. For example, to introduce a speed change, a node of the type “speed node”
has to be set at the desired position, or an existing node has to be changed. On the other
hand, node-orientated models require more complex algorithms to calculate all the attributes
on a certain link of the network. This task can be very challenging when considering the
kilometre markings in railway networks, which are often not well defined for historical reasons.
Changes of the kilometre marking direction (going up, going down) and the handling of erratic
kilometre values require significant effort. In general, link-orientated models do not need this
functionality.
In Fig. 3.3 an alternative approach to node modelling has been illustrated. The so-called colon-
graph allows an easy way to ensure the right utilisation of points during railway operation.

50
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 3.2 Graph Theory and its Application

On the right side, the duplication of the nodes can be seen. Applying the rule that a feasible
path contains “node-node-link-node-node-link-node-node” relation, it is obvious that only
paths from A via B to C or D are feasible; a path D – B – C is excluded.

Fig. 3.3: Principle of colon-graph

Fig. 3.4 illustrates the transition of raw data from track layout plans into a node-link graph
model. The depiction of a set of points (node {2}) can be seen on the left side. In practice, the
middle of the set of points is transformed to a node. The link {2-3 or 2-4} contains the speed
for the diverging route. On the right side of Fig. 3.4, the crossing allows the following paths ({1-
2} to {2-4} and {1-2} to {2-5} / {3-2} to {2-5} and {3-2} to {2-4}).

Fig. 3.4: Transition from track layout plan to a graph model

To summarise, it is proven by a large number of applications that graph theory can model
very complex and large railway infrastructure in a mathematical model. Furthermore, efficient
algorithms and concepts to store the data and to solve railway operational problems have
been developed.

51
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
3 Infrastructure Modelling

3.3 Macroscopic Models

In general, macroscopic models are usually preferred for long term planning tasks or special
routing problems. Macroscopic infrastructure models contain far fewer links and nodes in
comparison with microscopic models and have a more abstract view of the infrastructure. For
example, the entire German Railway infrastructure can be described in a macroscopic model
with approximately 10,000 links and 10,000 nodes. In 2002, the macroscopic network used
by the slot price information system TPIS (TrassenPreisInformationsSystem) required only
1,600 nodes and 6,900 links.
A node in a macroscopic model represents a station or a junction in the network. Fig. 3.5
illustrates the basic principle of a macroscopic railway network. The microscopic model of the
element “station”, “line” and “junction” may include a couple of hundred microscopic links and
nodes, depending on the complexity. For example, the “station” could be a simple regular stop
along a line with platforms on the main track, or the entire Hannover main station in Germany
including 14 platform tracks at the convergence of several railway lines. The macroscopic
model illustrated in Fig. 3.5 contains only two nodes and four links.

Fig. 3.5: Principle of macroscopic infrastructure modelling

Macroscopic data can be entered manually from various sources; even public (internet)
sources are sufficient in some cases. If a microscopic infrastructure database is available, a
direct transfer from the microscopic to the macroscopic model is preferable.
A typical macroscopic node would contain the following information:
– Geographical attributes (co-ordinates, names etc.)
– Type of node (station, shunting yard, junction etc.)
– Some operational information (terminal station, travel resistances)
Macroscopic links may hold the following information:
– Length
– Type of line (high speed, passenger, freight or mixed line)
– Line number
– Number of tracks
– Train availability (electrification, axle load, loading gauge)
– Average running time
– Average capacity (e. g. according to UIC 406)

52
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 3.4 Microscopic Models

One main application of macroscopic network models for long term infrastructure planning
and traffic assignment purposes is the search for train paths in networks without time
restrictions. Introducing new trains during traffic assignment, finding a path or rerouting
of freight trains always requires the search for a feasible path. In practice, this is a two-step
process, see Sewcyk (2007). The first search on a macroscopic network does not consider
all individual tracks on lines or stations, but determines a sequence of stations for the train
trip. Some fundamental criteria considered are, for example, axle load, electrification or other
operational rules like stopping patterns for passenger trains or preferred routes for high-speed
or freight trains. In complex networks, the computing time can be reduced. After finding an
initial solution (e. g. through applying the Dijkstra algorithm) the “fine” tuning of the train path
can be finished later on a microscopic network, working out all tracks on lines and through
stations. In addition, the platforming in stations can be determined.
It is possible to control the routing process, even without the exact depiction of single tracks
in stations, by using the method of the inverse graph, see Kettner (2005). This problem may
occur in terminal stations where the inverse graph allows the correct modelling of feasible
paths along the lines and the terminal station.
Macroscopic models are not suitable for all detailed railway operation planning tasks like
the calculation of running times or conflict detection. In general, these models do not
contain information about the exact maximum speed and track gradient or restrictions due
to the signalling system. Therefore, it is not possible to calculate a correct running time. The
approach of using an “average” gradient or speed in macroscopic models will mislead a user,
and running time calculations based on this data should not be used in serious research or
consultancy work.

3.4 Microscopic Models

A detailed node-link model can be used to model infrastructure in a so-called microscopic

infrastructure model. This model combines track information such as speed, gradient
or radius, with the signalling system (signals, block sections, release points) and some
operational information like routes, alternative platforms, timing points and availability (time
stamps). Every single change in one of the attributes considered above requires a new node,
splitting an existing link and generating a new one. Depending on the chosen approach, the
information is assigned to either the node or the link. In practice, the minimum length of a link
describing existing railway infrastructure in microscopic models has the length of one metre.
For modelling reasons, it is useful to work in a few cases with 0-m links (modelling of signalling
issues). Based on this granularity, the microscopic infrastructure model is not only suitable
but even mandatory for exact running time calculation, timetable construction, possession
planning and railway operational simulation, conflict detection and resolution.
A typical microscopic infrastructure model contains all tracks on lines and especially in
stations. Fig. 3.6 shows the how single links build a complete railway infrastructure. A block
section, defined as a section protected by signals, can be formed by a few links. The block
section itself has a starting and an ending node; both of these nodes belonging to existing
individual links. Therefore, in the model, no additional nodes to describe the block are
necessary.

53
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
3 Infrastructure Modelling

Fig. 3.6: Possible hierarchy in microscopic infrastructure models

Several block sections can be abstracted to route sections, which are sequences of links
guiding the trains through the station. They are provided by the interlocking system in the
respective station. Trains are assigned to route sections, which can be used for specifying
the correct train path through each station, finding a correct train path during timetable
construction or railway operational simulation, or assigning signal protected paths into a
complex station to a platform.
Therefore, a route section can only be created if this section is feasible from the operational
point of view. A route section, which is supposed to be used by electric locomotives,
must not be created in a model if some links are not electrified (or are temporarily without

54
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 3.4 Microscopic Models

electric supply). This simple approach can be used to transfer technical requirements (e. g.
electrification) or operational rules (alternative platforming taking into account the platform
length) into the model.
Slightly different modelling approaches use station sections for the definition of train paths.
Station sections are sequences of connected links within a single station, starting and ending
at so-called station boundaries. They are provided by the interlocking system in the respective
station. Trains are assigned to route sections to specify the correct train path through each
station. This concept has the advantage that complex train movements in stations can clearly
be described in the model.
In 1978, Schwanhäußer introduced a model of decomposing large railway networks into route
nodes, see Schwanhäußer (1978). Let S be the set of points and crossings of an investigated
railway network. Then the route nodes are defined as the respectively greatest possible
subsets Sk
Sk  S

so that every pair (i, j) of interlocking routes i and j using Sk are incompatible with each other.
All subsets are disjunctive to each other. This means that every set of points and crossings
belongs to exactly one route node (see Fig. 3.7).

Fig. 3.7: Example of route nodes

The route nodes can be interpreted as servers of a single server queue in order to assess
the capacity and to identify the bottlenecks of interlocking zones. Moreover, the route nodes
can be used in the dispatching process, as decisions on the ranking or priorities of trains are
related exactly to the route nodes.
The following list contains infrastructure attributes, sorted by importance for correct modelling
of infrastructure for the calculation of running times, timetable construction and simulation:
– Length of a link
– Gradient
– Permissible speed (variations for different train types)

55
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
3 Infrastructure Modelling

– Speed indicators and speed boards

– Electrification and different electrical systems
– Radius
– Signalling system (M/P, ATC, LZB, ETCS L1 to L3, various multipleaspect signalling)
– Overlaps
– Release contact and clearance location
– Track circuits
– Stop boards
– Blocks and routes
– Interlocking technique
– Exclusion of routes
– Availability of infrastructure
During recent years, the modelling of infrastructure has been extended to include the
availability (time stamp) of infrastructure elements. These time stamps can be assigned to a
single group or groups of nodes and links of a microscopic infrastructure model. It is evident
that with these time stamps in the infrastructure model, it is straightforward to automatically
generate the matching model of microscopic infrastructure and the related timetable. This
uses operating days in a very flexible way for a certain time period (day, week etc.) and is
sourced from the same database. Therefore, it is possible to maintain only one infrastructure
database to generate short term infrastructure for ad-hoc timetabling, as well for long term
infrastructure planning tasks.
Microscopic models have been used for more than 25 years in Germany and other European
countries for the computer based operational planning of networks, lines and junctions. From
the experience gained, it is clear that in order to obtain reliable results the interrelation of
the infrastructure, the rolling stock and the timetable should be assessed on a microscopic
network level, which includes all lines and stations (network wide, all tracks). The complexity
considered can include various signalling systems, different vehicle dynamics and train
scheduling rules, as well as additional operational conditions. All these issues are – in some
way – dependant on the reliable modelling of the railway infrastructure.
Furthermore, working in complex infrastructure networks (total or partial) is becoming
a standard approach. For example, the identification and evaluation of reliable capacity
indicators for railway lines requires the inclusion of at least the adjacent junctions and in a
better way the consideration of a (partial) network. It became apparent during the practical
application of timetable construction, possession planning and/or railway simulation methods,
that all branch lines from those junctions should be considered in the network model as
well, to ensure that the timetable construction, possession planning and/or simulation could
incorporate the impact of trains arriving at the junctions from those branch lines.

3.5 Differences between Infrastructure Models

The differences between microscopic and macroscopic models will be explained in the
following example. In Fig. 3.8 a small section of a railway line has been illustrated in both. The
vertical axis depicts the speed limit over the section; the horizontal axis describes the length of
the section. In total, the 8,600 m are separated by five microscopic links, which are different in
length.
As is typical in microscopic models, there is a new node at the exact position where an
attribute (here the speed limit) has been changed. In order to keep the example simple, all
other attributes have been removed from the illustration. The macroscopic model contains only

56
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 3.6 Migration between Infrastructure Models

Fig. 3.8: Relationship between macroscopic and microscopic infrastructure models

one link (merging the five microscopic links to one macroscopic), joining the start and end
node. The problem of assigning a speed limit to the macroscopic link is obvious. There are the
following options:
– Lowest microscopic speed limit (80 km/h)
– Highest microscopic speed limit (100 km/h)
– Average speed limit considering the proportionate length of the microscopic links
(96.51 km/h): the red line in Fig. 3.8 labelled as “macroscopic speed”
A comparison of running time calculations using the three options of macroscopic speed limits
with the microscopic speed limit showed variations between + 6 % (too fast) and – 20 % (too
slow), see Radtke (2006). This demonstrates why macroscopic models should not be used for
running time calculations, timetable construction or conflict detection in the planning process.

3.6 Migration between Infrastructure Models

A simplified example in Fig. 3.9 illustrates a transition between different infrastructure models.
The microscopic model generally contains far more links and nodes. These nodes are
symbolised by scale lines. In a “Bottom Up” approach, it is evident that all macroscopic
models can be derived directly from microscopic models, as they consist of less information
than microscopic models. The “Top Down” approach can be used for generating (artificial)
microscopic models, whose accuracy of detail depends on predefined rules and assumptions.
The “mesoscopic infrastructure model” is placed in-between both models. The model can be
generated for specific tasks, such as the simplified railway “simulation” of complex networks
to answer some strategic questions. Another field of application is a transfer of macroscopic
infrastructure information into a standardised microscopic infrastructure model. These
standardised models, in a further step, certainly require more input (e. g. gradient information
and speed profiles) to be acknowledged and used as complete microscopic models.

57
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
3 Infrastructure Modelling

Fig. 3.9: Multi scale infrastructure models

“Multi-scale” models contain both areas that are modelled on a microscopic level and areas
that are modelled on a macroscopic level. The advantage of such models is the minimisation
of effort for modelling aspects of a complex problem that are of insignificant relevance for the
overall outcome of the investigation (e. g. the individual operation of shunting yards or vehicle
depots for the planning of a network wide timetable).

Fig. 3.10: Mesoscopic infrastructure model

In Fig. 3.10 another approach of mesoscopic railway modelling can be seen. The first element
is the automatic generation of (artificial) microscopic data along a line with standard block
sections or headways. Furthermore, the modelling of the correct length between stations and
– with reasonable effort – between signals, could be achieved. However, the correct location of
signals and block length can only be modelled in microscopic models and, therefore, reliable
capacity figures are not achievable with mesoscopic infrastructure models.

58
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 3.6 Migration between Infrastructure Models

The second element of this approach is the more detailed depiction of the station (here called
a “mesoscopic station”). In comparison with a pure macroscopic model where stations are
modelled as a single node, here the individual station tracks and possible paths from the lines
to the station tracks can be modelled. It might be necessary to increase the level of detail of
multi scaled infrastructure models for specific applications. This can be achieved without any
problems if the correct microscopic data is available, see Radtke (2006).
The combination of microscopic and macroscopic models can be illustrated with the following
example of a successful implementation in tpractice. A microscopic timetable construction and
railway simulation model (here RailSys®) and a macroscopic Network Evaluation Model (here
NEMO) allow the generation of future timetables under consideration of traffic forecasts. This
approach allows the application of microscopic infrastructure elements in the process of long
term planning. Austrian Railways (ÖBB) achieved the successful combined application of both
models in 1999, see Sewcyk (2007).

Fig. 3.11: Transformation of microscopic into macroscopic network

Fig. 3.11 illustrates the transformation of a partial microscopic network into a macroscopic
model. All stations and junctions are transformed into a macroscopic node. These nodes are
connected with only one link. In the microscopic model on the left side of Fig. 3.11 however, a
different range of tracks can be seen:
– Single line
– Double lines
– Triple lines
Furthermore, all station tracks and all signals have been incorporated into the macroscopic
links and nodes. This approach of generating macroscopic infrastructure models out of a well-
maintained microscopic database has some significant advantages:
– Maintenance of only one corporate database, including the time stamps
– Provision of the same infrastructure data for various applications
– Reduction of redundancies during data entry

59
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
3 Infrastructure Modelling

– Utilisation of microscopic data for macroscopic models

– Correct running time calculation for the assignment of trains
– Calculation of minimal headways on lines and in stations
– Calculation of capacity figures (e.g. according UIC 406)
– Flexible data model to define and implement model enhancements for both directions
– Possibility of communicating between the models with modern technology (web services,
client server processes etc.)
As a result, the quality of results of macroscopic planning has increased significantly in recent
years.

3.7 Application of Infrastructure Models

The classification of when to apply the different infrastructure models can be seen in
Fig 3.12. The level of detail varies between low and high. Therefore, the effort to establish an
infrastructure database correlates with the level of detail.

Fig. 3.12: Application of infrastructure models

The macroscopic model can be applied for high-level tasks like vehicle scheduling, or the
planning of traffic demand and allocation. For both tasks, the route search is one of the
key elements. On the contrary, running time calculation, timetable construction, possession
planning and railway operational simulation require detailed infrastructure models.
The typical workflow to establish a microscopic infrastructure from scratch can be described in
the following way:
– Organising the raw data
– Quality check of that raw data
– Establishing the model in steps
– Definition of the area of investigation (network, station, lines)
– Main tracks, including length, gradient, speed, signalling systems, location of signals,
interlocking, time stamps)
– Main station tracks

60
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 3.7 Application of Infrastructure Models

– Sidings
– Points of interest (timing points, stopping boards, platforms)
– Additional points of interest (evaluation, dispatching)
– Shunting yards and depots
– Quality check of infrastructure model
– Visual check by experienced planners
– Compare running times of modelled trains with reality (using verified rolling stock data)
– Integration of database maintenance into the business process
The problem of third parties obtaining access to microscopic infrastructure data has to
be discussed at this stage. Today, it is not yet common practice that railway undertakings
possess and maintain microscopic infrastructure data of their entire network in a single
database. However, there is some progress in various countries around the world. Depending
on the policy of railway undertakings in their business processes, the range of access to that
data for third parties varies from very restricted to comparatively control free access.
In Fig. 3.13 a benchmark for the levels of importance of the different steps between
microscopic and macroscopic models can be seen. Macroscopic models hold advantages
in respect of the first four criteria. However, applying an established and well-maintained
microscopic database, those criteria may not be the most dominant points. The last criterion,
“accuracy of results”, is much more important when compared against all other criteria.
From previous experience, the value of correct results during the planning process should be
highlighted.

Fig. 3.13: Benchmark of microscopic and macroscopic network models

Microscopic infrastructure models have been used successfully in various countries around
the world. In countries such as Germany, Austria, United Kingdom, Sweden, Denmark or
Switzerland, microscopic infrastructure network models have been used since the beginning
of the 1990s. Other railway companies or government bodies (e. g. the European Union) have
started initiatives to follow this approach in the future.

61
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
3 Infrastructure Modelling

3.8 Outlook

The future of the utilisation of railway infrastructure models will probably include the following
three trends:

– Enhancement of railway infrastructure models with further attributes, time stamps and
functionality. Current research work is going to establish a link between railway planning
data with cost-orientated asset data or alignment data. This includes, for example, the
consideration of Life Cycle Costs of infrastructure asset elements and a financial evaluation
of infrastructure and train operation in respect to budget spending.
– The development of standard interface formats will help the exchange of infrastructure data
across railway undertakings or to third parties. However, due to the complexity of railway
infrastructure models in comparison with timetable exchange formats, there will be still a lot
of work necessary to define a practical “standard” infrastructure format in the near future.
For timetable information, an exchange format like RailML® or other XML-based formats
could be a reasonable approach. Nevertheless, even in timetable formats, there is still some
work necessary to enable an interchange of data without further interaction. Currently, there
are talks regarding a kind of certification of both interfaces for infrastructure and timetable
data based on the RailML® approach.
– Linking railway planning infrastructure data with Geographical Information System (GIS)
data (e. g. OpenStreetMap®). Railway planning systems should be able to support all
common graphic formats (BMP, JPEG, TIFF, etc.) and to set multiple images with individual
scale and position functionality. The link between both modelling approaches will enable
planners to better present complex results to decision makers and the public. Fig. 3.14
illustrates a simple example showing a train run in a Google Earth® map of Austria.

Fig. 3.14: Train route interface shown with Google Earth v4®

62
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

– Generating infrastructure models from public sources. These technologies may be used by
third parties who – for whatever reason – do not have access to the railway undertaking
infrastructure databases:
– Network statements of railway undertakings from the internet
– Electronic maps
– Internet maps (e. g. OpenStreetMap® or Google Earth®)
– Printed public maps
– GPS and video technology (collecting data by using trains)
– Establishing Corporate Network Models (CNM) through the integration of railway planning
data with asset management data, geographical information, as well as data from the
engineering, control and other departments.
The microscopic infrastructure for railway planning will be useful as an initial step in
the creation of the CNM. The main challenge at present is to maintain the microscopic
infrastructure model for purposes such as running time calculation and timetable planning.

References
Dijkstra E. W.: A note on two problems in connexion with graphs. Numerische Mathematik.
1 (1959), pp. 269 – 271.
Hauptmann D.: Automatische und diskriminierungsfreie Ermittlung von Fahrplantrassen in
beliebig großen Netzen spurgeführter Verkehrssysteme (Automatic and non-discriminatory
train path allocation in railway networks of arbitrary size). Dissertation am Institut für
Verkehrswesen, Eisenbahnbau und -betrieb (Schriftenreihe Nr. 54), Universität Hannover,
Hestra Verlag 2000.
Kettner M.: Netz-Evaluation und Engpassbehandlung mit makroskopischen Modellen des
Eisenbahnbetriebs (Network evaluation and bottleneck handling with macroscopic models
for railway operation). Dissertation am Institut für Verkehrswesen, Eisenbahnbau und -betrieb
(Schriftenreihe Nr. 65), Universität Hannover, Eurailpress 2005.
Radtke A.: EDV-Verfahren zur Modellierung des Eisenbahnbetriebs (Software tools to model
the railway operation). Habilitation am Institut für Verkehrswesen, Eisenbahnbau und -betrieb
(Schriftenreihe Nr. 64), Universität Hannover, Eurailpress 2005.
Radtke A., Watson R.: Railway simulation in the United Kingdom. RTR 1 (2007), pp. 24 – 28.
Schumacher A.: Ein hybrides Verfahren zur Umlaufplanung von Fahrzeugen des spurgeführten
Verkehrs (A hybrid method for vehicle allocation in rail traffic). Dissertation am Institut für
Verkehrswesen, Eisenbahnbau und -betrieb (Schriftenreihe Nr. 60), Universität Hannover,
Eurailpress 2004.
Schwanhäußer W.: Die Ermittlung der Leistungsfähigkeit von großen Fahrstraßenknoten und
von Teilen des Eisenbahnnetzes. AET, vol. 33 (1978), pp. 7 – 18.
Sewcyk, B:. Makroskopische Abbildung des Eisenbahnbetriebs in Modellen zur langfristigen
Infrastrukturplanung (Macroscopic display of railway operation in models for long-term
infrastructure planning). Dissertation am Institut für Verkehrswesen, Eisenbahnbau und -betrieb
(Schriftenreihe Nr. 61), Universität Hannover, Eurailpress 2004.
Sewcyk B., Radtke A., Wilfinger G.: Combining Microscopic and Macroscopic Infrastructure
Planning Models. Proceedings IAROR Hannover 2007.
http://www.openstreetmap.org/ Website of OpenStreetMap®, 31 January 2014
http://www.railml.org/ Website of RailML®, 31 January 2014

63
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.1 Introduction

4 Running Time Estimation

Olaf Brünger, Elias Dahlhaus

4.1 Introduction

4.1.1 Fundamentals

Trains are run to a schedule, so why is there a need to estimate their running time? The
physical process of accelerating a train is quite approximate in the time it takes, depending as
it does on the friction between wheel and rail, the overhead line voltage, and the actions of the
train driver. But it is even necessary to calculate running times for very short distances using
precise numerical calculations (with great attention afforded to the exactness of the arithmetic
and counting the seconds).
When compiling a timetable, it is necessary to estimate running times for different train
configurations using the current infrastructure, which is normally different from last year’s
infrastructure. It may even change within the timetable period. Examining the capacity of
nodes and routes, constraints like minimum headways are considered. These are derived from
the speed and time at which the trains approach and clear the critical block signals and track
sections. Normally, these times are not given in the timetable, so they have to be calculated.
A linear interpolation derived from departure and arrival times is wrong, so it has to be
recalculated based on how the train accelerates, runs, and brakes. Accurate running times are
used to achieve good dispatching results or to save energy by using margins in the timetable.
Most modern tools for analysis or timetable construction calculate running times on demand
for the train concerned. In more simple circumstances, it is enough to prepare tables of
running times or headways for certain train configurations and for different combinations of
station stops and passing through non-stop. (Note that in preparing these tables, very detailed
infrastructure and train data are also needed!)
Many influences on running times are not deterministic, such as human behaviour, weather,
train composition, infrastructure works, the rolling stock, and route condition. To cover this,
random influences are taken into consideration through recovery times, and the basic
calculation is done with standard values representing good conditions. Another approach is to
use probabilistic approaches as described in a later section of this Chapter.
However, some information is fundamental for estimating running times:
– route details with distance, speed limits, and gradient (and curves and tunnels),
– traction unit characteristics, with tractive effort and resistance,
– rolling stock characteristics, with mass, length, and resistance (mainly dependent on the
type of train), and
– given data of the operating cycle such as starting point, stops, and – as appropriate –
timetable restrictions.
Dependent on the safety devices used, information about brake effort or braking supervision
might also be necessary.
Historically, running times were estimated using manual graphic algorithms or mechanical
instruments like a Conzen-Ott-Machine (a mechanical analog integrator for running time
estimation) as shown in Fig. 4.1, which was capable of taking these values into account. This
book concentrates on running time estimation using computers. Reference may be made to
special literature as Wende (2003) for the more detailed aspects.

65
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

Fig. 4.1: Conzen-Ott-Machine for Estimating Running Times

4.1.2 Speed Profile

The speed profile of a train on a given route is often visualised in a speed/distance diagram. It
is useful to add a gradient/distance diagram to see how that influences the whole. The Authors
examine a typical speed profile as shown in Fig. 4.2, where the principal order of acceleration,
constant movement, braking, and coming to a stop can be seen.

Fig. 4.2: Sample Speed Profile

1. The horizontal axis shows the route length measured in kilometres. Both diagrams have
identical divisions.
2. The vertical axis of the speed/distance diagram shows the velocity measured in km
per hour (km/h) and metres per second (m/s). Both the maximum permitted and actual
speeds are shown in the same diagram, which is concerned with the front of the train.

66
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.1 Introduction

3. The vertical axis of the second diagram shows the gradient of the line measured per 1,000
– positive values for ascents, negative for descents. Note that railway gradient angles are
very slight, so it makes only a negligible difference whether the horizontal measurement is
truly so, or along the length of the track (tan  ≈ sin ).
4. Maximum speed of the train is 160 km/h.
5. The speed limit given by infrastructure constraints changes several times on the route. It
starts with 100 km/h and rises later to 140 and then to 200 km/h.
6. At km 15 there is a station where entry speed is reduced to 60 km/h. This is often caused
by restrictions on points. Note that the difference in running time estimation caused by
turnout speed is dependent on the actual track used by the train. In a detailed analysis
or timetable construction this difference has to be considered, as it can easily reach a
minute, or more.
7. Exit speed from the station and track is reduced to 40 km/h. Later, the train is allowed to
run at 100 km/h again.
8. At km 28 there is a special restriction to 80 km/h. There are many possible reasons. Work
sites with speed restrictions (for all trains), defects in the contact wire (for electric trains
only), tunnels (often for passenger trains only), wind danger (for empty freight trains). Also,
special increases of the speed limit for several trains exist, for example for tilting trains.
When calculating very precisely, there are also dependencies between different trains, so
that running time estimation is not for a single train only. For example, if the meeting of
trains in a tunnel is restricted to a certain (relative) speed limit for aerodynamic reasons, or
if the exit speed of a train from a station is dependent on the progression of the preceding
train.
9. In the other diagram, two small down gradients of 5 and 8 per 1,000 can be seen.
Dependent on the braking efforts of the train and the gradient characteristics of the line,
there might be special speed reductions.
10. There are also two ascents, one of which is quite steep (40 ‰, a typical value for
suburban train operations).
11. Considering the estimated speed of the train now. At first, it accelerates from 0 until it
reaches the actual speed limit of 100 km/h.
12. It remains at the speed limit by constant movement.
13. When the speed limit rises, it accelerates again to the new speed limit of 140 km/h. Note
that it starts accelerating a little later: Speed limits refer to the whole train, and the head
of the train is marked. The train driver (and the simulating computer model) has to wait to
accelerate until the back of the train has passed the point.
14. Constant movement.
15. Acceleration up to the speed limit of this train which cannot make full use of the speed
allowed for the line.
16. Constant movement.
17. The train starts braking. The position is dependent on the following.
18. At the point where the entry speed restriction takes place, the train must already be
running at the lower speed. The train driver (and the model) has to estimate where to start
braking to achieve this. This is done by calculating the position of intersection (no. 17)
between the braking curve and that of the constant movement beforehand. Of course in
practice this does not always occur exactly, but as mentioned before good conditions are
assumed when estimating running times (and the recovery times added subsequently).
19. Constant movement (in the entry section of the station).
20. Braking to stop. (When estimating complete scheduled times used by the train, the dwell
time has to be added here.)

67
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

21. Acceleration to the exit speed.

22. Constant movement.
23. Acceleration to the speed limit of the line, then constant movement again.
24. On the uphill gradient, the train loses velocity: The effort of the train does not compensate
fully for the sum of resistances, including the line resistance of the steep incline. As will
be seen later, the tractive effort increases as speed reduces, and resistances have the
opposite behaviour. Thus there may be a velocity where the train moves at constant
speed as in the example; otherwise, it would reduce speed until it stops on the incline.
25. When the top of the hill has been reached, the train accelerates again. Note that the effect
does not start with the head of the train but a little later. In a very detailed calculation, the
train is regarded as a (homogeneous) strip, whereas normally the impact of the gradient is
just shifted by half of the train length.
26. Because the special speed limit is valid for this train, it has to reduce its speed by braking.
It continues by constant movement and acceleration later.
This diagram shows no coasting (run-out) section of the train movement, because the
calculation is for the shortest running times in good conditions and recovery times are added
later. Of course in practice the train driver will make use of reserves by reducing power (while
the resistances continue), before he has to brake for a stop or a speed limit reduction. In a
computer model (like that implemented in the electronic display of energy saving driving in
trains of Deutsche Bahn), the energy-optimal position of starting run-out can be calculated,
dependent on the amount of recovery time left. Chapter 5 deals with this topic.

4.2 Infrastructure and Train Data

In the last section, it was seen in principle how the train behaves on a given route. To
calculate running times in the different sections, it is now necessary to look at the influences of
infrastructure and train, especially for calculating the acceleration section of the speed profile.
It is useful to look at four different components:
– Tractive effort (FT )
– Traction unit resistance (FRt )
– Rail vehicle or wagon resistance (FRw )
– Infrastructure data, especially line resistances (FRl )
The difference between tractive effort and the sum of the resistances is the effort available for
accelerating the train.
Note: the following formulas use the physical quantity with the SI-dimensions, including
– velocity v measured in [m/s] (3.6 v is the value in [km/h]) and
– mass m measured in [kg].
The quantity g is the earth gravity constant 9.81 m/s².

4.2.1 Tractive Effort FT

The locomotive or the power equipment of the multiple unit generates effort which is intended
to move the train. It is called induced tractive effort FTi [N].
However, not the whole amount of this effort can be used, due to three principle reasons:
– The internal power transmission of the locomotive or the multiple unit consumes between
2 % and 3 % of the effort.

68
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.2 Infrastructure and Train Data

– The effort can be limited to a constant maximum amount to prevent the power equipment
from overheating.
– The wheels will spin if the effort exceeds the maximum adhesion between wheel rim and
rail.
The last topic is described by the adhesion value µ (a factor without dimension, dependent
on driven speed v [m/s]) and the wheel load FL [N] of the driven wheels. The product of both
is the maximum effort which is transmitted without spinning wheels. In first instance µ may be
determined by the formula of Curtius and Kniffler (1950):
µ (v) = 7.5 / (3.6 v + 44) + 0.161

Concluding, the induced tractive effort is reduced by the transmission losses and limited by
traction motor overheating and the adhesion limit. Additionally, some (older) power units have
different characteristics, so that the effort cannot be deduced directly from the engine power.
Instead, the tractive effort at wheel rim FTr [N] (depending on speed v ) is normally given by a
diagram which often looks like that shown in Fig. 4.3.

Fig. 4.3: Tractive Effort at Wheel Rim Dependent on Speed

To use this diagram in running time estimations, it is normally described numerically in a group
of hyperbolic or parabolic formulas, each of which approximates to the actual tractive effort at
wheel rim at speed v [m/s] for a certain speed interval from vk to vk+1:
FTr (v) = c0,k + c1,k · v + c2,k · v2 , vk < v ≤ vk+1 (4.1)
FTr (v) = ch,k / v , vk < v ≤ vk+1 (4.2)

The tractive effort at wheel rim in this description is given by the factors c0,k [N], c1,k [N · s/m]
and c2,k [N · s²/m²] or ch,k [m/s/N] and the limits vk [m/s] (all together for a few k = 1, 2,…) is
one of the input parameters for the running time estimation.

69
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

4.2.2 Vehicle resistances

The locomotive and the vehicles or wagons of the train consume some of the power of the
engine because of
– air resistance (with quadratic dependence of the train velocity relative to the wind),
– rolling resistance caused by wheel rims, axle-boxes, adhesion, and similar reasons (parts of
which are constant, and parts which have approximately linear dependency on the velocity
v [m/s]).
In practice, these resistances are generally described by parabolas with coefficients ri which
depend on the train characteristics and the wind speed:
FR (v) = r0 + r1 · v + r2 · v2 (4.3)

As for a special train configuration, these coefficients need to be calculated from the data
known about the train. There are heuristic formulae to estimate the resistances from the train
characteristic as described in the following sections.
Note: The formulae mentioned here are those often used in middle Europe under typical
circumstances; many railways use other approximations of physical reality or have to cater for
different circumstances.
(In this section, the dimensions of the constant parameters are omitted. They have to be
transferred if speed or mass are used in different dimensions.)

Traction Unit Resistance FRt

For the traction unit (including multiple units), the resistance FRt is usually described with given
parameters a0 , a1 and a2 or a2r for the following formula:
FRt (v) = g · mT · ( a0 + a1 · v ) + a2 · v2 + a2r · vr2 (4.4)
with
– the mass of the traction unit mT [kg],
– the speed of the vehicle v [m/s],
– the relative speed between air and vehicle vr [m/s] usually assumed as v + 4.17 m/s (i. e.
headwind of 15 km/h).

Vehicle Resistance for Passenger Trains FRwp

For passenger trains, the parameters of formula (4.3) are mainly described by
– the mass of the vehicles mW [kg],
– a factor cb for the number of axles which can be assumed as 0.0025 for vehicles with 4
axles, 0.004 for those with 3 axles and 0.007 for those with 2 axles,
– the number of vehicles nW ,
– and a value Af [m²] which stands for the cross-sectional area of the vehicles weighted with
their aerodynamic behaviour. Normally, this value is assumed as 1.45 m².
Using the formula of Sauthoff (as an example for other approximations), the relationship
between the vehicle speed v [m/s] and the passenger vehicle resistance FRwp [N] has
theoretically and experimentally been determined as follows (leaving out a more detailed
splitting of the constants into several physical influences), see Sauthoff (1932):
FRwp (v) = 1000 mW · g · ( 1.9 + cb · 3.6 v ) + 0.0471 · (nw + 2.7) · Af · (3.6 vr)2 (4.5)

70
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.2 Infrastructure and Train Data

Vehicle Resistance for Freight Trains FRwf

For freight trains the formula of Strahl (1913) can be used to approximate the vehicle
resistance FRwf [N] dependent on the vehicle speed v [m/s]:
FRwf (v) = 1000 mW · g · (ca + (0.007 + cm) · (3.6 v)2/100) (4.6)
with the following parameters:
– the mass of the wagons mW [kg],
– a coefficient for axle adhesion ca (use 1.4 for roller bearings and 2.0 for older plain-bearing
axle-boxes),
– and a value for air resistance cm dependent on the kind of wagons:
– cm = 0.05 for mixed trains,
– cm = 0.032 for full train loads of coal or ore,
– cm = 0.04 for closed wagons,
– cm = 0.1 for empty open wagons.

4.2.3 Line Resistances FRl

Obviously, the main influence is the gradient of the line. The weight force on a slope of angle 
is described by
F = m · g · sin 

The gradients of railways are very slight, so sin  is similar to tan , which is usually measured
as n in per thousand [‰]. Measuring the mass m of the complete train in kg, the gradient line
resistance FRlg [N] comes to
FRlg = g · 1000 m · n (4.7)

Line resistance FRlc is also caused by sharp curves and is described approximately by
FRlc = g · 1000 m · 700 / r (4.8)

with the radius r [m] of the curve. Note that it has to be left out on (nearly) straight lines, and
note also that a curve of 700 m radius generates the same resistance as a slope of 1 ‰, so
the error caused by leaving out FRlc is relatively small.
Lastly, the influence of the air resistance as a function of the cross section and speed in
tunnels should be included. However, there is no formula of general acceptance, so it is left
out here. The most significant effect is caused by trains meeting each other in a tunnel, but
information on this is often not available when estimating the running time.

4.2.4 Combining Efforts and Resistances, Rotating Masses

The different resistances described can all be combined in a parabola formula dependent on
velocity as in (4.3) with some easy binomical transformation, if the characteristics of train and
line are known. The difference between tractive effort at wheel rim FTr and resistances FR is left
for accelerating the train.
However, the train contains some rotating parts which consume some of the effort. The usual
way to take care of this effect is a mass factor f for each part of the train. For the traction unit,
the factor fT should be given with the engine data (if not, take 1.09); for passenger vehicles
and freight wagons, fW ≈ 1.06. For the whole train, it comes out to
f = (fT · mT + fW · mW) / (mT + mW) (4.9)

71
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

Last, calculating the acceleration of a train is the same as solving the differential equation
FTr (v) – FR (v) = f · m · dv/dt (4.10)

with train mass m [kg] and speed v [m/s] (t in seconds).

It needs to be remembered that FTr is quadratic and linear or hyperbolic dependent on v, and
FR is quadratic and linear dependent on v.
The estimation of running time consists mainly of three tasks:
– Interpreting all given data of train and infrastructure
– Partitioning the route into pieces of equal characteristics and behaviours
– Solving the differential equation (4.10) for those places where the train accelerates

4.3 Moving Sections

As already seen in the speed profile in Section 4.1, the train moves in the order of acceleration,
constant movement, coasting, and braking. This order is followed in the running time
estimation along the route. The movement is determined by tractive effort, the resistance, and
of course by the route and timetable restrictions the train is intended to use. So the moving
sections are formed by two influences:
– Characteristic sections caused by changes of infrastructure or by timetable restrictions,
especially stops
– Behaviour sections formed by the movement of the train

Fig. 4.4: Moving Sections

72
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.3 Moving Sections

When partitioning the route into pieces of equal characteristic and behaviour, both speed and
length have to be considered:
– The tractive effort is given by several formulas, each of them valid only for a certain speed
interval, and acceleration of course has to finish when reaching the maximum velocity.
– On the other hand, changes of gradient and speed limit are given along the line at certain
kilometre positions.
So an algorithm must be able to calculate movement parts described by either speed or length
limits. In addition, some algorithms which solve the differential equation by approximation make it
necessary to divide the remaining steps again into small intervals of speed or length.

4.3.1 Characteristic Sections

Of course the positions where the train starts and stops have to be known.
If possible, the detailed route including the choice of tracks and the location and type of
points in stations should be given, because they cause different entry and exit speeds. When
collecting the speed limits, the design speed limit of the line and train have to be observed as
well as the special speed limits for the train as mentioned above. In some cases overspeed
control of the signalling system has to be considered, or the effects of low braking forces
(brake ratio deficiency) of the train at:
– short signalled braking distances, or
– slopes with ruling down-gradients.
The design of a computer model has to take care of breaks in the kilometrage and of the
situation at the start position of the train (for instance concerning the speed limit there). If the
head of the train is regarded as representative of the whole train, the length of the train can be
considered as follows:
– Changes of gradient and curve have effect on the whole train, so they will be shifted for
approximately half the train length in the forward direction.
– The speed limit must be observed for the whole train length, so a less severe speed limit is
shifted for the train length in forward direction.

4.3.2 Behaviour Sections

Acceleration section
In the acceleration section, all the available tractive effort is used as the shortest running times
calculated. The various resistances work against it. The different efforts and the acceleration
are not constant, so the simple use of the Newton formula for constant acceleration is not
appropriate. This has to be solved by the differential equation (4.10). Sections 4.4 to 4.7 will
deal with this.
Constant Movement (Cruising)
When the train has reached its speed limit and changes to constant movement, the simple
Newton formulae can be used for constant movement, as long as effort minus resistances
remains positive:
s=v·t (4.11)

with way s [m], velocity v [m/s] and time t [s]. Note that by reaching a gradient section the
resistances can increase, and it might be necessary to change the calculation to an
acceleration section with negative acceleration. (Downhill it can happen that the brakes have

73
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

to be used to restrain the maximum velocity. As long as the braking effort is sufficient, this has
no effect on running time estimation.)
Sometimes constant movement is calculated as a sequence of acceleration and run-out in a
small interval around the maximum speed.
Coasting section
If considering the coasting or run-out section as well, it is calculated like an acceleration
section without effort.
However it depends on the purpose of the calculation whether a coasting section is intended
at all, and how the start position is to be determined. Normally the coasting section is left out
when calculating for timetable construction, because the shortest running times are used with
certain recovery times added.
Braking section
Estimating the running time in the braking section is very controversial. Theoretically it can be
considered as an acceleration section replacing the tractive effort by the braking effort of the
train (note that the vehicles contribute much of this effort). Some other arguments however
must be taken into consideration:
– Under normal circumstances the train does not use its full braking effort, which is reserved
for an emergency stop. Instead, the train driver brakes in a comfort mode.
– Besides, when constructing a timetable, the locomotive type is more or less known.
However, information about all the vehicles with their braking effort is still lacking at that
stage.
– Last but not least, some safety systems in normal use just calculate with (segments of)
constant braking effort (like ETCS or German continuous train control system LZB).
The choice is constant braking for running time estimation, and so the formula for constant
acceleration ab [m/s2] can be used:
s = ½ ab · t2 (4.12)

Normally the following typical values for braking acceleration ab are used unless no more
detailed values (like “Minden formula” or others) are given:
– ab = 0.525 m/s2 for suburban trains (service braking)
– ab = 0.375 m/s2 for passenger trains (service or comfort braking)
– ab = 0.225 m/s2 for freight trains (service braking)
– ab = 0.7 m/s2 for suburban trains (sharp braking)
– ab = 0.5 m/s2 for passenger trains (sharp braking)
– ab = 0.3 m/s2 for freight trains (sharp braking)
If the train is fitted with an automatic system like LZB, the braking acceleration the system
applies for the train needs to be considered. Often service braking is 0.75 times the value for
sharp braking.
Note: Depending on the purpose of the running time estimation. much more detailed
calculations have to be made. This explanation does not satisfy safety requests.

4.4 Modelling Running Time Calculations

An objective might be to determine the minimum running time for a certain route. So there is
no coasting phase. The characteristic sections as described in Section 4.3.1 are provided. In
each of these sections, the gradient and the maximum speed are constant.

74
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.4 Modelling Running Time Calculations

4.4.1 Determination of Braking and Acceleration Sections

The first step is to determine the braking sections. Note that if the maximum speed of the
next characteristic section is less than the speed of the section just considered, the speed
reduction has to be completed by the time the train reaches the next section. For any
characteristic section, it is assumed that the maximum exit speed vex is given. Mostly, the
maximum exit speed is the maximum speed of the forthcoming section. If the maximum exit
speed is less than the maximum speed of the section, the procedure is as follows.
1. First, determine the braking distance sred to reduce from the maximum speed vmax of the
section to the maximum exit speed vex. Remember that the braking deceleration is denoted
by ab. Then the braking distance is just
sred = vmax2/(2ab) – vex2/(2ab) (4.13)

If the braking distance sred is shorter than the length of the section, it is split into two
sections. The first section is a constant maximum speed section where vmax applies, and
the second section is a braking section where the maximum entrance speed ventr is vmax
and the maximum exit speed is the maximum exit speed vex of the whole section. The
length of the second section is sred.
2. If the braking distance sred is at least as long as the length of the section, the maximum
entrance speed ventr of the whole section has to be determined. This is the maximum
permitted speed at the beginning of the section to reduce to vex at the end of the section.
Denote the length of the section by ssec. Then the maximum entrance speed is determined
as follows:
ventr = √2ab ssec + vex2 (4.14)

If the maximum speed of the previous section is greater than the maximum entrance speed,
the maximum entrance speed becomes the maximum exit speed of the previous section.
For this reason, it is recommended that the braking sections are determined in reverse
order, i. e. starting with the last section of the route and then considering the second last
section, and so on.
The route for which the running time is to be estimated is now partitioned into sections as
follows. Each section is either a constant maximum speed section or a braking section. In the
first case, the maximum entrance speed and the maximum exit speed are equal. In the second
case, the maximum exit speed is less than the maximum entrance speed.

Speed

train
position

Fig. 4.5: Braking Curves Derived from Speed Restrictions

The next step is to determine the acceleration phases. When a section is entered at a certain
speed and this speed is less than the maximum entrance speed, an acceleration phase is
entered. There are three possibilities.

75
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

1. The section is a constant maximum speed section and the maximum speed is reached in
this section.
speed

entrance
speed

train position

Fig. 4.6: Acceleration Phase, Exit Speed is Identical to the Maximum Speed

2. The acceleration phase is through the whole section and the exit speed is less than the
maximum exit speed (here it does not matter whether the section is a constant speed
section or a braking section).

Fig. 4.7: Acceleration Phase, Exit Speed is Less Than the Maximum Exit Speed

3. The section is a braking section and a braking phase follows immediately after the
acceleration phase.

Fig. 4.8: Acceleration Phase in a Braking Section

76
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.4 Modelling Running Time Calculations

The exit speed from one section always becomes the entrance speed of the next section.
As mentioned before, the relation between time and speed and the relationship between
speed and the length of the acceleration path are expressed by differential equations. The train
can be modelled as a mass point, or as a homogeneous strip.

4.4.2 Modelling the Train as a Mass Point

As mentioned in equation (4.10), the relation between time and speed during the acceleration
phase is expressed by the following differential equation:
FTr (v) – FR (v) = fp · m · dv/dt (4.15)

The speed v is given in m/s and the mass m is given in kg. The traction force FTr and the
resistance FR are considered as functions in v[m/s]. Also of interest is the relation of the speed
v and the length s of the acceleration path. Since v = ds/dt, the following differential equation
expresses the relation between v and s:
(FTr (v) – FR (v))/v = m · fp · dv/dt · dt/ds = m · fp · dv/ds (4.16)

Example
As an example, consider a diesel multiple unit set (DMU) with a tractive effort given by the
following table, where coefficients c0,k, c1,k, and c2,k are the coefficients of the polynomial
expressing the tractive effort as in equation (4.1) .

Speed interval in km/h c0,k[kN] c1,k[kN h/km] c2,k[kN h2/km2]

0 – 11.4 128.4644 -1.47513 -0.0045413
11.4 – 30 132.2741 -2.12073 0.00106938
30 – 60 132.4702 -2.16075 0.01200450
60 – 79,8 105.5959 -1.50635 0.00650599
79.8 – 99 20.0891 0.301952 -0.02009088
99 – 117 50.5504 -0.155587 -0.00046380
117 – 140 -4.107 0.483731 -0.00210228
140 – 160 35.0809 -0.025005 -0.00050031

The speeds are given in km/h, the coefficients apply to the speeds in km/h, and the resulting
tractive effort is the value in kN. If, for example, v is 70 km/h then the tractive effort is
105.5959 – 1.50635 · 70 + 0.00650599 · 702 kN.
The weight of the power car is 116 tonnes. The resistance of the DMU is assumed to be
1.39 · m[t] · g + 0.00905 · m[t] · g · v[km/h] + 0.0296 v[km/h]2 measured in Newton.
There is a need to determine the time taken to speed up the DMU from 40 km/h to 70 km/h.
The gradient is assumed to be zero.
First, the speeds have to be transformed into m/s. Since 1 km/h = 1/3.6 m/s and therefore
1 m/s = 3.6 km/h, the interval borders have to be divided by 3.6 to get them as speeds in
m/s. c1,k has to be multiplied by 3.6, and c2,k has to be multiplied by 3.62 = 12.96. Finally, the
coefficients have to be multiplied by 1,000 to get the tractive effort in Newton. This gives the
following table:

77
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

Speed interval in m/s c0,k c1,k c2,k

0 – 3.166 128464.4 – 5310.6 – 58.86
3.166 – 8.333 132274.1 – 7634.6 13.86
8.333 – 16.667 132470.2 – 7778.7 155.58
16.667 – 22.167 105595.9 – 5422.9 84.32
22.167 – 27.500 20089.1 1087.0 – 260.38
27.500 – 32.500 50550.4 – 560.1 – 6.01
32.500 – 38.889 – 4107 1741.4 – 27.25
38.889 – 44.444 35080.9 – 90.0 – 6.48

The intervals that are affected are 30 – 60 km/h and 60 – 79.8 km/h. When the speed measure
is converted to m/s, consider the interval 40 – 60 km/h that is identical to the interval 11,1 to
16,667 m/s, and the interval 60 – 70 km/h that is identical to the interval 16.667 to 19.4 m/s. In
the interval 11,1 to 16.667 m/s, the tractive effort in Newton is
132470.2 – 7778.7 v – 155.58 v2

and in the interval 16.667 to 19.4 m/s, the tractive effort is

105595.9 – 5422.9 v – 84.32 v2

Taking into account that the gravitation constant g is 9.81 m/s2, the resistance is
1581.8 + 37.075 v + 0.3836 v2

FTr – FR is therefore
130888.4 – 7815.775 v + 155.1964 v2

in the interval 11.1 to 16.667 m/s and

104014.1 – 5459.975 v + 83.9364 v2

in the interval 16.667 to 19.4 m/s.

The mass factor is given by fp = 1.09. Then the following differential equations apply to express
the relation between time and speed during the acceleration phase:
130888.4 – 7815.775 v + 155.1964 v2 = 116000 · 1.09 · dv/dt

for the interval 11.1 to 16.667 m/s and

104014.1 – 5459.975 v + 83.9364 v2 = 116000 · 1.09 · dv/dt

for the interval 16.6 to 19.4 m/s.

4.4.3 Modelling the Train as a Homogeneous Strip

Trains have a certain length. It appears to be closer to reality to model the train as a
homogeneous strip, i. e. the weight is distributed equally through the whole length of the
train. The essential difference from the model of the train as a mass point is to determine its
resistance when the train passes the border line between two sections of the route where
different gradients apply. More generally, this is when over the total length of the train, more
than one section with different gradients is covered. It is assumed that the train covers k
sections S1,…,Sk. The train length is denoted by L and in section Si a gradient of ni applies.
The mass of the train is distributed homogeneously over the sections S1,…,Sk. The first
section S1 and the last section Sk are covered only partially.

78
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.5 Solving the Differential Equations by a Difference Equation Approach

The length of that part of the train meeting S1 is denoted by L1, and by Lk the length of the
part of the train meeting Sk and by Li the length of Si where i = 2,…,k – 1. The sum L1 +…+ Lk
always remains L. The resistance is
FR = FRt(V) + FRw(V) + FRlg (4.17)

The gradient resistance FRig is determined as follows:

FRlg = gm1n1 + gm2n2 + … + gmknk (4.18)

where mi is just the mass of the part of the train that covers Si. Since the train is considered as
a homogeneous strip, for each i = 1,…,k
mi = m · Li/L

L1 decreases in the same grade as the train moves, i. e. the location s of the train increases,
and Lk increases in the same grade as s, i. e. there are constants s1 and sk , such that
L1 = s1 – s

and
Lk = sk + s

That means FRig can be written as follows:

FRig = gm (n1L1 + n2L2 + … + nkLk)/L
= gM (n1(s1 – s) + n2L2 + … + nk-1Lk-1 + nk(sk + s))/L
= gM (n1s1 + n2L2 + … + nk-1Lk-1 + nksk + (nk – n1)s)/L (4.19)

FRig depends linearly on s, i. e. there are constants FRig,1 and FRig,2, such that
FRig(s) = FRig,1 + FRig, 2 · s (4.20)

The acceleration phase is expressed by the following differential equations:

m · fp · dv/dt = FTr(V) – FRt(V) – FRw(V) – FRlg (s) (4.21)

m · fp · dv/ds = (FTr(V) – FRt(V) – FRw(V) – FRig (s))/v (4.22)

The differential equations are transformed as follows:

m · fp · dv/dt = FTr(V) – FRt(V) – FRw(V) – FRlg,1 – FRlg, 2 · s (4.23)

m · fp · dv/ds = (FTr(V) – FRt(V) – FRw(V) – FRlg,1 – FRlg, 2 · s)/v (4.24)

In forthcoming sections, the train is considered as a mass point. This is much easier to handle
and in most cases the computational results are sufficiently precise.

4.5 Solving the Differential Equations by a Difference Equation Approach

The goal of this section is to find out the dependency of time and speed, and time and
distance covered through the acceleration phase. For simplicity, FTr and FR are considered as
functions on v in m/s.
The aim is to determine the time and the length of the acceleration path to increase speed
from an initial speed v0 to a speed vmax. This can be ascertained by subdividing the interval
[v0, vmax] into a certain number of sub-intervals of a fixed length ∆v. For each of these sub-
intervals, it is assumed that dv/dt and dv/ds are constant. One initially sets s0 = 0 and t0 = 0,
and the initial speed is v0. Then determine s1,…,sn, t1,…,tn and v1,…,vn as follows:

79
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

1) vi is just v0 + i · ∆v.
2) ti+1 can be determined from ti by transforming the differential equation (4.15) into a
difference equation:
m · fp · ∆v/(ti + 1 – ti) = FTr(vi) – FR(vi) (4.25)
This leads to the following recursion formula:
ti + 1 = m · fp · ∆v/(FTr(vi) – FR(vi)) + ti (4.26)
3) si + 1 is determined from si by transforming the differential equation (4.16) into a difference
equation:
m · fp · ∆v/(si + 1 – si) = (FTr(vi) – FR (vi) )/vi (4.27)
This gives the following recursion formula:
si + 1 = m · fp · ∆v · vi /(FTr(vi) – FR (vi) ) + si (4.28)

Assume that vn is vmax, tn is the computed time and sn is the computed length of the path to
accelerate from v0 to vmax.
If the maximum speed is not reached in this particular section, divide the section that has a
length ssec into a certain number of intervals of the same length ∆s.
Determine intermediate values s1, …, sn, t1, …, tn and v1, …, vn as follows.
1) si is just i · ∆s.
2) Replacing ∆v by vi + 1 – vi, obtains from (4.27) the following recursion formula for vi:
vi + 1 = (FTr(vi) – FR (vi)) ∆s/(vi · m · fp) + vi (4.29)
3) A recursion formula for ti can be got with the help of equation (4.26):
ti + 1 = ∆s/vi + ti (4.30)

Assuming that sn = ssec, the exit speed for this section is vn, and the time to pass the section is
tn.
It is not always clear whether the maximum speed is reached within the section. In that case,
proceed in the same way initially as if the maximum speed is reached. If si + 1 is larger than
the length ssec of the section, ∆s is defined to be ssec – si and ti + 1 and vi + 1 are computed the
formulae (4.29) and (4.30).
Finally, consider the case that the section is a braking section. The braking curve is given by
the following equation:
s = ssec – v2/(2ab) – vex2/(2ab) (4.31)

Determine si, vi, and ti as in the case that ∆s is fixed until the braking curve is cut, i. e.
ssec – vi + 12/(2ab) – vex2/(2ab) < si + 1. To find the speed at which there is a change from the
acceleration phase to the braking phase, determine the cut point of the braking curve and the
line joining (si, vi) and (si + 1, vi + 1). Solve the following quadratic equation:
ssec – v2/(2ab) – vex2/(2ab) = (v – vi)(si + 1 – si)/(vi + 1 – vi) (4.32)

The solution for v is

v = – ab (si + 1 – si)/(vi + 1 – vi)
(4.33)
ˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉˉ
+√(a 2
(s – s )2/(v – v )2 + v 2 + 2a s + 2a v (s – s )/(v – v ) – 2a s )
b i+1 i i+1 i ex b sec b i i+1 i i+1 i b i

The value for s is just determined by equation (4.31).

80
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.6 Solving the Differential Equations Analytically

Example continued

Assume that the following differential equations describe the tractive force:
130888.4 – 7815.775 v + 155.1964 v2= 116000 · 1.09 · dv/dt

for the interval 11.1 to 16.667 m/s and

104014.1 – 5459.975 v + 83.9364 v2 = 116000 · 1.09 · dv/dt

for the interval 16.667 to 19.4 m/s.

Determine the time and the length of the path to accelerate from 11.1 m/s to 19.4 m/s.
Restrict the acceleration from 11.1 m/s to 16.67 m/s (that is around 16.7). The difference is
5.6, and an appropriate value for ∆v is 1.4. Equation (4.26) can be written as follows:
ti + 1 = 116000 · 1.09 · 1.4/(130888.4 – 7815.775 vi + 155.1964 vi2) + ti

Equation (4.28) is now as follows:

si + 1 = 116000 · 1.09 · 1.4· vi/(130888.4 – 7815.775 vi + 155.1964 vi2) +si

The following table shows the time to accelerate from 11.1 m/s to 16.7 m/s and all
intermediate values vi, ti, and si.
i vi ti si
0 11.1 0 0
1 12.5 2.8 31.02
2 13.9 5.9 69.5
3 15.3 9.6 126.4
4 16.7 13.7 194.1

In 13.7 seconds the diesel unit set passes a path of length 194.1 m if it speeds up from
11,1 m/s = 40 km/h to 16.7 m/s = 60 km/h.
Compute the time and the path length to accelerate from 16.6 m/s to 19.4 m/s = 70 km/h in
the same way. This may be left as an exercise.

4.6 Solving the Differential Equations Analytically

The results of numerical computation procedures are precise enough. But closed formulas
lead to computationally efficient algorithms. The whole section is more interesting for an
expert who is familiar with the mathematical background behind running time calculation. To
understand the contents of Sections 4.7 and 4.8, it is not necessary to be familiar with the
contents of this section.
This case is restricted to that where the train is modelled as a mass point. As mentioned
before, the motion of the train can be modelled as follows:
FTr (v) – FR (v) = m · fp · dv/dt
(FTr (v) – FR (v))/v = m · fp · dv/ds

Here, consider FTr and FR as functions dependent on v (the speed measured in m/s). The
equation can be transformed as follows:
dt/dv = m · fp /(FTr (v) – FR (v)) (4.34)

and

81
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

ds/dv = v · m · fp/(FTr (v) – FR (v)) (4.35)

Integration gives the time and the length of the path to accelerate from a speed v’ to a speed
v’’.
t = ∫ m · fp dv/(FTr (v) – FR (v)) (4.36)

and
s = ∫ v · m · fp dv/(FTr (v) – FR (v)) (4.37)

where the integral borders are v’ and v’’.

Thus the integrals can be expressed by closed formulae.
For the solution of the integrals, it is helpful to observe that the resistance FR(v) is a polynomial
of second order in v, i. e. FR(v) is of the form
FR (v) = r0 + r1 · v + r2 · v2

Both possible characteristic curves are discussed for the traction effort.

4.6.1 Traction Effort Depends Piecewise Parabolically on Speed

Here it is assumed that the tractive effort is given as follows (see equation (4.1)):
FTr (V) = c0,k + c1,k · v + c2,k · v2, vk < v ≤ vk + 1

Reference is made also to the article of Jentsch & Gröpler (1995) as the first paper dealing with
the solution of the integrals by closed formulas (see also Jentsch (2003)).
Consequently, the acceleration effort, i. e. the difference between tractive effort and the
resistances, can be written as a polynomial of second order in v.
FTr(v) – FR(v) = d0,k + d1,k · v + d2,k · v2 (4.38)

where d0,k = c0,k – r0, d1,k = c1,k – r1, and d2,k = c2,k – r2
The acceleration time is determined by the following integral:
t = ∫ m · fp dv/(FTr (v) – FR (v))
= ∫ m · fp dv/(d0,k + d1,k v + d2,k v2) (4.39)

The following cases have to be considered:

1) d1,k2 – 4 d2,kd0,k < 0. Then D is defined as the square root of 4 d2,kd0,k – d1,k2
The integral for t can be rewritten as follows:
∫ m · fp dv/(d0,k + d1,k v + d2,k v2) =
= m · fp · 4d2,k/D2 ∫ ((2d2,k v/D + d1,k/D)2 +1) dv
= m · fp · 4 · arctan(2d2,k v/D + d1,k/D) /D + C (4.40)
where C is some constant.
That means to accelerate from some speed v1 to v2, the following amount of time is taken.
t = m · fp · 4 ·
(arctan(2d2,k v2/D + d1,k/D) – arctan(2d2,k v1/D + d1,k/D))/D (4.41)

2) d1,k2 – 4 d2,kd0,k > 0. D is defined as the square root of d1,k2 – 4 d2,kd0,k.

Defining A = (d1,k + D)/2d2,k and B = (d1,k – D)/2d2,k , the acceleration effort can be
transformed as follows:
d2,k ((v + d1,k/2d2,k)2 – (d1,k2 – 4 d2,kd0,k)/4d2,k2) = d2,k (v + A)(v + B) (4.42)

82
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.6 Solving the Differential Equations Analytically

and the integral can be rewritten as follows:

∫ m · fp dv/(d0,k + d1,k v + d2,k v2)
= m · fp/(d2,k(B – A)) · ln((v + A)/(v + B)) + C (4.43)
where C is some constant.
That means to accelerate from some speed v1 to v2, the following amount of time is taken:
t = m · fp/(d2,k(B – A)) ·(ln((v2 + A)/(v2 + B)) – ln((v1 + A)/(v1 + B))) (4.44)

It remains to find a closed formula for s.

∫ v · m · fp dv/(FTr (v) – FR (v))
= ∫ m · fp · v dv/(d0,k + d1,k v + d2,k v2) (4.45)

When z is defined as z = d0,k + d1,k v + d2,k v2, dz/dv = 2d2,k v + d1,k and the integral is
transformed as follows:
∫ m · fp · v dv/(d0,k + d1,k v + d2,k v2) =
= ∫ m · fp · (2d2,k v + d1,k) dv/(2 d2,k(d0,k + d1,k v + d2,k v2))
– ∫ m · fp · d1,k dv/(2 d2,k(d0,k + d1,k v + d2,k v2)) (4.46)

First integral can be solved by substitution.

∫ m · fp · (2d2,k v + d1,k) dv/(2 d2,k(d0,k + d1,k v + d2,k v2))
= (m · fp/2 d2,k) ln(d0,k + d1,k v + d2,k v2) + C (4.47)

where C is some constant.

Second integral is just
∫ m · fp ·d1,k dv/(2 d2,k(d0,k + d1,k v + d2,k v2))
= (d1,k/2d2,k) ∫ m · fp dv/(d0,k + d1,k v + d2,k v2) (4.48)

That means if the train accelerates from a speed v1 to a speed v2 where v1 and v2 are inside
the interval [vl,k, vl + 1,k) and needs a time t to accelerate from v1 to v2, then the path length to
speed up from v1 to v2 is
(m · fp/2 d2,k) (ln(d0,k + d1,k v2 + d2,k v22)
– ln(d0,k + d1,k v1 + d2,k v12)) + (d1,k/2d2,k) · t (4.49)

Since there is a closed formula for t, given v1 and v2, there is also a closed formula for s, when
v1 and v2 are given.

4.6.2 Traction Depends Hyperbolically on Speed

Here only a sketch is given, with reference to Dahlhaus & Körner (2005).
The tractive effort is given by the speed v
FTr (v) = ch,k / v, vl,k < v ≤ vl + 1,k

The overall acceleration effort is

FTr (v) – FR (v) = ch,k / v – (a0 + a1 · v + a2 · v2)
= (ch,k – a0v – a1v2 – a2 v3) / v (4.50)

This leads to the following differential equations:

dt/dv = m · fp/(FTr (v) – FR (v))
= m · fp · v/(ch,k – a0v – a1v2 – a2v3) (4.51)

83
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

ds/dv = v · m · fp/(FTr (v) – FR (v))

= m · fp · v2/(ch,k – a0v – a1v2 – a2v3) (4.52)

Decompose the polynomial ch,k – a0v – a1v 2 – a2v 3 of third order into a linear factor and a
quadratic factor:
ch,k – a0v – a1v2 – a2 v3 = (v – v0)(e0 + e1v + e2v2) (4.53)

v0 is chosen, such that ch,k – a0v0 – a1v02 – a2 v03 = 0.

Therefore, the differential equations can be rewritten as follows:
dt/dv = m · fp · v/(v – v0)(e0 + e1v + e2v2)
= m · fp · v · ((Av + B)/(e0+ e1v + e2v2) – C/(v-v0)) (4.54)

where
C = 1/(e0 + e1v0 + e2v02),
B = (e1 + e2v0)/(e0 + e1v0 + e2v02)

and
A = e2/(e0 + e1v0 + e2v02)

Through some intermediate steps,

dt/dv = m · fp · ((e2v0v – e0)/(e0 + e1v + e2v2)
– v0/(v – v0))/(e0 + e1v0 + e2v02) (4.55)

t dependent on v is expressed by the following integral:

∫ m · fp· ((e2v0v – e0)/(e0 + e1v + e2v2) – v0/(v – v0))dv
= m · fp· (∫(e2v0v – e0)/(e0 + e1v + e2v2)dv – v0ln(v – v0)) (4.56)

The first integral can also be expressed by a closed formula, because the denominator is a
polynomial of second order and the enumerator is linear.
To determine the length of the acceleration path, solve the differential equation
ds/dv = m · fp · v · ((e2v0v – e0)/(e0 + e1v0 + e2v02) – v0/(v – v0))/(e0 + e1v0 + e2v02))
= m · fp · ((e0 – v0e1)v + e0v0)/(e0 + e1v0+ e2v02) – v02/(v – v0))/(e0 + e1v0 + e2v02)

s dependent on v is expressed by the following integral:

∫ m · fp · (((e0 – v0e1)v + e0v0)/(e0 + e1v0 + e2v02) – v02/(v – v0)) dv/(e0 + e1v0 + e2v02)
= (m · fp/(e0 + e1v0 + e2v02)) (∫((e0 – v0e1)v + e0v0)/(e0 + e1v0 + e2v02)dv –
v02 ln(v – v0) (4.57)

The first integral in the last expression can also be expressed as a closed formula by the same
arguments as in equation (4.56).
So the time and the length of the acceleration path from one speed to another can be
represented by a closed formula.

4.6.3 Passing a Section

The train enters a section with an entrance speed ventrance. Three cases have to be considered
as mentioned above.
1) The maximum speed is a constant vmax and the maximum speed is reached before the
section is left. Then integrate over all the speed intervals [vi,k, vi + 1,k) that intersect the interval

84
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.6 Solving the Differential Equations Analytically

[ventrance, vmax), where in the first interval, vi,k is replaced by ventrance and in the last interval,
vi+1,k is replaced by vmax.
2) The maximum speed is not reached before the section is left. There is a closed formula for
s dependent on v, but there is no closed formula for v dependent on s. Therefore numerical
methods apply to determine the exit speed. Most appropriate method is the procedure of
Newton. We have to solve an equation
G(v) = s
where G(v) is the length of the path that is needed to accelerate from ventrance to v. The
equation is transformed into
G(v) – s = 0
An initial value v0 is fixed and the values vi are defined recursively as follows
vi + 1 = vi – (G(vi) – s)/G’(vi) (4.58)
where G’ is the derivation of G to v. G(v) is expressed by an integral, and G’(v) is just the
function that has to be integrated.
speed

initial value

first approximation

second approximation

third approximation

exact solution

train
position

Fig. 4.9: Illustration of Newton’s Method

3) The section is a braking section, and the braking curve is cut. To determine the value of v,
where the acceleration curve cuts the braking curve, solve the equation
G(v) + v2/(2ab) – vexit2/(2ab) = s (4.59)
that is equivalent to the equation
G(v) + v2/(2ab) – vexit2/(2ab) – s = 0 (4.60)
Apply Newton’s method to get a solution for v.

85
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

4.7 Solving the Differential Equations by Gauss Quadrature Integral Formulae

In the past Sections, it has been shown that the integrals behind the differential equations
describing the acceleration movement of the train can be solved by closed formulae. But the
closed formulae contain basic functions as arcustangens and logarithm. For computational
reasons, it appears reasonable to express the integral formulas by polynomials (for further
readings, see Wegele (2005)).
In general, the differential equations have the form
dt/dv = m · fp /(FTr(v) – FR(v))

and
ds/dv = m · fp · v /(FTr(v) – FR(v))

Approximate the functions F(v) = m · fp/(FTr(v) – FR(v)) and G(v) = m · fp · v /(FTr(v) – FR(v)) by
polynomials as follows:
A support v1 < v2 < ….< vk is fixed and unique polynomials P and Q of order k – 1 are
determined, such that for each vi,
F(vi) = P(vi)
G(vi) = Q(vi)

They are just the Lagrange-polynomial determined as follows.

Li is defined as follows:
Li(v) = ∏j = 1,…,k, j ≠ I (v – vj)/(vi – vj) (4.61)

It is easy to realise that Li(vi) = 1 and Li(vj) = 0 if j is different from i. The Lagrange polynomials
P(v) and Q(v) are
P(v) = ∑i = 1,…,k F(vi) Li(v)
and
Q(v) = ∑i = 1,…,k G(vi) Li(v) (4.62)

Afterwards, the polynomials P and Q can be integrated.

It appears to be reasonable to determine an optimal support and to approximate the functions
F and G by
Pq (v) = ∑i = 1,…,k F(vi) Li(v)2
and
Qq (v) = ∑i = 1,…,k G(vi) Li(v)2 (Gauss Quadrature) (4.63)

First the zeros of the k th derivation of (x 2 – 1) k are determined. They are all in the interval (1,1),
and there are exactly k zeros, and these zeros are denoted by x1,…,xk. The zeros x1,…,xk are
now transferred to the interval (ventrance, vmax), i. e. the support v1,…,vk is defined as follows.
vi = (ventrance + vmax)/2 + xi(vmax – ventrance)/2 (4.64)

Pq and Qq are transformed into a standard polynomial representation

Pq(v) = p0 + p1v +…+ pk – 1vk – 1
Qq(v) = q0 + q1v +…+ qk – 1vk – 1 (4.65)

and the integral of Pq(v) is

∫ Pq(v) dv = p0v + p1v2/2 + …+ pk – 1vk/k + C
∫ Qq(v) dv = q0v + q1v2/2 + …+ qk – 1vk + 1/(k + 1) + C (4.66)

where C is some constant.

86
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 4.9 Calculating Blocking Times

Therefore the time and the path length to accelerate from a speed ventrance to a speed v have
therefore the following approximate representations as a polynomials of the form
t = p0(v – ventrance) + p1(v2 – ventrance2)/2
+ … + pk – 1(vk + 1 – ventrancek + 1)/(k +1)

and
s = q0(v – ventrance) + q1(v2 – ventrance2)/2
+ … +qk – 1(vk + 1 – ventrancek + 1)/(k +1) (4.67)

According to Wegele (2005), a support of four values appears to be sufficient for good
approximations.

4.8 A Probabilistic Approach

Steckel (1991) proposed the using of random distributions for running time calculations.
Examples of randomly distributed variables are
1. mass of freight trains,
2. mass factor,
3. weight of the tractive unit.
Some of the randomly distributed variables are discussed here in detail.
1. Mass of a freight train: The mass of a freight train may be different on each day.
Nevertheless the same timetable has to be provided for the particular train on each
occasion. It can be assumed that the weight of a particular freight train is normally or
Weibull distributed. Parameters can be estimated by statistics.
2. Mass factor: Irregularities in the axles cause additional resistance and are essentially
responsible for the fact that the mass factor is greater than one. For the tractive unit, the
mass factor is considered as equally distributed where the lower and the upper bounds
depend on the type of the tractive unit. Also, for each axle of the remaining train, the mass
factor is assumed to be equally distributed. The average mass factor over all cars of the
train appears to be very close to the ordinary constant fp.
3. Weight of the tractive unit: If the tractive unit is diesel, the weight is dependent on the
amount of fuel it carries. It can be assumed to be equally distributed.
The running time computation is repeated several times on the particular route, using
randomly generated train data. Dependencies of random variables discovered during data
collection have to be taken into account. It appears to be reasonable to fix a running time for
the particular route that is exceeded only by a very low percentage of the repeated running
time computations. For details, refer to the Ph.D. thesis Steckel (1991).

4.9 Calculating Blocking Times

In this last section, the running time estimation explained above is used as an example for
calculating blocking times. It is assumed that the running time can be calculated for a given
train for each position p of its route, denoting with t(p) the complete running time in seconds
from start position 0 to position p (for the front of the train), including any dwell times between
0 and p. From now on, no differentiation is made between the moving sections of the train.
87
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
4 Running Time Estimation

The components of the blocking time of a block section limited by signals S1 and S2 are listed
in Section 2.4.1.
– The first component, the time tc for clearing the signal S1, is not dependent on the
running time but on the signal box. It is normally assumed as a constant depending on the
interlocking technology (if not, take 12 s).
– The second component, the time tv for viewing the signal by the driver, is also assumed as
a constant value (12 s).
– The approach time ta is the first component which is dependent on the running time
estimation. It is measured between the signal that provides the approach indication (i. e.
distant signal D1 of signal S1) and the signal S1 at the entrance to the block section, so the
approach time begins at t(D1) and ends at t(S1).
– The time tb between the block signals begins at t(S1) and ends at t(S2).
– The clearing time to begins at t(S2). It ends when the train has left the overlap section after
signal S2 with its full length l. If the overlap section ends at position O2, the clearing time
ends at the time t(O2 + l). Note that this position is dependent on the length of the train.
– The release time tr is dependent on the interlocking like the first component. It can often be
assumed as 6 s.
So the blocking time between signals S1 and S2 begins at the time
t(D1) – tv – ta

and ends at the time

t(O2 + l) + tr

By calculating the blocking times for each block (and for each sectional release route locking, if
provided by the interlocking), the blocking time stairway can be built up.

Fig. 4.10: Blocking Time Calculation

88
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

References
Curtius E. W., Kniffler A. (1950): Neue Erkenntnisse über die Haftung zwischen Treibrad und
Schiene, Elektrische Bahnen 21 No. 9, pp. 210 – 210.
Dahlhaus E., Körner H. (2005): Exact determination of the travel time of rail vehicles, to appear.
Habich G., Eickmann G. (1990): DYNAMIS – ein Simulationsmodell zur Bearbeitung
fahrdynamischer Fragestellungen, Eisenbahntechnische Rundschau, year 1990, issue 1/2,
pp. 83 – 87.
Jentsch E., Gröpler O. (1985): Innovation der methodischen Basis für Fahrzeitermittlungen
und ihre Anwendung bei Zugfahrtsimulation, Schriftenreihe für Verkehrssystemtheorie und
Bahnverkehr, 1.8, pp. 14 – 36.
Jentsch E. (2003): Fahrzeitermittlung mit neuen Elementen der Zugfahrtsimulation, ZEVrail 127
No. 2, pp. 66 – 71.
Sauthoff F. (1932): Die Bewegungswiderstände der Eisenbahnwagen unter besonderer
Berücksichtigung der neueren Versuche der Deutschen Reichsbahn, Ph.D. Thesis (Dr.-Ing.),
Berlin, Technical University (Technische Hochschule), (1933 in: Berlin: VDI-Verlag).
Steckel J. (1991): Strategische Optionen für die zufällige Fahrzeit im Eisenbahnbetrieb, Ph.D.
Thesis (Dr.-Ing.), Dresden Institute for Traffic Research (Hochschule für Verkehrswesen).
Strahl (1913): Die Berechnung der Fahrzeiten und Geschwindigkeiten von Eisenbahnzügen
aus den Belastungsgrenzen der Lokomotiven, Glasers Annalen für Gewerbe- und Bauwesen
73, pp. 869 – 871.
Wegele S. (2005): Echtzeitoptimierung für Disposition im Schienenverkehr, Ph.D. Thesis
(Dr.-Ing.), Braunschweig, Institute for Traffic Safety and Automation Engineering.
Wende D. (2003): Fahrdynamik des Schienenverkehrs, Wiesbaden: B. G. Teubner Verlag.

89
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.1 Minimisation of Mechanical Energy Consumption

5 Energy-Efficient Railway Operation

Thomas Albrecht

In times of climate change and growing pollution, the energy efficiency of transportation
systems becomes more and more important. That includes the railways.
The main consumption of energy by the railways is that needed to move the trains. This
energy comprises the mechanical energy needed to overcome the driving resistances, and
the losses caused by the transmission and transformation of energy in the traction and power
supply chain. Both are very much influenced by the way a train is driven within the given
timetable.
This Chapter starts by introducing the fundamentals of mechanical energy consumption and
presents algorithms to minimise mechanical energy consumption during train operation.
Constraints from operations planning which influence energy consumption are discussed here.
Then, billing systems for energy consumption applied in practice are presented, as well as
methods not only to reduce mechanical energy consumption, but also to reduce energy costs
directly.
Finally, some of the existing solutions for implementing energy-efficient driving in practice will
be explained, with a special focus on driver support systems.

5.1 Minimisation of Mechanical Energy Consumption

Mechanical energy (often called traction energy) is needed to move a vehicle from one place to
another. It can be computed as the integral of mechanical power Pmech over time. Mechanical
power at the wheel-rail interface can be expressed as the product of tractive effort F and
speed v :
Emech = ∫ Pmech (t) dt with Pmech (t) = F(t)v(t) (5.1)

Another frequently used formula for traction energy consumption is the integral of tractive effort
F over space s
Emech = ∫ F(s) ds (5.2)

The above formulae demonstrate that basically every acceleration process of the train (positive
tractive effort F ) costs energy, whether it is planned (acceleration after a regular stop) or
unplanned (acceleration after a stop at a signal). The shorter the available running time until a
target point, the more the train has to accelerate, and the more energy is needed.

5.1.1 Energy-Efficient Driving in Undisturbed Operation

Assume that a train will be moved from one station to the next within the time given by the
timetable. Further assume that the track is presently unoccupied and that the route to the next
station is set (undisturbed operation).
The way of driving the train on this section depends on any delay on departure and the
planned running time. The primary goals when driving a train are safety and punctuality, see
Fig. 5.1.
The planned running time is the sum of the minimal running time and a running time
supplement, also known as running time reserve or slack (Section 2.3). If the train has a
delay on departure greater than the available running time reserve, it must be driven as fast

91
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

as possible within the constraint of the maximum speed permitted – time-optimal – in order
to reduce that delay as much as possible. See Chapter 4 for the computation of the minimal
running time.
If the delay on departure is less than the available running time reserve, there are different
possible ways to drive a train with the goal of arriving on time. Among those different driving
strategies there is one which is energy-optimal, in that it requires the minimum use of
mechanical energy. In this Section, this particular driving strategy – what tractive or braking
effort to apply and when – will be calculated.

Fig. 5.1: Understanding train control as a control loop

The solution can be obtained straightforwardly using the theory of optimal control. The first
publications on this approach date back to the 1970s, see Strobel et al. (1974). It has been
extended and improved by others over the years, see Howlett & Pudney (1995), Khmelnitsky
(2000), Liu & Golovitcher (2003), Wang et al. (2011), Albrecht, A. et al. (2013), but the basic
approach remains unchanged.

5.1.1.1 Derivation of the Optimal Driving Strategy for Simple Tracks

To model the process of train movement, process time t and the (time-dependent) state
variables,
– the position of the train x1 and
– the speed of the train x2
are introduced. The control variable u of the process (also varying over time) is the effort F
applied by the train at the traction wheels in relation to the absolute value of the maximum
braking effort of the train FB :
F
u= (5.3)
FB
Some specific conditions are made under which the task of energy-optimal train control is
regarded:
F
– The control variable u shall be kept within the limits −1≤ u ≤ T = umax where FT denotes
FB
the maximal tractive effort.

92
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.1 Minimisation of Mechanical Energy Consumption

– A single section with only one speed limit vm is considered. Because return travel is
excluded as well as exceeding the maximum speed, the travelling speed x2 must stay within
the bounds 0  x2  vm .
– The variation of the gradients of the section shall be limited so that a single average
gradient im on the section can be assumed.
– The train starts at time t = 0 with x1(0) = 0 and x2(0) = 0, it shall arrive at the next station at
time tf with x1(tf ) = x1f and x2(tf ) = 0.
These simple conditions assumed for the track (constant maximum speed and gradient) are
met in practice on most suburban railway lines. Under these conditions, train movement can
be described by the following system of differential equations (Chapter 4):
dx1
= x2 (5.4)
dt

dx 2
= uBu − fW (x 2 ) − img (5.5)
dt
F (x )
where fW (x 2 ) = Ax 22 + Bx 2 + C is the specific train running resistance (fW (x 2 ) = W 2 ), uB
F m
the specific maximum braking rate uB = B , and m the train mass. The influence of rotating
m
masses as described in Section 4.2.4 can be integrated into the given parameters.
The optimisation criterion will be the mechanical energy needed to drive the vehicle. It can
be computed from the state variable x2 (speed) and the control variable u as follows (cf.
Equation (5.1)):
tf tf tf
1 1
Q=
∫ f (x ,x ) dt = ∫
0 1 2
2
(F + | F |) x 2 dt = ∫ 2 (u + | u |)u m x
B 2 dt (5.6)
0 0 0

Note that during coasting (u = 0) and braking (u < 0), the term u + |u| becomes 0, i. e. no
traction energy is needed by the train in those regimes.
To solve the optimisation problem under the given simplified conditions, Pontryagin’s Maximum
Principle can be applied. It requires that the optimal control uopt (t) should be selected from the
various admissible control variables u, which maximise the Hamiltonian function
n

∑⎜⎜⎜⎝p dxdt ⎟⎟⎟⎠− f (x,u)

⎛ i⎞
H(x,u,p) = i 0 (5.7)
i=1
For the given problem, the number of states n equals 2, the state equations (5.4) and (5.5)
and the integrand of the optimisation criterion f0 from Equation (5.6) have to be used. The
Hamiltonian function then becomes
⎛ u + | u | ⎟⎞
H = p1x 2 + p2 (uBu – fW (x 2 )−img) – ⎜⎜ ⎟u mx (5.8)
⎜⎝ 2 ⎟⎟⎠ B 2

The so-called costate variables p1 and p2 must fulfil the following system of differential
equations:
dp1 ∂H dp2 ∂H dx1 ∂H dx 2 ∂H (5.9)
=− ; =− ; = ; =
dt ∂ x1 dt ∂ x 2 dt ∂p1 dt ∂p 2
From Equation (5.8), it can be seen, that the Hamiltonian reaches a maximum with respect to
the control variable u, if
u+|u|
up2 − mx 2 (5.10)
2

93
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

becomes a maximum. As a consequence, there are only five optimal driving regimes to be
applied depending on the value of the costate variable p2 :
⎧⎪ umax for p2 > x 2m Acceleration with maximal acc. rate
⎪⎪
⎪⎪
⎪⎪ [0,umax ] for p2 = x 2m (2nd singular solution) Cruising
uopt = ⎪⎨ 0 for 0 < p2 < x 2m Coasting (5.11)
⎪⎪
⎪⎪ [−1,0] for p2 = 0 (1st singular solution) Cruising
⎪⎪
⎪⎪ −1 for p2 < 0 Braking with max. braking rate
⎪⎩

The obtained optimal regimes can be explained in detail:

– Acceleration and braking must both use the maximum (operational) tractive and braking
effort, for the following reasons. The slower a train brakes, the more time it needs to come
to a standstill. So, in order to obtain the same running time with a lower braking rate, the
train must accelerate to a higher speed, which needs more energy. The maximum rate must
therefore be the most energy-efficient.
– During coasting, when no tractive or braking effort is applied (so-called free motion of the
train), the train just rolls forward on the infrastructure – no traction energy is needed. Thus,
the earlier that coasting can start, the more energy that can be saved.
– Cruising means holding a constant speed. There are two different cruising regimes, one
obtained with partial tractive effort, the other obtained with partial braking effort. The
choice depends on the running resistance of the train at the cruising speed. Under most
conditions, running resistance will be positive, so that partial acceleration will be needed
to hold the speed. Only with very steep downhill gradients does the running resistance
become negative, and the brakes must be applied partially in order not to exceed the
cruising speed, see Liu & Golovitcher (2003). The solution found by the application of the
Maximum principle does not allow a conclusion to be reached about the optimal cruising
speed itself. Howlett & Pudney (1995) have shown that it can vary depending on the
properties of the track and the timetable. However, in urban or suburban train operation
with short station distances and modest gradients, it has been shown by several Authors,
e. g. Strobel et al. (1974), that the application of cruising below the speed limit is not
necessary to save energy. It is sufficient to apply cruising at the maximum speed here.
Focussing on suburban and urban transport systems, the optimal sequence of driving
regimes is given in Fig. 5.2. After a stop at a station, acceleration with maximum tractive effort
must be applied. Then, depending on the available running time reserve, either the coasting
phase follows immediately, or the maximum speed is reached and therefore cruising has to
be applied. The coasting phase can also take place after cruising at maximum speed, as
illustrated in Fig. 5.2. Finally, the train has to brake with maximum operational deceleration rate
to reach the next station with speed 0 .

Fig. 5.2: Energy-optimal driving on a simple section with constant gradient and maximal speed

94
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.1 Minimisation of Mechanical Energy Consumption

On steep downhill gradients, coasting may precede cruising at maximum speed but, without
loss of generality, the first and more frequent case of cruising preceding coasting will be
explored further in this section.

5.1.1.2 Computation of the Switching Points

As the optimal sequence of the driving regimes on such a section is now known, the points
at which the switches are made have to be determined. This can be done by solving the
differential equations of motion, or by train running simulation through systematic variation of
the coasting start point. But when this approach was first presented, the available computing
power was low and the operation took too long. Consequently, a linear approximation of the
train model (especially the running resistance) was used to solve the differential equations of
motion analytically.
Of course, every approximation causes errors compared with the more precise train model
given in equations (5.4) and (5.5). Those errors are especially critical during the coasting
phase, where the running resistance is the only active force in the system. Thus, Strobel et al.
(1974) proposed to provide a linear output in such a way that the error becomes minimal in
the interval, where the coasting phase usually takes place. This interval vu  x2  vm is usually
limited by the maximum speed vm and the planned average speed on the section
x (t ) − x1(0)
vu = 1 f .
tf
The optimal linear approximation in this speed interval can be described as follows:
f w = ax 2 + b with
A B
a= ( vu + vm ) +
m m (5.12)
and
1 A 1 C
b=− (vm vu + (vm + vu )2 ) + + img (5.13)
3m 2 m
The approximation then leads to the following linear state equation for the train speed:
dx 2
= uuB − (ax 2 + b) (5.14)
dt
This ordinary linear differential equation can be solved analytically for constant values of the
control variable u, using the well-known techniques of first finding a solution for the related
homogeneous differential equation and then adding the particular solution for the non-
homogeneous equation.
To determine the point where coasting has to start, the trajectory of the train is computed
backwards beginning with its arrival at the target station. Therefore, T is introduced denoting
the time until arrival T = tf – t . The braking trajectory starts at T = 0 at target distance x1 = x1f
and speed x2 = 0. To compute it, the differential equations of motion have to be solved for
u = –1. One then obtains functions x2 (T) and by further integration x1 (T). By computing T(x2 )
and later substituting T in the equation x1 (T) the braking trajectory is obtained:

1 ⎡⎢ uB + b ⎛⎜ ax 2 ⎞⎟⎟ ⎤
⎥
x1(x 2 ) = x1f − ⎢ ln⎜⎜1+ ⎟⎟ − x 2 ⎥ (5.15)
a⎢ a ⎜⎝ uB + b ⎠ ⎥⎦
⎣

By solving the linear differential equations of motion for u = 0, formulae for time/space and
time/speed relations for the coasting phase can be found. Again, T can be substituted by T(x2 )
in the time-space-function. By using a generic time T1 as the switching time between coasting
95
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

and braking and consecutively as a boundary condition for solving the differential equations
of motion of the coasting phase, the following switching curve can be obtained to determine
where coasting must start (the so-called coasting point):
⎡ ⎤
uB ⎛⎜ uB + b − (ax 2 + b) e
−aT ⎞
1⎢ ⎟⎟ ⎥
x1(T,x 2 ) = x1f − ⎢bT − ln⎜⎜⎜ ⎟⎟ − x 2 ⎥
⎟⎟⎠
(5.16)
a⎢ a ⎜⎝ uB ⎥
⎣ ⎦
This equation represents the most important part of a closed-loop algorithm for energy
optimum train control on a section with constant gradient and maximum speed.
In the case of large changes to track gradients between two stations or in the case of different
maximum speeds, the sequence of optimal control regimes is still valid for each part of the
section with constant gradient and maximum speed (sub-section). The running time reserve
available for the whole section must be distributed among the different sub-sections in order to
minimise energy consumption. For instance, this can be done by a gradient method proposed
by Horn and Zinke (1990), which is based on the following:
1. On each part of the section with constant speed limit (and gradient), the above mentioned
optimal regimes are used in the optimal sequence depending on the gradient. Some of the
regimes may be left out as well.
2. The ratio between energy saving ∆E and increase in running time ∆t must be equalised for
every part of the section, with constant maximal speed and gradient (sub-section).
The algorithm starts to simulate the minimal running time for the whole section. Then, for every
sub-section, the effects of running time prolongation or a decrease of the target speed at the
end of a sub-section, are simulated. For the sub-section with the greatest gradient
E
t
the simulated measure is approved, which leads to an increase in overall running time for the
section. The algorithm is repeated until all the available reserve is distributed.
Fig. 5.3 illustrates the way the algorithm works by showing optimal trajectories for different
travel times on a section with no gradients, but five subsections with different maximum
speeds.

Fig. 5.3: Distributing the available running time reserve among the different subsections with equal
constraints

96
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.1 Minimisation of Mechanical Energy Consumption

There are other approaches described in the literature for finding the switching points between
the driving regimes, with detailed train and track models. As an example, Chang & Sim
(1995) make the assumption that all possible trajectories can be described by a sequence of
acceleration and coasting regimes. These are overridden by an assumed ATP controller which
guarantees that neither maximum speed limits nor target braking are omitted. They propose
the application of Genetic Algorithms (GA) to solve this problem, where the relative positions
of the switching points between the regimes are coded in a chromosome. As the number
of coasting regimes is not known in advance for complicated tracks with different maximum
speeds, the length of the chromosomes is not fixed but also subject to optimisation.
De Cuadra et al. (1996), uses direct search algorithms in combination with precise train
running simulation models to find the optimal trajectory of the train for a detailed track and
train model, without any restrictions to the regimes applicable.

5.1.2 Energy-Efficient Driving in Disturbed Operation

Conflicts cannot be avoided in practical railway operation, especially in complex networks

with heterogeneous traffic. These conflicts frequently cause trains to stop at signals showing
restrictive or stop aspects. This causes delays, additional energy consumption and wear
and tear on train and track. Different strategies to solve conflicts in railway operation by re-
scheduling are explained in Chapter 11. One of them is modifying the speed of the train before
the conflict, in order to avoid the unplanned stop. This solution strategy is sometimes called
bending of the time-distance-curve of the train.
To avoid unplanned stops and their negative consequences, the sequence of trains on each
track vacancy detection section and predicted earliest release times need to be known. Then,
time windows can be computed for signals and section release points such that none of the
trains needs to stop. See Albrecht T. et al. (2013). Jaekel & Albrecht (2013). Time windows
between two consecutive trains should be set in such a way that they follow each other, with
minimal buffer time at points where the trains involved take different routes.
In long-distance traffic, the time windows are chosen so that the train reaches the distant
signal before the common critical section when it has just been released by the previous train.
This distant signal then has to be passed at maximum speed in order to minimise train delay
(see Fig. 5.4).
To reach such a target point within a given time window, the optimal regimes introduced in the
previous Section should be used. In order to facilitate the practical implementation of those
driving strategies, the use of cruising instead of coasting is sometimes preferred. The optimal
trajectory of the train can be found using search methods, see D’Ariano & Albrecht (2006).
For a reliable solution, an accurate prediction of the driving of the trains in the network is
required. This can be done by precise train running simulation. To realise this kind of driving
strategy in practice requires dynamic network-related driver support systems, as described in
Section 5.3.2, or automatic train operation.

5.1.3 Operational Requirements for Energy-Efficient Driving

The timetable for each single train as well as those for the whole railway system influences
energy consumption and costs. The size and distribution of time reserves which are included
in the timetable (Section 2.3) largely influence the mechanical energy consumption of the
single train.

97
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

Fig. 5.4: Basic principle of anticipating train control in long-distance traffic

The influence of the size of running time reserves on the energy consumption on a
single section (start speed 0, target speed 0) is illustrated in Fig. 5.5. As can be seen, the
dependency of energy consumption on running time reserve resembles a hyperbolic function.
Thus the closer the planned running time is to the shortest possible running time, the more
energy can be saved with a given amount of running time reserve.

Fig. 5.5: Characteristic dependency of energy consumption on running time for a section between
two consecutive stations

As a consequence, the distribution of running time reserve among the different sections of a
line strongly influences the train’s traction energy consumption during its journey. In practical
operation, time reserves are often fixed according to some timetabling rule (Section 2.3).

98
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.1 Minimisation of Mechanical Energy Consumption

Unequal distribution of time reserves (e.g. putting most of the reserve before reaching a critical
station) may lead to higher energy consumption than almost equal distribution.
An energy-optimal distribution can be computed using Dynamic Programming, see Albrecht
& Oettich (2002), Binder & Albrecht (2012). The problem can be regarded as a multi-stage
decision process. For each section of line, the optimal amount of reserve to spend on that
section has to be found, given that the arrival time at the terminus is fixed. The operational
conditions on the track further ahead (probable disturbances by other trains or during dwell
time) must be predicted until that critical point is reached, for a reliable estimation of the
available running time reserve.
For the same reason, the driving strategy to be applied in the case of delays heavily influences
the traction energy consumption of a single train. If any delay which occurs must be recovered
immediately due to operational rules, the trains will consume a lot of energy. It may be that
a delay can only be recovered on reaching a critical infrastructure element (station) further
along the route. If time reserves are still available, an algorithm for dynamic train running time
regulation based on Dynamic Programming can be applied. This will find the energy-optimal
distribution of running time reserves until that critical point is reached.
In passenger transport systems, dwell times and published departure times from stations also
have an impact on energy consumption. Timetables and departure times therein are usually
given to the passenger with the precision of one minute, whereas they may be calculated with
a much higher precision in the order of a few seconds (Chapter 4). Depending on the strategy
of the Train Operating Company, trains may leave a station when the published departure time
(to the full minute) has been reached, which may be earlier than the operational departure time.
Thus the dwell time originally planned becomes available as additional running time reserve.

Fig. 5.6: Using dwell time reserves as running time reserves

The introduction of stops on demand on lines with little passenger demand may further
decrease traction energy consumption, due to less frequent and shorter acceleration. The
running time reserves resulting from not having to stop become available for the remainder of
the journey and can be used for reducing energy consumption there.

99
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

Furthermore, the buffer times incorporated in the timetable (Section 2.5) influence the number
and severity of conflicts in a network. The more buffer time is available, the less other trains are
disturbed by a delayed train and the fewer unplanned stops occur.
For the optimisation of freight trains, energy consumption is often not the only relevant factor
when optimising speed profiles or tractive effort. Two examples are in-train forces (between
wagons) in couplers or minimal speed restrictions, e.g. in the approach of steep inclines,
which need to be considered in the optimisation. Then driving regimes such as partial tractive
effort for acceleration might become relevant. See Zhang & Zhuan (2014).

5.2 Optimisation under Practical Billing Systems

In practice, mechanical energy is not used to determine the energy bill, because it cannot
be provided as such for operating trains. It must be converted from other sources of energy
(diesel fuel, electricity), which are to be purchased on the market. Therefore, different billing
systems are used in railway practice, which can be distinguished between train-based and
substation-based methods. Some of those billing systems require specific methods to reduce
energy costs. A few of these methods will be briefly introduced in this Section.

5.2.1 Energy Supply and Billing

The mechanical energy needed to move a vehicle is usually provided by either an internal
combustion engine or electricity.
Internal combustion engines in trains nowadays mainly work with diesel fuel. The fuel is burned
and thereby rotating energy is provided. This is transformed into the movement of the train at
the wheel-rail interface, and – using a generator – into electrical energy for on-board use. The
actual diesel consumption can be computed with the efficiency of the traction system , the
heating value of the fuel H and the power used for all hotel functions Phelp:
1 1
mfuel =
H ∫ (F,v) Fv + P
help dt (5.17)

The heating value H of diesel is approximately 42 MJ/kg. The volume of diesel consumed
depends on its density and therefore also on the temperature.
The diesel fuel is bought by the Railway Undertaking (RU). It is billed based on the actual fuel
consumption of each train. To reduce the energy costs of diesel powered trains in undisturbed
operation, it is necessary and sufficient to have a look on the behaviour of each train
individually.
Also electric locomotives can be powered by energy stored on board in chemical form (e. g. in
accumulators or fuel cells), but those systems – which are comparable to combustion vehicles
concerning the billing and cost reduction strategies – have no practical relevance today. The
most commonly used form of energy supply is an electric power supply network along the
track, either by catenaries or third rail.
Electric traction has several advantages over diesel. As well as high power and efficiency, less
noise, and no local emissions, electric trains are able to transform kinetic energy to electrical
energy (regenerative braking). This energy can be used to supply on-board hotel systems (air
conditioning etc.) or be fed back for the use of other trains within the power supply network.
The billing of electric energy varies substantially between the different operators. Some
mainline railway infrastructure managers include the energy costs in the price of a train slot

100
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.2 Optimisation under Practical Billing Systems

on the network, or allocate the energy costs in the whole network to the different operators
based on their respective traffic volumes (in tonne-km). Those billing strategies are the easiest
to implement in practice, but are usually not fair to some of the operators since they are not
dependent upon the actual energy consumption of individual fleets.
Furthermore, these simplistic demand-independent billing strategies are no incentives for
Railway Undertakings (RUs) to reduce their energy consumption, which is critical from an
environmental point of view.
To encourage RUs to use energy efficiently, demand-dependent energy billing must be used.
For this, the following physical measures are to be used:
– Electric power, e. g. at the pantograph of the train, can be computed as the product of
voltage U and current I as well as the product of mechanical power and the efficiency of the
traction system:
1
Pel= UI= Phelp (5.18)
 (F,v)
The measuring unit of power is Watt ( W ).
– Electric energy consumption: It can be computed as the integral of electric power Pel over
time t :
E el =
∫ P dt
el (5.19)

The standard SI-unit of energy consumption is the Joule (J). In the context of electrical
systems, however, the units kWh (kilowatt hour) or MWh (megawatt hour) are more
commonly used. 1 kWh corresponds to 3.6 · 10 6 J.
Based on those physical measures, there are two categories of billing systems for electrically
powered railways. See Fig. 5.7 and Hoffmann (2006) and Uher & Sathi (1983):
1. Energy is billed per train (Train-based billing). Here, an energy meter must be installed on-
board each electric locomotive or EMU, which measures the electric energy taken from
the power supply system as well as regenerative energy fed back into that network. The
energy bill consists of a price for the consumed energy and a reward for energy fed back
into the network during braking. The price of energy may vary depending on where and
when it is requested from the power supply network. One example for such a billing system
is Germany, where the tariff varies according to the time of day (high costs during daytime,
lower cost during night time). The reward paid for regenerative energy is approximately
70 % of the price of consumed energy (as of 2012). This billing method requires a lot of
effort (installation of energy meters, management of energy meter data), but it motivates
companies to reduce their energy consumption and is therefore the best in networks with
many different RUs. In undisturbed operation, energy costs can be reduced by optimising
each train individually.
2. Energy is billed per electric power substation (Substation-based billing). If the energy is
billed per substation, no on-board equipment has to be installed, which is an advantage
over billing per train. But, on the other hand, this method can only be applied in networks
operated by a single RU, e. g. in urban rapid transit systems (light rail, metro, tramway),
see Albert et al. (1997). Here, electricity costs are calculated from one or two demand-
dependent components:
a) Electric energy consumed at substation level:
This comprises the energy consumed by all the trains and the losses in the power
supply system (between substation and trains). In contrast to an energy bill per train,
regenerative energy from braking is not automatically rewarded. It can be used free of
charge by all trains in the network, but if no train is available to use regenerative power at
the time it is available, it cannot be used to reduce electricity costs of the RU.

101
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

b) Electric power at substation level:

The electricity supplier must provide sufficient equipment (generation, transmission and
distribution facilities) to match the maximum power requirements of the railway. This
will be when many trains are accelerating at the same time. To encourage the RUs to
decrease their power peaks, the highest power value makes up part of the electricity bill
charged by many electricity suppliers. In order not to charge very short power peaks,
the average power over a certain measured interval t* (10 or 15 min in metro or light rail
systems) is used as billing measure. This is referred to as 15-min-average power. It can
be computed from
⎛⎜ x+t* ⎞⎟
⎜⎜ 1 ⎟⎟
Pmax,av (t*) = max ⎜⎜
x∈[0;tm −t*]⎜ ⎜⎜ t * ∫ P(t) dt ⎟⎟⎟
2
⎟⎟
⎟⎟
(5.20)
⎝ t=x ⎠
for the time interval [0; tm ] within which the power is measured, see Biesenack et al.
(2006).
To reduce electricity costs here, an individual optimisation of each single train is not
sufficient. The whole system of moving trains within the network must be taken into
account, even in undisturbed operation. Optimisation in the case of billing per substation
is therefore a lot more complex compared with the case where every train is measured and
billed individually.

Fig. 5.7: Different billing models in electric railway systems

5.2.2 Optimisation under Train-Based Billing Systems

Train based billing methods can be applied to both diesel and electrically powered trains.
There are some scenarios, where the traction-dependent part of the energy consumption
of the train can be regarded as being roughly proportional to the mechanical energy
consumption. This is the case where the efficiency of the traction chain hardly varies with
speed or tractive effort and the regenerative energy from braking cannot be used (diesel or
older electric vehicles). In those cases, as in suburban railway operation with powerful trains
travelling at relatively low speed, the application of an algorithm as simple as introduced in
Section 5.1 can lead to significant energy savings.

102
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.2 Optimisation under Practical Billing Systems

At longer inter-station distances as with intercity passenger trains or freight trains, trains travel
at higher speeds more frequently. Therefore, the effects of non-constant traction efficiency as
well as the non-linearity of train resistance must be taken into account during the calculation
of the energy-optimal trajectory. The problem becomes highly non-linear and determining
switching points by solving the differential equations of motion analytically is no longer
possible.
Franke et al. (2002) proposed the introduction of other variables to better take into account
the train behaviour, especially at high speeds and over long distances. They use normalised
kinetic energy ekin = 0.5v 2 and time t as dynamic states and the position of the train s as an
independent variable in order to get the following system of differential equations:
dekin 1 (5.21)
= (Ftrac −immg− Aekin − C)
ds m
dt 1
=
ds 2ekin
where  is the factor for the rotating masses, A stands for the quadratic part of the running
resistance and C for the speed-independent part of it, cf. Eqn. (5.5).
With the choice of ekin as state variable, the Authors have eliminated some important non-
linearities in the model. The influence of air resistance becomes linear, which is very interesting
for trains travelling at high speed at which the quadratic part of air resistance plays a major
role. Additionally, even with the remaining non-linearity in the system, a piecemeal analytical
solution of the equations becomes possible, if everything but ekin , t and s is constant. Using
s as an independent variable simplifies the incorporation of track-related data such as speed
limits and gradients or tunnels. The main improvement of this approach compared with the
simplified model is that the optimisation goal is not mechanical energy, but electrical energy at
pantograph level:
sf

Q = E el =
∫ F v +P
el loss (Fel,v)ds ⎯⎯
→ min
Ftrac
(5.22)
s0

where the tractive effort Ftrac is the sum of an electrical part Fel for acceleration and
regenerative brake, and a mechanical part Fmech for the pneumatic brake:
Ftrac = Fel + Fmech

The efficiency of the traction and braking system is regarded as speed and effort dependent
and is represented by Ploss, which can be derived from a detailed propulsion model.
The optimal control problem can no longer be formulated in a continuous way, but is regarded
as a discrete, multistage decision problem. Therefore, the track is disaggregated into several
stages, whose length must guarantee that track and train parameters as well as the tractive
effort sought can be considered as constants. Using Discrete Dynamic Programming, the
problem can be solved in real-time. The resulting control program allows the handling of
small deviations during the journey without excessive re-computing. One result of high
practical relevance is that the consideration of the efficiency of the propulsion system leads
to a solution, where no maximum acceleration or braking efforts are required at high speeds
(due to their low efficiency). This is qualitatively different from the analytical solution explained in
Section 5.1 and especially relevant for long-distance traffic, where trains usually travel at high
speeds for a comparatively long time.
In diesel trains, the efficiency of the traction system is also speed dependent and must be
integrated into the optimisation process. Additionally, the reality of a limited set of control

103
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

settings (running notches) must be taken into account. Howlett & Pudney (1995) give a detailed
and extensive overview on the optimisation of diesel-powered trains. The most important
differences from the simple model introduced in 5.1 are
– the use of cruising below maximal speed,
– the substitution of some cruising phases by infinitesimally short applications of acceleration
and coasting phases (sawtooth-shaped curve of speed over time), and
– the determination of the optimal running notch setting during acceleration.

5.2.3 Optimisation under Substation-Based Billing Systems

In the case of electricity costs being calculated at power substation level, the interaction of
many trains influences electricity costs sometimes more than the individual mechanical energy
consumption of the trains. It must therefore be taken into consideration when striving to
reduce electricity costs.
Fig. 5.8 represents two situations, where both the trains involved have the same mechanical
energy consumption, but the power values measured at the substation are completely different
due to the interaction of the trains. In scenario A, both trains accelerate at the same time,
whereas in scenario B, train 2 starts accelerating when train 1 starts braking. In Scenario A,
the resulting power peaks are twice as big as in scenario B, and the regenerative energy of
neither of the trains is used. Scenario B represents an optimal situation, where power peaks
are low, regenerative energy is almost completely used by another train and therefore the
energy consumption at the power substation (PSS, lowest subfigure) decreases (integral of
power over time, i. e. area between abscissa and power curve above 0).
The timetable of a railway system strongly influences the interaction of many trains and
therefore power at the substations and the usability of regenerative energy from braking. In
advanced clock face timetables as applied in Switzerland, regenerative energy from braking
can hardly be used because many trains arrive by design at the main stations at the same time
(synchronous braking). Additionally, those timetables have a very high power demand because
many trains start at around the same time (e. g. around the full hour).
Conventional railway timetables cause significantly fewer power peaks and braking energy can
be recuperated frequently. The electrical interaction of many trains could even be taken into
consideration during timetabling. In mainline systems, the co-ordination of starts and stops in
order to use regenerative energy from braking is still not considered in practice, which may be
explained by the many stochastic influences on operations.
In highly automated, dedicated mass rapid transit systems, trains are able to adhere to
the timetable with the precision of a few seconds. Here, a slight modification of timetable
parameters such as train headways, times at line termini, or dwell and running time reserves
can lead to a reduction of electricity costs. One possibility is the application of Genetic
Algorithms to find a good distribution of time reserves, described in Albrecht (2004).
Miyatake & Ko (2010) not only vary the running times, but modify also the driving behaviour
between stations for optimal use of regenerative braking with on-board energy storage
devices. They apply dynamic programming or use a sequential quadratic problem approach
for this kind of substation-related optimisation process.
As the calculation of the cost function is in any case computationally very expensive, efficient
methods for on-line control of a fleet of trains still need to be developed.

104
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.3 Measures to Introduce Energy-Efficient Driving

Some railway operators with such a billing scheme prefer the installation of energy-storage
systems in the power substations, in order to store the regenerative energy from braking and
release it in times of high demand.

Fig. 5.8: Example for good (B) and bad (A) synchronisation of two trains running over a short
distance

5.3 Measures to Introduce Energy-Efficient Driving

Although the fundamentals of energy-efficient driving have been known for many years, the
interest of RUs in them has only started to grow with the significant increase of energy costs.
One reason is that punctuality is recognised, in general, as a first priority and arriving early is
estimated as being less critical than being late. The use of
– driver training,
– driver support systems, and
– automatic train operation
help realise the theoretically optimal train driving strategies in practical operation. Possible
effects are outlined at the end of this Section.

5.3.1 Driver Training

Teaching energy efficient train driving is part of the basic education of train drivers: Here,
they get to know the optimal regimes and get taught that coasting may reduce energy
consumption. On the other hand, finding the optimal coasting point is a difficult task, always
bearing in mind that being late is to be avoided. Only by repeatedly driving the same train with
the same timetable do the drivers learn that optimal point.

105
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

One way of speeding up learning is driving under supervision in a train simulator. Here, train
drivers get accustomed to energy-efficient driving without wasting real energy. Simulators are
part of regular driver training anyway.

5.3.2 Driver Support Systems

5.3.2.1 Technical Classification

Different tools for driver support have been developed, that may be characterised by the time
behaviour of their input data.
Static systems give offline information to the drivers based on the timetable, such as the
speed at which coasting should start on the basis that the train departed on time. These were
developed during the 1970s and 1980s, when the available processing power of on-board
computers was too low to run algorithms for energy-efficient train control in real-time. The
main disadvantage of this approach is that real-time deviations from the timetable cannot be
taken into account.
With the increasing computing power and availability of onboard measurement equipment,
dynamic train-related systems have been developed. These collect information about the
state of the train. Typical examples for input signals are speed or position of the train, or even
energy consumption. Dynamic train-related systems then give real-time information or advice
to the driver. They are able to react to perturbations of the train on which they are installed.
Typically, most of these are caused by the interaction between trains in the network. They
require action before the conflict occurs. Therefore, pure train-related information systems are
not suitable.
This gap can be closed by dynamic network-related systems. They possess information
about the state of rail traffic in the vicinity of the train concerned. These systems could only be
developed with the broad availability of standardised communication channels and equipment
such as GSM or GSM-R. Those channels are used to transmit data between the train and the
dispatching system in the central control. Thereby, the control centre can transmit information
about the time windows that the train needs to aim for, in order to avoid conflicts with other
trains by anticipatory driving.
Furthermore, the train could be sent information about the current state of traffic in the railway
network, as well as predictions on its future state, such as the position of other trains and
signal aspects. More information about train properties in the control centre, of train mass
or current speed, can lead to improved dispatching decisions, giving fewer delays and lower
energy consumption.
Advisory and information systems can also be classified according to the degree of integration
with safety-relevant systems like ATP. There are standalone (advisory) systems which work
independently from any safety system. Special provisions must be made such that the
information or advice given does not contradict with the goal of safe operation. There are also
systems which are partly or fully-integrated in safety components, which eliminates some or
most of the safety risks of standalone systems.
Increasing the volume and quality of data used by the systems (from static train-based
systems to dynamic network related systems) and the degree of integration with the safety
system, increases the accessible potential for energy saving. This is due to the fact that the
system can provide adequate support in more situations. On the other hand, the technical
requirements and the cost of the driver support also increase.

106
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.3 Measures to Introduce Energy-Efficient Driving

The final decision on what kind of system to install and apply depends on the specific
operating conditions such as the frequency of delays and of conflicts, the accessible technical
equipment and economic constraints.

5.3.2.2 The Role of the Train Driver

Drivers, in general, are controlling the trains. That is why it is important to make them
sensitive to energy consumption, energy costs, and the impact of their personal driving style
upon it. This sensitisation favours the acceptance of tools or instruments to reduce energy
consumption.
Once the drivers are sensitive to energy consumption, they will reclaim support systems for the
difficult task of energy-optimal train control. There are two basically different means of driver
support (Table 5.1):
1. Information systems provide the driver with raw process data.
2. Advisory or assistance systems use process data to compute actual control variables to
be used by the driver.

Driver Information System Driver Advisory System (DAS)

Raw process data, Advice calculated from process data, prescriptive
descriptive information information
Examples: Delay of own train, energy consumption Example: Currently optimal driving regime (e. g.
of own train, state of traffic in the network coasting), target or coasting speed, tractive or
braking effort
Human intelligence needed to translate information No intelligence needed
into control sequence
Learning phase to correctly interpret the given Advice easy to interpret
information
Working environment improves and motivation Eliminating degree of freedom in the working
increases with the additional information process may be disliked (degradation of the
drivers to simple executives)
Some drivers may not be willing, some may not be Some drivers may not be willing to use the system
able to use the system
Accessible potential for energy saving limited by Accessible potential for energy saving limited by
human intelligence (e.g. in complicated conflict the underlying algorithms and credibility of the
situations) given advice

Table 5.1: Comparison of Information and Advisory Systems

There is no general rule to decide on the optimal means of support; often a combination
of the two strategies is the most successful. As the driver plays a major role in the control
loop (Fig. 5.1), he should be actively involved in the decision on how he is to be supported.
Otherwise, his acceptance of the support system may be unsatisfactory and the benefits
limited. Organisational methods such as rewarding energy-efficient driving can contribute to a
high acceptance of the support systems.
The reliability of the given information and/or advice is important, because every time the
control strategy derived from the support system’s information turns out to be wrong, the
driver may resist using the system next time. Consequently, for DAS, all data relevant for the
computation of the advice must be reliable. That includes, for instance, data on temporary
speed restrictions (TSRs).

107
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

However, it should be noted that neither information nor advisory systems need to be safe
systems: In the case of standalone or partially integrated systems, they must always be
understood as support systems giving information and advice, but not safety relevant
information. The driver has the duty of not following any given advice that could lead to a
dangerous situation. Additionally, automatic train control systems must work independently to
react to errors by the driver.

5.3.2.3 Examples from Practice

Some examples of practically implemented advisory and information systems are given below.
Static systems give advice about where and when to start coasting (coasting point),
assuming the train departed on time. The means of giving this information may vary,
depending on the kind of traffic. In mainline operation, the coasting information can be given to
the drivers by additional information in the printed operational timetable.
Fig. 5.9 a) gives an example from the Dutch railways used during the 1980s: The hourly
Intercity train series 900 leaves Utrecht Centraal (UT) at minute 34. The coasting reference
point is defined at signal 204. If the train passed this signal at minute 52 of the hour or earlier, it
may start coasting. Under the assumed normal operating conditions, it will then arrive on time
at the next station s’Hertogenbosch (HT, minute 03 of the next hour).

Fig. 5.9: Coasting speed indicated in the timetable (a) – example from the Netherlands (1989) – and
coasting point indicated by a coasting board (b) near the track

In urban mass transit systems, train configuration and running times over a section are usually
constant. Consequently, information on coasting points can also be displayed on coasting
boards near the track (Fig. 5.9 b). This can be found for instance in the metros of Nuremberg
and Berlin in Germany.
In the context of Dynamic Train-Related Systems, Driver Advisory Systems (DAS) are state-
of-the-art and are applied in all kind of railway transportation systems. Technically, they mostly
apply a similar concept, which contains the following components as a minimum (Fig. 5.10):
108
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.3 Measures to Introduce Energy-Efficient Driving

1. a positioning or speed metering device,

2. a clock,
3. a database containing
– timetable: departure, pass and arrival times at all relevant points,
– infrastructure data: the maximum permitted speeds – including temporary speed
restrictions (TSRs), gradients and curve radii to determine track running resistance,
– rolling stock characteristics: acceleration and braking behaviour, train resistances, length
and mass, and
4. a Driver Machine Interface (DMI) to give the advice optically and acoustically and to receive
input from driver about for instance actual train configuration.
For the position detection (1), a combination of absolute positioning sensors like satellite
positioning systems (e. g. GPS) including some map-matching algorithms (digital track map
has to be available) and balise reception and evaluation systems, and relative sensors like an
odometer, radar or inertial measurement units can be used. Kalman filters can be applied for
data fusion. The current time (2) is available on each computing device. Measures must be
taken to ensure that this time is synchronous with trackside clocks at platforms.
The most critical point in a DAS is data integrity in the onboard infrastructure and timetable
database (3): Only gradients and curve radii can be regarded as constant. Timetable
information and speed restrictions usually vary during operation. Therefore processes and
mechanisms for data updates need to be implemented.

Fig. 5.10: Basic design and components of a Driver Advisory System

From the numerous advisory and information systems (see Mitchell (2009) for an overview) a
few typical representatives shall be presented in detail in the remainder of this section. Their
main properties and areas of application are summarised in Table 5.2. The DMIs reflect the
specific application conditions and are therefore distinct.

109
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

EBULa-ESF CATO RouteLint Trainguard MT

EED
Time behaviour of Dynamic, train- Dynamic, network- Dynamic, network- Dynamic, train-
input data related related related related
Advice Coasting Optimal tractive/ – All driving regimes,
countdown, braking effort, distance and time
coasting start optimal speed before start of next
profile regime
Information Energy Timetable, Height Current position, Timetable,
consumption, profile, Speed Timetable Supervised speed
Timetable Limits deviation, Position
deviation, of surrounding
Current Position, trains
Timetable, Speed
Limits
Kind of integration Integrated Standalone Display of Fully integrated in
in electronic signalling system on-board module
timetable display information of CBTC system
Fields of Suburban, Freight, (Suburban) Suburban, Urban and
application Regional, Regional, IC, suburban
Long distance (Freight)
passenger

Table 5.2: Comparison of different driver support systems

The German Railways, DB passenger transportation, use a dynamic train-related system

to support the driver with process information and advice. The system is called EBuLa-ESF
which stands for electronic timetable system including TSR, and energy saving driving advice
Netz 2005. The electronic timetable contains and displays infrastructure data and timetable
information (the latter is updated before the start of a train run by radio transmission).
Furthermore, the display shows characteristic dynamic data of the driving process, e.g.
position of the train (taken from GPS or odometer), deviation from timetable, and energy
consumption (taken from energy meter), cf. blue part in Fig. 5.11. The DAS module compares
the planned and actual time at a given position and if the train is early with respect to the
planned timetable, the ESF module integrated in the software simulates the train run from the
current state until the next target point (station). This simulation may come up with the result
that it is possible to start coasting within the next few seconds and still arrive on time at that
target point. Then a countdown is displayed which decreases from either 10 or 5 to 1. Right
after that, a point appears on the screen, which indicates to the driver that coasting should
start immediately.
The system has been in operation on all ICE trains of DB since May 2005 and was later
extended to the major part of DB passenger trains. The DMI is being amended for the future
application “Zuglaufregelung (ZLR)” as a network-related advisory system, for which the
implementation decision has been made after successful trials. The main objective of ZLR is
to guide trains along train-paths recalculated in real-time to avoid conflicts with other trains.
Trains are guided such that signals turn green before a train arrives. ZLR driving advice
displayed on EBuLa will therefore consist additionally of advice such as “reduce speed by” or
“drive at maximum permitted speed”.
The system CATO (Computer-Aided Train Operation) was originally developed for freight traffic
optimisation on single track lines and contains many aspects of a dynamic, network-related

110
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.3 Measures to Introduce Energy-Efficient Driving

Fig. 5.11: EBuLa-ESF System of Deutsche Bahn AG, Source: Deutsche Bahn AG

system. On single track lines, most energy can be saved if the crossing of trains is optimised,
which is arranged in the central system component “CATO TCC” (Traffic Control Centre).
Based on the current train positions (measured using GPS), target points are calculated
dynamically and transmitted by radio to the on-board systems of each of the trains involved.
The optimal trajectory is computed in the on-board system and displayed to the driver
together with a large amount of other data explained in detail in Fig. 5.12. Showing all this
data increases the awareness of the driver and the acceptance of the system. Because of its
success, it has been installed as train-related advisory system on the dedicated Arlanda airport
express passenger line.
A dynamic, network related driver information system has been developed by the Dutch
infrastructure manager ProRail. The system called “RouteLint” (“Route band”) shows process
data from the control centre on board the trains, Fig. 5.13.
– The middle column displays the journey (sequence of routes) the train is about to follow in
signal steps. Each row represents one signal along the route, thus AH-24A stands for signal
24A in Arnhem station. The last but one row gives the name of the last signal the train has
passed, which is the signal section on which the train is currently.
– The first column indicates the train numbers of services which currently travel that route.
The train 7543 (for which this RouteLint is regarded, last but one row in blue) follows train
2043 and is directly followed by train 47785 (indicated in the last black row). Further on the
route there are the trains 31148 and 3747.
– The rightmost column gives timetable deviation information from all trains on the screen as
available in the traffic control centre.
– The shape in the middle column indicates the relationship with the journeys of other trains.
Thus train 31148 will enter the route of 7543 in AH-24A and leave it already after the next
signal, train 3747 will cross the route after signal FV.

111
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

Fig. 5.12: CATO System DMI as applied on the Iron Ore line in Northern Sweden, Source: Transrail

– The colour of the middle column has the following meaning:

– White: route of own train, which has already been set (The signal at the entrance of this
section does not show stop). Note that the colour green is not used here, for safety
reasons.
– Red: route of own train, which has not yet been set (The signal at the entrance of the
section shows stop, the upper four sections in Fig. 5.15).
– Orange: a section, which is currently not occupied, but is already set for another train
(e.g. sections WF-1 and AY set for train 2043).
– Grey: automatic block section, whose exit signal does not show a stop aspect.
The main purpose of the RouteLint system is to avoid conflicts and stops before red signals,
but it has been shown to contribute also to safe driving by reducing the time in front of signals
indicating warning or other reduced speed aspects (Albrecht & van Luipen (2006)). The system
has been used on the entire network of Netherlands Railway from 2014.
Siemens developed a train-related driver advisory system which is fully integrated with the
Communications-based train control system (CBTC) Trainguard MT, see Rahn et al. (2013) and
as such can use the most comprehensive approach for driver advice without compromising
safety. Although the algorithm for energy-efficient driving (EED) is implemented in the non-vital
part of the CBTC, it can access all vital data and use it for computation of the advice. This is:
– position and speed as computed by the odometry module of the CBTC,
– track database including civil speed profile restrictions (there is always only one valid
version of the database available for all trains),
– the current movement authority, including the temporary speed restrictions and the
resulting braking intervention curves.

112
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 5.3 Measures to Introduce Energy-Efficient Driving

Fig. 5.13: DMI of the driver information system “RouteLint”, Source: ProRail

It can always be guaranteed that a given advice is compliant with the operating conditions.
The system displays all the driving regimes as stated in Section 5.1.1.1 to give maximal
support to the driver. The DMI given in Fig. 5.14 has been designed for the first application
of the system in the Copenhagen suburban railway and includes timetable information. It can
be seen that the speed advice for EED (target speed for application of current regime) is given
using the original CBTC speedometer display.

Fig. 5.14: DMI of the CBTC-system Trainguard MT with functionality for energy-efficient driver
advice, Source: Rahn et al., 2013

113
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

In heavy rail systems, the train control system ERTMS/ETCS offers all relevant technical
characteristics for fully integrated driving advice. However, a standard communication protocol
for sending operational speed control advice from traffic control centres to the trains still
needs to be developed. Thus solutions based on ETCS like the system Admirail in commercial
operation since 2007 in the Swiss Lötschberg tunnel cannot yet use the standardised driver
display to show speed advice, but must rely on a text message display Montigel et al. (2007).

5.3.3 Automatic Train Operation

In order to fulfil safety and operational requirements, a number of high-speed lines and
dedicated urban rapid-transit railway systems with very short headways are operated
automatically to a very high level (Automatic Train Operation – ATO). This requires
sophisticated (and expensive) equipment on-board the trains as well as in the control centre,
which would enable energy-efficient driving as well. Given the very high precision of the
automatic controllers, even challenging strategies like synchronisation of accelerating and
braking trains can be put into practice.
Of course, energy efficiency alone is not a sufficient reason to equip manually driven lines with
ATO. But once an operator has introduced DAS, the efficiency of this system may be improved
by further automation of the train operation using the existing intelligence and equipment of the
DAS. One example of this is the Czech system CRV & AVV, see Lieskovský & Myslivec (2008),
which started with the purpose of energy saving and ended up as a system for fully automatic
(though not driverless) train driving. This is in use on some of the major suburban railway lines
of the country.

5.3.4 Effects of Energy-Efficient Operation

The main effect of energy-efficient driving is of course a reduction in energy consumption. In

the literature, values between 5 % and 25 % savings are reported, see Albrecht (2005), Howlett
& Pudney (1995), Netz (2005), Mitchell (2009).
Another important effect is the reduction in variations of the actual running times and driving
styles. Driving becomes more predictable, which theoretically allows buffer times between
trains to be reduced and thus increase capacity. Generally, energy-efficient driving leads to
arriving later than conventional driving, which also reduces the risk of early departures for
passenger trains through less dwell time at stations. Long coasting phases usually reduce the
braking phase, which leads to less wear-and-tear of brakes, wheels and rails. All these effects
can be observed for all kinds of measures which introduce energy-efficient driving. The biggest
wins are to be obtained using dynamic network-related systems, as they can – on the contrary
to static or purely train-related systems – also reduce the number of unnecessary train stops.

5.4 Conclusions
Research on energy-efficient train operation has resulted in proven concepts and systems
which have shown their potential for energy saving in many practical tests. Dynamic network-
related driver advisory systems have been developed, which permit driver support under all
operating conditions.
However, the operational aspects discussed in this Chapter represent only some of the
possibilities to increase the energy-efficiency of the railways. An extensive database of other
measures is maintained by the International Union of Railways UIC (2013).

114
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

References
Albert H., Levin C., Vielrose E., Witte G. (Eds. (1997): Reducing Energy Consumption in
Underground Systems – an Important Contribution to Protecting the Environment – Volume
3, Part 2. Proc. of the 52nd International Congress of the UITP (International Union of Public
Transport), Stuttgart.
Albrecht A. R., Howlett P. G., Pudney P. J., Vu, X. (2013): Energy-efficient train control: From
local convexity to global optimisation and uniqueness. Automatica 49 (10), pp. 3072 – 3078.
Albrecht T., Oettich S. (2002): A new integrated approach to dynamic schedule
synchronisation and energy saving train control. In: Allan J. et al. (Eds.): Computers in Railways
VIII, pp. 847–856, WIT Press, Southampton.
Albrecht T., Gassel C., van Luipen J., Knijff J. (2009): Analysis of Energy consumption and
traffic flow by means of track occupation data. In: Proc of IET Conf on Railway Traction
Systems (IET RTS).
Albrecht T., van Luipen J. (2006): What role can a driver information system play in railway
conflicts? In: IFAC Control in Transportation Systems, Vol.11, pp. 251 – 256.
Albrecht T., Binder A., Gassel C. (2013): Applications of real-time speed control in rail-bound
public transportation systems. IET Intelligent Transport Systems 7(3), pp. 305 – 314.
Biesenack H., George G., Hofmann G. et al. (2006): Energieversorgung elektrischer Bahnen,
Teubner-Verlag.
Binder A., Albrecht T. (2012): Predictive Energy-Efficient Running Time Control for Metro Lines
In: CASPT12 – Conference on Advanced Systems for Public Transport.
Chang C. S., Sim S. S. (1997): Optimising train movements through coast control using
genetic algorithms. IEE Proc.-Electr. Power Appl., Vol. 144, No.1, pp. 65 – 73.
D’Ariano A., Albrecht T. (2006): Running time re-optimization during real-time timetable
perturbations. In: Allan, J. et al. (Eds.): Computers in Railways X, pp. 531 – 540, WIT Press,
Southampton.
De Cuadra F., Fernández A., de Juan J., Herrero M. A. (1996): Energy-saving automatic
optimisation of train speed commands using direct search techniques. In: Allan J. et al. (Eds.):
Computers in Railways V – Vol., pp. 337 – 346, WIT Press, Southampton.
Franke R., Meyer M., Terwiesch P. (2002): Optimal Control of the Driving of Trains. at-
Automatisierungstechnik, Vol. 50, No. 12, pp. 606 – 613.
Hoffmann C. (2006): Railway energy metering and billing in Europe. eb-Elektrische Bahnen,
No.5, pp. 204 – 208.
Horn P., Zinke R. (1990): Energy-optimal control of trains in long distance traffic. In: 2nd Mini-
Conference on Vehicle System Dynamics, Identification and Anomalies, Budapest.
Howlett P. G., Pudney P. J. (1995): Energy-Efficient Train Control. Springer.
Jaekel B., Albrecht T. (2013): Interfacing Conflict Resolution and Driver Advisory Systems in
Railway Operations. In: 3rd Int. Conference on Models and Technologies for ITS, Dresden.
Joborn M. (2012): Eco-driving boosts Swedish heavy-haul efficiency. IRJ – International
Railway Journal, February 2012, pp. 20 – 24.
Khmelnitsky E. (2000): On an optimal control problem of train operation. IEEE Transactions on
Automatic Control, 45(7), 1257 – 1266.
Lieskovský A., Myslivec I. (2008): CRV & AVV – modular control system for all kinds of vehicles
with special respect to energy savings, (www.azd.cz, retrieved Jan. 2014).

115
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
5 Energy-Efficient Railway Operation

Liu R., Golovitcher I. M. (2003): Energy-efficient operation of rail vehicles, Transportation

Research Part A, vol. 37, pp. 917 – 932.
Mitchell I. (2009): The Sustainable Railway: Use of Advisory Systems for Energy Savings. IRSE
News, no. 151, pp. 2 – 7.
Miyatake M., Ko H. (2010): Optimisation of Train Speed Profile for Minimum Energy
Consumption. IEEJ Transactions on Electrical and Electronic Engineering 5, pp. 263 – 269.
Montigel M., Kleiner C., Achermann E. (2007): Rescheduling in the ETCS Level 2 environment.
In: Proc. of the 2nd International Seminar on Railway Operations Modelling and Analysis
(RailHannover).
Netz M. (2005): Technische Assistenzsysteme der Deutschen Bahn unterstützen Energie
sparendes Fahren. ETR – Eisenbahntechnische Rundschau, Vol. 54 (10), pp. 595 – 598.
Rahn K., Bode C., Albrecht T. (2013): Energy-efficient driving in the context of a
Communications-based train control system (CBTC). In: 1st IEEE International Conference on
Intelligent Rail Transportation (ICIRT), Beijing.
Strobel H., Horn P., Kosemund M. (1974): Contribution to optimum computer-aided control
of train operation. In: 2nd IFAC/IFIP/IFORS Symposium on Traffic Control and Transportation
Systems, pp. 377 – 387, Monte Carlo.
Uher R. A., Sathi N. (1983): Reduction of peak-power demand for electric rail transit systems.
Technical Report 3, Transportation Research Board.
UIC (2013), Railway Energy Database, www.railway-energy.org.
Wang Y., Ning B., Cao F., De Schutter B., van den Boom T. J. J. (2011): A survey on optimal
trajectory planning for train operations, Proceedings of the 2011 IEEE International Conference
on Intelligent Rail Transportation (ICIRT 2011), pp. 589 – 594, Beijing, China.
Zhang L., Zhuan X. (2014): Optimal Operation of Heavy-Haul Trains Equipped With
Electronically Controlled Pneumatic Brake Systems Using Model Predictive Control
Methodology. IEEE Transactions on Control System Technology 22 (1), pp. 13 – 22.

116
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 6.1 Introduction and motivation

6 Queueing
Nils Nießen1

6.1 Introduction and motivation

Railway operations are subject to a wide range of random influences. Hence, use is
frequently made of stochastic models, notably the queueing theory, when dimensioning
railway infrastructure. Before setting out how the queueing theory can be specifically
applied to research into railway operations, a few of the key principles underpinning it will be
explained. Reference is made to the relevant specialist literature, e. g. Kleinrock & Gail (1996),
Gnedenko & König (1983), Fischer & Hertel (1990) for more detailed treatment of the subject.

Fig. 6.1: Queueing system

Fig. 6.1 shows the basic configuration of a queueing system, comprising a service channel
and a waiting space. Demands enter the service system with a service wish. If the service
channel is already occupied by another customer, the new demands join the queue in the
waiting space. Service is then effected later.
The model assumes that a customer k (k = 1, 2, 3, ...) enters the service system at time ta,k
where
0 < ta,1 < ta,2 < ta,3 < ... (6.1)

The inter-arrival time is the difference between successive arrival times:

t A,k = ta,k − ta,k−1; k > 1 (6.2)

The service time during which the customer has resort to the service channel is given as tB,k .
Let A(t) be the total number of customers who have arrived at the service system at time t
⎪⎧⎪ 0 ; t < ta,1
A(t) = ⎪⎨ (6.3)
⎪⎪ i ; ta, i ≤ t < ta, i+1
⎪⎩

then A(t) is referred to as the arrival process. The service process B(t) is defined by analogy as

⎪⎧⎪ 0 ; t < tB,1

⎪⎪
B(t) = ⎪⎨ i i+1
(6.4)
⎪⎪ i ; ∑ tB,j ≤ t < ∑ tB,j
⎪⎪ j=1 j=1
⎪⎩
For a large number of practical applications, it is now posited that interarrival times tA,k (k =
1, 2, 3, ...) are independent realisations of the random variable TA subject to probability
distribution function (pdf)
FA (t) = P(TA < t) (6.5)

1 The first edition of this Chapter (2008) was written by the late Ekkehard Wendler.

117
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
6 Queueing

An analogous assumption is made for service times tB,k (k = 1, 2, 3, ...), which are regarded as
being the independent realisations of the random variable TB with pdf
FB (t) = P(TB < t) (6.6)

If random variable TA is exponentially distributed with pdf

⎪⎧⎪ 0 ; t≤0
FA (t) = ⎪⎨ (6.7)
⎪⎪ 1– e−t/ tA ; t>0
⎪⎩

then A(t) is what is known as a Markovian arrival process. tA represents the mean inter-arrival
time between demands.
tA can be defined for random pdf’s by means of the following Stieltjes integral
∞
tA = ∫ t dFA (t) (6.8)
0

 = 1/ tA (6.9)

is termed the arrival rate. In the case of exponentially distributed service times with pdf
⎪⎧⎪ 0 ; t≤0
FB (t) = ⎪⎨ (6.10)
⎪⎪ 1– e−t/ t B ; t>0
⎪⎩

with a mean service time tB, the term Markovian service process is adopted by analogy.
As a general rule, tB can be computed as in (6.8).

 = 1/ t B (6.11)

is termed the service rate.

It is possible on the basis of these process definitions to classify service systems using what
is known as Kendall notation. Very generally, this involves representing a service system as
A/B/n/m. ‘A’ stands for the type of arrival process, ‘B’ for the type of service process;
‘n’ indicates the number of service channels and ‘m’ the size of the waiting space. Use is
frequently made of the following abbreviations to indicate the process types:
– M ... Markovian process, exponentially distributed process times
– D ... (deterministic) process with constant process times
– E ... Erlang process, Erlang-distributed process times
– GI ... general independent process, process times are identically independently distributed
(iid)
Systems in which n = 1 are called single-channel service systems, those in which n > 1
multiple-channel service systems. Reference is made to “loss systems” where m = 0, “waiting-
loss systems” where 0 < m < ∞, and “waiting systems” where m = ∞. No further consideration
will be given to loss and waiting-loss systems here.
The Kendall notation M/D/1/∞, for instance, indicates a single-channel waiting system with
exponentially distributed inter-arrival times and constant service times.
With the single-channel waiting system of the general form A/B/1/∞, the waiting time tW,k for
customer k is defined as follows:
⎪⎧⎪ 0 ; k =1
t W,k = ⎪⎨ (6.12)
⎪⎪ max (tB,k−1 + t W,k−1 − t A,k , 0) ; k >1
⎪⎩

118
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 6.2 Scheduled waiting times

The system’s mean waiting time tW in steady state is often of interest

1 k
tW,k = lim ∑ t W,k
k →∞ k i=1
(6.13)

For single-channel waiting systems a finite mean value obtains if

tB 
= = <1 (6.14)
tA 

 is defined as being the level of utilisation of the system.

For many waiting system types, formulae have been extrapolated for computing tW and other
parameters. Cited here is the famous formula by Pollaczek and Chintschin for computing the
mean waiting time in M/GI/1/∞ systems:

 ⋅ ETB2
tW,k = (6.15)
2(1−  )
∞
ETB2 = ∫ t2 dFB (t) (6.16)
0

represents the second moment of service time distribution FB (t) here.

6.2 Scheduled waiting times

Queueing theory approaches can be useful to calculate the scheduled waiting times on a
railway. Scheduled waiting times arise during the process of timetabling when different train
operating companies (TOCs) compete for the railway infrastructure available.
Railway infrastructure consists of lines and nodes, whereas a node can be divided into station
throats and station tracks. Methods how to calculate these scheduled waiting times for lines
and nodes will be highlighted.

6.2.1 Scheduled waiting times on railway lines

The time-distance graph in Fig. 6.2 depicts the blocking time sequences of two conflicting
train paths. The conflict can be resolved by repositioning the second path. The outcome is
a scheduled waiting time tW,k+1. This waiting time is referred to as “scheduled” because the
displacement is already contained in the running schedule. The scheduled waiting time
generally leads to an increase of total transit time (longer stopping time, in some cases longer
running time – “path bending”). If the whole path is displaced at its very starting point, this
does not necessarily give rise to an increase of the total transit time. If it is assumed, however,
that the TOC has chosen the best time of departure from the starting point for its customers
(passengers or freight), any displacement from this optimum train path requested will, at the
very least, result in a longer journey time for these customers.
The minimum headway time on a railway line is the time span a second train has to wait,
after the first departs (cf. Chapter 2). Pushing the blocking-time sequences for two trains on
a line section up close to each other allows the minimum headway time th to be ascertained
(Fig. 6.3). For capacity research purposes, the minimum headway time of two trains is derived
from the beginning of the blocking time of the first block section.

119
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
6 Queueing

Fig. 6.2: Train path conflict and scheduled waiting time

Fig. 6.3: Minimum headway time

It is now assumed that nT different types of trains are using the railway line concerned. In this
case, nT x nT different minimum headway times th,ij can be determined and presented in matrix
form
⎛⎜ t th,12  th,1nT ⎞⎟⎟
⎜⎜ h,11 ⎟⎟
⎜⎜ t t  th,2nT ⎟⎟⎟
Th = ( th,ij ) = ⎜⎜⎜
h,21 h,22
⎟⎟ (6.17)
⎜⎜     ⎟⎟
⎜⎜ ⎟⎟
⎜⎜⎝ th,nT 1 th,nT 2  th,nTnT ⎟⎟⎟
⎠

The train paths requested by the TOC can now be seen as customers of a queueing system.
Let ta,k be the requested start of the blocking time in the first block section for train path k
requested by the TOC and let  (k) be the type of this train path.

120
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 6.2 Scheduled waiting times

Assume that
ta,k < ta,k+1  k (6.18)

The inter-arrival times tA,k between consecutive requested train paths can be determined by
(6.2). It is vital that the tA,k are determined on the basis of train paths requested by the TOCs
instead of existing, already compiled timetables. If
t A,k < th,(k ),(k +1) (6.19)

the consecutive train paths are in conflict. Compiling a timetable involves making a decision
in the event of a conflict as to which train is to be left as it is and which is displaced, although
it can mean making an adjustment to both. For the moment it is assumed that this decision
is made on the basis of a first-in first-out discipline (FIFO) respecting the time order of the
requested paths in the first block section. Moreover, it is posited that paths may only be
inserted at a later point than the requested path, never at an earlier one. Now, the scheduled
waiting time tw,k+1 can be calculated by
t w,k +1 = max (th,(k−1)(k ) + t w,k−1 − t A,k , 0) (6.20)

which is similar to (6.12). Hence, the minimum headway times correspond to the service times
that were already proposed by Schwanhäußer (1978).
For long and medium term studies, neither the concrete values of the tA,k nor the concrete
order of the requested train paths are known in detail. That is why the tA,k must be modelled as
iid random variables. In particular, if there is a larger number of independent TOCs competing
for infrastructure capacity and if there are no (cyclic) dependencies between the requested
train paths, the arrival process can be modelled as a Markovian one with intensity . Thoughts
on how to model the arrival process are set out in Wendler (1998). The service time tB,k of the
k-th customer of type  (k) = i is described by a conditional pdf
P (tB,k ≤ t (k + 1) = j) = pij ⋅ th,ij (t) (6.21)

where
⎧⎪ 0 if t ≤ x
⎪
 x (t) = ⎪⎨ (6.22)
⎪⎪ 1 otherwise
⎪⎩
represents the D-distribution and the pij are the conditional train sequence probabilities
pij = P ((k + 1) = j (k) = i) (6.23)

The pij form a stochastic matrix

P = (pij) (6.24)

In practical applications often a “random mix” of trains is assumed; for this special case it is
posited that
pij = p j (6.25)

where pj is the probability of a train of type j appearing in the timetable configuration.

For many practical cases, the formula of Pollaczek and Chintschin already represents a
good steady state estimation of the mean scheduled waiting time. The parameters can be
calculated as follows: nT nT
tB = th = ∑ ∑ pE,i ⋅ pij ⋅ th,ij (6.26)
i=1 j=1
nT nT
ETB2 = ∑ ∑ pE,i ⋅ pij ⋅ th,ij
2
(6.27)
i=1 j=1

121
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
6 Queueing

where pE = (pE,i ) is the stochastic left eigenvector of P

pE = pE · P (6.28)

with the normalisation nT

∑ pE,i = 1 (6.29)
i=1
It may be noted that if the special case of eq. (6.25) is modelled, then
pE,i = pi ∀ i (6.30)

holds and equations 6.26 and 6.27 simplify to

nT nT
tB = th = ∑ ∑ pi ⋅ p j ⋅ th,ij (6.31)
i=1 j=1

nT nT
ETB2 = ∑ ∑ pi ⋅ p j ⋅ th,ij
2
(6.32)
i=1 j=1
In Wendler (2007) it was shown that (6.21) constitutes the kernel of a Semi-Markovian service
process (SM), so that the service system can be modelled as an M/SM/1/∞ queue. The mean
scheduled waiting time can be calculated by a closed solution derived by Neuts (1977). It
can be shown that (6.15) is a light-traffic approximation for the waiting time of the M/SM/1/∞
queue, see Wendler (1999).
In equation (6.15), a FIFO discipline without priorities is assumed. In practical train path
management, however, consideration will also be given, when deciding which path is to be
moved, to the priorities of the paths requested. There is a wide range of priority systems
to be observed in practical scheduling that are adopted with a greater or lesser degree of
consistency, or could be adopted in future:
– priority for special types of trains (e. g. freight paths on dedicated freight lines),
– speed bundling (priority for trains running with a harmonised speed),
– other legal priorities,
– priorities as a function of the infrastructure charge,
– so called “grandfather rights” (e. g. framework agreements).
Schwanhäußer (1978) derived an approximation dealing with train priorities by replacing
equation (6.27) by
nT nT 2
ETB2 = ∑ ∑ pE,i ⋅ pij ⋅(th,ij + td,ij ) (6.33)
i=1 j=1

where ⎧⎪ +t
⎪⎪ h,ji r(i) > r(j)
⎪⎪
td,ij = ⎨ 0 r(i) = r(j) (6.34)
⎪⎪
⎪⎪ −th,ij r(i) < r(j)
⎪⎩
r(i) ... rank number of train type i,
r(i) ∈ N ∀ i (6.35)

Note that it is assumed that a high rank number represents a high priority train. Sometimes, a
reverse definition is used in the literature.
tB calculated by (6.26) is not affected by the introduction of priorities.
Wakob refined Schwanhäußer’s approximation by moving from a Markovian arrival process to
-distributed inter-arrival and service times,

tW = ⋅ tB (6.36)
1− 

122
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 6.2 Scheduled waiting times

where  can be found recursively from the equation

⎛ 1−  ⎞⎟−k
 = ⎜⎜⎜1+ ⎟⎟ (6.37)
⎜⎝ k ⋅  ⎟⎠
The auxiliary variables are defined as follows:
1
 = 1− (6.38)
⎡  ⎤ y2
y1 ⎢
⎢ 2(1−) (
1+ vB2 )⎥⎥ +1
⎣ ⎦
y1 = 1− 0,8060(1− x A )(1− xB ) − 0,1940(1− x 0,2924
A )(1− xB0,2924 ) (6.39)

y 2 = 1− 0,1278((1− x A )(1− xB ))1,1375 (6.40)

2 ⋅ v 2A
xA = (6.41)
1+ v 2A

2 ⋅ vB2
xB = (6.42)
1+ vB2

The auxiliary variables y1 and y2 were estimated by Wakob using exact results for the E/E/1/∞
queue, see Wakob (1985). v 2A and v B2 , the squared coefficients of variations of the inter-arrival
and service times, can be derived from the mean and the second moment of the random
variables TA and TB, respectively:
ETA2 − t 2A
v 2A = (6.43)
t 2A

ETB2 − t 2B
vB2 = (6.44)
t 2B

For  g 1, Wakob’s formula converges to Gudehus’ well known heavy traffic approximation for
the GI/GI/1/∞ queue, see Gudehus (1976),

2 v 2A + vB2
tW = tA ⋅ ⋅ (6.45)
1−  2
Wakob’s model is used as the standard approach for the estimation of the scheduled waiting
times by means of queueing models in Germany. Wakob’s formula was implemented in the
software tool ALFA for the first time, and is nowadays used in the successor tool LUKS for
analytical capacity assessment of junctions and sub-networks, see Vakhtel (2002), Nießen (2013).

Fig. 6.4: Waiting time function

123
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
6 Queueing

The basic function curve for the waiting time functions

tw = f(nTr ) (6.46)

produced by waiting time models, where nTr is the number of trains, is shown in Fig. 6.4 for a
single-channel system. The function has a pole at
nTr,max = tI / tB (6.47)

where tI is the investigation period. At nTr,max the level of utilisation  of a single-channel

systems equals one.
tB (6.48)
=
 = 1
tA

nTr,max is designated as being theoretical capacity. Such theoretical capacity is of no relevance

for practical purposes, however, as the level of quality – scheduled waiting time in this instance
– is infinitely poor in theory. If, conversely, the admissible level for scheduled waiting times
,adm is known, a practical capacity nTr,opt can be specified. The admissible waiting time must
ttww,adm
be derived from the specific transport market conditions. The competitive situation with other
carriers on the route being examined impacts to a particularly great extent on the waiting times
that customers will accept before modal shift occurs.
The following formula for admissible scheduled waiting times (for line sections and route
nodes) is applied in Germany:
−1,3p PTr
0,479 ⋅ e
tw,adm = (6.49)

pPTr ... relative frequency of passenger trains.

0 ≤ pPTr ≤ 1 (6.50)

Formula (6.49) with its parameters was derived from correlation calculations based on Delphi
interviews with German railway dispatchers in the 1980s. The dispatchers were requested
to describe the levels of service on their railway lines; the answers were correlated with the
utilisation of the railway lines at that time.

6.2.2 Scheduled waiting times at railway nodes

A route node comprises a set of station tracks, including the switching zones in station
throats. Route nodes link the adjoining lines with the set of station tracks. Fig. 6.5 shows the
division of a node into route nodes and a set of station tracks in schematic form.

Fig. 6.5: Route nodes (RN) and station tracks (ST)

124
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 6.2 Scheduled waiting times

For reasons of capacity calculation, route nodes are divided into smaller units, so called serial
route nodes (cf. Chapter 3). Serial route nodes along which all movements are mutually exclusive
have the property of being single-channel service systems, see Schwanhäußer (1978).

Fig. 6.6: Subdivision into serial route nodes

Use is made of Wakob’s formula to assess the waiting time at serial route nodes. The
assumption that all serial route nodes act independently of each other holds for the
identification of infrastructure bottlenecks but is not valid to determine the overall waiting times
of a railway node. Schwanhäußer (1978) reduces a route node into an auxiliary single channel
system by introducing the concatenation value . Thereby the inverse of  identifies the
average number of effective channels, see Potthoff (1962). For computing the mean waiting
time of a route node the Pollaczek and Chintschin formula (6.15) is applied.
2
 ⋅ ETB,RN
tW = (6.51)
2(1−  ⋅ tB,RN )
with
nT nT
tB,RN = ∑ ∑ pij ⋅ th,ij (6.52)
i=1 j=1

nT nT
2
2
ETB,RN = ∑ ∑ pij ⋅(th,ij + td,ij ) (6.53)
i=1 j=1

⎪⎧⎪ p ⋅p if train i excludes train j

pij = ⎨ i j (6.54)
⎪⎪ 0 else
⎪⎩

The theoretical capacity of a route node can be calculated by using a multi-resource queue
approach, see Nießen (2007).
Let g be the number of different customer types occupying one or more channels in a route
node. Customer types are designated as j = 1... g. The service rate of customer type j is
denoted as j . The overall arrival rate of the system is  ; the arrival rate of customer type j is j
respectively.
g
 = ∑ j (6.55)
j=1
The proportion of the arrival rate of customer type j to the overall rate is described by the
factor fj .
j
fj = (6.56)


125
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
6 Queueing

The system of a route node comprises of a set of single servers as shown in Fig. 6.6. Either
the system is empty or else at least one customer is being served. Whether more than one
customer can be served at the same time depends on the customers’ requirements. Using
customer combinations c the system’s statuses can be described. A customer combination c
is possible if all customers of this combination can be served at the same time. Let m be the
number of all possible customer combinations with at least one customer. The trivial scenario
in which there is no customer in the system is defined as c0 . Combinations are numbered
consecutively as l = 0...m. Combination cl states which customers occur in combination l. A
combination is described in the form of
ccl l = (c1l,c2l,...,c jl,...,cgl ) (6.57)

where
⎪⎧ 1 if combination l involves customer of type j
c jl = ⎪⎨ (6.58)
⎪⎪ 0 otherwise
⎩
The trivial combination arises when
c j0 = 0 ∀ j = 1...g (6.59)

The system’s stationary status  can be described in terms of status probabilities. The system
will be in the status of combination l with a certain probability πl . The probability of there being
no customer in the system is denoted by means of π0 .
 = {0, 1, 2,..., l,..., m } (6.60)

In respect of status probabilities, it holds that

0 ≤ l ≤ 1 (6.61)

and m
∑ l = 1 (6.62)
l=0

The theoretical capacity of the system nTr,max can be derived from the maximum arrival rate max

nTr,max = max ⋅ tI (6.63)

where tI is the investigation period.

As an objective function the arrival rate  has to be maximized by means of linear optimisation.
The status probabilities πl depend on the maximum arrival rate. The optimisation problem can
be described as follows:
⎪⎧ j m
max = max ⎪⎨ |  = ⋅ ∑ c jl ⋅ l (max ) ∀ j = 1…g,
⎪⎪ fj l=1
⎪⎩
m ⎪⎫ (6.64)
0 ≤ l (max ) ≤1∀ l = 1…m, ∑ l (max ) = 1⎪⎬
⎪⎪
l=1 ⎭
Further details on assessing the theoretical capacity and waiting times of route nodes can be
found in Nießen (2008).

6.2.3 Set of tracks as multiple server queues

It is relatively easy to adopt queueing theory models for use in the science of railway
operations, as long as infrastructure elements can be modelled as single-channel service
systems (serial route node, section of line). For sets of platform tracks at stations and of

126
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 6.3 Estimating knock-on delays

reception and departure sidings in marshalling yards etc., it becomes more difficult. The scope
for reciprocally representing the individual tracks in a set of tracks leads to a reduction of
waiting times over a situation in which each movement is rigidly assigned to a given track.
It is possible to calculate the scheduled waiting times arising in a set of n tracks with the aid
of n-channel service systems. An approximation for the GI/GI/n/∞ waiting system after Hertel
(Fischer & Hertel, 1990) has asserted itself in railway operations science:
n ⋅  ⋅
tW = tA ⋅ p 0 ⋅ (6.65)
n!(1− )2
where

 (6.66)
= <1
n
2 (6.67)
=
v 2A + c ⋅ vB2

⎪⎧⎪ 1 ; v A ≥1
⎪⎪
c = ⎪⎨ 2
⎛  ⎞1−v A (6.68)
⎪⎪ ⎜⎜ ⎟⎟ ⋅(1+ v 2A )− v 2A ; v A <1
⎪⎪ ⎜⎝ n ⎟⎠
⎪⎩
⎡ n i n ⋅ ⋅ ⎤−1
p0 = ⎢⎢ ∑ + ⎥
⎥ (6.69)
⎢⎣ i=1 i! n!(1− ) ⎥⎦

p0 represents the empty probability of the system, i. e. the probability of all tracks being empty.

6.3 Estimating knock-on delays

The correlation between utilisation and scheduled waiting time represented in Fig. 6.4 can
likewise be observed during operations management. Trains suffering initial delays before
entering an area will, under certain circumstances, experience further delays originating within
that area. These delays may trigger knock-on effects to other trains within the timetable
structure. The amount of the knock-on delay, too, is dependent on the level of utilisation and
accordingly constitutes a utilisation-dependent quality parameter. This can, therefore, be used
to assess the robustness of a timetable on a given railway infrastructure.

Fig. 6.7: Propagation of delays

127
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
6 Queueing

Fig. 6.7 portrays a situation in which a late running train of train type i enters a line and hence
triggers a knock-on delay to the next train of train type j . Assuming that there is a buffer time tb
between the blocking times and assuming provisionally that the first train is suffering an initial
delay td1,i , the knock-on delay to the second train td2,j is calculated as follows:
⎪⎧⎪ td1,i − tb ; tb < td1,i ≤ tb + th,ij
td2,j = ⎪⎨ (6.70)
⎪⎪ 0 ; else
⎪⎩
If the initial delay is greater than tb + th,ij , a change of sequence occurs at the beginning of
the stretch of line that both trains are following. The upshot is that train i as opposed to train
j suffers a knock-on delay (it being posited here that the two trains are equal-ranking). If the
initial delay is smaller than buffer time tb or greater than tb+th,ij+ th,ji , no knock-on delay arises in
this two-train model.
For mid-term to long-term dimensioning tasks, however, neither the actual timetable
structure (inclusive of sequences of trains and levels of buffer times) nor the actual initial
delays anticipated are known here. Schwanhäußer has thus developed a formula for the non
timetable-dependent approximation of anticipated knock-on delays. In this, input variables
that have yet to be determined in detail are modelled as random variables, see Schwanhäußer
(1974). The mean knock-on delay to be expected on a railway line is arrived at:
⎛ 2
p2 ⎞⎟ td1
td2 = ⎜⎜⎜pd1 − d1 ⎟⎟⎟ ⋅
⎜⎝ 2 ⎟⎠ ⎛ −t / t ⎞
tb + td1 ⋅⎜⎜⎜1− e h d1⎟⎟⎟ (6.71)
⎝ ⎠
⎡ ⎛ −t /t ⎞
2 th,dr ⎛ −2 t /t ⎞ th ⎛ −t / t ⎞
2⎤
⋅ ⎢⎢per ⋅⎜⎜⎜1− e h,er d1⎟⎟⎟ + (1−per ) ⋅ ⋅⎜⎜⎜1− e h,dr d1⎟⎟⎟ + ⋅⎜⎜⎜1− e h d1⎟⎟⎟ ⎥⎥
⎢⎣ ⎝ ⎠ td1 ⎝ ⎠ tb ⎝ ⎠ ⎥
⎦
The auxiliary variables are defined as follows:
pd1 … probability of an initial delay
td1 … mean initial delay
th … mean minimum headway time
nT nT
th = ∑ ∑ pi ⋅ p j ⋅ th,ij (6.72)
i=1 j=1
per … probability of equal-ranking train succession scenarios occurring
nT nT
per = ∑ ∑ pi ⋅ p j ⋅ erij (6.73)
i=1 j=1

where ⎧⎪ 0 ;r(i) ≠ r(j)

⎪
erij = ⎪⎨ (6.74)
⎪⎪ 1 ;r(i) = r(j)
⎪⎩
th,er … mean minimum headway time for equal-ranking trains
nT nT
∑ ∑ pi ⋅ p j ⋅ erij ⋅ th,ij
i=1 j=1
tB,er = (6.75)
per
th,dr … mean minimum headway time for different-ranking trains
nT nT
∑ ∑ pi ⋅ p j ⋅(1− erij ) ⋅ th,ij
i=1 j=1
tB,dr = (6.76)
1− per

128
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 6.3 Estimating knock-on delays

The initial delays are taken as being random variable Td1 with the following pdf
⎧⎪ 0 ; t<0
⎪⎪
Fd1(t) = P(Td1 < t) = ⎨ (6.77)
⎪⎪ 1−p ⋅e −t/ td1
⎪⎪⎩ d1 ; t≥0
which represents a mixture of D(eterministic) and exponential distribution (Fig. 6.8). pd1 and
td1 form the parameters of this distribution. The initial delays are observed in model form as
occurring at the beginning of the line. This simplification is admissible for mid-term and long-
term studies as long as the railway line being considered is not more than about 100 km long
in total. In such cases, any original delays occurring along the line are combined with the initial
delays obtaining at entry onto the line as a makeshift measure.
For detailed discussion of delay distributions, reference is made to Chapter 11.

Fig. 6.8: Pdf of initial delays

t b represents the mean buffer time in the timetable structure where

tI
tb = − th (6.78)
nTr

tI ... duration of investigation period (e. g. one day).

Given that the buffer times are not as yet known in detail, exponentially distributed buffer times
Tb are assumed in (6.61),
⎧⎪ 0 ; t≤0
⎪⎪
Fb (t) = P(Tb < t) = ⎨ (6.79)
⎪⎪ 1− e−t/ tb ; t>0
⎪⎩⎪
(Schwanhäußer, 1974) discusses timetables with non-exponentially distributed buffer times,
too. This needs to be borne in mind, particularly if the future timetable is expected to involve a
high degree of clock face operation.
Fig. 6.9 shows the function curve for (6.61), which resembles that in Fig. 6.4. Knock-on delays
are dependent not only on the number of trains nTr but also on the extent of initial delays.
Where initial delays are great in extent, the knock-on delays increase accordingly (broken line)
whereas, if they are of a lesser magnitude, then the attendant knock-on delays are also less
pronounced (dotted line). As with scheduled waiting times, pole nTr,max is defined exclusively by
means of minimum headway times.

129
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
6 Queueing

Fig. 6.9: Function of knock-on delays

In Germany,
−1,3p
0,257 ⋅ e PTr
td2,adm = (6.80)


is set as an admissible level of knock-on delays for the assessing of railway lines. By analogy
with (6.49), formula (6.70) is also based on Delphi interviews with railway dispatchers (cf.
Chapter 6.2). (6.61) and (6.70) have been put to effect in the module STRELE of the LUKS
software tool (Janecek, 2012) widely used in Germany for performance capacity studies; (6.61)
is thus often dubbed the “STRELE formula”.

6.4 Summary and future prospects

Queueing approaches calculating the scheduled waiting times and/or the knock-on delays
as a utilisation-dependant quality parameter are nowadays used in the medium-term
infrastructure planning and in the long-term strategic network planning on a grand scale. The
queueing models show their power if the concrete parameters of future operation programmes
are not yet known in detail, and are afflicted with uncertainty. The random variables used in
the queueing approaches are best suited to derive sustainable results, despite the existing
uncertainties.
There is a limit to using the models based on scheduled waiting times in the case of a heavily
dependant arrival process. Such dependant arrival processes occur if integrated regular
interval timetables are requested by the TOCs. Nevertheless, the formulae mentioned above
can be used as good approximations, even in the case of regular interval timetables, if several
independent cyclic patterns are superimposed on a couple of non-cyclic train paths. This is
very often the case at large railway junctions.
The containment of dependencies will be one of the major challenges for the application of
queueing approaches in railway operations research in the future. The potential gains offered
by queueing theory are still not fully exploited.

130
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

References
Fischer K., Hertel G.: Bedienungsprozesse im Transportwesen. Transpress Verlagsgesellschaft,
Berlin (1990).
Gnedenko B. W., König D.: Handbuch der Bedienungstheorie. Vols I and II. Akademie-Verlag
Berlin (1983).
Gudehus T.: Staueffekte vor Transportknoten. Zeitschrift für Operations Research. 20 (1976),
B 207 – B 257.
Janecek D., Kuckelberg A., Nießen N.: Kapazitätsermittlung von Eisenbahnknoten und
Strecken. Eisenbahntechnische Rundschau (ETR) 61 (2012) 10, pp. 30 – 36.
Kleinrock L., Gail R.: Queueing theory. Wiley, New York (1996).
Neuts M. F.: Some explicite formulas for the steady-state behavior of the queue with Semi-
Markovian service times. Adv. Appl. Prob. 9 (1977), pp. 141 – 157.
Nießen N., Wendler E.: Grenzleistung von Gesamtfahrstraßenknoten. Verkehrswissen-
schaftliche Tage, Vol. 21, Dresden (2007).
Nießen N.: Leistungskenngrößen für Gesamtfahrstraßenknoten. RWTH Aachen (2008).
Nießen N., Schaer T.: Dimensionierung der Infrastruktur – Betriebswissenschaftliches Know-
How für den Bahnbetrieb. Deine Bahn 42 (2012) 5, pp. 42 – 45.
Nießen N.: Waiting and loss probabilities for route nodes. Proc. of the 5th International
Seminar on Railway Operations Modelling and Analysis, Copenhagen (2013).
Potthoff G.: Verkehrströmungslehre, Vol. 1: Zugfolge auf Strecken und Bahnhöfen. Transpress,
Berlin (1962).
Schwanhäußer W.: Die Bemessung der Pufferzeiten im Fahrplangefüge der Eisenbahn.
Veröffentlichungen des Verkehrswissenschaftlichen Instituts der RWTH Aachen, Vol. 20 (1974).
Schwanhäußer W.: Die Ermittlung der Leistungsfähigkeit von großen Fahrstraßenknoten und
von Teilen des Eisenbahnnetzes. AET, Vol. 33 (1978), pp. 7 – 18.
Schwanhäußer W.: The Status of German Railway Operations Management in Research and
Practice. Transp. Res. A, 28 (1994) 6, pp. 495 – 500.
Vakhtel S.: Rechnerunterstützte analytische Ermittlung der Kapazität von Eisenbahnnetzen.
Veröffentlichungen des Verkehrswissenschaftlichen Instituts der RWTH Aachen, Vol. 59 (2002).
Wakob H.: Ableitung eines generellen Wartemodells zur Ermittlung der planmäßigen
Wartezeiten im Eisenbahnbetrieb unter besonderer Berücksichtigung der Aspekte
der Leistungsfähigkeit und der Anlagenbelastung. Veröffentlichungen des Verkehrs-
wissenschaftlichen Instituts der RWTH Aachen, Vol. 36 (1985).
Wendler E.: Scheduled waiting time and uncertain traffic demand. Proc. of the Symposium
“Planning Traffic Networks for a Dynamic Transport Market – Rail Traffic Compared to Air
Traffic”, Utrecht, (1998), pp. 61– 68.
Wendler E.: Analytische Berechnung der planmäßigen Wartezeiten bei asynchroner
Fahrplankonstruktion. Veröffentlichungen des Verkehrswissenschaftlichen Instituts der RWTH
Aachen, Vol. 55 (1999).
Wendler E.: The scheduled waiting time on railway lines. Transp. Res. B, 41 (2007) 2,
pp. 148 – 158.

131
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 7.1 Introduction

7 Timetable Stability Analysis

Rob M. P. Goverde

7.1 Introduction

Reliable train operations require robust and stable timetables, with built-in slack to avoid and
reduce delays and their propagation. Informally, a stable timetable means that any delay in the
network can be settled after some time without adjusting the timetable. Stability thus refers to
the self-regulating behaviour of the timetable after disruptions. The issue of timetable stability is
rapidly gaining attention because of the increasingly saturated European railway infrastructure,
where a slightly delayed train may cause a domino effect of consecutive delays over the entire
network. European railways are typically operated according to a predetermined master
timetable, which represents a conflict-free co-ordination of train paths and includes slack
time to manage train delays. From a system point of view, the timetable can be viewed as a
steady state for railway traffic. Stability then refers to the ability to settle delays and return to
the (periodic) steady state after disruptions.
Timetable stability can be analysed using network-wide simulation, see Nordeen (1996) and
Middelkoop & Bouwman (2002), or analytically using max-plus algebra, see Braker (1993),
Goverde (2005) and Heidergott et al. (2006). The essential feature of scheduled railway traffic
is the synchronisation of train paths at stations and time separation of train paths on conflicting
routes. This can be captured by precedence constraints between train events, where an
event may occur if all preceding activities have been completed. The result is a discrete event
system that can be modelled efficiently using the so-called max-plus algebra. Within the max-
plus framework, timetable properties can be evaluated very effectively. Thus stability can be
assessed by computing an eigenvalue of the state matrix describing the scheduled railway
traffic in max-plus algebra. Moreover, the approach does not only answer whether or not
the timetable is stable, but also identifies the critical processes in the railway traffic network.
The network interdependencies and the effect of timetable slack hidden in the timetable
within running time supplements and buffer times become visible in the max-plus framework.
This results from critical path computations over the network defined by the precedence
constraints. An application of this is the propagation of an initial delay scenario over the
network in space and time, where the settling time is a (local) stability measure.
Max-plus algebra is a particular example of an idempotent semiring, or dioid, originating
from the 1960s as a formal algebraic structure for solving a range of path problems in graph
theory. Many graph algorithms are equivalent to classical algorithms for solving linear systems
of equations when formulated in the appropriate algebra, see Gondran & Minoux (1984).
Cuninghame-Green (1979) presented a pioneering treatment of max-plus algebra, showing
many counterparts to linear algebra with interpretations in path problems over graphs. Cohen
et al. (1985) showed that a subclass of Petri nets – the decision-free Petri nets or timed
event graphs – can be represented by (recursive) linear systems in max-plus algebra. They
also explained how the periodic steady-state behaviour of these discrete event systems are
characterised by solving an eigenproblem in max-plus algebra. This led to a renewed interest
in max-plus algebra accumulating to a max-plus linear system theory of discrete-event dynamic
systems, as presented in Baccelli et al. (1992). This was termed analogous to linear system
theory of dynamic systems governed by differential and difference equations. Heidergott et al.
(2006) is a recent introductory textbook on max-plus algebra and its applications.
In the early 1990s, Braker (1993) applied the max-plus approach to the modelling and
analysis of the timetable of the Dutch intercity network. This work started a series of PhD
133
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
7 Timetable Stability Analysis

researches towards the modelling and analysis of periodic railway timetables, see Subiono
(2000), Soto y Koelemeijer (2003) and Goverde (2005). The efficiency of a railway timetable
is typically limited by the railway infrastructure and safety constraints, which have a major
effect on train traffic dynamics. Goverde (2005) extended the max-plus modelling approach
by incorporating headway constraints, which led to very large-scale models since not only
the arrivals and departures at stations had to be modelled but also through times of trains
at all stations and junctions where trains could interact. Fortunately, Cochet-Terrasson et al.
(1998) had developed the breakthrough policy iteration algorithm that computes eigenvalues of
large-scale max-plus matrices in seconds. Together with efficient graph algorithms, large-scale
periodic railway timetables can be analysed for stability extremely fast as is demonstrated
in the software tool PETER, see Goverde & Odijk (2002). This contains the state-of-the-art
implementation of the max-plus model building and analysis algorithms, as well as a graphical
user-interface which allows users to change parameters and view the results in graphical
and tabular output. Max-plus algebra has also been applied to railway capacity assessment,
see Ekman (2004) and Van Egmond (1999, 2001). Burkholter (2005) and Herrmann (2006)
developed a heuristic iterative approach to compute stable railway timetables in complex
station areas, using max-plus algebra in combination with a station routing feasibility check.
The max-plus stability analysis can also be applied in a stochastic context. In this case, all the
matrix entries are stochastic variables representing the stochastic free-flow behaviour of the
train processes, while the secondary delays due to interactions between trains are modelled
by the constraints in the max-plus model, see Goverde et al. (2009). In the stochastic setting
the “Lyapunov exponent” of the stochastic matrix equals the expected cycle time of the
system, like in the deterministic setting the eigenvalue represents the minimum cycle time. If
all process times are deterministic, then the Lyapunov exponent reduces to the eigenvalue.
The Lyapunov exponent can be used to assess stability, taking into account the stochastic
variation of the process times. Goverde et al. (2011) give an efficient algorithm to compute an
unbiased estimate of the Lyapunov exponent for stochastic max-plus matrices.
Büker & Seybold (2012) also consider the stochastic network model, where the arcs represent
process times that are modelled as stochastic variables with an extended exponential
distribution. These stochastic process times generate primary delays that propagate through
the network by adding pairs of stochastic variables (convolution) or taking the maximum value
of them. The extended exponential distribution is defined such that the resulting distribution
after addition or taking the maximum is again an extended exponential distribution, with
adjusted parameters that represents the delay propagation. This analytical stochastic model
assumes independent processes and also neglects the dependences caused by combining
two processes that have a shared process upstream in the network. However, the resulting
errors are small if the timetable contains sufficient time allowances. This approach is able to
compute rapidly the stochastic delay propagation over large-scale networks and can be used
to compare the delay propagation of different timetable variants. It is implemented in the tool
OnTime and applied in Switzerland, amongst others.
This Chapter is outlined as follows. Section 7.2 defines some notions of stability. Section 7.3
introduces the max-plus algebra after which in Section 7.4 it is explained how to model a
periodic railway timetable within this framework. Section 7.5 explains the max-plus eigenvalue
problem and its use in max-plus stability analysis, while Section 7.6 considers the recovery
time matrix associated to a periodic railway timetable. Section 7.7 describes a delay
propagation algorithm using the max-plus model. Section 7.8 considers stochastic max-plus
systems. Finally, Section 7.9 gives some conclusions.

134
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 7.2 Stability

7.2 Stability

Timetable stability is a key performance issue associated with the ability of the railway system
to recover from delays. Two types of stability can be distinguished: local stability of an open
system (or sub-network) and global stability (or network stability) of a closed system. An open
system is a sub-network or railway line with trains entering and exiting. An area is locally
stable if the sum of output delays is smaller than the sum of input delays. Here input delays
include trains entering at the system border and trains starting within the system, while output
delays include trains exiting the system and trains terminating in the system. A closed system
is a closed network of train circulations. A system is globally stable if initial delays within the
system can settle in finite time. Local stability can also be applied to closed (sub)systems in
which case all initial delays start within the system and all delays must settle within the system.
Thus, in a closed system local and global stability are equivalent.

Fig. 7.1: Propagation and recovery of delays

To study stability it is essential to know how much slack is available for delay recovery within
a basic period. This problem is related to the UIC timetable compression method of railway
capacity analysis, although for the estimation of stability both the running time supplements
and buffer times between conflicting train paths need to be taken into account. Moreover,
for stability the compression of train paths also must take into account network constraints
such as connections between train lines. The network stability problem can be solved using an
eigenvalue problem in max-plus algebra as explained in Section 7.5. Here it will be shown that
the minimum cycle time that can be obtained by compressing all train paths with respect to all
precedence constraints is equivalent to the maximum mean path weight over all circuits in an
associated precedence graph. The system is stable if this minimum cycle time is smaller than
the timetable cycle time, so that some slack is available in each timetable period.
However, even if the system is globally stable, the distribution of timetable slack may be
disastrous for local stability. Local stability can be evaluated by computing the propagation of
initial delays using a delay propagation model such as Fig. 7.1. Initial delays of starting trains
or trains entering the system boundaries, plus primary delays of processes that require more
time than scheduled, can be (partially) recovered by time supplements in running and dwell
times. The remaining delays at the next stations are called follow-up delays. On the other
hand, if delays exceed the buffer time between train paths, then they generate new delays to
successor trains, called secondary delays. For closed local areas, initial delays should settle
within some finite settling time. For example, German rules recommend a stability norm that
10 min initial delay should settle within two periods, see Pachl (2002). Sensitivity to delays can
also be analysed using recovery time analysis, see Section 7.6.

135
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
7 Timetable Stability Analysis

The local stability tests are complementary to the global stability test. If a timetable is formally
unstable on a global level, it may still be locally stable as long as initial delays do not reach
the unstable area. Conversely, a globally stable timetable may be locally unstable when in
some areas the train paths have been scheduled very tightly. In this area delays then settle
only slowly, although on a network scale the delays are absorbed somewhere downstream
of the locally unstable area. Stability is also intimately related to the feasibility of the timetable,
see also Goverde (2010). A realisable timetable is either critical or stable, but never (globally)
unstable. On the other hand, an unrealisable timetable may be globally stable, although it is
very likely locally unstable in the area of conflicting train paths (depending on the boundaries of
the local area).
Timetable stability can be analysed by studying the properties of the timetable interconnection
structure. In particular, the natural language for studying timetables is the max-plus algebra in
which the scheduled railway operations can be modelled and analysed as a linear system. The
next section briefly introduces max-plus algebra before explaining the max-plus modelling and
analysis of periodic railway timetables.

7.3 Max-Plus Algebra

Max-plus algebra is the algebraic structure consisting of the real numbers extended with – ∞
equipped with an addition and multiplication operation, denoted by the symbols  and ,
which are defined by
a ⊕ b = max (a, b)
a ⊗b = a +b

This algebraic structure is an idempotent semiring, implying that this “addition” is commutative,
associative, idempotent (a  a = a) and has a zero element ( = – ∞), “multiplication” is
associative and has a unit element (0), and “multiplication” is distributive over “addition”.
Moreover, the zero element is absorbing (a  =  a = ). Like in conventional algebra,
multiplication is also simply written as ab instead of a  b if it is clear that multiplication is in
max-plus algebra. Since the zero element is = – ∞, a non-zero element will also be referred
to as a finite element.
The scalar max-plus operations are extended to matrices in a standard way similar to
conventional linear algebra but now in max-plus algebra. Let A = (aij ) and B = (bij ) be two
square n×n matrices with entries over the real numbers and = – ∞, then
(A ⊕B)ij = a ij ⊕ bij = max(a ij,bij )
n
(A ⊗B)ij = ⊕(a ik ⊗ bkj ) = max (a ik + bkj ).
k=1 k=1,…,n

The set of square matrices with the above addition and multiplication is again a dioid. The zero
matrix is the matrix with all entries equal to , which is also simply denoted by . The identity
matrix E = (eij ) is defined as the diagonal matrix with all diagonal entries equal to 0, i. e. eii = 0
and eij = for j ≠ i. Scalar multiplication of a max-plus matrix A is defined by component as
( c ⊗ A )ij = c ⊗ aij = c + aij for all real numbers c and .
The interest in max-plus algebra originates from the connection between powers of matrices in
max-plus algebra and critical paths in digraphs. For square matrices A in max-plus algebra, a
matrix power Al is defined recursively by A0 = E and Al = Al-1  A for all positive integers l.
The symbol  is added in the power notation to distinguish a power in max-plus algebra from
an ordinary power. Each square n×n matrix A corresponds to a directed graph G(A) with n

136
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 7.3 Max-Plus Algebra

nodes numbered 1,…, n and an arc (j,i) with weight aij for each aij ≠ , called the precedence
graph. Hence, A is the matrix whose finite entries correspond to arcs in the digraph G(A). A
path is a sequence of arcs in a digraph and the associated path weight is the sum of the arc
weights. By definition, A2 is defined as
n
(A ⊗2 )ij = ⊕(a ik ⊗ a kj ) = max (a ik + a kj )
k=1 k=1,…,n

which is just the maximum path weight over all paths with exactly two arcs from node j via any
node k to i. In general, Al is the matrix of the maximum weights of paths with exactly l arcs.
The critical path matrix is the matrix of critical path weights, i. e., the maximum weights over
all paths of any length, which can be defined as
∞
A + = ⊕ A ⊗l = A ⊕ A ⊗2 ⊕
l=1

If the precedence graph G(A) has no positive-weight circuits, then any critical path in a graph
of n nodes has at most n arcs and so the critical path matrix is defined by a finite sum of the
n
first n matrix powers, A+ = ⊕l=1 A⊗l . In an acyclic graph of n nodes critical paths have length
n−1
at most n-1, and therefore in this case A+ = ⊕l=1 A⊗l .

Fig. 7.2: The network for the critical paths example

Example. To illustrate the connection between matrix powers and critical paths in graphs,
consider the acyclic digraph of Fig. 7.2, corresponding to matrix
⎡     ⎤
⎢ ⎥
⎢ 4  2  ⎥
A=⎢ ⎥
⎢ 3    ⎥
⎢ ⎥
⎢  5 4  ⎥
⎣ ⎦

The matrix powers can be computed as

⎡     ⎤ ⎡     ⎤ ⎡     ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⊗2
⎢ 5    ⎥ ⊗3
⎢     ⎥ ⊗l
⎢     ⎥
A =⎢ ⎥, A = ⎢ ⎥, A = ⎢ ⎥ for all l ≥ 4
⎢     ⎥ ⎢     ⎥ ⎢     ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ 9  7  ⎥ ⎢ 10    ⎥ ⎢     ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦

For example, note that there are two paths of two arcs from node 1 to 4 with weight 7 (via
node 3) and 9 (via node 2) respectively, and so the maximum weight is 9 corresponding to
⊗2
A41 = 9. There is only one path with three arcs over the nodes 1-3-2-4 with weight 10, and
there are no paths with more than 3 arcs. The critical path matrix is now computed as
⎡     ⎤
⎢ ⎥
3 ⎢ 5  2  ⎥
A =⊕A
+ ⊗l
=⎢ ⎥
⎢ 3    ⎥
l=1 ⎢ ⎥
⎢ 10 5 7  ⎥
⎣ ⎦
137
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
7 Timetable Stability Analysis

The critical path matrix is an essential element in the theory of max-plus linear systems and is
in practice computed by efficient graph algorithms instead of computing matrix powers. The
critical path weights in an a cyclic digraph G(A) can be computed in linear time O(m) using a
topological ordering algorithm, see Ahuja et al. (1993), where m is the number of arcs in G(A) .
In general digraphs, all-pair shortest path algorithms are used. The Floyd-Warshall algorithm
computes A+ in O(n 3) time and for sparse digraphs Johnson’s repeated single-origin shortest
path algorithm, see computes A+ in O(nm + n 2 log n) time, see Ahuja et al. (1993). Note that
critical paths can be computed by shortest path algorithms by first redefining the arc weights
to –aij , solving the problem by shortest path algorithms and then again taking the negative
shortest path weights to find the critical path weights. The shortest path algorithms also can
be adapted easily to compute the critical paths directly, see Goverde (2005).

7.4 Max-Plus Modelling

The max-plus timetable stability analysis is based on a macroscopic deterministic model

of the scheduled railway traffic system. The variables of interest are the train event times at
stations, which are connected by precedence constraints. In general, these events include
arrivals, departures and passages of through trains at all stations, but since the running times
are deterministic through times and arrival times can be expressed as an extension of the
preceding departure times, so that we may consider departures only. A periodic timetable is
assumed, where each departure repeats at a regular interval called the cycle time T, which is
usually 60 minutes. Train lines with a frequency of more than once per T minutes are modelled
using separate variables. For instance, a train line with a frequency of four trains per hour is
modelled by four different departures at each station served.

7.4.1 The Higher-Order Max-Plus Model

Denote by xi (k) the k-th departure time of a train from train line Li at station Si . A train is not
allowed to depart before its scheduled departure time, which gives the timetable constraint
xi (k) ≥ di (k) (7.1)

where di (k) is the scheduled departure time of the k-th departure from train line Li at station
Si . In a periodic timetable with cycle time T, the scheduled departure times are given by
d i ( k ) = d i ( 0 ) + k ⋅ T , where di ( 0) is an initial scheduled departure time of event i .
Before a train is ready-to-depart it must satisfy some precedence constraints. First of all, the
train must have arrived from the preceding station and the passengers must have had time to
alight and board. This gives the constraint
xi (k) ≥ a ij + x j (k −ij )
(7.2)
where aij is the sum of the running time from the preceding station and the minimum dwell time
at station Si , and x j ( k −  ij ) is the departure time from the preceding station, which may have
occurred in the same period or one or more periods ago denoted by the period delay  ij . If
the preceding event j is scheduled in the same period k as event i then  ij = 0, if j is scheduled
in the previous period then  ij = 1, etc. If a train has to wait for a feeder train to secure
passenger transfers, then a second precedence constraint of the form (7.2) is defined, where
now j is the preceding departure of the feeder train and aij is the sum of the running time of the
feeder train and the minimum transfer time at station Si . Likewise, multiple connections are

138
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 7.4 Max-Plus Modelling

defined which may also include the transfer of rolling stock between trains, a direction reversal,
or crew transfers. Finally, conflicting train routes may also prevent a train from departing. This
is again a precedence constraint (7.2), where now j is the departure of a preceding train on a
conflicting train route and aij is the minimum headway after departure j before the conflicting
route is released for departure i. Note that such an infrastructure constraint also defines a train
order on a conflicting route. In general, infrastructure constraints may be defined between
departing, arriving and through trains, and therefore it may be helpful to model arrivals and
through events explicitly, see Goverde (2010).
Assuming that a train departs right time after all conditions are satisfied, we obtain
xi (k) = max(max(a ij + x j (k −ij )),di (k)) (7.3)
j

where j ranges over the predecessors of i . This recursive equation can be written in max-plus
algebra as n
xi (k) = ⊕(a ij ⊗ x j (k −ij )) ⊕ di (k) (7.4)
j=1

where n is the total number of events and aij = for any unconstrained event pair ( j,i) . To write
the equations in matrix notation, collect all the activity times aij in (7.4) associated with events
with period delay l in a matrix Al , with (Ai )ij = if no such term exists, and define the vectors
x(k) = (x1(k),…, xn(k))’ and d(k) = (d1(k),…, dn (k))’. Then (7.4) can be written in matrix notation as
p
x(k) = ⊕ A l ⊗ x(k −l) ⊕ d(k) (7.5)
l=0

where the maximal period delay p is called the order of the system. The timetable vector d(k)
can be written in max-plus algebra as
d (k) = d (0) ⊗ T ⊗k (7.6)

where scalar multiplication of a vector is defined per entry and the k-th power of a scalar
denotes k times repeated “multiplication”, i. e., T k = k · T.

Fig. 7.3: Time-distance diagram of the example railway network

Example. Consider a simple single-track line from station A to station C with intermediate
station B where trains from opposite directions meet and pass, see Fig. 7.3. Each hour
trains departs from both stations A and C, which meet halfway in station B. The scheduled
time-distance diagram is also shown in Fig. 7.3. The running time between A and B is 26
minutes and between B and C this is 25 minutes (in both directions). The minimum dwell times
in station B are 1 minute and the minimum layover times at the terminal stations A and C

139
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
7 Timetable Stability Analysis

are 3 minutes. After this the trains proceed in the opposite direction. The max-plus model is
derived with respect to the four departures: departure x1 from A to B, departure x2 from C to
B, departure x3 from B to A, and departure x4 from B to C. Departure x1 (k) must wait on the
preceding departure x3 (k-1) from station B in the previous period plus a running time of 26
minutes from B to A and 3 minutes layover time at station A, i. e., x1( k ) ≥ 3 ⊗ 26 ⊗ x 3 ( k −1),
and also on its scheduled departure time d1(k) = 0 + k · 60, which gives the max-plus equation
x1(k) = 29 ⊗ x 3 (k −1) ⊕ d1(k) (7.7)

Likewise, departure x2 (k) from station C waits on the preceding departure x4 (k-1) from B plus
25 minutes running time and 3 minutes layover time, and on the scheduled departure time
d2 (k) = 0 + k · 60, giving
x 2 (k) = 28 ⊗ x 4 (k −1) ⊕ d2 (k) (7.8)

The departure x3 (k) from B has three constraints. First, it has to wait on the preceding
departure x2 (k) from C which occurs at the beginning of the same period, plus 25 minutes
running time and 1 minute dwell time in B, resulting in x 3 ( k ) ≥1⊗ 25 ⊗ x 2 ( k ); second, the train
can only depart from B if the train from A in the opposite direction has arrived in B, giving
x 3 ( k ) ≥ 26 ⊗ x1( k ), since the running time from A to B is 26 minutes (it is assumed here that
a train can depart immediately after the arrival of the train in the opposite direction, so the
minimum arrival-departure headway is 0 minutes); third, the train must wait on its scheduled
departure time d3 (k) = 27 + k · 60. These three constraints give the max-plus equation
x 3 (k) = 26 ⊗ x1(k) ⊕ 26 ⊗ x 2 (k) ⊕ d3 (k) (7.9)

Likewise, departure x4 (k) from B towards C has three constraints. First, it has to wait on
the preceding departure from A plus 26 minutes running time and 1 minute dwell time in B,
resulting in x 4 ( k ) ≥ 27 ⊗ x1( k ); second, the train can only depart from B if the train from C
in the opposite direction has arrived in B, giving x 4 ( k ) ≥ 25 ⊗ x 2 ( k ); and third, the train must
wait on its scheduled departure time d4 (k) = 27 + k · 60. Together these three constraints give
the max-plus equation
x 4 (k) = 27 ⊗ x1(k) ⊕ 25 ⊗ x 2 (k) ⊕ d4 (k) (7.10)

The four equations (7.7) – (7.10) can be written in matrix notation as

⎡
⎢     ⎤
⎥
⎡
⎢   29  ⎥⎤
⎢     ⎥ ⎢
⎥ ⊗ x(k) ⊕ ⎢    28 ⎥⎥
x(k) = ⎢ ⊗ x(k −1) ⊕ d(k) (7.11)
⎢
⎢ 26 26   ⎥
⎥
⎢
⎢     ⎥⎥
⎢
⎣ 27 25   ⎥
⎦
⎢
⎣     ⎥⎦

with d( k ) = ( 0,0,27 ,27 )'⊗ 60⊗k . This is of the form (7.5) with system order p 1. This
example will be used throughout this Chapter to illustrate the theory.

7.4.2 The First-Order Max-Plus Model

Any higher-order max-plus linear system (7.5) can be transformed into a pure first-order linear
system of the form
x(k) = A ⊗ x(k −1) ⊕ d(k) (7.12)

for some matrix A and possibly expanded state vector x(k). This transformation is done in two
steps. First, all higher-order terms of the form (7.2) with  ij > 1 can be transformed to first-

140
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 7.4 Max-Plus Modelling

order constraints by expanding the state vector with  ij – 1 dummy variables and dividing
the activity over a sequence of these new variables. For example, consider a constraint
x 2 ( k ) ≥ a21 ⊗ x1( k − 2 ). Then define a new event 3 and write the precedence constraint by
two constraints x 3 ( k ) ≥ a31 ⊗ x1( k −1) and x 2 ( k ) ≥ a23 ⊗ x 3 ( k −1) with a23 ⊗ a31 = a21, e. g.
a31 = T and a23 = a21 −T . The first inequality will act as an equality in the system equations
by choosing the initial scheduled event time of the new event 3 tightly as d 3 ( 0 ) = a31 ⊗ d1( 0 ).
Applying this state augmentation technique to all precedence constraints with a period delay
higher than one leads to the first-order system (with still zero-order terms):
x(k) = A 0 ⊗ x(k) ⊕ A1 ⊗ x(k −1) ⊕ d(k) (7.13)

The second step is to remove the implicit zero-order term in (7.13). This can be done by
substitution as illustrated first for the example system. Note that the example system (7.11) is
already of the order p = 1 and so of the form (7.13).
Example. The equations (7.7) and (7.8) are already in a first-order representation, but (7.9) contains
zero-order terms. To remove the zero-order terms substitute (7.7) and (7.8) into (7.9), giving
x 3 (k) = 26 ⊗ x1(k) ⊕ 26 ⊗ x 2 (k) ⊕ d3 (k)
= 26 ⊗ 29 ⊗ x 3 (k −1) ⊕ 26 ⊗ d1(k) ⊕ 26 ⊗ 28 ⊗ x 4 (k −1) ⊕ 26 ⊗ d2 (k) ⊕ d3 (k) (7.14)
= 55 ⊗ x 3 (k −1) ⊕ 54 ⊗ x 4 (k −1) ⊕ d3 (k)

Note that the timetable terms are dominated by the scheduled departure time d3 (k), that is,
26 ⊗ d1( k ) ⊕ 26 ⊗ d 2 ( k ) = 26 ⊗ 60⊗k ≤ 27 ⊗ 60⊗k = d 3 ( k ) and hence can be removed from
the final equation. Timetables that satisfy this property are called realisable. Likewise, the
zero-order terms in (7.10) can be removed by substitution of (7.7) and (7.8), yielding
x 4 (k) = 27 ⊗ x1(k) ⊕ 25 ⊗ x 2 (k) ⊕ d4 (k)
= 27 ⊗ 29 ⊗ x 3 (k −1) ⊕ 27 ⊗ d1(k) ⊕ 25 ⊗ 28 ⊗ x 4 (k −1) ⊕ 25 ⊗ d2 (k) ⊕ d4 (k) (7.15)
= 56 ⊗ x 3 (k −1) ⊕ 53 ⊗ x 4 (k −1) ⊕ d4 (k)
In matrix notation we now have the pure first-order system of the form (7.12) given by
⎡
⎢   29  ⎤⎥
⎢    28 ⎥⎥
x(k) = ⎢ ⊗ x(k −1) ⊕ d(k) (7.16)
⎢
⎢   55 54 ⎥⎥
⎢   56 53 ⎦⎥
⎣
The above example showed that zero-order terms can be removed by substituting first-order
terms. This process of algebraic manipulation can be generalised in matrix notation: It can be
proven that (7.13) is equivalent to
x(k) = A 0* A1 ⊗ x(k −1) ⊕ d(k) (7.17)

where A0* = E ⊕ A0+ (Bacelli et al. (1992), Goverde (2005)). This is of the form (7.12) with
⎛n−1 ⎞⎟
A = A 0* A1 = (E ⊕ A 0 ⊕ A ⊗2 ⊗n−1
0 ⊕…⊕ A 0 ) ⊗ A1 = ⎜⎜⎜⊕A ⊗l ⎟⎟ ⊗ A1 (7.18)
⎜⎝l=0 0 ⎟⎟⎠
where it is assumed that A0 is acyclic, i. e. there is no sequence of activities from any event i
back to i with only zero period delays. It can be proven that the property of an acyclic A0 must
be satisfied for our timetable models; if this property is not satisfied then the system contains
a deadlock situation, see Goverde (2005). The matrix A0+ contains the critical path weights
over all paths in the acyclic digraph G(A0 ) defined by nodes 1,…, n and an arc ( j,i ) with weight
aij for all pairs ( j,i ) with  ij = 0, corresponding to the constraints with zero period delay. Recall
from Section 7.2 that the critical path matrix A0+ associated to an acyclic digraph G(A0 ) can be
computed in linear time. Hence, (7.17) can be computed from (7.13) efficiently using standard
algorithms.

141
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
7 Timetable Stability Analysis

Fig. 7.4: The precedence graph G(A0) associated to the activities with zero period delay of the
example system

Example. In our example system (7.11), G(A0 ) only contains paths of one arc, see Fig. 7.4,
and so A0* is computed as
⎡ 0    ⎤ ⎡     ⎤ ⎡ 0    ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢  0   ⎥ ⎢     ⎥ ⎢  0   ⎥
A 0* = E ⊕ A 0 = ⎢ ⎥⊕⎢ ⎥=⎢
⎥ ⎢
⎥
⎥
(7.19)
⎢   0  ⎥ ⎢ 26 26   26 26 0 
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢    0 ⎥ ⎢ 27 25   ⎥ ⎢ 27 25  0 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
Hence, (7.18) gives
⎡
⎢ 0    ⎤ ⎡
⎥ ⎢   29  ⎥⎤ ⎡⎢   29  ⎤⎥
⎢  0   ⎥ ⎢    28 ⎥⎥ ⎢⎢    28 ⎥⎥
A= A 0* A1 = ⎢ ⎥⊗⎢ = (7.20)
⎢
⎢ 26 26 0  ⎥ ⎢
⎥ ⎢     ⎥⎥ ⎢⎢   55 54 ⎥⎥
⎢
⎣ 27 25  0 ⎥ ⎢
⎦ ⎣     ⎥⎦ ⎢⎣   56 53 ⎥⎦

and by (7.17) we obtain again the pure first-order system (7.16).

7.5 Critical Circuit Analysis

The first focus in this section ison the behaviour of the max-plus linear system when all events
occur as early as possible respecting all precedence constraints, but discarding the timetable
vector. Hence, we consider the system equation
x(k) = A ⊗ x(k −1), x(0) = x 0 (7.21)

with given initial event time vector x0 . Since, (7.12) is a simple recursive equation, we can
compute the solution x(k) for k 1 as a function of the initial condition as follows. For k = 1 we
obtain x(1) = Ax( 0 ) = Ax 0 , for k = 2 we get x( 2 ) = Ax(1) = AAx( 0 ) = A⊗2 x 0 , and in general
we have x( k ) = A⊗k x 0 . It turns out that independent of the initial condition x0 the event times
x(k) approach a periodic regime, i. e., x i ( k +1)− x i ( k ) = i for all i and k K for some period
K, where i is the event cycle time satisfying
xi (k)
lim = i (7.22)
k →∞ k
If the precedence graph is strongly connected, i. e. if there is a path from any event i to any
other event j, then all events have the same unique cycle time . If the network contains
several strongly-connected components, then it can be partitioned into several subnetworks
with each having a unique cycle time. For stability analysis the subnetwork(s) with the
maximum cycle time are then crucial. In the sequel, it is assumed that the network is strongly
connected. For the general case, see Goverde (2005, 2007).
The main questions for stability analysis are:
a. What is the value of the cycle time  and how can it be determined?
b. What are the critical activities determining ?
c. How far is this “minimal achievable cycle time” from the timetable cycle time T ?

142
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 7.5 Critical Circuit Analysis

d. How many transient periods does it take before the system reaches the periodic regime?
The answers to these questions are related to a max-plus eigenvalue problem; the cycle time
 is the unique max-plus eigenvalue of the system matrix A. This can be explained as follows.
Let  be the minimal cycle time in which all the train trips can be realised while respecting
all the precedence constraints and let v be the associated vector of earliest possible event
times, where thus 0 ≤ v i <  for all events i. Hence, all events repeat at a minimal cycle time
 and therefore if x( 0 ) = v then x( k ) = ⊗k ⊗ v for any k, where the multiplication is by entry,
i. e. x i ( k ) = v i + k ·  for all i. Moreover, starting with initial vector v, the earliest possible event
times in the next period are given by A  v and if they are periodic this is equivalent to v.
Hence, the problem of determining  and an associated vector v can be written as the max-
plus eigenvalue problem
A ⊗v = ⊗v (7.23)

In conventional scalar notation this equation becomes

max(a ij + v j ) =  + vi for all 1≤ i ≤ n,
j

where the maximum is taken over all predecessors j of i. A solution  is called an eigenvalue
and v is an eigenvector associated to .
A fundamental theorem in max-plus algebra states that the eigenvalue  is equal to the
maximum cycle mean
w(C)
 = max (7.24)
C l (C)

where w( C ) = ⊗( j ,i )∈Caij is the circuit weight, l(C) is the circuit length (number of arcs in
circuit C ), and the maximum is taken over all elementary circuits C in the (strongly connected)
precedence graph G(A). Note that in a pure first-order system (7.21), l(C) is the number of
periods covered by circuit C. A circuit C on which the maximum (7.24) is obtained is called a
critical circuit. Hence, a critical circuit has the largest sum of activity times relative to the number
of periods required by these activities compared with any other circuit. This gives a direct
interpretation of a stability criterion, since  gives a lower bound on the cycle time of all events.
Theorem 7.1. A scheduled max-plus linear system (7.12) is structurally stable if  < T . It is
called critical if  < T , and unstable if  < T .
This theorem states that the intended timetable cycle time T should exceed the minimal
possible one, which justifies the interpretation of the eigenvalue as the minimal cycle time.
Hence, the eigenvalue enables a stability test. If  exceeds the timetable cycle time T then
the system is unstable. Instability here means that there is no slack on the critical circuit, by
which delays that reach the critical circuit can never settle, i. e. all trains on the critical circuit
will get delayed and possibly propagate this delay to other trains in the network. If  is close
to the cycle time T > , then delays on the critical circuit will settle only slowly and the system
will become unstable in case of temporary process time prolongations on the critical circuit.
In practice, a timetable will therefore only be considered stable if there is also “enough” slack
on the critical circuits. This practical stability can be defined by some norm such as  < T – ,
or equivalently, by  =  / T < crit = 1 –  / T. However, such stability norms are not yet defined
and may also differ between various countries or networks depending on the achieved
variance of train running times.
An eigenvector can be obtained by selecting any node on a critical circuit and computing
the critical path weights from this node to all other nodes modulo . An eigenvector is not
unique. For example, if v is an eigenvector, then so is c  v for any constant c. Note that this
multiplication corresponds to shifting the entire timetable by a constant c.

143
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
7 Timetable Stability Analysis

Fig. 7.5: The precedence graph of the example max-plus system

Example. Consider the example system (7.16). The precedence graph G(A) is shown in
Fig. 7.5. There are three elementary circuits and by (7.24) the eigenvalue is
⎛ 55 53 54 ⊗ 56 ⎞⎟
 = max ⎜⎜⎜ , , ⎟ = max(55,53,55) = 55
⎝ 1 1 2 ⎟⎠
There are two critical circuits: 3-4-3 and 3-3. The first corresponds to a complete train
circulation starting from station B via A back to B and via C back to B again. The second
critical circuit corresponds to the single track between station A and B, and includes a train
run from B to A, the layover time in A, the train run back to B, and the minimum headway with
the opposite train in B. An eigenvalue can be computed by the critical path weights from the
critical node 4 modulo 55 giving (28,28,54,0)’. To compare the eigenvector with the actual
timetable we can multiply it by a constant modulo , so that the first event has scheduled
departure time 0, that is, 27 ⊗ ( 28,28,54,0 )'mod 55 = ( 0,0,26,27 )'.
The eigenvalue is a unique characteristic of the state matrix A of a max-plus linear system
and thus depends on all (or the critical) precedence constraints that define the state matrix
A. Hence, the eigenvalue is a performance indicator of the traffic scenario characterised by
the network timetable, rolling stock, infrastructure, and signalling system. The most effective
algorithm to solve the max-plus eigenvalue problem is the policy iteration algorithm developed
by Cochet-Terrasson et al. (1998). This algorithm also finds the various strongly-connected
components – if the network is not strongly connected – and the eigenvalue for each of these
components, an eigenvector, and the critical activities that make up the critical circuit(s). For an
overview of algorithms, see Goverde (2005) or Heidergott et al. (2006).
The first three questions posed at the beginning of this section have now been answered. The
remaining question concerns the number of required periods before reaching the periodic regime.
Clearly, if we start with an eigenvector then we are immediately in the periodic regime. Otherwise,
this transient time can be arbitrarily large depending on the initial condition, see Soto y Koelemeijer
(2003) or Heidergott et al. (2006). However, in structurally stable scheduled max-plus linear
systems, the greater timetable cycle time T >  gives an additional mean slack time ∆ = T – 
on the critical circuits. If this stability margin is evenly distributed over the critical activities, then
the timetable can reduce any initial delay by this amount over each period. In general, however,
the settling time of any initial delay depends on the distribution of running time supplements and
buffer times over the timetable, which will be considered in Sections 7.6 and 7.7.

7.6 Recovery Time Analysis

A timetable generally contains slack time to prevent delays or to recover from delays. Slack
time can be incorporated within scheduled process times (running time supplements) or
between train movements (buffer times). For any activity (j,i) the slack time is the difference

144
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 7.6 Recovery Time Analysis

of the end of the activity dj (k – ij) + aij and the start of the new activity dj (k). In a periodic
timetable, this slack time is given by dj (0) – dj (0) aij + ij · T. In a realisable timetable, all
activities have non-negative slack time. However, the critical slack time from event j to i may
be less than the slack time of activity (j,i) if there is some other path from j to i with less slack
time. The critical slack time between any two events j and i will be called the recovery time
between the two events. In general, it is interesting to know how much recovery time is
available between any pair of events in the timetable, which can be represented in a recovery
time matrix associated with a given timetable. The recovery matrix can be computed by a
critical path algorithm over a digraph where each arc corresponds to an activity (j,i) and the arc
weights represent how much delay can be absorbed by (of after) this activity. Thus, this graph
is the precedence graph G(S) of the matrix S = (sij ) defined as
sij = d j (0) − di (0) + a ij − ij ⋅ T
which is the negative slack time of activity (j,i). The recovery matrix R is then defined as
R = −S+ (7.25)

Note that this recovery time matrix contains the minimum cumulative slack over any path
between two events. If there is no path from an event j to i then sij = and rij = ∞, where we
use the convention that – = ∞.
Various elements of the recovery time matrix have different interpretations. The i th column of
R gives the recovery times from event i to all later events. Hence, the columns of R can be
interpreted as delay impact vectors: if event i has some delay then all entries in column i
smaller than the delay correspond to events j that will be delayed. On the other hand, the i th
row of R gives the recovery times from all earlier events to event i. The rows can therefore
be interpreted as delay sensitivity vectors. If any event j in the network has a delay
exceeding the associated entry in row i, then event i will be delayed. Finally, the diagonal of R
corresponds to the feedback recovery times to a future occurrence of the same events. Any
delay larger than the associated diagonal entry in R will backfire (in reduced form) over some
cycle in the timetable. Note that the columns (rows, respectively) of the recovery matrix can
also be obtained by a single-origin (single-destination) critical path algorithm.
Example. For the example system we obtain
⎡
⎢   −4  ⎥⎤ ⎡
⎢ 5 5 4 10 ⎤⎥
⎢    −5 ⎥⎥ +
⎢ 5 7 9 5 ⎥⎥
S= ⎢ , R = −S = ⎢
⎢
⎢ −1 −1   ⎥⎥ ⎢
⎢ 1 1 5 6 ⎥⎥
⎢
⎣ 0 −2   ⎥⎦ ⎢
⎣ 0 2 4 7 ⎥⎦

From this recovery time matrix, the following may be observed. The zero recovery time
corresponds to the zero dwell buffer time between the departure from station A and the
departure in B towards C. The 1st column of R shows the delay impact vector of the departure
from station A: any delay will propagate to event 4 and delays larger than 1 minute also to
event 3. On the other hand, this departure is not that sensitive to delays of other events
as shown in the 1st row; any delay up to 4 minutes settles before reaching this event. The
diagonal entries reveal that trains between station A and B (events 1 and 3) have less recovery
time than the ones between station B and C (events 2 and 4), which agrees with the critical
circuit over the single-track between stations A and B.

145
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
7 Timetable Stability Analysis

7.7 Delay Propagation

The recovery matrix R = (rij ) gives the maximum delays that can be recovered between any
two pairs of events in the network. So if an initial delay of a train j exceeds the recovery time
rij then the delay will propagate to train i. Nevertheless, from the recovery matrix we still do
not know when this train will be delayed and whether the train is delayed more than once by
cyclic delay propagation or by initial delays of different trains. The max-plus dynamic system
equations (7.13) can however be used to compute the detailed propagation of some initial
delay scenario over time and space directly.
Let z0 be a vector of initial delays which we can assume without loss of generality to be in
period k = 0. Then we can compute the propagation of these initial delays in two steps: first
compute the propagation over the remaining events in period 0 which gives the initial state
vector x(0), and then compute the propagation over the subsequent periods k  1 using (7.12)
for the given initial condition x(0). The delays over the subsequent periods can be computed
from the max-plus model as z(k) =x(k) – d(k) for k 0, and the settling period is the earliest
period ks satisfying z(ks ) = 0, or equivalently, x(ks ) = d(ks ).
The initial condition can be computed as follows. Define the initial event time vector by
x 0 = z0 + d( 0 ) (using conventional vector addition). Then x( 0 ) = A0* x 0 = x 0 ⊕ A0+ x 0 . The first
term is the vector with initially known delays and the second term computes the possible
delayed event times within the remaining initial period. Note that ( A0+ )ij is the critical sum of
activity times over all paths covering only activities with zero period delays from j to i. Hence,
x i ( 0 ) ≥ ( A0+ )ij ⊗ x j ( 0 ) for all j.
The delay propagation algorithm using (7.12) with A = A0* ⊗ A1 can thus be summarized as
follows. First compute x 0 = z0 + d( 0 ) and x( 0 ) = A0* ⊗ x 0 . Then iterate for k 1
⎧⎪ d(k) = T ⊗ d(k −1)
⎪⎪
⎪
⎨ x(k) = A ⊗ x(k −1) ⊕ d(k)
⎪⎪
⎪⎪ z(k) = x(k)− d(k)
⎪⎩
until z(k) = 0, in which case set ks = k. An efficient graph implementation of these equations to
handle large-scale matrices is described in Goverde (2005, 2010).
Formally, a scheduled max-plus linear system (7.12) is (globally) stable if for any x(0) > d(0)
there exists a finite K > 0 such that x(k) = d(k) for all k ≥ K. In particular this implies that for any
event i a delay z i ( k 0 ) = x i ( k 0 ) − d i ( k 0 ) in some period k0 must have been (partially) absorbed
before it recurs to a next occurrence of this event xi (k) with k > k0. Thus, a railway timetable is
stable if each circuit in the associated precedence graph contains some slack time for delay
recovery or reducing delay propagation, respectively. This condition is identical to the test
 < T from Theorem 7.1 in Section 7.5. Hence, if a system is formally stable, i. e.,  < T, then
any initial delay scenario settles after some finite settling period ks and the delay propagation
algorithm terminates.
The delay propagation algorithm described above essentially computes the critical paths of
slack time over some successive periods as required to settle an initial delay vector. Recall
that for timetable stability analysis it is of prime interest to compute the delay propagation
while maintaining the timetable structure to evaluate how well the timetable is able to settle
delays. For large delays dispatchers will in practice change train orders or cancel certain
passenger connections to settle the delays quicker. The max-plus model can be adjusted
so that dispatching strategies are taken into account and it can also be embedded into an
optimization algorithm to compute optimal dispatching strategies, see Heidergott & De Vries
(2001) and Van den Boom & De Schutter (2006).

146
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 7.7 Delay Propagation

Example. Consider the example system (7.11) and assume that the initial departures
from stations A and C have 10 minutes initial delay. So x0 = (10,10,0,0)’ + (0,0,27,27)’ =
(10,10,27,27)’. Then
⎡ 0    ⎤ ⎡ 10 ⎤ ⎡ 10 ⎤ ⎡ 10 ⎤ ⎡ 0 ⎤ ⎡ 10 ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢  0   ⎥ ⎢ 10 ⎥ ⎢ 10 ⎥ ⎢ 10 ⎥ ⎢ 0 ⎥ ⎢ 10 ⎥
x(0) = A 0* x 0 = ⎢ ⎥⊗⎢ ⎥=⎢ ⎥ , z(0) = ⎢
⎥ ⎢
⎥−⎢
⎥ ⎢
⎥=⎢
⎥ ⎢
⎥
⎥
⎢ 26 26 0  ⎥ ⎢ 27 ⎥ ⎢ 36 36 27 9
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ 27 25  0 ⎥ ⎢ 27 ⎥ ⎢ 37 ⎥ ⎢ 37 ⎥ ⎢ 27 ⎥ ⎢ 10 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦

Note that in the initial period also events 3 and 4 get delayed. Now, we can compute the
delays using (7.16) until the delays settle:

Fig. 7.6: The delay propagation of the example network over the first two periods until the settling
time

⎡   29  ⎤⎥ ⎡⎢ 10 ⎤ ⎡ 60 ⎤ ⎡ 65 ⎤ ⎡ 65 ⎤ ⎡ 60 ⎤ ⎡ 5 ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢    28 ⎥⎥ ⎢⎢ 10 ⎥ ⎢ 60 ⎥ ⎢ 65 ⎥ ⎢ 65 ⎥ ⎢ 60 ⎥ ⎢ 5 ⎥
x(1) = ⎢ ⊗ ⎥⊕⎢ ⎥=⎢ ⎥ , z(1) = ⎢ ⎥−⎢ ⎥=⎢ ⎥
⎢   55 54 ⎥⎥ ⎢⎢ 36 ⎥ ⎢ 87 ⎥ ⎢ 91 ⎥ ⎢ 91 ⎥ ⎢ 87 ⎥ ⎢ 4 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢   56 53 ⎥⎦ ⎢⎣ 37 ⎥ ⎢ 87 ⎥ ⎢ 92 ⎥ ⎢ 92 ⎥ ⎢ 87 ⎥ ⎢ 5 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦

⎡   29  ⎤⎥ ⎡⎢ 65 ⎤ ⎡ 120 ⎤ ⎡ 120 ⎤ ⎡ 120 ⎤ ⎡ 120 ⎤ ⎡ 0 ⎤

⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢    28 ⎥⎥ ⎢⎢ 65 ⎥ ⎢ 120 ⎥ ⎢ 120 ⎥ ⎢ 120 ⎥ ⎢ 120 ⎥ ⎢ 0 ⎥
x(2) = ⎢ ⊗ ⎥⊕⎢ ⎥=⎢ ⎥ , z(1) = ⎢ ⎥−⎢ ⎥=⎢ ⎥
⎢   55 54 ⎥⎥ ⎢⎢ 91 ⎥ ⎢ 147 ⎥ ⎢ 147 ⎥ ⎢ 147 ⎥ ⎢ 147 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢   56 53 ⎥⎦ ⎢⎣ 92 ⎥ ⎢ 147 ⎥ ⎢ 147 ⎥ ⎢ 147 ⎥ ⎢ 147 ⎥ ⎢ 0 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦

The settling period is thus ks = 2. Fig. 7.6 shows the delay propagation in a time-distance
diagram.
As an example of a real case, consider the basic hour pattern of the Dutch 2007 railway
timetable. Fig. 7.7 shows the delay propagation of 10 minutes initial delay of the intercity
from Delft (Dt) to The Hague HS (Gv) with scheduled departure 58 each hour. This initial delay

147
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
7 Timetable Stability Analysis

causes in total 106 consecutive delayed trains and settles within 120 minutes. The last delay
is the intercity from Tilburg (Tb) to ’s Hertogenbosch (Ht) with scheduled departure 55 two
periods after the initial period, which has a departure delay of 18 seconds that settles shortly
after. Fig. 7.8 visualizes the delay impact (see Section 7.6) of the same train from Delft. The
colours corresponding to recovery time up to 10 minutes correspond to the area of the delay
propagation of the 10 minutes initial delay, see Fig. 7.7.

7.8 Stochastic systems

In practice, process times are not deterministic but have some variance around a mean
process time. The effect of this stochastic behaviour on stability can also be analysed using
max-plus stability analysis. The state matrix is now a stochastic matrix, where the entries are
stochastic variables so that each period k a slightly different realisation A(k) occurs according
to some distribution of all process times. The first-order stochastic linear system then is
x(k) = A(k) ⊗ x(k −1) ⊕ d(k) (7.26)

where { A( k ) : k ≥ 0 }is a sequence of matrices with each finite element aij ( k ) a stochastic
variable with a given (independent or joint) distribution function. Hence, { A( k ) : k ≥ 0 } has
fixed support, i. e., each entry is either a stochastic variable or equal to = –∞.
As in the deterministic case, the event times of the autonomous system
x(k) = A(k) ⊗ x(k −1), x(0) = x 0 (7.27)

approach a periodic regime independent of the initial condition, cf. (7.22). In the stochastic
setting the following result holds (Heidergott et al. (2006)): For any event i and any initial
condition x 0 ≠ , there exists a i such that
xi (k) E[xi (k)]
lim = lim = lim E[xi (k) − xi (k − 1)] = i (7.28)
k →∞ k k →∞ k k →∞

where E [ x i ( k )] denotes the expectation of event time x i ( k ) , and i is the steady-state cycle
time of event i. Again, if the precedence graph is strongly connected then all events have the
same steady-state cycle time , which is known as the Lyapunov exponent of the stochastic
matrix A( k ). The Lyapunov exponent can be computed numerically using the Lyapunov
Exponent Algorithm (LEA) of Goverde et al. (2011).
Example. Consider again the example system, but now let all process times in the state
matrix be stochastic variables with a shifted Gamma distribution. Interpret the given process
times now as the minimum process times (the shift parameters), and let the shape and scale
parameter ( and  ) of the Gamma distribution be determined by the method of matching
moments for given mean  ij and standard deviation ij of the delay in process time aij. Hence,
 ij = (  ij / ij )2 and ij =  ij / ij2 , see Goverde et al. (2009). Let the mean and standard deviation
be defined as a percentage of the minimum process time for each process, see Fig. 7.9. Then
we can analyse the effect of mean delay and standard deviation on the cycle time.

148
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 7.8 Stochastic systems

Fig. 7.7: The delay propagation of 10 minutes initial delay of an intercity from Delft (Dt) to The
Hague (Gv); the colours correspond to the largest delays at the stations and open tracks

Fig. 7.8: The delay impact of the intercity from Delft (Dt) to The Hague HS (Gv); dark colours
correspond to small recovery times, the lighter the colour the more recovery time; white stations
correspond to at least 25 minutes of recovery time

149
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
7 Timetable Stability Analysis

Fig. 7.9: Shifted Gamma density function for minimum process time 26 min and 5 % mean delay,
and standard deviation ranging from 1 % to 4 %

Fig. 7.10 shows the result of the computed steady-state cycle times with on the horizontal
axis the standard deviation ranging from 0 % to 10 % of the minimum process times and the
different curves indicating a mean delay from 0 % (the horizontal axis) to 9 %. A mean and
standard deviation of + 1 % and 2 % implies that all process times have a mean delay of + 1 %
and a standard deviation of 2%. The Figure shows that an increasing standard deviation
leads to a larger cycle time. With an integrated mean delay of 5 % the system stays stable
for standard deviations up to 10 %. If the integrated delay is 8 % then the system becomes
unstable for standard deviations larger than 3 %. Hence, a system that operates close to
its scheduled cycle time is better capable of managing larger delays with low variance than
smaller delays with high variance.
The delay propagation can be computed for a stochastic max-plus linear system (7.26) as
well. The stochastic delay propagation model is
⎪⎧⎪ d(k) = T ⊗ d(k −1)
⎪⎪
⎨ x(k) = A(k) ⊗ x(k −1) ⊕ d(k)
⎪⎪
⎪⎪ z(k) = x(k)− d(k)
⎪⎩
with x( 0 ) = A0* ( 0 ) ⊗ x 0 similar to Section 7.7. However, since the A(k) are now stochastic
matrices, all matrix elements from { A( k ) : k ≥ 0 } must be randomly selected from their
distribution functions resulting in the event times of one sample path. To estimate the expected
event times, or other statistics, the delay propagation model must be applied several times
where in each replication the process times are selected randomly from their distributions,
according to a Monte Carlo simulation set-up. This model is equal to the simulation part of
the robust timetabling problem of Chapter 8. One final remark is that computing the first-order
representation for the stochastic case leads to many stochastic dependent matrix elements.
It is therefore recommended to apply the original higher-order model with zero-order terms A0
directly using the delay propagation algorithm of Goverde (2010).

150
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 7.9 Conclusions

Fig 7.10: Dependence of steady-state cycle time on delay mean and standard deviation

7.9 Conclusions

Periodic railway timetables are effectively modelled as a linear system in max-plus algebra.
This algebraic approach combined with efficient graph algorithms enables a powerful stability
analysis of large-scale networks. Network stability can be analysed by solving an eigenvalue
problem in max-plus algebra, which not only gives an efficient stability test but also identifies
the critical sequences of activities within the timetable. The allocation of time allowances over
the timetable determines local stability, which can be analysed using recovery time analysis or
a delay propagation algorithm. Also the effect of variations in process times can be analysed
using stochastic max-plus linear systems. The max-plus modelling and analysis techniques
have been implemented in the software tool PETER (Performance Evaluation of Timed Events
in Railways), which is compatible to the Dutch timetable design system DONS.
In general, timetable stability can also be analysed using simulation, and in particular in
combination with dispatching rules to analyse the operational flexibility with respect to larger
delays. The max-plus algebra approach can be used to analyse basic properties of the
timetable structure quickly and to identify the critical activities from a network perspective. This
may lead to improvements in the initial timetable design or the selection of the best timetable
from a set of candidate timetables. In a second step, (stochastic) simulation can be applied to
analyse the stability of the timetable under larger delays when dispatching rules are also taken
into account. The identified critical activities or locally unstable sub-areas from the max-plus
analysis may then also guide the simulations to study potential delay sensitive areas in more
detail.

151
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
7 Timetable Stability Analysis

Some future research directions include (a) establishing norms on stability margins for stable
network timetables, (b) analysing the effect of different allocations of running time supplements
and buffer times over the timetable on stability and robustness, and (c) microscopic modelling
of network timetables at signal level to detect conflicting train paths directly instead of using
minimum headway constraints at station level.

References
Ahuja R. K., Magnanti T. L., Orlin J. B. (1993). Network Flows: Theory, Algorithms, and
Applications. Prentice-Hall, New Jersey.
Baccelli F. L., Cohen G., Olsder G. J., Quadrat J-P (1992). Synchronization and Linearity: An
Algebra for Discrete Event Systems. Wiley, Chicester.
Braker J. G. (1993): Algorithms and Applications in Timed Discrete Event Systems. PhD thesis,
Delft University of Technology, Delft.
Büker T., Seybold, B. (2012): Stochastic modelling of delay propagation in large networks.
Journal of Rail Transport Planning & Management, 2(1-2), pp. 34 – 50.
Burkolter D. M. (2005): Capacity of Railways in Station Areas Using Petri Nets. PhD thesis,
Swiss Federal Institute of Technology (ETH), Zurich.
Cochet-Terrasson J., Cohen G., Gaubert S., Mc Gettrick M., Quadrat J-P. (1998): Numerical
computation of spectral elements in max-plus algebra. In: Lafay, J.-F. (Ed.), Proceedings of the
IFAC Conference on System Structure and Control 1998, Nantes, July 8–10, pp. 699 – 706.
Cohen G., Dubois D., Quadrat J.-P., Viot M. (1985): A linear-system-theoretic view of discrete-
event processes and its use for performance evaluation in manufacturing. IEEE Transactions
on Automatic Control, AC-30(3), pp. 210 – 220.
Cuninghame-Green R. A. (1979): Minimax Algebra. Lecture Notes in Economics and
Mathematical Systems, vol. 166, Springer, Berlin.
Ekman J. (2004): Capacity estimation of new infrastructure based on a discrete event model
of train paths. In: Allan J., Brebbia C. A., Hill R. J., Sciutto G., Sone S. (Eds.), Computers in
Railways IX, WIT Press, Southampton, pp. 539 – 548.
Gondran M. & Minoux M. (1984): Graphs and Algorithms. Wiley, New York.
Goverde R. M. P. (2005): Punctuality of Railway Operations and Timetable Stability Analysis.
PhD thesis, Delft University of Technology, TRAIL Thesis Series, no. T2005/10, Delft.
Goverde R. M. P. (2007): Railway timetable stability analysis using max-plus system theory.
Transportation Research Part B, 41(2), pp. 179 – 201.
Goverde R. M. P. (2010): A delay propagation algorithm for large-scale railway traffic networks.
Transportation Research Part C, 18(3), pp. 269 – 287.
Goverde R. M. P., Heidergott B., Merlet G. (2009): Railway timetable stability analysis using
stochastic max-plus linear systems. In: Hansen I. A., Wendler E., Weidmann U., Lüthi M.,
Rodriguez J., Ricci S., Kroon L. (Eds.), Proceedings of the 3rd International Seminar on
Railway Operations Modelling and Analysis (RailZurich 2009), Zurich, February 11–13, 2009.
Goverde R. M. P., Heidergott B., Merlet G. (2011): A coupling approach to estimating the
Lyapunov exponent of stochastic max-plus linear systems. European Journal of Operational
Research, 210(2), pp. 249 – 257.
Goverde R. M. P., Odijk, M. A. (2002): Performance evaluation of network timetables using
PETER. In: Allan J., Hill R. J., Brebbia C. A., Sciutto G., Sone S. (Eds.), Computers in Railways
VIII, WIT Press, Southampton, pp. 731 – 740.

152
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

Heidergott B, De Vries R (2001): Towards a control theory for transportation networks. Discrete
Event Dynamic Systems, 11(4), pp. 371 – 398.
Heidergott B., Olsder G. J., Van der Woude J. (2006): Max Plus at Work Princeton University
Press, Princeton, New Jersey.
Herrmann T. M. (2006): Stability of Timetables and Train Routings through Station Regions.
PhD thesis, Swiss Federal Institute of Technology (ETH), Zurich.
Middelkoop D., Bouwman M. (2002): Testing the stability of the rail network. In Allan J.,
Hill R. J., Brebbia C. A., Sciutto G., Sone S. (Eds.), Computers in Railways VIII, WIT Press,
Southampton, pp. 995 – 1002.
Nordeen M (1996): Stability Analysis of Cyclic Timetables for a Highly Interconnected Rail
Network. PhD thesis, Swiss Federal Institute of Technology (EPFL), Lausanne.
Pachl J. (2002): Railway Operation and Control. VTD Rail Publishing, Mountlake Terrace.
Soto y Koelemeijer G. (2003): On the Behaviour of Classes of Min-Max-Plus Systems. PhD
thesis, Delft University of Technology, TRAIL Thesis Series, no. T2003/6, Delft.
Subiono (2000): On Classes of Min-Max-Plus Systems and their Applications. PhD thesis, Delft
University of Technology, TRAIL Thesis Series, no. T2000/2, Delft.
Van den Boom T. J. J., De Schutter B (2006): Modelling and control of discrete event
systems using switching max-plus-linear systems. Control Engineering Practice, 14(10),
pp. 1199 – 1211.
Van Egmond R. J. (1999): Railway Capacity Assessment: An Algebraic Approach. TRAIL,
TRAIL Studies in Transportation Science, no. S99/2, Delft.
Van Egmond R. J. (2001): An Algebraic Approach for Scheduling Train Movements. In Voss S.,
Daduna J. R. (Eds.), Computer-Aided Scheduling of Public Transport, LNEMS 505, Springer,
Berlin.

153
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 8.1 Introduction

8 Optimisation Models for Railway Timetabling

Leo Kroon, Dennis Huisman, Gábor Maróti

8.1 Introduction

This Chapter describes mathematical optimisation models and techniques that can be
used for determining cyclic and non-cyclic railway timetables. The current practice in many
European countries is that a railway infrastructure manager is responsible for the railway
infrastructure, while the railway operators are responsible for operating passenger or freight
trains on this infrastructure. The infrastructure manager is responsible for the timetable,
which is usually generated in co-operation with the railway operators. The latter propose their
preferred timetables. The infrastructure manager then combines these into a single timetable,
aiming to satisfy as many demands as possible.
The timetable of a railway operator is usually generated in a number of planning stages. The
process starts with determining the routes and lines concerned. A timetable for each generic
day of the week (Monday to Sunday) is created first. The generic timetables are adjusted
in later planning stages for particular calendar days in order to take into account irregular
maintenance of the railway infrastructure as well as additional trains because of special events
(e. g. sporting fixtures). The timetable can still be modified up to the last minute, by inserting
additional freight trains for example.
A railway timetable consists of two elements: (i) the arrival and departure times of the trains at
the stations and other relevant locations such as railway junctions, and (ii) the assignment of
each train to an appropriate platform track and corresponding inbound and outbound routes
at every relevant station.
Preferably, the generation of the arrival and departure times and the selection of the routes
through the stations are carried out simultaneously. However, when using a model-based
approach for creating a timetable, this would lead to models that are too large to be solved
by the currently available optimisation technology. Therefore, in model-based approaches, the
two steps are commonly split. The arrival and departure times are computed first, and then
the routings through the stations are determined. To get a timetable which is feasible overall,
several iterations of this process may be necessary. This Chapter focuses on the departure
and arrival times of trains. For a description of models for routing trains through railway
stations, see Chapter 9. For more detail of the timetabling models, see Cacchiani and Toth
(2012).
Timetabling is not the only planning problem faced by a railway operator. A railway operator
also has to deal with rolling stock and crew, see Fig. 8.1. An overview of the application of
mathematical optimisation models and other Operations Research models in passenger
railway transportation can be found in Huisman et al. (2005) and in Kroon et al. (2009).
Planning problems in a railway system are solved traditionally in a manual way, based on
the experience and skills of railway planners and possibly with the help of conflict detection
tools. Analytical and simulation models have been developed to support the planning process.
However, these models cannot generate solutions themselves; they can only evaluate those
that were generated elsewhere.
The latter contrast with mathematical optimisation models and techniques. These are able
to generate solutions themselves, guided by appropriately set parameters that define the
objective function(s). In the 1990s, it was recognised that optimisation models may lead to
better solutions, which can be obtained more quickly than by a manual approach. This
155
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
8 Optimisation Models for Railway Timetabling

Fig. 8.1: Planning problems in a railway system

recognition started the research and development of model based decision support tools for
generating railway timetables.
This transition was stimulated from two directions. First, railway operators are more and more
confronted with competition, in particular in Europe. Several of these companies operate
on a commercial or semi-commercial basis, which puts more pressure on efficiency and
operational excellence. The application of mathematical optimisation models may help in
realising these objectives. Second, the scientific community has spent a lot of effort to model
planning problems arising in the railway world in the most appropriate way. Together with
mathematical optimisation techniques – often from the field of mathematical programming –
and increased computing power, many of these problems can nowadays be solved relatively
quickly. Nevertheless, it remains important to balance the level of detail in the models and the
computation time needed for finding appropriate solutions.
Mathematical optimisation models form the basis for the advanced timetabling tools that
are currently used in the railway industry. It is also important that timetables are as robust as
possible. They need to be constructed in such a way that small disturbances in the real-time
operations have only a limited impact on the punctuality.
In this Chapter, the infrastructure between the stations is considered at a macroscopic level
and that within the stations only to a limited extent. Although line planning is often seen as part
of the timetabling process, the models discussed in this Chapter rely on the existence of a line
plan. A line is defined here as a direct railway connection between two terminal stations that is
operated at a given frequency by a given type of train.
Section 8.2 describes mathematical optimisation models for generating cyclic railway
timetables. The most commonly used model is explained in detail, together with an example.
In Section 8.3, mathematical optimisation models for generating non-cyclic timetables are
considered. These models can be used by an infrastructure manager to integrate the preferred
timetables of a number of railway operators. Section 8.4 discusses how the robustness of
a cyclic timetable can be improved, and reports on the results of an experiment in practice.
Section 8.5 concludes with some perspectives on further developments.

156
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 8.2 Cyclic Timetabling

8.2 Cyclic Timetabling

Cyclic or periodic timetable operation means that on all lines the trains run at least every T
minutes. In most cases the cycle time T is equal to 30, 60 or 120 minutes. If the passenger
trains are operated according to a cyclic timetable, then the time slots for freight trains are
often cyclic as well. However, not all planned time slots for freight trains may be actually used
in practice, thereby leading to some loss of railway capacity.
An advantage of a cyclic timetable is that passengers and staff can easily remember the
departure and arrival times of the trains. However, this advantage is quickly diminishing due to
the wide availability of smartphones providing up-to-date and real-time timetable information.
From a planning perspective, a cyclic timetable is less complex than a non-cyclic one; once
planned, you basically have the complete timetable. However, a drawback is that a cyclic
timetable cannot easily offer a large number of direct connections, since it is not flexible within
itself. Another drawback is its potential inefficiency in that the frequency of the services cannot
be adjusted for a lower or higher demand

8.2.1 A mathematical formulation

Most cyclic timetabling models are based on the model for the so-called Periodic Event
Scheduling Problem (PESP), initially developed by Serafini & Ukovich (1989). The PESP model
aims at the cyclical scheduling of a number of events e = 1,…, E. The cycle time of a timetable
is denoted by T. Thus, if an event takes place at time instant v, then a similar event takes place
at all time instants {…, v − 2T, v − T, v, v + T, v + 2T,…}.
Note that the PESP model is a pure scheduling model and not a routing model. Thus, if there
are several routing options as may be the case with a four track railway, then it is assumed that
the route selection has already been made.
With a railway timetable, the events are the arrivals and departures of the trains at the stations
and at the other relevant locations, such as railway junctions. The decision variable ve denotes
the time at which event e is scheduled within each cycle. It is common that all event times are
integers. Then ve should have an integer value in the interval [ 0, T − 1], and all computations
are carried out modulo the cycle time T.
The event times are the start and end times of certain processes in the timetable. For example,
trains have to run from one station to another, they dwell for a certain period of time at each
station, there has to be a certain headway time between two consecutive trains, trains may
have to be split or combined with others, or they may have to make a connection for the
benefit of passengers. The turnaround between the arrival of a train at its end station and
its subsequent departure on the return journey is also a process. Thus there are processes
related to single trains and those related to more than one train.
The set of processes is denoted by P. If a process starts with event e and ends with event e',
then it is denoted by (e, e'). PESP assumes that for each process (e, e') a lower bound Le,e'
and an upper bound Ue,e' for the corresponding process time are known. For example, the
running time of a train between two stations can be determined based on the details of the
infrastructure between the stations (including the signalling system) and the characteristics of
the rolling stock concerned. The lower and upper bound Le,e' and Ue,e' satisfy 0 ≤ Le,e' ≤ T − 1
and 0 ≤ Ue,e' − Le,e' ≤ T − 1. Note that in principle a process with a duration exceeding the
cycle time T does not fit within a cyclic timetable. Nevertheless, it may be useful to consider
upper bounds that exceed the cycle time T.

157
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
8 Optimisation Models for Railway Timetabling

The event times ve and ve' related to the start and end of process (e, e') are now connected by
the following constraints (8.1):

(v e' − v e − L e,e' ) mod T ∈ [0,Ue,e' − L e,e' ] for all (e,e') ∈ P (8.1)

In other words, constraints (8.1) state that for each process the extra duration of the process
above the minimum process time should be between zero and Ue,e' − Le,e' . Here the modulo
operator “mod T ” indicates that the timetable is cyclic, with cycle time T. For example, time
instant 55 is identical to time instant 175 in a cyclic timetable with cycle time T = 60. Also, if
the departure and the arrival time of a train on a certain trip are 55 and 12, respectively, then
the running time of the train on this trip equals (12−55) mod 60 = 17.
Since this modulo operator is rather difficult to handle directly in mathematical optimisation
models due to its strongly non-linear character, the cyclic character of the timetable is usually
modelled by introducing binary decision variables Qe,e' . These indicate whether the process
starting and ending with events e and e' crosses the end of the cycle. That is, constraints (8.1)
are replaced by the following constraints (8.2):
L e,e' ≤ v e' − v e + T × Qe,e' ≤ Ue,e' for all (e,e') ∈ P (8.2)

If the process starting and ending with events e and e' crosses the end of the cycle, then
Qe,e' = 1. If not, then Qe,e' = 0. Loosely speaking, the decision variables Qe,e' determine the
orders of the starting and ending times of the processes within each cycle. In the example
above, the running time of the trip that starts at 55 and ends at 12 equals 12 − 55 + 60 = 17.
Examples of constraints (8.2) are given in Section 8.2.2.
It should be noted that constraints (8.1) are not precisely equivalent to the following constraints
(8.3), which are also often used in the context of PESP models:
(v e' − v e ) mod T ∈ [L e,e', Ue,e' ] for all (e,e') ∈ P (8.3)

In general, constraints (8.3) are more restrictive than constraints (8.1). For example, if there is
a process with ve = 30 , v e' = 31, Le,e' = 55, and Ue,e' = 65, then (8.1) is satisfied but (8.3)
is not. Furthermore, in (8.3) a value Ue,e' > T − 1 does not really make sense, which is not the
case in (8.1). Note that there is only a difference between the constraints if Ue,e' > T − 1. If
Ue,e' ≤ T – 1, then constraints (8.1), (8.2), and (8.3) are equivalent, as can be seen easily.
PESP deals implicitly rather than explicitly with the details of the block occupation staircases
of the trains, described elsewhere in this book. Indeed, if the speed profiles of the trains are
known, then the block occupation staircases can be determined for all trains. Next, these
block occupation staircases can be used to compute the minimum headway times needed
between any pair of trains. This is done so that the planned headway times are respected if
and only if the block occupation staircases of the trains do not overlap each other.
If the speed profiles of the trains are not known in advance (e. g. if the running times are
not fixed), then the headway times can be based on certain assumptions. After a timetable
has been generated by the optimisation model, the correctness of the headway times can
be verified by computing the speed profiles based on the generated timetable. If necessary,
another optimisation run is carried out with modified headway times. In the end, a detailed
simulation run of the obtained timetable can be carried out to verify that the speed profiles and
the headway times are consistent, see also Harrod & Schlechte (2013).

158
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 8.2 Cyclic Timetabling

8.2.2 Example

Section 8.2.2 illustrates the different aspects of timetabling with a simple example, and
explains how they can be modelled in the PESP formulation. Liebchen (2006) and Liebchen &
Möhring (2008) give a further comprehensive overview of the modelling power of the PESP
model. The example described here considers the triangle Amersfoort (Amf) – Deventer (Dv)
– Zwolle (Zl) in the Netherlands, as depicted in Fig. 8.2. The other relevant stations in this
example are Apeldoorn (Apd), Harderwijk (Hd) and Olst (Ost). Between Deventer and Olst, the
infrastructure is single track only.

Fig. 8.2: The triangle Amersfoort – Deventer – Zwolle

The line plan is a minimum frequency of two trains per hour. In Table 8.1, the details of the line
plan and the minimum running times are described. The maximum running times per trip are
assumed to be equal to the minimum running times plus one minute. Note that in practice the
minimum running times may depend on the rolling stock type.
The columns in the table should be read as follows: the Intercity Amersfoort – Deventer dwells
in Apeldoorn and has running times of 24 or 25 minutes (Amersfoort – Apeldoorn) and 10 or
11 minutes (Apeldoorn – Deventer), respectively. It runs every 30 minutes. In the remainder of
this section, this train is denoted with train number 1a in the direction of Deventer and with
train number 1b in the reverse direction.
The dwell times at all stations in Table 8.1 should be 1 or 2 minutes. Stations not mentioned
in Table 8.1 are irrelevant in this example. Thus for the Regional train Amersfoort – Zwolle, only
the stop in Harderwijk is listed. The dwell times for the other stops are included in the running
times. The Intercity trains Amersfoort – Zwolle (2a, 2b, 3a, and 3b) do not stop intermediately.
However, in order to construct a timetable, the passing times in Harderwijk must be known,
since in Harderwijk the Regional trains can be overtaken by the Intercity services.
In the following, examples of constraints (8.2) are given. Because the timetable is repeated
every 30 minutes, the cycle time T is equal to 30 minutes in the model. Furthermore, arrst
and depst are defined as the arrival and departure event of train t at station s. For the sake
of simplicity, the symbol v is removed from the notation of the event times. The arrival and
departure time of train t at station s as is denoted as Ast and Dst .

159
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
8 Optimisation Models for Railway Timetabling

Train type Route Relevant stops Min. running time Frequency Train
Intercity Amf – Dv Apd 24 + 10 1 1a, 1b
Intercity Amf – Zl – 15 + 20 2 2a, 2b, 3a, 3b
Intercity Dv – Zl Ost 6 + 14 1 4a, 4b
Regional Amf – Zl Hd 28 + 26 1 5a, 5b

Table 8.1: Line plan for the example instance of PESP

The first constraints which need to be satisfied are those of the running time. Analogous to
the general definition, dep1a
Amf is the departure of train 1a (Intercity Amersfoort – Deventer) from
1a
Amersfoort, and arrApd is the arrival of the same train at Apeldoorn. The constraint to ensure
the required running time of 24 or 25 minutes on this trip is as follows:
24 ≤ A1a 1a
Apd − D Amf + 30 × Qe,e' ≤ 25 with e = dep1a 1a
Amf and e' = arrApd (8.4)

The next constraints considered are those of the dwell time. For example, a dwell time of 1 or
2 minutes in Apeldoorn for train 1a is modelled as follows:
1≤ D1a 1a
Apd − A Apd + 30 × Qe,e' ≤ 2
1a
with e = arrApd and e' = dep1a
Apd (8.5)

A required passenger connection between two trains may have to be scheduled. For example,
if on a trip from Apeldoorn to Zwolle, a transfer time between 2 and 5 minutes in Deventer
from train 1a to train 4a is required, then this can be guaranteed as follows:
4a
2 ≤ DDv − A1a
Dv + 30 × Qe,e' ≤ 5
1a
with e = arrDv 4a
and e' = depDv (8.6)

The next type of constraints considered are the headway constraints. For instance, the
minimum headway between an Intercity and a Regional train in both Amersfoort and
Harderwijk must be at least 3 minutes. This can be formulated as follows. (Note that this
constraint holds for both Intercity trains, but it is described only for train 2a).
15 ≤ D2a 5a
Amf − D Amf + 30 × Qe,e' ≤ 27 with e = dep5a 2a
Amf and e' = dep Amf (8.7)

The difference in the departure times should be at least 15 minutes, because the Regional
train cannot be overtaken by the Intercity between Amersfoort and Harderwijk; to guarantee
a time difference of at least 3 minutes in Harderwijk, the minimum difference in running times
should be added to the minimum headway time to obtain the right minimum difference in
departure time in Amersfoort (3 + (28 − 16) = 15). The difference in the departure times should
be at most 27 minutes, because otherwise the Regional train would depart too soon after the
Intercity. A similar constraint is needed for the arrivals of these trains in Harderwijk, see also
Kroon & Peeters (2003).
Other types of headway constraints can be formulated, such as single track and crossing
constraints in stations. An example of the first type involves the single track between Deventer
and Olst. The difference in the departure times of the Intercity from Deventer to Zwolle and that
in the opposite direction from Zwolle to Deventer should be at least 2 minutes on both ends of
the single track section. The constraint for station Deventer reads as follows:
4b 4a 4a 4b
14 ≤ A Dv − DDv + 30 × Qe,e' ≤ 28 with e = depDv and e' = arrDv (8.8)

The lower bound follows from the fact that, after the departure of train 4a from Deventer, at
least twice the minimum running time on the single track line (6 minutes) plus the minimum
headway time in Olst (2 minutes) should elapse before the arrival of train 4b in Deventer. A
similar constraint is needed at Olst.

160
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 8.2 Cyclic Timetabling

Note that in this section only a subset of the relevant constraints is described. For an overview
of all relevant constraints, see Dollevoet et al. (2007).

8.2.3 Objectives

Any feasible solution to the PESP system of constraints provides a valid timetable. This is often
an achievement in itself, but the usual aim is to get a timetable that is as good as possible.
“Good” can have many meanings here.
For instance, passengers prefer short travel times (on each single train, and on their overall
journey). This may be achieved by having short running times per trip, short dwell times,
and short transfer times. The railway operator may prefer short running times as well, since
these reduce both rolling stock and train crew requirements. The operator may also prefer
a timetable with short turn-around times. However, the foregoing may conflict with the
requirement of a robust timetable, which is the topic of Section 8.4. Thus there is a trade-off.
One of the advantages of mathematical optimisation models is that they can be used to make
an explicit trade-off between conflicting objectives. This can be accomplished by running the
model with different parameter settings.

8.2.4 Solving PESP

The binary decision variables Qe,e' that determine the orders of certain pairs of events in each
cycle make the PESP model quite hard to solve by standard branch-and-bound methods. Due
to the relatively large coefficient T in constraints (8.2), the Linear Programming (LP) relaxation
of a model based on this formulation is usually quite weak.
In the LP relaxation the binary decision variables Qe,e' are replaced by continuous decision
variables that should have a value between zero and one. LP relaxation is the most widely
used relaxation method in branch-and-bound procedures. A weak LP relaxation is a drawback
in a branch-and-bound procedure, since it makes it hard to estimate how far the solution
obtained differs from the optimal.
Therefore, Schrijver & Steenbeek (1994) developed an algorithm for solving PESP based on
a completely different approach, namely constraint propagation. Their algorithm, CADANS,
has been implemented in the DONS system. This system has become an indispensable tool
in the Dutch long term railway timetabling process in the Netherlands, see Hooghiemstra
et al. (1999). The constraint propagation approach first looks for a feasible solution only, but
Schrijver & Steenbeek (1994) also developed local optimisation techniques to improve a
feasible solution for fixed values of the variables Qe,e' in constraints (8.2). Instances with up to
250 trains (all those running in one hour of the Dutch timetable) can be solved usually within
reasonable computation times.
Nachtigall & Voget (1996) use PESP to generate cyclic timetables with minimal passenger
waiting times. Odijk (1996) uses PESP at a strategic level to determine the capacity of the
infrastructure around railway stations. Kroon & Peeters (2003) describe a PESP model with
variable running times. This gives a higher probability of obtaining a feasible model solution;
indeed, fixed running times may result in an over-constrained (thus infeasible) timetabling
problem.
In order to cope with the weak LP relaxation of models based on constraints (8.2), Nachtigall
(1999), Peeters (2003) & Liebchen (2006, 2008) also describe a formulation of PESP in terms
of cycle bases. This formulation does not use the event times as the decision variables, but

161
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
8 Optimisation Models for Railway Timetabling

the process times. Moreover, this formulation uses the fact that the sum of the process times
along each directed cycle of the so-called constraint graph is a multiple of the cycle time. This
constraint graph contains a node for each event time e, and an arc from node e to node e' for
each process (e, e' ).
The cycle base formulation is somewhat easier to solve than the standard PESP formulation
based on constraints (8.2), because of its lower number of integer variables and the better
LP relaxation. Several classes of cutting planes are described in the papers mentioned, which
intend to further improve the LP relaxation of the model. Liebchen (2006, 2008) gives an
extensive overview of these results. Liebchen used this model to generate a completely new
timetable for the Berlin Underground system.
Vansteenwegen, Van Oudheusden (2006, 2007) present a model that aims at minimising
the travel times of the passengers by focussing on their transfer times. The authors take into
account the stochastic nature of the running times and the dwell times of trains, but their
model does not include propagation of delays of individual trains. The method starts with an
existing timetable and aims at improving it by slightly modifying the arrival and departure times.
That implies that the basic structure of the timetable has already been decided, so the model
does not need the complicating binary Qe,e' decision variables. Thus the model can be solved
by using LP techniques only.

8.3 Non-cyclic Timetabling

This Section is devoted to mathematical optimisation models that can be used for generating
non-cyclic timetables. Several models have been proposed. This Section focuses on those
presented by Cacchiani et al. (2008, 2010) and Caprara et al. (2002). These are models to
be used by an infrastructure manager for integrating the preferred timetables of the different
railway operators, as described in Section 8.1. That is, the starting point of the models is a set
of preferred timetables of the individual operators. The goal is then to integrate them into one
timetable that fits on the available railway infrastructure. The objective is a combination of two
criteria: to maximise the financial benefit of the trains and to modify the preferred timetables
of the railway operators as little as possible. The financial benefit itself arises from the revenue
of running the train and from cost elements such as track access charges payable to the
infrastructure manager.

8.3.1 Graph representation of the timetable

The models for generating non-cyclic timetables use a directed graph to represent the
timetable, and a discretization of time, usually with a time unit of one minute.
The set of trains to be scheduled is denoted by T. The nodes in the graph are associated
with the arrival and departure events of the trains; thus each node corresponds naturally to a
station and to a time instant. The set of nodes is denoted by V.
The arcs in the graph are trip arcs and wait arcs. The trip arcs correspond to the movements
of the trains from one station to the next. A trip arc connects a node representing the
departure station and time of a train with the node representing the arrival station and time
of this train. The wait arcs connect consecutive relevant nodes at the same station with each
other. They correspond to trains waiting at a station. The complete set of arcs in the timetable
graph is denoted by A.

162
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 8.3 Non-cyclic Timetabling

The timetable of a single train is represented by a directed path in this timetable graph, from
the node corresponding to the origin station of the train and its departure time, to the node
corresponding to the destination of the train and its arrival time there. The timetable of several
trains is represented by the collection of the paths of the single trains.
In order to be able to modify the arrival and departure times of single trains, the graph above is
extended by adding nodes and arcs to it. The extended set of nodes represents the possible
arrival and departure times of the trains at the stations, while the extended trip arcs represent
the possible trips of the trains from one station to the next. The wait arcs again connect
consecutive nodes corresponding to the same station. The operations that may be applied
to the trips of a train are shift and stretch. A shift operation moves a trip forward or backward
in time, and a stretch operation increases or decreases the duration of the trip. The deviation
costs of the modified trips directly depend on the shifts and stretches applied to the operator’s
preferred trips.
The extended graph is a directed acyclic graph; this fact allows one to find shortest paths in
the graph (with respect to any length function) efficiently. The timetabling models based on the
extended graph can be divided into arc based models and path based models. These variants
are described in Sections 8.3.2 and 8.3.3 respectively.

8.3.2 Arc based timetabling model

Caprara et al. (2002) describe an arc based model for scheduling a number of trains on some
important Italian railway corridors. Cacchiani et al. (2010) use a similar model for scheduling
additional freight trains on the railway corridor between Kufstein in Austria and Rome in Italy.
These freight trains must be scheduled between the existing passenger trains. Thus for each
train, only a discrete set of time slots is available.
In the model of Caprara et al. (2002), a number of consecutive arcs of the extended timetable
graph is selected for each train in order to indicate its timetable. To that end, the model
contains binary arc variables Ya,t for each arc a in A and each train t in T. These variables
indicate whether arc a is traversed by train t. In addition, the binary variables Zv are used for
each node v in V. These variables describe whether node v is visited by any train. Based on
these variables, the model can be formulated as follows:
Maximise ∑ ∑ Ra,t Ya,t (8.9)
a∈A t∈T

The constraints of the arc based timetabling model can be expressed as follows:

∑ Ya,t ≤ 1 for all t ∈ T (8.10)

a∈St

∑ Ya,t = ∑ Ya,t ≤ 1 for all t ∈ T and v ∈ Vt (8.11)

a∈−
v a∈+
v

∑ (Ya,t + Ya ',t ) ≤ 1 for all (a,a') ∈ K a (8.12)

t∈T

∑ ∑ Ya,t = Z v for all v ∈ V (8.13)

a∈−
v
t∈T:v∈ Vt

Z v + Z v' ≤ 1 for all (v, v') ∈ K v (8.14)

All decision variables are binary (8.15)

163
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
8 Optimisation Models for Railway Timetabling

Here the function (8.9) specifies that the objective is to maximise the total profit of the
timetable. The total profit is defined as the sum of the individual profits Ra,t of the selected arcs.
The individual profit of an arc is computed from the financial benefits of running the train on the
arc as well as from the costs associated with shifting and stretching the arc.
Next, constraints (8.10) describe that for each train t in T at most one arc from the set of
starting arcs St of train t must be selected. Constraints (8.11) guarantee that, if a train t arrives
in an intermediate node v in the timetable graph, then it leaves from there also. Here the
symbols v− and v+ denote the sets of incoming and outgoing arcs of node v, respectively.
Constraints (8.12) forbid two conflicting arcs to be used simultaneously. Here the set K a is
the set of pairs of conflicting arcs. For example, two arcs representing one train overtaking
another on a single track are conflicting. Constraints (8.13) form the connection between the
arc variables and the node variables. Finally, constraints (8.14) show that two conflicting nodes
cannot be visited by two trains at once. Here the set K v is the set of pairs of conflicting nodes.
For example, two nodes representing two arrivals in a station that are too close after each
other are conflicting.
It should be noted that constraints (8.12) and (8.14) can be strengthened by including as many
variables as possible in their left-hand sides, see Cacciani et al. (2010).
The dimensions of the model (8.10)–(8.15) can be quite large in real life. Caprara et al.
(2002) describe a heuristic solution approach based on Lagrangian relaxation. They apply a
subgradient procedure for generating good Lagrangian multipliers, and they use a heuristic
procedure for generating solutions. The trains are scheduled one by one, based on
the Lagrangian costs. For each train this is a shortest path problem, which can be solved
efficiently in the directed acyclic extended timetable graph. Afterwards, the solution is
improved by applying local search heuristics.

8.3.3 Path based timetabling model

Cacchiani et al. (2008) describe a path based model for scheduling a number of trains on a
number of important Italian railway corridors. For each train t a set of possible timetables is
generated as a start. Each timetable for train t is represented by a path p in the extended
timetable graph. The set of all possible timetables for train t is denoted by Pt . The set P is
formed by the union of the sets Pt . In principle, these sets can be extremely large. However,
an aim of the infrastructure manager is to modify the preferred timetables of the railway
operators as little as possible. That in itself may limit the number of possible paths.
As described before, the costs of path p for train t depend on its differences with the original
timetable for train t. Therefore the operator may wish to cancel a train path, if its modification
is such that the costs of the necessary shifts and stretches outweigh the revenue of the train
path.
The timetabling problem now is to select for each train t at most one path in Pt , such that the
total costs of the selected paths is minimal, and that the combination of all paths results in a
feasible timetable. Note that at most one path per train needs to be selected and not exactly
one path, since it is allowed to cancel trains. This timetabling problem can be formulated as a
so-called Set Packing model, by introducing a binary decision variable Xp for each path p in P.
The objective is to maximise the total profit of the selected paths:

Maximise ∑ Rp Xp (8.16)
p∈P

164
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 8.4 Robust timetabling

Here Rp denotes the profit of path p ; this value depends on the financial benefit of the path
as well as on the shifts and stretches of the involves arcs. The constraints of the Set Packing
model can be expressed as follows:

∑ Xp ≤ 1 for all t ∈ T (8.17)

p∈Pt

X p + X p' ≤ 1 for all (p,p') ∈ K (8.18)

All decision variables are binary (8.19)

Constraints (8.17) describe the need to select for each train t at most one path p in the set
Pt . Furthermore, constraints (8.18) forbid conflicting paths to be selected simultaneously. The
set K consists of all conflicting pairs of paths. Note that all the details of the blocking time
staircases can be taken into account in the set K.
Constraints (8.18) are known to provide a weak LP relaxation. A stronger formulation is
achieved by including a larger number of mutually conflicting paths into the constraints
instead of just two. Improved constraints of this kind are obtained from cliques in the conflict
graph. The nodes of the undirected conflict graph are paths in the set P, and two nodes are
connected with each other if the corresponding paths are conflicting. A clique is a subset of
nodes with an edge connecting each pair of its elements; notice that each pair of conflicting
paths forms a clique. Letting  denote the set of cliques, Constraints (8.18) can be improved
upon by replacing them by:

∑ Xp ≤ 1 for all k ∈  (8.20)

p∈k

The strongest clique constraints arise from containment-wise maximal cliques, they in fact
imply all other clique constraints.
Cacchiani et al. (2008) describe a Branch-and-Cut-and-Price algorithm to produce an optimal
solution. Dynamic column generation is used to deal with the large number of decision
variables in the model. Starting with the set of paths corresponding to the given preferred
timetables, the set of paths is extended by looking for useful paths in the extended timetable
graph. This pricing problem can be formulated as a shortest path problem, which can be
solved efficiently in the acyclic extended timetable graph. In addition, the model handles the
very large number of constraints (8.18) or (8.20) by applying adequate separation algorithms,
see Cacchiani et al. (2008).

8.4 Robust timetabling

One objective is to create a timetable that is as robust as possible, to deal with relatively small
disturbances in real-time operations, see Jen (2004) and Sastry & Bodson (1989).
A robust timetable has one or more of the following properties:
(i) initial disturbances can be absorbed to some extent so that they do not lead to delays,
(ii) there are few knock-on delays from one train to another, and
(iii) delays disappear quickly by using straightforward train dispatching actions (e. g. by
changing the order of a few trains on a part of the infrastructure).
Both (i) and (iii) can be achieved by the insertion of time supplements (i. e. additional time on
top of the minimal travelling or dwelling times) in the timetable, and (ii) follows from placing
buffer times between consecutive trains at certain locations.

165
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
8 Optimisation Models for Railway Timetabling

Note that, being limited to straightforward dispatching measures, a timetable can only be
robust against small disturbances. It is impossible to create a timetable that is robust against
major disruptions, such as a track blockage or a train failure, without re-routing or cancelling
trains.
Several approaches have been developed to analyse the robustness of timetables. For
examples based on (i) simulation, see Medeossi et al. (2009), (ii) queuing theory, see Huisman &
Boucherie (2001) & Wendler (2007), (iii) Max-Plus algebra, see Goverde (2007), and (iv)
statistical analysis, see Yuan & Hansen (2007). Some of these approaches are described in
other Chapters of this book.
Approaches based on mathematical optimisation models include the Delay Resistant
timetabling model of Liebchen et al. (2009, 2010), the Light Robustness approach of Fischetti
et al. (2009), and the Flexible Periodic Event Scheduling Model by Caimi et al. (2011), see also
Vansteenwegen, Van Oudheusden (2006, 2007).
This Section is devoted to a Stochastic Optimisation Model, which modifies the time
supplements and the buffer times of an existing timetable slightly so that they can be as
effective as possible. The model described here assumes the initial timetable to be cyclic but,
mutatis mutandis, it can also be used for non-cyclic timetables.
The model requires an initial cyclic timetable as input and it is assumed that both this and the
improved timetable have a cycle time T of one hour. The term hours is used instead of the
term cycles in what follows.
The model contains separate parts for determining the improved timetable, and a simulation
part for evaluating it. To simulate the timetable, R realisations of the trains in the timetable are
operated subject to preselected initial stochastic disturbances. Here each realisation covers H
consecutive hours of the timetable. Each realisation can be considered as a simulation of the
timetable during a single day.
Whereas the PESP model described in Section 8.2 includes the problem of determining the
cyclic order of the trains on each part of the infrastructure, it is assumed here that these orders
are given in the initial timetable and cannot be modified by the model. This assumption helps
to keep the computation times within certain limits.
The simulation part of the model does not include active dispatching measures. In particular, it
is assumed here that the trains in the improved timetable retain the same order relative to each
other. That includes that transfer possibilities between passenger trains are maintained in each
realisation. Thus the trains depart as early as the initial disturbances, the accumulated delays,
and the relations with other trains permit.

8.4.1 Notation

As described in Section 8.2, a timetable consists of a number of processes. A process that

starts with event e and ends with event e' is denoted by (e, e'). In the given cyclic timetable,
the planned time of event e is described by the parameter Ve. This parameter denotes the
offset of the event time in each hour of the given cyclic timetable. Thus, in the given timetable,
process (e, e') starts and ends at time instants Ve and Ve' , respectively. A process (e, e') that
falls completely within the hour has Ve < Ve' , and a process (e, e') that crosses the end of the
hour has Ve > Ve' . Here it is assumed, without loss of generality, that the process times do not
exceed the cycle time T.
To distinguish these cases, for each process (e, e') a binary parameter Ke,e' is introduced. This
records whether or not the process (e, e') crosses the end of the hour in the initial timetable. In

166
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 8.4 Robust timetabling

other words, Ke,e' = 0 if and only if Ve > Ve' . The model is not allowed to modify the given cyclic
orders of the events. Note that the role of the parameters Ke,e' is similar to that of the decision
variables Qe,e' that were described in Section 8.2.
In the improved timetable, the planned time of event e is denoted by the decision variable ve.
The decision variable se,e' denotes the planned time supplement for the time of process (e, e').
In order to make sure that the end of the hour does not lead to undesirable restrictions for the
planned event times, the values ve are not restricted to the time interval [0, T − 1]. They may
take any (integer) value.
As mentioned, the timetable is evaluated during its generation by operating R realisations of
the timetable subject to a priori selected independent stochastic disturbances. Each realisation
covers H consecutive hours of the cyclic timetable. The stochastic disturbance of process
(e, e') in hour h of realisation r is denoted by the parameter e,e',r,h for all (e, e')  P and for all
r = 1,…, R, and h = 1,…, H. The realised time of event e in hour h of realisation r is denoted
by the decision variable we,r,h. The event times of each realisation are assumed to occur on a
linear time axis. Thus they are certainly not restricted to the time interval [0, T − 1].
The delays of the events corresponding to arrivals of trains are evaluated, but other delays
can be taken into account. Arrival events are, by definition, those whose delays are measured.
The set of arrival events is denoted by Ea. The delay of arrival event e in hour h of realisation
r is denoted by the decision variable ∆e,r,h. The average weighted delay of all trains is denoted
by ∆. Certain events, such as a departure of a train, should not precede their corresponding
planned event times. Such events are called departure events, the set of which is denoted
by Ed .

8.4.2 Timetabling part of the model

Most of the constraints to be satisfied in a cyclic timetabling model can be expressed in terms
of the planned event times and the planned process times. For each process (e, e') that may
include a variable amount of time supplement se,e', the following constraints must be satisfied:
Me,e' + se,e' = v e' − v e + T ×K e,e' for all (e,e') ∈ P (8.21)

The left-hand side of this equation describes the planned time of a process (e, e') as the sum
of the technically minimum process time Me,e' and the variable time supplement se,e' . The right-
hand side describes it as the time difference between the planned completion time and the
planned commencement time of process (e, e'), thereby taking into account that an event
might continue into the next hour by the term T × Ke,e' .
For all processes, a (possibly void) upper bound Ue,e' on the planned process time is specified.
This results in the following constraints:
v e' − v e + T ×K e,e' ≤ Ue,e' for all (e,e') ∈ P (8.22)

Other relevant constraints specify that, at each part of the infrastructure, the time difference
between the last and the first planned event time within each hour should be less than the
cycle time T. To that end, let e be the first planned event in an hour on a certain part of
the infrastructure, and let e' be the last planned event in an hour on the same part of the
infrastructure. Then the following constraint must be satisfied:
0 ≤ v e' − v e ≤ T − 1 (8.23)

Constraints (8.23) are important, since the event times are not restricted to the time interval
[0, T − 1], as was explained earlier. These constraints guarantee that, after the optimisation, all
the planned event times obtained can be transferred back into the time interval [0, T − 1].

167
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
8 Optimisation Models for Railway Timetabling

Next, in order to control the allocation of the time supplements to the planned process times,
Q subsets B1,…, BQ of processes are selected. Each subset Bq of processes is connected
with a certain given amount of time supplement Sq that is to be allocated to the processes in
the set Bq. Then the following constraints are to be satisfied:
∑ se,e' ≤ Sq for all q = 1,,Q (8.24)
(e,e')∈Bq

Such a constraint may indicate that a certain total amount of running time supplement is to be
allocated to the consecutive running times along the path taken by a single train. However, a
certain amount of time supplement may also be allocated to a number of lines together.
Note that for modelling the processes that involve pairs of trains as in (8.21) – (8.23), the
assumption that the order of the events should remain unchanged is essential. Indeed, if
the order of the events was previously unknown, then additional binary variables would be
required to model them. Finally, if the timetable has to be expressed in integer minutes, then
integral constraints have to be imposed on the planned event times.

8.4.3 Simulation part of the model

It will be recalled that the simulation part of the model does not include dispatching decisions.
The H hours of each realisation are operated consecutively. A process (e, e') with Ke,e' = 0 has
Ve < Ve' . Thus it is planned within a single hour. Therefore, it can be assumed that in the
realisations a process (e, e') with Ke,e' = 0 ends in the hour in which it started.
However, a process (e, e') with Ke,e' = 1 has Ve' < Ve. Obviously, in the realisations it is
impossible for a process to end earlier than it started. Thus it is assumed that a process
(e, e') with Ke,e' = 1 ends in a later hour than that in which it started. Thus a process (e, e')
with Ke,e' = 1 that starts in hour h of realisation r at time instant we,r,h ends in hour h + 1 of
realisation r at time instant we,r,h+1 .
This implies that the following constraints link the event times of the processes to the
technically minimum process times and the disturbances:
Me,e' + e,e',r,h ≤ w e',r,h+K e,e' − w e,r,h for all (e,e') ∈ P;r = 1,,R;h = 1,,H (8.25)

As a consequence, a delayed train in hour h of realisation r may influence the trains in hour
h + 1 of realisation r. Note that the realised process times do not have an upper bound, in
contrast with the planned process times that may have one. Indeed, the realised process
times should have the freedom to be extended indefinitely, depending on the sizes of the
disturbances in (8.25). Note further that (8.25) is an inequality and not an equality. Indeed,
(8.25) only deals with process times that are enlarged due to initial disturbances e,e',r,h.
However, trains may also suffer knock-on delays from interactions with other trains.
Departure events should not occur too early, and a delay corresponds to a late arrival event.
This results in the following constraints:
v e + h× T ≤ w e,r,h for all e ∈ E d;r = 1,,R;h = 1,,H (8.26)

w e,r,h − (v e + h× T) ≤ e,r,h for all e ∈ E a ;r = 1,,R;h = 1,,H (8.27)

Here the cyclic character of the timetable is used, since the planned time of event e in hour h
of realisation r equals ve + h × T. Imported delays of trains that enter the area studied can be
modelled by adding a disturbance term to the left-hand side of (8.26). Finally, all delay variables
∆e,r,h are non-negative. Positive delays to trains should not be compensated by negative delays
to other trains.

168
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 8.4 Robust timetabling

As was indicated earlier, the objective is to minimise the weighted total (or average) delay of
the trains. Thus the objective is to
R H
minimise  = ∑ ∑ ∑ cee,r,h (8.28)
e∈Ea r =1 h=1

Here the weights ce indicate the weights of the different delays. Delays are weighted, since
a delay of one train at a certain location may be harmful to more passengers than a delay of
another train at another location. Note that, besides average delays of trains, several other
aspects of the timetable may be incorporated in the objective function.

8.4.4 Results and timetable experiment in practise

Computational results obtained with the Stochastic Optimisation Model can be found in
Vromans (2005) and in Kroon et al. (2007, 2008). These show that the punctuality of a
railway system can often be improved significantly, even by applying only minor timetable
modifications.
In particular, with the same total time supplements, the optimised timetables showed fewer
knock-on delays in the simulation part of the model, and once delays of trains did occur, these
were absorbed more quickly. In these papers it is also shown that the guideline of UIC (2000,
2004) to allocate to each trip an amount of running time supplement that is a fixed percentage
of the technically minimum running time of the trip does not lead to a timetable with maximum
robustness.
In mid-2006, a timetable generated by the Stochastic Optimisation Model was tested
in practice on the Zaanlijn. On the “Zaanlijn” a mix of Intercity trains and Regional trains is
operated by Netherlands Railways.
The service had been notorious for its poor punctuality for several years. It was decided to
carry out a number of experiments by operating, temporarily, slightly different timetables arising
from the Stochastic Optimisation Model.
In the optimisation, the total running time supplement and the dwell times were both
unchanged. The journey times of the passengers changed only marginally.
Table 8.2 shows the average punctuality of the “Zaanlijn” per week as well as that on the
whole Dutch railway network. Punctuality is defined here as the percentage of trains arriving
with a delay of less than three minutes.

Week 22 23 24 25
Zaanlijn 89.7 % 87.5 % 80.5 % 89.5 %
Overall 88.6 % 87.9 % 82.6 % 88.8 %

Week 26 27 28 29 Average
Zaanlijn 85.3 % 74.6 % 90.2 % 85.8 % 85.4 %
Overall 86.5 % 75.4 % 87.3 % 81.2 % 84.8 %

Table 8.2: Punctuality of the Zaanlijn and overall punctuality on NS

During the experiment, the punctuality of the “Zaanlijn” turned out to be quite comparable with
that of the network as a whole. Moreover, the average punctuality of the “Zaanlijn” over the
first 13 weeks of 2006 was 79.4 %, compared with a network average of 86.5 %. Thus the
punctuality of 85.4 % on the “Zaanlijn” during the experiment was a clear improvement over

169
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
8 Optimisation Models for Railway Timetabling

the regular 79.4 %. It can be concluded that the application of the Stochastic Optimisation
Model had a positive effect on the punctuality of the “Zaanlijn”. Note that weeks 24, 27 and 29
had a number of major infrastructure problems, which had a negative effect on the punctuality
of the overall railway network.
In fact, the effects were very much as expected: at moments with only small initial
disturbances, the improved timetable is better able to deal with them. For large disturbances,
the improved timetable does not give a clear advantage over the original one.
Based on the expected positive effect, the Stochastic Optimisation Model was recently
implemented within the DONS system. This is used for generating cyclic timetables for one
hour for the Dutch railway system, see Section 8.2.4.

8.5 Conclusion

An advantage of the application of mathematical optimisation models for solving railway

planning problems, and in particular timetabling problems, is that it may lead to better
solutions in a shorter throughput time. Better solutions may be the result of the fact that the
application of the optimisation models enables the study of several solutions instead of just a
single one. Indeed, it is relatively easy to generate several solutions by running an optimisation
model under different parameter settings. This enables a detailed study of the trade-off
between costs and service in detail. Moreover, the application of optimisation models may lead
to improvements in several criteria at the same time. This is in contrast with manual planning
methods or heuristic approaches.
Optimisation models play an increasingly important role in the planning processes of railway
systems. For example, the December 2005 timetable of the Berlin Underground was created
completely using techniques based on the PESP model and the corresponding optimisation
techniques described in Section 8.2, see Liebchen (2006, 2008).
Similarly, the models described in this Chapter have developed into indispensable tools in the
planning process of the Dutch railway system. They have been implemented in the automatic
timetabling system DONS (Designer Of Network Schedules). This system has been used
extensively in the development of the Dutch railway timetable that went in operation from
December 2006. Whereas in earlier years new timetables were usually created incrementally
by modifying the existing one, this timetable was generated from scratch with the help of
the models described. In addition, the available models for rolling stock planning and crew
scheduling played a crucial role in this planning process. More details can be found in Fioole
et al. (2006), Abbink et al. (2005) and Kroon et al. (2009).
A principal objective was to improve the timetable’s robustness, which should translate into
better punctuality. This objective was attained and in 2007 the punctuality of the Dutch railway
system reached what was then an all time high of 87 % of trains arriving within 3 minutes of
time and 93 % within 5 minutes.
A major challenge now is to find a better way of integrating the models for determining the
arrival and departure times of trains, and those that have been developed for routing trains
through stations. This would allow more details of the routes through the stations being taken
into account in constructing the timetable.
Until now, the planning process has been the traditional application area of mathematical
optimisation models. However, decision support tools are also needed in disrupted situations
to decide in real-time which trains to cancel or to re-route in order to retain as much as
possible of the promised service, or how to adjust the rolling stock and crew schedules.

170
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

Disruption management requires a fundamentally different focus than the planning process.
Quick results are essential, and optimality of solutions is less relevant. Disruption management
is currently an active research area leading to many publications. Also the first applications are
being implemented in practice, but still many challenges are still to be solved.

References
Abbink E. J. W., Fischetti M., Kroon L. G., Timmer G. and Vromans M. J. C. M: Reinventing
Crew Scheduling at Netherlands Railways. Interfaces, 35: 393 – 401, 2005.
Cacchiani V., Caprara A., Toth P. A.: Column Generation Approach to Train Timetabling on a
Corridor. 4OR A Quarterly Journal of Operations Research, 6: 125 – 142, 2008.
Cacchiani V., Caprara A., Toth, P.: Scheduling Extra Freight Trains on Railway Networks.
Transportation Research Part B, 44: 215 – 231, 2010.
Cacchiani V., Toth P.: Nominal and Robust Train Timetabling Problems. European Journal of
Operational Research, 219: 727 – 737, 2012.
Caimi G., Fuchsberger M., Laumanns M., Schüpbach K.: Periodic Railway Timetabling with
Event Flexibility. Networks 57: 3 – 18, 2011.
Caprara A., Fischetti M., Toth P.: Modelling and Solving the Train Timetabling Problem.
Operations Research, 50: 851 – 861, 2002.
Dollevoet T., Van Jaarsveld W., Van der Hurk E.: Railway Logistics – Timetabling. Technical
Report for “Topics in OR and Logistics”. Erasmus University Rotterdam, 2007.
Fioole P.-J., Kroon L. G., Maróti G., Schrijver A.: A Rolling Stock Circulation Model for
Combining and Splitting of Passenger Trains. European Journal of Operational Research, 174:
1281 – 1297, 2006.
Fischetti M., Salvagnin D., Zanette A.: Fast Approaches to Improve the Robustness of a
Railway Timetable. Transportation Science 43: 321 – 335, 2009.
Goverde R.: Railway Timetable Stability Analysis Using Max-Plus System Theory.
Transportation Research B, 41: 179 – 201, 2007.
Harrod S., Schlechte T.: A Direct Comparison of Physical Block Occupancy Versus Timed
Block Occupancy in Train Timetabling Formulations, Technical Report of ZIB, Berlin,
pp. 13 – 18, 2013.
Hooghiemstra J, Kroon L G, Odijk M, Salomon M and Zwaneveld P J. Decision Support
Systems Support the Search for Win-Win Solutions in Railway Network Design. Interfaces, 29:
15 – 32, 1999.
Huisman D., Kroon L. G., Lentink R. M., Vromans M. J. C. M.: Operations Research in
Passenger Railway Transportation. Statistica Neerlandica, 59: 467 – 497, 2005.
Huisman T., Boucherie R. J.: Running Times on Railway Sections with Heterogeneous Train
Traffic. Transportation Research Part B, 35: 271 – 292, 2001.
Jen E.: Robust Design: A Repertoire of Biological, Ecological, and Engineering Case Studies.
Santa Fe Institute: Studies in the Sciences of Complexity. Oxford University Press, 2004.
Kroon L. G., Dekker R., Vromans M. J. C. M.: Cyclic Railway Timetabling: a Stochastic
Optimisation Approach. In: Geraets F., Kroon L. G., Schöbel A., Wagner D., Zaroliagis C.
(editors): Algorithmic Methods for Railway Optimisation, Lecture Notes in Computer Science.
Springer, Berlin, 2007.

171
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
8 Optimisation Models for Railway Timetabling

Kroon L. G., Huisman D., Abbink E. J. W., Fioole P.-J., Fischetti M., Maróti G., Schrijver A.,
Steenbeek and Ybema R. J.: The New Dutch Timetable: The OR Revolution, Interfaces, 39:
6 – 17, 2009.
Kroon L. G., Maróti G., Retel Helmrich M., Vromans M. J. C. M., Dekker R.: Stochastic
Improvement of Cyclic Railway Timetables. Transportation Research Part B, 42: 553 – 570,
2008.
Kroon L. G., Peeters L. W. P.: A Variable Trip Time Model for Cyclic Railway Timetabling.
Transportation Science, 37: 198 – 212, 2003.
Liebchen C.: Periodic Timetable Optimization in Public Transport. PhD thesis, TU Berlin,
Germany, 2006.
Liebchen C.: The First Optimized Railway Timetable in Practice. Transportation Science, 42:
420 – 435, 2008.
Liebchen C., Lübbecke M., Möhring R. M., Stiller S.: The Concept of Recoverable
Robustness, Linear Programming Recovery, and Railway Applications. Volume 5868 of
Lecture Notes in Computer Science, pp. 1 – 27. Springer Berlin/Heidelberg, 2009.
Liebchen C., Schachtebeck M., Schöbel A., Stiller S., Prigge A.: Computing Delay Resistant
Railway Timetables. Computers & OR, 37: 857 – 868, 2010.
Liebchen C., Stiller S.: Delay Resistant Timetabling. Public Transport, 1: 55 – 72, 2009.
Medeossi G., Hürlimann D., Longo G.: Stochastic Mirco-Simulation as a Timetable
Robustness Estimation Tool, Proceedings of the 3rd International Seminar on Railway
Research (IAROR), eds. Hansen I. A., Wendler E., Weidmann U., Rodriguez J., Ricci S., Kroon
L.G., Zürich, Switzerland, 2009.
Nachtigall K.: Periodic Network Optimization and Fixed Interval Timetables. Habilitation Thesis,
Deutsches Zentrum für Luft- und Raumfahrt, Braunschweig, 1999.
Nachtigall K., Voget S.: A Genetic Algorithm Approach to Periodic Railway Synchronisation.
Computers and Operations Research, 23: 453 – 463, 1996.
Odijk M.: A Constraint Generation Algorithm for the Construction of Periodic Railway
Timetables. Transportation Research, 30: 455 – 464, 1996.
Peeters L. W. P.: Cyclic Railway Timetable Optimisation. PhD thesis, Erasmus University
Rotterdam, Rotterdam School of Management, The Netherlands, 2003.
Sastry S., Bodson M.: Adaptive Control: Stability, Convergence, and Robustness. Prentice
Hall, 1989.
Schrijver A., Steenbeek A.: Dienstregelingontwikkeling voor Railned (Timetable construction for
Railned). Technical report, Center for Mathematics and Computer Science, Amsterdam, 1994.
In Dutch.
Serafini P., Ukovich W. A.: Mathematical Model for Periodic Event Scheduling Problems. SIAM
Journal of Discrete Mathematics, 2: 550 – 581, 1989.
U.I.C. Timetable Recovery Margins to Guarantee Timekeeping – Recovery Margins. Leaflet
451-1 of U.I.C. Paris, France, 2000.
U.I.C. Capacity. Leaflet 406 of U.I.C. Paris, France, 2004.
Vansteenwegen P., Van Oudheusden D.: Developing Railway Timetables which Guarantee a
Better Service. European Journal of Operational Research, 173: 337–350, 2006.
Vansteenwegen P., Van Oudheusden D.: Decreasing the Passenger Waiting Time for an
Intercity Rail Network. Transportation Research B, 41: 478 – 492, 2007.

172
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

Vromans M. J. C. M.: Reliability of Railway Systems. PhD thesis, Erasmus University

Rotterdam, Rotterdam School of Management, The Netherlands, 2005.
Wendler E.: The Scheduled Waiting Time on Railway Lines. Transportation Research B, 41:
148 – 158, 2007.
Yuan J., Hansen I.: Optimizing Capacity Utilization of Stations by Estimating Knock-on Train
Delays. Transportation Research B, 41:2:202 – 217, 2007.

173
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 9.2 The bi-level approach

9 Bi-level scheduling in highly utilised networks

Gabrio Caimi

9.1 Introduction

The last decades have brought substantial improvements in the field of automatic train
scheduling. For instance, Dutch Railways introduced a new timetable generated from scratch
with the massive use of operations research methods in all the steps of the planning process,
see Kroon et al. (2009). Close collaboration between researcher and the railway companies
was instrumental in reaching this goal, since only the combination of practical know-how with
formal modelling skills allows sophisticated algorithmic solution approaches to enter real-world
applications.
In this Chapter the Author presents a further step towards automatic train scheduling, which
was undertaken by collaboration between ETH Zurich and Swiss Federal Railways (SBB). SBB
operate the densest railway network in Europe, with about 3,000 track kilometres and 9,000
annual train movements. A total of 300 million passengers and 60 million tonnes of freight are
transported each year. The core of the network is already at the limits of its capacity, and traffic
is expected to grow even further. Construction of new tracks, however, is very expensive,
difficult to realise for many reasons and should be avoided as much as possible. Therefore,
new planning methods and operational processes are necessary to improve the performance
of the existing railway network. In this context the Author presents a comprehensive approach
for timetable planning, from the description of the commercial offer as a starting point, to the
conflict-free detailed train schedule (the production plan) as the outcome. In particular, there
is a focus on the scheduling problem in station areas, which are particularly complex but also
crucial for generating a detailed conflict-free solution.
The approach presented is particularly suited for generating conflict-free schedules in
highly utilised networks. In this case, the goal of having a detailed plan for each train, which
specifies at which time exactly which track is used, is of crucial importance to enable a stable
operation. Of course, this plan has to be conflict-free, i.e. the detailed schedule does not
generate any track occupation conflicts, given the safety and signalling systems that are in
force. For this purpose, a multi-level approach is used. The starting point is the formalisation
of the commercial services that the railway companies want to offer to the customers over a
whole day, including periodicity information and exceptions. From this input, different methods
and algorithms are used to reach the final goal of a network-wide, detailed conflict-free train
schedule for a whole day. Partial use is made of well-known methods from the literature,
combined with methods developed especially during this project and described in detail
elsewhere.
The multi-level approach is crucial in order to integrate the functions and properties of the
different levels, as discussed in detail in Section 9.2. For a more detailed description of the
different models and algorithms, see Caimi (2009) and Caimi et al. (2011-ii).

9.2 The bi-level approach

For sufficiently large railway networks, the amount of data makes it impossible to incorporate
all details of the railway topology in a single planning step. For networks with particularly
dense traffic and complex topologies, this can be already the case for regions with a radius of
25 – 30 km. A two-level approach is followed therefore for the train scheduling problem based

175
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
9 Bi-level scheduling in highly utilised networks

on two different description levels of the infrastructure. The upper macroscopic level considers
a simplified model of the topology for the entire network with the most relevant information,
while the lower microscopic level uses a detailed model of the topology, but it is restricted to a
locally bounded region. The goal is the generation of a detailed train schedule that is checked
to be conflict-free on the blocking time level for all parts of the network.

Fig. 9.1: Two-level approach for the train scheduling. First, macroscopic scheduling is applied
using a simplified topology model, then microscopic scheduling is applied locally for considering
all the details.

Train scheduling on the macroscopic level (macro timetabling) focuses on global

interdependencies over the entire network. These are for generating the most important
properties of the timetable (departure and arrival times in major stations, train sequences and
crossing points on open tracks) and thus has to avoid dealing with large amounts of detailed
information that is only relevant locally. Whether or not a given macro timetable yields a
corresponding feasible micro schedule is in general unclear and this can only be decided in a
subsequent planning step. This is particularly relevant for highly utilised networks that are close
to the capacity limits.
The existence of an operable production plan for a given macro timetable has to be checked
at the micro level by taking into account detailed information, in particular the exact track
paths and the corresponding blocking times for ensuring that the schedule is conflict-free.
Therefore, train scheduling on the micro level (micro scheduling) considers additionally the
details of the railway topology that are important for describing local conflicts but were not
relevant for the global structure of the timetable and therefore neglected on the macro level.
The safety system considered on the micro level follows the way the interlocking system
works. It computes the blocking times on each topology element based on the signalling
systems and ensures that these blocking time intervals do not overlap. Therefore, the topology
model on the micro level must include all switches and signal positions that are necessary for

176
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 9.3 Macro scheduling

the computation of blocking times on each topology element. Moreover the computed train
paths must specify a speed profile that the train driver can follow, given a reasonable tolerance
band.
Micro scheduling cannot be done for the whole network simultaneously, because the huge
amount of data would lead to both memory problems and prohibitively long computation
times. To overcome this problem the network is divided into different zones. For a suitable
subdivision, the network’s typically very heterogeneous structure is exploited, to distinguish
two types of zones:
– Condensation zones are the bottlenecks of the network, where the available unused
capacity is scarce. Therefore, the network needs to be used at the maximum possible
capacity. The scheduling policy for these zones is such that no time reserves are
introduced, and the specified track paths must be followed very precisely. Usually, but not
necessarily, condensation zones are situated around major stations, and cover a relatively
small area (radius up to 15 km). There are many switches and thus overtaking and train
crossings are both possible and quite common. An appropriate train routing is crucial in
order to solve the micro scheduling problem in this zone.
– Compensation zones connect the condensation zones and have typically more capacity
available as well as a much simpler topology (some parallel tracks with few crossing and
switches). Time reserves are introduced to improve timetable stability. Train overtaking
and crossings are seldom needed here, as the train sequencing is basically decided within
the condensation zones. Therefore, the choice of appropriate speed profiles is the most
important degree of freedom to be exploited.
To be able to schedule each zone independently and then to merge their micro schedules
seamlessly to a feasible global schedule, it is important to define and coordinate clearly the
interface between the zones. For this, the concept of a portal is introduced, which represents
the border between condensation and compensation zones. Once the boundary conditions
(passing time and speed) are fixed for each train and portal, the zones can indeed be treated
independently, with the assurance that the locally generated schedules are also globally
feasible by simply merging them.
Finding portal passing times (or time slots) is a coordination task and is therefore part of macro
scheduling. If the microscopic scheduling problem turns out not to be feasible in at least
one of the zones considered, it is clear that the timetable generated on the macro level was
also not feasible. But this does not imply that the given service intention, which was used as
the input for macro scheduling, is not feasible as well. It might well be possible that another
draft timetable, suboptimal from the macroscopic point of view, is feasible on the micro level,
implying that the service intention is actually feasible. It is therefore necessary in case of the
lack of feasibility on the micro level, that the method does not just give a negative answer but
also a feedback report that tells the macro scheduler how to adjust the timetable in the right
way. The logic of the entire multi-level scheduling approach, including the feedback between
the levels, is depicted in Fig. 9.1.

9.3 Macro scheduling

Macro scheduling accounts basically for the dependencies in timetabling that have network-
wide impact. It produces timetables for large networks by leaving aside detailed information
through simplifying the topology, safety system, and train dynamics. On the macro level, the
topology model only considers major stations and the number of parallel tracks connecting
them, the train dynamics only considers the travel times between them, and the safety system

177
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
9 Bi-level scheduling in highly utilised networks

is simplified by enforcing a minimum headway between two trains on the same track. The
variables are the departure and arrival times of the trains. They are chosen such that the
travel times and connections are fulfilled as specified in the service intention. Furthermore, the
timetable is assumed to be periodic.
The simplified macro topology is represented by a graph with nodes for relevant stations and
important check-points such as where tracks branch or cross each other. The nodes are
connected by edges if there is a direct track in between. Tracks may be used by trains in
one direction only, or both if the signalling so permits. The routing of the trains through the
macro topology graph is assumed to be fixed in advance. In particular, for sections with
multiple tracks, it is also required that the track choice at the portal is predetermined. This
is reasonable, because multiple parallel tracks are often separated according to simple and
widely used policies. The situation is different in station areas, where many route choices are
possible. This freedom will be used during the consideration of micro scheduling in Section 9.4
to avoid conflicts in the station areas.

9.3.1 The Periodic Event Scheduling Problem

A macro timetable consists of a set of event times for departure, arrival events at stopping
stations and for passage times at passing stations in the macro topology. The event time is
denoted by i for an event i . As the services are periodic with periodicity T, each event time
variable actually represents a set of event times {..., i - T, i, i + T, i + 2T, …}.
The macro timetabling problem can be formulated as a Periodic Event Scheduling Problem
(PESP). It was introduced first by Serafini & Ukovich (1989) and first applied to railway
timetabling by Schrijver & Steenbeek (1994). For a comprehensive review of the PESP model
and for solutions using mixed-integer programming, see Liebchen (2006).

9.3.2 Flexible PESP

The Flexible Periodic Event Scheduling Problem (FPESP) is a generalisation of the PESP that is
well suited for the multi-level scheme. It takes into account that the solution of the macro level
still is to be refined and checked at the micro level. The key idea is to produce a flexible output
from the macro level that yields a larger solution space on the micro level.
A larger solution space has many advantages. First, the risk that the generated macro
timetable is not feasible reduces, even much less if the flexibility is sufficiently large and well
distributed. Furthermore, it will be possible to find micro schedules of better quality compared
with the case with a fixed macro timetable. Finally, due to the greater correction potential, the
flexibility gives the possibility of considering a less detailed macro topology than when working
with the original PESP. For instance, the macro topology used by the module CADANS in the
Netherlands is very detailed, even in the vicinity of main stations, see Schrijver & Steenbeek
(1994). Using the FPESP, some of these details can be omitted on the macro level, which
in turn permits the consideration of even larger networks. Another possible benefit of added
flexibility for the purpose of real-time traffic management has been pointed out by D’Ariano,
Pacciarelli & Pranzo (2008).
Instead of determining exact event times i , the idea of a flexible macro timetable is to
generate time slots (i, i + i), for the events, where i is the flexibility range. Such flexible time
slots are shown in Fig. 9.2.

178
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 9.3 Macro scheduling

Fig. 9.2: (a) The generated timetable without event slots from the 2007 SBB service intention,
visualised in a time-space diagram. The horizontal axis represents the route from Arth-Goldau (GD)
towards the alps (Erstfeld, ER), whereas the vertical axis represents the time. (b) When using event
slots, each event gets an event time i and a flexibility i. In this diagram, the earliest possible line
and the latest possible line are filled in grey. Any choice for train trajectories in the grey zones are
feasible from the macro scheduling point of view.

The resulting macro schedule with the generated departure and arrival times fulfils the trip time
and connection requirements on a network-wide scale. It serves as an input to the detailed
micro scheduling. The presented FPESP extension makes sure that the data transmitted
between macro and micro level is not more restrictive than necessary, leading to a larger
solution space for the latter. The solution provided has the crucial property that each possible
choice of event times inside the generated time slot is feasible towards the given constraints.
This makes possible the separation of the problem in the subsequent micro scheduling, as
each choice of a concrete event time can be taken separately by keeping the constraints on
the macro level feasible.

179
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
9 Bi-level scheduling in highly utilised networks

9.4 Micro scheduling in condensation zones

In condensation zones the choice of an appropriate routing is the crucial degree of freedom
to exploit for finding a conflict-free train schedule, see Fig. 9.3 as an example. As no time
reserves are introduced on train trips in this zone, the complete dynamic of a train is given
exclusively by the choice of its route.
For this task the conflict graph model (CG) has become, among others, a commonly used
approach in the literature, see for instance Herrmann (2005). In this intuitive model, all potential
routes for each train are first computed and then represented as nodes in the conflict graph. If
two nodes block the same topology resource at some point in time, they are called conflicting
and are connected with an edge. In addition, all nodes representing the same train will be
connected to each other to form a clique. A conflict-free schedule in this model corresponds
to choosing exactly one node per train, forming a maximum stable set in this conflict graph.
Unfortunately, the LP-relaxation of this problem is very weak, such that large instances of this
model can only be solved heuristically as in Herrmann (2005) and Zwaneveld et al. (1996), or
with time-consuming pre-processing, see Zwaneveld et al. (2001).

Fig. 9.3: Micro topology of the core of the condensation zone Lucerne with its terminal station.
Source: Lüthi, Medeossi & Nash (2008)

It is possible to characterise the set of conflict-free routes by a different model in a much

more concise way. Instead of a simple binary relation between the nodes representing the
alternative train paths, each possible train path can be considered as a flow through the graph
of the different block sections (the resources). This can be represented as a tree structure (the
Resource Tree), where each branch in the Resource Tree corresponds to a switch where both
diverging directions lead to the designated target for the train (the platform or the exit portal),
as shown in Fig. 9.4.
The time slots that were generated at the macro level can now be used at the micro level to
generate a (discrete) set of alternative arrival and departure times inside the given time interval.
It is therefore necessary not only to assign a route to each train, but also to choose one of
these alternative times for each arriving and departing train. For each of these alternatives,
the complete Resource Tree of the possible routing is available and is always the same, yet
the passing times and, consequently, the blocking time intervals are different (shifted in time
exactly by the time difference of the input time alternatives).

180
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 9.4 Micro scheduling in condensation zones

Fig. 9.4: (a) An example track topology with two trains running on it: train 1 drives from
infrastructure resource 0 to 10 and train 2 drives from resource 1 to 9. (b) Resource Trees represent
the routes of the two trains given in Fig. 9.4.

Afterwards, each edge in the Resource Tree (i. e. each topology element used) is directly
associated with the corresponding resource in the railway topology indicating its blocking
time interval. It will be recalled that these blocking times are computed precisely by taking
into account the detailed topology and signal information. Given all these blocking times, it
is possible to compute the so-called conflict cliques, which limit the concurrent allocations
of a resource to at most one train, separately for each resource. In Velasquez et al. (2005)
and Lusby (2008), a similar idea was proposed, using a discretisation of the time axis and
forcing at most one allocation per time unit. This approach is not satisfactory, because on the
one hand a fine discretisation is too time consuming and will produce mostly redundant or
unnecessary constraints, whereas a coarse time discretisation could miss important conflicts
(while still generating mostly redundant constraints).
Here, a much more efficient detection of potential conflicts is used, which is fast, exact, and
does not generate redundancy. In fact, this detection algorithm probably finds all maximum
conflict groups (which are cliques in the corresponding intersection graph) for each resource in
non-periodic scheduling. Assuming that each blocking span is at most one third of the period
length (which is essentially always true in reality), it will find them also in the periodic case.
This conflict detection scheme is based on the idea that it is only necessary to check the end
times of the allocation time intervals, as in Caimi et al. (2011-i). The result for one example is
illustrated in Fig. 9.5.

181
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
9 Bi-level scheduling in highly utilised networks

Fig. 9.5: Several track paths allocate a resource during different time intervals (horizontal bars). The
dashed vertical bars illustrate conflict cliques among the allocation intervals formed by the conflict
grouping algorithm.

Finally, the set of alternative scheduling possibilities for each train and the detected conflict
groups can be merged to a multi-commodity flow formulation. Here, for each train, a path
through its Resource Tree is sought that meets the conflict constraints. This is called the
Resource Tree Conflict Graph (RTCG) model. The idea is represented schematically in Fig. 9.6.
On the left are the Resource Trees of two trains, augmented with a source s and a sink t for
each train. On the right, each resource of the infrastructure is shown and connected with the
edges of the Resource Tree that belong to a path through this resource.
Additional to the RTCG model, the Tree Conflict Graph (TCG) has also been implemented. This
is a hybrid model between CG and RTCG, which represents the possible paths of a train in
Resource Trees as in the RTCG model, but models conflict pairs as in the CG model instead of
using conflict cliques. This way, it is possible to understand better the added value given to the
RTCG model by the tree structure and by the conflict cliques.
Even though the main focus of micro scheduling is on feasibility, an objective function
related to the potential lack of feasibility of the micro problem can be defined. In case there
is no feasible solution within the bound given by the macro timetable, it is important to give
a feedback to the macro level to indicate where more flexibility might be necessary. This is
achieved in the RTCG model by providing further time alternatives to those trains that could
not be scheduled within their original time slots given by the macro timetable. These additional
alternatives are then penalised in the objective function. An optimal solution will then try to
assign all times inside the input time slots if there exist any conflict-free assignment; otherwise
it will deliver a solution with minimum deviation from these given time intervals.
The RTCG can be formulated as an integer linear program (ILP). Given the set of route weights
wzl for each route l of a train z, the set E of all edges in the Resource Tree, the set Sz of edges
connected to the source node of train z, the set Tz of edges connected to the sink node of
train z, the family C r of sets corresponding to conflict cliques of the infrastructure resource r
and the number of trains |Z|, the ILP can be written as shown in Fig. 9.7.
The RTCG problem can be solved with a commercial solver. The improvement in structure
achieved with the presented RTCG model is so strong that in many cases the LP-relaxation
was already integer, and realistic, large-scale instances could be solved in few seconds (with
up to many millions of variables). Some computational results are presented in Section 9.6,
and in Caimi et al. (2011-i) a detailed discussion based on exhaustive tests is given.

182
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 9.4 Micro scheduling in condensation zones

Fig. 9.6: Example of a Resource Tree Conflict Graph (RTCG)

Fig. 9.7: Integer linear problem formulation of the Resource Tree Conflict Graph (RTCG)

183
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
9 Bi-level scheduling in highly utilised networks

9.5 Micro scheduling in compensation zones

After having scheduled the condensation zones, the exact (portal) times inside the given time
interval from the macro level have been fixed definitively. The last remaining step is microscopic
scheduling in the compensation zones. In these zones, the important degree of freedom to
exploit is the speed profile. The idea is that for each train a set of alternative speed profiles
connecting the given times and speeds at the portals are first generated. One of these speed
profiles is then selected for each train, such that they are mutually non-conflicting and optimal
according to a given objective function, see Caimi et al. (2009). This might be the minimisation
of energy consumption or optimising the time reserve for distribution along the line.
With a proper choice of the parameters, a good, but manageable variety of speed profiles can
be achieved. Each generated speed profile is then assigned a quality value to measure criteria
like energy consumption and time reserve distribution.
A tree structure is built depicting the given speed profiles, grouping them together as long as
they are equal, and introducing a branch in the tree only where they start to differ. Even for this
way of creating the tree of different scheduling alternatives, it is possible to apply the same
RTCG modelling as for condensation zones, and to solve the resulting MIP formulation with a
commercial solver. Computational results show that the MIP formulation is almost as efficient
as in the condensation zone, and that the bottleneck for the CPU time as well as solution
quality lies in the generation of the speed profiles.
Microscopic scheduling in compensation zones is usually not critical for feasibility, as the zones
are less dense and the simplified safety model at the macro level is often good enough to
yield feasibility. This is in contrast to the condensation zones. Nevertheless, this step remains
very important for computing accurate conflict-free trajectories for the train drivers to follow.
Otherwise, they might be forced to stop at a red signal, thus generating delay even without
disruptions.

9.6 Computational results

The procedure described is illustrated using real data provided by SBB, in a test case for
the region between the cities of Lucerne, Zug, and Arth-Goldau, in central Switzerland. This
area has a radius of about 25 km. Although the region is not geographically large, it remains
challenging because it contains many changes between double and single track. Furthermore,
the traffic is very dense and consists of a mixture of long-distance passenger, local passenger,
and freight trains. Besides testing the entire approach on the Lucerne area, the RTCG model
for condensation zone has also been tested for the main station area of Berne. This is larger
than Lucerne and is therefore more challenging for validating the model. All mixed integer
programs (MIPs) are solved with the commercial solver CPLEX.
The start was by constructing a partial periodic service intention, corresponding to the
operated timetable for a typical Wednesday of 2008. It contains 33 train runs, yielding a total
of 196 train services for a whole day. This service intention is first projected to a reference time
period of one hour, generating 48 equivalence classes. The subsequent FPESP instance at the
macro level illustrated in Fig. 9.8 has 212 events and 647 arcs, after resolving arcs with zero
span. This instance results in a mixed integer programming formulation with 1,083 variables
(436 integer) and 1,730 constraints.
Depending on the chosen objective function, the amount of generated flexibility as well as the
CPU time can vary considerably. Table 9.1 presents the solution values for the total flexibility
as well as the total passenger trip times for several combinations of these objective functions.

184
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 9.6 Computational results

a)

b)

Fig. 9.8: (a) The considered macro topology of the test scenario in central Switzerland. It is partly
double track and partly single track and is used by regional and intercity trains as well as freight
trains. (b) A less precise macro topology of the network in central Switzerland. This version can be
used as an alternative to the one illustrated in Fig. 9.8a.

185
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
9 Bi-level scheduling in highly utilised networks

The first column lists the different objective functions tested, whereas in the second column
the necessary time to find the optimal solution according to this objective can be found. In the
third column, i stands for the total flexibility of the optimal solution found and in the fourth
column  xt,d,c stands for the sum of all trip, dwell and connection times. All computations
were repeated until an optimal solution was found.
Noflex is the objective function that minimises the total trip time of the solution, without
considering its flexibility. On the contrary, Maxflex only maximises the total flexibility, with the
effect that the resulting trip times are longer. These two objectives generate the best values
for the objectives considered individually. Mixflex  is an objective function that considers both
flexibility (with weight  (0,1)) and trip time (with weight 1−). The solution is guaranteed
Pareto-optimal and gives a compromise solution compared to the two previous ones. Finally,
Postopt is an approach that first solves Noflex and only in a second step introduces and
maximises the total flexibility by keeping all integer variables fixed (i. e. all train sequences
remains unchanged). This method is very fast, though its solution is not guaranteed to be
Pareto-optimal. The reported CPU time is without the time needed for Noflex.
The last line (Topology 2) displays the results of the FPESP by considering a less precise
macro topology (illustrated in Fig. 9.8), using the objective function Mixflex 1/2. Finally, it can
be observed that the consideration of a less precise topology (Fig. 9.8b) on the macro level
performs worse than the use of the more precise topology (Fig. 9.8a) as in all other cases. On
the other hand, computations on this less precise topology are clearly faster.

Objective CPU time [s] i  xt,d,c

NOFLEX 14 – 1914
MAXFLEX 336 330 2241
MIXFLEX ½ 57 189 1971
MIXFLEX 2/3 133 326 2124
MIXFLEX 9/10 157 329 2144
POSTOPT 0.3 216 2006
MIXFLEX ½ (Topology 2) 7 185 2008

Table 9.1: Results for the reference scenario with bounds on the flexibilities i  4

The resulting macro timetable is then fed as input to micro scheduling. Condensation zones
are solved first. The most interesting is the one around the main station of Lucerne, which is a
terminal station with 12 platforms. The condensation zone has a radius of about six kilometres,
consisting of roughly 40 switches, while serving five directions with 30 standard gauge trains
per hour. (There are separate metre gauge services).
The timetable obtained from the macro level gives rise to 1,217 different routes in the
condensation zone Lucerne, which results in an RTCG with 44,327 nodes, 43,487 blocking
edges, and 43,517 flow edges. The resulting MIP has 43,517 binary variables, but only 1,999
conflict constraints. After 10 seconds of pre-processing, the MIP solver was able to deliver a
relaxed solution that was directly integer.
Table 9.2 gives an overview of the behaviour of the RTCG model for different instances in
the condensation zones of Lucerne and Berne, comparing it with the standard conflict graph
approach (CG). The input used for the Lucerne scenario is the macro timetable with objective
function Mixflex 9/10, for Bern 2008 the operating one and for Bern East a manually generated
macro timetable. As an additional comparison, results for the TCG model are also presented.

186
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 9.6 Computational results

The first two columns show the scenario and the model that was tested. The two blocks
afterwards show the results before and after the standard pre-processing of the MIP-solver.
For both, the first column gives the number of variables of the considered MIP, the second
shows the number of constraints and the third indicates the gap between the value of the
relaxed solution of the corresponding MIP and the value of the (mixed-integer) optimal solution.
The strong structure using the RTCG model, which is always very close the optimal solution,
can be clearly noted. By contrast, the other models present quite large gaps. The last column
presents the total computation time for solving the problem using the pre-processing.

Before preprocessing After preprocessing

Solution
Scenario Model Binary Conflict Integrality Binary Conflict Integrality time [s]
variables constraints gap [%] variables constraints gap [%]
RTCG 23,477 1,999 0 119 79 0 <<1
Lucerne TCG 23,477 17,655 26.4 480 465 0 4
CG 661 27,897 47.8 14 12 0 1
RTCG 319,142 4,401 0 6,899 3,461 0 11
Berne East TCG 319,142 4,808,270 27.2 10,994 351,444 7.2 10,022
CG 6,820 6,822,627 48.4 6,245 3,808,221 7.3 6,426
RTCG 535,264 9,773 3.1 25,471 11,946 1.5 26
Berne 2008 TCG 535,264 38,983,513 27.8 30,544 2,155,718 8.44 19,300
CG 18,191 10,568,742 31.9 15,922 4,011,140 8.5 6,099

Table 9.2: Statistics of the three different models for the micro scheduling in condensation zones,
before and after pre-processing of the MIP-solver, for three tested scenarios. All computations
were continued until the optimal solution was found.

It is particularly interesting to observe that the use of a more precise topology at the macro
level results in better performance. It clearly offers more feasible possibilities than the use of
the less precise topology (Fig. 8a and 8b). This, not only at the macro level itself, but also
in the subsequent micro level, confirms the theoretical considerations of Section 9.2. In
particular, the neglected single track line between Lucerne (switch region FMUE in Fig. 9.3)
and Ebikon (EBI) in the condensation zone of Lucerne is very difficult to allocate if this structure
is not already considered in the macro topology. Indeed, in the macro timetable resulting from
the simpler Topology 2, the departure time of a train in Lucerne and the arrival time of another
train also travelling through the single track are set in a way that necessitates a crossing in the
middle of this single track. This is of course impossible. Even small values of flexibility were
not sufficient to adjust these times in order to make it feasible. Only a very large amount of
flexibility to the involved events could help to resolve the conflict. On the other hand, when
considering this single track previously at the macro level (Topology 1), a small amount of
flexibility was sufficient to find a feasible solution at the micro level.
Another example from the test case in Lucerne shows the risk posed by a too detailed macro
topology. If a node corresponding to the region of Gütsch (GTS, see Fig. 9.3) is introduced and
connected by double track to the station of Lucerne, it automatically means that each track is
used exclusively for traffic in one direction, preventing the result from being feasible. Indeed, in
order to fulfil all the connectional requirements of the service that are intended, it is sometimes
necessary that two trains enter or leave Lucerne in parallel, using the double track. This is
not allowed if these two tracks are modelled as double track on the macro level. It becomes

187
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
9 Bi-level scheduling in highly utilised networks

possible, however, if this double track is neglected at the macro level. At the micro level, many
routes that use both tracks will be considered at the micro level, which makes the parallel entry
(or exit) of two trains possible.
Finally, the compensation zones are solved. The longest is the line between Lucerne and Zug,
with a total length of about 25 km including two single track sections of roughly 5 km and
3 km, respectively. Twelve trains per hour travel on this line. For these 12 trains, a total of 873
speed profiles have been generated, which took about five seconds of computation time.
The construction of the RTCG model required 0.3 seconds, and the MIP-solver could find the
optimal solution for all three tested objective functions (energy consumption, time reserves
distribution or a mixed version) almost instantaneously. This was because the relaxed solution
was directly integer.

9.7 Conclusion and outlook

The Author has presented a bi-level approach for conflict-free timetabling in a railway network,
to be used particularly for dense services with heterogeneous periodicities. The presented
hierarchical procedure and the developed models for the macro and the micro scheduling
allow the planning of a system-wide train schedule, from the description of the commercial
offer to the detailed production plan. The approach allows the generation of dense timetables,
fulfilling large sets of requirements including trip times, connections, detailed safety constraints,
and station capacity. To reach this goal, known models and algorithms for each level have
been adapted and applied. These are based on state-of-the-art from the literature and from his
own previous and current work.
To increase the chances that a generated macro schedule can be extended into a conflict-free
micro schedule, an extension of the well-known PESP model has been used, with the ability to
generate flexible time slots in lieu of exact event times. These time slots guarantee that each
choice of the time inside the slot, which is now deferred to the micro level, is feasible from the
macroscopic point of view.
The network separation at the micro level appears to be a natural way to cope with the huge
amount of data. The division into zone types, with their own distinct properties, enables a
focus on different degrees of freedom. These are the routing in condensation zones and the
speed profiles in compensation zones. In both cases, the developed RTCG model is able
to describe conflicts much more efficiently than previous approaches. Thanks to the strong
structure of this model, the corresponding integer program can be solved in few seconds,
even for large realistic instances.

Acknowledgments

The research at the basis of this Chapter has been supported by the Swiss Federal Railways
(SBB), Infrastructure Division. The Author wishes to thank Martin Fuchsberger, Marco
Laumanns, and Kaspar Schüpbach for their collaboration in its production.

188
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

References
Caimi G. (2009): Algorithmic decision support for train scheduling in a large and highly utilised
railway network. PhD thesis, ETH Zurich.
Caimi G., Chudak F., Fuchsberger M., Laumanns M., Zenklusen R. (2011-i): A New Resource
– Constrained Multicommodity Flow Model for Conflict-Free Train Routing and Scheduling.
Transportation Science 45(2), pp. 212 – 227.
Caimi G., Fuchsberger M., Burkolter D., Herrmann T., Wüst R., Roos S. (2009): Conflict-free
train scheduling in a compensation zone exploiting the speed profile. In: Proceedings of the
3rd International Seminar on Railway Operations Modelling and Analysis (RailZurich 2009),
Zurich, Switzerland.
Caimi G., Fuchsberger M., Laumanns M., Schüpbach K. (2011-ii): A multi-level framework for
generating train schedules in highly utilised networks. Public Transport 3(1), pp. 3 – 24.
Caimi G., Fuchsberger M., Laumanns M., Schüpbach K. (2011-iii): Periodic railway timetabling
with event flexibility. Networks 57(1), 3–18 (2011). 10.1002/net.20379.
Caimi G., Laumanns M., Schüpbach K., Wörner S., Fuchsberger M. (2011-iv): The periodic
service intention as a conceptual frame for generating timetables with partial periodicity.
Transportation Planning and Technology 34(4), pp. 305 – 321.
Corman F., D’Ariano A., Pacciarelli D., Pranzo M. (2009): Evaluation of green wave policy in
real-time railway traffic management. Transportation Research, Part C 17(6), pp. 607–616.
Corman F., D’Ariano A., Pacciarelli D., Pranzo M. (2010-i): Centralized versus distributed
systems to reschedule trains in two dispatching areas. Public Transport 2(3), pp. 219 – 247.
Corman F., D’Ariano A., Pacciarelli D., Pranzo M. (2010-ii): A tabu search algorithm for
rerouting trains during rail operations. Transportation Research, Part B 44(1), pp. 175 – 192.
D’Ariano A., Pacciarelli D., Pranzo M. (2008): Assessment of flexible timetables in real-
time traffic management of a railway bottleneck. Transportation Research, Part C, 16(2),
pp. 232 – 245.
Herrmann T. M. (2005): Stability of Timetables and Train Routings through Station Regions.
PhD thesis, ETH Zurich.
Kepaptsoglou K., Karlaftis M. G. (2009): The bus bridging problem in metro operations:
conceptual framework, models and algorithms. Public Transport 1(4), pp. 275 – 297.
Kroon L. G., Huisman D., Abbink E., Fioole P. J., Fischetti M., Maroti G., Schrijver A.,
Steenbeek A., Ybema R. (2009): The New Dutch Timetable The OR Revolution. INTERFACES
39(1), 6–17. 10.1287/inte.1080.0409. URL ifundefinedselectfonthttp://interfaces.journal.
informs.org/cgi/content/abstract/39/1/6.
Liebchen C. (2006): Periodic Timetable Optimisation in Public Transport. PhD thesis,
Technische Universität Berlin.
Lusby R. M. (2008): Optimisation methods for routing trains through railway junctions. PhD
thesis, University of Auckland.
Lüthi M., Medeossi G., Nash C. A. (2008): Structure and simulation evaluation of an integrated
real-time rescheduling system for railway networks. Networks and Spatial Economics 9(1), pp.
103–121. 10.1007/s11067-008-9085-4.
Michaelis M., Schöbel A. (2009): Integrating line planning, timetabling, and vehicle scheduling:
a customer-oriented heuristic. Public Transport 1(3), pp. 211 – 232.

189
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
9 Bi-level scheduling in highly utilised networks

Peeters L., Kroon L. G. (2001): A cycle based optimization model for the cyclic railway
timetabling problem. In: S Voß, J Daduna (eds.) Proceedings Computer-Aided Scheduling of
Public Transport (CASPT 2000), vol. 505, pp. 275–296. Springer, Berlin.
Rodriguez J. (2007): A constraint programming model for real-time train scheduling at
junctions. Transportation Research, Part B 41(2), pp. 231 – 245.
Schrijver A., Steenbeek A. (1994): Dienstregelingontwikkeling voor Railned (timetable
construction for Railned, in Dutch). Tech. rep., C.W.I. Centre for Mathematics and Computer
Science, Amsterdam.
Serafini P., Ukovich W. (1989): A mathematical model for periodic scheduling problems. SIAM
J. Disc. Math. 2(4), pp. 550 – 581.
Törnquist J., Persson J. A. (2007): N-tracked railway traffic re-scheduling during disturbances.
Transportation Research, Part B 41(3), pp. 342 – 362.
Velasquez R., Ehrgott M., Ryan D., Schöbel A. (2005): A set-packing approach to routing
trains through railway stations. In: Proceedings of the 40th Annual Conference of the
Operations Research Society of New Zealand, pp. 305 – 314.
Zhang Y., Lam W. H. K., Sumalee A., Lo H. K., Tong C. O. (2010): The multi-class schedule-
based transit assignment model under network uncertainties. Public Transport 2(1-2),
pp. 69 – 86.
Zwaneveld P. J., Kroon L. G., Romeijn H. E., Salomon M., Dauzère-Pérès S., van Hoesel S. P.
M., Ambergen H. W. (1996): Routing Trains through Railway Stations: Model Formulation and
Algorithms. Transportation Science 30(3), pp. 181 – 194.
Zwaneveld P. J., Kroon L. G., Van Hoesel S. P. M. (2001): Routing trains through a railway
station based on a node packing model. European Journal of Operational Research 128(1),
pp. 14 – 33.

190
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 10.1 Introduction – the role of simulation

10 Simulation
Robert Watson, Giorgio Medeossi

10.1 Introduction – the role of simulation

Simulation has been defined by Robinson (1994) as “a model that mimics reality” and by
Gamerman & Lopez (2006) as “treatment of a real problem through reproduction in an
environment controlled by the experimenter”. Robinson notes that simulation can be more
appropriate than mathematical models where there are many variables and that mathematical
models are not good at assessing “knock on” impacts. For example, in railway operations,
“secondary delay” is caused by one train delaying another. In some circumstances, simulation
is undertaken by making a physical model; attempting to simulate the behaviour of railway
operations is however best undertaken using a computer.
This computer simulation is undertaken by firstly creating a detailed scale map or plan of the
infrastructure. This includes all important “features” (most obviously stations and junctions,
but also for many of the models much more detail, including for instance gradients and the
location of individual signals and track circuits). The model then needs to know details about
the physical characteristics of the trains that will be operated – for instance how fast they can
go, acceleration and braking rates. The intended operational plan – the timetable – then must
be provided, as the basis of the simulation.
Set the simulation running and if the model has been built correctly and the timetable is well
constructed, it will be possible to watch the trains run across the computer screen (either in
“real time” or “fast forward”), in the same way as it is possible to watch trains running on the
screen in a traffic control centre. Usually the computer model takes the decisions that the
signaller takes in real life, although with some models the modeller can make decisions as the
simulation runs. If the timetable has inaccuracies (e. g. conflicts between trains or unachievable
running times), then it will be possible to see these on-screen as the simulation runs or in the
model outputs. The usual final stage of the simulation process is to add delays to the trains
running in the model to properly “mimic reality” and to see what level of punctuality will be
achieved when problems occur (e. g. points failure, station dwell time exceeded).
The appropriateness of using simulation in a business situation has been described by Jones
(1972) as follows: “a company may feel that before committing itself to a plan of action it
would be useful to see what the effect of alternative actions would be – a sort of look before
you leap”. Robinson sets out the potential benefits as being:
– Risk reduction
– Greater understanding
– Operating cost reduction
– Lead time reduction
– Faster planned changes
– Capital cost reduction
– Improved customer service
Simulation of railway timetables is used to test the impact on performance of changing a range
of variables, including:
– Different timetable options (more trains, different timetable structures)
– Different train performance characteristics (mechanical and human factors)

191
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

– Different infrastructure configurations

– Changes to level and nature of delays/lateness
Typically, this is undertaken by a railway’s strategic planning department or “long term” train
planning organisation, focused on making the best possible decisions regarding the next major
timetable change or investment in infrastructure or rolling stock. In addition simulation is used
within railway operational planning to:
– undertake analyses of buffer times and allowances as an input to timetable construction.
– test draft timetables for conflicts between trains.
– generate timetables from scratch (although unless the problem is very simple it is usually
better to use other approaches – see for instance Chapter 2) or to search for additional
paths in an existing timetable (for instance for short notice freight services).
– modify timetables to accommodate engineering work.
– model train movements within maintenance depots.
– model pedestrian flows at stations, including understanding the interaction between trains
running late and congestion on platforms and escalators (outside the scope of this book).
– predict conflicts “real time” to support on the day decisions to amend the train service in
reaction to delays (see Chapter 13 for more on this use of simulation).
– help drivers and other train crew “learn the route” through driver simulation packages and
signaller practice making best use of the infrastructure to minimise train delay.
Simulation can be deterministic, where the results are entirely predictable, or stochastic (also
called probabilistic in some texts, e. g. Selia, Ceric & Tadikamalla (2003) where “behaviour
cannot be entirely predicted, though some statements may be made about how likely certain
events are to occur”, see Pidd (2004). Or, as Gamerman & Lopes (2006) put it, “stochastic
simulation is the area of science dealing with when some or all of its components are subject
to random variations”. This kind of random variation clearly occurs in railway operations – be
it caused by points” failures, extended station dwell times, driving variations etc. Monte Carlo
simulation is particularly appropriate in these circumstances, using “a method of sampling in
such a way that the sample represents the whole”, see Jones (1972). For railway timetable
simulation this method is used to sample delay events and use these delays to affect the
running of trains in a way which, if repeated over enough simulation days, will produce by the
average of all these days a good approximation of performance on a typical day. Section 10.2
considers different methods of simulation in more detail.
Most railway timetabling computer simulation software has started life as a university research
project and much important work to develop the accuracy and sophistication of simulation still
takes place in universities. Section 10.3 describes important developments in the academic
arena. There are now a number of commercially available simulation packages: some of these
are described briefly in Section 10.4.
Getting reliable results from simulation requires a considerable amount of effort. A significant
volume of data has to be collected and input, the resulting model needs to be tested to ensure
that as far as practicable it does replicate reality and considerable care needs to be taken
in analysing the model outputs to ensure that the correct conclusions and recommendations
emerge. Section 10.5 looks at simulation in practice, seeking to provide the reader with some
insight into how to undertake successful simulation projects; Section 10.6 then looks at some
of the current limitations of simulation software and discusses areas for development.
The EU funded project Optimal Networks for Train Integration Management across Europe
(ON-TIME) recently published an “Assessment of State-of-Art of Train Timetabling” (ON-TIME,
2013) and this covers similar ground to this Chapter.

192
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 10.2 Methods of simulation

10.2 Methods of simulation

This section considers a number of key differentiators in how the simulation is to be

undertaken and considers when and if each is appropriate.

Deterministic vs stochastic
Simulation is deterministic if all parameters are defined by the user and do not contain any
random components. Deterministic models are used to represent real systems which are
too complex to be evaluated analytically and where the impact of random (or quasi-random)
disturbances to the normal running of the system does not need to be assessed.
In stochastic simulations random components are introduced to better represent other
phenomena. Since stochastic simulations are used to evaluate the behaviour of a system with
some random factors, results of a single simulation run have no statistical relevance. As a
consequence, a number of multiple simulations must be performed.
Deterministic simulations support timetable planning or the design of new infrastructures, while
stochastic models allow timetable robustness or stability analysis.

Continuous vs discrete
In continuous simulation models, the value of state variables change continuously in time.
It is calculated therefore with continuous analytic resolution of state equations. In discrete
simulation models, state variables are calculated only at fixed time intervals (fixed-increment
time advance models) or when an event occurs, independently from the time-span between
two successive events (next-event time). Railway simulators are continuous when they solve
the motion equation with a continuous algorithm, while more simplified models use pre-defined
process and event times in a discrete approach.
Since some of the processes of rail operations are continuous, such as for example train
motion along a line, while other are discrete, such as the release of a block section or the
scheduled departure from a station, the state-of-the-art simulation models use a mixed
continuous-discrete approach.

Micro vs Macroscopic
Macro-simulation models use a simplified infrastructure model to reduce computational time
and therefore allow simulation of larger networks. Micro-simulation models offer a description
of infrastructure which reproduces the functionality of interlocking, safety and block systems.
Macroscopic models describe the network and the timetable as a directed graph of simple
nodes (stations) and links representing the line sections and including the frequency and the
running time of trains. Geographic infrastructure elements within microscopic models are also
represented by a graph, but it represents the complete layout of stations and lines including
the vertical and horizontal alignment, all switches and the signalling, interlocking and automatic
train control/protection systems. These data, combined with the technical characteristics of
the rolling stock and a detailed timetable, allow the reproduction of train movements in the
greatest detail.
Since some of the outputs of a microscopic simulation, such as the calculated running times,
are an input of macro-simulation, and the dimension of networks that can be simulated
in a reasonable time is completely different, as well as the accuracy of the outputs, the two
simulation methods can be considered as complementary. Microscopic models tend to be
more accurate because of the greater detail in the infrastructure model, and so have to date

193
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

been used to a very much greater extent than macroscopic models for railway operational
simulations. However, research and development work continues to expand the capabilities
of macroscopic models, e. g. mesoscopic simulation models. These aggregate some of the
microscopic infrastructure and timetable data of individual lines, so that the critical interfaces
on a network scale remain computationally tractable (see Chapter 3).

Synchronous and asynchronous simulation models

Microscopic simulation models can be either synchronous or asynchronous.
In synchronous simulation, all the motion equation of trains which are moving together on a
microscopic network are modelled with respect to interlocking and safety system functionality.
Differences between models are mainly in the flexibility to represent different technologies, in
the capability to receive automatic inputs or to generate specific outputs.
In a standard asynchronous simulation, trains with highest priority are simulated first,
and conflicts among them solved with a first-come – first-served strategy; the resulting
infrastructure occupations are stored. Then the process is repeated for each priority group,
more and more saturating the time windows which are still unused. Therefore no conflict
between trains with different priority is possible: high-priority services are never forced to brake
or to stop by conflicts with low-priority trains.
Asynchronous and synchronous models show different strengths and weaknesses, which
make them complementary tools to support railway planning. Synchronous simulation allows
a more realistic representation of train traffic, with all trains simultaneously moving on lines and
therefore interacting between each other, with secondary delays transferred from high-priority
to low-priority trains and vice-versa. In synchronous models it is also simple to implement a
wide variety of incidents, which may involve nearby all infrastructure and signalling elements,
also featuring stochastic variability, such as intermittent opening and closing of level crossings.
Asynchronous simulation is more suitable to support timetable construction, which is
performed stepwise and where low-priority trains are normally scheduled after other services.
It is also simple to define new courses within a pre-defined timetable, obtaining a conflict-free
timetable and an automatic allocation of station tracks.
Synchronous models are much more widely used, but there are some inherent issues.
There is a risk that synchronous models will not always manage the priority order correctly,
since priority may be based only on pre-reservation of block sections, causing unrealistic
hindrances to other trains and increasing occupation of the conflict sections. On the other
hand, in asynchronous tools trains with high priority are always preferred and displace trains
with lower priority. Because of the strictly hierarchical structure of an asynchronous simulation,
trains with low priority may receive more delays in an asynchronous simulation than they would
in reality.
Another issue with synchronous models is that of deadlocks. They normally occur on single-
track or bi-directional lines, when two or more trains may be allowed to block and/or enter
the same line section at the same time, potentially stopping “head to head” in front of each
other, unable to move and “deadlocked” until the end of the simulation. Deadlocks can mostly
be solved using simple algorithms, which are normally reliable and efficient, but in certain
situations more complex algorithms are needed – or manual intervention (see later section of
this Chapter). Where it is not possible to eliminate deadlocks by modifying the settings for the
model, it may be necessary to discard these simulation runs which can reduce the reliability of
the results.

194
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 10.3 Research and development

10.3 Research and development

Significant research work in the late 1990s and in the first decade of the 21st century has
led to the development of a significant number of software tools and the refinement of the
algorithms. Thanks to this, simulation is now able to represent nearly all “real life” cases with a
good level of accuracy. Since the most used commercial tools appear now as very complete
and reliable, there has been a reduced level of investigation in the last few years into new
simulation methods and algorithms and the research in the field has been aimed mainly at
improving the quality of input data, extending the traditional fields of application, improving the
validity of dispatching algorithms (see later section of this Chapter) or to its use as a decision
support system for dispatchers (see Chapter 13).
One piece of interesting recent work is that carried out at RWTH Aachen, especially by Bueker
(2010, 2012), which has led to an innovative macroscopic simulation algorithm. In this, delays
are represented as distributions functions and delay propagation is calculated analytically on
a rule-based infrastructure model. This combination leads to very short computation times,
which allow simulating even very large networks, such as the entire network of Germany in
a reasonable time, see Franke et al. (2013). The work was implemented in the OnTime
simulation tool.
This macroscopic simulation algorithm can be considered as complementary to the more
conventional synchronous microscopic simulation. The tools might also be combined to
analyse an entire network (macroscopically), but with a number of focuses on the critical
junction areas or regions, which are simulated microscopically. The two simulation tools can
be co-ordinated, in order to let the input delays of the microscopic simulation be coherent
with the delays obtained using the macroscopic simulation. The outputs of the two algorithms
are then aggregated, obtaining an estimation of the robustness of the timetable, see Maricaua
et al. (2013).
While the use of simulation has become widespread since the late 1990s, very little scientific
work was dedicated to assessing the quality of outputs and calibrating the stochastic inputs,
such as the performance of trains. De Fabris et al. (2011) first proposed using GPS data to
calibrate the motion equation within synchronous micro-simulation and estimated the benefits
of modelling departures, dwell times and train performances with growing accuracy. Further
studies at the University of Trieste, see Medeossi et al. (2011), led to the development of an
algorithm to estimate a set of performance factors based on GPS trackings collected on-
board. These performance factors allow representing the variability of running times within
microscopic simulations, see Chapter 11 for more details. More recently, Besinovic et al.
(2013) introduced a method to estimate a set of parameters of the motion equation from
observed track occupation/release data. A simulation-based optimisation approach is adopted
to calibrate input parameters of the equations describing the tractive effort, the motion
resistance, the braking effort and the cruising phase.
While these approaches aim at calibrating the motion equation using real data, Kirkwood et al.
(2013) presented the results of a comparison between the speed profiles calculated using two
commercial microscopic tools and the corresponding real profile measured on board using a
GPS device.
One of the most interesting potential fields of application of simulation is the validation
of dispatching solutions. First Luethi et al. (2007) and later Corman et al. (2010) and
Quaglietta et al. (2013) used synchronous microscopic simulation to test the benefits in terms
of delays or robustness of adopting optimal dispatching algorithms.

195
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

The application spectrum of simulation might also be extended by adding further functionalities
to the existing tools. The OpenPowerNet tool, see Stephan (2010), provides a very detailed
simulation of electric power supply systems. It works together with the OpenTrack simulator as
a so called “co-simulation”. OpenPowerNet is able to simulate all common AC and DC railway
power supply systems, taking into account the entire electrical network structure. It can be
used as an energy prognosis and analysis tool as well as for the planning and optimisation of
power supply installations.
Another interesting functionality that extends the outputs of simulation is the estimation of
costs. First Lienau (2007) presented an approach to include the construction and maintenance
costs of infrastructure; Beck et al. (2008) extended this method to a comprehensive cost-
benefit analysis tool, called RAILONOMICS, connected to the RailSys simulation software.
Finally, simulation algorithms might be used to support dispatchers in taking decisions.
Instead of calculating an optimal set of decisions, as performed by rescheduling algorithms
(see Chapter 13), simulation tools can predict operations by finding conflicts between trains.
These conflicts can be solved by means of conflict resolution algorithms. As with simulation,
these algorithms may be synchronous, as recently presented by Oetting et al. (2013) or
asynchronous, as proposed by Weymann (2011) and Kuckelberg (2011).

10.4 Commercial software

This section describes briefly some of the simulation packages that are available for
commercial use at the time of writing. Some of these are available primarily for research or
consultancy work by the owning organisation, whereas others can be purchased for use by
railway companies or third party consultancies.

Macro-simulation models
These have been developed mainly to simulate timetables over a wide network. To do this,
approximations are made to keep the model complexity and simulation run time manageable,
but at the cost of reduced precision and in many cases accuracy. These models tend to be
used at a strategic planning stage, sometimes in conjunction with micro-simulation models.
FASTA (EPFL 2013) is a discrete-event simulation system developed at the EPFL Lausanne.
The network is modelled as a directed graph where nodes correspond to stations and edges
to the line sections; infrastructure constraints are represented by minimum headway times.
Running times are calculated during the simulation, also considering random parameters in
stochastic simulations.
MERIT (Modelling the Reliability of Infrastructure and Timetable) (AEA Technology Rail, 2002)
is a discrete-event simulation tool owned by DeltaRail for simulations on large networks. The
tool is based on a significantly simplified model in which margins and technical times are
approximated from the planned values and instead of distributions of primary delays, it uses
randomly generated incidents to assess timetable robustness.
The deterministic macro-simulator NEMO (Netz-EvalutionsMOdell) (IVE, 2013) is a strategic
planning tool for the evaluation of infrastructure scenarios. NEMO allows the automatic
generation of future capacity requirements considering traffic forecasts. The tool considers
freight and passenger traffic separately. Freight traffic is given as an O/D matrix that
distinguishes different kinds of goods; NEMO calculates the corresponding number of loaded
freight wagons per segment and builds up block trains for higher traffic amounts and single
wagon traffic based on a production network for the resting parts. The number of passenger

196
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 10.4 Commercial software

trains needed on each route is calculated based on forecast traffic volumes and a modal split
model. Freight and passenger trains are then assigned to the network, leading to network
loads. http://www.ivembh.de/nemo/index.en.htm
A stochastic macro-simulation model has been developed at the RWTH Aachen, see Weidner
(2007), to be integrated in the strategic planning tools of DB. The model represents the impact
of real process time distributions on a given timetable, producing statistics about punctuality
and the number of missed connections. The low computation load of the model enables study
on a very large scale, such as on the entire SBB network, see Akermann & Bueker (2008). The
model was further developed in the OnTime software, see Franke et al. (2013) (http://ontime-
rail.com/index_en.html).
SIMONE is a discrete-event simulation tool used by ProRail, the Dutch Infrastructure
Manager since 1999, see Middelkoop & Bouwman (2001). Receiving automatic input from
the timetabling system DONS, SIMONE performs stochastic simulation to assess punctuality
with different timetable or infrastructure scenarios. The infrastructure model is made up of
timetable points and line sections; in each timetable point a train is assigned to a platform
group. A train must wait outside the station area if all possible platforms are already occupied.
Stochastic disturbances to dwell and running time to define primary delays can be inserted to
perform multiple simulations, producing a variety of statistics at different aggregation level and
considering the most commonly used measures.
TRAIL (Train Reliability Availability Infrastructure Logistics) is a railway system reliability
simulation tool supplied by DNV. A TRAIL software simulation requires the items of
infrastructure to be divided into sections. Each section is generally defined as a continuous
piece of track with no switches or crossings. Each section is described in terms of its
components and their failure and repair characteristics, making it straightforward to compare
scenarios relating to different asset types and their reliability, predicting at a macro level the
performance that can be expected with different failure rates (http://www.dnv.com/services/
software/products/maros_taro/trail/).
TTRA (The Timetable Robustness Analyser) is an element of Trapeze’s TrainPlan product.
It simulates a detailed timetable over an approximate network to test timetable quality and
robustness, see Stallybrass (2002). TTRA enables users to get a quick and coarse indication
of a detailed timetable’s performance over an approximate network (http://www.trapezegroup.
eu/solutions/rail-transport/simulation).
All of the above packages have been used for the most part on complex mixed-traffic railways
and originate from Europe. In other parts of the world the simulation problem can be rather
different, particularly where freight trains predominate, single track main lines are more
commonplace and train plans vary significantly day by day to accommodate the freight cars
that need to be moved. Simulation software that focuses more on this has been developed
including the following.
The MultiRail Suite, from the USA, which includes Trip Planner and its Advanced Network
Simulator. The Trip Planner simulates the movement of traffic through the overall operating
plan. Starting with a traffic movement’s release day and time, and based on its blocking
sequence, the Trip Planner simulates the selection of trains used to move the traffic. The
Advanced Network Simulator looks at a higher level, simulating the overall operating plan
highlighting high and low demands across the system at the yard-block, train and yard level
(http://rail.railplanning.com/multi-rail/).
CFlex Movement Planner, from Brazil, which can again simulate overall operations plans for
freight railways (http://cflex.com.br/site/en/menu/products-and-services/railroad-solutions/
movement-planner/).

197
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

Synchronous microscopic simulation models

These packages, the most commonly used by both academic and commercial organisations,
have considerable similarities in terms of approach. They have detailed infrastructure models
and calculate the location of trains as they are moving through the network, using knowledge
of the infrastructure and train performance characteristics. Key differences between the
models relate to how well they function with large networks and the sophistication of the
despatching algorithm (the heart of the simulator that makes train priority and routing
decisions in place of the signaller or train controller). Other factors are how user friendly the
user interfaces are, the extent to which the user can adjust the way the simulation runs or
manually override settings within the software, provision of interfaces with other systems and
the sophistication of the outputs. At the present time Opentrack and RailSys are the most
widely used of these software packages.
HERMES – Graffica offers the Holistic Environment for Rail Modelling and Experimental
Simulation. This relatively new simulator is being used on the European Union’s 7th Framework
Package ON TIME project (http://www.ontime-project.eu/home.aspx) to provide an integration
platform for the traffic management and driver advisory work packages and to provide support
for the timetable and emergency planning work packages (http://www.graffica.co.uk/rail-
traffic-management/simulation/).
OpenTrack (Huerlimann (2001)) is a simulation tool developed at the Institute for Transport
Planning and Systems (IVT) of the ETH Zurich and now supplied and refined by OpenTrack
Railway Technology Ltd. OpenTrack is used by railways, the railway supply industry,
consultancies and universities in a significant number of different countries, including for
the calculation of power and energy consumption of train services. It is often used also for
simulation of Tram/Streetcar, Light Rail, Metro/Subway/Underground systems and even
Maglev systems (such as Transrapid) (http://www.opentrack.ch/).
RailPlan is another detailed signal berth simulation model, which simulates infrastructure,
timetable and rolling stock information in detail (Comreco Rail, 1997). RailPlan was first
developed in 1987 and is of Swedish origin, now being developed and sold by Trapeze. An
“add-on” is PowerPlan, a post-processor for RailPlan that allows the modelling of electrical
networks (http://www.trapezegroup.co.uk/solutions/rail-planning/simulation-of-train-operations).
RAILSIM Network Simulator models train operations on many train control systems over
rail networks of multiple corridors and multiple modes. Features include: automatic conflict
detection and resolution, an “Optimiser” to diagnose model issues, cab and wayside signalling
(and wayside with cab), advanced interlocking processing and operational randomisation (for
instance, lateness entering the modelled area, dwell times, and tractive effort) (http://www.
railsim.com/).
RailSys (RMCon, 2008) is a comprehensive signal-berth simulation package, which also has
modules that facilitate infrastructure management, timetable construction and possession
planning. The software database structure allows simple storing of very large models, which
is reflected by its use by many infrastructure operators. The model features a timetable
construction system and an automatic slot search algorithm (Hauptmann, 2000). Refined
simulation technique to prevent deadlocks, a number of input and output capabilities and
the presence of variable process and event times made this tool suitable for large-network
stochastic simulations Rudolph & Demnitz, (2003) (http://www.rmcon.de/en.html).
VISION (Visualisation and Interactive Simulation of Railway Networks) is owned by DeltaRail
and Network Rail has a perpetual licence. It is used for specialist purposes by Network Rail
(where other microscopic simulators do not have a full understanding of British signalling), with

198
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 10.5 Simulation in practice

some licences also in Spain and Italy (RFI). The model focuses mainly on the infrastructure
design for small network sections. This is reflected in the simple interface for building models
and in the outputs. It has a high quality video capability for displaying simulations. A detailed
description of VISION is provided in Brown (2002). Where power modelling is required, there is
a two way interface with OSLO (Overhead System Loading).
As noted above, freight railways have different needs from mixed traffic railways, and
an example of software focused on this kind of simulation need is Rail Traffic Controller
(RTC). This simulates the movement of trains through rail networks at a detailed level
and has “meet-pass N-train logic™” incorporated in the software to model the single line
problem that many freight railways face, whilst also dispatching trains at the network level.
Like the other simulators described, it has an integrated train performance calculator to
take account of different equipment types, train consists, terrain and track conditions
(http://berkeleysimulation.com/rtc.php).

Asynchronous simulation models

While synchronous models have been developed in different countries and universities,
asynchronous simulation tools have been mainly studied at the Institute of Transport
Engineering (VIA) of the RWTH Aachen. In a standard asynchronous simulation, trains with
the highest priority are simulated first, and conflicts among them solved with a first-come –
first-served strategy. The resulting infrastructure occupations are stored. Then the process
is repeated for each priority group, saturating progressively the time-windows which are still
unused. Therefore, no conflict among trains with different priority is possible; high-priority
services are never forced to brake or stop by conflicts with low-priority trains.
At the same Institute, the infrastructure model called SPURPLAN (Bruenger, 1995) was
developed, which can be stored in a database and contains a detailed infrastructure
description. SPURPLAN is a graph, which contains not only the infrastructure and interlocking
systems, but also route priorities at each station and for each train category or family. This
allows automatic routing of trains and platform assignment at stations. The model was first
developed for the RUT timetabling system based on blocking time theory currently in use
in Germany. On the basis of the SPURPLAN graph, a software family has been developed
including a capacity assessment algorithm based on queueing theory ANKE (2.3.2, Vakhtel,
2002), the asynchronous simulation tool BABSI (BetriebABwicklungSImulation) – see Groeger
(2002) and the dispatching-support tool ASDIS (ASynchrones DISposition0 – see Jacobs
(2002).

10.5 Simulation in practice

This section focuses on the most commonly used simulation approach – synchronous
microscopic simulation and provides a step by step description of the key stages that need
to be undertaken in setting up, running and analysing the results from this type of simulation.
Fig. 10.1 shows the inputs and outputs of simulation. The stochastic outputs have to be
inserted only when performing stochastic simulation.

Base data and model set up

Collation and input of base data to a simulation package can often be a very time consuming
task. Potentially It can take many weeks to build a model covering 100 km or more of complex
rail network from scratch.

199
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

Fig. 10.1: Inputs and outputs of simulation

Considering infrastructure configuration first, most microscopic simulators work at the “signal
berth” level and the data to be collated and entered into the model includes:
– Track layout, including switches and crossings with location (to nearest metre)
– Block sections, signals and permissible routes with location (to nearest metre) together with
associated detail including number of signal aspects, overlap distance (to nearest metre),
set time and release time
– Speed restrictions by location and stock type
– Gradient changes
– Stations and other stopping or timing locations
– Platform length
– Other restrictions by line and route, for instance whether a line is electrified or not and
forbidden stock types (e. g. due to weight or curvature limits)
This data may be available electronically or may have to be painstakingly transcribed from
track plans and signalling diagrams and other sources.
Next train performance characteristics need to be procured and input. The basic laws of
physics regarding power to weight ratios, adhesion limits and braking rates apply, but the
detail for each different train type needs to be input. Manufacturers provide “as built” data, but
for a variety of reasons this may not be appropriate for timetable construction and simulation.
For instance, subsequent modifications can change characteristics significantly, safety
margins might be applied in practice which mean that maximum design braking rates are not
used and different driving practices can impact significantly as well. It is best to use actual
performance data derived from logging systems to at least cross check design-derived data.
Some simulation software uses simplified formulae to calculate running times (e. g. not fully
taking into account dynamic effects of long freight trains) and in some circumstances it will be
necessary to use yet more detailed software to calculate more exact train running times.

200
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 10.5 Simulation in practice

Fig. 10.2: Drawing of a station (top) and its model in OpenTrackTM (centre) and RailSysTM (bottom)

Once the infrastructure has been set up and train performance data input, the timetable
needs to be overlaid on that infrastructure. The timetable might have been developed using
a separate software package and be imported via a standard interface format (e. g. .xml or in
the UK PIF or CIF) or it might be developed from scratch using the timetable editor provided
in most commercial simulation software. This editor capability will usually permit the creation
and retiming of trains to meet the required specification and the subsequent assessment and
elimination of any conflicts created. The more sophisticated packages will have “drag and
drop” functionality to allow for rapid editing and will automatically highlight conflicts between
trains (for example, where signalling headways have been infringed). A decision needs to be
made on how to accommodate in the modelling trains that do not run every day. Particularly
in areas where there are significant numbers of freight trains, this is important to get right.
Different numbers and types of freight trains can materially impact on the performance of a
mixed-traffic railway.
Depending on the number of trains included in the simulation and the presence and detail
of the stochastic inputs, it is possible to run “single-train”, deterministic, stochastic and
“advanced” stochastic simulations. These four methods are compared in Fig. 10.3.

Run “single train” and “nominal” or “deterministic” timetable simulations

At this stage it is possible to run “single-train” and “nominal” or “deterministic” timetable
simulations.
In the “single-train” simulation a single train is considered each time. This simulation is used
only to verify the running time of that train and the correctness of the model.

201
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

Fig. 10.3: Single-train, Deterministic, Stochastic and Advanced stochastic simulations

The nominal timetable simulator can be used to highlight direct conflicts between trains (which
can occur as a result of inaccuracies or lack of precision in the scheduled timetable) and their
consequences. In this mode, the simulator is run without adding delays. Conflicts that will
be identified include headway conflicts, double occupation or crossing conflicts, as well as
any incidence where a train driver will see and react to a restrictive aspect signal. Running
a “nominal” simulation is also an essential step in ensuring that the model is set up correctly
– comparing the trains running in the simulation with the timetable and if possible real train
running data to find and resolve model set up errors.
Fig. 10.4 shows the animation produced by a commercial tool during the simulation is
performed. It is possible to recognise that the model correctly reproduces the 4-Aspect
signalling used in the UK in the combination of colors shown by signals.

Fig. 10.4: Animation of a train running under 4-Aspect signalling

202
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 10.5 Simulation in practice

Once errors have been resolved, the nominal timetable simulation is suitable for:
– Running time analyses, to check values used currently (although comparison with “real”
data is most important) and to provide new figures for future infrastructure and rolling stock
– Headway and junction margin calculations
– Conflict detection as part of the timetable construction process
– Capacity analyses (some models include calculation of UIC406 capacity utilisation indices
as part of their standard product)
– Generating timetables from scratch (although unless the problem is very simple it is usually
better to use other approaches – see for instance Chapter 2)
– Searching for additional paths in an existing timetable (for instance for short notice freight
services)
– Modelling and testing of current and future timetables for conflicts
– Certain types of what-if-analyses, for example checking whether the timetable can still be
operated in full without incurring delays when track possessions are planned

Stochastic Simulation
Whilst nominal simulations have their uses, most timetable simulation studies require as an
output a prediction of delay and lateness, which is achieved using stochastic simulation. In
stochastic simulation, random delays are inserted, and the simulation repeated several times
to obtain statistically relevant outputs.
The most useful outputs of stochastic simulation are the delay statistics, which give planners
an estimation of the capability of the system to recover the delays or longer dwell times.
The easiest way of inserting these variables is to use the functions already integrated in the
simulation tools, in which it is possible to select the mean initial delays and the variability of
dwell times for each train category. The tool will then automatically choose the delay and
dwell times to be assigned to each train and station based on an integrated distribution. The
use of these pre-defined functions allows them to be defined quickly, much reducing the
time required before starting a set of simulation. Moreover, since no real operational data is
required, it is possible to test the robustness of lines that are not yet in operation.
The stochastic simulation is suitable for any test in which it is possible to measure the
robustness of operations under realistic conditions:
– Testing of current and future timetables for robustness (and systematic conflicts)
– Capacity analyses in which the robustness corresponding to a certain level of saturation
must be considered explicitly
– Certain types of what-if-analyses, for example checking whether the timetable can still be
operated in full without incurring delays when track possessions are planned

“Advanced” Stochastic Simulation

When the highest accuracy in the estimation of robustness is required the stochastic variables
are modelled with more detail, to try to reproduce the real behaviour of trains on a certain
line or network. Since it is obviously impossible to forecast accurately the initial delays before
a new timetable is put into service, the best possible approximation is to use the existing
variations on the line. In this way it is also possible to perform a validation of the model, which
is obtained by simply comparing the result of the simulation with the existing delays. Once
the validation is performed, and the quality of the results is considered acceptable, any new
scenario can use the same variability as stochastic inputs. Since the scenarios will be stressed
under the same conditions, any difference in the output delays will be due only to the impact of
the different input (the timetable).

203
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

The advanced stochastic simulation appears as the most accurate method to reproduce rail
operations, since it combines a very accurate modelling of the network (including all features
that influence train traffic) and of the rolling stock with the real (measured) stochastic behaviour
of the system. It is suitable for the same purposes of the stochastic simulation described, plus
the ex-ante estimation of reliability and punctuality:
– Ex-ante punctuality estimation of future timetables
– Testing of current and future timetables for robustness (and systematic conflicts)
– Capacity analyses in which the robustness corresponding to a certain level of saturation
must be considered explicitly
– Certain types of what-if-analyses, for example checking whether the timetable can still be
operated in full without incurring delays when track possessions are planned.
The following sections describe the stochastic inputs to be used in simulation and how they
can be extracted from operational data collected by the infrastructure managers.

Base Model Delay and Lateness Data

The complexity of modern railways leads to variability in the process times and to potential
risks of disruptions due to malfunctions and deteriorations. In a few cases, devices such as
switches or signals which are particularly vulnerable to disruptions can be found and can
be simulated. On most networks, breakdowns are more randomly distributed, significantly
reducing the statistical relevance of the simulation of a specific case. Therefore, to validate a
model, major disturbances within the node have to be excluded, and only “normal” variability is
considered.
As primary variability sources the initial delays, the stop times and the running times are
considered. The initial delays represent the distribution of departures for each train from its
first station within the simulation area. After this first departure, trains move along their route
according to their running time distribution, and stop at stations for a variable time according
to the stop time distributions. If these data are estimated accurately, and the simulation model
correctly represents the interactions among trains, the departure distribution at each following
station can be quite accurate when compared to the real ones.

Initial Delays
When defining the expected initial delays for a new timetable on the basis of the past
data collected at each station, it is important to exclude all major perturbations and all
secondary delays (delays due to traffic conflicts). Some researchers use interlocking data
and sophisticated mathematical models, see Yuan & Hansen (2007), to identify primary and
secondary delays. However, these methods require using blocking times collected at block
sections of the network, which are not currently stored in most countries.
When such accurate methods for filtering out secondary delays cannot be used, a rough
filtering process can be deployed with the goal of defining a realistic initial delay distribution
(removing the effects of conflicts with arriving trains) rather than analysing in detail conflicts
based on track occupation. The method uses only train passing data, automatically collected
at each timetabling point by the train describers, which identify trains at their first station and
then keep track of progress along the route.
This simple model is relatively accurate for use in simple stations where complex conflict
chains hardly occur, while In other cases, for example at large stations, it introduces higher –
but still acceptable in the Authors’ experiences – initial delays.

204
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 10.5 Simulation in practice

The method consists of two steps. First, major perturbations are simply filtered out by
excluding all delays greater than 600 seconds (5 minutes). The 5-minute threshold was chosen
based on an extensive analysis of major stations in Italy showing that more than 99% of delays
due to passengers alighting and boarding or other “normal” phenomena are lower than 5
minutes. Therefore, this threshold can be used to filter out breakdowns that can be excluded
in timetable planning. (The 5-minute threshold would need to be reconsidered for application in
another context or country.)
Second, the conflicts between arriving and departing trains are identified using a mesoscopic
model at each relevant station. This process uses data on station arrival/departure time for
each train as well as the sequence of signals the train has passed and the corresponding
time. All trains with departure delays are checked using these data to determine if a conflicting
movement took place. In case of a conflict, the record is simply excluded from the data set.

Stop/Dwell Times
A precise estimation of stop time appears relevant on passenger lines, while it appears less
important on freight lines as well as on single-track lines, where the stopping times are more
influenced by the crossings than by the passengers boarding and alighting.
Stop time variability includes two phenomena affecting trains at stations: departure imprecision
and dwell time variability. Departure imprecision occurs when a train, which arrived early or
punctually at a station, does not depart punctually although all passengers are already on
board. Departure imprecision is generally short for regional trains, with mean values between
10 and 20 seconds, while it is often more than 30 seconds for long-distance trains. Dwell
time depends on the number of passengers boarding/alighting, the presence of late-arriving
passengers and on the physical layout of the train-set.
The two phenomena must be considered separately to model stop time variability correctly.
This process is complicated because often the only data available are arrival and departure
times. Longo (2012) proposes a detailed stop time calculation model, based on a previous
work by Buchmueller et al. (2008). The model can also be used to estimate the effect of an
increased demand or different rolling stock, but it requires a set of measures to be collected
manually at some stations. However, in most studies it is quite demanding to collect those
inputs, so a simpler method is presented here, see Medeossi et al. (2011). The method is
valid for services operated with a frequency greater than every 10 minutes and uses two
assumptions to divide the measured stop time into different parts:
1. For late-running trains, stop time has no departure imprecision and consists only of dwell
time.
if Da > 0 then Td – Ta = Ts = Tdw (10.1)

2. On early-running trains, the stop time includes both dwell time and departure imprecision.
if Da < 0 then Td – Ta = Ts = Tdw + Tdv (10.2)
where
Da = Arrival Delay
Ta = Arrival time
Td = Departure time
Ts = Stop time
Tdw = Dwell time
Tdv = Departure variability

205
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

In this method, dwell time (Tdw ) is obtained by considering only the delayed trains (where stop
time equals dwell time). The departure variability is obtained by using the estimated dwell time
in Equation (10.2) for trains that arrive early.
The departure time used in the model is the maximum of the departure time (which considers
departure variability) and the sum of arrival time and dwell time (which implicitly includes its
variability). Therefore, given the distributions, the model automatically considers dwell time
variability and departure imprecision.
Td = max ((Tdp + Tdv), (Ta + Tdw), (10.3)

where
Td = Departure time (to be estimated) and
Tdp = Departure time (planned)

The same train describer data used for defining the initial delays can be used as input to
estimate Td and Tdw.

Running times
More complex is the estimation of the variability of running times to be used in simulation.
It is modeled through a performance parameter in the simulation tools. The performance
parameters generally reduce train performance to represent real driving more accurately.
The estimation of performance parameters is by far more complex than that of the initial
delay and stop time distribution, since it requires finding the performance parameter that
best reproduces the measured running time. A method for estimation of the performance
parameter(s) is presented in Section 11.4 (Distribution fitting of train delays) of this book.
In a more practical use of simulation, it appears sufficient normally to use again the train
describer data and compare that to the results of simulation.
Train performance parameters are normally between 90 % and 98 %, with the lower end more
common on stopping trains and values quite close to the upper limit for high-speed services.

Grouping trains
Manually preparing the inputs for an advanced stochastic simulation might take a considerable
amount of time, especially when performed without specific tools such as TRENO Analysis.
Independently from the tool used to prepare the distributions, another task that has to be
performed is the definition of the groups of trains that will have the same stochastic behaviour.
In the experience of the Authors, the variability of running times appears quite related to the
kind of services operated, and therefore the category of the train. Thus a single distribution for
each train category can be used.
The initial delays are related to the train category, each station and often the time of the day.
Assigning a distribution to all trains of a certain category and starting at the same station
can be a good compromise, since defining different distributions to reproduce the peak and
off-peak hours does not often lead to significant differences. An analysis of operational data
before starting the simulation allows choosing a reasonable approximation.
Dwell times are the most difficult to group, since the effect of peak hours is particularly
important. Moreover, since each train usually has multiple stops, and at each stop the number
of passengers boarding and alighting is different, it is necessary to consider at least the
relative importance of each station. Only an analysis of real dwell times allows the decision

206
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 10.5 Simulation in practice

on how stations and trains can be aggregated and whether peak and off-peak hours must be
considered separately or not. Obviously, a higher number of distributions might require more
time not only to be prepared, but also to be assigned to each simulation scenario.

Calibration
It is best practice to compare the model outputs with actual delay and lateness data, where
the model is to be used to analyse changes from an existing situation, once all the data has
been input to the model as above and before looking at the future situation.
This will usually be the first use of the stochastic simulation capability of the software for a
particular model. The delay and lateness distributions are sampled using typically the Monte
Carlo statistics principle and the software creates a number of timetables to be perturbed by
overlaying these samples of primary delay and lateness. Simulating these timetables gives
punctuality results for, in effect, a range of possible operating days, with different primary
delays and hence different secondary delays. The number of timetable days required to
be generated so that there is a high likelihood that the overall average delay reflects reality
can vary between 30 and 250, depending on the nature of delays in the study area and the
complexity of the network and timetable. Higher levels of congestion usually require higher
numbers of runs as overall delay levels can be more sensitive to small changes in primary
delay.
The resulting outputs can then be compared with the existing actual average levels of delay
and lateness and a decision made on whether the level of correlation is good enough for
the study results to be valid or whether more refinement of the model is required. Deciding
what is “good enough” can be quite difficult – it depends on how critical the model accuracy
is to making planning decisions (in particular how close one option is to another, when fully
evaluated). An analysis of model reliability and an approach for the estimation of the stochastic
inputs and the comparison with the outputs for a series of Italian case studies has been
provided by Medeossi (2010).
It is important to understand that none of these days is likely to replicate an actual day exactly,
as the distributions have been sampled randomly. Should there be a reason to want to
understand a particular day, or as a further check that the model is closely modelling reality,
then the specific delays and entry lateness for that day will need to be input.

Setting up variant models

Once the modeller is completely satisfied that the base model is mimicking reality, the next
step is to set up the future options or “variants”. Depending on the scheme, this could
involve different timetable options (more trains, different timetable structures), different
train performance characteristics (mechanical and human factors), different infrastructure
configurations (e. g. provision of fly-overs, more tracks, changes to speeds); changes to
level and nature of primary delays/lateness to reflect anticipated benefits from more reliable
infrastructure or trains. Once the variants have been set up and tested for errors, stochastic
simulation is used to produce predictions of average levels of delay and lateness across the
number of timetable days that are simulated.

Results and analysis of outputs

Most simulation software packages provide a range of outputs, enabling comparisons
to be made between the base and variants in terms of their robustness, as measured in a
range of ways, including secondary delay caused, average minutes lateness and other

207
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

specific measures (e. g. in Great Britain, the Passenger Performance Measure is used, which
looks at the percentage of trains more than 5 or 10 minutes late for local and long distance
trains respectively). The data is typically “carved up” in a variety of ways, looking at different
locations, times of day, train type and direction, highlighting particular problems that could be
remedied by changing one of the input parameters. Graphs often provide a visual picture of
the results, including network maps showing which links have the worst delays on them and
showing how lateness changes along the routes trains take. Some example outputs follow:

Fig. 10.5: Delays in a network (RailSys)

Fig. 10.6: Lateness of simulated trains in train graph view (TRENO)

208
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 10.6 Limitations and areas for development of microscopic simulation

Fig. 10.7: Scenario comparison (RailSys)

10.6 Limitations and areas for development of microscopic simulation

At present, microscopic simulation is the most accurate way to mimic rail operations, since
infrastructure, the interlocking and safety systems as well as train movements are represented
in considerable detail. However there are limitations and areas that need further development.
They are discussed in this section.

Errors in model setup

The accuracy of data to set up the model is paramount but it is sometimes hard to find entirely
up to date sources. Signal plans may not be entirely up to date, train performance data might
be theoretical rather than realistic and delay data might have errors in location or content. To
some extent calibration can highlight these issues, but calibration is often not given sufficient
attention and significant defects remain in the model when it is used to compare variants. Over
time, it is essential that rail industry data repositories become increasing reliable, so that the
data source can be trusted.
Other inaccuracies are inherent in simplifications made in the configuration allowed by the
models. In particular, a high accuracy in the representation of the signalling systems appears
paramount to estimate running and blocking times correctly. The most flexible commercial
software already allows the representation of any kind of signalling system, including those
of tramways and the moving block or communication-based train control (CBTC)’ of metros.
Others offer an easier choice of standard signals but appear less easy to customise. This can
lead to errors in the results (a UK example is “approach controlled” signals, where trains are
slowed prior to the signal turning to “proceed”. Some software packages do not fully represent
the complexity of this, leading to run times being over or under stated).

209
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

Stochastic nature of inputs not fully recognized or modelled

When the deterministic inputs are set up correctly, the theoretical residual model error is
negligible, but the error becomes higher because the stochastic behaviour of some of the
inputs is not recognized in the simulation. The traction characteristics of trains and hence
journey times are not perfectly represented using conventional equations as they vary
depending for instance on the mechanical condition of the equipment (dependent on wear and
tear), the voltage being supplied in the case of electric traction (which can vary significantly for
DC powered trains) and the rail adhesion (depending on weather conditions).
Using the described simplified methods to calculate and apply delays and station dwell time
variability does not fully represent reality; in addition the nature of the distributions used is
frequently a considerable simplification.
Driving variability is an important issue that is hard to fully reflect in simulation; some models do
not even have the capability to take this into account.
Another question is whether the Monte Carlo method is always appropriate. It assumes
a degree of randomness which may not always adequately match reality (e. g. the lateness
of a sequence of trains entering the model area will often be inter-related rather than varying
independently). One approach to addressing this is to use lateness and delay data from actual
days rather than randomly sampling from distributions.

Dispatching rules
In microscopic simulation models a dispatching algorithm needs to be provided (equivalent to
a modern traffic management system), which can reschedule trains on the basis of the current
traffic status. However, in some simulation packages, instead of this being a realistic control
algorithm working at a network level, dispatching is based on the resolution of local traffic
conflicts sometimes based simply on a weighted FIFO (First in-first out). Each train features
a so-called pre-reservation distance (and time), which means that it can reserve a block
section a given time interval before passing the corresponding signal. The pre-reservation
distance is a function of the train category and delay. Conflicts are solved on the basis of the
pre-reservations and hence of train priorities, in line with railway practice in many countries.
But the fixed pre-reservation parameters do not fit well with the very different practical rules
used by traffic controllers, who are often better able to foresee the consequences of their
decisions.
For example, if a delayed train with low priority has stopped at a junction that is quite busy, it
will not be able to enter the main line until the route on which it is to travel or any with which
that movement conflicts is on a route which has been pre-reserved for another train. In the
corresponding real situation, the delayed train would probably be allowed to pass much earlier,
delaying another train but minimising overall delay. This and other dispatching issues can be
solved by integrating (or connecting) an automatic dispatching algorithm within the simulation.
At a given time interval (as frequent as possible whilst achieving an acceptable overall elapsed
time for running the model), the algorithm would receive the position of trains, find the set
of train routings and rescheduling actions that minimises delay (weighted by service group if
appropriate) and they would then be set and used in the simulation.
The lack of a higher-level traffic control is particularly noticeable on lines where real-time
dispatching systematically modifies the planned timetable in order to minimise delays. This
situation can be found, for example, on high-density, mixed-traffic lines, where freight trains,
and often also low-priority passenger services, are continuously rescheduled to minimise delay
for long-distance services.

210
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 10.6 Limitations and areas for development of microscopic simulation

Some extreme examples can be found in mixed-traffic lines where, to allow a higher flexibility
in traffic management, places where overtaking is possible are not fixed but are moved
continuously to the location chosen by controllers thanks to their experience. Moreover, the
time loss for them is often not concentrated at one station as stop/dwell time, but inserted as
running time supplements between three consecutive stations, forming the so-called “flying
overtakings”. The location of the overtaking is decided by the controllers autonomously. In
such cases, when performing a stochastic simulation, the simulator is not always able to avoid
conflicts for high-priority services. This can also lead to unrealistically high waiting times if a
train has to be overtaken and the overtaking tracks are already occupied.
Another weakness in traffic management can be found in the selection of alternative routes.
When a conflict is found, the simulator searches in a pre-defined itinerary list (with decreasing
priority) for an itinerary which is still free and blocks it. The itinerary is changed, even if the
pre-defined one is released few seconds later and if other conflicts are caused by the new
choice. Although this method to assign trains to routes is not realistic, experience gained
with a wide range of topologies shows that normally it does not lead to deadlocks nor to a
significant increase in unrepresentative delays – although of course in some circumstances it
might. Typically only localised rerouting is possible (for example from fast line to slow line) and
decisions on whether it might be better to divert trains so that they miss out one or more
stations en route are not a feature of the algorithms.
Platform change is not realistically representable with the existing algorithms, since it works
with the above-mentioned principle used to change the route. Change of platform forces
all passengers to move, often just some minutes before train departure. Sometimes this is
practicable and sometimes it is not, depending on a number of factors including passenger
numbers and distance between platforms. No current simulator package can fully take this
into account.
Since turnround times at terminus stations or where trains reverse are often quite long for
long-distance services, the platform change of one train can cause another change to a
second train scheduled to use that platform. As a result, during the day, a high number of
platform changes follow others in a sort of propagation, again often not accurately reflecting
reality. No exact rule exists that leads to decide a platform change, and in the simulation it can
be either performed anyway if the pre-defined track is occupied or totally deactivated. As a
result, the pre-defined tracks are used anyway or tracks are very often changed unrealistically.
Case studies have shown that realistic results in terms of delay per train can often be obtained
using the pre-defined platforms, although the real track allocation process is not reproduced.
It can be concluded that much more work is needed to improve dispatching algorithms in
simulation models (as it is to produce really effective traffic management systems for
incorporation in “real time” operations).

Effectively modelling seriously delayed trains

Seriously delayed trains represent a particular challenge for dispatchers, since a de facto new
slot has to be inserted, also routes found within nodes and large stations and with the lowest
impact on other trains. Moreover, if such trains are freight services, overtaking may have to
be managed. If they are long-distance passenger services, platforms have to be allocated at
terminal stations, and it is necessary to decide whether the train set maintains the planned
turnaround (delaying the outbound service) or not.
From the simulation point of view, the problem appears as a combination of the above issues
regarding complex dispatching on lines and platform allocation. With the lookforward rules
as currently applied in many simulation packages, it is impossible to ensure that simulation

211
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

represents such phenomena without generating inappropriately high waiting times for other
services.

Modelling major disruptions

When delays become too great, or in event of a major train breakdown or infrastructure failure,
in real life dispatchers might decide to cancel one or more services or let them terminate short.
Obviously, cancelling a service will imply re-programming the corresponding rolling stock and
personnel scheduling, which would require a dedicated algorithm. At present, simulators do
not cancel services automatically, often leading to unrealistically high output delays when
simulating significant initial delays or disruptions.
The normal variability of process times can be inserted easily in simulation using real collected
data as previously described. The simulation is then repeated a significant number of times,
with random choice of the input parameters based on the input delay distributions. Far more
complex is the definition of failures and other significant distributions due to the conceptual
complexity of defining an important failure scenario, given the relatively low frequency of similar
events and the dissimilar impact they might have. Moreover, once a scenario is defined, the
complex decisions required to reschedule trains are very difficult to be made by a simulator
automatically, and currently must therefore be inserted by the user. This requires a significant
amount of both time and experience.
A further issue with simulating major disruptions is that rolling stock and train crew reallocation
become critical parts of the problem. No microscopic simulations currently deal with these and
ignoring them can lead to simulators significantly understating the impact on performance of
particular scenarios.
As major disruptions represent a significant proportion of train lateness, this seriously limits
the usefulness of microscopic simulation. This represents a key area for development of
microscopic simulation packages if they are to meet more fully the needs of railway operators
and planners.

Complexity of use
The most modern simulation tools show a quite intuitive and user-friendly interface, which also
allows non-skilled users to perform a simulation. However, a very high level of understanding is
needed of the principles of simulation, the specific functionality of the software being used, the
signalling system being modelled and the practical day to day operation of the railway network
being simulated. This will ensure the accuracy of the model in reproducing a certain real
network. This includes for example representing correctly all details of interlocking, signalling
and ATP/ATC systems, or assigning the right priorities to train categories and individual train
services. The modellers who provide the best results from simulation are those who use the
software on an almost daily basis and become experts over several years, whilst also taking
the time to learn about practical railway operation. There are relatively few people of this
calibre.
Another key issue is the complexity and hence time taken to set up and run microscopic
simulations. Where models have been used for some time (e. g. RailSys in Germany and the
UK, OpenTrack in Switzerland and Italy) complete national infrastructure models have been
developed. This speeds the preparation of models, but careful checking that the models are
up to date at all times and of a high enough quality still needs to be undertaken. Interfaces
with other systems can speed the input of timetables but again often a considerable amount
of checking is required as the geographic details in different systems vary. There is not a

212
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

one to one relationship. This is a particular problem where timetables being imported into
the simulation package do not have enough detail (for instance of the specific tracks taken
through junctions or approaching major stations).

Using optimisation and simulation together

The sophisticated dispatch algorithms being developed to better replicate decisions made
by signalers and controllers are in fact optimisation routines. It can be anticipated that more
optimisation will gradually be built into simulators, including finding rolling stock and train crew
solutions that match with the simulated train lateness and making decisions about cancelling
trains or major rerouting. For more information on rescheduling models see Chapter 13.

Using macro- and microscopic simulation together

To facilitate modelling very large geographic areas, links are being made between macro- and
microscopic simulation. Cui & Martin (2011) discuss how this can be achieved and Huber &
Wilfinger (2006) provide a practical example of how this can be achieved, between Nemo
and RailSys, while Schlechte et al. (2011) demonstrate the combined use of OpenTrack and
optimal macroscopic network capacity allocation.

10.7 Conclusions

Simulation provides a very important tool for understanding railway operations and further
traffic growth will make it more so. Simulation provides the opportunity to understand the
performance implications of changes to the railway system in advance of implementing them.
Whilst its use is well established, considerable care needs to be taken to ensure that good
quality results are obtained and there are a number of important limitations which warrant
further research and development.

References
AEA Technology Rail (2002): MERIT Manual, Unpublished user manual, London.
Beck R., Scheier B., Jäger B. (2008): RAILONOMICS – determining investment strategies for
railway signalling through simulation. In: Allan, J. et al. (eds.), Computers in Railways XI, WIT
Press, Southampton, pp. 65 – 74.
Besinovic N., Quaglietta E., Goverde, R. M. P. (2013): A simulation-based optimization
approach for the calibration of dynamic train speed profiles. Journal of Rail Transport Planning
& Management 3 (2013).
Brown S. (2002): Railway network Simulation as a Systems Engineering Tool, MSc thesis.
University of Sheffield.
Brünger O. (1996): Konzeption einer Rechnerunterstützung für die Feinkonstruktion von
Eisenbahnfahrplänen. RWTH Aachen, PhD Thesis.
Buchmueller S., Weidmann U. et al. (2008): Development of a dwell time calculation model for
timetable planning. In: Allan J., Arias E., Brebbia C. A. et al., Computers in Railways XI, WIT
Press, Southampton, pp. 525 – 534.
Büker Th. (2010): Ausgewählte Aspekte der Verspätungsfortpflanzung in Netzen. RWTH
Aachen, PhD thesis.

213
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
10 Simulation

Bueker T., Seybold B. (2012): Stochastic modelling of delay propagation in large networks.
Journal of Rail Transport Planning & Management, 2(1–2): pp. 34 – 50.
Corman F., D’Ariano A., Longo G., Medeossi G. (2010): Robustness of advanced train
dispatching solutions under disturbances. Proceedings of the 12th World Conference on
Transport Research Society. Lisbon, July 11–15, 2010. (Paper 02816).
Cui Y., Martin U. (2011): Multi-Scale Simulation in Railway Planning and Operation. Traffic and
Transportation, Vol. 23, No. 6, pp. 511 – 517.
de Fabris S., Longo G. et al. (2008): Automated analysis of train event recorder data to
improve micro-simulation models. In: J. Allan, E. Arias, C. A. Brebbia et al., Computers in
Railways XI, WIT Press, Southampton, pp. 575 – 585.
EPFL – Swiss Federal Institute of Technology, Lausanne: Fasta – Stability analysis of timetables
for highly interconnected rail networks. http://www.fasta.ch. Version: 04.04.2013.
Gamerman D., Lopes H. F. (2006): Markov Chain Monte Carlo: Stochastic Simulation for
Bayesian Inference. Chapman and Hall, Boca Raton, USA.
Groeger T. (2002): Simulation der Fahrplanerstellung auf der Basis einen hierarchischen
Trassenmanagements und Nachweis der Stabilitaet der Betriebsabwicklung.
Veroeffentlichungen des Verkehrswissenschaftlichen Institutes der RWTH Aachen, PhD Thesis.
Hauptmann D. (2000): Automatische und diskriminierungsfreie Ermittlung von Fahrplantrassen
in beliebig großen Netzen spurgeführter Verkehrssysteme. Hamburg, Eurailpress, PhD Thesis.
Huber H-P., Wilfinger G. (2006): Integration-Enhancements for Microscopic and Macroscopic
Railway Infrastructure Planning Models. World Conference of Railway Research. Montreal,
Canada.
IVE (2013): NEMO – Netz Evaluations Modell. Ingenieurgesellschaft für Verkehrs- und
Eisenbahnwesen mbH, Hannover.
Jacobs J. (2003): Rechnerunterstützte Konfliktermittlung und Entscheidungsunterstützung bei
der Disposition des Zuglaufs. RWTH Aachen, PhD Thesis.
Jones G. T. (1972): Simulation and Business Decisions, Penguin Education, Harmondsworth,
Middlesex, England.
Kirkwood D., Roberts D., Schmid F. (2013): Validation of Railway Network Simulation Software.
In: Hansen I. A. et al. (eds.), Proceedings of 5th International Seminar on Railway Operations
Modeling and Analysis, Technical University of Denmark.
Kuckelberg A. (2011): Mikroskopische Disposition spurgebundener Verkehrsmittel unter
Echtzeitbedingungen. RWTH Aachen, PhD Thesis.
Lienau, C. (2007): Abbindung von Infrastrukturkosten in der Eisenbahnbetriebssimulation.
Hamburg, PhD Thesis.
Longo G., Medeossi G. (2012): Enhancing timetable planning with stochastic dwell time
modelling. In Computers in Railways XIII, WIT Press, Southampton, pp. 461 – 471.
Luethi M., Medeossi G. et al. (2007): Evaluation of an Integrated Real-Time Rescheduling
System for Railway Networks. 2nd International Seminar on Railway Operations Modelling and
Analysis (RailHannover 2007). Hannover.
Maricaua E., Scheerlinckb K., Tassenoya S., Büker T., Verboven S. (2013): Evaluating
the robustness of a railway timetable: a practical approach. In: Hansen I. A. et al. (eds.),
Proceedings of 5th International Seminar on Railway Operations Modeling and Analysis,
Technical University of Denmark.

214
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

Medeossi G., Longo G., de Fabris S. (2011): A method for using stochastic blocking times
to improve timetable planning. Journal of Rail Transport Planning & Management 1 (2011),
pp. 1 – 13.
Medeossi G. (2010): Capacity and Reliability on Railway Networks, A Simulative Approach.
University of Trieste, PhD Thesis.
Middelkoop D., Bouwman M. (2001): Simone: large scale train network simulations.
Proceedings of the Winter 2001 Simulation Conference, pp. 1042–1047.
Oetting A., Rittner M., Fey S. (2013): Synchronal algorithms for determination and evaluation
of conflict resolution scenarios for real-time rescheduling. In: Hansen, I.A., et al. (eds.),
Proceedings of 5th International Seminar on Railway Operations Modeling and Analysis,
Technical University of Denmark.
ON-TIME (2013): Task 3.1 Assessment of State-of-Art of Train Timetabling. Available on line at
www.ontime-project.eu.
Pidd M. (2004): Computer Simulation in Management Science. Wiley, Chichester, UK.
RMCon (2008): RailSys – Network wide Timetable and Infrastructure Management Integrated
network wide Operational and Timetable Simulation. R. M. Consultants, Hannover.
Robinson S. (1994): Successful Simulation: A Practical Approach to Simulation Projects.
McGraw Hill, London.
Rudolph R., Demnitz J. (2003): Simulation of large railway networks. Robustness test for the
timetable in NorthRhine Westphalia. Proceedings of World Conference on Railway Research
2003. Edinburgh: pp. 644 – 652.
Selia A. F., Ceric V., Tadikamalla P. (2003): Applied Simulation Modeling. Thomson, Belmont,
USA.
Schlechte T. et al. (2011): Micro–macro transformation of railway networks. Journal of Rail
Transport Planning & Management, Volume 1, pp. 38 – 48.
Stephan A. (2008): OpenPowerNet – simulation of railway power supply systems. In: Allan, J.
et al. (eds.), Computers in Railways XI, WIT Press, Southampton, pp. 449 – 458.
Vakhtel S. (2002): Rechnerunterstützte analytische Ermittlung der Kapazität von
Eisenbahnnetzen. RWTH Aachen, PhD Thesis.
Yuan J., Hansen I. A. (2007): Optimizing capacity utilization – stations by estimating knock-on
train delays. Transportation Research Part B vol. 41, pp. 202–217.
Watson, R. P. (2008): Train Planning in a Fragmented Railway – A British Perspective.
Loughborough University, PhD Thesis.
Weymann F. (2011): Qualität von Heuristiken in der Disposition des Eisenbahnbetriebs, RWTH
Aachen, PhD thesis.

215
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 11.1 Introduction

11 Statistical Analysis of Train Delays and Movements

Jianxin Yuan, Giorgio Medeossi

11.1 Introduction

Passenger train services in most countries are operated according to a planned timetable.
However, train running times between stations and dwell times at stations vary a lot,
depending on many factors from within the railway system and from exogenous sources.
There will always be a certain number of operational processes that exceed the scheduled
times. Therefore, some train delays are inevitable in real operations.
A detailed delay analysis enables the identification of the influencing factors, which helps
operators improve the level of train punctuality. Harris (1992) used linear regressions to get a
better understanding of the punctuality of train services in the UK and the Netherlands. Based
on the analysis results, he proposed to ensure sufficient time supplements for long-distance
services, and to enhance demand management for short-distance services.
Lin & Wilson (1992) established dwell time regression models for single car and two-car light
rail operations on the basis of data gathered on the Green Line of the Massachusetts Bay
Transport Authority. The resulting models show that both the numbers of passengers boarding
and alighting and the level of passenger crowding onboard the trains significantly affect the
dwell times. Evidence was also found for a non-linear increase of the marginal delay with
the number of standees. Important differences exist between dwell time estimation models
for one-car and two-car trains as a result of typically uneven distributions of passenger
movements and passenger loads between cars in a two-car train.
Goverde & Hansen (2000) presented a data mining tool TNV-Prepare based on train describer
records and modelled the dependence of the departure delays of a pair of interconnected
trains on their arrival delays. This was on the basis of train detection data recorded by the
Dutch railways (2001). Olsson and Haugland (2004) calculated the correlation coefficients of
a number of selected factors for the punctuality of trains in Norway. The calculation results
reveal that the rate of capacity utilisation determined by a timetable has a significant impact
on the punctuality of train operations, but it alone cannot explain all the variations in the
punctuality during a day. Additionally, it was found that a key success factor for punctuality
of local and regional trains in congested areas is the management of boarding and alighting
passengers.
Furthermore, Yuan (2006) analysed route conflicts, propagation of train delays between train
pairs and the impact on the level of train punctuality. Goverde & Meng (2011) introduced the
add-on analysis tool TNV-Statistics that automatically determines chains of train routes with
associated knock-on delays and rankings of signals according to number of conflicts, time
loss or delay jump. This information is used to identify and analyse structural and serious route
conflicts due to timetable flaws or capacity bottlenecks automatically.
In this Chapter, statistical analysis of train delays and movements is focussed on enabling a
deep insight into their variability. Based on this, suitable strategies can be applied to reduce
delays, improve punctuality and save energy. Moreover, the statistical analysis will result in
appropriate distribution models of train delays and a detailed representation of stochastic
running times for precisely identifying potential track conflicts and estimating the propagation
of train delays with either analytical models or simulation tools. This precise identification
and estimation is always required in the evaluation of timetable robustness and rescheduling
optimality.

217
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
11 Statistical Analysis of Train Delays and Movements

This Chapter is structured as follows. At first, a number of important terms widely used in the
field of railway operations research are outlined. Data collection, a prerequisite for statistical
analysis of train delays, running times and dwell times, is then introduced. Next, the optimal
fitting of random train delays to a probability distribution model is discussed. Afterwards,
the Authors present newly developed methods for estimating stochastic running times
between successive stations. At the end of this Chapter, there is a discussion concerning the
application of the methods presented.

11.2 Train delays, running times and dwell times

A train delay is the deviation from either a scheduled event or process time of this train. To
evaluate the performance of a railway system, aggregate measures such as punctuality are
generally applied. The punctuality of trains is usually defined as the percentage of the trains
that arrive at (depart from, pass) a location with a delay less than a certain time in minutes, see
Hansen (2001). For most European railway companies, trains that arrive less than five minutes
late are not considered to be delayed. In Switzerland, however, the threshold value for defining
the punctuality of train operations is four minutes, whereas this value is even limited to 10 to
15 seconds in Japan. The punctuality of train operations may be measured not only at terminal
stations, but also at intermediate stations and other timetable reference points.
Punctuality is an important measure of the performance of train operations, but it gives
only limited information about the extent of train delays. Thus a delay of five minutes can
be regarded as bad as a delay of 20 minutes. Furthermore, the delays which do not reach
the threshold value are not considered as delays in reports published by train operating
companies or infrastructure managers, while their impact on the quality of train operations is
considerable, see Hansen (2001); Goverde (2005); Yuan (2006). A comprehensive punctuality
analysis should be performed by including not only the level of train punctuality, but also the
mean and standard deviation of the delays. Apart from that, the percentage of scheduled trips
completed is a measure of the reliability of operations.
Train delays at stations may be referred to both arrival and departure events. The arrival delays
of trains can be negative, corresponding to early arrivals, since a certain amount of slack time
is generally incorporated in the scheduled running and dwell times. Early arrivals do not affect
the level of train punctuality, but they may lead to a less efficient use of the track capacity due
to the extended occupation of platform tracks. Departure delays are generally non-negative,
since trains are not allowed to depart from a station earlier than scheduled.
According to the locations and sources of generation, train delays can be classified as initial,
original and knock-on delays, see Carey & Kwieciński (1994); Hansen (2004). Initial delays
are those recorded first at the cordon of the network being investigated, when a train enters.
Original delays occur within the network due to technical failures, running at lower speeds
than scheduled, prolonged alighting and boarding times of passengers, and bad weather
conditions. When a train is delayed, it may hinder other trains by still occupying (part of) the
scheduled route and preventing other trains from crossing on single lines.
The delays transferred to other trains are called knock-on delays. Connecting trains at
interchange stations that are ready to depart also suffer knock-on delays if the trains that
feed them arrive late and the maximum scheduled waiting time allows holding the trains. The
effects of knock-on delays of trains reflect the stability of timetables and the reliability of train
operations in networks, depending on the size and distribution of the buffer times between
scheduled train pairs.

218
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 11.3 Data collection

When running between two stations, the movement of an unhindered train is a sequence of
the following modes: 1) acceleration; 2) cruising; 3) coasting; 4) braking. Conventional running
time calculators use a set of empirical parameters to solve the motion equation (Chapter 4).
These parameters have been measured over many years for different types of rolling stock,
see Wende (2003), and therefore it is possible to calculate very detailed train speed profiles.
The motion equation yields a precise running time. However, there are many non-deterministic
influences on running time such as electrical current supply, loading percentage, weather,
driver behavior and even different trains of the same series. These influences introduce
variability in running time. To cope with this variability, recovery times are inserted in timetables,
including implicitly an imprecision in the representation of train movements.
Dwell time of a train at a station consists mainly of two parts: the time needed for passengers
to alight and board the train and train dispatching time. Sometimes, coupling or uncoupling
time may need to be included. Furthermore, a part of buffer time may be added in the
timetable. The scheduled dwell times of trains are constant and must be determined carefully.
A tight dwell time is a source of delay, while long dwell time means extended travel time and a
low utilisation of platform track capacity.
To model the propagation of train delays in a railway network, initial delays are defined at
the boundaries of the network. Based on the distribution of the initial delays and those of
the original delays within the network, the distributions of knock-on delays suffered, the
distributions of arrival and departure delays at downstream stations in the network, as well as
the distribution of exit delays at the opposite boundary of the network can be estimated by a
delay propagation model, see Higgins & Kozan (1998); Carey & Carville (2000); Yuan & Hansen
(2007). The original delays of trains in a track section corresponds to the extended running/
dwell times excluding the knock-on effects. The initial and exit delays of trains at the cordons
of a network, as well as the arrival and departure delays of trains at the stations, may contain
both original and knock-on delays.

11.3 Data collection

Data are, of course, necessary for any statistical analysis of train delays. Manual records may
be used to collect data on train dwell times, as well as the characteristics of passenger flows
at stations, see Wiggenraad (2001); Daamen (2004). However, manual records of arrival and
departure times at stations are normally not very accurate. Automatic track occupancy and
release records from the existing signalling and safety system offer a very efficient and precise
manner to collect the data for analysing track blocking times and real capacity utilisation. GPS
and train event recorder data offers a continuous tracking of train movements and therefore
the possibility of reconstructing complete speed profiles. These can include the precise arrival
and departure times at stations. However, this is available only in modern rolling stock.

11.3.1 Train Describer Data

Train describer systems are important tools for efficient traffic management and provide the
essential backbone for traffic supervision, automatic train routing and passenger information
systems, see Bailey (1995). The train describer receives information regarding the position of
trains from the signalling system interlocking. Track circuits and/or axle counters detect the
presence of a train in a given section of a track. The train describer system holds a database
of the status of all the signalling information, containing train number berths. The database is
updated in real time, as train movements occur.

219
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
11 Statistical Analysis of Train Delays and Movements

In the Netherlands, the train describer, i. e. current TROTS (called TNV before) system keeps
track of the progress of trains at discrete steps along their routes. The daily log files contain
chronological information about all signalling and interlocking events in a traffic control area,
giving an accurate description of train movements with a maximal error of about a second.
However, track section related occupancy messages are not automatically matched with
individual train numbers.
Depending on the train describer and control systems, only the arrival and departure times
at stations or the occupancy and release times of all track sections may be stored by the
infrastructure managers and/or train operators. In the former case, the arrival and departure
times at a station are normally estimated using a correction term on the basis of the
occupancy and release times of the platform track sections. Moreover, the estimated arrival
and departure times are often rounded to 1/10 minute, to 30 seconds or even to the full
minute. Thus, the station passing data might include a combination of inaccurate estimation of
the arrival and departure itself and a rounding, leading to inaccurate statistics.
In the latter case, the train describer log-files are completely stored. Based on subsequent
analysis of the TNV-log files recorded by the Dutch train describer, Goverde & Hansen (2000)
developed the data mining tool TNV-Prepare. This couples the individual events of the
infrastructure elements to the corresponding train numbers. The TNV-Prepare output consists
of chronological TNV-tables per train line service and (sub) route. For each individual train,
the event times along the route are given, including train description steps, section entries
and clearances, signal changes, and direction switches of points. From this information the
other interesting process times, such as running times, blocking times, and headways can be
derived very precisely. A software package with a more user-friendly interface, the so-called
VTL-Tool was developed by the Dutch Railways, see Stam et al. (2007), which automatically
searches the route of a train in a yard from a number of available itineraries.
The arrival and departure times of trains at station platforms are not measured directly by any
train describer systems, as the measurement points are situated at the insulation joints rather
distant from the platforms. Hermann (1996) used train describer data of the railway network in
the area of Frankfurt for statistical delay analysis. He approximated the off-set times between
platforms and the closest measurement locations, i. e. home signals, by constant interpolation.
In Switzerland, station passing data collected by the train describer system is used, apart from
train path monitoring, for punctuality analysis by means of OpenTimetable, see Ullius (2004),
and even on-line checking of train positions accessible to passengers via the website of Swiss
Railway. To estimate the running times of trains between platforms and station home signals,
Swiss Railways use correction terms differentiated by route, activity, and train type. The
estimated running times are accurate with an absolute error of less than 20 seconds. Similar
correction terms are used in Italy and in Norway, where the TRENOAnalysis tool, see Longo
& Medeossi (2013), is deployed to analyse the operations, point out the weaknesses of the
timetable, choose the distributions to be used as input for the simulations and validate the
simulation models.
To estimate more accurately the running times of trains at station platforms based on data
generated by TNV-Prepare, Goverde (2005) developed the tool called TNV-Filter. In detail, the
speed trajectory of each train entering or departing from a station is estimated on the basis
of section occupancy and clearance times and the scheduled train characteristics. Estimation
errors due to rounding-off of train passing times and deceleration/acceleration variations
during the approach/departure are filtered by means of a nonlinear least-squares method,
taking into account the speed limits at, for instance, signals and switches. Subsequently,
the running times over the platform sections before and after coming to a standstill are

220
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 11.3 Data collection

estimated from the filtered inbound and outbound speed profiles, known standard maximum
deceleration and acceleration characteristics for the type of train, and the stop location at the
platform section (depending on the train length). The arrival, departure and dwell times of each
individual train, as well as the delays, are then calculated. The resulting running, arrival and
departure times, as well as the delay estimates are with a precision in the order of a second.

11.3.2 GPS and Train Event Recorder data

More complex and demanding in terms of computational effort required is the management
of GPS or train event recorder data, which offer a continuous tracking of the vehicle and
therefore the possibility of reconstructing its complete speed profile. This includes the precise
departure and arrival times. GPS devices are very widely used by road transport operators,
being a fundamental component of the Automatic Vehicle Location systems (AVL), while in rail
transport their usage is less common, since train describers already allow the location of trains
in real time. While this has limited the adoption of GPS devices in the former state-owned
national railways, in recent years most of the new operators installed them, since they allow
trains to be located quite accurately and independently from the services provided by the
infrastructure managers.
The analogue on-board recorders are gradually been replaced by digital Train Event Recorders
at most rail operators. A train event recorder (also called On-Train Monitoring Recorder (OTMR)
in the UK, Event Recorder System (ERS), Event Recorder Unit (ERU), or simply Event Recorder
(ER)) is a device that records data about the operation of train controls and performance
in response to those controls and other train control systems. It is similar to the flight data
recorder found on aircraft. Since they are integrated with most car-borne systems, they
allow enhanced diagnostic and control functions. For example they control penalty brake or
emergency brake systems, and speedometers.
Data storage is normally provided by non-volatile EEPROM or Flash Memory overwritten in
a FIFO continuous loop. The data is intended for use in the investigation of accidents and
incidents, but it is also used to monitor the performance of the train set, the competence of
drivers, and general state of the train over a period of time.
Since speedometers are very precise, train event recorders too store very accurate data.
Additionally, they provide a wide range of data including the speed limit calculated by the train
protection system, the signal aspects, the throttle and brake percentage use. These data do
not require any additional filtering, since tracking data are filtered in real time by comparing two
speedometers and using balises for discrete calibration of distances.
In contrast, GPS tracking data is significantly more difficult to prepare for use in the motion
equation calibration process. Medeossi et al. (2011) presented a procedure for filtering and
storing GPS data. The data preparation begins by post processing the collected tracking data
using a proven algorithm based on the Kalman filter, see Grower et al. (1997). The filtered data
are sufficiently precise, but do not contain any information on signal aspects, which is required
in order to reconstruct driver behaviour. To fill this gap, the recorded speed profile is compared
with the planned speed profile (in which the train brakes only for track speed reductions and at
stops). When unplanned braking actions are found, the planned speed profile is recalculated
by assuming that a yellow signal aspect is displayed at the corresponding signal.
On lines with discrete ATP, the braking curve based on the approach speed and distance to
next signal is computed, since the train is not allowed to re-accelerate before passing the
corresponding main signal balise. On lines equipped with continuous ATP, where the train is

221
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
11 Statistical Analysis of Train Delays and Movements

allowed to accelerate as soon as the driver-machine interface (DMI) shows the less restrictive
aspect again, the instant when the train re-accelerates is obtained by the local minimum of the
real speed profile. The planned speed profile is then modified to include the braking and re-
acceleration.

11.4 Distribution fitting of train delays

A negative exponential distribution of train delays was introduced first by Schwanhäußer

(1974), based on manual sample records by railway staff. More literature (Hermann, 1996;
Goverde et al., 2001; Yuan et al., 2002; Wendler & Naehrig, 2004; Goverde, 2005; Güttler,
2006; Yuan, 2006) exists on the analysis of delay distributions based on comprehensive
empirical operations data from automatic train describer systems. A variety of distribution
models have been used to capture the stochasticity of train delays. Yuan (2006) presented
an engineering approach for fine-tuning the parameters and assessed several probability
distribution models of train delays.
Distribution fitting of initial and original delays is required for estimating the distribution of the
knock-on delays and further the impact on train punctuality. The initial delays of trains at the
cordons of a railway network are the difference between the actual passing times recorded
and the scheduled ones. The original delays of trains in a certain track section cannot
be derived directly from track occupancy and release records, since the running and dwell
times of the trains may include knock-on delays. However, by comparing the passing time
of each approaching train at the distant signal of the block signal with the signal block and
route clearance time of the preceding train, the knock-on delay of the following train can be
identified.
If no knock-on delay occurs, the original delay can be computed as the difference between the
free running time and the scheduled one. In such a manner, the original delay of a train during
a dwell process can also be estimated. Note that it is difficult to derive the original delays of
every train in every individual track section, because some trains may suffer knock-on delays in
a certain track section even though longer headway times are scheduled. Fortunately, it is not
necessary to obtain all the original delays for every train in every track section in order to model
the delays with a standard probability distribution.
Having collected a number of data sets for statistical analysis of a type of delay, it is important
to check the homogeneity of the data. This check can be done for different train types,
different days of the week, and between peak and off-peak periods. The equivalence of the
distributions can be diagnosed statistically, e. g. by means of two-sample Kolmogorov-Smirnov
goodness-of-fit test, see Gibbons (1985); Mathsoft (1999). The data sets accepted to be
homogeneous by a statistical test can be combined. Based on a detailed statistical analysis
of train delays, Yuan (2006) revealed that even the original delays of trains at a station, i. e.
the excess of the scheduled dwell time in the absence of conflicts with other trains, depends
much on the arrival delays of the trains. The data set of the original delays of a train series at a
station should be split into at least two subsets, corresponding to both early and late arriving
trains, which are to be analysed separately.
For a statistical analysis of train delays, the sample size of a data set to be analysed must be
taken into account carefully. If the sample size of a data set is not big enough, the parameters
of a distribution model of the data cannot be estimated with sufficient accuracy. Moreover,
goodness-of-fit tests are not very sensitive to the disagreements between a fitted distribution
and the data in case of a small to moderate sample size. Even the chi-square goodness-of-fit

222
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 11.4 Distribution fitting of train delays

test is not valid in the case of a small sample size of for example less than 20, see Banks et al.
(2004).
This section, focusses on the optimal fitting of train delays to a probability distribution model. It
is first explained how to select suitable distribution models that may capture the real variability
of train delays. Two common statistical methods for estimating the parameters of a distribution
model are then introduced. Furthermore, standard methods are presented to the testing of the
goodness-of-fit of a hypothetical distribution. Finally, the Authors centre on the comparison of
a number of candidate distributions and the optimal fitting of train delays to a model by fine-
tuning the parameters.

11.4.1 Selection of statistical distributions

An appropriate distribution model of a data set can be chosen on the basis of the physical
properties of the corresponding random variable. However, it is often difficult to know all the
properties when the variable represents a result of a lot of complicated processes influenced
by human behavior, as in the case of railway operations. The distribution of a data set can also
be selected by inspecting the data. In this case, a histogram is a useful representation, since
it shows the shape of the distribution. Of course, it is best to use both approaches to ensure
that they are in agreement.
Train event times and process times generally have a lower bound, i. e. the earliest or shortest
time. The histograms are often skewed to the right, especially in the case of train event times,
which are subject to an accumulation of previous delays. Therefore, the (negative) exponential
distribution is often considered to be a valid model for train delays. Schwanhäußer (1974) is
probably the first who applied the exponential distribution to approximate the variability of
positive arrival delays. This distribution has been validated later on by more comprehensive
statistical analyses, see Goverde et al. (2001); Yuan et al. (2002); Wendler & Naehrig (2004).
The exponential distribution is also used to model departure delays, see Goverde (2005) and
original delays, see Ferreira & Higgins (1996).
Because of some slack times inserted in scheduled running and dwell times, as well as the
response of train crew, the peak of a delay histogram is generally not located at the minimal
value. For positive arrival delays and departure delays, the histogram peaks may be located at
zero, but the shape of the histograms is often steeper than that of the exponential distribution.
To better capture the variability of train delays, more flexible distribution models, e. g. the
gamma, Weibull and lognormal distributions, have also been applied, see Steckel (1991);
Hermann (1996); Yuan (2006). When the minimal value of a type of train delays is no-zero, a
location shift parameter has to be incorporated in the distribution models.
The arrival delays of trains at stations are subject to the alteration of train orders in real
operations, therefore the statistical histograms might have a heavy tail on the right. In this
case, a mixture distribution represents better the delay variability. Güttler (2006) adopted
the normal-lognormal mixture distribution to model stochastic running times between two
successive stations using empirical data from the German Railways. In case of a relatively
high probability of route conflicts, a mixture distribution can well represent the variability of
the resulting heterogeneous running times. However, finite mixtures will introduce additional
complexity because they have more parameters, whose estimation usually requires solving
a complex system of non-linear equations. Moreover, schedulers generally prefer to use
deterministic values or as simple as possible distribution models.

223
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
11 Statistical Analysis of Train Delays and Movements

Since a weighted exponential distribution was proposed by Schwanhäußer (1974), it has been
widely used to model arrival delays, see Meng (1991); Ullius (2004); Wendler & Naehri (2004);
Herrmann (2006). There, a parameter between 0 and 1 is introduced to represent the share of
delayed trains, while the probability of on-time arrivals is reflected with 1 minus this parameter.
The positive delays are further modelled by the exponential distribution. This weighted
exponential model is a special type of mixture distribution, which eases the modelling of the
propagation of train delays by an analytical approach, see Weigand (1981); Mühlhans (1990).

11.4.2 Common approaches to parameter estimation

Given a distribution model of train delays, the parameters can, of course, be specified
by experience only. However, it is not easy even for experienced professionals to make an
accurate specification. Therefore, the parameters of a delay distribution are preferably
estimated on the basis of empirical data. The common approaches to parameter estimation
include the method of moments and maximum likelihood method.
The method of moments is carried out by equating sample moments with population
moments and then solving those equations for the distribution parameters to be estimated.
The maximum likelihood method is based on a likelihood function of sample data. The former
method has the advantage of simplicity, but the parameter estimates fail to take into account
all the relevant information in a sample data. The latter method is used more widely because
Maximum Likelihood Estimator (MLE) is an unbiased estimator with the minimum variance as
the sample size increases, see Dekking et al. (2005).
The parameter estimates of a statistical distribution are influenced significantly by the data
outliers, which are the data points deviating much from the bulk of the data. A graphical
display called box-plot (Mathsoft, 1999) can be applied to detect the outliers of a data set
by showing a measure of the median, the dispersion, and the presence of possible outliers.
However, there exists no exact standard procedure in this aspect.

11.4.3 Testing the goodness-of-fit

After selecting one or more distribution models that may capture the stochasticity of train
delays, these distributions can be investigated to see how well they represent the true
underlying distribution of the data. Law & Kelton (2000) presented five main graphical
approaches for assessing the goodness-of-fit of a hypothetical distribution. These approaches
include density/histogram overplot, frequency comparison, distribution function differences
plot, probability-probability plot and quantile-quantile plot. For the first two approaches,
a histogram of the data being analysed is influenced by the number of the pre-specified
intervals. Therefore, it may be better to compare the density curve of a hypothetical distribution
with a kernel density estimate (Dekking et al., 2005) of the data. Graphical approaches give a
good guide as to the goodness-of-fit and enable to weed out poor candidate distributions.
Following a graphical analysis, it is important to perform a formal goodness-of-fit test to
assess whether a data set comes from an assumed distribution. The oldest goodness-of-fit
test is the chi-square test (Law & Kelton, 2000), which can be thought of as a more formal
comparison of a histogram with the fitted density function. Having divided the data into k bins,
the chi-square test statistic is defined as
k (Oi – Ei )2
2 =  (11.1)
i=1 Ei
where Oi and Ei are the observed and expected frequencies for bin i, respectively.

224
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 11.4 Distribution fitting of train delays

First, suppose that all parameters of a distribution have been specified without making use of
the data. The chi-square test statistic  2 converges then in distribution (as sample size n g )
to the chi-square distribution with k – 1 degrees of freedom if the null hypothesis is true. For
large n, the chi-square test with a significance level  is obtained approximately by rejecting
the hypothetical distribution if the realization of the chi-square statistic  20 exceeds the upper
1 –  critical point  2k-1,1- of the chi-square distribution with k – 1 degrees of freedom. In other
words, the hypothetical distribution is rejected, if the p-value of the chi-square test, which is
the probability of  2   20 , is smaller than the defined significance level.
Second, assume that the parameters of a distribution have been estimated from the data
directly. If the null hypothesis is true, then as n g  the distribution of  2 converges to a
distribution that lies between the chi-square distributions with k – 1 and k – m –1 degrees
of freedom (m represents the number of estimated parameters). If  21- were the upper 1 – 
critical point of the asymptotic distribution of  2, then
 2k-m-1,1-   21-   2k-1,1- (11.2)

It is often recommended that the hypothetical distribution be rejected only if  02   2k-1,1-,

since this is conservative (the rejection area is smaller than necessary). This choice, however,
will entail loss of power (probability of rejecting a false null hypothesis) of the test.
Another commonly used goodness-of-fit test is the Kolmogorov-Smirnov (K-S) test (Gibbons,
1985), which compares a hypothetical distribution function with the empirical (sample) one.
Let F (x) and Sn (x) be the hypothetical and empirical distribution functions, respectively. The
statistic of the K-S test is defined as
Dn = sup | F(x) – Sn(x) | (11.3)
x

Suppose there are r different values of observations for some r  n and these values are
sorted such that x1  x2  …  xr . The K-S statistic can be formulated by

Dn = max { max [  F(xi ) – Sn (xi ) , F (xi ) – Sn (xi-1 ) | ] } (11.4)

1ir

If all observations are different, the above equation reduces to

i i–1
Dn = max { max [  F(xi ) –  , F (xi ) – |]} (11.5)
1in n n

The K-S test is to reject the hypothetical distribution if Dn exceeds some constant dn,1-,
which is the upper 1 –  critical point of the distribution of Dn. If none of the parameters of the
hypothetical distribution are estimated from the data, a generic critical point dn,1- is available
for all continuous distribution types. If the distribution parameters are estimated from the
data directly, the critical point dn,1- can only be approximated, depending on the type of the
distribution.
The aforementioned goodness-of-fit tests have their own advantages and drawbacks.
Perhaps the main advantage of the chi-square test is that when unknown parameters must
be estimated from the data, a generic correction can be introduced in the statistic distribution
by reducing the number of degrees of freedom. However, the K-S test has several advantages
over the chi-square test, too. First, the K-S test does not require grouping the data in any way,
so no information is lost; this also eliminates the troublesome problem of interval specification
in the chi-square test. Furthermore, the K-S test is exact for any sample size and any
continuous distribution model when the parameters are not estimated from the data directly,
whereas the chi-square test is valid only in an asymptotic sense.

225
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
11 Statistical Analysis of Train Delays and Movements

11.4.4 Assessing distribution models by fine-tuning the parameters

For a data set of train delays, several distribution models may be accepted by goodness-of-
fit tests. It is interesting to know which distribution matches best with the data. For a number
of continuous distributions, if the parameters are specified without making use of the data
directly, the K-S test can be applied to rank the goodness-of-fit of those distribution models
accurately. However, if the parameters are specified grossly, those models may often be
rejected especially in case of a large sample size of the data.
Yuan (2006) presented a new engineering approach for fine-tuning the parameters of several
distribution models based on real-world data. The models include the normal, uniform, beta,
gamma, Weibull and lognormal distributions, most of which are applied widely in the literature.
The exponential distribution, which is a special type of the gamma and Weibull distributions,
has been taken into account, too. To deal with the impact of data outliers on the estimation
of distribution parameters, a heuristic procedure is incorporated in the parameter fine-tuning
approach. Based on the parameter refinement, the goodness-of-fit of those candidate
distribution models is further compared.
To assess those distribution models for a data set of train delays, the data set is first split into
two parts randomly, e. g. by assigning chronologically tabled train delays alternately into the
subsets. The parameters of each candidate distribution model are fine-tuned in the following
steps:
1. Obtain an initial estimate of the distribution parameters using the MLE based on one of the
split data subset;
2. Omit the outliers, i. e. spread large delays in this data subset iteratively one by one
estimating the parameters correspondingly using the MLE. The fitted distributions are
assessed with the K-S test where the empirical distribution is, however, obtained from the
other data subset. The iterative procedure terminates if the p-value of the K-S test cannot
be increased any more;
3. Locally fine-tune the distribution parameters again to further improve the estimation by
maximising the p-value of the K-S test, where of course the empirical distribution remains
unchanged.
For the split data subsets, the sample size has been halved. However, the critical value of the
K-S statistic can be estimated accurately and be applied to all the distribution models. This
enables an accurate assessment of those candidate distribution models. Randomly splitting a
data set into two subsets also ensures that a distribution fitted well to one of the split subsets
is the distribution we need for capturing the variability of the original data set.
To compare with those candidate distribution models, the K-S test is applied again. Each
model is ranked according to the p-value of the K-S test where the hypothetical distribution is
specified with the fine-tuned parameters. In detail, the distribution with the biggest p-value is
ranked the first and considered to be the best one among the candidate models.
Having analysed more than 10.000 track occupancy and release records measured at the
Dutch railway station The Hague HS, Yuan (2006) found that the lognormal distribution is in
most cases the best model among the candidate distributions for both initial delays at cordons
of the local network and arrival delays at the case station. The Weibull distribution appears
to be the best distribution model for positive arrival delays, departure delays and the original
delays during the dwell processes of most train series per direction at the station. These
findings have also been confirmed with a statistical analysis of recently recorded TROTS data,
see Kecman & Goverde (2012).

226
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 11.4 Distribution fitting of train delays

The best fitted lognormal and gamma density curves, kernel density estimate and empirical
histogram for the arrival delays of the northbound intercity trains IC2100N at the platform track
in The Hague HS station are shown in Fig. 11.1. The lognormal model appears better than
the gamma model for fitting a distribution to a data set with a right-skewed density estimate
including a steep peak and a long tail.

Fig. 11.1: Lognormal and gamma density fits, kernel density estimate and histogram of the arrival
delays of IC2100N at the platform track in The Hague HS (March – May 2010)

Fig. 11.2 shows the density curves of the best distribution fits, i.e. the Weibull and gamma
distributions and the empirical histogram for the departure delays of the southbound intercity
trains IC2100S. It appears that the Weibull fit matches with the histogram better than the
gamma fit. It has been found that the shape parameter of a Weibull distribution fit to positive
arrival delays and departure delays is generally between 0.5 and 1.5. If the estimated shape
parameter is around 1.0, the exponential distribution could be used as the best model.
Fig. 11.3 shows the density curves of the best fits, i. e. the Weibull and normal distributions,
kernel density estimate and empirical histogram for the dwell times of the northbound intercity
trains IC2200N which have an arrival delay greater than 30 seconds. The fitted Weibull
distribution matches with the kernel density estimate for the empirical data much better than
the normal fit with respect to overall shape and the tails of the distribution. Herewith, the
considered trains are in the absence of conflicts with other trains. Therefore, the differences
between the dwell times and the scheduled one are the corresponding original delays at the
platform track.

227
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
11 Statistical Analysis of Train Delays and Movements

Fig. 11.2: Fitted Weibull and gamma density curves and histogram of the departure delays of
IC2100S at the platform track in The Hague HS (March – May 2010)

The empirical histogram and kernel density estimate of the dwell times of late arriving trains are
generally right-skewed. These dwell times may be shorter than scheduled, but they are always
longer than the minimal time for opening the doors of a train, announcing the departure, and
closing the doors. As a result, the peak of an estimated density curve is not located at the
minimal value and the shape parameter of the Weibull fit is generally larger than 1.0.
Although the aforementioned distribution models are identified for the delays of trains at
the Dutch railway station The Hague HS, they are applicable, too, for the delays of trains in
other station areas. This theoretical expectation is based on the flexibility of those distribution
models. The statistical parameters of those distributions, of course, depend on the local
infrastructure and signaling constraints, timetable, rolling stock and possibly behavioral
characteristics.

228
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 11.5 Estimation of stochastic train running times

Fig. 11.3: Weibull and normal density fits, kernel density estimate and histogram of the dwell times
of late arriving trains of IC2200N at The Hague HS (March – May 2010)

11.5 Estimation of stochastic train running times

While a deterministic running time calculation is sufficient to plan timetables, a more detailed
representation is required in micro-simulation and for ex-ante evaluation of timetable
robustness, which aim at reproducing train behaviour with the highest detail. Two different
sources of data can be used to estimate the running times: the continuous train tracking
collected by means of GPS and train event recorders, as proposed by Medeossi et al. (2011),
or the occupation data collected by train describer systems at each track circuit as proposed
by Besinovic et al. (2013).
The disadvantage of the latter solution is that the speed profile of the train must be
reconstructed, adding implicitly an uncertainty in the estimation. However, train describer data
are remarkably easier to manage, are collected automatically for all trains and are normally
made available more easily by the train operators. However, these data are still not collected
by every train operator; on-board digital train event recorders are used only in a few countries,
while in others only the arrival and departure times at stations are collected by conventional
train describer systems.
Steckel (1991) developed a method for introducing randomly distributed train data into the
movement equation to overcome the weaknesses of deterministic running time calculation in
coping with stochastic processes. While this approach introduced probabilistic running time

229
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
11 Statistical Analysis of Train Delays and Movements

calculation, its complex method for estimating input data and lack of a simple ex-post data
validation method limited its further development.
In principal, it is possible to model any parameter of the movement equation as a stochastic
factor. However it appears easier to describe the variation of the train performance by a set
of additional parameters than through calibration of the parameters of the train resistance
equation itself. Medeossi et al. (2012) proposed inserting a set of stochastic performance
factors to the movement equation, one for each movement phase. The performance factors
are calibrated based on real data, while other parameters remain unchanged.
More recently, Besinovic et al. (2012), introduced a method to derive the most probable
speed profiles of train runs from observed track occupation/release data. The conventional
movement equation is used to represent the train, and a simulation-based optimisation
approach is adopted to calibrate input parameters of the equations describing the tractive
effort, the rolling resistance, the braking effort, and the cruising phase. These parameters
are finetuned by minimising the gap between observed and simulated running times, using
a genetic algorithm. Additionally, due to the unknown individual train composition, a model
for the estimation of the train length was developed. The approach is applied to trains on the
corridor Rotterdam-Delft in the Netherlands. The probability distributions of the different input
parameters for each class of train compositions have been calibrated based on a significant
set of observed train runs.

11.5.1 Calibration using on-board collected data and performance factors

This section presents briefly the method introduced by Medeossi et al. (2012) to calibrate the
running times. Since a single performance factor does not allow a precise representation of the
different movement phases, see de Fabris et al. (2008), separated parameters for acceleration,
cruising, coasting and braking are used. The following performance factors are inserted:
– During acceleration, the tractive effort is multiplied by a performance factor to simulate the
difference between theoretical and real acceleration, for example due to driving style or
unique train performance characteristics.
– During cruising, the speed limit is multiplied by a performance factor to simulate the
difference between actual speed and the speed limit.
– Coasting takes place before braking, when the throttle is de-activated and therefore the
train decelerates because of resistance. The performance factor for coasting is its duration.
– During braking, the performance factor is multiplied by the deceleration obtained using the
UIC formula (see Chapter 4 of this book).
As previously mentioned, the calibration can be performed based on GPS tracking files or on
the log-files of digital train event recorders. Evidence from a case study in Italy shows that
about 100 random events per train set type allow reliable results.
The best-fitting set of performance factors for a train trip can be achieved if the infrastructure
model contains the position of the signals, relative distances and speed limits. The calibrated
performance parameters of the train model must result exactly in the measured running
time. The two vectors compared in the equation (11.6) must have the same dimension to be
comparable. The integration period used in the running time calculation must correspond to
the sampling interval of the GPS or train event recorder data. The two arrays are compared by
a simple mean squared error estimator:

230
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 11.5 Estimation of stochastic train running times

N
MSE = ∑ (v GPS (t) − v c (t))2 (11.6)
t=1
that must be minimized.
N: number of samplings/integration steps between two stations
vGPS: vector of recorded speed
v c: vector of calculated speed
The native sampling interval of both the high-precision GPS and the Train Event Recorder
(10 Hz) is used. To compute a fixed length speed vector using the four performance factors,
one dependent variable and three independent variables are to be considered. The algorithm
uses cruising performance as a dependent variable, because:
– it has a small variation and relative high value (always higher than 85 % in the test case).
– cruising running phase is the longest during a train’s run, therefore a minimum variation of
its value has a great impact on total train running time.
– the value of cruising performance is inversely proportional to train running time.
The dependent variable value can be found using the bisection method.
The minimum mean square error giving four independent variables of the non-linear function is
found by resolving the equation (11.7):
⎪⎧⎪ N
⎪⎪min {∑ [v GPS (t) – v c (t)]2 }
⎪⎪ t=1 (11.7)
⎨
⎪⎪v c = f (a, b, c, d, e)
⎪⎪
⎪⎪⎩size (v GPS ) = size (v c ) = N

The coefficients a1, … an are the performance parameters.

The non-linear optimisation problem can be solved e. g. by Simulated Annealing (SA), see
Hillier & Liebermann (2005). The algorithm begins by finding a solution at a random point of
the feasible region; after that, many steps are performed in order to find a better solution. The
maximum number of steps to be performed can be fixed, to limit computation time. Each
step of the SA algorithm replaces the current solution by a random “nearby” solution, chosen
with a probability that depends on the difference between the corresponding function values
and on a global parameter T (called the temperature), that is gradually decreased during the
process. The current solution changes almost randomly when T is large and the more when T
approaches zero. The allowance for “uphill” moves saves the method from becoming stuck at
local optima. Both performance parameters are saved at the end of the optimisation process
into a custom file format for aggregate data analysis.

11.5.2 Calibration results

This paragraph describes how the proposed method was used to calibrate the movement
equation for use in a case study analysis. The case study data were collected on the Trieste
– Udine rail line in North East Italy. Data were collected from approximately 100 train runs
operated with the same rolling stock and used to estimate the parameters. However, even
this number of records was still insufficient to fully explain the very high variability of external
conditions (traffic conflicts, other delays) and driving styles.

Acceleration
The acceleration performance factor is used both for departures from stations and speed limit
increases. In about 5 % of cases the acceleration rate was slightly higher than the theoretical

231
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
11 Statistical Analysis of Train Delays and Movements

rate, although this was limited to a few seconds over the entire train run. This could be due
to conservative estimation of tractive effort, a very low number of passengers, or to very
good adhesion. Two acceleration factors were estimated using the case study data: one for
acceleration with re-acceleration (i. e. in cases where the train is already moving), and one
for acceleration without re-acceleration (i.e. in cases where the train is accelerating from
a stop). Fig. 11.4 presents the histogram of the acceleration performance factor without re-
acceleration and the corresponding best fitting distributions (log-logistic).

Fig. 11.4: Histogram and best fitting distribution excluding re-accelerations

Cruising
An analysis of the case study data showed that there was a very low variability of cruising
performance between on-time and delayed trains; cruising performance ranged from 94 % to
98 %. Trains running early often keep a constant but lower speed (up to 89 %). Generally, the
approach allowed a good fit of on-time and delayed train behaviour, while for trains running
early the variability is very high. For example some early trains run at normal speed and then
wait at stations, while others cruise at lower speeds on open track. Fig. 11.5 presents the
cruising performance factor histogram and the corresponding best fitting distribution (Pert).
Although it may be useful to distinguish between early, punctual and delayed trains to describe
and reproduce the different driving styles in more detail, this analysis indicates that this may be
a difficult task, especially concerning early running and punctual trains. In fact, some drivers –
interviewed during data collection – drive slowly from the beginning of a trip, already estimating
that the following stop will be reached early, while others accelerate normally and then cruise
at lower speeds. Therefore, it is difficult not only to determine a threshold to separate punctual
and early running trains, but also where it must be measured. A larger data set and more case
studies are needed to further improve the accuracy of driving behaviour modelling.

232
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 11.5 Estimation of stochastic train running times

Fig. 11.5: Histogram and best fitting distribution (Pert) of the cruising factor

Coasting
In the case study data, coasting appears generally to be limited to the time necessary to
deactivate the throttle and start the braking action, which includes a pneumatic delay. In
approximately 90 % of the records, coasting time is less than 10 seconds. The data shows
that some drivers use coasting on trains running early for 10 to 20 seconds, from 100 km/h
cruising speed.

Braking
Braking at speed reductions appears to be strongly influenced by significant advance
decelerations and by disturbances caused by early running trains. Often trains brake to a
speed lower than required and then quickly re-accelerate, which generates further noise in
the data. The estimated braking performance factor for speed reductions in about 95 % of the
instances is less than 55 %. Fig. 11.6 (on the left) presents the braking (at speed restrictions)
performance factor histogram and the corresponding best fitting distribution (log-logistic).

Fig. 11.6: Histogram and best fitting distribution (log-logistic) of the performance factor for braking
at speed restrictions (left) and braking at stations (right)

233
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
11 Statistical Analysis of Train Delays and Movements

A more precise analysis is carried out concerning braking at stations. The performance factors
appear very low (20 – 60 %) although delayed trains show higher rates (80 – 90 %). Fig. 11.6
presents the braking (at stations) performance factor histogram and the corresponding best
fitting distribution (lognormal). The data also shows that many trains use a two-step braking
procedure: first they brake to about 30 – 40 km/h, then they coast and finally they brake
again to reach the stopping point (Fig. 11.6 on the right). Although in the two braking steps
the deceleration rates are comparable to high performance factors, the intermediate coasting
leads to lower average values, since the UIC formula (assuming constant deceleration) is used
in the model.

11.6 Conclusions

Analysing train process times accurately appears fundamental to the understanding of railway
operations on a network, identifying the critical points, and correctly estimating in particular
the dwell times and running time margins to be used in the timetable. The fitted distributions
of train delays can be used as input distribution models for either analytical models or
simulation tools that are used to predict the delay propagation and the impact of stochastic
process times on train punctuality at stations. The distribution of train delays and particularly
the parameters may vary in time and space. Thus, regular statistical analysis of train delays at
different locations is necessary.
The variability of running times might show lesser changes over time; however, relying only
on the results of the deterministic movement equation plus standard timetable slack can
lead to an over- or underestimation of running times. The former does not allow exploitation
of the maximum infrastructure capacity, while the latter introduces slight delays, which may
propagate at critical points of the network. When timetables for a new infrastructure project
are designed and evaluated, real-world data is not available. In that case, the parameters of
the train delay distributions can be estimated based on data recorded in a similar existing
network, stations, and train service and by further refinement. A sensitivity analysis of the delay
distribution parameters and their impact on punctuality is recommended.

References
Bailey C. (1995): European Railway Signalling, A & C Black, London.
Banks J., Carson II J. S., Nelson B. L., Nicol D. M. (2004): Discrete-Event System Simulation,
Pearson Education, Inc., New Jersey.
Besinovic N., Quaglietta E., Goverde R. M. P. (2013): A Simulation-based Optimization
Approach for the Calibration of Dynamic Train Speed Profiles, Journal of Rail Transport
Planning & Management 3 (2013).
Carey M., Kwieciński A. (1994): Stochastic Approximation to the Effects of Headways on
Knock-on Delays of Trains, Transportation Research Part B 28(4), pp. 251 – 267.
Carey M., Carville S. (2000): Testing Schedule Performance and Reliability for Train Stations,
Journal of the Operational Research Society 51, pp. 666 – 682.
de Fabris S., Longo G., Medeossi G. (2008): Automated Analysis of Train Event Recorder Data
to Improve Micro-simulation Models, in: Allan J. et al. (eds.), Computers in Railways XI, WIT
Press, Southampton, pp. 575 – 583.
Daamen W. (2004): Modelling Passenger Flows in Public Transport Facilities, PhD thesis, Delft
University of Technology.

234
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

Daamen W., Goverde R. M. P., Hansen I. A. (2007): Non-Discriminatory Automatic and Distinct
Registration of Primary and Secondary Train Delays. In: Hansen, I.A., et al. (eds.), Proceedings
of 2nd International Seminar on Railway Operations Modeling and Analysis, Leibniz Universität
Hannover.
Dekking F. M., Kraaikamp C., Lopuhaä H.P., Meester L. E. (2005): A Modern Introduction to
Probability and Statistics, Springer, London.
Ferreira L., Higgins A. (1996): Modelling Reliability of Train Arrival Times, Journal of
Transportation Engineering 122(6), pp. 414 – 420.
Gibbons J. D. (1985): Nonparametric Methods for Quantitative Analysis, American Sciences
Press, INC., Columbus, Ohio.
Goverde R. M. P. (2005): Punctuality of Railway Operations and Timetable Stability Analysis,
PhD thesis, Delft University of Technology.
Goverde R. M. P., Hansen I. A. (2000): TNV-Prepare: Analysis of Dutch Railway Operations
Based on Train detection Data. In: Allan J. et al. (eds.), Computers in Railways VII, WIT Press,
Southampton, pp. 779 – 788.
Goverde R. M. P., Hansen I. A. (2001): Delay Propagation and Process Management at
Railway Stations, in: Proceedings of World Congress on Railway Research, Köln.
Goverde R. M. P., Hooghiemstra G.; Lopuhaä, H. P. (2001): Statistical Analysis of Train Traffic:
The Eindhoven Case, DUP Science, Delft.
Goverde R. M. P., Meng, L. (2011): Advanced Monitoring and Management Information of
Railway Operations, Journal of Rail Transport Planning & Management 1 (2011), pp. 69 – 79.
Grower Brown R. Hwang, P. Y. C., (1997): Introduction to Random Signals and Applied
Kalman Filtering, second ed. John Wiley & Sons, NavtechGPS.
Güttler S. (2006): Statistical modelling of Railway Data, MSc thesis, Georg-August-Universität
zu Göttingen.
Hansen I. A. (2001): Improving Railway Punctuality by Automatic Piloting. In: 2001 IEEE
Intelligent Transportation Systems Conference Proceedings, Oakland, USA, pp. 792–797.
Hansen I. A. (2004): Increase of Capacity through Optimised Timetabling. In: Allan, J. et al.
(eds.), Computers in Railways IX, WIT Press, Southampton, pp. 529 – 538.
Harris N. G. (1992): Planning Passenger Railways, chapter Punctuality and Performance,
Transport Publishing Company, pp. 130 – 142.
Hermann U. (1996): Untersuchung zur Verspätungsentwicklung von Fernreisezügen auf der
Datengrundlage der Rechnerunterstützten Zugüberwachung Frankfurt am Main, PhD thesis,
Technischen Hochschule Darmstadt.
Herrmann T. M. (2006): Stability of Timetables and Train Routings through Station Regions,
PhD thesis, Swiss Federal Institute of Technology.
Higgins A., Kozan, E. (1998): Modeling Train Delays in Urban Networks, Transportation
Science 32(4), pp. 346 – 357.
Kecman P., Goverde, R. M. P. (2012): Process Mining of Train Describer Event Data and
Automatic Conflict Identification, in: Brebbia, C. A. et al. (eds.), Computers in Railways XIII, WIT
Press, Southampton, pp. 227 – 238.
Law A. M., Kelton W. D. (2000): Simulation Modeling and Analysis, McGraw-Hill Higher
Education.

235
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
11 Statistical Analysis of Train Delays and Movements

Lin T. M., Wilson H. M. (1992): Dwell Time Relationships for Light Rail Systems, Transportation
Research Record 1361, pp. 287 – 295.
Longo G., Medeossi G. (2013): An Approach for Calibrating and Validating the Simulation of
Complex Rail Networks, Transportation Research Board Annual Meeting 2013 Compendium
of Papers, Washington DC.
Hillier F., Lieberman G., (2005): Introduction to Operations Research, 8th ed. McGraw-Hill
Science/Engineering/Math.
MathSoft (1999): S-Plus 2000, User’s Guide, Seattle, USA.
Meng Y. (1991): Bemessung von Pufferzeiten in Anschlüssen von Reisezügen, PhD thesis,
RWTH Aachen.
Medeossi G., Longo, G., de Fabris, S. (2011): A Method for Using Stochastic Blocking Times
to Improve Timetable Planning, Journal of Rail Transport Planning & Management 1 (2011),
pp. 1 – 13.
Mühlhans E. (1990): Berechnung der Verspätungsentwicklung bei Zugfahrten,
Eisenbahntechnische Rundschau ETR, 39 (7/8), pp. 465 – 468.
Olsson N. O. E., Haugland H. (2004): Influencing Factors on Train Punctuality – results from
some Norwegian studies, Transport Policy 11, pp. 387 – 397.
Schwanhäußer W. (1974): Die Bemessung der Pufferzeiten im Fahrplangefüge der Eisenbahn,
PhD thesis, RWTH Aachen.
Stam-Van den Berg B., Weeda V. A. (2007): VTL-Tool: Detailed Analysis of Dutch Railway
Traffic. In: Hansen I. A. et al. (eds.), Proceedings of 2nd International Seminar on Railway
Operations Modeling and Analysis, Leibniz Universität Hannover.
Steckel J. (1991): Strategische Optionen für die zufällige Fahrzeit im Eisenbahnnbetrieb, PhD
Thesis, Dresden Institute for Traffic Research (Hochschule für Verkehrswesen).
Ullius M. (2004): Verwendung von Eisenbahnbetriebsdaten für die Schwachstellen- und
Risikoanalyse zur Verbesserung der Angebots- und Betriebsqualität. PhD thesis, Swiss Federal
Institute of Technology.
Weigand W. (1981): Verspätungsübertragungen in Fernverkehrsnetzen, Eisenbahntechnische
Rundschau ETR, 30 (12), pp. 915 – 920.
Wende D. (2003): Fahrdynamik des Schienenverkehrs, Teubner Verlag, Wiesbaden.
Wendler E., Naehrig M. (2004): Statistische Auswertung von Verspätungsdaten,
Eisenbahningenieurkalender EIK 2004, pp. 321 – 331.
Wiggenraad, P. B. L. (2001): Alighting and Boarding Times of Passengers at Dutch Railway
Stations. In: Hansen I. A. (ed.), Train Delays at Stations and Network Stability, TRAIL Research
School, Delft.
Yuan J. (2006): Stochastic Modelling of Train Delays and Delay Propagation in Stations, PhD
thesis, Delft University of Technology.
Yuan J., Goverde R. M. P., Hansen I. A. (2002): Propagation of Train Delays in Stations. In:
Allan J. et al. (eds.), Computers in Railways VIII, WIT Press, Southampton, pp. 975 – 984.
Yuan J., Hansen I. A. (2007): Optimizing Capacity Utilization of Stations by Estimating Knock-
on Delays, Transportation Research, Part B 41(2), pp. 202 – 217.

236
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 12.1 Introduction

12 Train delay prediction

Pavle Kecman, Rob M. P. Goverde

12.1 Introduction

Daily variations and unforeseeable disruptive events are inevitable in modern railway systems
with many interacting processes that depend on human behaviour, technical devices, and
the environment. In busy and heavily utilised networks, deviations from the planned path
of a single train can easily propagate as a secondary delay to other trains with a logistical
connection or which run over the same infrastructure. Delay propagation could be prevented
or reduced if the traffic is managed proactively, so that controllers have a reliable prediction of
train path conflicts and can take measures to prevent them.
Typically, signallers do not have any intelligent decision support system to allow them to
estimate the expected running times. Currently, their awareness of train delays is further
limited due to imprecise measurements of actual delays. These are usually measured at home
or exit signals, adjusted with a correction term and rounded to full minutes. Traffic controllers
usually take the expected arrival delay of running trains equal to their current upstream
delay as they have no information about possible recovery times, see Fig. 12.1. Likewise, a
prediction of the future traffic situation is sometimes obtained by simple extrapolation. This
“linear shift” method neglects the fact that some trains make up for their delays by running in
a maximum performance regime and exploiting the running time supplements incorporated in
the timetable. On the other hand, other trains may get even more delayed due to possible time
losses in route conflicts.

Fig. 12.1: Linear shift method for delay prediction

A train typically crosses multiple traffic control areas, each with its own local traffic controllers
(signallers and/or dispatchers). However, the situational awareness of signallers is often limited
to their own control area; they have to handle the trains coming their way and dispatch them

237
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
12 Train delay prediction

to an adjacent interlocking area without further knowledge of the network-wide consequences

of their decisions, see Kauppi et al. (2006). This often leads to the creation of new conflicts in
adjacent areas and suboptimal effects at the network wide level.
Kecman et al. (2011) proposed a model-predictive control framework for railway traffic
management to tackle this problem. This approach supports traffic controllers by monitoring
train positions, predicting train running times, and detecting train path conflicts well in
advance. Thus the controllers have enough time to solve the path conflicts. Potentially
conflicting train paths in the actual working plan can be predicted in advance. These
predictions are based on accurate monitoring of train positions and infrastructure conditions
such as track unavailability or temporary speed restrictions affecting the running times. As a
result, unscheduled stops before red signals can be avoided by resolving such conflicts in
advance using retiming, reordering, rerouting, and speed advice to drivers. A predictive traffic
model should support route setting and traffic control decisions and must be used interactively
by signallers and traffic controllers. Also rolling stock and crew connections can be taken into
account as constraints to detect conflicts.
The theoretical predictive traffic control framework is presented in Fig. 12.2. Trains are
scheduled according to a master timetable containing a detailed process plan of all train paths
at route level. Train movements, infrastructure and other resources are monitored and the
actual traffic state (train positions and delays) is presented to traffic controllers. A predictive
traffic model continuously estimates the future traffic state and the propagation of delays. The
impact of any control decision can be inserted into the predictive model to check new conflict
and arrival time predictions. Finally, if a control decision leads to satisfying results it can be
implemented in the actual process plan.

Fig. 12.2: Framework of predictive traffic control

238
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 12.2 Monitoring

This chapter presents an overview of the recent advances in the field of real-time traffic
monitoring and prediction. Section 12.2 presents the general concepts of real-time traffic
monitoring. Section 12.3 explains the existing tools and models for traffic prediction. Special
focus in this chapter is on a data-driven approach for monitoring and traffic state prediction,
which is considered in Section 12.4, followed by conclusions in Section 12.5.

12.2 Monitoring

A monitoring system keeps track of all dynamic data required for traffic state predictions.
Tracking of train positions is a basic requirement for the monitoring of railway traffic.
Traditionally, train positions were monitored only at staffed stations and other timetable points.
However, development in sensory and communications technologies enabled a more detailed
observation of running trains.
Train positioning can be track-based or train-based, see Lüthi (2009). Fig. 12.3 illustrates the
track-based approach to train positioning. Train describers are the main example of track-
based train positioning, see Exer (1995). A train describer system keeps track of train positions
in discrete steps over the train’s route. This is based on messages received from track-clear
detection devices and point control machines, that indicate the state of track sections (clear,
occupied) and switches (left, right). Moreover, an important function of train describers is the
logging of incoming infrastructure element messages and the generated train number messages.

Fig. 12.3: Track-based train positioning

The second approach to train positioning is train-based where trains send their position at
regular intervals to the traffic control centre, see Fig. 12.4. The train position is determined in
the train itself using e. g. odometers and balises, such as in ETCS Levels 2 and 3, or using the
Global Positioning System (GPS). In this case, trains communicate their position to the traffic
control centre using a radio connection, such as GSM-R.
The train positions are used in traffic control centres for, among other things, automatic
route setting and current delay monitoring. Moreover, archives of train positions are used for
performance analysis of railway operations. The monitoring information can also be processed
further for predicting train traffic over some horizon.

239
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
12 Train delay prediction

Fig. 12.4: Train-based train positioning

12.3 Traffic State Prediction and Conflict Detection

Based on the current traffic situation provided by the monitoring system, a prediction model
forecasts the future evolution of traffic, i. e., the progress of trains and possible conflicting train
paths and their consequences. Traffic state prediction can be performed on a macroscopic
and microscopic level. Macroscopic models provide estimates of departure, arrival and
passage times at stations or other timetable points while incorporating a simple view of
infrastructure constraints. This enables traffic predictions at the national level. On the other
hand, microscopic models consider detailed constraints imposed by signalling and interlocking
systems and are very accurate.
Macroscopic models for delay prediction are mainly based on a fixed timetable and comprise
deterministic (Goverde, 2010) and stochastic (Büker & Seybold, 2012; Berger et al., 2011)
delay propagation models. These models are suitable for online applications with frequent train
position updates and were applied to large-scale networks such as the Dutch, German and
Swiss railway networks. The model of Berger et al. (2011) also incorporates waiting policies for
passenger connections.
Hansen et al. (2010) present a data-driven macroscopic traffic state prediction approach
based on the actual state of traffic and event-data records. A macroscopic model for
predicting train running times is calibrated from historical track occupation data. Robust
estimates of minimum running times are computed on the level of open-track sections
between stations, which are used to predict arrival times. A critical path algorithm is used to
predict the event times over the macroscopic graph starting from the last realised event. This
data-driven approach takes into account dependencies of running times on delays, peak-
hours, weather, and rolling-stock types.
Fukami, Yamamoto (2001) presented a time-distance diagram prediction system for a high-
speed line in Japan. The system follows the positions of all trains on the network using train
describer messages and estimates the speed of running trains by assuming a constant
velocity over track-clear detection sections. Train trajectories are simulated until the arrival
at the next station with respect to all microscopic operational constraints. However, the

240
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 12.4 Data-Driven Approach to Monitoring and Traffic State Prediction

arrival delays at the subsequent stations are extrapolated using the linear shift method. This
drastically reduces the complexity of the computationally demanding simulations but also
makes the system less applicable for long corridors or complex networks.
Microscopic traffic prediction models are described in the literature in connection with real-
time rescheduling models. Lüthi (2009) studied the role of real-time prediction systems in the
traffic control environment. In this context, the purpose of traffic prediction is twofold. First, a
prediction of train trajectories is required until the next controllable point in the network that
can accommodate rescheduling decisions such as reordering or rerouting. Second, traffic
prediction is required to evaluate the impact of the rescheduling decisions. D’Ariano et al.
(2007) present a graph-based real-time rescheduling model, where the graph represents
precedence relations between train events using blocking time theory and including reordering
decisions by so-called alternative arcs. Their model simplified the route setting and release
principles in station areas which require sequence-dependent setup times, see Corman
et al. (2009). In the latest version the sectional-release route locking principle is also properly
modelled, see Goverde et al. (2013).
An online prediction tool has been implemented in the Swiss Rail Control System RCS-
DISPO, see Dolder et al. (2009). The prediction model is based on a directed acyclic graph,
with nodes in the graph corresponding to arrival and departure events at timetable points and
signals, and arcs representing precedence relations between nodes corresponding to running,
dwell, headway and connection arcs. Arc weights are computed using train motion equations,
using a detailed description of infrastructure and train characteristics. After each train position
update, the running times until the next station are computed and a critical path algorithm
derives predictions of all event times on the graph. Prediction errors smaller than one minute
were reported for events within a 10 minute prediction horizon. However, the approach does
not explicitly model the train dynamics after route conflicts.

12.4 Data-Driven Approach to Monitoring and Traffic State Prediction

12.4.1 Methodological Framework

This section presents an approach to predict train traffic based on historical traffic realisation
data. Data archives generated by track or train based train positioning systems can be used
for building and calibrating data-driven railway traffic models. The historical data is used
to learn about train running and dwell times in various traffic conditions. The future train
trajectories on the network can then be predicted based on the actual traffic state using
running and dwell time estimates from similar circumstances in the past.
Kecman & Goverde (2013) present a real-time model for continuous online prediction of
train traffic using a microscopic graph model. All scheduled events and signal passages are
captured in the model. Moreover, all precedence relations between events, such as minimum
running and dwell times, connections, and minimum headways are taken into account. Robust
estimates for process times are derived dynamically from pre-processed historical traffic data.
The model is updated instantaneously when new information becomes available on train
positions or traffic control decisions. The realisation times of all events within a rolling horizon
are predicted by considering the usage of running time supplements and buffer times, as well
as time loss due to route conflicts based on a conflict detection scheme within the prediction
algorithm. Moreover, an adaptive scheme improves the accuracy of predictions by checking
the prediction errors of updated train positions and adjusting the downstream arc weights
accordingly. This approach is explained in more detail below.

241
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
12 Train delay prediction

The methodological framework of the prediction model is illustrated in Fig. 12.5. The approach
consists of two parts. The offline part comprises data processing of a training set of raw train
describer data and a statistical analysis of the resulting traffic data. The historical event data
contain track occupation data logged by a train describer system. The event data records
are processed using a process mining tool that extracts the running times over successive
blocks, the dwell times at stops, minimum headway times between train paths, and all realised
route conflicts. In the data analysis, a set of predictors and their impact on process times is
determined using a statistical model based on the training set of historical data for running
and dwell times. Robust regression analysis is here used to determine correlations between
process times and explanatory variables such as train line, delay and time of day. Also
quantified is the dependence of time loss due to braking and re-acceleration on the duration
of a route conflict. The resulting regression coefficients are stored in the database of historical
data.

Fig. 12.5: Methodological framework of data-driven traffic prediction

The prediction is performed in the online part (dashed box) by propagating the actual traffic
state (current train positions and upstream delays) through a graph model of rail traffic. The
actual process plan, i. e., routes, timetable, connections, as well as the current positions of
all trains is used to build the microscopic graph model. The graph topology is updated after
each change of the actual process plans such as adding new trains, modifying train routes, or
updating connections. Microscopic operational constraints are reflected in the graph structure,
which captures all train interactions due to signalling or connection constraints. The graph is

242
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 12.4 Data-Driven Approach to Monitoring and Traffic State Prediction

calibrated dynamically, in real-time, using historical data. There is a predetermined dependency

of running and dwell times on departure and arrival delays, respectively. When an update
of the train positions is received from the real-time data stream, a depth-first search based
algorithm traverses the graph, computes robust estimates of the arc weights and predicts all
event times within the prediction horizon.
This approach detects route conflicts and incorporates the time losses, due to braking and
re-acceleration of hindered trains, in the predictions. Moreover, an adaptive scheme exploits
feedback information from all realised. Trains with process times that deviate continuously from
computed estimates in a certain pattern are detected. Downstream process times are then
adjusted to minimise the expected prediction error.

12.4.2 Process Mining of Train Describer Event Data

Process mining is a method for discovering processes and extracting and analysing
information about them from event data using a process model (Van der Aalst, 2011). It
combines data mining with domain knowledge about the specific processes that are analysed.
The principle idea of the concept is to extract the necessary information from large data sets
and obtain an output containing clean and structured data ready for analysis.
Blocking time theory provides the logic for building the process model from the data records.
Signal passages are events that initiate processes such as blocking a part of the infrastructure
and running over a block. Each complete train run can thus be represented as a graph,
discovered and built online by sweeping through the file once. Moreover, route conflicts can
be identified simultaneously by determining the time difference between relevant events and
verifying if the train separation principles are respected, see Daamen et al. (2009).
Fig. 12.6 shows the methodological framework for process mining of train describer event
data. This method can be applied both for processing a live stream of incoming train describer
messages for the purpose of monitoring traffic, and processing archives of data records to
extract the structured historical traffic realisation data. The core of the tool is an environment
containing section, signal, block and train objects. Each infrastructure object keeps track of
occupation, release and passing times of all trains that are reported by the train describer
system. Similarly, train objects store the occupation and release time of each section and the
passing time of each signal along route of the train. Information passing between different
object types and methods within the same object type reflect the operational constraints of
railway traffic such as the route setting and release principles and train separation on open
tracks according to blocking time theory. The output of the process mining include realised
running and dwell times, route conflicts, realised departure and arrival times, as well as entire
realised train paths (blocking time or time distance diagram).

12.4.3 Robust Process Time Estimates

The data processing step extracts the conflict-free running times at the level of block sections,
the dwell times in stations, and the minimum headway times between trains that run over the
same infrastructure, while the running times of the hindered train runs with identified route
conflicts are filtered out. This data can be used to derive robust estimates of the process
times. A statistical model has been developed to determine and quantify the impact of
explanatory variables on running, dwell and headway times. By monitoring the value of the
explanatory variables in real time, the robust estimates of process times can be conditioned to
derive more accurate predictions.

243
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
12 Train delay prediction

Fig. 12.6: Process mining of train describer event data

The process times need to be analysed separately for each train line. It is assumed that all
trains on the same train line have identical stopping patterns in terms of scheduled stops
and scheduled running and dwell times. Different routes are identified in the data processing
and dealt with separately. A major item in the data analysis is the hypothesis that the running
and dwell time of a train depends on the current delay. Delayed trains may run with full
performance in order to use the running time supplements to reduce the delay. On the other
hand, trains running on time or ahead of their schedule run in a lower performance regime
thus avoiding early arrivals and achieving energy-efficient driving, see Hansen et al. (2010).
Similarly, dwell times of trains in stations may depend on arrival delay. Since under normal
circumstances trains cannot depart from a station before the scheduled departure time, early
trains have longer dwell times than scheduled in order to avoid early departures. On the other
hand, trains with a positive arrival delay can depart after a minimum dwell time in order to
reduce the departure delay. These general assumptions can be tested on the data set of track
occupation data. The purpose of the analysis is to test the correlation between running and
dwell times with actual delays.
The hypothesis on the different behaviour of delayed (delay exceeding 60 s) and punctual
trains can be tested. The set of observed running and dwell times is separated into
corresponding sets of delayed and punctual trains and the Wilcoxon rank sum test is applied
to verify whether the subsets have equal distributions. The null hypothesis is that the samples
have identical continuous distributions with equal medians. Robust regression can be used to
derive a functional relationship between the running or dwell time and delay and assessing its
performance. In particular, Least Trimmed Squares (LTS) regression (Rousseeuw & Driessen,

244
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 12.4 Data-Driven Approach to Monitoring and Traffic State Prediction

2006) has been shown to work well. LTS regression is a robust regression method that can
identify and exclude a specific percentage of outliers in the data set and perform the analysis
on the remaining data points.

Fig. 12.7: Dependence of running time on delay (left) and running time distributions conditional on
delay (right)

We present an example of applying these statistical methods to the running time of train
line 2200 from The Hague to Delft over the approaching block before a scheduled stop in
Delft station. A weak regression correlation was found between the running time over the
approaching block and the previous departure delay from The Hague HS (R 2 = 0.0376), see
Fig. 12.7 (left). The Wilcoxon rank sum test rejected the null hypothesis of equal distributions
with p ≈ 0. The box plots in Fig. 12.7 (right) show small differences in distributions of six data
samples conditional on the value of departure delay. The analysis shows that delayed trains
indeed tend to run faster to recover the departure delay. However, for punctual, early or
slightly delayed trains no significant correlation with the departure delay can be established.
Earlier analyses of running times over sections between two scheduled stops, conducted in
the Netherlands (Hansen et al., 2010) and Switzerland (Lüthi, 2009) show similar results. This
shows some indifference of drivers to small deviations from their path.

Fig. 12.8: Dependence of dwell time on delay (left) and dwell time distributions conditional on delay
(right)

245
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
12 Train delay prediction

The dependence of dwell times on arrival delays can be quantified in the same manner.
Fig. 12.8 (left) shows the dependency of dwell time on arrival delays for train line 2200
in station Delft. The horizontal dashed line represents the 10th percentile of all dwell times,
whereas the red line represents the robust LTS fit for early trains. The scheduled dwell time
is 60 seconds. A strong correlation (R 2 = 0.8704) was captured for the early trains. The
Wilcoxon rank sum test rejected the null hypothesis of equal distributions (p ≈ 0) and different
distributions of dwell times for punctual and late trains are also clear from the box plots in
Fig. 12.8 (right). However, the variation of dwell times for delayed trains needs to be explained
by other factors and therefore the data set is divided into a set of punctual and delayed trains.
The variability of dwell times for delayed trains can be explained by a dependence on the
time of the day. Dwell times of delayed trains are theoretically equal to the minimum dwell
time required for passenger alighting and boarding and route setting. We hypothesised that
passenger volumes and thus boarding and alighting time increase during peak hours. Fig. 12.9
shows the dwell times on work days relative to the scheduled arrival time for train line 2200 in
Delft. The solid line indicates the median dwell time. The increase in dwell times during peak-
hours is clearly visible. The distinct causes of dwell time variability for punctual and delayed
trains result in a bimodal approach to the robust conditional prediction of dwell times. Hence,
for early trains, dwell time can be predicted based on the correlation with arrival delay. On the
other hand, dwell time for a delayed train can be estimated from historical data based on dwell
times of the same train number and its adjacent train numbers from the same train line as a
moving median, see Hansen et al. (2010).
Finally, historical traffic data can be used to analyse the time loss due to a route conflict
depending on the conflict duration and the route and running time of the conflicting train. The
typical situation that occurs in practice is the “conflict wave” for two successive trains over
the same route, where the hindered train keeps meeting yellow signals and thus is unable to
reaccelerate to full speed, see Goverde & Meng (2011). This analysis therefore considers the

Fig. 12.9: Dependence of dwell time on scheduled dwell time over the day

246
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 12.4 Data-Driven Approach to Monitoring and Traffic State Prediction

time loss due to re-acceleration only after the hindered train passes a green signal again. Since
the running time estimates are computed based on the free running times, the arc weights
modelling the running processes over affected blocks need to be adjusted for every predicted
route conflict. This takes into account braking, possible waiting time in rear of a red signal,
running at a restricted speed, and re-acceleration. The duration of a conflict is computed
using the process mining conflict identification tool on a set of historical traffic data (Kecman &
Goverde, 2012) and the resulting time loss can be obtained as the difference between the
realised running time over a block and the conflict-free running time derived conditional on the
current train delay.

Fig. 12.10: Dependence of time loss on conflict duration

Fig. 12.10 shows an example of a regression analysis that was performed based on
20130 data points split into conflicts shorter (Fig. 12.10, left) and longer than 150 seconds
(Fig. 12.10, right). A robust quadratic fit resisting 25 % of the outliers showed the best
performance in terms of coefficient of determination R 2 = 0.79 for conflicts shorter than
150 seconds. Even though the data points are scarce for conflict durations longer than 150
seconds, the linear regression line (R 2 = 0.92) can be interpreted easily as the waiting time
in rear of the signal, which is the greater part of time loss in long conflicts. This approach
enables the adjustment of running time estimates for hindered trains after a route conflict and
its duration have been predicted.

12.4.4 Online Prediction of Train Event Times

Running and dwell times can be predicted accurately by using the up-to-date information
about actual train positions and delays provided by the monitoring system, and the robust
estimators and dependencies of process times. A timed event graph (Goverde, 2010) is a
convenient approach to represent the microscopic operational constraints of railway traffic,
described by blocking time theory including the route setting and release principles in station
areas. The graph topology is built and updated based on the actual timetable, train orders,
routes and connection plans. A distinction is made between signal events (a train passing a
signal) and scheduled station events (arrival/departure at/from a platform track). Train events
are connected by running and dwelling arcs. Interdependence between trains is modelled by
signal headway and connection arcs.
An important parameter of the prediction approach is the length of the rolling prediction
horizon. When a train position update becomes available, the prediction horizon is moved
forward in time and the event times of all events within the horizon are predicted. The
prediction algorithm traverses the graph in topological order, assigns arc weights dynamically
based on the actual (predicted) delay and computes predicted event times for all events within

247
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
12 Train delay prediction

the prediction horizon, see Kecman & Goverde (2013). The microscopic character of the
model enables prediction of route conflicts. Bearing in mind that the running time estimates are
computed based on representative conflict-free running times, the estimates of running times
of hindered trains need to be adjusted to incorporate the time loss due to braking and re-
acceleration. The running time adjustment is computed depending on the predicted time loss
using the functional dependence established from historical traffic realisation data as explained
in Section 12.4.3.
The presented prediction model is event driven, i. e., all event times are predicted when an
update on the current position of a running train becomes available. The estimated running
times over block sections depend on the departure delay from the last scheduled stop. An
adaptive scheme of the prediction model keeps track of the actually realized running times of a
train and adjusts the estimates until the next scheduled stop. A simple way of doing this is by
a moving average smoothing method to incorporate the prediction error observed during the
train run into future predictions until the next stop. This adaptive prediction strategy enables a
correction for a slow running individual train due to a particular driving style, defective rolling-
stock, or temporary speed restrictions.

12.4.5 Model Performance

An example of the performance of the presented prediction model is shown in Fig. 12.11.
The time-distance diagram shows the predicted train paths (in magenta and blue). The
realised train paths are presented in black. The prediction is performed at the departure of
train ST5025 from The Hague HS (GV) at 07 : 13. The complete routes from The Hague HS

Fig. 12.11: Time-distance diagram of realised and predicted train paths between The Hague HS
and Rotterdam Central (at 07:13)

248
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 12.4 Data-Driven Approach to Monitoring and Traffic State Prediction

(GV) to Rotterdam Central (RTD) of the seven trains that entered the network within a 30
min prediction horizon are included in predictions. The average prediction error for the 161
predicted events (including signal passages) is 19 seconds, while the maximum prediction
error is 69 seconds.
The main advantage of the presented model for traffic controllers is the ability to predict all
route conflicts within the prediction horizon. We use the principle of overlapping blocking
times to predict and visualise route conflicts. Fig. 12.12 and Fig. 12.13 show the predicted
and realised blocking time diagram, respectively. Local trains are presented in magenta and
intercity trains in blue. Overlaps in blocking times indicating route conflicts are presented in
red. Three out of four major route conflicts that occurred were predicted by the model: one
in Schiedam (SDM) and two in Rotterdam (RTD). The fourth conflict which was not captured
occurred in Schiedam more than 30 minutes after the moment of prediction due to a train
running faster than usual. This conflict is detected at a later prediction time when this fast
running is identified and corrected by the adaptive scheme.
For the evaluation of the data driven approach to train event time prediction we consider a
real life case study of one day of traffic along a busy corridor in the Netherlands. Fig. 12.14
(left) shows the distributions of prediction error for different prediction horizons. The standard
deviation and median of the prediction error reduce with decreasing prediction horizon
length. For a prediction horizon of 10 minutes the mean prediction error is 24 seconds and
the standard deviation is 58 seconds. This increases for longer horizons but for a two hour
prediction horizon the prediction error is with a mean of 59 seconds and standard deviation
of 94 seconds still very good. Fig. 12.14 (right) presents the prediction error distribution for
running and dwell times. It is clear that dwell times are the major source of inaccuracy which

Fig. 12.12: Realised blocking time diagrams between The Hague HS and Rotterdam Central

249
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
12 Train delay prediction

Fig. 12.13: Blocking time diagrams between The Hague HS and Rotterdam CS predicted at 07:13

Fig. 12.14: Box plots of prediction error for different prediction horizon length (left) and different
process times (right)

is also propagated throughout the network. The running times are estimated very well with
prediction errors less than ±10 seconds. More advanced predictive models for dwell times
are thus needed to reduce the uncertainties in the dwell times and further improve the
overall performance of the traffic predictions. Note that the dwell times are predicted here

250
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

without information on passenger flows. Finally, even for the longest prediction horizon, the
algorithm execution takes less than one second which makes it fully applicable in a real-time
environment.

12.5 Conclusions

Predictive railway traffic management gives traffic controllers an opportunity to anticipate on

route conflicts and deviations from the train schedules. It enables proactive traffic control
where route or connection conflict. The resulting secondary delays can be prevented or
reduced. A predictive traffic model therefore provides effective decision support to signallers
and traffic controllers. It also contributes to a better utilisation of railway infrastructure,
improved reliability of train services, and more reliable and dynamic passenger information. The
developed model can be embedded in a closed-loop model-predictive railway traffic control
framework, where online optimisation algorithms will automatically resolve detected conflicts
and propose control decisions to traffic controllers together with the predicted conflicts.
This way an intelligent railway traffic management system will be feasible, which proactively
supports traffic controllers with decisions that optimize the traffic on a network level, beyond
the traditional local control areas.
The data-driven approach to monitoring and traffic state predictions relies on historical train
describer data, thus reflecting all phenomena of railway traffic captured by the train describer
systems and accompanying processing tools. The estimation of process times can be
performed in real time using the identified functional dependencies on delays and peak-hours.
This way, computationally demanding running time calculations based on train dynamics
can be avoided without loss of prediction accuracy. Furthermore, the graph-based model
allows the application of fast algorithms to compute event time predictions even for large and
dense networks. Train interactions are modelled with high accuracy by inclusion of the main
operational constraints. The model relies on validated minimum headway times obtained from
historical data rather than on theoretical values.
The application results indicate that more accurate modelling of dwell times is required to
increase the reliability of predictions. The presented model relies solely on track occupation
data which provide sufficient information to derive robust estimates for running times.
However, more information on dwell time processes is required, such as platform and rolling-
stock properties, passenger flow counts and passenger distributions along the platform.

References
Berger A., Gebhardt A., Müller-Hannemann M., Ostrowski M. (2011): Stochastic delay
prediction in large train networks. In: Caprara, A., Kontogiannis, S. (Eds.), 11th Workshop on
Algorithmic Approaches for Transportation Modelling, Optimization, and Systems (ATMOS
2011), Saarbrücken, September 8, 2011, pp. 100 –111.
Büker T., Seybold B. (2012): Stochastic modelling of delay propagation in large networks.
Journal of Rail Transport Planning & Management, 2(1-2), pp. 34 – 50.
Corman F., Goverde R. M. P., D’Ariano A. (2009): Rescheduling dense train traffic over
complex station interlocking areas. In: Ahuja R. K., Möhring R. H., Zaroliagis C. D. (Eds.),
Robust and Online Large-Scale Optimization: Models and Techniques for Transportation
Systems, LNCS 5868, Springer, Berlin, pp. 369 – 386.

251
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
12 Train delay prediction

D‘Ariano A., Pranzo M., Hansen I. A. (2007): Conflict resolution and train speed coordination
for solving real-time timetable perturbations. IEEE Transactions on Intelligent Transportation
Systems, 8 (2), pp. 208 – 222.
Daamen W., Goverde R. M. P., Hansen I. A. (2009): Non-discriminatory automatic registration
of knock-on train delays. Networks and Spatial Economics, 9 (1), pp. 47– 61.
Dolder U., Krista M., Völcker M. (2009): RCS–Rail Control System: Realtime train run
simulation and conflict detection on a net wide scale based on updated train positions. In:
Proceedings of the 3rd International Seminar on Railway Operations Modelling and Analysis
(RailZurich 2009), Zurich, February 11 – 13, 2009.
Exer A. (1995): Rail traffic management. In: Bailey, C. (Ed.), European Railway Signalling.
Institution of Railway Signal Engineers, A&C Black, London, pp. 311– 342.
Fukami K., Yamamoto H. (2001): A new diagram forecasting system for the Tokaido-Sanyo
Shinkansen. In: Proceedings of the World Congress on Railway Research (WCRR 2001), Köln,
November 25 – 29, 2001.
Goverde R. M. P. (2010): A delay propagation algorithm for large-scale railway traffic networks.
Transportation Research Part C, 18(3), pp. 269 – 287.
Goverde R. M. P., Corman F., D’Ariano A. (2013): Railway line capacity consumption of
different railway signalling systems under scheduled and disturbed conditions. Journal of Rail
Transport Planning & Management, available online 28 December 2013.
Goverde R. M. P., Meng L. (2011): Advanced monitoring and management information of
railway :operations. Journal of Rail Transport Planning & Management, 1(2), pp. 69 – 79.
Hansen I. A., Goverde R. M. P., Van der Meer D. J. (2010): Online train delay recognition
and running time prediction. In: 13th International IEEE Annual Conference on Intelligent
Transportation Systems (ITSC 2010), Funchal, September 19–22, 2010, pp. 1783 – 1788.
Kauppi A., Wikström J., Sandblad B., Andersson A. W. (2006): Future train traffic control:
control by re-planning. Cognition, Technology & Work, 8 (1), pp. 50 – 56.
Kecman P., Goverde R. M. P., Van den Boom A. J. J. (2011): A model-predictive control
framework for railway traffic management. In: Proceedings of the 4th International Seminar on
Railway Operations Modelling and Analysis (RailRome 2011), Rome, February 16 – 18, 2011.
Kecman P., Goverde R. M. P. (2012): Process mining of train describer event data and
automatic conflict identification. In: Brebbia C. A., Tomii N., Mera J. M. (Eds.): Computers in
Railways XIII, WIT Transactions on The Built Environment, Vol. 127, WIT Press, Southampton,
pp. 227 – 238.
Kecman P., Goverde R. M. P. (2013): Adaptive, data-driven, online prediction of train event
times. In: Proceedings of the 16th International IEEE Annual Conference on Intelligent
Transportation Systems (ITSC 2013), The Hague, October 6-9, 2013, pp. 803 – 808.
Lüthi M. (2009): Improving the efficiency of heavily used railway networks through integrated
real-time rescheduling. PhD thesis, Swiss Federal Institute of Technology (ETH), Zurich.
Rousseeuw P. J., Van Driessen K. (2006): Computing LTS regression for large data sets. Data
Mining and Knowledge Discovery, 12 (1), pp. 29 – 45.
Van der Aalst W. M. (2011): Process mining: discovery, conformance and enhancement of
business processes. Springer, Berlin.

252
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 13.1 Introduction

13 Rescheduling
Andrea D’Ariano, Francesco Corman, Dario Pacciarelli

13.1 Introduction

13.1.1 Background

Timetable design is a complicated task addressed by railway operators many months in

advance of actual operations. During operations, however, unforeseen events may disrupt
the timetable and cause conflicts between train paths. These must be resolved in real time.
Once a delayed train deviates from its original plan, it may hinder subsequent trains that are
scheduled over the same railway infrastructure, or there may be conflict when passing or
meeting other trains. Hence, a delayed train may propagate its delay to other trains, due to
infrastructure, signalling or timing conflicts.
Current operational traffic management is mostly reactive. Traffic management can be seen as
multiple levels interactions, as follows:
A train driver tries to adhere to the original schedule and responds to restrictive signalling
aspects when they indicate that the route ahead is occupied.
Local traffic controllers (dispatchers) can update orders and routing decisions within an area of
limited geographical size (dispatching area). Traffic controllers have a very limited knowledge of
the current status of the railway network, mostly limited to the block section where the train is
at present. They have no precise information on the train’s positioning, speed or acceleration.
For this reason, dispatchers can only update the plan when a considerable delay has
accumulated. A reference situation for a traffic controller is shown in Fig. 13.1.
At the highest hierarchical level, network traffic controllers co-ordinate the traffic over large
areas; they become active only when rail traffic is already highly disrupted. Reactive control
typically results in sub-optimal performance of the rail service as a whole, especially in
congested networks, or when planned events or unplanned events (such as maintenance
or disruptions) temporarily reduce the capacity of some portion of the rail network. In such
situations, any small initial delay can easily spread in space and time over a network, causing
consecutive or knock-on delays to other traffic.
Many consecutive delays can be prevented or reduced drastically with proactive traffic
management. With proactive control, each train driver receives an advisory speed to follow,
while dispatchers and network traffic controllers focus on preventing traffic disturbances rather
than solving them. Proactive traffic management requires the following:
– the precise monitoring of the current train positions,
– the prediction of train speed profiles or running times in a defined geographical area and for
a defined time window,
– the detection of the effects of disturbances to train traffic in terms of conflicts,
– the rescheduling of trains in real time, such that consecutive delays are minimised, by
adjusting orders, routes set, times, and finally
– the communication of the advisory location-time-speed targets to train drivers.
The focus of this Chapter is on the most promising models and algorithms for advanced
proactive rescheduling, from the recent technical literature. However, reactive models as well
as system architectures are also discussed briefly.

253
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
13 Rescheduling

Fig. 13.1: A Traffic Control Center. Source: Network Rail

13.1.2 Problem description

Signals, interlocking and Automatic Train Protection (ATP) systems control the train traffic by
imposing a minimum safety separation between trains, setting up train routes and enforcing
speed restrictions on train running. Fixed block ATP systems ensure safety through the
concept of the block section, a part of the infrastructure that is assigned exclusively to at most
one train at any one time. Train movements can be modelled by a set of characteristic times,
as follows. The running time of a train through a block section starts when its head (the first
axle) enters the block section and ends when the train reaches the end of the block section.
Safety regulations impose a minimum separation between consecutive trains travelling through
the same block section, which translates into a minimum headway time between the start of
the running times of two consecutive trains on the same block section. This time depends
on the length of the block section, as well as on other factors like the speed and length of
the trains. It includes the time between the head of the train entering a block section and the
exit of its tail (the last axle) from the previous one. Time margins must be added to allow the
release of the occupied block section and to take into account the signal sighting distance for
the driver.
Proactive rescheduling must take into account the following. At stations, a train is not allowed
to depart from a platform stop before its scheduled departure time and is considered late
when arriving at the platform later than its scheduled arrival time. At a platform stop, the
scheduled stopping time of each train is called dwell time.
Disturbances affect rail traffic. There is a need to distinguish between traffic perturbations
from neighbouring dispatching areas, caused by a set of delayed trains at the entrance

254
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 13.1 Introduction

of a dispatching area, and timetable disruptions. These latter are much more substantial
modifications of scheduled train speeds and routes, and for which route changes may be
necessary. This may be caused by some block sections not being immediately available.
Other kinds of disturbances include extensions to dwell times due to passengers boarding,
passenger connections, or technical problems. Running times may be prolonged because of
headway conflicts between consecutive trains, or technical failures.
Delays may be propagated between trains when potential route conflicts are solved. Namely a
potential conflict between two trains arises if both request the use of the same block section
and the minimum distance headway between them, which is needed for safety reasons
and smooth operation, is not available. This conflict is resolved by regulating the order of
trains over the block section. In this case, one of the approaching trains might be forced to
decelerate and hence experience a knock-on delay. The unscheduled braking and stopping of
trains increases further the running time and contributes to additional delays. Similarly, trains
can be held at stations due to the lack of availability of routes and/or platforms. In general,
delays may propagate causing a domino effect of increasing disturbances.
A train speed profile has further to comply with the acceleration/braking rates usually
performed by the train (Fig. 13.2). A dispatching solution is therefore feasible if each train
has an acceptable speed profile, train distance headways are respected and there are no
unsolved potential conflicts between trains. Thus trains are allowed to pass yellow signals at
the required approach speed, but must stop on reaching red signals.

Fig. 13.2: Typical driver behaviour during disturbed operations

Real time train traffic management can therefore be defined as meeting the following
requirements. It is assumed that a railway network, a set of train routes and passing/stopping
times at each relevant point in the network are known, also, the position and speed of each
train at a given starting time t0.
The task is to find an updated plan of operations that:
– solves all potential conflicts between trains,
– does not result in deadlock situations (trains that are all waiting for each other, making any
planned movement impossible),
– is compatible with the initial positions of all trains,
– is such that the selected train routes are not blocked,
– the speed profiles are acceptable,
– no train appears in the network before its expected entrance time (including the entrance
delays),
– no train departs from a relevant point before its scheduled departure time, and
– trains arrive at the relevant points with the smallest possible knock-on delay.

255
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
13 Rescheduling

13.1.3 Current dispatching practice

Network dispatchers regulate railway traffic by sequencing the train movements and setting
the routes. Their aim is to ensure smooth train movements and to limit existing delays as much
as possible. Due to the strict time limit available for computing a new plan of operation this
is incompatible with the complexity of existing timetabling tools. Operators usually restrict
themselves to a few manual modifications of the original timetable. Experienced dispatchers
have developed strategies which allow them to foresee simple forms of possible disturbance
and their effects in advance, and to take compensatory control actions based on local
information. In general, minor changes are preferred to extensive rescheduling, in order to alter
the original timetable as little as possible.
Basic dispatching systems have been developed to aid traffic controllers, whose computer
support is often limited to graphical interfaces and simple automatic route setting systems.
(These are systems which provide the automatic setting of routes when a train approaches a
signal). The usual policy still consists of scheduling trains following their order in the timetable,
or according to pre-determined dispatching rules. A practical use of simple local measures,
e. g. train reordering at crossing points on a first come, first served, basis has recently been
introduced into the Dutch railway network (see for example Hemelrijk, Kruijer & De Vries
(2003)).
Automatic route setting systems (such as the ARI system, see Berends & Ouburg (2005))
might include some limited conflict detection and resolution measures. However, such systems
are automated only in the sense of no or little disturbance to traffic patterns, and do not
provide effective support when dealing with heavy disturbances, or in complex networks.
In summary, existing support system for dispatchers rarely go beyond monitoring the situation,
and they are able to provide acceptable solutions only for small scale problems or for simple
perturbations. Real-time traffic management is still mainly under the control of human
dispatchers, who usually do not have precise information about how the train traffic is evolving.
Thus their chosen actions for traffic control may often be sub-optimal, see Mannino & Mascis
(2009), Lamorgese & Mannino (2013).
One of the possible reasons is the complexity of finding effective dispatching decisions. These
require a good compromise between the quality of the solution, the time horizon (time span)
for which the traffic prediction is considered, and the computational effort. The consequences
of the measures adopted should also be known in advance. An accurate prediction of the
effects of delays and other disruptions requires the modelling of how train traffic evolves in
sufficient detail. This will reflect the actual state of the network, the dynamic behaviour of the
trains on that network, and the dispatching measures used to control traffic.

13.1.4 Real-time railway traffic management

Automated real-time train traffic management has, so far, received rather limited attention
in the technical literature. Possible reasons are the inherent complexity of the real time
process and the strict time limits for taking decisions. These leave small margins for using
computerised decision support systems. Another reason is that the inability to effectively
monitor and forecast operations limits the possibility for real-time schedule adjustments.
However, the service quality offered is related directly to the efficiency of the real time traffic
control, and any improvement here has a direct impact on customer satisfaction.
The design and implementation of advanced mathematical models is a prerequisite to the
development of innovative decision support systems, which are used for solving the real time
256
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 13.2 Review of the related literature

train traffic management problem. For this latter, a railway network can be seen mathematically
as the combination of potential train conflict detection (forecast operations and detection of
potential conflicts) and its resolution (find effective orders, routes, updated times that avoid
or solve the potential conflict). This conflict detection and resolution (CDR) problem has to
guarantee a conflict-free and deadlock-free schedule. Thus running time and headway time
constraints between operations need to be satisfied and to be compatible with the actual
status of the network. No train can be allowed to enter the network before its expected
entrance time, and the trains on that network must arrive and depart with the smallest possible
expected consecutive delay.
Fig. 13.3 shows an example where trains A and B can possibly be re-ordered in a simple
network.

Fig. 13.3: A simple network with two trains

13.2 Review of the related literature

This section reviews mathematical models and algorithms to support the task of local traffic
controllers (dispatchers) and network traffic controllers. The revision is limited to analysis of
the real-time traffic control. It does not include papers on timetabling, which are reviewed in
Chapter 8. Next, the rescheduling problems are classified in centralised approaches, and
co-ordinated approaches for large networks. The complexity of the networks considered
ranges from a simple junction to a set of dispatching areas. All approaches have a range of
dispatching actions that might include one or more of the following:
– Changes of times, orders, local routes in stations or interlocking areas
– Global routes across a wide area
– Cancelling train connections
– Cancelling or short turning train services
Moreover, the approaches can model potential conflicts explicitly (disjunctive formulations),
resulting in continuous time models, or implicitly (cumulative formulations), resulting in time-
indexed formulations. The solution approach can be finally exact or heuristic.

13.2.1 Centralised approaches

Among the centralised approaches for managing a dispatching area, Şahin (1999) formulates
a meet and pass problem as a job shop scheduling problem. Conflicts between up to 20 trains
are solved in the order that they appear at 19 meeting points. An algorithm based on look-
ahead measures detects potential delays and takes ordering decisions at merging or crossing
points, in order to minimise the average delays.
Wegele et al. (2007) use genetic algorithms to reschedule trains, with the objective of
minimising passenger annoyance from delays, changes of platforms and missed connections.

257
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
13 Rescheduling

The dispatching strategies adopted are dwell time modifications, adaptation of train speeds on
corridors and local rerouting within stations. Examples of its application on a large part of the
German railway network are reported for a single delayed train.
Rodriguez (2007) focuses on a real-time train conflict resolution problem and proposes a train
routing and scheduling system based on constraint programming. The experiments show that
a truncated branch and bound algorithm can find satisfactory solutions quickly for a railway
junction considering a few kilometres of tracks, traversed by up to 24 trains.
Törnquist & Persson (2007) introduce a model for dispatching trains in a railway network
with several merging and crossing points. A mixed integer linear programming problem is
formulated and solved with commercial software packages. Heuristic scheduling strategies are
proposed, to reduce the search space by restrictions of reordering and local rerouting actions,
by up to 80 trains for a 90-minute horizon of traffic prediction. Experiments are presented for
various disturbance settings on a Swedish railway network with over 250 track segments
traversed.
An exact method has been proposed by D’Ariano et al. (2007) for a train scheduling problem
with fixed routing. Their computational experiments, carried on the Dutch railway bottleneck
around Schiphol International Airport and for multiple delayed trains, show that optimal
or near-optimal solutions can be found within a short computation time. In two follow-up
papers D’Ariano et al. (2008) and Corman et al. (2010), this algorithm is incorporated in more
sophisticated meta-heuristic frameworks to manage multiple train rerouting options.
Caimi et al. (2012) and Fuchsberger (2012) report on the successful implementation of a
model predictive control approach for rescheduling trains at the central railway station area
of Berne, Switzerland. To reschedule train traffic, they solve an integer program taking into
account a number of operational constraints. Their objective is the maximisation of customer
satisfaction.
Meng & Zhou (2011) address the problem of taking robust meet-pass decisions where the
situation is uncertain. A stochastic programming with recourse framework is solved that
minimises the delays in the current train schedule due to disturbances. It also computes the
additional delay to be expected under different forecasted operational conditions, namely a
disruption with probabilistic duration.
Lusby et al. (2013) adopt a set packing inspired method for routing trains at a junction in real-
time. To solve the problem, the Authors develop a branch and price algorithm and show that
practical problems can be solved to near-optimality in a reasonable time.
Pellegrini et al. (2014) develop a comprehensive MILP model that considers multiple
scheduling and routing decisions simultaneously, with fine granularity. An evaluation of multiple
test cases and scenarios shows that it is possible to compute the optimal solution for up to 20
minutes of traffic quickly.

13.2.2 Co-ordinated approaches

The problem of co-ordinating the tasks of different dispatchers has seen less scientific
interest in recent years. Among the few papers on this subject, Jia & Zhang (1994) present an
approach based on fuzzy decision-making for distributed railway traffic control. A multi-level
decision process is described that consists of several regional decision centres that need to be
co-ordinated. The test case is a main Chinese network with 12 stations and 12 trains.
Lamma et al. (1997) propose a distributed advisory system which helps traffic controllers in
traffic management and control within railway stations and along railway branches. The

258
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 13.2 Review of the related literature

scheduling of trains along a railway line with up to 31 trains is performed by traffic control
modules, each one controlling a limited number of block sections and solving train conflicts
based on local decision rules. These modules need to exchange information about the local
traffic frequently and to verify the feasibility of the solution.
Chou et al. (1997) propose a distributed control system and study a number of railway areas
that are influenced mutually. A time-shift co-ordination strategy between neighbouring traffic
control areas is proposed for collaborative train rescheduling. Distributed control techniques in
neighbouring regions are applied in a fictitious network and evaluated in terms of delay cost.
Methodologies for railway traffic regulation and co-ordination of local areas were developed
within the European project COMBINE 2, see Pacciarelli (2003). Mazzarello & Ottaviani (2007)
report on the implementation of these methodologies for two test cases on the Dutch railway
network. They also report on a practical pilot carried out for one of the two test cases. Based on
the same architecture, Strotmann (2007) presents a two-level approach for rescheduling trains
between multiple areas. The computational complexity is divided between a lower level solver,
computing traffic control measures in each area, and a higher level solver, checking whether
neighbouring areas have consistent solutions. A co-ordinator graph is used to find out solutions
that are not feasible between neighbouring areas, which will result in additional constraints to
the first level solver. A solution at the higher level is found by imposing train ordering constraints
until a feasible schedule to the global problem is found, or it is proven that no globally feasible
schedule exists. Fictitious examples are studied with up to 16 trains and 73 block sections.
Corman et al. (2011) develop further ideas from the COMBINE 2 project to solve the co-
ordination problem for a complex and busy railway station divided into two dispatching areas,
and compare distributed and centralised systems. Recently, a branch and bound algorithm
has also been proposed by Corman et al. (2012), in order to compute optimal train scheduling
solutions for the distributed approach generalised to n > 2 areas. Optimal solutions are
computed for a number of disturbance handling scenarios. This approach works well for short
time horizons, and it was tested in the absence of heavy disruption. Corman et al. (2014)
develop this approach further by investigating the impact of advanced heuristics, together with
centralised and decomposed approaches and their effects on both the feasibility and quality
of solutions. These are assessed at both local and global levels. The approach is also able to
deal with the relevant case of disrupted traffic over multiple areas.

13.2.3 Discussion

In general, it can be observed that most of the existing approaches lack a thorough
computational assessment and limit the analysis to simple networks or simple perturbation
patterns. In fact, the analysed delay patterns are often quite specific, for example only one
train is delayed or the problem is limited to a single junction or to a straight line. Moreover, the
models used in the literature for the assessment are often simplified and do not capture the
consequences of delays and other disturbances entirely. For instance, when dealing with large
networks and dense traffic, possibly with heavy disruption, the risk of deadlock is relevant and
should not be ignored by real-time models. Safety rules must be modelled accurately in order
to avoid deadlock situations in the network. This is especially the case for what concerns the
reservation and occupation of each block section in busy railway corridors and in complicated
station areas. Precision should reach the level of seconds (not minutes) and a block or track
sections, for heavily used networks. The alternative graph model is among the few that
incorporate this microscopic level of detail within an optimisation framework.

259
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
13 Rescheduling

13.3 A real-time train traffic management approach

This section describes the functional implementation of a state of the art real time train traffic
management system, called ROMA (Railway traffic Optimisation by Means of Alternative
graphs), see D’Ariano & Pranzo (2009). It is designed to support dispatchers in the task of
managing disturbances, and offer support during disruptions. The real-time train traffic
management problem is divided into six sub-problems:
(i) data loading and exchange of information with the field,
(ii) assigning a passable route to each train in order to avoid blocked tracks,
(iii) detecting potential conflicts, solving potential conflicts by defining optimal train timings
and orders,
(iv) and routes,
(v) as well as the exact arrival and departure times at stations and at relevant points in the
network, such as junctions and passing points, and
(vi) ensuring minimum distance headway between trains while maintaining acceptable speed
profiles.
ROMA is able to address the resolution of all these sub-problems.

Fig. 13.4: ROMA train traffic management support system architecture

260
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 13.3 A real-time train traffic management approach

Fig. 13.4 presents the overall ROMA dispatching support system architecture, which is
composed of inter-related procedures. A human dispatcher can interact with the system by
adding or removing constraints, or changing the timetable. The function of each procedure is
as follows.
– Data loading (sub-problem (i)): Collect data from the field such as the current infrastructure
status, the existing timetable, the actual position and speed of all trains, and provide
an expected value to describe future events such as entrance times of the trains in the
network, dwell times, etc.
– Disruption recovery (sub-problem (ii)): Given a default route and a prioritised set of rerouting
options, find a passable routing for each train by avoiding, eventually, the tracks blocked in
the area under study.
– Traffic optimisation (sub-problems (iii), (iv) and (v)): Given a set of possible dispatching
actions such as updating orders, routes and departure times, find a new deadlock-free
and conflict-free schedule by rescheduling and/or rerouting trains. This relates to detecting
conflict (iii); adjusting the schedule in terms of orders and times (iv) and adjusting routes
for trains (v). In this phase, the minimum time for a train to traverse a block section is
considered fixed.
– Train speed co-ordination (sub-problem (vi)): Given the schedule computed by the
previous procedure, check its consistency with the train dynamics. If the blocking times
of trains in some block sections overlap, some restrictive signal will be shown to trains,
their speed profile has to be updated. In such a case, acceptable train speed profiles need
to be computed on the basis of the actual signal aspects, infrastructure and rolling stock
characteristics.

13.3.1 Load information

The data loading procedure periodically collects all the information from the field, which is
required by the other procedures (sub-problem (i)). The primary condition for calculating the
future train movements is the availability of a detailed and accurately updated data set. Real
time data are gathered from the field during operations. Clearly, a continuous and reliable
communication with the trains has to be assured; that is a real time data processing unit on
board and in the traffic control centre is necessary. Among the real-time data, the current
operational situation has to be included. Required are the actual positions and speeds of the
trains operating at the beginning of the time horizon being considered (at time t0). Additional
data relates to the expected entrance time/route for each incoming train, time windows
of availability for all block sections/platforms, possible temporary speed limits occurring at
some block sections, and additional scheduled stops on open tracks and stations with their
scheduled arrival and departure times.
Off-line (planning) data consists of detailed information about the infrastructure, timetable
and rolling stock characteristics. The timetable contains a list of arrival/departure times (time
windows of minimum/maximum arrival/departure times) for a set of relevant points in the
network, including all the station platform tracks visited by each train. The infrastructure is
described in terms of a set of available block sections delimited by signals. For each block
section the status, length, grade, speed limitations, traversing direction and maximum speed
are given. The route release and switching times are also known off-line.
The specific technical characteristics of the rolling stock of each train (power, train length and
expected weight, maximum speed and friction rates between rails and wheels) are considered
to be known off-line, in order to enable a re-calculation of the required minimum running time.

261
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
13 Rescheduling

The data associated with each train includes a prioritised list of routing options (the most
evident and frequently used alternative routes are selected by the dispatcher and given to
the support system), plus the (expected) acceleration and braking rates in order to be able to
compute the required speed profiles. The driver behaviour is assumed to be known in advance
and follows standard braking and acceleration profiles. The weather conditions, the train load
(number of passengers) and weight are assumed to be defined a priori. Although some data
may be different from day to day, for the purpose of rescheduling they are computed as off-line
data. If there is substantial real-time variation of these factors, a more accurate estimation of
the train’s movements should be considered.
Two types of speed profiles are distinguished. These are scheduled (used during the planning
phase) and operational (adopted in the conflict detection and resolution phase). Operational
speed profiles, used in the rescheduling process, assume that trains travel at their maximum
speed according to the train characteristics, infrastructure speed restrictions and adopted
standard driver behaviour, and are obtained by using off-line data.

13.3.2 Disruption recovery

The disruption recovery procedure checks if there are any block sections in the network that
are not available for use. This activity corresponds with the resolution of sub-problem (ii).
For each train, this procedure discards routes that are infeasible, sorts the passable routing
options on the basis of a priority list (given by traffic controllers) and then assigns the one with
the highest priority, called the default routing. The default routing of each train and the set of
remaining passable routes are then given to the real-time railway traffic optimisation procedure
of the ROMA dispatching support system. If no passable route is available for a train, the
system requests external support by the human dispatcher.

13.3.3 Real-time traffic optimisation

The real-time railway traffic optimisation procedure is the kernel of the decision of the train
dispatching support system. This is responsible for detecting and solving train conflicts,
while minimising the train delay propagation and maximising the dynamic utilisation of railway
capacity. Given all the necessary information by the data loading procedure and (at least) a
passable route for each train by the disruption recovery procedure, the following problems are
tackled.
Conflict detection (sub-problem (iii)). A conflict detection procedure detects potential conflicting
train paths (and possible deadlocks) in a given period of traffic prediction (e. g. 15 minutes
ahead). The conflict resolution problem (sub-problems (iv), (v)) addresses the predicted
conflicts, computing a new feasible schedule (deadlock-free and conflict-free) that solves the
predicted conflicts and is compatible with the actual train delays, and the network status. This
involves choice of orders and times (iv) and choice of routes (v) for all trains.
The conflict detection and resolution problem can be formulated as a job shop scheduling
problem, with no-store and no-swap constraints, see D’Ariano et al. (2007). This, in turn, can
be modelled through an alternative graph, representing the routes of all trains in a given control
area along with their precedence constraints (minimum headways). The main value of the
alternative graph is the detailed but flexible representation of the network topology, at the level
of railway signal aspects and operational rules. In case of fixed block signalling, each block
signal corresponds to a node in the alternative graph and the arcs between nodes represent

262
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 13.3 A real-time train traffic management approach

times between two operations. A train schedule corresponds to the set of the start time of
each operation.
Since a train must traverse the block sections in its route sequentially, a route is modelled in
the alternative graph with precedence constraints and a chain of associated nodes. At this
stage, the speed profile of the train is assumed to be fixed, thus giving a fixed running time for
each block section, except for a possible additional waiting time between operations in order
to solve train conflicts.
Since a block section cannot host two trains at the same time, a potential conflict occurs
whenever two or more trains require the same block section. In this case, a passing order
must be defined between the trains which are modelled in the graph by introducing a suitable
pair of alternative arcs for each pair of trains traversing a block section. A deadlock-free and
conflict-free schedule is finally obtained by selecting one of the two alternative arcs from each
pair, in such a way that there are no positive length cycles in the graph (i. e. deadlock). In
other words, the alternative arcs represent operational choices such as the order of trains at a
crossing or merging section. The real-time railway traffic optimisation procedure is in charge of
computing the first feasible schedule and is then looking for better solutions in terms of delay
minimisation. In order to evaluate a schedule, ROMA uses the maximum consecutive delay as
performance indicator of a solution, which is the maximum delay introduced when solving train
conflicts in the dispatching area. This is caused by the propagation of the input delays of late
trains to the other trains in the railway area.
Given a timetable, a set of passable routes associated with each train and the current status
of the network, the train scheduling procedure returns a feasible schedule for each train. This
defines its entrance time on each block section. Firstly, the default routings are considered by
the disruption recovery procedure. If a feasible schedule is found, the train rerouting procedure
verifies whether local rerouting options may lead to better solutions. For each changed route
the running times and setup times are modified accordingly. Whenever some route is replaced,
the train scheduling procedure computes a new deadlock-free and conflict-free timetable, by
rescheduling the train movements. The combined scheduling and rerouting procedure returns
the best solution found when a time limit of computation is reached, or no local rerouting
improvement is possible. The scheduling algorithms for sub-problems (iv) and (v) can be
grouped into three main categories.
Rule-based approaches: Well-known simple decision rules (for example a completely
automated version of the ARI system, see Berends & Ouburg (2005), using what-if scenarios),
timetable orders and routes, and priority rules are evaluated as the base case that simulates
the Dutch practice of traffic management.
Heuristics: Greedy heuristics are able to compute incrementally a set of train orders for
practical size instances within a short computation time. Concerning routings, local search and
tabu search procedures explore the solution space by trying train routes potentially leading to
a better schedule.
Optimisation-based approaches: Exact methods are explored with the objective of
computing near-optimal train schedules and/or to prove their quality, for practical size
instances within a short computation time.
If the real-time railway traffic optimisation procedure is unable to find a deadlock-free and
conflict-free schedule, the dispatcher has to carry out other types of timetable modifications.
These may include the introduction of new train routes, applying short-turning of trains in case
of track blockage, or even cancelling of train services at some stations. In fact, emergency
situations, such as extensive blockages, might need consideration of more effective actions,

263
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
13 Rescheduling

than the adjustment of train orders and routes. The challenge is the evaluation of conflict and
deadlock prone situations across large areas and multiple objective functions, representing the
many stakeholders involved. An example of the operational steps, models and a system for
automated support to follow in those situations is given in Corman & D’Ariano (2012).

13.3.4 Train speed co-ordination

The ROMA dispatching support system follows the common assumption of solving sub-
problems (iii), (iv), (v) with blocking times based on fixed speed profiles and the computed
waiting times. This does not model explicitly the dynamic consequences of braking and
subsequent acceleration imposed to avoid conflicts in the network. For the standard three-
aspect signaling system (such as the Dutch NS’54), this means that the time taken by each
train to traversing each block section is computed assuming green signal aspects, and that
trains have infinite deceleration and acceleration capabilities when facing/departing from a red
signal.
A speed profile feasibility check procedure verifies whether or not the schedule is compatible
with the actual train dynamics and signal aspects. The latter case might for instance
correspond to the situation which is not feasible, in which a train traverses a block section at
yellow (or other restrictive) signal aspect, without decelerating to an approach speed. In that
case, the minimum safety distance headways are not being respected and there is an overlap
of the blocking times assumed in the model.
The speed updating procedure adjusts the speed profile of a train in such a situation
according to typical driver behaviour in the case of variable signal aspects, and to the
dynamics of the rolling stock.
Feasibility checks and speed updates address and solve in succession all overlaps, possibly
addressing also new conflicts arising as a consequence of the adjustments. Through the
procedure, feasible computed orders and routes are considered, until all the overlaps of the
assumed blocking times are solved. The blocking times are then correctly recomputed. This
can be done in three possible ways:
Iterative scheduling strategy. The real-time optimisation, speed profile feasibility check
and speed updating procedures are performed iteratively. At each iteration, the real-time
optimisation procedure solves sub-problem (iii), using a scheduling algorithm, and calls the
feasibility check and speed updating procedures for the resolution of sub-problem (iv). For
any speed profile which is not feasible, another run of the real-time optimisation procedure is
performed to provide an acceptable speed profile for this train. The iterative procedure cannot
result in an endless sequence of adjustments.
Single scheduling strategy. This is a simplification of the iterative scheduling strategy. The
sub-problems (iii) and (iv) are solved in cascade. The real-time optimisation procedure solves
sub-problem (iii), using a chosen scheduling algorithm, and calls the speed profile feasibility
check and speed updating procedures until admissible speed profiles are calculated for every
train. Following this strategy, the sequencing obtained by the real-time optimisation procedure
is therefore not modified during the resolution of sub-problem (iv).
Energy-efficient strategy. Any braking due to a restrictive signal aspect and subsequent re-
acceleration to the planned speed result in a loss of time (with additional traffic delay) and
increased energy consumption. In order to reduce those negative effects, energy-optimal train
trajectories can be computed. The underlying idea is that the train is slowed down slightly
some time before such a possible conflict. This action may give way to the re-acceleration of

264
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 13.4 Traffic optimisation model

the train without it being hindered, so that it reaches the conflict area with an optimal distance
behind the train which is causing the conflict.

13.4 Traffic optimisation model

The Authors propose here the mathematical model used for problems (iii), (iv), and (v). The
sub-problem (iv) is next denoted as conflict detection and resolution with fixed routes
(CDRFR). The sub-problem (v) generalises the CDRFR setup by allowing for modification of
some train routes.

13.4.1 Alternative graph model

The alternative graph model, see Mascis & Pacciarelli (2002), generalises the disjunctive graph
formulation and is able to take into account several constraints arising in real-world scheduling
applications. In order to define formally an alternative graph, the following definitions are
introduced. A triple G = (N, F, A) is an alternative graph, where N is a set of nodes, F is a set of
fixed directed arcs and A is a set of pairs of alternative directed arcs.
N(F) denotes a subset of N such that for each node i  N(F) there is at least an arc of F
adjacent (connected) to node i. Set N includes two dummy nodes 0 and n, called start and
finish respectively, such that for each node i  N(F) there is a directed path from node 0 to
node i and from node i to node n in the graph G(Ø) = (N, F, Ø). A node is isolated if it is an
element of N but not an element of N(F).
Each node i is associated with an operation oi. There are a set of operations o0, o1, … , on to
be performed on m machines m1,m2, … ,mm, where o0 and on are (dummy) operations with
zero processing time. Note that the dummy nodes 0 and n are isolated only if there are no
other nodes in N. Operations are processed by resources; and a resource mi can process
only one operation at a time. An operation oi cannot be interrupted from its start time ti to its
completion.
The length of arc (i, j)  F is a given quantity bij. Each arc (i, j), either fixed or alternative,
represents a precedence relation constraining the start time (tj ) of oj with respect to the start
time ti of oi, i. e., tj ≥ ti + bij. Set F also includes outgoing arcs from node 0 and ingoing arcs to
node n.
In general, the alternative arcs of each pair ((i, j), (h, k))  A are decision variables in the
alternative graph model, i. e. a decision is made by choosing arc (i, j) V (or) arc (h, k). A graph
selection S is a set of alternative arcs obtained from A by choosing at most one arc from each
alternative pair. Given a pair ((i, j), (h, k))  A, arc (i, j) is selected in S if (i, j)  S, whereas arc
(i, j) is forbidden in S if arc (h, k)  S. Finally, the pair is called unselected if neither (i, j) nor (h, k)
is selected in S.
Given a selection S from A, G(S) indicates the graph (N, F  S, A). The selection S is
consistent if the graph G(S) has no positive length cycles. Given two nodes i, j  N(F), l S(i, j)
denotes the length of the longest path from node i to node j in G(S). The quantity ri = l S(0, i) is
the head of node i, and the quantity qi = l S(i, n) is the tail of node i. The longest path from the
start node 0 to the finish node n in G(S) is called the critical path of the graph. A selection S is
complete if exactly one arc from each alternative pair is selected. Given a consistent selection
S, an extension of S is a complete consistent selection S' such that S is part of the set S', if
this exists. A solution is a complete consistent selection S, and its value is therefore l S(0, n).
An optimal alternative graph solution is indicated as S*, while an empty selection is denoted

265
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
13 Rescheduling

as SØ. The objective function is the minimisation of the start time tn of operation on, i. e. the
makespan.
To summarise, the alternative graph formulation corresponds to the following mathematical
problem:
min tn − t0
s.t. tj − ti ≥ bij (i, j)  F
(tj − ti ≥ bij ) V (tk − th ≥ bhk ) ((i, j), (h, k))  A

13.4.2 CDRFR problem formulation

In the formulation of the CDRFR problem, each train (job) must pass through a prescribed
sequence of block sections (machines). An operation oi corresponds with the traversing of a
block section Bu for a train Ty, i. e. it is associated with the pair < Bu, Ty >. The route of Ty is
therefore a sequence of operations < B1, Ty >,< B2, Ty >, . . . ,< Bk, Ty >, where the operation
oi associated with the pair < Bu, Ty > is the successor of < Bu-1, Ty > and the predecessor of
< Bu+1, Ty >. The travelling of Ty through Bu (an operation) is a node of the alternative graph.
Any route of Ty is therefore a chain of nodes to be processed in sequence. Given an operation
oi, operations o(i) and oµ(i) are associated to the following and preceding block sections to be
travelled by the train on its route. The two additional dummy operations o0 and on represent
the start and end of the schedule, and each train route is modelled with a chain of fixed arcs
from node 0 to node n.
Train routes are represented with fixed arcs that correspond to the precedence constraints
between consecutive operations. For each operation oi of a train route, there is a fixed arc
(µ(i), i)  F with length bµ(i)i = pµ(i) and pµ(i) denotes the processing time of operation oµ(i).
The running time constraint imposes that oi can start only after at least pµ(i) time units from
the completion of oµ(i), i. e. ti ≥ tµ(i) + pµ(i). Furthermore, fixed arcs are used to impose other
constraints (e. g. release times, rolling stock constraints, connecting train services) and to
compute the train delays.

13.4.2.1 Setup time constraints

For each block section to be traversed by at least two trains, the set of all associated
operations is considered. Since a block section cannot host two trains at the same time,
alternative arcs are adopted to model the ordering constraint of the CDRFR problem, i. e. are
used to determine the train orders. By letting oi and oj be the two conflicting operations at the
railway junction, the two possible processing orders are modelled with the pair of arcs ((k, j),
(h, i))  A. If oi is scheduled before oj, the alternative arc (k, j) is selected and TA precedes TB.
The length of the alternative arc (k, j) corresponds to the headway time aij between oi and oj
(tj ≥ tk + aij). Similarly, if oj is scheduled before oi, then TB precedes TA and the length of (h, i)
is the headway time aji (ti ≥ th + aji ). If ti < tj < tk + aij or tj < ti < th + aji, then the two trains are
claiming the block section concurrently, and the solution of the conflict will correspond to a red
signal shown to one of the two trains. Specifically, if oi is processed first, then oj must wait for
the completion of the running time pi and the starting of o(i) plus the headway time aij between
oi and oj. The potential conflict between TA and TB has been detected and solved by choosing
alternative arc (k,j) for the CDRFR problem. It might thus result in increase of the start time tj of
operation oj.

266
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 13.4 Traffic optimisation model

Fig. 13.5 shows the alternative graph formulation for two trains (TA and TB) at a junction. A
fixed arc (i, k) (depicted with solid arrows in Fig. 13.5) represents the travelling of TA through
the junction, i. e. the block section delimited by two signals. The arc length bik corresponds to
its running time pi. Similarly, the arc length bjh corresponds to the running time pj of TB through
the junction. Since the junction cannot host TA and TB at the same time (i. e. two jobs in the
alternative graph require the same resource), there is a potential conflict.

Fig. 13.5: The alternative graph of two trains at a junction

A processing order must be defined between the operations in potential conflict, and this
constraint is modelled by a suitable pair of alternative arcs from the set A (depicted with
dashed arrows in Fig. 5). Each alternative arc models a possible precedence relationship
between two operations, and the length of an alternative arc is the minimum time headway
between the associated trains.

13.4.2.2 Objective Function and Scheduled Stop Formulation

Typically, the punctuality of railway operations is measured at relevant points in the network, on
the basis of the timetable times. The Authors show in what follows that by choosing suitable
lengths on the arcs entering node n, representing the due times, the maximum consecutive
delay of the solution is l S(0, n). Given an initial selection SI, potentially empty, the objective of
the CDRFR problem is to find a complete consistent selection S, such that SI is part of S and
the length of the longest path from node 0 to node n in G(S) is minimised. The selection SI
represents the precedence constraints due to the initial positions of the trains at time t0 and/or
to their order of entering the network.
Let yu be the scheduled arrival time of a train Ty at a block section Bu in the timetable, which
might not be met in case of real-time disturbances, i. e. the train will be late. Let yu be the
earliest possible arrival time of Ty at Bu. This is computed according to its initial position,
initial speed, assigned route and following a maximum speed profile (allowed by the train
characteristics and infrastructure) in the empty network (i. e. by disregarding the presence of
other trains). yu does not take into account possible conflicts with other trains, and therefore
yu is a lower bound on the feasible arrival time of Ty at the relevant point Bu. The total delay of
Ty at Bu is defined as the difference between its actual arrival time ti and yu.
The total delay can be divided into two parts as follows. If yu > yu, then the quantity
yu − yu is an unavoidable delay that cannot be recovered by real-time rescheduling of train
operations. The quantity max{0, yu − yu} is the primary (initial) delay of Ty at Bu. The quantity
max{0, ti − max{yu, yu}} is the consecutive (secondary) delay of Ty at Bu, which is the
additional delay due to the solution of conflicts between Ty and the other trains running in the
network. In order to avoid accounting for the initial delays in the scheduling phase, a modified
due date is defined for Ty at Bu as max{yu, yu}. Adding an arc from node i  N(F) to node n

267
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
13 Rescheduling

(with length − max{yu, yu}) for each relevant point, the makespan l S(0, n) corresponds to the
maximum consecutive delay.
The formulation of the objective function for arrival and departure at a scheduled stop is further
explained in Fig. 13.6. Two nodes i and h are required for a train TA stopping at a scheduled
stop Q. Let oi be the operation associated to the entrance of TA at the block section (relevant
point) Bu with the scheduled stop Q, and pi the running time of TA at Bu until the complete stop
at the end of the block section. Let oh be the operation associated with the scheduled dwell
time sh of TA at Bu, i. e. the scheduled stopping time at operation oh. Let ok be the operation
associated with the entrance of TA at its next block section, with running time pk.

Fig. 13.6: The alternative graph formulation of a scheduled stop

In Fig. 13.6, the value dAQ [AQ] represents the scheduled departure [arrival] time of TA at Q,
respectively. To be precise, TA is not allowed to depart from Q before dAQ and is considered
late if arriving after AQ. The constraint imposed by the timetable on the scheduled departure
time dAQ (tk ≥ dAQ) is obtained by adding the fixed arc (0, k) with length dAQ.
Similarly to the previous subsection, by adding a fixed arc (h, n) from node h [ N(F) to n,
having length − max{AQ, AQ}, the length of the path from node 0 to n passing through (h, n) is
equal to the consecutive delay of TA at Q.

13.4.2.3 Solution

A solution to the CDRFR problem is a feasible start time ti for each operation oi (i. e. the exact
time at which each train will enter each block section) such that all fixed precedence relations
are satisfied, exactly one of each pair of alternative precedence relations is selected and the
resulting alternative graph has no positive length cycles. Note that a positive length cycle
represents a deadlock situation which is not feasible. Fig. 13.7 shows a traffic example with
three trains, the time-distance plot of a solution and the corresponding alternative graph.

13.4.2.4 Additional railway constraints

The alternative graph formulation allows the modelling of situations more general than the ones
already mentioned. For example, Oliveira & Smith (2000) point out the relevance of precedence
and meeting constraints in the management of railway operations. A precedence constraint
between two trains at a station may be necessary if one train is carrying the crew or a vehicle
necessary for another train service (rolling stock connections). In such cases the latter train
must wait for the arrival of the former one before departing. This constraint can be modelled by
simply adding one fixed arc of suitable length from the node associated with the arrival of the
former train to the node associated with the departure of the latter.
A meeting constraint allows the specification of a minimum dwell time, during which two given
trains must be together at a given station for exchanging passengers or goods (passengers

268
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 13.4 Traffic optimisation model

Fig. 13.7: Time-distance plot and the graph for a schedule with three trains

transfer). This can be modelled by adding two fixed arcs of suitable length. These consist of
an arc from the node associated with the arrival of the former train to the node associated with
the departure of the latter one, another arc from the node associated with the arrival of the
latter train to the node associated with the departure of the former one. A unilateral passenger
connection is also possible.
Further extensions of the alternative graph are to model train traffic running according to a
green wave strategy that allows only dwell time modifications and no change to the speed
profile of trains. Thus no yellow and no red signals are shown to train drivers.

13.4.3 CDR problem formulation

An alternative graph represents all the scheduling alternatives when train routes are defined.
As far as the CDR problem is concerned, changing a route implies changing the set F and A
in G. If a train can be assigned to different routes, then F is a decision variable and A depends
on the choice of F. Note that each arc of F corresponding to the traversing of a block section
has a length equal to the running time of the associated train on that block section. In fact, this
depends also on the train speeds on the previous/following block sections. Changing part of a
route for a train may therefore cause adjustment to its running time on several block sections
that have not been modified. In such cases, new fixed arcs replace all the arcs of F that are
associated with a different block section, or to the same block section but with a different
running time. Once a route-set F is defined, S(F) is a complete consistent selection obtained
from A in G, and G(N, F  S(F), A) is the alternative graph solution. A solution to the CDR
problem can be therefore be expressed compactly as the pair (F, S(F)).
The length lF,S(F)(0, n) of the critical path in G(F, S) is the maximum consecutive delay. This value
can be reduced by either changing the train sequencing (i. e. the set S(F), the CDRFR problem)
or modifying the train routes (i. e. the set F ). An optimal CDR solution (F*, S*(F*)) is thus to find
an optimal route-set F* and an optimal graph selection S*(F*).

269
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
13 Rescheduling

The next illustration shows the concept based on the example of Fig. 13.8, a network with 14
block sections and three trains.

Fig. 13.8: A small network with three trains

In the timetable, the default route of TA is given by the sequence of operations A1, A2, A3,
A9, A12, A13, A14, even if the alternative routes A1, A2, A3, A9, A10, A5, A13, A14 and
A1, A2, A3, A4, A5, A13, A14 are available for this train in order to reach the exit at block
section 14. The route of TB is B7, B8, B9, B10, B5, B6 and the route of TC is C11, C8, C9,
C10, C5, C6. TB and TC share the same path from block section 8 to 6, which is crossed
by the default route of TA on block section 9. TB is a slow train entering block section 7 at
release time 0. Its running time on each block section is 20 time units. TA and TC are fast trains
requiring a running time of 10 time units on each block section. Their release time is 60 and
40, respectively. There are only two points that are relevant for the timetable, namely the exit of
block sections 6 and 14. The timetable requires that TA, TB and TC exit the network within time
131, 160 and 122, respectively. Each train, individually, would be able to exit the network on
time, i. e. the initial delay of each train at its relevant point is zero.

Fig. 13.9: Alternative graph formulation of the proposed example

Fig. 13.9 shows the alternative graph model of the CDRFR problem of Fig. 13.8. A node is
denoted with the pair (train, block section) of the associated operation or with the pair (train,
exit point), except for dummy nodes. Alternative pairs are depicted using dashed arcs and
for simplicity their length (i. e. the setup time) is assumed to be always equal to zero. The
three fixed arcs departing from node 0 model the release time of each train, whereas the
arcs entering node n model the objective function. Note that the current alternative graph G
= (N, F, A) has three isolated nodes (A4, A5 and A10) that can be used in order to implement
alternative routes.
Fig. 13.10 shows the optimal solution to the alternative graph of Fig. 13.9. In this schedule
TB always precedes TC, while TA follows the other trains on block section 9. The longest path
passes through the nodes 0, B7, B8, B9, B10, B5, B6, Bout, C6, Cout and n. The optimal
value lF,S(F)(0, n) is 8.
Consider now the CDR problem, in which train routes (i. e. the set F of G) can be changed,
and choose the new route A1, A2, A3, A4, A5, A13, A14 for TA (see the network of Fig. 13.8).

270
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 13.5 Conclusions

Fig. 13.10: Optimal rescheduling solution

Clearly, a different alternative graph G' = (N, F', A') has to be adopted to represent the new
CDRFR problem, in which TA crosses the routes of TB and TC at block section 5. In the new
alternative graph, the isolated nodes are A9, A10 and A12. Fig. 11 shows the optimal solution
to G', in which TC precedes TB, while TA follows TC and precedes TB. The longest path, of
length 0, passes through the nodes 0, C11, C8, C9, B8, B9, B10, B5, B6, Bout and n. This
solution is also optimal for the compound CDR problem, i. e., (F*, S*(F*)).

Fig. 13.11: Optimal rescheduling and rerouting solution

13.5 Conclusions
In the traffic management of railway networks, the problem of finding conflict-free and
deadlock-free schedules is faced by railway practitioners both in the timetable design and
during train operations. This Chapter focuses on the real-time traffic management aspects of
the problem. When train operations are perturbed, a new plan of feasible arrival and departure
times is computed, such that deviations from the original are minimised.
This Chapter illustrates detailed discrete optimisation models. The train rescheduling problem
is viewed as a job shop scheduling problem with no-store and other additional constraints. A
careful estimation of space and time separation between trains is modelled via the alternative
graph formulation, which is able to treat efficiently the no-store aspect of the train scheduling
problem for trains travelling in the same or opposite directions, while satisfying the signalling
and safety constraints.
The models presented can be extended to include further railway constraints in the traffic
optimisation model, such as constraints on the maximum time allowed for consecutive delays,
different block signalling technologies, and others. However, the models are deterministics,
i.e., they assume full knowledge of future operations and full compliance of train traffic to the

271
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
13 Rescheduling

predictions. In other words, stochastic running and setup times, as well as the uncertainty
causing disturbances and delay can be modelled only to a limited extent. A practical
implementation of such a system should consider carefully those aspects, by means of closing
the control loop around operations. Moreover, those systems should be implemented in real
life as decision support systems for dispatching, by presenting a carefully designed interface,
and proposing stable and consistent solutions to the human operators, especially in highly
disturbed operations. Finally, the rescheduling decisions should be always taken sufficiently
in advance and communicated timely to the involved stakeholders (including dispatchers, rail
operators and passengers) to leave them enough time to smoothly react to the updated train
orders and routes.

References
Berends N., Ouburg N. (2005): Beschrijving ARI-functionaliteit (Technical Report No. 20).
Utrecht, the Netherlands: ProRail Internal Specification (in Dutch).
Caimi G., Fuchsberger M., Laumanns M., Lüthi M. (2012): A model predictive control
approach for discrete-time rescheduling in complex central railway station areas, Computers
and Operations Research 39 (11), pp. 2578 – 2593.
Chou Y. H., Weston P. F., Roberts C. (2007): Dynamic Distributed Control for Realtime
Rescheduling of Railway Networks. In: Hansen I. A., Radtke A., Pachl J., Wendler E. (Ed),
Proc. of the 2nd International Conference on Railway Operations. Modelling and Analysis,
Hannover, Germany.
Corman F., D’Ariano A., Pacciarelli D., Pranzo M. (2010): A tabu search algorithm for rerouting
trains during rail operations. Transportation. Resesarch. Part B 44 (1), pp. 175 – 192.
Corman F., D’Ariano A., Pacciarelli D., Pranzo M. (2011): Centralised versus distributed
systems to reschedule trains in two dispatching areas. Public Transport 2 (3), pp. 219 – 247.
Corman F., D’Ariano A., Pacciarelli D., Pranzo M. (2012): Optimal inter-area coordination of
train rescheduling decisions, Transportation. Research. Part E 48 (1), pp. 71 – 88.
Corman F., D’Ariano A., Pacciarelli D., Pranzo M. (2014): Dispatching and coordination
in multi-area railway traffic management. Computers and Operations Research (44),
pp. 146 – 160.
Corman F., D’Ariano A. (2012): Assessment of advanced dispatching measures for recovering
disrupted railway situations. Transportation Research Records: Journal of the Transportation
Research Board. 2289/2012.
D’Ariano A., Corman F., Pacciarelli D., Pranzo M. (2008): Reordering and local rerouting
strategies to manage train traffic in real-time. Transportation Science 42 (4), pp. 405 – 419.
D’Ariano A., Pacciarelli D., Pranzo M. (2007): A branch and bound algorithm for scheduling
trains in a railway network. European Journal of Operational Research 183 (2), pp. 643 – 657.
D’Ariano A., Pranzo M. (2009): An advanced real-time train dispatching system for minimising
the propagation of delays in a dispatching area under severe disturbances. Networks and
Spatial Economics 9 (1), pp. 63 – 84.
Fuchsberger M. (2012): Algorithms for rail traffic management in complex central station areas,
Dissertation. ETH No. 20398, ETH Zurich.
Goverde R. M. P. (2005): Punctuality of Railway Operations and Timetable Stability Analysis.
PhD thesis, TU Delft.

272
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

Hemelrijk R., Kruijer J., De Vries D. K. (2003): Schiphol tunnel 2007. Description of the
situation (Technical Report No. 22). Utrecht: Holland Railconsult.
Jia L.-M., Zhang X.-D. (1994): Distributed intelligent railway traffic control: A fuzzy-decision
making-based approach. Engineering Applications of Artificial Intelligence 7 (3), pp. 311 – 319.
Lamma E., Mello P., Milano M. (1997): A distributed constraint-based scheduler. Artificial
Intelligence in Engineering 11 (2), pp. 91 – 105.
Lamorgese, L., Mannino, C. (2013): The track formulation for the train dispatching problem,
Electronic Notes in Discrete Mathematics 41, pp. 559 – 566.
Lusby R. M., Larsen J., Ehrgott M., Ryan D. M. (2013): A set packing inspired method for real-
time junction train routing, Computers and Operations Research 40 (3), pp. 713 – 724.
Mannino C., Mascis A. (2009): Optimal real-time traffic control in metro stations. Operations
Research 57 (4), pp. 1026 – 1039.
Mascis A., Pacciarelli D. (2002): Job shop scheduling with blocking and no-wait constraints.
European Journal of Operational Research 143 (3), pp. 498 – 517.
Mazzarello M., Ottaviani E. (2007): A traffic management system for real-time traffic
optimisation in railways. Transportation Research. Part B 41 (2), pp. 246 – 274.
Meng L., Zhou X. (2011): Robust single-track train dispatching model under a dynamic and
stochastic environment: a scenario-based rolling horizon solution approach. Transportation
Research Part B 45 (7), pp. 1080 – 1102.
Oliveira E., Smith B. M. (2000): A Job-Shop Scheduling Model for the Single-Track Railway
Scheduling Problem (Technical report No. 21). England: School of Computing, University of
Leeds.
Pacciarelli D. (2003): Deliverable D3: Traffic Regulation and Cooperation Methodologies – code
WP4UR DV 7001 D. Project COMBINE 2 “enhanced COntrol centres for fixed and Moving
Block sIgNalling systEms – 2” Number: IST-2001-34705.
Pachl J. (2002): Railway Operation and Control. Mountlake Terrace: VTD Rail Publishing.
Pellegrini P., Marlière G., Rodriguez J. (2014): Optimal train routing and scheduling for
managing traffic perturbations in complex junctions. Transportation. Research. Part B 59,
pp. 58–80.
Rodriguez J. (2007): A constraint programming model for real-time train scheduling at
junctions. Transportation Research Part B 41 (2), pp. 231 – 245.
Şahin I. (1999): Railway traffic control and train scheduling based on inter-train conflict
management. Transportation Research Part B 33 (7), pp. 511 – 534.
Strotmann C. (2007): Railway scheduling problems and their decomposition. PhD thesis,
Universität Osnabrück, Germany.
Törnquist J., Persson J. A. (2007): N-tracked railway traffic rescheduling during disturbances.
Transportation Research Part B 41 (3), pp. 342 – 362.
Wegele S., Slovak R., Schnieder E. (2007): Real-time decision support for optimal dispatching
of train operation. In: Hansen I. A., Radtke A., Pachl J., Wendler E. (Ed.), Proceedings of
the 2nd International. Conference on Railway Operations Modelling and Analysis, Hannover,
Germany.

273
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 14.1 Aims of Performance Evaluation

14 Performance Evaluation
Ullrich Martin

14.1 Aims of Performance Evaluation

The organisational structure of rail transport in Europe has changed fundamentally within the
last few years. The separation of infrastructure managers and train operators increases the
complexity of the railway system, as reliability, availability and efficiency depend on the co-
operation of different companies and requires a subtly differentiated analysis of the interests of
the various parties.
Typically, the railway’s performance is examined through technical, business-oriented and
overall economic evaluations. Although these evaluations are inter-dependent, there is no
standardised assessment base for measurements and indicators. Depending on the aim, the
railway and the country concerned, different approaches as well as divergent definitions of
efficiency and service quality are used for performance evaluation. In general, reliable, punctual
and fast transport of passengers and goods at minimal cost are essential for increasing the
competitiveness of railways. Apart from a robust technological design, economic aspects are
very important for successful operation.
Performance evaluation consists of a model of the factors influencing the effect of rail
transport by means of criteria and representative indicators. A comparison is made with
reference to some subjective or valid standards for a given period of time. This may include the
comparison standards to be used, the degree to which the goal should be achieved, as well
as the way the factors will be recorded, measured or estimated. The substantial changes in
customer relations due to open access lead to a modification of the basic aim in a system for
performance evaluations (Fig. 14.1).

Fig. 14.1: Modified Aim System and Target Groups

275
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
14 Performance Evaluation

The aims of the different target groups are interdependent and reliability and availability are
essential requirements to increase the efficiency of railways. A funded efficient infrastructure
is a prerequisite for high performance by the infrastructure manager, while high capacity,
availability and reliability of the track, power supply, signalling and telecommunications
is a basis for the customer-oriented performance of the train operating companies. Only
satisfactory services by the train operators lead to more railway passengers and shippers. The
more the railway service is used, the lower are the negative environmental effects of road traffic
in general, and it makes more effective use of limited funding resources.
Global or clustered partial goals of performance evaluation are to be defined in the context of
the target groups. Quality of operation is an indicator that may be used globally, or related to
specific target groups. If the quality of operation is poor, the entire railway system is unlikely
to be efficient. The European Directive statement is rather vague: “The Infrastructure Manager
and the Railway Undertaking shall have processes in place to monitor the efficient operation of
all the services concerned”, see EU (2006). Most performance evaluation methods applied in
rail transport are directly or indirectly based on quality of operation.
Performance evaluations vary a lot and are not clearly differentiated from each other. In the
following sections, those related to railway operation with more clearly defined goals are
discussed, while the appraisal of infrastructure investment is only considered marginally. See
Chapter 15 for explanation of the various approaches that can be taken to measuring the
economic performance of railways and the issues that arise in applying them.

14.2 Methods of Performance Evaluation

14.2.1 General Remarks

Typical features of the railway are very high fixed costs of the infrastructure and the longevity
of the capital goods. The performance of railways, however, is defined by its product, which
is the scheduled movement of persons and goods by rail, during much shorter periods.
The evaluation of transport performance, in general, is on the basis of inputs, outputs and
consumption, see Lan et al. (2006). The inputs consist of the number of personnel and the
technical characteristics, rolling stock, power supply and its distribution, while the quantitative
outputs are given by the amount of transport work delivered, such as passenger train-km or
freight train-km. Effectiveness is expressed by the services rendered (number of passengers,
passenger-km, tonnes lifted and tonne-km), and the associated revenues. Efficiency of
the transport services results from certain input-output relationships, e. g. the number of
passenger journeys made or passenger-km per employee and per train respectively, as well as
the average revenue per employee, per train and per train-km respectively.
However, quantitative criteria alone are not sufficient to describe the performance of the
railways and their external effects (e. g. environmental) related to the expenditure. Therefore, a
comprehensive evaluation of the performance of railways should also take the qualitative input,
output and consumption parameters into account. See Chapter 15 and Hansen et al. (2012).
In many cases, it is not sufficient to look only at railway traffic performance, which is described
easily by quantitative indicators for its effectiveness and efficiency. The conditions under
which it is provided must also be investigated. Speed, punctuality and reliability are common
qualitative criteria for railway traffic. The use of statistical indicators, process analyses and
simulations are typical approaches to carry out performance evaluations.

276
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 14.2 Methods of Performance Evaluation

14.2.2 Applying Effectiveness and Efficiency as Indicators

Technical and economic indicators are similarly taken into account in the case of combined
output-outcome evaluation methods. Monitoring of non-monetary performance information
is a basic principle of these approaches. The combined evaluation model takes both the
non-monetary components and monetary economic characteristics into account. Using
a reasonable abstraction of indicators and considering the specific framework conditions
in the form of parameters, several enterprises of different structures can also be compared
with the help of performance evaluations. Effectiveness and efficiency are typical general
and comprehensive indicators. Effectiveness is the ratio of the result of the work (output),
to the achieved effect (outcome), and means that the defined goals are fully accomplished
by the implemented actions. Efficiency either means the achievement of the best possible
result within the given resources, or obtaining a defined result with the least amount of effort
(economic principle).
Fig. 14.2 shows a possible combined performance evaluation model that can be used for the
comparison of different railway companies, railway systems, or of railways and other means
of transport in the passenger sector. With the inclusion of tonne-kilometres as a reference
quantity, the model can also be applied to goods. These reference quantities usually already
exist or can be derived easily. The combination of technical, economic and ecological effects
facilitates a strongly-abstract, comprehensive evaluation.

Fig. 14.2: Performance Evaluation based on Effectiveness and Efficiency

Lan et al. (2006) suggest implementing a comprehensive model for performance evaluation
based on stochastic distance functions that consider inefficiency and ineffectiveness
effects. They apply this approach to compare 39 worldwide railway systems based on their
effectiveness and efficiency. A set of Key Performance Indicators that can be used for a
comprehensive benchmark analysis is recommended in Hansen et al. (2012). The approach
includes relevant technical and economic key performance criteria and is based on standard
parametric (regression) analysis of empirical technical and commercial data.

277
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
14 Performance Evaluation

The influence of punctuality is considered indirectly. If punctuality deteriorates then the train
operator loses customers, and may want to reduce the frequency if the franchise allows for
it. Thus, the consumed infrastructure capacity is decreasing (see also Section 2.7). Finally, the
efficiency of infrastructure management and use of public funds is also declining.
If sufficient statistical data is available, the dependencies between different partial indicators
and the effectiveness or the efficiency may be determined. Combined output-outcome
performance evaluations are particularly suitable for assessing the demand-oriented quality of
transport services. An additional advantage of output-outcome performance evaluation is that
the partial indicators do not have to be transformed into a unified measurement. Effectiveness
and efficiency are assessed in a multi-scale model. If the economic partial indicators
are excluded, then this method can also be used in a simplified manner without monetary
influence factors.

14.2.3 Assessing Punctuality

Punctuality is a particularly suitable measure of the quality of operation or service, and seems
to be quite easy to apply. However, there are significant differences for using punctuality
as evaluation parameter. Punctuality can be related to trains or passengers. If a train is on
schedule, this does not mean yet that all passengers inside this train are also punctual. A
passenger is delayed inside a punctual train if he has missed the previous connection due to a
delayed feeder train. However, it is also possible that a passenger trip is punctual when using
a previously delayed train instead of the originally scheduled one.
It has also to be defined where punctuality is to be recorded. The “reporting points”, see EU
(2006), should be distributed over the whole network and not limited to terminal stations.
If the punctuality of a train is only recorded on arrival at the final destination, a high level of
punctuality may be the result of sufficient recovery time on previous route sections, although
a large number of passengers may have experienced considerable delays at earlier stations.
Therefore, additional reporting points should be situated at least at major intermediate stations
where the trains have high occupation rates. If passenger counting devices are available or
seat reservation is obligatory, actual train occupation rates may be used for estimating the time
loss experienced by the passengers.
There is no standard definition for punctuality yet. On German railways, a long-distance train
is delayed if it is more than five minutes late at final destination, whereas the corresponding
measure in Great Britain is delays of 10 minutes and more. In the Netherlands, the critical
threshold is fixed at three minutes. The measurements are often based on automatic track
occupation records of train numbers at home signals corrected by a certain standard off-set
due to the estimated running time until the platform stop position. For the determination of the
punctuality threshold value, the following aspects should be considered:
– Minimum headways
– Recovery times and the times between trains
– Technical equipment for a standard time measurement
– Service level expected by the passengers (related to punctuality at destination using
absolute or relative measures, depending on distance travelled as well as the number of
transfers from one train to another)
Furthermore, attention has to be paid to the trade-off between punctuality and speed. The
more recovery times are included in the schedule, the lower is the probability of delays but the
longer is the overall time taken. The longer the scheduled times between trains, the lower is

278
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 14.2 Methods of Performance Evaluation

the probability of delay propagation from one train to another. Lower, too, is the capacity of the
infrastructure.
For a scientific assessment and modelling, it is necessary to calculate not only the total time
loss or percentage of trains that are delayed by more than a certain threshold value, but also
to estimate the mean and analyse the shape of the underlying distribution of delays. As initial
delays and original delays can hardly be influenced by the operational processes taking place
inside the area investigated, consecutive delays are used as a separate measure of quality of
operation. Due to the scheduling connected with the required timetable recoverability, delays
are not equally distributed random events. Short delays occur very often, longer delays seldom
(see Chapter 11 for more details).
The performance evaluations that directly use punctuality as a central indicator can be carried
out by direct-demand or modal-split approaches. The direct-demand models consider only
the conditions within the railway system, and the change of the demand is derived from that,
see WABB Maidment (1988). Modal-split models include also the boundary conditions of
the other means of transport in order to determine the change of demand (Walther, 1991).
Although the modal-split approaches are more comprehensive, direct-demand approaches
may be very useful if the boundary conditions of the other means of transport are not known,
or the influences for changing of these conditions can not be determined. Modal-split models
can be distinguished into two variants:
– The supply components are assessed according to the benefit for the passengers. The
mode that offers the maximum benefit is selected.
– The supply components are transformed into (traffic) resistances. The expected traffic flows
are estimated according to the sums of resistances.
German Federal State Railways introduced a system of three service levels to evaluate the
quality of operation (Table 14.1) that is still being used today, see DB AG (2008). However, the
different service levels describe merely the ability of the infrastructure and timetable to recover
(or not) from initial and original delays. The impact on passenger delays is not considered
directly.

Levels of quality of operation

Delay Grade
Ratio of the sum of Decreasing strongly Better than necessary
delays of trains leaving (DC < 0.5)
the investigation are
Decreasing clearly Economically optimal Fully
and delays of trains
(0.5 ≤ DC < 1.0) acceptable
terminating within
the investigation area Without change or Acceptable
divided by the sum of increasing slightly fraught with
initial and original delays (1.0 ≤ DC ≤ 1.3) risk
within the investigation
area Increasing clearly Insufficient
(1.3 < DC)

Table 14.1: Levels of quality of operation according to Martin et al. (2013) based on DB AG (2008)
(DC: delay-coefficient)

Ackermann (1998) used a comprehensive direct-demand approach for assessing the influence
of travel time, cost and punctuality on the performance of the railway system by means of
utility assessment or Conjoint-Measurement of the direct benefit of railway passengers, see
Hamman et al. (1994). The method comprises three steps:

279
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
14 Performance Evaluation

– Measurement of preferences
– Analysis of preferences
– Forecast of behaviour
More than 7,000 railway passengers were interviewed concerning the partial indicators of
journey time, transfer need from one train to another, fare paid, and delay. The recorded data
were evaluated, weighted and fitted with statistical functions. The assessment may be based
on a linear, an exponential, as well as a logistic function. The logistic function fitted best to
empirical data but three parameters are to be estimated:
1
F(x) = (14.1)
1
+ a ⋅ bx
k
k limit to which the function converges
a, b additional parameters that have to be estimated

In addition, the arrival delays of more than 10,000 trains at final destination were considered.
The weighted data of all partial indicators were transformed into monetary values. The
summarised amounts of the partial indicators were used for the evaluation of the quality of
operation with direct consideration of the benefit for the railway passengers. The evaluation is
based on the combination of the loss of receipts and additional operating costs due to train
delays and can be applied separately for all target groups.
The ratio between the real (objective) delays and the delays perceived by the passengers,
called the general weighting factor of delay, amounted to about 2.8 with long-distance railway
journeys. This means that if the actual delay has been reduced by about three minutes, then
a reduction in delay of just one minute is perceived by the passengers. Based on the general
weighting factor of delay, passenger-optimal recovery times could be derived. The train
operators and infrastructure managers are encouraged to calculate the loss of receipts due
to delays. The business-oriented loss of receipts is shown as costs per minute of delay in
Table 14.2.

Kind of train Loss of receipts Marginal costs of Sum

[EUR/min.] operation [EUR/min.] 1995
1995 (2012) [EUR/min.] (2012)
1995 (2012)
InterCityExpress 137.50 (177.81) 5.50 (7.11) 143.00 (184.92)
InterCity/EuroCity 66.00 (85.35) 5.50 (7.11) 71.50 (92.46)
InterRegional 22.00 (28.45) 3.50 (4.53) 25.50 (32.98)

Table 14.2: Weighted business-oriented costs of delays per minute Ackermann (1998) – converted
into EUR, price level 1995 (updated to price level 2010) based on DESTATIS (2013)

With the inclusion of external influences such as emissions and the impact of accidents, the
overall economic weighted costs of delay in railway transport can be determined (Table 14.3).
The total costs of a delay minute can be calculated by the sum of the values from the
Tables 14.2 and 14.3.
Finally, determining the weighted costs of delay for the partial indicators can also be used for
the potential estimation of measures for improving the quality of operation. The costs saved by
reducing delays can be compared directly with the necessary expense for it.

280
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 14.2 Methods of Performance Evaluation

Kind of train Costs of Emissions Accidents Travel time Sum

operation (railway and (motorised (railway [EUR/min.]
(motorised motorised individual transport) 1995 (2012
individual individual transport) [EUR/min.]
transport) transport) [EUR/min.] 1995 (2012)
[EUR/min.] [EUR/min.] 1995 (2012)
1995 (2012) 1995 (2012)
InterCityExpress 150.00 39.00 11.00 36.50 236.50
(193.98) (50.43) (14.22) (47.20) (305.83)
InterCity/ 78.00 22.00 6.00 24.50 130.50
EuroCity (100.87) (28.45) (7.76) (31.68) (168.76
InterRegional 33.00 12.00 2.50 12.50 60.00
(42.67) (15.52) (3.23) (16.16) (77.59)

Table 14.3: Weighted overall economic costs of delays per minute Ackermann (1998) – converted
into EUR, price level 1995 (updated to price level 2010) based on DESTATIS (2013)

The weighted costs of delay may be used as criterion for the evaluation of traffic management
measures, see Martin (1994), Schlaich (2002), Hlawenka (2003), Cui (2005). Computer-based
traffic management requires powerful tools for assessing the present situation and expected
future situations due to the means used to dispatch trains. The consequences are evaluated
on the basis of certain objectives (e. g. minimum of unscheduled waiting time weighted by
the costs of delay) in order to recommend the most beneficial option to the dispatcher. A
complete optimisation procedure would be needed for automation of the traffic management
process. The complexity of the problem to solve increases exponentially with an extension of
the infrastructure network size, number of lines and trains scheduled, as well as the amount
of preview time. In addition, the impact of shunting movements within the terminals must
be taken into account. At present, the performance of modern computers only suffices for
semi-automatic dispatching in rather restricted infrastructure areas. For more information see
Chapter 13.
There is a considerable advantage of performance evaluations with the direct assessment
of punctuality as an essential criterion. Thus such methods can be used for network and
timetable improvement and analysis of weaknesses of the railway operation. A disadvantage is
that absolute monetary values are used. Also, the results of an investigation become obsolete
if a value adjustment for the influencing factors of the partial indicators becomes necessary.
Therefore, a periodic calibration of the models is inevitable.

14.2.4 Use of Capacity Research

Capacity research methods are technology-oriented types of performance evaluations and

used for the determination of adequate
– track infrastructure and signal systems design,
– operating performance,
– quality of operation.
The dimensions of infrastructure, operating performance and quality of operation are all inter-
dependent. If two components are known, the third one can be derived. Safety requirements
and general economic framework conditions as well as environmental constraints are provided
from outside sources (Fig. 14.3).

281
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
14 Performance Evaluation

Fig. 14.3: General Influence Factors of Capacity Research

In principle, there are two methods for capacity research, neglecting the seldom used
graphical methods:
– Analytical approach
– Simulation approach
When using analytical methods for capacity research, an estimation of the line exploitation rate
and the waiting time may be carried out based on minimum line headway times and certain
initial delays. The railway operation process is described mathematically in a deterministic
or stochastic form. The exploitation rate of a line can be easily ascertained deterministically.
For the determination and assessment of the exploitation rate, see Chapter 2.7 of this book.
Queuing theory may be used for the estimation of waiting times if only the frequency of trains
is known (see Chapter 6). Performance investigations based on analytical methods are well
suited for easy assessment of simple, homogeneous track arrangements and separate line
sections. Complex structures or those which are not homogeneous must be subdivided in
order to get meaningful results without exerting a non-justifiable amount of effort. For detail on
simulation methods see Chapter 10, Breuer et al. (2008) and Martin et al. (2010). Simulation
enables the evaluation of different variants of infrastructure and operating programs.
Fig. 14.4 shows methods, components and results of capacity research from different
perspectives described on macro-, meso- and microscopic scale. The quality of operation of
an operating program is expressed as the Delay-coefficient, which can be calculated according
to Table 14.1 or by observing the change of delays at certain points of the infrastructure within
a defined time period. The macro- or mesoscopic performance of an operating program is
described by the Recommended area of traffic flow (for detailed description see below). As a
global indicator, it also gives a sense of the capacity of complex infrastructure layouts. Both,
Delay-coefficient and Recommended area of traffic flow always should be considered together.
On microscopic scale, the usage and behaviour of infrastructure elements can be determined
absolutely or relatively by calculation of the level of hindrance depending on the occupancy
rate of that infrastructure element, see Martin et al. (2012). The result of the bottleneck-
analysis shows the infrastructure elements which cause the bottlenecks, the infrastructure
elements which are affected by bottlenecks as well as reserves in the form of infrastructure
elements without of hindrance and low occupancy rate. For more information see Martin et al.
(2012), Hantsch et al. (2013) and Martin et al. (2013).

282
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 14.2 Methods of Performance Evaluation

Fig. 14.4: Methods, Components and Results of Capacity Research, see Martin et al. (2012)

Capacity represents the average of the maximum number of trains that may be operated using
a given part of the infrastructure at the same time. It is a theoretical limiting value that is not
reached in practice. The measure of quality of operation is defined by unscheduled waiting
times that result from hindrances during the operational process (consecutive delays). The
unscheduled waiting times depend strongly on the number of trains in the network and their
headways. The more trains that occupy the infrastructure, the higher is the sum of waiting
times. The (maximum /throughput) capacity is the right boundary of the graph of the waiting
time function where the sum of waiting times converges to infinity (Fig. 14.5). If the structure
of the operating program is maintained strictly, the waiting time function is limited by the
Throughput capacity LD, see Chu (2014) and Martin et al. (2014). The ratio of the amount of
waiting time per train improves the comparability of different infrastructure or timetable options.
An important aim of capacity research is the determination of the Recommended area of traffic
flow (Fig. 14.6). The lower (left) limit of the Recommended area of traffic flow is represented by
the traffic flow which minimises the relative sensitivity function of the waiting times. The upper
(right) limit is the traffic flow corresponding to the maximum value of the so-called traffic energy
function, see Hertel (1992). Concrete estimation examples are given in Bosse et al. (1995) and
Schmidt et al. (2007).
The approach has been further developed by Schmidt (2009) and Chu (2014) to take into
account the requirement of keeping an unmodified structure of the operating program, even
when traffic density changes. Essential conditions for this include, among others, ensuring that
the same number of trains enter and leave the investigation area within the investigation period
and that the ratio of different types of trains is retained within appropriate time slices, see Chu
et al. (2012). Accordingly the Throughput capacity LD should be used for capacity research.

283
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
14 Performance Evaluation

Fig. 14.5: Waiting Time Diagram and Capacity, Maximum capacity LF: with impaired consideration
of the structure of the operating program, Throughput capacity LD: with retention of the structure
of the operating program

Fig. 14.6: Recommended Area of Traffic Flow based on Throughput capacity LD

284
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 14.2 Methods of Performance Evaluation

Several steps are necessary for determining the Recommended area of traffic flow by means
of synchronous simulation. First, the limits of the investigation area and the limitation of the
track arrangement to be scrutinised are to be defined such that delays may be compensated
within the investigation area. If necessary, the limits of the investigation area will have to be
widened iteratively until no more significant changes concerning the capacity of the simulation
results of the track arrangement in question occur. The operating program (number, type,
sequences, origin and destination of the trains during rush hours) must be generated or
adapted. The resulting formation of train groups covers similar types of trains (class, velocity,
weight etc.). The base schedule in which the time-dependent train paths are modelled must
not include any delays, nor conflicts between the train paths.
The waiting time function and the Throughput capacity LD are determined by means of
synchronous simulation (see Chapter 10). The base schedule is used for calibration of the
simulation model by overlaying known real or empirical delay distribution functions. The results
are checked using available reference values. The base schedule may also be scrutinised for
stability by varying initial delays and the prolongation of the dwell times (see also Chapter 7).
Single points of the waiting time function are generated by a stepwise increase of the train
density, while retaining the structure of the base schedule representing the operating program.
The unscheduled waiting times resulting from simulation are recorded. If the number of trains
that leave the investigation area during the specified time period cannot be increased further,
Maximum capacity LF is reached.
Based on the Maximum capacity LF and single points of the waiting time function recorded
from the simulation, the Throughput capacity LD and the waiting time function can be
estimated by using the optimisation method according to Chu (2014) based on the equalising
method of Bosse (1994). Chu (2014) also states that the required stationary phase often
cannot be reached during the simulation process and is described by:
a ⋅  ⋅ (1− c )
ETW = t W = (14.2)
(1− )b
ETW: (statistical) expected waiting time function
tW: average waiting time per train (because of standardisation)
a, b, c: parameters of the waiting time function
: utility factor of capacity

The utility factor of Throughput capacity LD results from:

Traffic flow nZug / T [number of trains /period of time]
= (14.3)
Throughput capacity LD [number of trains /period of time]

The parameters a, b and c of Formula 14.2 are determined by an optimisation method (e.g.
trust-region-algorithm), which minimises the sum of errors between the single points of the
waiting time function recorded from the simulation and the waiting time function itself.
The relative sensitivity EMPFrel of waiting time and the utility factor of capacity  corresponding
to the minimum value of this function can be calculated with the following formulae:
t W' 1 c ⋅ c−1 b
EMPFrel = = − + (14.4)
tW  1− c 1− 
tW: average waiting time per train (because of standardisation)
tW’: derivative of average waiting time per train (for explicit form see formula 14.9)

−1 (c2 − c) ⋅ c−2 c2 ⋅ 2c−2 b !

EMPFrel' = 2
− c
− c 2
+ 2
= 0 → Minimum (14.5)
 1−  (1−  ) (1− )
285
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
14 Performance Evaluation

The result is the lower limit of the Recommended area of traffic flow for the investigated track
arrangement, under given operational conditions. The traffic energy and the utility factor
of capacity corresponding to the maximum value of this function are determined with the
formulae:

nZug ⋅ vb
Eb = (14.6)
T
Eb: traffic energy
T: period of time
nZug: number of trains
vb: average transport velocity
 
Eb = = (14.7)
t ⎛⎜ a ⋅ ⋅(1– c ) ⎟⎞
1+ W ⎜⎜ ⎟⎟
tb ⎟⎟
⎜⎜⎝ (1− )
b
⎟⎠
1+
tb
tW: average waiting time per train
tb: average transport time per train
 utility factor of Throughput capacity LD
a, b, c: parameters of the waiting time function

t W t W'
1+ − ⋅ !
tb tb
Eb' = 2
= 0 → Maximum (14.8)
⎛⎜ t W ⎞⎟⎟
⎜⎜1+ ⎟
⎜⎝ b t ⎟⎠

With a ⋅ ⎡⎢1–(c +1)⋅ c + b ⋅ ⋅(1− c )⋅(1− ) ⎤⎥

−1

t W' = ⎣ ⎦ (14.9)
(1− )b

The solution of the Equation (14.8) is the upper limit of the Recommended area of traffic flow
for the investigated track arrangement under given operational conditions.
Experience shows that the traffic flow in reality should rather be orientated at the lower (left)
limit of the Recommended area of traffic flow if no further special arrangements are provided.
The greater the Recommended area of traffic flow, the higher the flexibility of the investigated
track arrangement with respect to the chosen operating program. Besides the Recommended
area of traffic flow, the shape of the traffic energy function (the maximum value itself as well
as the slope to the right of the maximum) and the slope of the relative sensitivity function to
both sides of the Recommended area of traffic flow, should be taken in consideration. In case
of congestion, the slope to the right of the maximum of the traffic energy can be used as
an indicator for the behaviour of the track arrangement. The lower the slope of the relative
sensitivity function to the right of the limit of the Recommended area of traffic flow, the more
robust are the results of the investigation. In addition, the ability to reduce overall delays within
the investigation area and at specific locations can be evaluated.

286
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 14.2 Methods of Performance Evaluation

14.2.5 Performance-dependent Funding of Infrastructure

Adequate infrastructure is necessary for providing the expected quality of railway services
related to all target groups. A special feature of the railway is the very strong technical and
operational interdependence of infrastructure, rolling stock, timetable and railway operation.
Therefore, an arbitrary separation of infrastructure management and the operating business
is not straightforward. Availability and reliability, as well as environmental impacts of the whole
railway system, are directly influenced by infrastructure investments. These must be evaluated
due to their overall economic and social significance with respect to business management as
well as considering of the benefits and costs for passengers and the general public.
In Germany, there are several legal regulations concerning the expenditure of public
funds for railway infrastructure investments (e. g. Bundesverkehrswegeplan – Federal
Infrastructure Development Plan, Gemeindeverkehrsfinanzierungsgesetz – Municipality
Transportation Financing Act). Based on these laws, positive overall economic assessment
by standardised evaluation methods is a precondition for public funding. Typical German
standardised evaluation procedures are the Overall Economic Evaluation Method of the
Federal Infrastructure (BVU et al., 2005) and the Standardised Evaluation of Infrastructure
Investments for Urban Public Transport (Intraplan et al., 2006) which is also occasionally
used in other countries (e. g. Austria, China, Italy, Norway, South Africa and Switzerland).
These evaluations are based on cost-benefit investigations (CBI), which include cost-
benefit analysis (CBA), value-benefit analysis (VBA) and cost-effectiveness analysis (CEA).
Table 14.4 gives an overview.

Cost Benefit
Cost Benefit Analysis (CBA) monetary monetary
Value Benefit Analysis (VBA) credit points credit points
Cost Effectiveness Analysis (CEA) monetary credit points

Table 14.4: Methods of Cost Benefit Investigations (CBI)

The German method “Standardised Evaluation of Infrastructure Investments for Urban

Public Transport” (SEIIUPT) is outlined in the following section. This method is stipulated for
investments of more than 25 million EUR, and it consists of two parts:
– Economic evaluation
– Subsequent cost analysis
While inter-regional comparability and compilation of the project impacts as a whole are
dominant in the economic evaluation, the financial impacts (subdivided into infrastructure
and operation, or if necessary into different public authorities or transport enterprises) are
determined in the subsequent cost analysis. The request for inter-regional comparability
requires an extensive standardisation of the rates to be used for the overall economic
evaluation by a uniform obligatory price level. Global standard price rates do not always
correspond to local rates of the area investigated. Therefore, the subsequent cost analysis
should be based on most current local rates.
The results of SEIIUPT determine the significance of the evaluated project and the ranking of
project alternatives or competitive projects. It is a multi-level, formal method with calculation
of the expected traffic flows, the determination of the evaluation indicators and calculations
of the impacts. The evaluation method is based on a modal-split model (Section 14.2.3) as
well as the so-called “with/without case principle”. For the assessment of the investment plan,

287
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
14 Performance Evaluation

those changes that are caused by the implementation of the planned project (“with case”
according to the investment plan) are to be determined compared to the conditions that will
occur without implementation of the project (“without case”). The results are always presented
in the form of balances, whose interpretation is supported by the absolute values of the
different cases (“with and without cases”). Fig. 14.7 shows the different steps of the SEIIUPT
procedure.

Fig. 14.7: Standardised Evaluation of Infrastructure Investments for Urban Public Transport
(SEIIUPT) Procedure in Germany

288
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 14.2 Methods of Performance Evaluation

The main goal of the evaluation method is to optimise the benefits of public transportation
infrastructure investments. A catalogue of detailed aims captures the essential effects due to
public transport investment measures. The economic assessment of a project consists of
– overall economic benefit that results from the project during regular service,
– overall economic benefit that results from service for special large-scale events (optional),
– improvement of quality of service based on punctuality of trains (optional),
– debt service for the complete infrastructure investment.
The general evaluation indicators, based on numerical measurable partial indicators, support
the technical/economic appraisal. Based on CBI for quantifiable partial indicators such as
balance of maintenance costs of the infrastructure, or noise pollution,
– a CBA-indicator (economic evaluation indicator E1),
– a VBA-indicator (economic evaluation indicator E2),
which depend partly on each other, are derived from quantifiable partial indicators (Table 14.5).

Economic Additional
Influences evaluation verbal
indicator explanations
(partial indicators) E1 E2 V
Monetary form or transferable into
monetary values [EUR/year] X X
Cardinal measurable components
Not tranferable into monetary values X
[credit point scale]
Not quantifiable components X

Table 14.5: General Economic Evaluation Indicators of SEIIUPT

For the final decision, further non-monetary and non-quantifiable criteria are recorded verbally.
A scale transformation may be needed for those partial indicators that cannot be determined
from numerical measurands. The problem of their transformation into a cardinal credit
point scale for the VBA and CEA, or into amounts of money for the CBA is still not solved
satisfactorily. The consideration of non-quantifiable components as a supplementary decision
support is therefore carried out in form of an additional verbal explanation (evaluation indicator
V, cf. Table 14.5).
The sum of the benefit components is compared to the debt service when estimating the
general economic evaluation indicator E1, for the public transport infrastructure in the “with
case”. Because all of the components are available in monetary form or transferable into
monetary values, it is possible to determine
– the cost benefit difference,
– the cost benefit ratio as a decision support.
calculable benefits
E1 = (14.10)
costs of infrastructure

An investment project is accepted only if the general economic indicator E1 is higher than one.
The results of the subsequent cost analysis should make clear
– the public share of the infrastructure investments,
– the expected changes in the operational expenses that the transport companies will have
to pay, due to the investment project,
– the expected income changes due to the investment project.
289
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
14 Performance Evaluation

The subsequent cost analysis provides three kinds of identification values for every involved
party:
– The cash-flow balances for every year of the review period represent a clue about the
financial requirement in each respective year. This financial requirement is calculated from
the balance of the receipts and expenditure, and should not be confused with the annual
profit or deficit that arises from the balance of the yield and costs.
– The net present value of the cash-flow balances is calculated from the sum of the
discounted cash-flow balances. If the net present value is positive, the investment project
is worthwhile for the involved parties. If it is less or equal to zero, there is no reason for the
involved parties to invest in the project based on their own economic motives.
– The changes of the annual figures for select years of the review period give a clue about
the project impacts on the success of an involved enterprise according to a profit and loss
account.
The impacts of CBA, CEA and VBA are first recorded generally in the original measurands
(measure, number and dimension, quantities or qualities). The difficulties associated with
the necessary scale transformations for the transformation and the adaptation of the original
measures into a uniform assessment scheme (EUR/year, credit points) are present for all of
the procedures. Therefore, sensitivity analyses are used to check the consequences of the
different assessment and weighting approaches.
Availability and reliability of the transport services are indirectly considered in the SEIIUPT in
two ways:
- by the partial indicators of the modal-split changes, while assessing the overall economic
benefit, and
- by the indicator punctuality in the optional part for improving the quality of service. Because
the general values for assessing the input factors such as wage costs and price of fuel
change with time, the base survey data of the model has to be updated about every five
years.

14.3 Conclusion and Further Development

Performance evaluations are becoming more important. Increasingly, political bodies also
demand a continuous analysis of the railway performance, see EU (2012). The greater the
number of involved parties for developing, financing and running the railway system, the more
difficult become the target-group-oriented performance evaluation. Therefore, deregulation
and privatisation of the railway system require a further development of the existing
approaches of the performance evaluation.
Ever since the invention of the railway, its organisation has alternated between being privately-
run and publicly-run. Depending on the socio-political priorities (e. g. reduction of national
debt, increase of general public welfare, rise of the modal-split in favour of the railway),
public-private partnerships may also be applicable. The political decision making body
should assess objectively the different possible organisational models of the railway system
and their impact on performance based on standard criteria. Reasonable simplifications and
feasible generalisations, including parameters for railway performance, are needed in complex
evaluation models.
Punctuality, whether assessed directly or indirectly, is one of the most important reliability
indicators of the performance evaluation. Therefore, a general travel-time-dependent threshold
value of punctuality has to be determined on a scientific base. The method for customer-

290
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

oriented recording and assessing of delays should be unified, as a precondition for a monetary
based reward-penalty system that reflects the service quality and provides an incentive for the
train operators as well as for the infrastructure manager.
Innovative funding models for railway lines and regional or urban public transport are gaining
more interest. At the same time, the goal is to reduce the amount of public financing and to
develop private financing possibilities. Banks are interested in performance evaluations to
secure their loans, see European Investment Bank (2005). Individual economic evaluation
models for railway systems are being developed. Publicly funded projects will pay more
attention to performance criteria. When a company is chosen directly to provide regional
or urban public railway transportation services without an invitation to bid, the company
is bound to European Union placement guidelines, which means that the company in
question must be able to prove its performance objectively compared with other public
transportation undertakings, see European Court (2003). A further development and European
standardisation of the combined performance evaluation procedures based on effectiveness
and efficiency would create a sound objective basis.
If several public financial backers are involved in an investment project, or if ordered railway
transportation services are financed by several backers, then a corresponding assignment of
expenses and benefits would be requested. To reach this, the existing methods of evaluation
must be further developed, so that a breakdown of the benefits and an allocation of the costs
is possible, see Umbach (2006).
The railway companies strive to improve their networks, timetables and quality of operation
in order to secure a competitive advantage. Capacity research methods can be used very
effectively for this purpose. A better correlation between models and reality, as well as an
improvement of real-time processing of large amounts of data in connection with standardised
interfaces, will create a broader range of applications for performance evaluation methods.

References
Ackermann T. (1998): Die Bewertung der Pünktlichkeit als Qualitätsparameter im
Schienenpersonenverkehr auf Basis der direkten Nutzensmessung. PhD thesis
Forschungsarbeiten des Verkehrswissenschaftlichen Instituts an der Universität Stuttgart,
Bericht 21.
Bosse G. (1994): Leistungsfähigkeit und Leistungsverhalten in Abhängigkeit von der
Anlagengestaltung und dem Betriebsprogramm. MSc thesis, Institut für Verkehr,
Eisenbahnwesen und Verkehrssicherung der TU Braunschweig.
Bosse G., Martin U., Pachl J. (1995): Anwendung des Simulationsprogramms UX-SIMU zur
Leistungsuntersuchung von Strecken. Schriftenreihe des Instituts für Verkehrssystemtheorie
und Bahnverkehr der TU Dresden, Band 1.
Breuer P., Li X., Martin U., Schmidt C. (2008): An Innovative Approach of Capacity Research of
Railway Infrastructure.
Networks for Mobility (2008): FOVUS 4th International Symposium, Universität Stuttgart,
Stuttgart.
BVU, IVV, Planco (2005): Die gesamtwirtschaftliche Bewertungsmethodik. Bundesverkehrs-
wegeplan 2003. Anwenderhandbuch. Bundesministerium für Verkehr, Bau und Wohnungs-
wesen (Hrsg.), Berlin Bonn Freiburg Aachen Essen.
European Investment Bank (2005): Evaluation of EIB Financing of Railway Projects in the
European Union. Evaluation Report. Operations Evaluation (EV), Luxemburg.

291
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
14 Performance Evaluation

Chu Z., Martin U. (2012): Dynamisierung von Zeitscheiben in Betriebsprogrammen bei

Leistungsuntersuchungen. Eisenbahntechnische Rundschau ETR, 61 (2012) 5, pp. 40 – 45.
Chu Z. (2014): Modellierung der Wartezeitfunktion bei Leistungsuntersuchungen im
Schienenverkehr unter Berücksichtigung der transienten Phase. PhD thesis, Institut für
Eisenbahn- und Verkehrswesen der Universität Stuttgart, Stuttgart.
Cui Y. (2005): Implementation of the Optimisation Theory for User Oriented Automatic
Dispatching Systems in Railway Transport. Master Thesis am Institut für Eisenbahn- und
Verkehrswesen der Universität Stuttgart, Stuttgart.
DB AG (2008): KoRil 405 Fahrwegkapazität – unpublished – 2008.
DESTATIS Statistisches Bundesamt (2013): Verbraucherpreisindizes für Deutschland.
Jahresbericht 2012, Wiesbaden.
EU (1991): COUNCIL DIRECTIVE of 29 July 1991 on the development of the Community’s
railways (91/440/EEC), Brussels.
EU (2006): COMMISSION DECISION of 11 August 2006 concerning the technical specification
of interoperability relating to the subsystem “Traffic Operation and Management” of the trans-
European conventional rail system (TSI) (2006/920/EC), Brussels.
EU (2012): EUROPEAN COMMISSION. REPORT FROM THE COMMISSION TO THE
COUNCIL AND THE EUROPEAN PARLIAMENT. Third report on monitoring development of
the rail market. COM(2012) 459 final, Brussels.
European Court (2003): Judgment of the Court of 24 July 2003 in Case C-280/00
(Reference for a preliminary ruling from the Bundesverwaltungsgericht): Altmark Trans GmbH,
Regierungspräsidium Magdeburg v Nahverkehrsgesellschaft Altmark GmbH (Regulation
(EEC) No 1191/69 – Operation of urban, suburban and regional scheduled transport services
– Public subsidies – Concept of State aid – Compensation for discharging public service
obligations).
Hansen I., Wiggenraad P., Wolff J. (2013): Performance Analysis of Railway Infrastructure and
Operations. 12th WCTR, July 15–18, 2013, Rio de Janeiro, Brazil.
Hantsch F., Li X., Martin U. (2013): Methoden zur Engpassanalyse bei der
Infrastrukturbemessung im Schienenverkehr. Eisenbahntechnische Rundschau ETR, 62 (2013)
3, pp. 30 – 33.
Hertel G. (1992): Die maximale Verkehrsleistung und die minimale Fahrplanempfindlichkeit auf
Eisenbahnstrecken, Eisenbahntechnische Rundschau ETR, Vol. 41, No. 10, pp. 665 – 671.
Hlawenka A. (2003): Entwicklung und Implementierung eines Dispositionswerkzeuges zur
kybernetischen Prozessoptimierung betriebsbedingter Wartezeiten im schienengebundenen
Verkehr. MSc project, Institut für Eisenbahn- und Verkehrswesen der Universität Stuttgart.
Intraplan Consult GmbH, VWI Verkehrswissenschaftliches Institut Stuttgart GmbH
(2006): Standardisierte Bewertung von Verkehrswegeinvestitionen des ÖPNV und
Folgekostenrechnung. Version 2006, erstellt im Auftrag des Bundesministers für Verkehr, Bau
und Stadtentwicklung (Hrsg.), Berlin München Stuttgart.
Jacobs J. (2003): Rechnerunterstützte Konfliktermittlung und Entscheidungsunterstützung
bei der Disposition des Zuglaufs. PhD thesis,Verkehrswissenschaftliches Institut der RWTH
Aachen.
Hamman P., Erichson B. (1994): Marktforschung, 3. Auflage, Gustav Fischer Verlag, Stuttgart
Jena.

292
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

Lan L. W., Lin E. T. J. (2006): Performance Measurement for Railway Transport: Stochastic
Distance Functions with Inefficiency and Ineffectiveness Effects. Journal of Transport
Economics and Policy, Vol. 40, Part 3, pp. 383 – 408.
Martin U. (1994): Verfahren zur Bewertung von Zug- und Rangierfahrten bei der Disposition.
PhD thesis, Schriftenreihe des Instituts für Verkehr, Eisenbahnwesen und Verkehrssicherung
der TU Braunschweig, Heft 52.
Martin U., Schmidt C. (2010): Erhöhung der Effektivität und Transparenz bei
Leistungsuntersuchungen mit Simulationsverfahren. Eisenbahntechnische Rundschau ETR,
59 (2010) 7+8, pp. 463 – 468.
Martin U., Li X., Warninghoff C.: Bewertungsverfahren für Knotenelemente bei der
Infrastrukturbemessung – RePlan. In: ETR, 61 (2012) 11, pp. 38 – 43.
Martin U., Chu Z., Hantsch F., Li X., Cui Y. (2013): Bahnhofskapazität, RePlan AP4.
DB Netz AG Programmschwerpunkt „System Schiene 2010“ des Bundesministeriums für
Wirtschaft und Technologie. Abschlussbericht, Stuttgart.
Martin U., Chu Z. (2014): Direkte experimentelle Bestimmung der maximalen
Leistungsfähigkeit bei Leistungsuntersuchungen im spurgeführten Verkehr. Deutsche
Forschungsgemeinschaft DFG, MA 2326/6-1. Neues Verkehrswissenschaftliches Journal des
Verkehrswissenschaftlichen Instituts an der Universität Stuttgart e. V., Ausgabe 7.
Schlaich J. (2002): Beitrag zur Entwicklung eines Dispositionswerkzeuges zur Optimierung
betriebsbedingter Wartezeiten im schienengebundenen Verkehr. Draft, Institut für Eisenbahn-
und Verkehrswesen der Universität Stuttgart.
Schmidt C., Martin U., Christiani V. (2007): Leistungsuntersuchungen für den
schienengebundenen ÖPNV. Der Nahverkehr, 25 (2007) 7– 8, pp. 32 – 38.
Schmidt C. (2009): Beitrag zur experimentellen Bestimmung der Wartezeitfunktion
bei Leistungsuntersuchungen im spurgeführten Verkehr. PhD thesis. Neues
Verkehrswissenschaftliches Journal des Verkehrswissenschaftlichen Instituts an der Universität
Stuttgart e. V., Ausgabe 3.
Umbach U. (2006): Die Finanzierung von Neu- und Ausbaumaßnahmen der
Eisenbahninfrastruktur – Schlüsselzahlen für die Ableitung von Beiträgen der
Finanzierungsbeteiligten. PhD thesis. Neues Verkehrswissenschaftliches Journal des
Verkehrswissenschaftlichen Instituts an der Universität Stuttgart e. V., Ausgabe 2.
WABB Maidment, D. (1988): Zuverlässigkeitsmanagement bei den Britischen Eisenbahnen,
Vorabdruck des Beitrages, Preprint 27. Sitzung des WABB and Passenger Demand
Forecasting Handbook. BR (1986) – unpublished.
Walther K. (1991): Maßnahmereagibler Modal-Split für den städtischen Personenverkehr.
Veröffentlichung des Verkehrswissenschaftlichen Instituts der RWTH Aachen, Band 45.

293
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 15.2 Background

15 Economic performance assessment in railways

Phill Wheat, Chris Nash, Andrew Smith

15.1 Introduction

The assessment of the economic performance of railways is of growing interest for two main
reasons. Firstly, the very different reforms of the railway sector in different countries, including
privatisation and deregulation in the US and vertical separation and the introduction of on track
competition in the European Union, lead to a wish to compare performance to try to determine
which is the most suitable path to take, see Nash (2006). Secondly, regulation of the natural
monopoly element of the rail system, the infrastructure, is leading to a need to benchmark its
costs, whilst continued concern about the costs of dominant incumbent train operators leads
to a similar need.
The aim of this Chapter is to explain the various approaches that can be taken to measuring
the economic performance of railways and the issues that arise in applying them. The Authors
first discuss the background and some of the characteristics that complicate performance
analysis. They then consider how to measure the outputs that railways produce and the
inputs used in producing them. Following this, they outline the general issues with measuring
productivity (the ratio of outputs to inputs) and the limitations thereof. Then the Authors
proceed to describe methods to measure the often more useful concept of efficiency, dealing
with technical and cost efficiency in turn. Finally, they reach their conclusions.

15.2 Background

For many industries, economic performance may be equated with profitability. Measures of
profitability compare the value of outputs produced with that of inputs used to produce them.
In a competitive market, the value of the outputs will represent what consumers are willing
to pay for them, and the value of inputs their opportunity cost (what they would be worth in
alternative uses).
Of course, the presence of monopoly power and external effects (impacts such as pollution
or congestion that effect people other than producers and consumers) often prevent such
a simplistic approach. In the case of railways the problems are particularly acute. Prices,
particularly for passenger services, are often regulated by government. Costs may be affected
by how capital investment is financed (grants or subsidised loans and by government write
offs). Moreover, there may be social benefits to users or reduced externalities due to diversion
of traffic from road or air, whilst railways themselves impose externalities. The technique of
cost benefit analysis enables the above mentioned user benefits and externalities to be
included to get a measure of social profitability of the railway, or of certain services. Cost
benefit analysis includes these non-market factors valued at what the recipients would be
willing to pay for them in the case of benefits, or what they would accept as compensation to
put up with them in the case of costs. There are numerous studies of what people are willing
to pay for time savings and reduced accident risk, but rather less in the case of environmental
externalities. Nevertheless the EU has published a handbook valuing these for all modes of
transport, see Maibach et al. (2008), whilst UIC has commissioned specific estimates by
country, see CE Delft et al. (2008).
Profitability is also heavily influenced by the environment in which the railway is operating.
Provision of railway infrastructure on a particular route requires a high fixed cost (at least a

295
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
15 Economic performance assessment in railways

single track railway must be provided before any output at all can be produced) and the more
traffic that is available in general, the lower the unit cost. The traffic available will be influenced
by factors such as the volume and location of the population, their incomes, the structure of
production and the price and quality of competing modes of transport.
Railways also produce multiple outputs – e. g. passenger and freight traffic, long distance
and commuter transport – using much the same assets. It is often desired to examine the
performance of the railway in these various markets. Examining the costs of any particular
type of service brings in the problem of joint costs. Different services may share locomotives
and staff, for instance working passenger trains by day and freight by night. But the most
acute joint cost problems concern the track. Wear and tear costs may be directly estimated
for different services using econometric or engineering formulae, see Wheat & Smith (2008).
Certain elements of the costs of providing the track may also be directly accountable to certain
types of service. For instance, passing loops may be avoided if there were no freight on the
line; additional slow lines if there were no suburban passenger (but both may also be avoided
if there were no fast passenger services; the same cost may be avoidable to more than one
type of service). Some routes may be only used by one type of service (e. g. freight or regional
passenger). But what is left is a genuine joint cost, necessarily incurred if any type of service
is provided. Thus whilst it may be possible to estimate the marginal cost of different types of
outputs, performance analysis really needs to take place at the level of the system as a whole
rather than for individual service types.
Of course, different outputs may require different amounts of inputs. Trains carrying bulk
commodities such as coal typically carry more tonnes of cargo per wagon than trains carrying
containers. Long distance passenger trains typically carry more passengers per train, but
fewer per vehicle, than regional trains (long distance trains typically offer more space per seat
and more facilities such as luggage space and refreshment facilities than regional trains).
Commuter services will be heavily loaded in the peak but very lightly loaded off peak.
Another very important characteristic of trains from the point of view of costs is the speed at
which they operate. Indeed, since labour is paid for by the hour and the cost of ownership
of assets is also time related, vehicle or train hours may be a better measure of output than
vehicle or train kilometres. Few studies have such data available, though Wheat & Smith (2014)
is an example. In its absence, introducing an estimate of average speeds as a characteristic of
outputs is a way of taking it into account.
All these characteristics of railways point to a need for a method of performance assessment
that can take account of the heterogeneity of outputs and of differences in the environment in
which the railway is operating. Failing that, care must be taken only to compare railways that
produce similar outputs in similar circumstances.

15.3 Outputs and inputs of rail systems

Rail systems essentially use inputs of labour, capital and fuel to carry passengers and freight.
Outputs may thus be considered to be passenger kilometres and freight tonne kilometres.
However, as noted above, this is clearly a gross oversimplification.
Network industries can be viewed as producing many different heterogeneous outputs. In
particular transport networks, given the non-storability of the product, the large number of
origin and destination combinations and the many different trip purposes, produce a very
large number of outputs. At the limit railways could be thought as producing individual travel
opportunities, by time, space and purpose. Clearly, such a disaggregation of outputs is

296
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 15.4 Productivity analysis

likely to be too extreme to undertake meaningful analysis of performance. As such, a more

pragmatic approach has to be taken in specifying outputs.
Several common features of the output specification can be considered. First, two general
classifications of outputs are common Oum and Yu (1994). One set is termed “available
outputs” which are measures of the service that the railway produces (capacity supplied)
which are available to customers to consume. Examples include train-km, vehicle-km and
seat-km. The second set is termed “revenue outputs” which are measures of consumed
outputs. Examples include passenger-km and tonne-km of freight hauled. These two sets
could also be thought of as intermediate versus final outputs of the railway system, although it
must also be borne in mind that the demand for rail services is often a derived demand.
When choosing whether to use available or revenue outputs, it is important to consider what
is required to be measured in the analysis and whether the implicit assumptions on what is
under the firm’s control versus what is exogenous is reasonable. For example, using available
outputs can be justified when considering the performance of a railway manager where the
required outputs from the railway are heavily prescribed by a regulator or government. As such
the railway manager does not have much discretion as to how many train-km, vehicle-km etc.
can be run. This is instead set by the regulator. However if analysis of the effect of government
policy is the aim of a study, then it is more appropriate to adopt revenue output measures, as
policy makers have discretion in the specification of railway services to best meet demand.
Of course, available outputs might be used in this context, alongside revenue outputs, as a
measure of characteristics (in particular quality) of the revenue output. A higher level of train
and vehicle kilometres per passenger kilometre generally equates to a more comfortable and
convenient service.
Measured inefficiency from models reflects both inefficiency of the railway manager and of policy
makers or regulators, see Oum & Yu (1994). Lan & Lin (2006) cite Fielding et al. (1985) who
define specific terms for these concepts. They define the degree of sub-optimal transformation
of inputs into intermediate outputs as “technical inefficiency”, while they define degree of sub-
optimal transformation of inputs into final outputs as “technical ineffectiveness”. They define
a further concept, “service ineffectiveness”, as the degree of sub-optimal transformation of
intermediate outputs into final outputs. They point out that it is the non-storability property of
railway outputs which requires such distinctions. Regarding inputs, there is a similar distinction
between intermediate and final inputs. Ultimately, all inputs comprise labour, capital and natural
resources. However, when examining railways, it is often useful to consider intermediate
inputs such as kilometres of track and numbers of locomotives, coaches and wagons. This is
necessary because of the very long life of railway assets; the volume and nature of the capital
employed in a railway may be very slow to adjust to changing circumstances.

15.4 Productivity analysis

One obvious and useful approach to measuring efficiency is to look at the output achieved
per unit of input – measures such as train kilometres per member of staff or vehicle kilometres
per vehicle. A whole range of interlocking partial productivity measures are in common use
in comparing railway efficiency. Since labour, infrastructure and rolling stock are the major
cost categories railways face, it makes sense to look at the efficiency with which these are
used to produce output. But the dependence of the results on the nature of the output and
the operating environment must always be borne in mind. A recent example of a study using
a whole range of such measures whilst trying to ensure that the railways concerned are
comparable is Hansen, Wiggenraad & Wolff (2013).

297
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
15 Economic performance assessment in railways

Economists (e. g. Oum et al. (2009)) worry that partial productivity measures considering the
efficiency of use of one input fail to allow for the levels of other inputs which may be used
to substitute for it (e. g. high investment in infrastructure and rolling stock may reduce labour
requirements). So it is argued that it is necessary to consider simultaneously all inputs.
Similarly, railways produce different types of output, so again it is necessary to aggregate
these. The result is measures of total factor productivity (TFP). But how are the different
inputs and outputs weighted? Various possibilities exist. Again, it might be that market prices
are used, in which case total factor productivity becomes the same thing as profitability.
Alternatively, the parameters of cost or production functions may be used to weight different
outputs by their marginal costs and different inputs by the value of their marginal product. For
a survey of such index number approaches see Coelli et al. (2005).
Ultimately, however, productivity is often not of central interest to economists. Instead the
determinants of it are of primary interest. TFP is determined by three factors. Firstly, there are
the scale (and density) properties of the technology. Thus depending on the scale of operation
the ratio of output to input can differ due to economies or diseconomies associated with the
technology. Secondly, technology can change over time. Thus TPF is likely to be different
(hopefully higher, all other things equal) as time progresses. Finally TFP is determined by the
technical efficiency of the firm. Technical efficiency is defined as the proportion of realised
costs that are required if the firm adopted (technical) best practice, all other things equal.
Thus technical efficiency is a measure of the extent of ‘waste’ by inefficient coordination of the
production process. Such concepts are illustrated graphically when cost analysis is considered
in Section 15.6.
Depending on the regulatory environment and the policy question under consideration,
different characteristics of TFP may be of interest. For example, in recent years competitive
tendering in railway operations has begun to be implemented. Thus the optimal size of
franchises (scale characteristics) is of interest. Other questions relate to the efficiency of
firms in the industry; clearly technical efficiency is of interest. In addition, as will be seen in
Section 15.6, analysis of cost data is often more appropriate as this allows for outputs to be
exogenous and cost efficiency (the product of both technical and allocative efficiency) to be
measured. This is often of central interest to economic regulators.

15.5 Measuring technical efficiency

As discussed above, measuring technical efficiency is often of interest rather than TFP.
There are two broad sets of methods to do this: statistical (or parametric) and mathematical
programming (or non-parametric). Statistical approaches estimate production or distance
frontiers (distance functions are useful for incorporating multiple outputs into a production
process). Statistical frontiers are returned to in Section 15.6 in the context of cost frontiers.
It suffices to note that statistical frontier techniques have been applied to measure technical
efficiency in many railway applications e. g. Cowie & Riddington (1996), Deprins & Simar
(1989), Lan & Lin (2006) and Tsionas & Christopolous (1999).
However, the most common approach in the current railway literature is to use the technique
known as data envelopment analysis (DEA), which is a mathematical programming technique.
DEA is a non-parametric technique in the sense that a function is not estimated in the statistical
sense. Instead, mathematical programming is used to envelop the data and in particular to
determine which observations are on the efficiency frontier and which are not. Farrell (1957)
was the first to propose the idea of enveloping data using a piecewise linear-hull i. e. a multi-

298
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 15.6 Measuring Cost Efficiency

dimension frontier, however it was not until 1978 that Charnes, Cooper & Rhodes (1978)
proposed the first methodology termed DEA for the constant returns to scale case.
Essentially, the computation of a DEA model is by means of a linear programming model in
which the objective is to maximise the efficiency measure of firm i by choosing output and
input weights. This is subject to all efficiency measures for all firms calculated using the
same output and input weights being less than 1 and all weights should be greater than or
equal to 1. That is all firms are constrained to have efficiency measures between 0 and 1,
with the same set of weights. The assumption of constant returns to scale means that firm
i can be compared with all other firms, even if the scale of production of some other firms
is substantially different. This is because the ratio of efficient input to output is invariant with
scale. However the technique can be extended to deal with non-constant returns to scale.
This problem has to be solved for each of the N firms (each will potentially have different
weights). DEA has a limitation since it does not allow explicitly for statistical noise in the data.
Instead, a set of linear programming problems are solved for each firm and for each time
period. This determines whether, in relation to other firms, the firm is on the frontier. This forms
a frontier which is piecewise linear.
A big advantage of DEA over statistical techniques is that there is no need to specify a
functional form for the cost and production relationships under consideration. Instead only a
relatively weak assumption that the frontier can be represented by a piece-wise-linear convex
hull is required. But note that efficiency measures are computed rather than estimated in DEA,
so that no standard errors and thus no confidence intervals can be constructed for them.
Because of this, while measures of efficiency can be constructed, it cannot be determined the
likely reliability of such measures.
DEA studies of railways are common – examples are Cantos et al. (2010), Cowie (2009) and
Merkert et al. (2009). They are attractive because they only require physical data, which is
generally easier to obtain and more directly comparable than financial data. However, in the
rail industry, there is a serious problem that different degrees of contracting out (for instance
of track or rolling stock maintenance or cleaning) exist for different railways, and this may
seriously affect comparisons of different railway companies in that some part of the labour
input does not appear in the data.
This problem may be overcome by switching from considering physical inputs to considering
costs. Cost data should include all costs whether functions are undertaken by the organisation
itself or bought in. Moreover techniques exist for estimating cost functions that permit
significance tests and controlling for the impact of other factors such as geography. Thus
there are strong arguments for favouring the use of cost functions over DEA for measuring
railway efficiency. However, as against that, it should be remembered that there may be
serious difficulties in getting comparable cost data particularly for capital costs, where different
methods of financing (loans versus grants), different conventions regarding depreciation
and the impact of past debt write-offs may all influence what appears in the accounts. Also,
international comparisons will be influenced by differences in purchasing power; it is necessary
to allow for this by the use of purchasing power parity exchange rates.

15.6 Measuring Cost Efficiency

Cost functions represent the relationship between costs, outputs and input prices. The first
empirical application of the formal cost function as it would be recognised today can be traced
back to Nerlove (1963). His seminal study on the costs of generating electric power was

299
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
15 Economic performance assessment in railways

the first to utilise an analytic relation (developed by Shephard (1953)) between the structure
of a firm’s costs and the structure of the firm’s production function. In particular, subject to
the behavioural assumption that the firm was cost minimising and the firm faced exogenous
input prices and output levels, Nerlove showed that a cost function could be derived from
a production function and input price information. Importantly, the resulting (total) cost
function was a function of both output levels and input prices. This work established the cost
function as having meaningful economic interpretation, which provided a strong motivation for
estimating such functions.
More generally, the economic model for firm production is the transformation function, which is
in itself a multi-output generalisation of the production function. This relates inputs to outputs,
i. e. it shows the input required to produce an amount of outputs, see Chambers (1988 p. 260).
Thus the cost function shows the minimum cost for achieving a certain amount of outputs,
given the prices of inputs. The question then arises as to whether the cost function can be
used to learn about the underlying production technology of the firm, such as returns to scale
and density and the effects of technological change. If so, the cost function is said to be ‘dual’
to the transformation function. A set of ‘regularity conditions’ needs to hold for this to be the
case (see Fuss and McFadden (1979) for full details of these conditions). It is important to test
or at least check that these conditions hold for an estimated cost function.
Some basic characteristics of the function are illustrated in Fig. 15.1.
Box A illustrates the fact that if we have estimated the minimum cost for producing a given
level of output, then any particular observation (such as point A or B) that lies above the
frontier must correspond to inefficient production.
Box B illustrates the phenomenon of economies of scale; as output increases so average cost
decreases
Box C illustrates how technological change may shift the cost frontier downwards over time.
Box D shows the measure of the efficiency of the firm: the ratio of the firm’s actual cost to the
minimum cost possible if the firm adopted best managerial practice – i. e. the gap between
the firm and cost frontier. Thus some firms do not produce at minimum possible cost and
are thus above the cost frontier. Cost efficiency is related to technical efficiency introduced
in the previous Section. In particular cost efficiency is the product of technical efficiency and
allocative efficiency. Allocative inefficiency is the sub-optimal combination of inputs, which
results in higher cost relative to the cost minimising combination of inputs. Thus cost efficiency
will be lower than (or equal to) technical efficiency. Finally, it should be noted that measuring
cost efficiency is often more appropriate than technical efficiency for economic regulators.
They are concerned with how well the firm is delivering its outputs, irrespective of whether it
is sub optimising with respect to how it transforms input into output (technical efficiency), or
choosing which inputs to combine (allocative efficiency).
In analysing transport cost functions, it is common to distinguish between scale and density
effects. Density effects comprise the effect on costs of increasing all outputs (in equal
proportion) while holding network size constant. Scale effects comprise the effect on costs
of increasing all outputs and network size in equal proportion. Railway cost functions typically
display increasing returns to density but constant returns to scale. That is to say, adding more
traffic to an existing network reduces unit cost whilst equal proportionate increases in traffic
and network size (holding traffic density constant) leaves unit cost unchanged.
It is important to emphasise that network size is viewed in railways as a characteristic of
railway outputs, since the size of the network affects the scope of travel opportunities available
to users.

300
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 15.6 Measuring Cost Efficiency

Fig. 15.1: Economic concepts with respect to a cost frontier

The need to distinguish between scale and density effects or the choice between revenue
versus available outputs is only part of the wider issue of how to account for the heterogeneity
of railway outputs, as introduced in Section 15.3. One way to deal with the heterogeneity of
outputs is to group outputs into a small number of groups (e. g. long distance and regional
passenger) and then to include additional variables to characterise the outputs within those
groups. Relevant variables would be mean train loads (in terms of numbers of passengers
or tonnes of freight) and mean train speeds and length of trains. Clearly this could lead to a
model with a large number of parameters to be estimated, however hedonic cost functions
allow for incorporation of quality characteristic variables in a parsimonious way (Spady &
Friedlaender (1978) and see Wheat & Smith (2014) for an application to train operations).
Understanding scale and density properties of railways is important, particularly given the trend
towards competitive tendering which requires the design of contracts. In such circumstances,
one consideration will be choosing services to bundle together into contracts that minimise the
cost of operation and/or state subsidy required. Wheat & Smith (2014) provide some evidence
on the scale and density properties of train operations in Britain and, crucially, how these are
affected by the heterogeneity (output characteristics) within a contract.
In terms of formally stating the economic model to account for inefficiency of firms, it should
be noted that the optimisation of the constrained optimisation discussed above yields the
minimum cost frontier, while in practice firms may fail to optimise. Thus ideally an estimation
technique is needed which can allow for an additional component representing the difference
between actual cost and minimum cost. A number of methods exist for estimating such
models. One method to correct for this is to re-interpret the residual in the regression as
firm inefficiency rather than statistical noise. Estimation can be undertaken by “shifting” the
estimated OLS regression line down. There are several techniques to do this. One such is

301
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
15 Economic performance assessment in railways

Corrected Ordinary Least Squares (COLS). Greene (1980) discusses the statistical properties
of the estimators resulting from this shift. While this and other deterministic frontier techniques
do improve on standard least squares methods by explicitly recognising firm inefficiency,
they do so at the cost of assuming away any legitimate statistical noise. On the other hand
the ‘stochastic frontier’ formulation recognises the presence of both unobserved noise and
inefficiency. This does come at a cost, namely the use of strong distributional assumptions
when cross sectional data is used, resulting in results being potentially sensitive to such
assumptions. However such limitations can be overcome when panel data is available
since persistency in inefficiency over time can be used as a means of identifying noise from
inefficiency (see Coelli et al. (2005) and Kumbhakar & Lovell (2000) for a survey).

15.7 An example

As an example of the use of cost frontier methods, the econometric efficiency analysis of
Network Rail formed a very important part of the 2008 Periodic Review determinations by the
British rail regulator (Office of Rail Regulation – ORR). The primary piece of analysis utilised
a data set collected by the UIC and previously analysed for the Lasting Infrastructure Cost
Benchmarking (LICB) project. This was data for a selection of railway infrastructure managers
who were members of the UIC over a number of years.
Work sponsored by ORR undertook econometric efficiency analysis of the dataset Smith
(2012). Utilised explanatory variables were train-km (passenger and freight), route-km,
proportion single track-km and proportion of track electrified. While this model is clearly an
abstraction, it does contain variables which allow separation of returns to scale and density
(route-km and train-km respectively) and includes some key characteristics of output
(namely passenger and freight for the usage output and electrification for the quality of the
infrastructure). Further PPP exchange rates were used as a means of controlling for input price
variation.
The measurement of efficiency in this model exploits the panel data structure of the data. In
particular the data set was a (relatively) long (11 year) panel, which permitted the modelling of
the evolution of inefficiency over time (in a parametric manner). This was a desirable feature
of the model given that it was widely acknowledged that Network Rail (or its predecessor,
Railtrack) had in the past suffered a cost shock and was only over time bringing costs down to
more efficient levels. The efficiency variation for Network Rail is reproduced in Fig. 15.2 below.
Three models are presented; a model for maintenance expenditure only, a model for the
summation of maintenance and renewals expenditure and a model for the summation of
maintenance and renewals expenditure but with an additional adjustment to expenditure
to take into account variations in the volume of track renewals. The latter model is a crude
attempt to account for the lumpy nature of reinvestment expenditures in railway infrastructure.
A clear picture emerges from these results. Post 2000, efficiency performance by Network Rail
deteriorated considerably. Only in 2004 did performance start to improve, however there was
still substantial inefficiency in 2006. The models indicate that Network Rail was approximately
60 % efficient in the final year. Taking this model as literal implies that Network Rail can
continue to operate the same length network, accommodate the same traffic and maintain
the same quality, but reduce costs to 60 % of their realised value if they adopted best practice
from other railways.
Of course this model is just that, a model. As such it is an abstraction of reality and care
must be taken in relying solely on this evidence to make a regulatory determination (about

302
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 15.8 Conclusions

Fig. 15.2: Network Rail efficiency profile

the funding level required for Network Rail). As such the ORR considered many sources
of evidence, both econometric and bottom up engineering analysis, to come to its final
conclusions (see Smith (2012) for a full discussion). It ended up setting a 21 % efficiency target
(cost saving target) over the period.
Overall, econometric efficiency frontier analysis can be used to great benefit in economic
regulation. Whilst it may not be able to provide the fine detail of how and why costs differ
between railways – only detailed bottom up analysis can do that – it provides objective
evidence based on the actual costs experienced by the firm and industry. Crucially the
availability and quality of the data generally determines the weight that can be placed on the
outcomes of the analysis in a regulatory determination.

15.8 Conclusions

A number of different economic approaches to measurement of performance have been

discussed in this Chapter and all have their uses. It is important to distinguish between
efficiency in turning inputs into intermediate outputs and effectiveness in producing sold output
(passenger kilometres and freight tonne kilometres). Partial productivity measures may be
helpful in understanding particular aspects of performance of individual railways, but care must
be taken to compare similar systems in terms of the nature of outputs and the environment
in which the railway is operating. Data Envelopment Analysis has become a popular way of
examining performance since it can simultaneously allow for multiple inputs and outputs whilst
avoiding the need for financial data, which can be problematic in terms of availability and
comparability. However, it has an important drawback in that no statistical tests exist for its
results and no allowance for random noise. Moreover, differing degrees of subcontracting can
distort measures of the labour force of the companies concerned.
Thus in general the Authors conclude that cost frontier methods provide the most appropriate
efficiency measures for railways. However, they are not without their problems, both technical

303
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
15 Economic performance assessment in railways

(in that strong assumptions are needed to estimate them) and practical (particularly in the
need for comparable data on capital costs). Thus the analyst of railway performance must be
pragmatic and use a variety of methods according to circumstances, always bearing in mind
the factors that may lead to misleading results.

References
Battese G. E., Coelli T. J. (1995): A Model for Technical Inefficiency Effects in a Stochastic
Frontier Production Function for Panel Data, Empirical Economics, 20, pp. 325 – 332.
Canto, P., Pastor J. M., Serrano L. (2010): Vertical and Horizontal Separation in the European
Railway Sector and its effects on productivity, Journal of Transport Economics and Policy, 22
(2), pp. 139 – 160.
C E Delft, Infras, Fraunhofer ISI (2011) External Costs of Transport in Europe Update Study for
2008 Delft, CE Delft.
Chambers R. G. (1988): Applied production analysis. Cambridge University Press, UK.
Charnes A., Cooper, W. W., Rhodes E. (1978): Measuring the Efficiency of Decision Making
Units. European Journal of Operational Research 2, pp. 429 – 444.
Coelli T. J., Rao D. S. P., O’Donnel C. J., Battese G. E. (2005): An Introduction to Efficiency
and Productivity Analysis. Second Edition, Springer.
Cowie J. (2009): The British Passenger Rail Privatisation: Conclusions on Subsidy and
Efficiency from the First Round of Franchises, Journal of Transport Economics and Policy,
43(1), pp. 85 – 104.
Cowie J., Riddington G. (1996): Measuring the Efficiency of European Railways, Applied
Economics, 28, pp. 1027 – 1035.
Deprins D., Simar L. (1989): Estimating Technical Inefficiencies With Correction For
Environmental Conditions, Annals of Public and Cooperative Economics, pp. 81 – 102.
Farrell M. J. (1957): The Measurement of Productive Efficiency, Journal of the Royal
Statistical Society, Series A (general), 120 (3), pp. 253 – 290.
Fielding G. J., Babitsky T. T., Brenner M. E. (1985): Performance Evaluation for Bus Transit,
Transportation Research, 19A, pp. 73 – 82.
Fuss M., McFadden, D. (1979): Production economics: a dual approach to theory and
applications, Edited by Melvyn Fuss and Daniel McFadden, North Holland: Amsterdam.
Greene W. H. (1980): Maximum Likelihood Estimation of Econometric Frontier Functions,
Journal of Econometrics 13:1 (May), pp. 27 – 56.
Hansen I. A., Wiggenraad P. B. L., Wolff J. W. (2013): Benchmark Analysis of Railway
Networks and Undertakings. 5th International Seminar on Railway Operations Modelling and
Analysis, Copenhagen.
Kumbhakar S. C., Lovell C. A. K. (2000): Stochastic Frontier Analysis, Cambridge University
Press, Cambridge UK.
Lan L. W., Lin E. T. J. (2006): Performance Measurement for Railway Transport: Stochastic
Distance Functions with Inefficiency and Ineffectiveness Effects. Journal of Transport
Economics and Policy, 40 (3), pp. 383 – 408.
Maibach M., Schreyer C., Sutter D., van Essen H. P., Boon B. H., Smokers R., Schroten
A., Doll C., Pawlowska B., Bak M., (2008): Handbook on estimation of external costs in the
transport sector. Brussels: European Commission.

304
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 References

Merkert R., Smith A. S. J., Nash C. A. (2009): Benchmarking of train operating firms – A
transaction cost efficiency analysis, Journal of Transportation Planning and Technology.
Nash C. A. (2006): Alternative models for restructuring. In: Gomez-Ibanez J. A. and de Rus G.
(eds.) Competition in the Railway Industry. An international comparative analysis, Edward Elgar,
Cheltenham, pp. 25 – 48.
Nerlove M. (1963): Returns to Scale in Electricity Supply, In Measurement in Economics:
Studies in Mathematical Economics and Econometrics in Memory of Yehuda Grunfeld.
Stanford University.
Oum T. H., Yu C. (1994): Economic Efficiency of Railways and Implications for Public Policy:
A Comparative Study of the OECD Countries’ Railways, Journal of Transport Economics and
Policy, 28, pp. 121 – 138.
Oum T. H., Waters W. G. (II), Yu C. (1999): A Survey of Productivity and Efficiency
Measurement in Rail Transport, Journal of Transport Economics and Policy, 33 (I), pp. 9 – 42.
Shephard R. W. (1953): Cost and Production Functions. Princeton: Princeton University
Press.
Smith A. S. J. (2012): The application of stochastic frontier panel models in economic
regulation: Experience from the European rail sector, Transportation Research Part E, 48,
pp. 503 – 515.
Smith A. S. J., Wheat P. (2012): Evaluating alternative policy responses to franchise failure:
Evidence from the passenger rail sector in Britain, Journal of Transport Economics and Policy,
46 (1), pp. 25 – 49.
Spady R. H., Friedlaender A. F. (1978): Hedonic cost functions for the regulated trucking
industry, The Bell Journal of Economics, 9 (1), pp. 159 – 179.
Tsionas E., Christopoulos D. K. (1999): Determinants of Technical Inefficiency in European
Railways: Simultaneous Estimation of Firm-Specific and Time-Varying Inefficiency,
KONJUNKTURPOLITIK, 45, pp. 240 – 256.
Wheat P., Smith, A. (2008): Assessing the Marginal Infrastructure Maintenance Wear and
Tear Costs for Britain’s Railway Network, Journal of Transport Economics and Policy, 42 (2),
pp. 189 – 224.
Wheat P., Smith A. (2014): Do the usual results of railway returns to scale and density hold in
the case of heterogeneity in outputs: A hedonic cost function approach. Journal of Transport
Economics and Policy.

305
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 16 Conclusions

16 Conclusions
Ingo Arne Hansen, Jörn Pachl

This book contains a concise overview of the current methods of designing, analysing, optimising,
simulating and evaluating the quality and efficiency of railway timetables and operations.
The timetabling, of course, must aim to model as realistically as possible the intended
arrival and departure times at the stations. This is both to comply with the transport service
promised and to satisfy the expectations of the customers. As the railway traffic depends on
many internal and external constraints and forces, it can never follow the original schedule
deterministically, but the train trajectories vary. The art of design and construction of railway
timetables is to develop a consistent, stable and robust model of train operations that
supports the railway personnel and the technical equipment involved efficiently. Timetables
must also match market demands and generate economic and social benefits.
The prime keys for high quality timetabling are:
– good representation of the track alignment and other infrastructure,
– incorporation of signalling constraints,
– state-of-the-art capacity modelling, based on blocking time theory,
– high precision in the estimation of train running times,
– application of proven energy-efficient driving standards,
– stability and robustness analyses, focusing on conflict points, bottlenecks and networks,
– combination of deterministic, stochastic analytical and simulation models,
– regular monitoring, analysis, the evaluation of and feedback from, operations.
The Editors are convinced that conflict-free timetabling is essential for achieving high
performance and reliability of train operations. The potential points of conflict in a railway
network are not limited to stations and platform tracks, but also where lines and routes
merge or cross. All need to be integrated into a consistent infrastructure model and the
corresponding scheduled running times along the various routes should be included. None
of this has any direct relevance to the customers, only to the infrastructure manager and train
operating companies.
Thus a distinction must be made between the published external timetable and the in-house
working timetable. The first is derived from the second, but it contains much less detail and
may even allow for time windows between the (earliest) departure times and (latest) arrival
times. This improves the flexibility of operations and traffic control.
Currently, a lot of timetables are still based only on deterministic running times and (minimal)
headway times between the trains, rounded to full minutes and represented by simple train
diagrams. These may contain hidden conflicts between the train paths if the standard minimal
headway times do not match the infrastructure constraints. This may be due to the specific
length of the signal blocks (sections), local speed restrictions, train deceleration rates or train
lengths. Thus the minimum headway time between individual pairs of trains may not only be
greater than those scheduled, but may also vary at the same location due to different driver
behaviours or weather impacts.
The first and necessary step to guarantee the existence of a conflict-free timetable is the
estimation of blocking times. Their added value is that the corresponding blocking time
graphs indicate easily any conflict in a timetable by means of an overlap between different
trains, estimating precisely the remaining buffer time between two train paths (in seconds!) and
guaranteeing a smooth running of trains without unnecessary deceleration when the schedule
is followed strictly.

307
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
16 Conclusions

The challenge for future research is to develop further models that enable estimation of the
optimal size of timetable allowances for running times, dwell times and buffer times. They
need also to distribute the margins over the lines and network such that the trains recover
quickly from initial delays and reduce knock-on delays as much as possible. The quality of a
timetable would then be expressed as a function of the consumption of scheduled capacity,
the distribution of blocking times and allowances, the periodic synchronisation of feeder
and distributor lines at the network nodes and the ability to save energy and recover from
perturbations.
The stability, feasibility and robustness of timetables for individual dedicated lines such as
high-speed or metro lines can be proven easily unless the rate of track capacity use does not
exceed 75 % during peak hours. Optimal periodic network timetables offering reliable short
transfer connections between different lines at nodes can be designed, provided that conflict-
free routes with running and dwell times corresponding to an integer multiple of half the cycle
time between the nodes exist. However, timetabling of lines along (partly) single track sections,
merging/diverging tracks with at-grade crossings, and of passenger and freight trains with
different speed profiles is very complicated. User-friendly tools for automatic accurate train
scheduling in networks of interconnected lines at different speed levels and frequencies are still
to be developed.
Concerning the estimation of running times, the available tools allow fairly precise and fast
calculations for each scheduled train, with a precision of seconds. This is provided that
detailed data on infrastructure and rolling stock is available and up-to-date. However, the
power, load, length and weight of trains may vary, in particular, if different train operators use
the same routes, whereas the current timetables are still based on standard rolling stock data.
Here, traction and braking force/speed characteristics and resistance coefficients are often
treated as confidential. The resistance between wheel rim and rail also depends a lot on actual
adhesion weight and weather conditions, while the tools for the calculation of running times
are still based on deterministic parameters.
Although train operations are monitored automatically, the realised running times, so far, are
insufficiently analysed with regard to the determination of statistically valid distributions. The
fine-tuning of the input parameters is needed, such that the output of analytical and simulation
models corresponds even better with the revealed railway process times. This enables a more
accurate prediction of actual train delays. Automatic route conflict detection tools based on train
describer and track occupation/clearance records have been developed in the past decade.
However, their practical implementation and use as decision support for traffic and network
controllers is still limited to a few networks. The ongoing modernisation of outdated signalling
and safety systems through the introduction of ETCS/ERTMS, electronic interlocking and
CBTC respectively, as well as train positioning via GPRS, generates a large treasure of empirical
data. This needs to be exploited by means of intelligent data mining, on-line analysis, real-time
information and decision support tools, for customers, drivers and traffic control centre staff alike.
In future, train driver assistance systems will enable the transmission of actual train speed,
weight, length and position data from the trains to the monitoring system of traffic control
centres. This can be used also for real-time prediction of train running times, arrival and
departure delays, as well as for automatic conflict detection. The accuracy of the calculated
advisory speed information, its mode of presentation to the driver (without compromising
safety by being a distraction) and the real-time computation of the start of coasting and
braking, are challenges for current research and development. These are needed in order to
improve the effectiveness of rescheduling measures, increase punctuality and reduce energy
consumption in the case of train delays.

308
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 16 Conclusions

For the estimation of capacity in bottlenecks, especially in complicated yards and at major
stations, the existing analytical approaches need to be developed further. The propagation
of delays between serial route nodes ought to be modelled depending on the signalling and
safety constraints, as well as on the variation of the routes and process times of individual
trains operating on different lines. More reliable, exact and non-discriminatory automatic
registration tools of initial and knock-on delays, as well as on-line delay prediction models
for lines, corridors and larger networks are necessary. This will allow a better insight into the
reasons for disturbances and to forecast their duration better.
More realistic short-time predictions of process times in the case of perturbations are
needed. These must take account of standard dispatching measures, support network traffic
management effectively, improve the quality of dynamic passenger information and satisfy
customer demand. The existing off-line deterministic models for stability analysis of timetables
are to be upgraded. This offers fast on-line decision support for the evaluation of suitable
dispatching measures and the selection of that judged the most efficient.
Optimisation models are proving their power to generate (near) optimal solutions for the
scheduling problems of trains in heavily occupied, complex and large networks. Stochastic
optimisation experiments show that standard fixed percentages to supplement running
times do not contribute as much as they might to the robustness of train services. These
need therefore to be varied in size and space according to the specific characteristics of the
infrastructure, rolling stock, timetable and expected traffic. The same conclusion is valid for
dwell time margins, synchronisation times and buffer times. This will require more innovative
research. Optimal network scheduling models for periodic service intentions must be able to
cope also with individual on-demand (freight) train path requests. As the computation effort
for the resolution of scheduling problems in large railway networks is still costly, the aim is to
find intelligent and fast algorithms. These must be able to handle signalling constraints in an
effective manner, without any loss of feasibility in practical operating matters.
Recently, new real-time rescheduling models for large networks have been developed. These
are different from offline scheduling models, in that they must be able to compute near-
optimal solutions within very short timescales. They must then generate feasible traffic plans
during the transition period of light disturbances and heavy disruption until complete recovery.
These need to exploit fully the available running and dwell time margins, plus reordering and
rerouting alternatives, as well as cancelling and reversing short of destination. These must
all take place in order to comply with temporary track capacity, rolling stock and crew duty
constraints. Successful test running and demonstration of the performance of real-time
rescheduling models in the daily practice of network traffic control is still a major challenge.
Automatic rescheduling tools, including reordering and local rerouting measures in station
areas, as well as reliable train speed advices for train drivers, would be of great benefit. These
can help to solve regular local disturbances that do not need the manual intervention of traffic
controllers. In the case of major disruption, track blockages and train breakdowns, however,
the decisions of transport, traffic and network controllers would be supported by an efficient
real-time rescheduling system. This would present clearly the costs and benefits of alternative
dispatching measures for larger networks.
The existing simulation models contain a great variety of features and performance in railway
timetabling and traffic management support. Their power stems from the ability of microscopic
modelling of the railway infrastructure. This includes the signalling system, the exact
computation of the rolling stock dynamics, and the impact assessment of different timetable
options based on a number of random realisations. As the microscopic simulation of different
train schedules in large and heavily occupied networks is not feasible online due to high

309
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
16 Conclusions

computation costs and the complexity of output evaluation in the short term, an integrated
micro-macroscopic simulation and global network optimisation model has been developed.
It has been applied successfully in recent years. The innovative micro-macro simulation and
optimisation approach needs to be expanded through fast routines, user-friendly graphical
interfaces and performance assessment modules. It will then be accepted in practice by
professional timetable designers and experienced traffic managers.
The growing power of simulation models may entice users to overestimate the outcome
if they are not aware of the limitations. The results are valid only for the cases investigated,
the set of input data and the algorithms applied. It is not sufficient to compare the results of
different simulation experiments with each other, in order to select the best timetable option.
It is essential to prove the consistency of the simulation model with respect to the expected
results of physical laws describing the real-world train process. The need is to calibrate the
model parameters, such that the simulated output corresponds to empirical measurements.
The correctness of the simulation results needs to be verified under different conditions, by
comparison with separately recorded and computed data. As the kernel of micro-simulation
tools consists of (multiple) deterministic computations of individual train movements, signalling
and interlocking states respectively, the behaviour of drivers, dispatchers and the technical
subsystems may be modelled differently from that observed in reality. It follows that the
derivation and application of sound statistical distributions, reflecting the stochasticity of train
operations in a given railway network, is a prerequisite for a thorough evaluation of timetables
and traffic management strategies by means of simulation.
The performance of railway systems depends first and foremost on the availability, reliability
and robustness of the hardware subsystems, as well as on the quality of the timetable and the
level of service offered and realised. Close collaboration of planners, engineers, managers and
researchers from many scientific disciplines is necessary to make continuous improvements in
the attractiveness of railway transport for regular, occasional and new customers.
The effectiveness and efficiency of timetabling, operations, and maintenance are key economic
criteria in the performance evaluation of the railway business. A comprehensive service quality
monitoring and management system can show the individual contribution of the railway
personnel, staff members and companies in reaching the goals of the industry and to sensitise
the people involved and stakeholders. It is hoped that the contents of this book will offer
useful information and appropriate scientific methods to stimulate professionals, students and
researchers for further innovations in railway timetabling and traffic management.

310
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Glossary

Glossary

A
Adhesion value A factor that describes the maximum tractive effort that
can be applied until the wheels start to slip, according to
the coefficient of friction. The formula of Curtius & Kniffler
describes a relationship between adhesion value and velocity.
Algorithm A well-defined procedure to find a feasible solution to a
problem in a finite number of steps.
Alternative graph A generalisation of the job shop scheduling problem with
additional real-world constraints.
Analytical model A description in terms of mathematical relationships between
certain entities that are relevant in a certain problem setting.
Analytic solution The solution of a differential equation or of an integral which
(closed formula) is given by a formula that uses arithmetic operations and
standard functions (as exponentiation, logarithm, sinus,
arcus-sinus etc), also called a closed formula.
Arrival process In queueing theory, a (stochastic) process modelling the
arrivals of the customers (trains) at a service system.
Arrival delay A deviation of the arrival time from the scheduled arrival time
at a station or other specified location.
Arrival rate Reciprocal of the mean inter-arrival time, measured in arrivals
per unit time.
Automatic route setting A system which provides the automatic setting of a given
route when a train approaches a signal, provided that the
route is currently available.
Availability Degree to which a system, subsystem, or component is
operable for a specified period of time, or the proportion of
time a system, subsystem, or component is functional. It
describes the duration of functionality.

B
Behaviour section The movement of a train consists of acceleration, constant
movement, coasting or run-out, and braking. These sections
of the movement describe the behaviour of the train. It is
dependent indirectly on the characteristic sections (by speed
limits, gradient etc.), efforts, resistances, and the train driver’s
actions.
Blocking time The time interval during which a section of track is allocated
for the exclusive use of one train and is therefore blocked to
all other trains.
Block section A section of track in a fixed block system, which a train may
enter only when it is not occupied by another train or vehicles.

311
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
Glossary

Buffer time Extra time that is added to the minimum line headway to
avoid the transmission of small delays.

C
Cab Signalling A signalling system that displays the movement authority on
the driver’s desk, rather than the use of lineside signals.
Capacity of infrastructure The maximum number of trains that may in theory be
operated over a defined part of the infrastructure over a
certain period of time, though this limiting value is unlikely to
be reached in practice.
Coasting speed The train speed at which the driver switches off traction
power to go into coasting mode.
Conflict detection The system to find potentially conflicting train movements in a
given time period.
Conflict resolution The system to solve the potential conflicts in train movements
on the basis of the most suitable and robust dispatching
options.
Consecutive delay Another term for a knock-on delay.
Constraints In an optimisation model, the constraints describe which
combinations of values for the decision variables are allowed.
Corrected Ordinary Least A simple statistical technique to measure the efficiency of
Squares firms, based on Ordinary Least Squares regression.
Crossing An assembly of rails that enables two tracks to cross each
other at grade.
Cyclic timetable A timetable which repeats itself over a certain period of time
(usually one hour).

D
Data Envelopment Analysis A mathematical programming technique to measure the
efficiency of a firm.
Deadlock A traffic situation involving a set of trains each claiming a
resource which is not available, leading to the impossibility of
further movement by any train in the set.
Decision variables Variables that are used to describe a solution in an
optimisation problem. The decision variables obtain a value
by applying an algorithm.
Delay The deviation from either a scheduled event or process time
of a train.
Difference equation approach A method of handling a system of differential equations by
replacing each appearing differential quotient by a difference
quotient. This can provide an approximate solution, even if an
analytical solution appears to be impossible.

312
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Glossary

Dijkstra algorithm Widely used calculation of shortest paths in graph networks

between two nodes according to different criteria.
Dispatcher An alternative term for a traffic controller or (mainly in North
American use) a signaller.
Disruptions Blockades or major failures that need large scale service
adjustment.
Distant signal A lineside caution signal that cannot itself show a stop
indication but which provides the driver with the status of the
line ahead as displayed by the next signal in advance.
Disturbed train operation An operating situation in which, due to a conflict with another
train, a train will encounter a signal showing a restrictive
aspect and will therefore have to modify its speed. See also
Undisturbed train operation.
Driving regimes The driving modes that may be applied by the driver. There
are four driving regimes: Acceleration, Cruising (going at
constant speed, realised by partial acceleration or braking),
Coasting (no traction or braking applied), Braking.
Dwell time The total elapsed time from when a train stops in a station
until it resumes moving. Note that dwell time includes that
needed to open and close doors and dispatch the train; it is
not just that during which passengers may join or alight.
Dwell time reserve Amount of time which is added to the calculated dwell times
in order to compensate for randomness of real dwell times or
to obtain departure times “to the full minute”.

E
(Economic) efficiency A proportionate measure which indicates the potential cost
saving from adopting managerial best practice.

F
Feasible solution An assignment of values to the decision variables such that
none of the constraints is violated.
First-in first-out discipline A principle which describes the behaviour of a queue. The
customers (trains) are served in the order in which they arrive,
rather than any predetermined priority of one type of train
over another.
Frontier function A mathematical function which represents the boundary
of a feasible production or cost set. For example, a cost
frontier shows the minimum cost of producing a given output
at given input prices. In practice firms can produce on this
boundary or at a higher cost (if they are less efficient).

313
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
Glossary

G
Gauss Quadrature Integral A special method to compute integrals numerically. The given
Formula function is approximated by a polynomial in a special way as
also described in the book.
Goodness-of-fit test A type of statistical test used to assess whether a data set
comes from a hypothesised distribution.
Gradient The rate at which a railway track rises or falls in relation to the
horizontal, usually measured as ‘per thousand’ rather than
‘per cent’. Thus 1 in 200 becomes 5 ‰.
Grandfather rights The continuation of any existing rights, such as that to
continue to use a train path in a future scheduling period
when it was already in use.

H
Headway The space or time interval between two successive trains.
Heavy traffic approximation Approximation formula for parameters of a waiting system
which provides good results, especially for high utilisation.
Homogeneous strip A calculation model in which the train is considered as a strip
of a fixed length. As an approximation, the weight of the train
is assumed to be distributed equally over its entire length.

I
Initial delay Any delay recorded at the cordons of an investigated network
when the train enters that network.
Integer Programming model A mathematical model in which all relations between the
decision variables in the constraints and in the objective
function are linear, and in which some of the decision
variables are required to obtain an integer value.
Inter-arrival time Difference between succeeding arrival times of trains or train
paths.
Interlocking An arrangement by which points and signals are electrically
or otherwise interconnected, which prevents different parts of
the system from giving contrary indications or allowing unsafe
operational conditions to be created.

J
Junction margin The minimum permissible time between one train passing a
junction and a second train passing the junction, where the
second train crosses the path of the first.

314
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Glossary

K
Knock-on delay A delay caused by other trains due to either short headway
times or late transfer connections.

L
LeiDis-S/K Leitsystem Disposition Strecke/Knoten, a system for
computer aided dispatching in DB Netz AG’s operations
control centres in Germany.
Linear Programming A mathematical model that is obtained from an Integer
relaxation Programming model by relaxing the requirement that some of
the decision variables should obtain an integer value.
Line resistance The traction effort of the train that is consumed by the
infrastructure, especially by gradients and curves.
Link A connection between two nodes.
Link-orientated models An infrastructure model in which each link contains
all relevant information (e.g. speed, gradient, radius,
electrification, direction).
Loss system In queueing theory, a service system where customers leave
the system if all service channels are busy.

M
Macroscopic infrastructure Node-link-models that contain aggregated information on
model nodes and links.
Main signal A lineside signal that authorises a running movement to enter
a section of line.
Main track A track which is used for train running movements, as
opposed to sidings.
Mass point A calculation model in which the train is considered as a point
that has a certain mass. Running time calculations become
easier if the train is considered as a mass point. Usually the
position of the mass point is approximated to the middle of
the train.
Maximum likelihood method A method for parameter estimation of data distribution,
based on a likelihood function of data sample.
Mechanical energy Energy needed to accelerate the train, to overcome the
consumption running and track resistances.
Mesoscopic infrastructure Node-link-models as syntheses of microscopic and
model macroscopic infrastructure models. Signal blocks and
headways are modelled only in stations and interlocking
areas, and not on links.
Method of moments A method for parameter estimation of data distribution,
carried out by equating sample moments with population
moments.

315
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
Glossary

Microscopic infrastructure Node link models that contain, depending on their purpose,
model the highest possible level of detail on nodes and links.
Minimum headway time Minimum inter-arrival time between two train paths,
respecting the requested speed profile of both trains. The
minimum headway time is determined in the first block
section of that part of the infrastructure which is used
in common by both trains. The first block section is not
necessarily the relevant block section where the blocking
times would touch without tolerance.
Modelling of railway The transition of real world railway infrastructure into a
infrastructure computerised infrastructure model.
Multi scale infrastructure Multi scale models that contain both areas that are modelled
models on a microscopic level and areas that are modelled on a
macroscopic level.

N
Newton’s method An approximation method to find zeros of a function. This
method uses derivatives of the given function.
Node A representation of an arbitrary location in a railway network.
Node-orientated models An infrastructure model in which special node types indicate
the change of attributes (e.g. speed by a speed change
node, gradient by a gradient change node etc.).

O
Objective function A mathematical representation of the objective that is aimed
at in terms of the decision variables.
Optimal solution A feasible solution to an optimisation problem is optimal if
another feasible solution with a better (higher/lower) objective
value does not exist.
Optimisation model A mathematical representation of a problem that aims at
finding feasible solutions to instances of the problem that are
as good as possible in terms of the objective function used.
The ability of a system or component to perform as designed.
Original delay A delay generated within the investigated network and not
caused by other trains.
Outlier A data point deviating substantially from the bulk of the data.
Overlap A certain length of track beyond a signal that must remain
unoccupied as long as a moving train is in rear of that signal.

P
Periodic Event Scheduling This is a standard model for representing scheduling
Problem (PESP) problems in which a number of events have to be scheduled
in a periodic way, in particular timetabling problems.

316
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Glossary

Perturbations Modifications and delays to scheduled operations due

to external sources. Small ones can be recovered by the
timetable plan, larger ones need intervention by traffic
controllers.
Points Trackside elements at junctions that may be aligned to one of
two positions depending on the route of a train.
Possession A period of time when engineering work is being undertaken
to the railway infrastructure, preventing trains from being run
or reducing the number of trains that can be run.
Productivity The ratio of output to input in a production process.
Productivity varies between firms according to their efficiency,
their scale of operation, and technical changes.
Punctuality The percentage of trains which arrive at (depart from, pass) a
location with a delay less than a certain time in minutes.

R
Random mix A model describing the traffic mix in a service system by
means of probabilities.
Recovery time A time supplement that is added to the pure running time to
enable a train to make up small delays.
Reliability The ability of a system or component to perform as designed.
Resource Tree Conflict Graph A model for representing in an efficient way the scheduling
(RTCG) possibilities of a train in a main station area. A path from the
root to a leaf represents a schedule for a train.
Returns to density A measure of how average costs change as the utilisation of
a network with fixed infrastructure is increased. Increasing
returns to density indicate that average costs fall as utilisation
increases, while decreasing returns to density indicate that
average costs rise as utilisation increases.
Returns to scale A measure of how average costs change as the size of a firm
is increased. Increasing returns to scale indicate that average
costs fall as the size of the firm increases, while decreasing
returns to scale indicate that average costs rise as the size of
the firm increases.
Robustness The ability or otherwise of a system or component to
withstand model errors, parameter variations, or changes in
operational conditions.
Route node A set of station tracks including switches and crossings in a
station throat which adjoins lines and directions of operation.
Running line British term for a main track.
Running movement A locomotive or train with authority to operate on a main
track under the control of main signals.

317
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
Glossary

Running time The time required to run over a given stretch of track. A
distinction can be made between planned running time
(including recovery time) and minimum operational running
time (trains run at the maximum speed permitted to recover
delays).
Running time reserve See Recovery time.

S
Scheduled waiting time An artificial increase in the overall timing of a train which is
caused by the resolution of conflicts during the scheduling
process.
Serial route node A part of the infrastructure which can be used by one train at
a time, used for capacity analysis.
Service channel In queueing theory, a part of the infrastructure which is
occupied exclusively by one customer.
Service process In queueing theory, a (stochastic) process modelling the
behaviour of customers in a service channel.
Service rate Reciprocal of the mean service time, measured in customers
per unit time.
Service system In queueing theory, a certain infrastructure where customers
request a service.
Service time The time a customer occupies a certain part of the
infrastructure for service. In railway operations research the
service times are often represented by the minimum headway
times or, in some special cases, by the blocking times.
Shunting movement A movement for making up trains, moving vehicles from one
track to another, or a similar purpose.
Shunting signal A lineside signal that authorises a shunting movement.
Signal berth The track section on the approach to a signal.
Signaller A person responsible for issuing movement authorities and
for operating signalling appliances.
Simulation model A mathematical representation of a system that can be
used to mimic the processes of the system under varying
circumstances. Usually, a simulation model is operated
subject to stochastic disturbances.
Speed bundling A scheduling principle where trains run at harmonised
speeds in order to make best use of the railway line capacity,
commonly found on Metro systems.
Stability The ability of a system or component to compensate for
delays and return to the desired state.
Station Place designated as a stopping place, by name, in the
timetable. Note that in UIC code 406, the term station is only
used for points of a network where overtaking, crossing, or
direction reversal is possible.
318
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Glossary

Steady state A state of a stochastic service system in which the recently

observed behaviour of the system will continue into the
future. The probability that a certain state will be repeated will
remain constant.
Stochastic Frontier Analysis A set of statistical techniques which explicitly models
efficiency relative to a frontier function.
Stochastic Optimisation A combination of an optimisation and a simulation model.
model The optimisation part of the model is guided by the results of
the simulation part of the model.
Stochastic process A process with some indeterminacy in its future evolution
described by probability distributions.
Substation Part of the power supply network, where the electricity is
transformed down to that needed by the trains and rectified if
necessary (Voltage, Frequency).
Switching points Places where the driving regime of the train changes; can be
(Train driving) characterised by speed, time, position or a combination of
those. See also Coasting speed.

T
Technical change The change in costs or productivity resulting from
improvements (or otherwise) to the technology available
for production. This is a temporal effect and unrelated to
changes in the scale of the firm or the efficiency of the firm.
Traction energy consumption Energy needed to accelerate the train, see also mechanical
energy consumption. Note that neither the energy
needed to operate auxiliary equipment on-board nor the
regenerated energy from braking are part of the traction
energy consumption. Some sources define traction energy
consumption as including losses caused by the traction
system. Here, it is used as a synonym for mechanical energy
consumption.
Traction unit resistance The vehicle resistance of a locomotive or multiple unit.
Tractive effort The effort of a locomotive or a multiple unit which is intended
to move the train. The induced tractive effort is the whole
effort generated. The tractive effort at wheel rim FTr is
the remaining part after considering restrictions of power
transmission or overheating.
Traffic controller A person in charge of dispatching, i.e. changing train speeds,
orders and routes. Controllers can be further organised into
a hierarchy, taking care for instance of network and local
issues.

319
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
Glossary

Traffic diagram A time-distance diagram that contains the train paths of all
trains that run on a line.
Train dispatching system A decision support system for railway traffic controllers. In
proactive traffic management systems, traffic controllers aim
to prevent disturbed situations through conflict detection and
resolution systems.
Train path That part of the capacity of the railway infrastructure which
is necessary to schedule or to run a train with a requested
speed profile and a given performance capability.
Transport manager A train operating company supervisor responsible for train
running and for managing vehicles and staff.
Transport time The total of all the running and dwell times in a train’s end to
end journey (train related) or the total hours during which the
service is operated (passenger related).
Travel time Sum of all the time taken by a passenger from origin to final
destination of a trip, including access.

U
Undisturbed train operation A train can go from one planned stop to the next planned
stop according to the timetable. It will not encounter any
restrictive signal aspects due to conflicts with other trains.
See also Disturbed train operation.
Upper/Lower bound The limits of the objective function in an optimisation model.
If the upper and the lower bound are equal, then the optimal
solution is found.

V
Vehicle resistance The resistances caused by air and wheels. They work against
the tractive effort of the traction unit.

W
Waiting-loss system In queueing theory, a service system where customers have
either to wait in a waiting space or leave the system if all
service channels are busy.
Waiting system In queueing theory, a service system where customers have
to wait in a waiting space if all service channels are busy.

320
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Keywords

Keywords

A Conflict(s) 97, 255

Acceleration 73 Conflict-free 175
Adhesion value 69 Conflict graph model (CG) 180
Alternative graph 265 Consecutive delay 279, 283
Analytical capacity 123 Constant movement 73
Approach time 24, 88 Consumed capacity 38
Arrival process 117 Conzen-Ott-Machine 65
Arrival rate 118 Cost-benefit analysis 287

Asynchronous simulation 194, 199 Cost-effectiveness analysis 287

Automatic train operation 114 Costs of delay 280, 281

Crossing 15
Cruising 73
B
Cycle time 43
Behaviour sections 73
Cyclic timetable 43, 138, 157
Bending 97
Blocking time(s) 23, 87
D
Blocking time stairway 25
Deadlock 255
Block section(s) 16, 22, 88
Delay(s) 135, 146, 150, 204
Bottleneck-analysis 282
Delay-coefficient 279, 282
Braking distance 75
Delay propagation 146
Braking forces 73
Derailers 15
Buffer path 37
Deterministic simulation 193, 201
Buffer time(s) 35, 100, 135
Diesel fuel 100
Direct-demand approach 279
C
Dispatching rules 210
Cab signalling 23, 27 Disruptions 255
Calibration 207 Distant signal(s) 88
Capacity 278, 279, 282, 283, 284, 285 Distribution fitting 222
CBTC 112 Distribution of delay 279
CDR 269 Driver Advisory System (DAS) 107
CDRFR 266 Dwell time(s) 20, 87, 99, 285
Coasting 74 Dynamic programming 99, 103, 104
Compensation zones 177, 184
Compression method 38 E
Computer-based scheduling 42 Economic evaluation 287, 289, 291
Condensation zones 177, 180 Effectiveness 276, 277, 278

321
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
Keywords

Efficiency 276, 277, 278 M

Electronic timetable 110 Macroscopic level 176
Energy consumption 91 Macroscopic node-link-models 48
ERTMS 114 Macroscopic simulation 193, 195
ETCS 25, 74, 114 Main signal 16
Evaluation parameter 278 Manual scheduling 40
Exchange format 62 Mass Point 77
Exploitation rate 282 Maximum capacity 284, 285
Maximum Principle 93
F Max-plus algebra 133, 136, 138, 142, 144,
Fixed block 16, 22 146, 148
Flexible Periodic Event Scheduling Problem Mesoscopic node-link-models 48
(FPESP) 178 Microscopic infrastructure model 53
Microscopic level 176
G Microscopic node-link-models 48
Gauss quadrature integral formulae 86
Microscopic simulation 193, 198, 209, 213
Genetic Algorithms 97, 104
Migration between infrastructure models 57
Geographical information system 62
Minimum headway 139, 278, 282
GPS 111, 219, 221
Minimum headway time 119
Gradient 66, 71, 73, 74
Modal-split approach 279
Monitoring 239, 240
H
Monte Carlo simulation 192
Headway(s) 26, 34
Motion equation 219, 221
Homogeneous strip 78
Movement authority 15, 23
Moving block 23, 29
I
Multiple-channel service systems 118
Infrastructure model 47
Multiple server queues 126
Initial delay(s) 129, 135, 146, 204, 218,
219, 222, 279, 282, 285
Interlocking 17, 30 N
Network traffic controllers 253
K Newton’s method 85
Knock-on delay(s) 127, 218, 219, 222
O
L Online Prediction 247
Line resistances 71 Operating performance 281
Lineside signals 16, 27 Operating program 282, 283, 285, 286
Local traffic controllers 253 Original delay(s) 218, 219, 222, 279
LZB 74 Overlap 22, 31

322
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Keywords

P S
Parameter estimation 224 Scheduled waiting times 119
Periodic Event Scheduling Problem Secondary delay 135
(PESP) 178 Serial route nodes 125
Perturbations 254 Service level 278, 279
Points 15 Service process 117

Primary delay 135 Service rate 118

Probabilistic approach 87 Shunting movements 15

Shunting signals 17
Process Mining 243
Simulated Annealing (SA) 231
Process time estimation 244
Simulation 191
Protected zone model 33
Single-channel service systems 118
Public sources 63
Speed profile 66, 184
Punctuality 218, 220, 222
Stability 133, 135, 142, 143
Standardised Evaluation of Infrastructure
Q 287
Quality of operation 276, 278, 279, 281, Station areas 14, 31
282, 283 Station track 17
Quality of service 278, 287, 289, 290 Steady state 133, 148, 150
Queueing theory 117 Stochastic blocking times 32
Stochastic simulation 193, 203
R Stochastic system 148

Rail vehicle 68 Subsequent cost analysis 287, 289, 290

Substation 101
Real-time railway traffic optimisation 262
Switching points 95
Real-time train traffic management 255, 256
Symmetry time 44
Recommended area of traffic flow 282, 283,
Synchronous simulation 194
285, 286
Recovery time(s) 20, 67, 144, 278, 280
T
Relative sensitivity of waiting time 283, 285
Theoretical capacity 124
Resource Tree 180
Throughput capacity 283, 284, 285
Resource Tree Conflict Graph (RTCG) 182
Time stamps 62
Rotating masses 71
Time supplement 135
Route 14 Timetable 133, 138, 139
Route conflicts 246 Tracks 14
Running movements 15 Traction unit resistance 70
Running time 20, 65, 74, 98 Tractive effort 71, 72
Running time supplement 91, 135 Traffic diagrams 18

323
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
Keywords

Traffic energy 283, 286 U

Traffic State Prediction 240, 241 Utility factor of capacity 285, 286
Train delay(s) 218, 222
Train describer data 219 V
Train event recorder data 219, 221 Value-benefit analysis 287
Train paths 21 Vehicle resistances 70
Train priorities 122
Train running resistance 93 W
Train separation 21 Wagon resistance 68
Train speed co-ordination 264 Waiting time 281, 282, 283, 284, 285
Tree Conflict Graph (TCG) 182 Waiting time function 283, 285
Turnout 15 Weighting factor of delay 280
Two-level approach 175 With/without case principle 287, 288

324
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 The Authors

The Authors

Dr.-Ing. Thomas Albrecht

Studied Traffic Engineering and Transport Operation Technology at
Technische Universität Dresden (TUD), where he received his doc-
toral degree in 2005. He completed a Post-Doc Fellowship at Delft
University of Technology in 2006 and then went back to Dresden
to continue working in the field of energy-efficient railway operation
and driver advisory systems at the Chair of Traffic Control Systems
and Process Automation of TUD. Since 2014 he works as a Busi-
ness Consultant in the area of Rail Traffic Management Systems for
CSC, a global leader of information technology (IT) services and so-
lutions.

Dipl.-Inform. Dr. rer. nat. Olaf Brünger

Studies of mathematics and computer science at University of Tech-
nology Aachen, 1989 diploma in computer science. Assistant sci-
entist at Institute of Traffic Science Aachen 1989 – 1994. Promotion
1995, dissertation about computer-assisted construction of detailed
railway timetables – this concept was later introduced by Deutsche
Bahn as FAKTUS and RUT. Since 1994 business engineer, consul-
tant and project manager at DB Systel GmbH (former TLC/DB Sys-
tems), working on projects concerning IT for railway timetable and
railway operations.

Dr. Gabrio Caimi

Studied Mathematics with core area operations research from 1998 to
2004 at ETH Zurich. During his diploma thesis at the Institute for Ope-
rations Research (IFOR) he worked in a running research project in col-
laboration with Swiss Federal Railways (SBB). From 2005 to 2009 he
performed a PhD thesis at ETH focusing on the development of new
methods and algorithms for the automatic generation of schedules in
large and highly utilised railway networks. In 2009 he moved to Bern-
Lötschberg-Simplon-Bahn (BLS). Since 2012 he is head of Project
Portfolio of BLS, in charge of the financing of the infrastructure, efficient
assignment and use of the available financial means.
325
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
The Authors

Dr. Francesco Corman

Assistant Professor at the Section of Transport Engineering and Lo-
gistics, Delft University of Technology. His research career started in
Italy at Roma Tre University, he got a PhD at Delft University of Tech-
nology and Post-Doc at Katholieke Universiteit Leuven, Belgium. He
has received a number of awards for his master thesis, a best paper
award by RAS, INFORMS, numerous nominations for awards for his
PhD thesis and papers. His research interests include optimisation
in railway traffic control, scheduling of discrete event systems, study
of dynamics of uncertainty in transport systems.

Dipl.-Math Dr. rer. nat. Elias Dahlhaus

Studies of mathematics and computer science at Berlin Univer-
sity of Technology,1978 diploma in mathematics. PhD in compu-
ter science 1983. Habilitation degree in computer science, 1991.
Since 1991, adjunct lecturer (Privatdozent) at University of Bonn.
1992 – 1995: Senior lecturer at University of Sydney. 1996 – 1999:
Acting professor at University of Cologne. 2000: Guest professor at
Vienna University of Technology. Since October 2000 consultant at
DB Systel GmbH (former TLC and DB Systems). Since 2002, ad-
junct lecturer at Darmstadt University of Technology.

Dr. Andrea D'Ariano

Assistant Professor in Operations Research at Department of
Engineering, Roma Tre University (Italy) since 2011. In April 2008,
he got the PhD title at Department of Transport and Planning, Delft
University of Technology (The Netherlands). His general research
interests include scheduling theory and its application to the ma-
nagement of public services. His publications were acknowledged
by TRAIL Research School, IAROR, INFORMS Railway Applica-
tions Section, IEEE as best papers; by IEEE Intelligent Transporta-
tion Systems Society as best PhD thesis; by AIRO as best Master
thesis. He got the 2012 Italian University Professorship Qualification.

Dr. Rob M.P. Goverde

Received an MSc in Mathematics from Utrecht University in 1993, a
Master of Technological Design in Mathematical Modelling and De-
cision Support from Delft University of Technology in 1995, and a
PhD in Railway Transportation from Delft University of Technology in
2005. Since 1996, he is affiliated with the Department of Transport
and Planning, Delft University of Technology, successively as PhD
student, Post-Doc, Assistant Professor, and currently as Associate
Professor. His research interests include railway operations, railway
timetable planning and analysis, railway traffic management, max-
plus algebra and mathematical optimisation.
326
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 The Authors

Prof. Dr.-Ing. Ingo A. Hansen

Studied Civil Engineering with a major of Transportation Planning
at the Technical University Hannover. PhD in 1978. He worked 17
years as transportation planner, project manager and vice-president
for different German consulting companies. From 1994 to 2011 he
was full professor for the design of transport facilities at Delft Univer-
sity of Technology. Currently he is visiting professor at Beijing Jiao-
tong University. His research focuses on the analysis and modelling
of delay propagation in railway networks. He is founding member and
President of the IAROR since 2005, editor-in-chief of the Journal of
Rail Transport Planning & Management and member of the Editorial
Advisory Board of the journal Transportation Research Part B.
Dr. Dennis Huisman
Part-time Associate Professor at the Econometric Institute of Eras-
mus University Rotterdam. Moreover, he works part-time as Exper-
tise Manager Logistic Processes at Netherlands Railways. He com-
bines these two jobs since 2004 after obtaining his PhD at Erasmus
University Rotterdam. His current research focuses on Operations
Research models and techniques that can support the planning and
operations of a railway operator in relation to complex logistical is-
sues. Moreover, he was a member of Netherlands Railways team
that won the Franz Edelman Award in 2008. Finally, he is one of the
founders of the Erasmus Center for Optimization in Public Transport.

Dr. Pavle Kecman

Studied traffic and transport engineering at the University of Bel-
grade, Serbia. In 2010 he joined the Department of Transport and
Planning at Delft University of Technology. His PhD research was a
part of the project: “Model-predictive railway traffic management”
sponsored by the Dutch Technology Foundation STW. After defen-
ding his PhD thesis in 2014, he started the post-doctoral research
as a part of EU project Capacity4Rail at the Department of Sci-
ence and Technology, Linköping University, Sweden. His research
interests include railway operations and application of data mining
and operations research to traffic and transport related problems.

Prof. Dr. Leo Kroon

Studied mathematics at the Free University in Amsterdam. In 1990
he defended his PhD thesis at Erasmus University Rotterdam. Cur-
rently he is a professor of Quantitative Logistics at the Rotterdam
School of Management of Erasmus University Rotterdam. He is also
a part-time logistic consultant within NS (Netherlands Railways), the
main operator of passenger trains in the Netherlands. His main re-
search interest is the development of model based decision support
systems for planning and real-time operations of railway systems,
including passenger movements. Leo Kroon is an associate editor
of the journal Transportation Science.
327
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
The Authors

Dr. Gábor Maróti

Studied mathematics at the Eötvös Univeristy in Budapest, Hunga-
ry, and received his PhD degree in Operation Research in 2006 at
the Technical University of Eindhoven, The Netherlands. Currently,
he is a researcher at Netherlands Railways, and a part-time assis-
tant professor at the VU University in Amsterdam, The Netherlands.
His research focuses on the application of operations research me-
thods for railway timetabling and rolling stock planning problems.

Prof. Dr.-Ing. Ullrich Martin

Studied Transportation Systems Engineering, HfV, Dresden.
1989 – 1992 Railway engineer and deputy head of the railway ope-
rations department in the regional administrative office Halle of the
German National Railways, 1992 – 1995 Researcher at TU Braun-
schweig, Institute of Railway Systems Engineering and Traffic Safety.
1995 Doctor degree from TU Braunschweig. 1995 – 1998 Railway
project manager, 1998 – 2001 Professor of Transportation Systems
Design at the University Leipzig. Since 2001 head of the Institute of
Railway and Transportation Engineering and director of the Institute
of Transportation Research at the University Stuttgart.

Dr.-Ing. Giorgio Medeossi

Studied Transportation Systems Engineering (2001 – 06) and recei-
ved a PhD at the University of Trieste (2010), where he is currently
researcher and contract professor. He was involved in several re-
search projects in the field of rail operations analysis, modelling and
simulation. With the paper “A method for using stochastic blocking
times to improve timetable planning” he was awarded the “Young
Researcher Award” of the IAROR in 2011. Joint founder and pre-
sident of the spin-off LIFT-Railway and Traffic Laboratory, he was
involved in several railway simulation studies in Italy and in other
countries in Europe, Africa, South America and Australia.

Prof. Dr. Chris Nash

Research Professor in the Institute for Transport Studies, University of
Leeds (ITS); Director of ITS for seven years. He works primarily on road
and rail transport economics and policy, including pricing, regulation
and project appraisal. He has been coordinator of six EU framework
programme research projects including UNITE, GRACE and Imprint-
Europe, and contributed to many more, including currently SUSTRAIL.
He has acted as specialist advisor to many organisations, including the
European Commission High Level Group on Infrastructure Charging,
OECD and the Transport and European Union Select Committees of
the British Parliament.
328
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 The Authors

Prof. Dr.-Ing. Nils Nießen

Studied Civil Engineering with a major of Transportation Planning
and Economic Geography at the RWTH Aachen University. Resear-
cher at the Institute of Transport Science, RWTH Aachen University;
2002 – 2007. PhD in 2008. Consultant at HaCon Ingenieurgesell-
schaft; 2007-2008. Founder and managing director of the company
VIA Consulting & Development GmbH, a spin-off of the Institute of
Transport Science specialising in consulting and software develop-
ment for railway operation; 2008-2013. Since 2013 Professor for
railway engineering and transport economics at the RWTH Aachen
University and Head of the Institute of Transport Science.

Prof. Dr. Dario Pacciarelli

Full Professor of Operations Research at Roma Tre University, De-
partment of Engineering, since 2012. He teaches “Operations Re-
search I” and “Logistics Optimisation”. His research interests are in
areas involving efficient utilisation of resources, computer algorithms
and discrete optimisation. He did work on routing and scheduling
theory. Applications work in the last 10 years mainly focused on real
time control of railway traffic. He is author or co-author of more than
80 publications in international journals, books and conference pro-
ceedings.

Prof. Dr.-Ing. Jörn Pachl

Study of Transportation Systems Engineering at Hochschule für
Verkehrswesen, Dresden; 1989 – 1991 researcher at Hochschule für
Verkehrswesen; 1991 – 1996 Managing position in the infrastructure
planning department of the Hamburg region of German Railways;
1993 external graduation to Doctor of Engineering at the Braun-
schweig Technical University; since 1996 Professor at the Braun-
schweig Technical University and Head of the Institute of Railway
Systems Engineering and Traffic Safety.

Priv.-Doz. Dr.-Ing. habil. Alfons Radtke

Graduation Diploma in Civil Engineering 1990, Doctorate in Civil
Engineering 1995, Post-doctoral lecture qualification (Habilitation)
for Railway and Transport 2005, all at University of Hanover, joint
founder and managing director of the software developing and con-
sulting companies IVE mbH and RMCon, specialising in consulting
and modelling of railway operation and software development for
timetable construction, railway simulation and optimisation prob-
lems around the world. Lecturer at the Institute of Transport, Rail-
way Construction and Operation (IVE) of the Technical University of
Braunschweig.
329
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
The Authors

Dr. Andrew Smith

Senior Lecturer in Transport Regulation and Economics and
Research Group Leader for the Economics and Discrete Choice
Research Group at the Institute for Transport Studies, University of
Leeds. He is a leading academic in the field of cost and efficien-
cy modelling and rail regulation. He has led a number of research
projects in these areas funded by UK research councils, the EU,
economic regulators and the Department for Transport. Andrew is
also academic advisor on econometric benchmarking issues for the
water regulator, OFWAT. He has a visiting research position at the
Centre for Transport Studies in Stockholm.

Dr. Robert Watson

Robert Watson studied for a BSc. in Transport Management and
Planning at Loughborough University (1977 – 1980) and then a Master
of Philosophy degree in Management Studies at Oxford University
(1980 – 1982) before embarking on a career in railway operations
and marketing. In 1998 he returned to Loughborough University as
Lecturer in Logistics and Rail Transport before setting up his own
consultancy which specialises in railway scheduling and in particu-
lar railway timetable simulation. He was awarded his PhD in railway
operational planning processes and systems in 2009 and continues
to teach and research alongside his consultancy work.

Dr. Phillip Wheat

Completed his doctoral degree at the Institute for Transport Studies
at the University of Leeds, where he is presently a Senior Research
Fellow. He is a regulatory economist and applied econometrician
whose primary area of application is in the rail and highways sec-
tors. His research is a mixture of applied econometric analysis, de-
velopment of econometric performance measurement techniques
and regulatory economics. He also teaches econometrics at mas-
ters level.

Dr. Jianxin Yuan

Received his BSc in mathematics from Tianjin University in 1988,
a MSc in hydraulic engineering from UNESCO-IHE in 2000, and
a PhD from Delft University of Technology in 2006. He was, from
2006 to 2008, a Post-Doc researcher at the Transport & Planning
Department, Delft University of Technology. Since October 2008, he
has worked by Arcadis Nederland as railway consultant. His main
research interests include railway capacity evaluation, modelling
and simulation of railway operations, signalling and traffic control,
applied probability, statistics, and operations research in the field of
railway transportation.
330
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt
 9 9
160 mm mm 21 mm mm 160 mm

The performance of many railway networks and the quality of

Hansen • Pachl
service offered is becoming more and more critical. The main
issues to be addressed are the increasing traffic volumes and making
the best use of the available capacity, at the same time resolving train
scheduling and management problems.

This is an updated, revised and extended edition of ‘Railway Timetable

& Traffic’, published in 2008. It describes the state-of-the-art methods
of railway timetabling and optimisation, capacity estimation, train Editors: Ingo Arne Hansen · Jörn Pachl
operations analysis and modelling, simulation, rescheduling and

Railway Timetabling

Railway Timetabling & Operations

performance assessment. The intention is to stimulate their broader
application in practice and to highlight current and future research

& Operations
areas. It is directed at academics, Masters and PhD students, as well
as professionals from the railway industry. It will also be of interest to
the public authorities that tender, monitor and perhaps fund railway
service provision. The overall aim is to improve the attractiveness and
efficiency of the train services which can be offered to the public. Analysis · Modelling · Optimisation
Simulation · Performance Evaluation
The key to achieving a higher efficiency and quality of train opera-

246 mm
246 mm

tions is an awareness of the impact of availability, reliability and

robustness of the subsystems on train processes. A deeper insight
into the probability of incidents and the propagation of train delays
depends on a thorough analysis of real world railway operations
and the feedback obtainable. This leads to an optimisation of the
timetable and a network-wide improvement in traffic management
performance.

This know-how should increase the efficiency of the railway system,

making it more attractive for regular, occasional and new customers,
and ensure that the railways continue to innovate. They will then be
able to make the maximum contribution possible to the transport
needs of the future.

iti en ed
ISBN 978-3-7771-0462-1

d
ed ext vis
on de
d re
an 2nd
Persönliches Exemplar für [email protected]
© PMC Media House GmbH. Weitergabe an Dritte urheberrechtlich untersagt