Transportation Research Part C 18 (2010) 807820

Contents lists available at ScienceDirect

Transportation Research Part C

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

Issues on railway wagon asset management using advanced

information systems
Athanasios Ballis *, Loukas Dimitriou
School of Civil Engineering, National Technical University of Athens, Athens, Greece

a r t i c l e

i n f o

Article history:
Received 15 December 2008
Received in revised form 2 June 2009
Accepted 2 September 2009

Keywords:
Railway empty wagon management
System surveillance
Estimated time of arrival
Optimal allocation
Mathematical simulation
Genetic Algorithms

a b s t r a c t
Wagon eet is an essential component of the railways business. Motivated by the recent
concern on freight wagon management in Europe and based on the experience from the
deployment and pilot demonstrations of an advanced wagon management system in Europe, the current work presents main technological and organizational issues associated with
a wagon pool, created and used by a railway operator partnership. The results include the
identication of the essential hardware and software requirements for the implementation
of such a system (user/wagon reservation, track-and-trace and energy-saver equipment,
loading/unloading sensors, wagon dispatching and delay management functionalities) as
well as the outcome of the investigation of signicant organizational aspects of the system
(imbalance between wagon offer and wagon demand, fair allocation of benets among the
users of the system, savings that can be achieved thanks to estimated time of arrival information, consideration of reliability metrics into the optimal empty wagons allocation
problem).
2009 Elsevier Ltd. All rights reserved.

1. Introduction
The revitalization of rail freight transport is among the key priorities of the European Transport Policy (EC, 2001a). To this
aim, various initiatives have been undertaken for the improvement of the various parameters that affect the performance of
the railway system, among them the management of the railway wagon eet. European wagon eet consists of Railway
owned wagons and privately owned wagons. Railway owned eets are owned and managed by the railway operators themselves. Wagons are subject to an agreement providing for common use. There is a presumption that an empty railway owned
wagon should be sent, preferably loaded, back towards its home railway. Furthermore, wagon pools have been set up for
railway owned wagons (e.g. the EUROP pool) where participating railways put in wagons which are then used in common.
Privately owned wagons are the property of their owners and are managed by their wagon owners. Today, some 180,000
privately owned railway wagons operate on Europes railway networks. They must be attached to the eet of a UIC (the
worldwide organisation of cooperation for railway companies) member railway if they need to run freely across other
UIC railways (EC, 2001b). A railway operator that receives a private wagon must send, after unloading, the wagon to its owner, as soon as possible.
The return to sender policy reduces the efciency of wagon redistribution while on the other, the common use of wagons has its own drawbacks: the RIV Regulation governs the reciprocal use of wagons, and also concerned with the use of
railway assets and so lays down arrangements for re-use of railway owned wagons and arrangements for payment for hire
of railway owned wagons by other railways. Nevertheless, the fact that the railway with possession of wagons (owned by
* Corresponding author. Tel.: +30 210 772 1235; fax: +30 210 772 2404.
E-mail addresses: abal@central.ntua.gr (A. Ballis), lucdimit@central.ntua.gr (L. Dimitriou).
0968-090X/$ - see front matter 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.trc.2009.09.003

808

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

another railway) has control meant that railways investing in wagons had no effective control over their assets. They will be
paid for the use of their wagon under the RIV charging system, but not for loss of opportunity (EC, 2001a).
Signicant effort is currently given to improve the above unsatisfactory situation, towards a wagon management schema
based on private sector nance criteria with bilateral agreements on fair prices being paid to the owner for the use of his
wagons by others. Synergies among railway operators have been identied as an opportunity to increase rail freight productivity. To this aim, the European Commission (EC) funded a number of research works among them the WagonLink, CroBIT
and F-MAN projects in order to investigate the deciencies of the wagon management system and propose solutions for its
advancement. WagonLink developed a wagon reservation system which offers an interactive map of wagon searching which
should be completed within one hour. This way, the participating railroad companies have the possibility to communicate
their wagon demands and offers. The Cross Border Information Technology (CroBIT) project introduced a data exchange platform along European rail corridors (door-to-door approach) that will offer tracking and tracing of goods as well as automatic
notication in case of deviations from the schedule (EC/DG Research, 2003b).
The current work is focusing on the research work and the ndings of the F-MAN project that investigated the potential
and requirements of a system offering real-time track-and-trace and wagon management functionalities (see Section 2). The
project concluded in the technical and organizational requirements of such a system and deployed a prototype that has been
tested by French, Portuguese and Slovenian Railways. In addition, model-based mathematical analysis (see Section 3) has
been used to allow for the understanding of the wagon management system potential and risk. This analysis has been initiated within F-MAN workload and complemented through the academic work of the authors. The conclusions are presented
in the nal section.

2. Organizational and technical requirements for the management of wagon eets

A typical rail transport action involves many actors: In the beginning of the chain there is the customer (sender, consignee, shipper or forwarder) that cooperates with the Order Manager providing him all relevant information for the desired
transport of goods (e.g. nature of goods, origin and destination of goods, date and place of loading and departure, freight type
and quantity or wagon type and quantity, etc.). The Order Manager is responsible to translate the customer request in a precise and coded order to the Fleet Manager as well as to inform his customer in case of problems during transport, to check
the time that wagons are kept by the consignee, to do a repartition of charges and ensure the transport in cooperation with
the Fleet Manager. The Fleet Manager receives the customers order and proceeds with the required actions for the coordination of the specic transport. His main responsibility is to optimize wagons productivity, thus to satisfy the demand of
empty wagons by balancing the wagon resources on the network based on information concerning empty wagon locations
(idle or foreseen idle at a specic date), timetables, train delays, trains rerouting, etc. Part of this information is obtained from
the Operational Manager (who is responsible to operate a transport service on the railway and to guarantee the transport of
wagon) as well as from the Train Operating Company that in cooperation with the Infrastructure Manager allocates timetable slots, locomotives, drivers and schedules the inspection of wagons before their departure (EC, 2002).
Wagons are provided by railways or private wagon operators upon customer request, unless the customer owns his own
wagons. They may be supplied for single journeys, or leased to the customer for dened periods of time. Wagon management
is performed by authorized railways personnel. In its everyday operations the Fleet Manager has to exchange information
with the various actors involved in the realization of a transport order, negotiate bookings and perform last-minute changes
by making phone calls, sending and receiving faxes and to allocate the idle wagons to the required transport requests according to their attributes (wagon type in relation to cargo characteristics, availability, mileage, time till maintenance, cost, etc.)
based on his experience and, if available, on the support of custom made software applications. Fleet Management is further
complicated by the fact that there are many spatially separated and independent Fleet Managers each one focusing on their
own needs and responsibilities. As a result, a Fleet Manager that has sent a number of his wagons in international trips is
anxiously waiting for their return (to be used in other transport requests) while these wagons may remain idle at the disposal of other Fleet Managers.
At present, no integrated wagon eet management exists in Europe. Tracking is generally based on data provided by train
control management, while the distribution of empty wagons is carried out by dispatchers relying on past experience rather
than real time data. This works relatively well at the national level since it is based on established procedures between national operators and infrastructure managers to assure empty wagon supply. As soon as wagons cross borders, however, the
number of interfaces increases, leading to poor wagon utilization, especially when the players speak different languages
(Kuhla et al., 2003). This is quite typical in European transport as, for example, a rail trip from Holland/Netherland to Greece
must cross Germany, Austria, Hungary, Romania and Bulgaria or alternatively Germany, Austria, Slovenia, Croatia, Bosnia
Hercegovina, Serbia and F.Y.R.O. Macedonia.
In order to cope with the above deciencies the F-MAN project investigated the requirements of a system that offers realtime track-and-trace and wagon management functionalities for a wagon pool accessible by its registered users (EC, 2003a).
The research concluded that the essential components of such a system are a wagon reservation scheme, on-board equipment, a prediction mechanism for the time of the wagon arrival at destination station and a suitable asset management software. These components are presented in Fig. 1 and analytically discussed in the following sections.

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

809

Track and trace device

Cumulative probability of wagon arrival

Cumulative
ofterminal
wagon arrival
times at theprobability
destination
times at the destination terminal

Loading/unloading sensor

Time margin
Time margin
Estimated Time
functionality

of

Arrival

System users (Fleet managers, wagon leasing companies)

Fig. 1. Overview of a eet wagon management system.

2.1. Wagon reservation

A wagon reservation system can be used either by one wagon owner/lessee or by a group of wagon owners/lessees operating under a partnership agreement. The system developed within the F-MAN project follows the latter (common wagon
pool) option: wagon owners/lessees and Fleet Managers can use the system to offer wagons to the pool, or to search of available idle wagons. A Fleet Manager can be a wagon owner and lessee at the same time, offering idle wagons of Type A while
seeking wagons of Type B that are available for loading. All participants have to be registered, and must accept the contractual conditions. Wagons in the pool should be identied, and cannot be used outside the pool without being withdrawn under the proper procedure.
The software that implements the reservation scheme should handle the interface with users, maintains the systems
database and assures security through passwords valid at various levels (access to specic functions is restricted to authorized users). At the top of the access pyramid are the Administrator of the system and one Master-User for each company.
These Master-Users can assign access rights to other staff in their company (e.g. allowing them to view only track-and-trace
information or to view but not modify wagon leasing agreements, etc.).

810

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

2.2. On-board devices for track-and-trace, communication and loading status identication
Track-and-trace functionality is a modern monitoring tool that can be achieved by various technical approaches: (a) by a
ground-based network of transponders installed along the railway lines that read electronic tags mounted on wagons, (b)
by GPS tracking of the locomotive position, in combination with information systems that know which wagons have been
assigned to this locomotive and (c) by GPS and communication devices mounted in each wagon. This (later) approach has
been selected by F-MAN due to its autonomy and exibility. The deployment of the system included an on-board terminal
device (OTB) that integrate a satellite positioning receiver, a GSM cellular radio communication terminal (that sent and received SMS messages), a microcontroller (responsible for communication and data acquisition) and a power management
utility.
Power management is a critical element of the system as (in contradiction with the truck where the battery of the car is
used to power the track-and-trace device) the freight wagon does not have electrical supply (brakes are operate by a pneumatic system). Furthermore, a systematic battery charging/replacement process is not feasible, except from the time that
wagon comes for maintenance. This imposes a requirement for a 6 year life time that can be coped only by an efcient power
management scheme: communication intervals and protocols restrict energy consumption by limiting transmission of data
to the minimum necessary: the wagon timetable is stored in the memory of the OBT memory in a form of waypoint (e.g. railway stations, border crossings, etc.) and associated times (check-times). Most of the time the system is in sleep mode and
wakes at every check-time to compare the coordinates of its waypoint with the actual wagon coordinates. In case of deviation, an alarm-message is send via SMS to the systems headquarters. The in-memory timetable can be revised remotely to
adapt to new situations like for example in case that the train has been rescheduled due to a long delay or in case that the
wagon had missed a connection and been sent forward on a later train. The pilot demonstration of the system has been performed by wagons belonging to Slovenian, French and Portuguese railways and revealed that: (a) in case that the train itineraries include a signicant number of intermediate stations, the battery life is reduced below the expected time duration
and (b) a number of messages are reaching their destination with unacceptable delay (a sample of 2850 messages was exchanged during the pilot run of the system). These problems have been tackled (in later versions of the OBT device) thanks to
improvements of the battery technology and to software modications that conrm the recipient and resend the message if
required.
The tracking of wagon location is not the only information required for wagon management. Many sensor types that can
monitor the status of the wagon or the transported cargoes (e.g. temperature, shocks and door-opening) can be mounted in a
wagon to send the status/alarm conditions through the OBT device. Among them, the need for a loading sensor that indicate
when the wagon is loaded/unloaded has been appraised (by railway experts) to be very signicant for the wagon management process. Within the pilot demonstration of F-MAN, a loading sensor that calculates the wagon loading/unloading status
according to the distance of the wagon oor from the head of the railways has been successfully tested.
2.3. Estimated time of arrival
Freight trains in Europe face delays mainly because passenger trains get priority (EC, 2002). An initial hold-up delay often
leads to a build-up of delays once the allocated timetable path has been lost. The knowledge of estimated time of arrival
(ETA) of trains/wagons to their destination is valuable information as it may trigger proactive actions to cope with problems
associated to the delayed train/wagons (e.g. to inform the forwarder to sent the cargo by truck). A system that offers information about ETA and wagon availability should be based on a comprehensive organizational structure, which will guarantee the monitoring of trains and the associated wagons (preferable also by automatic transfer of information among the
information systems of the various railways involved in the transport process). This is among the targets set by the railway
Organizations in Europe, but much effort should be devoted in order to archive such a global system. In the mid time, Internet-based information systems belonging to individual companies/Organizations and/or satellite-based GPS applications, are
foreseen as the means for information exchange. It must be noted that the satellite-based tracking will be used as a tool for
additional precaution for specic cargoes, even in the case that the railway community will establish a communication
channel.
A track-and-trace system can provide the train/wagon location and the deviation from schedule but such information are
meaningful only when they can be converted in estimations about the train/wagon arrival time in the destination (ETA) or in
certain cases in borders (ETB, Estimated Time in Borders). A reliable ETA at the destination station is essential for effective
wagon deployment. In F-MAN system, the ETA calculation has been based on statistical records of the past experience of
trains running in the same itinerary that are stored in a form of probability functions in the central computer hosted in
the systems headquarters. When a deviation Di from the timetable at a waypoint i is reported (through SMS communication)
the conditional density function f(Ddest|Di) where Ddest is the delay at destination station, is retrieved from the systems database. The associated distribution function F(x) provides the probability P dest Ddest 6 Mdest that the wagon delay will not exceed Mdest which is the time margin between the scheduled wagon arrival at destination and the next scheduled transport
activity of the wagon. If Pdest is below a threshold set by the wagon dispatchers, then proactive actions should initiated (e.g.
search for another wagon to substitute the delayed one in its next transport activity, organize transport by road, etc.).
The validation of the above method was performed using data, provided by the Portuguese Railways, for the SpainPortugal corridor that passes from the train stations of Alcantara, Lisboa, Entrocamento, Pamhilhosa, Gaia and Leixoes. The ETA

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

811

predictions were performed at the Entrocamento station and concerned the train departure time at the station of Leixoes.
The data set included 259 records and concerned the period from January 2002 to March 2004. Part of these data was used
for the calibration of ETA functions (formation of statistical records) while the rest was used for test cases. The 85% of the ETA
predictions had been proved accurate (EC, 2004). Nevertheless, the above test cannot be considered as a proof for the reliability of the method and more tests are required, in various parts of the European network in order to have an overall conclusion for its applicability. A similar study performed in Greece for the freight trains running in the ThessalonikiAthens
itinerary, concluded in similar success rates. Another test is foreseen for 2008 within the framework of a research project
that investigates the operating conditions of rail services in the Netherlands Greece/Turkey corridors. An extensive presentation of the above ETA approach as well as the analysis of the impact of ETA information in the management of the wagon
eets exceeds the scope of the current work (see Ballis, 2005).

2.4. Asset management

The term asset management is quite generic and can include various decision support tools and/or automatic processes.
The F-MAN system included the following functionalities: bid and offer, delay management and wagon dispatching.
In bid and offer functionality, the wagon offer is performed during the wagon registration in the system and includes
price level and the use terms (for example, the wagon must not be used outside central European countries). The wagon
bid concerns the selection of the appropriate wagons taking into account their technical characteristics, price and constraints. This module should also provide statistical information to allow pool members to dene their pricing strategy.
The wagon delay management functionality is required when ETA information indicates a delay, and the relevant Fleet
Manager should seek (through an appropriate user interface) for available wagons or communicate with other actors for
support.
The wagon dispatching functionality allows for the scheduling of wagon movements (for the consecutive transport
phases from the idle wagon location, to the station where needed, then to its nal destination and back) as well as for
the dispatching of idle units to wagon demand either in a simple form of decision support (that depicts wagon demand
and offer in clusters according to their characteristics and allows for sorting, ltering, etc. operations) or in an automatic
mode that suggests the wagon allocation solution. Given that the concept of the common wagon pool that operates under
private sector nancial criteria is new, there is lack of experience concerning the dispatching of idle wagons to transport demand. In F-MAN project, the automatic mode for wagon dispatching was left for a future software version and instead mathematical modeling was used to increase the understanding of the systems performance and behavior. This work has been
continued by the authors in an attempt to dene mathematical formulation for improved wagon management tactics, utilizing methods of stochastic and chance constraint programming. The relevant investigation is presented in the following
section.

3. Network optimization formulations for the management of empty wagons

A thorough survey on optimization models for train operations has been offered by Cordeau et al. (1998). In particular,
wagon eet management concerns a large scale optimization problem which even in its static case (xed wagon demand
in a single study period) that can be addressed with standard mathematical programming, exhibits difculties. The problem
gets even more complicated when considering that:
 The production and demand for empty wagons is a dynamic process with under-dened or stochastic attributes.
 The performance gains that can be achieved by a common wagon pool (as compared with the performance of individual
wagon eets) must be fairly allocated among the participating railway operators.
 Updated information (e.g. through the ETA information system) can be used in the wagon dispatching process.
 Statistical information on wagon delays can be used to improve the reliability of the system performance (reliable optimal
empty wagon allocation problem).
In the following sections a more detailed description of the problem and requirements of the formation of a realistic system for the optimal allocation of empty wagons are outlined.
3.1. The static case of wagon management
In its static form (this with known wagon supply and demand) the problem of the allocation of empty wagons from production points (locations that wagons are unloaded or empty wagons depots), to demand points (production areas) in order
to minimize the total cost can be solved as a linear programming problem (classically termed as the Transportation Problem)
in which assuming that xij are the number of empty wagons from i to j and cij is the associated unitary transportation cost,
then the objective is the minimization of the total transportation cost, subject to demand, supply and physical constraints
(Hillier and Lieberman, 2002):

812

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

minimize
subject to

Z0
X

cij xij

ij

xij ai

for all i 2 I;

xij bj

for all j 2 J;

xij P 0 for all i 2 I and j 2 J;

where ai stands for the supply availability in empty wagons in point i, bj for the demand in empty wagons at point j forming a
constraints set that provides operational and physical meaning.
The problem of distributing rail freight cars from unloading to loading points has often been formulated as a linear programming transportation problem with minimum travel distance, time or cost objective, as presented, e.g. by White and
Bomberault (1969), Mixa (1972), Kim and Kinkle (1982), Mendiratta and Turnquist (1982), Glickman and Sherali (1985), Cordeau et al. (1998) and Lim et al. (2006). As can be identied in the relevant studies, the standard Transportation Problem
formulation has been used as benchmark and starting point when considering an advanced model that can provide optimal
treatment of the equipment. Nevertheless, depending on the operating conditions, many additional considerations/modications or even reformulations may be required to provide a pragmatic solution. Some of the most important considerations
for employing the above formulation to realistic situations will be drawn in the following sections.
3.2. Imbalance between wagon demand and supply forecasting inefciencies
P
P
As Eqs. (2) and (3) imply, i ai j bj means that the Transportation Problem provides solutions when the supply is equal
to demand. If not, the introduction of dummy demand nodes or dummy supply nodes can solve the mathematical part of
the problem, but in practice that means that either some transport requirements will be not satised, or that some wagons
will not be used and therefore the associated owners will not be paid. A description of a relevant problem (termed as transportation problem with minimum quantity commitment) is offered by Lim et al. (2006) leading to a mixed-integer programming formulation, while a branch-and-cut heuristic algorithm is proposed for addressing it.
This problematic situation can be smoothed by business rules that can maintain the fair treatment of similar users inside
the pool. To this aim, the system should keep records for the non-satised demands in order to compensate in a similar future case. According to F-MAN project, business rules are also required for gain allocation (see the following section), for protecting regular users versus occasional users of the pool without eliminating the exibility for occasional participation as
well as for rewarding consistent and reliable users against other users causing frequent wagon delivery delays (Kuhla
et al., 2003).
Kikuchi (1985) identied additional drawbacks in applying the standard techniques of linear programming to realistic
situations and provided a framework for the allocation of empty wagons in a common pool environment. Under the proposed framework it is possible to handle the demand forecasting inefciencies and the additional imposed terminal handling
costs, on a day-to-day basis. It should be noted that the problem of forecasting empty ows has been identied as an issue of
increased complexity (Holguin-Veras and Thorson, 2003) making the optimal programming of the dynamic optimal empty
wagons allocation a demanding task.
Also, Powell and Carvalho (1998) presented a dynamic (real-time) framework for optimization of the atcars able to be
used for addressing forecasting inefciencies. Such suitable problem formulations can be addressed by adopting mixed-integer programming methods. It should be also noticed that since in real-time implementations time-consuming optimization
algorithms maybe not operationally suitable, a reasonable extension could be the introduction of algorithms able to be parallelized (utilizing technologies of distributed computing) leading to procedures that can be implemented in realistic
situations.
3.3. Fair allocation of gains
The incentive of an individual railway operator to join a common wagon pool is the anticipated gains from doing so. Such
gains can be obtained by the optimal resource allocation achieved by utilizing information technologies (e.g. the above described track-and-trace equipment and ETA-based information platforms) in combination with network models and Operations Research techniques. While for the case of a single operator the Transportation Problem formulation is able to provide
optimal solutions for the empty wagon management, that does not hold for the case of management of common-pooled
empty wagons since although the above formulation is able to provide an overall optimal solution (system optimum) it
may be occur that some members of the common pool agreement may be gaining unreasonably low benets from participating to the common pool, or even to perform worse than the no-participating scenario.
Glickman and Sherali (1985) assembled a set of data from actual operating records to analyzed/compare pooled empty
freight cars. In an indicative example, the overall cost of the systems has been reduced by 31.5% but 3 out of the 15 operators
involved in the wagon pool resulted is up to 68% losses in relation to their cost when performed individually. The authors
proposed two formulations/criteria aiming to improve the distribution of gains of the pool among the participants: The rst
criterion was to maximize the minimum (maximin) relative saving of all the railroads, and the other was to minimize the

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

813

mean absolute deviation (MAD) sum of the relative savings. Both formulations were solvable by techniques of linear programming accompanied with a heuristic rounding procedure for providing the eligible integer number of transferred empty
wagons.
Various alternative formulations have been presented in the literature in order to provide acceptable optimal solution by
all participants to the common pool. The complexity of the optimal wagons handling, under multi-participants schemes is
further described in Adamidou et al. (1993), where the optimal allocation of empty wagons of time-varying demand proles
is modeled as a non-cooperative game among the railroad companies while the performance of the model is tested against
realistic scenarios.
All the above formulations used for addressing the optimum common-pooled empty wagons, are utilizing conventional
algorithms of linear programming (like the simplex method) which typically converge to the global optimum. Since in realistic implementations of optimum strategies in large scale, computation time matters, other algorithms for addressing linear
programming problems with improved computational time can be used, like the class of various algorithms of the Interior
Point Method (e.g. the projection, afne-scaling and primaldual methods).
Optimal allocation of common-pooled empty wagons is, actually, a mixed-integer problem and its transformation to a
linear one can be viewed as a relaxation of the original formulation. By relaxing the problem to a continuous linear one
can be more easily addressed, since linear programming algorithms are faster and able to provide global optimum solutions.
Although the linear programming algorithms are accompanied with rounding procedures in order to provide integer solutions these rounding procedures can alter the solution so much, resulting to a setup that could be far from optimality (Glickman and Sherali, 1985). So methods for addressing the linear integer programming formulation of the problem can be
implemented based on decomposition tactics, like the Branch-and-Bound and the method of Lagrangian Relaxation. It should
be noted that the results provided by rounding procedures may be signicantly departed from optimality. That is why the
employment of algorithms able to address the mixed-integer form of the problem is preferred by several researchers of the
eld (for a comprehensive review see Dejax and Crainic, 1987).
Nevertheless, another angle of view to this problem is to decouple the issue of the optimization of the physical system
from the prot/losses allocation and implement an accounting system that will take into account real costs and incomes
and performed the fair allocation of net prots (the analysis of such a system is outside the scope of the current work).
3.4. The use of ETA information
According to the current railway practice, wagon availability (supply) is dened by the number of idle wagons existing in
the rail terminal. Reliable ETA information for the forthcoming wagons is considered as of great value for the Fleet Management (the term ETA-wagons will be used hereinafter for this wagon category). Nevertheless, the answer to the question
Why? is not obvious. As a general rule, additional information means better results yet, the way that this information should
be used and the exact additional benets is not always known. A Fleet Manager for example that has 100 wagon demands
and only 80 idle wagons in his station, can use ETA information to identify 20 additional wagons (e.g. to be arrive in the next
hours). Nevertheless, these wagons will be missing from the plan of the next day. So the prot of the rst day becomes a loss
for the next one. In order to have a positive outcome, the following conditions/strategies should be met:
(a) Wagon offer and wagon demand patterns are uctuating. In such a case, there are days when the offer of wagons at a
station is higher than the wagon demand, thus ETA information can be used to utilize the excess wagon availability.
(b) An appropriate wagon dispatching strategy is used. Such a strategy (that can contribute positively even in case that
the wagon offer is lower than the wagon demand) is as follows: if W is the number of idle wagons in the station
and Z the total wagons that can be used thanks to ETA information (Z = {existing idle} + {ETA-wagons}), then select
the best W wagons out of the population of the Z wagons. The best is interpreted as the ones that optimize the
total wagon-km savings.
In order to obtain an order of magnitude for the above wagon-km savings, a simulation model (incorporating wagon supply and demand generators as well as a solver of the classic Transportation Problem, see Eqs. (1)(4)) was developed and a
hypothetical case was investigated. The input variables and the assumptions are:
s The geographical area is a 2000 km  2000 km square.
s Within the geographical area, four stations (demand point) asking for wagons (wagon requests) exist. The coordinates
of these stations are set by the program user. Two cases with the stations symmetrically and asymmetrically distributed in the area are presented in Fig. 2.
s Within the geographical area, 25 stations with idle wagons (wagon offers) exist. The coordinates of these stations are
generated automatically through random numbers (following uniform distributions), thus these stations are randomly
dispersed. The term supply point stands for these stations.
s The mileage between supply and demand points is assumed to be the Euclidean distance between the associated
stations.
s The total number of wagon requests is set by the program user. These wagon requests are equally allocated to the
demand points.

814

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

Fig. 2. Comparison of wagon-km savings resulting from ETA information for two spatial allocation cases of the demand-stations.

s The total number of wagon offers is set by the program user. The wagon offers are allocated to the supply points
according to the following technique: A (user dened) Pi percentage is equally distributed among supply points, while
the remaining (1  Pi) percentage is allocated randomly. In the scenarios presented, in Fig. 2, a Pi = 50 had been
considered.
s Each wagon request generates a transport activity (if adequate supply points exist). Wagon requests that are satised
by wagon offers in the same station are not considered.
Fig. 2 presents the results of more than 50 simulation runs, each one containing 14 replications (each replication includes
30 days). According to the simulation results a 20% percentage of additional wagons offered to the system due to ETA information results, is about 15% savings in the total wagon-km. Similarly, a 40% percentage results in 22% savings. These gains
have been increased when the demand points/stations were asymmetrically distributed. This can be explained by the fact
that in this case the average transport distance is increased and thus the savings can be greater.

3.5. Travel time uncertainties alternative formulations of the stochastic optimal empty wagon allocation
In real-world situations one of the most signicant problems dealing with the optimal allocation of empty wagons, is related (as previously mentioned) to the uncertainty pertaining railway systems. In particular, since the railroad companies are
committed on the delivery time (sometimes even by contract), the estimation of the risk of these commitments is of particular interest, related to the concept of value-at-risk for providing transportation services. The implementation of track-and-

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

815

trace and ETA functionalities in the wagon management may also provide useful statistics that can use to smooth the effects
of travel time uncertainties. For example the role of ETA information (presented in Section 2) is essential for estimating the
statistical properties of network service time, an element that is crucial for the whole system performance. The more reliable
the estimation of service time is, the more the system performance is expected to be improved. The effect of this element will
be discussed hereinafter.
In order to take into consideration the uncertainty of service times to the problem of optimal allocation of the empty wagons, two formulations are presented in this section. Uncertainty and risk are elements closely related to management of reallife operations. In order to take into account here the inuence of uncertainty on optimal management of the empty wagon
resources, two formulations will be presented. The rst one is aiming to introduce the systems uncertainty to the total service cost, quantifying its negative effect to the system performance. Following the standard formulation of the transportation
problem presented earlier, this is made by introducing an alteration on the objective function, such as:

minimize
subject to

Z0
X


X
t
cij xij cijij xij Ptij P T ij

ij

xij 6 ai

for all i 2 I;

xij bj

for all j 2 J;

X
i

xij P 0 for all i 2 I and j 2 J;

where cijij is the unitary (marginal) delay cost for late deliverance of empty wagons and Ptij P T stands for the probability
that the actual train/wagon arrival time at destination tij will be greater than a predened threshold Tij (e.g. Tij could stand for
the service time of the second quickest service path). This particular formulation that introduces an additional cost for delivery inefciencies (delays) by imposing a penalty in the pertained uncertainty, is regarded (and termed) as soft constraint of
the problem. Under this formulation the economic effect of the uncertainty can by estimated and taken into consideration to
the systems performance. It should be noted that the ETA functionalities are providing all the necessary information regarding the estimation of Pt ij P T.
An alternative approach would be the consideration of some more strict rules for describing optimality conditions. In particular, since the railways operation is composed of a system of inter-related sub-functions, a failure to one stage of operation (sub-function) may inuence the system performance as a whole. Focusing in the uncertainty of service times,
inefciencies in keeping up with tight time schedules could lead to major system malfunctions. In order to overcome such
circumstances, a mesh of operational rules (ordinance) could be introduced, constraining the negative effect of the service
time uncertainty to the system performance. In particular, by considering the travel time tij (instead of the travel cost cij) a
formulation of the following form could be adopted:

minimize
subject to

Z0
X

8 i 2 I and j 2 J

tij xij

ij

xij 6 ai ;

8 i 2 I;

10

xij bj ;

8 j 2 J;

11

8 i 2 I and j 2 J;
8 i 2 I and j 2 J;

Pt ij P T ij 6 Z ij ;

12

xij P 0;

13

adding a tolerance Zij for the probability of the travel time to be greater than a threshold Tij. This formulation enables the
estimation of the optimal resource of the empty wagons allocation, not only in terms of cost minimization but by selecting
also an accepted risk rate to achieve it. In this case, the results of the uncertainty have been taking into consideration, hard
constraining its effect and introducing metrics of reliability into the solution plan of the optimal empty wagons allocation
problem.
The later two formulations of the optimal empty wagon allocation problem (Eqs. (5)(13)) complicates the solution procedure, since now this problem belongs to the non-linear stochastic integer programming class of problems, that typically
requires more complex solution procedures. For simplied cases optimal solution can be obtained by inspection or by estimating analytically the reliability requirements and then introducing them into the standard formulation. In cases where the
size of the network is increased or the stochastic variables are correlated, then no analytical solutions can be obtained. Also,
the use of exact methods (like the Branch-and-Bound) is not feasible here too, since their search tactics are depended on
information of deterministic nature. As so, an extension is proposed here related to the hybridization of optimization methods in order to take into account the stochasticities involved in the problem variables. In particular the optimal solution of
such problems is aimed to be obtained here by hybrid heuristic algorithms, able to internally treat stochastic variables within the optimization process. Heuristics are considered as a valid optimization framework, since their search tactics are based
only on the information gained on repeatedly (but appropriately selected) executed alternative trials of plausible solutions,
aiming in (stochastically) approaching optimality when, for a predened number of trials, no improvement is gained,

816

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

disregarding the processes that each trial involves. Such paradigms of hybridizing optimization procedures come up with the
combination of Mathematical Simulation with standard Genetic Algorithms (GAs). A framework of a hybrid GA able to tackle
problems of that class will be described in the following section.
3.5.1. Solution procedure for the stochastic optimal empty wagons allocation
GAs are nature-inspired stochastic approximation procedures, utilizing an initial population of feasible solutions and
mechanics of genetic evolution in order to approach optimality (Holland, 1975). Since textbooks concerning GAs are
widely available (Goldberg, 1989; Mitchell, 1998) a parsimonious description of their fundamental attributes will be presented here. So, a set of candidate solutions (individuals) is initially selected and coded in strings (chromosomes) in order
to comply with the problem at hand (Initialization). The evolution of a typical GA is based on the performance (tness) of
the objective function for each individual of the population. After all individuals of the population have been tested, the
most powerful are selected forming an intermediate (parent) population (Selection Operation). Then an exchange of parts
between randomly coupled individuals from the parent population is performed resulted in a new, genetically improved
population (Crossover Operation). Finally, random parts of the new individuals strings are altered in order to avoid primitive convergence (Mutation Operation) forming a new candidate population (Generation). The process iterates until
convergence.
As observed from this short description of the GAs mechanics, one of their cardinal characteristic is that the convergence
process relies on information only about the performance of the objective function in the current generation, making the
process gradient-free. Also GAs are able to treat the various constraints through their coding. These characteristics are making the GAs attractive for tackling problems of increased complexity as that of non-linear integer stochastic programming
like the present one.
In order to estimate the stochastic properties of the travel time and to introduce their risk attributes within the solution
process, GAs can be hybridized with a simulation procedure. In particular, the estimation of the risk can be made with the
use of random sampling procedures of numbers from pre-specied statistical distributions, which in turn have been estimated from historical observations. In the current study a Latin Hypercube sampling procedure has been selected. Adopting
this sampling method, both correlated and uncorrelated random numbers can be produced based on Cholesky decomposition technique of the correlation matrix (Iman and Conover, 1982). It should be recalled, that the problem which is aimed to
be addressed here by the proposed optimization framework, concerns the incorporation of the travel time stochasticity into
the problem solution. It is essential to be noted that in real-world situations, travel time disturbances is frequently correlated
between consecutive (or neighboring) network links, since the uctuations may be related to similar causing forces (adverse
weather conditions, network incidents, etc.). The proposed simulation method used here, can facilitate the incorporation of
such assumptions into the modeling framework.
The hybrid schema adopted here for solving the current problem is presented in gure (Fig. 3). In the following section
results of the proposed framework application to an articial network, able to provide an illustrative example, will be
presented.

Initialization
(Production of an initial population of
feasible solutions)

Performance evaluation for each

populations individual

NO
Convergence?

Mutation

Crossover

Selection

YES
Optimal empty wagons management
plan
Fig. 3. Flowchart of Hybrid Genetic Algorithm.

Genetic Evolution of Solutions

For each of the populations

individual, perform risk assessment
by means of Mathematical Simulation

817

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

3.5.2. Computational results

The above described framework has been applied to an appropriately composed network (Fig. 4) able to expose the properties of the solution procedure, while keeping all characteristics of a typical railway system at the same time. The network is
composed of three major terminals (A, B and C) where the empty wagons are stored (empty wagons supply nodes) forming a
corridor that serves the demand nodes for empty wagons (nodes 111 bracketed in squares), resulting to a network of 14
nodes and 19 links. As cost function has been selected the travel time, indicated for all network links.
For this example, the demand is selected to represent the total number of empty wagons that will be requested by each
node for the period considered. In order the example to be more demonstrative it is selected that the demand for each node
is homogenous, so a number of 1000 wagons/period has been considered. Also, each of the A, B and C node is considered to
have available 7000 wagons/period, so the availability of empty wagons in each terminal node probably fully covers the
demand.
At rst, the problem of allocation of empty wagons from the three terminals to the demand nodes in its standard form has
been solved by using the formulation presented in Eqs. (1)(4). The algorithm used is the interior point algorithm presented
in Sanchez et al. (2005).
The results are presented in Table 1, and are representing the global optimum solution (system optimum) for allocating
the empty wagons. In terms of computational effort, the algorithm is able to converge to global optimum in fractions of a
second in a PC of modest computational resources.
By using the above empty wagons allocation plan as benchmark, an optimal solution plan will be investigated by assuming that the travel time of certain links are not xed, but they exhibit a rather stochastic nature. In particular, these links are:
(A-2), (A-3), (B-6) and (B-7), indicated in dashed lines at Fig. 2.
In the current study the travel time of those links has been considered that follows a log-normal distribution. Lognormal
distribution is frequently chosen for reliability analysis purposes, since skewed distributions are shown to adequately describe stochastic variables that can be considered as a composition of events following normal distributions, having large
variances (compared with its means) and non-negative values (Limpert et al., 2001; Martn and Prez, 2009). As so, log-normal distribution is selected here for modeling links travel time in the particular test railway network.

3.5
1.5

2
4.

4.5

7.6

4.2

6.7

6.8

4.1

6.9

3.5

11

7
1.1

3.4

5.
4

3 0.9

1.4

8
4
3.

10
Fig. 4. Network conguration.

Table 1
Optimum empty wagons allocation plan for xed travel times.
Origin

Destination
1

10

11

A (wagons)
B (wagons)
C (wagons)

1000

1000

1000

1000

1000

1000

1000

1000

1000

1000

1000

Travel cost (h)

7500

4000

3000

4200

3900

4500

3500

3400

4100

3400

8800

Total cost (h)

50,300

818

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

The properties of log-normal distribution for each stochastic link is taken to be as Log-N(E(LTT), 1.7E(LTT)), where E(LTT)
stands for the expected value of the link travel time (LTT) and 1.7E(LTT) for its variance. The probability distribution of each
link travel time is estimated by a Latin Hypercube sampling procedure (200 samples).
It is noted that estimating the properties of the distribution of the travel time in cases where service paths are composed
of multiple links that they exhibit uctuations in travel time from the log-normal distribution, no exact analytical formulation is available (El Faouzi and Maurin, 2007). The results of the simulation of travel times are presented in Fig. 5. As far as
the setup of the GA used here, a population of 30 binary coded individuals (candidate solutions), two genetic operations, i.e.
crossover and mutation with typically used rates of 75% and 5% for each operation, respectively, has been adopted.
Then, following the formulation expressed in Eqs. (9)(13), the probability of the service time to remain optimal compared with the next best (shortest) service time is constrained by a pre-specied reliability level. In particular by setting
a reliability requirement for a node service time to be less than the service time of the next best alternative for the 80%
of the cases, the constraint expressed in inequality (12) is written as:

Pt ij P t 0ij 6 0:20;

8i 2 I and j 2 J;

14

where t ij stands for the shortest travel time, while t0ij for the second best travel time and Zij = 0.2.
By adopting this constraint the solution obtained now assigns the service of nodes 6 and 7 to terminal C, since it is found
that the probability of travel time from terminal B (4.5 h for node 6 and 3.5 h for node 7) although better, exceeds those from
node C (5.5 for node 6 and 4.5 h for node 7) for a probability of nearly 22% and 21%, respectively, resulting to a system total
travel time equal to 52,300 h. When making the travel time reliability levels stricter (i.e. dropping the Zij value) more nodes
are serviced by alternative routes of higher travel time in favor of reliability gains. For example by setting Zij = 0.15 node 2 is
serviced by terminal B (instead of A) since the probability of the travel time from terminal A to exceed that from terminal B is
almost 18%, resulting to 54,000 h of total system travel time. Finally by setting Zij = 0.10, node 3 is now served by terminal B
(instead of A). The diagram of reliability requirements (percentage of service time to be better compared with that of the
next best alternative route) and total system travel time is presented at Fig. 6, showing the system losses (in terms of total
travel time or cost) in order reliability gains to be obtained. In particular and for the example network used here, by increasing the reliability requirements from 80% to 90% (12.5% increment on the base value) the resulted system travel time increases by almost 8% (from 52,300 to 56,300).
From the computational point of view the current framework for the estimation of the optimal empty wagons allocation
can be considered as computational intensive, since the effort for combining the Latin Hypercube simulation procedure with
the GA requires increased computational recourses. Although the convergence of the algorithm can be obtained within 50
generations, the convergence time for the small network shown in Fig. 3, is about 1.5 CPU hours on a modest 2.5 GHz PC.

Distribution for E(LTT)=4.5

0.15

X <=0.33
5%

X <=15.49
95%

Mean = 4,541741

0.10
0.05
0.00
0

X <=0.29
5%

0.25

Probability

Probability

0.20

Distribution for E(LTT)=4.0

0.30

0.25

X <=13.44
95%

0.20
Mean = 3,936688

0.15
0.10
0.05

10

20

30

40

50

60

0.00

70

10

20

Travel Time (Hrs)

Distribution for E(LTT)=3.5

0.25

X <=0.26
5%

40

50

60

70

60

70

Distribution for E(LTT)=3.0

0.40

X <=12.03
95%

X <=0.22 X <=10.23
95%
5%

0.35

0.20

0.30

Probability

Probability

30

Travel Time (Hrs)

0.15
Mean = 3,640489

0.10

0.25

Mean = 3,007329

0.20
0.15
0.10

0.05

0.05
0.00
0

0.00
10

20

30

40

Travel Time (Hrs)

50

60

70

10

20

30

40

Travel Time (Hrs)

Fig. 5. Probability distributions for the link travel time of the stochastic links.

50

819

Total System Travel Time

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

57000
56000
55000
54000
53000
52000
51000
50000
0.8

0.85

0.9

Probability that each selected route will be the shortest in terms of travel time
Fig. 6. Total system travel time subject to reliability requirements.

It should be noted that the implementation of the above presented optimization framework to realistic large-scale circumstances, could possibly require the use of advance computational architectures, like parallel or distributed computing.
4. Conclusions
The liberalization in railway sector, the introduction of new wagon ownership regimes and the technological evolution in
GPS, computing, mobile communication and Internet offer new potentialities but also impose higher productivity targets for
the wagon management. This work investigated technological and organizational requirements for the operation of a system
that offers real-time wagon monitoring and management functionalities. The relevant research concluded that the essential
components of such a system are: (a) a user/wagon reservation scheme, on-board track-and-trace equipment with autonomous battery and advanced energy-saver apparatus and a loading/unloading sensor and (b) a suitable software platform that
include a prediction mechanism for the time of the wagon arrival at destination station and wagon asset management (to
allow for the wagon bid and offer, the dispatching operations and to facilitate the wagon delay management).
In addition, certain wagon management issues have been further investigated, namely:

s The adequacy of the classic Transportation Problem formulation to solve the wagon dispatching process and the requirements to tackle the imbalance between wagon offer and wagon demand as well as to allow for the fair allocation of benets among the users of the system. The implementation of a data recording system for the non-satised wagon
dispatching cases in order to compensate in a similar future case, has been foreseen as a potential solution to this problem.
s The benets that can be achieved thanks to ETA information. An appropriate wagon dispatching strategy has been proposed that can improve the performance of the system, in terms of wagon-km. An order of magnitude of these saving
has been identied through model-based investigation of hypothetical wagon offer and demand scenarios.
s The trade-off between wagon availability (in terms of reliable wagon arrival at demand nodes) and systems performance
(in terms of total travel time of wagons). An optimization framework based on hybrid GAs has been developed and applied
in a hypothetical network for addressing the reliable optimal empty wagons allocation problem. The results revealed that
reliability considerations may signicantly alter the systems performance.

References
Adamidou, E.A., Kornhauser, A.L., Koskosidis, Y.A., 1993. A game theoretic/network equilibrium solution approach for the railroad freight car management
problem. Transportation Research Part B 27B (3), 237252.
Ballis, A., 2005. Freight Railways Transport: Wagon Fleet Management Aspects. 2nd International Congress: Development of Railway Transport Sector.
Hellenic Institute of Transportation Engineers, Athens, Greece.
Cordeau, J.-F., Toth, P., Vigo, D., 1998. A survey of optimization models for train routing and scheduling. Transportation Science 32 (4), 380404.
Dejax, P., Crainic, T., 1987. Models for empty freight vehicle transportation logistics. Transportation Science 21, 227247.
EC, 2001a. White Paper on Transport. European Transport Policy for 2010. Time to Decide, 2001.
EC/DG INFSO, 2002. F-MAN Research Project, Deliverable D3: Users Needs for a Future Fleet Management, Brussels.
EC/DG INFSO, 2003a. F-MAN Research Project, Deliverable D6: Description of the F-Man Tools, Brussels.
EC/DG INFSO, 2004. F-MAN Research Project, Deliverable D11.2: Verication Report, Brussels.
EC/DG RESEARCH, 2003b. CroBIT Research Project, Brussels.
EC/DG-TREN, 2001b. Symonds Group. A Study of Single Wagonload Rail Trafc. Final Report.
El Faouzi, N.E., Maurin, M., 2007. Reliability metrics for path travel time under log-normal distribution. In: van Zuylen, H.J. (Ed.), Proc. INSTR2007.
Glickman, T., Sherali, H., 1985. Large-scale network distribution of pooled empty freight cars over time, with limited substitution and equitable benets.
Transportation Research Part B 19 (2), 8594.
Goldberg, D., 1989. Genetic Algorithms in Search, Optimization and Machine Learning. Addison-Wesley, Reading, MA.
Hillier, F., Lieberman, G., 2002. Introduction to Operations Research, seventh ed. McGraw-Hill.

820

A. Ballis, L. Dimitriou / Transportation Research Part C 18 (2010) 807820

Holguin-Veras, J., Thorson, E., 2003. Modeling commercial vehicle empty trips with a rst order trip chain model. Transportation Research Part B 37, 129
148.
Holland, J.H., 1975. Adaptation in Natural and Articial Systems. University of Michigan Press (Second edition: MIT Press, 1992).
Iman, R.L., Conover, W.J., 1982. A distribution-free approach to inducing rank correlation among input variables. Communications in Statistics B11, 311
334.
Kikuchi, S., 1985. Empty freight car dispatching model under freight car pool concept. Transportation Research Part B 19 (3), 169185.
Kim, T.J., Kinkle, J.J., 1982. A model for statewide freight transportation planning. TRB 61st Annual Meeting, Washington, DC.
Kuhla, E., Ballis, A., Guigon, M., 2003. F-MAN will boost wagon productivity in Europe. Railway Gazette International Journal 159 (11), 720722.
Lim, A., Wang, F., Xu, Z., 2006. A transportation problem with minimum quantity commitment. Transportation Science 40 (1), 117129.
Limpert, E., Stahel, W.A., Abbt, M., 2001. Lognormal distributions across the sciences: keys and clues. Bioscience 51 (5), 341352.
Martn, J., Prez, C.J., 2009. Bayesian analysis of a generalized lognormal distribution. Computational Statistics and Data Analysis 53, 13771387.
Mendiratta, V.B., Turnquist, M.A., 1982. Model for management of empty freight cars. Transportation Research Record 838, 5055.
Mitchell, M., 1998. An Introduction to Genetic Algorithms. MIT Press.
Mixa, S.C., 1972. Linear programming of empty wagon disposition. Rail International 151158.
Powell, W.B., Carvalho, T.A., 1998. Real-time optimization of containers and atcars for intermodal operations. Transportation Science 32 (2), 110126.
Sanchez, I.G., Romero, R., Mantovani, J.R.S., Garcia, A., 2005. Interior point algorithm for linear programming used in transmission network synthesis. Electric
Power Systems Research 76, 916.
White, W.W., Bomberault, A.M., 1969. A network algorithm for empty freight car allocation. IBM Systems Journal 8, 147169.