Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

Computers and Operations Research 115 (2020) 104670

Contents lists available at ScienceDirect

Computers and Operations Research


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

Scheduling of maintenance windows in a mining supply chain rail


network
Thomas Kalinowski a,b, Jason Matthews b, Hamish Waterer b,∗
a
School of Science and Technology, University of New England, Australia
b
School of Mathematical and Physical Sciences, University of Newcastle, Australia

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

Article history: Rail infrastructure forms a critical part of the mining supply chain in Australia due to the high weight to
Received 27 July 2017 volume ratio of the product and the long distances between the mines and the ports. Across Australia, rail
Revised 14 February 2019
infrastructure has been steadily expanding to account for the growth in export volumes and the move-
Accepted 27 March 2019
ment of mining operations further inland, and so the efficient and effective management of this critical
Available online 30 March 2019
infrastructure is vitally important. Maintenance plays a crucial role in this management as it ensures that
Keywords: the infrastructure assets are in a condition that allows safe, reliable, and efficient transport.
Rail infrastructure
In this paper we consider the annual planning of maintenance for Australia’s largest coal rail network, the
Maintenance scheduling
Maintenance windows
Central Queensland Coal Network (CQCN), that is owned, operated, and managed, by Aurizon Holdings Pty
Mixed integer programming Ltd. The current planning approach at Aurizon uses the concept of a maintenance access window (MAW)
Matheuristic which provides a train-free time window across geographically contiguous track locations that define a
maintenance zone. These train-free time windows facilitate the scheduling of specific maintenance tasks
at specific track locations within zones closer to day of operation and forms the basis for a planning
framework.
A MIP model is introduced which facilitates the planning of different maintenance resources across this
network to schedule MAWs. The model takes into account maintenance requirement forecasts as well
as the availability of resources. Candidate solutions are compared using a proxy for network throughput
capacity. Due to the long computation times required to solve the MIP model at the annual planning
horizon a matheuristic is developed and two variants are tested. On average 80% less computational time
is required to find a good solution (average gap of 5%) using the matheuristic compared to solving the
MIP model directly (average gap of 1.5%).
The MIP model and associated matheuristic provides a suitable framework for semi-automated mainte-
nance planning and is being integrated into the current suite of decision support tools used by Aurizon.
© 2019 Published by Elsevier Ltd.

1. Introduction Australia. Rail infrastructure forms a critical part of coal and iron
ore export supply chains due to the high weight to volume ratio
The mining sector plays an important role in the Australian of the product and the long distances between the mines and the
economy and contributed an estimated AUD 143 billion to its net ports. Across Australia, rail infrastructure has been steadily expand-
exports in the 2016 financial year (Bureau of Resources and Energy ing to account for the growth in export volumes and the move-
Economics, 2016). With an annual net export of AUD 40 billion ment of mining operations further inland, and so the efficient and
from coal and AUD 50 billion from iron ore, these two commodi- effective management of this critical infrastructure is vitally impor-
ties account for approximately 62% of the total mineral exports in tant.

1.1. Maintenance planning



Corresponding author.
E-mail addresses: tkalinow@une.edu.au (T. Kalinowski),
Maintenance plays a crucial role in the management of rail in-
jason.matthews@newcastle.edu.au (J. Matthews), hamish.waterer@newcastle.edu.au frastructure as it ensures that the infrastructure assets are in a con-
(H. Waterer). dition that allows safe, reliable, and efficient transport. However,

https://doi.org/10.1016/j.cor.2019.03.016
0305-0548/© 2019 Published by Elsevier Ltd.
2 T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670

the planning of rail infrastructure maintenance is complex as sev- available to perform maintenance activities, the time period during
eral, often competing, objectives need to be accounted for and off- which infrastructure capacity is reduced may be shortened, that is,
set against each other. These objectives include: (a) ensuring that an increase in expenses may result in a reduction in loss of rev-
the condition of the assets are suitable for safe and effective use; enues. Also, by scheduling maintenance activities in adjacent parts
(b) the operational costs of maintenance activities is minimised; of the rail network consecutively, it may be possible to reduce
and (c) the impact of maintenance activities on the throughput of crew and equipment repositioning expenses. However reduced ca-
the rail network is minimised. pacity in specific parts of the rail network for prolonged periods of
The maintenance of infrastructure assets requires resources, time may result, which can unfairly disadvantage specific infras-
such as equipment and personnel, which implies an expense to the tructure users. How to find the right trade-off is far from obvious.
infrastructure provider. At the same time, the maintenance of in- Short-term maintenance planning uses up-to-date information
frastructure assets temporarily reduces the throughput of the rail on the condition of rail infrastructure assets and the volume and
network, which can result in a loss of revenue for the infrastruc- timing of demand of the rail infrastructure to adjust the long-term
ture provider. Moreover, the cost and impact depends on the loca- maintenance schedule. Maintenance activities may be adjusted or
tion and the timing of the maintenance activities, because infras- rearranged to respond to wear (deterioration) above or below what
tructure usage varies over time and the repositioning of personnel was expected, for example, increasing or reducing the amount of
and equipment costs money. Also, the location and timing of main- a certain type of maintenance or even postponing a maintenance
tenance activities needs to be balanced with the frequency of those activity altogether.
activities. Reducing the time between consecutive maintenance ac-
tivities on an asset results in an increase in costs. 1.2. Maintenance access windows
The planning of maintenance is further complicated by inher-
ent uncertainty. The amount of maintenance required in any given To obtain secure access to a rail infrastructure asset for main-
period cannot be projected perfectly, as the deterioration in the tenance a resource needs to obtain a (work) possession. A posses-
condition of rail infrastructure assets depends on factors, such as sion restricts train access to the corresponding track section(s) and
weather and rail traffic density, that are not known with certainty may also induce traffic or speed restrictions on neighbouring track
in advance. The volume and the timing of product that has to be sections. It is the impact of these possessions which creates the
transported in any given period cannot be projected perfectly ei- conflict between maintaining the condition of the assets and the
ther, particularly as spot market sales are becoming an increasingly throughput of the rail network. One approach to scheduling these
important part of the portfolio of mining companies. possessions is to use maintenance access windows (MAWs) which
An additional difficulty is that the planning of rail infrastruc- are seen as train-free time windows scheduled on geographically
ture maintenance in supply chains cannot be viewed in isolation contiguous track locations that define a maintenance zone, before
of the facilities at either end, namely, the mine sites and the ports. the train timetable is constructed. These train-free time windows
Maintenance schedules of the load point infrastructure at the mine facilitate the scheduling of specific maintenance tasks at specific
sites, of the terminal infrastructure at the ports, and of the rail in- track locations within zones closer to day of operation and forms
frastructure, need to be coordinated. This is a demanding, intricate the basis for a planning framework.
process, involving multiple stakeholders and therefore is conducted In this paper we consider the use of MAWs in the annual main-
relatively infrequently, which implies a need for long-term mainte- tenance plan for Australia’s largest coal rail network, the Central
nance plans, often covering a period of one to two years. However, Queensland Coal Network (CQCN), which is owned, operated, and
given that the rate at which the condition of rail infrastructure as- managed, by Aurizon Holdings Pty Ltd. Fig. 1 illustrates the layout
sets deteriorate, and the volume and timing of demand on the rail of the CQCN which consists of approximately 2700 km of track.
infrastructure are uncertain, these long-term maintenance plans The distance between the most southern (Gladstone) and most
can only provide a framework to guide the construction of more northern (Abbot Point) ports is approximately 700 km. The ma-
detailed short-term maintenance schedules. Short-term schedules jority of the coal exported from these ports is shipped to Asian
must reflect more accurate up-to-date knowledge of the condition markets in China, Japan, and South Korea. Aurizon has partitioned
of, and demand on, the rail infrastructure assets. the CQCN into maintenance zones consisting of multiple track sec-
The information available when constructing long-term main- tions and junctions. Fig. 2 illustrates this partitioning with zones
tenance plans includes — for each rail infrastructure asset and for ranging in size from 19 km to 388 km of track. Each zone consists
each type of maintenance, based on maintenance policy and anal- of a trunk (or main line) and adjacent balloon loops which serve
ysis of historical data — an estimate of the amount of maintenance the load points at the mines and the dump stations at the coal ter-
that has to be performed within a given (repeating) period, for ex- minals. This main line structure is illustrated by an extract from a
ample, 12 h of rail grinding in each 12-week period. The types and network line diagram shown in Fig. 3. The primary purpose of the
amounts of resources required to perform that maintenance, and MAWs is to align resources and consolidate maintenance activities
the cost and availability of each resource type over time, can also on the main line which is viewed as the critical path of the zone. It
be estimated. The maintenance required over a given repeating pe- is assumed that maintenance on balloon loops can be planned in-
riod must be delivered in short time intervals, each scheduled at a dependently of MAWs as the amount of maintenance required and
specific date and time. The resulting patterns of maintenance spec- the impact on the overall network capacity from this maintenance
ify a set of time intervals during which all or part of the track at is much less. We therefore only consider maintenance on the main
that location is not available for regular operations, for example, line of a zone when scheduling MAWs.
3 h intervals repeated every week starting at 11 am on the first The zones form the geographical building blocks for MAW
Monday of the planning horizon. Both the maintenance pattern schedules as the train-free time window defined by a MAW will
and the specific maintenance activities performed in each time in- apply to an entire zone. Once a MAW has been scheduled for a
terval in the pattern must be determined so that the impact on zone, maintenance planners can schedule individual possessions
available throughput capacity is minimised and all required main- at specific track locations within that zone during their short-
tenance is carried out at minimum cost with the available mainte- term maintenance planning phase. Once a short-term maintenance
nance resources. schedule is completed, around four weeks from day of operation,
Several trade-offs have to be considered when constructing trains are routed around the individual possessions where possible
such long-term maintenance plans. By increasing the resources to increase the throughput of the network.
T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670 3

Fig. 1. A map of the Central Queensland Coal Network (CQCN) illustrating the scale of the rail network infrastructure maintained by Aurizon.

Aurizon identified three types of possession (A, B, and C), illus- in capacity. All possessions scheduled on single track areas are C-
trated in Fig. 4, each of which have a different estimated capac- type possessions.
ity impact. The estimated capacity impact of a possession is based The different types of possessions lead to Aurizon defining
on the complexity of the manoeuvres and the amount of track re- three corresponding types of MAWs (A, B, and C) in an effort to
quired to coordinate the passing of trains in the up and down di- manage the collective impacts of the possessions. Any possession
rections through the surrounding stations. A-type possessions oc- can be scheduled within a C-type MAW, B-type MAWs can accom-
cur on the track sections between consecutive stations and require modate A- and B-type possessions, and A-type MAWs are limited
the smallest amount of track to coordinate the passing of trains. It to A-type possessions.
is estimated that on average an A-type possession results in a 50% In this paper we present a model for developing an annual
reduction of capacity within the proximity of the possession. B- MAW schedule for the CQCN which seeks to:
type possessions require approximately three times as much track
to coordinate the passing of trains compared to A-type possessions 1. Ensure that there is a sufficient number of MAWs per zone to
and are estimated to result in a 75% reduction in capacity. Finally, achieve all of the required maintenance for each resource type.
C-type possessions are those which do not allow for any traffic to 2. Ensure that there is sufficient flexibility to accommodate
pass through in either direction and so result in a 100% reduction changes to a maintenance plan closer to day of operation.
4 T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670

Fig. 2. A graphical illustration of Aurizon’s partitioning of the Central Queensland Coal Network (CQCN) into zones of different sizes.

Fig. 3. An extract from a line diagram illustrating the main line and balloon loops in a zone. Included in the extract is a portion of the main line consisting of duplicated
track segments and three balloon loops servicing the loadouts at several mines.

3. Minimise the impact of maintenance windows on the through- day of operation with limited integration with other resource man-
put (or capacity) of the network. agers.

Scheduling too few maintenance windows results in the plan-


ning of work outside of these windows, which further reduces net- 1.3. Overview of the paper
work capacity. Too many maintenance windows can result in inef-
ficient use of maintenance windows. This inefficiency is caused by The main contribution of this paper is the development of a
resource managers scheduling possessions independently closer to prototype MIP model and a matheuristic solution approach for
T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670 5

Fig. 4. An illustration of the different types of possessions. The rectangular blocks indicate the extent of the possessions. The dashed lines indicate the movement of trains
when negotiating the possessions. A-type possessions occur on either the up or down track within a duplicated track section. B-type possessions occur on either the up or
down track at a station within a duplicated track section. C-type possessions are any possessions which reduce throughput completely and are the only type of possessions
which may occur on single track.

scheduling MAWs for the Central Queensland Coal Network. This fic, scheduling minor possessions which occur in between traffic
model extends the PACE suite of decision tools that have facili- and in the shadow of major work, and scheduling train-free slots
tated improved decision support within Aurizon (Savelsbergh et al., or maintenance windows before train timetabling occurs.
2014). The MIP model and associated matheuristic provides a suit- Lidén (2015) identifies several major trends in medium
able framework for Aurizon to generate MAW schedules which term major possession scheduling namely: the construction and
align resource availability and reduce the impact on network ca- scheduling of regular possession patterns; the coordination of
pacity. maintenance tasks; and the adjustment of a given mainte-
The remainder of the paper is structured as follows. We discuss nance plan. den Hertog et al. (2005) and van Zante-de Fokkert
the relevant literature in Section 2 and provide a formal defini- et al. (2007) construct possession patterns which provide regular
tion of the MAW scheduling problem in Section 3. In Section 4 we track access. In these studies the network is broken into smaller
describe a MIP model for the problem and prove its validity. A working zones and regular possessions planned for each work-
matheuristic for finding good quality solutions to the problem is ing zone. Budai-Balke (2009) and Pouryousef et al. (2010) fo-
presented in Section 5 and the results of our computational exper- cus on coordinating maintenance projects/tasks to reduce posses-
iments are presented in Section 6. Our conclusions are presented sion time and maintenance costs. Boland et al. (2011) and Boland
in Section 7. et al. (2013) adjust a given maintenance plan in an effort to im-
prove the overall capacity of the network. Here, possessions are
2. Literature review scheduled into predetermined time windows while maximising
network capacity using a network flow model.
In this section we review some relevant literature, focusing on There has also been some attention given to short term pos-
maintenance window and possession scheduling in rail networks, session scheduling (planning horizon of days to weeks). In most
and highlight both the similarities and differences to the problem cases the scheduling of possessions is done around a given train
faced at Aurizon. timetable with a limited scope to adjust train schedules them-
Lidén (2015) summarises a list of rail maintenance planning selves. Higgins (1998) schedules maintenance crews and pos-
problems some of which have received little attention in the litera- sessions while minimising the impact on the train schedule.
ture. Among the problems in the medium term (planning horizons Lake et al. (20 0 0) proposes a model for short term maintenance
from weeks to one year) is the possession scheduling problem. The scheduling once train schedules have already been put in place.
author describes several facets of the possession scheduling prob- The model seeks to minimise maintenance costs and schedules in-
lem including scheduling major possessions which impact rail traf- dividual maintenance crews to activities.
6 T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670

Albrecht et al. (2013) introduces a problem space search heuris- J The set of resource types. Each resource has a single associated
tic to simultaneously schedule train paths and maintenance pos- resource type j ∈ J.
T The set of time periods. A typical instance consists of a planning
sessions. The heuristic was tested using a long haul rail network
horizon of one year and each time period t ∈ T has a duration of one
over a one day planning horizon. The results illustrate the potential hour, i.e., |T | = 365 × 24.
for such a heuristic to be used for on demand possession schedul- uj The duration in hours of a MAW with resource type j ∈ J. The duration
ing in the short term. Forsgren et al. (2013) introduces a mathe- of a MAW is solely dependent upon its resource type.
Tzp j A subset of T . The set Tzp j is the set of permissible time periods that a
matical model which schedules possessions within given time win-
MAW in zone z ∈ Z, with maximium capacity reduction p ∈ P and
dows while simultaneously scheduling train traffic in an effort to resource type j ∈ J, can start at the beginning of.
minimise train delays and cancellations. This model was tested us-
ing several scenarios in a 24 h planning horizon.
Definition 1. A maintenance access window (MAW) is an element
Budai et al. (2006) proposes a mathematical formulation to
(z, p, j, n, t) of the set Z × P × J × N>0 × T and denotes a MAW in
build a possession schedule for preventative maintenance activi-
zone z with a maximum capacity reduction of p that starts at the
ties. The objective was to group maintenance activities together to
beginning of time period t and requires n resources of resource
reduce the number of possessions and the impact of possessions
type j to be allocated to the MAW for the duration uj .
on rail traffic. Several heuristic approaches were proposed to solve
the mathematical formulation. The models were tested using a sin- Definition 2. A MAW schedule is a set M of MAWs satisfying the
gle rail link only, but may be expanded to a network using the con- following two conditions:
cept of single track grids introduced by den Hertog et al. (2001).
Lidén and Joborn (2017) focus on simultaneously scheduling (z, p, j, n, t ) ∈ M ⇒ t ∈ Tzp j , (1)
maintenance activities and trains on the Swedish rail network.
Maintenance activities are scheduled as maintenance windows (z, p, j, n, t ), (z, p , j, n , t  ) ∈ M ⇒ |t − t  |  u j . (2)
where a maintenance window is seen as a regular train-free slot
which is determined before the train timetable is constructed. Condition (1) ensures that MAWs start at permissible times.
These windows mimic the impact of maintenance on train activity Condition (2) ensures that MAWs in the same zone for the same
before the procurement of maintenance contracts takes place. Con- resource type do not overlap.
tractors schedule the contracted work within these maintenance Informally, a feasible MAW schedule is one that allows the
windows and in most instances use some of the time to setup be- scheduling of all work activities without overusing any resources.
fore, and packup after, the actual maintenance activity. This results More specifically, it has to satisfy the following conditions:
in a trade-off between the cost of completing the maintenance ac-
tivities and the impact on train traffic depending on the mix (dura- 1. The number of resources required by the scheduled MAWs at
tion and quantity) of maintenance windows. Both the cost of main- any point in time must not exceed the number of resources
tenance activities as well as the cost to delay or cancel passenger available for each resource type. Assigning resources to MAWs
and freight train services was taken into account. is the resource allocation problem.
The segmentation of the CQCN into zones shows some similar- 2. The number of hours of work achievable within the scheduled
ity to the work described by den Hertog et al. (2005), however, MAWs must exceed the amount required to achieve all of the
the zones in our problem consist of much larger portions of the forecast maintenance activities to be scheduled closer to day
network. Moreover, in our setting both work and train activity can of operations. Assigning the forecast work to the MAWs is the
occur within a maintenance window depending on where the pos- work allocation problem.
sessions are eventually scheduled in the zone on day of operation.
In the remainder of this section we provide a formal definition
Lidén and Joborn (2017) apply maintenance windows to indi-
for a feasible MAW schedule. In Sections 3.1 and 3.2 we describe
vidual sections, however, in our paper a maintenance window cov-
formulations for the resource allocation problem and the work al-
ers a larger portion of the network, namely a zone. Moreover, we
location problem respectively. The formal definition of a feasible
replace the scheduling of individual trains by proxy measures for
MAW schedule corresponding to the two conditions listed above is
network capacity due to the delayed scheduling of possessions and
then given in Section 3.3.
the length of the planning horizon.
The coordination of work for different resource types into
the same maintenance window resembles the framework by 3.1. Resource allocation problem
Budai et al. (2006). However, Budai et al. (2006) does not con-
sider the coordination of resources which must be shared between The number of resources required for each MAW must be con-
zones and focuses on projects at the more granular section level sistent with the actual availability of resources. In most cases re-
once resource allocations have already been made. sources are shared between zones and may move from a base of
The MAW scheduling problem in this paper shows some sim- operation to any location within their jurisdiction. The collection
ilarities to these studies in the literature, but differs due to the of zones that defines the jurisdiction of each resource for each re-
planning horizon, network zoning, and resource coordination. source type, is referred to as a zone cluster. For each resource type,
the corresponding set of zone clusters covers the set of zones, that
3. MAW scheduling is, each zone is a member of one or more zone clusters associated
with that resource type. Moreover, the same zone cluster can de-
In this section we describe a new approach for scheduling fine the jurisdiction of resources for more than one resource type.
MAWs. Before defining a MAW schedule we introduce the follow- Fig. 5 illustrates the zone clusters used to manage one of the re-
ing notation: source types. Note that for this resource type Zone 4 can be ser-
viced by resources from two clusters.
Z The set of zones. Each zone z ∈ Z is a set of geographically contiguous
Maintenance resources also follow distinct rosters depending
track locations. The set Z partitions the rail network.
P The set of maximum capacity reduction percentages, one for each on the resource type and other non-maintenance related activi-
MAW type. Each MAW type has a single associated percentage p ∈ P. ties such as training days. The combination of zone clusters and
MAW type A corresponds to p = 50, type B corresponds to p = 75, and resource rosters provides the necessary input to address resource
type C corresponds p = 100.
allocations in the planning process.
T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670 7

Fig. 5. A schematic representation of the zone clusters for one of the resource types. There are five zone clusters each serviced by one or more independent resources of
this type. Zone 4 is serviced by resources from both of the overlapping clusters B and C.

Before defining the resource allocation problem we introduce Given a MAW schedule M and a resource availability vector b,
the following notation: the resource allocation problem is to find a resource allocation yˆ.
Condition (3) ensures that each MAW is allocated the correct num-
K The set of zone clusters where K ⊆ 2Z . Each zone cluster γ ⊆ Z is a
collection of zones that defines the possible jurisdiction of one or
ber of resources from the zone clusters which overlap with that
more resources. The same zone cluster can define the jurisdiction of zone. Condition (4) ensures that the total number of resources al-
resources for more than one resource type. located to MAWs from a particular zone cluster at any point in
Kz A subset of K. The set Kz = {γ ∈ K : z ∈ γ } is the set of zone clusters time does not exceed the total number of resources available from
that contain zone z ∈ Z.
that zone cluster.
bγ jt The number of resources from zone cluster γ ∈ K that have resource
type j ∈ J and are available during time period t ∈ T . These resources
can be utilised by a MAW (z, p, j, n, s ) ∈ M if the zone z ∈ γ and the 3.2. Work allocation problem
start time s ∈ (t − u j , t].
yˆγ zp jnt The number of resources from cluster γ ∈ K allocated to the MAW
(z, p, j, n, t ) ∈ M. Before MAWs can be scheduled the amount of preventative
maintenance and renewal work to be done on the network must
Definition 3. For a MAW schedule M ⊆ Z × P × J × N>0 × T and be estimated. This data is then used to calculate the number of
|K ||J||T | resource hours within MAWs to be made available and timing
a resource availability vector b ∈ Z0 , a resource allocation is
|M |
thereof to guarantee the completion of the work tasks taking into
a vector yˆ = (yˆγ zp jnt ) ∈ Z0 satisfying the following two condi- account some expected deviations. The resulting set of track ac-
tions: cess demands must then be met by the MAW schedule and forms
 the input when developing a MAW schedule. The track access re-
yˆγ zp jnt = n, (z, p, j, n, t ) ∈ M, (3)
quirements for each resource type take on one of five forms: (1)
γ ∈Kz
 weekly (52 weekly requirements); (2) monthly (13 4-weekly re-
yˆγ zp jns  bγ jt , γ ∈ K, j ∈ J, t ∈ T . (4) quirements); (3) quarterly (four 13-weekly requirements); (4) an-
(z,p, j,n,s )∈M: nual (one annual requirement); (5) frequency (regular asset inspec-
z∈γ , s∈(t −u j ,t ]
tions).
8 T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670


Definition 4. A track access demand is an element (z, p, j, ts , te , d) TS = Tδ
of the set Z × P × J × T × T × R>0 and denotes a demand for d re- δ ∈S
source hours of resource type j in zone z within the time window
pδ = min p
[ts , te ) with a minimum capacity reduction of p. (z,p, j,n,t )∈M (δ )

Before defining the work allocation problem we introduce the pS = min pδ


δ ∈S
following notation:
Dz j = {S ⊆ Dz j : ∀ S , S ⊆ Dz j where S = S ∪ S ⇒ TS ∩ TS = ∅}
The set of track access demands where D ⊆ Z × P × J × T × T × R>0 . For
Dz j = {S ∈ Dz j : ∀ S ∈ Dz j \ 2S where TS ⊆ TS ⇒ pS > pS }
D
convenience we also denote a track access demand (z, p, j, ts , te , d ) ∈ D
by δ, i.e. δ ∈ D.
We further define the following variables and constraints:
Tδ The time window [ts , te ) ⊆ T associated with demand
δ = (z, p, j, ts , te , d ) ∈ D. xzp jt A binary variable indicating whether a MAW with a capacity reduction
M (δ ) A subset of M. The set M (δ ) = { (z, q, j, n, t ) ∈ M : q  p, t ∈ Tδ } is the of p ∈ P starting at the beginning of time period t ∈ Tzp j in zone z ∈ Z
set of MAWs that can fulfil demand δ ∈ D. is activated for resource type j ∈ J.
dδ The number of resource hours d of resource type j required by yγ zp jt An integer variable indicating the number of resources of type j ∈ J
demand δ = (z, p, j, ts , te , d ) ∈ D. assigned to zone z ∈ Z with a capacity reduction p ∈ P starting at the
yˆzp jnt δ The number of resources from MAW (z, p, j, n, t ) ∈ M allocated to track beginning of time period t ∈ Tzp j ∩ Tγ j from the pool of resources
access demand δ ∈ D. available for zone cluster γ ∈ K.
Yγ jt An integer variable indicating the number of resources of type j ∈ J
Definition 5. For a MAW schedule M and a set D ⊆ Z × P × J × T × assigned to a MAW starting at the beginning of time period t ∈ Tγ j
from cluster γ ∈ K.
T × R>0 of track
 access demands, a work allocation is a vector yˆ =
|M ( δ )|
(yˆzp jnt δ ) ∈ Z0δ∈D satisfying the following two conditions:
  
 xzp jt   1, z ∈ Z, j ∈ J, t ∈ Tzp j ,
yˆzp jnt δ = dδ /u j , δ ∈ D, (5) p∈P t  ∈Tzp j ∩(t −u j ,t ] p∈P
(z,p, j,n,t )∈M (δ )
(7)
 0  yγ zp jt  aγ jt xzp jt ,
yˆzp jnt δ  n, (z, p, j, n, t ) ∈ M. (6) 
δ ∈D:(z,p, j,n,t )∈M (δ ) j ∈ J, γ ∈ K, z ∈ γ t ∈ Tzp j : t ∈ Tγ j ,
p∈P
Given a MAW schedule M and track access demands D, the
work allocation problem is to find a work allocation yˆ. Condi- (8)

tion (5) ensures that a sufficient number of resource hours have Yγ js  bγ jt ,
been assigned to each track access demand. Condition (6) ensures s∈(t −u j ,t ]∩Tγ j
that the number of resources assigned to track access demands
from a MAW does not exceed the number of resources available.
j ∈ J, γ ∈ K, t ∈ T : (t − u j , t] ∩ Tγ j = ∅,
(9)
3.3. MAW scheduling problem
 
yγ zp jt = Yγ jt , j ∈ J, γ ∈ K, t ∈ Tγ j ,
z∈γ p∈Pz jt
Now that we have formally defined the resource allocation and
work allocation problems, we can formally define a feasible MAW (10)
schedule. The MAW scheduling problem is then to find a feasible
   
yγ zp jt  dδ /u j , S ∈ Dz j , z ∈ Z, j ∈ J. (11)
MAW schedule M.
δ ∈S (z,p, j,t )∈Mδ γ ∈Kz δ ∈S
Definition 6. Given track access demands D, resource availability Constraint set (7) ensures that only one MAW is activated for a
b, and permissible MAW starting time periods Tzpj for all (z, p, specific resource type in a specific zone at any point in time which
j) ∈ Z × P × J, a feasible MAW schedule is a MAW schedule M satis- ensures property (2) of a MAW schedule. Constraint (8) ensures
fying the following two conditions: that if resources are allocated to that zone during a time period
then a MAW is activated. The allocation of available resources is
1. There exists a resource allocation for M and b.
done by constraints (9) and (10). Constraint (11) ensures that all of
2. There exists a work allocation for M and D.
the track access demands are met.
An objective function is now introduced by considering the ca-
4. MIP model pacity impact of MAWs on each zone. We define the following sets,
indices, and parameters:
In this section we describe a MIP formulation for the MAW Pzt A subset of P. The set Pzt is the set of possible capacity reductions
scheduling problem described in Section 3. Before describing the which can occur in zone z ∈ Z during time period t ∈ T .
MIP formulation we introduce the following notation. For a zone cz The capacity of zone z ∈ Z.
czpt The marginal proportional decrease in capacity by having a MAW with
z ∈ Z, a zone cluster γ ∈ K, a resource type j ∈ J, a time period t ∈ T,
a capacity reduction of at least p ∈ P active over time period t ∈ T in
a track access demand δ ∈ D, and a subset of demands S⊆D, we de- zone z ∈ Z.
fine:
We now introduce the following additional variables and con-
aγ jt = min bγ js straints:
s∈[t ,t +u j )
xzpt A binary variable indicating whether a MAW with a capacity reduction
Tγ j = {t  ∈ T : aγ jt  > 0} of at least p ∈ P is active over time period t ∈ T in zone z ∈ Z.

Pz jt = { p ∈ P : t ∈ Tzp j }  
Dz j = {(z , p, j , ts , te , d ) ∈ D : z = z, j = j} xzq jt   xzpt , z ∈ Z, j ∈ J, t ∈ T , p ∈ Pzt .
q∈Pzt :q p t  ∈Tzq j ∩(t −u j ,t ]
Mδ = {(z, q, j, t ) : q ∈ P, q  p, t ∈ Tδ }
(12)
T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670 9

Constraint set (12) determines the highest capacity reduction (i) (a) xzp jt = 0 ⇒ yγ zp jt = 0, and
impact across all maintenance windows overlapping time period
(b) M(x, y) is a MAW schedule, and
t ∈ T. The overall weighted capacity reduction is then minimised in
(c) yˆγ zp jnt is a feasible solution for the resource allocation
the following objective:
 problem, and
Minimise cz czpt xzpt . (13) (d) there exists a feasible solution (yˆzp jnt δ ) for the work allo-
z∈Z p∈P t∈T cation problem.

4.1. Model validity (ii) The vector (x, y, Y) satisfies (7)–(11).

To show that formulation (7)–(12) produces a feasible MAW Proof.


schedule for the given objective function we need only prove that
the formulation (7)–(11) provides a feasible MAW schedule as con- (i)⇒(ii). Constraints (7) follow from (2) and (b). For (8), fix (z,
straint (12) is used to determine an objective function value. The p, j, n, t) ∈ M(x, y), γ ∈ Kz , and let t  ∈ [t, t + u j ) such that
following lemma shows that constraints (11) for S ∈ Dz j imply that bγ jt  = aγ jt . Then
the inequality is satisfied for every subset S⊆Dzj .  (4 )
yγ zp jt = yˆγ zp jnt  yˆγ zp jns  bγ jt  = aγ jt
Lemma 1. Every vector y = (yγ zp jt ) which satisfies (11) also satis-
(z,p, j,n,s )∈M:
fies z∈γ , s∈(t  −u j ,t  ]
   
yγ zp jt  dδ /u j , S ⊆ Dz j , z ∈ Z, j ∈ J. and together with the condition (a) this implies (8). For (9),
δ ∈S (z,p, j,t )∈Mδ γ ∈Kz δ ∈S
  
(14) Yγ js = yˆγ zp jns
s∈(t −u j ,t ]∩Tγ j s∈(t −u j ,t ]∩Tγ j (z,p, j,n,s )∈M (x,y ):
Proof. Suppose y satisfies (11) and violates (14) for some S⊆Dzj . Let z∈γ
S be a minimal counterexample, i.e., (14) is satisfied for all proper  (4 )
subsets of S. This implies S ∈ Dz j , because otherwise S = S ∪ S for = yˆγ zp jns  bγ jt .
(z,p, j,n,s )∈M:
some S , S ⊆Dzj with TS ∩ TS = ∅, and S or S is a smaller coun- z∈γ , s∈(t −u j ,t ]
terexample. From S ∈ Dz j it follows that there exists S ∈ Dz j with
S ⊆ S, TS ⊆ TS and pS  pS . Consider the set S = S ∪ S ∈ Dz j . Note Constraints (10) follow from
that
  (z, p, j, n, t ) ∈ M (x, y ) ⇒ xzp jt = 1 ⇒ p ∈ Pz jt ,
Mδ = Mδ ,
δ ∈S δ ∈S and therefore
  
and therefore yγ zp jt = yγ zp jt = Yγ jt .
      
yγ zp jt = yγ zp jt < dδ /u j  z∈γ p∈Pz jt (z,p, j,n,t )∈M (x,y ):
z∈γ
δ ∈S (z,p, j,t )∈Mδ γ ∈Kz δ ∈S (z,p, j,t )∈Mδ γ ∈Kz δ ∈S
 For (11), fix j ∈ J, z ∈ Z, and S⊆Djz . Then
< dδ /u j .
δ ∈S      
yγ zp jt = yˆγ zp jnt
So constraint (14) is violated for S as well. Iterating the step from
δ ∈S (z,p, j,t )∈Mδ γ ∈Kz δ ∈S (z,p, j,n,t )∈M (δ ) γ ∈Kz
S and S to S , if necessary, we can assume that S ∈ Dz j , and this
(3 )  
is the required contradiction.  = n
δ ∈S (z,p, j,n,t )∈M (δ )
We use constraint set (11) instead of constraint set (14) as it
showed better computational performance during preliminary test- (6 )   
 yˆzp jnt δ 
ing.
δ ∈S (z,p, j,n,t )∈M (δ ) δ  ∈D:
For notational convenience when proving model validity, we (z,p, j,n,t )∈M (δ  )
define   (5 ) 
   yˆzp jnt δ = dδ /u j .
A1 = |Tzp j |, A2 = |Tzp j ∩ Tγ j |, δ ∈S (z,p, j,n,t )∈M (δ ) δ ∈S
z∈Z p∈P j∈J γ ∈K z∈γ p∈P j∈J
 (ii)⇒(i). Condition (a) follows from (8). Condition (b) follows
and A3 = |Tγ j |.
from (7) which implies (2). In order to verify that (c) holds,
γ ∈K j∈J
we need to show that (3)(4) hold. If we fix (z, p, j, n, t) ∈ M(x,
A
For a vector (x, y ) ∈ {0, 1}A1 × Z20 , we define a set M(x, y) of y) then (3) follows from the definition of M(x, y):
MAWs by  
  yˆγ zp jnt = yγ zp jt = n.
 γ ∈Kz γ ∈Kz
M (x, y ) = (z, p, j, n, t ) : xzp jt = 1, n = yγ zp jt ,
γ ∈Kz For (4), fix γ ∈ K, j ∈ J, t ∈ T, and observe
and we set  
yˆγ zp jns = yγ zp js
yˆγ zp jnt = yγ zp jt , (z, p, j, n, t ) ∈ M (x, y ), γ ∈ Kz ,
 (z,p, j,n,s )∈M: (z,p, j,n,s )∈M:
z∈γ , s∈(t −u j ,t ] z∈γ , s∈(t −u j ,t ]∩Tzp j
Yγ jt = yγ zp jt , γ ∈ K, j ∈ J, t ∈ Tγ j .
(z,p, j,n,t )∈M (x,y ):    (10 )  (9 )
z∈γ = yγ zp js = Yγ jt  bγ jt .
A A s∈(t −u j ,t ]∩Tγ j z∈γ p∈Pz js s∈(t −u j ,t ]∩Tγ j
Lemma 2. For a vector (x, y, Y ) ∈ {0, 1}A1 × Z20 × Z30 , the follow-
ing two statements are equivalent. For (d), fix j ∈ J and z ∈ Z. For every S⊆Dzj ,
10 T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670

  (3 )    iteratively fixing carefully selected subsets of variables and search-


n = yγ zp jt
ing the induced large scale neighbourhood by solving the resulting
δ ∈S (z,p,n, j,t )∈M (δ ) δ ∈S (z,p,n, j,t )∈M (δ ) γ ∈Kz
smaller, more tractable, MIP. Rather than focusing on the subsets
   (14 ) 
= yγ zp jt  dδ /u j . of variables to be fixed, we consider the complementary view of
δ ∈S (z,p, j,t )∈Mδ γ ∈Kz δ ∈S decomposing the problem into subproblems, each of which can be
induced by fixing those variables not in the subproblem of inter-
By Hall’s theorem, this implies that there exists a feasible est. Two orthogonal strategies for decomposing our problem in a
solution for the system (5) and (6). Since the work alloca- meaningful way are to consider subproblems induced by a single
tion problem decomposes into independent problems for the resource type, and to consider subproblems induced by a restricted
pairs (z, j) ∈ Z × J this concludes the proof. planning horizon.
During preliminary testing it was observed that the computa-
 tional time required to solve a MIP instance to optimality for a sin-
gle resource type was relatively short. This lead to the first decom-
4.2. Additional constraints position in which the problem is split into resource type specific
subproblems Ij defined as the MIP model (7)–(13), (15), and (16),
An additional operational side constraint is also introduced restricted to the resource type j ∈ J. This was achieved by fixing the
which allows the user to incorporate pre-planned MAWs when de- variables associated with all the other resource types j ∈ Jࢨ{j} to
veloping a schedule. We achieve this with parameter mzpt ∈ {0, 1}, their current values.
which indicates whether there is a planned MAW overlapping time The second decomposition splits the problem temporally into
period t ∈ T in zone z ∈ Z of at least capacity reduction p ∈ P, and the subproblems Ie defined as the MIP model (7)–(13), (15), and (16),
constraint with the planning horizon restricted to a single week e ∈ E where
the set E is defined to be the set of weeks (Monday to Sunday).
mzpt  xzpt z ∈ Z, t ∈ T , p ∈ Pzt . (15)
This was achieved by fixing the variables associated with t ∈ Tࢨ{e}
We further include the aggregate demand constraint below to their current values. Such a decomposition exploits the struc-
which improves the preprocessing performance of the solver and ture of the track access demands δ ∈ D as the time windows Tδ are
the overall computational performance of the MIP model. multiples of weeks.
    Our matheuristic has both a construction and an improvement
aγ jt xzp jt  dδ /u j , S ∈ Dz j , z ∈ Z, j ∈ J. phase. In the construction phase, an initial MAW schedule for each
δ ∈S (z,p, j,t )∈Mδ γ ∈Kz δ ∈S resource type is determined sequentially by solving each resource
(16) type subproblem in turn. In the improvement phase, we first up-
date the current MAW schedule for each resource type subproblem
before refining the resulting MAW schedule for the whole problem
5. Matheuristic by solving each week subproblem in turn. A high-level overview of
the matheuristic framework is given in Algorithm 1.
During development various alternate MIP models using dif-
ferent variable spaces were tested on problem instances with
planning horizons of 12 weeks. While the model introduced in Algorithm 1: Matheuristic framework.
Section 4 showed the best computational performance, subse- /* Construction phase */
quent testing on problem instances with an annual planning hori- 1 Solve MIP subproblem I j for each resource type j ∈ J
zon revealed that it was difficult to find good feasible solutions /* Improvement phase */
2 while Stopping criteria for improvement phase not met do
to larger instances of the problem early in the solution proce- /* Resource type subproblem phase */
dure which was an important requirement of Aurizon. To this end, 3 while Stopping criteria for resource type subproblems notmet do

a simple matheuristic was developed. Matheuristics are “model- 4 for j ∈ J sorted by descending| z∈Z Dz j | do
based” heuristics that exploit an underlying mathematical pro- 5 Solve MIP subproblem I j
6 end
gramming model (see, for example, Boschetti et al., 2009). Due to
7 end
significant recent advances in mathematical programming software /* Weekly subproblem phase */
(Bixby and Rothberg, 2007), matheuristics are becoming increas- 8 for e ∈ E do
ingly widely used to find high quality solutions quickly to chal- 9 Solve MIP subproblem Ie
lenging MIPs (see, for example, Archetti and Speranza, 2014; Erera 10 end
et al., 2013; Hewitt et al., 2010; Papageorgiou et al., 2018). 11 end

As is the case with many matheuristics, we maintain a single


MIP model and seek good feasible solutions relatively quickly by

Table 1
A description of the major differences between the problem instances.

Instances Number of Number of Planned Amount of


resource types MAW types MAWs track access demand

Aq , Ay 9 3 No Normal
Bq , By 9 1 No Normal
Cq , Cy 7 3 No Normal
Dq , Dy 6 3 Yes Normal
Eq , Ey 9 3 No Low

Instances Time horizon

A q , Bq , Cq , Dq , E q 12 weeks (1 quarter)
A y , By , Cy , Dy , E y 52 weeks (1 year)
T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670 11

Table 2
A description of the types of track access demand profiles present in each problem instance.

Instance Weekly Monthly Quarterly Annually Frequency

Aq , Ay X X X X X
Bw , By X X X X X
Cq , Cy X X X X
Dq , Dy X X X
Eq , Ey X X X X X

Several variations of the matheuristic framework in the performances. All computations were carried out on a Dell
Algorithm 1 were tested by changing the stopping criteria of PowerEdge R710 with dual hex core 3.06 GHz Intel Xeon X5675
the resource type subproblem phase within the improvement processors and 96 GB RAM, running Red Hat Enterprise Linux 6.
phase. Two resource type subproblem phase stopping criteria were All mathematical models were solved using Gurobi 7.0.1 using four
tested which resulted in two variants of the matheuristic. The first threads. The gaps in Tables 3 and 4 and Figs. 6 and 7 are calcu-
variant (IN-FT-LP) would iterate over the resource type subprob- lated with respect to the best lower bound found within the time
lems once per iteration of the improvement phase. The second limits.
variant (IN-LP-FT-LP) would repeat the resource type subproblem The performance of the MIP model and matheuristic variants
phase until no improvement was made during this iteration of for instances with a 12 week planning horizon is illustrated in
the improvement phase or the time limit was exceeded. During Table 3. In all cases a feasible solution is found within the 40 min
preliminary testing these two variations significantly outperformed time limit by the MIP model with a gap ranging from 0.3% to
other variations of the matheuristic framework that were tested. 1.5%. The two matheuristic variants perform equally well with gaps
For both variants the matheuristic would stop if a time limit ranging from 3.5% to 5.1% and achieve an initial solution, in most
was exceeded (the same time used to test the MIP model) or cases, within 10% of the time required by the MIP model. How-
no improvement was made during the previous iteration (lines 3 ever, the time required by the IN-LP-FT-LP variant to find the
to 10). best solution is up to three times more than that of the IN-FT-LP
variant.
6. Computational results The effectiveness of the matheuristic is further highlighted us-
ing a one year planning horizon. Here, the MIP model did not pro-
To test the performance of the proposed MIP model and vide a feasible solution within the 2 h time limit for the instances
matheuristic several realistic instances were generated using actual with a normal amount of track access demand and included all re-
maintenance profiles, maintenance forecasts, and resource rosters source types and MAW types. Moreover, in most cases it took in
obtained from Aurizon. The instances differ in terms of the number excess of one hour to find an initial feasible solution. The quality
of resource types considered; the number of MAW types allowed; of solutions achieved by the two matheuristic variants are similar
inclusion of planned MAWs; amount of track access demand; the and have a gap ranging between 2.98% and 6.69%. In most cases
types of track access demand profiles; and the length of the plan- the time required to find the best solution for the IN-FT-LP vari-
ning horizon. These instances are described in Tables 1 and 2. ant is half the time required for the MIP model to find an initial
Both matheuristic variants as well as the MIP model were ap- solution.
plied to all the instances. Tables 3 and 4 illustrate the results af- The progression of convergence to a good solution by the two
ter solving all the instances with a 12 week and one year plan- matheuristic variants is illustrated in Figs. 6 and 7. These graphs
ning horizon, respectively. A maximum runtime of 40 min and suggest that there is no marginal benefit for multiple resource type
two hours was set for the instances with a 12 week and one year subproblem loops in the matheuristic framework. Moreover, we
planning horizon respectively. For each instance and solution ap- can see the benefits of finding initial solutions quickly using the
proach combination, five independent tests were run to compare matheuristic.

Table 3
A comparison between the different heuristic approaches and the MIP model for instances with 12 week planning horizons. Each instance was solved five times with each
solution approach. The results in each column are represented as: average (minimum, maximum).

Number of Time to first Time to best


Data Model iterations Runtime solution solution First solution Best solution Gap

Aq MIP 2400 (2400, 2400) 83 (81, 84) 113 (111, 114) 1180 (1180, 1180) 1167 (1167, 1167) 0.4% (0.4%, 0.4%)
Aq IN-FT-LP 4 (4, 4) 85 (81, 88) 16 (14, 18) 73 (69, 76) 1409 (1409, 1409) 1203 (1203, 1203) 3.5% (3.5%, 3.5%)
Aq IN-LP-FT-LP 2 (2, 2) 110 (109, 111) 16 (14, 17) 87 (85, 88) 1409 (1409, 1409) 1216 (1216, 1216) 4.6% (4.6%, 4.6%)
Bq MIP 2400 (2400, 2400) 55 (55, 56) 81 (81, 82) 1011 (1011, 1011) 1003 (1003, 1003) 0.3% (0.3%, 0.3%)
Bq IN-FT-LP 4 (4, 4) 52 (51, 54) 8 (7, 8) 45 (43, 46) 1224 (1224, 1224) 1043 (1043, 1043) 4.3% (4.3%, 4.3%)
Bq IN-LP-FT-LP 3 (3, 3) 66 (65, 67) 8 (8, 8) 51 (50, 53) 1224 (1224, 1224) 1051 (1051, 1051) 5.1% (5.1%, 5.1%)
Cq MIP 2400 (2400, 2400) 79 (78, 81) 160 (158, 162) 1191 (1191, 1191) 1160 (1160, 1160) 0.4% (0.4%, 0.4%)
Cq IN-LP-FT-LP 3 (3, 3) 157 (152, 161) 12 (11, 13) 134 (130, 136) 1400 (1400, 1400) 1197 (1197, 1197) 3.6% (3.6%, 3.6%)
Cq IN-FT-LP 5 (5, 5) 93 (92, 94) 12 (11, 12) 70 (69, 71) 1400 (1400, 1400) 1201 (1201, 1201) 4.0% (4.0%, 4.0%)
Dq MIP 2400 (2400, 2401) 27 (26, 27) 32 (32, 32) 1174 (1174, 1174) 1160 (1160, 1160) 0.4% (0.4%, 0.4%)
Dq IN-FT-LP 4 (4, 4) 52 (51, 53) 6 (6, 6) 32 (30, 32) 1386 (1386, 1386) 1202 (1202, 1202) 4.1% (4.1%, 4.1%)
Dq IN-LP-FT-LP 4 (4, 4) 112 (110, 114) 6 (6, 6) 99 (97, 101) 1386 (1386, 1386) 1203 (1203, 1203) 4.2% (4.2%, 4.2%)
Eq MIP 2400 (2400, 2400) 125 (124, 127) 1212 (1199, 1229) 818 (818, 818) 806 (806, 806) 1.5% (1.5%, 1.5%)
Eq IN-FT-LP 4 (4, 4) 85 (83, 86) 18 (18, 19) 73 (72, 74) 921 (921, 921) 827 (827, 827) 4.2% (4.2%, 4.2%)
Eq IN-LP-FT-LP 3 (3, 3) 149 (145, 151) 14 (13, 14) 124 (122, 126) 921 (921, 921) 830 (830, 830) 4.5% (4.5%, 4.5%)
12 T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670

Fig. 6. A graph illustrating the convergence of solutions by the MIP model, and the IN-FT-LP and IN-LP-FT-LP matheuristic variants for a single test for each instances with
a 12 week time horizon.
T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670 13

Fig. 7. A graph illustrating the convergence of solutions by the MIP model, and the IN-FT-LP and IN-LP-FT-LP matheuristic variants for a single test for each instances with
a one year time horizon.
14 T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670

Table 4
A comparison between the different heuristic approaches and the MIP model for instances with one year planning horizons. Each data set was solved five times with each
solution approach. The results in each column are represented as: average (minimum, maximum). A dash indicates that no feasible solution was found within the time limit
for all attempts.

Number of Time to first Time to best


Data Model iterations Runtime solution solution First Solution Best solution Gap

Ay MIP 7201 (7200, 7201) – – – – –


Ay IN-FT-LP 13 (13, 13) 3301 (3262, 3361) 232 (222, 244) 3233 (3192, 3292) 6196 (6196, 6196) 5131 (5131, 5131) 4.1% (4.1%, 4.1%)
Ay IN-LP-FT-LP 6 (6, 6) 6622 (6526, 6697) 245 (216, 264) 6397 (6299, 6472) 6196 (6196, 6196) 5139 (5139, 5139) 4.2% (4.2%, 4.2%)
By MIP 7200 (7200, 7200) 3702 (3499, 6697 (6531, 7181) 4372 (4372, 4372) 4335 (4331, 4338) 1.3% (1.2%, 1.4%)
4112)
By IN-FT-LP 12 (12, 12) 1620 (1601, 1639) 140 (133, 147) 1487 (1472, 1505) 5434 (5434, 5434) 4528 (4528, 4528) 5.8% (5.8%, 5.8%)
By IN-LP-FT-LP 5 (5, 5) 2224 (2192, 2242) 146 (133, 160) 2068 (2031, 2086) 5434 (5434, 5434) 4536 (4536, 4536) 6.0% (6.0%, 6.0%)
Cy MIP 7200 (7200, 7200) 5594 (5180, 6665 (6411, 7099) 4995 (4995, 4995) 4972 (4969, 4976) 1.4% (1.4%, 1.5%)
6042)
Cy IN-FT-LP 7 (7, 7) 1415 (1385, 1441) 142 (131, 150) 1343 (1312, 1368) 6156 (6156, 6156) 5114 (5114, 5114) 4.3% (4.3%, 4.3%)
Cy IN-LP-FT-LP 6 (6, 6) 4621 (4561, 4682) 144 (135, 150) 4411 (4357, 4470) 6156 (6156, 6156) 5141 (5141, 5141) 4.9% (4.8%, 4.9%)
Dy MIP 7201 (7200, 7201) 1494 (1386, 4840 (4683, 5031) 4872 (4872, 4872) 4862 (4862, 4862) 0.6% (0.6%, 0.6%)
1613)
Dy IN-LP-FT-LP 4 (4, 4) 4150 (4128, 4178) 154 (144, 168) 3989 (3966, 4018) 5948 (5948, 5948) 4980 (4980, 4980) 3.1% (3.1%, 3.1%)
Dy IN-FT-LP 10 (10, 10) 1446 (1407, 1478) 161 (143, 180) 1286 (1249, 1319) 5948 (5948, 5948) 4991 (4991, 4991) 3.3% (3.3%, 3.3%)
Ey MIP 7203 (7200, 7214) 7008 (6817, 7022 (6831, 7186) 3955 (3955, 3955) 3950 (3950, 3950) 4.6% (4.6%, 4.6%)
7172)
Ey IN-FT-LP 16 (16, 16) 4218 (4161, 4278) 232 (203, 266) 3930 (3870, 3990) 4628 (4628, 4628) 3985 (3985, 3985) 5.5% (5.5%, 5.5%)
Ey IN-LP-FT-LP 6 (6, 6) 6442 (6412, 6469) 227 (210, 256) 6148 (6107, 6180) 4628 (4628, 4628) 4001 (40 01, 40 01) 5.9% (5.9%, 5.9%)

7. Conclusions and future work We would like to acknowledge the valuable contributions of
Mike Backhouse, Matt Dall, Steve Straughan, Scott Thomas, and
The effective scheduling of maintenance in a rail network is a Miles Till, from Aurizon. Without their patience, support, and feed-
complex task and requires a resolution to conflicts between the back, this research could not have occurred. We also thank Aurizon
desire for low maintenance costs and the need for adequate net- and the Australian Research Council for their joint funding under
work capacity. Aurizon makes use of maintenance access windows the ARC Linkage Grant LP1401010 0 0.
(MAWs) across the different zones defined within their rail net-
work to manage this conflict. The current MAW scheduling ap- References
proach is manual and based on spreadsheets which makes it dif-
ficult for the decision maker to consider the two major aspects of Albrecht, A., Panton, D., Lee, D., 2013. Rescheduling rail networks with maintenance
disruptions using problem space search. Comput. Operat. Res. 40 (3), 703–712.
MAW scheduling, namely resource availability and network capac- Archetti, C., Speranza, M.G., 2014. A survey on matheuristics for routing problems.
ity impact. EURO J. Comput. Optim. 2 (4), 223–246.
A formal definition of MAWs and a MAW schedule was pro- Bixby, R., Rothberg, E., 2007. Progress in computational mixed integer program-
ming—a look back from the other side of the tipping point. Ann. Oper. Res. 149
vided building on the definitions of a resource allocation and work
(1), 37–41.
allocation problem. A MIP model was introduced and is shown to Boland, N., Kalinowski, T., Waterer, H., Zheng, L., 2011. An optimisation approach to
generate a MAW schedule which provides a feasible work and re- maintenance scheduling for capacity alignment in the Hunter Valley Coal Chain.
In: Baafi, E.Y., Kininmonth, R.J., Porter, I. (Eds.), Proceedings of the 35th APCOM
source schedule. The model takes into account resource rosters,
Symposium. The Australasian Institute of Mining and Metallurgy, pp. 887–897.
track access demands based on high level estimations and network Boland, N., Kalinowski, T., Waterer, H., Zheng, L., 2013. Mixed integer programming
capacity impacts at a zone level. based maintenance scheduling for the Hunter Valley Coal Chain. J. Schedul. 16
Due to the long computation times required to solve the MIP (6), 649–659.
Boschetti, M.A., Maniezzo, V., Roffilli, M., Bolufé Röhler, A., 2009. Matheuristics:
model at the annual planning horizon a matheuristic was de- Optimization, simulation and control. In: Blesa, M.J., Blum, C., Di Gaspero, L.,
veloped to solve the model and two variants were tested. The Roli, A., Sampels, M., Schaerf, A. (Eds.), Hybrid Metaheuristics. Springer Berlin
matheuristic variants provided good solutions (average 5% gap) Heidelberg, Berlin, Heidelberg, pp. 171–177.
Budai, G., Huisman, D., Dekker, R., 2006. Scheduling preventive railway maintenance
within reasonable time frames (20-60 minutes for instance with activities. J. Operat. Res. Soc. 57 (9), 1035–1044.
an annual planning horizon). The MIP model and associated Budai-Balke, G., 2009. Operations Research Models for Scheduling Railway Infras-
matheuristic provides a suitable framework for generating MAW tructure Maintenance. Erasmus University Ph.D. thesis.
Bureau of Resources and Energy Economics, 2016. Resources and energy quarterly,
schedules within a decision support tool for MAW scheduling September quarter. Technical Report. Australian Government, Commonwealth of
across this large scale coal rail network. Australia.
A natural extension of this research is to explore more complex Erera, A., Hewitt, M., Savelsbergh, M., Zhang, Y., 2013. Improved load plan design
through integer programming based local search. Transp. Sci. 47 (3), 412–427.
objective functions taking into account the interactions of trains
Forsgren, M., Aronsson, M., Gestrelius, S., 2013. Maintaining tracks and traffic flow
between zones. This may require a more complex structure to cap- at the same time. J. Rail Trans. Plann. Manag. 3 (3), 111–123.
ture an objective function and different solution approaches. In ad- den Hertog, D., van Zante-de Fokkert, J.I., Sjamaar, S., Beusmans, R., 2005. Optimal
working zone division for safe track maintenance in The Netherlands. Accident
dition, the physical design of the zones has not been extensively
Anal. Prevent. 37 (5), 890–893.
analysed and this tool may be used to help design and evaluate al- den Hertog, D., van Zante-de Fokkert, J.I., Sjamaar, S.A., Beusmans, R., 2001. Safe
ternative zone configurations taking into account the robustness of Track Maintenance for the Dutch Railways, Part I: Optimal Working Zone Divi-
MAW schedules and overall network impact. sion. University of Tilburg. Unpublished paper
Hewitt, M., Nemhauser, G.L., Savelsbergh, M.W.P., Stewart, M.H., 2010. Combining
exact and heuristic approaches for the capacitated fixed-charge network flow
Acknowledgments problem. INFORMS J. Comput. 22 (2), 314–325.
Higgins, A., 1998. Scheduling of railway track maintenance activities and crews. J.
Oper. Res. Soc. 49 (10), 1026–1033.
This research was supported by the ARC Linkage Grant Lake, M., Ferreira, L., Murray, M., 20 0 0. Minimising costs in scheduling railway track
LP1401010 0 0. maintenance. WIT Trans. Built Environ. 50.
T. Kalinowski, J. Matthews and H. Waterer / Computers and Operations Research 115 (2020) 104670 15

Lidén, T., 2015. Railway infrastructure maintenance - A survey of planning problems scenarios. In: Proceedings of the Joint Rail Conference. American Society of Me-
and conducted research. Transp. Res. Procedia 10, 574–583. chanical Engineers, ASME, pp. 317–326.
Lidén, T., Joborn, M., 2017. An optimization model for integrated planning of railway Savelsbergh, M., Waterer, H., Dall, M., Moffiet, C., 2014. Possession assessment and
traffic and network maintenance. Trans. Res. Part C Emerg. Technol. 74, 327–347. capacity evaluation of the Central Queensland Coal Network. EURO J. Transp.
Papageorgiou, D.J., Cheon, M.-S., Harwood, S., Trespalacios, F., Nemhauser, G.L., 2018. Logist. 4 (1), 139–173.
Recent Progress Using Matheuristics for Strategic Maritime Inventory Routing. van Zante-de Fokkert, J.I., den Hertog, D., van den Berg, F.J., Verhoeven, J.H.M., van
Springer International Publishing, Cham, pp. 59–94. den Berg, F.J., Verhoeven, J.H.M., 2007. The Netherlands schedules track mainte-
Pouryousef, H., Teixeira, P., Sussman, J., 2010. Track maintenance scheduling and nance to improve track workers’ safety. Interfaces (Providence) 37 (2), 133–142.
its interactions with operations: dedicated and mixed high-speed rail (HSR)

You might also like