Sei sulla pagina 1di 35

Accepted Manuscript

Optimizing rolling stock assignment and maintenance plan for passenger rail-
way operations

Yung-Cheng Lai, Dow-Chung Fan, Kwei-Long Huang

PII: S0360-8352(15)00125-4
DOI: http://dx.doi.org/10.1016/j.cie.2015.03.016
Reference: CAIE 3991

To appear in: Computers & Industrial Engineering

Received Date: 21 February 2014


Revised Date: 27 September 2014
Accepted Date: 25 March 2015

Please cite this article as: Lai, Y-C., Fan, D-C., Huang, K-L., Optimizing rolling stock assignment and maintenance
plan for passenger railway operations, Computers & Industrial Engineering (2015), doi: http://dx.doi.org/10.1016/
j.cie.2015.03.016

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Optimizing Rolling Stock Assignment and Maintenance Plan
for Passenger Railway Operations

Yung-Cheng Lai1,a, Dow-Chung Fan1,b, Kwei-Long Huang1,c,*

1
Railway Technology Research Center, National Taiwan University, Room 313, Civil Engineering
Building, No 1, Roosevelt Road, Sec 4, 10617, Taipei, Taiwan
2
Institute of Industrial Engineering, National Taiwan University, Room 102, IYC Building, No 1,
Roosevelt Road, Sec 4, 10617, Taipei, Taiwan
a
yclai@ntu.edu.tw, b r00521522@ntu.edu.tw, c craighuang@ntu.edu.tw

Abstract
The efficient use of railway rolling stock is an important objective pursued in a
railway agency or company because of intensive capital investment in rolling stock.
Daily rolling stock assignment assigns appropriate equipment to cover a given set of
utilization paths in the utilization schedules while considering practical requirements,
such as maintenance, depot capacity, and circulation rules. Experienced railway
practitioners can generally produce a feasible assignment plan; however, this manual
process is time consuming, and an optimal solution is not guaranteed. This research
develops an exact optimization model to improve the efficiency in rolling stock usage
with consideration of all necessary regulations and practical constraints. Compared to
the manual process, a hybrid heuristic process is also developed to improve solution
quality and efficiency. Empirical results demonstrate that the heuristic process can
successfully increase the efficiency of rolling stock use by about 5% and significantly
reduce the solution time from 3 hours to 11.2 seconds. Using this decision support
tool can help railways with similar characteristics to improve the efficiency in rolling
stock usage and productivity in rolling stock assignment process.

Keywords: Rail transportation; Rolling stock assignment planning; Maintenance


planning; Heuristics; Decision-support model

Optimizing Rolling Stock Assignment and Maintenance Plan


for Passenger Railway Operations

Abstract

*
Tel: 886-2-33669508
Email: craighuang@ntu.edu.tw
1
The efficient use of railway rolling stock is an important objective pursued in a
railway agency or company because of intensive capital investment in rolling stock.
Daily rolling stock assignment assigns appropriate equipment to cover a given set of
utilization paths in the utilization schedules while considering practical requirements,
such as maintenance, depot capacity, and circulation rules. Experienced railway
practitioners can generally produce a feasible assignment plan; however, this manual
process is time consuming, and an optimal solution is not guaranteed. This research
develops an exact optimization model to improve the efficiency in rolling stock usage
with consideration of all necessary regulations and practical constraints. Compared to
the manual process, a hybrid heuristic process is also developed to improve solution
quality and efficiency. Empirical results demonstrate that the heuristic process can
successfully increase the efficiency of rolling stock use by about 5% and significantly
reduce the solution time from 3 hours to 11.2 seconds. Using this decision support
tool can help railways with similar characteristics to improve the efficiency in rolling
stock usage and productivity in rolling stock assignment process.

Keywords: Rail transportation; Rolling stock assignment planning; Maintenance


planning; Heuristics; Decision-support model

1. Introduction

1.1 Background

Rolling stock, wheeled vehicles that move on a railway, is a capital-intensive


component for a railway operator. The sole conventional railway operator in Taiwan,
Taiwan Railways Administration (TRA), uses and manages a number of different
rolling stock with complicated maintenance and inspection regulations. Therefore, the
efficient use of rolling stock is one main objective in railway operation (Abbink et al.,
2004, Alfieri et al., 2006, Caprara et al., 2007, Cheng, 2010).
The general operational planning process of passenger railways usually includes
the determination of a timetable, a rolling stock schedule, rolling stock assignments, a
crew schedule, and crew rosters (Caprara et al., 2007). Fig. 1 demonstrates the
planning process at TRA. Based on service demand, timetabling process creates an
appropriate timetable containing all the train services (i.e., train trips) with specific
departure and arrival times at every station on route. According to a specific
timetable, a rolling stock utilization schedule (a.k.a., utilization schedule) and a crew
schedule are produced through the rolling stock scheduling and crew scheduling

2
processes by connecting and covering all train trips in the timetable. The utilization
schedule has multiple utilization paths and each contains either a sequence of train
trips within a defined period or a specific maintenance task. The train trips in a
utilization path are connected according to several practical requirements, such as
rolling stock availability, turnaround time, and station capacity. Given the utilization
schedule, appropriate rolling stock are assigned to all the utilization paths in the
utilization schedule through the rolling stock assignment process in the daily
operation, subject to maintenance requirements, depot capacity, and circulation rules.
The crew rostering process assigns crew to serve the crew schedule and the
corresponding train trips.

Fig. 1: The operational planning process at TRA

Fig. 2 shows an example of a utilization schedule with four utilization paths (E3,
E4, E7, and E9). Each utilization path is composed of several trips and coordinated
departure/arrival time and places. The combinations of trips provide information on
how long the entire path is, how many days it takes to finish this path, which types of
rolling stock should be used, how many of them are required, and where and when
they start, finish, and execute inspections. The length of utilization paths at TRA can
vary from one to four days. For example, E3 is a one-day utilization path, whereas E9
is a four-day utilization path. Each path on each day requires a particular train formed
by multiple train-sets combined (a train-set is a set of passenger cars and power
vehicle(s) permanently connected together as a unit).

Fig. 2: Sample set of utilization paths

Based on the utilization schedule, rolling stock planners determine how to


assign appropriate rolling stock to cover all paths every day (i.e., the rolling stock
assignment process). For TRA, this process has to consider several practical
constraints and special operational characteristics, such as multiple levels of
inspection/maintenance, time- and mileage-based inspection rules, multi-day
utilization paths (one to four days), maintenance capacity, multiple types of rolling
stock, and uncertainty in train-set availability. Similar to other railway systems in the
world, this process is still manual at TRA because of its complexity and stochastic
nature (Cadarso and Marín, 2010, Cadarso and Marín, 2011). Experienced
practitioners can formulate effective and feasible assignments. However, this process
is time consuming, and the optimal solution is not guaranteed. Moreover, considering
only a short-term process may lead to a myopic decision far from the global optimum.
Consequently, it is necessary to improve the efficiency and quality of this daily
3
assignment process.

1.2 State of the art


Several previous studies on rolling stock planning in passenger rail have
focused on the development of rolling stock utilization schedules with optimal paths.
This process is often called rolling stock circulation, scheduling, or routing in
literature. Different planning objectives, such as minimizing operational cost and the
number of train-sets used, have been considered (Schrijver, 1993, Zhao et al., 2002,
Alfieri et al., 2006, Fioole et al., 2006, Giacco et al., 2011, Giacco et al., 2014). Few
studies have also accounted for passenger demand as an input by minimizing the
shortage of seat supply (Fioole et al., 2006, Peeters and Kroon, 2008). Other
considerations include rolling stock composition, shunting cost, connection time
between trips, and deadhead (i.e., non-revenue movement of rolling stock) (Brucker et
al., 2003, Fioole et al., 2006, Peeters and Kroon, 2008, Xie et al., 2010, Giacco et al.,
2011, Zhu et al., 2013). However, most of these studies did not include maintenance
cost in the planning objective. In addition, the assignment between specific rolling
stock and utilization paths is not determined in this process.
In contrast to the problem described in this research, rolling stock assignment
was also used in the past to determine the type and composition of the train used in
each trip (Cacchiani et al., 2010, Cadarso and Marín, 2011). Maroti and Kroon (2005,
2007) developed maintenance routing models to adjust rolling stock utilization path
according to necessary maintenance in the coming three days. However, multiple
types and levels of maintenance and inspection are not considered, nor the effect of
deadhead movements. In addition, the demand of maintenance was pre-determined by
the model according to the conditions of specific rolling stock.

1.3 Contribution

Previous studies focused on the determination of rolling stock utilization


schedule instead of rolling stock assignment. According to Table 1, these studies
consider only single-level inspection, and fixed duration of utilization paths (e.g., a
day). Consequently, developing a comprehensive rolling stock assignment
optimization process and automating this tedious process are necessary to increase the
efficient use of rolling stock and reduce operating cost. This research develops an
exact optimization model to improve the efficiency in rolling stock usage with
consideration of all regulations and practical constraints, including multiple types of
rolling stock, multiple units of rolling stock, multiple levels of inspections, flexible
duration of utilization paths, and operations constraints on deadhead and replacement.
A hybrid heuristic process is also developed to significantly improve solution
4
efficiency while maintaining solution quality. The optimization model and heuristic
method proposed in this research can also be applied to railways with similar
operation characteristics to improve the efficiency in rolling stock usage and
productivity in rolling stock assignment process. In the following sections, we first
demonstrate the current train-set assignment process of TRA and then present an
operational train-set assignment model to optimize the train-set assignment process,
followed by empirical case studies and conclusions.

Table 1: The characteristics of different study


Rolling stock Inspection Utilization path Operation
Authors/Years
Multiple types Multiple units Single-level Multi-level Fixed duration Flexible duration Deadhead Replacement
Schrijver (1993) v v v
Zhao et al. (2002) v v v
Brucker et al. (2003) v v v
Abbink et al. (2004) v v
Maróti and Kroon (2005) v v
Alfieri et al. (2006) v v v
Fioole et al. (2006) v v v
Maróti and Kroon (2007) v v
Peeters and Kroon (2008) v v v
Budai et al. (2010) v v v
Cacchiani et al. (2010) v v v
Cadarso and Marín (2010) v v v
Cheng (2010) v v
Xie et al. (2010) v v v
Cadarso and Marín (2011) v v v v v
Giacco et al. (2011) v v v v
Nielsen et al. (2012) v v
Zhu et al. (2013) v
Giacco et al. (2014) v v v v
Kroon et al. (2014) v v
This paper v v v v v v

2. Rolling Stock Assignment Problem at TRA


The rolling stock assignment process at TRA entails assigning each rolling stock
to a utilization path in the utilization schedule. Every depot serves as the home of a set
of rolling stock, and planners at the depot are responsible for this daily assignment
process. To increase efficiency in rolling stock usage, planners aim to minimize
possible deadheads and rolling stock replacement and eliminate unnecessary
inspections. At the same time, this process has to follow a set of constraints, including
(1) demand, (2) connection, and (3) maintenance/inspection constraints. The majority
of TRA rolling stock for passenger services are multiple-unit train-sets. A train-set
usually consists of three to four units of power and non-power rolling stock
permanently connected together. Therefore, each unit within a train-set operates daily
tasks and undergoes inspections together. Depending on the demand of train trips, a
train can be formed by one to three train-sets or even more.
5
2.1 Demand

According to the utilization schedule, a utilization path has a desired set of


rolling stock and a required quantity to fulfill the carrying capacity. If the desired type
of rolling stock is unavailable, an alternative rolling stock can be used (i.e.,
replacement) but with considerable penalty because of the difference in seat
arrangements (a.k.a., replacement cost).

2.2 Connection

In terms of connection, a TRA utilization schedule can cross multiple days; thus,
the continuing connection between days within a particular path should be accounted
for. Moreover, the minimum time between assignments to different paths (i.e., buffer
time - usually 2 hours) for a rolling stock has to be complied. In other words, after
being used by a particular utilization path, the rolling stock has to stay at the station or
depot for at least 2 hours for cleaning and inspection before being assigned to the next
path.

2.3 Maintenance/Inspection

Similar to the usage of other transportation mode, maintenance/inspection


requirements are another type of complex constraints that the assignment process has
to consider (Clarke et al., 1997, Barnhart et al., 1998, Gopalan and Talluri, 1998, Talluri,
1998, Boland et al., 2000, Mak and Boland, 2000, Sriram and Haghani, 2003, Jünger et
al., 2003, Sarac et al., 2006). Regular inspections can generally be divided into four
levels: daily inspection (DI, level 1), monthly inspection (MI, level 2), bogie
inspection (BI, level 3), and general inspection (GI, level 4).
Table 2 shows the inspection requirement for the electric multiple unit (EMU),
the most common train-set in the TRA system, in terms of cumulative operating days,
cumulative operating mileage, and inspection location. The inspection requirement
sets the limit on how much time and/or distance a train-set can operate before the next
mandatory maintenance. For example, a train-set must go through MI after operating
for 90 days or for 90,000 km before it can be assigned to the next operation utilization
path with actual train trips. Therefore, planners have to keep track of the cumulative
operating days and mileage for each type of maintenance for each train-set.
High-level maintenance, such as BI and GI, is scheduled in advance for each train-set
during tactical maintenance planning because it requires longer inspection than DI
and MI and the main workshop has limited capacity. By contrast, low-level
maintenance, such as DI and MI, must be considered during the train-set assignment
process in the operational level along with restrictions on inspection location and
6
capacity. The process of DI is about an hour and is usually performed during the
connection between consecutive paths or during the overnight period in a multi-day
utilization path. On the other hand, MI requires one day so cannot be done during the
connection or embedded in a utilization path. In addition, the maintenance procedures
of higher-level maintenance include all maintenance procedures in lower-level
maintenance; therefore, after one class of maintenance process, all cumulative
operating days and mileage associated with that class and the corresponding
lower-level classes of maintenance are set to zero. For example, all associated
cumulative operating days and mileage for MI and DI should be set to zero after MI.
This special characteristic has been ignored in past research, but it is an essential part
of rolling stock assignment in practice. Aside from these requirements on rolling stock,
every depot has only a certain capacity for the number of inspections that can be
performed per day.

Table 2: Inspection regulations for EMU in TRA


Implementing
Inspection level Inspection requirement
authorities
DI Mileage 1,800 km
Rolling stock depot
Level 1 Operating days 3 days
MI Mileage 90,000 km
Rolling stock depot
Level 2 Operating days 3 months
Mileage 500,000 km
BI Operating days 1.5 years
Workshop
Level 3 Mileage 1,000,000 km
Operating days 3 years
Mileage 2,000,000 km
GI Operating days 6 years
Workshop
Level 4 Mileage 4,000,000 km
Operating days 9 years

An optimal assignment plan is one that can cover the utilization paths and
satisfy related requirements and limitations with minimum cost throughout the
decision horizon. At TRA, planners usually manually determine the assignment every
day for the next coming 2 days. Experienced practitioners can generate effective and
feasible assignments. However, this process is time consuming, and an optimal
solution is not guaranteed. Moreover, considering only the short-term process may
lead to a myopic decision far from the global optimum.

7
3. Mixed-Integer Programming Model (MIP) for Rolling Stock Assignment

The optimization model is presented in this section. To formulate this problem,


a single day is divided into two time slots (i.e., 0:00~11:59 and 12:00~23:59) because
utilization paths can begin and end either in the morning or in the afternoon at TRA.
Moreover, because of multiple day paths, a path may be divided into several partial
paths as shown in Fig. 3. A train-set assigned to a particular path should complete
service in all partial paths within this utilization path. For example, Fig. 3 shows that
E11, a 2-day utilization path, is divided into four partial paths over four time slots. A
tilde (~) indicates “overnight,” which means that the rolling stock stays in Changhua
depot at the end of day one.

Fig. 3: Time slots and partial paths


(Left: original utilization path; right: paths divided into time slots)

In practice, the total number of available train-sets every day must be able to
cover that of the utilization paths in a depot. The excess can be used as spare in case
of accidents or can stay for inspection. In our model, a train-set has four possible
states: assigned to a utilization path without DI, assigned to a utilization path with DI,
assigned to MI, and no assignment. The optimization model aims to minimize the
total operating cost associated with rolling stock assignment. The expected output
ought to include two parts, namely, rolling stock assignment plans within the planning
period and maintenance/inspection arrangements. To display these two plans, eight
types of decision variables are used in the model. These variables include four
positive variables and four binary variables. Variable xi,v,k is a binary variable that
keeps track of the status of each rolling stock and equals to 1 when rolling stock v
operates path i in time slot k or 0 otherwise. Variable qi,u,k is a binary variable that
represents whether path i is operated by rolling stock of type u in time slot k. Binary
variables yv,k and zv,k denote whether a train-set v undergoes MI or DI, respectively, in
time slot k. If yv,k or zv,k is equal to one, train-set v cannot be assigned to any path (i.e.,
xi,v,k = 0). Finally, four positive variables dv,kA1, dv,kA2, dv,kB1, and dv,kB2 are used to
control whether the rolling stock should be sent to inspection. These variables store
the DI and MI cumulative mileage and operating days for each rolling stock at the end
of time slot k. When either value of rolling stock v reaches the upper bound, which is
set based on TRA inspection regulations, the corresponding inspection is triggered.
These values return to zero after inspection. The detailed calculation and
zero-returning mechanism are explained later in the associated constraints. The
notations, including the indices, sets, and parameters, used in the optimization model
are listed in Table 3. Decision variables that appear in the model are listed in Table 4.
8
Table 3: Notation of indices, sets, and parameters of MIP model
Notation Description
indices:
i, j the utilization path; i,j I
k a particular time slot in the planning period; k K
u the type of train-sets; u U
v the available train-sets; v V
sets:
K set of time slots within the planning period
1
K set of starting time slots of each day
k
I set of utilization paths without inspection operation in time slot k
k
IA set of utilization paths comprising daily inspection in time slot k
k set of any two paths (i , j ) indicating that path j in time slot k connects
O
with path i in time slot k −1. The pair of paths should be assigned to the
same train-sets.
k set of all possible connections between two paths (i ,j ) indicating that a
L
train-set could go from path i in time slot k −1 to path j in time slot k .
U set of types of train-sets
V set of all train-sets
V' set of train-sets borrowed from other depots; V ' V
u u
V set of train-sets of type u; V V
paramters:
A
C the daily inspection cost
B
C the monthly inspection cost
E the deadhead cost incurs when a train-set is sent from one depot to
C
another due to the train-set replacement operation
C i,v,k train-set replacement operation costs in time slot k
A1
D the cumulative operating time upper-bound for daily inspection
A2
D the cumulative mileages upper-bound for daily inspection
B1
D the cumulative operating time upper-bound for monthly inspection
B2
D the cumulative mileages upper-bound for monthly inspection
F i,k the ending time of path i in time slot k
G the monthly inspection capacity in the depot
H the minimum buffer time for train-set transition
M a relatively large positive number
N i,u,k the number of train-sets of type u required for path i in time slot k
Pi the operation time of path i
Q the minimal number of monthly inspections needed per week
Si mileages of path i
T j,k the starting time of path j in time slot k
W a relatively small positive number

9
Table 4: Notation list of decision variables of MIP model
Notation Description
decision variables:
A1 non-negative integer, indicates daily inspection cumulative operating time of train-
d v,k
set v at the end of time slot k
A2 non-negative integer, indicates dailly inspection cumulative mileages of train-set v
d v,k
at the end of time slot k
B1 non-negative integer, indicates the monthly inspection cumulative operating time
d v,k
of train-set v at the end of time slot k
B2 non-negative integer, indicates the monthly inspection cumulative mileages of
d v,k
train-set v at the end of time slot k
q i,u,k binary integer, indicates whether path i in time slot k is operated by type u or
not. =1 if yes; =0 otherwise.
x i,v,k binary integer, indicates whether train-set v operates path i in time slot k or not.
=1 if yes; =0 otherwise.
y v,k binary integer, indicates whether train-set v executes monthly inspection in time
slot k or not. =1 if yes; =0 otherwise.
z v,k binary integer, indicates whether train-set v executes daily inspection in time slot
k or not. =1 if yes; =0 otherwise.

The mixed-integer linear mathematical optimization model is shown as follows:

Objective function

(1)
Min C A ∑∑ zv ,k +C B ∑∑ yv ,k +C E ∑ ∑ ∑ xi ,v ,k + ∑∑ ∑ Ci ,v ,k xi ,v ,k
v∈V k∈K v∈V k∈K k∈K v∈V ' i∈I k ∪ I Ak k∈K v∈V i∈I k ∪ I Ak

+ W ∑∑ ( d vA,1k + d vA,k2 + d vB,1k + d vB,k2 )


v∈V k∈K

Subject to

∑q i ,u , k =1 ∀k ∈ K , i ∈ I k ∪ I Ak (2)
u∈U

∑x i ,v , k = N i ,u ,k qi ,u ,k ∀k ∈ K , u ∈U , i ∈ I k ∪ I Ak (3)
v∈V u

x j,v,k = xi,v,k −1 ∀k ∈ K , v ∈V , i ∈ I k −1 ∪ I Ak −1 , (4)


j ∈ I ∪ I , (i, j ) ∈ O
k k
A
k

Tj ,k − Fi ,k −1 ≥ H − M (2 − x j ,v,k − xi ,v,k −1 ) ∀k ∈ K , v ∈V , i ∈ I k −1 ∪ I Ak −1 , (5)


j ∈ I ∪ I , (i, j ) ∈ L
k k
A
k

d vA,1k ≥ d vA,1k −1 + ∑ Pi xi ,v ,k − Myv ,k − Mzv ,k ∀k ∈ K , v ∈ V (6)


i∈I k

10
d vA,k2 ≥ d vA,k2−1 + ∑ Si xi ,v ,k − Myv ,k − Mzv ,k ∀k ∈ K , v ∈ V (7)
i∈I k

d vB,1k ≥ d vB,1k −1 + ∑ Pi xi ,v ,k − Myv ,k ∀k ∈ K , v ∈ V (8)


i∈I k

d vB,k2 ≥ d vB,k2−1 + ∑ Si xi ,v ,k − Myv ,k ∀k ∈ K , v ∈ V (9)


i∈I k

R = A1, A2, B1, B 2,


dvR,k ≤ DR (10)
∀k ∈ K , v ∈ V

∑ xi ,v ,k ≤ 1 − yv ,k ∀k ∈ K , v ∈ V (11)
i∈I k ∪ I Ak

∑x i ,v ,k ≥ zv ,k ∀k ∈ K , v ∈ V (12)
i∈I Ak

∑y
v∈V
v,k ≤G ∀k ∈ K (13)

∑∑ y
v∈V k∈K
v,k ≥Q (14)

yv,k = yv,k +1 ∀k ∈ K 1 , v ∈ V (15)

xi,v,k ∈{0,1} ∀ k ∈ K , v ∈ V , i ∈ I k ∪ I Ak (16)

yv,k , zv,k ∈{0,1} ∀k ∈ K , v ∈ V (17)

R = A1, A2, B1, B 2,


dvR,k ≥ 0 (18)
∀k ∈ K , v ∈ V

To improve the efficiency in rolling stock usage, the MILP model is constructed
by minimizing operating costs, which consist of five terms in Equation (1). The first
two terms ( C A ∑∑ zv,k and C B ∑∑ yv ,k ) represent the DI and MI costs,
v∈V k∈K v∈V k∈K

respectively, during the planning horizon. The third term ( C E ∑ ∑ ∑ xi ,v ,k ) is the


k ∈K v∈V ' i∈I k ∪ I Ak

deadhead cost incurred from non-revenue service/transportation during rolling stock


replacement or unexpected accidents/failures. The fourth term ( ∑ ∑ ∑ Ci ,v ,k xi ,v ,k )
k ∈K v∈V i∈I k ∪ I Ak

is the replacement cost of undesired train-sets assigned to a particular utilization path.


The last component of the objective function is a dummy cost which combined with
constraints (6) to (9) ensure that cumulative operating days and mileage are set to zero
after specific maintenance. Thus, W is a very small number and the value of the last
component is less than 1 so that this dummy cost would not affect the optimal
11
solution.

Constraints (2) and (3) guarantee that every path in each time slot is assigned to
one type of trans-sets and is allocated required number of train-sets, respectively. In
addition, because a multiple-day path is divided into several partial paths, constraint
(4) ensures that all partial paths from the same path are served by the same train-sets.
Note that variables xi,v,k for time slot k = 0 are input status of the train-set assignment
right before planning. Constraint (5) ensures a train-set has sufficient time for
transition when it is assigned from one utilization path to another. This transition
enables cleaning staff to go through the train-sets and train crew to perform crew
changes.
Constraints (6) to (10) handle DI and MI for train-sets. Constraints (6) and (7)
ensure that cumulative operating days and mileage with respect to DI are correctly
recorded, whereas constraints (8) and (9) are for MI. Once DI or MI is performed on a
train-set, the counters of cumulating operating days and mileage associated with the
DI of that train-set are reset to zero because higher-level maintenance includes all
maintenance procedures in lower-level maintenance. Similarly, the counters of a
train-set associated with MI return to zero when MI is conducted on that train-set. The
operating days for each path, Pi, is represented in the unit of the planning time slot,
and mileage Si is precisely indicated in kilometers. Constraint (10) conforms every
cumulative value of the train-sets to TRA inspection regulations.
Constraint (11) confirms that a train-set under MI cannot be assigned to any
path. Moreover, the constraint limits that every train-set can be assigned only to a path
within a time slot. Constraint (12) handles DI. Because DI takes only 1 to 2 hours, the
inspection operation is generally combined with some utilization paths. That is, when
a train-set requires DI, it has to be assigned to a path comprising DI operation (IAk).
Constraint (13) ensures that the number of train-sets under MI in one time slot does
not exceed depot capacity. Constraint (14) guarantees that the minimum number of
MIs is conducted within a planning horizon to balance the MI workload and avoid
idleness or overwork of inspection crews. In addition, because MI requires one
working day to complete, constraint (15) ensures that inspections start in the first time
slot of a day and end in the last time slot of that day. Finally, constraints (16) to (18)
specify the type of each decision variable.

4. Hybrid Heuristic Model for Rolling Stock Assignment

A hybrid heuristic model is developed to enhance solution efficiency because


the optimization model may be inapplicable to larger-scale practical applications. The
hybrid heuristic model is designed by combining heuristic rules and the MIP model
12
presented in previous section. The lower the number of inspections (especially for MI
due to the long duration and higher maintenance cost), the higher the efficiency in
rolling stock usage and the lower the maintenance cost. To reduce the number of MIs,
this assignment process aims to arrange rolling stock assignment by sending train-sets
to MI only if necessary (i.e., very close to the inspection requirement in Table 2).
Based on this concept, we proposed a heuristic algorithm to seek an excellent
arrangement for inspections. DIs are mostly incorporated into utilization paths; thus,
the practice in TRA repeats the same utilization path (as well as the DI patterns) for a
particular train-set. This concept can be adopted in the heuristic process for train-sets
that require no MI. However, this repeated pattern has to be interrupted if an MI is
required. According to the practice, the indicator of cumulative operating days is
always the determining factor for MI because of the slower operating speed
(compared with high-speed rail); therefore, rules and criteria can be established to
assign train-sets to MI with a minimum gap between the regulated inspection
threshold (i.e., 90 days) and the actual cumulative operating days.
The proposed hybrid heuristic model consists of four main phases, as described
below. The notations used in the heuristic process are summarized in Table 5.

Table 5: Notations of the hybrid heuristic model


Notation Description
i, j utilization path
k time slot, k K
v train-set, v V
k set of paths in time slot k
I
V set of all available train-sets
U set of all urgent train-sets during assignment
V
U' set of all regular train-sets during assignment
V
R set of all remaining train-sets
V
B1
D the cumulative operating time upper-bound for monthly inspection
F i,k the ending time of path i in time slot k
Pi the operation time of path i
T j,k the starting time of path j in time slot k
B1
d v,k non-negative integer, indicates the monthly inspection cumulative
operating time of train-set v at the end of time slot k

Phase 1: Assignment of urgent train-sets

Each train-set is prioritized based on its cumulative operating days for MI. A
train-set whose cumulative operation days are closer to the regulated inspection value
(DB1) has a higher priority for being assigned to a utilization path. That is, this
train-set has more choices of paths to be assigned so that MI regulation would not be

13
violated. Therefore, train-set v in time slot k with cumulative days (DB1 – dv,kB1) less
than the longest path (max{Pi}) is assigned to set VU and is planned first in Phase 1;
otherwise, the train-set is placed in set VU’ and is planned in Phase 2. The following
steps explain how assignment works in this phase.

Step 1.1: For time slot k, the train-sets in VU are sorted in the ascending order of
DB1 – dv,kB1.
Step 1.2: Train-sets with DB1 – dv,kB1 = 0 are assigned to MI.
Step 1.3: For each train-set v in VU, the path assigned to v in time slot k–1 is
assumed to be i.
Step 1.3.1: If Pi ≤DB1 – dv,kB1, Train-set v is assigned to path i and remove
path i from Ik;
Step 1.3.2: Else if Pj = DB1 – dv,kB1 and the minimal connecting buffer time
(min{Tj,k – Fi,k } ≥ 2), v is assigned to path j and remove path j from Ik;
Step 1.3.3: Else if l = arg max{
k
Pj } < DB1 – dv,kB1, v is assigned to path l and
j∈I

remove path l from Ik;


Step 1.3.4: Else v is placed in set VR.

Phase 2: Assignment of regular train-sets

Phase 2 deals with train-sets of lower urgency, that is, those that require no MI
in the next few days. The principle of this assignment is to repeat the previous
utilization paths for these train-sets if possible.

Step 2.1: For time slot k, set VU’ and utilization paths (Ik) are initialized.
Step 2.2: For each train-set v in VU’, the path assigned to v in time slot k–1 is
assumed to be i. Train-set v is assigned to path i if i ∈ Ik. Otherwise, v is placed in
set VR.

Phase 3: Optimization of unassigned train-sets

After Phases 1 and 2, some unscheduled train-sets (VR) and paths may remain.
These resources are left unscheduled because of the change in assignment pattern or
the arrangement of MI. The assignment of these remaining resources in this phase is
to execute the optimization model introduced in the previous section.

Step 3.1: VR and Ik are parameters of the optimization model.


Step 3.2: The assignment results of time slot k are output from the optimization
model.
14
Phase 4: Updating of cumulative values

The cumulative values of each train-set are updated after all assignments are
determined for time slot k.

Step 4.1: For each train-set v, if v is assigned to MI, set dv,kB1 = 0.


Otherwise, dv,kB1= dv,kB1+Pi , where Pi is the time length of the path assigned to v.
Step 4.2: The assignment for current time slot k (k ∈ K) is accomplished.
Step 4.3: All phases are repeated until the last time slot.

The following example illustrates the procedure of the proposed heuristic


process. Assume there are four train-sets and the operating status of each train-set is
summarized in Table 6 (a). For example, train-set v1 is currently assigned to path P2
and it is in the service of the second day of path P2. In addition, three utilization paths
need to be assigned and the corresponding length of operating days are listed in Table
6 (b).

Table 6: The data of train-sets and utilization paths

(a) Train-sets (b) Utilization paths


Monthly inspection Operating Operating
Train-sets Path
cumulative days status days
v1 89 P2(2) P2 2
v2 88 P4(4) P3 3
v3 75 P3(1) P4 4

v4 26 −
∗−means no assignment

The assignment of next day (say, Day 1) is demonstrated in the four phases of
the proposed heuristic process:

Phase 1: A train-set is considered as urgent if its cumulative operating days would


exceed upper bound of MI (i.e., DB1) after completing one more path. If DB1 is
assumed to be 90 and the longest path is P4 with 4 operating days (max{Pi}),
train-sets whose cumulative operating days are larger than or equal to 86 (DB1 −
max{Pi} = 90 −4) will be placed in set VU, resulting in set VU={v1, v2.}. On the other
hand, set VU’={v3, v4} contains train-sets with more slack time for MI and is planned
in Phase 2. Within VU, train-set v1 is planned first and its remaining day before MI is
1 (DB1 – dv,kB1 = 90 – 89). However, the shortest operating days of all paths is 2 so that
v1 is placed at set VR according to Step 1.3.4. Next, for train-set v2, its remaining days
before MI is 2. It cannot be assigned to its previous operating path P4 (Step 1.3.1) but
15
is assigned to path P2 (Step 1.3.2).

Phase 2: This phase processes the non-urgent train-sets, set VU’={v3, v4}. Train-set v3
continues its second-day service of path P3. For train-set v4, it is placed at set VR
because no path was assigned to it in previous day.

Phase 3: Set VR={v1, v4} and the set with unassigned paths Ik ={P4} which are
planned by using MIP proposed in Section 3. The solution shows that v1 is sent to MI
and v4 is assigned to P4.

Phase 4: The last phase is to update status of all train-sets so that the assignment on
Day 1 is completed. Day 2 assignment will continue by repeating these four phases.
The assignment of Days 1~3 obtained by the proposed heuristic process is
summarized in Table 7.

Table 7: The assignment result of the example


Day 1 Day 2 Day 3
Train Cumulative Operating Cumulative Operating Cumulative Operating
days status days status days status
v1 0 MI 0 − 1 P2(1)
v2 89 P2(1) 90 P2(2) 0 MI
v3 76 P3(2) 77 P3(3) 78 P3(1)
v4 27 P4(1) 28 P4(2) 29 P4(3)

The train-set assignment plan can cross multiple days, and to have an
appropriate decision horizon is often desirable to avoid myopic decisions. The usage
of train-sets is supposed to follow this plan; however, depending on the reliability of
the system, the daily operational condition for the train-sets is somewhat different
because of unexpected failures or accidents. Therefore, the original plans may need to
be adjusted according to the operational status of the train-sets. Consequently, to
implement this hybrid heuristic model, we proposed a rolling horizon planning
process in this study.
Two decisions need to be made for the rolling horizon optimization process: the
length of the (1) decision horizon and (2) implementation horizon of a single
optimization. The first decision (the length of the decision horizon) has to be made by
considering the balance between solution quality and computational time. The second
decision (the length of the implementation horizon) is based on the degree of
uncertainty in train-set availability.
In principle, maintaining a long decision horizon and a short implementation
horizon and conducting optimization once new information is available are desirable.
Therefore, we suggest setting the decision horizon and the implementation horizon to
16
7 days and 1 day, respectively, and conducting optimization every day (Fig. 4). In this
example, each optimization result determines a 7-day train-set assignment, but only
the result for the first day is implemented. This procedure continues and repeats until
the entire assignment plan is generated. As suggested, the rolling horizon method not
only promotes planning efficiency by increasing the decision horizon but also
provides great flexibility to deal with unexpected incidents.

Fig. 4: Example of rolling horizon

This rolling horizon process is responsive to uncertainty in the usage of


train-sets because it optimizes assignment on a daily basis. The process also avoids
the myopic decisions made in the manual process by considering the specified
decision horizon.

5. Empirical Case Study

To demonstrate the potential use of the rolling stock assignment process, we


implemented the proposed method to such a process at Hsinchu depot, the largest
depot in the TRA system. Three case studies were selected and analyzed. Case I, a
reduced-size scenario with fewer train-sets and utilization paths, compares the
performance of manual assignment, exact optimization based on the MIP model, and
the heuristic method. In Case II, the heuristic method is used to solve a real-size case
(which cannot be solved by the MIP model) to demonstrate its applicability to
real-world problems. A stochastic planning problem by considering rolling stock
failure is also used to demonstrate the benefit of this developed process in Case III.
Finally, a sensitivity analysis was conducted to evaluate the effects of the variation in
required buffer time, allowable cumulative days before MI, and available MI capacity
at the depot.
Hsinchu depot is the home depot of 55 EMU train-sets, that is, 220 cars serving
721 trips in a week from Luodong to Chiayi. Among these trains, there are 6 sets of
EMU500 trains, 9 sets of EMU600 trains, and 40 sets of EMU700 trains. Table 8
shows the 12 utilization paths that Hsinchu depot has to maintain for a week. Each
utilization path has its own requirement on type and number of train-sets, mileage and
operating days, starting time and ending time, and occurrence in a week. Some paths
can be operated only by the EMU700 series, some others can be operated only by the
EMU500 or EMU600 series, and others can use any train-sets. However, train-sets of
different types can never be coupled together to operate. For example, path E5
suggests that it requires two sets of either the EMU500 or the EMU600 series,
indicating that the path can be operated by two sets of the EMU500 or EMU600 series.
17
The operating days range from 1 to 4 days. Thus, this set of utilization paths actually
requires 48 train-sets per day on average, leaving less room for rolling stock
maintenance. All train-sets have to be used efficiently to cover these utilization paths.

Table 8: Utilization paths of Hsinchu depot


Required Required Operating Operating Operating frequencies
Path no.
type quantity days (day) mileages (km) (days of occurrences)
E3 500/600 1 1 269 Mon. Tue. Wed. Thu. Sun.
E4 500/600 2 2 1,338 Everyday
E5 500/600 2 1 434 Everyday
E6 All 2 3 1,783 Everyday
E7 700 2 3 1,839 Everyday
E8 700 2 2 1,269 Everyday
E9 700 2 4 1,637 Everyday
E10 700 2 3 1,916 Mon. Tue. Fri. Sat. Sun.
E10* 700 2 2 1,703 Wed. Thu.
E11 700 2 2 1,618 Everyday
E12 All 2 2 1,144 Mon. Tue. Wed. Thu. Sun.
E12* All 2 1 775 Fri. Sat.
E13 700 2 1 802 Everyday
E16 600 1 2 770 Everyday

The inspection regulations for EMU train-sets are listed in Table 2. Hsinchu
depot has nine tracks. Tracks 1 to 3 are outdoor tracks, mainly for DI and car wash.
Tracks 4 to 7 are for MI, although DI can be operated in these tracks if necessary.
Tracks 8 and 9 cannot be used for inspection because they have no catenary for these
electric train-sets. As MI or higher-level inspection is conducted for at least 1 day,
tracks 4 to 7 have capacity limitations. Each of these tracks can serve at most two sets
of rolling stock; thus, the total capacity for MI is eight train-sets per day.
Table 9 shows the initial condition of all train-sets at the beginning of the
planning process, including the initial cumulative mileage, operating days, and the
finished and unfinished utilization paths at the end of the previous day. For example,
E9(1), E9(2), E9(3), and E9(4) indicate the first, second, third, and fourth days,
respectively, of utilization path E9.

18
Table 9: Initial conditions of all train-sets in Hsinchu depot
Initial 1A Initial 2A Initial 1A Initial 2A
Initial 1A Initial 2A Initial 1A Initial 2A
inspection inspection inspection inspection
inspection inspection Operating inspection inspection Operating
Train-sets cumulative cumulative Train-sets cumulative cumulative
cumulative cumulative status cumulative cumulative status
mileages mileages mileages mileages
days (day) days (day) days (day) days (day)
(km) (km) (km) (km)
EMU542 0 500 71 42,600 E12(1) EMU714 0 500 66 39,600 E11(3)
EMU543 0 500 82 49,200 E6(1) EMU715 0 500 54 32,400 E8(1)
EMU544 0 500 61 36,600 E5(1) EMU716 0 500 54 32,400 E8(1)
EMU545 0 500 54 32,400 E6(1) EMU717 0 500 64 38,400 E9(4)
EMU546 0 500 76 45,600 E12(1) EMU718 0 500 64 38,400 E9(4)
EMU547 0 500 42 25,200 E5(1) EMU719 0 500 41 24,600 E4(1)
EMU601 1 1,000 44 26,400 Spare EMU720 0 500 41 24,600 E4(1)
EMU602 1 1,000 71 42,600 E16(2) EMU721 0 500 79 47,400 E6(2)
EMU603 1 1,000 49 29,400 Spare EMU722 0 500 79 47,400 E6(2)
EMU604 1 1,000 56 33,600 Spare EMU723 0 500 60 36,000 E9(1)
EMU605 1 1,000 74 44,400 Spare EMU724 0 500 60 36,000 E9(1)
EMU606 1 1,000 49 29,400 Spare EMU725 0 500 65 39,000 E8(2)
EMU607 1 1,000 55 33,300 Spare EMU726 0 500 65 39,000 E8(2)
EMU608 1 1,000 52 31,200 E3(1) EMU727 0 500 47 28,200 E4(2)
EMU609 1 1,000 61 36,600 E16(1) EMU728 0 500 47 28,200 E4(2)
EMU701 0 500 83 49,800 E10(1) EMU729 0 500 48 28,800 E7(3)
EMU702 0 500 83 49,800 E10(1) EMU730 0 500 48 28,800 E7(3)
EMU703 0 500 69 41,400 E6(3) EMU731 0 500 67 40,200 E7(1)
EMU704 0 500 69 41,400 E6(3) EMU732 0 500 67 40,200 E7(1)
EMU705 0 500 77 46,200 E11(1) EMU733 0 500 43 25,800 E9(2)
EMU706 0 500 77 46,200 E11(1) EMU734 0 500 43 25,800 E9(2)
EMU707 0 500 61 36,600 E10(2) EMU735 0 500 84 50,400 E9(3)
EMU708 0 500 61 36,600 E10(2) EMU736 0 500 84 50,400 E9(3)
EMU709 0 500 54 32,400 E10(3) EMU737 0 500 22 13,200 E12(2)
EMU710 0 500 54 32,400 E10(3) EMU738 0 500 22 13,200 E12(2)
EMU711 0 500 78 46,800 E7(2) EMU739 0 500 80 48,000 E13(1)
EMU712 0 500 78 46,800 E7(2) EMU740 0 500 80 48,000 E13(1)
EMU713 0 500 66 39,600 E11(3)

Several criteria can be used to evaluate the quality of assignment: (1) the
number of deadhead movement and/or replacement, (2) the number of MIs, (3) the
usage rate (in terms of MI), and (4) solution time. A well-planned assignment should
accommodate all utilization paths with no deadhead or replacement, minimum MI, the
maximum usage rate, and acceptable solution time.

5.1 Case I: Comparison among Manual, MIP, and Heuristic Methods

Case I is a smaller case that can be solved by both the exact optimization model
and the proposed heuristic method. This case primarily aims to ensure the solution
quality of the heuristic method compared with that of the exact optimization. In this
case, we select 20 EMU700 train-sets (EMU701–EMU720) to cover five types of
utilization paths, E7, E8, E10, E10*, and E13, which require 18 train-sets per day on
average. The planning horizon is set to 7 days. For the rolling horizon process, the
decision horizon for each optimization is 4 days, and the implementation horizon is 1
day. In other words, we implement this process seven times for the seven-day
19
planning. Each optimization for each day looks ahead for 4 days, and the optimal
result is implemented for the current day.
Both the MIP model and heuristic solution process for this case study are coded
in GAMS (GAMS, 2011) on a 3.10GHz PC with Intel i5 Processor and 12GB RAM.
This case study includes 21,843 constraints and 38,881 variables in the MIP model.
The manual result is based on an actual planner’s assignment, and the optimal results
are based on the MIP model solved by CPLEX and the heuristic method. The detailed
numerical results are shown in Table 10. All assignments incur no deadhead or
replacement. The average usage rate of the train-sets is 100%, which means all
train-sets going through MI at 90 cumulative days. This is the best scenario we can
obtain for this process since there is no gap between the MI inspection requirement
(i.e., 90 days), and the actual cumulative days of these train-sets right before
performing MI. Compared to results from manual assignment, the developed method
improves the usage rate by 4.05% based on the optimal results of the exact
optimization and heuristic models. The solution time for the exact model is
approximately 40 seconds, whereas the heuristic model requires less than 5 seconds.
These results show that the quality of the heuristic method is as good as the MIP
model (which has optimal results), but its solution efficiency is considerably better.

Table 10: Assignment results of case I


Models # of deadhead Number of Average MI Usage rate Improved Solution
& replacement MIs cumulative days effects time
Manual 0 2 86.49 days 96.10% - ~ 1 hr
Exact 0 1 90.00 days 100.00% 4.05% 42.4 sec.
Heuristics 0 1 90.00 days 100.00% 4.05% 4.7 sec.

5.2 Case II: Empirical Case at Hsinchu Depot

Case II is a real-size problem at Hsinchu depot with 55 train-sets and 12


utilization paths (Tables 8 and 9). The planning period is set to 90 days, the decision
horizon for each optimization is 7 days, and the implementation horizon is 1 day. A
90-day duration is the complete cycle of MI so we can evaluate the benefit of this
decision support tool compared with manual assignment.
With 206,933 constraints and 154,951 variables, this case is too complicated to
be solved by using CPLEX within reasonable time; therefore, the problem is solved
by the proposed heuristic method. Heuristics yields a feasible solution with a usage
rate of 99.8% in 11.2 seconds on a 3.10GHz PC with Intel i5 Processor and 12GB
RAM. Since the best possible solution is at the usage rate of 100% (i.e., all train-sets

20
going through MI at 90 cumulative days), a usage rate of 99.8% from the heuristic
method is highly efficient and very close the best scenario. The results based on the
heuristic method are compared with the manual assignments under the same initial
conditions at Hsinchu depot (Table 11). The number of MIs based on the heuristic
method is lower than that of the manual assignment. The efficiency of rolling stock
usage, represented by the average cumulative operating days and usage rate, increases
by about 5%. Heuristic method is able to reduce the solution time from 3 hours to
11.2 seconds, resulting in a substantial improvement on solution efficiency.

Table 11: Assignment results of case II


Models # of deadhead Number of Average MI Usage rate Improved Solution
& replacement MIs cumulative days effects time
Manual 0 37 85.77 days 95.30% - ~ 3 hrs
Heuristics 0 35 89.82 days 99.80% 4.98% 11.2 sec.

Fig. 5 demonstrates the cumulative days before MI for all train-sets at Hsinchu
depot at the beginning of the planning horizon. The heuristic method clearly
outperforms manual assignment as the cumulative days in heuristic assignment are all
close to the inspection limit, resulting in an average of 89.8 days. Because of myopic
consideration, planners in the manual process usually insert certain buffers in the
assignment so they can have a better opportunity (and flexibility) to identify feasible
assignment in the rolling horizon environment. By contrast, the heuristic method can
push the rolling stock to the limit because it is making decisions for a much longer
decision horizon. The method can also easily adjust and efficiently reassign train-sets
if the system has any disruptions and thus requires a new rolling stock assignment
plan.

Fig. 5: Cumulative days before MI

Based on this 90-day result, we can estimate that the annual savings in the
number of MIs from this proposed approach are eight times, equivalent to 5.41% of
annual saving in MI cost. More inspections indicate that more depot facilities and
human resources are required, whereas few rolling stock resources can be assigned
when dealing with either daily operational planning or emergency backup planning.
This result also demonstrates that a 7-day decision horizon is long enough to prevent
myopic decisions with a 99.8% usage rate, with 0.2% gap to the best scenario at 100%.
The verification of the resulting assignment plan by TRA experts demonstrates that
the proposed hybrid heuristic method is capable of solving large-scale practical

21
problems and the resulting assignment and maintenance plan are reasonable in
practice.

5.3 Case III: Stochastic Planning

Case II demonstrates the results of the manual assignment and the heuristic
method under the environment without unexpected accidents or failures. However,
unexpected events, such as rolling stock failures, do happen during actual operations.
Some of the failures can be remedied within a day and others may take more time to
recover. During rolling stock failure, planners have to make instant decisions to
replace defected rolling stock with spare ones, and come up with alternative
assignment plan right away to minimize possible delay, and inconveniences to the
passengers.
In this case study, we estimated the number and duration of possible rolling
stock failures in a 90-day planning horizon based on historical failure rate obtained
from TRA. We then randomly allocate these 18 failures in the 90-day horizon (Table
12). Each failure is associated with a train-set with specific date and duration in the
planning horizon.

Table 12: Failures in the 90-day planning horizon


Failures Train-sets Failure date Duration (days)
1 EMU602 6 1
2 EMU603 10 3
3 EMU735 14 2
4 EMU601 20 1
5 EMU717 25 3
6 EMU607 31 3
7 EMU713 35 3
8 EMU547 39 1
9 EMU707 45 3
10 EMU546 49 2
11 EMU711 55 1
12 EMU701 61 2
13 EMU542 65 1
14 EMU608 70 1
15 EMU544 75 1
16 EMU705 80 2
17 EMU739 86 2
18 EMU604 90 1

The results based on the heuristic method are compared with the manual

22
assignments. According to Table 13, the heuristic method yields a usage rate of
99.78% slightly less than the result from Case II; whereas the manual assignment
results in a usage rate of 94.33%, about 1% less than that from Case II. Due to the
unexpected failure, the number of replacement and deadhead have been significantly
increased compared to that from Case II. During this process, what manual
assignment can do is always to replace the failure rolling stock with the spare unit
first, and then update the complete assignment plan later because a complete
assignment plan would take few hours of manual work. On the other hand, the
developed heuristic method can incorporate the new condition (i.e., failures in this
case) and generate a new assignment plan less than a minute so it has more flexibility
in this process resulting in better performance in both usage rate and number of
replacements.

Table 13: Assignment results of case III


Models # of deadhead Number of Average MI Usage rate Improved Solution
& replacement MIs cumulative days effects time
Manual 13 38 84.90 days 94.33% - ~ 4 hrs
Heuristics 12 35 89.80 days 99.78% 5.45% 11.5 sec.

This case study also demonstrates the necessity to keep a short implementation
horizon, such as one day, in the rolling horizon scheme. As it can be seen in Table 12,
rolling stock failures happen quite often (every few days), and the length of the failure
duration is uncertain, depending on the condition of the rolling stock. For these
defected rolling stock, planners usually would not know the availability of the rolling
stock until the last minute (before the actual assignment being implemented). Since
the condition of the rolling stock changes dynamically, a short implementation
horizon can maintain the highest flexibility of this process.
Moreover, the proposed heuristic method can be used to pre-generate multiple
assignment plans according to the possibility of the rolling stock availability
beforehand. With these pre-generated plans, the actual implementation can choose the
best one based on the availability of the rolling stock right before the actual
assignment. This is another important benefit to automate and improve this process
with great solution efficiency.

5.4 Sensitivity Analysis

A sensitivity analysis was performed to evaluate the effects of the variation in


three major parameters in the rolling stock assignment process: (1) the required buffer
23
time during connection between consecutive assignments for the same rolling stock;
(2) the allowable cumulative days before MI according to the MI inspection
requirement; and (3) the available MI capacity at the depot.
A buffer time between consecutive assignments is necessary for rolling stock
turnaround, clean up, and DI. The shorter the required buffer time the more the
flexibility in rolling stock assignment since train-sets would be available sooner. This
can be done by increasing number of inspectors and/or upgrade the inspection
technology. The current required buffer time at Hsinchu is 2 hours, and we evaluate
the impact of reducing this buffer time to 1 hour or increasing it to 3 hours in this
analysis.
The number of the allowable cumulative days before MI is also a very important
parameter in this planning process. The higher the value the more flexibility in the
process and the fewer the number of MIs; on the contrary, a lower allowable MI
cumulative days requires more inspections so as to reduce the availability of rolling
stock. In this analysis, we test the effect of setting the allowable cumulative days to 80
days (10 days less than the current value), and 100 days (10 days more than the
current value).
MI capacity at the depot can be adjusted by reallocating human and
infrastructure resources. The results from Case II show that the highest number of MI
required for a day is around 4 MIs, considerably lower than the available MI capacity
at Hsinchu (i.e., 8 trains per day). Therefore, we examine the scenarios with only 2 or
4 available MI capacity.
According to the variation in the three important parameters, eight scenarios
were created in this sensitivity analysis (Table 14). For each scenario, we compare the
performance in usage rate under manual and the proposed heuristic processes. In
Table 14, we can see the lower the required buffer time, the better the usage rate for
both manual and heuristic processes due to the flexibility. Variation in allowable MI
cumulative days has significant impact to the efficiency of rolling stock usage,
especially for manual assignments, with 1.5~2% difference in usage rate. Finally, the
impact of reduction in MI capacity is not as sensitive as the other two parameters.
Adopting the developed heuristic method in this process further reduces the difference
in less than 0.05%. Therefore, TRA can consider to reallocate or reduce the human
and infrastructure resources in MI capacity.

24
Table 14: Sensitivity analysis on the effects of the variation in required buffer time,
allowable MI cumulative day, and available MI capacity
Scenario Buffer time Allowable MI MI capacity Usage rate Improved
(hrs) cumulative day (days) (trains) effects
Manual Heuristics
1 1 80 2 93.05% 99.71% 6.66%
2 1 80 4 93.09% 99.73% 6.64%
3 1 100 2 94.75% 99.79% 5.04%
4 1 100 4 95.00% 99.80% 4.80%
5 3 80 2 93.00% 99.67% 6.67%
6 3 80 4 93.03% 99.71% 6.68%
7 3 100 2 94.67% 99.79% 5.12%
8 3 100 4 94.84% 99.80% 4.96%

The developed heuristic process performed very well in all scenarios. The best
results (99.80%) are from scenarios with higher flexibility in the allowable MI
cumulative day and MI capacity (scenarios 4 and 8). The worst result (99.67%) is
from the most constraint scenario (Scenario 5) with the long required buffer time, and
the low allowable MI cumulative day and capacity. The difference in usage rate
between the best and the worst scenarios obtained from the heuristic method is only
0.13%. Compared to the manual process, the proposed heuristic method provides a
clear advantage with improvement rate ranging from 4.8% to 6.68%. Using this tool is
highly beneficial in increasing the efficiency of rolling stock usage.

6. Conclusion

This research develops an exact optimization model to improve the efficiency in


rolling stock usage with consideration of all necessary regulations and practical
constraints. Compared to the manual process, a hybrid heuristic process is also
developed to improve solution quality and efficiency. Empirical results demonstrate
that the heuristic process can successfully increase the efficiency of rolling stock use
by about 5% and significantly reduce the solution time from 3 hours to 11.2 seconds.
Using this decision support tool can help railways with similar characteristics to
improve the efficiency in rolling stock usage and productivity in rolling stock
assignment process. What have been done in this research is to optimize rolling stock
assignment and maintenance plan for DI and MI in the daily operational level.
Future research may investigate the possibility to optimize the maintenance plan for
BI and GI.

25
References

Abbink, E., Van der Berg, B., Kroon, L., & Salomon, M. (2004). Allocation of
Railway Rolling Stock for Passenger Trains. Transportation Science, 38, 33-41.
Alfieri, A., Groot, R., Kroon, L., & Schrijver, A. (2006). Efficient Circulation of
Railway Rolling Stock. Transportation Science, 40(3), 378-91.
Barnhart, C., Boland, N. L., Clarke, L. W., Johnson, E. L., Nemhauser, G. L., &
Shenoi, R. G. (1998). Flight String Models for Aircraft Fleeting and Routing.
Transportation Science, 32(3), 208-220.
Boland, N. L., Clarke, L. W., & Nemhauser, G. L. (2000). Asymmetric Traveling
Salesman Problem with Replenishment Arcs. European Journal of Operational
Research, 123(2), 408-427.
Brucker, P., Hurink, J. L., & Rolfes, T. (2003). Routing of Railway Carriages. Journal
of Global Optimization, 27, 313-332.
Cacchiani, V., Caprara, A., & Toth, P. (2010). Solving a Real-world Train-unit
Assignment Problem. Mathematic Programming, 124, 207-231.
Cadarso, L., & Marín, Á. (2010). Robust Routing of Rapid Transit Rolling Stock.
Public Transport, 2, 51-68.
Cadarso, L., & Marín, Á. (2011). Robust Rolling Stock in Rapid Transit Networks.
Computers & Operations Research, 38, 1131-1142.
Caprara, A., Kroon, L., Monaci, M., Peeters, M., & Toth, P. (2007). Passenger
Railway Optimization. In Barnhart C, Laporte G, (Eds.), Handbooks in Operations
Research and Management Science (pp. 129-87). Netherlands: Elsevier.
Cheng, Y. H. (2010). High-Speed Rail in Taiwan: New Experience and Issues for
Future Development. Transport Policy, 17(2), 51-63.
Clarke, L., Johnson, E., Nemhauser, G., & Zhu, Z. (1997). The Aircraft Rotation
Problem Source. Annals of Operations Research, 69, 33-46.
Fioole, P. J., Kroon, L., Maróti, G., & Schrijver, A. (2006). A Rolling Stock
Circulation Model for Combining and Splitting of Passenger Trains. European
Journal of Operational Research, 174, 1281-1297.
Giacco, G., D'Ariano, A., & Pacciarelli, D. (2011). Rolling Stock Rostering
Optimization under Maintenance Constraints. Proceedings of 2nd International
Conference on Models and Technologies for Intelligent Transportation Systems.
Leuven, Belgium.
Giacco, G., D'Ariano, A., & Pacciarelli, D. (2014). Rolling Stock Rostering
Optimization under Maintenance Constraints. Journal of Intelligent Transportation
Systems: Technology, Planning, and Operation, 18(1), 95-105.
Gopalan, R., & Talluri, K. T. (1998). The Aircraft Maintenance Routing Problem.
Operations Research, 46(2), 260-271.
26
Jünger, M., Elf, M., & Kaibel, V. (2003). Rotation Planning for the Continental
Service of a European Airline. Mathematics — Key Technology for the Future,
675-689.
Mak, V., & Boland, N. (2000). Heuristic Approaches to the Asymmetric Travelling
Salesman Problem with Replenishment Arcs. International Transactions in
Operational Research, 7, 431-447.
Maróti, G., & Kroon, L. (2005). Maintenance Routing for Train Units: the Transition
Model. Transportation Science, 39, 518-525.
Maróti, G., & Kroon, L. (2007). Maintenance Routing for Train Units: the Interchange
Model. Computers & Operations Research, 34, 1121-1140.
Peeters, M., & Kroon, L. (2008). Circulation of Railway Rolling Stock: a
Branch-and-price Approach. Computers & Operations Research, 35, 538-556.
Schrijver, A. (1993). Minimum Circulation of Railway Stock. CWI Quarterly. 3:
205-217.
Sarac, A., Batta, R., & Rump, C. M. (2006). A Branch-and-price Approach for
Operational Aircraft Maintenance Routing. European Journal of Operational
Research, 175(3), 1850-1869.
Sriram, C., & Haghani, A. (2003). An Optimization Model for Aircraft Maintenance
Scheduling and Re-assignment. Transportation Research Part A: Policy and Practice,
37(1), 29-48.
Talluri, K. T. (1998). The Four-Day Aircraft Maintenance Routing Problem.
Transportation Science, 32(1), 43-53.
Xie, M., Mao, B., Ho, T., Chen, Z., & Bai, Y. (2010). Modeling Circulation of
Train-set with Multiple Routing. Journal of Transportation Systems Engineering and
Information Technology, 10, 50-57.
Zhao, P., Tomii, N., Fukumura, N., & Sakahuchi, T. (2002). An Algorithm for
Train-set Scheduling Based on Probabilistic Local Search. Computers in Railways
VIII, 817-826.
Zhu, C., Liu, D., Cheng L., & Li, H. (2013). Study on Optimization of Railway
Passenger Train Sets Assignment. Information Technology Journal, 12(6), 1251-1256.

27
Figure(s)

Service Demand

Timetabling
Operational
Plan
Rolling Stock Crew
Scheduling Scheduling

Rolling Stock Crew


Assignment Rostering
Figure(s)

MI Cumulative Days (day)


0 10 20 30 40 50 60 70 80 90
EMU542
EMU544
EMU546
EMU601
EMU603
EMU605
EMU607
EMU609
EMU702
EMU704
EMU706
EMU708
EMU710
Heuristic
Train-sets

EMU712 Manual
EMU714
EMU716
EMU718
EMU720
MI
EMU722 Requirement
(90days)
EMU724
EMU726
EMU728
EMU730
EMU732
EMU734
EMU736
EMU738
EMU740
Acknowledgements
The authors are grateful to the editor and the anonymous referees for their
constructive comments and suggestions. This research was sponsored by the Ministry
of Science and Technology of Taiwan, under the grant: NSC 101-2221-E-002-157.

28
Highlights:
 Assignment between rolling stock and daily utilization schedule is optimized.

 Practical constraints in operations and multi-level maintenance are considered.

 A heuristic process is proposed to significantly improve solution efficiency.

 Our optimization process increases utilization of rolling stock by around 5%.

 The empirical study shows that monthly maintenance cost reduces by 5.41%.

29

Potrebbero piacerti anche