Transportation Research Part C: Xin Wu, Lei Nie, Meng Xu

Transportation Research Part C 74 (2017) 212244
Contents lists available at ScienceDirect
Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc
Designing an integrated distribution system for catering

services for high-speed railways: A three-echelon location
routing model with tight time windows and time deadlines q
Xin Wu a,b, Lei Nie a,b,, Meng Xu b
a
b
School of Traffic and Transportation, Beijing Jiaotong University, 100044, PR China

State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing 100044, PR China
a r t i c l e
i n f o
Article history:
Received 30 January 2016
Received in revised form 26 October 2016
Accepted 8 November 2016
Available online 30 November 2016
Keywords:
Catering service
High-speed rail
Public railway plan
Location routing model
Cross entropy algorithm
Time-oriented neighbor heuristic
a b s t r a c t
An emerging task in catering services for high-speed railways (CSHR) is to design a distribution system for the delivery of high-quality perishable food products to trains in need.
This paper proposes a novel model for integrating location decision making with daily rail
catering operations, which are affected by various aspects of rail planning, to meet timesensitive passenger demands. A three-echelon location routing problem with time windows and time budget constraints (3E-LRPTWTBC) is thus proposed toward formulating
this integrated distribution system design problem. This model attempts to determine
the capacities/locations of distribution centers and to optimize the number of meals delivered to stations. The model also attempts to generate a schedule for refrigerated cars traveling from distribution centers to rail stations for train loading whereby meals can be
catered to trains within tight time windows and sold before a specified time deadline.
By relaxing the time-window constraints, a relaxation model that can be solved using an
off-the-shelf mixed integer programming (MIP) solver is obtained to provide a lower bound
on the 3E-LRPTWTBC. A hybrid cross entropy algorithm (HCEA) is proposed to solve the 3ELRPTWTBC. A small-scale case study is implemented, which reveals a 9.3% gap between the
solution obtained using the HCEA and that obtained using the relaxation model (RM). A
comparative analysis of the HCEA and an exhaustive enumeration algorithm indicates that
the HCEA shows good performance in terms of computation time. Finally, a case study considering 156 trains on the Beijing-Shanghai high-speed corridor and a large-scale case
study considering 1130 trains on the Chinese railway network are addressed in a comprehensive study to demonstrate the applicability of the proposed models and algorithm.
2016 Elsevier Ltd. All rights reserved.
1. Introduction
Many supply chains, from those providing medical blood, which must be consumed while in good condition (Shen et al.,
2003; Shu et al., 2005), to perishable food products with limited lifespans (Federgruen et al., 1986; Zhang et al., 2003), are
currently operated in time-critical modes in which tasks must be executed within a tight time frame. In regard to the delivery of perishable food products, whose quality deteriorates when they are subjected to extended travel times and frequent
q
This article belongs to the Virtual Special Issue on Integrated optimization models and algorithms in rail planning and control.
Corresponding author at: School of Traffic and Transportation, Beijing Jiaotong University, 100044, PR China.
E-mail addresses: lnie@bjtu.edu.cn, lnie8509@yahoo.com (L. Nie).
http://dx.doi.org/10.1016/j.trc.2016.11.006
0968-090X/ 2016 Elsevier Ltd. All rights reserved.
X. Wu et al. / Transportation Research Part C 74 (2017) 212244
213
stops (Hsu et al., 2007), it is critical to design a distribution system that considers time constraints to maintain the food supply chain (FSC) (Nagurney et al., 2013).
In China, with the emerging development and construction of high-speed passenger lines, the China Railway Corporation
is increasingly focusing on developing a distribution system for rail catering operations. It is estimated that in 2014, China
Railway catered more than 1000 trips per day and purchased/produced approximately 171 million meals (Science and
Technology Division of Chinas Ministry of Railways, 2010). The task of providing catering services for high-speed railways
(CSHR) is thus an important component of passenger rail transport services (Wu et al., 2015). Furthermore, in China, food
products for CSHR are currently divided into two classes: (1) ambient meals, with a lifespan of 6 months, and (2) cold chain
meals, with a lifespan of 24 h. Meals of both types should receive microwave heating before being sold. Generally, cold
chain meals taste better and satisfy passengers needs better than do ambient meals. However, cold chain meals can spoil
if the appropriate time deadlines are not met. To increase the level of service provided to train passengers and to improve the
competitiveness of high-speed trains on the intermodal transportation marketplace, it is logical to begin to realize a timecritical cold food chain by searching for a well-designed distribution system; in this paper, the related problem is termed the
distribution system design problem for CSHR (DSDP-CSHR).
The systematic integration of different optimized operation decisions is important for providing better passenger train
services. Two types of integrated problems are important in this field. One is the integration of planning for the transitional
processes that occur throughout the entire rail transport plan (RTP), such as line planning, timetabling and rolling stock planning (Goossens et al., 2006). The other is the integration of external support systems (e.g., a distribution system for catering
services, as considered in this paper) with the components of the transportation system to better meet various passenger
demands. These external support systems are closely related to the passenger service performance of a transportation system. For example, Ho and Leung (2010) demonstrated the importance of airline catering services for the Hong Kong International Airport and indicated that flight catering is strongly affected by the utilized flight plan. Tong et al. (2015b) discussed
train trip package transportation that is realized by a specialized train. In the field of road transportation, Ruan et al. (2016)
explained how to systematically optimize the locations of park-and-ride stations, the number of parking lots, and the schedules for bus rapid transit. The DSDP-CSHR is also an integrated optimization problem of the latter type. Compared with airline catering, the service locations for rail catering pose a more complex problem. The meals required for a flight are usually
catered at its origin port (Goto et al., 2004); by contrast, in rail catering, a train may be catered either at the origin of a trip or
at stations along the trains route. Hence, the locations of distribution centers (DCs) also constitute a prominent issue in
CSHR. To the best of our knowledge, few studies have addressed these topics in the context of CSHR.
The planning for the distribution of perishable meals throughout a railway network should integrate the problem of distribution system design with the consideration of daily rail catering operations as affected by the RTP to ensure that timesensitive passenger demands are met. An RTP consists of three components: a line plan, a train timetable and an electric multiple unit (EMU) circulation plan. A line is defined as a path between an origin and a destination along a given route on a
railway network. A line plan specifies the frequency of the line services and their halting patterns, which define the rail stations along a lines route at which a train serving that line dwells (Bussieck, 1998; Goossens et al., 2006; Fu et al., 2015). A
timetable specifies the arrival and departure times for each line in the form of a series of trips (Peeters, 2003; Zhang and Nie,
2016). An EMU circulation plan assigns trains to each trip (i.e., train path) in the timetable (Peeters and Kroon, 2008). The RTP
gives rise to special characteristics of CSHR. Compared with a normal distribution system design problem, the time-critical
aspect of the DSDP-CSHR is that the longest lead time from a food supplier to any accessible trips destination is subject to
time-deadline restrictions that limit the length of the routes. Furthermore, the meals must eventually be stowed on the
trains within tight time windows as imposed by the RTP.
The purpose of this paper is to introduce a location routing model for integrating location-allocation decisions with daily
catering operations in the DSDP-CSHR, whereby meals can be delivered to and stowed on trains within tight time windows
and sold before a given time deadline. Based on this model, we can analyze the impacts of optimization on passenger service
performance.
1.1. Previous related studies
Managerial decision making is classified into three hierarchical levels: strategic, tactical and operational. In past studies,
the DSDP and other network design problems have been investigated as problems of strategic decision making (Geoffrion,
1974; Gelders et al., 1987; Pooley, 1994; Wouda et al., 2002; Zhang et al., 2003; Liu and Zhou, 2016). Most such studies consider only the determination of location decisions through the application of a mixed integer programming (MIP) model.
Their objective functions minimize the total cost for opening alternative facilities and shipping products (Melo et al.,
2009). Relevant studies mainly address facility location/network design problems (FLNDPs), which were initially introduced
as generalizations of classical facility location problems (FLPs). Klose and Drexl (2005) provided an excellent review of the
various types of FLNDP models. Adding budget constraints to the p-center problem gives rise to the center FLNDP with budget constraints (CFLNDB). The goal of the CFLNDB is to minimize the maximum travel time from any customer to that customers allocated facility rather than to minimize the overall cost, as in typical FLNDPs (Contreras et al., 2012).
Several authors have studied location-allocation problems for specific FSC problems (Akkerman et al., 2010). Geoffrion
(1974) described a DSDP using an MIP model to determine the locations of intermediate facilities between plants and customers for a food firm. Pooley (1994) and Wouda et al. (2002) studied a DSDP for a case study of a dairy industry. Another
214
study of a beer industry was also conducted by Gelders et al. (1987). Zhang et al. (2003) discussed a distribution system for
frozen and chilled food products with time-dependent quality degradation in facilities and vehicles. However, no studies
have focused on the DSDP in the context of CSHR.
The vehicle routing problem (VRP) represents operational decision making in logistic management. This problem is
defined as the determination of an optimal set of routes for a fleet of vehicles that serve customers with known demands
from a depot (Chen et al., 2009; Mahmoudi and Zhou, 2016). Toth and Vigo (2002) reviewed several VRP variants, including
versions with time windows, multiple vehicles and backhauls. Moreover, the VRP with time windows (VRPTW) has been closely investigated because of its applicability to real-life situations (Tas et al., 2014). The VRPTW has been applied to a wide
variety of food industries, e.g., in the fields of catering (Chung and Norback, 1991), dairy (Adenso-Daz et al., 1998) and meat
production (Belenguer et al., 2005). In particular, Ho and Leung (2010) studied manpower scheduling for airline catering
using the VRPTW. Many heuristics have also been applied to the hard problems derived from the VRP (Prins et al.,
2014; Montoya et al., 2015). To the best of our knowledge, however, VRP variants concerning daily rail catering operations
have yet to be studied.
Many studies have integrated decision making at different levels to obtain more reasonable solutions (Perl and Daskin,
1985; Miranda and Garrido, 2004, 2008; Verma et al., 2012; Meisel et al., 2013). The locationrouting problem (LRP) arises
from the integration of location and routing decisions (Berger et al., 2007). The LRP is defined as a special case of the VRP
with the intent of simultaneously determining the optimal numbers and locations of DCs while specifying the vehicle routes
from depots to customers (Ahmadi-Javid and Seddighi, 2013). To date, LRP research has attracted less attention. However,
the LRP is important because of its real-world applications (Govindan et al., 2014). For example, Watson-Gandy and
Dohrn (1973) used an LRP model to address various issues in food distribution. Berger et al. (2007) demonstrated that
the LRP with distance constraints (LRPDC) is especially well suited to location problems for perishable goods and other
time-critical deliveries. The LRP with time windows (LRPTW) is another extended version of the LRP that is applicable in
the field of perishable food delivery (Govindan et al., 2014). However, the LRP has yet to be studied in the context of CSHR,
and few studies have focused on the consideration of both time windows and deadlines in the LRP model.
The LRP is an NP-hard problem and thus is difficult to solve using exact algorithms (Perl and Daskin, 1985). Heuristic
approaches generally become more viable alternatives as the problem size increases (Nagy and Salhi, 2007). Bruns et al.
(2002) concluded that the available heuristic methods for solving the LRP can be classified into three types: sequential methods, iterative methods and approximation methods. Nagy and Salhi (2007) reviewed a series of applications of the LRP using
these three types of methods and concluded that most practical studies use heuristic approaches to solve large-scale cases.
1.2. Contributions and objectives

The DSDP-CSHR, which is the focus of this paper, can be considered to be a routing problem with location-allocation decisions, time windows and time budget constraints. Because all three components of the RTP involve mid-term decisions, with
a planning horizon ranging from one month up to one year, another challenge in CSHR is that the capacities and location of
the facilities cannot be changed as readily as can the RTP. Hence, in this problem, we adopt the following fundamental
assumptions (see Fig. 1).
The trips, characterized by train numbers in a train timetable, are served as the customers in our distribution system. The
timetable may vary over time; thus, we assume that the cold chain meals demanded over a given planning horizon are
outsourced to contracted food suppliers (FSs). The operators of the DCs select appropriate cities for their locations from
among potential sites and rent sufficient capacities at these DCs for storage, rather than building expensive plants for food
production. Then, the operators order the demanded meals from cooperative FSs and deliver the meals to rail stations using
refrigerated cars (RCs). The CSHR process thus begins at the FSs, proceeds to DCs in various cities, and ends with catering to
high-speed trains at stations. The process integrates the following four layers with three echelons: food suppliers (FSs), distribution centers (DCs), rail stations and high-speed trains.
The establishment of such a distribution system faces several considerable challenges, such as (i) the organizational effort
required to coordinate the activities on the various layers (three echelons) to restrict the lead times from each food supplier
(FS) to the destinations of its accessible trips to remain within a given time deadline, (ii) the organizational effort required to
Fig. 1. The catering service process in the supply chain for CSHR.
215
deliver the meals to the trains within the time windows imposed by the arrival and departure times of trips at stations based
on the train timetable and EMU circulation plan, and (iii) the need to simultaneously optimize the number, locations and
capacities of facilities and the routes of the RCs to integrate the spatial consolidation of meals (i.e., where meals are produced
and collected) with their temporal consolidation (i.e., when meals are collected to facilitate rail catering) (Meisel et al., 2013).
To obtain a deeper understanding of the DSDP-CSHR, this paper investigates this problem from the following perspectives:
(1) To formulate the problem, we define two correlated networks: a physical distribution network (PDN) and an operational distribution network (ODN). The PDN, which is a collection of network elements, including nodes and links, is
generated based on the restrictions imposed by the RTP (Section 2.2). The ODN, which is derived based on the PDN,
explicitly reflects the rail catering process and is used as the basis on which the operators of the DCs determine a set of
RC routes to serve trips with known demands (Section 2.3).
(2) The time deadlines that are imposed to guarantee food safety are reformulated as series of time budget constraints on
a rooted forest in our proposed model (Section 2.4).
(3) To integrate the multiple layers of the problem, a three-echelon location-routing problem with time windows and
time budget constraints (3E-LRPTWTBC) is introduced (Section 3.1). By linearizing and relaxing certain constraints
in the 3E-LRPTWTBC, a relaxation model (RM) is obtained to provide a lower bound as a measure of the solution quality for the 3E-LRPTWTBC (Section 3.2).
(4) A hybrid cross entropy algorithm (HCEA) embedded with a time-oriented nearest-neighbor heuristic (TNNH) is presented for solving the 3E-LRPTWTBC (Section 4).
(5) A comprehensive study based on two real train timetables from China is also implemented (a case consisting of 156
trips along the Beijing-Shanghai corridor and a case consisting of 1130 trips throughout the entire rail network) to
demonstrate the performance of the models and algorithm (Section 5).
Table 1 summarizes our contributions in comparison with several other problems addressed in the literature.
2. Problem statement with network representations

In an FSC, the catering process is conventionally divided into two stages. In the first echelon, meals are prepared in the
kitchens of food suppliers (FSs). Distribution centers (DCs) order the meals, and the food products are shipped from the FSs to
the DCs via line haul. In the second echelon, the meals are packed into refrigerated cars (RCs) to serve retail entities with in
the range of each DC. As illustrated in Fig. 2, in the DSDP-CSHR, the retail entities are rail stations on the rail network.
However, this conventional system cannot adequately describe the DSDP-CSHR for the following reasons.
(1) The demands of the stations are unpredictable when information about the line plan is lacking. A given trip needs to
be catered at only one of the stations it passes according to the line plan. A trip passing a station does not necessarily
indicate the existence of a demand at the station.
(2) The number of meals required for a trip is influenced by a variety of time factors in the train timetable. A long-distance
trip usually creates greater demand than does a short trip. In addition, a trip that has a greater overlap with a meal
time creates a greater food demand.
(3) For each trip, the time required for the train to travel from one of its stations to its destination is important for guaranteeing food safety. Based on this time, we are able to judge whether any of the travel times from an FS to the destinations of its accessible trips exceeds the foods lifespan.
(4) Meals must be delivered to trains at platforms within the time windows imposed by the train timetable and the EMU
circulation plan. These time windows greatly impact the transportation cost; thus, it might be misleading to neglect
them.
Table 1
Overview of important related studies.
Paper
Zhang et al.
(2003)
Berger et al.
(2007)
Belenguer et al.
(2005)
Govindan et al.
(2014)
Our paper
Model
Algorithm
FSC
Tabu search
heuristic
LRPDC
Branch and
price
VRPTW
Tabu search
heuristic
2E-VRPTW
Hybrid genetic
algorithm
3E-LRPTWTBC
Hybrid cross entropy
algorithm
Number of echelons
Considers location decisions
Considers vehicle routing
Considers deadline restrictions
Considers time-window
constraints
Considers restrictions imposed by
the RTP
1
p
p
p
1
p
2
p
p
3
p
p
p
p
p
216
Fig. 2. Illustration of the food distribution system for CSHR.
As a result, in the DSDP-CSHR, the classical distribution system should be extended. First, a third echelon is added to the
conventional system to represent the task of station-to-trip assignment based on the information given by the line plan and
the timetable (see the rightmost part of Fig. 2). This extended network is called the physical distribution network (PDN). Second, a network solution is obtained based on the PDN, and an operational distribution network (ODN) is derived to ensure
that the graph abides by the structure of the LRP model. Third, the time-deadline constraints are reformulated as time budget
constraints to limit the maximum accumulated travel time (MAT) on any path to a trips destination from its root supplier.
2.1. Notation and assumptions

We now begin to formally define the DSDP-CSHR as follows. The distribution system for the CSHR begins at the FSs, proceeds to the DCs and railway stations, and ends at the high-speed trains in the rail network. The goal of the problem studied
in this paper is to make joint decisions regarding the locations of the DCs and the routes of the RCs to satisfy the demands of
the trips specified in a train timetable while minimizing the total setup and transportation cost, subject to the special
requirements imposed by the RTP, capacity constraints, deadline restrictions, etc. The relevant notation used in this paper
is defined in Tables 2 and 3. Table 2 does not include all of the notation for the sets used in this paper, such as the different
sets of links in the PDN, the different sets of arcs in the ODN, or the sets of nodes and edges in the rooted forest. All other sets
are defined with respect to the fundamental sets listed in Table 2 and are discussed in detail in Sections 2.22.4. Table 3 also
does not include all of the notation for the parameters used in our proposed algorithm, which is discussed later in Section 4.
In the DSDP-CSHR, an RTP for operation is specified in advance. The food demands for each trip are assumed. For each
potential DC, the opening, expanding, and closing costs and the maximum holding time are given. The handling cost at each
DC is combined with its opening cost for simplicity. The transportation cost and travel time for each link of the PDN are
known. In this paper, the stretch of each period is one day, and the costs are all daily costs. Because line-haul and urban
transportation represent much higher costs and much longer times compared with the transportation costs and loading
times at stations, in this paper, we neglect the cost/time of the loading and unloading processes at rail stations. For simplicity, the problems of trash and the collection/transportation of unsold meals are also neglected in our study.
Strategically, the goals of the DSDP-CSHR are to (i) determine the locations of new DCs from among a set of candidate DCs
(i.e., N D ), (ii) establish a series of links between facilities and (iii) determine the amounts of food products transported
between the FSs in set N P and the DCs in set N D . Operationally, the goals of the DSDP-CSHR are to (iv) determine the routes
to be followed by the RCs between each d in N D and each s in N S d, where N S d is the set of stations corresponding to distribution center d and N S [d2N D N S d. The decisions listed above must be made under the following restrictions: (i) the DCs,
FSs and RCs have limited capacity; (ii) the cold chain meals delivered for each trip cannot exceed their defined lifespan
before the train serving the trip arrives at the trips destination station; and (iii) the food products should be delivered to
the stations within the time windows imposed by the RTP.
217
Table 2
Sets and indexes used in formulating the DSDP-CSHR.
Sets and indexes
Descriptions
Sets in the PDN
N NP [ ND [ NS [ NT
L
NP
ND
NS
N S d
NT
The
The
The
The
The
The
The
set
set
set
set
set
set
set
of
of
of
of
of
of
of
nodes in the PDN

links in the PDN
food suppliers (FSs) (the first layer)
distribution centers (DCs) (the second layer)
stations (the third layer)
stations served by DC d in a feasible network solution
trips that should be satisfied (the fourth layer)
Sets in the ODN
No NP [ ND [ NW [ NT
A
NW
N T s
V
Vd
The
The
The
The
The
The
set
set
set
set
set
set
of
of
of
of
of
of
nodes in the ODN

links in the ODN
dummy nodes corresponding to DCs in the ODN
trip nodes that are catered at station s in the ODN
RCs
RCs that are available from DC d, where V [d2N D Vd
Indexes used in both the PDN and ODN
i; j
p; p0 ; p1 ; p2 ; p3 . . .
0
d; d ; d1 ; d2 ; d3 . . .
s; s0 ; s1 ; s2 ; s3 . . .
t; t0 ; t 1 ; t 2 ; t3 . . .
The
The
The
The
The
general node indexes in both the PDN and ODN

indexes for FSs in both the PDN and ODN
indexes for DCs in both the PDN and ODN
indexes for stations in both the PDN and ODN
indexes for trips that should be satisfied in both the PDN and ODN
Indexes used only in the ODN
w; w0 ; w1 ; w2 ; w3 . . .
0
wd; wd ; wd1 ; wd2 . . .
v; v0; v1; v2; v3 . . .
D j
D j
The indexes for dummy nodes in the ODN

0
The indexes for dummy nodes corresponding to DCs d, d , d1 , d2 ; . . . in the ODN
The indexes for RCs in the ODN.
The set of nodes with head j in the ODN
The set of nodes with tail j in the ODN
Table 3
Given parameters of the DSDP-CSHR.
Parameters
Descriptions
General parameters
Lspa
TT i;j , TC i;j
tti;j , tci;j
The lifespan of the cold chain meals (min)

The travel time/transportation cost from node i to node j in the PDN (min)
The travel time/transportation cost from node i to node j in the ODN (yuan/vehicle)
Parameters for FSs
Capp
The capacity of FS p (units)

The capability of RCs from FS p (units)
Cp
Parameters for DCs
HT d
The
The
The
The
The
The
ds; t
Es; t
Ls; t
Js; t
DWT s;t
MST s
Dt
If ds; t 1, trip t can be catered at station s; otherwise, it cannot be catered at station s

The earliest start time for train loading at station s for trip t (min)
The latest time at which train loading can end at station s for trip t (min)
The journey time from station s to the destination for trip t (min)
The dwell time at station s for trip t (min)
The maximum storage time permitted at station s (min)
The demand that must be satisfied for trip t (units)
OC d
UC d
UCapd
Kd
Cd
Parameters for stations and trips
opening cost of a potential DC d (yuan)

cost to rent one unit of capacity at DC d (yuan)
capacity for units of storage at DC d (units)
maximum number of units that can be rented at DC d
capacity of RCs originating from DC d (units)
maximum holding time in the warehouse at DC d (min)
The strategic decisions are made based on the PDN, as discussed in Section 2.2. Then, the operational decisions are made
based on the ODN, as discussed in Section 2.3. The decision variables considered in the DSDP-CSHR are listed in Tables 4 and
5. The other relevant assumptions adopted in this paper are listed in Table 6. Regarding these assumptions, we would like to
offer the following three remarks to support the claim that they will not unduly hinder the applicability of our study to realistic scenarios.
Remark 1. Assumptions (A1.a)(A1.e) define a limited structure for the distribution system. For example, A1.b requires that
only one FS can supply each DC (i.e., each DC signs a contract with only one FS). Furthermore, A1.e requires that each trip be
catered only once per day. There are both practical and theoretical reasons for assuming such a limited system. Practically,
the assumptions ensure that all meals for a given trip are produced and provided by only one FS and thus are subject to the
same production processes and standards. This is usually required for safe and convenient CSHR management. Furthermore,
because FSs are dispersed distantly on a large railway network, it is usually practical that a DC only orders its required meals
from a nearest FS. Theoretically, assumptions (A1.a)(A1.e) are essential for formulating the time budget constraints in our
model to limit the lead time to within a given time deadline, especially when the number of path flows in the network is
unknown during the design phase (see Section 2.4 for details).
218
Table 4
Strategic decision variables to be optimized.
Decision variables
Descriptions
zd
ypd
qds
xst
kd
f pd
Binary variable; =1 if DC d 2 N D is established, =0 otherwise

Binary variable; =1 if DC d 2 N D orders its meals from FS p 2 N P , =0 otherwise
Binary variable; =1 if DC d 2 N D serves station s 2 N S , =0 otherwise
Binary variable; =1 if trip t 2 N T is catered at station s 2 N S , =0 otherwise
The number of units of capacity rented at DC d 2 N D
The amount of food products delivered from FS p to DC d, where p 2 N P and d 2 N D . For all p 2 N P and d 2 N D , f pd 2 R [ f0g
Table 5
Operational decision variables to be optimized.
Decision variables
Descriptions
r dijv
dv
i
Binary variable; =1 if RC v originating from DC d traverses link i; j 2 A n APD , where v 2 Vd and d 2 N D , =0 otherwise
The start time of service by RC v from DC d at node i 2 N D [ N W [ N T , where v 2 Vd and d 2 N D . For all i, sdi v 2 R [ f0g
Table 6
Relevant assumptions adopted in this paper.
A1. Assumptions regarding FSs, DCs, rail stations and trips
(A1.a) FSs cannot directly send food products to stations. Furthermore, the DCs satisfy only the stations demands, and the stations satisfy only the
trips demands
(A1.b) The meals that are required over a planning horizon are all outsourced to FSs. A DC can purchase its food products from any contracted FS. Each
DC selects only one FS. The capacity of all FSs is sufficient to satisfy all trains in the network
(A1.c) There is at most one DC in a city (thus, a DC is synonymous with a city). Each DC can provide food products only to its corresponding rail
stations in the city (e.g., the DC in Beijing can serve Beijing South, Beijing North, Beijing West and the main Beijing station)
(A1.d) The trips specified in the train timetable are treated as the customers in our distribution system. Each trip is required to be served only once, at
its origin station or at an intermediate halting station. A single type of cold chain meal is sold on the trains, and all trains have sufficient cold
storage capacity to store the delivered cold chain meals
(A1.e) The one-time catering for each trip must be performed before the corresponding meal time. If the trip overlaps only with the meal time for
lunch, it must be catered before lunchtime. Similarly, if the trip overlaps only with the meal time for supper, it must be catered before suppertime.
Finally, if the journey time of the trip overlaps with both lunchtime and suppertime, then it must be catered before lunchtime. If the train departs
from its origin station during lunchtime or suppertime, then the train must be catered at its origin station
A2. Assumptions regarding the vehicle routing problem for RCs
(A2.a) We consider only the routing problem for urban areas between the DCs and their corresponding stations
(A2.b) Each RC originating from an FS delivers all meals to the active DCs and directly returns to the same FS
(A2.c) Each RCs route should begin and end at an active DC for a fleet of RCs that is serving a set of trips stopping at stations
(A2.d) The food products can be temporarily stored at the stations for some acceptable time before they are stowed on the trains. However, they
cannot wait an excessive amount of time
(A2.e) Only one type of RC departs from each DC d, with a capacity of Cd . For all trips t 2 N T and all DCs d in N D , Dd 6 Cd . Similarly, only one type of
RC, with capacity Cp , departs from FS p for all p 2 N P
(A2.f) jVdj is no fewer than the minimum number of RCs needed
Remark 2. Assumptions (A2.a)(A2.b) are adopted because of the special context of CSHR. Assumption (A2.a) implies that
an RC originating from a DC in a given city will not provide food products for stations in other city. This is because, in realworld situations, the distance between two stations in different cities is usually excessively long for the purposes of CSHR. As
a result, the intercity delivery of cold chain meals (which have a short lifespan of 24 h) from a DC is quite rare. Hence, we
simplify our problem by means of assumption (A2.a). Furthermore, because two DCs in different cities are also far away from
each other and line-haul operations from FSs to DCs usually cover a long travel time, considering the short lifespan of cold
chain meals assumption (A2.b) allows us to simplify the VRP for an RC from each FC visiting different DCs in different cities
such that we need only formulate the VRP from each DC to its corresponding stations in an urban area.
Remark 3. Assumptions (A2.c), (A2.d), (A2.f) are generally accepted in VRPs (Toth and Vigo, 2002).
2.2. Physical distribution network and network solution
2.2.1. Physical distribution network
Consider an existing distribution system consisting of the set of FSs denoted by N P , the set of candidate DCs in different
cities denoted by N D , and the set of stations denoted by N s . Furthermore, we define the following notation:
219
Notation 1. A line is a path in a railway network, and a line pool is a set of lines defined in terms of their frequencies and
halting patterns. A trip pool N T ft1 ; t2 ; . . . ; t jN T j g denotes the set of trips that are to be satisfied in our problem. Each trip in
the trip pool must correspond to one type of line in the line pool.
Notation 2. A train timetable is defined as a space-time graph that describes the arrival and departure times for each trip
t 2 N T at each station s 2 N S .
Notation 3. Based on the train timetable, a matrix Ls; t can be generated to describe the departure time for each trip t 2 N T
at each station s 2 N S . The symbol L here is used to signify the latest time for train loading, in that the departure time for
trip t at s is the latest time at which catering can be performed for this trip at s.
Based on the timetable, a possible station-to-trip assignment matrix ds; t is initialized, where s 2 N S and t 2 N T . If trip t
halts at station s, then ds; t 1; otherwise, ds; t 0. Furthermore, the ds; t values can be updated based on assumption
(A1.e) in Table 6. By the above principle, we can generate the following parameters:
ds; t
1 if trip t can be catered at station s

0 otherwise
Moreover, the following constraints describe the restrictions imposed by the line plan and the train timetable.
xst 6 ds; t
8s 2 N S ;
t 2 NT
Now, let a directed graph G GN; L, which represents the PDN, be the collection of all elements of the CSHR distribution
system, where N N P [ N D [ N S [ N T is the set of nodes. The trips in N T are represented by trip nodes in the PDN.
L LPD [ LDS [ LSS [ LST is the set of links between nodes. Considering assumptions (A1.a)(A1.e), we construct the set of links
as follows:
LPD N P N D
LDS [ fdg N S d
d2N D
LSS [ N S d N S d
d2N D
LST # N S N T
(1) LPD denotes the set of physical links from the FSs (N P ) to the DCs (N D ).
(2) LDS denotes the set of physical links between the DCs and the stations in N S . The definition indicates that each DC can
provide food products only to its corresponding rail stations in the city.
(3) LSS denotes the set of physical links between the stations in a city.
(4) LST denotes the set of possible station-to-trip assignment links determined by ds; t: If and only if ds; t 1, s; t 2 LST .
For each link i; j 2 L, there exists a tuple TC i;j ; TT i;j that represents the corresponding transportation cost and travel
time, respectively. Each illegal link is assigned some very large transportation cost and some very large travel time. Let
M be a large positive number. If s R N S d, then TT d;s M and TC d;s M. For each DC d 2 N D , if s 2 N S d and s0 R N S d, then
TT s;s0 M and TC s;s0 M.
Consider a rail company that serves a series of trips (denoted by t1 ; t2 ; t3 ; t4 ; t5 ; t6 ; t7 ; t8 ) using the distribution system represented in Fig. 3. Fig. 3(A) illustrates a distribution system that consists of two FSs p1 ; p2 , four DCs d1 ; d2 ; d3 ; d4 , and nine
stations s1 ; s2 ; s3 ; s4 ; s5 ; s6 ; s7 ; s8 ; s9 for a train timetable with eight trips. High-speed railways generally operate on double
tracks; therefore, we consider only unidirectional trips here. The meal time for lunch is from 11:00 to 13:00, and the meal
time for supper is from 16:00 to 20:00. Thus, we obtain the possible station-to-trip assignment depicted in Fig. 3. For example, trips t1 and t 2 dwell at s2 , so they can be catered at s2 . Therefore, ds2 ; t1 1 and ds2 ; t 2 1. Trip t 1 can only be catered
at s2 because trip t1 departs from its origin station (i.e. s2 ) during lunchtime. Trip t2 can be catered at either s2 or s4 because it
passes both s2 and s4 before suppertime. Fig. 3(A) can then be transformed into the graph depicted in Fig. 3(B), namely, the
PDN, which contains two FSs p1 ; p2 , four DCs d1 ; d2 ; d3 ; d4 , six stations s1 ; s2 ; s3 ; s4 ; s5 ; s6 (because s7 , s8 and s9 have no trip
assignments) and eight trip nodes, the latter of which are transformed from the trips (train paths) specified in the timetable.
2.2.2. Network solution
Based on the PDN obtained as described above, the planner can make the strategic decisions summarized in Table 4. A set
of values determined for the strategic decision variables is called a network solution r z; k; y; f; q; x:
(1) Selecting the DCs to be opened in N D :
z zd ; d 2 N D :
220
Fig. 3. Illustrative example of a possible station-to-trip assignment and its corresponding PDN.
(2) Determining the capacities of the DCs:
k kd ; d 2 N D :
(3) Selecting the FSs from N P that will provide the food products for each DC in N D :
y ypd ; p 2 N P ; d 2 N D :
(4) Determining the flow patterns between the FSs and DCs:
f f pd ; p 2 N P ; d 2 N D :
221
(5) Selecting the stations from set N S to be served by each DC:
q qds ; d 2 N D ; s 2 N S :
(6) Determining the station at which each trip in set N T will be catered:
x xst ; s 2 N S ; t 2 N T :
Given a feasible network solution, there exists a quadruplet Dt; Es; t; Ls; t; Js; t associated with each trip t 2 N T at a
station s 2 N S , which represents the demand that must be satisfied, the earliest start time for train loading, the latest time at
which train loading can end, and the journey time of trip t from station s to its destination.
(1) Dt: The demand for trip t. The method used in this paper to estimate demand is presented in Appendix A.
(2) Es; t and Ls; t:
Based on the train timetable, a matrix Ls; t can be generated to describe the departure time for each trip t 2 N T at each
station s 2 N S . Furthermore, based on the EMU circulation plan, each trip dwelling at a given station is labeled as an inbound
train, a halting train, a turnaround train or an outbound train, each of which have different types of dwell times at stations
(i.e., DWT s;t ): (i) An inbound train, which is assigned only an arrival time and remains at the depot overnight after it arrives at
the station, does not need to be catered. (ii) An outbound train is a train that has arrived at a station the day before or even
earlier and is assigned only a departure time for starting a new trip. A train of this type has a long dwell time to prepare for
the new trip (approximately 1530 min). (iii) A halting train is assigned both an arrival time and a departure time. The train
departs from the station on the same day as its arrival. A halting train continues with its current trip after a short dwell time
without changing its number; therefore, such a train has a short dwell time (approximately 24 min). (iv) A turnaround train
represents the intermediate case between and outbound train and a halting train (with a dwell time of approximately
15 min). In contrast to a halting train, a turnaround train changes its train number to begin a new trip after dwelling at
the station.
Fig. 4 displays an example of an EMU circulation plan on a space-time graph. This EMU circulation plan requires two days
to complete. We assume that one train is used to serve both trip t1 and trip t2 in Fig. 4. On the first day, the train that serves
trip t1 from s2 to s5 changes its train number at s5 and performs a round trip returning to s1 . The train remains overnight at
Depot 1. On the morning of the second day, the train serves trip t2 from s1 to s5 and then returns to Depot 2. At s1 and s2 , the
train is an inbound train. The train is an outbound train for trips t 1 and t 2 at the beginning of each day. At s5 , the train acts as a
turnaround train. At s4 , the train is a halting train on either trip t1 or trip t 2 .
After the meals are delivered to the stations, some maximum storage time (MST), denoted by MST s , is permitted at each
station. Finally, for each t 2 N T at station s 2 N S ,
Es; t Ls; t DWT s;t MST s
8s 2 N S t 2 N T
Then, a time window Es; t; Ls; t is defined for train loading, where Es; t; Ls; t 2 0; 1440.
(3) Js; t: The train timetable also gives rise to a matrix Js; t, which represents the remaining journey time for trip t from
s to its destination.
Fig. 5(A) shows the PDN derived from Fig. 3 when trips t1 ; t2 ; t3 ; t4 ; t5 ; t 6 ; t 7 and t8 are transformed into trip nodes. A feasible network solution is presented in Fig. 5(B). Let Dt 500 units for all trips. We allow an outbound or turnaround train
to have a dwell time of 15 min and set a dwell time of 3 min for halting trains. MST s 25 min for all stations. Based on the
network solution depicted in Fig. 5(B), we have Dt 1 500 units, Es2 ; t 1 650 min ( 690 min 15 min 25 min),
Ls2 ; t1 690 min (11:30), and Js2 ; t 1 = 300 min, and we also have Dt6 500 units, Es3 ; t6 542 min
( 570 min 3 min 25 min), Ls3 ; t6 570 min (9:30), and Js3 ; t 6 = 300 min.
The quadruplets associated with other trips are also shown in the figure.
2.3. Operational distribution network and operation scheme
Given a feasible network solution r z; k; y; f; q; x, an ODN is generated. Then, the operators of the DCs can determine
the amounts of product to order from the FSs and create a time schedule for the RCs based on the ODN. Consider the directed
graph G GN o ; A, where N o N P [ N D [ N W [ N T . Each DC d is represented by a node d in N D and a corresponding dummy
node wd in N W . Thus, a feasible RC route is a vehicle path in the ODN starting from a node d in N D and ending at its dummy
node wd in N W . Moreover, A APD [ ADT [ ATW [ A1TT [ A2TT [ ADW :
APD # N P N D
ADT # [
[ fdg N T s
d2N D s2N S d
222
Fig. 4. An example of the time windows imposed by an EMU circulation plan.
Fig. 5. Illustrative example of a PDN and a possible feasible network solution in CSHR.
ATW # [
[ fwdg N T s
d2N D s2N S d
A1TT # [ N T s N T s
s2N S

A2TT # [
d2N D

[
s;s0 2N S d
N T s N T s0
ADW # N D N W
(1) APD denotes the set of operational links from the FSs to the DCs.
(2) ADT and ATW represent the set of operational links between the DCs and the trips. Because each DC can serve only stations in its territory, if there is a path from d in N D to t in N T via any s in N S according to network solution r (i.e.,
qds xst 1), then there exists an operational link represented by d; t 2 ADT and t0 ; w 2 ATW , where t0 can equal
or do not equal to t.
(3) A1TT denotes the set of operational links between the trips catered at a station. If the number of meals needed for trip t
is satisfied by DC d via station s, then this trip is included in N T s, where s is in N S d. For example, in Fig. 5(B),
N T s1 ft3 ; t4 ; t5 g, N T s2 ft1 ; t2 g, and N T s3 ft6 ; t7 ; t8 g.
223
(4) A2TT denotes the set of operational links between trips dwelling at different stations.
(5) ADW denotes the fictitious link set.
Given TT ij and TC ij , which represent the travel time and transportation cost, respectively, from node i to node j in the PDN,
we define the travel time tt ij and the transportation cost tcij for each link in the ODN as follows.
(1) The set of operational links from the FSs to the DCs (APD ):
ttp;d TT p;d
8p 2 N P ;
d 2 ND
tcp;d TC p;d
8p 2 N P ;
d 2 ND
(2) The set of operational links between the DCs and the trips (ADT and ATW ):
ttd;t
TT d;s xst
8d 2 N D ; t 2 N T
TC d;s xst
8d 2 N D ; t 2 N T
s2N S d
tcd;t
X
s2N S d
ttt;wd tt d;t
8d 2 N D ;
t 2 NT
tct;wd tcd;t
8d 2 N D ;
t 2 NT

(3) The set of operational links between the trips catered at a station A1TT :
ttt;t0 e 8t; t 0 2 N T s;
8s 2 N S
tct;t0 e 8t; t0 2 N T s;
8s 2 N S
where e is some sufficiently small value. In this paper, the loading/unloading times and transportation costs at rail stations
are neglected. Therefore, if a trip dwelling at a station has already been catered by an RC, then other trips that pass through
that station can also be catered by the RC without incurring any additional transportation cost or travel time.

(4) The set of operational links between trips dwelling at different stations A2TT :
ttt;t0
X X
TT s;s0 xst xs0 t0
8t; t 0 2 N T
TC s;s0 xst xs0 t0
8t; t 0 2 N T
d2N D s2N S ds0 2N S d
tct;t0
X X
d2N D s2N S ds0 2N S d
where s and s0 are different stations served by the same DC d.

(5) The fictitious link set (ADW ):
ttd;wd 0 8d 2 N D
ttd;w M
8w 2 N W ; wwd;
8d 2 N D
Fig. 6 illustrates the process of deriving an ODN G GN o ; A from the network solution r presented in Fig. 5(B) (i.e., Fig. 6
(A)), where N o fp1 ; p2 ; d1 ; d2 ; w1 ; w2 ; t1 ; t2 ; t3 ; t4 ; t5 ; t6 ; t7 ; t8 g. Because t1 and t 2 are catered at s2 from d1 , the physical links
d1 ; s2 , s2 ; t2 , and s2 ; t 1 in the PDN are replaced with the operational links d1 ; t 2 and d1 ; t 1 . Similarly, the physical links
d2 ; s3 , s3 ; t 6 , s3 ; t7 and s3 ; t8 are represented by the following set of parallel operational links: d2 ; t6 , d2 ; t 7 and d2 ; t 8 .
If a trip dwelling at a station has already been catered by an RC, then other trips that pass through that station may be
catered by the same RC without incurring any additional transportation cost or travel time. Fig. 6(B) shows, for all trips
in N T s1 ; N T s2 and N T s3 , the trips that are connected to each other by green links. Given that N S d1 fs1 ; s2 g
andN S d2 fs3 g, the red1 links between the trips in N T s1 and N T s2 are also displayed in Fig. 6(B). Fictitious links
d1 ; wd1 and d2 ; wd2 are also shown in the figure, where wd1 and wd2 are dummy nodes corresponding to d1 and d2 .
1
For interpretation of color in Figs. 6,15,16,18 and 19, the reader is referred to the web version of this article.
224
Fig. 6. An example of an ODN representation and a possible operation scheme.
Based on the ODN, the operators of DCs make the tactical and operational decisions summarized in Table 5. An operation
scheme o r; s is thus obtained, subject to time-window constraints, capacity constraints and time-deadline constraints.
These sub-problems can be treated as the VRPTW with time budget constraints (VRPTWDC).
(1) The routes of the RCs from any DCs:

r rdijv ; i; j 2 N o n N P ; v 2 Vd; d 2 N D :
(2) The schedules of the RCs from any DCs:
s sdi v ; i 2 N o n N P ; v 2 Vd; d 2 N D :
225
Consider an RC from DC d1 that first caters food products for trips t 3 and t 4 , then traverses link s1 ; s2 in the PDN to provide meals for t1 , and finally returns to DC d1 . This process can be represented by a route in the ODN such as
d1 ! t 4 ! t 3 ! t1 ! wd1 (route in Fig. 6(C)). If a trip dwelling at a station has already been catered by an RC, then other
trips that pass through that station may be also catered by the RC without incurring any additional transportation cost or
travel time. Therefore, if d1 ! t 4 ! t 3 ! t 1 ! wd1 is a feasible route, then the order of t3 and t4 can be reversed, and
d1 ! t 3 ! t 4 ! t1 ! wd1 is also a feasible route representing the same process. Fig. 6(C) also shows RC routes , ,
and , which represent a feasible operation scheme.
Let the maximum holding time at a DC be 100 min, and let the lifespan of the food products be 1440 min. For each trip, in
Fig. 6(C), we find that the lead time from the FS to the trips destination does not exceed the given time deadline. Thus, Fig. 6
(C) truly represents a feasible operation scheme o r; s. In Section 2.4, we discuss how, given an operation scheme
o r; s, to guarantee its feasibility using a series of time budget constraints.
2.4. Reformulating time deadlines as time budget constraints
The DSDP-CSHR involves one additional difficulty with respect to classical center location problems because of the need
to trace the paths to determine the lead times for rail catering. Given a network solution r z; k; y; f; q; x, a feasible operation scheme o r; s should ensure that none of the lead times from an FS to the destinations of its accessible trips exceeds
a given deadline (Ho and Leung, 2010). Referring to Contreras et al. (2012) and the conventional p-center problem, in this
section, we reformulate time deadlines as a series of time budget constraints (Tong et al., 2015a).
The key idea guiding the formulation of these time deadline constraints is based on a rooted forest that is defined based
on assumptions (A1.a)(A1.e) in Table 6. Based on these assumptions, it is known that a solution to the DSDP-CSHR exists,
which can be represented by a rooted forest. Because we assume that each DC is associated with only one FS in N P , each tree
in this forest contains a single FS as its root node. The tree structure makes us possible formulate time deadlines as time
budget constraints to restrict the longest accumulated travel time among all paths from an FS (a root node) to the destinations of its accessible trips, especially when the path flows in the network are unknown.
For simplicity, we use only a two-echelon network solution to illustrate this process. In Fig. 7(A) and (B), we show two
paths whose travel times are longer than the specified time deadline (i.e., 10 min). Now we apply the time deadline to
restrict the longest accumulated travel time among all paths from A0, A1 and A2 (root nodes) to their accessible ends.
The first network solution is clearly infeasible since the travel time for the path A0 ? B0 ? C0 is longer than 10 min. By contrast, the feasibility of the second network solution cannot be determined when we do not know path flows in the network.
For example, if the path flows on A2 ? B ? C2 are positive, then the network solution is clearly infeasible. However, if only
the path flows on A1 ? B ? C2 and A2 ? B ? C1 are positive while the path flows on A2 ? B ? C2 and A1 ? B ? C1 are
zero, then the network solution is feasible, although a path with an 11 min travel time also exists.
It means that if the distribution process is not based on a tree structure, the time-budget constraints to restrict the longest
accumulated travel time among all paths from a root node will incorrectly remove some feasible network solutions and
operation schemes. Because of the difficulty of knowing the path flows during the network design phase, assumptions
(A1.a)(A1.e) are adopted in this paper to simplify the problem.
Let R denote the rooted forest for a network solution r z; k; y; f; q; x and an operation scheme o r; s. Let K be the
total number of FSs that supply products to the DCs, and let K 6 jN P j. For the rooted forest R, we write
R fT 1 ; T 2 ; . . . ; T k ; . . . ; T K g, where
(1) T k N k ; Ek is a rooted tree, where N k fpk g [ N D k [ N V k [ N T k [ N S k, for k 1; 2; . . . K.
(2) The root nodes are denoted by pk 2 N P , for k 1; 2; . . . K.
(3) N D k # N D is the second-layer node set. {N D 1; N D 2; . . . ; N D k; . . . ; N D K} defines a partition of the set of DCs in
the network solution r z; k; y; f; q; x.
(4) N V k # V is the third-layer node set. {N V 1; N V 2; . . . ; N V k; . . . ; N V K} defines a partition of the set of RCs used in
the operation scheme o r; s.
(5) N T k # N T is the fourth-layer node set. {N T 1; N T 2; . . . ; N T k; . . . ; N T K} defines a partition of the set N T .
(6) N S k # N S is the fifth-layer node set. {N S 1; N S 2; . . . ; N S k; . . . ; N S K} defines a partition of the set of destination
stations for the trips in N T .
(7) For k 1; 2; . . . K, Ek # fpk g N D k [ N D k N V k [ N V k N T k [ N T k N D k [ N T k N S k.
First, we note that each trip has only one destination station. Thus, for each trip node t 2 N T , we can define the travel time
for a path in R to express the catering lead time for trip t from its root FS to its destination station. The destination station of
trip t is represented by an end leaf of the rooted forest R. The restrictions that limit the lead time for rail catering to within a
given time deadline can be reformulated as time budget constraints on the rooted forest R. Fig. 8 displays the rooted forest
corresponding to the network solution and operation scheme given in Fig. 6(C).
The question is how to describe the edges in the rooted forest R in terms of the decision variables in the network solution
r z; k; y; f; q; x and the operation scheme o r;s. From the depiction of the rooted forest R in Fig. 8, it can be seen that
the path defining the catering lead time from FS p to the destination of trip t via DC d consists of five types of edges. For
226
Fig. 7. Two simplified two-echelon network solutions.
Fig. 8. A rooted forest corresponding to the network solution and operation scheme depicted in Fig. 6(C).
k 1; 2; . . . K and for each e 2 Ek , there exists a corresponding travel time, which can be described in terms of the decision
variables in the network solution and the operation scheme. For k 1; 2; . . . K, Table 7 lists the five edge sets in the rooted
forest and the mathematical descriptions of their travel times. Each type of edge (edge set) corresponds to a head set and a
tail set. For example, the edge set that is named the travel time from DC d to trip t via RC v contains a series of edges
whose heads are in the set N V k and whose tails are in the set N T k (see Table 7).
To keep track of the longest accumulated travel time among all paths that originate from an FS (a root node), we define a
series of continuous variables g to represent the travel times. For k 1; 2; . . . K and every node in N k , the corresponding variable gives the MAT on any path in R from that node to one of its accessible leaves (i.e., the destination of a trip node in N T )
(see Fig. 8). Table 8 lists the additional time decision variables referenced in this section.
Then, given a network solution and an operation scheme, we propose the following time budget constraints:
gt
Js; txst
s2N S
"
dv
t
8t 2 N T
1
#0
X

X dv
dv A
@
sdt v sdt v sddv
Ls; txst
r tj
rjt
g t pdt v M 1
8t 2 N T ;
j2D t
j2D t
s2N S
v 2 Vd; 8d 2 N D
r djtv
j2D t
g dv HT d M1 zd 6 g d
6 g dv
8t 2 N T ;
8v 2 Vd;
g d tt p;d ypd M1 ypd 6 g p
8d 2 N D ;
v 2 Vd; 8d 2 N D
8d 2 N D
p 2 NP
227

Table 7
List of edge sets in a rooted forest.
Edge set name (Ek )
Head
Tail
Mathematical description in terms of r and o
The travel time from FS p to DC d

The holding time at DC d
The travel time from DC d to trip t via RC
fpk g
N D k
N V k
N D k
N V k
N T k
tt p;d ypd
HT d zd
Part 1: The travel time from DC d to the
station
Part 2: The storage time at the service
station
P
s2N S Js; txst
The journey time from the service station to the

destination
N T k
N S k
P
sdt v sddv j2D t r djtv
hP
s2N S Ls; txst
iP
dv
j2D t r tj
sdt v
Table 8
List of additional time variables.
Decision
variable
Corresponding vertexes
in R
Description
gt
g dv
N T k
N V k
gd
gp
N D k
fpk g
N/A
The time for which the meals are stored on the train serving trip t in N T . For all t 2 N T , g t 2 R [ f0g
The MAT from DC d to the destination of any accessible trip via RC v . For all v 2 Vd and d 2 N D ,
g dv 2 R [ f0g
The MAT from DC d to any accessible trips destination. For all d 2 N D , g d 2 R [ f0g
The MAT from FS p to any accessible trips destination. For all p 2 N P , g p 2 R [ f0g
An intermediate variable
pdt v
g p 6 Lspa 8p 2 N P
Constraint (3) defines the storage time of the meals in a train. Constraints (4) and (5) are used to calculate the MAT on any
path in R from a DC to the destination of one of its accessible trips via a specific RC v . Constraint (6) is used to calculate the
MAT on any path in R from a DC to the destination of one of its accessible trips. Constraint (7) is used to calculate the MAT on
any path in R from a root FS to the destination of one of its accessible trips. Constraint (8) states that the MATs on all paths
from any root FS should be less than a given time deadline Lspa, as shown in Fig. 8. It should be noted that the above constraints are no longer correct when we relax the basic assumptions (A1.a)(A1.e), since they may remove some feasible network solutions and operation scheme.
3. Model formulation
3.1. Formulation of the 3E-LRPTWTBC for the DSDP-CSHR
The DSDP-CSHR addressed in this study can be formulated as a mathematical model called the three-echelon locationrouting problem with time windows and time budget constraints (3E-LRPTWTBC).
Model 1 (3E-LRPTWTBC):
min Zr; o
X X X
XX
X
X
f pd
tcp;d
PDtci;j r dijv
OC d zd
UC d kd
p2N P d2N
Cp
d2N D v 2Vdi;j2AfA
d2N D
d2N D
The objective function given in (9) minimizes the total cost of the distribution system. The first term is the cost of transportation from the FSs to the DCs, where df pd =Cp c is the number of RCs on link p; d in the ODN. The second summation represents the cost of the RCs transporting the meals from the DCs to the trips in the ODN. Considering that this term is required
to determine the accumulated transportation cost tci;j rdijv for each link i; j 2 A n APD , the summation also includes the cost of
the RCs returning to the DCs after delivering the meals to the trips, as noted in Fig. 6. The third term represents the opening
cost for the DCs. The fourth term represents the cost of the rented capacity at each DC.
This objective function is minimized subject to the following constraints:
(1) Capacity constraints for facilities in the PDN:
X X
Dtxst qds 6 UCapd kd
8d 2 N D
10
s2N S dt2N T
kd 6 K d
X
d2N D
8d 2 N D
f pd 6 Capp
8p 2 N P
11
12
228
Constraint (10) represents a restriction whereby the number of meals provided by a DC cannot exceed its capacity. Constraint (10) represents a class of non-linear constraints in this model. Constraint (11) represents a restriction whereby the
number of units of capacity to be rented at any DC cannot exceed a certain upper bound. Constraint (12) ensures that the
capacities of the FSs are respected.
(2) Generalized
upper bounds in the PDN:
X
xst 1 8t 2 N T
13
s2N S
qds 6 1 8s 2 NS
14
d2N D
qds 0 8s R NS d;
X
8d 2 N D
15
ypd 6 1 8d 2 ND
16
p2N P
Constraint (13) indicates that each trip is required to be served only once. Constraints (14) and (15) state that each DC can
provide meals only to their corresponding stations and that there is only one DC in a given city. Constraint (16) states that
each DC is supplied by at most one FS.
(3) Constraints imposed by the line plan and train timetable in the PDN: Constraint (1).
(4) Flow balance constraints in the PDN:
X X
Dtxst qds
s2N S dt2N T
8d 2 N D
f pd
17
p2NP
Constraint (17) ensures flow conservation at each DC. Constraint (17) also represents a class of non-linear constraints.
(5) Vehicle routing constraints in the ODN:
X X X
r dtjv 1 8t 2 N T
18
d2N D v 2Vdj2D t
rddjv 1 8v 2 Vd;
8d 2 N D
19
j2D d
rdjtv
j2D t
r dtjv
8t 2 N T ;
v 2 Vd; 8d 2 N D
20
wd 2 N W ;
21
j2D t
v 1 8v 2 Vd;
r djwd
8d 2 N D
j2D wd
Dt
t2N T
8v 2 Vd;
rdtjv 6 Cd
8d 2 N D
22
j2D t
The above constraints are similar to those in a VRPTW model (Toth and Vigo, 2002). Notably, the above constraints can only
be used when the length of each arc in the ODN is greater than zero. Thus, in Section 2.3, we defined
ttt;t0 e 8t; t 0 2 N T s;
8s 2 N S
tct;t0 e 8t; t0 2 NT s;
8s 2 N S
ttt;t0
X X
d2N D s2N S
tct;t0
ds0 2N
X X
d2N D s2N S
TT s;s0 xst xs0 t0 8t; t 0 2 N T
S d
ds0 2N
TC s;s0 xst xs0 t0 8t; t 0 2 N T
S d
where e is a very small positive number and s and s0 are different stations served by the same DC d. Constraint (18) restricts
the assignment of each trip to exactly one RC route. Constraints (19)(21) characterize the paths followed by the RCs. Constraint (22) ensures that the number of meals carried by an RC does not exceed the RCs capacity.
(6) Time-window constraints in the ODN:
sdi v tti;j sdj v 6 1 rdijv M 8i; j 2 A n APD ; v 2 Vd; 8d 2 N D
23
sdi v tti;j sdj v P 1 rdijv M 8i; j 2 A n APD ; v 2 Vd; 8d 2 N D
24
1
#0
X
d
v
Es; txst @
r tj A 6 sdt v
s2N S
j2D t
"
1
#0
X
dv A
@
P sdt v
Ls; txst
r tj
s2N S
j2D t
"
229
8t 2 N T ; v 2 Vd;
8d 2 N D
25
8t 2 N T ; v 2 Vd;
8d 2 N D
26
We define a decision variable sdi v to represent the start time for service by RC v from DC d at i in N D [ N W [ N T . The time sdi v is
accumulated from the departure of v from DC d to the arrival of v at a dwelling station for trip t using constraints (23) and
(24). Sub-tour elimination can be naturally achieved by means of constraint (23) and (24) in the VRPTW through the use of
the small parameter e. Constraints (25) and (26) guarantee the feasibility of the time windows. They ensure that sdt v 0
whenever trip t is not visited by v .
(7) Time-deadline constraints in the rooted forest: Constraints (3)(8).
(8) Relationships between the variables in the PDN and ODN:
kd 6 Mzd
8d 2 N D
27
zd 6 Mkd
8d 2 N D
28
xst 6 M
t2N T
qds
8s 2 N S
29
8s 2 N S
30
d2N D
qds 6 M
xst
t2N T
d2N D
f pd 6 Mypd
8p 2 N P ;
d 2 ND
31
ypd 6 Mfpd
8p 2 N P ;
d 2 ND
32
xst
s2N S d
X X
v 2Vdj2D t
r dtjv
8t 2 N T ;
d 2 ND
33
Constraints (27) and (28) state that a candidate DC is opened if and only if some storage capacity has been rented at that DC.
Constraints (29) and (30) state that a station should be served by a DC if and only if trains are to be catered at that station.
Constraints (31) and (32) state the relationships between pairs of decision variables f and y. Constraint (33) states the relationships between pairs of decision variables x and r.
In summary, all three types of graphs (the PDN, ODN and rooted forest) represent different organizational forms of the
basic elements of the DSDP-CSHR. Each of them serves as the basis for the formulation of constraints in the 3ELRPTWTBC (see Fig. 9).
3.2. A lower bound on the 3E-LRPTWTBC
The LRP has been proven to be an NP-hard problem, and the 3E-LRPTWTBC for the DSDP-CSHR is even more difficult to
solve. Therefore, we propose a relaxation model (RM) to provide a lower bound on the solution to the 3E-LRPTWTBC for the
DSDP-CSHR to estimate the gap with respect to the global optimum.
The underlying idea of the RM is to relax the constraints related to urban transportation:
(1) Time-window and time-deadline constraints.
(2) All other constraints related to the rail stations and RCs.
Consequently, we delete the variables xst , qds , rdijv ,
sdi v , and g dv in the 3E-LRPTWTBC and introduce a new variable:
xdt 1 if DC d 2 N D serves trip t 2 N T

xdt 0
otherwise
Furthermore, each additional time variable is replaced with a corresponding lower-bound variable:
(1) For all t 2 N T , g t is replaced with g t , where g t 6 g t .
(2) For all d 2 N D , g d is replaced with g d , where g d 6 g d .
(3) For all p 2 N P , g p is replaced with g p , where g p 6 g p .
230
Fig. 9. Relationships between the different graphs and the constraints in the 3E-LRPTWTBC.
With the introduction of these new decision variables, some constraints in the 3E-LRPTWTBC are linearized. The RM further relaxes the time windows and the time-deadline constraints in the 3E-LRPTWTBC. The RM thus becomes a classical MIP
model that can be solved using any off-the-shelf MIP solver.
Model 2 (RM):
min ZL
XX
TC p;d
p2N P d2N D
f pd
Cp

P
X
X
X
t2N T xdt Dt
2 min TC d;s
OC d zd
UC d kd
d2N D
Cd
s2N S d
d2N D
34
d2N D
The best optimum of the objective function represents a lower bound on the objective function of the 3E-LRPTWTBC:
Proposition 1. For all network solutions s and operation schemes o, min Zr; o P min ZL .
Proof. It is evident that
XX
TC p;d
p2N P d2N D
f pd
Cp
XX

f pd
tcp;d
Cp
p2N P d2N D
We observe the following:
(i) Because mins2N S d TC d;s is the cheapest transportation cost from DC d to the stations it serves in its city, 2 mins2N S d TC d;s
is the lower bound on the transportation cost for a delivery by any RC from DC d to the stations and the subsequent
return to DC d.
P

x Dt
t2N T dt
is the minimum number of RCs that must be used to satisfy the demand served by DC d.
(ii) Moreover,
C
d
Evidently,
t2N T xdt Dt
Cd
P
6
t2N T xdt Dt
Cd
Because of (i) and (ii), we know that

P
X X
X
t2N T xdt Dt
6
2 min TC d;s
d2N D
Cd
s2N S d
d2N D v 2Vdi;j2AnAPD
tci;j r dijv
Consequently, we have min Zr; o P min ZL . h

The RM is subject to the following constraints:
(1) Capacity constraints for facilities:
Dtxdt 6 UCapd kd
8d 2 N D
35
t2N T
kd 6 K d 8d 2 N D
X
f pd 6 Capp 8p 2 N P
d2N D
36
37
231
With the introduction of the decision variables xdt , the RM directly relates the DCs to the trips and treats the distribution system
as a two-echelon network. Thus, the non-linear constraint (10) in the 3E-LRPTWTBC is linearized to yield constraint (35).
(2) Generalized
upper bounds:
X
xdt 1 8t 2 N T
38
ypd 6 1 8d 2 N D
39
d2N D
p2N P
(3) Flow
constraints:
X balanceX
Dtxdt
t2N T
f pd
8d 2 N D
40
p2N P
Constraint (40) ensures flow conservation at each DC. This constraint is also linearized.
(4) Constraints
imposed by the line plan and train timetable:
xdt 6 1 9 s 2 N d s s:t: ds; t 1

xdt 0
otherwise
(5) Relaxed
X time-deadline constraints:
xdt min Js; t 8t 2 N T
gt
d2N D
s2N S d
g t min tt d;s HT d 6 g d
s2N S d
8t 2 N T ; d 2 N D
g d tt p;d ypd M1 ypd 6 g p
8d 2 N D ; p 2 N P
g p 6 Lspa 8p 2 N P
41
42
43
44
45
The underlying structure of constraints (41)(44) is similar to that of constraints (3)(8) for the rooted forest.
We observe the following:
(i) For all t 2 N T ,
X
X
xdt min Js; t 6
Js; txst
d2N D
s2N S d
s2N S
where mins2N S d Js; t is the shortest travel time from a station to the destination of train t among all stations in the city
served by DC d.
(ii) Furthermore, pdt v in the 3E-LRPTWTBC is replaced with mins2N S d ttd;s g t , and
min tt d;s g t 6 pdt v ;
s2N S d
where mins2N S d tt d;s is the shortest travel time from DC d to the stations it serves in the city.
Because of (i) and (ii), it is implied that the RM has looser deadline restrictions compared with the 3E-LRPTWTBC.
(6) Relationships between variables:
kd 6 Mzd
8d 2 N D
46
zd 6 Mkd
8d 2 N D
47
f pd 6 Mypd
8p 2 N P ;
d 2 ND
48
ypd 6 Mfpd
8p 2 N P ;
d 2 ND
49
Eight groups of constraints are listed in Table 9 to better illustrate the model structure. The different formulations for the 3ELRPTWTBC and the RM are also compared. Groups 1, 2, 4, and 8 consist of constraints for the discrete network design problem
with flow assignment (Meng and Wang, 2011). Groups 5 and 6 are routing constraints for the VRPTW (Toth and Vigo, 2002).
Group 7 contains the time accumulation constraints, which are similar to the constraints in the CFLNDB (Contreras et al., 2012).
4. Solution algorithm
This paper proposes a metaheuristic algorithm called the hybrid cross entropy algorithm (HCEA), which combines the
cross entropy algorithm (CEA) (De Boer et al., 2005) with a time-oriented nearest-neighbor heuristic (TNNH). First, we define
the fitness function as follows:
232
Hr; o Zr; o
Pen maxkd K d ; 0
50
d2N D
d2N D
where Pen represents positive penalty coefficients. The HCEA is motivated by the iterative heuristic approach used to solve
the LRP (Bruns et al., 2002). This approach serves as the solution procedure through two iterative steps. The first step generates a network solution r z; k; y; f; q; x, and the second step generates an operation scheme o r; s based on the generated r. We then evaluate the fitness function and iteratively expand the search process.
4.1. Hybrid cross entropy algorithm
The sample set is initialized and further prepared to be updated and optimized using probability vectors, namely,
^ p
^st js 2 N S ; t 2 N T , where p
^st denotes the probability that trip t is catered at station s.
p
The main loop of the HCEA.
c0 pc
0 js 2 N ; t 2 N are given at
Input: Let n = 0 and let iter 0. Suppose that a PDN and an initial probability vector p
T
S
st
iter 0.
Output: A near-optimal network solution and operation scheme.
cn , randomly determine a population consisting of g distinct samples, denoted by the set
(1) Initialization:
Based on p
n

n
n
n

S rm m 1; 2; . . . ; g , where the bit string rnm znm ; km ; ynm ; f m ; qnm ; xnm represents a network solution. The
detailed process is as follows:
pbn
(1.1) For every t 2 N T , randomly select one station as the service station in N S with probability P st
:
pbn
s2N S st

select another
station. Thus, xnm xnm
If xnm
st 6 ds; t, accept the choice; otherwise,
st js 2 N S ; t 2 N T is obtained.

(1.2) Based on xnm , generate qnm qnm
jd
2
N
;
s
2
N
based
on constraints (14)(15) and (29)(30).
D
S
ds

P
P
qnm > 0, then zd 1: Then, znm znm
(1.3) For d 2 N D , if s2N S d t2N T Dtxnm
d jd 2 N D .
stP ds P

nm nm
Dtxst qds
nm

nm
s2N S d
t2N T
n
(1.4) For s 2 N S d and d 2 N D , kd
. Then, km kd d 2 N D .
UCap
d
(1.5) Assign flows between the FSs and DCs by approximately solving the following integer linear programming
model:
lf nm m
XX
X X tc nm
p;d
tcp;d Cpdp
min
Cp f pd
p2N P d2N D
p2N P d2N D
subject to
P
P
P
nm
nm nm
p2N P f pd
s2S d
t2T Dtxst qds
P
nm
8p 2 N P
d2N D f pd 6 Capp
f pd 6 Mypd
8p 2 N P ;
8d 2 N D
d 2 ND
ypd 6 Mfpd 8p 2 N P ; d 2 N D
P
nm
8d 2 N D
p2N P ypd 6 zd
nm
f pd 2 R [ f0g
8p 2 N P ; d 2 N D

n
nm
Thus, we obtain f m f pd jp 2 N P ; d 2 N D and ynm ynm
pd jp 2 N P ; d 2 N D .
(2) Sub-routing problem: Generate an ODN for each rni in Sn . Then, for each d 2 N D , solve the VRPTWTBC to obtain

onm rnm ; snm , where
nm

rdijv
j 8i; j 2 A n APD ; v 2 Vd; 8d 2 N D
rnm
nm

snm sdi v ; 8i 2 No n NP ; v 2 Vd; 8d 2 ND

(4) Fitness calculation: Calculate the evaluation values Hnm H rnm ; onm .
(5) Normalization: Compute the fitness of each sample by normalizing its evaluation value using the following
n
n
n
min h
~ n Hrmn ;om H
, where Hnmax and Hnmin are the maximum and minimum evaluation values, respectively,
equation: H
m
Hmax Hnmin h
among all samples included in Sn and h 2 [0, 1].
~n P H
~n P H
~n P H
~ n . Let c
^n represent the
(6) For i 1; 2; . . . ; g, order the fitnesses from largest to smallest: H
1
2
m
g
~n
^n : H
(1q) sample quintile of the performances as follows: c
1qg

n
c
cn
(7) Renew
Pg probability vectors: Use the samples to renew p pst js 2 N S ; t 2 N T .
I ~ n cn g Ifxnm 1g
m1 fHm 6^
st
n
Pg
pc
8s 2 NS ; t 2 NT , where Ifpropositiong = 1 if the proposition is true and =0 otherwise.
st
I ~ n cn g
m1 fHm 6^
cn becomes a 01 binary vector or if jc

^n1 c
^n j < e, then iter iter 1; otherwise, iter 0.
(8) If p
(9) Stopping criterion: Stop if iter > S1 or if n > S2 , where S1 and S2 are arbitrarily generated values. Otherwise, set
n n 1 and go to (1).
233

Table 9
Comparison of the model formulations.
Group
index
Constraint type
3E-LRPTWTBC
RM
Capacity
constraints for
facilities
Generalized upper
bounds
Constraints
imposed by the RTP
Flow balance
constraints
Vehicle routing
constraints
Time-window
constraints
Time-deadline
constraints
Non-linear constraints that refer to variables xst and qds Linear

constraints that refer to variables f pd
Linear constraints that refer to variables xdt Linear

constraints that refer to variables f pd
Refer to variables xst , ypd and qds
Refer to variables xdt and ypd
Refer to variables xst
Refer to variables xdt
Non-linear constraints that refer to variables xst , qds and f pd
Linear constraints that refer to variables xdt and f pd
Refer to variables r dijv
N/A
2
3
4
5
6
7
Relationships
between variables
Non-linear constraints that refer to variables xst ,
dv
i
and
r dijv
Non-linear constraints that refer to variables g t , g dv , g d and g p
8t 2 N T , d 2 N D , p 2 N P , g t 6 g t , g d 6 g d , g p 6 g p
Refer to variables kd , zd , f pd , ypd , r dijv ; xst and qds
N/A
Linear constraints that refer to variables g t , g d and
gp
Refer to variables kd , zd , f pd and ypd
4.2. Time-oriented nearest-neighbor heuristic

The HCEA approach is hybridized with a time-oriented nearest neighbor heuristic (TNNH) to solve the sub-routing problem with time windows and time budget constraints (i.e., the VRPTWTBC). The TNNH is a popular route construction heuristic for the VRPTW because of its ease of use and good problem-solving performance (Solomon and Desrosiers, 1988). In the
HCEA, this simple heuristic is implemented for each d 2 N D in a given ODN to obtain rnm for every sample m and every generation n.
The TNNH belongs to the class of tour-building algorithms. For all d 2 N D , the TNNH starts every route by finding an
unvisited customer closest to d in the ODN. In every iteration, the heuristic searches for the trip closest to the last trip added
to the route and adds it to the end of the route. Given a network solution r = (z, k, y, f, q, x), the time window for catering
hP
i
P
P
t 2 N T is known, i.e.,
s2N S Es; txst ;
s2S Ls; txst . The storage time in each t 2 N T is also known, i.e.,
s2N S Js; txst . The
capacity of the RCs originating from DC d 2 N D is given. Thus, for each d 2 N D , the search for a route for RC v 2 Vd in
the ODN among all feasible trips can be performed with respect to the relevant time windows, RC capacity constraints
and time budget constraints. The following formula should be checked for each trip t 2 N T :
g t 6 Lspa
0
dv
t
tt d;p ydp HT d p
p2N p
M @1
1
xst qds A 8v 2 Vd;
8d 2 N D ;
t 2 NT
51
s2N S d
If this condition is satisfied, then the time budget constraint is satisfied. Trip t is added to RC v originating from DC d. Then, a
new route is started whenever the search fails to find a feasible place for insertion, unless no unvisited customers remain for
the DC.
4.3. Method to avoid becoming trapped

The HCEA occasionally becomes trapped in a local optimum after accepting worse nearby solutions. We present an example in which such a trap leads to an unacceptable situation in our approach. Fig. 10 presents a distribution system subject to
the following considerations: The system consists of s1 , s2 , s3 , and s4 as well as d1 and d2 . Let d1 serve s1 , and let d2 serve s2 and
s3 . The utilized line plan, which consists of only one line type, is shown in Fig. 10. The line corresponds to 10 trips, any of
which can be catered at either s1 , s2 or s3 . All other costs, except the opening cost for each DC, are neglected. The global opti

c
0
p0 p
0:5js 2 N t 2 N , only
mum is the solution whereby all trips are catered at s , with a total cost of 1. However, if c
1
st
(1/2)30 percent of feasible solutions will escape from the local optimum.
A method that can be applied to avoid becoming trapped to some extent is as follows. First, we label all indispensable DCs,
which are required to cater certain trips. If DC d is not an indispensable DC, then in each iteration, we generate at least one
sample for which d is closed. This means that no trips are catered at the stations served by d. This approach can be
extended by shutting down pairs of closable DCs to generate some samples in each iteration.
234
Fig. 10. An example of a system for which the HCEA becomes trapped.
5. Numerical examples
In our experiments, all programs were coded in MATLAB 2012 and run under Windows 7 on a system with an Intel (R)
Core (TM) i7-4700 MQ CPU at 2.4 GHz and 8 GB of RAM. The MIP solver in CPLEX v. 12.3 was used to solve the RM, using the
default parameters. In the following studies, we allowed an outbound/turnaround train to have a dwell time of 15 min and
set a dwell time of 3 min for halting trains. We assumed the meal time for lunch to be 11:0013:00 and that for supper to be
16:0020:00. The method used to estimate the demand for each trip is presented in Appendix A. Although these values were
arbitrarily set, they do not affect the applicability of our method.
5.1. Small-scale case study
To illustrate the performance of proposed method, the results of the HCEA are compared with the solutions obtained
using the RM based on a hypothetical network with 8 stations, 6 DCs and 2 FSs in Fig. 11. The transportation costs and travel
times between the nodes are shown. A train timetable with 4 types of lines is also displayed. The cost to open a DC is 50
yuan/day, and the cost to rent one unit of capacity is 3 yuan/day. For all p and d, UCapd 10 units, K d 200,
HT d 150 min Cd 3000 units, Capp 30000 units, and Cp 6000 units, MST s 300 min, and Lspa 1440 min.
We first solved the RM to obtain a lower bound on the solution to the 3E-LRPTWTBC. Then, the HCEA was executed to
solve the 3E-LRPTWTBC with parameters of g 500, q 0:5, S1 20, S2 100, and e 0:01. The computed results of the
RM and the 3E-LRPTWTBC are compared in Table 10. The objective values for the two models are denoted by OPTL and
OPT, respectively. The CPLEX solver obtained the optimal solution to the RM in 0.05 s. The HCEA yielded a solution to the
3E-LRPTWTBC after running for 100 generations (540 s). We found that 12 maximum iterations (64.8 s) were sufficient
for the HCEA to obtain a solution with a 9.3% deviation from the objective value for the RM. This gap was calculated as follows: % gap RM = 100OPT-OPTL =OPT. As indicated by Table 10, the difference between the results obtained for the RM and
the 3E-LRPTWTBC mainly arises from the costs for transportation in urban areas. Compared with the solution obtained using
the RM, the distribution system obtained using the 3E-LRPTWTBC uses four RCs in total, with a total transportation cost of
120 yuan. Thus, we find the following:
P
P

Dtxd5 t
t2N T Dtxd2 t
25:28 6 30 4 120
2 min tcd2 ;s
min tcd5 ;s t2T
s2N S d2
Cd 2
s2N S d5
Cd5
In this case study, both d2 and d5 use two RCs to facilitate urban delivery. The routes of each vehicle are shown in Fig. 11.
The parameters (a, b, c) are shown for each node in Fig. 12, where the first value denotes the time schedule of an RC, the
second value indicates the MAT from the node to the destination of one of its accessible trips, and the third value indicates
the amount of product passing through the node. The capacity utilization ratios of the RCs and DCs are also shown in Fig. 11.
We find that the capacity ratios for the RCs are quite small (all < 0.5). This finding demonstrates the influence of the time
windows. The maximum lead times from p1 and p2 are explicitly calculated. All catering services are provided within the
given deadline of 1440 min. However, the maximum lead times from both p1 and p2 are longer than the time consumptions
calculated using the RM:
g p1 400 6 g p1 686;
g p2 400 6 g p2 628
Two conclusions can thus be drawn. (1) The 3E-LRPTWTBC more explicitly calculates the transportation costs in urban
areas compared with the RM, which is meaningful for network design. (2) More precise estimates of time consumption
are also obtained using the 3E-LRPTWTBC compared with the RM, which is critical in the CSHR context.
An exhaustive enumeration algorithm (EEA) with a 01 search tree with 310 28 nodes at the final depth was used to
obtain a strategic network solution. The TNNH was again used in the sub-routing problem. The computed results obtained
using the EEA + TNNH and the HCEA are compared in terms of computation time and objective value in Table 11.
235
Fig. 11. A hypothetical network and train timetable.
Table 10
Comparative analysis of the distribution system planning results obtained using the two models.
CPLEX for the RM
HCEA for the 3E-LRPTWTBC
CPLEX for the RM
HCEA for the 3E-LRPTWTBC
zd1
zd2
zd3
zd4
zd5
zd6
kd1
kd2
0
1
0
0
1
0
0
166 units
0
1
0
0
1
0
0
179 units
kd3
kd4
kd5
kd6
f p1 d2
f p2 d5
g p1 =g p1
g p2 =g p2
0
0
87 units
0
1659 items
869.8 items
400
400
0
0
75 units
0
1783.3 items
745.5 items
686
628
Line haul
Rental cost
65.32 yuan
759 yuan
65.31 yuan
762 yuan
Urban transport cost

Opening cost
25.28 yuan
100 yuan
120 yuan
100 yuan
OPT and OPTL
949.61 yuan
1047.3 yuan
% gap RM
9.3
Fig. 12. The time schedule for each RC in the illustrative example.
236

Table 11
Comparative analysis of the two algorithms.
Algorithm
HCEA (CEA + TNNH)
EEA + TNNH
CPU time consumption

Evaluation function value
64.8 s
1047 yuan
45.3 h
1047 yuan
Table 12
Capabilities of FSs (measured in thousands of units).
ID
FS
Capp
ID
FS
Capp
ID
FS
Capp
1
2
3
4
5
6
7
Beijing FS
Chengdu FS
Fuzhou FS
Guangzhou FS
Harbin FS
Hangzhou FS
Hefei FS
90
20
50
90
20
50
50
8
9
10
11
12
13
14
Jinan FS
Nanchang FS
Nanjing FS
Nanning FS
Shanghai FS
Shenyang FS
Shijiazhuang FS
50
20
60
20
90
20
20
15
16
17
18
19
20
21
Taiyuan FS
Wuhan FS
Xi an FS
Changchun FS
Changsha FS
Zhengzhou FS
Chongqing FS
20
50
20
20
50
50
20
Table 11 records the best solution obtained by the EEA after running 45.3 h (not necessarily global optimum) and the
HCEA produces the same solution. However, an increase in the number of trips to be catered results in an exponential
increase in computation time for the EEA. The results indicate that the HCEA provides a reasonably good solution in terms
of both computation time and solution quality.
5.2. Application of the HCEA in two large-scale cases
There are many FSs that produce or purchase food products at multiple production sites in China. These products are then
sold to DCs in different cities. Many additional DCs are planned to be built across China to obtain food products from various
FSs and deliver them to stations on rail networks. In this study, we focus on the design of a distribution system for cold
chain meals with a 24-h lifespan for developing CSHR markets to offer attractive passenger services. After the DCs have been
built, a schedule is generated to inform the RCs of where and when to pick up and deliver food products. The orders and
distribution plans of the DCs are defined based on the meals required for trips specified by train timetables.
In this paper, we provide relevant information regarding 21 potential production sites in China. Each of these sites can
produce/purchase cold chain meals with a shelf life of 24 h (1440 min). The cooperative FSs and their daily capacities
Capp are listed in Table 12.
For all p and d, UCapd 10 units, K d 5000, HT d 150 min, Cd 3000 units, Cp 6000 units, MST s 300 min, and
Lspa 1440 min. The method used to estimate the transportation cost and travel time for each link is presented in Appendix
B. For the application of our method to realistic scenarios, the micro-locations of the potential DCs in a city and careful calibration of the demand functions for each trip are important because they may impact the outputs of the DSDP-CSHR. For
this study, we relied on the approximate method presented in Appendices AC; however, we expect to further address this
issue in future work. Then, we implemented our experiments based on the network depicted in Fig. 13.
Case 1 is based on a timetable for June 2015, consisting of 156 trips on the Beijing-Shanghai high-speed corridor. For this
case, we considered 31 stations on the Beijing-Shanghai corridor, as listed in Table A.1 in Appendix C. The DCs corresponding
to these stations are also listed in Table A.1, as are the OC d and UC d values for the DCs. A planning period of one day was
considered. The DCs were labeled using the names of the cities. Furthermore, only trips starting and ending at stations on
the Beijing-Shanghai high-speed line were considered in this case. All over-line trips were neglected. Finally, only four
FSs were considered: Beijing FS, Jinan FS, Nanjing FS and Shanghai FS.
Case 2 is based on a timetable for April 2014, consisting of 1130 trips on Chinas high-speed rail network, as illustrated in
Fig. 13. Tables A.2 and A.3 list the 85 stations and 33 potential DC sites in the network that were considered in this case. The
opening and rental costs are also listed in the latter table. Although high-speed rail lines had been built in the eastern region
as of 2014, 2 unfinished exclaves were present in the western region. The separate parts of the Shanghai-WuhanChongqing high-speed railway had not been connected. The Hengyang-Liuzhou line had not been established. Moreover,
the Lanzhou-Urumqi line is not shown in Fig. 13. This may lead to an underestimation of activity at some stations, such
as Xian, Chengdu, and Nanning. Furthermore, any local lines without over-line trips were neglected (e.g., Beijing-Tianjin
and Guangzhou-Zhuhai). In this case, all 21 potential FSs were considered.
We conducted three experiments to (1) evaluate the impact of the maximum storage time (MST) at rail stations on the
decisions made based on Case 1, (2) identify the benefits of integrating system design with the sub-routing problem based on
Cases 1 and 2, and (3) perform a capacity analysis of each DC based on Cases 1 and 2.
5.2.1. Analysis of the effect of the maximum storage time
A total of 21 problems based on Case 1 were formulated to conduct analyses in terms of the MST. The HCEA was executed
with g 100, q 0:5, S1 20, S2 60, and e 0:01. First, we conducted a series of tests with different MST values. For
237
Fig. 13. Locations of alternative DCs.
every station s, the MST was set to 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 200, 250, 300, 400 and
1000. For example, MST 0 means that for all s 2 N S , MST s 0, which indicates that all trips passing through station s are
required to be catered within time windows of Ls; t DWT st ; Ls; t. MST 40 means that for all s 2 N S , MST s 40, which
indicates that the time window for each trip t at station s is Ls; t DWT st 40; Ls; t. Fig. 14 illustrates that the objective
value decreases with a decreasing number of RCs. The impact of the MST on the objective value is summarized in Fig. 14.
Second, for each MST value, Fig. 15 displays the change in the minimum objective value with respect to the number of
generations in the HCEA. The relationship between the MST and the gap between the results of the RM and the 3ELRPTWTBC is shown in Fig. 15. This gap decreases with an increasing MST. This means that the objective value of the 3ELRPTWTBC asymptotically approaches the objective value of the RM as the time windows are relaxed.
Third, we varied the MST from 0 to 1000 and calculated the total urban transportation cost in the 21 cities and the MATs
from the FSs, namely, Beijing FS, Jinan FS, Nanjing FS and Shanghai FS, as shown in Fig. 16. The results indicate that the urban
transportation cost decreases as the MST increases and that the MATs from the FSs improve. As indicated in Fig. 16, allowing
an MST at stations of 40 min decreases the urban transportation cost by 50% (from 6000 yuan to 3000 yuan). However, the
incidental improvement in the MATs from the FSs has a negative effect on pursuing a short lead time (or a high food quality).
5.2.2. Benefits of integrating distribution system design with the sub-routing problem
This section investigates to what extent the rail company benefits from integrating distribution system design with the
sub-routing problem. For this purpose, we compare two design concepts: sequential design and integrated design. Sequential design serves as the reference because it represents the general method for designing a distribution system. Here, in a
first decision stage, the network solutions are determined using the RM. Then, in a second stage, the RC routes from each DC
are determined using the TNNH. To find the minimum-cost solutions, we also apply the MIP solver in CPLEX v. 12.3 to solve
the RM. For comparison, the integrated design concept is applied by means of the HCEA with g 500, q 0:5, S1 20,
S2 60, and e 0:01. The MST was set to 30 min in the following experiments.
We defined a parameter / to represent a discount on the rental cost. Based on Case 1, we conducted a series of tests with
different rental costs, including / = 0.9, / = 0.8 and / = 0.7. For example, / = 0.9 indicates that the rental cost for each DC d,
UCd , is 90 percent of the value given in Table A.1. Table 13 shows that integrated design already leads to a 99 yuan cost savings compared with sequential design when / = 0.9. As the value of / decreases, greater cost savings can be obtained. Finally,
when / = 0.4, a cost savings of 475 yuan is achieved. Hence, as indicated by Table 13, the sum of the individual costs corresponding to each of the two sequential optimization steps is larger than the overall cost of integrated design implemented
using the HCEA.
238
Fig. 14. Relationship between the MST and the objective value/number of RCs.
Fig. 15. Relationship between the MST and the gap between the results of the 3E-LRPTWTBC and the RM.
Fig. 16. Urban transportation cost and MATs versus the MST.
Let / = 1. Table 14 details the best solutions found using the two approaches. This table includes the total cost of the system together with the relative change with respect to the reference solution obtained using solely the RM. The subsequent
rows in this table show the cost contributions from DC opening, DC rental, line-haul transportation and urban transportation.
239

Table 13
Performance measures for the Case 1 solutions obtained using different design concepts and for different rental costs.
/ = 0.9
/ = 0.8
/ = 0.7
/ = 0.6
/ = 0.5
/ = 0.4
RM
Sequential design (RM + TNNH)
Urban integrated design (HCEA)
30,544
33,121
33,022
27,489
30,068
29,899
24,416
26,946
26,807
21,341
23,871
23,704
18,267
20,797
20,605
15,409
17,927
17,452
Cost savings
99
169
139
167
198
475
Table 14
Cost configurations for solutions obtained for both Case 1 and Case 2 using different design concepts.
Case 1 (/ = 1)
Case 2 (/ = 1)
RM
Sequential design (RM + TNNH)
Integrated design (HCEA)
Cost savings
Total cost
% gap RM
33,592
36,138
7.05
36,137
7.04
1
0.01
Opening cost
Rental cost
1190
30,464
1190
30,464
1102
30,652
88
188
Line haul cost

Urban transportation cost
937
1001
937
3547
929
3454
8
93
Total cost
% gap RM
245,607
269,752
8.9
268,625
0.5
1127
0.4
Opening cost
Rental cost
3198
224,520
3198
224,520
3198
224,968
666
448
Line-haul cost
Urban transportation cost
8262
9627
8262
33,772
8328
32,131
66
1641
Fig. 17. Cost distributions in the minimum-cost solutions for Case 2.
When sequential design is replaced with integrated design using the HCEA, 0.01% and 0.4% of the total costs can be saved in
Case 1 and Case 2, respectively. To achieve these savings, an increase in rental costs is needed. Fig. 17 presents the shares of
the total cost in the minimum-cost solutions for Case 2. We observe that the urban transportation cost is an important contributor. The urban transportation cost represents nearly 13% of the total cost in sequential design, whereas it accounts for
12% of the total cost in our proposed integrated design. In other words, the HCEA is an effective method of reducing urban
transportation costs. Interestingly, integrated planning does not affect the rental cost share compared with sequential planning. This is because integrated planning does not shift the DCs to sites with lower accessibility to trips and FSs.
5.2.3. Capacity analysis of distribution centers
Fig. 18 depicts the capacity of and number of meals provided by each DC in Case 1. In Case 1, trips on the Beijing-Shanghai
high-speed corridor are considered, neglecting over-line trips. In this case, it is assumed that all cities on the corridor serve as
240
Fig. 18. Capacity distribution among the DCs based on Case 1.
Fig. 19. Capacity distribution among the DCs based on Case 2.
potential DC sites. We find that two large DCs are operated in Shanghai and Nanjing. This is explained by the meal demands
for local trips between Shanghai and Nanjing, which have journey times of 80120 min, and other intercity trips.
Fig. 19 similarly depicts the capacity of and number of meals provided by each DC in Case 2. In Case 2, we consider only 33
important cities as potential sites for DCs. We observe that the solutions obtained using both the RM and the 3E-LRPTWTBC
indicate that all of these cities should provide some capacity for CSHR. The four largest DCs are in Shanghai, Beijing, Guangzhou, and Wuhan. In Case 1, Nanjing and Beijing have nearly the same capacity. Here, from the perspective of the entire network, the capacity required in Beijing exceeds the capacity required in Nanjing. Notably, our solution underestimates the
importance of Nanning, Xian, Chongqing, etc. because of the shortcomings of our data, as mentioned above.
6. Conclusions
In this paper, we propose two models for addressing the distribution system design problem in the context of catering
services for high-speed railways (DSDP-CSHR). The DSDP-CSHR is fundamentally different from other DSDP models for
the following reasons. First, the problem is influenced by decisions concerning the rail transport plan (RTP), e.g., the line plan,
the train timetable and the EMU circulation plan. Thus, the routing problem for refrigerated cars (RCs) transporting food
products in urban areas should be introduced into the distribution system design to integrate the passenger-demand-side
aspects of passenger train service. The RCs from the distribution centers (DCs) must serve the trips specified in the timetable,
which are treated as the customers, within the time windows imposed by the RTP. Our problem also involves a limited lifespan for the food products, which is enforced by means of time-deadline constraints.
241
To formulate the problem, we introduce the information provided by the line plan and train timetable into a conventional
distribution system model, which is called the physical distribution network (PDN). Based on the PDN, a planner attempts to
generate a network solution. Next, an operational distribution network (ODN), which represents the RC routing choices for
urban transportation, is derived. An operation scheme is obtained based on the ODN. Then, the time deadlines are reformulated as time budget constraints on a rooted forest corresponding to the network solution and the operation scheme. Based
on these concepts, this study makes three contributions to the literature.
(1) The paper proposes a model, termed the three-echelon location routing problem with time windows and time budget
constraints (3E-LRPTWTBC), in the context of CSHR. Based on the PDN and the ODN, the 3E-LRPTWTBC jointly determines (i) the locations of the DCs, (ii) the capacity plan for the DCs, (iii) the food supplier (FS) contracted by each DC,
(iv) the station at which each trip in the train timetable is served, (v) the volumes of food products transferred
between the FSs and DCs, (vi) the detailed routes and schedules for the RCs, and (vii) the maximum accumulated travel
times (MATs) from the FSs to their accessible destinations.
(2) The paper also describes a relaxation model (RM). This model can be solved using exact solution methods and provides
a lower bound on the solution to the 3E-LRPTWTBC. A hybrid cross entropy algorithm (HCEA), which integrates the
cross entropy algorithm (CEA) with the time-oriented nearest-neighbor heuristic (TNNH), is proposed to solve the
3E-LRPTWTBC. It is also demonstrated how this metaheuristic algorithm can be helped to escape from local minima
via some simple modifications.
(3) The proposed approach is applied to one hypothetical example and two real-world cases. Based on the hypothetical
case, we demonstrate the advantages of considering time windows by performing a comparative analysis of the 3ELRPTWTBC and the RM. Based on the two large-scale cases, we conduct a parametric analysis based on the maximum
storage time (MST) at stations and illustrate the importance of determining a proper MST. We also reveal the cost savings achieved by means of the HCEA and the role of urban transportation. The capacity distributions are also
presented.
The presented approach still has limitations, which may be addressed in future research. For example, the model could be
extended to address uncertain passenger demands, or the relaxation of the assumption that each DC can be supplied by only
one FS. In addition, the possibility of multiple catering instances for a given trip is also beyond the scope of the current
research.
Acknowledgments
We appreciate the insightful comments from a number of scholars, especially professor Ingo Hansen at Delft University of
Technology and associate professor Xuesong Zhou at Arizona State University. This research was supported by the following
funds: Fundamental Research Funds for the Central Universities (Beijing Jiaotong University) (Grant No. 2013YJS050 and
Grant No. 2014JBZ008), Evaluation and system research on line planning for high-speed railways (Grant No. 2014X010A), and Research on passenger service management and technology for the high-speed railway system - a study of extended
services and their quality guarantees (Grant No. 2011X014-E). The work presented in this paper remains the sole responsibility of the authors.
Appendix A. The method used to estimate the demand for each trip
The demand for trip t is estimated as follows:

Dt

kst
dst
DRLt
DRSt SCtxt 8t 2 N T
ELT SLT
EST SST
a:1
Let kst be the overlap between the journey time of t and the meal time for lunch, and let dst be the overlap between the
journey time of t and the meal time for supper. SLT and SST denote the start times of lunch and supper, respectively. ELT and
EST denote the end times of lunch and supper, respectively. DRLt and DRSt denote the lunch rate and supper rate, respectively, on trip t; SCt is the seat capacity of the train serving trip t; and xt is the load factor for trip t.
Throughout this paper, we assume that the meal time for lunch is 11:0013:00 and that the meal time for supper is
16:0020:00, i.e., SLT = 660 min (11:00), ELT = 780 min (13:00), SST = 960 min (16:00) and EST = 1200 min (20:00). Thus,
in Fig. 3, we have kst1 90 min, dst1 30 min; kst 2 0 min, dst2 130 min, etc. In all our experiments, we set
DRLt 0:35; DRSt 0:35 in the small case study, and DRSt 0:4 in the other case studies; SCt 1200 persons;
and xt 1.
Appendix B. The method used to estimate the transportation cost and travel time for each link
(1) The costs for line-haul transportation between two cities for each vehicle were estimated based on the road distance
between the two cities using data from Baidu Maps for 2015. Let the line-haul transport rate be 0.62 yuan/km.
242
Distance km Road distance between two cities
p
Built-up land area of the city K
Maximum holding time in the warehouse

where K is a factor for converting the holding time into an equivalent transportation distance.
Travel time min
Distance km 60
60 km=h
Transportation cost yuan 0:62 yuan=km Distance km

(2) In the cases considered in our study, there are at most three stations in each city. Consequently, we can assume that the
p
distance between any two stations in a given city is equal to Built-up land area of the city (analogous to an equilateral
triangle). The built-up land area for each city was gathered from the China City Statistical Yearbook (2012). In an urban area,
let the transport rate be 3 yuan/km.
Travel time between two stations
p
Built-up land area of the city km 60
20 km=h
p
Transportation cost between two stations yuan 3 yuan=km Built-up land area of the city km
Thus, the distance from a DC to the stations it covers can be estimated using the law of cosines as follows:
s
2
Built-up land area of the city
km
Distance from DC d to its covered stations
21 cos2p=Numd
where Numd is the number of stations covered by DC d.
s
2
60
Travel time from DC d to its covered stations
km
20 km=h
21 cos2p=Numd
s
2
Transportation cost from DC d to its covered stations 3 yuan=km
km
21 cos2p=Numd
Appendix C. Case data

See Tables A.1A.3.
Table A.1
Stations considered in Case 1 and information related to their corresponding DCs.
Station
ID
Name of
station
DC
ID
Name of
DC
OC d
(yuan)
UC d
(yuan)
Station
ID
Name of station
DC
ID
Name of
DC
OC d
(yuan)
UC d
(yuan)
Bengbu South
Bengbu
906
50
16
12
Shanghai
1845.2
69
2
3
Beijing South
Cangzhou
West
Changzhou
Changzhou
North
Danyang
Danyang
South
Dezhou East
Dingyuan
Jinan West
Langfang
Nanjing
Nanjing South
Qufu East
Shanghai
2
3
Beijing
Cangzhou
1799.1
878.8
66
46
17
18
Shanghai
Hongqiao
Suzhou
Suzhou North
13
13
Suzhou
Suzhou
1080.1
1080.1
58
58
4
4
Changzhou
Changzhou
1080.1
1080.1
58
58
19
20
Suzhou Yuanqu
Taian
13
14
Suzhou
Taian
1080.1
938.5
58
14
5
5
Danyang
Danyang
1080.1
1080.1
58
58
21
22
Tengzhou East
Tianjin South
15
16
Tengzhou
Tianjin
938.5
1411.6
15
16
6
7
8
9
10
10
11
12
Dezhou
Dingyuan
Jinan
Langfang
Nanjing
Nanjing
Qufu
Shanghai
938.5
878.8
1080.1
938.5
1080.1
1080.1
938.5
938.5
53
50
53
46
58
58
53
69
23
24
25
26
27
28
29
30
31
Tianjin West
Wuxi
Wuxi East
Wuxi Xinqu
Suzhou East
Xuzhou East
Zaozhuang
Zhenjiang
Zhenjing South
16
17
17
17
18
19
20
21
21
Tianjin
Wuxi
Wuxi
Wuxi
Suzhou
Xuzhou
Zaozhuang
Zhenjiang
Zhenjiang
1411.6
1080.1
1080.1
1080.1
906.0
1080.1
938.5
1080.1
1080.1
16
17
17
17
18
19
20
21
21
4
5
6
7
8
9
10
11
12
13
14
15
243

Table A.2
Stations considered in Case 2 and their corresponding DCs.
ID
Station
DC
ID
Station
DC
ID
Station
DC
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
Bengbu South
Baoji South
Beidaihe
Beijing
Beijing South
Beijing West
Cangnan
Changzhou
Chengdu
Chengdu East
Dalian
Dalian North
Fuding
Fuzhou
Fuzhou South
Guangzhou
Guangzhou East
Guangzhou South
Guilin
Harbin West
Handan East
Hankou
Hangzhou
Hangzhou East
Hefei
Huainan East
Jilin
Jinan
Bengbu
Baoji
Beidaihe
Beijing
Beijing
Beijing
Cangnan
Changzhou
Chengdu
Chengdu
Dalian
Dalian
Fuding
Fuzhou
Fuzhou
Guangzhou
Guangzhou
Guangzhou
Guilin
Harbin
Handan
Hankou
Hangzhou
Hangzhou
Hefei
Huainan
Jilin
Jinan
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
Jinan West
Jiangshan
Jinhua West
Jingzhou
Jiujiang
Lichuan
Lingbao West
Liuan
Longyan
Luoyang Longmen
Nanchang
Nanchang West
Nanjing
Nanjing South
Nanning
Ningbo
Qinhuangdao
Qingdao
Qingdao North
Sanming North
Xiamen North
Shangqiu
Shanghai
Shanghai Hongqiao
Shenzhen
Shenzhen North
Shenyang
Shenyang North
Jinan
Jiangshan
Jinhua
Jingzhou
Jiujiang
Lichuan
Lingbao
Liuan
Longyan
Luoyang
Nanchang
Nanchang
Nanjing
Nanjing
Nanning
Ningbo
Qinhuangdao
Qingdao
Qingdao
Sanming
Xiamen
Shangqiu
Shanghai
Shanghai
Shenzhen
Shenzhen
Shenyang
Shenyang
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
Shiyan
Shijiazhuang
Suzhou
Taiyuan
Taining
Tianjin
Tianjin South
Tianjin West
Wenzhou South
Wuchang
Wuhan
Xi an north
Xiangyang
Xinxiang East
Xinyang East
Xuzhou
Xuzhou East
Yanan
Yichang East
Yiwu
YueYang East
Zhangzhou
Changchun
Changchun West
Changsha
Changsha South
Zhengzhou
Zhengzhou East
Chongqing North
Shiyan
Shijiazhuang
Suzhou
Taiyuan
Taining
Tianjin
Tianjin
Tianjin
Wenzhou
Wuchang
Wuhan
Xi an
Xiangyang
Xinxiang
Xinyang
Xuzhou
Xuzhou
Yanan
Yichang
Yiwu
YueYang
Zhangzhou
Changchun
Changchun
Changsha
Changsha
Zhengzhou
Zhengzhou
Chongqing
Table A.3
Information related to the DCs considered in Case 2.
DC ID
1
2
3
4
5
6
7
8
9
10
11
Bengbu
Beijing
Chengdu
Dalian
Fuzhou
Guangzhou
Guilin
Harbin
Hangzhou
Hefei
Jinan
OC d (yuan)
UC d (yuan)
DC ID
906
1799.1
640.0
800.0
820.0
1610
900
910
1080
1080
1080
50
66
49
57
51
69
46
50
61
50
53
12
13
14
15
16
17
18
19
20
21
22
Jiujiang
Nanchang
Nanjing
Nanning
Ningbo
Qingdao
Xiamen
Shanghai
Shenzhen
Shenyang
Shijiazhuang
OC d (yuan)
UC d (yuan)
DC ID
860
750
1080
900
830
710
820
1845.1
810
800
670
55
55
57
46
60
53
51
69
69
57
47
23
24
25
26
27
28
29
30
31
32
33
Taiyuan
Tianjin
Wenzhou
Wuhan
Xi an
Xuzhou
Yichang
Changchun
Changsha
Zhengzhou
Chongqing
OC d (yuan)
UC d (yuan)
650
1411.6
830
1411.6
830
720
760
740
1080
1080
770
41
57
61
54
44
58
55
54
52
43
52
References
Adenso-Daz, B., Gonzlez, M., Garca, E., 1998. A hierarchical approach to managing dairy routing. Interfaces 28 (2), 2131. http://dx.doi.org/10.1287/
inte.28.2.21.
Ahmadi-Javid, A., Seddighi, A.H., 2013. A location-routing problem with disruption risk. Transport. Res. Part E 53, 6382. http://dx.doi.org/10.1016/j.
tre.2013.02.002.
Akkerman, R., Farahani, P., Grunow, M., 2010. Quality, safety and sustainability in food distribution: a review of quantitative operations management
approaches and challenges. OR Spectrum 32 (4), 863904. http://dx.doi.org/10.1007/s00291-010-0223-2.
Belenguer, J.M., Benavent, E., Martnez, M.C., 2005. RutaRep: a computer package to design dispatching routes in the meat industry. J. Food Eng. 70 (3), 435
445. http://dx.doi.org/10.1016/j.jfoodeng.2004.02.045.
Berger, R.T., Coullard, C.R., Daskin, M.S., 2007. Location-routing problems with distance constraints. Transport. Sci. 41 (1), 2943. http://dx.doi.org/10.1287/
trsc.1060.0156.
Bruns, A., Klose, A., Sthly, P., 2002. Restructuring of Swiss parcel delivery services. Oper. Res. Spectrum 22, 285302.
Bussieck, M.R., 1998. Optimal Lines in Public Rail Transport PhD Thesis. Technical University Braunschweig, Brauschweg, Germany.
Chen, H., Hsueh, C., Chang, M., 2009. Production scheduling and vehicle routing with time windows for perishable food products. Comput. Oper. Res. 36 (7),
23112319. http://dx.doi.org/10.1016/j.cor.2008.09.010.
China City Statistical Yearbook, 2012. China Statistics Press, Beijing.
Chung, H.K., Norback, J.P., 1991. A clustering and insertion heuristic applied to a large routing problem in food distribution. J. Oper. Res. Soc. 42 (7), 555564.
http://dx.doi.org/10.1057/jors.1991.114.
Contreras, I., Fernndez, E., Reinelt, G., 2012. Minimizing the maximum travel time in a combined model of facility location and network design. Omega 40,
847860. http://dx.doi.org/10.1016/j.omega.2012.01.006.
244
De Boer, P., Kroese, D.P., Mannor, S., Rubinstein, R.Y., 2005. A tutorial on the cross-entropy method. Ann. Oper. Res. 134 (1), 1967. http://dx.doi.org/
10.1007/s10479-005-5724-z.
Federgruen, A., Prastacos, G., Zipkin, P.H., 1986. An allocation and distribution model for perishable products. Oper. Res. 34 (1), 7582. http://dx.doi.org/
10.1287/opre.34.1.75.
Fu, H., Nie, L., Meng, L., Sperry, B.R., He, Z., 2015. A hierarchical line planning approach for a large-scale high speed rail network: the China case. Transport.
Res. Part A 75, 6183. http://dx.doi.org/10.1016/j.tra.2015.03.013.
Gelders, L.F., Pintelon, L.M., Van Wassenhove, L.N., 1987. A location-allocation problem in a large Belgian brewery. Eur. J. Oper. Res. 28 (2), 196206. http://
dx.doi.org/10.1016/0377-2217(87)90218-9.
Geoffrion, A.M., 1974. Lagrangian relaxation for integer programming. In: Approaches to Integer Programming. Springer, Berlin, pp. 82114.
Goossens, J., van Hoesel, S., Kroon, L., 2006. On solving multi-type railway line planning problems. Eur. J. Oper. Res. 168 (2), 403424. http://dx.doi.org/
10.1016/j.ejor.2004.04.036.
Goto, J.H., Lewis, M.E., Puterman, M.L., 2004. Coffee, tea, or ellipsis? A Markov decision process model for airline meal provisioning. Transport. Sci. 38 (1),
107118. http://dx.doi.org/10.1287/trsc.1030.0022.
Govindan, K., Jafarian, A., Khodaverdi, R., Devika, K., 2014. Two-echelon multiple-vehicle location-routing problem with time windows for optimization of
sustainable supply chain network of perishable food. Int. J. Prod. Econ. 152, 928. http://dx.doi.org/10.1016/j.ijpe.2013.12.028.
Ho, S.C., Leung, J.M.Y., 2010. Solving a manpower scheduling problem for airline catering using metaheuristics. Eur. J. Oper. Res. 48 (3), 903921.
Hsu, C., Hung, S., Li, H., 2007. Vehicle routing problem with time-windows for perishable food delivery. J. Food Eng. 80 (2), 465475. http://dx.doi.org/
10.1016/j.jfoodeng.2006.05.029.
Klose, A., Drexl, A., 2005. Facility location models for distribution system design. Eur. J. Oper. Res. 162 (1), 429. http://dx.doi.org/10.1016/j.
ejor.2003.10.031.
Liu, J., Zhou, X., 2016. Capacitated transit service network design with boundedly rational agents. Transp. Res. Part B 93, 225250. http://dx.doi.org/10.1016/
j.trb.2016.07.015.
Mahmoudi, M., Zhou, X., 2016. Finding optimal solutions for vehicle routing problem with pickup and delivery services with time windows: a dynamic
programming approach based on statespacetime network representations. Transport. Res. Part B 89, 1942. http://dx.doi.org/10.1016/j.
trb.2016.03.009.
Meisel, F., Kirschstein, T., Bierwirth, C., 2013. Integrated production and intermodal transportation planning in large scale production-distributionnetworks. Transport. Res. Part E 60, 6278.
Melo, M.T., Nickel, S., Saldanha-da-Gama, F., 2009. Facility location and supply chain managementa review. Eur. J. Oper. Res. 196 (2), 401412. http://dx.
doi.org/10.1016/j.ejor.2008.05.007.
Meng, Q., Wang, X.C., 2011. Intermodal hub-and-spoke network design: incorporating multiple stakeholders and multi-type containers. Transport. Res. Part
B 45, 724742.
Miranda, P.A., Garrido, R.A., 2004. Incorporating inventory control decisions into a strategic distribution network design model with stochastic demand.
Transport. Res. Part E 40 (3), 183207. http://dx.doi.org/10.1016/j.tre.2003.08.006.
Miranda, P.A., Garrido, R.A., 2008. Valid inequalities for Lagrangian relaxation in an inventory location problem with stochastic capacity. Transport. Res. Part
E 44 (1), 4765. http://dx.doi.org/10.1016/j.tre.2006.04.002.
Montoya, A., Guret, C., Mendoza, J.E., Villegas, J.G., 2015. A multi-space sampling heuristic for the green vehicle routing problem. Transport. Res. Part C 70,
113128. http://dx.doi.org/10.1016/j.trc.2015.09.009.
Nagurney, A.Y., Yu, M., Masoumi, A.H., Nagurney, L.S., 2013. Network against Time: Supply Chain Analytics for Perishable Products. Springer-Verlag, New
York.
Nagy, G., Salhi, S., 2007. Location-routing: issues, models and methods. Eur. J. Oper. Res. 177 (2), 649672. http://dx.doi.org/10.1016/j.ejor.2006.04.004.
Peeters, L., 2003. Cyclic Railway Timetable Optimization Thesis. Erasmus Research Institute of Management, Erasmus University Rotterdam.
Peeters, M., Kroon, L., 2008. Circulation of railway rolling stock: a branch-and-price approach. Comput. Oper. Res. 35 (2), 538556. http://dx.doi.org/
10.1016/j.cor.2006.03.019.
Perl, J., Daskin, M., 1985. A warehouse location-routing problem. Transport. Res. Part B 19 (5), 381396.
Pooley, J., 1994. Integrated production and distribution facility planning at Ault foods. Interfaces 24 (4), 113121. http://dx.doi.org/10.1287/inte.24.4.113.
Prins, C., Lacomme, P., Prodhon, C., 2014. Order-first split-second methods for vehicle routing problems: a review. Transport. Res. Part C 40, 179200.
Ruan, J.M., Liu, B., Wei, H., Qu, Y., Zhu, N., Zhou, X., 2016. How many and where to locate parking lots? A spacetime accessibility-maximization modeling
framework for special event traffic management. Urban Rail Transit 2 (2), 5970. http://dx.doi.org/10.1007/s40864-016-0038-9.
Science and Technology Division of Chinas Ministry of Railways, 2010. Business administration pattern of catering service on EMU trains.
Shen, M.Z.J., Coullard, C., Daskin, M.S., 2003. A joint location-inventory model. Transport. Sci. 37 (1), 4055. http://dx.doi.org/10.1287/trsc.37.1.40.12823.
Shu, J., Teo, C., Shen, Z.M., 2005. Stochastic transportation-inventory network design problem. Oper. Res. 53 (1), 4860. http://dx.doi.org/10.1287/
opre.1040.0140.
Solomon, M.M., Desrosiers, J., 1988. Survey paper time window constrained touring and scheduling problem. Transport. Sci. 22 (1), 113. http://dx.doi.org/
10.1287/trsc.22.1.1.
Tas, D., Dellaert, N., Van Woensel, T., de Kok, T., 2014. The time-dependent vehicle routing problem with soft time windows and stochastic travel times.
Transport. Res. Part C 48, 6683. http://dx.doi.org/10.1016/j.trc.2014.08.007.
Toth, P., Vigo, D., 2002. The Vehicle Routing Problem. Society for Industrial and Applied Mathematics, Philadelphia.
Tong, L., Zhou, X., Miller, H.J., 2015a. Transportation network design for maximizing spacetime accessibility. Transport. Res. Part B 81, 555576. http://dx.
doi.org/10.1016/j.trb.2015.08.002.
Tong, L., Nie, L., He, Z., Fu, H., 2015b. Optimization of train trip package operation scheme. Math. Prob. Eng. 2015, 472591. http://dx.doi.org/10.1155/2015/
472591.
Verma, M., Verter, V., Zufferey, N., 2012. A bi-objective model for planning and managing rail-truck intermodal transportation of hazardous materials.
Transport. Res.: Part E: Logist. Transport. Rev. 48 (1), 132149. http://dx.doi.org/10.1016/j.tre.2011.06.001.
Watson-Gandy, C., Dohrn, P., 1973. Depot location with van salesmena practical approach. OMEGA Int. J. Manage. Sci. 1 (3), 321329. http://dx.doi.org/
10.1016/0305-0483(73)90108-4.
Wouda, F.H.E., Van Beek, P., Van der Vorst, J.G.A.J., Tacke, H., 2002. An application of mixed-integer linear programming models on the redesign of the supply
network of Nutricia Dairy & Drinks Group in Hungary. OR Spectrum 24 (4), 449465. http://dx.doi.org/10.1007/s002910200112.
Wu, X., Nie, L., Xu, M., 2015. Service station evaluation problem in catering service of high-speed railway: a fuzzy QFD approach based on evidence theory.
Math. Prob. Eng. 2015, 404926. http://dx.doi.org/10.1155/2015/404926.
Zhang, G., Habenicht, W., Spie, W.E.L., 2003. Improving the structure of deep frozen and chilled food chain with tabu search procedure. J. Food Eng. 60 (1),
6779. http://dx.doi.org/10.1016/S0260-8774(03)00019-0.
Zhang, X., Nie, L., 2016. Integrating capacity analysis with high-speed railway timetabling: a minimum cycle time calculation model with flexible overtaking
constraints and intelligent enumeration. Transport. Res. Part C 68, 509531. http://dx.doi.org/10.1016/j.trc.2016.05.005.

Transportation Research Part C: Xin Wu, Lei Nie, Meng Xu

Enviado por

Dados do documento

Título original

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Transportation Research Part C: Xin Wu, Lei Nie, Meng Xu

Enviado por

Direitos autorais:

Formatos disponíveis

Transportation Research Part C 74 (2017) 212244

Contents lists available at ScienceDirect

Transportation Research Part C

Designing an integrated distribution system for catering

School of Traffic and Transportation, Beijing Jiaotong University, 100044, PR China

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

1.2. Contributions and objectives

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

2. Problem statement with network representations

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

Fig. 2. Illustration of the food distribution system for CSHR.

2.1. Notation and assumptions

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

Sets in the PDN

nodes in the PDN

Sets in the ODN

nodes in the ODN

Indexes used in both the PDN and ODN

general node indexes in both the PDN and ODN

Indexes used only in the ODN

The indexes for dummy nodes in the ODN

The lifespan of the cold chain meals (min)

Parameters for FSs

The capacity of FS p (units)

If ds; t 1, trip t can be catered at station s; otherwise, it cannot be catered at station s

opening cost of a potential DC d (yuan)

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

Binary variable; =1 if DC d 2 N D is established, =0 otherwise

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

1 if trip t can be catered at station s

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

(2) Determining the capacities of the DCs:

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

(5) Selecting the stations from set N S to be served by each DC:

Es; t Ls; t DWT s;t MST s

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

Fig. 4. An example of the time windows imposed by an EMU circulation plan.

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

TT s;s0 xst xs0 t0

TC s;s0 xst xs0 t0

d2N D s2N S ds0 2N S d

d2N D s2N S ds0 2N S d

where s and s0 are different stations served by the same DC d.

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

Fig. 6. An example of an ODN representation and a possible operation scheme.

(2) The schedules of the RCs from any DCs:

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

Fig. 7. Two simplified two-echelon network solutions.

g d tt p;d ypd M1 ypd 6 g p

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

Mathematical description in terms of r and o

The travel time from FS p to DC d

The journey time from the service station to the

s2N S Ls; txst

Dtxst qds 6 UCapd kd

X. Wu et al. / Transportation Research Part C 74 (2017) 212244

TT s;s0 xst xs0 t0 8t; t 0 2 N T

TC s;s0 xst xs0 t0 8t; t 0 2 N T

sdi v tti;j sdj v 6 1 rdijv M 8i; j 2 A n APD ; v 2 Vd; 8d 2 N D

sdi v tti;j sdj v P 1 rdijv M 8i; j 2 A n APD ; v 2 Vd; 8d 2 N D

X. Wu et al. / Transportation Research Part C 74 (2017) 212244