Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Expert Systems With Applications 223 (2023) 119850

Contents lists available at ScienceDirect

Expert Systems With Applications


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

Hub-and-spoke network design for container shipping in inland waterways


Saiqi Zhou a , Bin Ji a ,∗, Yalong Song a , Samson S. Yu b , Dezhi Zhang a , Tom Van Woensel c
a
School of Traffic & Transportation Engineering, Central South University, Changsha, 410075, Hunan Province, China
b
School of Engineering, Deakin University, Waurn Ponds, 3216, Victoria, Australia
c
Department of Industrial Engineering and Innovation Sciences, Eindhoven University of Technology, Eindhoven, The Netherlands

ARTICLE INFO ABSTRACT

Keywords: The enormous growth in the volume of river transportation and the upgrade of inland waterways make inland
Hub-and-spoke network design shipping more and more important in transportation and logistics. This work proposes a hub-and-spoke network
Inland waterway container shipping design (HSND) for container shipping in inland waterways based on the tree-like structure river. Firstly, the
Integer linear programming model
characteristics of the hub-and-spoke network in the inland waterways are presented in detail. Then, an integer
Decomposition-based math-heuristic
linear programming model is proposed to simultaneously determine the optimal hub location, feeder port
Enhanced genetic algorithm
allocation, and fleet deployment to minimize the total cost of ships, transportation, and transshipment. A
decomposition-based math-heuristic method and an enhanced genetic algorithm are then proposed to solve
the HSND. A case study based on the traffic on the Yangtze River and extensive computational experiments
are conducted to verify the effectiveness of the proposed models and methods. In addition, the impacts of the
number of hub ports and the economies of scale resulting from the hub-and-spoke network are investigated
from the economic and network structure perspectives, which leads to some managerial insights.

1. Introduction and the navigability on diverse segments is restricted by many con-


ditions (Bu & Nachtmann, 2021). With the upgraded waterways and
Inland waterway shipping, as a means of environment-friendly and the development of the regional economy on tributaries, evidenced
economical transportation, plays increasingly important roles in mod- by Jiang et al. (2018), Li et al. (2021), Radojčić et al. (2021) and Tang
ern transportation and logistics systems and has thus received extensive et al. (2022), incorporating the tributary system into the design of
research and industry attention. For example, according to the Statisti- the shipping network became an integral part of the comprehensive
cal Bulletin of Transportation Industry Development (https://xxgk.mot.
development of an inland waterway transportation system. Inland wa-
gov.cn), the navigable length of inland waterways in China reached
terway shipping in this study aims to serve a tree-like river structure
127,700 kilometers in 2020, which still increases year by year. The
consisting of a trunk waterway and several tributaries. In this river
cargo throughput of 300,000 TEU (twenty-foot equivalent units) was
completed in inland river ports, and investment in fixed assets of system, low water levels or lock and bridge obstacles limit the access
inland rivers reached 70.4 billion RMB, with an increase of 14.8% in of large ships on some segments, especially that of tributaries (Ji et al.,
2019. Policies and measures for developing inland waterway shipping 2022). The river streamflow’s direction affects ships’ navigation, lead-
have also been established in many countries. For example, the United ing to different transportation costs for downstream and countercurrent
States government approved the Louisiana International Gulf Transfer navigation. Container OD (origin–destination) flows are distributed in
Terminal in 2013 to support the Mississippi River and its tributaries, different regions. Ship routes may pass through several segments with
and the European Union developed the Next Generation of European different ship access restrictions and water flow directions to finish
Inland Waterway Ship and Logistics System in 2013 to support the the transportation task of the OD flows. To achieve this, the hub-and-
full integration of waterborne transport (Tan et al., 2018). All of these spoke network is designed in this study for container shipping in inland
indicate that the inland shipping system has been developed rapidly, waterways, where the hub location, feeder port allocation, and fleet
which puts forward higher requirements for shipping technology. deployment are determined simultaneously.
Unlike ocean shipping, a river system usually has a tree-like struc- The network design problems in inland waterway shipping have
ture consisting of a trunk (or mainstem) waterway and numerous
recently gained increased attention. The modes of point-to-point direct
tributaries. There is no horizontal connection among different branches,

∗ Corresponding author.
E-mail addresses: zhou_saiqi@csu.edu.cn (S. Zhou), cumtjibin@126.com (B. Ji), 214211035@csu.edu.cn (Y. Song), S.yu@ieee.org (S.S. Yu),
dzzhang@csu.edu.cn (D. Zhang), t.v.woensel@tue.nl (T. Van Woensel).

https://doi.org/10.1016/j.eswa.2023.119850
Received 22 November 2022; Received in revised form 6 March 2023; Accepted 8 March 2023
Available online 15 March 2023
0957-4174/© 2023 Elsevier Ltd. All rights reserved.
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

route (An et al., 2015), calling at multiple ports (An et al., 2015; Yang 2. An integer linear programming model is proposed to minimize
et al., 2014), and hub-and-spoke network (Zheng & Yang, 2016) in the cost of transportation, transshipment, and the number of
inland waterway shipping were studied preliminarily. Some exclusive ships. The hub location, feeder port allocation, and fleet deploy-
impacts of geographical and economic features in inland waterway ment are determined simultaneously in the model.
shipping network design were investigated. However, the existing stud- 3. A decomposition-based math-heuristic method is proposed to
ies are conducted on the trunk waterway with a line structure without solve the HSND problem. The problem is decomposed into
considering the tributaries. the main problem of hub-location determination and two sub-
The design of hub-and-spoke networks for tree-like river systems problems of determining the feeder ports allocation and fleet
differs from other classical hub-and-spoke networks problems arising in deployment. Then a cut-based approximate model of the HSND
maritime transportation or in long-haul road transportation. In other is proposed to achieve high-quality solutions for the main prob-
classical hub-and-spoke network problems in maritime transportation lem iteratively, which, as an input of the original HSND model,
or in long-haul road transportation, ports/stations are usually dis- can lead to optimal solutions of two sub-problems. Meanwhile,
tributed in a clustered manner in areas that are far apart, which usually
an enhanced genetic algorithm with sophisticatedly designed
has a star-star network structure, and allows simple direct transport
operators and a diversity enhancement strategy is also proposed
to be used between the stars (Fadda et al., 2020; Roni, 2013). In
in this work to solve the HSND for quick solutions.
the star-star networks, hubs are directly connected to a central hub
4. A real-world case study on the Yangtze River and extensive
with a single allocation for non-hubs (Labbé & Yaman, 2008; Yaman,
computational experiments are conducted to demonstrate the
2008). Compared to other classical hub-and-spoke networks problems,
effectiveness of the proposed models and solution methods. In
decisions of hub locations and feeder allocations in tree-like river
addition, the impacts of the number of hub ports and the hub-
systems are influenced by more complex factors. Specifically, ports are
evenly distributed along the trunk waterway and tributaries, which to-hub transportation discount factor are investigated from the
makes the spatial distribution of ports more complicated. The decision economic and network structure perspectives, and managerial
of hub ports is affected not only by trunk waterway or tributaries, but insights are provided based on the obtained results.
also by the combinatorial ship routes of trunk waterway and tributaries.
This paper is organized as follows. Section 2 reviews the liter-
Meanwhile, the navigation conditions of different sections of rivers are
ature on the network design in inland waterway shipping and the
different, and the choice of the type of ship can affect the shipping
HSND-related problems and solution methods. Section 3 describes the
network. In addition, the specificity of OD flow distribution and river
characteristics of inland waterway shipping with tributaries, hub-and-
direction also have impacts on the shipping network. However, the
characteristic tree-like structure of the river system leads to a specified spoke network in inland waterway shipping, and routes of OD flows.
shipping route. For example, by identifying the ports that ships pass Section 4 formulates the mathematical model of the problem, and
through, the only fixed shipping route can be identified accordingly. an integer linear programming model is constructed. Section 5 de-
This feature is employed to construct the model of the network design tails the proposed decomposition-based math-heuristic method and
problem and develop the solution methods. enhanced genetic algorithm. Section 6 presents a real-world case study
The hub-and-spoke network has been studied deeply from the struc- on the Yangtze River and extensive computational experiments, with
ture and solution method perspectives (Alumur et al., 2021; Martins de sensitivity analysis performed. Section 7 concludes the paper.
Sá et al., 2015). The tree-star, star-star, and cycle-star hub networks
have been proposed and investigated. Among them, the cycle-star hub 2. Literature review
networks, where the hubs are connected with a cycle with a single
allocation of non-hubs, were proposed by Contreras et al. (2017). This There is limited relevant research on the network design in inland
structure seems more suitable for inland shipping network design since waterway container shipping. Yang et al. (2014) proposed an integer
the hub ports are usually located at the trunk waterway with a line programming model to determine the optimal container liner network
structure allowing cycle shipping routes to travel to and from the on the Yangtze River, where the shipping routes, calling ports, calling
upstream and downstream hub ports. The star part suitably serves sequences, numbers, and types of ships can be determined simultane-
the feeder ports located on tributaries. However, the cycle-star hub ously. An et al. (2015) studied the construction of a system service
networks cannot be directly used in the inland waterway shipping network to determine the optimal routing and fleet deployment simul-
network design. The tree-like structure and flow direction of rivers, taneously for inland water container transportation. The experimental
the types and access restrictions of ships, the distribution of hub ports, study on the Yangtze River shows that the point-to-point direct route
and OD flows are needed to be comprehensively considered (Bu & is the most economical mode for large cargo flows between ports, and
Nachtmann, 2021; Notteboom et al., 2020). routes calling at multiple ports can achieve economies of scale when
To our knowledge, there are no published studies research on the the flow of containers is small. Afterward, Zheng and Yang (2016)
network design problem for container shipping in inland waterways explored the economies of scale and studied the design of the hub-and-
considering the tributaries. Meanwhile, no published studies design spoke network on the Yangtze River for a shipping company, which
a cycle-star hub-and-spoke network for container shipping in inland discussed the advantages of the hub-and-spoke structure of the network
waterways with complex application constraints. Solution methods
in inland shipping. Some research investigates the exclusive impacts
for hub-and-spoke network design (HSND) problems have been abun-
of geographical and economic features in inland waterway shipping
dantly developed. Contreras et al. (2017) presented a branch-and-cut
network design, such as the impacts of non-identical streamflow speed
algorithm and a greedy randomized adaptive search metaheuristic
and the uncertain dam transit time (Tan et al., 2018), purchasing price
specifically for solving the cycle hub location problem. However, the
and transportation cost (Li et al., 2019), the profit of the shipping
methods of solving the cycle-star HSND for container shipping in inland
company (Alfandari et al., 2019), and energy savings and emission
waterways need to be studied.
reduction (He et al., 2021; Zhu et al., 2021). However, all the above
This study addresses the HSND for container shipping in inland
research is conducted on the trunk waterway with a line structure, and
waterways. The main contributions of the paper are fourfold:
the tributaries are not considered in the network.
1. The network design for container shipping in tree-like inland The advantages of hub-and-spoke networks are proven not only
waterways is conducted. Considering the low water levels or lock for inland shipping (Zheng & Yang, 2016) but also for ocean ship-
and bridge obstacles that limit the access of large ships on some ping (Gelareh et al., 2013; Imai et al., 2009; Msakni et al., 2020;
segments, inland waterway container shipping is considered a Tu et al., 2018) and intermodal network design (Meng & Wang,
cycle-star hub-and-spoke network. 2011), which mainly includes: (i) lower transportation costs owing to

2
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

economies of scale, (ii) lower costs due to building a sparser network needing the visit sequence due to the linear structure. For example,
to connect many scattered OD (origin–destination) flows, and (iii) given the river structure in Fig. 1(a), ships must pass through Ports 2
better service thanks to consolidated flows that allow more frequent and 3 from Port 1 to Port 4. This is different from the direct connection
connections (Alumur et al., 2021). Some preliminary research has been of any two ports in ocean shipping. However, when jointing the ports
conducted on the structure of hub network (Alumur & Kara, 2008; on tributaries, inland waterway shipping cannot be regarded as a
Klincewicz, 1998). For example, Contreras et al. (2009) and Con- complete line structure, which increases the complexity and research
treras et al. (2010) examined the tree-star hub networks, where a significance of the problem.
tree connects hubs with a single allocation for non-hubs. Labbé and Usually, a river can be divided into the lower, middle, and upper
Yaman (2008) and Yaman (2008) explored the star-star networks, reaches based on the depth of the draft and regional development
where hubs are directly connected to a central hub with a single plans (Notteboom et al., 2020). This limits the navigability of diverse
allocation for non-hubs. Fadda et al. (2020) designed a two-hub-and- trunk waterways and tributary segments, requiring different types of
spoke network for the Mediterranean short-sea-shipping market, where ships with a specific transportation cost of a unit flow. In addition,
the fleet sizing, the frequency of each service were scheduled. The small ships are usually navigable on canals of tributaries, and both
two-hub-and-spoke network is a successful application of the star- small and large ships are navigable on trunk waterways.
star networks. Recently, Contreras et al. (2017) studied cycle-star A river’s topography forms a fixed direction of its streamflow. As for
hub networks in which the hubs are connected with a cycle with inland waterway shipping, the direction of the shipping route opposite
a single allocation of non-hubs. This structure is useful in modeling the river streamflow, or counter-current shipping, costs more than co-
applications in transportation systems where setting requirements make current shipping (shipping routes run downstream) (Tan et al., 2018).
cycle topology a prominent network architecture. A shipping route on the trunk waterway may consist of segments
The class of HSND models is known to be significantly difficult with different sailing directions. When jointing the ports on tributaries,
to handle due to the flow conservation constraints and additional the shipping routes become more complex and may have numerous
design variables. The model-based approaches have been developed segments of different sailing directions.
to solve complex models from two main streams. One is to use the
complete model to efficiently explore solutions’ neighborhood. For
example, Fischetti and Lodi (2003) introduced the local branching cuts 3.2. Hub-and-spoke network in inland waterway shipping
to improve the heuristics, where local branching cuts can alter high-
level strategic branchings to define the solution neighborhood, and The application of hub-and-spoke network structure in inland water-
low-level tactical branchings are used to explore them. This efficient way shipping mainly covers three aspects in this study: the hub location
local branching strategy has been inserted into variable neighborhood problem, the feeder ports allocation problem, and the fleet deployment
search and other heuristics by Hansen et al. (2006) and Jalal et al. problem.
(2022). The other is to construct models to solve sub-problems or a Numerous berths and efficient processing ability are required for
relaxed version of the original problems. For example (Gelareh et al., hub ports to accommodate huge throughput and frequent handling
2015) proposed an efficient decomposition scheme for solving the work (Jiang et al., 2012; Nam & Song, 2011). Restricted by the natural
multi-period hub location with budget constraints by decomposing the geographical conditions and social development level, the chosen hub
complex original problem into simpler subproblems. Martins de Sá ports are usually located on the trunk waterway, which is also the
et al. (2015) presented a Benders-branch-and-cut algorithm and several case in this study. As shown in Fig. 1(b), in inland waterway shipping,
heuristic algorithms to solve the hub networks design problem with once the hub ports are determined, a circular shipping route connecting
multiple lines, which can solve instances with up to 100 nodes. Alibeyg the hub ports (i.e., the main ship route) operated in the form of a
et al. (2018) proposed a branch-and-bound algorithm, which turns out liner (Wang & Zhao, 2021) can also be determined due to the structure
to be an efficient exact algorithm. Real et al. (2021) developed two of the trunk waterway.
decomposition strategies for solving multimodal hub network design We assume each spoke port should be allocated to exactly one hub
problems with flexible routes, reducing the problem’s complexity. port, and connections between two ports are not allowed in this study.
For more practical constraints and decision variables involved in In inland waterways with a tree-like structure, the container routes
the network design problem, some other meta-heuristics have been between the origin and destination ports can be determined as long as
developed, such as ant colony optimization (He et al., 2021), variable the decision of hub location and feeder port allocation has been made.
neighborhood search (Zhu et al., 2021), particle swarm optimization Specifically, given the origin and destination ports, the container routes
method (Özgün-Kibiroğlu et al., 2019), two-step heuristic method (Ko- must begin at the origin port, travel along the arc connected by the hub
rani & Eydi, 2021), and genetic algorithms (Azizi et al., 2018; Meng
ports corresponding to the origin and destination ports, and end at the
& Wang, 2011; Zheng et al., 2015). Genetic algorithms are favored
destination port. A possible example can be seen in Fig. 1(c).
because of their superior global optimization ability. For more details
Ships have different capacities and transportation costs of a unit
about other state-of-art hub network design-related problems and solu-
flow, and there are strict traffic restrictions for ships in inland wa-
tions, interested readers can be referred to Martins de Sá et al. (2015)
terways, especially on tributaries. After the shipping routes are deter-
and Alumur et al. (2021).
mined, the fleet needs to be deployed to meet the transportation needs
at the lowest cost (Tan et al., 2021). We assume that there is an infinite
3. Problem statement number of available ships that third-party companies can provide, and
each ship is associated with a fixed cost. Given a circular ship route
Before formulating the mathematical model for the hub-and-spoke connecting the hub ports (i.e., the main ship route), determining the
network in inland waterway container shipping, we first introduce its fleet deployment on the circular main ship route should meet the
characteristics in this section. demands of all segments. As for each feeder, one type of ship is chosen
to travel back and forth between the corresponding hub port and spoke
3.1. Characteristics of inland waterway shipping with tributaries port (i.e., the feeder ship route). It should be noted that the hub
location, feeder ports allocation, and fleet deployment problems are
In inland waterway shipping, container routes on the trunk water- interconnected, which leads to high complexity when trying to achieve
way can be directly determined according to the visit ports without the global optimal solution for the HSND problem.

3
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

Fig. 1. Schematic diagram of HSND in a tree-like river.

Table 1
Notations and variables.
Sets and parameters:
N: Set of all ports.
N𝑠 : Set of ports on tributaries, N𝑠 ∈ N.
𝑤𝑜𝑑 : Number of containers transported from origin ports 𝑜 ∈ N to destination ports 𝑑 ∈ N, i.e., the OD
flow,in a given period.
𝑝: Given number of hub ports.
𝑎: Hub-to-hub transportation discount factor.
V: Set of ship types.
𝑞𝑣 : Capacity of ship type 𝑣 ∈ V.
𝑐̄𝑣 : Fixed cost of ship type 𝑣 ∈ V.
𝑐𝑣 : Cost of transporting per container per kilometer of ship type 𝑣 ∈ V.
𝑒𝑖𝑗𝑣 : Whether ship type 𝑣 ∈ V is allowed to travel from port 𝑖 ∈ N to port 𝑗 ∈ N, which equals 1 if it is;
otherwise, 0.
𝑐:
̌ Unit cost of each transshipment of a container.
𝑑𝑖𝑗 : Transportation distance from port 𝑖 ∈ N to port 𝑗 ∈ N.
𝑏𝑖𝑗𝑚𝑙 : Parameter matrix, which equals one if a ship from port 𝑖 ∈ N to port 𝑗 ∈ N must travel through port
𝑚 ∈ N to port 𝑛 ∈ N; otherwise, 0.
𝑀: A large positive constant.
Decision variables:

𝑋𝑜𝑑𝑖𝑗 : Binary variable, which equals 1 if OD flow 𝑤𝑜𝑑 (𝑜, 𝑑 ∈ N) is transported from port 𝑖 ∈ N to port 𝑗 ∈ N
on the main ship route; and 0, otherwise.
𝑠
𝑋𝑜𝑑𝑖𝑗 : Binary variable, which equals 1 if OD flow 𝑤𝑜𝑑 (𝑜, 𝑑 ∈ N) is transported from port 𝑖 ∈ N to port 𝑗 ∈ 𝑁
on feeder ship routes; and 0, otherwise.
𝑌𝑖𝑘 : Binary variable, which equals 1 if port 𝑖 ∈ N is allocated to port 𝑘 ∈ N; and 0, otherwise. If 𝑌𝑘𝑘 is 1,
the port 𝑘 ∈ N is a hub port.
𝑇𝑖𝑣 : Binary variable, which equals 1 if ships with type 𝑣 ∈ V are assigned to the feeder ship route
connected by port 𝑖 ∈ N ∪ {𝑛 + 1} and its corresponding hub port; and 0, otherwise. The case 𝑖 = 𝑛 + 1
respects the main ship route. If 𝑖 is a hub port, 𝑇𝑖𝑣 , ∀𝑣 ∈ V is 0, which is explained in the following
formulation.
𝛩𝑖𝑣 : Integer variable, representing the number of ships with type 𝑣 ∈ V assigned to the feeder ship route
connected by port 𝑖 ∈ N ∪ {𝑛 + 1} and its corresponding hub port.

3.3. Container routes of OD flows port that receives containers from other ports (Wang & Slack, 2000). In
other words, container routes can be between any two ports (i.e., OD
In the inland waterway container shipping system, each port can be flows) in a given period. The transshipment may occur on hub ports
an origin port that exports containers to other ports and a destination due to different ships used on hub-to-hub transportation and feeder

4
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

transportation. Due to the economies of scale, the transportation cost


( ) ∑
of a unit flow of OD flows on hub-to-hub arcs is usually cheaper than 𝑀 1 − 𝑇𝑖𝑣 + 𝑞𝑣 𝛩𝑖𝑣 ≥ 𝑠
𝑤𝑖𝑑 𝑋𝑖𝑑𝑖𝑘 , ∀𝑣 ∈ V, 𝑖, 𝑘 ∈ N, 𝑘 ≠ 𝑖, (16)
those on the feeders. 𝑑∈N
( ) ∑
𝑠
𝑀 1 − 𝑇𝑖𝑣 + 𝑞𝑣 𝛩𝑖𝑣 ≥ 𝑤𝑜𝑖 𝑋𝑜𝑖𝑘𝑖 , ∀𝑣 ∈ V, 𝑖, 𝑘 ∈ N, 𝑘 ≠ 𝑖. (17)
4. Mathematical model of the Inland waterway container shipping 𝑜∈N
network
The objective function (1) aims to minimize the total cost of ships,
transportation, and transshipment of OD flows. Constraints (2)–(5)
This section proposes an integer programming model for the HSND
model the hub location, wherein constraint (2) limits that ports can
of container shipping in inland waterways with a tree-like structure, only be connected to one hub port, constraint (3) ensures that each port
which is expanded from the classic model of hub location problems is connected to another port, constraint (4) guarantees that there are 𝑝
in Alumur et al. (2021) and Contreras et al. (2010). The HSND problem hub ports in this network and constraint (5) limits that hub ports are
aims to determine the optimal hub location, feeder port allocation, located on trunk waterway. Constraints (6)–(8) model that the OD flows
and fleet deployment scheduling scheme to minimize the total cost of transported on the hub-and-spoke network must begin at the origin
ships, transportation, and transshipment of OD flows. First, we present port, pass the arc connected by two hub ports corresponding to the ori-
a four-index model for the HSND problem with a nonlinear term in the gin and destination port, and finish at the destination port. Constraints
objective function. In light of this, a set of five-dimensional variables (9)–(11) model the fleet deployment on the main ship route, where
is introduced to linearize the objective function, and an integer linear constraints (9)–(10) guarantee that exactly one allowable type of ship
programming model for the HSND problem is formulated. on the main ship route is chosen, and constraint (11) ensures that the
Notations and variables for describing the model are given in Ta- number of the chosen ships can meet the requirements of all OD flows
ble 1. transported on each segment on the main ship route. Constraints (12)–
(17) model the fleet deployment on the feeder ship routes: constraints
4.1. The integer non-linear programming model (M1) (12)–(15) guarantee choosing only one allowable type of ships to travel
back and forth through the feeder, and constraints (16)–(17) ensure
Let 𝑍 denote the objective value of the problem, the integer non- that the number of chosen ships meets the requirements of all OD flows
linear programming model (M1) is represented as follows. transported on the feeder ship routes.

minimize 𝑍 = 𝑐̄𝑣 𝛩𝑖𝑣 4.2. The integer linear programming model (M2)
𝑖∈N∪{|N|+1},
𝑣∈V

+ ℎ
𝑐𝑣 𝑤𝑜𝑑 𝑑𝑖𝑗 (𝑎𝑋𝑜𝑑𝑖𝑗 𝑠
𝑇(|N|+1)𝑣 + 𝑋𝑜𝑑𝑖𝑗 𝑇𝑖𝑣 )+ There are nonlinear terms in the objective function (1), which
𝑜,𝑑,𝑖,𝑗∈N, (1) makes the model in Section 4.1 difficult to solve. In light of this, a set
𝑣∈V
∑ of five-dimensional variables is introduced to linearize the nonlinear

𝑐𝑤
̌ 𝑜𝑑 𝑋𝑜𝑑𝑖𝑗 , terms of the model. Let 𝑋̌ 𝑜𝑑𝑖𝑗𝑣
ℎ be a binary variable which is one if OD
𝑜,𝑑,𝑖,𝑗∈N,
𝑖≠𝑗∩𝑖≠𝑜∪𝑗≠𝑑 flow 𝑤𝑜𝑑 (𝑜, 𝑑 ∈ N) is transported by ships with type 𝑣 ∈ V on segments
∑ from port 𝑖 ∈ N to port 𝑗 ∈ N on the main ship route; and 0, otherwise.
subject to 𝑌𝑖𝑘 ≤ (|N| − 𝑝 + 1) 𝑌𝑘𝑘 , ∀𝑘 ∈ N, (2) Let 𝑋̌ 𝑜𝑑𝑖𝑗𝑣
𝑠 be a binary variable which is one if OD flow 𝑤𝑜𝑑 (𝑜, 𝑑 ∈ N)
𝑖∈N
∑ is transported by ships with type 𝑣 ∈ V on segments from port 𝑖 ∈ N
𝑌𝑖𝑘 = 1, ∀𝑖 ∈ N, (3) to port 𝑗 ∈ N on feeder ship routes; and 0, otherwise. Let 𝑍 ′ denote
𝑘∈N the objective value, the integer linear programming model can then be

𝑌𝑘𝑘 = 𝑝, (4) re-formulated as follows:
∑ ∑
𝑘∈N minimize𝑍 ′ = 𝑐̄𝑣 𝛩𝑖𝑣 + 𝑤𝑜𝑑 𝑐𝑣 𝑑𝑖𝑗 (𝑎𝑋̌ ℎ + 𝑋̌ 𝑠 )+
𝑜𝑑𝑖𝑗𝑣 𝑜𝑑𝑖𝑗𝑣
𝑖∈N∪{|N|+1}, 𝑜,𝑑,𝑖,𝑗∈N,
𝑌𝑘𝑘 = 0, ∀𝑘 ∈ N𝑠 , (5) 𝑣∈V 𝑣∈V
∑ (18)

𝑐𝑤
̌ 𝑜𝑑 𝑋𝑜𝑑𝑖𝑗 ,
𝑠
𝑀𝑋𝑜𝑑𝑜𝑘 ≥ 𝑤𝑜𝑑 𝑌𝑜𝑘 , ∀𝑜, 𝑑, 𝑘 ∈ N, (6) 𝑜,𝑑,𝑖,𝑗∈N,
𝑖≠𝑗∩𝑖≠𝑜∪𝑗≠𝑑

𝑠 subject to (2)–(17),
𝑀𝑋𝑜𝑑𝑘𝑑 ≥ 𝑤𝑜𝑑 𝑌𝑑𝑘 , ∀𝑜, 𝑑, 𝑘 ∈ N, (7)


( ) ℎ ℎ
𝑀𝑋𝑜𝑑𝑖𝑗 ≥ 𝑤𝑜𝑑 𝑌𝑜𝑖 + 𝑌𝑑𝑗 − 1 , ∀𝑜, 𝑑, 𝑖, 𝑗 ∈ N, 𝑖 ≠ 𝑗, (8) 𝑋̌ 𝑜𝑑𝑖𝑗𝑣 ≥ 𝑋𝑜𝑑𝑖𝑗 + 𝑇𝑛𝑣 − 1, ∀𝑜, 𝑑, 𝑖, 𝑗 ∈ N, 𝑣 ∈ V, (19)
∑ 𝑠 𝑠
𝑇(|N|+1)𝑣 = 1, (9) 𝑋̌ 𝑜𝑑𝑖𝑗𝑣 ≥ 𝑋𝑜𝑑𝑖𝑗 + 𝑇𝑖𝑣 − 1, ∀𝑜, 𝑑, 𝑖, 𝑗 ∈ N, 𝑣 ∈ V, (20)
𝑣∈V
( ) 𝑠 𝑠

𝑇(|N|+1)𝑣 ≤ 𝑒𝑖𝑗𝑣 + 1 − 𝑋𝑜𝑑𝑖𝑗 , ∀𝑜, 𝑑, 𝑖, 𝑗 ∈ N, 𝑣 ∈ V, (10) 𝑋̌ 𝑜𝑑𝑖𝑗𝑣 ≥ 𝑋𝑜𝑑𝑖𝑗 + 𝑇𝑗𝑣 − 1, ∀𝑜, 𝑑, 𝑖, 𝑗 ∈ N, 𝑣 ∈ V. (21)
Constraints (19)–(21) define the relationship between the five-
∑ index variables and four-index variables. Unlike the former integer

𝑀(1 − 𝑇(|N|+1)𝑣 ) + 𝑞𝑣 𝛩(|N|+1)𝑣 ≥ 𝑤𝑜𝑑 𝑏𝑖𝑗𝑚𝑙 𝑋𝑜𝑑𝑖𝑗 , ∀𝑣 ∈ V, 𝑚, 𝑙 ∈ N,
𝑜,𝑑,𝑖,𝑗∈N linear programming models of hub-and-spoke networks in shipping
networks (Fontes & Goncalves, 2021; Yang et al., 2014; Zheng & Yang,
(11) 2016), the proposed model uses a special parameter matrix 𝑏𝑖𝑗𝑚𝑙 rather
∑ ∑ than the flow conservation constraints to gather the volume on the
𝑠
𝑀(𝑌𝑖𝑖 + 𝑇𝑖𝑣 ) ≥ 𝑤𝑖𝑑 𝑋𝑖𝑑𝑖𝑘 , ∀𝑖 ∈ N, (12) main ship route. This reduces the number of variables as well as the
𝑣∈V 𝑑,𝑘∈N constraints. Meanwhile, the fleet deployment is considered, and the
∑ ∑
𝑀(𝑌𝑖𝑖 + 𝑇𝑖𝑣 ) ≥ 𝑤𝑑𝑖 𝑋 𝑠𝑑𝑖𝑘𝑖 , ∀𝑖 ∈ N, (13) induced cost of transportation, transshipment, and the number of ships
𝑣∈𝑉 𝑑,𝑘∈𝑁 are jointly optimized. However, this idea is only suitable when dealing
( 𝑠
) with shipping on inland waterways with tree-like structures, and the
𝑇𝑖𝑣 ≤ 𝑒𝑖𝑘𝑣 + 1 − 𝑋𝑖𝑑𝑖𝑘 , ∀𝑖, 𝑘, 𝑑 ∈ N, 𝑖 ≠ 𝑘, 𝑣 ∈ V, (14) hub ports are located on the trunk. Since the line structure can be
( ) regarded as a simplified case of the tree-like structure, the idea can
𝑠
𝑇𝑖𝑣 ≤ 𝑒𝑘𝑖𝑣 + 1 − 𝑋𝑑𝑖𝑘𝑖 , ∀𝑖, 𝑘, 𝑑 ∈ N, 𝑖 ≠ 𝑘, 𝑣 ∈ V, (15) also be applied to inland waterways with a line structure.

5
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

5. Solution method

With numerous constraints and high dimensional variables of the in-


teger linear programming model presented in Section 4.2, commercial
solvers such as CPLEX can hardly solve the model optimally reasonably,
especially for large-scale problems. To solve the stiff HSND optimiza-
tion problem, a decomposition-based math-heuristic method and an
enhanced genetic algorithm are developed, which can effectively obtain
high-quality solutions.

5.1. Parameter initialization

To reduce the computational complexity of the model, a complete


set of parameters, including distances 𝑑(𝑖, 𝑗), allowable types of ships
𝑎(𝑖, 𝑗, 𝑣), the parameter set of 𝑏𝑖𝑗𝑚𝑙 and between any two ports are re-
quired in advance. Given an inland waterway with a tree-like structure,
the distances and allowable types of ships between adjacent ports are
easy to capture. Therefore, the required parameters can be cumula-
tively calculated based on the information of the directly connected
adjacent ports.
Fig. 2. The flowchart for the decomposition-based math-heuristic method.
5.2. Decomposition-based math-heuristic method

The set of five-index variables is introduced to linearize the model


that another set of hub ports can be taken as parameters of model M2
M1 presented in Section 4.1, which leads to an integer linear model
for generating new optimal solutions of sub-problems. The following
M2 in Section 4.2. However, introducing the five-index variables sig-
relation holds:
nificantly increases the number of constraints and variables in the ∑
model, which is still very difficult to solve due to the enormous possible 𝑌𝑘𝑘 < 𝑝. (23)
combinatorial schemes. Since optimal solutions to large-scale complex 𝑘∈𝑆3
problems can hardly be achieved, math-heuristic solutions may be more The procedure stops when the pre-specified number of iterations is
realistic. Indeed, for some complex problems, the solution to a simpli- reached or no better solution to the original problem can be obtained
fied problem by relaxing some non-destructive conditions or constraints in a certain number of consecutive iterations. In the solution approach
may arguably reach an approximate or even the optimal solution of above, the reason for choosing hub ports as cut in each iteration is to
the original complicated problem (Ruocco et al., 2017). In light of this, expand the solution space as much as possible and force the commercial
approximate model-based deterministic methods (Ruocco et al., 2017) solver to find a better solution in a reasonable time. The flowchart of
and model-based heuristics (Ömer Nedim Kenger & Ozceylan, 2023; the decomposition-based math-heuristic approach is described in Fig. 2.
Perboli et al., 2011; Sel & Bilgen, 2014) have been developed for com- Note that the basic combinatorial benders cut approaches (Codato
plex combinatorial optimization problems and obtained satisfactory & Fischetti, 2006) usually incrementally add the partially relaxed con-
results. straints in the form of feasible or optimal cuts to the master problems.
Careful examination of model M1 shows that the condition causes However, the proposed decomposition-based math-heuristic method
the nonlinear term that different ship types have different transporta- decomposes the original problem into hub location main problem and
tion costs per container per kilometer (𝑐𝑣 , 𝑣 ∈ V) In other words, if all feeder ports allocation and fleet deployment sub-problems. Compared
ships have the same unit transport costs (i.e., 𝑐𝑣 is a constant), model to the original problem, the main problem is slightly relaxed but has
M1 can be converted to an integer linear model (M3) as follows: a similar network structure. The hubs solved by the main problem
∑ ∑
minimize 𝑍 ′′ = 𝑐̄𝑣 𝛩𝑖𝑣 + ℎ
𝑐𝑣 𝑤𝑜𝑑 𝑑𝑖𝑗 (𝑎𝑋𝑜𝑑𝑖𝑗 𝑠
+ 𝑋𝑜𝑑𝑖𝑗 )+ are then input into sub-problems as parameters to obtain complete
𝑖∈N∪{|N|+1}, 𝑜,𝑑,𝑖,𝑗∈N feasible solutions. In addition, cutting the explored network is used to
𝑣∈V
∑ (22) iteratively search for better solutions. This means that the proposed

𝑐𝑤
̌ 𝑜𝑑 𝑋𝑜𝑑𝑖𝑗 . decomposition-based math-heuristic method only obtains feasible so-
𝑜,𝑑,𝑖,𝑗∈N,
𝑖≠𝑗∩𝑖≠𝑜∪𝑗≠𝑑 lutions approaching the optimal.

subject to (2)–(17). 5.3. Genetic algorithm


The simplified approximate-model M3 can be more efficiently solved by
commercial solvers such as CPLEX and Gurobi, compared with the five- The genetic algorithm searches for the optimal solution to problems
index model M2. Moreover, since the only difference is the condition by imitating the evolution process of the population in nature. It has
of ships’ unit transportation costs, the solution to model M3 can be been successfully applied to many HSND problems in maritime trans-
regarded as a high-quality solution to the original model M1. To ensure portation (Azizi et al., 2018; Meng & Wang, 2011; Zheng et al., 2015).
that the solution to the simplified model M3 is close enough to the The population characteristics of a genetic algorithm can maintain good
optimal solution of the original model M1, an extensive search needs genes along evolution, and the genetic operators can meet the wide
to be implemented. To achieve this, the HSND problem is decomposed searching need of rigorous and complex gene structure to construct the
into the main problem of determining the locations of hub ports and hub-and-spoke network. However, it is easy to fall into local optimality.
two sub-problems that determine the feeder ports allocation and fleet In this study, an enhanced genetic algorithm is proposed to solve the
deployment. The hub ports interpreted from the solution (𝑆3 ) of M3 are HSND problem for inland waterway container shipping. The enhanced
regarded as parameters of the model M2, and the optimal feeder ports genetic algorithm has a similar solution framework to the classical
allocation and fleet deployment results can be obtained by optimally genetic algorithms. However, the crossover and mutation operators are
solving model M2. Afterward, the hub ports obtained by solving M3 are divided into two levels to cater to the characteristics of this problem,
defined by formula (23) as a cut to add to M3 for the next iteration so i.e., the hub ports level and the feeder ports level. The use of these

6
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

Thirdly, for each parent, the value of each gene is replaced according
to the corresponding mapping relation to generating two offspring.
The feeder ports crossover is a process that swaps randomly selected
segments of two offspring chromosomes. Specifically, two segments at
the same position in offspring chromosomes are swapped, wherein the
length of each segment is generated randomly.
Fig. 3. Schematic diagram of the encoding. After the above crossover process, there may be a case where the
hub port is not connected to itself, representing an infeasible solution.
Therefore, the repair procedure is implemented to avoid this issue.

two level operators is also elaborately designed according to the perfor- Mutation operator
mance of the evolutionary solutions. Meanwhile, the repair procedure Similarly, the mutation operator consists of hub ports mutation and
is implemented once the solution is infeasible after the evolution. In feeder ports mutation processes. The repair procedure mentioned above
addition, a diversity enhancement strategy is introduced to improve is also used to ensure the feasibility of the updated chromosomes.
the convergence rate. The encoding, initialization process, and genetic In the hub ports mutation process, a hub port is randomly selected
operators in the enhanced genetic algorithm are specially designed from each offspring to be associated with another port that was not
according to the characteristics of the HSND problem. originally associated with this hub port. Then the repair procedure is
used to correct the mutated gene.
Encoding The feeder ports mutation process consists of 2 parts, i.e., the point
The integer encoding (Azizi et al., 2018) is used in the proposed mutation adapted from Zheng et al. (2015) and the segment reverse
enhanced genetic algorithm to represent the connection of ports. Specif- mutation adapted from Meng and Wang (2011), which were success-
ically, the length of the gene is equal to the number of ports 𝑛, the fully applied in relative shipping problems. In the point mutation,
sequence of each gene in the chromosome corresponds to the encoding a feeder port is randomly selected and associated with a different
port, and the value on each gene corresponds to the hub port associated hub port. In the segment reverse mutation, a chromosome segment
with the encoding port. As shown in Fig. 3, Port 0 is associated with is randomly selected. Afterward, the sequence of genes is reversed
hub port 1, and Port 1 is associated with itself, which is a hub port. in this segment and reinserted into the chromosome. Similarly, the
A similar case can be seen for Port 3. Given such a chromosome, the repair procedure is used after the mutations to correct any infeasible
fleet deployment is optimally solved by enumeration according to the chromosome.
network structure, and the volume of OD flows.
With the two mutation processes, an elaborate framework shown
in Fig. 5 is designed to organize them and enhance their effectiveness.
Initialization process Specifically, if the offspring generated by crossover is not a new best
Population initialization is important in leading the enhanced ge-
solution, the hub ports mutation is performed with a probability of
netic algorithm to search solution space efficiently. Therefore, a ran-
𝑃 𝑚ℎ . If hub ports mutation does not happen, the feeder ports mutation
dom initialization process is utilized to diversify the initial population.
is performed with a probability of 𝑃 𝑚𝑠 ; otherwise, if the fitness of
Specifically, 𝑝 hub ports are randomly selected from all ports on the
the new offspring generated by the hub ports mutation has not been
trunk waterway with equal probability. Other ports are regarded as
increased, the feeder ports mutation is performed. In addition, if the
feeder ports and randomly assigned to one of the selected hub ports
population falls into the local optimum, i.e., all individuals are the
with equal probability.
same, the hub ports mutation is performed on each individual before
the next iteration but leaves one individual. This can also be called
Selection operator
group mutation.
The roulette selects two parents according to the fitness of each
individual in the population with the size of 𝛺. As in conventional en-
hanced genetic algorithms, the proposed algorithm regards the inverse Diversity enhancement strategy
of objective value as fitness. Through repeated tests, it is found that the algorithm converges
quickly but easily falls into local optima, like generic algorithms.
Crossover operator As such, a diversity enhancement strategy is proposed to delay the
The chromosomes represent feasible solutions, including informa- converges. Specifically, when the selected parents are the same, one is
tion on hub ports and feeder port allocation results. Therefore, as randomly initialized with a probability of 𝑃𝑑 to continue the evolution.
shown in Fig. 4, the crossover operator consists of hub ports crossover
and feeder ports crossover processes with probability 𝑃𝑐 , and a re-
6. Numerical experiments
pair procedure is then used to ensure the feasibility of the solution
of chromosomes since crossover procedures may generate infeasible
solutions. In this section, numerical examples based on the Australia Post (AP)
It is believed that a candidate hub port associated with more feeder data set and the traffic and shipping data on the Yangtze River are
ports is deemed to have a stronger radiation capacity and more poten- implemented to confirm the effectiveness of the proposed models and
tial to become a hub port. In this regard, the hub ports crossover process solution methods, and managerial insights are provided based on the
can be divided into three steps. Firstly, 𝑝 hub ports are selected by the obtained results. The parameters of the enhanced genetic algorithm
roulette according to the number of times the hub ports are associated were tuned for all instances through orthogonal experiments. The
with feeder ports in two parents. Then, the selected hub ports are sorted values with the best average results were chosen for the enhanced
according to their number of times in descending order (as shown on genetic algorithm, which was set to 𝛺 = 50, 𝑃𝑐 = 0.45, 𝑃 𝑚ℎ = 0.60,
the ‘‘Sorting table’’ in Fig. 4). Secondly, the hub ports are sorted for two 𝑃 𝑚𝑠 = 0.10, 𝑃𝑑 = 0.75. In addition, CPLEX 12.9.0 is used for solving the
parents with the same rule. Therefore, two mapping relations can be integer linear programming models, and all experiments are conducted
constituted for the selected hub ports in each parent with the hub ports on a Windows PC with Intel(R) Core (TM) i5-8500 CPU and 8.00 GB
in two parents, respectively (as shown in ‘‘Mapping relations’’ in Fig. 4). RAM, which runs MS Visual C ++ 2019.

7
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

Fig. 4. Schematic diagram of crossover operator.

Fig. 5. Framework of the mutation process.

6.1. Data description Table 2


Information of ships.

The AP data set consists of 10 to 200 nodes based on a postal Type Capacity (TEU) Fixed cost (RMB) Unit rates
(RMB/TEU/km)
delivery in Sydney, Australia. Since the proposed enhanced genetic
algorithm can be applied to solve hub location problems by adapting 0 200 2000 8.5
1 500 3500 7.5
the objective function and relaxing constraints, we use the instances 2 1000 4800 7
for the uncapacitated single allocation 𝑝 hub median problem (US-
ApHMP) with 25, 50, 75, 100, 150, and 200 nodes to verify the
performance of the enhanced genetic algorithm, where the USApHMP
defined by O’kelly (1987) is to choose 𝑝 hubs and assign every 𝑖 ∈ N shown in Table 4, where the throughput of ports from January 2019 to
to them to minimize the overall transport cost with no hub cost and no September 2021 is released by the Ministry of Transport of the People’s
capacity constraint involved. This leads to a similar physical solution Republic of China (https://www.mot.gov.cn/).
structure with the HSND for container shipping in the inland waterway.
As shown in Fig. 6, 22 inland ports along the Yangtze River, 6.2. Results and performance analysis of the solution methods
i.e., Luzhou, Chongqing, Yichang, Jingzhou, Chenglingji, Wuhan,
Huangshi, Jiujiang, Anqing, Tongling, Wuhu, Maanshan, Nanjing, Zhen-
The results of the enhanced genetic algorithm for solving the AP
jiang, Jiangyin, Nantong, Shanghai, Changsha, Nanchang, Hefei, Jin-
instances are depicted in Table 5, where 𝑛 denotes the number of nodes,
ing, and Hangzhou are selected. The candidate hub ports can only
and 𝑝 denotes the number of used hubs. 𝑂𝑝𝑡. 𝑂𝑏𝑗. denotes the optimal
be on the trunk waterway. All data is collected from the practical
investigation, and data such as the unit transportation costs of ships objective for each instance summarized by Meier and Clausen (2018).
have been modified for confidentiality. Since the buffer time and the 𝑂𝑏𝑗. denotes the objective of the solution by the enhanced genetic
passage time of ships from Luzhou port to Shanghai port is about two algorithm, where 𝑎𝑣𝑔 and 𝑏𝑒𝑠𝑡 denote the average value for ten runs
weeks, we design the shipping network for two weeks. We consider and the best value, and 𝑡𝑖𝑚𝑒 denotes the average computational time
three main types of ships, and the bi-weekly fixed costs and transporta- in seconds. 𝐺𝑎𝑝 denotes the relative deviation of objectives between
tion costs of a unit flow of ships are shown in Table 2. The practical the 𝑂𝑝𝑡. 𝑂𝑏𝑗 and 𝑏𝑒𝑠𝑡 (i.e., 𝐺𝑎𝑝 = (𝑏𝑒𝑠𝑡 − 𝑂𝑝𝑡. 𝑂𝑏𝑗.)∕𝑂𝑝𝑡. 𝑂𝑏𝑗. × 100%),
distances and allowable ship types between all adjacent ports are shown and 𝐴𝑣𝑔 denotes the average value of the corresponding column. The
in Table 3. According to the speed of the Yangtze River current and results show that the proposed enhanced genetic algorithm produces
ships (Tan et al., 2018), we assume that the co-current and counter- satisfactory solution performance for solving the hub location-related
current transportation cost of a unit flow along the Yangtze River is problems with up to 200 nodes, where most of the instances can be
0.9 and 1.1, respectively. The OD demands of each port are met based solved to optimality, and the objective of all the instances is only 0.03%
on the data of the throughput of ports and practical investigations as lower than the optimal objective.

8
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

Fig. 6. Schematic diagram of the ports on Yangtze River.

Table 3
The basic distances and allowable ship types between all adjacent ports.
Starting Ending Distance Allowable Starting Ending Distance Allowable
ports ports (km) ship types ports ports (km) ship types
Luzhou Chongqing 254 0 Chongqing Yichang 660 0, 1
Yichang Jingzhou 167 0, 1 Jingzhou Chenglingji 323 0, 1
Chenglingji Wuhan 232 0, 1 Wuhang Huangshi 143 0, 1
Huangshi Jiujiang 117 0, 1 Jiujiang Anqing 154 0, 1
Anqing Tongling 92 0, 1 Tongling Wuhu 104 0, 1
Wuhu Maanshan 47 0, 1 Maanshan Nanjing 49 0, 1
Nanjing Zhenjiang 87 0, 1, 2 Zhenjiang Jiangyin 117 0, 1, 2
Jiangyin Nantong 60 0, 1, 2 Nantong Shanghai 128 0, 1, 2
Changsha Chenglingji 158 0 Nanchang Jiujiang 164 0
Hefei Wuhu 134 0 Jining Zhenjiang 550 0
Hangzhou Zhenjiang 310 0

Table 4 Table 5
Average OD demands in two weeks. Comparison results of AP instances for USApHMP.
Origin–destination 𝑤/TEU Origin–destination 𝑤/TEU 𝑛 𝑝 𝑂𝑝𝑡. 𝑂𝑏𝑗. 𝑂𝑏𝑗. Time Gap
Luzhou-Wuhan 769 Luzhou-Nanjing 1923 Avg Best (s) (%)
Chongqing-Wuhan 4615 Chongqing-Jiujiang 1153
25 2 175541.98 175541.98 175541.98 27 0.00
Chongqing-Zhenjiang 769 Chongqing-Jiangyin 1153
3 155256.32 155256.32 155256.32 35 0.00
Yichang-Wuhan 384 Yichang-Nanjing 1923
4 139197.17 139197.17 139197.17 42 0.00
Jingzhou-Nanjing 1923 Chenglingji-Chongqing 1153
5 123574.29 123574.29 123574.29 49 0.00
Chenglingji-Nanjing 3461 Chenglingji-Nantong 2307
Wuhan-Yichang 769 Wuhan-Jingzhou 384 50 2 178,484.29 178484.29 178484.29 77 0.00
Wuhan-Anqing 769 Wuhan-Wuhu 4230 3 158,569.93 158569.93 158569.93 104 0.00
Wuhan-Zhenjiang 1923 Wuhan-Jiangyin 2307 4 143,378.05 143378.05 143378.05 131 0.00
Wuhan-Nanchang 769 Wuhan-Hefei 1923 5 132,366.95 132422.22 132422.22 156 0.04
Jiujiang-Chongqing 1153 Jiujiang-Wuhan 2692 75 2 180,118.91 180118.91 180118.91 160 0.00
Jiujiang-Nantong 2307 Anqing-Wuhan 769 3 161,056.74 161056.74 161056.74 219 0.00
Tongling-Nanjing 384 Wuhu-Chongqing 3076 4 145,734.21 145778.81 145772.43 276 0.03
Wuhu-Jiujiang 1153 Wuhu-Nanjing 5769 5 136,011.35 136053.72 136053.72 330 0.03
Wuhu-Nantong 4230 Maanshan-Wuhan 769
100 2 180,223.80 180,223.80 180,223.80 275 0.00
Nanjing-Luzhou 1923 Nanjing-Chongqing 6153
3 160,847.00 160,847.00 160,847.00 381 0.00
Nanjing-Chenglingji 3461 Nanjing-Wuhan 7307
4 145,896.58 145896.58 145896.58 483 0.00
Nanjing-Anqing 1923 Nanjing-Tongling 769
5 136,929.44 137411.58 136948.78 577 0.01
Nanjing-Zhenjiang 3076 Nanjing-Jiangyin 3846
Nanjing-Nanchang 1923 Nanjing-Hefei 3076 150 2 180,898.84 180898.84 180898.84 606 0.00
Zhenjiang-Wuhan 1923 Zhenjiang-Wuhu 384 3 161,490.48 161490.48 161490.48 839 0.00
Jiangyin-Chongqing 1153 Jiangyin-Wuhan 2692 4 146,521.33 146521.33 146521.33 1059 0.00
Jiangyin-Nantong 2307 Nantong-Luzhou 769 5 137,425.91 138178.13 137528.94 1272 0.07
Nantong-Jingzhou 384 Nantong-Chenglingji 2307 200 2 182,459.25 182459.25 182459.25 1070 0.00
Nantong-Anqing 384 Nantong-Wuhu 4230 3 162,887.03 163157.98 162954.83 1484 0.04
Nantong-Zhenjiang 1538 Nantong-Jiangyin 2307 4 147,767.30 147963.31 147910.96 1882 0.10
Nantong-Hefei 1538 Shanghai-Nanjing 384 5 140,062.65 141001.07 140617.16 2270 0.40
Changsha-Nantong 769 Nanchang-Wuhan 769
Hefei-Chongqing 769 Hefei-Wuhan 1923 𝐴𝑣𝑔 3.5 154695.82 154811.74 154738.50 575 0.03
Jining-Nanjing 384 Hangzhou-Wuhan 384

solutions of 20 trials achieved by the enhanced genetic algorithm.


The solution results obtained by different methods for solving the ‘‘Gap’’ denotes the relative deviation of solution costs between the
HSND problem mentioned above are depicted in Table 6, where 𝑝 and decomposition-based math-heuristic (DM) method and the enhanced
𝑎 denote the number of hub ports and the transportation cost of a unit genetic algorithm (EGA) (i.e., 𝐺𝑎𝑝 = (𝐸𝐺𝐴 − 𝐷𝑀)∕𝐷𝑀 × 100%).
flow, respectively. Symbol ‘‘Best’’ denotes the best solutions obtained The results in Table 6 demonstrate that the proposed DM method
by the enhanced genetic algorithm, and ‘‘Avg’’ represents the average shows significant advantages in solution quality compared with the

9
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

Table 6
Solution results and comparison of different methods.
𝑎 𝑝 Math-heuristic Genetic algorithm Gap 𝑎 𝑝 Math-heuristic Genetic algorithm Gap
method (%) method (%)
Cost Time Cost (RMB) Time Cost Time Cost (RMB) Time
(RMB) (s) Best Avg (s) (RMB) (s) Best Avg (s)
3 1.35E+09 5319 1.35E+09 1.35E+09 28 0 3 1.53E+09 6636 1.53E+09 1.56E+09 29 0
4 1.27E+09 8347 1.28E+09 1.28E+10 31 0.27 4 1.46E+09 8484 1.47E+09 1.48E+09 33 0.26
5 1.23E+09 6202 1.24E+09 1.24E+09 34 0.28 5 1.42E+09 8323 1.43E+09 1.43E+09 35 0.28
0.7 6 1.19E+09 5659 1.20E+09 1.20E+09 45 0.32 0.85 6 1.40E+09 7619 1.40E+09 1.41E+09 45 0
7 1.16E+09 5827 1.17E+09 1.17E+09 60 0.43 7 1.38E+09 7820 1.39E+09 1.39E+09 54 0.4
8 1.16E+09 5373 1.16E+09 1.16E+09 65 0 8 1.38E+09 8184 1.38E+09 1.38E+09 70 0
9 1.15E+09 5414 1.16E+09 1.16E+09 97 0.79 9 1.37E+09 8040 1.38E+09 1.38E+09 99 0.69
3 1.41E+09 5925 1.41E+09 1.43E+09 28 0 3 1.59E+09 6415 1.59E+09 1.60E+09 29 0
4 1.34E+09 6976 1.34E+09 1.35E+09 31 0 4 1.53E+09 9345 1.53E+09 1.55E+09 36 0
5 1.30E+09 6473 1.30E+09 1.30E+09 34 0 5 1.49E+09 9855 1.49E+09 1.49E+09 34 0
0.75 6 1.26E+09 6333 1.26E+09 1.27E+09 40 0 0.9 6 1.47E+09 8735 1.47E+09 1.48E+09 41 0
7 1.24E+09 6396 1.24E+09 1.24E+09 51 0 7 1.45E+09 8829 1.46E+09 1.46E+09 59 0.39
8 1.23E+09 5885 1.24E+09 1.24E+09 73 0.54 8 1.45E+09 7862 1.46E+09 1.46E+09 73 0.56
9 1.22E+09 6631 1.23E+09 1.23E+09 87 0.68 9 1.44E+09 8706 1.45E+09 1.46E+09 92 0.73
3 1.47E+09 5929 1.47E+09 1.49E+09 30 0 3 1.63E+09 6441 1.63E+09 1.63E+09 28 0
4 1.40E+09 7700 1.40E+09 1.41E+09 32 0 4 1.59E+09 10482 1.59E+09 1.59E+09 31 0
5 1.36E+09 7838 1.36E+09 1.36E+09 35 0 5 1.55E+09 10796 1.55E+09 1.55E+09 34 0
0.8 6 1.33E+09 7198 1.33E+09 1.34E+09 43 0 0.95 6 1.54E+09 9641 1.54E+09 1.54E+09 40 0
7 1.31E+09 6992 1.31E+09 1.32E+09 56 0 7 1.53E+09 9695 1.53E+09 1.54E+09 50 0
8 1.30E+09 7074 1.31E+09 1.31E+09 65 0.52 8 1.52E+09 9493 1.53E+09 1.53E+09 66 0
9 1.30E+09 5319 1.31E+09 1.31E+09 87 0.62 9 1.52E+09 9298 1.53E+09 1.53E+09 100 0.67
Avg 1.28E+09 6486 1.29E+09 1.84E+09 50 0.4 Avg 1.49E+09 8605 1.49E+09 1.50E+09 51 0.39

On the one hand, as the number of hub ports increases, the net-
work’s total cost and transportation cost show a declining trend, and
the shipping cost and transshipment cost fluctuate but also display
an overall downward trend. It is not surprising that the more hub
ports, the less the total cost of the designed network. However, when
the number of hub ports reaches 7, increasing the number of hub
ports cannot contribute to a significantly higher economic benefit.
This result denies the subjective consciousness of the linear relation
between the number of hub ports and economic benefit. Administrative
and operational costs must be included for building and maintaining
hub ports, in reality, (Kavirathna et al., 2018). Therefore, the optimal
number of hub ports should be further determined according to other
relative restrictions such as geographical conditions, regional policies,
and knowledge infrastructure capabilities.
Fig. 7. Relative deviation of solutions between M3 and DM method. Nevertheless, this study can still be used as a supporting technique
for the decision-making of hub port locations. On the other hand, as the
hub-to-hub transportation discount factor increases, the network’s total
cost and transportation cost show a quick downward trend. Still, the
enhanced genetic algorithm but costs much more computational time. shipping cost and transshipment costs keep fluctuating. This shows that
The proposed enhanced genetic algorithm also shows good solutions, the hub-to-hub transportation discount factor predominantly affects the
and the relative deviation from the solutions of the DM method is cost of transportation in the shipping network. The higher the discount,
only 0.40%. Moreover, the average computational time of the enhanced the lower the cost of the whole network. In practice, the hub-to-hub
genetic algorithm is only 51 s, which is much shorter than solving the transportation discount factor may be influenced by the market and
model with commercial software. policy (Wetzstein et al., 2021). In addition, the formation of large-scale
To illustrate that the solution of the approximate model (M3) is batches and innovation of ship capacity can also increase the discount
close enough to the solution of the HSND model (M2), Fig. 7 gives of the main route transportation proverbially and indirectly (Zhang
the relative deviation between the best solutions and the solutions after & Zhao, 2015). The mathematical model in this study determines
entering partial optimal solution variables of M3 into M2. It shows that the optimal container flows and fleet deployment on a hub-and-spoke
entering the optimal solution variables of M3 into M2 can obtain 33 structure, which makes it conducive to economic benefits in inland
out of the 42 best solutions, and the relative deviations of others are waterway container shipping.
below 1.00%, which proves the effectiveness of solving the original Fig. 8 also shows that the transportation cost contributes to most of
problem by using the approximate model. Meanwhile, the extensive the total cost of the shipping network, followed by the shipping cost.
search mechanism of the proposed decomposition-based math-heuristic This infers that fleet deployment should be considered when designing
method is significantly helpful for further improving the solutions. an inland waterway shipping network. The proportion of transshipment
cost is relatively small, which explains why many studies (Hellsten
6.3. Economics assessment et al., 2021) ignored it in inland waterway network design.

The line charts of the total cost, ship cost, transportation cost, and 6.4. Analysis of the structure
transshipment cost under different combinations of the number of hub
ports (𝑝) and the discount of the main route transportation (𝑎) are The selected hub ports under different combinations of the number
shown in Fig. 8. of hub ports (𝑝) and the hub-to-hub transportation discount factor (𝑎)

10
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

Fig. 8. The line charts of cost under different combinations of 𝑝 and 𝑎.

Table 7
Hub ports under different combinations of 𝑝 and 𝑎.
𝑝 𝑎
0.7 0.75 0.8 0.85 0.9 0.95
3 1,5,12 1,5,12 1,5,12 1,5,12 5,7,12 5,7,12
4 1,5,7,12 1,5,7,12 1,5,7,12 1,5,7,12 1,5,7,12 5,7,10,12
5 1,5,7,12,14 1,5,7,10,12 1,5,7,10,12 1,5,7,10,12 1,5,7,10,12 1,5,7,10,12
6 1,5,7,10,12,14 1,5,7,10,12,14 1,5,7,10,12,14 1,5,7,10,12,14 1,5,7,10,12,14 1,4,5,7,10,12
7 1,4,5,7,10,12,14 1,4,5,7,10,12,14 1,4,5,7,10,12,14 1,4,5,7,10,12,14 1,4,5,7,10,12,14 1,4,5,7,10,12,14
8 1,3,4,5,7,10,12,14 1,3,4,5,7,10,12,14 1,4,5,7,10,12,13,14 1,4,5,7,8,10,12,14 1,4,5,7,8,10,12,14 1,4,5,7,8,10,12,14
9 1,3,4,5,7,10,12,14,15 1,3,4,5,7,10,12,14,15 1,4,5,7,8,10,12,13,14 1,4,5,7,8,10,12,13,14 1,4,5,7,8,10,12,13,14 1,4,5,7,8,10,12,13,14

are listed in Table 7. The requirement of increasing a hub port will not It shows that the ports on the trunk waterway are usually associated
reshuffle the existing structure of hub ports. Instead, it will select one with the nearest hub port downstream. For example, 4 out of 6 feeder
spoke port from the original network and add it to the set of hub ports. ports associated with the Wuhan port are located upstream of the
As for the hub-to-hub transportation discount factor, different values Wuhan port. The cost of co-current transportation can explain it is
may lead to slight differences in choosing hub ports. Market regulation cheaper than that counter-current transportation. This conclusion is
regarding input and output (Zolfani et al., 2023) may be needed to also supported by the observation presented in Zheng and Yang (2016),
stabilize the hub-to-hub transportation discount factor to maintain the which studied HSND for container shipping along the trunk waterway
benefits of existing hub ports. However, careful examination shows that of the Yangtze River. Ports on tributaries are usually associated with
some hub ports remain unchanged under different hub-to-hub trans- the nearest hub port downstream, which is usually the same hub port
portation discount factors and the required number of hub ports, such of the port on the trunk located at the junction of the trunk and
as the Wuhan port (5) and Nanjing port (12). This can be explained by corresponding tributaries. In this case, the ports on tributaries and the
the fact that these ports have much more OD flows due to the high level corresponding ports on the junction can be merged as a single port for
of regional development. Additionally, these ports are located in the HSND. However, when there is no port at the junction of the trunk
middle of the nearest tributaries, allowing them to serve both feeders. and tributary, the port’s allocation on the tributary must be determined
This is also a factor contributing to being selected as hub ports. interdependently. In addition, based on the topographical structure of
To investigate the impacts of the tree-like structure of inland water- the Yangtze River waterway system, the navigation capacity of each
ways on the HSND, we select three ports that appear most frequently in segment, port layout, and real-world OD flows, the proposed models
Table 7 (i.e., Chongqing, Wuhan, and Nanjing ports) as hub ports. We and methods for HSND are of practical significance and match the
have obtained optimal feeder port allocation and the fleet deployment current status and trends of the shipping development of the Yangtze
results by solving model M2. We have found that the optimal feeder River.
port allocation results remain the same no matter how the hub-to- This paper studies from the perspective of optimization rather than
hub transportation discount factor changes, which shows that the main data analysis. The above results and discussions confirm some con-
factor affecting the structure of the hub-and-spoke shipping network is clusions from the data analysis. However, Fig. 9 shows that not all
not the freight charges. The obtained shipping network with these three ports will assign to their nearest downstream hub port. For example,
specific hub ports is shown in Fig. 9. Jiujiang and Nanchang port do not assign to their nearest hub (Wuhan

11
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

Fig. 9. The designed hub-and-spoke shipping network on Yangtze River with three hub ports.

Port) since more containers need to be shipped downstream; Huangshi CRediT authorship contribution statement
and Anqing ports not assign to their nearest downstream hub port
(Nanjing Port) since more containers need to be shipped upstream. In Saiqi Zhou: Investigation, Methodology, Software, Data curation,
other words, a subjective empirical law should not be used directly Writing – original draft, Formal analysis. Bin Ji: Conceptualization,
to determine the shipping network in the inland waterway. Because Methodology, Validation, Supervision, Writing – original draft, Writing
the shortest route may not be selected for some OD flows, which, – review & editing, Funding acquisition. Yalong Song: Methodology,
however, can reduce the overall costs of the transportation network. Software, Writing – original draft. Samson S. Yu: Writing – review &
The proposed model and methods can be used more scientifically in editing, Visualization. Dezhi Zhang: Data curation, Writing – review &
decision-making. editing. Tom Van Woensel: Writing – review & editing, Visualization.

7. Conclusions Declaration of competing interest

In this study, we addressed the HSND problem in an inland water- The authors declare that they have no known competing finan-
way with a tree-like structure. An integer linear programming model cial interests or personal relationships that could have appeared to
has been proposed to minimize the total operational costs by de- influence the work reported in this paper.
termining the optimal hub location, feeder port allocation, and fleet
deployment. A decomposition-based math-heuristic approach and an Data availability
enhanced genetic algorithm-based approach are proposed to solve the
problem. No data was used for the research described in the article.
Numerical results and a case study on the Yangtze River have
been conducted, which verifies the functionality and effectiveness of Acknowledgments
the proposed models and solution methods. Experimental results have
demonstrated that the operational costs of shipping decrease signifi- This work was supported by the National Natural Science Founda-
cantly with the increase in the number of hub ports until it reaches a tion of China under Grant No. 72001216, Central South University,
certain value. The hub-to-hub transportation discount factor directly af- China under Grant No. 202045007, National Natural Science Founda-
fects transportation costs. Lower total operational costs can be achieved tion of China under Grant No. 71672193, and National Natural Science
with a higher discount, which is well-known and further analyzed from Foundation of China under Grant No. 62003091.
the optimization perspective. The transportation cost contributes most
to the total operational cost of the shipping network, followed by the References
shipping cost, with the second highest percentage of the total cost.
Alfandari, L., Davidović, T., Furini, F., Ljubić, I., Maraš, V., & Martin, S. (2019). Tighter
Meanwhile, the results show that the existing hub ports do not
MIP models for barge container ship routing. Omega, 82, 38–54.
change noticeably when the number of hub ports increases, which Alibeyg, A., Contreras, I., & Fernández, E. (2018). Exact solution of hub network design
is consistent with the practical situation. Careful examinations of the problems with profits. European Journal of Operational Research, 266(1), 57–71.
results demonstrate that the chosen hub ports have larger throughput Alumur, S. A., Campbell, J. F., Contreras, I., Kara, B. Y., Marianov, V., & O’Kelly, M.
E. (2021). Perspectives on modeling hub location problems. European Journal of
due to the high level of regional development. In addition, we can
Operational Research, 291(1), 1–17.
observe that feeder ports on the trunk and tributaries are usually but Alumur, S., & Kara, B. Y. (2008). Network hub location problems: The state of the art.
not always associated with the nearest downstream hub port, which is European Journal of Operational Research, 190(1), 1–21.
synthetically influenced by the direction of the river stream, network An, F., Hu, H., & Xie, C. (2015). Service network design in inland waterway liner
structure, and OD flow distribution. transportation with empty container repositioning. European Transport Research
Review, 7(2), 9.
Further extensions could be implemented in the future by con- Azizi, N., Vidyarthi, N., & Chauhan, S. S. (2018). Modelling and analysis of hub-
sidering the ship speed optimization and carbon emission to obtain and-spoke networks under stochastic demand and congestion. Annals of Operations
a more reasonable model that can further reduce operational costs Research, 264(1), 1–40.
and emissions. Considering uncertainties in models (Khalilpourazari & Bu, F., & Nachtmann, H. (2021). Literature review and comparative analysis of inland
waterways transport:‘‘Container on Barge’’. Maritime Economics & Logistics, 1–34.
Hashemi Doulabi, 2022; Özmen et al., 2011; Savku & Weber, 2020)
Codato, G., & Fischetti, M. (2006). Combinatorial Benders’ cuts for mixed-integer linear
or inventory management (Tirkolaee et al., 2022) in ports are other programming. Operations Research, 54(4), 756–766.
practical directions. In addition, more sophisticated solution methods Contreras, I., Fernández, E., & Marín, A. (2009). Tight bounds from a path based
including stochastic programming (Kropat et al., 2020; Savku, 2022) formulation for the tree of hub location problem. Computers & Operations Research,
are worth developing to solve the HSND problem more efficiently. 36(12), 3117–3127.
Contreras, I., Fernández, E., & Marín, A. (2010). The tree of hubs location problem.
Other realistic factors that have not been considered in this study but European Journal of Operational Research, 202(2), 390–400.
exist in practice can be included when designing the hub-and-spoke Contreras, I., Tanash, M., & Vidyarthi, N. (2017). Exact and heuristic approaches for
network, such as speed optimization or time-dependent scheduling. the cycle hub location problem. Annals of Operations Research, 258(2), 655–677.

12
S. Zhou et al. Expert Systems With Applications 223 (2023) 119850

Fadda, P., Fancello, G., Mancini, S., Pani, C., & Serra, P. (2020). Design and Notteboom, T., Yang, D., & Xu, H. (2020). Container barge network development
optimisation of an innovative Two-Hub-and-Spoke network for the Mediterranean in inland rivers: A comparison between the Yangtze River and the Rhine River.
Short-Sea-Shipping market. Computers & Industrial Engineering, 149, Article 106847. Transportation Research Part A: Policy and Practice, 132, 587–605.
Fischetti, M., & Lodi, A. (2003). Local branching. Mathematical Programming, 98(1), O’kelly, M. E. (1987). A quadratic integer program for the location of interacting hub
23–47. facilities. European Journal of Operational Research, 32(3), 393–404.
Fontes, F. F. d. C., & Goncalves, G. (2021). A variable neighbourhood decomposition Ömer Nedim Kenger, & Ozceylan, E. (2023). A hybrid approach based on mathematical
search approach applied to a global liner shipping network using a hub-and-spoke modelling and improved online learning algorithm for data classification. Expert
with sub-hub structure. International Journal of Production Research, 59(1), 30–46. Systems with Applications, 218, Article 119607.
Gelareh, S., Maculan, N., Mahey, P., & Monemi, R. N. (2013). Hub-and-spoke net- Özgün-Kibiroğlu, Ç., Serarslan, M. N., & Topcu, Y. İ. (2019). Particle swarm optimiza-
work design and fleet deployment for string planning of liner shipping. Applied tion for uncapacitated multiple allocation hub location problem under congestion.
Mathematical Modelling, 37(5), 3307–3321. Expert Systems with Applications, 119, 1–19.
Gelareh, S., Monemi, R. N., & Nickel, S. (2015). Multi-period hub location problems in Özmen, A., Weber, G. W., Batmaz, İ., & Kropat, E. (2011). RCMARS: Robustification of
transportation. Transportation Research Part E: Logistics and Transportation Review, CMARS with different scenarios under polyhedral uncertainty set. Communications
75, 67–94. in Nonlinear Science and Numerical Simulation, 16(12), 4780–4787.
Hansen, P., Mladenović, N., & Urošević, D. (2006). Variable neighborhood search and Perboli, G., Tadei, R., & Vigo, D. (2011). The two-echelon capacitated vehicle routing
local branching. Computers & Operations Research, 33(10), 3034–3045. problem: Models and math-based heuristics. Transportation Science, 45(3), 364–380.
He, W., Jin, Z., Huang, Y., & Xu, S. (2021). The inland container transportation problem Radojčić, D., Simić, A., Momčilović, N., Motok, M., & Friedhoff, B. (2021). The Danube
with separation mode considering carbon dioxide emissions. Sustainability, 13(3), waterway with its tributaries. In Design of contemporary inland waterway vessels (pp.
1573. 17–33). Springer.
Hellsten, E. O., Sacramento, D., & Pisinger, D. (2021). A branch-and-price algorithm Real, L. B., Contreras, I., Cordeau, J.-F., de Camargo, R. S., & de Miranda, G. (2021).
for solving the single-hub feeder network design problem. European Journal of Multimodal hub network design with flexible routes. Transportation Research Part
Operational Research. E: Logistics and Transportation Review, 146, Article 102188.
Imai, A., Shintani, K., & Papadimitriou, S. (2009). Multi-port vs. Hub-and-Spoke port Roni, M. S. (2013). Analyzing the impact of a hub and spoke supply chain design for
calls by containerships. Transportation Research Part E: Logistics and Transportation long-haul, high-volume transportation of densified biomass. Mississippi State University.
Review, 45(5), 740–757. Ruocco, E., Wang, C., Zhang, H., & Challamel, N. (2017). An approximate model for
Jalal, A. M., Toso, E. A., Tautenhain, C. P., & Nascimento, M. C. (2022). An integrated optimizing Bernoulli columns against buckling. Engineering Structures, 141, 316–327.
location–transportation problem under value-added tax issues in pharmaceutical Savku, E. (2022). Fundamentals of market making via stochastic optimal control.
distribution planning. Expert Systems with Applications, 206, Article 117780. Operations Research, 136–154.
Ji, B., Zhang, D., Zhang, Z., Samson, S. Y., & Van Woensel, T. (2022). The generalized Savku, E., & Weber, G.-W. (2020). Stochastic differential games for optimal investment
serial-lock scheduling problem on inland waterway: A novel decomposition-based problems in a Markov regime-switching jump-diffusion market. Annals of Operations
solution framework and efficient heuristic approach. Transportation Research Part Research, 1–26.
E: Logistics and Transportation Review, 168, Article 102935. Sel, C., & Bilgen, B. (2014). Hybrid simulation and MIP based heuristic algorithm
Jiang, X., Lee, L. H., Chew, E. P., Han, Y., & Tan, K. C. (2012). A container yard storage for the production and distribution planning in the soft drink industry. Journal of
strategy for improving land utilization and operation efficiency in a transshipment Manufacturing Systems, 33(3), 385–399.
hub port. European Journal of Operational Research, 221(1), 64–73. Tan, Z., Liu, Q., Song, J., Wang, H., & Meng, Q. (2021). Ship choice and shore-power
Jiang, Y., Lu, J., Cai, Y., & Zeng, Q. (2018). Analysis of the impacts of different modes service assessment for inland river container shipping networks. Transportation
of governance on inland waterway transport development on the Pearl River: The Research Part D: Transport and Environment, 94, Article 102805.
Yangtze River Mode vs. the Pearl River Mode. Journal of Transport Geography, 71, Tan, Z., Wang, Y., Meng, Q., & Liu, Z. (2018). Joint ship schedule design and sailing
235–252. speed optimization for a single inland shipping service with uncertain dam transit
Kavirathna, C., Kawasaki, T., Hanaoka, S., & Matsuda, T. (2018). Transshipment hub time. Transportation Science, 52(6), 1570–1588.
port selection criteria by shipping lines: the case of hub ports around the bay of Tang, M., Ji, B., Fang, X., & Yu, S. S. (2022). Discretization-strategy-based solution for
Bengal. Journal of Shipping and Trade, 3(1), 1–25. berth allocation and quay crane assignment problem. Journal of Marine Science and
Khalilpourazari, S., & Hashemi Doulabi, H. (2022). A flexible robust model for blood Engineering, 10(4), 495.
supply chain network design problem. Annals of Operations Research, 1–26. Tirkolaee, E. B., Goli, A., Ghasemi, P., & Goodarzian, F. (2022). Designing a sustainable
Klincewicz, J. G. (1998). Hub location in backbone/tributary network design: a review. closed-loop supply chain network of face masks during the COVID-19 pandemic:
Location Science, 6(1–4), 307–335. Pareto-based algorithms. Journal of Cleaner Production, 333, Article 130056.
Korani, E., & Eydi, A. (2021). Bi-level programming model and KKT penalty function so- Tu, N., Adiputranto, D., Fu, X., & Li, Z.-C. (2018). Shipping network design in a
lution approach for reliable hub location problem. Expert Systems with Applications, growth market: The case of Indonesia. Transportation Research Part E: Logistics and
184, Article 115505. Transportation Review, 117, 108–125.
Kropat, E., Weber, G.-W., & Tirkolaee, E. B. (2020). Foundations of semialgebraic Wang, J. J., & Slack, B. (2000). The evolution of a regional container port system: The
gebe-environment networks. Journal of Dynamics & Games, 7, 253. Pearl River Delta. Journal of Transport Geography, 8(4), 263–275.
Labbé, M., & Yaman, H. (2008). Solving the hub location problem in a star–star Wang, S., & Zhao, Q. (2021). Probabilistic tabu search algorithm for container
network. Networks: An International Journal, 51(1), 19–33. liner shipping problem with speed optimisation. International Journal of Production
Li, W., Wang, D., Yang, S., & Yang, W. (2021). Three Gorges Project: benefits and Research, 1–18.
challenges for shipping development in the upper Yangtze River. International Wetzstein, B., Florax, R., Foster, K., & Binkley, J. (2021). Transportation costs:
Journal of Water Resources Development, 37(5), 758–771. Mississippi River barge rates. Journal of Commodity Markets, 21, Article 100123.
Li, F., Yang, D., Wang, S., & Weng, J. (2019). Ship routing and scheduling problem Yaman, H. (2008). Star p-hub median problem with modular arc capacities. Computers
for steel plants cluster alongside the Yangtze River. Transportation Research Part E: & Operations Research, 35(9), 3009–3019.
Logistics and Transportation Review, 122, 198–210. Yang, Z., Shi, H., Chen, K., & Bao, H. (2014). Optimization of container liner network
Martins de Sá, E., Contreras, I., & Cordeau, J.-F. (2015). Exact and heuristic algorithms on the Yangtze River. Maritime Policy & Management, 41(1), 79–96.
for the design of hub networks with multiple lines. European Journal of Operational Zhang, H., & Zhao, X. (2015). Quantitative analysis of organizational behavior of
Research, 246(1), 186–198. container shipping in the upper and middle reaches of the Yangtze River based
Meier, J. F., & Clausen, U. (2018). Solving single allocation hub location problems on on hub-and-spoke network. Journal of Coastal Research, (73 (10073)), 119–125.
euclidean data. Transportation Science, 52(5), 1141–1155. Zheng, J., Meng, Q., & Sun, Z. (2015). Liner hub-and-spoke shipping network design.
Meng, Q., & Wang, X. (2011). Intermodal hub-and-spoke network design: incorporating Transportation Research Part E: Logistics and Transportation Review, 75, 32–48.
multiple stakeholders and multi-type containers. Transportation Research, Part B Zheng, J., & Yang, D. (2016). Hub-and-spoke network design for container shipping
(Methodological), 45(4), 724–742. along the Yangtze River. Journal of Transport Geography, 55, 51–57.
Msakni, M. K., Fagerholt, K., Meisel, F., & Lindstad, E. (2020). Analyzing different Zhu, S., Gao, J., He, X., Zhang, S., Jin, Y., & Tan, Z. (2021). Green logistics oriented
designs of liner shipping feeder networks: A case study. Transportation Research tug scheduling for inland waterway logistics. Advanced Engineering Informatics, 49,
Part E: Logistics and Transportation Review, 134, Article 101839. Article 101323.
Nam, H.-S., & Song, D.-W. (2011). Defining maritime logistics hub and its implication Zolfani, S. H., Görçün, Ö. F., Çanakçıoğlu, M., & Tirkolaee, E. B. (2023). Efficiency
for container port. Maritime Policy & Management, 38(3), 269–292. analysis technique with input and output satisficing approach based on Type-2
Neutrosophic Fuzzy Sets: A case study of container shipping companies. Expert
Systems with Applications, 218, Article 119596.

13

You might also like