See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/360313350

Clustering algorithm for a vehicle routing problem with time windows

Article in Transport · May 2022

DOI: 10.3846/transport.2022.16850

CITATIONS READS

13 2,379

4 authors:

Chau Thi Diem Le Duc Duy Nguyen

University of Debrecen Ho Chi Minh City University of Technology
12 PUBLICATIONS 50 CITATIONS 35 PUBLICATIONS 119 CITATIONS

SEE PROFILE SEE PROFILE

Judit Oláh Pakurár Miklós

John von Neumann University University Doctoral School of Management and Bu… University of Debrecen
285 PUBLICATIONS 7,361 CITATIONS 23 PUBLICATIONS 1,039 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Chau Thi Diem Le on 02 May 2022.

The user has requested enhancement of the downloaded file.

 TRANSPORT
ISSN 1648-4142 / eISSN 1648-3480
2022 Volume 37 Issue 1: 17–27
https://doi.org/10.3846/transport.2022.16850

CLUSTERING ALGORITHM FOR A VEHICLE ROUTING

PROBLEM WITH TIME WINDOWS

Thi Diem Chau LE1, 2, Duc Duy NGUYEN3, Judit OLÁH1, 4*, Miklós PAKURÁR5
1, 5Faculty
of Economics and Business, University of Debrecen, Debrecen, Hungary
2Dept of Quality Assurance, Ho Chi Minh City University of Technology – VNUHCM, Ho Chi Minh, Vietnam
3School of Business and Management, LUT University, Lappeenranta, Finland
4Faculty of Applied Sciences, WSB University, Dabrowa Górnicza, Poland

Submitted 11 October 2021; resubmitted 14 January 2022; accepted 10 March 2022

Abstract. The demand for daily food purchases has increased dramatically, especially during the Covid-19 pandemic. This
requires suppliers to face a huge and complex problem of delivering products that meet the needs of their customers on
a daily basis. It also puts great pressure on managers on how to make day-to-day decisions quickly and efficiently to both
satisfy customer requirements and satisfy capacity constraints. This study proposes a combination of the cluster-first –
route-second method and k-means clustering algorithm to deal with a large Vehicle Routing Problem with Time Windows
(VRPTW) in the logistics and transportation field. The purpose of this research is to assist decision-makers to make quick
and efficient decisions, based on optimal costs, the number of vehicles, delivery time, and truck capacity efficiency. A dis-
tribution system of perishable goods in Vietnam is used as a case study to illustrate the effectiveness of our mathematical
model. In particular, perishable goods include fresh products of fish, chicken, beef, and pork. These products are packed
in different sizes and transferred by vehicles with 1000 kg capacity. Besides, they are delivered from a depot to the main 39
customers of the company with arrival times following customers’ time window. All of the data are collected from a logis-
tics company in Ho Chi Minh city (Vietnam). The result shows that the application of the clustering algorithm reduces the
time for finding the optimal solutions. Especially, it only takes an average of 0.36 s to provide an optimal solution to a large
Vehicle Routing Problem (VRP) with 39 nodes. In addition, the number of trucks, their operating costs, and their utiliza-
tion are also shown fully. The logistics company needs 11 trucks to deliver their products to 39 customers. The utilization
of each truck is more than 70%. This operation takes the total costs of 6586215.32 VND (Vietnamese Dong), of which, the
transportation cost is 1086215.32 VND. This research mainly contributes an effective method for enterprises to quickly
find the optimal solution to the problem of product supply.
Keywords: vehicle routing problem, logistics, time window, cluster, route, k-means clustering, transportation.

Notations
Variables and functions:
c – transportation unit cost per km [VND/km]; ei – the earliest time that customer can receive
C – set of customers; the products;
Ck – the be a set of K centers in the k-means clus- f – refrigeration unit cost per km [VND/km];
tering model; jj – the latest time that customer can receive the
CTk – maximum permissible load of truck k [kg]; products;
d ‒ dimensional real vector; G – graph;
dij – distance from node i to node j [km]; Kij – maximum spending time between 2 nodes;
Di – demand of customer i; L – labour cost per truck per day [VND];
dist (di, ck) – the Euclidean distance between a data point pi – processing time of customer i;
di and the cluster center ck; Sj – the set of samples that belong to the kth
Ei – estimated volume of products of customer i; cluster;

*Corresponding author. E-mail: juditdrolah@gmail.com

Copyright © 2022 The Author(s). Published by Vilnius Gediminas Technical University

This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unre-
stricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
18 T. D. C. Le et al. Clustering algorithm for a vehicle routing problem with time windows

Sik – an integer variable that present the time of from academics and managers Zhang et al. 2019). In the
vehicle k starts to service a customer i; practical, VRP appears as several classes of additional
tij – moving time from node i to node j; constraints, such as considering customer time windows,
(ji, ei) – the time window at customer i where ji the route lengths, the limits on vehicle capacity, etc. In which,
earliest time that customer i can receive the the VRPTW is an important variant of the VRP and has
products and ei is the latest time that cus- been paid attention from researchers (Wang et al. 2020).
tomer i can receive the products; The VRPTW aims to determine the optimal set of routes
n – speed [km/h]; for a set of restricted capacity identical vehicles and sat-
V – vehicle; isfy the time window constraints (El-Sherbeny 2010).
VTk – maximum volume of truck k [m3]; Each customer is served exactly once by a vehicle within
xijk – a binary variable that equals to 1 if vehicle k the time window defined by the earliest and latest times.
drives directly from node i to node j, other- Each vehicle trip starts and finishes at the given depot.
wise, it is 0; The locations, demand, and time windows of customers
X – a set of observations. are known as a priority. In the VRPTW, the vehicle could
Abbreviations: arrive at the customer’s location before the earliest time
ASEAN – Association of South-East Asian Nations; and wait without cost or a penalty cost for example park-
GDP – gross domestic product; ing slots cost until service becomes allowable. However,
IMF – International Monetary Fund; the vehicle must not arrive at customers after the latest
LIFO – last in first out; time. The time window constraints in the VRPTW are re-
MILP – mixed integer linear programming; quired strictly satisfied, however, in many practical prob-
MIP – mixed integer programming; lems, these constraints can be violated. There are 2 types
VND – Vietnamese Dong; of VRPs that allow violating the time windows constraints:
VRP – vehicle routing problem; (1) VRP with flexible time windows and (2) VRP with soft
VRPTW – VRP with time windows. time windows.
There are different methodologies for determining the
Introduction solutions of VRP and its variants such as exact algorithm,
heuristic methods, and hierarchical method (Comert et al.
Logistics costs are costs associated with the process of dis- 2018). In the hierarchical method, the problem is system-
tribution and circulation of goods, which have a direct atically split into different levels for determining solutions.
impact on the efficiency of businesses, localities, and the In this research, one of the hierarchical methods named
national economy. Logistics costs are paid by businesses cluster-first – route-second was employed. This study in-
for supply activities in the market. Therefore, it can ac- troduces a practical problem of vehicle routing in the field
celerate or inhibit product consumption and influence of logistics and transportation. This problem mainly takes
the competitiveness of enterprises and the whole econo- into account three constraints, including time window,
my. Logistics costs mainly consist of transportation costs, vehicle capacity, and product type. The aims of this study
warehousing, packing and losses, inventory, and order are to determine the number of vehicles and minimize to-
processing, and administrative (Rushton et al. 2022). tal logistics costs. In which, logistics costs include labour,
In Vietnam, logistics costs are much higher than in fuel, and refrigeration costs. The MIP model is built to
many countries ASEAN region and around the world such describe the product delivery network with 39 nodes. The
as Thailand, Singapore, and so on. According to the data clustering algorithm is used to solve the proposed MIP
of the World Bank, logistics costs including transporta- model. The results show that 11 trucks are used to serve
tion, storage, customs clearance, and so on in Vietnam 39 customers and the truck utilization efficiency is more
are about 20.9…25.0% of GDP (VILAS 2021). In which, than 70%. These optimal results are found in an average
transportation costs account for about 60%, a high cost of 0.36 s.
compared to developed countries. This cost is higher than This paper consists of several sections as follows. Sec-
Thailand’s 6%, Malaysia’s 12%, and 3 times higher than tion 1 is the literature review, which describes the over-
Singapore’s. According to the IMF, in the US economy, lo- view of the research gaps of previous studies. A math-
gistics costs account for 9.9% of the country’s GDP ($921 ematical model of the VRP and methodology are shown
billion in 2000). High logistics cost situations could cause in Sections 2 and 3, respectively. Section 4 presents a case
a low level of enterprises’ competitiveness. Therefore, re- study, including data collection, assumptions, and results.
ducing logistics costs, especially transportation costs, is a The finally is the conclusions.
big challenge for Vietnam’s logistics industry.
VRP is formed to minimize the transportation cost. 1. Literature review
The development of VRP is based on the multiple trave-
ling salesman problems (Toth, Vigo 2014; Xu et al. 2018). VRP is an important topic considered by many managers
VRP played a vital role in the field of logistics and trans- in the logistics and transportation fields. According to Cao
portation. Consequently, VRP attracts a lot of attention and Yang (2017), there are two basic types of VRP, includ-
Transport, 2022, 37(1): 17–27 19

ing pure VRPs and their variations. In which, pure VRP hicles in their objective functions. Kantawong and Praves-
refers to as the distribution of cargo from one destination jit (2020) implemented the optimization of distance and
to many other locations (Yao et al. 2019). The solutions of vehicle service cost. Londoño et al. (2021) also optimized
VRP bring more benefits for companies such as shortest distance in addition to minimizing transportation costs
routes and time, the number of necessary vehicles, and op- and penalty fees. Qi and Hu (2020) found the solution
timal vehicle capacity. These results contribute to reduce to optimize three types of cost in their cost function, in-
the total costs of logistics and increase businesses’ profit cluding the costs of fuel, refrigeration, and cargo damage.
(Le et al. 2020; Zhu et al. 2020). Simultaneously, they also optimize the number of vehicles
In recent years, VRP has gained more interest from used in their problem. Tasar et al. (2019) determined the
academics. There are many characteristics of VRP that are optimal number of vehicles to minimize the fixed and fuel
solved by different mathematical models, and their results costs. Similarly, Aggarwal and Kumar (2019) established
meet the aims of businesses (Table 1). Table 1 shows that the objective to drop their transportation cost and cost
the differences in VRP characteristics between previous for the availability of vehicles and determine the optimal
studies focus on objectives, problem description, and vehicle number. Unlike introduced studies, Afifi et al.
mathematical model and algorithm. First of all, the cost (2016) and Pérez-Rodríguez, Hernández-Aguirre (2019)
is the most concern of previous studies while time, dis- focused on reducing time and distance in their problem,
tance, and the number of vehicles have less appearance in respectively.
their objective function. Zhu and Hu (2019) considered Secondly, problem description in almost previous stud-
the cost of opening routes in the objective equation (Ruiz ies includes network configure, constraint, and data. Net-
et al. 2019). Zhang et al. (2017) and Spliet and Desaulniers works configure from previous studies are sets of nodes
(2015) introduced the transportation cost as a factor to be ranged from 6 nodes to 200 nodes. These problems mainly
reduced in their objective function. In a more complicated deal with two constraints: time window and vehicle capac-
way, Birim (2016) considered both transportation cost and ity. In which, a time window is a more popular constraint
fixed vehicle cost. Zhu and Hu (2019) minimized three than the vehicle capacity. According to Bachu et al. (2021),
types of costs that include fuel cost, labour cost, and ve- building a model to respond to time constraints is a major
hicle depreciation cost. Some studies consider both costs challenge. Thus, almost all previous studies consider one
and other factors such as distance and the number of ve- of two constraints when determining the optimal solution.

Table 1. VRPs of previous studies (source: current study)

Problem description
Objective
network Constraint Data Mathematical
Study
no of dis- configuration time vehicle kind of experi- model and algorithm
cost time reality
vehicle tance (N – nodes) window capacity products ence
Qi, Hu (2020) ü ü 13 N ü MIP, heuristic
Kantawong, MIP, artificial bee
ü ü 100 N ü ü
Pravesjit (2020) colony
Londoño et al.
ü ü 51 N ü MILP, local search
(2021)
Aggarwal, Kumar
ü ü 60 N ü ü MIP
(2019)
Pérez-Rodríguez,
an estimation of
Hernández- ü 6N ü ü
distribution
Aguirre (2019)
Tasar et al. (2019) ü ü 100 N ü MIP, heuristic
MILP, response
Zhu, Hu (2019) ü 200 N ü ü
surface method
MIP, biased
Ruiz et al. (2019) ü 14 N ü ü
random – key genetic
tabu search, the
Zhang et al. (2017) ü 27 N ü ü
artificial bee colony
MILP, simulated
Birim (2016) ü 10 N ü ü
annealing
MIP, simulated
Afifi et al. (2016) ü 30 N ü ü
annealing
Spliet, Desaulniers MIP, exact branch –
ü 60 N ü ü ü
(2015) price – cut algorithm
MIP, clustering
Current study ü ü 39 N ü ü ü ü
algorithm
20 T. D. C. Le et al. Clustering algorithm for a vehicle routing problem with time windows

For example, Kantawong and Pravesjit (2020), Aggarwal 2. Methodology

and Kumar (2019), or Pérez-Rodríguez and Hernández-
Aguirre (2019) solved their problem with time window. In this study, the vehicle routing time windows problem is
Zhu and Hu (2019) or Ruiz et al. (2019) presented the ve- formulated as MIP firstly. In particular, the main features
hicle capacity constraint in their model. Especially, Spliet of the problem are presented to provide a general descrip-
and Desaulniers (2015) determined the solution for their tion of the model being built. The objective function and
problem when considering both time window and vehi- constraints, as well as the key assumptions, are also for-
cle capacity. However, there are some studies solving their mulated to specifically describe the mathematical model.
problem without constraints such as Qi, Hu (2020) and In order to solve large-scale problems, we introduce the
Tasar et al. (2019). All VRPs are solved using data from cluster-first – route-second method. This is a hierarchical
experience or practice. Empirical data are used in more approach that applied dead with a large-scale instance by
research than data from in reality. Particularly, empirical using an exact algorithm in a reasonable time. In the 1st
data are used to solve problems involving multiple nodes phase of the method, a clustering algorithm is applied to
in a network configuration. Besides, it is also common in form the clusters. After that, the capacity of each cluster
problems containing one or two constraints, even prob- is controlled by adjusting members in clusters. In the 2nd
lems with no constraints. phase, each subproblem for each cluster is solved by us-
Last, but not least, mathematical models and algo- ing the branch and bound algorithm. k-means clustering
rithms are used to find the suitable solution for each VRP algorithm and capacitated k-means clustering algorithm
in previous studies. Previous studies have built two math- are used to cluster customers. This approach is applied to
ematical models for their problems, MIP and MILP. In ad- a real case study given in Section 3.1.
dition, they apply various algorithms to find the solution
quickly and efficiently. Qi and Hu (2020) applied a heuris- 2.1. Mathematical model of the VRPTW
tic to find solutions for MIP models with 13 nodes. Simi-
larly, Tasar et al. (2019) used heuristic algorithms to deal The VRPTW is a generalization of the VRPs. The model
with a MIP consisting of 100 nodes in the network config- not only aims to minimize the cost of routes but also satis-
ure. Birim (2016) and Afifi et al. (2016) adopted simulated fy the time window (Wang et al. 2020). The customer must
annealing to solve MILP and MIP, respectively. Some oth- be served within a specified time window with bound is
er studies use the algorithm of artificial bee colony (Kanta- the earliest time and the latest time. In our research, the
wong, Pravesjit 2020), local search (Londoño et al. 2021), model “one-to-all”, all vehicles need to start and return at
response surface method (Zhu, Hu 2019), biased random a depot (Figure 1). In Figure 1, the warehouse icon de-
key genetic (Ruiz et al. 2019), or exact branch-price-and- picts where the products are stored and where the vehi-
cut algorithm (Spliet, Desaulniers 2015). Especially, Zhang cles depart. The direction of the arrows also indicates the
et al. (2017) applied both tabu search and the artificial bee direction of each truck in a route. Customers on different
colony for their problems. routes are presented with polygons of different colours.
Similar to some previous studies, MIP is built to solve Ordered products will be placed on the vehicle to ship to
the VRP in order to determine the required number of ve- the respective customer. The number and weight of prod-
hicles and minimize the total logistics costs in this study. ucts must meet the storage capacity of each vehicle. Based
The vehicles considered in both this study and most pre- on the customer’s order requirements, two main assump-
vious studies were trucks designed to transport products
(Tretjakovas, Čereška 2021). The objective function also Node / customer
considers three types of costs, including labour, fuel, and
refrigeration costs. However, unlike previous studies, this t – traveling time
study deals with a real VRP so the data collected from
the real world will be used in our mathematical model. In
addition, due to being a real problem, the mathematical
model of this study has more constraints than previous (ji, ei)
studies. Time window, vehicle capacity, and kind of prod- Time window at customer
Depot
ucts are considered in the proposed mathematical model.
Besides, the logistics network in this study is also larger
than the network of some previous studies using real data.
For example, Pérez-Rodríguez, Hernández-Aguirre (2019)
and Ruiz et al. (2019). Furthermore, there is a difference
in algorithmic application between this study and related
studies. A clustering algorithm is applied in our study to
solve the VRP with 39 nodes in the network structure. Figure 1. An example about VRPTW (Wang et al. 2020)
Transport, 2022, 37(1): 17–27 21

tions are considered in this mathematical model. They presents that customers need to be visited exactly once by
include: (1) the current total number of vehicles meet the a vehicle. On other words, one customer is only served
total customer demand per day and (2) each customer by one truck. Equation (3) states that the used vehicle
orders the number of products for the day not to exceed could not exceed the number of available vehicles. Equa-
the capacity of one vehicle. As a result, each vehicle is ar- tions (4) and (5) mean vehicle capacity need to be satis-
ranged on a route to serve customers and carry a variety of fied. This means that transporting products to customers
suitable products. This means that each vehicle can serve is not exceed the maximum capacity of a truck. In which,
many customers in its trip. By contrast, one customer only the vehicle capacity is calculated by both permissible load
receives products from only one vehicle at one time. with the unit of a [kg] and volume with the unit of [m3].
The VRP time windows is the VRPTW is formed by Equation (6) ensures that every vehicle leaves at the de-
a set of V vehicles, a set of nodes that includes a set of pot zero to a customer (Equation (7)) and then finishes
customer C and a driving-out depot, and a returning de- the trip at the depot n + 1 (Equation (8)). Equation (9)
pot. All of the nodes together generate a directed graph presents that every vehicle from i to j cannot arrive at j
G = (V , C ). Customers are denoted from 1 to n with n is before Sik + tij. Where tij is the spending time from i to
the number of customers, the driving – out depot and re- j that is equal to total moving time from i to j and the
turning depot are presented by 0 and n + 1, respectively. moving time at customer j. The moving time is calculated
The network is a set of N nodes that consist of C + 2 based on the distance between i and j and vehicle velocity
nodes. An arc ( i, j ), where i ≠ j is an arc from node i to (Equation (10)). Equation (11) defines a value of Kij that
node j. No outgoing arc at node n + 1 and no arc come to helps linearize a non-convex optimization. Finally, Equa-
node 0. However, in some cases, node 0 and node n + 1 tion (12) ensures that time windows are observed. Equa-
are in the same location. tions (13) and (14) show type of variables.
min ∑∑∑ (c + f ) ⋅ dij ⋅ xijk + L ⋅ ∑∑ x0 jk (1)
2.2. k-means clustering algorithm
k∈V i∈N j∈N k∈V j∈C
subject to: k-means clustering is a well-known clustering method
that is used to form n observations to k clusters, where
∑∑ xijk = 1, ∀i ∈ C ; (2) k is known as a priority. The objective of the algorithm is
k∈V j∈C to minimize the objective function and to separate each
∑∑ x0 jk ≤ V , ∀k ∈V , ∀j ∈ N ; (3) compact class as far as possible. k-means clustering could
be explained as below (Khan, Ahmad 2004).
k∈V j∈C
Given X = {x1 , x2 , ..., xn } is a set of observations,
∑ Di ⋅ ∑ xijk ≤ CTk , ∀k ∈V ; (4)
and each observation is a d ‒ dimensional real vec-
i∈C j∈C
tor. CCk = {c1 , c2 , ..., cn } is the be a set of K centers.
∑ Ei ⋅ ∑ xijk ≤ VTk , ∀k ∈V ; (5) S j = {d | d is member of cluster k} be the set of samples
i∈C j∈C that belong to the kth cluster.
∑ x0 jk = 1, ∀k ∈V ; (6)
n

∑ dist (di , ck ), (15)

j∈C
i =1
∑ xihk − ∑ xhjk =
0 , ∀h ∈ C , ∀k ∈ V ; (7) where: dist ( di , ck ) is the Euclidean distance between a
i∈N j∈N data point di and the cluster center ck.
∑ xi,n+1,k = 1, ∀k ∈V ; (8) The procedure for the k-means algorithm consist of
4 steps:
i∈C
Sik + tij − K ij ⋅ (1 − xijk ) ≤ S jk , ∀i, j ∈ C , ∀k ∈ V , (9) »» step 1: a set of ck was initialized by using random
sampling;
where:
dij »» step 2: decide the members of each cluster based on
t=
ij + pi ; (10) the minimum distance from cluster center criteria;
n »» step 3: ck will be calculated as Equation (16). Sk is
= (
K ij max ei + tij − j j , 0 ; ) (11) the number of data items in the kth cluster:
n
ai ≤ Sik ≤ bi , ∀k ∈ V ; (12)
∑ di
xijk ∈ {0, 1}, ∀i, j ∈ N , ∀k ∈ V ; (13) ck =
di∈Sk
; (16)
Sk
Ski ∈ {0, 1}, ∀i ∈ N , ∀k ∈ V . (14)
»» step 4: step 2 and step 3 could be repeated until the
The model aims to determine the required numbers of objective is optimal, and the algorithm reaches the
vehicles to meet the needs of customers in order to mini- maximum number of iterations.
mize both labour cost and traveling cost (fuel cost and re- However, the classical means clustering based on the
frigeration cost) as described in Equation (1). Equation (2) distance among members and centers. In the VRP, the ca-
22 T. D. C. Le et al. Clustering algorithm for a vehicle routing problem with time windows

pacity of the truck needs to be taken into account. There- circles. The notations for customers in the map begin with
fore, Comert et al. (2018) introduced the equation for C, for example, C1 is the name of customer 1. The distance
determining the number of k in k-means clustering in a matrix among depots and customers is calculated based
capacitated VRP problem. The number of clusters is based on latitudes and longitudes of their locations from Google
on the capacity of trucks and demand: Maps.
total demand The logistics company has 15 1000 kg reefer trucks
number of clusters = . (17) that can flexibly deliver goods in Ho Chi Minh city urban
truck capacity
in rush hours and off-peak hours. Every day, the planning
The number of clusters is round up to an integer department in the warehouse get orders from customers
number. For example, in this research, the total demand and assigns customers for trucks. Next, the driver sorted,
is 10020 kg, and a truck that has a capacity is 1000 kg. As loaded, and delivered to customer. The product in a truck
a result, the number of clusters is calculated is 10.020 kg is arranged and grouped by customers and follows the
then the decimal rate is round up to 11. Once the number LIFO rule. The velocity of each truck is assumed 45 km/h.
of clusters is determined, the k-means algorithm is per- The working time is from 8:30 to 15:30 h. The earli-
formed in the data. The related customers are assigned to est time and the latest time for serving by trucks and the
clusters has total demand could not exceed the capacity. processing time are varied by customers.
Set number base value is specified by considering the con- In this research, the truck cost is focused on labour
dition: truck utilization is 70% truck. After the initial set cost, fuel cost, and refrigeration cost. Each truck is op-
of members in each cluster is generated from the k-means erated by a driver, where the salary was assumed to be
clustering algorithm based on the customer locations. The 500000 VND/day. The salary is calculated based on the
cluster demand is checked whether it is satisfied with the salary provided by the human resource department. Based
truck capacity and truck utilization condition (less or on the manufacturer’s technical report, each truck with
more than 700 to 1000 kg), the member of each group full load cost 2551 VND/km for fuel. The refrigeration
could be modified and capacity controlled repeated until cost per km is assumed to equal 20% of the fuel cost.
set satisfied the conditions.
3.2. Assumptions
3. A case study
Assumptions are:
3.1. Data collection »» the problem is static;
»» the model uses one type of reefer truck;
In this research, a transportation network from a logis- »» the products of customers are packed and arranged
tics company in Ho Chi Minh city (Vietnam) is consid- before being shipped;
ered. The company provides warehousing and distribu- »» the number of customers, customer time window,
tion services for 39 main customers in Ho Chi Minh city. customer demand, processing time at customer lo-
As shown in Figure 2, a red star is used to represent the cations are deterministic and they are known as a
depot, which also is the warehouse and parking place, priority;
whereas the locations of customers are presented by blue »» the fuel cost and refrigeration cost are deterministic;
»» customer demand is changed day-by-day;
»» every truck can start at the depot at j0 = 0 and re-
turn the depot at e0 = 420 (7 working h);
»» each customer may be serviced by one tractor truck;
»» backhauls are not permitted;
»» 15 vehicles can meet the total demand of customers
every day;
»» each customer’s demand per day is less than or equal
to 1000 kg.

3.3. Results and analysis

The VRPTW model was implemented by IBM ILOG
CPLEX Optimization Studio (version 12.10.0.0, https://
www.ibm.com/products/ilog-cplex-optimization-studio)
and the k-mean clustering algorithm was applied in MAT-
LAB R2019a (https://www.mathworks.com/products/mat-
lab.html). The results from the model were visualized by
using Tableau Desktop 2019 version (https://www.tableau.
com/products/desktop). All of the software was run on a
computer with Intel I7-10510U-1.80GHz (8 CPUs), Win-
Figure 2. Customer and depot locations (source: current study) dow 10 operation system.
Transport, 2022, 37(1): 17–27 23

3.3.1. k-means clustering Table 3. Classical k-means clustering results

(source: current study)
In order to solve a large – scale VRPTW problem, clus-
tering methodology was applied to dispatch the original No of Demand
problem to subproblems that be able to determine in a Cluster No of node
cluster [kg] [%]
practical time with a limited computational system. When
k=1 1 39 10020 100.00
applying the classical k-means clustering algorithm, the
1 20 5589 55.78
number of clusters is an important factor that needs to k=2
be considered. Besides, the number of clusters affects the 2 19 4431 44.22
number of members in each cluster, therefore, it affects the 1 20 5051 50.41
overall solutions of the VRPTW problem after dispatching. k=3 2 14 4459 44.50
A set of data for the k-means clustering algorithm that 3 5 510 5.09
includes latitude and longitude of customers was used as 1 18 3574 35.67
input data, there were altogether 39 customers. It is also
2 5 2021 20.17
noted that each customer has a demand that was satisfied k=4
3 5 510 5.09
by a truck. The output from the classical k-means cluster-
ing algorithm was the cluster indices of customers. Table 2 4 11 3915 39.07
shows an example about input and output of k-means 1 9 3065 30.59
clustering algorithm with the number of clusters is set to 2 5 510 5.09
equal to 2. For instance, customer 1 and customer 3 are in k=5 3 3 1250 12.48
cluster 1, customer 2, and customer 38 in cluster 2. 4 17 3174 31.68
Table 3 presents a summary result from performing 5 5 2021 20.17
classical k-means clustering with different values of the
1 3 1250 12.48
number of clusters (value of k). The algorithm was ex-
ecuted in MATLAB. Typically, the objective function from 2 9 3065 30.59
k-means clustering contains local optimal solutions; there- 3 3 374 3.73
k=6
fore, the number of replications was set to 10, and the 4 9 2124 21.20
number of interactions in each replication was set to 100. 5 6 2133 21.29
When increasing the number of replications and number 6 9 1074 10.72
of interactions in each replication beyond these values, the 1 2 128 1.28
objective was not improved.
2 9 2124 21.20
When the number of clusters is equal to 1, all of the
3 9 1074 10.72
customers are in the sample cluster. The number of clus-
ters is increased from k = 2 to k = 11. The increasing k=7 4 3 1250 12.48
number of clusters (number of k) makes the number of 5 9 3065 30.59
members in a cluster decreased. For example, when the 6 5 2021 20.17
number of clusters is equal to 2 (k = 2), the maximum 7 2 358 3.57
number of members in a is 20 while number the number 1 2 366 3.65
of clusters is equal to 11 (k = 11), the maximum number
2 9 3065 30.59
of customers in a cluster is 6. This is the key point to help
3 8 2100 20.96
reduce the size of the network.
4 2 128 1.28
k=8
5 3 1250 12.48
Table 2. Input-output of the classical k-means clustering
algorithm (source: current study) 6 5 2021 20.17
7 9 1074 10.72
Input Output
8 1 16 0.16
customer latitude longitude cluster indices
1 3 1250 12.48
1 10.74168068 106.6298856 1
2 2 128 1.28
2 10.80059579 106.743338 2
3 2 366 3.65
3 10.8034963 106.6812906 1 4 6 1520 15.17
4 10.80097871 106.6607492 1 k=9 5 8 2665 26.60
5 10.86402657 106.6483003 1 6 9 1074 10.72
… … … … 7 1 16 0.16
38 10.81670269 106.7064081 2 8 3 980 9.78
39 10.78474569 106.757709 2 9 5 2021 20.17
24 T. D. C. Le et al. Clustering algorithm for a vehicle routing problem with time windows

End of Table 3 Table 4. Capacitated k-means clustering results

(source: current study)
No of Demand
Cluster No of node
cluster [kg] [%] Demand
Cluster No of node
1 3 980 9.78 [kg] [%]
2 3 1250 12.48 1 4 1000 9.98
3 6 1520 15.17 2 2 995 9.93
4 8 2665 26.60 3 3 720 7.19
5 6 524 5.23 4 1 1000 9.98
k = 10
6 1 16 0.16 5 6 840 8.38
7 2 128 1.28
6 4 982 9.80
8 4 1708 17.05
7 3 980 9.78
9 2 366 3.65
8 3 829 8.27
10 4 863 8.61
9 1 1000 9.98
1 6 1520 15.17
2 1 16 0.16 10 4 970 9.68
3 2 128 1.28 11 8 704 7.03
4 3 1250 12.48 Note: % represents the ratio between the demand of each cluster
5 6 2455 24.50 and the total demand of customers.
k = 11 6 2 366 3.65
7 2 580 5.79 ent colours and shapes. Customers in the same cluster or
8 4 1708 17.05
group are presented with the same icon and colour. For
example, customers 4, 7, 21 and 27 are in a cluster repre-
9 3 610 6.09
sented by red squares.
10 4 863 8.61
11 6 524 5.23 3.3.2. VRPTW
The clusters and number of members in each cluster were
k-means clustering algorithm aims to minimize the determined and explained in detail in Section 3.3.1. Then,
distance from members to centers. Therefore, without the routing problem was solved by applying the branch
considering the demand of each cluster, the proportion and bound algorithm that was formulated in CPLEX. The
of demand of clusters varies due to a different number of running time is limited to 2 hours, the MIP gap is set to
members. For example, at k = 2, the demand of cluster 1 is default setting.
55.78% and the demand of cluster 2 is 44.22%. Demands The computational results from a different number of
between the 2 clusters are not significantly different. How- clusters from classical k-means clustering are shown in Ta-
ever, at k = 3, the highest cluster demand is 50.41%, while ble 5. Results from solving mathematical model include
the lowest cluster demand is 5.09%. Without consider- sub-objectives and overall objectives that are transporta-
ing demand and truck capacity, the model cannot satisfy tion costs, the number of used trucks, and running time
the truck utilization condition that is at least 70%. For for determining a solution.
instance, at k = 9, sum of demand of cluster 1 is equal As mentioned in Table 3, without considering demand
to 1250 kg while the requirement is 2 1000 kg trucks. So and truck capacity, k-means clustering forms clusters and
that the truck utilization is equal to 62.50% (truck capacity causes a low truck utilization level. Therefore, when the
is 1000 kg). Similarly, the sum of demand of cluster 6 is number of clusters increased, the number of trucks was
equal to 16 kg while truck capacity is 1000 kg. Therefore, used to deliver products is tended to increase. At k = 2, the
truck utilization is equal 0.16%. model decides to use 12 trucks and at k = 11, the model
Capacitated k-means clustering algorithm was applied uses 17 trucks. As a result, the overall cost is from k = 2
to reduce the bias among clusters by considering demand to k = 11 increases from 6661671.46 to 10036219.62 VND.
and truck capacity (Khan, Ahmad 2004). Table 4 presents By dispatching a large-scale problem to subproblems,
the results from capacitated k-means clustering that num- the running time is decreased from k = 1, the problem could
ber of clusters equal to 11. be solved in a limited computational system, to k = 11,
As shown in Table 4, the demand of each cluster was the problem needs [ms] to generate the results.
from 704 to 1000 kg, which can satisfy the truck capacity An interesting finding was that the solving time by ap-
at least 70%. Although cluster 4 and cluster 9 have only plying branch and bound in CPLEX not only depend on
one customer in a cluster, their demand already fulfilled the size of problems (number of nodes) but also the data
truck capacity. Figure 3 presents customer and cluster set. The average solving time varies due to the size of prob-
locations from capacitated k-means clustering. There are lems, at k = 2, the model contains 2 subproblems that take
11 clusters or 11 groups of customers depicted in differ- an average solving time is 7214.43 s. The model stops solv-
Transport, 2022, 37(1): 17–27 25

Figure 3. Customer and cluster locations from capacitated k-means clustering (source: current study)

Table 5. Computational results for different numbers of k from Table 6. An example demand set for problems have
classical k-means clustering algorithm (source: current study) 9 customers (source: current study)

No of No of Traveling Overall costs Solving Demand [kg]

cluster truck cost [VND] [VND] time [s] Problem 1st 420 480 320 100 100 80 100 24 500
k=1 ‒ ‒ ‒ ‒ Problem 2nd 915 1000 60 60 300 80 150 400 100
k=2 12 661671.46 6661671.46 7214.43
k=3 13 801425.99 7301,425.99 2548.67
Table 7. Computational results according to capacitated
k=4 12 939921.40 6939921.40 1804.98 k-means clustering algorithm (source: current study)
k=5 14 1012314.21 8012314.21 1445.70
k=6 15 1275519.08 8775519.08 57.44 Traveling cost Overall costs Solving time
No of truck
[VND] [VND] [s]
k=7 16 1376399.53 9376399.53 46.64
k=8 17 1473159.71 9973159.71 15.65 11 1086215.32 6586215.32 0.36
k=9 16 1417516.89 9417516.89 9.39
k = 10 16 1459383.50 9459383.50 0.31 has 1229 rows (constraints), 1200 columns(variables), and
k = 11 17 1536219.62 10036219.62 0.26 7630 non-zero variables while a reduced MIP for problem
2 has 979 rows, 950 columns, and 5970 non-zero variables.
ing due to meet the time limit condition. The MIP gaps Therefore, the solving time for problem 1st and problem
were recorded that 20.98% for the first cluster and 49.11% 2nd is 224.12 and 16.90 s, respectively.
for the second cluster. At k = 11, there are 11 subproblems Table 7 presents a summary result from VRPTW ac-
in the model, and they spend [ms] to generate a solution cording to capacitated k-means clustering algorithm. The
with a MIP gap equal to 10 – 4 (Equal to the default set- model uses 11 trucks to deliver products. The traveling
ting in CPLEX). Although the problems contain the same cost and overall costs are lower than these costs from
number of nodes, the solving time could be different. In the solution at k = 2 for classical k-means clustering. The
some cases, if more arcs will be considered, in an optimal subproblems take an average of 0.36 s for generating a
solution a larger number of variables will take a non-zero solution. Although applying capacitated k-means cluster-
value that causes a larger number of solutions to need ing algorithm and branch and bound was not guarantee
to be considered, therefore, an increase in the running optimality, capacitated k-means clustering for VRPTW is
time of CPLEX will be occurred (Ekşioğlu et al. 2009). suitable for the VRPTW problem due to the acceptable
For example, there is a problem (Table 6) that has 9 cus- cost and practical running time. The data set for VRPTW
tomers and 1 deports that has 1320 variables (110 integer could be changed every day, especially demand. Therefore,
variables and 1210 binary variables) and 1679 constraints. the decision-maker needs to generate solutions in a practi-
Though a presolve problem, a reduced MIP for problem 1 cal time with a limited configuration computer.
26 T. D. C. Le et al. Clustering algorithm for a vehicle routing problem with time windows

Conclusions »» although applying capacitated k-means clustering

algorithm and B&B was not guarantee optimality,
This research presents cluster-first – route-second ap- capacitated k-means clustering for VRPTW is suit-
proach for a large – scale VRPTW. The customers were able for the VRPTW problem due to the acceptable
clustered by using clustering algorithms such as classical cost and practical running time.
k-means clustering algorithms and capacitated k-means The data set for VRPTW could be changed every day,
clustering algorithms. Classical k-means clustering algo-
especially demand. Therefore, the decision-maker needs
rithms cluster the customer locations and capacitated k-
to generate solutions in a practical time with a limited
means clustering algorithms take into account both cus-
configuration computer. So that, the methodology of this
tomer’s location and truck capacity. By clustering process,
study is a fundamental foundation for similar practical
the scale of VRPTW was reduced. Therefore, branch and
instances. Delivering essential food and products during
bound algorithm could be executed to generate the solu-
Covid-19 situation is a good example. When the Covid-19
tion in a practical time and limited laptop configuration.
pandemic occurred, the governments of many countries
The effectiveness of proposed method is demonstrat-
ordered a lockdown to eliminate the spread of the virus.
ed by the application of this method in a real case study
As a result, product distribution is severely affected by
from Vietnam. In particular, the considered logistics com-
pany has 39 customers in Ho Chi Minh city, and 1 type strict timing in each country, city, and locality. Based on
of truck. The truck utilization must be more than 70%. our method, it can significantly reduce the vehicle sched-
This distribution network is considered to be larger than uling time and the required products can be delivered
some previous studies using real data, which are clearly to the customer on time. Another example is school bus
presented in the literature review. The results show that routing problems. These issues face diverse needs that can
without considering demand and truck capacity, classical change as the timetable depends on the student’s grades.
k-means clustering forms bias clusters and causes a low While the vehicle capacity is limited, re-routing is neces-
truck utilization level. Besides, from results obtained from sary to have a low operation cost while maintaining con-
classical k-means clustering for VRPTW, an interesting straint satisfaction. Therefore, cluster-first – route-second
finding was that the solving time by applying branch and is a method that can generate the solution a practical time
bound in CPLEX not only depend on the size of problems to deal with the abundance of customers’ demand sets.
(number of nodes) but also the data set. In the capacitated However, the limitations of this research have thrown up
k-means clustering algorithm for VRPTW, the model uses many questions in need of further investigation. In our
11 trucks to deliver products and can keep truck utiliza- study, there are only one type of truck was used. Hence,
tion more than 70%. Each subproblem takes an average of the method to adjust the cluster needs to be investigated.
0.36 s for generating a solution. Besides, the model could be extended when the demand
Rapidly finding optimal solutions plays an essential of a customer is more than truck capacity. Hence, the re-
role in making the right and correct decisions related to search could adjust the network. Besides, the rising envi-
logistics problems. Based on the results, managers can use ronmental concern has opened an opportunity to develop
and coordinate resources appropriately. The solution pro- a mathematical model to consider multiple objectives be-
vides the necessary number of vehicles and vehicles trips. sides the total costs. In many practical situations such as
Thus, managers can plan and distribute them more effi- newspaper delivery, non-emergency parcel delivery, furni-
ciently. Simultaneously, the manager can understand how ture delivery, the time window constraints can be relaxed
many vehicles need to be used to serve how many cus- with a customer unsatisfied penalty. That could be another
tomers and who and where the customers are. The time research direction.
for finding a solution is extremely short, which can sup-
port managers in planning resources to deliver products Author contributions
to customers quickly and exactly. This can help companies
improve customer service and increase their competitive For research papers with several authors, a short para-
advantage in today’s dynamic market. On the other hand, graph specifying their individual contributions must be
the total costs are also fully calculated. Managers can use provided.
them to compare with other possible alternatives or to Thi Diem Chau Le and Judit Oláh conceived and de-
compare current financial resources to make more suit- signed the experiments.
able decisions. Thi Diem Chau Le and Duc Duy Nguyen collected and
The findings from this study make several contribu- analysed data.
tions to the current logistics and supply chain manage- Thi Diem Chau Le performed the experiments.
ment: Judit Oláh contributed reagents/materials/analysis
»» VRPTW was applied by a real case study in Viet- tools.
nam, therefore, the VRPTW model considers the Miklós Pakurár contributes methodology.
constraints of logistics problems in reality. These re- Thi Diem Chau Le and Duc Duy Nguyen wrote original
sults contribute in several ways to our understand- draft.
ing of VRPTW problems and provide a basis for Miklós Pakurár and Judit Oláh adjusted and completed
cluster-first – route-second approach. the paper.
 Transport, 2022, 37(1): 17–27 27

Funding Qi, C; Hu, L. 2020. Optimization of vehicle routing problem

for emergency cold chain logistics based on minimum loss,
This paper is supported by EFOP-3.6.3-VEKOP- 16-2017- Physical Communication 40: 101085.
00007 – “Young Researchers for Talent” – supporting ca- https://doi.org/10.1016/j.phycom.2020.101085
reers in research activities in higher education program. Ruiz, E.; Soto-Mendoza, V.; Ruiz Barbosa, A. E.; Reyes, R. 2019.
Solving the open vehicle routing problem with capacity and
distance constraints with a biased random key genetic al-
References gorithm, Computers & Industrial Engineering 133: 207–219.
Afifi, S.; Dang, D.-C.; Moukrim, A. 2016. Heuristic solutions for https://doi.org/10.1016/j.cie.2019.05.002
the vehicle routing problem with time windows and synchro- Rushton, A.; Croucher, P.; Baker, P. 2022. The Handbook of Logis-
nized visits, Optimization Letters 10(3): 511–525. tics and Distribution Management: Understanding the Supply
Chain. 7th Edition. Kogan Page. 824 p.
https://doi.org/10.1007/s11590-015-0878-3
Spliet, R.; Desaulniers, G. 2015. The discrete time window as-
Aggarwal, D.; Kumar, V. 2019. Mixed integer programming for
signment vehicle routing problem, European Journal of Op-
vehicle routing problem with time windows, International
erational Research 244(2): 379–391.
Journal of Intelligent Systems Technologies and Applications
https://doi.org/10.1016/j.ejor.2015.01.020
18(1–2): 4–19. https://doi.org/10.1504/IJISTA.2019.097744
Tasar, B.; Türsel Eliiyi, D.; Kandiller, L. 2019. Vehicle routing with
Bachu, A. K.; Reddy, K. K.; Vanajakshi, L. 2021. Bus travel time
compartments under product incompatibility constraints,
prediction using support vector machines for high variance
Promet – Traffic & Transportation 31(1): 25–36.
conditions, Transport 36(3): 221–234.
https://doi.org/10.7307/ptt.v31i1.2670
https://doi.org/10.3846/transport.2021.15220
Toth, P.; Vigo, D. 2014. Vehicle Routing: Problems, Methods, and
Birim, Ş. 2016. Vehicle routing problem with cross docking: a
Applications. 2nd Edition. SIAM – Society for Industrial and
simulated annealing approach, Procedia – Social and Behav- Applied Mathematics. 481 p.
ioral Sciences 235: 149–158. https://doi.org/10.1137/1.9781611973594
https://doi.org/10.1016/j.sbspro.2016.11.010 Tretjakovas, J.; Čereška, A. 2021. The truck trailer suspension ax-
Cao, W.; Yang, W. 2017. A survey of vehicle routing problem, les failure analysis and modelling, Transport 36(3): 213–220.
MATEC Web of Conferences 100: 01006. https://doi.org/10.3846/transport.2021.14964
https://doi.org/10.1051/matecconf/201710001006 VILAS. 2021. Thực trạng chi phí Logistics Việt Nam năm 2020.
Comert, S. E.; Yazgan, H. R.; Kır, S.; Yener, F. 2018. A cluster Vietnam Logistics and Aviation School (VILAS). Available
first-route second approach for a capacitated vehicle routing from Internet: https://vilas.edu.vn/chi-phi-logistics-tai-viet-
problem: a case study, International Journal of Procurement nam-2020.html (in Vietnamese).
Management 11(4): 399–419. Wang, Y.; Wang, L.; Chen, G.; Cai, Z.; Zhou, Y.; Xing, L. 2020.
https://doi.org/10.1504/IJPM.2018.092766 An improved ant colony optimization algorithm to the peri-
Ekşioğlu, S. D.; Acharya, A.; Leightley, L. E.; Arora, S. 2009. Ana- odic vehicle routing problem with time window and service
lyzing the design and management of biomass-to-biorefinery choice, Swarm and Evolutionary Computation 55: 100675.
supply chain, Computers & Industrial Engineering 57(4): https://doi.org/10.1016/j.swevo.2020.100675
1342–1352. https://doi.org/10.1016/j.cie.2009.07.003 Xu, X.; Yuan, H.; Liptrott, M.; Trovati, M. 2018. Two phase heu-
El-Sherbeny, N. A. 2010. Vehicle routing with time windows: an ristic algorithm for the multiple-travelling salesman problem,
overview of exact, heuristic and metaheuristic methods, Jour- Soft Computing 22(19): 6567–6581.
nal of King Saud University – Science 22(3): 123–131. https://doi.org/10.1007/s00500-017-2705-5
https://doi.org/10.1016/j.jksus.2010.03.002 Yao, Y.; Zhu, X.; Dong, H.; Wu, S.; Wu, H.; Tong, L. C.; Zhou, X.
Kantawong, K.; Pravesjit, S. 2020. An enhanced ABC algorithm 2019. ADMM-based problem decomposition scheme for
to solve the vehicle routing problem with time windows, ECTI vehicle routing problem with time windows, Transportation
Transactions on Computer and Information Technology 14(1): Research Part B: Methodological 129: 156–174.
46–52. https://doi.org/10.37936/ecti-cit.2020141.200016 https://doi.org/10.1016/j.trb.2019.09.009
Khan, S. S.; Ahmad, A. 2004. Cluster center initialization al- Zhang, D.; Cai, S.; Ye, F.; Si, Y.-W.; Nguyen, T. T. 2017. A hybrid
gorithm for k-means clustering, Pattern Recognition Letters algorithm for a vehicle routing problem with realistic con-
25(11): 1293–1302. straints, Information Sciences 394–395: 167–182.
https://doi.org/10.1016/j.patrec.2004.04.007 https://doi.org/10.1016/j.ins.2017.02.028
Le, T. D. C.; Nguyen, D. D.; Oláh, J.; Pakurár, M. 2020. Optimal Zhang, H.; Zhang, Q.; Ma, L.; Zhang, Z.; Liu, Y. 2019. A hybrid
vehicle route schedules in picking up and delivering cargo ant colony optimization algorithm for a multi-objective vehi-
containers considering time windows in logistics distribution cle routing problem with flexible time windows, Information
networks: a case study, Production Engineering Archives 26(4): Sciences 490: 166–190.
174–184. https://doi.org/10.30657/pea.2020.26.31 https://doi.org/10.1016/j.ins.2019.03.070
Londoño, J. C.; Tordecilla, R. D.; Martins, L. C.; Juan, A. A. Zhu, D.; Mortazavi, S. M.; Maleki, A.; Aslani, A.; Yousefi, H.
2021. A biased-randomized iterated local search for the ve- 2020. Analysis of the robustness of energy supply in Japan:
hicle routing problem with optional backhauls, TOP 29(2): Role of renewable energy, Energy Reports 6: 378–391.
387–416. https://doi.org/10.1007/s11750-020-00558-x https://doi.org/10.1016/j.egyr.2020.01.011
Pérez-Rodríguez, R.; Hernández-Aguirre, A. 2019. A hybrid esti- Zhu, L; Hu, D. 2019. Study on the vehicle routing problem con-
mation of distribution algorithm for the vehicle routing prob- sidering congestion and emission factors, International Jour-
lem with time windows, Computers & Industrial Engineering nal of Production Research 57(19): 6115–6129.
130: 75–96. https://doi.org/10.1016/j.cie.2019.02.017 https://doi.org/10.1080/00207543.2018.1533260

View publication stats