mathematics

Article
Digital-Twin-Driven AGV Scheduling and Routing in
Automated Container Terminals
Ping Lou , Yutong Zhong, Jiwei Hu , Chuannian Fan and Xiao Chen *

School of Information Engineering, Wuhan University of Technology, Wuhan 430070, China

* Correspondence: [email protected]

Abstract: Automated guided vehicle (AGV) scheduling and routing are critical factors affecting the
operation efficiency and transportation cost of the automated container terminal (ACT). Searching
for the optimal AGV scheduling and routing plan are effective and efficient ways to improve its
efficiency and reduce its cost. However, uncertainties in the physical environment of ACT can make
it challenging to determine the optimal scheduling and routing plan. This paper presents the digital-
twin-driven AGV scheduling and routing framework, aiming to deal with uncertainties in ACT. By
introducing the digital twin, uncertain factors can be detected and handled through the interaction
and fusion of physical and virtual spaces. The improved artificial fish swarm algorithm Dijkstra
(IAFSA-Dijkstra) is proposed for the optimal AGV scheduling and routing solution, which will be
verified in the virtual space and further fed back to the real world to guide actual AGV transport.
Then, a twin-data-driven conflict prediction method is proposed to predict potential conflicts by
constantly comparing the differences between physical and virtual ACT. Further, a conflict resolution
method based on the Yen algorithm is explored to resolve predicted conflicts and drive the evolution
of the scheme. Case study examples show that the proposed method can effectively improve efficiency
and reduce the cost of AGV scheduling and routing in ACT.

Keywords: digital-twin-driven; AGV scheduling and routing; conflict prediction; conflict resolution;
IAFSA-Dijkstra

Citation: Lou, P.; Zhong, Y.; Hu, J.; MSC: 68Txx

Fan, C.; Chen, X. Digital-Twin-Driven
AGV Scheduling and Routing in
Automated Container Terminals.
Mathematics 2023, 11, 2678. https:// 1. Introduction
doi.org/10.3390/math11122678
In recent years, economic globalization has led to an increase in international trade,
Academic Editors: Fan Zhang, and maritime transport has become an essential means of transportation. Ports serve as
Songhe Feng, Yongsheng Zhou and a crucial hub for both sea and land transportation and play an integral role in facilitating
Junlin Hu international trade. However, with the growing demand for maritime transportation, ports
Received: 28 April 2023
are under immense pressure to increase efficiency and reduce costs. To address these
Revised: 26 May 2023
challenges, the development of automated container terminals (ACTs) has emerged as a
Accepted: 29 May 2023 new trend. ACTs leverage advanced technologies, such as robotics, artificial intelligence,
Published: 13 June 2023 and the Internet of Things, to automate container handling processes. This technology
can significantly increase port productivity, reduce labor costs, and improve operational
efficiency. Automatic guided vehicles (AGVs) are significant types of equipment in an ACT,
which can walk along a pre-set guidance path to complete a series of horizontal transport
Copyright: © 2023 by the authors. operations [1]. AGV scheduling and routing are crucial problems in the ACT. The purpose
Licensee MDPI, Basel, Switzerland. of AGV scheduling is to allocate a batch of container transportation tasks to a set of AGVs
This article is an open access article and specify the sequence and time of task execution to achieve specific goals under given
distributed under the terms and
constraints. AGV routing is to determine an optimal path from the start position to the
conditions of the Creative Commons
end position for the AGV, complete AGV’s assigned transport tasks in order, and avoid
Attribution (CC BY) license (https://
collision with other AGVs. Reasonable scheduling and routing for AGVs can not only
creativecommons.org/licenses/by/
improve the efficiency of ACT operation but also reduce the traffic accident rate in the ACT
4.0/).

Mathematics 2023, 11, 2678. https://doi.org/10.3390/math11122678 https://www.mdpi.com/journal/mathematics

Mathematics 2023, 11, 2678 2 of 25

and improve the reliability of transportation [2]. Therefore, AGV scheduling and routing
are considered comprehensively in this work.
Research on AGV scheduling and routing in an ACT starts with static problems,
exploring optimal or near-optimal solutions under deterministic ACT environments. There
are many studies about AGV static problems. Luo and Wu [3] develop an integrated
modeling approach for AGV and YC static scheduling problems in ACTs that minimizes
the ship’s berth time; they design a genetic algorithm to solve the problem. Xu et al. [4]
propose a reinforcement learning-based hyper-heuristic genetic algorithm to solve the
integrated static scheduling problem of AGVs and other facilities in the U-shaped ACT;
their method can avoid AGV conflicts. Lu et al. [5] propose an ant colony system-improved
grey wolf optimization to solve the fourth-party logistics routing problem. Yan et al. [6]
present a hybrid metaheuristic algorithm of discrete particle swarm optimization and Harris
Hawks optimization to solve the location problem and vehicle routing problem. Lu et al. [7]
design a bi-level whale optimization algorithm to solve a bi-level multi-objective schedule
risk management model. However, a real-world ACT environment is full of uncertain
events, such as AGV failure, ship arrival time delay, and weather change; thus, the static
problem is impractical. In static problems, the mathematical model establishment usually
relies on predefined constraints (e.g., constant AGV speed). However, the occurrence
of uncertain events can destroy these constraints, which causes the scheme to become
unfeasible and even induce AGV conflicts.
In order to respond to uncertain events, dynamic problems are developed. Further-
more, an increasing number of researchers are beginning to study dynamic problems. For
example, in order to respond to the uncertainty caused by new job arrival, Cai et al. [8]
propose a rescheduling new arrival jobs (RNJ) policy and rescheduling combination of
new and unexecuted jobs (RCJ) policy for container-transportation task allocation in the
ACTs. Jian et al. [9] develop a multi-objective scheduling model that uses a symmetric
triangular fuzzy number to describe the AGVs’ operation time distribution; they build an
improved genetic algorithm to solve the problem effectively. Yue and Fan [10] introduce a
dynamic scheduling process to respond to the QC waiting caused by AGVs’ delay in an
uncertain environment.
Uncertain events can trigger a series of chain reactions that can disrupt the entire ACT
operation, causing deterioration of ACT efficiency [11]. Therefore, a timely response to
dynamic events during the plan execution becomes a vital issue that needs to be addressed
urgently. The emergence of the digital twin (DT) provides a new idea to meet the above
challenges. DT is an integrated system that can simulate, monitor, calculate, regulate, and
control the system status and process [12], with the characteristics of real-time reflection,
interaction and convergence, and evolution and iteration.
This work proposes a digital-twin-driven AGV scheduling and routing framework that
copes with uncertainties in automated container terminals. This paper’s main contributions
are as follows:
(1) A digital-twin-driven AGV scheduling and routing framework is proposed. Based on
this framework, an initial scheduling and routing plan is generated first. The AGV
transport process is continuously monitored, and dynamic events are fed back into
the physical space, enabling timely responses to changes in the environment.
(2) The improved artificial fish swarm algorithm Dijkstra (IAFSA-Dijkstra) is proposed to
solve the bi-level mixed integer programming model and obtain an optimal solution.
The task combination encoding is presented to reduce the encoding length of the
IAFSA-Dijkstra, and the adaptive parameter adjustment operator is used to improve
the global optimization capability of the IAFSA-Dijkstra. The optimal plan is verified
and fed into the physical space to guide AGV transport.
(3) Twin-data-driven conflict prediction and resolution: A twin-data-driven AGV conflict
prediction method is explored to predict conflicts by comparing physical and virtual
data. A conflict resolution method based on the Yen algorithm is presented to resolve
these conflicts and drive timely revision and evolution of the initial plan.
Mathematics 2023, 11, 2678 3 of 25

The remainder of this paper is organized as follows. Section 2 reviews the relevant
literature. Section 3 describes the AGV scheduling and routing problem and establishes a
bi-level mixed integer programming model. The DT-based AGV scheduling and routing
framework is discussed in Section 4. Section 5 gives examples to verify the effectiveness of
the proposed method. Finally, Section 6 concludes the paper and points out future works.

2. Related Works
2.1. AGV Scheduling
There have been many studies on the AGV scheduling problem in recent years. A
brief review of previous works is provided on AGV scheduling.
Existing studies on AGV static scheduling in the ACT have provided very valuable
information. Rashidi and Tsang [13] propose a novel algorithm, NSA+, to solve the AGV
scheduling problem considering minimizing const flow in container terminals. Ma et al. [14]
used an improved mathematical model to describe the multi-load AGV scheduling problem
in the ACT; they built a shuffled frog leaping algorithm with a mutant process (SFLAMUT)
to solve the problem. Given the high potential synergy between automated quayside cranes
(AQCs) and AGVs, Zhao et al. [15] built a collaborative model for AQCs and AGVs; they
used a two-stage taboo search algorithm to solve the scheduling problem.
AGV static scheduling usually is infeasible and impractical due to uncertainties; thus,
AGV dynamic scheduling has become a growing concern. Angeloudis and Bell [16] propose
a novel algorithm structured around a cost/benefit concept to solve the AGV dis-patching
problem under an indeterministic ACT environment, mainly including uncertain task
durations for future events. Xin et al. [17] propose a rescheduling method to rearrange
AGVs, quay cranes (QCs), and automated stacking cranes (ASCs) in ACT, including two
types of methods: the time-efficient schedule and the energy-efficient schedule. Sahin
et al. [18] develop a multi-agent-based system to simultaneously schedule flexible machine
groups and AGVs under an uncertain manufacturing environment. Xu et al. [19] present
a response method with AGVs based on the mode of “request-scheduling-response” to
solve the logistics dynamic scheduling problem in an intelligent manufacturing workshop
(IMW), which aims to minimize the finish time with the minimum AGVs and limited time.
In order to deal with the AGV scheduling problem in complex material handling in smart
factories, Zhang et al. [20] develop a dynamic scheduling method for self-organized AGVs
(SAGV), aiming to minimize the delay and reduce the cost of logistics systems.

2.2. AGV Routing

The earliest research on AGV routing starts with single-AGV path planning. For
the single robot path planning problem, Wang et al. [21] improve the traditional rapidly-
exploring random tree (RRT) algorithm and propose an autonomous routing algorithm of
node control (NC-RRT) with higher exploration efficiency and faster convergence speed.
Lee and Jeong [22] use the Q-Learning and Dyna-Q algorithms to solve the mobile robot
path planning problem in a warehouse environment.
In the actual environment, the situation of using one AGV to complete the task is
extremely rare; multiple AGVs complete tasks in cooperation with each other. There have
been many previous studies on multi-AGV path planning in a static environment. Oboth
et al. [23] present a route-generation technique that can realize conflict-free path planning
for multiple AGVs with varying speeds. Yuan et al. [24] propose a bi-level path planning al-
gorithm; global and local paths are obtained using the A* and RRT algorithms, respectively.
The above research studies focus on routing problems in a static environment. In order
to respond to uncertainties, dynamic routing is developed. Xu et al. [25] present an im-
proved dynamic window method to deal with the routing problem in a three-dimensional
dynamic environment, outperforming traditional DWA algorithms in terms of efficiency,
smoothness, and security. To deal with complex and variable situations during the multi-
robot operation process, Bae et al. [26] combine deep Q-learning with the CNN (convolution
neural network) algorithm, which is more flexible and efficient than conventional methods.
Mathematics 2023, 11, 2678 4 of 25

In order to respond to dynamic obstacles and incomplete maps, Bai et al. [27] combine an
improved Q-learning path optimization algorithm and an improved genetic algorithm; the
proposed method has favorable performance in narrow working environments and highly
congested situations. Onoufriou et al. [28] propose a new hybrid parallelism deep learning
framework, which can provide a solution and unification of deep learning techniques using
a common interface.
Some researchers have begun to study the AGV scheduling and routing problems
simultaneously. Zhong et al. [29] tried to minimize ships’ loading and unloading time via
integrated scheduling of AGVs and other facilities with the hybrid GA-PSO algorithm
with adaptive autotuning approaches by a fuzzy control solution. Fazlollahtabar et al. [30]
studied the simultaneous scheduling and routing problem for AGVs in flexible manu-
facturing systems, obtaining the optimal plan by a modified network simplex algorithm
(NSA). Miyamoto and Inoue [31] propose local/random search methods to deal with the
dispatching and routing problem for capacitated AGV systems (DCFRPC). Desaulniers
et al. [32] present an exact method for AGV dispatching and conflict-free routing in the flex-
ible manufacturing system, implicitly considering congestion and blocking problems. Xing
et al. [33] propose a novel tabu search algorithm to solve the conflicts that happen when
multiple AGVs work at the same time. Liang et al. [34] present a three-stage integrated
scheduling algorithm for AGV routing planning to deal with the locking problem of AGV.

2.3. Digital Twin

The above works have made an outstanding contribution to the AGV scheduling and
routing problem but fail to address the challenges confronting AGV scheduling and routing
mentioned in Section 1. Recently, DT has been widely applied in various fields, including
predictive maintenance [35], factory design [36], assembly [37], prognostics [38], health
care [39], traffic management [40], and so on.
Some researchers attempt to combine DT with other methods to address scheduling
problems. For example, by introducing DT, Zhang et al. [41] address three critical problems
in dynamic job-shop scheduling: machine availability prediction, disturbance detection,
and performance evaluation. Wang and Wu [42] develop a planning and scheduling system
by combining DT with planning and scheduling to effectively manage and control the
disturbances in workshop scheduling. Negri et al. [43] present a DT-based scheduling
framework, which includes equipment health predictions in the scheduling activity; their
framework embeds a field-synchronized Equipment Health Indicator module into the DT
simulation to synchronize the simulation model to the physical system. In AGV scheduling,
Han et al. [44] developed a DT-based dynamic AGV scheduling (DTDAS) method, which
can reasonably arrange the charging process to effectively improve AGV transportation
efficiency in a workshop.
Several researchers have started to apply DT to routing problems. For example, Wang
et al. [45] built a digital twin platform for gantry robots and present a multi-objective
three-dimensional routing method for gantry robots based on the NavMesh algorithm
based on the platform, improving production efficiency and safety. Zohdi [46] designed
a digital twin framework for firefighting to help aerial first responders rapidly perform
flight routing in risky fire environments. Guo et al. [47] constructed a digital twin system
for product assembly and present a modified Q-learning algorithm to solve the routing
problem based on the system. To address the shortcoming of supervised learning of
requiring large-scale training data, Vasanthan and Nguyen [48] introduced a digital twin of
the vessel to generate enough training data to solve the vessel’s routing problem, effectively
detecting potential collisions. Gao et al. [49] developed a digital-twin-enabled automated
storage yard scheduling framework for uncertain port dispatching.
DT brings a chance to deal with the AGV scheduling and routing problem. By intro-
ducing DT technology, the real-time data generated by the ACT operation process can be
fused with data from the virtual space to drive the dynamic evolution of the initial plan
 requiring large-scale training data, Vasanthan and Nguyen [48] introduced a digital twin
of the vessel to generate enough training data to solve the vessel’s routing problem,
effectively detecting potential collisions. Gao et al. [49] developed a digital-twin-enabled
automated storage yard scheduling framework for uncertain port dispatching.
Mathematics 2023, 11, 2678 DT brings a chance to deal with the AGV scheduling and routing problem. 5 of By
25
introducing DT technology, the real-time data generated by the ACT operation process
can be fused with data from the virtual space to drive the dynamic evolution of the initial
planand
first, first,then
andthe
then the novel
novel plan isplan is simulated,
simulated, verified,
verified, and, further,
and, further, timely timely transmitted
transmitted to the
to the physical space to control the AGV
physical space to control the AGV operation. operation.

3. Problem
3. Problem Description
Description and and Formulation
Formulation
3.1.
3.1. AGV
AGV Scheduling
Scheduling andand Routing
Routing Problem
Problem
The
The devices
devicesof ofthe
theACT
ACTin inAGV
AGVscheduling
schedulingand androuting
routinginvolve
involveAGVs,
AGVs, quay
quaycranes
cranes
(QCs),
(QCs),and
andyard
yardcranes
cranes(YCs)
(YCs)(see
(seeFigure
Figure1).
1).AGVs
AGVs are
are used
used for
for moving
moving containers
containersbetween
between
QC
QC and
andstorage
storageyards.
yards. QCs
QCs are
are to
todeliver
delivercontainers
containers ononAGVs
AGVs from
from aa ship
shipor
orcontainers
containers
from AGVs to a ship. The storage yard is an area for the temporary storage
from AGVs to a ship. The storage yard is an area for the temporary storage of containersof containers
for
fortransferring
transferringto to
destinations. YCsYCs
destinations. are used to deliver
are used containers
to deliver to their storage
containers to theirlocations
storage
in storage yards or from their storage locations in storage yards.
locations in storage yards or from their storage locations in storage yards.

Ship

Container

Quay Crane
Yard Crane

AGV

AGV Transportation Swap Storage

Area Area Yard
Figure1.1. The
Figure Thelayout
layoutof
ofan
anautomated
automatedcontainer
containerterminal.
terminal.

TheAGV
The AGV transferring
transferring areaarea has
hasaaregular
regularroute.route. In
Inorder
ordertotoconveniently
convenientlydescribe
describethe the
AGVscheduling
AGV scheduling andand routing
routing problem
problem of an of ACT,an the
ACT, the transferring
transferring area isdescribed
area is formally formally
as a weighted
described as adirected
weightedgraph, in which
directed graph,there are 10
in which nodes
there are (see Figure
10 nodes 2a).Figure
(see The node n3
2a). The
indicates the
node n3 indicates QC, the node n8 is the storage yard, the set {n1, n2, n4, n5, n6,
QC, the node n8 is the storage yard, the set {n1, n2, n4, n5, n6, n7, n9, n7, n9, n10}
isn10}
theispath
the node set, and
path node set,the
andset
the setl2,{l1,
{l1, , l24}
. . .l2, is the
…, l24} is path set, set,
the path whose lengths
whose are d1,
lengths are d2,
d1,
.d2,
Mathematics 2023, 11, x FOR PEER REVIEW d24,d24,
. . , …, respectively. TheThe
respectively. direction of each
direction arrow
of each expresses
arrow the direction
expresses of AGV
the direction that
of AGV they
6 ofthat
25
can
theytravel on theonpath.
can travel Figure
the path. 2b shows
Figure 2b showsthe corresponding
the correspondingadjacency matrix,
adjacency in which
matrix, the
in which
symbol
the symbol ∞ means
∞ means
therethere
are noare
ways between
no ways betweentwo nodes.
two nodes.

(a) (b)
Figure
Figure2.
2.The
Theweighted
weighteddirected
directedgraph
graphand
andadjacent
adjacentmatrix.
matrix.(a)
(a)AAweighted
weighteddirected graph;
directed (b)(b)
graph; anan
adjacent matrix.
adjacent matrix.

The AGV scheduling and routing problem is satisfied under the following
assumptions:
• The location of all containers’ delivery points and pickup points are fixed and known.
This work does not distinguish whether a task is a loading or an unloading task.
• All AGVs are in good condition without failure.
Mathematics 2023, 11, 2678 6 of 25

The AGV scheduling and routing problem is satisfied under the following assumptions:
• The location of all containers’ delivery points and pickup points are fixed and known.
This work does not distinguish whether a task is a loading or an unloading task.
• All AGVs are in good condition without failure.
• Each AGV can only transport one container at a time, and each container can only be
assigned to one AGV for transporting.
• The operation time for both QCs and YCs is fixed.
• All AGVs are homogeneous and travel at the same velocity, and AGV speed remains
unchanged during turning.
• All containers have the same priority.
• The initial position of each AGV is known.
Notations used in the model of the studied problem are listed in Appendix A.

3.2. The bi-Level Mixed Integer Programming Model

In order to describe the AGV scheduling and routing problem, the bi-level mixed
integer programming model is established. The upper-level scheduling model focuses
on AGV scheduling and the lower-level routing model deals with AGV routing. The
upper-level scheduling model determines the overall horizontal transportation operation
optimization results; the lower-level route model determines the actual completion time of
each operation task for each AGV.

3.2.1. The Upper-Level Scheduling Model

The objective of AGV scheduling is to minimize the completion time of all tasks. The
AGV scheduling optimization model can be formulated as follows:

minCmax = minmax (Ck ) (1)

k ∈V

Subject to:
Ck = tkf , ∀k ∈ V (2)

∑ αkp = 1, ∀p ∈ C (3)
k ∈V

∑ βk0p = 1, ∀k ∈ V (4)
p∈C

∑ βkp f = 1, ∀k ∈ V (5)
p∈C

∑ βkpp0 = ∑ βkp0 p , ∀ p ∈ C, ∀k ∈ V (6)

p0 ∈C ∪{0, f } p0 ∈C ∪{0, f }

 
TOk k L p ≤ tkp + M 1 − βk0p , ∀ p ∈ C, ∀k ∈ V (7)

 
p p
tkp + τL + TLk p U p + τU + TUk p L 0 ≤ tkp0 + M 1 − βkpp0 , ∀ p ∈ C, p0 ∈ C ∪ { f }, ∀k ∈ V (8)
p

αkp , βkpp0 ∈ (0, 1), ∀ p, p0 ∈ C, ∀k ∈ V (9)

p p
tkp ≥ 0, τL > 0, τU > 0, TL p U p > 0, TU p L p0 ≥ 0, TOk L p ≥ 0, ∀ p, p0 ∈ C, ∀k ∈ V (10)

Equation (1) is the optimization goal to minimize the completion time of all tasks.
Equation (2) indicates that the task completion time for each AGV is the completion
time of its last task, which is the start time of its virtual last task.
Mathematics 2023, 11, 2678 7 of 25

Equation (3) indicates that each container is transported by one and only one AGV.
Equation (4) and equation (5) ensure that the initial task of each AGV is virtual task 0
and the ending task is virtual task f.
Equation (6) indicates the order in which each AGV performs its tasks. For each AGV,
there is only one container transport task before and after each container transport task.
Equations (7) and (8) reflect the time correlation between the two adjacent tasks
performed by the same AGV. Equation (7) indicates that the start moment of the first
task for the AGV is the time from its start position to the pickup point of its first task.
Equation (8) indicates that the start moment of one task for the AGV is the time from its
start position to the pickup point of its first task. The start moment of the current task is
equal to the sum of the start moment of the previous task, the loading time of the previous
task, the transport time of the previous task, the unloading time of the previous task, and
the time from the delivery point of the previous task to the pickup point of the current task.
Equation (9) is binary constraints for some decision variables.
Equation (10) is the non-negativity constraints for some variables.

3.2.2. The Lower-Level Routing Model

The objective of AGV routing is to minimize the delivery time from the start node to
the end one. The AGV routing optimization model can be formulated as follows:
 
min ∑ µijk × tijk , ∀i, j ∈ N, ∀k ∈ V (11)
(i,j)∈ G

Subject to:
tijk = tkout,ij − tin,ij
k
, ∀(i, j) ∈ G, ∀k ∈ V (12)

tkout,ij ≥ tin,ij
k
+ dij /v0 , ∀(i, j) ∈ G, ∀k ∈ V (13)


 1, i = s
∑ µkji − ∑ µijk 0 = −1, i = e , ∀k ∈ V (14)
( j,i )∈ G (i,j0 )∈ G 0, o.w.


0
k
tin,si k
≥ tin,sj + Ds /v0 , ∀s ∈ NP , ∀k, k0 ∈ V, ∀(s, i ), (s, j) ∈ G (15)

µijk ∈ (0, 1), ∀(i, j) ∈ G, ∀k ∈ V (16)

tijk > 0, tkout,ij ≥ 0, tin,ij

k
≥ 0, ∀(i, j) ∈ G, ∀k ∈ V (17)
Equation (11) is the optimization objective to minimize the delivery time from the start
node to the destination one.
Equation (12) indicates that the travel time of the AGV on one path is the time differ-
ence between the moment of leaving that path and the moment of entering that path.
Equation (13) reflects the time correlation of one AGV entering and leaving one path.
Equation (14) places restrictions on the out-in degree of nodes in the current transport
process. For the start node s of the current transport process, the out-degree is greater than the
in-degree by 1. For the end node e of the current transport process, the in-degree is greater
than the out-degree by 1. For other nodes, the out-degree and the in-degree are equal.
Equation (15) is related to AGV conflict. Equation (15) indicates that a safe distance
should be maintained when AGVs visit the same path node.
Equation (16) is binary constraints for the decision variable.
Equation (17) is the non-negativity constraints for some variables.
 Equation (15) is related to AGV conflict. Equation (15) indicates that a safe distance
should be maintained when AGVs visit the same path node.
Equation (16) is binary constraints for the decision variable.
Equation (17) is the non-negativity constraints for some variables.
Mathematics 2023, 11, 2678 8 of 25

4. Digital-Twin-Driven Scheduling and Routing

4.1. Overall Framework
4. Digital-Twin-Driven Scheduling and Routing
4.1. Overall Framework
The framework of DT-based AGV scheduling and routing for an ACT consists of two
The framework of DT-based AGV scheduling and routing for an ACT consists of two
parts, involving the physical space and virtual space (see Figure 3).
parts, involving the physical space and virtual space (see Figure 3).

Figure 3. The framework offramework

Figure 3. The the digital-twin-driven AGV scheduling
of the digital-twin-driven and and
AGV scheduling routing.
routing.

In the physical space, the physical entities mainly involve AGVs, QCs, and YCs. The
In the physical space, the physical entities mainly involve AGVs, QCs, and YCs. The
data sensing the conditions of various devices and its surrounding are collected and also
data sensing the conditions
transmitted ofvirtual
to the various devices
space. and itsphysical
In addition, surrounding
devices are
will collected
execute in and also
real time
transmitted toaccording
the virtual
to thespace. Infeedback
decision addition, fromphysical
the virtualdevices
space. will execute in real time
according to the decision feedback from the virtual space. the service modules, algorithms
The virtual space consists of four parts, involving
library, evaluation verification platform, and data center. The functions of the data center
are mainly to store data from the physical and virtual space and to process this data to
provide data support for service modules. With the assistance of the algorithm library,
the function of the service modules primarily consists of (1) generating an initial AGV
scheduling and routing plan to guide the AGV transport process; (2) monitoring the ACT
operation process and predicting AGV conflicts by comparing simulation data with real-
time data; and (3) resolving conflicts caused by uncertain events, revising the initial plan,
and feeding it back into the physical space. The purpose of the evaluation verification
platform is to verify the feasibility and efficiency of the plan by simulation.
The workflow of the proposed framework is shown in Figure 4. The specific processes
are as follows.
 operation process and predicting AGV conflicts by comparing simulation data with real-
time data; and (3) resolving conflicts caused by uncertain events, revising the initial plan,
and feeding it back into the physical space. The purpose of the evaluation verification
platform is to verify the feasibility and efficiency of the plan by simulation.
Mathematics 2023, 11,The
2678 workflow of the proposed framework is shown in Figure 4. The specific 9 of 25

processes are as follows.

Figure 4. The workflow

Figure 4.ofThe
digital-twin-driven AGV scheduling
workflow of digital-twin-driven and routing.
AGV scheduling and routing.

Step 1: By introducing the digital twin, the AGV transport process is continuously
Step 1: By introducing the digital twin, the AGV transport process is continuously
monitored, including AGV speed, AGV acceleration, and AGV position. The real-time data
monitored, including AGV speed,
are transmitted AGVspace.
to the virtual acceleration, and AGV position. The real-time
data are transmittedStep
to the virtual space.is proposed for the optimal AGV scheduling and routing
2: IAFSA-Dijkstra
Step 2: IAFSA-Dijkstra
solution, uncertainis factors
proposedcan befor the optimal
detected and handledAGV scheduling
through and and
the interaction routing
fusion
of physical and virtual spaces, and a conflict resolution method based
solution, uncertain factors can be detected and handled through the interaction and fusion on the Yen algorithm
is explored to resolve predicted conflicts and drive the evolution of the scheme.
of physical and virtual spaces, and a conflict resolution method based on the Yen
Step 3: The evaluation verification platform evaluates the updated plan. The results of
algorithm is explored to resolve
scheduling and pathpredicted
planning areconflicts and drive
fed into physical thetoevolution
space guide AGVof forthe
pathscheme.
planning,
Step 3: Thewhich
evaluation verification
timely responds platform
to changes in the evaluates
environment. the updated plan. The results
of scheduling and path planning are fed into physical space to guide AGV for path
4.2. IAFSA-Dijkstra Algorithm
planning, which timely responds to changes in the environment.
A bi-level mixed integer programming model is established in Section 3. The upper
level of the model pertains to AGV scheduling, while the lower level of the model concerns
4.2. IAFSA-Dijkstra Algorithm
collision-free AGV routing. To solve the model, an improved artificial fish swarm algorithm-
Dijkstra algorithm (IAFSA-Dijkstra), is presented in this section. The IAFSA is used to
optimize AGV scheduling, which determines the start and endpoints of the path. The
Dijkstra algorithm is used for collision-free routing, which affects the AGV scheduling
process. The flowchart of IAFSA-Dijkstra is shown in Figure 5.
 collision-free AGV routing. To solve the model, an improved artificial fish swar
algorithm-Dijkstra algorithm (IAFSA-Dijkstra), is presented in this section. The IAFSA
used to optimize AGV scheduling, which determines the start and endpoints of the pa
Mathematics 2023, 11, 2678 The Dijkstra algorithm is used for collision-free routing, which affects the AGV scheduli
10 of 25
process. The flowchart of IAFSA-Dijkstra is shown in Figure 5.

FigureFigure
5. The flowchart
5. The ofofIAFSA-Dijkstra.
flowchart IAFSA-Dijkstra.

AFSA AFSA
hashas better
better global optimization
global optimization capability owing
capability to its parallel
owing search and
to its parallel the and t
search
nature of controlling the search direction [50]. There is a more substantial search capability
nature of controlling the search direction [50]. There is a more substantial search capabil
in the early stage of optimization, but the search capability weakens, and it is easily trapped
in thein early stage
the local of optimization,
extremum but
in the later stage. Thus,theansearch
adaptivecapability weakens, method
parameter adjustment and it is eas
trapped in the
is used local extremum
to improve in the later
AFSA. Meanwhile, stage. Thus,
the encoding methodan inadaptive
this work is parameter adjustme
related to the
number of container transport tasks. Therefore, a task combination
method is used to improve AFSA. Meanwhile, the encoding method in this work approach is adopted to is relat
combine container tasks according to specific rules to reduce the encoding length.
to the number of container transport tasks. Therefore, a task combination approach
Encoding and decoding method: A random number encoding method is adopted.
adopted to combine
The value above each container tasks according
position denotes the containerto no.specific
The value rules to reduce
after rounding the encodi
in each
length.
position denotes the AGV no., and the value range is [0.5, Nagv + 0.5), where Nagv denotes
the number of
Encoding and AGVs. The initial
decoding value in A
method: each positionnumber
random is randomly generated.
encoding method is adopte
Assume 8 containers and 3 AGVs, and the encoding
The value above each position denotes the container no. The value after method is as shown in Figure 6a. in ea
rounding
Taking the first position as an example, 1.181 is rounded to 1. Therefore, container 1 is
position denotes
assigned to AGV the 1.
AGV no.,
In the sameand theasvalue
way for therange is [0.5, 𝑁AGV+ 10.5),
other positions, where 𝑁for deno
is responsible
the number of AGVs.
transporting The1, initial
containers 7, and 8,value
AGV 2in each position
is responsible is randomly
for transporting generated.
containers 4 and 6,
Assume 8 containers and 3 AGVs, and the encoding method is as shown
and AGV 3 is responsible for transporting containers 2, 3, and 5. For containers in Figure
of each
AGV, the order after sorting by their value is the transportation order
Taking the first position as an example, 1.181 is rounded to 1. Therefore, container 1 of AGV. Taking AGV
1 as an example, the values of containers 1, 7, and 8 are 1.181, 1.211, 1.203, respectively.
assigned to AGV 1. In the same way as for the other positions, AGV 1 is responsible f
Since 1.181 < 1.203 < 1.211, the transportation order of AGV 1 is container 1–container
transporting
8–containercontainers
7. 1, 7, and 8, AGV 2 is responsible for transporting containers 4 a
6, and AGV 3 is responsible for transporting containers 2, 3, and 5. For containers of ea
AGV, the order after sorting by their value is the transportation order of AGV. Taki
AGV 1 as an example, the values of containers 1, 7, and 8 are 1.181, 1.211, 1.2
respectively. Since 1.181 < 1.203 < 1.211, the transportation order of AGV 1 is container
container 8–container 7.
 time. The cost of the method is that the number of solution spaces will be reduced.
Task combination is a strategy of combining multiple tasks into one task. If the pickup
point of container p and the delivery point of container p’ are the same point, two tasks
will be combined into one task. The combined task point is the pickup of container p’, the
delivery point of container p’ (viz., the pickup point of container p), and the delivery point
Mathematics 2023, 11, 2678 11 of 25
of container p in sequence. The task combination with 8 containers as an example is shown
in Figure 6b.

1 2 3 4 5 6 7 8
Container 1 Container 2 Container 3 Container 4

1.181 3.486 2.632 1.574 2.679 2.311 1.211 1.203 Pickup Point
Delivery Point
1
3
Pickup Point
Delivery Point
4
6
Pickup Point
Delivery Point
5
2
Pickup Point
Delivery Point
3
8

Container 5 Container 6 Container 7 Container 8

Pickup Point 2 Pickup Point 6 Pickup Point 2 Pickup Point 9
Delivery Point 6 Delivery Point 7 Delivery Point 10 Delivery Point 11

AGV1 AGV3 AGV3 AGV2 AGV3 AGV2 AGV1 AGV1 Task Combination

AGV1: 1, 7, 8 AGV1: 1 − 8 − 7 Containers Task Points

Task 1 1→4 1→3→8
AGV2: 4, 6 Sorting AGV2: 4 − 6 Task 2 2→6 4→6→7

AGV3: 3 − 5 − 2
Task 3 3→5 5→2→6
AGV3: 2, 3, 5 Task 4 7 2→10
Task 5 8 9→11

(a) (b)
Figure
Figure 6.
6. The
The decoding
decoding method
method and
andtask
taskcombination
combinationstrategy.
strategy.(a)
(a)The
Thedecoding
decodingmethod;
method;(b)
(b)Task
Task
Combination strategy.
Combination strategy.

Adaptive parameter
Task combination adjustment
strategy: strategy:
Considering In AFSA,
the above Visualmethod,
encoding and Step are essential
encoding length
parameters
is related to the number of container transportation tasks. The higher the numberaof
that affect the movement of the AF. In the early stage of optimization, larger
con-
Visual and Step can make the AF move quickly and jump out of
tainer transportation tasks, the longer the encoding length, which causes a more extended the local extremum, thus
converging faster. Nevertheless,
algorithm running time. Therefore, in athe later
task stage of optimization,
combination a larger Visual
strategy is introduced and Step
to reduce the
will leadof
number tocontainer
the optimal solution being
transportation skipped,
tasks, which the
shortening is not conduciverunning
algorithm’s to optimization.
time. The
In the
cost of traditional
methodAFSA, is that Visual and Step
the number are fixed,spaces
of solution whichwillmakes it difficult to reconcile the
be reduced.
requirements in the early
Task combination and later
is a strategy stages ofmultiple
of combining optimization.
tasks intoAnone adaptive parameter
task. If the pickup
adjustment method is used to address the above problem. The
point of container p and the delivery point of container p’ are the same point, two tasks value of Visual and Step
gradually
will be combined become
into one smaller
task.asThethecombined
number of iterations
task point isincreases.
the pickup of container p’, the
Visual
delivery andofStep
point can be p’
container calculated
(viz., theaccording
pickup point to Expressions
of container(18) andthe
p), and (19).
delivery point
of container p in sequence. The task combination with 8 containers as an example is shown
𝑉𝑖𝑠𝑢𝑎𝑙(𝑔𝑒𝑛) = 𝑉𝑖𝑠𝑢𝑎𝑙 × 𝑔𝑒𝑛 ( ⁄ )⁄ ( ) (18)
in Figure 6b.
Adaptive parameter adjustment strategy: ( In AFSA, ⁄
Visual
)⁄
and
(
Step
)
are essential
parameters that affect 𝑆𝑡𝑒𝑝(𝑔𝑒𝑛) = 𝑆𝑡𝑒𝑝 of ×
the movement 𝑔𝑒𝑛
the AF. In the early stage of optimization, a larger
(19)
Visual
where and𝑔𝑒𝑛Step can make
denotes the AF move
the number quickly and
of iterations, 𝑉𝑖𝑠𝑢𝑎𝑙
and jump 𝑉𝑖𝑠𝑢𝑎𝑙
out of, the , 𝑆𝑡𝑒𝑝
local extremum, , thus
and
𝑆𝑡𝑒𝑝
converging faster.
denote theNevertheless,
maximum and in the later stage
minimum of ofVisual
optimization,
and Step,a larger Visual and
respectively. 𝑔𝑒𝑛Step
will leadthe
denotes to the optimal number
maximum solution of being skipped,𝑉𝑖𝑠𝑢𝑎𝑙(𝑔𝑒𝑛)
iterations. which is notand 𝑆𝑡𝑒𝑝(𝑔𝑒𝑛)
conducive to optimization.
denote Visual
In traditional
and Step AFSA, Visual
with the number and Step are fixed, which makes it difficult to reconcile
of iterations.
the requirements
The procedure inofthe early is
IAFSA and laterin
shown stages of optimization.
Algorithm An adaptive
1. First, combine parameter
tasks according to
adjustment method is used to address the above problem. The value
task combination strategy and initialize the artificial fish (AF) population. Then, calculate of Visual and Step will
gradually
𝑉𝑖𝑠𝑢𝑎𝑙 andbecome
𝑆𝑡𝑒𝑝 smaller
based on as the number (18)
Equations of iterations
and (19). increases.
Meanwhile, evaluate swarming
Visual
behavior and Step can be
andfollowing calculated
behavior andaccording
select the to Expressions
optimal behavior(18)toand (19). The fitness
execute.
values of these two behaviors can be obtained based on Algorithm 2. Repeat
log(Visualmin /Visualmax )/log( genmax )
the above
Visual ( gen ) = Visual max ×
steps until the stopping criterion condition is satisfied.gen (18)
The Dijkstra algorithm is a path search algorithm to determine the optimal path in
one given path network. The traditional Dijkstra algorithm cannot prevent potential AGV
Step( gen) = Stepmax × genlog(Stepmin /Stepmax )/log( genmax ) (19)
conflict; thus, a criterion called “Node visit time (𝑁𝑇)” is presented to record the time at
where gen denotes
which AGVs thenodes,
visit all number of iterations,
expressed and Visual
by Equation max𝑁𝑇
(20). , Visual
makes , Step
minit max , and
possible Stepinto
to take min
denote the maximum and minimum of Visual and Step, respectively. genmax denotes the
maximum number of iterations. Visual ( gen) and Step( gen) denote Visual and Step with the
number of iterations.
The procedure of IAFSA is shown in Algorithm 1. First, combine tasks according to
task combination strategy and initialize the artificial fish (AF) population. Then, calculate
Visual and Step based on Equations (18) and (19). Meanwhile, evaluate swarming behavior
and following behavior and select the optimal behavior to execute. The fitness values of
these two behaviors can be obtained based on Algorithm 2. Repeat the above steps until
the stopping criterion condition is satisfied.
Mathematics 2023, 11, 2678 12 of 25

The Dijkstra algorithm is a path search algorithm to determine the optimal path in
one given path network. The traditional Dijkstra algorithm cannot prevent potential AGV
conflict; thus, a criterion called “Node visit time (NT)” is presented to record the time at
which AGVs visit all nodes, expressed by Equation (20). NT makes it possible to take into
account the path information of other AGVs when planning paths for one AGV, effectively
avoiding potential AGV conflicts.

NT = {nt1 , . . . , nti , . . . ,}, nti = {si , k i , ti } (20)

where nti denotes AGV k i visits to node si at moment ti .

The procedure of the Dijkstra algorithm is shown in Algorithm 2. First, the task
assignment result is obtained based on the decoding result. Based on the task assignment
results, the start and end nodes of the AGV can be determined so that the optimal route of
the AGV can be obtained by Dijkstra algorithm. The plan of AGV k can be denoted by Skv ,
expressed by Expression (21).
n        o
Skv = nvk,1 , tvk,1 , nvk,2 , tvk,2 , nvk,3 , tvk,3 , τk,L,1
v
, . . . , nvk,6 , tvk,6 , τk,U,1
v
,..., (21)

where the explanations of nvk,1 , tvk,1 , and τk,U,1

v are in Appendix B.

Algorithm 1: IAFSA in AGV scheduling.

Input: Problem data, IAFSA parameters.
Output: The optimal solution planbest , the completion time of all tasks Ybest .
1: Ybest ← ∞, planbest ← ∅
2: Task combination.
3: Initialize AF population [X1 , . . . , X NAF ]
4: while gen ← 1 to genmax do
5: calculate Visual and Step by Expressions (18) and (19)
6: for i ←1 to NAF do
7: [ Xi1 , Yi1 , plani1 ] ← swarm( Xi ) // Evaluate swarming behavior,
8: [ Xi2 , Yi2 , plani2 ] ← f ollow( Xi ) // Evaluate following behavior
9: [ Xi , Yi , plani ] ← (Yi1 < Yi2 )?[ Xi1 , Yi1 , plani1 ] : [ Xi2 , Yi2 , plani2 ]
10: end for
11: Ymin ← min{Y1 , Y2 , . . . , YNAF }
12: [Ybest , planbest ] ← (Ymin < Ybest )?[Ymin , planmin ] : [Ybest , planbest ]
13: end while

Then, NT is updated based on the routing results and sorted in chronological order. If
two AGVs visit the same path nodes and their visit moments do not satisfy Equation (15),
their infornamtion are recorded into C P . The set C P records the events that do not satisfy
Equation (15), expressed by Equation (22).

C P = {e1 , e2 , . . . , ei , . . .}, ei = {si , k i,1 , ti,1 , k i,2 , ti,2 } (22)

where ei denotes AGV k i,1 visits to node si at moment ti,1 , AGV k i,2 visits to node si at
moment ti,2 , and |ti,1 − ti,2 | < Ds /v0 .
If C P is empty, the time relationship of all AGVs visiting the path node satisfies
Equation (15), and there is no potential conflict. Otherwise, if e1 involves two AGVs, the
passage priority is decided according to Equation (23), and the low-priority AGV waits
for (∆t + Ds /v0 ), where ∆t denotes the time gap between the high-priority AGV and the
low-priority AGV to visit the path node. If e1 involves multiple AGVs, Equation (23) may
lead to a circular waiting problem; thus, Equation (24) is used to decide the passage order,
and the low-priority AGV waits for (∆t + Ds /v0 ).
Mathematics 2023, 11, 2678 13 of 25

Ck1 > Ck2 → Fk1 > Fk2 (23)

tk1 > tk2 → Fk1 > Fk2 (24)

where tk1 denotes the moment when AGV k1 visits the conflicting node and Fk1 denotes the
passage priority of AGV k1.

Algorithm 2: Dijkstra algorithm in AGV routing.

Input: Coding results of IAFSA, problem data
Output: The fitness value of IAFSA Y.
1: Decode to obtain task assignment results.
2: for k∈V do
3: planning path for AGV k using Dijkstra algorithm, recorded in Skv
4: end for
5: while True do
6: C P ← ∅ , NT ← ∅(s ∈ N )
7: update NT based on {. . . , Skv , . . .}, and sort NT in chronological order
8: for nti , nti+1 ∈ NT do
9: if si == si+1 and |ti+1 − ti | < Ds /v0 and si ∈ NP then
10: record {si , k i , ti , k i+1 , ti+1 } into C P
11: end if
12: end for
if isEmpty C P then


13:
14: break;
15: end if
16: sort C P in chronological order, tm ← t1,2
17: for ei ∈ C P (i > 1) do // update e1
18: if si == s1 and ti,1 − tm < Ds /v0 then
19: e1 = e1 ∪ ei , tm ← ti,2
20: end if
21: end for
22: if e1 involves two AGVs then
23: decide the pass priority according to Equation (23) and update Skv
24: else
25: decide the pass priority according to Equation (24) and update Skv
26: end if
27: end while
28: Y ← max (C1 , . . . , Ck , . . .) (k∈V)

4.3. Twin-Data-Driven AGV Conflict Prediction Method

During the AGV transport process, the occurrence of disturbance events (e.g., AGV
failure, AGV speed fluctuations, and uncertain QC operation times) will influence the
execution of the scheduling and routing plan, resulting in the deviation between the actual
plan and the expected plan. As the AGV operation is a continuous process, the accumu-
lation of deviations may affect the subsequent AGV operation and cause AGV conflicts.
Therefore, it is necessary to predict the subsequent potential AGV conflicts to facilitate early
countermeasures to deal with them. A twin-data-driven AGV conflict prediction method is
explored to monitor the deviation between the actual and original plans and then predict
AGV conflicts. The parameters in this section are shown in Appendix B. The workflow of
the twin-data-driven AGV conflict prediction method is shown in Figure 7.
Mathematics
Mathematics2023,
2023,11,
11,x 2678
FOR PEER REVIEW 15 14
of of2525

Figure
Figure7.7.The
Theworkflow
workflowofoftwin-data-driven
twin-data-drivenAGV
AGVconflict
conflictprediction
predictionmethod.
method.

4.4. AGV Conflict

During theResolution
AGV transport Method Based on
process, each Yen’sAGV Algorithm
reports its operation state information
to the virtual space in real-time, expressed
In this section, a conflict resolution method is proposed by Expression (25).
for resolving AGV conflicts
and revising the original plan. 
in f o = k, t, v , ak,t , posk,t , phasek,t (25)
The deviation between the initial plank,tand the revised plan can affect the stability of
the ACT operation.
If AGV Therefore,
k is located on thethe path
objectiveat momentof the AGV t (i.e.,conflict
posk,t resolution
/ N), the method
∈ time to is to
reach
minimize the time deviation of all AGVs’ task completion
its next node is predicted according to Expression (26). Expression (27) is used for con- time between the initial plan
and
flictthe revised plan, as inAGV(32), kwhile being subject to Equations (12)–(17).
prediction. When and n AGV q satisfy (27), the otwo AGVs will conflict in
their next node. The conflict event s, k, tk,t N , pos , q, t N , pos
k,t q,t q,t is recorded to the conflict
min (𝐶 − 𝐶 ) n (32)o
set CsN . Assume the two conflicts in CsN ,∈denoted as s, k1, tk1,t N , pos N
k1,t , k2, tk2,t , posk2,t ,
n o    
N
For tthose conflicts N 𝐶, pos N , tN − max N ,the N AGVs
s, k3, k3,t , pos k3,t , k4, in
tk4,t , the
k4,tAGVs
; if are minlocatedtk3,t on the
k4,t path, and
tk1,t tk2,t < will
Ds /vconflict
0 , these
attwo
their next node. Since the AGV transportation area
N is a directed
events are combined into one event {s, k1, tk1,t , posk1,t , k2, tk2,t , posk2,t , k3, tk3,t , posN graph, AGVs N can only ,
k3,t
travelN in one direction on a certain path. The conflicts in 𝐶 cannot be resolved by
k4, tk4,t , posk4,t }. C N records the conflict events for all AGVs’ next path nodes, C N =
planning a new route—only by the waiting strategy. All conflicting AGVs visit the conflict
C1N ∪ C2N ∪ . . . ∪ CsN ∪ . . . , s ∈ NP .
node sequentially according to the time order, and the time interval between two AGVs
visiting the conflicting node  is 𝐷 ⁄N𝑣 . Taking one conflict {𝑠, 𝑘1, 𝑡 , , 𝑝𝑜𝑠 , , 𝑘2, 𝑡 , ,
Dk,t /vk,t , i f ak,t = 0
𝑝𝑜𝑠 , , 𝑘3, 𝑡 , , 𝑝𝑜𝑠 , } as  an example, assume vthat 𝑡 , < 𝑡 , < 𝑡 , ; then, AGV k1

 2
∞, i f k2 < 0 and
ak,tvisits k,t
≤ Dk,tN

visits node s at moment N 𝑡
tk,t = , , AGV node2×s |at ak,tmoment
| , 𝑡∀k , ∈+V𝐷 ⁄𝑣 with waiting (26)
time 𝑡 , − 𝑡 , + 𝐷 ⁄𝑣 , and 𝑡 + 2𝐷 ⁄ 𝑣
q


 AGV v k,t k3
2+ 2 ×visits
a k,t × D Nnode
k,t − v k,t s at moment , with
waiting time 𝑡 , − 𝑡 , + 𝐷 ⁄𝑣 .

ak,t , o.w.
The conflicts in 𝐶 can be  resolved by re-routing for AGVs. Yen’s algorithm is an
N N < D /v
algorithm for obtaining multiple k,t − tq,t paths
 tshortest s from 0 a starting point to an endpoint with

the idea of a deviated path algorithm s = nk,t = nq,t , ∀method.
in the recursive k, q ∈ V An AGV conflict resolution (27)
based on Yen’s algorithm is explored to resolve conflicts in 𝐶 . The procedure of Yen’s

s ∈ NP

algorithm is shown in Algorithm 3. In Algorithm 3, m paths are planned from the current
If AGV k is located on the node at moment t (i.e., posk,t ∈ N), the deviations between
node to the phase endpoint for AGV k if the conflicting phase is the current phase of AGV
the plan and the actual operation will be measured to predict the occurrence of conflicts.
k. Otherwise, m paths are planned from the phase start point to the phase endpoint. The
Aiming to measure the deviation, a criterion called “Consistency deviation (∆TD)” is
waiting strategy is used to resolve the conflict if the conflict still exists in the conflicting
phase after re-routing the path. Otherwise, adopt Yen’s algorithm to resolve the
subsequent conflicts.
Mathematics 2023, 11, 2678 15 of 25

presented, defined as the deflection between the planned AGV operation state and the
actual AGV operation state, expressed by Expression (28).

∆TD = {∆TD N , ∆TDL , ∆TDU } (28)

Due to the fluctuating AGV speed during the AGV transport process, the actual time
of the AGV visiting one node will deviate from the planned time. The consistency deviation
of node visit time ∆TD N is calculated by Expression (29).

∆TD N = nrk,i − nvk,i (29)

Due to the uncertain operation time of QCs and YCs, one container’s actual loading
and unloading time will deviate from the planned time. The consistency deviation of
loading ∆TDL and unloading time ∆TDU is calculated by Expressions (30) and (31).
r v
∆TDL = τk,L,j − τk,L,j (30)

r v
∆TDU = τk,U,j − τk,U,j (31)
When ∆TD 6= 0, the current AGV’s times of visiting subsequent nodes need to
superimpose ∆TD. Assume that AGV k has a consistency deviation ∆TD N from the
expected
n  planwhen
 visiting the ith node. o k can be denoted as
  The updated plan ofAGV
. . . , nrk,i , trk,i , nvk,i+1 , tvk,i+1 + ∆TD N , nvk,i+2 , tvk,i+2 + ∆TD N , . . . .

NT will be updated as the AGV plan is updated and sorted in chronological order.
Then, if the time of visiting node si by two AGVs in NT satisfies |ti,1 − ti,2 | < Ds /v0 , the
conflict event {si , k i,1 , ti,1 , k i,2 , ti,2 } will be recorded to the conflict set C P . Assume the two
conflicts in C P , denoted as {si , k i,1 , ti,1 , k i,2 , ti,2 }, si , k j,1 , t j,1 , k j,2 , t j,2 ; if t j,1 − ti,2 < Ds /v0 ,


these two events are combined into one event si , k i,1 , ti,1 , k i,2 , ti,2 , k j,1 , t j,1 , k j,2 , t j,2 . C P


records the conflict events of all AGVs’ non-next path nodes.

4.4. AGV Conflict Resolution Method Based on Yen’s Algorithm

In this section, a conflict resolution method is proposed for resolving AGV conflicts
and revising the original plan.
The deviation between the initial plan and the revised plan can affect the stability of
the ACT operation. Therefore, the objective of the AGV conflict resolution method is to
minimize the time deviation of all AGVs’ task completion time between the initial plan and
the revised plan, as in (32), while being subject to Equations (12)–(17).
!
min ∑ Ck − Ck
0


(32)
k ∈V

For those conflicts in C N , the AGVs are located on the path, and the AGVs will conflict at
their next node. Since the AGV transportation area is a directed graph, AGVs can only travel
in one direction on a certain path. The conflicts in C N cannot be resolved by planning a new
route—only by the waiting strategy. All conflicting AGVs visit the conflict node sequentially
according to the time order, and thentime interval between two AGVs visiting the conflicting
o
N , pos
node is Ds /v0 . Taking one conflict s, k1, tk1,t N N
k1,t , k2, tk2,t , posk2,t , k3, tk3,t , posk3,t as an
example, assume that N
N < t N ; then, AGV k1 visits node s at moment t N ,
< tk2,t
tk1,t AGV
k3,t k1,t
N + D /v with waiting time t N − t N + D /v , and
k2 visits node s at moment tk1,t AGV
s 0 k1,t k2,t s 0
N + 2D /v with waiting time t N − t N + D /v .
k3 visits node s at moment tk1,t s 0 k1,t k3,t s 0
Mathematics 2023, 11, 2678 16 of 25

The conflicts in C P can be resolved by re-routing for AGVs. Yen’s algorithm is an

algorithm for obtaining multiple shortest paths from a starting point to an endpoint with the
idea of a deviated path algorithm in the recursive method. An AGV conflict resolution based
on Yen’s algorithm is explored to resolve conflicts in C P . The procedure of Yen’s algorithm
is shown in Algorithm 3. In Algorithm 3, m paths are planned from the current node to the
phase endpoint for AGV k if the conflicting phase is the current phase of AGV k. Otherwise,
m paths are planned from the phase start point to the phase endpoint. The waiting strategy
is used to resolve the conflict if the conflict still exists in the conflicting phase after re-routing
the path. Otherwise, adopt Yen’s algorithm to resolve the subsequent conflicts.

Algorithm 3: Yen’s algorithm in AGV conflict resolution.

Input: Problem data, real-time information for all AGVs, NT, Skv , C P
Output: the revised plan plannew
1: sort C P in chronological order, and update e1
2: for k∈ e1 do
3: obtain AGV k’s phase at the occurrence of e1 , denoted as phaseC
4: if phaseC is the current phase of AGV k then
5: arrange m paths from the current nodento the phase endpoint
o for AGV k
v , . . . , Sv
based on Yen’s algorithm, and update Sk,1 k,m
6: else
7: arrange m paths from the phase start point
n to the phaseoendpoint for AGV
v , . . . , Sv
k based on Yen’s algorithm, and update Sk,1 k,m
8: end if
9: for i ←1 to m do
10: while True do
11: update NT, and
C P .
12: if is Empty C P then

13: f k,i ← ∑k∈V Ck0 − Ck
14: break;
15: end if
16: if e1 ∈ phaseC then
17: adopt a waiting strategy to resolve the conflict, and update Sk,i v .

18: else
19: adopt Yen’s algorithm to resolve the conflict, and update Sk,i v .

20: end if
21: end while
22: end for
23: end for
24: select the plan plannew corresponding to min ( f 1,1 , f 1,2 , . . . , f k,i , . . .)

4.5. Evaluation Verification

Before the initial plan is executed in the physical space, the ACT evaluation verification
platform simulates the ACT operation process, thus evaluating the efficiency and feasibility
of the plan. The process of AGV evaluation verification is as follows [44].
First, the ACT evaluation verification platform is built based on Unity2021.2.7f1c1, as
shown in Figure 8. Then, the process flow and data interface of the ACT are established, and
real-time data collected from the physical space is transmitted to the simulation platform
through the data interface. Finally, the ACT operation process can be mimicked via the
simulation platform, based on the plan, to evaluate the task completion time of each
AGV, conflict incidence rate, etc. The evaluation indicators are an essential guide for ACT
operation of the physical space. Meanwhile, since these models in the virtual space can be
continuously updated according to real-time data, the simulation results can better reflect
the actual AGV operation situation.
x FOR PEER REVIEW 17 of 25

Mathematics 2023,
Mathematics 2023,11,
11,x2678
FOR PEER REVIEW 17 of 25 17

Figure 8. The ACT evaluation verification platform in Unity.

5. Experiments
Figure
In the section, some
Figure 8. 8. The
experiments
The ACTACT evaluation
are conducted
evaluation verification
verification platform
to verify
platform the in
in Unity. Unity. of the proposed
benefits
method. The first experiment
5. Experimentsof AGV scheduling and routing is aimed to verify the
5. Experiments
performance of the IAFSA and calculate
In the section, somethe optimalare
experiments scheduling
conducted and paththeplanning
to verify benefits ofresults,
the proposed
and based on the result
method. In The
of the section,
thefirst
first some of
experiment,
experiment experiments
AGV the secondare conducted
scheduling experiment
and routing is toaimed
verify
of AGV tothe benefits
conflict
verify of the prop
the perfor-
method.
mance of The
the IAFSAfirstandexperiment
calculate theof AGV
prediction and resolution is conducted to verify the efficiency of DT-based conflict
optimal scheduling
scheduling and and
path routing
planning is aimed
results, andto verify
based onAll
performance
prediction and resolution. the experiments
result of the
of the first
IAFSA experiment,
are and the on
calculate
performed second experiment
theWindows
optimal 10,ofIntel
schedulingAGV(R) conflict
and pathpredic-
Core planning res
tion
and and resolution
based on16GB is conducted
the result to verify the efficiency of DT-based conflict prediction andAGV co
(TM) i7-10750H CPU @ 2.60GHz, RAMofMATLAB the first experiment,
2018a. the second experiment of
resolution. All experiments are performed on Windows 10, Intel (R) Core (TM) i7-10750H
As shown in Figureprediction and resolutionarea is conducted tointo
verifythe the efficiency of DT-based co
CPU @9,2.60GHz,
the transportation
16GB RAM MATLAB is abstracted
2018a. weighted directed
graph, including 18 prediction
nodes andand
As shown 44 resolution.
paths.
in Figure 9,n3,
the n5,Alland
experiments area isare
n7 indicate
transportation performed
the three
abstracted on weighted
intonodes
the Windows
of QCs, 10, Intel (R)
directed
(TM) including
graph,
n12, n14, and n16 represent i7-10750H
the three 18CPU
nodes
[email protected],
of 44 16GB
paths.
storage n3, RAM
n5,
yards, and
andMATLAB
n7the
indicate
other 2018a.
the three are
nodes nodes of QCs,
the
As
n12, n14, shown in Figure
n16 represent 9, the
threetransportation area
path nodes. The number of QCs and YCs is 3. The safe distance 𝐷 = 15 m, and the velocityare
and the nodes of storage is
yards, abstracted
and the other into the
nodes weighted
the dire
path nodes.
graph, The number
including of QCs and
18 nodes andYCs44 is 3. Then3,
paths. safen5,
distance
and n7 Ds indicate
= 15 m, and thethree
the velocity
nodes of
𝑣 = 5 m/s. vn12,
0 = 5 n14,
m/s. and n16 represent the three nodes of storage yards, and the other nodes ar
path nodes. The number of QCs and YCs is 3. The safe distance 𝐷 = 15 m, and the vel
𝑣 = 5 m/s.

Figure
Figure 9. The AGV path 9. The
in an ACT.AGV path in an ACT.

5.1. Verification of AGV Scheduling and Routing

5.1. Verification of AGV Scheduling
Figure and path
The9.experiment
The AGV Routing
in an
of AGV ACT.
scheduling and routing is used to verify the performance of
The experimentthe ofIAFSA
AGV and calculate the
scheduling and optimal
routingscheduling
is usedand path planning
to verify results.
the performance of
5.1. Verification
The of
experimental AGV
data Scheduling
involved in and
this
the IAFSA and calculate the optimal scheduling and path planning results. Routing
Section is shown in Table 1, including container
no. and the container transportation tasks. The experiment consists of 30 container trans-
The experimental The experiment
data involved inof by
AGVSection
this schedulingshown and routing is used to verify the performan
portation tasks transported 6 AGVs. The is start nodes ofintheTable
6 AGVs 1,are including
[n1, n5, n3, n15,
container no. and n7, then6].
the IAFSA
container and transportation
The IAFSA calculate
parameters theare
optimal
tasks.
set scheduling
based The and path
experiment
on preliminary planning
testsconsists
and are set results.
ofas30
follows:
container transportation The
tasks
Visualmin experimental
transported
= 0.1, Visualmax = 1.5, data
by Stepinvolved
6 AGVs.
min = The
0.1, in
Step this
start
max = Section
nodes
1.5, genof is
the
max = shown
6
300, AGVs in Table
are
try_number = 10,1, inclu
container
= 0.618, and
[n1, n5, n3, n15, n7, n6]. The IAFSA
δ no.
N and
= 20.the container transportation tasks.
AF parameters are set based on preliminary tests and are The experiment consists o
set as follows: 𝑉𝑖𝑠𝑢𝑎𝑙 container
= 0.1,transportation
𝑉𝑖𝑠𝑢𝑎𝑙 =tasks transported
1.5, 𝑆𝑡𝑒𝑝 by 𝑆𝑡𝑒𝑝
= 0.1, 6 AGVs. The = 1.5,start𝑔𝑒𝑛nodes=of the 6 AGV
300, try_number = 10,[n1,
𝛿 =n5, n3, n15,
0.618, andn7, 𝑁 n6]. The IAFSA parameters are set based on preliminary tests an
= 20.
set as follows: 𝑉𝑖𝑠𝑢𝑎𝑙
As shown in Figure 6b, the task combination = 0.1, 𝑉𝑖𝑠𝑢𝑎𝑙is adopted
strategy = 1.5, 𝑆𝑡𝑒𝑝
to combine = 0.1, 𝑆𝑡𝑒𝑝
container = 1.5, 𝑔𝑒𝑛
300, try_number = 10, 𝛿 = 0.618, and 𝑁 = 20.
tasks according to specific rules to reduce the encoding length. Through the task
As shown in Figure 6b, the task combination strategy is adopted to combine cont
combination strategy, the 30 container transportation tasks are combined to 16 post-
tasks according to specific rules to reduce the encoding length. Through the
combination tasks, as shown in Table 2, including new task no., task nodes, and post-
Mathematics 2023, 11, 2678 18 of 25

Table 1. Container transportation tasks.

No. 2023, 11,Task

Mathematics No.
x FOR PEER REVIEW Task No. Task No. Task No. Task
18 of 25
1 [n3, n16] 7 [n3, n14] 13 [n14, n3] 19 [n3, n14] 25 [n14, n3]
2 [n3, n12] 8 [n7, n14] 14 [n14, n7] 20 [n3, n12] 26 [n7, n14]
3 [n16, n5] 9 [n14, n5] 15 [n12, n3] 21 [n14, n5] 27 [n5, n14]
4 [n16, n7] Table 1. Container
10 [n7, n16] transportation
16 tasks.
[n7, n16] 22 [n3, n16] 28 [n3, n12]
5 [n5, n16] 11 [n5, n14] 17 [n12, n7] [n7, n16] [n12, n7]
N Task No. Task No. Task No.23 Task 29
No. Task
6 [n7, n12] 12 [n16, n3] 18 [n12, n5] 24 [n16, n3] 30 [n16, n3]
1 [n3, n16] 7 [n3, n14] 13 [n14, n3] 19 [n3, n14] 25 [n14, n3]
2 [n3, n12] 8 [n7, n14] 14 [n14, n7] 20 [n3, n12] 26 [n7, n14]
3 [n16, n5] 9 As shown
[n14, n5] in Figure
15 6b, the task combination
[n12, n3] 21 strategy is adopted
[n14, n5] to
27 combine[n5,container
n14]
4 [n16, n7] 10 tasks according
[n7, n16] to specific
16 rules to reduce
[n7, n16] the encoding
22 length.
[n3, n16] Through
28 the task
[n3,combina-
n12]
tion strategy, the 30 container transportation tasks are combined to 16 post-combination
5 [n5, n16] 11 [n5, n14] 17 [n12, n7] 23 [n7, n16] 29 [n12, n7]
tasks, as shown in Table 2, including new task no., task nodes, and post-combination
6 [n7, n12] 12 [n16, n3] 18 [n12, n5] 24 [n16, n3] 30 [n16, n3]
task no.
Table
Table 2.
2. Post-combination
Post-combination tasks.
tasks.
No. Nodes Task No. Nodes Task No. Nodes Task
No. Nodes Task No. Nodes Task No. Nodes Task
1 [n3, n16, n5] 1,3 7 [n7, n16, n3] 10,24 13 [n3, n12, n7] 20,29
1 [n3, n16, n5] 1,3 7 [n7, n16, n3] 10,24 13 [n3, n12, n7] 20,29
2 [n3, n12, n3] 2,15 8 [n5, n14, n7] 11,14 14 [n14, n3, n16] 25,22
2 [n3, n12, n3] 2,15 8 [n5, n14, n7] 11,14 14 [n14, n3, n16] 25,22
33 [n16,
[n16, n7, n12]
n7, n12] 4,6
4,6 9 9 [n7,
[n7,n16,
n16, n3]
n3] 16,30
16,30 1515 [n7,
[n7, n14]
n14] 2626
44 [n5,
[n5, n16, n3]
n16, n3] 5,12
5,12 1010 [n12,
[n12,n7,
n7, n16]
n16] 17,23
17,23 1616 [n3,
[n3, n12]
n12] 2828
55 [n3, n14, n5]
[n3, n14, n5] 7,9
7,9 1111 n5, n14]
[n12,n5,
[n12, n14] 18,27
18,27
[n7, n14, n3] [n3, n14, n5]
66 [n7, n14, n3] 8,13
8,13 1212 [n3, n14, n5] 19,21
19,21

The
TheAGV AGV scheduling
schedulingGanttGanttchart
chartisis shown
shown according
according to to the
the scheduling
scheduling and and routing
routing
plan
plan in
in Figure
Figure 10a.
10a. The
Thenumbers
numbersin ineach
eachrectangle
rectangleindicate
indicatethe
thecontainer
containerno.,
no.,the
thecontainer
container
loading completion time,
loading completion time,andandthethe container
container unloading
unloading start start
time intime in sequence.
sequence. The
The comple-
completion
tion time fortime for allistasks
all tasks is 386.
386. The The
path is path
shownis according
shown according to the scheduling
to the scheduling and
and routing
plan in Figure
routing plan in10b.
FigureThe10b.
time interval
The for two for
time interval AGVs
twoto visit to
AGVs thevisit
samethenode
sameisnode
greater than or
is greater
equalortoequal
than 3 s (Dtos /v = 15
3 s0 (𝐷 ⁄𝑣m/(5
= 15m/s)
m/(5=m/s)
3 s),=which satisfies
3 s), which Constraint
satisfies (15). Thus,
Constraint the AGV
(15). Thus, the
routing
AGV is conflict-free.
routing is conflict-free.

(a) (b)
Figure
Figure 10.
10. AGV
AGV scheduling
scheduling and
and routing result of
routing result of 66 AGVs
AGVs handling
handling 30
30containers
containers(a)
(a)AGV
AGVscheduling
scheduling result; (b) AGV routing result.
result; (b) AGV routing result.

To verifythe
To verify the effectiveness
effectiveness of improvement
of improvement strategies,
strategies, the time
the running running time and
and completion
completion
time of IAFSA time of AFSA
and IAFSAare and AFSA areWe
compared. compared. We solved
solved each each 10
experiment experiment
times and10 times
took the
and tookvalue
average the average valueThe
as a result. as aresults
result.of
The
theresults of the are
experiment experiment
shown inare shown
Table in running
3, the Table 3,
the running
time denotestime
the denotes the algorithm’s
algorithm’s execution timeexecution
when thetime when thenumber
maximum maximum number of
of iterations is
iterations is reached,
reached, and and completion
completion time denotes time
thedenotes
maximum the maximum
completion completion timewhen
time of tasks of tasks
the
when
maximumthe maximum
number ofnumber
iterationsofisiterations is reached.
reached. The The GAP
GAP represents therepresents the relative
relative improvement
improvement
percentage of percentage
the IAFSA. In of Table
the IAFSA. In Tablethe
4, NC denotes 4, Nnumber
C denotesof the numberNof
containers, containers,
T denotes the
NT denotes the number of tasks after task combination, and the relative improvement
percentage GAP was employed to measure the performance of the IAFSA. The GAP can
be calculated by Equation (33):
𝑍 −𝑍
Mathematics 2023, 11, 2678 19 of 25

number of tasks after task combination, and the relative improvement percentage GAP
was employed to measure the performance of the IAFSA. The GAP can be calculated by
Equation (33):
Z − ZIAFSA
GAP = AFSA × 100%, (33)
ZAFSA
where ZAFSA denotes the running time or the completion time of the AFSA and ZIAFSA
denotes the running time or the completion time of the IAFSA.

Table 3. Results comparison between AFSA and IAFSA.

Running Time (s) Completion Time (s)

No. NC /NT AGV
AFSA IAFSA GAP AFSA IAFSA GAP
1 20/12 4 702 293 58.26% 430 398 7.44%
2 30/17 5 1394 565 59.47% 529 467 11.72%
3 30/16 6 1443 573 60.29% 452 408 9.73%
4 50/26 6 2136 684 67.98% 751 656 12.65%
5 100/53 8 6250 1837 70.61% 1174 1072 8.69%
6 100/55 10 8182 2401 70.66% 966 897 7.14%
7 200/103 10 23638 5215 77.94% 1930 1691 12.38%
8 200/103 15 35110 9820 72.03% 1338 1221 8.74%

Table 4. Completion time comparison between different algorithms.

Completion Time (s)

No. NC /NT AGV
GA PSO GWO IAFSA
1 10/6 2 409 386 396 367
2 20/11 3 572 541 546 493
3 30/15 4 643 618 626 550
4 30/15 5 539 502 517 445
5 50/26 6 793 759 769 681
6 80/42 8 975 930 933 871
7 80/42 10 841 790 805 718
8 100/51 10 1035 984 1006 897

For different problem scales, the running time of the AFSA is between 702 and 35110 s,
and the running time of the IAFSA is between 293 and 9820 s; the completion time of AFSA
is between 430 and 1930 s, and the completion time of IAFSA is between 398 and 1691 s. The
running time average GAP between AFSA and IAFSA in the eight experiments is 67.15%.
The completion time average GAP between AFSA and IAFSA in the eight experiments
is 9.81%.
To verify the effectiveness of IAFSA, IAFSA was compared with other algorithms
in terms of completion time, including the genetic algorithm (GA), particle swarm opti-
mization algorithm (PSO), and gray wolf optimization algorithm (GWO). We solved each
experiment 10 times and took the average value as a result. The results of the experiment
are shown in Table 4. For different problem sizes, the solution quality of IAFSA is better
than the other algorithms compared.

5.2. Verification of AGV Conflict Prediction and Resolution

In order to verify the effectiveness of the AGV conflict prediction and resolution
method proposed in Section 4, some disturbance events are simulated. The disturbance
events considered in this work are uncertain container loading and unloading times. As-
sume that each container’s loading and unloading time obeys a uniform distribution
U (7, 13).
As shown in Figure 11, which shows the loading and unloading times of the containers
handled by the 6 AGVs, the horizontal coordinates indicate the loading and unloading
 Mathematics 2023, 11, 2678 20 of 25

process, (1, L) represents the loading node of the first container, (2, U) represents the
Mathematics 2023, 11, x FOR PEER REVIEW
unloading 20 of 25
node of the second container, and so on. The vertical coordinates indicate the
loading/unloading time.

Figure11.
Figure Theloading
11.The loadingand
andunloading
unloadingtimes
timesofofthe
thecontainers
containershandled
handledbybythe
the6 6AGVs.
AGVs.

Theuncertain
The uncertaincontainer
container loading/unloading
loading/unloading timestimes lead
lead to
to AGV
AGV conflicts.
conflicts.The
Thedetails
detailsof
some predicted
of some predicted AGV conflict events are shown in Table 5. At the 100th
AGV conflict events are shown in Table 5. At the 100th second, second, the loading
the
of AGV2 0 s second container is completed, and the virtual space predicts that there will
loading of AGV2′s second container is completed, and the virtual space predicts that there
be be
will AGVAGVconflicts. The
conflicts. Thefirst
firstpredicted
predictedAGV
AGVconflict
conflictisis Conflict
Conflict 11 in
in Table
Table 5.
5.For
ForConflict
Conflict
1, AGV20 s route to transport its second container is re-routed to avoid the conflict. The
1, AGV2′s route to transport its second container is re-routed to avoid the conflict. The
route of AGV2 to transport its second container before re-routing is . . . → n3 → n4 → n15
route of AGV2 to transport its second container before re-routing is …→ n3 → n4 → n15
→ n16 → . . . , and the route of AGV2 after re-routing is . . . → n3 → n2 → n17 → n16 →
→ n16 →…, and the route of AGV2 after re-routing is …→ n3 → n2 → n17 → n16 →…. A
. . . . A new conflict is caused after re-routing, as in Conflict 2 in Table 5. For Conflict 2,
new conflict is caused after re-routing, as in Conflict 2 in Table 5. For Conflict 2, AGV5′s
AGV50 s route to transport its third container is re-routed to avoid the conflict. The route of
route to transport its third container is re-routed to avoid the conflict. The route of AGV5
AGV5 to transport its third container before re-routing is . . . → n14 → n15 → n4 → n5 →
to transport its third container before re-routing is …→ n14 → n15 → n4 → n5 →…, and
. . . , and the route of AGV5 after re-routing is . . . → n14 → n13 → n6 → n5 → . . . . At the
the route of AGV5 after re-routing is …→ n14 → n13 → n6 → n5 →…. At the 154th second,
154th second, the loading of AGV10 s third container is completed, and the virtual space
the loading of AGV1′s third container is completed, and the virtual space predicts that
predicts that there will be AGV conflicts. The first predicted AGV conflict is Conflict 3 in
there will be AGV conflicts. The first predicted AGV conflict is Conflict 3 in Table 5. For
Table 5. For Conflict 3, both AGV1 and AGV2 keep the original path. AGV1 waits for 1 s
Conflict 3, both AGV1 and AGV2 keep the original path. AGV1 waits for 1 s after0 AGV2
after AGV2 visits n4 before visiting n4. At the 261st second, the loading of AGV4 s fourth
visits n4 before
container visiting n4.
is completed, andAtthethevirtual
261st second, the loading
space predicts of AGV4′s
that there will befourth containerThe
AGV conflicts. is
completed, and the virtual space predicts that there will be AGV conflicts.
first predicted AGV conflict is Conflict 4 in Table 5. For Conflict 4, both AGV4 and AGV6 The first
predicted
keep the AGVoriginalconflict
path.isAGV4
Conflict 4 infor
waits Table 5. ForAGV6
2 s after Conflict 4, both
visits AGV4visiting
n13 before and AGV6 n13.keep
the original path. AGV4 waits for 2 s after AGV6 visits n13 before visiting n13.
Table 5. The description of AGV conflicts.
Table 5. The description of AGV conflicts.
Conflict Time Node AGV Description
Conflict Time Node AGV Description
1 AGV5 and AGV2 happen conflict in the
1 112–113 n4
112–113 5-2 AGV5
n4 and AGV2 happen
5-2 conflict in the 112th–113th second in node n4.
112th–113th second in node n4.
2 166–168 n15 2-5 AGV2 and AGV5 happen conflict in the 166th–168th second in node n15.
AGV2 and AGV5 happen conflict in the
3 2 182–184 166–168
n4 2-1 n15 and AGV1 happen
AGV2 2-5 conflict in the 182th–184th second in node n4.
166th–168th second in node n15.
4 270–271 n13 6-4 AGV6 and AGV4 happen conflict in the 270th–271st
AGV2 and AGV1second
happen in nodeinn13.
conflict the
3 182–184 n4 2-1
182th–184th second in node n4.
AGV6 and AGV4 happenanconflict
AGV in the
4 270–271 In order to handle
n13 the AGV conflicts
6-4 described above, introduce conflict
270th–271st second in node n13.
resolution method based on Yen’s algorithm. The procedure of the conflict resolution
method is shown in Algorithm 3, and it can solve the path of AGV conflicts well. As shown
in Figure 12, the
In order to revised
handle path of 6 AGVs
the AGV handing
conflicts 30 containers
described is the result
above, introduce an AGVof conflict
conflict
resolution.
resolutionAfter
methodthe plan revision,
based thealgorithm.
on Yen’s task completion times for each
The procedure AGV
of the were 284
conflict s, 278
resolution
s,method
378 s, 382 s, 358 s, and 344 s.
is shown in Algorithm 3, and it can solve the path of AGV conflicts well. As shown
in Figure 12, the revised path of 6 AGVs handing 30 containers is the result of conflict
resolution. After the plan revision, the task completion times for each AGV were 284 s,
278 s, 378 s, 382 s, 358 s, and 344 s.
Mathematics 2023, 11, x FOR PEER REVIEW
Mathematics 2023, 11, 2678 21 of 25

Mathematics 2023, 11, x FOR PEER REVIEW 21 of 25

Figure
Figure 12.
Figure 12.
12. Path of
Path
Path of6of
6AGVs handling
6 AGVs
AGVs 3030containers
handling
handling after
afterconflict
30 containers
containers resolution.
after
conflict conflict resolution.
resolution.

Uncertain
Uncertaincontainer
containerloading
loadingororunloading
unloading timetimeis simulated to verify
is simulated the efficiency
to verify of
the efficiency
Uncertain
DT-based conflict
of DT-based container
prediction
conflict prediction loading
andand or unloading
resolution in in
resolution different
different time is simulated
problem
problem scales.
scales.The
Thetoresults
verify
results the effici
areare
DT-based
shown conflict
shown in Figure
Figure 13,the
13, prediction
the and
traditionalmethod
traditional methodresolution
totoconflict
conflict isin
thedifferent
is the waiting
waiting problem
strategy.
strategy. scales.
TheThe DT- The resu
DT-based
based
shown method uses
in Figure
method uses DT technology to
13, thetotraditional
DT technology predict and
method
predict and resolve resolve
AGV AGV
toconflictsconflicts
conflictwhenis the when the
waiting
the plan plan
strategy. T
execution
execution
deviates. deviates.
The The axis
vertical vertical axis indicates
indicates the time the time deviation
deviation between between
the the initial
initial and and
revised
based method uses DT technology to predict and resolve AGV conflicts when th
revised plans.
plans. The The difference
difference in theofeffect
in the effect of the
the two two methods
methods on the
on the time time deviation
deviation is not
is not apparent
execution
apparent
when the when
deviates.
problemthescales The
problem vertical
scales As
are small.
axis
arethe indicates
small. As the
number
the
number
of AGVs
time
and
deviation
of containers containersthe the init
between
AGVs andincreases,
revised
DT-basedplans.
increases, the Thehas difference
DT-based
method amethod
smallerhas
timein the effect
a smaller
deviationtime of
than thetraditional
deviation
the two
thanmethods
themethod. on the
traditional time deviatio
method.
apparent when the problem scales are small. As the number of AGVs and con
increases, the DT-based method has a smaller time deviation than the traditional m

Figure 13.
Figure Comparisonbetween
13. Comparison betweenDT-based
DT-basedmethod
methodand
andtraditional
traditionalmethod.
method.

6. Conclusions
6. Conclusions
In this
In thispaper,
paper,digital-twin-driven
digital-twin-driven AGVAGVscheduling and conflict-free
scheduling routingrouting
and conflict-free is proposed
is
to solve the uncertainties in the physical environment of ACT. The
proposed to solve the uncertainties in the physical environment of ACT. The bi-level bi-level mixed integer
programming
mixed integer model of the AGV
programming model scheduling
of the AGV andscheduling
routing problem is established.
and routing problem Theis
improved
Figure 13. artificial
Comparisonfish swarm
betweenalgorithm Dijkstra
DT-based (IAFSA-Dijkstra)
method and is proposed
traditional
established. The improved artificial fish swarm algorithm Dijkstra (IAFSA-Dijkstra) method.to findisthe
optimal AGV
proposed scheduling
to find and routing
the optimal plan. A twin-data-driven
AGV scheduling and routing plan.conflict prediction method
A twin-data-driven
is used for
prediction method is used for AGV conflict prediction caused conflict
AGV conflict prediction caused by disturbance events, and a resolution
6. Conclusions
conflict by disturbance
method is presented for AGV conflict resolution. The experimental
events, and a conflict resolution method is presented for AGV conflict resolution. results in different
The
problemIn this
scales paper,
show digital-twin-driven
that the average solution AGV
efficiencyscheduling
and average
experimental results in different problem scales show that the average solution efficiency and conflict-free
solution quality rou
of IASFA
proposed
and average toare better than
solvequality
solution those of AFSA
the uncertainties by
of IASFA are better 9.81%
inthan and
thethose 67.15%,
physical
of AFSArespectively.
environment Meanwhile,
of ACT. The
by 9.81% and 67.15%,
IAFSA has better solution quality compared with GA, PSO, and GWO. However, theand
port
mixed integer
respectively. programming
Meanwhile, model
IAFSA has better of the
solution AGV
quality scheduling
compared with GA,and
PSO, routing prob
environment
GWO. However, is characterized
the port by continuous,
environment is complex,
characterized andby interlocking
continuous, components,
complex, and
and
established.
the destruction The
of anyimproved
link may artificial fish swarmofalgorithm Dijkstraprocess.
(IAFSA-Dijk
interlocking components, and thelead to the deterioration
destruction of any link may the leadentire
to theoperation
deterioration of
proposed
Therefore, to future
our find the work optimal
is AGV the
to consider scheduling
integrated and routing plan.
andAother
twin-data
the entire operation process. Therefore, our future work isproblem of AGVs
to consider the integrated
conflict
problem of AGVs and other equipment in the port environment to be more in line with by distu
equipment prediction
in the port method
environment is used
to be for
more AGV
in line conflict
with the prediction
actual situation caused
of the port.
events,
the actualand a conflict
situation resolution method is presented for AGV conflict resolutio
of the port.
experimental results in different problem scales show that the average solution effi
and average solution quality of IASFA are better than those of AFSA by 9.81% and
respectively. Meanwhile, IAFSA has better solution quality compared with GA, PS
Mathematics 2023, 11, 2678 22 of 25

Author Contributions: Conceptualization, P.L. and Y.Z.; methodology, P.L., Y.Z. and X.C. validation,
P.L., Y.Z. and X.C.; formal analysis, C.F. and X.C.; investigation, P.L. and Y.Z.; writing—original draft
preparation, P.L. and Y.Z.; writing—review and editing, X.C. and C.F.; visualization, Y.Z.; supervision,
P.L. and X.C.; project administration, P.L. and J.H.; funding acquisition, P.L. and J.H. All authors have
read and agreed to the published version of the manuscript.
Funding: This research was funded by the National Natural Science Foundation Committee (NSFC)
of China, grant number 52075404, the National Key Research and Development Project of China,
grant number 2020YFB1710804, and the Application Basic Frontier Special Project of Wuhan Science
and Technology Bureau, grant number 2020010601012176.
Data Availability Statement: Not applicable.
Acknowledgments: The authors are grateful to the subjects for their contributions to the experiment.
Conflicts of Interest: The authors declare no conflict of interest.

Appendix A. Notations and Description in Problem Formulation

Notations Description
V Set of AGVs.
C Set of containers.
N Set of nodes. N = NQ ∪ NY ∪ NP .
NQ Set of QCs’ nodes.
NY Set of storage yards’ nodes.
NP Set of path nodes.
G Set of directed line segment between the nodes.
k Index of AGVs, k ∈ V.
p, p0 Index of container transport task, p, p0 ∈ C.
0 The virtual initial container transport task.
f The virtual last container transport task.
s, e, i, j Index of nodes, s, e, i, j ∈ N.
Cmax All task completion time.
Ck The completion time of the last task of AGV k.
Lp The pickup point of container p.
Up The delivery point of container p.
Ok The initial node of AGV k.
v0 The AGV velocity.
Ds The safe distance between AGVs.
dij The distance from node i to node j, in which (i, j) ∈ G.
tkp The moment when AGV k starts handling container p.
Tijk The travel time for AGV k on the connected path of node i and j.
k
tin,ij The moment when AGV k enters the connected path of node i and j.
k
tout,ij The moment when AGV k leaves the connected path of node i and j.
The travel time for AGV k from the initial position of AGV k to the pickup point of
TOk k L p
container p.
TLk p Up The travel time for AGV k when transporting container p.
The travel time for AGV k from the delivery point of container p to the pickup point of
TUk p L
p0 container p’.
p
τL The time to load container p.
p
τU The time to unload container p.
M A very large positive number.
αkp =1 if container p is assigned on AGV k for transporting; =0 otherwise.
βkpp0 =1 if container p and container p’ are transported by AGV k consecutively; =0 otherwise.
µijk =1 if AGV k passes through path from node i to node j, in which (i, j) ∈ G; =0 otherwise.
Mathematics 2023, 11, 2678 23 of 25

Appendix B. Parameters in Conflict Prediction Method

Parameters Descriptions
nvk,i The ith node that AGV k plans to visit.
nrk,i The ith node that AGV k actually visits.
tvk,i The planned time for AGV k to visit its ith node.
trk,i The actual time for AGV k to visit its ith node.
vk,t The speed of AGV k at moment t.
ak,t The acceleration of AGV k at moment t.
posk,t The position of AGV k at moment t.
The phase of AGV k at moment t. The traveling process from the pickup point of one
container to its delivery point, from the delivery point of one container to the pickup
phasek,t
point of the next container, or from the initial position of AGV k to the pickup point
of its first container is recorded as one phase.
N The distance of AGV k from its next node at moment t, which can be calculated
Dk,t
based on the values of posk,t and phasek,t .
The next node of AGV k at moment t, which can be obtained based on the values of
nk,t
posk,t and phasek,t .
v
τk,L,j The planned loading time of the jth task of AGV k.
r
τk,L,j The actual loading time of the jth task of AGV k.
v
τk,U,j The planned unloading time of the jth task of AGV k.
r
τk,U,j The actual unloading time of the jth task of AGV k.

References
1. Zhong, M.; Yang, Y.; Dessouky, Y.; Postolache, O. Multi-AGV scheduling for conflict-free path planning in automated container
terminals. Comput. Ind. Eng. 2020, 142, 106371. [CrossRef]
2. Hu, H.; Yang, X.; Xiao, S.; Wang, F. Anti-conflict AGV path planning in automated container terminals based on multi-agent
reinforcement learning. Int. J. Prod. Res. 2021, 61, 65–80. [CrossRef]
3. Luo, J.; Wu, Y. Modelling of dual-cycle strategy for container storage and vehicle scheduling problems at automated container
terminals. Transp. Res. Part E Logist. Transp. Rev. 2015, 79, 49–64. [CrossRef]
4. Xu, B.; Jie, D.; Li, J.; Yang, Y.; Wen, F.; Song, H. Integrated scheduling optimization of U-shaped automated container terminal
under loading and unloading mode. Comput. Ind. Eng. 2021, 162, 107695. [CrossRef]
5. Lu, F.; Feng, W.; Gao, M.; Bi, H.; Wang, S. The Fourth-Party Logistics Routing Problem Using Ant Colony System-Improved Grey
Wolf Optimization. J. Adv. Transp. 2020, 2020, 8831746. [CrossRef]
6. Yan, T.; Lu, F.; Wang, S.; Wang, L.; Bi, H. A hybrid metaheuristic algorithm for the multi-objective location-routing problem in the
early post-disaster stage. J. Ind. Manag. Optim. 2023, 19, 4663–4691. [CrossRef]
7. Lu, F.; Yan, T.; Bi, H.; Feng, M.; Wang, S.; Huang, M. A bilevel whale optimization algorithm for risk management scheduling of
information technology projects considering outsourcing. Knowl.-Based Syst. 2022, 235, 107600. [CrossRef]
8. Cai, B.; Huang, S.; Liu, D.; Dissanayake, G. Rescheduling policies for large-scale task allocation of autonomous straddle carriers
under uncertainty at automated container terminals. Robot. Auton. Syst. 2014, 62, 506–514. [CrossRef]
9. Jian, W.; Zhu, J.; Zeng, Q. An Optimization Model of Integrated AGVs Scheduling and Container Storage Problems for Automated
Container Terminal Considering Uncertainty. Symmetry 2021, 13, 1904. [CrossRef]
10. Yue, L.; Fan, H. Dynamic Scheduling and Path Planning of Automated Guided Vehicles in Automatic Container Terminal.
IEEE/CAA J. Autom. Sin. 2022, 9, 2005–2019. [CrossRef]
11. Xu, B.; Wang, L.; Li, J. Propagation of Uncertain Events in Multilevel Handlings at Container Terminals from the Perspective of
Hypernetwork. Discret. Dyn. Nat. Soc. 2021, 2021, 6611181. [CrossRef]
12. Zheng, Y.; Yang, S.; Cheng, H. An application framework of digital twin and its case study. J. Ambient. Intell. Humaniz. Comput.
2019, 10, 1141–1153. [CrossRef]
13. Rashidi, H.; Tsang, E.P.K. A complete and an incomplete algorithm for automated guided vehicle scheduling in container
terminals. Comput. Math. Appl. 2011, 61, 630–641. [CrossRef]
14. Ma, X.; Bian, Y.; Gao, F. An Improved Shuffled Frog Leaping Algorithm for Multiload AGV Dispatching in Automated Container
Terminals. Math. Probl. Eng. 2020, 2020, 1260196. [CrossRef]
15. Zhao, Q.; Ji, S.; Guo, D.; Du, X.; Wang, H. Research on Cooperative Scheduling of Automated Quayside Cranes and Automatic
Guided Vehicles in Automated Container Terminal. Math. Probl. Eng. 2019, 2019, 6574582. [CrossRef]
Mathematics 2023, 11, 2678 24 of 25

16. Angeloudis, P.; Bell, M.G.H. An uncertainty-aware AGV assignment algorithm for automated container terminals. Transp. Res.
Part E Logist. Transp. Rev. 2010, 46, 354–366. [CrossRef]
17. Xin, J.; Negenborn, R.R.; Lodewijks, G. Rescheduling of interacting machines in automated container terminals. IFAC Proc. Vol.
2014, 47, 1698–1704. [CrossRef]
18. Sahin, C.; Demirtas, M.; Erol, R.; Baykasoğlu, A.; Kaplanoğlu, V. A multi-agent based approach to dynamic scheduling with
flexible processing capabilities. J. Intell. Manuf. 2017, 28, 1827–1845. [CrossRef]
19. Xu, W.; Guo, S.; Li, X.; Guo, C.; Wu, R.; Peng, Z. A Dynamic Scheduling Method for Logistics Tasks Oriented to Intelligent
Manufacturing Workshop. Math. Probl. Eng. 2019, 2019, 7237459. [CrossRef]
20. Zhang, L.; Yan, Y.; Hu, Y.; Ren, W. A dynamic scheduling method for self-organized AGVs in production logistics systems.
Procedia CIRP 2021, 104, 381–386. [CrossRef]
21. Wang, X.; Luo, X.; Han, B.; Chen, Y.; Liang, G.; Zheng, K. Collision-Free Path Planning Method for Robots Based on an Improved
Rapidly-Exploring Random Tree Algorithm. Appl. Sci. 2020, 10, 1381. [CrossRef]
22. Lee, H.; Jeong, J. Mobile Robot Path Optimization Technique Based on Reinforcement Learning Algorithm in Warehouse
Environment. Appl. Sci. 2021, 11, 1209. [CrossRef]
23. Oboth, C.; Batta, R.; Karwan, M. Dynamic conflict-free routing of automated guided vehicles. Int. J. Prod. Res. 1999, 37, 2003–2030.
[CrossRef]
24. Yuan, Z.; Yang, Z.; Lv, L.; Shi, Y. A Bi-Level Path Planning Algorithm for Multi-AGV Routing Problem. Electronics 2020, 9, 1351.
[CrossRef]
25. Xu, C.; Xu, Z.; Xia, M. Obstacle Avoidance in a Three-Dimensional Dynamic Environment Based on Fuzzy Dynamic Windows.
Appl. Sci. 2021, 11, 504. [CrossRef]
26. Bae, H.; Kim, G.; Kim, J.; Qian, D.; Lee, S. Multi-Robot Path Planning Method Using Reinforcement Learning. Appl. Sci. 2019,
9, 3057. [CrossRef]
27. Bai, Y.; Ding, X.; Hu, D.; Jiang, Y. Research on Dynamic Path Planning of Multi-AGVs Based on Reinforcement Learning. Appl.
Sci. 2022, 12, 8166. [CrossRef]
28. Onoufriou, G.; Bickerton, R.; Pearson, S.; Leontidis, G. Nemesyst: A hybrid parallelism deep learning-based framework applied
for internet of things enabled food retailing refrigeration systems. Comput. Ind. 2019, 113, 103133. [CrossRef]
29. Zhong, M.; Yang, Y.; Zhou, Y.; Postolache, O. Adaptive Autotuning Mathematical Approaches for Integrated Optimization of
Automated Container Terminal. Math. Probl. Eng. 2019, 2019, 7641670. [CrossRef]
30. Fazlollahtabar, H.; Hassanli, S. Hybrid cost and time path planning for multiple autonomous guided vehicles. Appl. Intell. 2018,
48, 482–498. [CrossRef]
31. Miyamoto, T.; Inoue, K. Local and random searches for dispatch and conflict-free routing problem of capacitated AGV systems.
Comput. Ind. Eng. 2016, 91, 1–9. [CrossRef]
32. Desaulniers, G.; Langevin, A.; Riopel, D.; Villeneuve, B. Dispatching and Conflict-Free Routing of Automated Guided Vehicles:
An Exact Approach. Int. J. Flex. Manuf. Syst. 2003, 15, 309–331. [CrossRef]
33. Xing, L.; Liu, Y.; Li, H.; Wu, C.-C.; Lin, W.-C.; Chen, X. A Novel Tabu Search Algorithm for Multi-AGV Routing Problem.
Mathematics 2020, 8, 279. [CrossRef]
34. Liang, C.; Zhang, Y.; Dong, L. A Three Stage Optimal Scheduling Algorithm for AGV Route Planning Considering Collision
Avoidance under Speed Control Strategy. Mathematics 2022, 11, 138. [CrossRef]
35. Liu, Z.; Meyendorf, N.; Mrad, N. The Role of Data Fusion in Predictive Maintenance Using Digital Twin. In Proceedings of
the 44th Annual Review of Progress in Quantitative Nondestructive Evaluation, Provo, UT, USA, 16–21 July 2017; p. 020023.
[CrossRef]
36. Guo, J.; Zhao, N.; Sun, L.; Zhang, S. Modular based flexible digital twin for factory design. J. Ambient. Intell. Humaniz. Comput.
2019, 10, 1189–1200. [CrossRef]
37. DebRoy, T.; Zhang, W.; Turner, J.; Babu, S.S. Building digital twins of 3D printing machines. Scr. Mater. 2017, 135, 119–124.
[CrossRef]
38. Tao, F.; Zhang, M.; Liu, Y.; Nee, A.Y.C. Digital twin driven prognostics and health management for complex equipment. CIRP
Ann. 2018, 67, 169–172. [CrossRef]
39. Bruynseels, K.; Santoni de Sio, F.; van den Hoven, J. Digital Twins in Health Care: Ethical Implications of an Emerging Engineering
Paradigm. Front. Genet. 2018, 9, 31. [CrossRef]
40. Kumar, S.A.P.; Madhumathi, R.; Chelliah, P.R.; Tao, L.; Wang, S. A novel digital twin-centric approach for driver intention
prediction and traffic congestion avoidance. J. Reliab. Intell. Environ. 2018, 4, 199–209. [CrossRef]
41. Zhang, M.; Tao, F.; Nee, A.Y.C. Digital Twin Enhanced Dynamic Job-Shop Scheduling. J. Manuf. Syst. 2021, 58, 146–156. [CrossRef]
42. Wang, Y.; Wu, Z. Model construction of planning and scheduling system based on digital twin. Int. J. Adv. Manuf. Technol. 2020,
109, 2189–2203. [CrossRef]
43. Negri, E.; Ardakani, H.D.; Cattaneo, L.; Singh, J.; Macchi, M.; Lee, J. A Digital Twin-based scheduling framework including
Equipment Health Index and Genetic Algorithms. IFAC-PapersOnLine 2019, 52, 43–48. [CrossRef]
44. Han, W.; Xu, J.; Sun, Z.; Liu, B.; Zhang, K.; Zhang, Z.; Mei, X. Digital Twin-Based Automated Guided Vehicle Scheduling: A
Solution for Its Charging Problems. Appl. Sci. 2022, 12, 3354. [CrossRef]
Mathematics 2023, 11, 2678 25 of 25

45. Wenna, W.; Weili, D.; Changchun, H.; Heng, Z.; Haibing, F.; Yao, Y. A digital twin for 3D path planning of large-span curved-arm
gantry robot. Robot. Comput. Manuf. 2022, 76, 102330. [CrossRef]
46. Zohdi, T.I. A digital twin framework for machine learning optimization of aerial fire fighting and pilot safety. Comput. Methods
Appl. Mech. Eng. 2021, 373, 113446. [CrossRef]
47. Guo, X.; Peng, G.; Meng, Y. A modified Q-learning algorithm for robot path planning in a digital twin assembly system. Int. J.
Adv. Manuf. Technol. 2022, 119, 3951–3961. [CrossRef]
48. Vasanthan, C.; Nguyen, D.T. Combining Supervised Learning and Digital Twin for Autonomous Path-planning. IFAC-
PapersOnLine 2021, 54, 7–15. [CrossRef]
49. Gao, Y.; Chang, D.; Chen, C.-H.; Xu, Z. Design of digital twin applications in automated storage yard scheduling. Adv. Eng.
Inform. 2022, 51, 101477. [CrossRef]
50. Zhang, C.; Zhang, F.; Li, F.; Wu, H. Improved Artificial Fish Swarm Algorithm. In Proceedings of the 2014 9th IEEE Conference
on Industrial Electronics and Applications, Hangzhou, China, 9–11 June 2014; pp. 748–753.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.