Available online at [Link].

com

ScienceDirect
IFAC PapersOnLine 55-10 (2022) 3172–3177
Machine Learning
Machine Learning Models for for Efficient Port
Port
Machine Learning Models Models for Efficient
Efficient Port
Terminal
Machine Operations:
Learning
Terminal ModelsCase of Vessels’
for Efficient Port
Terminal Operations:
Operations: CaseCase of of Vessels’
Vessels’
Arrival
Terminal Times Prediction
Operations:
Arrival Case of Vessels’
Arrival Times
Times Prediction
Prediction
Arrival Times Prediction
Sara El Mekkaoui ∗∗ Loubna Benabbou ∗∗
∗∗ Abdelaziz Berrado ∗
Sara El Mekkaoui ∗ Loubna Benabbou ∗∗ Abdelaziz Berrado ∗∗
Sara El Mekkaoui Loubna Benabbou Abdelaziz Berrado
Sara
∗ El Mekkaoui
Équipe AMIPS,
∗
ÉcoleLoubna Benabbou
Mohammadia
∗∗
Abdelaziz
d’Ingénieurs, MohammedBerrado V
∗
∗
Équipe AMIPS, École Mohammadia d’Ingénieurs, Mohammed V
∗
Équipe AMIPS, University École in Mohammadia
Rabat, d’Ingénieurs, Mohammed V
Équipe University
AMIPS, École Rabat, Rabat,
in Mohammadia Rabat, Morocco
Morocco (e-mail:
(e-mail:
∗
University in Rabat, Rabat, d’Ingénieurs,
saraelmekkaoui@[Link]; Morocco Mohammed V
(e-mail:
berrado@[Link]).
saraelmekkaoui@[Link];
University in Rabat,
∗∗ saraelmekkaoui@[Link]; Rabat, berrado@[Link]).
Morocco (e-mail:
Département Sciences Sciences de de la
la Gestion, berrado@[Link]).
Gestion, Université
Université du du Québec
Québec à à
∗∗
Département
∗∗ saraelmekkaoui@[Link]; berrado@[Link]).
Département
Rimouski, Lévis, Sciences
Qc, Canada de la Gestion,
(e-mail: Université du Québec à
loubna_benabbou@[Link])
Rimouski,
∗∗
DépartementLévis, Qc, Canada
Sciences de la(e-mail:
Gestion, loubna_benabbou@[Link])
Université du Québec à
Rimouski, Lévis, Qc, Canada (e-mail: loubna_benabbou@[Link])
Rimouski, Lévis, Qc, Canada (e-mail: loubna_benabbou@[Link])
Abstract:
Abstract: Port Port terminals
terminals are are critical
critical nodes
nodes in in the
the maritime
maritime transport
transport network
network and and play
play aa
Abstract: role
significant Portinterminals
the global aresupply
critical nodesHowever,
chain. in the maritime
they still transport
suffer from network
many and play a
disruptions
significant
Abstract: role
Port interminals
the global aresupply
criticalchain.
nodes However,
in the they still transport
maritime suffer from many and
network disruptions
play a
significant
entailed by role
their in the global supply chain. However, they still suffer from many disruptions
entailed
significant theirincomplex
by role complex
the
environment
environment
global supply
leading
leading
chain.
to
to many
However,many challenges.
challenges.
they still
With
With
suffer from
the maritime
themany
maritime digital
digital
disruptions
entailed by their
transformation, ports complex
ports and environment
and ships
ships produce leading
produce significant to many
significant amounts challenges.
amounts of of data With
data offering the
offering an maritime
an opportunity digital
opportunity to to
transformation,
entailed by their complex environment leading to many challenges.
transformation,
use Machine ports
Learning and ships
techniques produce
to significant
address some amounts
issues and of data With
support offering
port theanmaritime
opportunity
terminal digital
operations to
use Machine
transformation, Learning
ports and techniques to
ships produce address some
significant issues
amounts and support port terminal operations
use Machine Learning
management. This techniques to address some issues and of data
support offering
port an opportunity
terminal operations to
management.
use Machine This paper
Learning paper addresses
addressestothe
techniques the problem
problem
address some
of
of vessel
vessel
issues
arrival
arrival
and
times prediction
times port
support prediction to
to destination
terminal destination
operations
management.
ports using This
using Machine paper
Machine Learning addresses
Learning models the problem
models and of
and vessels’ vessel arrival
vessels’ historical times
historical trajectories prediction
trajectories data. to destination
data. This
This paper
paper
ports
management. This paper addresses the problem of vessel arrival times prediction to This
destination
portsprovides
also using Machine
a structured Learning
overviewmodelsof and vessels’
research work historical
concerning trajectories
the data.
contribution of paper
Machine
also
ports provides
using a structured
Machine overview
Learning modelsof research
and vessels’work concerning
historical the contribution
trajectories of
data. This Machine
paper
also provides
Learning a structured
techniques in overview
handling portof terminal
research work
[Link]
The the contribution
existing literature of Machine
shows that
Learning
also provides techniques
a structuredin handling
overview portof terminal
research concerns.
work The existing
concerning literature of
the iscontribution shows that
Machine
Learning
related worktechniques
has tackledin handling
different port terminal
problems, but concerns.
further The existing
development literature shows
needed. that
related
Learning work has
techniques tackled different
in handling problems,
port terminal but further
concerns. development
The existing is needed.
literature shows that
related work has tackled different problems, but further development is needed.
related
Copyright work has tackled different problems, but further development is needed.
Copyright ©©
Copyright ©20222022
2022TheIFACAuthors. This is an open access article under the CC BY-NC-ND license
IFAC
Copyright © 2022 IFAC /licenses/by-nc-nd/4.0/)
([Link]
Copyright © 2022 IFAC
Keywords: Machine
Keywords: Machine Learning,
Learning, Port Port Terminals,
Terminals, Automatic
Automatic Identification
Identification System,
System, Estimated
Estimated
Keywords:
Time of Machine
Arrival, Learning,
Intelligent Port Terminals,
transportation Automatic
systems, Identification
Transportation System, Estimated
logistics
Time of
Keywords: Arrival,
Machine Intelligent
Learning, transportation
Port Terminals, systems, Transportation
Automatic Identificationlogistics
System, Estimated
Time of Arrival, Intelligent transportation systems, Transportation logistics
Time of Arrival, Intelligent transportation systems, Transportation logistics
1.
1. INTRODUCTION
INTRODUCTION management and
management and using
using multiple
multiple datadata sources.
sources. One One ofof the
the
1. INTRODUCTION management
promising data dataand using in
sources multiple
maritimedataresearch
[Link] the of the
Auto-
1. INTRODUCTION promising
management and sources
using in maritime
multiple data research
sources. is the
One Auto-
of the
promising data
matic Identification
Identificationsources
Systemin maritime
(AIS), research
which is is the track-
Auto-
The maritime transportation system is the backbone of matic
promising data System
sources in (AIS),
maritime which is aa is
research
vessel
vessel
the track-
Auto-
The maritime transportation system is the backbone of ing matic
ing systemIdentification
system providing System
providing frequent (AIS),
frequent updates which
updates on is a
on vessel vessel
vessel movementtrack-
movement
The maritimetrade
international transportation
as 80% of system
global istrade
the backbone
volumes of matic
are ing Identification
system providing System
frequent (AIS),
updateswhich is a vessel
onother
vessel track-
movement
international
The maritime trade as
transportation80% of global
system trade
istrade
the volumes
backbone are
of ing and
and other
other relevant
relevant ship
ship voyage
voyage data
data to
to other parties
parties on aa
on
international
carried by trade
seas. Ports as play
80%a ofcritical
globalrole in volumes
the maritimeare and system providing
otherbasis.
relevant frequent
ship voyage updates
data toonother
vessel movement
parties onre-a
carried by seas. Ports play a critical role in the maritime real-time
real-time basis. Information
Information can
can be
be exchanged
exchanged between
between
international
carried by system
transport
transport
trade
seas.
system
Ports
by as play
80%a ofcritical
by handling
handling
globalrole
intermodal
intermodal
trade volumes
intransport
the maritime
transport
are real-time
con- and
con- ceivingother relevant
basis.
stations ship voyage
Information
onboard data
canalong
ships, to otherand
be exchanged
coasts, parties
between onre-
re-a
by satel-
satel-
carried by seas.
transportassystem
nections Ports
linking play
by nodes.
handling a critical
The role
intermodal
complexity in the maritime
transport
of the ceiving
con- real-time
port stations
basis. onboard
Information ships,
can along
be coasts,
exchanged and by
between re-
nections assystem linking ceiving
lite. AIS stations
AIS data
data is onboard
is intended
intended for ships, along
safety and coasts,
security and by satel-
navigation
transport
nectionsand
system as linking
the by nodes.
handling
nodes.
increasing
Theintermodal
The
traffic
complexity
complexity
make the
of the port
transport lite.
con- ceiving
ofoperations
the port lite. AISstations
data onboard
ishave
intended
for safety
ships,
for along
safety
andcoasts,
and
security
security and navigation
by satel-
navigation
system
nections and
as the
linkingincreasing
nodes. traffic
The make
complexity the ofoperations
the port purposes,
purposes, but
but have been
been collected
collected by
by governmental
governmental and
and
system
management and the
more increasing
and more traffic make
challenging the
and operations
can turn lite. AISorganizations
purposes, data
butishave
intended
been for safetyinand
collected by security
governmentalnavigation
and
management
system and the more and
increasing more challenging
traffic make and
the can turn private
private organizations and
and served
served in different
different case
case studies
studies
management more and more challenging andoperations
can turn
ports
ports into
into transshipment
management transshipment
more and more
bottlenecks.
bottlenecks. Improving
Improving
challenging and can
port per- purposes,
per-
port turn private
(Svanberg
(Svanberg
but [Link]
organizations
et al.
et
been
(2019)).
(2019)).
and collected
served in by governmental
different and
case studies
ports into can
formance transshipment
be achieved bottlenecks.
by advanced Improving
analytics port per- private
solutions (Svanberg organizations
et al. (2019)).and served in different case studies
formance
ports into can be achievedbottlenecks.
transshipment by advancedImproving analytics portsolutions
per- (Svanberg
formance
to analyze can be achieved
different factors by influencing
advanced analyticstheir solutions
operational Using AIS
Using AIS et data, we address
address one
al. (2019)).
data, we one ofof the
the port
port terminal
terminal op- op-
to analyze
formance candifferent
be factors
achieved by influencing
advanced their
analytics operational
solutions Using AIS data, we address one of the port terminal op-
to analyze
performance. different factors influencing their operational erations
erations planning
planning and scheduling
andaddress
scheduling issues
issues related
related to vessels’
to vessels’
performance.
analyze different factors influencing their operational Using AIS data,
to
performance. erations
Estimated
Estimated Time we
planning
Time
and
of
of Arrivals
Arrivals
one
scheduling of
(ETA).
(ETA).
the
In port
issues
In fact, terminal
related
fact, ship
ship
op-
to vessels’
arrival
arrival
The literature
performance. related to simulation
The literature related to simulation and optimization times and optimization erations
Estimated planning
times variations Time and
of
variations affect scheduling
Arrivals
affect the (ETA).issues
the performance In
performance of related
fact, to
ship
of port vessels’
arrival
port termi-
termi-
The literature
methods for portrelated
terminal to operations
simulation has andbeen optimization
abundant Estimated
times TimeWe
variations of Arrivals
affect the (ETA).
performance In different
fact, shipMachine
of port arrival
termi-
methods
The for
literature port terminal operations has been abundant nal
nal operations.
operations. We use
use in
in this
this paper
paper different Machine
methods portrelated
for (2013, terminal to operations
simulation hasandbeen optimization
abundant
(Carlo
(Carlo et
methods et al.
al. (2013, 2014a,b);
2014a,b); Bierwirth
Bierwirthhas and
andbeenMeisel
Meisel (2015); times
(2015); nal
Learning
Learning
variations
operations.
methods,
methods,
Weaffect
use the
including
including
performance
in this
Neural
Neural
paper of port
different
Networks,
Networks,
termi-
Machine
Tree-based
Tree-based
(Carlo et for
Dragovic al.
et port(2017)).
(2013,
al. terminal
2014a,b); operations
Along Bierwirth
with theanddeveloped
Meiselabundant
(2015);
opti- nal operations.
Learning methods, We use
includingin this
Neuralpaper different
Networks, Machine
Tree-based
Dragovic
(Carlo et (2013,
et and
al. al. (2017)).
2014a,b); Along with the
Bierwirth anddeveloped methods and
opti- methods
Meisel (2015); and Support
Support VectorVector Regression,
Regression, together
together withwith
Dragovic
mization et al. (2017)).
simulation Along
models, with the
Machine developed
Learning opti- Learning
tech- methods methods,
andpredict
Supportincluding
Vector Neural Networks,
Regression, Tree-based
together with
mization
Dragovic and simulation
et help
al. (2017)). models,
Along Machine
with the Learning
developed tech-
opti- AIS
AIS data
data to
to predict vessels’
vessels’ ETA.
ETA. For
For that,
that, we
we explore
explore the
the
mization and simulation models, Machine Learning tech-
niques
niques can
mization canand help analyze the
the impact
analyze models,
simulation impact of
of some
Machine some factors
factors
Learning on methods
on
tech-
AIS
case data
case
andpredict
to
of multiple
of multiple Support
vessels
vessels
Vector
vessels’ ETA.
travelling
travelling
Regression,
For
between
between
together
that, different
we explore
different
with
the
ports.
ports.
niques can helpperformance.
analyze theMoreover,impact ofthe some factors on
port
port terminals
niquesterminals
can help performance.
analyze theMoreover,
impact ofthe some
digitization
digitization
factors of AIS
of
on
case
Using
Using
data
ofonly
only
to apredict
multiple
a limited
vessels’
vessels
limited set of ETA.
settravelling
of
For
variables
variables
that,
between
from
we
AISexplore
fromdifferent
AIS data
data
the
ports.
and
and
portmaritime
the terminals performance.
sector generates Moreover,
voluminous theanddigitization
useful data,of caseUsing of multiple
only a vessels
limited settravelling
of between
variables from different
AIS dataports.
and
the maritime
port terminals sector generatesMoreover,
performance. voluminous theand useful data,
digitization given aa vessel
given
of Using vessel position,
position, we we try
try to
to predict
predict thethe transit
transit time
time
the maritime
which can sector
support thegenerates
adoption voluminous
of Machine and useful based
Learning data, given aonly
vessela limited
position,setweoftry variables
to predictfromthe AIS data time
transit and
which
the can
maritime support
sector the adoption
generates of Machine
voluminous Learning
and useful based
data, given to
to destination.
destination.
which can support
solutions. This canthe adoption
serve in of Machine
building Learning
robust based
predictive a vessel position, we try to predict the transit time
to destination.
solutions.
which This
can support canthe serve
adoptionin building
of parameters
Machine robust
Learningpredictive
based
solutions. This can serve in building robust predictive Thedestination.
next section
section of of this
this paper
paper provides
provides an an overview
overview of of
tools,
tools, and help
help to
and This to control
control some
some parameters variation
variation to to
to The
The
next
next section of this
solutions.
tools,
assist and
port help can
to
managers serve
control
in the in
some building
parameters
decision-making robust predictive
variation
process. to the the existing
existing literature
literature onpaper
on Machine
Machine
provides an overview
Learning
Learning techniques
techniques
of
assist port
tools, and managers
help to controlin the some
decision-making
parametersprocess. variation to The the next section
existing of this
literature paper
onoperations
Machineprovides an overview
Learning techniques of
assist port managers in the decision-making process. to address
to address port
port terminal
terminal operations challenges.
challenges. The third
The third
The
assistexisting applications
port managers
The existing applicationsin the of Machine
of decision-makingLearning
Machine Learning techniques
process.
techniques section the
to existing
address
section discusses literature
port terminal
discusses some on
some research Machine
operations Learning
challenges.
research directions.
directions. Section techniques
The
Section four third
four isis
The port
for existing applications
terminal operations of Machine
concern Learning
different techniques
areas of to address
section portstudy
discusses terminal
some operations
research challenges.
directions. SectionThefour
thirdis
for
The port terminal
existing operations
applications concern
of Machine different
Learning areas of about
about the
the case
case study of
of predicting
predicting ship
ship arrival
arrival times,
times, giving
giving
for port
research terminal
ranging operations
from the concern differenttechniques
areas of section discusses some research directions. Section
times, four is
research
for port ranging
terminalfrom the seaside
operations seaside to
to land
concern land gates
gates operations
operations about the case study of predicting ship arrival giving
research ranging from the seaside to landdifferent areas of about the case study of predicting ship arrival times, giving
gates operations
research
2405-8963 ranging
Copyrightfrom© 2022 the seaside
The Authors. toThis
landis gates
an openoperations
access article under the CC BY-NC-ND license.
Peer review under responsibility of International Federation of Automatic Control.
10.1016/[Link].2022.10.217
 Sara El Mekkaoui et al. / IFAC PapersOnLine 55-10 (2022) 3172–3177 3173

more details about AIS data, and reporting the study Table 1. Classification of the research areas and
results. Finally, the last section concludes our paper. topic
Research Research topic References
2. MACHINE LEARNING MODELS FOR PORT area
TERMINAL OPERATIONS Operating Customs Hoshino et al. (2011)
methods
2.1 Literature review Vessels traffic fore- Wang et al. (2017)
casting Zhang et al. (2019)
Wang et al. (2020)
Machine learning algorithms have been used to deal with
different port terminal issues. We review in this section, Cargo throughput Gosasang et al. (2011)
related works to provide an overview of this research area. forecasting Xiao et al. (2014)
We extracted 37 studies addressing port terminal chal- Geng et al. (2015)
lenges with Machine Learning techniques from different Moscoso-López et al. (2016)
databases including Taylor & Francis, Springer, Science Chan and Xu (2017)
Direct, SAGE, ASCE, IEEE Xplore, and JSTOR. The Milenkovic et al. (2021)
Baochai (2020)
evolution of this research field is depicted in Figure 1,
where we can see that research has been active since 2006. Seaside Berth allocation de León et al. (2017)
area
Quay crane opera- Jakovlev et al. (2021)
tions Bandong et al. (2021)

Port congestion Abualhaol et al. (2018)

Vessel arrival time Fancello et al. (2011)

Kim et al. (2017)
Yu et al. (2018)

Efficiency metrics Gómez et al. (2016)

Sutrisnowati et al. (2015)
Fig. 1. Number of papers per year Grida and Lee (2018)
Štepec et al. (2020)
We established a classification scheme of the extracted Wang et al. (2020)
studies research areas and topics to understand the chal- Martinčič et al. (2020)
Wang et al. (2021)
lenges addressed by Machine Learning to improve port
terminal operations efficiency. As proposed in Table 1, Storage Containers Kang et al. (2006)
related works are classified as belonging to four research yard area reshuffling
areas, which are operating methods, seaside area, storage
yard area, and transport area. For each research area, we Yard crane schedul- Fotuhi et al. (2013)
specify a topic reflecting the problem addressed by the ing
papers. Operating methods include studies on activities
and practices needed for operations execution within port Efficiency metrics Sutrisnowati et al. (2015)
terminals. The studies belonging to this area employ Ma- Moini et al. (2012)
Kourounioti et al. (2016)
chine Learning methods to support customs containers
Maldonado et al. (2019))
control by predicting whether a container has been fu-
migated and improve operations planning by forecasting Transport Vehicle routing and Zeng et al. (2009)
vessels’ traffic and cargo throughput. The seaside area area dispatching
related papers apply Machine Learning methods to select
the best algorithm for berth allocation based on historical Inland transporta- van Riessen et al. (2016)
performance, improve quay crane operations by predict- tion attribution
ing the spreader speed, predict port congestion, estimate
vessels’ arrival times, and forecast efficiency metrics of Truck traffic fore- Xie and Huynh (2010)
casting Chu (2011)
vessels’ service and turnaround times. The storage yard
Hill and Böse (2017)
research area includes studies using Machine Learning to
improve containers reshuffling by predicting their weights, Efficiency metrics Sutrisnowati et al. (2015)
increase terminal capacity by predicting containers dwell Huynh and Hutson (2008)
and handling time, reduce truck waiting times by solving Pham et al. (2011)
the yard cranes scheduling problem, and improve termi- Wasesa et al. (2017)
nals capacities by predicting efficiency metrics related to Hill and Böse (2017)
containers dwell and handling times. Finally, the transport
area includes studies to solve the yard trailers scheduling Various Machine Learning techniques were applied to deal
problem, choose the optimal inland transportation mean, with port terminal operations challenges as shown in Fig-
forecast truck traffic and turnaround time to reduce road ure 2. A single paper may use different techniques and
and gate congestion, and improve transportation reliabil- compare their performances. The most employed tech-
ity. niques are Neural Networks, Decision Trees, Linear Meth-
3174 Sara El Mekkaoui et al. / IFAC PapersOnLine 55-10 (2022) 3172–3177

3. CASE STUDY: PREDICTING SHIP ARRIVAL

TIMES

3.1 Problem description

Dealing with arriving ships at port terminals involves

several decision-making problems (Bierwirth and Meisel
(2015)). These problems are interconnected and form a
logistic process starting with ship arrival times. They
are solved using optimization methods but usually make
assumptions regarding the arrival times and ignore the
variations of ETA provided by ships. For this reason, we
need solutions to predict ship arrival times to improve port
terminal operations planning and scheduling. This could
help minimize ship turnaround times, avoid unoccupied
Fig. 2. Frequency of Machine Learning techniques berths, reduce vehicle waiting time and decrease energy
consumption and emissions. Some studies tried to predict
ship arrival times to destination using Machine Learning
techniques (Fancello et al. (2011); Kim et al. (2017); Yu
ods, and Support Vector Machines. Neural Networks meth- et al. (2018)). The study by Fancello et al. (2011) is consid-
ods can deal with large unstructured data and learn non- ered to be the first work using Machine Learning models
linear and complex relationships. Decision Trees, though, to determine ship arrival times to a container terminal,
exhibit better interpretability and can be simple and easy proposing a Neural Network for short-term forecasting to
to understand. Most of the problems were formulated as optimize human resource allocation. Kim et al. (2017) ex-
supervised learning tasks, while few papers dealt with plored AIS data and information extracted from container
unsupervised learning and reinforcement learning tasks. shipping documents to predict early and real-time delays
Multiple data sources were employed, for example, port of ships using a modified Case-Based Reasoning method.
terminal systems, port authorities, border services, truck Yu et al. (2018) discretized ships’ delays into intervals
and logistic companies, and AIS. and considered a classification problem. Using data from a
container terminal, they applied Random Forests, Neural
Networks, and Classification And Regression Trees and
used the predicted arrival times in an optimization model
to solve the berth allocation and quay crane assignment
2.2 Opportunities for Machine Learning methods to improve problem. We address in this study the problem of predict-
port terminal operations ing ship arrival times to destination based on AIS historical
trajectories data provided by MarineTraffic, a popular
Ports are complex systems playing a crucial role in the online ship tracking solution with a large network of AIS
global supply chain. Understanding and solving port ter- receivers. We use Machine Learning techniques to predict
minal challenges improve the interface between sea and the travel times to the destination given information about
land transportation. The maritime digital transformation ship’s position, speed, and course, as provided by AIS data.
has made available various data sources (Heilig and Voß
(2017)), which can support the development of Machine 3.2 Material and methods
Learning based solutions for several port terminal op-
erational issues. In that respect, the existing literature The AIS is a vessel tracking system broadcasting ships
addressed different port terminal challenges but still can be locations and other information using a radio channel.
extended to other cases to provide accurate information to AIS data is globally used by harbors and port authorities
port actors and improve the decision-making process. For through Vessel Traffic Services (VTS), which is a real-time
instance, further research can be undertaken to deal with control system similar to air traffic control, to improve
uncertainty in scheduling floating equipment (tugboat, the safety and efficiency of vessel traffic in related waters.
pilot, etc.), warehouse management, environmental condi- The increasing relevance of AIS data in addressing many
tions (weather, tides, etc.), navigation channel occupancy, problems has motivated many commercial companies to
anchorage congestion, and cargo inspection, and more. It build a large global AIS network. The network is composed
would also be important to investigate the integration of of coastal receivers enhanced by satellites for a better
Machine Learning techniques and optimization algorithms coverage.
in the port context to build models closer to reality. More-
over, related studies focused on container terminals, while AIS information is classified into four categories which are
other terminal types such as bulk, passengers, and general static, dynamic, voyage, and safety messages (Bole et al.
cargo require some attention. For instance, bulk terminals (2013)). Static information is about the ship’s characteris-
suffer from many disruptions engendered by events such tics such as identification number, length, beam, and ship
as last-time charters, ship substitutions, and vessel avail- type. Voyage-related data include details about the current
ability. It would also be insightful to explore new regions voyage such as draught, ETA, and destination. Dynamic
such as Africa and the Middle East, as the emphasis has messages report ships’ movements including position (lat-
been on European, Asian, and American ports. itude and longitude), course over ground, and speed over
 Sara El Mekkaoui et al. / IFAC PapersOnLine 55-10 (2022) 3172–3177 3175

ground. Finally, the safety text message can be sent when data normalization, discretization, and feature engineering
required. (Kotsiantis et al. (2007). We first removed duplicates and
idle vessels to only keep messages reflecting port-to-port
Although AIS is meant to be a safety and security tool, the voyages, each of them being a trip of a single ship from
system provides important information on ship’s motions, an origin port to a destination port. The new dataset is
which is useful in analyzing maritime traffic patterns. AIS composed of 13,557 rows and 9 features, comprising 15
data have been well explored in studying shipping motion different vessels and 17 voyages. Figure 4 represents the
patterns for trajectories extraction and estimation, and final dataset positions plot.
anomaly detection (Svanberg et al. (2019)). The objective
is to deal with safety and security issues, especially in busy
areas near ports, straights, and inland waterways. The use
of AIS is not limited to that, but it can also contribute to
improving ship routing and scheduling and building good
systems for the prediction of arrival times. However, AIS
data can have some quality issues (Harati-Mokhtari et al.
(2007)).
Figure 3 represents the conceptual framework of this study.
We start with a preprocessing step to inspect data qual-
ity and prepare the predictive features. We use as input
features the ship’s position (longitude and latitude), speed
over ground, course, heading, and origin and destination Fig. 4. Visualization of dataset positions over the Mediter-
ports coordinates and predict the transit time to destina- ranean Sea
tion in hours. The ship’s ETA is then calculated by adding
the travel time to the position timestamp. The prepared For the Neural Network (NN), we use a fully connected
dataset is used to train and test different Machine Learn- feedforward network implemented using the Python pack-
ing models, including Neural Networks (Bishop (1995)), age Keras (Chollet et al. (2015)), for which we selected
Support Vector Regression (Smola and Schölkopf (2004)), the best hyperparameters using a random search over the
Random Forests (Breiman (2004)), Gradient Boosting number of layers, the number of neurons per layer, and
(Friedman (2001)) and Ridge Regression (James et al. the learning rate. The Support Vector Regression (SVR),
(2021)). A Machine Learning model has parameters that Random Forests (RF), Gradient Boosting (GB), and Mul-
control its behavior. The model learns the parameters tiple Ridge Regression (RR) models were performed us-
that minimize the training error using the training set. ing Scikit-Learn (Pedregosa et al. (2011)), with a grid
The validation set is used to tune the hyperparameters by search to select their hyperparameters. Table 2 reports
estimating the generalization error during or after training. the performance of the different models measured by the
The test set is a separate set from the training and valida- Mean Absolute Error (MAE) and the Mean Squared Error
tion sets, used to evaluate the Machine Learning algorithm (MSE) on the test set.
on data not seen before. We use 80% for training and 10% Table 2. Comparison of the Machine Learning
for validation and 10% for the test. models performance on the test set
NN SVR RF GP RR
MAE 15.6 17.2 23.4 27.2 18.2
MSE 356.8 730.9 889.7 862.3 534.9

The results show that the Neural Network model perform

better than the other models. However, we still need to
further lower the prediction error for the NN or SVM
models to be practical. We can improve the models’
performance by considering a larger dataset and including
more factors influencing ships’ behavior such as weather
Fig. 3. General approach for the prediction of ships’ ETA data or additional information related to ships or their
using Machine Learning models voyages, which we plan to explore in future research.

3.3 Experimental results 4. CONCLUSION

We use in this study a dataset composed of more than
Today, port terminals can benefit from the maritime data
180,000 rows of AIS data as collected by a coastal network.
sources made available thanks to the maritime sector
The provided features are ship identifier, speed, ship’s po-
digitization. Machine Learning techniques can particularly
sition given by the longitude and latitude, course, heading,
serve to gain insights from such data and develop tools to
timestamp, and name of the last visited port. The coverage
improve port terminal operations efficiency.
area is the Mediterranean Sea for the period from 1th April
till 28th April 2015. As mentioned before, a preprocessing In this study, we address the problem of ships’ ETA pre-
step is required to make the data suitable for the regression diction, which is an important component in port terminal
task. Data preprocessing operations include techniques operations planning and scheduling. We consider the case
such as outliers’ detection, dealing with missing values, of vessels’ voyages between different origin and destination
3176 Sara El Mekkaoui et al. / IFAC PapersOnLine 55-10 (2022) 3172–3177

ports. We use information from the AIS about vessels’ his- Carlo, H.J., Vis, I.F., and Roodbergen, K.J. (2014a). Stor-
torical trajectories and movement to build Machine Learn- age yard operations in container terminals: Literature
ing models and compare different techniques. However, it overview, trends, and research directions. European
is challenging to build reliable ships’ ETA forecasting so- Journal of Operational Research, 235(2), 412–430. doi:
lutions as their activity is subject to many considerations. 10.1016/[Link].2013.10.054.
The study is based on a limited set of AIS information and Carlo, H.J., Vis, I.F., and Roodbergen, K.J. (2014b).
could be further developed by including more information Transport operations in container terminals: Literature
on ships and their voyages and considering a larger dataset. overview, trends, research directions and classification
Accurate predictions of ships’ ETA can improve berths scheme. European Journal of Operational Research,
allocation, equipment and resources utilization, and the 236(1), 1–13. doi:10.1016/[Link].2013.11.023.
interface between sea and land transportation. This can Chan, H.K. and Xu, S. (2017). Intelligent port data
enhance operational efficiency and also reduce the envi- management systems to improve capability. In 2017 In-
ronmental impact of shipping. ternational Conference on Service Systems and Service
Management, 1–3. doi:10.1109/ICSSSM.2017.7996283.
We also conduct in this paper a literature review to un-
Chollet, F. et al. (2015). Keras.
derstand how Machine Learning techniques are adopted
[Link]
to leverage port terminals’ operational efficiency. Previ-
Chu, H.C. (2011). Empirical method for predicting
ous studies have addressed different port terminal issues
internal-external truck trips at a major port. Journal
related to operating methods, seaside, storage yard, and
of Transportation Engineering, 137, 496–508.
transport areas, but further development is still needed
de León, A.D., Lalla-Ruiz, E., Melián-Batista, B., and
to tackle more challenges. It will be important that fu-
Moreno-Vega, J.M. (2017). A machine learning-based
ture research expand the application of Machine Learning
system for berth scheduling at bulk terminals. Expert
techniques to deal with seaside area challenges such as
Systems with Applications, 87, 170–182.
improving pilotage and towage services, dredging plan-
Dragovic, B., Tzannatos, E., and Park, N.K. (2017). Simu-
ning, and marine weather forecasting. As the literature
lation modelling in ports and container terminals: litera-
has been focusing on the container terminal case, future
ture overview and analysis by research field, application
studies might also address other port terminal types.
area and tool. Flexible Services and Manufacturing
ACKNOWLEDGEMENTS Journal, 29, 4–34.
Fancello, G., Pani, C., Pisano, M., Serra, P., Zuddas,
The authors would like to thank MarineTraffic for provid- P., and Fadda, P. (2011). Prediction of arrival times
ing the dataset used in this study. and human resources allocation for container terminal.
Maritime Economics & Logistics, 13, 142–173.
REFERENCES Fotuhi, F., Huynh, N.N., Vidal, J.M., and Xie, Y. (2013).
Modeling yard crane operators as reinforcement learning
Abualhaol, I., Falcon, R., Abielmona, R.S., and Petriu, agents. Research in Transportation Economics, 42, 3–12.
E.M. (2018). Mining port congestion indicators from Friedman, J.H. (2001). Greedy function approximation:
big AIS data. 2018 International Joint Conference on A gradient boosting machine. Annals of Statistics, 29,
Neural Networks (IJCNN), 1–8. 1189–1232.
Bandong, S., Nazaruddin, Y.Y., and Joelianto, E. (2021). Geng, J., Li, M.W., Dong, Z.H., and Liao, Y.S. (2015).
Container detection system using CNN based object Port throughput forecasting by mars-rsvr with chaotic
detectors. 2021 International Conference on Instrumen- simulated annealing particle swarm optimization algo-
tation, Control, and Automation (ICA), 106–111. rithm. Neurocomputing, 147, 239–250.
Baochai, D. (2020). Research on prediction of port cargo Gómez, R., Camarero, A., and Molina, R. (2016). De-
throughput based on PCA-BP neural network combi- velopment of a vessel-performance forecasting system:
nation model. 2020 5th International Conference on Methodological framework and case study. Journal of
Information Science, Computer Technology and Trans- Waterway Port Coastal and Ocean Engineering-asce,
portation (ISCTT), 518–523. 142, 04015016.
Bierwirth, C. and Meisel, F. (2015). A follow-up sur- Gosasang, V., Chandraprakaikul, W., and Kiattisin, S.
vey of berth allocation and quay crane scheduling (2011). A comparison of traditional and neural net-
problems in container terminals. European Jour- works forecasting techniques for container throughput
nal of Operational Research, 244(3), 675–689. doi: at bangkok port. The Asian Journal of Shipping and
10.1016/[Link].2014.12.030. Logistics, 27, 463–482.
Bishop, C.M. (1995). Neural networks for pattern recog- Grida, M. and Lee, C.Y. (2018). An empirical model for
nition. estimating berth and sailing times of mega container
Bole, A.G., Wall, A., and Norris, A. (2013). Automatic ships. Maritime Policy & Management, 45, 1078 – 1093.
identification system (AIS). In Radar and ARPA Man- Harati-Mokhtari, A., Wall, A., Brooks, P., and Wang,
ual, 255–275. Elsevier. J. (2007). Automatic identification system (AIS):
Breiman, L. (2004). Random forests. Machine Learning, Data reliability and human error implications.
45, 5–32. Journal of Navigation, 60(3), 373–389. doi:
Carlo, H.J., Vis, I.F.A., and Roodbergen, K.J. (2013). 10.1017/S0373463307004298.
Seaside operations in container terminals: literature Heilig, L. and Voß, S. (2017). Information systems in
overview, trends, and research directions. Flexible Ser- seaports: a categorization and overview. Information
vices and Manufacturing Journal, 27(2), 224–262. doi: Technology and Management, 18(3), 179–201. doi:
10.1007/s10696-013-9178-3.
 Sara El Mekkaoui et al. / IFAC PapersOnLine 55-10 (2022) 3172–3177 3177

10.1007/s10799-016-0269-1. Cournapeau, D., Brucher, M., Perrot, M., and Duches-

Hill, A. and Böse, J.W. (2017). A decision support system nay, E. (2011). Scikit-learn: Machine learning in Python.
for improved resource planning and truck routing at lo- Journal of Machine Learning Research, 12, 2825–2830.
gistic nodes. Information Technology and Management, Pham, Q., Huynh, N.N., and Xie, Y. (2011). Estimating
18, 241–251. truck queuing time at marine terminal gates. Trans-
Hoshino, R., Coughtrey, D., Sivaraja, S., Volnyansky, I., portation Research Record, 2222, 43 – 53.
Auer, S., and Trichtchenko, A. (2011). Applications and Smola, A. and Schölkopf, B. (2004). A tutorial on support
extensions of cost curves to marine container inspection. vector regression. Statistics and Computing, 14, 199–
Annals of Operations Research, 187(1), 159–183. doi: 222.
10.1007/s10479-009-0669-2. Sutrisnowati, R.A., Bae, H., and Song, M. (2015).
Huynh, N.N. and Hutson, N. (2008). Mining the sources Bayesian network construction from event log for late-
of delay for dray trucks at container terminals. Trans- ness analysis in port logistics. Computers & Industrial
portation Research Record, 2066, 41 – 49. Engineering, 89, 53–66.
Jakovlev, S., Eglynas, T., and Voznák, M. (2021). Ap- Svanberg, M., Santén, V., Hörteborn, A., Holm, H., and
plication of neural network predictive control methods Finnsgård, C. (2019). AIS in maritime research. Marine
to solve the shipping container sway control problem in Policy.
quay cranes. IEEE Access, 9, 78253–78265. van Riessen, B., Negenborn, R.R., and Dekker, R. (2016).
James, G., Witten, D., Hastie, T., and Tibshirani, R. Real-time container transport planning with decision
(2021). An Introduction to Statistical Learning : with trees based on offline obtained optimal solutions. De-
Applications in R. Springer, New York. cision Support Systems, 89, 1–16.
Kang, J., Ryu, K.R., and Kim, K.H. (2006). Deriving Štepec, D., Martinčič, T., Klein, F., Vladušič, D., and
stacking strategies for export containers with uncertain Costa, J.P. (2020). Machine learning based system for
weight information. Journal of Intelligent Manufactur- vessel turnaround time prediction. 2020 21st IEEE
ing, 17, 399–410. International Conference on Mobile Data Management
Kim, S., Kim, H., and Park, Y. (2017). Early detection (MDM), 258–263.
of vessel delays using combined historical and real- Wang, C., Zhang, X., Chen, X., Li, R., and Li, G. (2017).
time information. Journal of the Operational Research Vessel traffic flow forecasting based on BP neural net-
Society, 68, 182–191. work and residual analysis. 2017 4th International Con-
Kotsiantis, S.B., Kanellopoulos, D.N., and Pintelas, P.E. ference on Information, Cybernetics and Computational
(2007). Data preprocessing for supervised leaning. Social Systems (ICCSS), 350–354.
World Academy of Science, Engineering and Tech- Wang, N., Cao, J., Huang, H., Ding, Y., and Ling, Q.
nology, International Journal of Computer, Electrical, (2021). A strategy to optimize the timing of prediction
Automation, Control and Information Engineering, 1, information release for container ships’ operation pro-
4104–4109. cess. 2021 8th International Conference on Behavioral
Kourounioti, I., Polydoropoulou, A., and Tsiklidis, C. and Social Computing (BESC), 1–5.
(2016). Development of models predicting dwell time Wang, N., Shen, S., Cao, J., Ding, Y., and Xiao, Y. (2020).
of import containers in port container terminals – an A system for container terminal operation prediction.
artificial neural networks application. Transportation 2020 IEEE International Conference on Progress in
Research Procedia, 14, 243–252. Informatics and Computing (PIC), 407–411.
Maldonado, S., González-Ramírez, R.G., Quijada, F., and Wasesa, M., Stam, A.W., and van Heck, E. (2017). The
Nafarrate, A.R. (2019). Analytics meets port logistics: seaport service rate prediction system: Using drayage
A decision support system for container stacking oper- truck trajectory data to predict seaport service rates.
ations. Decision Support Systems, 121, 84–93. Decision Support Systems, 95, 37–48.
Martinčič, T., Štepec, D., Costa, J.P., Čagran, K., and Xiao, J., Xiao, Y., Fu, J., and Lai, K.K. (2014). A transfer
Chaldeakis, A. (2020). Vessel and port efficiency met- forecasting model for container throughput guided by
rics through validated AIS data. Global Oceans 2020: discrete PSO. Journal of Systems Science and Com-
Singapore – U.S. Gulf Coast, 1–6. plexity, 27, 181–192.
Milenkovic, M., Milosavljević, N., Bojovic, N.J., and Val, Xie, Y. and Huynh, N.N. (2010). Kernel-based machine
S. (2021). Container flow forecasting through neural net- learning models for predicting daily truck volume at sea-
works based on metaheuristics. Operational Research, port terminals. Journal of Transportation Engineering,
1–33. 136, 1145–1152.
Moini, N., Boilé, M., Theofanis, S., and Laventhal, W. Yu, J., Tang, G., Song, X., Yu, X.H., Qi, Y., Li, D.,
(2012). Estimating the determinant factors of container and Zhang, Y. (2018). Ship arrival prediction and its
dwell times at seaports. Maritime Economics & Logis- value on daily container terminal operation. Ocean
tics, 14, 162–177. Engineering.
Moscoso-López, J.A., Turias, I.J., Come, M.J.J., Ruiz- Zeng, Q., Yang, Z., and Lai, L. (2009). Models and
Aguilar, J.J., and del Mar Cerbán, M. (2016). Short- algorithms for multi-crane oriented scheduling method
term forecasting of intermodal freight using ANNs and in container terminals. Transport Policy, 16, 271–278.
SVR: Case of the port of algeciras bay. Transportation Zhang, Z.G., Yin, J., Wang, N.N., and Hui, Z. (2019).
Research Procedia, 18, 108–114. Vessel traffic flow analysis and prediction by an im-
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., proved PSO-BP mechanism based on AIS data. Evolving
Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Systems, 1–11.
Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,