Available online at www.sciencedirect.

com

ScienceDirect
IFAC PapersOnLine 55-31 (2022) 1–6
Adaptive
Adaptive
Adaptive Learning
Learning
Learning of
of
of Inland
Inland
Inland Ship
Ship
Ship Power
Power
Power
Adaptive
Adaptive Learning
Learning
Propulsion underof
of Inland
Inland Ship
Ship
Environmental Power
Power
Propulsion
Adaptive Learning
Propulsion under
underof Environmental
Inland Ship
Environmental Power
Propulsion
Propulsion under
under
DisturbancesEnvironmental
⋆
⋆
Environmental
Propulsion Disturbances ⋆⋆
under Environmental
Disturbances
Disturbances
Disturbances ⋆⋆
Nicolas Disturbances
Nicolas Dann,
Dann, Pablo
Pablo Segovia,
Segovia, Vasso
Vasso Reppa
Reppa
Nicolas
Nicolas Dann, Dann, PabloPablo Segovia,Segovia, Vasso Vasso Reppa Reppa
Nicolas
Nicolas Dann,
Dann, Pablo
Pablo Segovia,
Segovia, Vasso
Vasso Delft Reppa
Reppa
Department
Department of
of Maritime
Maritime and
and Transport
Transport Technology,
Technology, Delft University
University of of
Department
Technology,
Department of
of Maritime
Delft,
Maritime The and Transport
Netherlands
and Transport Technology,
(e-mail:
Technology, Delft
{n.dannruiz@student.,
Delft University
University of
of
Technology,
Department of Delft,
Maritime The Netherlands
and Transport (e-mail:
Technology, {n.dannruiz@student.,
Delft
Technology,
Department
Technology, of Delft,
Maritime The Netherlands
and
p.segoviacastillo@,
Delft, The Transport
Netherlands
p.segoviacastillo@,
(e-mail:
Technology,
v.reppa@}tudelft.nl).
(e-mail:
v.reppa@}tudelft.nl). Delft University
{n.dannruiz@student.,
{n.dannruiz@student.,University of of
Technology,
Technology, Delft, Delft, The Netherlands
p.segoviacastillo@,
The Netherlands (e-mail:
v.reppa@}tudelft.nl).{n.dannruiz@student.,
(e-mail: {n.dannruiz@student.,
p.segoviacastillo@,
p.segoviacastillo@, v.reppa@}tudelft.nl).
p.segoviacastillo@, v.reppa@}tudelft.nl).
v.reppa@}tudelft.nl).
Abstract:
Abstract: This
This paper
paper presents
presents an
an adaptive
adaptive approximation-based
approximation-based scheme
scheme for for learning
learning aaa partially
for learning partially
partially
Abstract:
known ship
Abstract: shipThispower
This paper
paper presents
propulsion
presents an
plant
an adaptive
under approximation-based
adaptive environmentalscheme
various environmental
approximation-based conditions.
scheme for Considering
learning a the
partially
known
Abstract:
known ship power
This
power propulsion
paper presents
propulsion plant
an
plant under
adaptive
under various
approximation-based
various environmental conditions.
scheme for
conditions. Considering
learning a
Considering the
partially
the
Abstract:
effect
known
effect of
ofship This
water
power
water paper
depth presents
on
propulsion
depth on the
the an adaptive
engine
plant
engine power,
under
power, approximation-based
aa dynamic
various environmental
dynamic model
model scheme
is for learning
defined
conditions.
is defined comprised a partially
Considering
comprised of
of the
the
known
effect
known ship
ofdynamics
ship power
water
power propulsion
depth on
propulsion the plant
engine
plant under
power, various environmental
aa dynamic model conditions.
is defined Considering
comprised of the
engine
effect
engine of
effect
engine of water
dynamics
water
dynamics
and
depth
and
depth
and
the
on
the
on
the
1-DoF
the
1-DoF
the
1-DoF shipunder
ship
engine
engine
ship
power, various
manoeuvring
manoeuvring
power,
manoeuvring a
environmental
dynamics.
dynamic
dynamics.
dynamic
dynamics.
model
model
The
The
The
conditions.
is
is
modelling
defined
modelling
defined
modelling
Considering
challenge
comprised
challenge of
comprised
challenge
is
is the
of
is the
the
effect
engine ofdynamics
determination
determination waterofof depth
ship
and
ship on 1-DoF
the the engine
resistance.
resistance. To
ship
To meet
meet power,
this
manoeuvring
this a dynamic
challenge
challenge dynamics. model
analytical
analytical The is modelling
definedof
modelling
modelling of comprised
ship
ship of
resistance
challenge is
resistance theis
is
engine
determination
engine
combined dynamics
with of
dynamics and
anship
and the 1-DoF
resistance.
the 1-DoFonline
error-filtering ship
To
ship
onlinemeetmanoeuvring
this
manoeuvring
learning challenge
(EFOL) dynamics.
analytical
dynamics.
scheme for The
The
for modelling
modelling
modelling
computing of challenge
ship
an is
resistance
challenge the
is the
approximation is
determination
combined
determination with of
an
of ship
ship resistance.
error-filtering
resistance. To
To meet
meet this
learning
this challenge
(EFOL)
challenge analytical
scheme
analytical modelling
computing
modelling of
of ship
an
ship resistance
approximation
resistance is
is
combined
determination
of the
combined with
with of
unmodeled an error-filtering
anship
part resistance.
of ship
error-filtering online
To meet
resistance
online learning
related
learning (EFOL)
this challenge
to wind
(EFOL) scheme
analytical
and
schemeair. for
After computing
formodelling
simulations
computing an
ofanshipapproximation
resistance
under multiple
approximation is
of the
of the unmodeled
combined unmodeled
with an part of ship
ship resistance
error-filtering
part of resistance
online related
learning
related to wind and
(EFOL)
to wind and
schemeair. After
air. After
for simulations
computing
simulations an under multiple
approximation
under multiple
combined
weather
of the
weather with
unmodeled an
conditions,
conditions, error-filtering
the
part of trained
ship
theoftrained
trained online
model
resistance
model was learning
was
related
was (EFOL)
demonstrated
to
demonstratedwind scheme
and toair.
toair. for
efficiently
After
efficiently computing
estimate
simulations an
estimate the approximation
the unmodelled
under
the multiple
unmodelled
of
of
partthe
weather
theof unmodeled
conditions,
unmodeled
the ship part
the
part ship
oftrained
resistance ship
for resistance
model
resistance
an inland related
related
vessel. to
to wind
demonstratedwind andand toair. After
After simulations
efficiently estimate
simulations under multiple
unmodelled
under multiple
weather
part
weather of conditions,
the ship
conditions, the
resistance
the for
trained anmodel
inland
model was
was demonstrated
vessel.
demonstrated to
to efficiently
efficiently estimate
estimate the
the unmodelled
unmodelled
part
weather of the ship
conditions, resistance
the for
trained an inland
model was vessel.
demonstrated to efficiently estimate the unmodelled
part of
Copyright
part of the © ship
2022 resistance
The Authors. for an
This inland
is an open access
vessel. article under the CC BY-NC-ND license
part of the ship
ship resistance
the On-line resistance
(https://creativecommons.org
Keywords: learning
for
for an
an inland
scheme, inland
surface
vessel.
vessel.
/licenses/by-nc-nd/4.0/) vehicles, speed-power prediction, ship resistance,
Keywords:
Keywords: On-line
On-line learning
learning scheme,
scheme, surface
surface vehicles, speed-power
vehicles, speed-power prediction, prediction, ship ship resistance,
resistance,
shallow
Keywords:
shallow water.
On-line
water. learning scheme, surface vehicles, speed-power prediction, ship resistance,
Keywords:
shallow
Keywords: On-line
water. learning scheme, surface vehicles, speed-power
On-line learning scheme, surface vehicles, speed-power prediction, ship resistance, prediction, ship resistance,
shallow
shallow water.
water.
shallow 1. water.
INTRODUCTION lation is the procedure
1.
1. INTRODUCTION lation is the procedure presented presented in
in Holtrop
Holtrop (1984)
(1984) and
and re-re-
1. INTRODUCTION
INTRODUCTION lation
cently
lation
cently
is the
the procedure
isvalidated
validatedprocedure in
in
presented
presented in
Nikolopoulos
Nikolopoulos and
in
and
Holtrop (1984)
(1984) and
Boulougouris
Holtrop
Boulougouris (2019)
and
(2019)
re-
re-
1.
1. INTRODUCTION
INTRODUCTION lation
cently is the
validatedprocedure in presented
Nikolopoulos in
and Holtrop (1984)
Boulougouris and
(2019)re-
Increasing environmental
Increasing environmental
environmental concerns concerns
concerns and and global
and global warming
warming and
global warming lation
and
cently is the
Grabowska
validated
Grabowska procedureand
in
and presented
Szczuko
Nikolopoulos
Szczuko in
(2015).
and
(2015). Holtrop
In (1984)
regards
Boulougouris
In regards to
to and
inland
(2019)
inlandre-
Increasing cently
and
cently
waterways, validated
Grabowska
validated
shallow in
and
in Nikolopoulos
Szczuko
Nikolopoulos
water depths and
(2015).
and Boulougouris
In regards
Boulougouris
significantly to
increase (2019)
inland
(2019)
ship
have prompted
Increasing
have prompted
prompted international
environmental
international regulations
concerns and
regulations on
global energy
on energy warming
energy effi-
effi- and and Grabowska
waterways, shallow and Szczuko
water depths(2015). In regards
significantly to
increase inland
ship
Increasing
have environmental
international concerns and global
regulations global
on warming effi- and Grabowska
waterways,
Grabowska
resistance, shallow
making and
andwaterSzczuko
Szczukodepths
propulsion (2015).
(2015). In
In regards
significantly
estimation regards
highly to
increase
to inland
ship
inland
intricate
Increasing
ciency
have for
ciencyprompted environmental
ocean-going
for ocean-goinginternational
ocean-going concerns
vessels.
vessels. and
According
regulations
According to
on
to the warming
Interna-
energy
the Interna-effi- waterways,
resistance, shallow
making water depths
propulsion significantly
estimation highlyincrease ship
intricate
have
ciency prompted
for international
vessels. regulations
According on
to energy
the Interna-effi- waterways,
resistance,
waterways,
(Zeng, 2019).shallow
making
shallow
A water
water
common depths
propulsion
depths
practice significantly
estimation
significantly
in the highlyincrease
increase
literature ship
intricate
ship
is to
have
tional
ciency
tional prompted
Maritime
for international
Organisation
ocean-going
Maritime vessels.
Organisation regulations
(IMO),
According
(IMO), on
maritime
to
maritime energy
the transport
Interna-
transport effi- resistance,
(Zeng, making
2019). A propulsion
common estimation
practice in the highly
literature intricate
is to
ciency
tional for ocean-going
Maritime vessels.
Organisation According
(IMO), to
maritime the Interna-
transport resistance,
(Zeng,
resistance,
apply a making
2019).
making
correction A propulsion
common
propulsion
on either theestimation
practice in
estimation the
propulsion highly
literature
highly
power intricate
is
intricate
or to
the
ciency
emits
tional
emits for
aroundocean-going
Maritime
around 940
940 millionvessels.
Organisation
million tonnes
tonnes According
of
(IMO),
of CO
CO to
maritime the
annually Interna-
and
transport
2 annually and is is (Zeng,
apply a 2019).
correction A common
on eitherpractice
the in the
propulsion literature
power or is to
the
tional
emits Maritime
around 940 Organisation
million tonnes(IMO),
of COmaritime
2
2 annually transport
and is (Zeng,
apply
(Zeng,
ship a 2019).
correction
2019).
velocity ofA
A a common
on
common either
deepwater practice
the
practice
method. in
in the
propulsion
the
The literature
power
literature
most or
renownedis
is to
the
to
tional
emits Maritime
responsible
around
responsible for
for 940 Organisation
about
about 2.5%
million
2.5% of
tonnes
of (IMO),
global
of
global COmaritime
greenhouse
annually
greenhouse
2 transport
gas
gas emis-
and
emis- is apply
ship a correction
velocity of on
aa deepwatereither the
method.propulsion
The mostpower or
renowned the
emits around
responsible for 940
about million
2.5% tonnes
of of
global CO annually
greenhouse
2 gas and
emis- is apply
ship
apply
shallow a
a correction
velocity of
correction
water resistanceon
on either
deepwater
either the
method.
the
correction propulsion
The
propulsion
methods mostpower
power
can be or
renowned
or
arguedthe
the
emits
sions around
(GHGs).
responsible
sions (GHGs).
(GHGs). for 940
At
about
At million
the current
2.5%
the current
currenttonnes
of pace, of
global
pace, CO
the annually
annual
greenhouse
the annual GHGs
2 GHGs
GHGsgas andfrom
emis- is
from ship ship
shallow velocity
water of a deepwater
resistance method.
correction The
methods most
can renowned
be argued
responsible for about 2.5% of global greenhouse gas velocity of aapresented
deepwater method. The most renowned
sions
responsible
maritime
sions
maritime for At
transport
(GHGs). At
transport
the
about
theinin2.5%
2050
current
2050
pace,
ofare
global
pace,
are
the
estimated
the
estimated
annual
greenhouse
annual to
to gas emis-
exceed
GHGs
exceed
from
emis-
the
from
the
shallow
ship
to be
shallow
to be
water
velocity
the
water
the ones
ones
resistance
of deepwater
resistance
presented
correction
inmethod.
correction
in Raven
Raven
methods
The
methods
(2016),
can
most
(2016), be
be argued
renowned
canSchlichting
argued
Schlichting
sions (GHGs).
maritime At
transport the incurrent
2050 pace,
are the
estimatedannual to GHGs
exceed from
the shallow
to be
shallow
(1934), water
the
water
and ones resistance
presented
resistance
Lackenby correction
in
correction
(1963). Raven methods
(2016),
methods can
can be argued
Schlichting
be argued
sions
total (GHGs).
shipping
maritime
total At
transport
shipping the
emissions
emissions incurrent
in
2050
in 2008pace,
are
2008 by the
estimated
by annual
90-130%,
90-130%, to GHGs
undermining
exceed
undermining from
the to be
(1934), the
and ones
Lackenby presented
(1963). in Raven (2016), Schlichting
maritime
total transport
shipping emissions in 2050
in are
2008 estimated
by 90-130%, to exceed
undermining the to
to be
(1934),
be the
and
the ones
Lackenby
ones presented
(1963).
presented in
in Raven
Raven (2016),
(2016), Schlichting
Schlichting
maritime
the
total
the 50% transport
reduction
shipping
50% reductionemissions in
imposed
imposed2050
in byare
2008
by the
the estimated
by Paris
90-130%,
Paris to
Agreement exceed (IMO,
undermining
Agreement (IMO,the (1934), and Lackenby (1963).
total shipping emissions in Limitations
(1934), and in
and Lackenbyterms of the
(1963). accuracy and applicability of
the
total50%
2020).
the 50%
2020).
reduction
shipping
Thereby,
reduction
Thereby, theimposed
emissions
the in 2008
shipping
imposed
shipping
by
2008
by theby
industry
the
industry
90-130%,
byParis
90-130%,
Parisis striving
is striving undermining
Agreement
Agreement (IMO,
(IMO, Limitations
undermining
to employ
to employ (1934),
Limitations in
in terms
Lackenby
terms of
of the
(1963).
the accuracy
accuracy and
and applicability
applicability of
of
the 50%
2020). reduction
Thereby, the imposed
shipping by the
industryParisis Agreement
striving to (IMO,
employ analytical
Limitations
analytical methods
in
methods terms mainly
of
mainly the rely on
accuracy
rely on their
and
their empirical
applicability
empirical nature.
nature. of
the 50%
measures
2020).
measures reduction
for
Thereby,
for fuel
fuel the imposed
efficiency,
shipping
efficiency, by
that
thatthe Paris
depends
industry
dependsis Agreement
on
striving
on the
the (IMO,
powering
to employ
powering Limitations
analytical in
methods terms of
mainly the accuracy
rely on and
their applicability
empirical nature. of
2020).
measuresThereby,
for fuel the shipping
efficiency, industry
that dependsis striving
on the to employ
powering Limitations
As a result
analytical ofin
methods terms
being of
based
mainly the
on accuracy
model
rely on and
tests
their applicability
carried
empirical out under
nature. of
2020). Thereby,
performance
measures
performance for of
fuel
of the
vessels.
vessels.shippingthat
efficiency, industry
dependsis striving
on theto employ As
powering aa result
analytical
As result of being
methods
of being based
mainly
based on
on model
rely
modelon tests
their
tests carried
empirical
carried out
out under
nature.
under
measures
performance for fuel
of efficiency,
vessels. that depends on the powering analytical
design
As a
design result methods
conditions,
of
conditions,being mainly
when
based
when onrely
implemented
model
implementedon their
testsin empirical
incarried
real-lifeout
real-life anature.
a under
rough
rough
measures
performance for of
fuel efficiency, that depends on the powering As
vessels. aa result of being based on
Propulsion performance design
As
estimate conditions,
result of of
the being
actual when
based on model
implemented
propulsion testsin
modelperformance
tests carried
real-life
incarried isout
outa under
rough
a under
achieved
performance
Propulsion
performance
Propulsion
of vessels. is
of vessels.
performance
performance is
is
a
a
a
measure
measure
measure
of
of
of
the
the
the
energy
energy
energy
con-
con-
con-
design
estimate
design
estimate
conditions,
of the
conditions,
of the actual
actual
when
when
implemented
propulsion
implemented
propulsion in
real-life
performance
real-life
performance is
is a
rough
achieved
rough
achieved
sumption
Propulsion
sumption at
at a certain
performance
a certain state,
is
state, a and
measure
and can
can be
of
be studied
the energy
studied as
as the
con-
the design
(Pedersen
estimate
(Pedersen conditions,
ofand
the
and Larsen,
actual
Larsen, when implemented
2009).
propulsion
2009). Furthermore, in real-life
performance
Furthermore, as
as is a rough
calculations
achieved
calculations
Propulsion
sumption atperformance
a certain is
state, a measure
and can of
be the energy
studied as con-
the estimate
(Pedersen
estimate
are based of
ofonthe
and
the the actual
Larsen,
actual propulsion
2009).
propulsion
parametrisation performance
Furthermore,
performance
of the hullas is
is
shape, achieved
calculations
achieved
solely
Propulsion
relation
sumption
relation atperformance
between
between a a ship’s
certain
a ship’s is
speed
state,
speeda measure
and
and
and the
can
the of
be the
required
required energy
studiedpropulsion
as
propulsioncon-
the (Pedersen
are based on and Larsen,
the 2009).
parametrisation Furthermore,
of the hull as calculations
shape, solely
sumption at (Pedersen and Larsen, 2009).
2009). Furthermore, as calculations
relation
sumption at a
between
power (Pedersen
relation
power (Pedersen
between a certain
a ship’s
certain
a and
ship’s
and
state,
speed
state,
Larsen,
speed
Larsen,
and
and
and
2009).
and
2009).
can
the be
be studied
required
canDuring
the During
requiredstudied
the as
as the
propulsion
lifetime
propulsion
the lifetime
are
are based
the (Pedersen
ships
ships within
based
within
on
and
on aathe
the
set
parametrisation
Larsen,
set of specification
parametrisation
of specification
of
of the
Furthermore,
limits
themay
limits may
hull
hullas
be
be
shape,
shape, solely
calculations
considered.
solely
considered.
relation
power between
(Pedersen a ship’s
and speed
Larsen, and
2009).the required
During propulsion
the lifetime are
ships
are based
within
based
According on
on
to athe
set
the
Bertram parametrisation
of specification
parametrisation
(2012), of
of the
limits
the
analytical mayhull
hullbe
methodsshape, solely
considered.
shape, aresolely
thus
relation
of a
power
of a between
ship, speed
(Pedersen
ship, speed aisship’s
and
is bound
bound speed
to
Larsen,
to and
decrease
2009).
decreasethe required
for
During
for the
the propulsion
same
the
same input
lifetime
input ships
Accordingwithin to a set
Bertram of specification
(2012), limits
analytical may be
methods considered.
are thus
power
of a (Pedersen
ship, speed and
is boundLarsen,
to 2009).
decrease During
for the the
same lifetime
input ships
According
ships
bound within
within
to to
becomea
a set
Bertram
set of
of specification
(2012),
specification
invalid for modernlimits
analytical
limits may
may
vessel be
methods
be considered.
are
considered.
shapes. thus
With
power
powers
of a
powers (Pedersen
as
ship,
as a result
speed
a result and
is of
bound
of Larsen,
performance
to
performance 2009).
decrease During
reduction.
for
reduction. the As
Asthe
same lifetime
ships
ships are
input
are According
bound to to
become Bertram (2012),
invalid for analytical
modern methods
vessel shapes. are thus
With
of a
powersship,
as speed
a result is bound
of to
performance decrease for
reduction. the Assameships input
are According
bound
According
respect to
to to
become
to
inlandBertram
Bertram (2012),
invalid for
(2012),
waterways, analytical
modern
analytical
Zeng (2019) methods
vessel shapes.
methods
and are
are thus
With
Schlichtingthus
of a
powersship,
subject
subject to
as
to speed
a external
result
external is bound
of factors to
performance
factors decrease
such
such as for
wind
reduction.
as wind the
and
As
and samewaves,
ships
waves,input
areit
it bound
respect to
to become
inland invalid
waterways, for modern
Zeng vessel
(2019) and shapes. With
Schlichting
powers
subject as
to a result
external of performance
factors such reduction.
as wind As
and ships are bound
respect to
to become
inland invalid
waterways, for modern
Zeng vessel
(2019) and shapes. With
Schlichting
powers
is
subject
is as
unlikely
unlikelytoathat
resulttwo
external
that two ofidentical
performance
factors
identical reduction.
operational
such as
operational As waves,
windscenarios
and
scenarios ships
occur,
waves,
occur,areit
it boundrespect to
(1934)
(1934) to become
have
have further
inland
further invalid
claimed
waterways,
claimed for shallow
modern
Zeng (2019)
shallow vessel
water
water and shapes.
methods With
to
Schlichting
methods to be
be
subject
is unlikelyto external
that two factors
identical such as
operational wind and
scenarios waves,
occur, it respect
(1934)
respect
physically to
have
to inland
further
inland
weak waterways,
claimed
waterways,
and questionableZeng
shallow
Zeng (2019)
water
(2019)
given and
and
their Schlichting
methods to
Schlichting
dependency be
subject
making
is unlikely
making to
the
the external
estimation
that two
estimation factors
of
of such
propulsion
identical
propulsion as wind
performance
operational
performance and
scenarios waves,
complex.
occur,
complex. it (1934)
physically have weak further and claimed shallow
questionable water
given theirmethods
dependencyto be
is unlikely
making the that two
estimation identical
of operational
propulsion scenarios
performance occur,
complex. (1934)
physically have weak further and claimed shallow
questionable water
given theirmethods
dependencyto be
is unlikely
making thethat two identical
estimation operational
of propulsion scenarios
performance occur, (1934)
complex. on
physically
on have
deepwater
deepwater weak further and claimed
calculations.
calculations. shallow
questionable water
given theirmethods
dependencyto be
Analytical
making
Analyticalthe mathematical
estimation
mathematical of models
propulsion
models have
have been
performance
been predominantly
complex.
predominantly physically
on deepwater weak and
calculations.questionable given their dependency
making
Analyticalthemathematical
estimation of propulsion
models have performance
been complex. physically
predominantly on deepwater weak and questionable given their dependency
calculations.
used
used to
Analytical
to address
address powering
mathematical
powering performance
models have been
performance by
by quantifying
predominantly
quantifying the
the With
on
With
on computational
deepwater
computational
deepwater calculations.
calculations. techniques
techniques becoming
becoming increasingly
increasingly
Analytical
used to
Analytical
speed-power mathematical
address powering
mathematical
relation based models
models
on have
performance
have
speed been
by
been
loss due predominantly
quantifying
predominantly
to ship the
resis- With
powerful computational
and efficient, techniques
data-driven becoming
methods increasingly
for analysing
used to address
speed-power powering
relation based performance
on speed lossbyduequantifying
to ship the
resis- With
powerful computational
and efficient, techniques
data-driven becoming
methods increasingly
for analysing
used to address
speed-power powering
relation based performance
on speed lossbyduequantifying
to ship the
resis- With
powerful
With
the computational
and
computational
powering efficient,
performance techniques
data-driven
techniques
of shipsbecoming
methods
becoming
using increasingly
for analysing
increasingly
measurement
used
tance.
tance.to address
For
speed-power
For powering
ocean-going
relation
ocean-going based performance
vessels,
on
vessels, a
speed
a primarily
loss
primarilybyduequantifying
utilised
to ship
utilised the
calcu-
resis-
calcu- powerful
the and
powering efficient,
performance data-driven
of ships methods
using for analysing
measurement
speed-power
tance. For relation
ocean-going based on
vessels, speed
a loss
primarily due to ship
utilised resis-
calcu- powerful
the
powerful
data have and
powering
and efficient,
performance
efficient,
attracted data-driven
of
data-driven
increased ships methods
using
methods
attention (Yoo for analysing
measurement
for analysing
and Kim,
speed-power
tance.
⋆ This For
research
relation
ocean-going
is supported
based by
on
vessels,
the
speed
a loss
primarily
project
due
“Novel
to ship
utilised
inland
resis-
calcu-
waterway the
data powering
have performance
attracted increased of ships
attentionusing (Yoomeasurement
and Kim,
tance.
⋆
⋆ This For ocean-going
research is supported vessels,
by the a primarily
project “Novel utilised
inland calcu-
waterway the
data
the
2019). powering
have
powering performance
attracted
performance
Particularly, increased
machine of
of ships
attention
ships
learning using
using (Yoo
(ML) measurement
and
measurement Kim,
techniques
tance.
⋆ This
transport
This
For ocean-going
research is
concepts
research is
supported
for
supported
vessels,
moving by
by
the a
freight
the
primarily
project “Novel
effectively
project “Novel
utilised
inland calcu-
waterway
(NOVIMOVE)”.
inland waterway
data
2019). have attracted
Particularly, increased
machine attention
learning (Yoo
(ML) and Kim,
techniques
transport
⋆ concepts for moving freight effectively (NOVIMOVE)”. data
2019).
data
have have
have
been attracted
Particularly,
attracted
shown toincreased
machine
increased
improve attention
learning
attention
the (Yoo
(ML)
(Yoo
accuracy and
and
of Kim,
techniques
Kim,
propul-
transport
⋆
Funded
transportby concepts
This research
the is
concepts
This research is
for moving
supported
European
for moving
supported
freight
by the
Union’sfreight
by the
effectively
project
Horizon “Novel
2020
effectively
project “Novel
(NOVIMOVE)”.
inland waterway
research
(NOVIMOVE)”.and
inland waterway in- 2019).
have beenParticularly,
shown tomachine
improve learning
the (ML)
accuracy techniques
of propul-
Funded
Funded
transportby
by the
the
conceptsEuropean
European
for Union’s
Union’s
moving Horizon
Horizon
freight 2020
2020
effectively research
research
(NOVIMOVE)”.and
and in-
in- 2019).
have
2019).
sion beenParticularly,
shown
Particularly,
performance tomachine
improve
machine
estimation. learning
the
learning
In the (ML)
accuracy
(ML)
paper by techniques
of propul-
techniques
de Geus-
novation
Funded
transport
novation byprogramme
the
concepts
programmeEuropean
for under grant
Union’s
moving
under agreement
Horizon
freight
grant No
2020
effectively 858508
research and
(NOVIMOVE)”.and the
in- have
sion been
performance shown to improve
estimation. In thethe accuracy
paper by ofde propul-
Geus-
novation by
Funded
Researchlabprogramme
the European
Autonomous grant agreement
underShipping
Union’s agreement
Horizon
(RAS)
No
No
2020
of
858508
858508
research
Delft
and
and the
and
University the
in-
of
have
sion
have
Moussault
sion
been
performance
been et
performance
shown
shown
al. to
to improve
estimation.
(2021), improve
an
estimation. In the
Inthe
artificial the accuracy
paper
accuracy
neural
the paper by
network of propul-
byofde Geus-
depropul-
(ANN)
Geus-
novation
Funded by
Researchlabprogramme
the European
Autonomous underShipping
grant agreement
Union’s Horizon
(RAS) of No
2020Delft858508
research and
and the
University in-
of Moussault et al. (2021), an artificial neural network (ANN)
Researchlab
novation
Technology. Autonomous
programme underShipping
grant (RAS)
agreement of Delft
No University
858508 and of
the sion
sion performance
Moussault et
performance al. estimation.
(2021), an
estimation. In
artificial
In the
the paper
neural
paper by
network
by de
de Geus-
(ANN)
Geus-
Researchlab
novation
Technology. Autonomous
programme underShipping
grant (RAS)
agreement of Delft
No University
858508 and of
the Moussault et al. (2021), an artificial neural network (ANN)
Technology. Autonomous Shipping (RAS) of Delft University of
Researchlab
Technology. Autonomous Shipping (RAS) of Delft University of
Researchlab
Moussault
Moussault et
et al.
al. (2021),
(2021), an
an artificial
artificial neural
neural network
network (ANN)
(ANN)
Technology.
2405-8963 Copyright © 2022 The Authors. This is an open access article under the CC BY-NC-ND license.
Technology.
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2022.10.400
2 Nicolas Dann et al. / IFAC PapersOnLine 55-31 (2022) 1–6

and a convolutional neural network (CNN) were developed density of water [kg/m3 ], D is the propeller diameter [m],
for the speed prediction of ocean-going ships, conclud- Ip is the total inertia of the propeller [kg · m2 ], ηS is the
ing on their superior prediction accuracy, particularly by mechanical efficiency from the engine to the propeller [-],
the ANN. Similarly, Moreira et al. (2021) introduced an Qp is the propeller torque [N · m], m is the ship’s mass
ANN for propulsion performance estimation by predicting [kg], ma is the ship’s added mass [kg], VS is the ship
speed and fuel consumption, while Abebe et al. (2020) speed through water [m/s], and R is the ship resistance
implemented a variety of regression methods such as forest [N]. The latter will be further derived as a function fr of
and gradient boosting regressors. The shorter computation the ship speed VS and wind conditions W in Section 3.
time and higher accuracy of ML techniques compared to In this paper, the inland vessel and propeller parameters
analytical methods make ML an appealing technology, of “Ship 2” presented in the study by Nuij (2021) will
particularly given its easy implementation on board any be considered. Based on equations (1a)-(1g), the following
ship comprising the designated set of sensors (Abebe et al., dynamic model is derived:
2020). ηS P KQ ρD5 2
n˙p = − np
Although ML techniques can be powerful in estimating 4π 2 Ip np 2πIp
the power propulsion performance of a ship, deriving (2)
1 1
guarantees for their performance considering properties V˙S = KT ρD4 np 2 − fr (VS , W ).
m + ma m + ma
like convergence is not possible. There has been significant
research activity on techniques that integrate model-based Ship manoeuvring and its required set of sensors are con-
and data-driven techniques in adaptive learning schemes sidered out of the scope of this study, therefore simplifying
(Farrell and Polycarpou, 2006), (Reppa et al., 2016). These ship motions to 1 DoF, such that lateral (i.e sway), vertical
schemes are developed by carrying out stability analysis (i.e heave), and rotational motions are neglected.
that offers valuable information about the properties of In the case of shallow water, the water depth affects
the adaptive scheme and a systematic way to select the the ship’s propulsion performance. This is considered by
design parameters. correcting the engine power P [W] in (2) to incorporate
the power increase effects of shallow water operation as
The goal and the main contributions of this work are: (Raven, 2016):
(1) the analytical ship power propulsion modelling for P ∆RV VS
inland waterways based on state-of-the-art propulsion and P → − . (3)
resistance calculation methods, and (2) a novel error- Rsink ηDid
filtering online learning (EFOL) scheme based on a ra- Thereafter, (2) becomes:
dial basis function neural network (RBFNN). The EFOL ηS P ηS ∆RV (H) VS KQ ρD5 2
scheme integrates an adaptive nonlinear approximator of n˙p = − − np
4π 2 Ip Rsink (H) np 4π 2 Ip ηDid np 2πIp
the unmodelled effect of ship resistance and an estima-
tor that utilizes analytical models (i.e. prior knowledge) 1 1
of the ship power propulsion and filters some signals to V˙S = KT ρD4 np 2 − fr (VS , W ), (4)
m + ma m + ma
limit the effects of noise. On a wider note, this paper
aims to contribute towards the development of a tool to where the viscous resistance increase due to shallow water,
be implemented onboard a ship for real-time propulsion the resistance due to sinkage, and the propulsive efficiency
performance prediction. coefficient in ideal condition are respectively defined as:
This paper is organized as follows. The analytical propul- ∆RV = RV deep 0.57(T /H)1.79 , (5)
sion and resistance models are presented in Sections 2
Rsink = (1 + δ )2/3 ,
∆
and 3, respectively. In Section 4, the developed adaptive (6)
learning scheme is presented, which is later tested in a case 1 − τid
ηDid = ηO ηRid , (7)
study in Section 5. Finally, Section 6 addresses concluding 1 − wSid
remarks and future research avenues. where T is the ship draft [m], H is the water depth [m],
τid is the ideal thrust deduction factor [-], wid is the ideal
2. SHIP POWER PROPULSION MODEL full-scale wake fraction [-], ηO is the propeller’s open-
water efficiency [-], and ηRid is the ideal factor for relative
In this work, ship propulsion is considered to be described rotative efficiency [-]. The deepwater viscous resistance and
based on the following equations (Yoo and Kim, 2017): additional displacement due to sinkage are calculated as:
Engine Speed : nb = GR np , (1a) 1
RV deep = Cv′ ρVS2 S, (8)
Engine power : P = 2πnb QB , (1b) 2
∆ ∆
δ = d(sinkage)AW / , (9)
Propeller thrust : T = KT ρn2p D4 , (1c)
where Cv′ is the viscous resistance coefficient [-] derived as
Propeller torque : Qp = KQ ρn2p D5 , (1d) a function of the Reynold’s number (Raven, 2016), AW
Engine dynamics : 2πIp n˙p = ηS GR QB − Qp , (1e) is the frontal projected area [m2 ], and d(sinkage) [m] is
Ship dynamics : (m + ma )V˙S = T − R, (1f) determined as:  
Ship resistance : R = fr (VS , W ), (1g) BTM CB F rh2 2
F rhd
d(sinkage) = 1.46  −  ,
where GR is the gear ratio between the propeller shaft Lpp 1 − F rh2 2
1 − F rhd
and ship engine [-], np is the propeller speed [Hz], nb is the (10)
engine speed [Hz], QB is the engine torque [N · m], ρ is the with:
 Nicolas Dann et al. / IFAC PapersOnLine 55-31 (2022) 1–6 3

VS The additional bulbous bow resistance RB is computed

F rhd =  according to:
0.3gLpp
(11) 
VS F ri 3 (−3.0PB −2 )
F rh = √ , RB = 0.11ρg( ABT )3 e , (20)
gH 1 + F ri 2
where Lpp denotes the ship’s length between perpendicu- where F ri denotes the immersion Froude number [-].
lars [m], B is the ship molded beam [m], TM is the draught The immersed transom resistance RT R is further com-
at midship [m], CB is the Block coefficient [-], and g is the puted as:
gravitational acceleration constant [m/s2 ]. 1
RT R = ρVS 2 AT c6 . (21)
The objective of this work is to define the ship resistance in 2
(1g) by integrating analytical modelling and an adaptive Additionally, the Holtrop (1984) method accounts for a
learning scheme. correlation allowance resistance RA to include the effects
of roughness and additional phenomena not captured in
3. ANALYTICAL MODELLING OF SHIP the previous resistance components:
RESISTANCE
 
1 
2
The deepwater resistance method derived in Holtrop RA = ρVS (CA + ∆CA ) S + SAP Pi , (22)
(1984) computes a dimensional total resistance [N] (12) 2 i
based on the ship’s speed and its principal dimensions.
where CA and ∆CA are correlation allowance coefficients
R = (1+k)RF +RAP P +RW +RB +RT R +RA +RAA . (12) [-] specified in Holtrop (1984).
The frictional resistance RF (13) is multiplied by the hull Finally, the shape of the ship has a direct influence on
form factor k and can be computed as: its aerodynamics, which is quantified as air and wind
1 resistance RAA , defined by:
RF = ρVS2 SCF , (13)
2 1 2
RAA = ρA Vw,ref Cw AW , (23)
where S is the wetted surface area [m2 ] and CF is the 2
model–ship correlation line coefficient [-], which may be where Cw is the wind coefficient [-] and Vw,ref is the
calculated as a function of the Reynold’s number Re (1957 relative wind speed [m/s], calculated as the difference
ITTC Standards): between the ship’s forward speed through water VS and
0.075 the wind velocity uw in the x-direction [m/s]:
CF = . (14)
(log10 Re − 2)2 uw = Vw cos(ψw ), (24)
The appendage resistance RAP P is calculated as the sum of where Vw is the sensed wind speed [m/s] and ψw denotes
the resistance due to the appendages and the bow thruster the wind direction, with ψw = 0 indicating headwind.
resistance, quantified by the expression: The complexity of parameterising a ship’s geometry and
  computing the frontal projected area AW using analytical
1 2 i (1 + k2 )SAP Pi
RAP P = ρVS  CF SAP Pi + RT H , methods makes RAA highly intricate to estimate. There-
2 i SAP Pi i after, RAA will be considered as the function to be approxi-
(15) mated, with the objective of learning the unknown relation
where the (1 + k2 ) values for each appendage i are pre- between its inputs through an approximation scheme.
sented in Holtrop (1984), SAP Pi denotes the surface area
of each considered appendage i [m2 ], and the resistance 4. ADAPTIVE LEARNING SCHEME
due to the bow thruster tunnel opening RT H is computed
as:
The dynamic system in (4) can be re-written in state-space
RT H = ρVS2 πd2T H CD,T H , (16)
form by considering x = [np VS ]T and u(t) = P as:
with the drag coefficient CD,T H for the thruster tunnel  
assuming values between 0.003 and 0.012 (Birk, 2019), and ηS 1 u ∆RV (H) x2 KQ ρD5 2
x˙1 = − − x
with dT H being the thruster tunnel’s diameter [m]. 2
4π Ip Rsink (H) x1 ηDid x1 2πIp 1
Furthermore, the wave-making resistance RW is estimated 1 1
as a function of the Froude number Fr [-] (Holtrop, 1984): x˙2 = KT ρD4 x1 2 − [(1 + k)RF + RAP P
m + ma m + ma
• for F r < 0.4: 1
+ RW + RB + RT R + R A ] − RAA .
d −2 m + ma
RW a (F r) = c1 c2 c5 ρgV e[m1 F r +m4 cos(λF r )] , (17) (25)
• for F r > 0.55:
d −2
Equation (25) can be expressed as:
RW b (F r) = c17 c2 c5 ρgV e[m3 F r +m4 cos(λF r )] , (18) x˙1 = f0,1 (x1 , x2 , u, d3 ), (26a)
• for 0.4 < F r ≤ 0.55, an interpolation of (17) and (18) x˙2 = f0,2 (x1 , x2 ) + f2∗ (x2 , d1 , d2 ), (26b)
applies:
where f0,1 and f0,2 are considered as known functions,
(20F r − 8) while f2∗ is an unknown function to be learned. The
RW (F r) = RW a (0.4)+ [RW b (0.55) − RW a (0.4)] , disturbance due to V , ψ , and H are denoted d , d ,
3 w w 1 2
(19) and d3 , respectively.
where the parameters denoted c and m are dimensionless An adaptive learning scheme is designed considering only
calculation coefficients presented in Holtrop (1984). (26b) since (26a) does not contain any uncertainty. In
4 Nicolas Dann et al. / IFAC PapersOnLine 55-31 (2022) 1–6

addition, it is assumed that x1 , x2 , d1 , d2 , d3 are all Estimator

measured and the output of the sensors are:
y2(t)
yd1 = d1 + w3 λ
y1 = x1 + w1
, yd2 = d2 + w4 , (27) -
y 2 = x2 + w 2 +
y1(t) +
yd3 = d3 + w5
y2(t) f0,2 ( y1, y2 )
+
+ +
λ +
+
- e(t)
where w denotes sensor measurement uncertainty. The s +λ
states x1 and x2 can be measured by a differential GPS +
+
and an RPM sensor, respectively. In regards to the dis-
turbances, d1 and d2 can be respectively measured by an
y2(t)
f 2̂ ( y2; θ ̂ )
anemometer and weathervane, while an echo sounder can
be used to measure d3 .
·
The developed learning scheme is comprised of a paramet- θ (t )
ric model and an error filtering online learning scheme.

4.1 Parametric Model Approximator

The parametric model is defined as: Fig. 1: Block diagram configuration of the EFOL scheme
χ2 (t) = fˆ2 (x2 (t); θ∗ ) + δ2 (t), (28) Applications of RBFs in on-line learning approximation
where χ2 (t) is given by: schemes can be seen in the works of Liu et al. (2021),
Zhang (2016), Wu et al. (2012), and Gorinevsky (1993).
χ2 (t) = ẋ2 (t) − f0,2 (x1 (t), x2 (t)), (29) In this paper, Gaussian RBFs will be introduced, re-
and fˆ2 is an adaptive approximation model (e.g radial defining (32) as:
basis function, sigmoidal neural network) which, based on fˆ2 (x2 (t); θ∗ ) = θ∗T ϕ(x, c, σ), (34)
unknown parameter weights θi∗ , intends to decrease the
minimum functional approximation error (MFAE) δ2 (t), where the regression vector ϕ(x, c, σ) consists of Gaussian
defined as: radial basis functions, with each RBF i being defined as:
δ2 (t) = f2∗ (x2 (t), d1 (t), d2 (t)) − fˆ2 (x2 (t); θ∗ ). (30) 1 ||x − ci ||2
ϕi (x, c, σ) = exp(− ), (35)
2 σ2
4.2 Error Filtering Online Learning Scheme where ci denotes the centre location for the i−th Gaussian
function and σ is the standard deviation or spread of the
Due to the fact that ẋ2 is not available for measurement, functions. To balance the trade-off between computational
a filtering technique is applied. By filtering both sides of complexity and prediction accuracy, 12 neurons are consid-
(28) with a first-order filter, it yields: ered, that is, i ∈ {1, ..., 12}. To increase the approximation
λs λ accuracy, the fixed centre locations for the RBFs are evenly
χ̂2 (t) = [x2 ] − [f0,2 (x1 , x2 )] spaced over the operating velocity range of x2 , which is
s+λ s+λ (31) found to be {0, 5.5}, while the spread is tuned to σ = 0.61
λ  
χ̂2 (t) = fˆ2 (x2 (t); θ∗ ) + δ̂2 (t), for all RBFs using a brute-force approach.
s+λ In regards to learning, the output error signal of the
where δ̂2 (t) is the filtered version of the MFAE and s is scheme e(t) = χ̂2 (t)−χ2 (t) is used to adjust the parameter
the Laplace operator. The filters are described as transfer weights of the RBFNN. The Lyapunov synthesis method
functions, using the relation x˙2 = s[x2 (t)]. In this work, a is typically utilised for EFOL schemes given its inherent
linearly parametrized approximator is considered, i.e, stability properties (Farrell and Polycarpou, 2006). The
qθ following parameter adaptive law is thus derived for θ:
fˆ2 (x2 (t); θ∗ ) = θi∗ ϕi (x2 (t)). (32) ˙
θ̂ = −Γe(t)ϕ(y2 , c, σ), (36)
i=1
where Γ denotes the adaptive gain matrix, or learning rate,
The unknown parameters θi∗ are the adjustable parameters
which will be simplified to Γ = γI, implying that each
that minimise the MFAE and ϕi denotes an element of the
element in θ(t) uses the same adaptive gain.
regression vector. Considering the measurements of x2 , the
EFOL scheme is designed as:
5. SIMULATION EXPERIMENT
λs λ
χ̂2 = [y2 ] − [f0,2 (y1 , y2 )]
s+λ s+λ (33) 5.1 Case Study
λ  T 
χ̂2 = θ ϕ(y2 ) ,
s+λ The inland vessel and propeller parameters of “Ship 2”
The structure of the implemented EFOL scheme is pre- studied in Nuij (2021) are used for this case study. The
sented in Fig. 1. value of the design parameter λ is set to 1, while the
efficiency coefficients are set to: ηS = 0.95, ηDid = 0.95,
4.3 Adaptive Law ηO = 0.37. The thrust, torque, wind resistance, and bow
thruster coefficients are set to: kT = 0.1965, kQ = 0.03316,
The update law for the unknown set of parameters θ∗ (t) Cw = 0.6, and CD,T H = 0.07. Furthermore, the mass of
must be further established. For this purpose, a lin- the ship is assumed to be m = 106 kg while the added mass
early parametrized RBFNN approximator is implemented. ma is neglected.
 Nicolas Dann et al. / IFAC PapersOnLine 55-31 (2022) 1–6 5
In regards to training the RBFNN, the root mean square
error (RMSE) of the prediction fˆ2 with respect to f2∗ is
used to define the scheme’s accuracy:

N 2
i=1 (x2,i − x̂2,i )
RM SE = , (37)
N
where N is the number of data points, x2,i are the actual
values of f2∗ , and x̂2,i are the estimated values by fˆ2 . To
optimise the training process, the EFOL scheme was run
for a range of values of γ = {0.05, 0.1, ..., 1} for 40 epochs
each, and input power and disturbances P = 600kW and
d = [0, 0, 4.5]. On the one hand, small learning rates can
cause the process of learning to get stuck, whereas, on
the other hand, large learning rates can cause the model
to converge quickly to a sub-optimal solution. A value of
γ = 0.3 was found to increase this trade-off, and was used
to train the model for 400 epochs.
5.2 Design of the Experiment
An experimental scenario aiming to mimic realistic wind
conditions was designed to test the trained model.
Medium-scale winds of up to 36km/h in the form of
wind gusts were replicated by introducing a pulse and
a ramp function for the sensed wind speed signal yd1 (t)
(Fig. 2). To imitate sensor noise, Gaussian white noise
was introduced. White noise is expressed by a sequence
of Gaussian variables, with their variance reflecting the
power or intensity at each time step (Miller and Childers,
2012). A power of E = 10−4 was selected, corresponding
to a standard deviation of 0.01N . Gaussian white noise
was introduced to the sensed input signals yd1 (t), yd2 (t),
and yd3 (t). Furthermore, headwind (d2 = 0) together with
an input power of u = 600kW were considered. The
simulation of the EFOL scheme was then run for 104 s.
5.3 Results and Discussion
The resulting approximation of the unknown function
f2∗ (t) by the function approximator fˆ2 (t) and the error
signal of the EFOL scheme e(t) are plotted in Fig. 3(a) and
Fig. 3(b) for the pulse and ramp yd1 (t) inputs, respectively.
In the plots, f2∗ (t) can be seen to be closely replicated by
the approximator function fˆ2 (t), with the latter showing
slight fluctuations due to sensor measurement noise.
For a pulse wind input signal (Fig. 3(a)), a total RMSE of
11.5329N is obtained, caused by overshoots of fˆ2 (t) at the
pulse input changes of yd1 (t) and by approximation error
due to sensor noise. Nevertheless, the error signal rapidly
decreases after each overshoot, with each spike elapsing
around 3 seconds.
In the case of yd1 (t) being described by a ramp function
(Fig. 3(b)), a smaller RMSE of 10.0073N is obtained,
reflected in a smooth approximation of the unknown
function f2∗ (t). Note that, despite an increase in the
sensed wind speed d1 (t), the unknown air resistance f2∗ (t)
decreases at around halfway of the function’s ramps. This
is a result of RAA being computed based on the wind
speed relative to the ship’s velocity. In reality, sensed wind
should increase with forward ship motion, which was not
accounted for in the theoretical yd1 (t) signal.
In essence, the obtained results show the ability of the
learning model to rapidly adapt to a varying input and
closely learn the unknown system dynamics. Given the
magnitude of the resistance values at hand, the obtained
RMSE values can be argued to be significantly small.
6. CONCLUSIONS AND FURTHER RESEARCH
This paper studied the speed-power prediction of inland
ships, a fundamental aspect of their powering performance,
for which an approach combining an EFOL and a RBFNN
was proposed. Considering a partially-known system, a
set of dynamic equations were derived analytically for
the known fraction of the system, while the air and wind
resistance was considered an unknown function of onboard
sensor measurements.
Simulation experiments indicate that the proposed method
can closely approximate unknown dynamics, presenting
itself as a feasible speed-power approximation model. For
future validation, the model could be tested using real
sensor data. This work allows for a variety of further ex-
tensions, such as considering the influence of ship manoeu-
vring by including the set of sensors required to describe
ship motion in 6 DoFs.
REFERENCES
Abebe, M., Shin, Y., Noh, Y., Lee, S., and Lee, I. (2020).
Machine learning approaches for ship speed prediction
towards energy efficient shipping. Applied Sciences,
10(7), 2325.
Bertram, V. (2012). Practical ship hydrodynamics. Else-
vier.
Birk, L. (2019). Fundamentals of Ship Hydrodynamics:
Fluid Mechanics, Ship Resistance and Propulsion. John
Wiley & Sons.
de Geus-Moussault, S., Buis, M., and Koelman, H. (2021).
A convolutional neural network developed to predict
speed using operational data. In COMPIT’21: 20th
International Conference on Computer and IT Appli-
cations in the Maritime Industries, 246–264. Technische
Universität Hamburg-Harburg.
Farrell, J.A. and Polycarpou, M.M. (2006). Adap-
tive approximation based control: unifying neural, fuzzy
and traditional adaptive approximation approaches, vol-
ume 48. John Wiley & Sons.
Gorinevsky, D.M. (1993). Adaptive learning control using
affine radial basis function network approximation. In
Proceedings of 8th IEEE International Symposium on
Intelligent Control, 505–510. IEEE.
Grabowska, K. and Szczuko, P. (2015). Ship resistance
prediction with artificial neural networks. In 2015 Signal
Processing: Algorithms, Architectures, Arrangements,
and Applications (SPA), 168–173. IEEE.
Holtrop, J. (1984). A statistical re-analysis of resis-
tance and propulsion data. International shipbuilding
progress, 31(363), 272–276.
IMO (2020). Fourth imo greenhouse gas study. Interna-
tional Maritime Organisation (IMO).
Lackenby, H. (1963). The effect of shallow water on ship
speed. Shipbuilder and Marine Engine Builder, 70(672).
Liu, X., Liu, X., Zhou, Z., and Hu, L. (2021). An
efficient multi-objective optimization method based on
the adaptive approximation model of the radial basis
function. Structural and Multidisciplinary Optimization,
63(3), 1385–1403.
6 Nicolas Dann et al. / IFAC PapersOnLine 55-31 (2022) 1–6

(a) (b)
12 12

10 10

8 8
Wind Speed (m/s)

Wind Speed (m/s)

6 6

4 4

2 2

0 0

-2 -2
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Time (s) Time (s)

Fig. 2: Measured wind velocity yd1 (t) input signals: (a) Pulse function, (b) Ramp function
(a) (b)
900 900

800 800
Wind and Air Resistance, R AA (N)

Wind and Air Resistance, R AA (N)

700 700

600
600
500
500
400
400
300
300
200
200
100

100 0

0 -100
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Time (s) Time (s)

Fig. 3: Simulation results for: (a) Pulse yd1 (t) signal, (b) Ramp yd1 (t) signal

Miller, S.L. and Childers, D. (2012). Chapter 12 - sim- Ships and Other Floating Structures PRADS.
ulation techniques. In S.L. Miller and D. Childers Reppa, V., Polycarpou, M.M., Panayiotou, C.G., et al.
(eds.), Probability and Random Processes (Second Edi- (2016). Sensor fault diagnosis. Foundations and
tion), 517–546. Academic Press, Boston, second edition Trends® in Systems and Control, 3(1-2), 1–248.
edition. Schlichting, O. (1934). Ship resistance in water of limited
Moreira, L., Vettor, R., and Guedes Soares, C. (2021). depth-resistance of sea-going vessels in shallow water.
Neural network approach for predicting ship speed and Jahrbuch der STG, 35, 127–148.
fuel consumption. Journal of Marine Science and Engi- Wu, Y., Wang, H., Zhang, B., and Du, K.L. (2012). Using
neering, 9(2), 119. radial basis function networks for function approxima-
Nikolopoulos, L. and Boulougouris, E. (2019). A study tion and classification. International Scholarly Research
on the statistical calibration of the holtrop and mennen Notices, 2012.
approximate power prediction method for full hull form, Yoo, B. and Kim, J. (2017). Powering performance analy-
low froude number vessels. Journal of Ship Production sis of full-scale ships under environmental disturbances.
and Design, 35(01), 41–68. IFAC-PapersOnLine, 50(1), 2323–2328.
Nuij, G.J. (2021). Introduction of pem fuel cells on inland Yoo, B. and Kim, J. (2019). Probabilistic modeling of ship
ships. powering performance using full-scale operational data.
Pedersen, B.P. and Larsen, J. (2009). Modeling of ship Applied Ocean Research, 82, 1–9.
propulsion performance. In World Maritime Technology Zeng, Q. (2019). A method to improve the prediction
Conference WMTC2009. of ship resistance in shallow water. Delft University of
Raven, H. (2016). A new correction procedure for shallow- Technology.
water effects in ship speed trials. In Proceedings of the Zhang, Q. (2016). Multilevel Adaptive Radial Basis Func-
13th International Symposium on PRActical Design of tion Approximation using Error Indicators. Ph.D. the-
sis, University of Leicester.