16 h In e na ional Symposium on P ac ical Design o Ships and O he Floa ing S uc u es PRADS 2025
Ann A bo , MI, USA, Oc obe 19 h-23 d 2025
Machine Lea ning Based Con ol Co-Op imiza ion o Geome y and
Powe Take-o o a Poin Abso be Wa e Ene gy Con e e
Weihan Lin1, And ew Fang2, Lisheng Yang1and Lei Zuo1,*
1Uni e si y o Michigan, Ann A bo , MI, USA
2Vi ginia Tech, Blacksbu g, VA, USA
Abs ac . The loa ing buoy geome y and he powe ake-o (PTO) sys em o a poin abso be wa e
ene gy con e e (WEC) we e he wo main ac o s ha in luence he o al powe ex ac ion pe o -
mance. This pape aims o p esen a machine lea ning based con ol co-op imiza ion me hod o geom-
e y and PTO o a poin abso be , wi h a discussion o pe o mance by using a ious aining models
including Suppo Vec o Machines (SVM) and Neu al Ne wo ks (NN). The machine lea ning model in
his s udy is de eloped o eplace he con en ional equency domain bounda y elemen me hod (BEM)
hyd odynamics so wa e wi h housands imes as e simula ion speed and wi h e o ma gin om 0.5%
o 2%. The ade-o be ween he numbe o aining samples and he accu acy o he machine lea ning
model is discussed o u he imp o e he collec ion speed o aining samples. The ad an ages and
disad an ages o a ious ypes o machine lea ning models, SVM and NN, a e discussed and iden i ied.
Finally, he con ol co design o he PTO o he poin abso be is conside ed du ing geome y op i-
miza ion s age, which allows he es ima ion o he elec ical powe o become mo e ealis ic o be used
in he main objec i e unc ion. In his p ocess, PTO pa ame e s a e op imized a he same ime as he
loa ing buoy geome y, achie ing global op imal pe o mance.
Key wo ds: Machine lea ning, Renewable ene gy, Poin abso be , Suppo ec o
machines, Neu al ne wo ks, Powe ake-o
1. INTRODUCTION
As global demand g ows o clean and sus ainable ene gy sou ces, in e es in enewable ene gy, pa -
icula ly om sola , wind, and ocean wa es, has signi ican ly inc eased. Renewable ene gy is expec ed o
accoun o o e 40% o global elec ici y demand wi hin he nex 15 yea s, ising om he cu en 25%,
as o al elec ici y usage expands om 22,200 TWh/yea o 35,500 TWh/yea [1]. Among enewable e-
sou ces, ocean wa e ene gy s ands ou due o i s high ene gy densi y (10-100 kW/m [2]), no ably su passing
ha o sola (100-150 W/m² [3]) and wind (200-400 W/m² [4]).
Theo e ical es ima es o ocean wa e ene gy po en ial ange om 8,000 o 80,000 TWh/yea [5], wi h
he b oade po en ial o all ocean ene gy o ms es ima ed a a ound 151,300 TWh/yea [6]. Al hough hese
heo e ical po en ials exceed cu en echnical capabili ies, ongoing scien i ic and echnological ad ance-
men s p omise imp o ed e iciencies. Wa e Ene gy Con e e s (WECs) a e de ices designed o ans o m
ocean wa e mo ion in o elec ical ene gy, sui able o esiden ial, communi y, and indus ial applica ions.
The poin abso be ype o WEC cap u es ene gy h ough he ela i e mo emen be ween a buoy—ei he
loa ing o subme ged—and a s a iona y o oscilla ing e e ence [7]. This concep , ini ially p esen ed by
Budal and Falnes [8][9], has spu ed ex ensi e esea ch in o hyd odynamics [10][11], s uc u al in eg i y
[12], and powe ake-o (PTO) sys em design [13][14].
Addi ionally, signi ican e o s ha e gone in o op imizing he geome y o poin abso be buoys o
enhance hei ene gy cap u e capabili ies [15]. Gillo eaux and Ringwood op imized cylind ical buoy di-
mensions by adjus ing diame e and d a o single-body hea ing abso be s [16]. De Backe explo ed he
hyd odynamic pe o mance o cylind ical buoys wi h a ious bo om shapes, such as conical and sphe ical
designs, iden i ying op imal geome ies o speci ic sea condi ions [17]. Mo ing beyond simple geome -
ic shapes, McCabe in oduced bi-cubic B-spline pa ame ic modeling combined wi h gene ic algo i hm
*Co espondence o: [email p o ec ed]
1
(GA) op imiza ion o maximize powe cap u e o su ge-and-pi ch WECs [18]. Ga cia-Te uel e al. u he
analyzed a ious op imiza ion s a egies, applying hem o mul iple buoy shapes, including ba ge, hemi-
sphe e, e ical cylinde , and bi-cubic B-spline-de ined su aces [19]. To educe compu a ional complexi y,
Lin e al. used 12 non-uni o m a ional B-spline (NURBS) con ol poin s and employed neu al ne wo ks
o signi ican ly speed up hyd odynamic e alua ions wi hin GA op imiza ions, cu ing compu a ion imes
d ama ically [20].
P e ious WEC shape op imiza ion s udies o en simpli ied powe calcula ions by using an idealized PTO
model as a passi e dampe , neglec ing eal-wo ld losses and dynamic adap abili y o changing wa e condi-
ions. This app oach leads o wo p ima y inaccu acies: i s , i assumes an ideal PTO wi h no powe losses,
and second, passi e damping only achie es maximum e iciency a esonance. Recen s udies ha e high-
ligh ed he bene i s o inco po a ing PTO losses and ac i e con ol in o op imiza ion amewo ks, p o iding
a mo e ealis ic es ima ion o elec ical powe ou pu om WECs [21, 22]. Accu a e powe assessmen is
c i ical, as op imis ic es ima ions based on simpli ied models can lead o ine ec i e de ice designs in la e
de elopmen s ages.
Add essing hese limi a ions, his pape in oduces a comp ehensi e and ealis ic app oach o WEC
op imiza ion h ough machine lea ning-based con ol co-op imiza ion. Two main con ibu ions se his
wo k apa om ea lie s udies. Fi s , PTO cha ac e is ics and ac i e con ol s a egies a e in eg a ed in o
he shape op imiza ion loop, enabling apid and ealis ic e alua ions o elec ical powe ou pu in he e-
quency domain. Second, a machine lea ning app oach, using SVM o Neu al Ne wo ks, quickly p edic s
hyd odynamic pa ame e s based on geome ic inpu s, signi ican ly accele a ing he gene ic algo i hm-based
op imiza ion p ocess. As gene a ing aining da a h ough bounda y elemen me hod (BEM) simula ions
can be compu a ionally in ensi e, his pape also in es iga es how he accu acy o SVM and NN models
a ies wi h he numbe o aining samples.
2. MODELING
In his sec ion, we p esen he modeling me hods o he loa ing buoy used in he WEC design. A
pa ame ic geome y based on Non-Uni o m Ra ional B-Splines (NURBS) con ol poin s is adop ed, as
ou lined by Lin e al. [20]. Dynamic models o he wo-body poin abso be unde wa e and cu en
condi ions a e de eloped o e alua e bo h powe pe o mance and hyd odynamic d ag. Addi ionally, we
de ail he modeling o he PTO sys em, highligh ing i s mechanical pe o mance. Finally, he machine
lea ning app oaches, including SVM and neu al ne wo ks, a e in oduced.
2.1. Pa ame ic Modeling
The buoy geome y is de ined using a pa ame ic NURBS model in oduced by Lin e al. [20], as shown
in Fig. 1.. This model uses 12 pa ame e s, o ganized in o he shape ec o s:
s= [sx, sy, s ](1)
and,
sx= [x1, x2, x3, x4]
sy= [y1, y2, y3, y4]
s = [ x41, x42, y41, y42]
(2)
By adjus ing hese pa ame e s, complex buoy geome ies can be c ea ed. Howe e , ce ain cons ain s mus
be me o a oid in easible deck lines, examples o which a e shown in Fig. 2..
2.2. Powe Take-o Modeling
In his s udy, we es ic ed he analysis o he hea e DOF o he wo bodies and and used a linea ,
equency domain BEM o ge he added mass, adia ion damping, wa e exci a ion, and hyd os a ics o
each candida e shape. Viscous e ec s in wa es and second-o de d i a e no sol ed in he BEM. The
iscous cu en d ag is handled sepa a ely h ough he objec i e in Sec. 3. The de ice is sel – eac ing, so
2
Figu e 1.: Pa ame ic modeling o he loa ing buoy using NURBS con ol poin s [20]
Figu e 2.: Examples o (a) a easible deck line (b) an in easible deck line
hyd odynamic coupling o a sepa a e e e ence s uc u e is no modeled, and he only in e ac ion o he wo
bodies is h ough he PTO. Moo ing dynamics is also no conside ed o he equency domain solu ion.
The dynamics o his wo-body sys em can be desc ibed as:
(m1+A1)¨x1=k1x1+c1˙x1+ e1+ PTO (3)
(m2+A2)¨x2=k2x2+c2˙x2+ e2− PTO (4)
He e, m1,2a e body masses, A1,2a e added masses, k1,2a e hyd os a ic s i nesses, c1,2a e adia ion
damping coe icien s, and e1,2a e wa e exci a ion o ces. PTO ep esen s he PTO o ce, and x1,2a e
displacemen s om equilib ium posi ions.
Figu e 3.: The ee-body diag am o he poin abso be p oposed in his s udy
3
The PTO consis s o a ball sc ew coupled o a gene a o , modeled as equi alen mass and damping wi h
a linea elec omagne ic o que ela ionship. The dynamics a e desc ibed by:
mb¨x=cb˙x+NK iL+ PTO (5)
He e, x=x1−x2 ep esen s ela i e displacemen , mband cb ep esen he ball sc ew p ope ies, K is
he gene a o o que cons an , Nis he gea a io, and iLis he load cu en . Posi i e PTO o ce is de ined
as upwa d h us .
T ans o ming he equa ions in o he equency domain, we de ine in insic impedance Zi o he WEC:
Zi=−c1+jω(m1+A1−k1/ω2),0
0,−c2+jω(m2+A2−k2/ω2)(6)
The ball sc ew impedance is de ined as Zb=−cb+jωmb, and he gene a o impedance is de ined as
Zw=Rs +jωLq, whe e Rsis he s a o esis ance and Lqis he synch onous induc ance. A PI con olle
is assumed o con ol he PTO o ce such ha PTO =Ki(x1−x2) + Kp( ˙x1−˙x2). The con olle
impedance is hus de ined as Zc=Kp+Ki/jω. T ans o m Eq. 5 o he equency domain yields:
Zb(V1−V2)−NK IL=FPTO (7)
Uppe cases a e used o ep esen alues in he equency domain. V1and V2a e equal o jωX1and jωX2.
F om his PTO dynamics we can de i e he equi ed gene a o load impedance ZLin o de o achie e he
desi ed con olle impedance. No ice ha Zc=FPTO/(V1−V2), we hen ha e:
Zc=Zb−NK IL
V1−V2
(8)
Ano he equa ion can be es ablished using he back EMF o he gene a o . Speci ically, o que cons an K
is also he linea coe icien be ween gene a o o a ional eloci y and he back EMF. Thus, we ha e he
elec ical pa and mechanical pa coupling equa ion:
IL(ZL+Zw) = NK (V1−V2)(9)
Plug Eq. 9 in o Eq. 8 and ea ange he equa ion, we ge he co esponding load impedance o he con olle :
ZL=N2K2
Zb−Zc
−Zw(10)
Finally, he elec ical powe abso bed by he gene a o load is exp essed as he a e age ac i e powe o he
load:
Pe=1
2R(ZL)|iL|2(11)
This a e age powe is used in he op imiza ion objec i e in he ollowing sec ion.
2.3. Building he Da ase
Gi en he no el y o his wo k, we c ea ed a da ase comp ising 120,000 dis inc buoy hull shapes
o machine lea ning aining. Each sample includes 13 pa ame e s—12 de ining he buoy geome y and
one speci ying he wa e pe iod. To manage compu a ional esou ces, we di ided he da ase in o smalle
subse s sui able o e icien aining. The da ase was gene a ed in ba ches, each con aining 100 andom
shapes adhe ing o p ede ined geome ic cons ain s om Lin e al. [20]. To educe he sample coun while
main aining p edic ion accu acy, alida ion es ing de e mines he op imal numbe o samples pe ba ch.
4
2.4. Suppo Vec o Machines
To selec he bes model o p edic ing hyd odynamic pa ame e s, wo machine lea ning me hods a e
conside ed: SVM and neu al ne wo ks.
Suppo Vec o Reg ession (SVR), a a ian o SVM cus omized o con inuous alue p edic ions, is
used [23]. SVR ope a es simila ly o linea eg ession by i ing a line o da a poin s, bu wi h wo main
di e ences: i s unique op imiza ion p ocedu e and he applica ion o ke nel unc ions.
In SVR op imiza ion, conside a da ase wi h inpu ec o s xNand co esponding con inuous ou pu s
yN. The SVR aims o ind an op imal line de ined as (x) = wx +bby pe o ming wo op imiza ions.
Fi s , i minimizes he loss unc ion, de ined in Eq. (12), wi h an allowable ma gin ϵ, which o his p oblem
is se a 0.0554 based on he in e qua ile ange o he da ase . Second, i en o ces la ness o he eg ession
line by minimizing he no m o he weigh ec o , as shown in Eq. (13).
L=(0|y− (x)|< ϵ
|y− (x)| − ϵ o he wise (12)
minimize :J(w) = 1
2||w||2(13)
Subjec o:
∀n:|y− (x)|< ϵ (14)
He e, w ep esen s model weigh s, ||w|| is hei Euclidean no m, and bis a bias e m. This op imiza ion
ocuses on da a poin s ou side he ma gin ϵ, known as suppo ec o s, and u ilizes Sequen ial Minimal
Op imiza ion (SMO).
SVRs also employ ke nel unc ions o handle da a no sui able o linea i ing. These unc ions map
inpu s in o highe -dimensional spaces o allow linea sepa a ion. Th ee ke nels es ed in his s udy a e
linea , polynomial, and Gaussian, as desc ibed in Eq. (15):
linea :K(xi, xj) = xi·xj
polynomial :K(xi, xj) = (1 + xi·xj)q
gaussian :K(xi, xj) = e−||xi−xj||2
(15)
Whe e: xiand xja e wo di e en sample ec o s om he inpu xN.Kis he ke nel unc ion. qis
some exponen such ha i is an in ege and q≥2.
Fo op imal ke nel selec ion, 10- old c oss- alida ion is applied, di iding he da a in o en subse s. Each
subse is sequen ially used o alida ion while aining he model wi h he emaining subse s. Resul s
showed ha a linea ke nel bes p edic s added mass and exci a ion o ce, while a polynomial ke nel mos
e ec i ely p edic s damping. Consequen ly, h ee sepa a e SVRs a e es ablished, each wi h 13 inpu pa-
ame e s co esponding o he buoy shape and wa e pe iod, o p edic added mass, damping, and exci a ion
o ce sepa a ely.
2.5. Neu al Ne wo ks
The second me hod conside ed is a eed o wa d neu al ne wo k (NN). This NN includes an inpu laye ,
wo hidden laye s wi h 15 neu ons each, and an ou pu laye . Unlike SVRs, a single neu al ne wo k p edic s
all h ee hyd odynamic pa ame e s simul aneously. The NN s uc u e and neu on coun s a e based on
p e ious op imiza ion s udies by Lin e al. [20].
Calcula ions o hidden laye s a e pe o med acco ding o Eq. (16), and he ou pu laye employs he
an-sigmoid ans e unc ion as shown in Eq. (17):
h(a)
j=σ(Xw(a)
ij h(a−1)
i+b(a)
j)(16)
σ(θ) = ansig(θ) = eθ−e−θ
e−θ+eθ(17)
5
Whe e, h(a)
jis he j- h neu on a he a- h laye . w(a)
ij is he weigh o he j- h neu on a he a- h laye
om he i- h neu on alue a he (a−1)- h laye . h(a−1)
iis he i- h neu on a he (a−1)- h laye . b(a)
jis
he bias o he j- h neu on a he a- h laye . σis he an-sigmoid ans e unc ion [24].
The NN aining employs a backp opaga ion algo i hm, adjus ing weigh s and biases o minimize p e-
dic ion e o s. The neu al ne wo k’s de ailed s uc u e is illus a ed in Fig. 4..
Figu e 4.: The neu al ne wo k s uc u e
3. RESULTS AND DISCUSSIONS
3.1. Valida ion Tes s
To e alua e he e ec i eness o suppo ec o machines (SVMs) and neu al ne wo ks (NN) in p edic -
ing hyd odynamic pa ame e s, alida ion es s we e conduc ed using a da ase o 18,240 samples no used
in he ini ial aining. The pe o mance o hese wo models was compa ed ac oss a ying numbe s o ain-
ing samples pe ba ch (2, 10, 50, and 100), co esponding o o al aining inpu s o 2,400, 12,000, 60,000,
and 120,000 espec i ely. These alida ion es s help e i y op imal model pe o mance and iden i y any
o e i ing o unde i ing issues.
P edic ion accu acy is measu ed using he pe cen age e o o mula:
E o =BEMResul s −MLResul s
BEMResul s ∗100% (18)
The alida ion esul s e eal subs an ial di e ences in pe o mance be ween he wo me hods. Ini ially,
when using only 2 samples pe ba ch, SVMs ou pe o m neu al ne wo ks in gene al. Howe e , while SVM
pe o mance emains ela i ely cons an wi h inc eased aining samples, neu al ne wo ks demons a e sig-
ni ican ly imp o ed accu acy as aining sample size g ows. By he ime he ba ch size eaches 50 samples,
neu al ne wo ks su pass SVMs o p edic ing all h ee pa ame e s: added mass, damping, and exci a ion
o ce.
Figu es 5.–7. illus a e he e o dis ibu ion h ough boxplo s o each pa ame e . Boxplo s depic
he median as he cen al line, wi h he op and bo om edges ep esen ing he 75 h and 25 h pe cen iles
espec i ely. Ou lie s, indica ed by ed c osses, all ou side 1.5 imes he in e qua ile ange om he box
edges, and whiske s ex end o he maximum and minimum non-ou lie alues.
A c i ical limi a ion o SVMs is hei di icul y p edic ing adia ion damping, pa icula ly wi h smalle
aining samples. E o s can each as high as 150% wi h jus wo samples pe ba ch. Al hough inc easing
he aining da a size educes his e o o app oxima ely 25%, he damping da a’s complexi y p e en s
accu a e modeling using linea , polynomial, o Gaussian ke nels. Thus, despi e pe o ming adequa ely o
added mass and exci a ion o ce, SVMs consis en ly s uggle wi h damping p edic ions.
In con as , neu al ne wo ks achie e signi ican ly lowe e o a es ac oss all h ee pa ame e s, e en
wi h small sample sizes. Wi h only wo samples pe ba ch, neu al ne wo ks main ain e o s unde 6%,
imp o ing d ama ically o unde 2% wi h inc eased aining samples. A 50 samples pe ba ch, neu al
ne wo k p edic ions exhibi e o s o less han 0.5%, demons a ing high accu acy and sui abili y o his
applica ion.
6
Figu e 5.: Added mass e o box plo o SVM and NN model ained using 2, 10, 50 and 100 samples pe
ba ch
Figu e 6.: Radia ion damping e o box plo o NN model (le ) and SVM model ( igh ) ained using 2, 10,
50 and 100 samples pe ba ch.
Figu e 7.: Exci a ion o ce e o box plo o SVM and NN model ained using 2, 10, 50 and 100 samples
pe ba ch
Based on hese alida ion esul s, neu al ne wo ks clea ly o e supe io pe o mance compa ed o
SVMs, especially o complex hyd odynamic p edic ions. Fo he mos accu a e p edic ions, aining neu al
ne wo ks wi h 50 samples pe ba ch (60,000 o al samples) is ecommended, achie ing e o s below 0.5%.
Fo as e aining while main aining accep able accu acy, educing he aining size o 10 samples pe ba ch
7
(12,000 o al samples) is easible, wi h p edic ed e o s emaining below app oxima ely 2%.
3.2. Op imal Shapes and Pe o mance
Two dis inc objec i e unc ions we e analyzed o op imize he WEC: one maximizing a e age powe
ou pu , and he o he balancing no malized powe ou pu and cu en d ag pe o mance. The objec i e
unc ions, wi h he goal o max (x), a e de ined as:
1=P= {x1, x2, ... y41, y42, T, N, mb, Lq}
2=P
Pmax
+Fmax −F
Fmax
= {x1, x2, ... y41, y42, T, N, mb, Lq}(19)
These unc ions a e e alua ed wi h ixed equali y cons ain s:
m= 80kg;H= 0.5m; = 2m/s;(20)
whe e P ep esen s he a e age powe ou pu (W), F he d ag o ce (N), mbuoy mass, Hwa e heigh , and
cu en speed.
Addi ionally, he ollowing inequali y cons ain s guide he buoy shape op imiza ion:
case1 :
si(j1)≤si(j1+ 1)
whe e, i ={x, y}, j1={1,2,3}
case2 :
si(j1)≤si(j1+ 1)
si(j2)≥si(j2+ 1)
whe e, i ={x, y}, j1={1,2}, j2={3}
case3 :
si(j1)≤si(j1+ 1)
si(j2)≥si(j2+ 1)
whe e, i ={x, y}, j1={1}, j2={2,3}
case4 :
si(j1)≥si(j1+ 1)
whe e, i ={x, y}, j1={1,2,3}
(21)
plus:
0.05 ≤sx≤0.5
0.05 ≤sy≤0.3
0.01 ≤s (1,2) ≤0.5
0.01 ≤s (3,4) ≤0.3
0.5≤a ea ≤0.6
5≤T≤10
(22)
whe e wi h shape ec o [sx, sy, s ]uni o m, and a ea uni o m2, and wa e pe iod Tuni o sec.
The op imal PTO pa ame e s and co esponding pe o mance a e p esen ed in Table 1., while Fig. 8.
illus a es he op imized buoy shapes unde he di e en objec i e unc ions. The esul s indica e ha op i-
mal PTO pa ame e s emain consis en ac oss objec i es due o hei di ec in luence on powe ou pu alone.
Howe e , signi ican shape di e ences occu because d ag pe o mance is conside ed only in one objec i e.
Maximizing a e age powe only ( 1) pushes he sea ch owa d ulle sec ions ha aise exci a ion nea
he a ge pe iod. When we add he no malized d ag e m ( 2), he shape shi s o a slimme unde wa e
p o ile wi h smoo he shoulde s. Tha cu s p ojec ed a ea and c oss- low leng h, so cu en d ag d ops,
while wa e plane and olume s ay high enough o good hea e esponse. This ade-o explains why he
igh igu e in Fig. 8 looks mo e “planing hull” and gi es a be e alue o 2e en hough he PTO se ings
in Table 1. ba ely change. In sho , geome y ca ies mos o he d ag educ ion, while he PTO s ill a ge s
elec ical powe . In oducing cos conside a ions, including buoy and PTO sys em cos s, could u he e ine
8
he op imiza ion, bu assigning cos weigh s is a bi a y and dependen on p ojec -speci ic needs, and hus
was excluded om his s udy.
Compa ed o Lin e al.’s op imal boa -shaped buoy [20], he op imized shape using he combined powe
and d ag objec i e closely ma ches hei esul , bu o e s g ea e ealism due o he inclusion o de ailed
PTO componen modeling and con ol op imiza ion.
Table 1.: Op imiza ion pe o mance o objec i e unc ion 1(powe , W) and 2(no malized powe plus
d ag) a H= 0.5m, T∈[5,10] s, = 2 m/s
Objec i e unc ion Op imal PTO pa ame e s Pe o mance
1N= 80, mb= 0.001, Lq= 0.0588 167
2N= 80, mb= 0.001, Lq= 0.0494 1.49
Figu e 8.: Op imal shape sol ed using a e aged powe as objec i e unc ion (le ) and using no malized
powe plus cu en d ag pe o mance as objec i e unc ion ( igh ).
4. CONCLUSION
In his s udy, we in oduced a p ac ical app oach using machine lea ning me hods o op imize bo h he
shape and PTO con ol o a poin abso be WEC. Two popula machine lea ning models, SVMs and neu al
ne wo ks, we e es ed o p edic key hyd odynamic pa ame e s, including added mass, adia ion damping,
and exci a ion o ce.
Ou alida ion showed ha neu al ne wo ks p o ided be e o e all pe o mance, especially o adia ion
damping. Al hough SVMs handled simple p edic ions like added mass and exci a ion o ce well when
ewe aining samples we e used, hey s uggled signi ican ly wi h he mo e complex damping da a.
By in eg a ing de ailed PTO componen modeling and con ol in o he op imiza ion p ocess, we we e
able o c ea e mo e ealis ic and e ec i e WEC designs han adi ional me hods. Addi ionally, conside ing
bo h d ag o ces and powe ou pu du ing op imiza ion esul ed in no ably di e en buoy shapes, highligh -
ing he impo ance o balancing mul iple objec i es o p ac ical designs.
Ou esul s indica ed ha neu al ne wo ks ained on la ge da ase s (50 samples pe ba ch, o aling
60,000 samples) achie ed he highes accu acy, keeping e o s below 0.5%. Howe e , when aining speed
is a p io i y, neu al ne wo ks ained on smalle da ase s (10 samples pe ba ch, o aling 12,000 samples)
s ill p o ided easonably accu a e esul s, wi h e o s a ound 2%.
In conclusion, his esea ch demons a es ha machine lea ning me hods can signi ican ly imp o e he
e iciency and ealism o WEC op imiza ion, o e ing enginee s a aluable ool o design mo e e ec i e
ma ine ene gy echnologies.
Acknowledgmen s
The au ho s exp ess hanks o he unding suppo p o ided by he O ice o Na al Resea ch (ONR)
h ough g an s # N00014-23-1-2100 and # N00014-21-1-2152.
9
