Renewable Energy 97 (2016) 504e517

Contents lists available at ScienceDirect

Renewable Energy
journal homepage: www.elsevier.com/locate/renene

Prediction and optimization of wave energy converter arrays using a

machine learning approach
Dripta Sarkar a, c, *, Emile Contal b, Nicolas Vayatis b, Frederic Dias a, b
a
UCD School of Mathematics and Statistics, University College Dublin, Belfield, Dublin-4, Ireland
b
Centre de Mathematiques et de Leurs Applications (CMLA), Ecole Normale Superieure de Cachan, CNRS, Universit

e Paris-Saclay, 94235 Cachan, France
c
Department of Engineering Science, University of Oxford, Parks Road, Oxford OX13PJ, United Kingdom

a r t i c l e i n f o a b s t r a c t

Article history: Optimization of the layouts of arrays of wave energy converters (WECs) is a challenging problem. The
Received 16 May 2015 hydrodynamic analysis and performance estimation of such systems are performed using semi-analytical
Received in revised form and numerical models such as the boundary element method. However, the analysis of an array of such
4 May 2016
converters becomes computationally expensive, and the computational time increases rapidly with the
Accepted 25 May 2016
Available online 13 June 2016
number of devices in the system. As such determination of optimal layouts of WECs in arrays becomes
extremely difficult. In this paper, a methodology involving multiple optimization strategies is presented
to arrive at the solution to the complex problem. The approach includes a statistical emulator to predict
Keywords:
Wave-energy
the performance of the WECs in arrays, followed by an innovative active learning strategy to simulta-
Oyster neously explore and focus in regions of interest of the problem, and finally a genetic algorithm to obtain
OWSC the optimal layouts of WECs. The method is extremely fast and easily scalable to arrays of any size. Case
Optimization studies are performed on a wavefarm comprising of 40 WECs subject to arbitrary bathymetry and space
Active learning constraints.
Genetic algorithm © 2016 Elsevier Ltd. All rights reserved.
Gaussian process

1. Introduction is to understand the effect of the interactions on the performance of

Arrays of wave energy converters (WECs) have been extensively
studied in the literature starting with the pioneering work of [1].
Notably, Budal [1] used the point-absorber approximation which
assumes the dimensions of the WECs are much smaller than the
wavelength of the incident wave field and therefore the diffracted
wave field can be considered to be negligible. Since then, advances
in numerical and analytical techniques (see e.g. Refs. [2e5]) in the
analysis of wave-structure interactions have enabled the investi-
gation of the behavior of arrays of WECs of arbitrary shapes, taking
into account the effects of both the diffracted and the radiated wave
fields. Numerical boundary element codes (e.g. NEMOH and
WAMIT developed at Ecole Centrale de Nantes and MIT, respec-
tively) are widely used to understand the effects of the hydrody-
namic interactions and to quantify the performance of the WECs
and the array as a whole. However, such methodologies involve
increased computational cost and resources. The general objective
* Corresponding author. Department of Engineering Science, University of
Oxford, Parks Road, Oxford, OX13PJ, United Kingdom.
E-mail address: dripta.sarkar@eng.ox.ac.uk (D. Sarkar).
the WECs and to determine layouts which would maximize the
power captured from the whole system. In order to quantify the
effects of the interactions on the performance of the array, Budal [1]
defined the q factor which is the ratio of the net power captured
(ideally the maximum possible) by the array to the power absorbed
by the same number of WECs in isolation. Budal [1] observed that it
is possible to have q factors much larger than 1 for particular wave
frequencies. But the peaks in the q factor are accompanied with
wide troughs in its variation, and since a real ocean is poly-
chromatic, Budal suggested that a properly designed array config-
uration should minimize the effect of the destructive influences.
However, the identification of optimal layouts for a particular
wave-climate is still a big challenge. The complexity of the opti-
mization problem is manifold. The number of WECs in arrays can
vary from one site to another, and separate optimizations needs to
be performed in each case. Every single evaluation of the
numerical/semi-analytical models has a computational cost which
increases with the size of the array. In addition, there can be various
constraints to such a problem (e.g. bathymetry variations for
nearshore WECs). Child and Venugopal [6] presented a parabolic
intersection method and a genetic algorithm to arrive at the

http://dx.doi.org/10.1016/j.renene.2016.05.083
0960-1481/© 2016 Elsevier Ltd. All rights reserved.
 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517 505

optimal layouts of arrays. While the parabolic intersection
approach uses simple calculations for a quick estimate of the array
layouts, the genetic algorithm requires many evaluations of a semi-
analytical (or numerical) method which is computationally
expensive. Such a direct application of sequential optimization
techniques also implies that if the number of WECs is changed, a
new set of evaluations needs to be computed and analysed. In this
work we propose a fast and scalable approach to address the
challenge of determining the best layout for any number of WECs
and arbitrary bathymetry constraints. We first use an active
learning strategy to train a statistical emulator of the individual
WECs inside the array. We then predict the performance of the
whole wavefarm by evaluating only the quasi-instantaneous
emulator. The optimization of the layout under the various con-
straints is then performed on the predicted performances with a
genetic algorithm designed specifically for this task.
The analysis in this work is performed on a well known WEC-
the Oscillating Wave Surge Converter. A number of studies are now
available which have looked at various aspects of the device (see
e.g. Refs. [7,8]). Most recently, a mathematical model was devel-
oped in Ref. [9] to analyse the behavior of the OWSCs in a wave farm
and the simulations in this work will use this model. The overall
performance of the array is decomposed as the sum of the powers
captured by each individual WEC. In a realistic scenario, the seas are
highly irregular, and the interaction effects on a particular WEC due
to WECs located away from it are largely diminished. It is reason-
able to assume that individual WECs in an array are predominantly
influenced by those located very close to them. Our approach tar-
gets the approximated performance where we simplify the model
of the individual WECs in order to take into account only a limited
number of interactions. A WEC located inside the array is strongly
influenced by the two WECs which are nearest to it, i.e. one on each
side (see Fig. 2). To model the behavior of such a WEC, we consider a
3-WEC cluster and focus on the behavior of the central WEC. On the
other hand, the edge WECs are modelled using a 2-WEC cluster.
Our prediction function is based on Gaussian process regression
trained with an active learning strategy, the Gaussian Process
Upper Confidence Bound with Pure Exploration algorithm (GP-
UCB-PE) [10]. Since the final goal of the work is to determine
optimal layouts of an array, we are more interested in layouts which
result in good performance of the individual WECs. The active
learning strategy determines sequentially which configurations of
the 3-WEC and 2-WEC clusters should be evaluated aiming to
improve the predictions of the optimal clusters using the least
number of evaluations possible. The algorithm performs a trade-off
between exploitation and exploration looking for the maximum
while exploring clusters with high uncertainty.
In our optimization methodology, we incorporate some con-
straints which are relevant to the problem. The OWSCs are near-
shore WECs with depth specific designs, and as such bathymetry
will play a significant role in deciding their locations. We consider
an upper and a lower bound on the bathymetry contours, within
which the placement of the centre of the OWSCs is restricted.
Although the mathematical model (for simulations) is based on a
constant water depth assumption, the bathymetry constraints in
the optimization problem take account of the spatial limitations
imposed by the depth variations at real locations, in the placement
of the WECs. In a practical situation, proper utilisation of space is
also an important consideration and to account for that an upper
bound in the distance separating the centres of the first and the last
WEC in the y0 direction (see Fig. 1) is fixed at a particular value.
The methodology of the simulations, statistical emulator and
the machine learning algorithm will be illustrated in the following
sections. Later in x5, case studies are performed on the optimization
of a wavefarm comprising of 40 OWSCs, followed by discussions
and conclusion.
2. Simulation model
The simulations for the various layouts are performed using the
mathematical model of [9]. A wavefarm comprising of M WECs (see
Fig. 1) is considered, and so the 2 M variables that need to be
optimized are x01 ; y01 ; …x0M ; y0M , which are the horizontal coordinates
of the centres of the OWSCs in the x0 and y’ directions respectively.

Fig. 1. Geometry of the physical system (a) side view of one of the WECs (b) top view of a wavefarm comprising of M WECs. The right most OWSC is considered to be the first WEC,
with the numbering increasing towards the left. The horizontal spatial coordinates of the centres of the OWSCs are: (x01 ; y01 ), (x02 ; y02 ), …..., (x0M ; y0M ).

Fig. 2. The methodology considers only the interaction of the WECs with their neighbouring ones. As such the two edge WECs located at the two extremes of the wavefarm can be
modelled with a simplified 2-WEC cluster configuration while those lying in the interior with a 3-WEC cluster arrangement. The right most OWSC is considered to be the first WEC
of the array. The horizontal spatial coordinates of the centres of the OWSCs are: (x01 ; y01 ), (x02 ; y02 ), …..., (x0M ; y0M ).
506 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517

The wave farm is considered to be located at a constant water depth

h0 . The OWSCs have width w0 and they perform pitching motion i X
þ∞ XM

about a horizontal axis which is located at a distance c0 above the

4ðaÞ ðx; y; zÞ ¼  kn Zn ðzÞ wm ðx  xm Þ
8 n¼0 m¼1
sea bottom. The problem is formulated in the framework of linear
potential flow theory which assumes inviscid fluid and the flow to X
P Z1  1=2
ðaÞ
be irrotational and incompressible. After non-dimensionalization  apnm 1  u2 Up ðuÞ
p¼0
of the physical variables, the time and vertical dependence are 1
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi !
separated out from the governing system of equations which yields  2
ð1Þ
kn ðx  xm Þ2 þ y  uwm2þ2ym
C
a 2D system that needs to be solved. Using Green’s integral equa- H1
tion formulation and employing Green’s function, the 2D system sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 du;
yields hyper-singular integrals in terms of the jump in potential  2
ðx  xm Þ2 þ y  uwm2þ2ym
C
across the two sides of the OWSCs, which are then solved using the
Chebyshev polynomial of the second kind. The governing equations
of the problem are: (7)

e the Laplace equation in the fluid domain where xm and ym are the x and y co-ordinates of the centre of the m-
th OWSC, Up(u) is the Chebyshev polynomial of the second kind,
V02 F0 ¼ 0; (1) aapnm are the numerically obtained solutions using a collocation
scheme (see Ref. [9]), Zn(z) are the well-known vertical eigenmodes
where V0 f 0
¼ ðf;x0 0 ; f;y0 0 ; f;z0 0 Þ is the nabla operator; subscripts with and kn are the solutions to the dispersion relation
commas denote differentiation with respect to relevant variables; u2 ¼ gkntanh(knh).
Once the solution is known, the hydrodynamic parameters are
e the free surface kinematic-dynamic boundary condition obtained which are then utilized to solve for the body equation of
motion of the WECs. Note, these solutions are for a wave of specific
F0;t0 t0 þ gF0;z0 ¼ 0; z0 ¼ 0; (2) period and amplitude. The behavior in a wave spectrum is obtained
by discretizing the spectrum into a particular number of fre-
where g is the acceleration due to gravity; quencies and then performing linear superposition of the regular
wave outputs [11]. The analysis is performed using a Bretschneider
e the no-flux boundary condition on the sea bed spectrum

 4 
F0;z0 ¼ 0; z0 ¼ h0 ; (3) 2 4 5
Sðf Þ ¼ 0:257H1=3 T1=3 f exp  1:03 T1=3 f (8)

e the kinematic boundary condition in the WEC boundaries (see Ref. [12]) to model the wave climate of the most probable sea


 state at the Isle of Lewis in Scotland, a planned site for wave farms,
F0;x0 ¼ q0m;t0 z0 þ h0  c0m H z0 þ h0  c0m ; with the significant wave height H1/3 ¼ 1 m and the significant
x0 ¼ x0m ±ε0 ; ε0 /0; (4) wave period T1/3 ¼ 8.24 s. Note in (8), f is the frequency of the
0
0 0 incident wave. The spectrum was covered with 51 equispaced fre-
yAm <y < yBm ; m ¼ 1; ::; M; quencies starting at 0.05 Hz and ending at 0.5 Hz. The power from
0 each OWSC was computed according to the relation
where x0m is the x0 coordinate of the centre of the mth flap, yAm and
0
yBm are the y0 coordinates corresponding to the two edges of the
1XN
device and H is the Heaviside step function. Pi ¼ npto Siv ðfl ÞDf (9)
2
The non-dimensional parameters of the problem are chosen to l¼1
be
where Siv ðfl Þ is the spectral density of velocity of the ith OWSC at
pffiffiffiffiffiffiffiffiffiffiffi
ðx; y; z; r Þ ¼ ðx ; y ; z ; r Þ=w ; t ¼ g=w0 t 0 ;
0 0 0 0 0 frequency fl, N ¼ 51 and npto is the optimal power take-off damping

pffiffiffiffiffiffiffiffi 0   (5) coefficient which is obtained using an optimization procedure
F ¼ F0 gw0 AI ; qm ¼ q0m ε; aiming to maximize the power capture by an isolated OWSC.
The effect of the interactions on the performance of the indi-
with ε ¼ A0I =w0 being the small parameter. Assuming the motions of vidual WECs can be quantified using the qmod factor (see Ref. [13])
the WEC to be harmonic, the velocity potential and the angular which is defined as
deflections can be expressed as F ¼ Re{4(x,y,z)eiut} and
qm ¼ RefQm eiut g; m ¼ 1; ::; M respectively, with u being the P i  Pisolated
angular velocity. Thanks to linearity, the spatial potential 4 can be qmod;i ¼
Pisolated
decomposed into
where Pi is the power captured by the i-th WEC in the array, while
X
M
4 ¼ 4I þ 4ð0Þ þ Va 4ðaÞ ; (6) Pisolated is the power captured by the same WEC in isolation. The q
a¼1 factor (ratio of the total power captured from the array to that by
the same number of WECs in isolation) and qmod factor can be
where 4(I) is the incident wave potential, 4(0) is the solution to the related as
diffraction potential when all the OWSCs are held fixed, Va ¼ iuQa
while 4(a), a ¼ 1,2,..,M indicates the solution to the a-th radiation 1 X
M

problem where the a-th OWSC is oscillating and the others are
q¼1þ qmod;i :
M
i¼1
stationary.
The solution to the spatial potential is obtained as ~mod and q
We denote by q ~, the approximate form of the qmod and
 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517 507

q factor, considering only the effects due to the local interactions.

The additive model for q~ can be expressed as

 1
0 
~ x01 ; y01 ; …x0M ; y0M ¼ 1 þ
q ~mod
q2 x2  x01 ; y02  y01
M

0 
~mod
þq2 xM  x0M1 ; y0M  y0M1
X
M1
0
þ ~mod
q3 xi  x0i1 ; y0i
i¼2
!

 y0i1 ; x0iþ1  x0i ; y0iþ1  y0i

where q ~mod
2
~mod
and q 3 are respectively the models for the 2-WEC
clusters (see Fig. 4) and the 3-WEC clusters (see Fig. 3). We pre-
dict the interaction factor of the individual WECs using two sta-
~mod
tistical emulators, one for q3
~mod
and one for q 2 .

Fig. 4. The WEC-2 (in red) models the behavior of the edge WEC of the array. The
3. Function estimation through Gaussian process approach position of WEC-1 (in blue) is varied such that its left edge is located within the striped
square box. (For interpretation of the references to colour in this figure legend, the
reader is referred to the web version of this article.)
We describe here how we predict the unknown values of a
function f : X /ℝ given a data set of evaluations of f. This approach
~mod
will be used for predicting q the interaction factor of a WEC with
3 model. The GP is defined in terms of its mean m and covariance
two neighbours (where X ⊂ℝ4 , the relative coordinates of the function k [14] as
neighbours inside two 100 m boxes as shown in Fig. 3) and q ~mod
2 the
interaction factor of a WEC with a single neighbour (where X ⊂ℝ2 f  G P ðm; kÞ (10)
as shown in Fig. 4).
with m: X /ℝ and k: X  X /ℝ. Note in (10) the symbol ~ means
3.1. Probabilistic setup ‘distributed according to’. For any set of input vectors x1,x2,…,xl in
X , the value of the function is given by a multivariate Gaussian
We assume that the unknown function f is a random sample random variable
from a Gaussian process (GP), which can be viewed as a general
form of the multi-normal distribution. GP is a natural choice for our ðf ðx1 Þ; f ðx2 Þ; …; f ðxl ÞÞ  N ðm; KÞ
problem based on the observation that there is high correlation
where the mean vector m and the covariance matrix K are expressed
between layouts which are strongly similar. Let {(x1,y1),…,(xl,yl)} be
as
a training set of l noisy observations, where yt ¼ f(xt) þ εt, t ¼ 1,2,..,l
and X is an arbitrary vector space in ℝd . We consider that the
m½xi  ¼ mðxi Þ
variables εt are distributed according to independent Gaussians
N ð0; s2 Þ, they represent the numerical noise of the semi-analytical and

Fig. 3. The general layout of the 3-WEC configuration. The position of the central WEC is fixed while the positions of WEC-1 and 3 vary such that the left side edge of WEC-1 lies
inside the striped square box and the right edge of WEC-3 inside the dotted box.
508 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517


 the semi-analytical model we can compute in parallel on a cluster

K xi ; xj ¼ k xi ; xj : of B computers. The methodology of the active learning algorithm is
illustrated in the sequel.
The widely used squared exponential kernel function for the
covariance function can be written as
 
kðx1 ; x2 Þ ¼ exp  ðx1  x2 ÞT Dðx1  x2 Þ (11) 3.2.1. The GP-UCB-PE algorithm
We first build a high confidence interval within which the
function f is expected to be located with high probability as follows
where D is a diagonal matrix and the inverse square root of the
diagonal terms defines the characteristic length scale. The squared þ
exponential kernel function is simple and yet can capture the b
f t ðxÞ ¼ m
b t ðxÞ þ bt b
s t ðxÞ;

structures in the data necessary in the present application. Some b
f t ðxÞ ¼ m
b t ðxÞ  bt b
s t ðxÞ
other choices for kernel function include Matern, periodic [15],
þ 
spectral-mixture [16]. The Matern kernel function can model where bf t and b f t are the upper and lower confidence bound on f
rougher variations (less smooth functions) and is ideally suitable in and the bt parameter regulates the width of the confidence region
practical data sets (possessing sharp changes), while the periodic (see Ref. [10]). We then define the region <t we are interested in,
kernel function is appropriate for periodic datasets. Spectral mix- which contains the best configuration with high probability. Let y⋄t

tures kernel can model a wide range of problems, however it be the maximum of the lower confidence bound b ft
typically has a number of hyperparameters (which needs to be

determined) higher than that of the squared exponential kernel. In y⋄t ¼ max b
f t ðxÞ
our analysis, we use a modified version of the squared exponential x2X

kernel function so as to incorporate the symmetrical nature of the The value for y⋄t describes the lower confidence bound for the
problem with respect to the x0 axis. As a matter of fact, the behavior optimum at each iteration and the region <t is defined as
of the particular WEC which is modelled (e.g. WEC-2 in Fig. 3)
doesn’t change if the y0 coordinates of the WECs on its two sides are (  )
 þ
 b
interchanged. <t ¼ x2X  f t ðxÞ  y⋄t
 
Conditioned on the observations
 Y, we have that the distribution
of f is a GP such that f ðxÞY  N ð m s 2 ðxÞÞ, where
b ðxÞ; b
as shown in Fig. 5. The first of the B query points is chosen such that
h i1
m
b ðxÞ ¼ kðX; xÞT K þ s2 I Y (12) þ
x0t ¼ arg max b
f t ðxÞ:
x2X
and
Using the fact that the uncertainty does not depend on Y but
h i1
only on X, once we know x0t we can update it to s b 2t;1 before having
s 2 ðxÞ ¼ kðx; xÞ  kðX; xÞT K þ s2 I kðX; xÞ
b (13)
the associated observation. The remaining B  1 points are chosen
by Pure Exploration inside <t . We greedily select the configuration
are the prediction and uncertainty respectively. Note in (12) and b 2t;b
with highest updated uncertainty s
(13), X indicates the training point and k(X,x) is the vector of size l
of the covariance terms between X and x. The hyperparameters of
the setting, which are the diagonal terms of the matrix D in (11),
and the noise variance s2, are chosen so as to maximize the pseudo-
likelihood of our training data set (see Ref. [14]). The optimization
of the hyperparameters is performed with a Quasi-Newton method
using the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm
[17].

3.2. Search of design points

The larger the data set of observations is, the better the pre-
dictions are. But each training point requires to run a computa-
tionally expensive semi-analytical method. Since we are more
interested in configurations leading to good performance, we use a
sequential strategy to select the most informative training points,
the Gaussian Process Upper Confidence Bound with Pure Explora-
tion algorithm (GP-UCB-PE) [10]. This algorithm proceeds in an
iterative way so as to reduce the uncertainty of the prediction of the
best configuration. It has recently been used in the estimation of
maximum wave run-up of tsunami waves behind conical islands
[18]. At iteration t after having computed m b t and s b 2t , we simulta-
neously explore the configurations with high uncertainty and focus
on the ones with high prediction. We choose a batch of B unknown
configurations using a tradeoff between m b t and bs 2t , we then query
the semi-analytical model on those B configurations to improve our
prediction, and continue to iteration t þ 1. The value of B is arbi-
trarily equal to 20 in our setting, it is typically the number of runs of
Fig. 5. A simple illustration of the GP-UCB-PE strategy showing the first two query
points. The horizontal axis represents the x coordinate while the vertical axis indicate
the function value. The red continuous line is the prediction m b while the black
continuous lines bound the high confidence region. The lower confidence bound on
the optimum is shown by the blue dotted line which is the maximum of the lower
confidence bound. The intersection of the blue dotted line and the upper confidence
bound defines the limits of our region of interest < which is shown in blue in the
figure. The first point x0 is chosen to be the maximum of the upper confidence bound.
The dotted line in black shows the updated variance confidence after having selected
x0. The next query point x1 is chosen such that it has the maximum updated uncer-
tainty inside <. (For interpretation of the references to colour in this figure legend, the
reader is referred to the web version of this article.)
 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517 509

Table 2
2
s t;b ðxÞ;
xbt ¼ arg max b 2-WEC cluster iterations.
x2X
Iteration no. t 2t ct Þ
maxð m
with b varying from 1 to B  1. Under mild probabilistic assump- 1 0.00320 0.116084
tions the queries selected with this active learning strategy are 2 0.00001 0.121031
shown to converge to the global optimum of the unknown function
at a rate (tB)½ (with omission of logarithmic factors).
approximate the power captured by any layout in a fast and scalable
3.2.2. Experimental design manner. Once we trained the two statistical emulators of the per-
Since the GP-UCB-PE algorithm requires to fix the hyper- formance of a WEC inside or at an edge of an array, it is easy to sum
parameters of the GP, we first optimize their value on an initial data the outputs for the M WECs to get a prediction bq of the q ~ factor.
set constructed following a classical experimental design approach.
A simple choice would be a regular grid in which every coordinate

 1 mod
0 
is divided into a particular number of equally spaced intervals. But q x01 ; y01 ; …x0M ; y0M ¼ 1 þ
b b
q2 x2  x01 ; y02  y01
such an approach is highly inefficient because of the high compu- M
tational costs associated with it, e.g. if there are l input variables, mod
0 
þb
q2 xM  x0M1 ; y0M  y0M1
each of which is divided into p equidistant, non-overlapping in-
tervals, then the total number of evaluations required to cover the X mod

M1

input space is pl. In our analysis, we employ the latin hypercube þ b

q3 x0i  x0i1 ; y0i
i¼2
sampling (LHS) in which the range of each input variable is divided !
into a particular number (say p) of equally probable, non- 
 y0i1 ; x0iþ1  x0i ; y0iþ1  y0i
overlapping intervals. Then values are chosen randomly from
each of the intervals for every variable and random combinations
are made to arrive at the sample points. We then use the maxi- mod mod
where b q2 and bq3 are the predictions given by the statistical
mization of the minimum distance as an additional constraint in
emulators. In order to optimize the array given some bathymetry
order to achieve maximal coverage of the input space. We chose
constraints, we explore the space of valid layouts using a genetic
800 sample configurations by LHS to initialize the 3-WEC cluster
algorithm [19]. The optimization routines of a genetic algorithm
predictor and 200 for the 2-WEC cluster.
(GA) are inspired from the evolutionary ideas of natural selection
and genetics. At each step of the iteration (called generation), a
3.2.3. Stopping criterion
proportion of the existing population of solutions (called in-
We stop the GP-UCB-PE algorithm when we know that is has
dividuals) is selected based on their individual scores (called fitness
discovered the best configuration with high probability, that is
value), to breed the next generation. In this particular problem, the
when the region <t shrinks to a single point. For the 3-WEC cluster
fitness function is b q . The optimization was performed using the
predictor the active learning strategy ended after six iterations with
genetic algorithm solver of the MATLAB Global Optimization
B ¼ 20, which means 120 evaluations in addition to the initial 800
Toolbox. The population size of each generation of the GA is chosen
from the LHS. In the case of the 2-WEC cluster predictor the algo-
to be 211, since the computational time per unit evaluation ceased to
rithm ended after a single iteration with B ¼ 20 which is explained
~mod factor with a single improve beyond it. Two important operations in a GA are crossover
by the highly predictable value of the q
and mutation. In crossover, two parents having good performance
neighbouring WEC. In order to quantify the extent of the region <t ,
are selected from the previous population and combined to create a
a factor 2t is defined as the ratio of points inside the region of in-
child for the present population. While in the mutation, the later is
terest <t to the total number of uniform design points. Tables 1 and
ct Þ and 2t at each iteration t for the 3- obtained by making random changes to a selected parent. The
2 show the values of maxð m
initial population of the WEC layouts is selected such that the ba-
WEC and 2-WEC cluster, respectively. The number of uniform
thymetry constraints are all satisfied. In our approach, we utilize
points was set to 106 for the 3-WEC case and 105 for the 2-WEC
custom crossover and mutation operators such that all the con-
case. With each iteration, the value of 2t decreases, signifying the
straints are kept satisfied. We illustrate our methodology for
reduction in the spatial extent of <t , while the maximum of the
initialization, crossover and mutation in the following paragraphs.
predictions of the performance factor q ~mod converges.
Selection of Initial Points: The initial population of the WEC
layouts is chosen randomly, while satisfying the required con-
4. Optimization of the estimated additive model
straints at the same time. We first choose the spacings between the
neighbouring WECs in the y0 direction according to a uniform dis-
The determination of the optimal layouts is a challenging task,
tribution satisfying the space constraint (total extent of the wave-
due to the number of variables (2 M), and the various constraints.
farm in the y0 direction). Once the y0 coordinates of the centre of the
Thanks to the quasi-instantaneous emulator, we are able to
WECs are fixed, the x0 coordinates of the WECs are selected
sequentially in the following way (illustrated in Fig. 6). Let us
Table 1 suppose the x0 coordinate of the m-th WEC is already determined,
3-WEC cluster iterations. we sample the x0 coordinate of the m þ 1-th WEC uniformly in the
Iteration no. t 2t ct Þ
region satisfying both the bathymetry constraints and the square
maxð m
box constraints relative to the m-th WEC. Here and in the following
1 0.057526 0.158511 the bathymetry constraints are given by lower and upper limits of
2 0.008354 0.218310
the feasible region. If it is not possible to satisfy all the constraints,
3 0.001943 0.225993
4 0.000540 0.234827 we start again the operation from scratch.
5 0.000021 0.236200 Crossover: In crossover operation two good parent solutions
6 0.000003 0.236036 from the present generation are combined to produce a child for
7 0.000001 0.238663
the future generation. The y0 coordinates of the centre of the WECs
510 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517
Fig. 6. Illustration of methodology for the selection of initial points and the mutation
operation. The y0 coordinates of the WECs are selected randomly such that the distance
between the edges of the neighbouring WECs is less than 100 m. Once the y0 co-
ordinates of the centre of the WECs are fixed, the x0 coordinates are selected
sequentially. The figure shows the selection of the coordinates of the (m þ 1)-th WEC
with the position of the m-th WEC already determined. The x0 coordinate must lie
within the upper and lower limits of the bathymetry while also satisfying the square
box constraint. The interval which satisfies both the mentioned constraints is shown in
red in the figure, with the red circles indicating the limits, and the position of the WEC
is selected randomly in this interval. (For interpretation of the references to colour in
this figure legend, the reader is referred to the web version of this article.)
are taken to be the average of the ones of the two parents. The
coordinates are then determined in a manner, such that the centre
of the WECs satisfies the bathymetry constraints for all situations.
The x0 coordinates of the WECs of the parents are first normalised
with respect to the upper and lower bounds of the bathymetry. The
mean of the normalised x0 coordinate of the WECs of the two par-
ents is then computed and projected on to the already determined
y0 coordinate of the child, with the bathymetry constraints being
the bounds again, and this gives the position of the WECs of the
child (see Fig. 7).
Mutation: The mutation operation is done in the following way.
Random alterations are made on n consecutive WECs of a particular
parent, with the number n chosen from a Beta distribution. The
Beta distribution is described over a finite interval and is parame-
trized by two positive shape parameters, which for the present
problem are chosen such that the mean value of n for a 40 WEC
array is 8. Given n, we choose randomly n consecutive WECs from
the array and replace them in a similar manner to the selection of
the initial points, where we take care of the box constraints for the
eventual neighbours of the n WECs (see Fig. 8).
5. Results
The results and analysis are presented in three parts in this
section. In the first part, the efficiency of the statistical emulator of
the 3-WEC cluster and 2-WEC cluster are discussed. Then the
prediction function is used to describe the variation of b q for a
wavefarm comprising of 40 WECs in inline and staggered config-
urations with uniform spacings and symmetrical arrangements.
Lastly, optimizations of layouts of WECs in a wavefarm, subject to
various constraints, are performed using GA.
5.1. 3-WEC and 2-WEC clusters
We use the well-known ‘leave-one-out’ strategy to validate the
statistical model for predictions. One of the l training examples is
left out and the statistical emulator is trained using the remaining
points. The left-out evaluation is then utilized for validation. The
same process is repeated for all the l possible cases and the oper-
ation enables to quantify the accuracy of the predictions. Figs. 9 and
mod
10 show the prediction ( b q ) from the statistical emulator versus
those obtained from the simulations of the mathematical model
~mod ) for the 3-WEC and 2-WEC clusters, respectively. The diago-
(q
nal continuous line corresponds to perfect predictions. The circles
plotted in the figure are the predictions while the vertical dotted
lines are the 95% confidence intervals. For the 3-WEC cluster, there
is good agreement given the small number of training examples
and the complexity of the prediction problem. Almost all the actual
results lie within the high-confidence interval of the predictions.
The accuracy of the predictions is extremely good for values of q ~mod
around zero. Note that in those figures the statistical model is
trained using only the 800 and 200 points from the LHS for the 3-
WEC and 2-WEC clusters respectively. We can remark that the
prediction of the best 3-WEC is slightly under-estimated, which
justifies the need of the GP-UCB-PE algorithm to obtain better
predictions for good clusters. In the case of the 2-WEC cluster,
almost perfect predictions are observed with a smaller number of
simulation points. This is largely due to the smaller dimension of
the 2-WEC cluster problem (dimension two) than that of the 3-
WEC cluster (dimension four).
5.2. Wavefarm of 40 WECs: predictions
In this subsection, we analyse the variation of b
q for two simple
configurations of 40 WEC arrays in the inline and staggered con-
x0
figurations. In both cases, we consider uniform spacing and sym-
metrical arrangement of the WECs for simplicity, so that the output
can be easily quantified in terms of the input variables which are
the distances of separation between the neighbouring WECs in the
x0 and y’ directions (see Fig. 11).
5.2.1. Inline
This case corresponds to the situation in which the pitching axes
of all the WECs are coaligned i.e. there is no separation in the x0
direction. The only variable is the lateral distance of separation in
the y0 direction - y0d . In Fig. 12, the prediction bq is plotted as function
of the lateral spacing between the neighbouring WECs. There is a
clear maximum of b q which occurs at a separation distance of 76.5 m
between the centres of the neighbouring WECs for the spectrum
considered in this study. The inline system of a finite array of
OWSCs is well known to result in a near-resonant phenomenon
[20]. In fact resonance occurs for an infinite array of OWSCs or an
OWSC in a channel at/near cut-off frequencies of the symmetric
transverse sloshing modes in a channel. An extensive literature
now exists on the inline system and the readers are referred to
Refs. [8], [21], [20], [22]. In Ref. [9], a single simulation of a 40 OWSC
inline wave farm in monochromatic waves was performed, with a
separation of 20 m (y0d ¼ 46 m) between the edges of the neigh-
bouring OWSCs. The random waves simulation of the case resulted
in a q factor of 0.9. Using the prediction algorithm presented in this
paper, a bq factor of 0.89 is observed with the same separation of the
inline case.
5.2.2. Staggered
A staggered array has two rows of OWSCs. Here again we
 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517 511

Fig. 7. Illustration of the crossover operation - (a) position of the m-th WEC of parent P1, (b) position of the m-th WEC of parent P2 and (c) position of the m-th WEC of the child.
(x0 (,m,UB), y0 (,m,UB)) and (x0 (,m,LB), y0 (,m,LB)) are the coordinates of the upper bound and lower bound at the location of the m-th WEC. The y0 coordinate of the m-th WEC of the child is
simply the average of the same of the parents. The x0 coordinate of the m-th WEC of the parents is first normalized with respect to the upper and lower bounds of the bathymetry,
such that the normalized values of x0 (,m,LB) and x0 (,m,UB), are 0 and 1, respectively. The average of the normalized x0 coordinate of the two parents is then computed and projected at
the y0 coordinate (y0C;m ) of the child and un-normalized, to obtain the location of the m-th WEC.

consider uniform spacing between the neighbouring devices.
However, there are now two parameters instead of one - x0d and y0d .
The variation of b q with respect to the two spacing parameters is
plotted in Fig. 13. For small distances of separation, the cumulative
effect of the interactions on the performance of the array is mostly
destructive. The highest value of b q is not much different from that
of the inline case. In fact in Fig. 13, the x0 ¼ 0 m line corresponds to
the inline configuration. However, by staggering the array config-
urations, there is now a wider possible range of layouts by which
the maximum of b q can be achieved.
5.3. Optimal layouts
The optimization of the layouts of a wavefarm comprising of 40
WECs is considered in this work. In order to understand the general
arrangement of the WECs and to demonstrate the robustness of the
presented methodology, we consider five different arbitrary ba-
thymetry constraints which take account of the limitations in the
placement of the WECs due to depth variations. The centre of the
WECs must be located within the upper and lower bounds of the
bathymetry constraints. The upper bound for the spatial extent of
the wavefarm in the y0 direction (perpendicular to the direction of
incident waves) is set at 2500 m.
5.3.1. Genetic algorithm
Fig. 14 shows five layouts of wavefarms of 40 WECs with
different bathymetry constraints, obtained after 100 (2 M) ¼ 8000
generations, that is over 16 million predictions. The optimum value
of bq obtained for all the cases is similar. The layouts are well
distributed along the y0 direction, so that the full spatial extent is
512 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517

Fig. 8. In the mutation operation, n consecutive WECs (n ¼ 4) are randomly selected and replaced by a new distribution of the same number of WECs. In this figure, four consecutive
WECs, shown in blue in (a), are replaced by the new arrangement, shown in red in (b). The dashed red vertical lines, which are the edges of the WECs before and after the n
consecutive WECs indicate the y0 coordinate bounds within which the new configuration of the WECs is to be placed. (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)

mod
Fig. 9. Prediction of the performance ( b
q ) of the central WEC for the 3-WEC configuration using the leave-one-out scheme with 800 training examples.

properly used. Even in the case of the perfect inline and staggered while planning wavefarms and can be considered as a guideline for
layouts (see Figs. 12 and 13), we observe that close spacings lead to designers. Also note that b
q for the optimized layouts of the five
strong destructive influence on the overall performance of the bathymetry configurations is marginally less than the maximum b q
array. This suggests that clustering of devices should be avoided for the inline and staggered cases (maximum of b q  1.03, see
 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517 513

mod
Fig. 10. Prediction of the performance ( b
q ) of a WEC for the 2-WEC configuration using 200 training examples.

Fig. 11. The layout of an (a) inline and (b) staggered layout of an OWSC array with uniform spacings. The parameters x0d and y0d are the distances between the neighbouring WECs in
the x0 and y’ direction respectively.

Fig. 13. Variation of the b

q factor versus the spacings between the centres of the
neighbouring WECs in the x0 direction - x0d and that in the y0 direction - y0d , for a
uniformly staggered array comprising of 40 WECs.

Fig. 12. Variation of b

q versus the distance of separation between the centres of the
Figs. 12 and 13). This is due to the constraints imposed by the ba-
WECs for an inline configuration of 40 WECs with uniform spacings.
thymetry contours and the total spatial extent of the array in the y0
direction, which restricts the selection of layouts having higher
values of bq . In general, the WECs are oriented in arrangements
514 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517

Fig. 14. The optimal configuration of a wavefarm comprising of 40 OWSCs subject to various constraints, obtained with the presented methodology using genetic algorithm. The
OWSCs are shown in red. The blue and the pink coloured arbitrary curves indicate the upper and lower bathymetry constraints, respectively, and the centres of the WECs must lie
between them. In addition, the total extent of the wavefarm in the y0 direction must be less than 2500 m. The arrow from the top indicates the direction of incident waves. (For
interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

which can be considered as a combination of inline and slightly
staggered layouts, while following the bathymetry contours.
To check the accuracy of the results, simulations of the identified
optimal arrangement of WECs are performed using the mathe-
matical model of [9], where the interactions between all the WECs
are accounted for. Table 3 shows the comparison of b q (predictions)
and q (simulations) for the optimal configurations identified by the
GA. The good agreement between b q and q demonstrates the
robustness of the methodology presented in this paper, despite the
use of a simplified interaction approach.
5.3.2. Monte Carlo optimization
In order to check the effectiveness of the proposed optimization
methodology using GA, the same optimization is also performed
Table 3
Comparison of b q (predictions) and q (simulations) for the optimal configurations
identified by the presented methodology. The q factors are obtained from the sim-
ulations of the mathematical model which considers the interactions between all
the WECs, while bq are the predictions using the simplified interaction approach.
Bathymetry no. b
q q
1 1.021 1.005
2 1.016 1.006
3 1.012 1.006
4 1.014 1.003
5 1.02 1.004
using a Monte Carlo type method. In this approach, b q is evaluated
for a particular number of randomly selected arrangement of the
WECs. The methodology to obtain the random distribution of the
WECs is the same as that used for the selection of initial points and
mutation operation of the GA. The layout corresponding to the
maximum value of b q is considered to be the optimal arrangements
of the WECs. Over 16 million evaluations (same number as that
required with the GA approach) of the prediction function were
used for each case of the bathymetry. Fig. 15 plots the layout of the
optimal configurations obtained using the Monte Carlo approach.
The maximum of b q , corresponding to the optimal configurations
was always less than those obtained using the GA (see Table 4).
Notably, the arrangement of the WECs is much more arbitrary than
that of the latter (see Fig. 14).
In the recent work of [23], optimizations were performed on
small OWSC arrays (5 WECs) with equidistant spacings for some
specific layouts of the WECs. A different wave spectrum and a
slightly disparate configuration of the OWSCs were considered in
the latter analysis, and optimal spacings of x0d z50m and y0d z65m
were determined. These results are qualitatively close to the region
of constructive layouts identified in Fig. 13, although our analysis is
for larger wavefarms (40 OWSCs with a different configuration).
6. Discussion
The general approach of developing a statistical emulator and
then performing an optimization using genetic algorithm could be
 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517
Fig. 15. The optimal configurations obtained using a Monte Carlo type approach, with the number of evaluations of the prediction function being the same as that required using
GA. The maximum values of b q are less than that obtained using GA for all the five cases.
Table 4
Optimal b
q factor with 16 million predictions.
Bathymetry no. Genetic algorithm Monte Carlo
1 1.021 0.988
2 1.016 0.986
3 1.012 0.987
4 1.014 0.987
5 1.02 0.99
pursued for determining optimal layouts of other type of WECs (e.g.
Heaving WECs). The Heaving WECs would however require a
different simplified interaction model. For example, in case of
OWSCs, the back-to-back configurations (one OWSC directly
behind another) were ignored as such situations are not expected
in practical layouts. For Heaving WECs, interactions of all devices
within a certain radial distance could be considered. To develop a
statistical emulator, one could consider a simple 2-WEC cluster
with the sampling space being a semi-circle around a fixed WEC.
Such an approach could actually work for all WECs.
The OWSC has a broad bandwidth of energy capture and is less
sensitive to changes in the wave spectrum [24]. The dynamics of
this device is dominated by the diffraction phenomenon, with its
characteristics strongly governed by the variations in the excitation
torque [8]. The PTO dynamics being strongly dependent on the
radiation characteristics of the system, does not undergo much
modifications as the mutual interactions in the radiation problems
are less important for OWSCs. And so for simplicity, the PTO
515
damping is considered optimized (isolated OWSC case) prior to the
array optimization. However for certain other designs (e.g. modular
OWSC [see Ref. [25]) and applications (e.g. Venice gates [see
Refs. [26,27]) of the flap-type systems, the dynamical characteris-
tics due to the radiation phenomenon can become important when
any of the eigenfrequencies of the system of oscillating flaps lies
within the frequency range of the incident wave field, resulting in
large motions of the flaps. There are devices like the Heaving WECs
which work on the resonance principle and their performance
parameters are strongly dependent on the radiation characteristics.
The mutual interactions in an array of such devices have consid-
erable influence on the power take-off characteristics. Optimization
in such cases can require the inclusion of power take-off damping-
inertia as extra input parameters to the problem (in addition to the
spatial coordinates). Another aspect is the geometry of the device
which is generally considered to be fixed and is a standard
assumption. Manufacturing constraints do restrict the geometry to
a single design of the device, usually optimized to capture the
maximum power in isolation. From an optimization perspective, all
the above variables (be it geometrical or dynamical) can be
considered as extra input parameters. In general, our model can be
extended to more complex cases e.g. the orientation angle of the
OWSCs, the angle of incidence of waves, spectral parameters being
additional input parameters. However, it should be noted that
addition of new input parameters increases the dimension of the
problem and thereby the complexity of the optimization problem.
In our problem, it can be observed that the predictions for the 2-
WEC cluster (dimension-2) are better than in case of 3-WEC
516 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517
Fig. 16. The method can be extended to include the effect of the interaction of two of the nearest neighbours for a particular WEC. For such a system, a 5-WEC cluster models the
behavior of a WEC located well inside an array, a 4-WEC cluster models the second most extreme WEC and a 3-WEC cluster is used for the edge WECs. The coloured lines with
arrowheads indicate the interactions between the WECs.
clusters (dimension-4), although the number of samples used in
the latter case is much higher (3-WEC: 800 samples; 2-WEC: 200
samples). Increase in dimensionality and computational con-
straints can be a hindrance in performing more complex optimi-
zation tasks.
Our interaction model takes account of the most dominant
interaction effects (from each side) on any particular OWSC. In
general, it provides fairly accurate predictions, however there can
be a reduction in the prediction accuracy for closely spaced array
configurations i.e. negligible separation between the edges of the
neighbouring OWSCs. Even in such situations, the interaction ef-
fects on particular OWSCs are strongest due its neighbouring de-
vices, with the influence of next-to-neighbouring devices being
significantly less. Construction of a better prediction model would
involve the consideration of 5-WECs system (see Fig. 16) to model
WECs well inside the array (2 WECs on each side instead of 1 as for
3-WECs system), 4-WECs system to predict the performance of
WECs just beside the ones at the edges and a 3-WECs configuration
for prediction of the edge WECs. Fig. 16 illustrates the interaction
model. Such an analysis would obviously require much computa-
tional resources, time and increase in complexity of machine
learning algorithm due to the increase in size of the input param-
eter space. The latter approach would have been necessitated if the
effect of the interactions was such that very close spacings led to
better cumulative performance of the OWSCs. However, in the
present problem, close spacings led to destructive interaction effect
on the array performance, and the optimization algorithm auto-
matically selects locations with moderate spacing between the
devices (demonstrating constructive interaction effects).
The developments in this work can have several implications.
Besides providing a procedure for optimization, the analysis can
offer some practical guidelines to wavefarm developers in
designing arrays - closed spacings are to be avoided and slightly
staggered configurations should be preferred. The presented
methodology can also support inclusion of experimental observa-
tions into the optimization procedure e.g. experimental analysis
can be performed on small OWSC (2e3 WEC clusters) and the re-
sults could then be used to optimize large arrays. In general, many
problems related to ocean engineering/wave energy require eval-
uation of costly functions (simulation models). Surrogate models
based on GP can be used to mimic the outcome of the simulations.
7. Conclusion
A methodology for the prediction and optimization of layouts of
WECs in a wave farm is presented using an innovative active
learning approach. The predictions are performed using a GP
regression method. Such a strategy is relevant when the perfor-
mance of similar layouts of WECs are strongly correlated. The
problem is simplified into an additive model with local interaction
using 3-WEC and 2-WEC clusters. The evaluations of the statistical
emulators are quasi-instantaneous and so once the hyper-
parameters of the GP are computed, one can immediately predict
the variation for an unknown arrangement of the WECs. A recently
developed active learning strategy, the GP-UCB-PE algorithm, was
then used to train the predictor in regions of interest of the prob-
lem. The algorithm focuses on the maxima and regions of high
uncertainty of the function. This in-turn enables the algorithm to
make better predictions in the process of determining the optimal
layout of the WECs. The predictions of the statistical emulator are
reasonably good despite the complexity of the problem and small
number of training data used.
Variation of the performance of an inline and a staggered array
of 40 WECs and uniform spacing is analysed using the prediction
function. The b q factor for the inline configuration has a maximum
for a separation distance of 76.5 m between the centres of the
neighbouring WECs. Optimization of the layout of a wavefarm
comprising of 40 WECs with arbitrary bathymetry constraints and
space limitations is then performed using a genetic algorithm. It is
shown that it is possible to achieve a b q factor greater than one even
with various restrictions. It is also suggested that clustering of de-
vices is to be avoided while designing layouts of wave energy
converter arrays.
Another significant advantage of the presented methodology is
its scalability to large problems. The prediction formulation once
developed needs no modification, and optimization using GA can
be performed for a wavefarm with an arbitrary number of WECs
and constraints. A similar approach of decomposing a problem into
smaller problems and applying machine learning/optimization
techniques may be pursued in solving other problems in the field of
marine renewable energy research.
Acknowledgements
The authors would like to thank Mr. Cedric Malherbe of ENS-
Cachan for fruitful discussions. The first author would like to
thank ENS-Cachan for hosting him during the period of this
research work. This work is supported by Science Foundation
Ireland (SFI) under the research project High-end computational
modelling for wave energy systems (Grant number SFI/10/IN.1/
12996) in collaboration with Marine Renewable Energy Ireland
(MaREI), the SFI Centre for Marine Renewable Energy Research
-Grant number SFI/12/RC/2302.
References
[1] K. Budal, Theory for absorption of wave power by a system of interacting
bodies, J. Ship Res. 21 (4) (1977).
[2] M. Simon, Multiple scattering in arrays of axisymmetric wave-energy devices.
Part 1. A matrix method using a plane-wave approximation, J. Fluid Mech. 120
(1982) 1e25.
[3] H. Kagemoto, D. Yue, Interactions among multiple three-dimensional bodies
 D. Sarkar et al. / Renewable Energy 97 (2016) 504e517
in water waves: an exact algebraic method, J. Fluid Mech. 166 (1) (1986)
189e209.
[4] S. Mavrakos, P. Koumoutsakos, Hydrodynamic interaction among vertical
axisymmetric bodies restrained in waves, Appl. Ocean Res. 9 (3) (1987)
128e140.
[5] S. Mavrakos, Hydrodynamic coefficients for groups of interacting vertical
axisymmetric bodies, Ocean. Engg 18 (5) (1991) 485e515.
[6] B. Child, V. Venugopal, Optimal configurations of wave energy device arrays,
Ocean. Engg 37 (16) (2010) 1402e1417.
[7] T. Whittaker, M. Folley, Nearshore oscillating wave surge converters and the
development of Oyster, Philos. T Roy. Soc. A 370 (1959) (2012) 345e364.
[8] E. Renzi, F. Dias, Resonant behaviour of an oscillating wave energy converter
in a channel, J. Fluid Mech. 701 (2012) 482e510.
[9] D. Sarkar, E. Renzi, F. Dias, Wave farm modelling of oscillating wave surge
converters, Proc. R. Soc. A 470 (2167) (2014) 20140118.
[10] E. Contal, D. Buffoni, A. Robicquet, N. Vayatis, Parallel gaussian process opti-
mization with pure exploration, in: Proceedings of ECML, 2013.
[11] D. Sarkar, E. Renzi, F. Dias, Wave power extraction by an oscillating wave
surge converter in random seas, in: Procedings of the ASME 2013 32nd In-
ternational Conference on Ocean, Offshore and Artic Engineering (OMAE
2013), Nantes, France, 2013.
[12] Y. Goda, A comparative review on the functional forms of directional wave
spectrum, Coast. Eng. J. 41 (1999) 1e20.
[13] A. Babarit, Impact of long separating distances on the energy production of
two interacting wave energy converters, Ocean. Engg 37 (2010) 718e729.
[14] C. Rasmussen, C. Williams, Gaussian Processes for Machine Learning, the MIT
Press, 2006.
[15] D.J. MacKay, Introduction to gaussian processes, NATO ASI Ser. F Comput. Syst.
Sci. 168 (1998) 133e166.
517
[16] A.G. Wilson, R.P. Adams, Gaussian process kernels for pattern discovery and
extrapolation, in: Proceedings of the 30th International Conference on Ma-
chine Learning, Atlanta, Georgia, USA, 2013.

, Quasi-newton methods, motivation and theory, SIAM
[17] J.E. Dennis Jr., J.J. More
Rev. 19 (1) (1977) 46e89.
[18] T. Stefanakis, E. Contal, N. Vayatis, F. Dias, C. Synolakis, Can small islands
protect nearby coasts from tsunamis? an active experimental design
approach, Proc. R. Soc. A 470 (2172) (2014) 20140575.
[19] M. Mitchell, An Introduction to Genetic Algorithms, MIT press, 1998.
[20] E. Renzi, A. Abdolali, G. Bellotti, F. Dias, Wave-power absorption from a finite
array of oscillating wave surge converters, Renew. Energy 63 (2014) 55e68.
[21] E. Renzi, F. Dias, Relations for a periodic array of oscillating wave energy
converters, Appl. Ocean. Res. 39 (2013) 31e39.
[22] E. Renzi, F. Dias, Motion-resonant modes of large articulated damped oscil-
lators in waves, J. Fluids Struct. 49 (2014) 705e715.
[23] I. Noad, R. Porter, Optimisation of arrays of flap-type oscillating wave surge
converters, Appl. Ocean Res. 50 (2015) 237e253.
[24] E. Renzi, F. Dias, Hydrodynamics of the oscillating wave surge converter in the
open ocean, Eur. J. Mech. B-Fluids 41 (2013) 1e10, http://dx.doi.org/10.1016/
j.euromechflu.2013.01.007.
[25] D. Sarkar, K. Doherty, F. Dias, The modular concept of the oscillating wave
surge converter, Renew. Energy 85 (2016) 484e497.
[26] P. Sammarco, S. Michele, M. d’Errico, Flap gate farm: from Venice lagoon
defense to resonating wave energy production. Part 1: Natural modes, Appl.
Ocean Res. 43 (2013) 206e213.
[27] S. Michele, P. Sammarco, M. d’Errico, E. Renzi, A. Abdolali, G. Bellotti, et al.,
Flap gate farm: from Venice lagoon defense to resonating wave energy pro-
duction. Part 2: Synchronous response to incident waves in open sea, Appl.
Ocean Res. 52 (2015) 43e61.