16th Mediterranean Conference on Control and Automation

Congress Centre, Ajaccio, France

June 25-27, 2008

Wave Prediction and Fuzzy Logic Control of Wave Energy

Converters in Irregular Waves
Marco P. Schoen, Member, IEEE, Jørgen Hals, and Torgeir Moan

Abstract— This paper presents a short comparison of varying sea condition. Results of the proposed algorithm
different existing and proposed wave prediction models, included prediction of waves for long time horizon, i.e. 30
which in turn are used to control wave energy converters seconds using a 0.5 sample time, resulting into predictions
(WEC) in irregular waves. The objective of the control with adequate results. Noise considerations were given only
action is to increase the energy conversion. The to the extent that outliers are removed through the standard
prediction models are compared based on the outlier detection rule. Forsberg’s model results into an
contribution to the prediction of the individual filter extensive model order to capture the nature of irregular
weights, in an effort to reduce the filter order. The waves. Halliday et al., [2], looked at using the Fast Fourier
controller is based on adjusting the short term stiffness Transform (FFT) to generate prediction of waves over
and damping of the Power Take Off (PTO) of the WEC. distance. The distance factor is used to avoid perturbations
The control action, which is a Fuzzy Logic (FL) based of the buoy to the measurements. This approach necessitates
design, is supported by utilizing nominal damping and the use of directional data. The simulation results presented
stiffness values determined for the given sea state and indicate reasonable success for long distances. One of the
instantaneous wave using a simple Genetic Algorithm draw-backs of using Halliday et al.’s approach is that the
(GA). The WEC chosen is comprised of a spherical buoy spectral estimates of the FFT based models are less sensitive
and is modeled using the bond-graph technique. The to sharp peaks and dips compared with autoregressive (AR)
simulation carried out indicates the proposed algorithm models, [3]. In addition, the FFT is for strictly stationary
increases the energy conversion of WEC in irregular systems, and can not handle the temporal fluctuations, which
waves compared to the no control case. are an essential component in the characteristics of irregular
waves.
I. INTRODUCTION The work presented in the first part of this paper describes
the pursuit of finding predictive models that work well in a

T he energy absorption of Wave Energy Converters

(WEC) is influenced by the use of the knowledge of future
stochastic environment, are of low model order (parametric
models), and have a sufficient accuracy that allows the
algorithm to be used for control of wave energy converters
wave data in the control loop. The prediction of future wave operating in irregular waves. In the second part of this
data is not a trivial undertaking since real waves are paper, we introduce the WEC model, and propose a simple
irregular in nature. In this paper, we will propose a hybrid Fuzzy-Logic control approach that utilizes the prediction
Kautz/AR predictive model as well as a pure predictive model in order to increase the energy conversion of the
Kautz model and compare them to a predictive WEC.
Autoregressive (AR) and Autoregressive with Moving
Average (ARMA) algorithm. II. WAVE PREDICTION MODELS
A number of contributions have been made in forecasting
2.1 AR Predictive Model:
irregular waves. In example, Forsberg, [1], introduced an
Consider a simple Finite Impulse Response (FIR) AR
ARMA model with fixed time horizon prediction, utilizing
model to characterize irregular waves:
the maximum likelihood estimation to describe ocean waves p
off the coast of Sweden. The algorithm included the on-line yˆ ( k k − 1) = ∑ a j y ( k − j ) , (1)
recursive form with an adaptation constant for slowly time j =1

where the output y ∈ n x1 is the wave height and is

o

Manuscript received January 26, 2008. measured up to time t = k − 1 , p is the model order, and
Marco P. Schoen was with the Norwegian University of Science and n ×n
a∈ o
are the influence coefficient matrices. One of the
o

Technology (NTNU), Center for Ships and Ocean Structures (CeSOS),

7491, Trondheim, Norway while on leave from the Idaho State University, main problems in fitting such a time series to the gathered
Pocatello, ID 83209 USA (phone: 208-282-4377; fax: 208-282-4538; e- data is the determination of the model order. Based on the
mail: schomarc@ isu.eduj). Wold decomposition theorem, an AR model can fit a
Jørgen Hals is with the Norwegian University of Science and
Technology (NTNU), Center for Ships and Ocean Structures (CeSOS), variety of data characteristics if a sufficient large model
7491, Trondheim, Norway, (e-mail: [email protected]). order is selected, [4].
Torgeir Moan is with the Norwegian University of Science and
Technology (NTNU), Center for Ships and Ocean Structures (CeSOS),
7491, Trondheim, Norway, (e-mail: [email protected]).

978-1-4244-2505-1/08/$20.00 ©2008 IEEE 767

 To construct a variable horizon prediction formulation on p p

yˆ ( k + i k − 1 ) = ∑a y (k − j ) + ∑ bm ε (k − m ) , (9)
(i) (i)
the basis of Equation (1) the following i-step-ahead j =1
j
m =1
predictor can be established (see [5]):
p
where the predictive parameters for the past residual terms
yˆ ( k + i k − 1 ) = ∑a (i)
y (k − j) , (2)
are defined as
a (ji ) = a1( i −1) a (0) ( i −1)
j + a j +1
j

j =1 for j = 1, 2, 3, …
where we made use of the residual, which is defined by the and a k = 0 for k > p
difference of the estimate and the true output, i.e. ε = y − yˆ . The predictive parameters for the moving average terms are
The corresponding predictive coefficient matrices are: defined as
( i −1) ( i −1)
a j = a1 a j + a j +1 . (3)
(i )
b (mi ) = a1(i −1) b (0) ( i −1)
m + b m +1 for m = 1, 2,3,…
One can easily argue that by any standard the above and b k = 0 for k > p
introduced predictive model possesses a rather large model
The estimation of these parameters can be done for example
order. Hence, the corresponding correlation matrix utilized
by following the Hannan-Rissanen algorithm, [7].
in the estimation algorithm induces a large computational
burden during the operation of the filter, in particular when
2.3 Orthogonal Basis Function Prediction Model
repeated estimations are needed to update to the current
One of the main reasons AR and ARMA models require
wave conditions. This is true to some degree also for the
large model orders is that they are of the type FIR filter and,
recursive least squares (RLS) method, even though no
they do not have a structure that allows for inclusion of prior
inversion of the correlation matrix P is required, the model
knowledge into the filter before the estimation occurs. AR
order dictates the number of computations needed. This
models belong to the class of Orthogonal Basis Functions
property of large model order is a general characteristic of
(OBF), which are generalizations of FIR models. The basis
FIR filters due to the missing feedback term in the filter
function of FIR models are delayed Dirac impulses
structure. We can investigate the contribution of each filter
expressed with the q-operator (delay operator). In the pursuit
weight to the predicted output by computing the transformed
of reducing the required model order while still describing
filter weights as follows: Starting with the formulation for
the future waveform adequately, we propose to choose
the estimated filter parameter matrices (least squares),
θˆ AR = (ΦTAR Φ AR ) −1 ΦTAR YAR , (4) orthonormal filters Li ( q ) such that Equation (1) becomes
where θˆ AR is a matrix containing the estimated parameter p

matrices, Φ AR is the information matrix, and YAR is the y ( k + i k − 1) = ∑c L (q) y (k ) .

i i
(10)
i =1
respective output vector. Defining X ss = ΦT YAR the updated
To keep the chosen OBF model linear in the parameters,
parameter estimate is Θˆ ( k + 1 k ) = P ( k + 1) × X ss ( k + 1) .
Li ( q )
are chosen before the estimation of the filter
Introducing the eigenvector matrix Q and diagonal
coefficient matrices ci occurs. Li ( q ) ’s are chosen with the
eigenvalue matrix Λ of the correlation matrix P = ( ΦT Φ ) ,
purpose of incorporating prior knowledge of the system to
the transformed filter weight vector results into:
be modeled. Since the waves exhibit an under-damped
Θ ( k + 1 k ) = Θ θˆ ( k + 1 k ) ,
/ T
(5) dynamical characteristic, we choose Kautz filters for the
which can be used to compute the transformed filter weight Li ( q ) ’s. Kautz filters allow us to incorporate conjugate
vector based on past data as follows: complex pole pairs into the model. Considering a given
X /ss ( k + 1) = ΛΘ / ( k + 1 k ) . (6) conjugate complex pole pair (α ± i β ) , the Kautz filter
Using xi/ as the elements in X /ss and λi the corresponding coefficients are computed by
eigenvalues found on the diagonal of Λ , each individual 2α
transformed filter weight contribution is given by c= and d = − (α 2 + β 2 ) . (11)
1+α + β
2 2
/
x
ai/ (0) = i
for 1 ≤ i ≤ p , (7) With this, the filter Li ( q ) can now be computed using the
λi
following orthonormalization process:
2.2 ARMA Predictive Model:
Adding to the AR formulation a Moving Average (MA) (1 − c )(1 − d )
2 2

L1 ( q, c, d ) = , (12)
part enhances the flexibility of the AR model, [6]. Consider q 2 + c ( d − 1) q − d
a typical ARMA model in the output error formulation:
(1 − d ) ( q − c )
2

L2 ( q, c, d ) =
p p

yˆ ( k k − 1) = ∑ a j y ( k − j ) + ∑ b m ε ( k − m ) , (8) . (13)
j =1 m=0
q 2 + c ( d − 1) q − d
where a j and b j are the ARMA filter matrices. The And the higher order filter parameters by using the recursive
moving average associated with the b j matrices is nonlinear formula as given below:
due to the prediction error formulation. The model can be − dq 2 + c ( d − 1) q + 1
L2 i −1 ( q, c, d ) = L2 ( i −1) −1 ( q, c, d ) ,
constructed through back-substitution of the output into the q 2 + c ( d − 1) q − d
original filter model formulation. An i-step ARMA predictor
thus can be given as:

768
 − dq 2 + c ( d − 1) q + 1 filter that are non zero are corresponding to the AR
L2 i ( q, c, d ) = L2( i −1) ( q, c, d ) . coefficients, hence the first entry can be considered as a
q + c ( d − 1) q − d
2

simple AR filter.
The limits on the coefficients c and d are given as: The correlation of the AR filter is respectable, but using
−1 < c < 1 and −1 < d < 1 . c ≠ 0 , the simulations indicate that c is to be selected
In addition to having a purely predictive Kautz filter, one around 0.9. Using the step-by step procedure outlined above,
can incorporate as well some simple AR terms, resulting into the average c value that would be selected by determining
hybrid Kautz/AR filters. The step-by-step procedure for the dominant frequency in the wave spectrum is 0.9979. c’s
constructing and estimating a predictive Kautz or hybrid relationship to the Kautz filter is that it sets its dominant
Kautz/AR filter is given below. A recursive formulation can frequency to the dominant frequency of the waves.
be developed using the RLS approach, which is left for Considering the second parameter in the proposed wave
future work. prediction model, the damping, we are interested in
Step 1: Collect a sufficient amount of wave height observing the influence of this parameter on the predicted
measurements y ( k ) , with a fixed sampling time. Normalize time series. Utilizing the same simulation parameters as
y (k ) above listed, with a Pierson-Moskowitz wave spectra
the wave height time series using y (k ) = , where corresponding to a wind speed of 46.3 km/h [25 knots], the
y max
statistics provided in Fig.1 b) is generated. The parameter c
{ }
ymax = sup y ( k ) . Step 2: Perform a spectral analysis and is computed based on the dominating wave frequency,
identify the frequency ω1 with the largest contribution or which resulted in c = 0.9973. Fig. 1 b) indicates that the
damping term d has little sensitivity to the quality of the
magnitude in the spectrum. Step 3: Compute the Kautz filter prediction. Intuitively, we would imagine that smaller
coefficients c and d based on the identified frequency ω1 , damping fits the nature of the waves. This perception is
also assume a damping ratio ξ . The computation of the validated ever so slightly by Fig. 1 b). No damping results
filter coefficients c and d require the real and imaginary into the more accurate wave prediction results.
parts of the dominant root location, which are computed as
follows: α = ξω1 and β = sin ( cos −1 ξ ) ω1 . Step 4: Estimate 2.4 Comparative Simulation Based Analysis of AR, ARMA,
Kautz/AR, and Kautz Predictive Filter Performance:
the model parameters ci using Simulations with four predictive filters are carried out in
order to compare their performance. Table 1 was compiled
{ } ΦTKautz YKautz ( M − i ) ,
−1
ˆ ( i ) = ΦT Φ
Θ Kautz Kautz Kautz
using a sea state of moderate gale 55.5 km/h wind speed, a
where Φ Kautz is the information matrix, moving prediction horizon of 20 data points, and a model
( )
⎡ L q, c, d y ( p )
⎢ 1
L3 ( q, c, d ) y ( p ) … Lp ( q, c, d ) y ( p ) ⎤
⎥
order of 35 for the Kautz/AR hybrid filter, the AR, and the
Φ Kautz ( q, c, d ) ( ) (
⎢ L q, c, d y p + 1
=⎢ 1
) L3 ( q, c, d ) y ( p + 1) … Lp ( q, c, d ) y ( p + 1) ⎥⎥ ARMA filter. The Kautz filter is based on a model order of
⎢
⎢
⎥
⎥ 25 odd terms. Table 1 details the performance for different
( ) (
⎢ L q, c, d y M − 1
⎣ 1 ) L3 ( q, c, d ) y ( M −1) … Lp ( q, c, d ) y ( M −1)⎥
⎦ additive measurement noise conditions (N) in % noise
where M is the data length, and variance and is based on 51 simulation runs.
Θ Kautz = ⎡⎣b1 , b 2 , b 3 , … , b p ⎦⎤
T
YKautz = [ y ( p + i + 1) , … y ( M + i )] .
T From the table we recognize that for the noise-free case,
all filters perform less well than when some noise is
Step 5: Compute the future output using Equation (10). included. This behavior can be traced back to the used
estimation algorithm. Comparing the performance of the
The above procedure for the computation of the hybrid hybrid Kautz/AR filter indicates that this filter, though of
Kautz/AR filter involves two coefficients that need to be equal filter weight length, performs equivalently well to the
selected. In the following we clarify the reasoning for pure AR filter. The AR filter seems to be the most accurate
executing Step 2. For this we use a set of simulations where one among the four filters investigated for cases where noise
the average correlation coefficient is computed between the is present. The comparison with the Kautz filter is to be
predicted wave height and the actual wave height. The made with caution, since it has fewer filter weights (25)
simulations are carried out with a hybrid model that contains compared to the other filters (35 filter weights). The
four Kautz filter terms and one AR term, i.e. the model order employed ARMA filter has a similar performance as the
is five, where p = 5 corresponds to L5 = q . The additive proposed Kautz filter for the cases where noise is included.
measurement noise is characterized by a signal to noise ratio The standard deviation (STD) and average (Ave) correlation
of 1000 (0.1% noise variance), and waves are computed coefficient reported in Table 1 indicate that the AR and
based on the Pierson Moskowitz wave spectrum with a Kautz/AR filter seem to be less receptive to noise then the
corresponding wind speed of 55.5 km/h [30 knots]. Fig. 1 a) ARMA and Kautz filters. Another characteristic noted in the
gives the average and standard deviation for a varying c simulations is that the Kautz filter and to some degree the
coefficient, corresponding to the frequency of the Kautz ARMA filter offer a smoother prediction curve than the AR
filter. The statistics is based on a set of 50 simulation runs. based filters. Correspondingly, from observing simulation
First we observe that when c = 0 , the only portions of the runs, the Kautz filter has a tendency to miss smaller peaks in

769
the forecast. A relatively simple analysis on the Taking the hydrostatic stiffness force to be linear is
effectiveness to include additional information into the filter equivalent to assuming a constant water-plane area Aω for
in order to reduce the number of filter weights is to look at the buoy. The hydrostatic stiffness coefficient is then given
the transformed filter weights, see Equation (7). Fig. 2 by Cb = ρ gAω . As long as the incoming waves are small and
depicts the four filter’s transformed filter weight the oscillation amplitude is restricted to y ≤ 3 [m], this
distribution. In comparison using Fig. 2, the proposed approximation keeps force errors within ±6%, [11].
predictive Kautz filter requires fewer terms to describe the 1.8
Transformed AR- and MA-. filter weight contribution 2. 5
Transformed AR filter weight contribution

prediction. a) 1.6
-. MA
-AR b)
2
1.4

Magnitude - absolute value

Correlation coef. for different c in Kautz/AR filter
a)

Magnitude - absolute value

1.2

1 1
1.5
Correlation Coef.

0.8
1

0.6
Correlation Average 0.4 0.5
0.5
0.2

Standard Deviation 0
0 5 10 15
Model Order Index
20 25
0
0 5 10 15
Model Order Index
20 25

Transformed Kautz filter weight contribution

Transformed Kautz/AR filter weight contribution 40
0 c) 1.8
d)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
b) 1.6 35

c-Value 1.4 30
Correlation coef. for different d in Kautz/AR filter

Magnitude - absolute value

Magnitude - absolute value
1.2
25
1
1
20
Correlation Coef.

Correlation Average 0.8

15
0.6
10
0.4
0.5
5
0.2

Standard Deviation 0 0
0 5 10 15 20 25 0 5 10 15 20 25
Model Order Index Model Order Index

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 2: Transformed filter weight distribution for a) ARMA
d-Value prediction, b) AR prediction, c) Kautz/AR prediction, and d)
Kautz prediction.
Fig. 1: Influence of Kautz parameters on prediction.
TABLE I The bond graph has two energy-storing elements in
CORRELATION FOR DIFFERENT PREDICTIVE FILTERS UNDER VARYING NOISE integral causality, which indicates that the model has two
ENVIRONMENT state variables: the linear momentum pb of the buoy, and its
KAUTZ KA./AR AR ARMA vertical position y . Also we see that the force of the added
N=0%, Ave 0.6560 0.6434 0.6044 0.5097 mass is in differential causality, which will provide us with
N=0%, STD 0.1110 0.1309 0.3264 0.2641 an implicit equation for the inertia force.
N=0.1%,Ave 0.6867 0.7532 0.7475 0.7203
N=0.1%,STD 0.1536 0.1156 0.1231 0.1216
N=1.0%,Ave 0.6229 0.7051 0.7080 0.6598
N=1.0%,STD 0.1801 0.1177 0.1152 0.1665
N=10%,Ave 0.5438 0.5788 0.5907 0.5537
N=10%,STD 0.1476 0.1652 0.1619 0.2083

III. BUOY MODEL

Fig. 3: Bond graph for the heaving sphere.
The chosen WEC model consists of a spherical heaving
buoy reacting against a fixed reference. The model includes Deriving state equations from the bond graph yields
only the heave mode with vertical excursion y , while the m∞ , b
pb = − pb − Cb y − Fl − Fr + Fe , (14)
water depth is assumed to be infinite, and the usual linear mb
approximation of the hydrodynamic forces are used.
y = pb / mb .
(15)
Assuming that the gravity force is balanced by the average
buoyancy force, our system can be modeled by the bond The hydrodynamic radiation force can be given as
graph shown in Figure 3, [8,9]. The following forces are t pb
Fr = ∫ κ (t − τ ) dτ , [10]. Here, the impulse response
included: the wave excitation force Fe , the inertia forces −∞
mb
due to buoy mass and added mass at infinite frequency, mb y function κ (t ) is a memory function linking the current
and m∞ ,b y . Furthermore, the hydrostatic stiffness force radiation force to past and current values for the buoy
Fc = −Cb y , the damping force due to wave radiation Fr , velocity, pb / mb . The convolution integral Fr may be
and the load force from the power take-off machinery Fl is
accounted as well in the model. As may be seen, the inertia approximated by a finite order state space model, [12]:
pb
force due to added mass at infinite frequency m∞ ,b has been ψ ( t ) = A kψ ( t ) + B k , Fr (t ) = Ckψ (t ) (16)
separated from the rest of the radiation force Fr , in mb
accordance with the usual time-domain representation, [10].

770
Which will require to find the additional state vector ψ . structural damping and stiffness of the WEC serving as the
Rearranging Equation (14), one arrives at: control parameters. To accomplish a fuzzy inference system
mb for buoy control, the following system structure is proposed,
pb = − ( −C
b
y − Fl − Fr + Fe ) . (17) see Fig. 5. The wave energy device (WEC) is defined by the
mb + m∞ , b
model given above. The system identification (SI) provides
Defining the new state vector x = ⎡⎣ pb y ψ T ⎤⎦ , one can a state space model to simulate the dynamics of the buoy for
given predicted wave motions.
now write the entire system, including the radiation force Wave motion -. and Buoy motion -
approximation, as 5
Buoy motion
⎡ − mb Cb ⎤ 4

⎢ 0 m +m −C k ⎥ ⎡ − mb mb ⎤ 3

⎢ b ∞ ,b
⎥ ⎢m + m m + m ⎥
⎢ b ⎥ ⎡F ⎤

Magnitude of Elevation [m]

∞ ,b b ∞ ,b 2
⎢1 ⎥ Wave motion
x=⎢ 0 0 ⎥x+⎢ 0 0 ⎥⎢ l⎥ 1

⎢ mb ⎥ ⎢ 0 0
⎥ ⎣ Fe ⎦ 0

⎢ Bk 0 Ak ⎥
⎢ ⎥
⎢ ⎥
-1
⎢ ⎥ ⎣ ⎦
⎣ ⎦ -2

-3
⎡ Fl ⎤
= Ax + B ⎢ ⎥ . (18) -4

⎣ Fe ⎦ -5
0 200 400 600 800 1000 1200
For the studied heaving sphere we have used data from Time Index k [0.1 s]

[13,14] to compute the memory function κ (t ) . A reasonable Fig. 4: Excessive buoy excursion.
approximation for the convolution term can be achieved
with a state model of order six using Hankel singular value z E
decomposition to reduce the model order, [12]. This gives a GA
total of eight states for the heaving buoy model, taking the c1* , k1* E
wave excitation force and the machinery load force as inputs z
WEC y
to the system.
Predictor
IV. HEAVING WAVE ENERGY DEVICE CONTROL Parameters
ẑ ′
Assuming the sea state is stationary in the wide sense, we
ẑ Eˆ
can optimize the tuning parameters of the WEC based on the Filter
ŷ FL
immediate past data to the current sea condition in order to z [ A B C D ] Model
i i i i
y
maximize the energy production. This type of tuning of the SI
device will lead to having the buoy operate in its resonance
mode. As mentioned above, the optimization utilizes a Fig. 5: System architecture for control implementation.
Genetic Algorithm (GA) for this purpose, see [15]. The
optimization of the stiffness and the damping of the external The predictor in Fig. 5 can be any of the prediction
mechanism may cause excessive amplitudes of the buoy, models introduced in the earlier section of this paper. The
which is depicted in Fig. 4. The spherical buoy has a radius Mamdani type Fuzzy Inference System (FIS, see [16]) is
of 5m and hence the oscillation depicted in Fig. 4 indicates given by three inputs (velocity, amplitude, and phase of
that at times the buoy is almost leaving the water plane. For buoy) and two outputs (damping and stiffness). The three
the optimization, the GA utilized a data length of 2000 inputs are processed using a set of input membership
points (20 seconds). The objective function for the GA is to functions. The output damping and stiffness are then
maximize the energy conversion without having the buoy computed as a change from the nominal values, which are
exceed its amplitude limitations. For practical reasons we the optimal coefficients computed using a GA based on past
want to limit these excessive excursions to ±3m . The PTO input/output data. The input functions are linked to the
adapts its damping and stiffness force based on the controller output by a rule base as depicted in Fig. 6. The
computed optimum values from the GA, in the extreme case, computed optimal damping and the optimal stiffness, using a
halting the motion in order to limit the excursions is GA, are taken as the initial values when the controller is
allowed. We will use a fuzzy logic controller with the switched on. After that, the controller adapts the new
purpose to adapt the immediate damping and stiffness damping and stiffness for the immediate future only. Fig. 7
generated by the PTO to control the buoy motion. The depicts the response of the buoy over time. The controller is
control output is computed based on the prediction of the turned on at time k = 650. The immediate reaction of the
future buoy motion. The horizon for this controller is 1 device is to oscillate in resonance. This oscillation always
second. In Fig. 4, z denotes the wave height, y the buoy stays within the excursion limits of ±3m . Fig. 8 depicts the
motion, E the produced energy, c1 and k1 are the controller output, the damping and the stiffness. The
converted energy is correspondingly increasing on a much

771
steeper curve then when no control is used, see Fig. 9. The as nominal values shows improved energy production of
nominal damping and stiffness for the simulations are WEC. Further work will address sea changing conditions in
c1 = ( 6.5147e + 5 ) Ns / m and k1 = ( 7.8974e + 5 ) N / m . The the optimization loop.
optimized structural coefficients take on the following 3
x 10
7
Energy Production Comparison

values: c1* = (16.429e + 5 ) Ns / m and k1* = (1.0742e + 5 ) N / m Energy with Damping Control

respectively. Also included in Fig. 9 is the comparison 2.5

Energy without Control
Energy with Damping and Stiffness Control
between the damping controller and the damping and
stiffness controller. The rule base was reduced to 8, in order 2

to accommodate the loss of one output variable. Fig. 9

Magnitude
indicates that a small improvement of the energy output is 1.5

noticed due to the help of adapting the stiffness of the PTO

during the simulations. The change is not significant, but 1

indicates that an instantaneous update of the control

parameters can have some influence. 0.5

0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Time Index k[Sampling time =0.1 seconds

Fig. 9: Comparison of Energy generated.

REFERENCES
[1] J.L. Forsber, “On-line identification and prediction of waves,”
Hydrodynamics of Ocean-Wave Energy Utilization, D.V. Evans, and
A.F. de O Falcao (eds.) Springer-Verlag, Berlin, 1986, pp. 185-193.
[2] J.R. Halliday, D.G. Dorell, and A. Wood, “A Fourier Approach to
Short term Wave Prediction,” Proc. Of the 16th International Offshore
and Polar Engineering Conference, San Francisco, CA, 2006, pp. 444
– 451.
Fig. 6: Rules with corresponding membership functions and [3] R.H. Jones, “Identification and Autoregressive Spectrum Estimation,”
firing strength of output. IEEE Transaction on Automatic Control, No. 6, 1974, pp. 894-897.
[4] S.M. Key and S.L. Marple, “Spectral analysis – a modern
perspective,” Proc. Of the IEEE, vol. 69 (11), 1981, pp. 1380-1419.
[5] U.A. Korde. M.P. Schoen, and F. Lin "Time Domain Control of a
Single-Mode Wave Energy Device," Proc. of ISOPE, 2001,
Stavanger, Norway
[6] O. Nelles, Nonlinear System Identification from Classical Approaches
to Neural Networks and Fuzzy Models, Springer, Berlin, 2001.
[7] Brockwell P.J. and Davis R.A., “Introduction to Time Series Analysis
and Forecasting,” Springer, New York, 1997.
[8] D.C. Karnopp, D.L. Margolis, and R.C. Rosenberg. System Dynamics:
Modeling and Simulation of Mechatronic Systems. Wiley, 4th edition,
2006.
[9] H. Engja and J. Hals, “Modeling and simulation of sea wave power
conversion systems.” Proceedings of the 7th European Wave and
Tidal Energy Conference, Porto, Portugal. 2007.
[10] J. Falnes, Ocean waves and oscillating systems: linear interactions
Fig. 7: Time history of buoy motion and wave motion. including wave-energy extraction, Cambridge University Press,
Cambridge, 2002.
[11] J. Hals, T. Bjarte-Larsson, and J. Falnes, “Optimum reactive control
and control by latching of a wave-absorbing semisubmerged heaving
sphere.” Proceedings of the International Conference on Offshore
Mechanics and Arctic Engineering - OMAE, 2002, pp. 415 – 423.
[12] R. Taghipour, T. Perez, and T. Moan, Hybrid frequency–time domain
models for dynamic response analysis of marine structures. To appear
in Ocean Engineering, doi:10.1016/j.oceaneng.2007.11.002.
[13] T. Havelock, “Waves due to a floating sphere making periodic
heaving oscillations,” Proc. Royal Society, 1955, 231A:1–7.
Fig. 8: Damping [Ns/m] and Stiffness [N/m] computed [14] A. Hulme, “The wave forces acting on a floating hemisphere
using the fuzzy logic controller. undergoing forced periodic oscillations,” Journal of Fluid Mechanics,
vol. 121, 1982, pp: 443–463.
V. CONCLUSION [15] R.L. Haupt and S.E. Haupt, Practical Genetic Algorithms, John Wiley
& Sons, Inc., New York, 1998.
The two proposed prediction algorithms (Kautz/AR and the [16] H. Zhang and D. Liu, Fuzzy Modeling and Fuzzy Control, Birkhäuser,
Kautz) indicate to perform in a competitive fashion when Berlin, 2006.
compared to existing prediction models. The proposed
simple FL controller using optimized structural parameters

772