Short-Term Wave Forecasting with AR models in

Real-Time Optimal Control of Wave Energy

Converters
Francesco Fusco and John V. Ringwood
Electronic Engineering Department
National University of Ireland Maynooth
Maynooth, Co. Kildare, Ireland
Email: [email protected], [email protected]

Abstract—Time domain control of wave energy converters through a force Fc . The optimal reference velocity, ûd , is
requires knowledge of future incident wave elevation in order to computed by an algorithm that requires future values of the
approach conditions for optimal energy extraction. Autoregres- incident wave elevation η, that must be provided by a predictor,
sive models revealed to be a promising approach to the prediction
of future values of the wave elevation only from its past history. which in this case is based only on the past history of the signal
Results on real wave observations from different ocean locations itself. Depending on the specific device operating principle,
show that AR models allow to achieve very good predictions as well as on the general control strategy adopted and on the
for more than one wave period in the future if the focus is put available instrumentation, the involved quantities may vary, but
on low frequency components, which are the most interesting the logic can always be described by a structure as the one
from a wave energy point of view. For real-time implementation,
however, the lowpass filtering introduces an error in the wave proposed in Fig. 1, where a certain physical quantity needs
time series, as well as a delay, and AR models need to be designed to be predicted in order for the reference to be as close as
so to be as robust as possible to these errors. possible to the optimality condition. This quantity may be
the wave excitation force or the oscillation velocity for an
I. I NTRODUCTION oscillating body, the dynamic air pressure inside the chamber
Wave energy conversion in most devices is based either on for an oscillating water column, the overtopping water for an
relative oscillation between bodies or on oscillating pressure overtopping device, and all of them are ultimately dependent
distributions within fixed or moving chambers. Oscillators on the incident wave elevation.
generally have pronounced resonances, which enable efficient The approach to prediction from only the past history of
power absorption in certain wave conditions. In order, how- the signal itself presents several advantages with respect to
ever, to cope with the variations of wave spectra a control the reconstruction of the wave field from distant observations,
system can be designed to alter the oscillator dynamics such which requires an array of measurements (to deal with multi-
that the efficient energy conversion occurs in a wide range of directionality and to separate incident from refracted waves)
wave conditions [1]. and complex prediction models, thus resulting in increased
The control approach in the early stage of wave energy, instrumentation costs. Linear Autoregressive (AR) models
consisted of frequency domain relationships regulating the have been proposed as an efficient and effective solution for
dynamics of the system to be tuned for maximum energy the prediction of the oscillation velocity [4] and the wave
absorption at different peak frequencies corresponding to elevation [5],[6],[7]. Particularly in [6] and [7], AR models
different incoming wave spectra [1],[2]. Although being an were shown to outperform even neural networks, with very
advantageous approach in real sea spectra, it does not generally accurate predictions obtained for more than one period on
allow control on a wave by wave basis, that is a real-time real wave elevation time series, when the high frequency
control that could significantly raise the device productivity components were filtered out.
and therefore its economical viability. Real-time optimal con- In this paper, the identification and estimation of AR models
trol can be directly derived from the aforementioned optimal for short term wave forecasting is presented in section II,
frequency relationships, and its main difficulties raise from the together with a methodology for the characterisation of the
fact that their transformation in the time domain results in non probability distribution of the predictions through confidence
causal transfer functions, so that the conditions for optimal intervals. More exhaustive results on real observations from
power absorption can be realised only if future motion of the different ocean locations and in a variety of climates are
device, or future incident wave, are known [1],[2],[3]. presented in section III. A real-time implementation of this
As an example, Fig. 1 represents a possible control scheme prediction procedure, however, requires the design of a real-
for a generic oscillating body, where the controlled variable is isable lowpass filter (so far, only ideal off-line lowpass filters
its oscillation velocity, u, and the control action is performed were applied). This issue and the consequences on the AR

978-1-4244-6392-3/10/$26.00 ©2010 IEEE 2475

η(k) η̂(k + l|k) ûd(k)
+ Fc(k) u(k) in longer-term prediction [9]. A multi-steps ahead cost func-
predictor optimal reference
generator G WEC
- tional is therefore applied, referred to as long-range predictive
identification (LPRI) [10]:
N2
N X
X 2
JLP RI = [η(k) − η̂(k|k − j)] (4)
Fig. 1. Time domain control of wave energy converters requires future
k=1 j=N1
knowledge of the incident wave elevation to approach optimality
This functional can be minimised using a standard algo-
rithm for solving nonlinear least-squares problems, such as
model are discussed in section IV. the Gauss-Newton approach, initialised with the results from
regular least squares [10]. In our specific case we set N1 = 1
II. AUTOREGRESSIVE MODEL and N2 as the maximum prediction horizon (N2 = 50 for
the Galway Bay data and N2 = 25 for the Pico Island data),
The wave elevation η(k) is supposed to be linearly depen-
because we want to optimise the prediction accuracy at any
dent on a number n of its past values:
lead time.
n
X Note that, from (3), the frequencies are related to the poles,
η(k) = ai η(k − i) + ζ(k) (1) so if the parameters are kept constant, the frequencies are
i=1 constant as well. An adaptivity mechanism based on the LPRI
where a disturbance term ζ(k) has been also included. If the cost function (4) could then be considered, which is also
parameters ai are estimated and the noise is supposed to be proposed in [10]. However, as it will be shown in the results,
Gaussian and white, the best l-step ahead prediction of the section III, a static AR model keeps its validity for long time
future wave elevation at instant k, η̂(k + l|k), is given by: after being estimated (more than 2 hours), in spite of wave
spectral variations. Adaptive AR models do not represent,
n
X therefore, a big issue at the moment and are not considered in
η̂(k + l|k) = âi (k)η̂(k + l − i|k) (2) this paper.
i=1
B. Identification
where, obviously, η̂(k + l − i|k) ≡ η(k) if k + l − i ≤ k (i.e.
information already acquired and no need of prediction). An indication about the appropriate order n of the AR model
The general shape of the prediction function is completely to predict the wave elevation, can be obtained through min-
determined by the poles, pi , of the corresponding transfer imisation of the classical criteria AIC, proposed by Akaike
function, and its particular realisation is determined, at each [11], and BIC, proposed by Schwarz [12] :
sampling instant, by the past values of the time series [8]. It is 2
AIC = log(σ̂ 2 ) + r (5)
particularly interesting to analyse the shape of the forecasting N
function (2) in the case of m/2 (when m is even) couples of log(N )
BIC = log(σ̂ 2 ) + r (6)
complex-conjugate poles pi and p∗i : N
Here, r = n + 1 is the number of parameters to be estimated,
m/2
X σ̂ 2 is the value of the minimised functional (4), and not,
η̂(k + l|k) = ci (k)|pi |l sin(∠pi k + ϕi (k)) (3)
as normally utilised, the one-step ahead error variance, a
i=1
modification which is coherent with the estimation criterion
Thus, an AR model with only complex-conjugate poles is and with the use of the AR model for multi-step ahead
implicitly a cyclical model, where the frequencies are related prediction purposes.
to the phase, ∠pi , of each pole and the amplitudes and
C. Confidence intervals
phases of the harmonic components are related to the last n
measurements of each time instant k, so that they adapt to the The predictions alone, as given by (2), do not give a
observations [7]. complete enough information about the future of the wave
elevation, because they are inevitably affected by an estimation
A. Estimation error, so that it would be fundamental to have an indication
about the entity of this error and about how reliable the
The AR coefficients can be estimated from a number N
predictions are.
of batch wave elevation observations. The standard approach,
If the l-steps ahead prediction error is Gaussian:
under the assumption of Gaussian error, is to minimise the
ê(k + l|k) = η(k + l) − η̂(k + l|k) ∼ ℵ 0, σl2


sum of squares of the one-step ahead prediction error, through (7)
regular least squares. Because the model will be utilised for
then the variance, σl2 , is all we need in order to define its
multi-steps ahead forecasting, however, either the estimation
probability distribution. Moreover, we can assume a confidence
and the identification problems turn out to be ill-posed, since
interval where the error is contained with a probability α:
many models can be found which are almost equivalent in
terms of one-step ahead prediction, but fare quite differently −p α2 ≤ ê(k + l|k) ≤ +p α2 (8)

2476
 1 0.05 0.05
Here, p α2 is the value of the probability distribution such that: data set G
1
data set G2 data set G
3

S(ω) [m2 s/rad]

Hs ≈ 0.31 m

S(ω) [m2 s/rad]

Hs ≈ 2.31 m

S(ω) [m2 s/rad]

Hs ≈ 0.34 m
Z +p α wide banded narrow banded predominant wind waves
2 0.5
P {−p α2 ≤ ê(k + l|k) ≤ +p α2 } = p(y)dy = α (9)
−p α
2
0 0 0
0 1 2 3 4 0 1 2 3 4 0 1 2 3 4
where p(·) is a Gaussian probability density function. ω [rad/s] ω [rad/s] ω [rad/s]

The estimate of the variance σl2 could be calculated from (a)

the specific model parameters and from the statistics of its 0.1 4 20
data set P1 data set P2 data set P3
estimation algorithm [8]. The problem with this approach is Hs≈0.77 m Hs≈2.51 m H ≈5.92 m
s

S(ω) [m2 s/rad]

S(ω) [m2 s/rad]

S(ω) [m2 s/rad]

that the confidence intervals can result to be misleading if the wide banded narrow banded narrow banded
slightly non Gaussian
model is not perfectly appropriate and if its accuracy changes 2 10

accordingly with the wave conditions. A more straightforward

alternative is adopted at this stage, where the estimate of the
0 0 0
variance of the forecasting error is based on the past history 0 1 2
ω [rad/s]
3 4 0 1 2
ω [rad/s]
3 4 0 1 2
ω [rad/s]
3 4

of the predictions:
(b)
N
1 X
Fig. 2. Spectrum of some significant data sets: (a) Galway bay; (b) Pico
σ̂l2 = ê(k + l|k)2
 
(10)
N −1 island. Hs is the significant wave height.
k=1

where N is the number of past observations available. This Here η(k + l) is the wave elevation and η̂(k + l|k) is its
estimate can also be recursively updated as soon as new prediction based on the information up to instant k. A 100%
observations become available [13]. value for F(l) means that the wave elevation time series
is perfectly predicted l steps into the future. Note that the
III. R ESULTS WITH IDEAL LOWPASS FILTERS quantity F has a direct correspondence with the variance of the
prediction error, that, as discussed in section II-C, is utilised to
The wave forecasting AR models are tested against real characterise the forecasts confidence intervals (the gaussianity
wave observations from different sea locations: of the multi-steps ahead prediction error distribution has been
1) The Irish Marine Institute provided real observations verified through analysis of skewness and kurtosis).
from a data buoy located in Galway Bay, on the Focus has been put on the low frequency components, as
West Coast of Ireland, at approximately 53o 1300 N , they are the more interesting from a wave energy point of view,
9o 1800 W (water depth nearly 20 m). Data consist of and this allows for a significant improvement of the prediction,
20 minute records sets for each hour, collected at a as shown through a predictability analysis in [6],[7]. The cutoff
sampling frequency of 2.56 Hz. frequency ωc has been chosen accordingly to the spectral
2) Wave elevation time series are also available from distribution of the specific sea state.
the Atlantic Ocean at the Pico Island, in the Azores Fig. 3 shows the very good accuracy achievable with AR
archipelago, at approximately 38o 330 0 N , 28o 3400 W models when narrow banded sea states occur, with F(l) >
(water depth around 40 m). The Pico data are collected 80% for predictions more than 15 s in the future. In the case
in the form of two contiguous 30 minute records sets of wide banded sea states, G1 and P1 , the accuracy is still
for each hour, with a sampling frequency of 1.28 Hz F(l) > 80% for prediction horizons of about 10 − 12 s. The
[14],[15],[16]. sea state G3 , dominated by wind waves and affected by strong
Some significant data sets, reflecting different wave cli- non linearities (detected through the Bispectrum [6]), gives
mates, were picked up from all the ones available, and their more problems and neural networks gave better results in this
spectrum is shown in Fig. 2. In particular one wide banded and case [6],[7].
one narrow banded sea state from the two sites is considered. It is interesting to note how the poles of the AR models
Then, a situation where wind waves predominate is picked up are estimated very close to the unit circle of the complex
from the Galway Bay data and a very high energy wave sys- plane, and the corresponding frequencies map the training
tem, where the sea bottom slightly affects the wave symmetry wave spectrum, as shown in Fig. 4.
(this was analysed through higher order spectral analysis and For the sake of completeness, a screenshot comparing the
skewness and kurtosis indices [6]), is chosen from the Pico 25-steps ahead prediction from an AR model and the real
Island data. wave, together with the 90% confidence interval, is plotted
The accuracy of some AR models on the selected wave time in Fig. 5.
series is measured, for any lead time l, through the following IV. E FFECTS OF REAL TIME FILTERS
index of fitness:
 qP  In order to improve the prediction accuracy and the fore-
2 casting horizon a lowpass filtering of the wave elevation is
k [η(k + l) − η̂(k + l|k)]
F(l) = 1 − pP  · 100% (11) performed on the original signal. This procedure is reasonable
2
k η(k) in view of the fact that wave energy devices, and therefore

2477
 100
low frequencies, as shown in section III and in [6],[7]. In
real time applications, however, lowpass filtering inevitably
fit [%]

set G1, ωc=1.2 rad/s, n=32 introduces an error, in terms of signal amplitude and delay
50
set G2, ωc=1 rad/s, n=24 and its effects on the forecasts need to be addressed.
set G3, ωc=1 rad/s, n=32

0
0 2 4 6 8 10 12 14 16 18 20
A. Digital filters: error and delay
forecasting horizon [s]
If the observed time series representing the wave elevation
(a) is η o (k), then the prediction will be focused on the filtered
100 wave:
η(k) , F (ejω )η o (k) (12)
fit [%]

set P1, ωc=1.2 rad/s, n=24

50
set P2, ωc=0.7 rad/s, n=16 where F (ejω ) is a digital filter.
set P3, ωc=0.7 rad/s, n=16 Ideally, F (ejω ) would have a perfectly zero phase response
0 and a piecewise constant amplitude response being 1 in the
0 2 4 6 8 10 12 14 16 18 20
forecasting horizon [s] pass band and 0 in the stop band. This, however, leads to
(b) a non causal impulse response and, therefore, to a filter
non computationally realisable in real time [17]. Inevitably, a
Fig. 3. Performance: (a) Galway bay; (b) Pico island. realisable digital filter will introduce two inconvenient issues:
1) A non zero phase response, which in the time domain
poles AR model spectrum
Z compex plane unit circle
6 training data spectrum means a delay. In the best of the hypothesis this delay
1 frequencies of AR model is a constant, if the phase response is linear in the
S(ω) [m2 s/rad]
imaginary part

0.5 4 frequency, otherwise different frequency components of

0 the wave elevation will be delayed by a different time.
−0.5
2 In any case it is important that the delay is very small,
because any sample of delay introduced by the filter is
−1
0 a sample lost in the achievable prediction horizon.
−1 −0.5 0 0.5 1 0 0.5 1
real part ω [rad/s] 2) A transition band in the amplitude response, which
(a) (b) causes frequency components in the stop band being not
completely removed by the filter and components in the
Fig. 4. Poles and frequency response of an AR model trained on a wave
elevations data set.
pass band being attenuated.
The two traditional approaches to digital filters design are
their control algorithms, are to be optimised at the waves the discretisation of an analogue filter, which generates an
that contain the most energy, and they are the low frequency Infinite Impulse Response (IIR) filter, or the direct synthesis of
ones [6]. Lowpass filtering is also reasonable because what is a Finite Impulse Response (FIR) discrete time filter (frequency
really required to be predicted, in the control of wave energy response with only zeros). The big advantage of FIR filters is
converters, may be one of the consequences of the incident the possibility to obtain a perfect linear phase response and so
wave elevation on the device, such as the wave excitation a constant delay. IIR filters, on the other hand, allow to obtain
force, and they are all filtered by the lowpass dynamics of far better amplitude responses with very low orders, although
the device itself. they have a non-linear phase response.
AR models achieved great performance in predicting the Several FIR and IIR filters were designed and their perfor-
mance was quantified through:
real wave η(k)
forecasts η(k+25/k) • The delay τ , which is perfectly known for FIR filters.
lower bound confidence 90% In the case of IIR filters it is identified by matching the
upper bound confidence 90% output of the filter with the ideal filtered signal.
wave elevation [m]

0.4 • The Mean Error, M E, between the filtered time series,

η(k), and the correspondent ideally filtered time series,
0 η i (k), normalised with respect to the standard deviation
of the ideally filtered time series:
PN −τ
−0.4 i
1 k=tr |η(k + τ ) − η (k)|
8000 8100 ME =
N − τ − tr + 1
qP
1 N i 2
sampling instant [Ts = 0.7813 s] N k=1 η (k)
(13)
Fig. 5. Confidence interval and 25-steps-ahead predictions of an AR model Here N is the length of the filtered signal, τ is the
with order n = 24 on the data set P2 , filtered with cut-off frequency ωc = estimated delay (or exact for a FIR filter) and tr is the
0.7 rad/s. transient time (in samples), which in theory is infinitely

2478
 Butterworth conditions. From (2), in fact, the predictions are computed
type I Chebyshev from the free evolution of the model based on the last n
type II Chebyshev measurements, where n is the order. In order to analyse the
40 sensitivity, it is handy to write the state space form of the AR
Elliptic
ME [%]

predictor, where the state vector is defined as:

20
xk (j) , [η̂(k + j|k) η̂(k + j − 1|k) ...
T
0 ... η̂(k + j + 1 − n|k)] (14)
0 2 4 6 8 10
estimated delay [s] Here, η̂(k + j|k) is the j-steps ahead prediction of the wave
(a) elevation based on the information up to instant k and η̂(k +
j/k) = η(k + j) if k + j ≤ k, as the information is already
Kaiser window acquired.
40 At any instant k, the evolution of the state vector xk (j)
ME [%]

represents the evolution of the j-steps ahead wave prediction

20 starting from the initial state xk (0) and it is regulated by the
following iterative form:
0 xk (j + 1) = Axk (j)
0 10 20 30 40 50 60 (15)
delay [s] η̂(k + j|k) = Cxk (j)
(b)
where
 
Fig. 6. Performance of digital filters on P2 data set: (a) IIR filters; (b) FIR a1 a2 ... an
filters. 1 0 ... 0
A= (16)
 
long for IIR filters, but that tends towards zero if they are .. .. 
. . . . ... . 
stable so it can be negligible after a certain point. 0 ... 1 0
Figure 6 shows the distribution of a number of digital C = [1 0 ... 0] (17)
filters, with cutoff frequency ωc = 1 rad/s, with respect
to τ and percentage M E, tested against the wave elevation It follows that the direct relationship between the initial state
data set from Pico island, P2 in Fig. 2. The FIR filters were xk (0) and the j-steps ahead prediction is:
designed through Kaiser windows with different specifications
for the transition band and the filter’s order [17], while the η̂(k + j|k) = CAj xk (0) (18)
IIR filters are discretisation, through, the bilinear transform where the initial state contains the last n observations available
of classical analogue filters. In particular, from figure 6(a), at instant k:
the best IIR filters seem to be low order Elliptic, Butterworth T
and type I Chebyshev filters. The higher the order, in fact, xk (0) = [η(k) η(k − 1) . . . η(k + 1 − n) . . .] (19)
the stronger the non-linearity of the phase, so that the phase
An error in the wave observations η(k), such as the one
distortion raises the error, in spite of a sharper transition of
introduced by a real time digital filter, means an error in
the amplitude response. In the case of FIR filters, figure 6(b),
the initial state xk (0), named ∆xk (0), which will have some
the M E decreases with increasing order (directly proportional
additive effect ∆η̂(k + j|k) on the j-steps ahead prediction:
to the delay) because of the perfectly linear phase. The delay,
η̂(k + j|k) + ∆η̂(k + j|k) = CAj xk (0) + ∆xk (0) (20)
 
however, quickly gets very high, so that only very small order
FIR filters can be acceptable.
In any case, it seems hard to obtain a M E < 10% with By exploiting some results about linear algebra, it can be
a delay τ < 5 s so that it is important for the prediction shown that, for any error ∆xk (0) and for any norm operator
algorithm, in this case Autoregressive models, to be robust || · ||:
||∆η̂(j + k|k)||
enough in response to errors of this order of magnitude. Of σ(CAj ) ≤ ≤ σ(CAj ) (21)
course, deeper research may be led in the direction of finding ||∆xk (0)||
an optimal filter minimising both the delay and the M E, where σ(·) and σ(·) are respectively the minimum and the
but the focus at the moment is put on the ability of wave maximum singular value operators.
forecasting AR models to properly deal with such errors in In our specific case CAj ∈ R1×n so that only 1 singular
the observations time series. value is non zero and will be considered as a measure of
sensitivity. In Fig. 7 the singular value of CAj is shown, in
B. AR model wave forecasts sensitivity dB, where the matrix A corresponds to an AR model of order
The sensitivity of the forecasts of an AR model to errors n = 16, trained on the wave elevation data set P2 , as from
in the observations is the same as the sensitivity to its initial figure 2(b). The sensitivity is clearly very high and gets worse

2479
 singular value [dB]
delay. A preliminary analysis of the sensitivity of the estimated
200 AR models, in section IV-B, showed how they can diverge for
very small deviations of the initial conditions, and therefore
100 of the observations. The issue of designing a more robust AR
models was highlighted and further work needs therefore to be
0 done in this direction, in order to make this promising wave
0 5 10 15 20 forecasting approach ready for use in a time domain control
forecasting horizon [samples]
framework for wave energy converters.
Fig. 7. Maximum singular value, in dB, of the matrix CAj ACKNOWLEDGMENT
and worse while the forecasting horizon increases, as the error The authors would like to thank the Irish Marine Institute for
accumulates: providing the data from the Galway Bay. The data from Pico
Island is available by courtesy of the Centre of Climate, Me-
||∆η̂(k + 1|k)|| ≤ 1.5402 × 104 ||∆xk (0)|| teorology and Global Change of the University of the Azores,
||∆η̂(k + 5|k)|| ≤ 3.0381 × 107 ||∆xk (0)|| projects CLIMAAT; CLIMARCOST; MacSIMAR (FEDER-
PIC Interreg IIIB & MAC; MAC2.3/A3; 03/MAC/2.3/A5;
||∆η̂(k + 10|k)|| ≤ 3.7629 × 109 ||∆xk (0)||
05/MAC/2.3/A1; MAC/1/AO89). The research was funded
by the Irish Research Council of Science, Engineering and
A very small error in the observations has a huge effect on Technologies (IRCSET) under the Embark Initiative.
the prediction and in fact this AR model heavily diverges if a R EFERENCES
very small noise is added to the observations. This is a serious
[1] U. A. Korde, “Control system applications in wave energy conversion,”
issue to deal with and further work should be focused on the Proceedings of the OCEANS 2000 MTS/IEEE Conference and Exhibi-
design of robust AR models, which are able to reject certain tion, vol. 3, pp. 1817–24, 2000.
kind of disturbances, in particular the ones introduced by a [2] J. Falnes, Ocean Waves and Oscillating Systems. Cambridge University
Press, 2002.
real time filter, while still maintaining a good accuracy in the [3] ——, “A review of wave-energy extraction,” Marine Structures, vol. 20,
wave predictions. pp. 285–201, 2007.
[4] U. A. Korde, “Efficient primary energy conversion in irregular waves,”
V. C ONCLUSION Ocean Engineering, vol. 26, pp. 625–651, 1999.
This study was focused on the problem of short term [5] M. P. Schoen, J. Hals, and T. Moan, “Wave prediction and fuzzy logic
control of wave energy converters in irregular waves,” Proceedings of
wave prediction, which is a central topic in the wave energy the 16th Mediterranean Conference on Control and Automation, Ajaccio,
field, in order to allow a better effectiveness and economic France, pp. 767–772, 2008.
viability of any WEC. An analysis of AR models, in section [6] F. Fusco and J. Ringwood, “A study on short-term sea profile prediction
for wave energy applications,” Proc. of the 8th European Wave and Tidal
II, highlighted how they implicitly represent cyclical models Energy Conference (EWTEC), pp. 756–765, Uppsala, Sweden, 2009.
where the frequencies are easily estimated with linear least [7] ——, “Short-Term Wave Forecasting for Real-Time Control of Wave
squares (as they are related to the regression coefficients). Energy Converters,” accepted for publication on IEEE Trans. on Sus-
tainable Energy, 2010.
The amplitudes and phases of each harmonic component is, [8] G. E. P. Box, G. M. Jenkins, and G. C. Reinsel, Time Series Analysis:
moreover, implicitly adaptive to the recent observations due Forecasting and Control. Prentice-Hall, 1994.
to the regression terms of the model. [9] L. Piroddi and W. Spinelli, “Long-range nonlinear prediction: A case
study,” Proceedings of the 42nd IEEE Conference on Decision and
Results on real wave elevation data from the Galway Bay Control, Maui, Hawaii USA, pp. 3984–3989, 2003.
and from the Pico Island, in section III, showed how an AR [10] D. S. Shook, C. Mohtadi, and S. L. Shah, “Identification for long-
model estimated on a batch training data set, offered very good range predictive control,” IEEE Proceedings D (Control Theory and
Applications), vol. 138, no. 1, pp. 75–84, 1991.
accuracy up to 15 seconds (in some cases even 20 seconds) [11] H. Akaike, “A new look at the statistical model identification,” IEEE
predictions for the low frequency components of the waves. Transactions on Automatic Control, vol. 19, pp. 716–723, 1974.
It was shown also how the frequencies are automatically [12] G. Schwarz, “Estimating the dimension of a model,” The Annals of
Statistics, vol. 6, no. 2, pp. 461–464, 1978.
estimated in the significant range of the wave spectrum of the [13] R. Isermann, Fault-Diganosis Systems. An Introduction from Fault
training set. The possibility to adapt the model in order to track Detection to Fault Tolerance. Springer, 2006.
variations of the wave spectrum through variable frequencies [14] R. Esteves, A. Valente, M. Costa, F. Reis, and E. Azevedo, “Dados
direccionais de agitao martima nos aores (projecto climaat-clima e
in the model, was also analysed. Because, however, static AR meteorologia dos arquiplagos atlnticos),” 6łs Jornadas Portuguesas de
models were shown to maintain their prediction ability for Engenharia Costeira e Porturia, PIANC, Funchal, 2009.
long times (no performance decrease for 2 hours simulations), [15] E. Azevedo and M. Rodrigues, “Projectos climaat & climarcost clima
e meteorologia dos arquiplagos atlnticos, clima martimo e costeiro,”
their adaptivity is not seen as a main issue at the moment, Workshop Intenacional sobre Clima e Recursos Naturais nos Pases de
and also a simple periodic batch estimate may be a feasible Lngua Portuguesa, WSCRA08, 2008.
solution, or the use of a set of AR models, estimated from [16] C. Barrera, E. Azevedo, M. Rueda, M. Gelado, and L. O., “Real-time
monitoring network in the macaronesian region as a contribution to
different sea conditions, and a switching logic deciding which the coastal ocean observations panel (COOP),” Journal of Operational
one is the more appropriate in real time. Oceanography, vol. 1, pp. 59–69, 2008.
In real-time applications, however, the lowpass filtering [17] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing.
Prentice Hall International Inc., 1989.
introduces an error in the wave elevation time series and a

2480