Linear model identification of the Archimedes Wave Swing

Pedro Beirão1 Duarte Valério2 , José Sá da Costa

Instituto Superior de Engenharia de Coimbra Technical Univ. of Lisbon, Instituto Superior Técnico
Department of Mechanical Engineering Department of Mechanical Engineering – GCAR
Rua Pedro Nunes, 3030-199 Coimbra, Portugal Av. Rovisco Pais 1, 1049-001 Lisboa, Portugal
[email protected] {dvalerio,sadacosta}@dem.ist.utl.pt

Abstract— This paper uses Levy’s identification method to The paper is organised as follows: section II briefly
build linear, second-order models for the Archimedes Wave presents the AWS; section III introduces Levy’s identifica-
Swing (AWS), an off-shore, fully-submerged, point absorber tion method, of which the results are given in section IV;
wave energy converter, expected to behave much like a mass-
spring-damper system, though with relevant non-linearities. conclusions are drawn in section V.
Since very few experimental data is available, data from an
accurate non-linear simulator of the AWS was used. One of
II. T HE AWS
the identified models yields a satisfactory performance, and The AWS is an off-shore, fully-submerged (43 m deep
can now be used for the development of control strategies for underwater), point absorber (that is to say, of neglectable
the AWS.
size compared to the wavelength) WEC. Its main two parts
are the silo (a bottom-fixed air-filled cylindrical chamber)
I. I NTRODUCTION
and the floater (a movable upper cylinder). Due to changes
Sea waves may become in a near future an important in wave pressure, the floater heaves (Fig. 2). When the AWS
source of renewable energy only if devices capable of is under a wave top, the floater moves down compressing the
competing with other proven technologies (like wind energy air inside the AWS. When the AWS is under a wave trough,
or solar energy) are developed. In the current stage of devel- pressure decreases and consequently the air expands and the
opment such devices require yet a great deal of research. This floater moves up [1].
paper about the Archimedes Wave Swing (AWS)—a wave The floater’s heave motion is converted into electricity by
energy converter (WEC) of which a 2 MW prototype has an electric linear generator (ELG). The AWS can hence be
already been built (Fig. 1), tested at the Portuguese northern expected to behave much like a mass-spring-damper system,
coast during 2004, and then decommissioned—intends to though with relevant non-linearities.
be a contribution towards that objective, by identifying
III. AWS LINEAR DYNAMIC MODEL IDENTIFICATION
(using a so-called wave frequency-amplitude analysis) an
approximate linear model for this WEC, fit for conceiving An accurate non-linear simulator of the AWS, the AWS
and testing control strategies. Time Domain Model (TDM), has already been developed
and implemented in Matlab [4], [6], [7]. But before thinking
about the extensive use of a non-linear model of the AWS
for control purposes, a linear model approximation of that
same WEC should be identified in the first place.
System identification deals with the construction of math-
ematical models of dynamical systems using measured data.
In the particular situation of the AWS, due to operational
problems of the prototype mentioned above, very few exper-
imental data is available. Thus, a different approach has been

Fig. 1. The 2 MW AWS prototype

1 Pedro Beirão was partially supported by the “Programa do FSE-UE,

PRODEP III, acção 5.3, III QCA”.
2 Duarte Valério was partially supported by grant SFRH/BPD/20636/2004
of FCT, funded by POCI 2010, POS C, FSE and MCTES.
Research for this paper was partially supported by POCTI-SFA-10-46-
IDMEC. Fig. 2. AWS working principle
 −90 −90
gain / dB

gain / dB
−120 −120

−150 −1 0 1
−150 −1 0 1
10 10 10 10 10 10
ω / rad⋅s−1 ω / rad⋅s−1
180 0
phase / o

o
90

phase /
−90
0

−90 −1 0 1
−180 −1 0 1
10 10 10 10 10 10

Fig. 3. Bode diagram of model (22); dots mark data used for identification Fig. 5. Bode diagram of model (24); dots mark data used for identification

1.5 1.5

1 1

0.5 0.5
Imag

Imag
0 0

−0.5 −0.5

−1 −1

−1.5 −1.5
−0.5 0 0.5 −0.5 0 0.5
Real Real

Fig. 4. Pole-zero map of model (22) Fig. 6. Pole-zero map of model (24)

followed. The AWS TDM was used as an emulator of the Minimising this norm (or its square) would be an obvious
real non-linear AWS WEC. Based on simulation results from but difficult way of adjusting the parameters of (1). Instead
the AWS TDM, a linear dynamic model of the AWS was of this, Levy’s method minimises the square of the norm of
estimated. Since sea waves are periodic oscillations (even
if not sinusoidal), an identification method in the frequency (ω)D(ω) = G(jω)D(ω) − N (ω) (4)
domain, such as the classical method of Levy, is the obvious
(which is easier). Let us call this new variable E (and omit
choice. It should be noticed, however, that due to industrial
the frequency argument ω to simplify the notation); we will
secrecy reasons several parameters of the AWS TDM have
have
been modified.
Levy’s identification method [3] is as follows. Let us sup- E = GD − N
pose we have a plant G with a known frequency behaviour.
= [Re(G) + jIm(G)] (σ + jτ ) − (α + jβ)
Let us suppose we want to model it using a transfer function

m = [Re(G)σ − Im(G)τ − α]
k
b0 + b1 s + b2 s2 + . . . + bm sm k=0 bk s
Ĝ(s) = =
n + j [Re(G)τ + Im(G)σ − β] (5)
1 + a1 s + a2 s2 + . . . + an sn 1 + k=1 ak sk
(1) The square of the norm of E is
The frequency response of (1) is given by
2

m k |E|2 = [Re(G)σ − Im(G)τ − α]

k=0 bk (jω) N (ω) α(ω) + jβ(ω) 2
Ĝ(jω) = n = = + [Re(G)τ + Im(G)σ − β] (6)
1 + k=1 ak (jω) k Dω σ(ω) + jτ (ω)
(2) From (2) we see that
where N and D are complex-valued and α, β, σ and τ (the
m

real and imaginary parts thereof) are real-valued. The error  
between model and plant, for a given frequency ω, will be α(ω) = bk Re (jω)k (7)
k=0
N (ω) m
 
(ω) = G(jω) − (3) β(ω) = bk Im (jω)k (8)
Dω
k=0
 Regular wave, amplitude 0.5 m, period 10 s Regular wave, amplitude 1.0 m, period 8 s
1.5 2
force−position force−velocity non−linear
1.5
1
1
position / m

position / m
0.5
0.5

0
0

−0.5
−0.5
−1

−1 −1.5
300 320 340 360 380 400 300 320 340 360 380 400
time / s time / s
Regular wave, amplitude 0.75 m, period 10 s Regular wave, amplitude 1.0 m, period 12 s
2 2.5

1.5 2

1.5
1
1
position / m

position / m
0.5
0.5
0
0
−0.5
−0.5

−1 −1

−1.5 −1.5
300 320 340 360 380 400 300 320 340 360 380 400
time / s time / s
Regular wave, amplitude 1.0 m, period 10 s Regular wave, amplitude 1.0 m, period 14 s
2.5 3

2
2
1.5

1
position / m

position / m

1
0.5

0 0

−0.5
−1
−1

−1.5 −2
300 320 340 360 380 400 300 320 340 360 380 400
time / s time / s
Regular wave, amplitude 1.25 m, period 10 s
3
Fig. 7. Floater’s position for AWS TDM and identified linear models (100 s
long period out of 600 s)
2

If we differentiate |E|2 with respect to one of the coefficients

position / m

1
bk and equal the derivative to zero, we shall have
0 ∂|E|2
= 0⇔
∂bk
 
⇔ [Re(G)σ − Im(G)τ − α] Re (jω)k
−1
 
−2
+ [Re(G)τ + Im(G)σ − β] Im (jω)k = 0 (11)
300 320 340 360 380 400
time / s
If we differentiate |E|2 with respect to one of the coefficients
ak and equal the derivative to zero, we shall have
n
   ∂|E|2
σ(ω) = 1 + ak Re (jω)k (9) = 0⇔
   ∂ak 
k=1 2 2
n
 ⇔ σ [Im(G)] + [Re(G)] Re (jω)k
    
τ (ω) = ak Im (jω)k (10) 2 2 
+ τ [Im(G)] + [Re(G)] Im (jω)k
k=1
      f
+ α Im(G)Im (jω)k − Re(G)Re (jω)k (12)    
     gl,1 = −Re (jωp )l
+ β −Im(G)Re (jω)k − Re(G)Im (jω)k = 0 p=1

2 2
The m+1 equations given by (11) and the n equations given {Re [G(jωp )]} + {Im [G(jωp )]} ,
by (12) form a linear system that may be solved so as to find l = 1...n (21)
the coefficients of (1). Usually the frequency behaviour of
the plant is known in more than one frequency (otherwise The input and the output of the system must be chosen in
it is likely that the identified model will be rather poor). advance. Two possibilities were explored for the AWS.
Let us suppose that it is known at f frequencies. Then the The first one considers the wave excitation force Fexc as
system to solve, given by (11) and (12) written explicitly on the input and the floater’s velocity Ξ̇ as the output. Fexc and
coefficients a and b, is Ξ̇ data provided by the AWS TDM for regular (sinusoidal)
waves with periods from 8 s to 14 s (this is the range the
A B b = e AWS was conceived for [6]) was used. According to wave
(13)
C D a g data provided by ONDATLAS software [5] for the Leixões-
buoy location (41o 12.2
N, 9o 5.3
W), near the test site where
where
the AWS prototype was submerged (at the Portuguese coast,
f
 5 km offshore Leixões), the most frequent significant wave
  
Al,c = −Re (jωp )l Re [(jωp )c ] height Hs (from trough to crest) is admitted as being equal
p=1 to 2 metres. Hence several waves with a 1 m amplitude
  
−Im (jωp )l Im [(jωp )c ] , (half of Hs ) and different periods were assumed for the
l = 0...m ∧ c = 0...m (14) simulations. (Notice that an approximation is involved here,
f since these waves used for identification are regular, while
   
Bl,c = Re (jωp )l Re [(jωp )c ] Re [G(jωp )] those addressed by ONDATLAS are real, irregular waves.)
p=1 To apply the Levy identification method, Matlab’s function
  levy was used [8]. The data found in Table I was used in
+Im (jωp )l Re [(jωp )c ] Im [G(jωp )]
  that process.
−Re (jωp )l Im [(jωp )c ] Im [G(jωp )] Regarding Levy’s identification method, all combinations
  
+ Im (jωp )l Im [(jωp )c ] Re [G(jωp )] , of values for the numerator and denominator orders m and n
l = 0...m ∧ c = 1...n (15) from 0 to 5 were tried. Only identified models with two poles
f or more and one zero or more reproduced the wave frequency
   
Cl,c = −Re (jωp )l Re [(jωp )c ] Re [G(jωp )] behaviour correctly. The identified model structure
p=1
  Ξ̇(s) 2.171 × 10−6 s − 6.759 × 10−7
+Im (jωp )l Re [(jωp )c ] Im [G(jωp )] = (22)
  Fexc (s) 0.967s2 + 0.5874s + 1
−Re (jωp )l Im [(jωp )c ] Im [G(jωp )]
   with one (non-minimum phase) zero and two (stable, com-
− Im (jωp )l Im [(jωp )c ] Re [G(jωp )] , plex conjugate) poles is the one that reproduces the AWS
l = 1...n ∧ c = 0...m (16) TDM responses making use of as few parameters as possible.
f
 By adding an extra pole at the origin, model
2 2
Dl,c = {Re [G(jωp )]} + {Im [G(jωp )]}
Ξ(s) 2.171 × 10−6 s − 6.759 × 10−7
p=1
   = (23)
Fexc (s) s (0.967s2 + 0.5874s + 1)
Re (jωp )l Re [(jωp )c ]
   relating the wave excitation force to the floater’s position Ξ is
+Im (jωp )l Im [(jωp )c ] ,
found. Figures 3 and 4 show, respectively, the Bode diagram
l = 1...n ∧ c = 1...n (17)
  and the pole-zero map of model (22). This model’s major
b0 drawback is its complexity. So another solution was looked
 
b =  ...  (18) for.
bm The second possibility is to consider the wave excitation
 force Fexc as the input and the floater’s position Ξ as the
a1
  output and provide this data (see Table I) to Levy’s identi-
a =  ...  (19) fication method. Since it was found that the former period
an range had insufficient data to allow a good identification,
f
    it had to be enlarged to 4 s to 14 s in order to obtain an
el,1 = −Re (jωp )l Re [G(jωp )] acceptable model. Under this new assumption, the identified
p=1 model, a second order transfer function, was
  
− Im (jωp )l Im [G(jωp )] , Ξ(s) 2.259 × 10−6
= (24)
l = 0...m (20) Fexc (s) 0.6324s2 + 0.1733s + 1
 TABLE I
DATA USED IN THE IDENTIFICATION

period / s 4 5 6 7 8 9 10 11 12 13 14
Fexc ampl. / kN 31.88 108.57 202.64 290.46 365.29 423.35 467.48 501.30 527.62 548.44 565.10
Ξ amplitude / m 0.1195 1.1359 1.1905 1.1939 1.2972 1.3573 1.5558 1.4766 1.4132 1.3803 1.4023
Ξ gain / dB 108.53 99.61 104.62 107.72 109.20 109.88 109.56 110.62 111.44 111.98 112.11
Ξ phase / o −160.20 −85.44 −26.80 −15.77 −14.10 −27.20 −30.72 −16.08 −9.00 −8.31 −11.31
Ξ̇ ampl. / ms−1 — — — — 1.0341 1.0535 1.0686 1.0071 0.8813 0.7999 0.7488
Ξ̇ gain / dB — — — — −110.96 −112.08 −112.82 −113.94 −115.54 −116.72 −117.56
Ξ̇ phase / o — — — — −111.60 −112.00 −97.92 −81.82 −78.60 −78.09 −76.11

TABLE II
C HARACTERISTICS OF SEVERAL IRREGULAR WAVES ACCORDING TO ONDATLAS

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
Hs / m 3.2 3.0 2.6 2.5 1.8 1.7 1.5 1.6 1.9 2.3 2.8 3.1
Te,min / s 5.8 5.8 5.2 5.5 5.0 4.7 4.6 5.0 5.2 5.3 5.5 5.3
Te,max / s 16.1 14.5 13.7 14.8 12.2 9.7 11.1 10.5 12.0 12.6 13.3 14.2

Figures 5 and 6 show, respectively, the Bode diagram and (Ξ̂ being the estimate of the floater’s position) are given,
the pole-zero map of this last identified model. for these simulations and others similar thereto, in Tables III
Even though the input of both transfer functions (22) and and IV.
(24) is the wave excitation force Fexc , the outputs take into From these results, it is seen that model (24) reproduces
account the effects of the radiated force as well, since it was the AWS TDM behaviour more accurately; it also requires
included in the AWS TDM. less parameters than (22), and its structure is similar to the
one normally assumed in the literature (e.g. [2]). Actually,
IV. R ESULTS (22) performs slightly better than (24) for regular waves
600 s (10 min) long simulations were carried out, employ- of low period and high amplitude. But these cases are a
ing the AWS TDM (for the non-linear case) and Simulink im- minority, and simulations with irregular waves (with which
plementations of (22) and (24) (for the linear cases). Models (24) is systematically better) are deemed more important
were submitted to several regular (sinusoidal, with different since they are expected to reproduce the behaviour of real
periods and amplitudes) and irregular incident waves with sea waves more accurately.
amplitudes and periods within the ranges expected to occur, There is an additional reason to prefer model (24), related
based in data supplied by ONDATLAS software for the to the resistance R, which is the real part of the inverse of
Leixões-buoy location. the transfer function from the wave excitation force to the
In what concerns irregular waves, Pierson-Moskowitz’s floater’s velocity:
spectrum, that accurately models the behaviour of real sea Fexc (jω)
waves [2], was used. This is given by R(ω) = Re (27)
Ξ̇(jω)
 
A B R may be frequency dependent, but it is physically impos-
S(ω) = 5 exp − 4 (25)
ω ω sible that it be negative [2]. Indeed, R is always positive
where S is the wave energy spectrum (a function such for (24) (actually in this particular case it does not even
+∞ depend on ω). But, for some frequencies, (22) leads to a
that 0 S(ω)dω is the mean-square value of the wave
elevation). The numerical values A = 0.780 (SI) and B = negative value of R, as seen in Fig. 9. This seems to denote
3.11/Hs2 were used. Values for the significant wave height an inaccurate identification in the case of (22).
Hs (from trough to crest) and for the limits of the frequency For these reasons, model (24) was the one chosen.
range (corresponding to the maximum and minimum values
of the wave energy period Te ) were those provided by V. C ONCLUSIONS
ONDATLAS for the twelve months of the year (see Table II). From the last section it can be seen that the identified
From these simulations, 100 s slices corresponding to second order linear model approximation (24) yields a sat-
seven regular waves and two significant months are high- isfactory performance. This model is a first step towards the
lighted in Figures 7 and 8; the root mean-square errors, given development of control strategies for the AWS. This would
by be difficult with a non-linear model only. Now this simpler
 600
1 2 model can be used for controller design and testing, and the
RM S = Ξ − Ξ̂ dt (26)
600 0 non-linear model for validation.
 Irregular wave for March Irregular wave for June
2 1

1.5

1 0.5

0.5
position / m

position / m
0 0

−0.5

−1 −0.5

−1.5
non−linear force−position force−velocity
−2 −1
300 320 340 360 380 400 300 320 340 360 380 400
time / s time / s

Fig. 8. Floater’s position for AWS TDM and identified linear models (100 s long period out of 600 s)

TABLE III
ROOT MEAN - SQUARE ERRORS FOR THE SIMULATIONS IN F IG . 7 AND OTHERS SIMILAR THERETO

Wave Wave period / s

amplitude / m Model 8.0 10.0 12.0 14.0
(22) 0.2001 0.2417 0.3528 0.4005
0.5
(24) 0.0886 0.1203 0.0674 0.0604
(22) 0.2717 0.3376 0.4912 0.5642
0.75
(24) 0.1919 0.2523 0.1476 0.1326
(22) 0.3380 0.4293 0.6100 0.7059
1.0
(24) 0.3319 0.4211 0.2639 0.2400
(22) 0.4105 0.5206 0.7139 0.8283
1.25
(24) 0.5181 0.6376 0.4259 0.3911

TABLE IV
ROOT MEAN SQUARE ERRORS FOR THE SIMULATIONS IN F IG . 8 AND OTHERS SIMILAR THERETO

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
model (22) 0.5490 0.5039 0.5096 0.5316 0.3677 0.4161 0.3296 0.3283 0.3776 0.4576 0.4490 0.6602
model (24) 0.4227 0.3472 0.2622 0.2386 0.1357 0.1375 0.1177 0.1229 0.1423 0.1991 0.2986 0.3828

ACKNOWLEDGMENT R EFERENCES
The authors acknowledge the very useful comments and [1] P. Beirão, D. Valério, and J. Sá da Costa. Phase control by latching
suggestions made by Professor Johannes Falnes (Professor applied to the Archimedes Wave Swing. In Proceedings of the 7th
Portuguese Conference on Automatic Control, Lisbon, 2006.
Emeritus at the Norwegian University of Science and Tech- [2] J. Falnes. Ocean waves and oscillating systems. Cambridge University
nology) on the identification process. Press, Cambridge, 2002.
[3] E. Levy. Complex curve fitting. IRE transactions on automatic control,
5
4:37–44, 1959.
x 10 [4] P. Pinto. Time domain simulation of the AWS. Master’s thesis,
15
Technical University of Lisbon, IST, Lisbon, 2004.
[5] M. T. Pontes, R. Aguiar, and H. Oliveira Pires. A nearshore wave
energy atlas for Portugal. Journal of Offshore Mechanics and Arctic
10
Engineering, 127:249–255, August 2005.
[6] J. Sá da Costa, P. Pinto, A. Sarmento, and F. Gardner. Modelling and
−1

simulation of AWS: a wave energy extractor. In Proceedings of the

R / N⋅s⋅m

5 4th IMACS Symposium on Mathematical Modelling, pages 161–170,

Vienna, 2003. Agersin-Verlag.
model (24) [7] J. Sá da Costa, A. Sarmento, F. Gardner, P. Beirão, and A. Brito-
0 Melo. Time domain model of the Archimedes Wave Swing wave energy
converter. In Proceedings of the 6th European Wave and Tidal Energy
model (22) Conference, pages 91–97, Glasgow, 2005.
[8] D. Valério and J. Sá da Costa. Identification of fractional models from
−5 −2
10 10
0
10
2 frequency data. In J. Tenreiro Machado, J. Sabatier, and O. Agrawal,
−1 editors, Advances in Fractional Calculus: theoretical developments and
ω / rad⋅s
applications in Physics and Engineering. Springer-Verlag, 2006.
Fig. 9. Evolution of R for both models