This article has been accepted for inclusion in a future issue of this journal.
Content is final as presented, with the exception of pagination.
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY
Linear Optimal Noncausal Control of
Wave Energy Converters
Siyuan Zhan
Abstract— This paper addresses the fundamental theoretical
development of a linear optimal noncausal control for wave
energy converters (WECs) in a closed analytic form. It is well
known that the WEC control is a noncausal control problem,
i.e., the future wave information contributes to the present control
action. This paper provides a reliable, efficient, and simple linear
optimal controller with guaranteed stability for the WEC control
problem. The proposed WEC linear optimal control consists of
a causal linear state feedback part and an anticausal linear
feedforward part to incorporate the influence of future incoming
waves. The stability of the closed-loop WEC control system with
the proposed linear optimal controller is proven. The contribution
of the noncausal term using wave prediction information to the
optimal control and the energy output is analyzed quantitatively.
The proposed linear controller can be more preferred when
constraint satisfaction becomes a less important issue for mild sea
states and some well-designed WECs with an ample operation
range. The proposed optimal control strategy can be extended
for a generic class of energy maximization problems. Numerical
simulations are presented to justify the efficacy of the proposed
WEC optimal control.
Index Terms— Optimal control, wave energy, wave prediction
technique.
I. I NTRODUCTION
A SUBSTANTIAL amount of renewable energy is con-
tained in ocean waves. The subject of harnessing wave
energy has been extensively studied in recent decades, espe-
cially after the oil crisis in 1973 [1], and many different
types of wave energy converters (WECs) have been invented
since then. However, the WEC technology is still immature
compared with other renewable energies, such as wind energy
and solar energy [2].
It is well known that control plays an important role
in increasing the energy conversion efficiency of WECs.
Conventional WEC control strategies are based on the
impedance matching principle that energy output can be
maximized when the resonance frequency of a WEC matches
the dominant frequency of the incoming waves [3]. The typical
Manuscript received August 9, 2017; revised December 1, 2017; accepted
February 8, 2018. Manuscript received in final form February 23, 2018.
This work was supported in part by Newton Mobility Grant under
Grant IE150833 and in part by the Newton Advanced Fellowship under
Grant NA160436 through the Royal Society, U.K. Recommended by Associate
Editor L. Wang. (Corresponding author: Guang Li.)
The authors are with the Department of Engineering and Material Sci-
ence, Queen Mary University of London, London E1 4NS, U.K. (e-mail:
s.zhan@qmul.ac.uk; g.li@qmul.ac.uk).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TCST.2018.2812740
1063-6536 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1
and Guang Li , Member, IEEE
conventional WEC control strategies include latching
control [4], [5], phase control [6], and declutching control [7].
However, most of these control strategies are only effective
for regular waves and become very complicated to implement
for irregular waves in real sea conditions.
Optimal control of WECs is essentially different from
the conventional optimal control problems. First, a WEC
controller aims to maximize energy output, whereas con-
ventional optimal control has been developed for reference
tracking or equilibrium regulation problems. Because a WEC
is subject to the persistent disturbance (excitation force from
incoming waves), there are no equilibrium or fixed references
for tracking. Second, it has been proven that to achieve
an optimal operation of a WEC, the future information of
the incoming wave is needed for the implementation of a
WEC controller, which results in a noncausal optimal control
problem [3]. These differences invalidate a direct application
of the conventional optimal control strategies.
Some linear causal suboptimal control strategies have been
recently developed for the WEC control problem, e.g., a non-
standard linear quadratic Gaussian control [8] and the non-
causal control proposed in [9]. Although the implementation
of these causal optimal controllers avoids the hardware cost
for wave prediction, this extra hardware cost can be trivial
compared with the increased energy output gained by incorpo-
rating wave prediction into control in some scenarios. In [10],
we demonstrate that the noncausal optimal WEC control with
wave prediction based on the concept of model predictive
control (MPC) can at least double the energy output than
the energy converted by the traditional causal WEC control
methods. Recent progress in wave prediction techniques, e.g.,
the deterministic sea wave prediction (DSWP) technology
developed in [11], makes the employment of wave prediction
in a WEC control system more economically viable.
In this paper, we develop a generic linear noncausal optimal
control strategy for the WEC control problem. Compared
with the existing causal suboptimal controllers, this controller
can explicitly incorporate wave prediction information into
the WEC control problem. At the price of not explicitly
incorporating constraints into WEC optimal control problem,
the resulting optimal controller takes a compact closed ana-
lytic form, which provides theoretical insights into the WEC
optimal control. The control signal is composed of two terms:
one feedback term determined by the current measurement of
state, and one feedforward term incorporating the influence
of current and future waves. The feedback and feedforward
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

Fig. 2. Free body diagram of the float.

Fig. 1. Schematic of a point absorber.

The energy absorbed during period [T1 , T2 ] is expressed as

gains used to calculate the feedback and feedforward control  T2
signals can be efficiently determined off-line. The stability P(t)dt. (1b)
T1
of the proposed noncausal optimal control is also analyzed,
which provides a theoretical basis toward the stability analysis For safety operation, we restrict the float’s heave motion so
of MPC of WECs when constraints are active. The reliable, that the constraint can be expressed as
efficient, and simple linear optimal controller with guaranteed |z v | ≤ max (2a)
stability can be more preferred when the potential constraints
violation becomes a less significant issue for mild sea states where max is the float heave motion limits. The WEC is also
and some well-designed WECs with an ample operation range, subject to control input limitation
e.g., the M4 device reported in [12]. As an alternative simple
controller to the MPC, the proposed linear controller can be | f u | ≤ u max . (2b)
easily tuned and implemented on economically viable hard- The controller design objective is to maximize the energy (1b)
ware. The proposed optimal control strategy can be extended subject to the state constraint (2a) and input constraint (2b).
to cope with a generic class of direct energy output maxi- Since the linear optimal control (LOC) cannot explicitly
mization problems where optimal references are difficult to incorporate hard constraints into optimization, we use soft
be calculated, or even do not exist, e.g., solar energy and constraints instead. The WEC LOC problem is solved via
wind energy control problems. This method has more potential the Bellman principle [16]. The implementation of the WEC
especially when the energy source can be predicted. LOC is based on the assumption that full-state information is
Although the control method is generic and can be used available from a state observer.
for a wide variety of WECs, we select a particular WEC The rest of this paper is organized as follows. The dynamic
as a case study for demonstration purpose. The device to modeling of a single WEC device is presented in Section II.
be investigated in this paper reflects the working principle Section III provides the WEC optimal control formulation.
of a point absorber whose schematic is shown in Fig. 1. Simulation results are demonstrated in Section IV. Finally,
To present the WEC control problem, we use the models this paper is concluded in Section V.
similar to those used in [13]. The excitation wave force drives
the float, resulting in the relative heave motion between the
II. M ODEL S ETUP
piston fixed to the buoy and the cylinder fixed to the seabed
(or anchored through mooring lines, but with a negligible The free body diagram of the float is shown in Fig. 2.
heave motion of antiheave plate compared with that of the By applying Newton’s second law, we have
float). This relative motion creates fluid flow that drives the
m s z̈ v = − f s − fr + f e + f u (3)
power takeoff (PTO) system to generate power; the typical
PTO mechanism includes direct linear generators [14] and where m s is the float mass; the restoring force f s is given by
hydraulic motors connected to electricity generators [15].
The external manipulated force f u acting on the piston is the f s = ks z v (4)
control input. z w and z v represent the sea surface level and the with the hydrostatic stiffness ks = ρgs, and ρ as water density,
heave position of the middle point of the buoy, respectively. g as standard gravity, and s as the cross-sectional area of the
By defining the direction as shown in Fig. 2, the power output float. fr is the radiation force determined by
of the WEC at time t is expressed as  ∞
fr = m ∞ z̈ v + h r (τ )ż v (t − τ )dτ (5)
P(t) = − f u ż v . (1a) −∞
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHAN AND LI: LINEAR OPTIMAL NONCAUSAL CONTROL OF WECs 3

where m ∞ is the added mass; h r is the kernel of the radiation Here, the following criteria hold.
force that can be computed via hydraulic software packages 1) (1/2)x kT Qx k is used to penalize the state. The weight
(e.g., WAMIT [17]). Following [13], the convolutional term Q influences the stability of the control system and
∞
in (5), f R := −∞ h r (τ )ż v (t − τ )dτ , can be approximated by can be used as a tuning parameter to handle the state
a causal finite dimensional state-space model: constraint (2a).
ẋr = Ar xr + Br ż v (6a) 2) −z k u k represents the power that can be captured by the
 t PTO mechanism, i.e., minimizing z k u k is equivalent to
f r = Cr x r ≈ h r (τ )ż v (t − τ )dτ (6b) maximizing the energy output.
−∞ 3) (1/2)r u 2k is used to penalize the input. The weight r
where ( Ar , Br Cr , 0) and xr ∈ Rnr are the state-space influences the stability of the control system and can
realization and the state, respectively. Following [13], the wave be used as a tuning parameter to handle the input
excitation force f e can be determined by: constraint (2b).
 ∞
Remark 1: If the constraints (2a) and (2b) are not explicitly
fe = h e (τ )z w (t − τ )dτ (7) included in the optimization, an analytical solution can be
−∞
derived. The weights Q and R can be used to penalize x
where h e is the kernel of the radiation force and the state-space
and u in the objective function and can be used as the tuning
approximation is given by
parameters to avoid constraint violations.
ẋ e = Ae x e + Be z w (8a) The linear optimal noncausal control policy for the WEC
 t
control is given in the following theorem.
fe = Ce x e ≈ h e (τ )z w (t − τ )dτ (8b) Theorem 1: The linear optimal noncausal controller for the
−∞
control problem (12d) is
where ( Ae , Be Ce , 0) and x e ∈ Rne are the state-space
realization and the state, respectively. u k = K x,k x k + K w,k wk + K s,k sk+1 (13a)
With the realizations of (6) and (8a), the state-space model
of (3) can be represented by where
  −1  
ẋ = Ac x + Buc u + Bwc w K x,k = − r + BuT Vk+1 Bu C z + BuT Vk+1 A (13b)
(9)
z = Cz x  −1 T
K w,k = − r + BuT Vk+1 Bu Bu Vk+1 Bw (13c)
where w := z w , z := z v , y := ż v , x := [z v , ż v , xr , x e ], and  −1 T
T
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ K s,k = − r + Bu Vk+1 Bu Bu (13d)
0 1 0 0 0 0
⎢ ks Ce Cf ⎥ ⎢ ⎥ ⎢1⎥ and Vk and sk+1 can be calculated via backward iterations
⎢− 0 − ⎥ 0⎥ ⎢ ⎥
Ac = ⎢ m m m ⎥ Bwc = ⎢ ⎣ 0 ⎦ Buc = ⎢ m⎥  T
⎣ 0 Br Ar 0 ⎦ ⎣0⎦
Be Vk = Q k + A T Vk+1 A − C z + BuT Vk+1 A (14a)
0 0 0 Ae 0  −1  
C z = [0 1 01×(nr +ne ) ] (10) × r + Bu Vk+1 Bu
T
C z + Bu Vk+1 A
T

sk = (A + Bu K x,k )T (Vk+1 Bw wk + sk+1 ) (14b)

with m := m a + m ∞ .
The continuous time model (9) can be converted to a with the boundary conditions
discrete time model

x k+1 = Ax k + Bu u k + Bw wk VN = 0 (15a)
(11)
z k = Cz xk s N = 0. (15b)
The WEC LOC development in the remaining paper will be
Proof: Define the optimal cost-to-go (from k to N) to
based on this discrete time model.
be v(x, k). Since the dynamics are linear and the stage cost
are quadratic, a reasonable guess of the cost-to-go function
III. WEC O PTIMAL C ONTROL F ORMULATION can be quadratic on the current state x k
A. Optimal Controller Derivation of WEC
1 T
The optimal control strategy of a WEC can be written as v(x, k) = x Vk x k + x kT sk + ak (16a)
 N−1 2 k
1  where Vk , sk , and ak do not depend on x k and Vk ≥ 0. From
min L k (x k , u k ) (12a)
u N
k=0 1 T
s. t. x k+1 = Ax k + Bu u k + Bw wk (12b) v(x, N) = x VN x N + x NT s N + a N = 0 (16b)
2 N
z k = Cz xk (12c)
we have the boundary conditions VN = 0, s N = 0, and
where N is the number of prediction steps and L k is a stage a N = 0 for any x N . From the Bellman optimality
cost principle [16], we have
1 1
L k = x kT Qx k + z k u k + r u 2k . (12d) v(x, k) = min {L(x k , u k ) + v(x, k + 1)} . (16c)
2 2 uk
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

Algorithm 1 Implementation of Theorem 1 B. Steady-State Solution

In this section, we develop a steady-state solution so that
the stability can be proved and Algorithm 1 can be further
simplified for implementation. We start with the backward
recursion of the discrete-time Ricatti equation (14a).
Assumption 1: The system (A, Bu ) is stabilizable. The
coefficients
 Q CT  of the cost function are chosen such that
Cz r
z > 0. The control horizon is infinite, i.e., N → ∞
in the cost function (12a).
Lemma 1 (Convergence and Stability): If Assumption 1
Define P(x k , u k , wk ) := L(x k , u k ) + v(x, k + 1). Since holds, the solution of the Ricatti equation (14a) converges as
(∂ 2 P(x k , u k , wk ))/(∂u 2k ) = r + BuT Vk+1 Bu > 0, we have N → ∞, i.e., V0 → V and V satisfies the discrete algebraic
∂ P(x k , u k , wk ) Ricatti equation (DARE)
uk =
∂u k V = AT V A + Q
  T  −1  T 
= −(r + BuT Vk+1 Bu )−1 (BuT Vk+1 A + C z )x k − BuT V A + C z r + BuT V Bu Bu V A + C z (17)

+ BuT Vk+1 Bw wk + BuT sk+1 and the solution of the DARE (17) is unique. The control
which results in (13d). Then, we replace u k in (16c) action to be implemented can be simplified as
with (13d). Since (16c) needs to be satisfied for all x k , after u = K x x + K w w + K s s1 (18a)
some calculation and rearrangement, we have
 T where
Vk = Q + A T Vk+1 A − C z + BuT Vk+1 A (16d)  −1  
 −1   K x = − r + BuT V Bu C z + BuT V A (18b)
× r + Bu Vk+1 Bu
T
C z + Bu Vk+1 A
T
 −1 T
K w = − r + BuT V Bu Bu V Bw (18c)
sk = (A + Bu K x,k )T (Vk+1 Bw wk + sk+1 ) (16e)  −1 T
T
K s = − r + Bu V Bu Bu (18d)
and ak satisfies
and s1 is the solution of the online backward recursion (14b).
ak = ak+1 + M1 − M2 + M3 (16f)
With this control law, the system is stable, i.e., A + Bu K x has
where eigenvalues strictly inside the unit circle.
1 Proof: We begin the proof by introducing a nominal
M1 = (Bw wk )T Vk+1 (Bw wk ) (16g) system
2
1  
M2 = (K w,k wk + K s,k sk+1 )T BuT Vk+1 Bu + r x̄ k+1 = A x̄ k + Bu ū k (19)
2
×(K w,k wk + K s,k sk+1 ) (16h) and an optimization problem based on this nominal system
M3 = wkT BwT sk+1 (16i) JN−k (x̄ N−k ) = min x̄ NT Q N x̄ N
ū N−k ,...,ū N−1
which completes the proof.   T   

N−1
x̄ i Q C zT x̄ i
Remark 2: From Theorem 1, we can see that the optimal + . (20)
control consists of two parts: ū i Cz r ū i
i=N−k
1) the causal linear part K x,k x k + K w,k wk , i.e., the part
depending on current state and wave information; By assuming Jk (x̄ N−k ) = x̄ N−k
T V̄N−k x N−k where V̄N−k ≥ 0
2) the anticausal part K s,k sk+1 , where sk+1 depends only and with the Bellman optimality principle [16], we have

on the future wave information provided by some wave JN−k (x̄ N−k ) = min JN−k+1 (x̄ N−k+1 )
u N−k
prediction techniques. This enables one to quantitatively  T   
investigate the influence of anticausal term on energy x̄ N−k Q C zT x̄ N−k
+ .
output. ū N−k Cz r ū N−k
The coefficients K x,k , K w,k , and K s,k can be determined (21)
off-line.
Remark 3: The calculation of the optimal controller (13d) Following the similar method to (14b), we have:
 T
can be partially computed offline for the gains K x,k , K w,k , V̄k = Q + A T V̄k+1 A − C z + BuT V̄k+1 A
and K s,k and partially computed online for sk incorporating the  −1  
× r + BuT V̄k+1 Bu C z + BuT V̄k+1 A (22)
wave prediction information. The control algorithm is shown
in Algorithm 1. with boundary condition V̄N = Q N , with optimal control
Note that the term s1 calculated from the online backward action
recursion can be viewed as the accumulation of the impact
ū ∗k = K̄ x,k x̄ k (23)
from the incoming wave prediction on the optimal control
action. where K̄ x,k − (r + BuT V̄k+1 Bu )−1 (BuT V̄k+1 A + C z ).
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHAN AND LI: LINEAR OPTIMAL NONCAUSAL CONTROL OF WECs 5

By defining Sk = V̄N−k , the backward recursion (22) can Combining (26) and (27), we have
be rewritten as
 T lim Pk∗ (0) = lim Pk∗ (Q N ) = P∞
∗
(0) (28)
Sk+1 = Q + A T Sk A − C z + BuT Sk A k→∞ k→∞
 −1  
× r + BuT Sk Bu C z + BuT Sk A (24) Assume that Ṽ is another solution of (17) other than
V . Then, we have
with S0 = Q N . With (20)–(24), after rearrangement (replacing
x̄ N−k with x̄ 0 ), we have lim Pk∗ (Ṽ ) = x̄ 0T Ṽ x̄ 0 . (29)
k→∞

x̄ 0 Sk x̄ 0 := Pk∗ (Q N ) = min Pk (Q N ) (25) We also have

u 0 ,...,u k−1
lim Pk∗ (Ṽ ) = lim Pk∗ (0̃) = x̄ 0T V x̄ 0 . (30)
where Pk∗ (Q N ) denotes the optimal value of Pk (Q N ) and k→∞ k→∞
k−1  T 
  
With (29) and (30), we have Ṽ = V , i.e., the solution
x̄ i Q C zT x̄ i
Pk (Q N ) := x k Q N x k +
T
. of DARE (17) is unique. 
ū i Cz r ū i
i=0
Remark 4: The control law proposed in Lemma 1 is prac-
Then, we move to the main part of the proof. tically intractable, because solving s1 in (18a) using backward
1) Convergence: By assuming Q N = 0, (22) becomes recursion (14b) requires infinite wave prediction horizon as
identical with (16d). Since (A, Bu ) is stabilizable, there N → ∞. However, only the finite length of wave prediction
exists a feedback control sequence Mk such that the is achievable from the wave prediction technique.
state x 0 can be regulated to the origin in n x steps, Remark 5: Recent studies [18], [19] show that the require-
 T 
where n x is the dimension of the state x k , with control ment for the stage cost CQ Crz to be positive definiteness can
z
sequence u k = Mk x k , Pk (0) = c1 for k > n x , where c1 be relaxed, so that greater weight can be imposed on the energy
is a constant. output term, and the DARE still yields a unique stabilizing
Meanwhile, since solution.
  To make the control law proposed in Lemma 1 practically
Q C zT
> 0, Pk∗ (0) implementable, in the remaining of this section, we will show
Cz r
that if the wave prediction horizon is finite but long enough,
is increasing, i.e., Pk∗ (0) ≤ ∗ (0). Also,
Pk+1 the control law proposed in Lemma 1 can be further simplified.
Pk (0) = x̄ 0 Sk x̄ 0 is upper bounded by Pk∗ (0) ≤
∗ T Assumption 2: The wave prediction technique can provide
Pk (0), which means that Pk∗ (0) is converging, sufficiently long n p steps of future wave prediction.
i.e., lim n→∞ Pk∗ (0) = P∞
∗ (0). Lemma 2: If Assumptions 1 and 2 hold, the control law
Since the above-mentioned condition holds for all x̄ 0 , proposed in Lemma (1) can further be simplified by linear
both the solution of (24) with Q N = 0 and (16d) are optimal noncausal control law
converging, i.e., V0 = S∞ → V as N → ∞ and can
be calculated by (17) and (13d) converges to (18a). u = K x x + Kd w (31a)
2) Stability: From limk→∞ Pk∗ (0) = P∞ ∗ (0), we have
where
 T     −1  
x̄ k Q C zT x̄ k K x = − r + BuT V Bu C z + BuT V A (31b)
lim = 0.  −1 T
k→∞ ū k Cz r ū k
K d = − r + BuT V Bu Bu  (31c)
 Q C zT   T T
Since Cz r
> 0, we have limk→∞ ū k = 0 and T
V = A V A + Q − Bu V A + C z
 −1  T 
limk→∞ x̄ k = 0. The system (19) with optimal control × r + BuT V Bu Bu V A + C z (31d)
action ū ∗ = limk→∞ K̄ x,k x̄ = K x x̄ with K x defined
in (18a) is stable. Then, we have A + Bu K x which has where x is the current state; w := [w0 , w1 , . . . , wn p −1 ]T
all eigenvalues strictly inside the unit circle. is the sequence of predicted future wave heave velocities;
3) Uniqueness of Solution: Assume that u ∗k is the  = (A + Bu K x )T ;  := [V Bw , V Bw , . . . , n p −1 V Bw ].
corresponding optimal control action of Pk∗ (0), and Proof: From Lemma 1, as N → ∞, the coefficients
Q N is an arbitrary positive semidefinite function. K x,k and Vk+1 converge for 0 ≤ k ≤ n p − 1, i.e., K x,k = K x ,
Immediately, we have Vk+1 = V . The backward recursion of sk (16e) can be
rewritten in a compact form
Pk∗ (0) ≤ Pk∗ (Q N ) ≤ Pk (Q N ). (26)
sk = sk+1 + V Bw wk (32)
By assuming Pk (Q N ) using the control action u ∗k ,
we have Pk (Q N ) = Pk∗ (0) + x̄ kT Q N x̄ kT . Since where  = (A + Bu K x )T .
limk→∞ x̄ k = 0, we have Define a sequence of n p steps of predicted wave heave ele-
vation profile as w := [w0 , w1 , . . . , wn p −1 ]T . The backward
lim Pk (Q N ) = lim Pk∗ (0) + lim x̄ kT Q N x̄ kT recursion (32) can be represented by
k→∞ k→∞ k→∞
= lim Pk∗ (0) = P∞
∗
(0). (27)
k→∞ s1 = n p−1 sn p + [0, V Bw , . . . , n p −1 V Bw ]w. (33)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

Substituting (33) into (18a) gives

u = K x x + K s s1 + K w w0
= K x x + K s n p−1 sn p + K d w (34)
where K d , K x , and  are defined in (2).
Note that since  has all the eigenvalues strictly inside
the unit circle, the impact of future waves on the control
input diminishes as the power of  increases. This implies
that the wave prediction of near future has more significant
impact on control actions than that of the far future and thus
justifies that the term K s n p −1 sn p in (34) can be negligible
with a sufficiently large wave prediction horizon n p in real-
time implementation. Then, we have (2), which completes
the proof. 
Fig. 3. WEC linear optimal noncausal control framework.
C. Sea Wave Prediction Techniques
The efficacy of wave prediction techniques plays a key role TABLE I
in the proposed noncausal optimal WEC control strategy. Dif- PARAMETERS U SED FOR THE WEC M ODEL IN S IMULATION S ET A
ferent wave prediction techniques have been developed, which
can be generally classified into two categories. One class of
wave prediction techniques are based on the past sea wave data
measured at the same point of the sea surface where the WEC
is located. This class of methods for the prediction are based
on statistical methods, such as autoregressive models, cyclical
models, and extended Kalman filters, as summarized in [20].
The other class of techniques are based on the measurements
of sea wave elevations at multiple nearby locations with certain information is available. However, in practice, it may not
distances away from the WEC. The information of wave always be realistic or economically viable to measure all the
propagation and wave directions can be used to predict the states of a WEC, in which case, a state observer is required.
sea wave profile at the WEC’s location with tens of seconds We assume that only some outputs can be measured
into the future. A representative wave prediction method in this
category is the DSWP method. We briefly describe its principle y(k) = C x(k) + v(k) (36)
for completeness, and more details can be found in [11] where v represents the measurement noise and the pair (A, C)
and the references therein. Because of the requirement of is assumed to be observable. Then, a standard Kalman observer
fast real-time wave prediction in the linear noncausal optimal can be designed to estimate the full-state information of a
control for WECs, we use a standard linear oceanographic WEC (see [23] for more details). The WEC optimal control
wave model [21], [22]. The weight elevation at the spatial framework is shown in Fig. 3. The future wave profile is
coordinates ( px , p y ) at time t has the form of predicted by, e.g., DSWP [10], [11]. The control input consists
N p Rs of: 1) the anticausal feedforward part from the predicted wave

h( px , p y , t) = A(ωn , θr ) cos[kn px cos(θr )
n=1 r=1
+kn p y sin(θr ) − ωn t + (ωn , θr )]
where N p is the number of frequencies used; Rs is the number
of significant storm directions which can be assumed to be
modest, i.e., Rs < 10; kn is the wavenumber, which is
computed via kn = 2π/λn , where λn is the wave length;
A(ωn , θr ) is the directional magnitude spectrum, where θr is
the propagation direction of an individual wave component
and ωn is the angular frequency; (ωn , θr ) is a phase-lift.
The principle of the DSWP method is to use the wave data
collected from the remote wave sensors at multiple fixed
positions to estimate the parameters in wave model (35) and
use the model to predict the future wave profile.
D. Implementation
In Sections III-A and III-B, we have designed the WEC
optimal control based on the assumption that full-state
information and 2) the causal feedback part from the estimated
states x̃ provided by a state estimator. The WEC optimal
(35) control with estimation can be implemented by Algorithm 2.
Based on the separation principle, the whole system is stable
if both the observer and the controller are stable.
IV. N UMERICAL E XAMPLES
In this section, we demonstrate the efficacy of the proposed
methods using two sets of numerical examples based on dif-
ferent WEC designs. The WEC model used in the simulation
has the form as described in (10). The optimal control policies
for WECs are generated in Algorithm 2.
A. WEC Model With Parameters Used in [13]
and No Model Mismatch
The parameters of the WEC model used in Section IV-A
are adopted from those used in [13], which are summarized
in Table I.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHAN AND LI: LINEAR OPTIMAL NONCAUSAL CONTROL OF WECs 7

Algorithm 2 Implementation of WEC Linear Optimal

Noncausal Control

Fig. 4. Coefficient on sn p is decreasing with the increase of n p , which

means that the term K s n p −1 sn p can be negligible for sufficiently large
wave prediction length n p .

After discretizing the model with a sampling time Ts =

0.1s, we formulate an optimal control with the objective
function as described in (12d). The weights Q and r are
tuned: 1) to be sufficiently small for maximizing the energy
output; 2) to guarantee that a unique stabilizing solution can
be derived from DARE (17); and 3) to ensure that input
and heave constraints are not violated. Here, we choose
Q = diag(6, 9.8, 0, . . . , 0) and r = 0.08, respectively. The
observer weights are tuned at Rv = 0.01 and Rw = 1, Fig. 5. Impact of future wave predictions on control.
respectively.
Then, we move to the formulation of the LOC for WEC
using the finite wave prediction with the prediction time hori-
zon t p := Ts n p , and n p represents the prediction horizon steps.
Note that in designing the WEC LOC, the term K s n p −1 sn p
described in (34), where sn p summarizes the future wave
information for t > t p , is neglected in designing the LOC
for WEC.
Fig. 4 shows that the closed-loop system is stable,
i.e.,  = A + B K x has eigenvalues strictly inside the unit
circle. With such properties, we can see that the coefficient
Fig. 6. 50-s period of wave profile (wave elevation) is used in simulation.
on sn p is decreasing with the increase in n p , which means
that the term K s n p −1 sn p can be negligible for a sufficiently
large wave prediction horizon. The real wave data gathered off the coast of Cornwall, U.K.,
Fig. 5 shows the change of the gain K d = [K d (0), are used for numerical simulations. The 50 s of the heave
K d (1), . . . , K d (n p − 1)] over the prediction steps. Here, elevation used in simulation is shown in Fig. 6.
K d (i ) represents the gain of wave prediction at step n = i , Four cases with different controllers are calculated with dif-
i.e., wave prediction at time t = Ts i . We can see that the ferent wave prediction horizons (0, 1, 2, and 3 s). Figs. 7 and 8
magnitude of K d (i ) decreases with the increase of predic- show the control inputs u = K x x + K d w and the wave
tion step i , which means that the impact of future wave feedforward part signals K d w for the four cases, respectively.
prediction decreases with time in wave feedforward control The limit on control input magnitude is |u| ≤ 8 kN. Note that
part K d w. the LOC with wave prediction horizon of 0 s is the causal
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

Fig. 7. Control input magnitude with different wave prediction horizons Fig. 9. Float heave displacement with different wave prediction horizons.
(0 s corresponds to no wave prediction). Inclusion of feedward control for
wave prediction does not change control input magnitude significantly.

Fig. 8. Feedforward control input magnitude with different wave prediction

horizons (0 s corresponds to no wave prediction). Fig. 10. Power output time simulation (a 27% efficiency increase between
the linear causal control and the linear noncausal control with a 3-s prediction
horizon).

LOC and can be used for comparison with the noncausal

optimal controllers proposed in this paper. By comparing
the control input without wave prediction (i.e., the 0s line)
with other control inputs with wave prediction in Fig. 7,
we can see that the magnitude of a maximal control input
signal does not significantly change when feedforward term is
added. This feature is helpful, because it not only provides
a fair comparison basis but also reduces controller design
complexity, because the WEC control designers do not have
to retune the feedback part when adding the wave feedforward
part.
Fig. 9 shows buoy heave displacement response of the WEC.
Fig. 11. Energy output with different wave predictions.
The constraints on maximal buoy heave displacement is 0.5 m.
We can see that although the trajectories of state responses
of system with different controllers are slightly different, prediction into optimal WEC control with a horizon of 3 s.
the maximal heave displacement magnitudes are similar and Fig. 10 shows that the noncausal optimal WEC control with
the state constraint violation is avoided. wave predictions can help the WEC system to get a better
Fig. 10 shows the power output and energy output of energy conversion efficiency without increasing the state and
system. The energy output without wave prediction is 19.58 kJ, control input magnitudes, i.e., introducing potential constraint
whereas the energy output with wave prediction horizons violations.
of 1, 2, and 3 s are 22.55, 24.36, and 24.94 kJ, respectively, Fig. 11 shows the energy output of the system with different
which represents a 27% energy increase by incorporating wave wave perdition horizons. We can see that the energy output
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHAN AND LI: LINEAR OPTIMAL NONCAUSAL CONTROL OF WECs 9

Fig. 13. Power and energy output of the WEC.

TABLE II
PARAMETERS U SED FOR THE WEC M ODEL IN S IMULATION S ET B

Fig. 12. Control input and buoy position magnitude of the WEC.

increases significantly when the wave prediction horizon is

within 3 s. We also find that the WEC LOC with longer wave
prediction has a better performance; however, the benefit of
further increasing wave prediction horizon beyond 3 s becomes
less obvious.

B. WEC Model With Parameters Used in [13] With

Parametric Uncertainties
To test the robustness of the proposed nonlinear LOC,
we assume that the term m ∞ from the radiation force (5) is
incorrectly modeled at 100 kg instead of the true value 83.5 kg
while the rest of the parameters remain the same as those used
in Section IV-A.
We tune the linear noncausal optimal controller using
Algorithm 2 based on the inaccurate model with
m ∞ = 100 kg. The controller weights in (12d) are
tuned to be Q = diag(6, 10, 0, . . . , 0) and r = 0.09 such that
the control input and state constraints are satisfied for the
incorrect modeled system. The observer weights Rw and Rv
remain the same. The 3-s wave prediction length is used. Fig. 14. Maximal eigenvalue of i decreases slower with the increase in i
After the offline computation of control coefficient compared with that of the smaller WEC in Section IV-A.
K x and K d and Kalman gain L using the incorrect
model, we implement the observer with incorrect estimator performance degradation using the estimator and controller
updates (41) and the controller on the WEC model with designed based on an inaccurate model, which indicates
m ∞ = 83.5 kg. that the linear noncausal optimal control method proposed
Fig. 12 shows the comparison of the control inputs and buoy in this paper has good robustness properties against model
position responses for the cases with and without parametric uncertainties.
uncertainties. Note that in the simulation with parametric
uncertainties, we use the inaccurate model for the designs of
a controller and an observer. We can see that there are no C. Simulation Comparison With a Bigger WEC
significant differences between the two cases (41). To show the universal efficacy of the proposed control
Fig. 13 shows the energy output. We can see that the energy strategy, we do simulations on a bigger WEC than that used
output drops from 24.94 to 24.52 kJ when the parametric in Section IV-B. The parameters of this WEC model is shown
uncertainties are taken into account. There is no significant in Table II.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
10
Fig. 15. Control input magnitude with different wave prediction horizons
(0 s corresponds to no wave prediction).
Fig. 16. Feedforward control input magnitude with different wave prediction
horizons (0 s corresponds to no wave prediction).
By comparing Fig. 14 with Fig. 4, we can see that the
closed-loop system of WEC is less strongly stable compared
with the WEC in Section IV-A. By “less strongly stable,” we
mean that the eigenvalues of closed-loop system A + B K x are
close to the unit circle, which result in i = (A + B K x )i
decreasing slower than that in Section IV-A with the increase
of i . Such properties will have the following impacts.
1) The WEC linear noncausal optimal control requires a
longer prediction horizon for this term to be small
enough and negligible.
2) The wave feedforward control part K d w has a larger
impact on the WEC performance.
Next, we present the time simulations of the WEC with
a designed linear noncausal optimal controller with the same
wave profile used in Section IV-A as shown in Fig. 6.
Three cases with different controllers are calculated with
different wave prediction horizons (0, 30. and 42 s).
Figs. 15–17 show the control inputs u = K x x + K d w,
the feedforward control signals, and buoy heave displacement
for the three cases, respectively.
By comparing Figs. 15 and 16 with Figs. 7 and 8, we can see
that the ratio of the maximum magnitude of the feedforward
part (Fig. 16) to that of the total control (Fig. 15) is much
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY
Fig. 17. Float heave displacement with different wave prediction horizons.
Fig. 18. Power output time simulation (a 37% energy increase between the
linear causal control and the linear noncausal control with a 42-s prediction
horizon).
bigger than the ratio for the smaller WEC case. This means
that the wave prediction plays a more important role in the
control signal in the less stable WEC design. Meanwhile,
the maximal control input signals and maximal buoy heave
displacement are similar over different prediction horizons.
Fig. 18 shows the power output and energy output of
system. The energy output of system without wave prediction
is 5575 kJ, whereas the energy output of system with wave
prediction of 30 and 42 s is 6492 and 7637 kJ, respectively.
From Figs. 15–18, we can see that if a closed-loop system
is less strongly stable, the benefits of incorporating wave
predictions become greater, i.e., 37% of energy conversion rate
increase compared with 27%, while the state and control input
magnitudes are not changed significantly. However, to achieve
a better performance, much longer wave prediction horizons
are required.
Fig. 19 shows the energy output of system with differ-
ent wave perdition horizons. We can see that the energy
output keeps increasing until the wave prediction horizon
reaches 40 s, which validates our analysis that for less
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ZHAN AND LI: LINEAR OPTIMAL NONCAUSAL CONTROL OF WECs
Fig. 19. Energy output with different wave predictions.
stable systems, longer wave predictions are required and
the impact of the wave prediction on systems’ performance
is greater. We can also find some fluctuation that occurs
between 0 and 10 s of the wave prediction horizons. This is
because with an insufficient wave prediction length, as shown
in Fig. 14, the term K s n p −1 sn p is nonnegligible so that
Algorithm 2 with insufficient wave prediction horizons has
a loss of optimality for the optimal control policy for WEC.
V. C ONCLUSION
In this paper, we have developed a linear noncausal opti-
mal control strategy for WECs. The optimal control strategy
consists of a linear causal state feedback part and a linear
anticausal feedforward part to incorporate the influence of
wave prediction. This linear noncausal optimal controller can
directly maximize the energy output with a sufficiently low
computational load, since the gains of the controller can be
calculated offline. It is also easy to tune and find the tradeoff
between the energy output, robust stability, and the satisfaction
of constraints.
This paper also studies the impact of incoming wave predic-
tions on the WEC LOC performance. The simulation shows
that the feedforward part for incorporating wave prediction
information can increase the systems performance without
introducing the issue of constraint violation. The explicit
analysis and demonstration of incoming wave prediction with
different horizons provide a general guideline for finding the
tradeoff between extra cost caused by the requirement of wave
predication and the LOC performance. The results of this
paper also provide the basis to analyze the stability issues
of MPC of WECs. In our future work, we will investigate
the theoretical basis of the robustness of the proposed control
against WEC dynamic uncertainties and wave prediction error.
R EFERENCES
[1] S. H. Salter, “Wave power,” Nature, vol. 249, no. 5459, pp. 720–724,
1974.
[2] J. V. Ringwood, G. Bacelli, and F. Fusco, “Energy-maximizing control
of wave-energy converters: The development of control system technol-
ogy to optimize their operation,” IEEE Control Syst., vol. 34, no. 5,
pp. 30–55, Oct. 2014.
[3] J. Falnes, Ocean Waves and Oscillating Systems. Cambridge, U.K.:
Cambridge Univ. Press, 2002.
[4] A. Babarit and A. H. Clément, “Optimal latching control of a wave
energy device in regular and irregular waves,” Appl. Ocean Res., vol. 28,
no. 2, pp. 77–91, 2006.
[5] U. A. Korde, “Latching control of deep water wave energy devices using
an active reference,” Ocean Eng., vol. 29, no. 11, pp. 1343–1355, 2002.
11
[6] K. Budal, “Model experiment with a phase-controlled point absorber,”
in Proc. 2nd Int. Symp. Wave Tidal Energy, 1981, pp. 191–206.
[7] A. Babarit, M. Guglielmi, and A. H. Clément, “Declutching con-
trol of a wave energy converter,” Ocean Eng., vol. 36, nos. 12–13,
pp. 1015–1024, 2009.
[8] J. T. Scruggs, S. M. Lattanzio, A. A. Taflanidis, and I. L. Cassidy,
“Optimal causal control of a wave energy converter in a random sea,”
Appl. Ocean Res., vol. 42, pp. 1–15, Aug. 2013.
[9] S. R. K. Nielsen, Q. Zhou, M. M. Kramer, B. Basu, and Z. Zhang,
“Optimal control of nonlinear wave energy point converters,” Ocean
Eng., vol. 72, pp. 176–187, Nov. 2013.
[10] G. Li, G. Weiss, M. Mueller, S. Townley, and M. R. Belmont, “Wave
energy converter control by wave prediction and dynamic programming,”
Renew. Energy, vol. 48, pp. 392–403, Dec. 2012.
[11] M. R. Belmont and J. Christmas, “An examination of the feasibility of
linear deterministic sea wave prediction in multidirectional seas using
wave profiling radar: Theory, simulation, and sea trials,” J. Atmos.
Ocean. Technol., vol. 31, no. 7, pp. 1601–1614, 2014.
[12] H. Santo et al., “Extreme motion and response statistics for survival of
the three-float wave energy converter M4 in intermediate water depth,”
J. Fluid Mech., vol. 813, pp. 175–204, Feb. 2017.
[13] Z. Yu and J. Falnes, “State-space modelling of a vertical cylinder in
heave,” Appl. Ocean Res., vol. 17, no. 5, pp. 265–275, 1995.
[14] S. R. Huang, H. T. Chen, C. H. Chung, C. Y. Chu, and G. C. Li,
“Multivariable direct-drive linear generators for wave energy,” Appl.
Energy, vol. 100, pp. 112–117, Dec. 2012.
[15] G. Weiss, G. Li, M. Mueller, S. Townley, and M. Belmont, “Optimal
control of wave energy converters using deterministic sea wave predic-
tion” in Fuelling the Future: Advances in Science and Technologies for
Energy Generation, Transmission and Storage. Boca Raton, FL, USA:
Brown-Walker Press, 2012, pp. 396–400.
[16] R. Bellman, “The theory of dynamic programming,” Bull. Amer. Math.
Soc., vol. 60, no. 6, pp. 503–515, 1954.
[17] C. H. Lee, “WAMIT theory manual,” Dept. Ocean Eng., MIT,
Cambridge, MA, USA, Tech. Rep. 95-2, 1995. [Online]. Available:
http://www.wamit.com/manualv7.2/wamit_v72manualli2.html
[18] O. I. Olanrewaju and J. M. Maciejowski, “Implications of dissipativity
on stability of economic model predictive control—The indefinite linear
quadratic case,” Syst. Control Lett., vol. 100, pp. 43–50, Feb. 2017.
[19] M. Zanon, S. Gros, and M. Diehl, “Indefinite linear MPC and approx-
imated economic MPC for nonlinear systems,” J. Process Control,
vol. 24, no. 8, pp. 1273–1281, 2014.
[20] F. Fusco and J. V. Ringwood, “Short-term wave forecasting for real-time
control of wave energy converters,” IEEE Trans. Sustain. Energy, vol. 1,
no. 2, pp. 99–106, Jul. 2010.
[21] B. Kinsman, Wind Waves: Their Generation and Propagation on the
Ocean Surface. North Chelmsford, MA, USA: Courier Corporation,
1965.
[22] M. J. Tucker and E. G. Pitt, Waves in Ocean Engineering, vol. 5.
Amsterdam, The Netherlands: Elsevier, 2001.
[23] G. Li and M. R. Belmont, “Model predictive control of sea wave energy
converters—Part I: A convex approach for the case of a single device,”
Renew. Energy, vol. 69, pp. 453–463, Sep. 2014.
Siyuan Zhan was born in Yingtan, China, in 1992.
He received the bachelor’s degree from Shanghai
Jiao Tong University, Shanghai, China, in 2013,
and the master’s degree from the University of
Pennsylvania, Philadelphia, PA, USA, in 2014. He is
currently pursuing the Ph.D. degree with the Queen
Mary University of London, London, U.K.
His current research interests include constrained
optimal control, model predictive control, and con-
trol applications in renewable energy.
Guang Li (M’09) received the Ph.D. degree in
electrical and electronics engineering, specialized in
control systems, from the University of Manchester,
Manchester, U.K., in 2007.
He is currently a Lecturer in dynamics mod-
eling and control with the Queen Mary Univer-
sity of London, London, U.K. His current research
interests include constrained optimal control, model
predictive control, adaptive robust control, and con-
trol applications in renewable energies and energy
storage.