You are on page 1of 4

Ship stabilization control using an adaptive input disturbance

predictor
J ingyang. Liu, Robert. Allen, Hong. Yi, Yufang. Zhang

AbstractWhen ships travel on the oceans, changes in the sea
states and the sailing conditions will induce significant uncertain
hydrodynamics, leading to a deterioration in the performance of
traditional stabilization systems. To overcome this problem, a
combination of Model Predictive Control (MPC) and an
adaptive input disturbance predictor is proposed. This
combination predicts the wave disturbance force by using a
predictive model of the input disturbance and then
compensating for the predicted disturbance within the MPC
framework. This has the advantages of MPC and the adaptive
model, and avoids the complicated robust tuning of a state
observer which is commonly used within the MPC framework
to reject output disturbances. Model Predictive Control is better
than classical control at dealing with constraints, and the
adaptive disturbance model enhances the ship adaptability
when traveling in varying sea conditions. Very good predictions
of the ship motion are obtained with less degradation under
changes of sailing conditions, thus achieving very good closed
loop performance in the MPC framework. An adaptive input
disturbance predictor based on the time series Auto Regressive
(AR) model is used in a numerical simulation which shows that
this combination works very well.
I. INTRODUCTION
HIP stabilization control has been studied for many years.
The book of Perez [1] gives us a comprehensive picture
of the development of roll stabilization control since the
1960s. Good roll stabilization performance has been achieved
by various control strategies through active fin and rudder
stabilizers [2]-[4], and several commercial systems can now
be found on the market.
Despite this development, much work remains to be done.
For ships traveling on the oceans, changes in the sea states
and sailing conditions will induce significant uncertain
hydrodynamics which cannot be described accurately by the
nominal model used in control system design [5]. To deal
with such a problem, some conservative designs, such as PID
and H-infinity can produce a less sensitive performance at
specified frequencies [6]. They are, however, based on the
assumption that the wave power spectral density will not
change, which is often not true. Another more
straightforward approach is to measure the approaching
waves and use a feedforward channel to compensate.
However, this is restricted by the lack of measurement
methods, which are still not ready for practical applications.
Oda, Ohstu and Hotta [7], [8] proposed an adaptive rudder
roll stabilizer based on an Auto Regressive Moving Average
(ARMA) model which actually falls into the framework of
General Predictive Control(GPC). Perez and Goodwin [1], [9]
proposed a new framework based on Model Predictive
Control (MPC). Although his main purpose was to use the
advantages of MPC to deal with the constraints, it did give a
solution to address the changes of sea states and sailing
conditions. His main idea was to identify a wave disturbance
model to predict the wave induced motion and then embed it
within the MPC framework to compensate. Whenever the
sailing condition changes, or every 20 minutes, it updates the
state-space disturbance model using the Kalman filter, during
which time the control action is switched off. This is good
enough for large ships, such as oil tankers, which have
moderate speed and a relatively constant sailing course. But
for high speed ships such as fast ferries or naval vessels,
sailing conditions can be regarded as constantly changing.

Manuscript received October 6, 2009.
J . Liu is with the School of Naval Architecture, Ocean and Civil
Engineering, Shanghai J iao Tong University, Shanghai, 200240 China
(phone: +86(0)2134207163; fax: +86(0)2134207163; e-mail:
jy_liu@sjtu.edu.cn).
R. Allen is with Institute of Sound and Vibration Research, University of
Southampton, Southampton, SO17 1BJ , UK. (e-mail: ra@soton.ac.uk).
H. Yi is with the School of Naval Architecture, Ocean and Civil
Engineering, Shanghai J iao Tong University, Shanghai, 200240 China
(e-mail: yihong@sjtu.edu.cn).
Y. Zhang is with the School of Naval Architecture, Ocean and Civil
Engineering, Shanghai J iao Tong University, Shanghai, 200240 China
(e-mail: zhangyf@sjtu.edu.cn).
Moreover, as Maciejowski [10 pp. 56-61] points out, the
system model augmented with a disturbance model can be
used to reject the immeasurable disturbance in MPC, but this
requires a well designed state observer to estimate the
disturbance states. If it is not, uncertainties in both the plant
model and the disturbance model will induce large errors in
the estimated states which are likely to degrade the
performance of the MPC. Efforts to create an effective
adaptive disturbance predictor could also be rendered
ineffective by these errors. To enhance the robustness, robust
observers can be chosen [11], or we can use the methods
proposed by Lee and Yu [12]. However, such methods will
increase the complexity of the design.
This paper will give an alternative method to avoid the use
of the augmented model. It simply calculates the wave
induced force or moment instead of the wave induced motion.
Section 2 will first extend the basic MPC algorithm to deal
with the input disturbance in time series form. Section 3
derives the input disturbance predictor based on an AR model
S
2010 8th IEEE International Conference on
Control and Automation
Xiamen, China, June 9-11, 2010
FrD3.6
978-1-4244-5196-8/10/$26.00 2010 IEEE 2158

and section 4 presents the control algorithm. Section 5
shows the numerical results of the closed loop performance
and the discussed. Conclusions are drawn in section 6.
I. COMBINATION OF MPC AND INPUT DISTURBANCE
Assume that a dynamic system has this discrete state space
form:
( 1) ( ) ( )
( ) ( )
x k Ax k Bu
y k Cx k
+ = +
=
k
(1)
The MPC controller should also respect constraints on control
variables as well as the constrained outputs,
min max
( ) u u k u

min max
( ) u u k u
(2)
min max
( ) y y k y

In which are the control increments. ( ) ( ) ( 1) u k u k u k =
The core of the MPC algorithm is to minimize the
following cost function [10 Ch. 4]:
1 1
2 2
0 0
( ) ( | ) ( ) ( | )
p u
H H
Q
i i

R
J k y k i k r k i u k i k

= =
= + + + +

(3)
In the cost function, are the predicted outputs at
time k based upon and are the predicted
control inputs, are the input references at the time
stepk ; Here we use rather than which is common in
Linear Quadratic control. Matrices are the weighting
matrices which are assumed to be constant over the prediction
horizon.
( | y k i k + )
)
)
i + k ( | u k i k +
( r k i +
i + u u
, Q P
p
H and
u
H
are the prediction and control horizons.
The cost function (3) can be written as
2
( ) ( ) ( )
Q
2
R
J k Y k k U = +
(4)
Where
( | )
( )
( 1|
p
y k k
Y k
y k H k

=


+

(5)
( | )
( )
( 1|
p
r k k
k
r k H k


=


+


(6)
( | )
( )
( 1|
p
u k k
U k
u k H k


=


+


(7)
(Q, ,Q) Q diag = , are the weighting
matrices. By deriving the prediction expressions, we can
write
(R, ,R) R diag =
( ) ( ) ( 1) ( ) Y k x k u k U k = + + (8)
Where
2
1
p
H
C
CA
CA
CA





=



(9)
2
0
0
p
H
i
i
CB
CAB CB
C A

=
B



+
=




(10)
2 1
0 0
0 0 0
0 0
0
p p
H H
i i
i i
CB
CAB CB
C A B C A

= =
u
H
B




+
=




(11)
Set the error as
( ) ( ) ( ) ( 1) E k k x k u k = (12)
Then change it to
( ) ( ) ( ) ( 1) k E k x k u k = + + (13)
Substitute it and (8) into equation (4) and use these prediction
expressions (5)-(7), we can get
T T T
J U H U U G E QE = + (14)
Where
2 (
T
G QE = ) k
(15)
T
H Q R = +
(16)
This is a standard Quadratic Programming (QP) problem.
Thus, the optimal control signals can be obtain by minimize
( ) J k under certain constrains.
argmin
opt T T T
U
U U H U U G

= + E QE
(17)
. . s t L U M
L , M are the constraints matrices.
The real ship motion in seaways suffers a continuous wave
disturbance in a force-superposition form, take the roll
motion as an example, can be expressed as:
b b
fins wave
I p Dp G
p

+ + = +
=

(18)
In which, denotes the roll angle, and p is the roll rate.
The roll moment generated by the fins and waves are denoted
as
b
fins
and
b
wave
on the right hand side of (18).
The relation from the fin angle to
b
fins
can be expressed as:
1 2
b
fins
K p K = + (19)
Details of obtaining the coefficients of (18) and (19) (the
inertia I

, damping D , static restoring matrix G and


1
K ,
2
K ) be find in [1].
Substitute (19) into (18) and rearrange it, we can get the
FrD3.6
2159

standard continues state space form as:
1
1 1 1
1 2
0 1 0
( )
b
wave
I
I G I D K I K p p





= +


+ (20)
Using the zero-order hold method and instead with
and
u
1 b
wave
I

with , we obtain the standard discrete form as:


( 1) ( ) ( ) (
( ) ( )
) x k Ax k Bu k
y k Cx k
k + = + +
=
(21)
where, [ ]
T
x p = , .
2 2
C I

=
If the future prediction of can be gotten from some
kinds of predictive methods. Then we can embed them within
the MPC algorithm. Apply the same route above; we get the
MPC prediction expression as:
( ) ( ) ( 1) ( ) ( 1) Y k x k u k U k k = + + + (22)
Here, is the prediction of input disturbance which
can be expressed as:
( 1) k
2
0
( 1)
( 1) ( | 1)
( 1)
( 1 |
p
H
i
p
i
C k
CA k C k k
T k
C A k H i k

+
=
+



1)

T
(23)
Set the error as
( ) ( ) ( ) ( 1) ( 1) E k k x k u k k
c
=
(24)
The input disturbance is embedded within the MPC
framework directly, and we can still write the cost function as
a quadratic programming form
T T
J U H U U G E QE = +
The only differences from the basic MPC algorithm are that
we need to calculate the past time series of from the sensor
measurements and use a predictive algorithm to get the
prediction . ( 1) k
II. ADAPTIVE INPUT DISTURBANCE PREDICTOR
In the system (20) , means the state vector can be
measured directly. From this we can easily obtain:
2 2
C I

=
( 1) ( ) ( 1) ( ) k x k Ax k Bu k = (25)
This expression reveals that the time series of the
disturbance can be calculated from the attitude
measurements. Although the uncertainties of A and also
induce errors, they will not accumulate to the next time step.
Consequently the estimation will never get too far away from
the real disturbance, which makes using the adaptive method
more practical.
B
In the derivation of section 2, the prediction of the wave
induced motion is not involved. Hence the MPC algorithm
does not require a specific form of the predictor but just the
future time series of the input disturbance. Thus, any
predictor can be included into the above framework when
necessary. From this point, the AR model, which is very
suitable for ship motion prediction [13] is selected here.
An AR model has the form:
1 1
( ) ( 1) ( 2) ( ) ( )
n n
y k y k y k y k n e

k = + + + +

(26)
n is the order of the model which can be obtained by the
minimum AIC Estimate procedure [14]. is Gaussian
noise. The model coefficients can be updated at every
sample using the least mean square method. By doing this
recursively, an adaptive predictive model is produced.
( ) e k
i

III. COMBINATION OF MPC AND ADAPTIVE INPUT


DISTURBANCE PREDICTOR
As mentioned previously, to combine the MPC with input
disturbances, the only modification of the MPC algorithm is
to append the matrix (23) in which all of the unknown are
obtained from a predictive model. Here, we use the AR model
described in section 3.
So the full control system design routine is, at every
sample step:
1) Get the ship motion measurements and current
control input signal .
( ) y k
( ) u k
2) Estimate the past sample ( 1) k by using (25).
3) Update the AR predictive model with the new ( 1) k .
4) Predict sequence ( | 1) ( 1| 1)
p
k k k H k + using
the updated AR predictive model (26).
5) Solve the MPC problem (17) using the algorithm
described in section 2 to obtain the optimal control sequence.
6) Adopt the first control variable of the optimal control
sequence to update the control command.
IV. NUMERICAL SIMULATION
For comparison, we adopted the same nonlinear vessel
model used by Perez and Goodwin [1], [9]. The nonlinear
dynamic feature of fins is discarded here for simplification.
Since though, constraints are considered in the MPC design
to include the mechanical characteristics of fins.
The parameters of the simulation are: Wave spectrum:
ITTC;
1/3
4 H m =
; 7.5 T s = ; Ship speed: 15kt; sample time:
0.25s; MPC predicted horizon: 10 samples. To change the
sailing conditions, the wave direction begins to vary at 50s
from 60 degrees to 45 degrees.
Fig. 1 shows the simulation results. The open loop
performance is the ship response to waves without active
control. The combination of MPC with an adaptive
disturbance model achieves a very good closed loop
performance even the sailing conditions are changing. The
case of MPC with a state space disturbance model which was
pre-identified with a wave angle as 60 degrees degrades
significantly after the sailing conditions depart from the
design conditions at which the model parameters were
estimated. This demonstrates the advantages of the
combination of MPC and an adaptive input disturbance
predictor in dealing with changing sailing conditions.
FrD3.6
2160


The following simulations investigate the tolerance to
model uncertainties of these two methods. Both methods are
carried out with one accurate model and one inaccurate model
which was derived by increasing the diagonal elements of the
state matrix A by 20%. The simulation parameters are the
same as in Fig. 1, except that the wave angle is now kept
constant as 90 degrees.


Fig. 2 shows simulation results of the combination of MPC
and an adaptive input disturbance model. As section 3
discussed, the use of an input disturbance model avoids the
state estimation. It will not induce large errors due to the
model uncertainties, and thus achieves a good robustness.
The use of an output disturbance model needs a well designed
observer to decrease the state estimation errors due to the
model uncertainties. Fig. 3 shows that the inaccurate model
caused an unacceptable close loop performance. In this case a
Kalman filter observer was used to estimate the states but
without robustness consideration.
V. CONCLUSION
In this paper, a general framework combining MPC with an
arbitrary adaptive input disturbance predictor is proposed.
Instead of the output disturbance, the input disturbance can be
calculated directly from the measurements. This avoids the
observer design and will not introduce accumulated errors.
By this simple extension, the MPC algorithm is then modified
to be able to deal with the input disturbance.
The whole study is based on the linear MPC framework,
which is usually a good approximation in ship motion control.
In some cases, if the ship motion faces large amplitude
oscillation, nonlinear MPC may be used with this kind of
predictor and this will be investigated in future studies.
Simulation results show that this method works better than
the state space output disturbance model case when the
sailing condition varies, and the avoidance of the observer
can more easily achieve a better robustness to model
uncertainties.
Although the AR model based predictor is the main
method adopted here, other predictive methods based on the
past time series can also be used within this framework. State
space expression of the predictive model is not essential.
0 50 100 150 200 250 300
-20
-15
-10
-5
0
5
10
15
20
Time (sec)
R
o
ll a
n
g
le
(d
e
g

Open loop
MPC with AR disturbance model
MPC with State space disturbance model
Fig. 1. Comparison of control performance with different disturbance
models with changing sailing conditions (Wave direction varies from60 to
45 degrees).
The simulations in this paper only used fin roll stabilizers
for demonstration. Indeed, this approach can be applied to
other stabilization problems such as longitudinal motion
stabilization of catamarans, and rudder roll stabilizations.
0 50 100 150 200 250 300
-8
-7
-6
-5
-4
-3
-2
-1
0
1
2
Time (sec)
R
o
ll a
n
g
le
(d
e
g
)


Accurate model
Inaccurate model
Fig. 2. Comparison of a combination of MPC and an adaptive input
disturbance model with both accurate and inaccurate models
REFERENCES
[1] T. Perez, Ship Motion Control, Springer, 2005.
[2] J .E. Conolly. Rolling and its stabilization by fins, Transactions of the
Royal Institution of Naval Architects, 1969.
[3] M. Blanke, P. Haals, and K.K. Andreasen. Rudder roll damping
autopilot experience in Denmark, in Proc. Of IFAC work shop
CAMS89, Lyngby, Denmark, 1989.
[4] J . van Amerongen, P. van der Klugt, and H. van Nauta Lemke, Rudder
roll stabilization for ships, Automatica, vol. 26, pp. 679-690, 1990.
0 50 100 150 200 250 300
-15
-10
-5
0
5
10
Time (sec)
R
o
ll a
n
g
le
(d
e
g
)


Accurate model
Inaccurate model
Fig. 3. Comparison of a combination of MPC and a state space output
disturbance model with both accurate and inaccurate models.
[5] A.R.M.J . Lloyd. Roll stabilization by rudder, 4th Ship Control
System Symposium-SCSS, The Netherlands, 1975.
[6] M.Blanke, J . Adrian, K. Larsen, and J . Bentsen. Rudder roll damping
in coastal region sea conditions, in Proc. Of 5th IFAC Conference on
Manoeuvring and Control of Marine Craft, MCMC2000, 2000.
[7] H. Oda, K. Ohtsu, and T. hotta. A study on roll stabilization by rudder
control, Journal of Japan Institute of Navigation, 1995.
[8] H. Oda, K. Ohtsu, and T. hotta. Statistical analysis and design of a
rudder roll stabilization system, Control Engineering Practice, vol. 4,
pp. 351-358, 1996.
[9] T. Perez, G. C. Goodwin, Constrained predictive control of ship fin
stabilizers to prevent dynamic stall, Control Engineering Practice, vol.
16, pp.482-494, 2008.
[10] J . M. Maciejowski. Predictive control with constrains. Prentice Hall,
2002,
[11] Z. Gao, S. Hu and F. J iang. A novel motion control design approach
based on active disturbance rejection, Proceedings of the 40th IEEE
conference on Decision and Control, Orlando, Florida USA, December
2001, pp. 4877-4882.
[12] J . H. Lee and Z. H. Yu. Tuning of model predictive controllers for
robust performance, Computers in Chemical Engineering, vol. 18,
pp.15-37, 1994.
[13] I. Yumori, Real Time Prediction of Ship Response to Ocean Waves
Using Time Series Analysis, OCEANS 1984, vol.13, pp. 1082-1089.
[14] H. Akaike, T. Nakagawa, Statistical analysis and control of dynamic
systems, Kluwer Academic Publishers, 1994.

FrD3.6
2161

You might also like