Chapter 2
OPTIMAL CONTROL APPLICATION
TO POWER SYSTEMS
We must realize by now that design of generator
controllers as we know them today has followed
“classical” lines. Root-locus methods, block-diagram
representation, intuition, trail and error, empiricism are
but a few of the “classical” design tools. The high degree
of reliability and general excellent functioning of today’s
power systems is proof of the soundness of design. In view
of this fact is there really any room for improvements?
The phenomenon of intermachine oscillations
demonstrates that our present power systems are not
immune to instabilities. It also accentuates an inherent
weakness of classical design-its inability fully to
anticipate and adequately to cope with all dynamic
features in systems of high dimensionality.
A modern gigawatt generator with its multistage reheat
turbine, including its ALFC and AVR controllers, is
characterized by an impressive complexity. When all its
nonnegligible dynamics are taken into account, including
cross-coupling between control channels, the overall
dynamic model may by of 20th order.
This “dimensionality barrier” can be nicely overcome
by means of computer-aided optimal control design
methods and to apply these methods to power systems.
A recent survey classifies the many contributions that
have later followed. “Optimal control” covers today a
widening spectrum of computer-oriented techniques of
which the so-called optimum linear regulator (OLR)
design has proven particularly useful.
Compactly stated, the OLR design results in a
controller that minimizes both transient variable
excursions and control effors. In practical power
systems terms this means optimally damped oscillations
with minimum wear and tear of control valves. OLR
design proceeds as follows:
Step 1 Casting the system dynamic model in state
variable form and introduction of appropriate
control forces.
Step 2 Choosing an integral-squared-error control
index, the minimization of which is the control
goal.
Step 3 Finding the structure of the optimal controller
that will minimize the chosen control index.
We demonstrate briefly the use of OLR design by
applying it to the single-area ALFC system depicted in
Fig. 2-1 we propose to find a control force u (depicted in
Fig. 2-1) which in an “optimal” manner steers or
controls the system. We define “optimal” later; presently
we develop an appropriate model for the system.
Figure (2.1) OLR control of primary ALFC loop.
2-1 Putting the Dynamic Model in State Variable Form
OLR design is based entirely upon the availability of a
dynamic system model in so-called state-variable form. We
demonstrate how this form can be obtained from the block
diagram model in Fig. 2-1.
For this system we have the transform equations:
1 1
Pv s u s f s
1 sTH R
1
PT s Pv s (2.1)
1 sTT
Kp
f s PT s D s
1 sTp
Or, if expressed in the time domain,
d
Pv TH Pv u 1 f
dt R
d
PT TT Pv Pv (2.2)
dt
d
f Tp f K p PT K p PD
dt
At this juncture we introduce the three state variables x1, x2, and
x3 forming the state vector
x1 Pv
x x 2 PT (2.3)
x 3 f
We also define the disturbance force:
p Δ Δ PD (2.4)
Equations (2.2) can, upon rearrangement, be put in the form
x 1 1 0
1 x1 1
TH RTH 0
TH
1
x
1
0 x 0 u 0 p
2 TT TT 2
K p
Kp
T
0
1 p
x 3 Tp Tp x 3 0
This is the sought “state-variable form”. Using matrix
and vector symbols it can be put in more compact form:
x Ax Bu p (2.6)
The single-area system had only one control force and
also one disturbance force.
The single-area system had only one control force and
also one disturbance force.
Figure (2.2) Block diagram of linear state model. Dashed
feedback portion represents the “optimum” controller
In the general case we shall assume x to be n-
dimensional and u of dimension m.
Figure (2.2) depicts a block symbolic representation of
Eq. (2.6).
2.2 Optimum Control Index I
Our next concern is to select an optimum index. OLR
design is based upon an integral-squared-error index of
the form.
I q1 x12 q n x 2n r1u12 rm u 2m dt (2.7)
0
where the q’s and r’s are positive penalty factors.
For example, in our single-area example system the I
index would read
I q1 Pv 2 q 2 PT 2 q 3 f 2 ru 2 dt (2.8)
0
As we shall find in the next section, the whole idea of
“optimum” control is to minimize the I index. Consider
for example the term q3(f)2. Due to the squaring a
frequency error will contribute to I independent of its
sign. Furthermore if f is doubled its contribution to I
will quadruple. The integral causes f to add to I during
its entire duration. The penalty factors qi distribute the
penalty weight among the state-variable errors. Should
the error in a certain state variable be of little practical
signifiance then we simply set its penalty factor equal to
zero.
What has been said about state errors applies equally to
control-force increments. The penalty factor ri
distribute the penalties among the m control forces. It is
important that none of the r’s be set equal to zero.
Should this be the case the corresponding control force
could assume infinite magnitude without affecting I. An
infinite control force could do its correcting job in zero
time. This would obviously be a very unrealistic
regulator.
If all the q’s and r’s constitute the diagonal elements of
the two penalty matrices.
q1 0 r1 0
r2
q2
Q R
and (2.9)
0 q n 0 rm
Then the I index in Eq. (2.7) can be written in the
compact form
I x T Qx u T Ru dt (2.10)
0
2.3 OLR Structure – The Riccati Equation
1. The linear regulator is optimal if the m control forces
are linearly dependent upon the n state variables, i.e.,
u1 k11 x1 k1n x n
(2.11)
u m k m1x1 k mn x n
Or in compact vector-matrix notation
uopt = Kxopt (2.12)
where the gain matrix K is of order m n. Equation
(2.12) settles immediately the question of structure. The
system in Fig. (2.2) is obviously turned into an OLR by
“closing the loop” via K as shown dashed in the figure.
2. The elements of the gain matrix K are computed by
first solving the matrix Riccati equation
PA + ATP – PBR-1BTP + Q = 0 (2.13)
for the unknown n n matrix P. With P known the K
matrix is then obtained from the matrix product
K = - R-1BTP (2.14)
Note that all matrices A, B, Q, and R that appear in
Eqs. (2.13) and (2.14) have been earlier defined.
Superscript ( )T means transposition.
Example: Find the gain matrix K that will make the
system
x Ax Bu
Or
x 1 0 1 x1 0
x 0 0 x 1 u
2 2
an OLR
Solution: We have
0 1 0
A and B
0 0 1
0 0
A
t
and Bt 0 1
0 1
Before we can solve the problem we must settle for
suitable penalty factors. We choose for simplicity equal
penalty weight on x1, x2, and u, and thus obtain.
R-1 = r-1 = 1
(Note that R and R-1 are scalars in this case).
Before we can solve the problem we must settle for
suitable penalty factors. We choose for simplicity equal
penalty weight on x1, x2, and u, and thus obtain.
1 0
Q R r 1
0 1
R-1 = r-1 = 1
(Note that R and R-1 are scalars in this case).
The triple product BR-1BT appearing in the third term
of the Ricatti equation equals
0 0 0
1 1 0 1 0 1
The Riccati equation thus takes on the form.
p11 p12 0 1 0 0 p11 p12
p
21 p 22 0 0 1 0 21
p p 22
p11 p12 0 0 p11 p12 1 0 0 0
0 1 0 0 (2.15)
p 21 p 22 0 1 p 21 p 22
After performing the multiplications called for this 2 2
matrix equation reduces to a set of the four scalar
equations: p p 1 0
12 22
p11 p12 p 22 0
p11 p 21 p 22 0 (2.16)
p 21 p12 p 222 1 0
Their solution is simple and yields matrix
3 1
P
1 3
We can now compute the K matrix:
3 1
K 1 0 1
1 3 (2.17)
1 3
The optimal feedback controller thus is characterized by
the control strategy
u u opt 1
x1
3 x1 3 x 2 (2.18)
x 2
It is left as an exercise for the reader to determine and
then plot the optimal trajectory that will transfer the
1 0
system from to . How would the
1 0
response look in the time domain?