seminar_presentation
TRANSCRIPT
In the name of the
most merciful
Seminar on:
Pole Placement and Robust
Adjustment of Power Systems
Stabilizers(PSSs) through Linear
Matrix Inequalities(LMIs)
Presented by: Mahsa Rezaei (909238)
Supervised by: Dr. Maryam Dehghani 2
The damping of inter-area oscillations is an
important problem in electric power systems.
Inter-area oscillations provide restrictions on the
transferred power.
To enhance damping, power system stabilizers
(PSSs) have been adopted.
PSSs provide phase compensation as
supplementary signals into the excitation
systems of the generators.
3
How to control Power systems?
Generally power Systems are modeled as large scale
systems composed of a set of small interconnected
subsystems.
Complication of incorporation of many feedback loops
into the controller design for large scale interconnected
systems is a motivation for the development of
decentralized control.
On the other hand, Varying operating condition of
power systems and also the nonlinear nature of them,
limits the applications of conventional PSSs.
There is a need for controllers which are robust to
changes in the system operating conditions.
4
This work presents the application of LMIs to the optimal
adjustment of PSSs with pre-defined structure, much common in
the power industry.
Results of some tests show that gain and zeros adjustments are
sufficient to guarantee Power Systems robust stability and
performance with respect to various points of operation.
Making use of the flexible structure of LMIβs, the algorithm permits
one to choose in the compromise between low gain and high
damping factor.
The technique used here is the pole placement design together
with gain limitation, which places the poles of the closed loop
system in a specific region of the complex plane.
5
Outline:
Power System Modeling
Power System Stabilizers and the Closed loop
Structure
Pole placement through linear matrix
inequalities
Experimental results and discussion
More general case
6
Power System Modeling:
The power system electromechanical stability problem can be
represented by a set of differential equations together with a set of
algebraic equations to be solved simultaneously with each other:
π₯ = π( π₯, π§ )0 = π( π₯ , π§ )
(1)
Where x is the state vector and z is a vector of algebraic variables.
Small disturbance stability analysis normally involves the linearization
of (1) around the system operating point.
(2)
7
β indicates a little variation of the variable.
The sub-matrix π¨ππ represents the dynamic description of the
system;
π¨ππ represents the link between the dynamic of the system
and its algebraic variables;
π¨ππ and π¨ππ represent the link between state and algebraic
variables;
π©π and π©π relate the states and the dynamic of the system to
its inputs,
πͺπ and πͺπ relate measurements to state variables. 8
Power System Stabilizer and the Close loop Structure:
The model described in (2) can be transformed in a more
appropriate form:
π₯ = π΄. π₯ + π΅. π’ (3)
π¦ = πΆ. π₯
The PSS structure is fixed, This fixed structure is given by:
(4)
In this scheme, we work with fixed poles and free zeros and gain;
so, we have to obtain the gain (πΎπΊ) and the zeros (given by the
values of π2 and π1) of the controller.
The values chosen for the poles of the PSS are of practical
interest. 9
10
For the decentralized structure, the controllers transfer function
matrix will assume the following form:
(5)
n is the number of machines, and each π²π(π) has the structure
given by (4).
Each PSS will be linked to a machine of the system. The control
structure given by (5) can be transformed in an state space form:
(6)
11
where π΄πΆ, π΅πΆ , πΆπΆ and π·πΆ are matrices that define the structure
of the controller given in (4), such that:
(7)
By using state space representation:
(8)
(9)
12
Applying the control structure (6) to the system described by (3),
we have the following description for the closed loop system:
(10)
By defining the matrices:
(11)
(12)
The modified system, which is equivalent to (10), is given by:
(13)
13
Where π₯π= π₯ π₯πΆ π
and the control law π’π = πΎπΆ . π¦ = πΎπΆ . πΆπ. π₯π
Using the state space description of the system and the
controller fixed matrices π΄πΆ and πΆπΆ , we can calculate
matrices π΄π, π΅π and πΆπ.
Then, we must determine the static output feedback gain
matrix π²πͺ that places the closed loop poles of the system in
a determined region of the complex plane.
14
Pole Placement through Linear Matrix Inequalities:
A. LMI Regions
An LMI region is any subset D of the complex plane that can be
defined as
(14)
where L and M are real matrices such that πΏπ= L.
Important characteristics of LMI regions are:
1. Intersections of LMI regions are LMI regions;
2. A real matrix A is D-stable, that is, has all its eigenvalues in
the LMI region D if and only if a symmetric matrix Q exists
such that:
(15)
15
LMI regions of interest in control applications for pole
placement purposes are the following:
1. Half-plane Re(z) < β πΌ : L = 2.πΌ and M = 1;
2. Disk centered at the origin with radius r:
3. Conic sector with apex at the origin and inner angle 2π:
We will place the poles of the power system in a LMI region that
is the intersection of these three regions. It ensures a minimum
decay rate πΌ, a minimum damping ratio π = cos (π), and a
maximum undamped natural frequency ππ = π. sin π for
the closed loop system.
16
B. Output feedback control and Pole Placement through LMIβs
Applying an output feedback control law π’π = πΎπΆ . π¦ to the
system (13), we get the following closed loop system:
(16)
Our objective is to place the poles of the closed loop system in a
region of the complex plane that is given by the intersection of
the three regions described. To do this, we have to apply (15) to
the closed loop system.
If we substitute the closed loop matrix π΄ππ into equation (15), the
term π΄ππ.Q will not be convex, Because:
(17)
17
To solve this problem and transform (17) into a LMI, we
proceed the following change of variables:
(18)
Substituting (18) into (17), we have:
(19)
Thus, now we have a LMI problem. To recover the controller
gain matrix, we do the following:
(20)
The existence of the inverse in (20) is guaranteed if matrix πΆπ is
full row rank. After getting matrix M, we calculate the gain
matrix πΎπΆ:
(21)
18
Using this change of variables, equation (15) can be applied to
the closed loop system (16) in order to place the poles of this
system in an adequate region of the complex plane:
(22)
Substituting the values of L and M for these three regions yields:
1. Half plane Re(z) < β πΌ:
(23)
2. Disk with radius r:
(24)
3. Conic sector with inner angle π:
(25)
19
where * denotes symmetric term, and:
(26)
20
Summarizing the procedure:
1. Firstly, Define π΄π, π΅π and πΆπ from (11)
and the specifications of performance for the closed loop system
(that is, the values of π, πΌ and π);
2. solve the system of LMIβs given by (23), (24), (25) and Q > 0
in the variables Q and N;
21
3. Calculate matrix M, given by (20);
4. Calculate the static gain matrix πΎπΆ = π.πβ1;
5. Recover the controller variables π·πΆ and π΅πΆ , considering (12);
6. Calculate the transfer function of each decentralized controller,
considering (7), (8) and (9).
22
C. Constraining the Static Gain
As we are working with a static controller formulation, it is
interesting to bound the norm of the static gain matrix πΎπΆ in
order to avoid infeasible values for the controller parameters.
To bound the Euclidean norm of the controller matrix
πΎπΆ = π.πβ1 (with M given by (20)), we do the following:
(27)
In other words, we are restricting at the same time the norms of
matrices N and πβ1 . Rewriting (27), we have:
(28)
C. Constraining the Static Gain
As we are working with a static controller formulation, it is
interesting to bound the norm of the static gain matrix πΎπΆ in
order to avoid infeasible values for the controller parameters.
To bound the Euclidean norm of the controller matrix
πΎπΆ = π.πβ1 (with M given by (20)), we do the following:
(27)
In other words, we are restricting at the same time the norms of
matrices N and πβ1 . Rewriting (27), we have:
(28)
23
The values of π²π΅ and π²πΈ are project parameters and have
to be chosen carefully, in order to avoid large controller
parameter values and, at the same time, not restrict too
much the controller gain matrix.
24
D. The Robust Procedure
It is necessary to ensure that the power system will present good
performance even in case of variations in the operating point. To
deal with this problem, we will make use of the so called
polytopic models.
let the i-th power system model linearized around a specific
operating point be denoted by the triple ( π΄π , π΅π , πΆπ ); considering
the model used here, matrices B and C do not vary with changes
in the operating points; so, the triple that defines the system will
be simply (π΄π , B , C). A polytope is the set Ξ© defined below:
(29)
where n is the dimension of matrices π΄π and m is the number of
operating points. The matrices π΄π are called vertices of the
polytope.
25
To ensure that the poles of any closed loop system associated to a
matrix π΄ β Ξ© will lie in the region of the complex plane described
previously, we have to solve m LMIβs jointly in the same variables
Q and N, that is:
(30)
26
for i = 1, 2, β¦ , m , where m is the number of operating points
for the power system, and:
(31)
(32)
π¨π, i = 1, 2, β¦ , m are the state transition matrices that define the
power system model, and these matrices represent nonlinear
models linearized around one specific operating point.
The procedure for recovering the controller gain matrix is the
same described before.
27
Experimental Results and Discussion: The studied case is the New England system, that consists of nine
synchronous machines with PSSs and one machine (number 39)
modeled as an infinite bus.
28
where ππ composed by variables ππ, ππΉ , ππ¬πβ² and ππ½ππ«.
πz is the algebraic variables vector, composed by variables
ππ°π , ππ°π, ππ½π and ππ½π. πy is the output vector (or measurements vector), composed by
measurement of variable ππ.
π is the synchronous machine rotor velocity, π½π is the voltage
magnitude at the real axis, π½π is the voltage magnitude at the
imaginary axis, π½ππ« is the field voltage, π°πand π°π are the currents
at the machine coordinates, πΉ is the synchronous machine load
angle, π¬πβ² is the voltage magnitude of the synchronous machine
during the transitory state and β indicates a little variation of the
considered variable. βu is the input vector (or control vector), given by variable
βπ½πΉπ¬π (it represents a reference input voltage, that is present in
the automatic voltage regulator model).
29
Eight Operating Conditions that were used in this work:
30
The eigenvalues of the power system before control are presented in
figure 1. As can be seen, there are many unstable eigenvalues, as
well as low damped eigenvalues.
31
Considering the operating conditions given and the following
performance specifications:
32
33
34
35
Now Let us consider a more general case,
We state a Robust PSS design method in which by the
help of pre-defined structure of controller and the
three interesting regions for poles to be placed on,
could change to a LMI problem.
But What if we want to design a Robust free structure
controller whose poles are placed on a desired region?
We will investigate this problem through another paper
and try to solve it with ILMI.
36
The design of a PSS can be cast as a static output control synthesis
using a system augmentation technique. Thus, we deal only with
static output feedback problems. First, we discuss the design of a
static output controller that locates its closed-loop poles within a
specified region. We then describe its extension to simultaneous
stabilization of a finite collection of LTI plants.
Let us consider the LTI system:
π₯ = π΄. π₯ + π΅. π’ , π¦ = πΆ. π₯ (33)
By pole-placement using static output feedback, we mean to address
the problem of searching for a static control law π’ = πΎπ¦ such that
all the eigenvalues of the closed-loop matrix π΄ + π΅πΎπΆ are placed in
a specified complex-plane described by:
(34)
37
Lemma 1:
For the LTI plant (33), we can find a static control π’ = πΎπ¦ that
places all its closed-loop poles in region (34) if and only if there
exist symmetric matrices W β π (π+π)Γ(π+π) , Wβ₯ 0 and P βπ πΓπ,
P β» 0 that satisfy the following LMI and rank condition,
(35)
(36)
If W satisfying (35) and (36) is obtained, K can be computed as
(37)
Assuming
(38)
38
Proof of lemma 1: Let π΄ β πΆπΓπ be a given complex matrix and let
π· = π β πΆ βΆ 1π
β π ππβ π
1π < 0
denote a given open region of the complex plane, where Hermitian matrix
π ππβ π β πΆ2Γ2
has one strictly negative eigenvalue and one strictly positive eigenvalue,
and the star denotes transpose conjugate.
Theorem 1 (Lyapunovβs Inequality): Matrix A has all its eigenvalues
in region D if and only if there is a matrix π = πβ > 0 βπΆπΓπ such that
πΌπ΄
β ππ πππβπ ππ
πΌπ΄
< 0 (39)
Matrix inequality is referred to as Lyapunovβs inequality.
39
What we almost always work with from Lyapunovβs inequality is
in the case that D is the open left half-plane, i.e., p= r = 0, q = 1
and inequality (34) becomes π΄ππ + ππ΄ < 0.
Sometimes this synthesis is called D-stabilization. The following
lemma states a necessary and sufficient condition for the
existence of a pole-placement static output feedback controller.
40
According to Lyapunov stability theory , the existence of a pole-placement
static output feedback controller is equivalent to the existence of a positive
definite matrix P β π πΓπ such that the following inequality holds
(40)
for all x and u satisfying the static control law
(41)
where R β π πΓπ is any nonsingular matrix. Therefore,
(42)
where Ο > 0 is a scalar. We thus obtain (35), (36) and (37) by letting
(43)
41
Now consider the simultaneous stabilization of the system
described by a set of state-space equations
(44)
Each triplet ( π΄π , π΅π , πΆπ ), i = 1, . . . ,N represents an LTI system
linearized at a certain equilibrium point. In the same manner, we
seek a single control law π’ = πΎ. π¦ such that all
matrices π΄π+π΅π . πΎ. πΆπ, i = 1, . . . ,N have their eigenvalues in a
prescribed region.
We can easily extend Lemma 1 to simultaneous stabilization with
closed-loop pole constraints described by the intersection of
multiple regional constraints.
42
Lemma 2:
For the finite set of plants (44), we can find a single control law
π’ = πΎ. π¦ that simultaneously stabilizes all N plants with all the
closed-loop poles located in the intersection of
regions π·(ππ , ππ , ππ), j = 1, . . . ,L if and only if there exist
symmetric matrices W βπ (π+π)Γ(π+π) , W > 0 and πππ β
π πΓπ , πππβ» 0 that satisfy the LMIs for i = 1, . . . ,N, j = 1, . . .
,L.
(45)
and the rank condition
ππππ π = π. (46)
If W satisfying (45) and (46) is obtained, K can be computed
by:
43
Proof of lemma 2:
Let us assume L = 1 without loss of generality. Sufficiency is trivial by
Lemma 1.
To prove necessity, suppose that a pole placement static controller K is
given for all N plants. Then, by setting R = I in (43), there exist ππβ» 0
and ππ> 0 such that the following LMIs are feasible for i = 1, . . . ,N
(47)
Where
Since (47) holds for all Ο β₯ ππ β max{ππ}, a common W for (47) can be
taken as
This completes the proof.
44
Next, We want to state a numerical method for solving Rank-constrained
LMI problems like the one we encounter here.
Penalty Function Method For Rank-constrained LMI Problems:
With a slight abuse of notation, the problem to be solved has the
following generic form:
(48)
Where π is the convex set
(49)
x is the decision vector, and W(x) β π πΓπ and R(x) β π πΓπ are matrices
that are affine functions of x. Also, r is assumed to be less than n.
The core of the PFM is to convert the problem (48) into an ordinary LMI
optimization problem by representing the difficult rank condition as a
penalty function to be minimized.
45
The rank condition is satisfied if and only if the π β π eigenvalues of
W are zero and that the partial sum of the eigenvalues of a
symmetric matrix is bounded above by a weighted trace of the
matrix.
Let the eigenvalues of W be π1β€ Β· Β· Β· β€ππ. The sum of the π β π
smallest eigenvalues of W is then bounded above:
(50)
where V β π πΓ(πβπ) is an arbitrary matrix such that πππ = πΌπβπ.
Thus, we define the penalty function as
(51)
further we introduce the following penalized objective function
(52)
where ΞΌ is the positive penalty parameter. Notice that the trace of W
plays an important role in minimizing the value of p(x;V) by placing
the relative weights on the eigenvalues of W since π‘π π = πππ1 .
46
We can now obtain a solution to problem (48) by sequentially
solving the following convex LMI optimization problem for k = 0,
1, . . .
(53)
where ππ β₯ ππβ1and ππ is constructed from the eigenvectors of
W(π₯π) since the eigenvectors of W can be taken to be orthonormal
to each other.
47
References:
[1] V. A. F. de Campos, J. J. da Silva; L. C. Zanetta Jr., "Pole
Placement and Robust Adjustment of Power System Stabilizers
through Linear Matrix Inequalities", Proceedings of the 2006 IEEE
Power Systems Conference and Exposition, Atlanta, 2006.
[2] F. E. Scavoni, A. S. e Silva, A. Trofino Neto and J. M. Campagnolo,
"Design of Robust Power System Controllers using Linear Matrix
Inequalities", in Proc. 2001 IEEE Porto Power Tech Conference.
[3] M. Chilali, P. Gahinet and P. Apkarian, "Robust Pole Placement in
LMI Regions", IEEE Trans. Automatic Control, vol. 44, no. 12, pp.
2257- 2269, Dec. 1999.
[4] M. Chilali and P. Gahinet, "Hβ Design with Pole Placement
Constraints: An LMI approach", IEEE Trans. Automatic Control, vol.
41, no. 3, pp. 358-367, Mar. 1996.
[5] L. C. Zanetta Jr. and J. J. da Cruz, "An Incremental approach to
Power Systems Stabilizers tuning using mathematical programming",
IEEE Trans. Power Systems, vol. 20, no. 2, pp. 895-902, May 2005.
48
[6] Kim S., Kwon S., Moon Y., Low-order robust power system
stabilizer for single-machine systems: an LMI approach. Int. J.
Elect. Power Energy Syst., 8 (2010) No. 3, 556-563.
[7] N. Martins and L. T. G. Lima, "Determination of suitable locations
for Power System Stabilizers and static var compensators for damping
electromechanical oscillations in large scale Power Systems", IEEE
Trans. Power Systems, vol. 5, no. 4, pp. 1455-1469, Nov. 1990.
[8] J. M. Maciejowski, Multivariable Feedback Design, Addison-
Wesley Publishing Company, 1989.
[9] D. Henrion and G. Meinsma, βRank-one LMIs and Lyapunovβs
inequality,β IEEE Trans. Autom. Control, vol. 46, no. 8, pp. 1285β
1288, 2001.
49
Thank You