Title: EPRI/NSF Workshop presentation 4/02 Cancun
1Participation Factors, Actuator Placement, and
Some Nonlinear Control Issues in Power Systems
Presentation at EPRI/NSF Workshop on Global
Dynamic Optimization of the Power Grid April
10-12, 2002, Playcar, Mexico
Eyad H. Abed Elec. and Comp. Eng. and Inst.
Systems Res. University of Maryland
2Summary
- In this talk, the notion of participation factors
is revisited, the relation to issues of
controller placement in power networks is
considered, and some related nonlinear control
issues are discussed. Participation factors are
an important element of Selective Modal Analysis
(SMA), a methodology introduced in 1982 by
Perez-Arriaga, Verghese and Schweppe 1-2. SMA
is a very popular tool for system analysis, order
reduction and actuator placement in the electric
power systems area. Related concepts occur in
other engineering disciplines. Participation
factors, a key element of SMA, provide a
mechanism for assessing the level of interaction
between system modes and system state variables. - Following 1-2, participation factors are
considered in two basic senses. In the first
sense, a participation factor measures the
relative contribution of a mode to a state. In
the second, a participation factor measures the
relative contribution of a state to a mode. To
motivate their original definition of
participation factors, the authors of 1-2
assumed a particular initial condition and
calculated the relative contribution of modes to
a state variable or of state variables to a mode
(depending on the sense of participation factor
being considered). It isn't clear at the outset
that these two senses should lead to identical
formulas for participation factors. However, the
precise definitions in 1-2 for these two
senses of participation factors did indeed result
in identical mathematical expressions. The same
conclusion is found to apply in the present work,
under assumptions valid for a large class of
problems. - The new definitions of participation factors
given here (see 3 for a detailed discussion)
involve averaging operations with respect to the
initial state, which means that the initial
condition is taken to be an uncertain vector. The
averaging operation can be a set-theoretic
average or a probabilistic average, which is
tantamount to assuming a probability distribution
for the initial state vector and then taking an
expectation of a measure of the relative
contribution of modes to a state variable or of
state variables to a mode. The more versatile of
these two averaging approaches is found to be the
probabilistic approach.
3- When the probability law of the initial state is
symmetric with respect to the coordinate axes,
the new definition reduces to the original
definition of 1-2. However, the new and
original definitions generally do not coincide
otherwise. - The next issue discussed in the presentation is
the use of participation factors in controller
and sensor placement. The issue of controller
placement is of great significance in electric
power networks, especially with respect to
placement of expensive FACTS devices. It is
important to have methodologies that lead to
optimal, or nearly optimal, placement of
controllers in the sense that a maximum impact
results on the system modes whose damping level
needs to be improved. One idea that appears in
this context is that placement of controllers in
the physical vicinity of state variables with
high participation in a particular mode leads to
maximum damping improvement when using a local
controller which is feeding back local
information. This fact is proved in the talk.
Also, the applicability of an optimal control
technique based on low authority controllers and
convex optimization is discussed. The issue of
combined controller and sensor placement in power
networks and other systems is considered in 4.
That reference also gives a generalized notion of
participation factors, taking into account the
inputs, outputs and state variables along with
the system modes. - The last topic in the presentation deals with
connections with robust nonlinear control for
systems close to their stability boundary and
with detection of such situations in a
nonparametric (i.e., non-model based) way. The
detection of near instability conditions using
signal-based methods is very important since
systems close to their stability boundary are
highly sensitive to modeling errors, and since
crossing the stability boundary can lead to
system disintegration. Using a concept of noisy
precursors, it may be possible to detect
impending instability prior to its onset 5.
Stabilization using well-placed sensors and
controllers can then be triggered to move the
system away from the stability boundary.
4- References
- 1 I.J. Perez-Arriaga, G.C. Verghese and F.C.
Schweppe, Selective modal analysis with
applications to electric power systems, Part I
Heuristic introduction, IEEE Trans. Power
Apparatus and Systems, Vol. 101, 1982, pp.
3117-3125. - 2 G.C. Verghese, I.J. Perez-Arriaga and F.C.
Schweppe, Selective modal analysis with
applications to electric power systems, Part II
The dynamic stability problem, IEEE Trans. Power
Apparatus and Systems, Vol. 101, 1982, pp.
3126-3134. - 3 E.H. Abed, D. Lindsay and W.A. Hashlamoun,
On participation factors for linear systems,
Automatica, Vol. 36, No. 10, October 2000, pp.
1489-1496. - 4 H. Yaghoobi and E.H. Abed, Optimal actuator
and sensor placement for modal and stability
monitoring, Proc. American Control Conference,
June 2-4, 1999, San Diego, pp. 3702-3707. - 5 T. Kim and E.H. Abed, Closed-loop monitoring
systems for detecting impending instability,
IEEE Trans. Circuits and Systems, - I
Fundamental Theory and Applications, Vol. 47, No.
10, October 2000, pp. 1479-1493.
5Participation Factors
- The main subject considered here is (modal)
participation factors an important element of
Selective Modal Analysis (SMA) (Verghese,
Perez-Arriaga and Schweppe, 1982). - SMA is a very popular tool for system analysis,
order reduction and actuator placement in the
electric power systems area. Related concepts
occur in other engineering disciplines. - We will revisit the concept of participation
factors, and consider why it is useful in
actuator placement.
6Basic Background and Original Definition
- Consider a linear time-invariant system
- dx/dt Ax(t),
- where x2 Rn, and A is n n with n distinct
eigenvalues (l1,l2,,ln). - It is often desirable to quantify the
participation of a particular mode (i.e.,
eigenmode) in a state variable. If the states are
physical variables, this lets us study the
influence of system modes on physical components.
7- Tempting to base the association of modes with
state variables on the magnitudes of the entries
in the right eigenvector associated with a mode. - Let (r1,r2,,rn) be right eigenvectors of the
matrix A associated with the eigenvalues
(l1,l2,,ln), respectively. - Using this criterion, one would say that
- the mode associated with li is significantly
involved in the state xk if rik is large.
8- Two main disadvantages of this approach
- (i) It requires a complete spectral analysis of
the system, and is thus computationally
expensive - (ii) The numerical values of the entries of the
eigenvectors depend on the choice of units for
the corresponding state variables. - Problem (ii) is the more serious flaw. It renders
the criterion unreliable in providing a measure
of the contribution of modes to state variables.
This is true even if the variables are similar
physically and are measured in the same units.
9- In SMA, the entries of both the right and left
eigenvectors are utilized to calculate
participation factors that measure the level of
participation of modes in states and the level of
participation of states in modes. - The participation factors defined in SMA are
dimensionless quantities that are independent of
the units in which state variables are measured. - Let (l1,l2,,ln) be left (row) eigenvectors of
the matrix A associated with the eigenvalues
(l1,l2,,ln), respectively.
10- The right and left eigenvectors are taken to
satisfy the normalization li rj dij (Kronecker
delta). - Verghese, Perez-Arriaga and Schweppe define the
participation factor of the i-th mode in the k-th
state xk as the complex number - pki lik rik
- Original logic/motivation for definition
- x(t) eAtx0
- where x0 is the initial condition. This yields
- x(t) å (li x0) exp(lit) ri
11- Now suppose the initial condition x0 is ek, the
unit vector along the k-th coordinate axis. Then
the evolution of the k-th state becomes - xk(t) åi (lik rik) exp(lit)
- This formula indicates that pki can be viewed as
the relative participation of the i-th mode in
the k-th state at t0. - A similar motivation was given for viewing pki as
the relative participation of the k-th state in
the i-th mode at t0.
12New Approach and New Definitions
- The linear system
- dx/dt Ax(t)
- usually represents the small perturbation
dynamics of a nonlinear system near an
equilibrium. - The initial condition for such a perturbation is
usually viewed as being an uncertain vector of
small norm. - We take two approaches to define participation
factors accounting for uncertainty in initial
condition
13Two approaches to handling uncertainty in initial
condition
- Set-valued uncertainty and average relative
participation at time 0 - Probabilistic uncertainty and mean participation
at time 0 - The second approach is more generally applicable.
14Set-valued uncertainty in initial condition
- Take x0 to lie in a connected set S containing
the origin x0 2 S - The case of greatest interest is when SRn.
- Sets S that are symmetric in the sense of the
next definition are of particular significance. - Definition 1. The set S is symmetric if it is
symmetric with respect to each of the hyperplanes
xk0, k1,,n. That is, for any k21,,n and
z(z1,,zk,,zn)2 Rn, z2 S implies that - (z1,,-zk,,zn)2 S.
15- Note that the average contribution at time t0 of
- the i-th mode to state xk vanishes and so is not
useful as a notion of participation factor - avg (li x0) rik 0
- Definition 2. The participation factor for the
mode associated with li in state xk with respect
to a symmetric uncertainty set S is - pki avgx0 2 S (li x0) rik / x0k
- whenever this quantity exists. Here, avgx0 2 S is
an operator that computes the average of a
function over the set S (in the sense of Cauchy
principal value).
16- This quantity measures the average relative
contribution at time t0 of the i-th mode to
state xk. In the definition, the i-th mode is
interpreted as the eli t term in the expansion
for xk. - Also, the denominator on the right side of the
definition is simply the sum of the contributions
from all modes to xk(t) at t0.
17- Next, we evaluate Definition 2 for pki under the
assumption of a symmetric uncertainty set S. - Let Vol(S) sx0 2 S dx0
- denote the volume of the set S. If S has infinite
volume, then the construction below is performed
for a finite symmetric subset, and then a limit
is taken as discussed in Definition 2. From Def.
2, - pki avg (lik x0k) rik/x0k avgx0åj¹ k (lij
x0j) rik/x0k - lik rik sx0 2 S åj¹ k (lij x0j)
rik/x0k dx0 / Vol(S) - lik rik åj¹ k lij rik sx0 2 S x0j/x0k dx0
/ Vol(S) - lik rik
18- The last step follows from the observation that,
because S is symmetric according to Definition 1, - sx0 2 S x0j/x0k dx0 0
- for any j¹ k, where the integral is interpreted
in the sense of Cauchy principal value. - Thus Definition 2 for pki reduces to the
original 1982 Verghese/Perez-Arriaga/Schweppe
definition in the case of a symmetric uncertainty
set S. - Problem For nonsymmetric S, Def. 2 isnt useful.
19Probabilistic uncertainty in initial condition
- Assumption 1. The components x0j, j1,,n,
- of the initial condition vector x0 are
independent random variables with probability
density functions fX0j(x0j). - Definition 3 (Participation of modes in states).
Suppose Assumption 1 holds. Define pki, the
participation at time t0 of the mode li in - state xk, as the expectation
- pki E (li x0) rik / x0k
- whenever this expectation exists.
20- In applications of this definition to specific
problems, it is useful to rewrite the defining
formula as follows - pki E (li x0) rik / x0k
- E åj1n (lij x0j) rik / x0k
- E (lik x0k) rik / x0k E åj¹ k (lij
x0j) rik / x0k - lik rik åj¹ k lij rik E x0j / x0k .
() - Note It is straightforward to verify that this
formula for pki is dimensionless and doesnt
depend on the units chosen for state variables.
21- The following well-known fact from probability
theory will be useful in the examples below. - Lemma 1. (Law of Iterated Expectation) Let X and
Y be random variables and let g(X,Y) be a
function of X and Y. Then - Eg(X,Y)EY EX g(X,Y)Y ,
- where EX and EY emphasize that the inner
expectation is conditioned on Y and taken with
respect to X, and the outer expectation is
unconditional and taken with respect to Y.
22- For the purposes of this paper, the following
observation, based on Lemma 1, is particularly
valuable. - Remark. If X and Y are independent random
variables and the probability density of at least
one of X or Y is symmetric with respect to the
origin, then - E XY 0, and
- E X / Y 0.
23- Example 1. Suppose that the marginal densities
- fX0j(x0j) are symmetric with respect to x0j0,
i.e., that they are even functions of - x0j, for j1,,n. With this assumption, Eq. ()
gives - pki lik rik åj¹ k lij rik E x0j / x0k
- lik rik
- Thus, again the new definition reduces to the
original 1982 Verghese/Perez-Arriaga/Schweppe
definition in the case of symmetric uncertainty,
here embodied in the symmetry and independence of
the marginal densities.
24- Example 2. Suppose that the state variables are
restricted to be nonnegative, and, more
specifically, that the density functions
fX0j(x0j) are Rayleigh - fX0j(x)x/bj exp -x2/2bj
- if xx0j 0, and fX0j(x)0 otherwise.
- The mean of X0j is EX0j p bj / 20.5,
j1,,n. - After some calculations, pki for this example is
found to be - pki lik rik åj¹ k (p/2) (bj/bk)0.5 lij rik
- Thus, this time the new definition does not
reduce to the original 1982 Verghese/Perez-Arriaga
/Schweppe definition --- note that the
uncertainty was nonsymmetric.
25Participation of states in modes
- Denote by V the matrix of right eigenvectors of
A - Vv1 v2 vn
- Because of the normalization of right and left
eigenvectors, we have that V-1 has the lj as its
rows. - Perform the change of variables
- z V-1 x.
- Then z follows the dynamics
- dz/dt (t) V-1AVz(t)
- L z
- where Ldiag l1,, ln
26- The new states zi, i1,,n evolve according to
- zi(t) eli t z0i,
- To define the participation of the original
states xk, k1,,n in the mode zi, z0i is written
in terms of x0 and then an appropriate
expectation is taken - zi(t) eli t li x0
- eli t åj1n (lij x0j).
- This equation, which shows the contribution of
each component of the initial state x0j, j1,,n,
to the i-th mode, motivates the following
definition for the participation factor governing
participation of states in modes.
27- Definition 4 (Participation of states in modes).
Suppose Assumption 1 holds. Define pki, the
participation at time t0 of the state xk in the
mode li, as the expectation - pki E lik x0k / z0i
- whenever this expectation exists.
- After some work, this gives
- pki lik rik åj¹ k lij rjk E z0j / z0i
- Again, in the case of symmetric uncertainty, we
can show that this formula gives the original
notion of participation.
28Controller Placement Implications?
- Consider the linear control system
- dx/dt Ax(t) Bu(t)
- where u(t) is a vector of controls, uk(t),
k1,,nu. - Which of the controls has the highest
effectiveness in moving a given system mode? - Conventional wisdom
- Choose controller(s) near the physical state
variable(s) that participates most in the mode of
interest.
29An Analytical View of Controller Placement
vis-à-vis Participations
- Following Hassibi, How and Boyd (1999), we focus
on the the setting of low-authority controller
(LAC) design the actuators have limited
authority, and hence cannot significantly shift
the eigenvalues of the system. - In this setting, the closed-loop eigenvalues can
be approximated analytically using perturbation
theory for linear operators.
30- A Naïve Calculation
- Suppose BI, nun, and each control uk-gk xk.
Then - dx/dt Ax(t) - G x(t),
- where G diag g1,,gn .
- What is the influence of each gain gk on each li?
- We need the following result from Kato (1982)
Let AA(g) have distinct e-vals depending
smoothly on the gk. Then - li(g) liåk1n (li Ak ri / li ri) gk
O(g2) - where
- Ak A(0) / gk, k1,,n.
31- For this setting, this quickly leads to
- li(g) liåk1n (pki / li ri) gk O(g2)
- li åk1n pkigk / li ri O(g2)
- So the k giving a larger pki ostensibly leads to
a greater effectiveness per unit gain in
providing eigenvalue mobility.
32- Notes
- Low authority controller setting was needed in
this calculation. - This result is related to an interpretation of
participation factors as eigenvalue sensitivities
to diagonal entries of state dynamics matrix (Van
Ness et al.). - More interesting control structures can be
considered input/output participation factors
(PhD thesis of Yaghoobi, Univ. Maryland, 2000)
paper of Boyd et al. noted above seeks sparse
controller structures via optimization.
33- (4) Other participation factor concepts have been
introduced by other groups - Fouad, Vittal et al. introduced a concept of
higher order participation factors - Verghese et al. considered participation factors
for limit cycles of nonlinear systems - Control energy concept etc.
34The Input/Output Selection Problem
- The I/O selection problem concerns making
decisions on the - number,
- placement, and
- type
- of actuators and sensors
- A nice review of techniques on input/output
selection is given by van de Wal and de Jager
(Automatica 2001).
35The Input/Output Selection Problem, Cntd.
- In the electric power field, we need a careful
re-examination of sensor/actuator placement
issues in light of related activities in other
fields. - The computational complexity requires innovative
techniques. - This is a hard problem.
36Border of Stability Issues/Control
- Nonlinear systems are pervasive in engineering
and natural sciences. - Nonlinear systems exhibit amazingly rich behavior
compared to linear systems which cannot even
exhibit robust periodic orbits. Interest in
nonlinear systems has been steadily growing
because of the success in modeling a variety of
phenomena across many disciplines. - Nonlinear systems show changes in steady state
behavior as parameters vary. This is known as
bifurcation.
37- Technically, bifurcation is a change in the
number and/or nature of steady state solutions as
a parameter is slowly changed. - Bifurcations are triggered by a loss of
linearized stability. - Bifurcations are possible in both smooth and
nonsmooth systems. - Chaos is an intriguing dynamical behavior -
seemingly random behavior for a deterministic
system. Bifurcation and chaos are linked since
chaos often arises through a cascading sequence
of bifurcations.
38- Important to remember
- Existence of bifurcation can be checked by linear
analysis around nominal operating condition.
However, - Severity of a bifurcation in terms of
implications for system operation depends
strongly on system nonlinearities.
39Supercritical stationary bifurcation
Bifurcated equilibria
State x
Nominal equilibrium
Bifurcation parameter m
40Subcritical stationary bifurcation
Bifurcated equilibria
State x
Nominal equilibrium
Bifurcation parameter m
41- Notes
- Subcritical bifurcations generally lead to
departure from nominal operation and possibly
hysteresis or chaotic or hunting motions. Also
called dangerous or hard bifurcations. - Supercritical bifurcations are less severe. Also
called safe or soft bifurcations. - Local bifurcation control can entail using
feedback to render supercritical an originally
subcritical bifurcation.
42An Example from Another Field Nonlinear Active
Controlof Jet Engine Stall
- The drive toward lighter jet engines is
motivating studies into operability of the
engine's axial flow compressor close to its
maximum pressure rise. - The compressor is only marginally stable in this
condition, and can stall under small or moderate
flow disturbances.
43Compressor stall and bifurcations
44Compressor Stall Control, Cntd.
- Active control laws were designed for stabilizing
jet engine stall, by applying bifurcation control
to nonlinear models of axial flow compressors. - The controllers employ one dimensional actuation
in the form of throttle feedback.
45Compressor Stall Control, Cntd.
- The project has shown the superiority of
nonlinear bifurcation-theory based methods for
design of simple one-dimensional throttle
controllers. - The controls allow operation up to and past the
stall stability limit ("stall line" in
schematic). - The control designs have been been expanded on
and implemented at university and industry
laboratories.
46Detection of Impending Instability using Noise
Bifurcation Precursors
- Some observations
- Noisy precursors give a robust nonparametric
indicator of impending instability. - Resonant perturbations delay or advance
bifurcations - Supercritical bifurcations delayed
- Subcritical bifurcations advanced
- Chaotic signals containing a resonant frequency
have a similar effect - White noise can have such an effect, but it is
less pronounced