Distributed Integral Action: Stability Analysis and

Frequency Control of Power Systems

Martin Andreasson , Henrik Sandberg , Dimos V. Dimarogonas and Karl H. Johansson
ACCESS Linnaeus Center, KTH Royal Institute of Technology, Stockholm, Sweden.

AbstractThis paper analyzes distributed proportional- not only for power systems, but also for other distributed or
integral controllers. We prove that integral action can be decentralized dynamical systems with first- or second-order
successfully applied to consensus algorithms, where attenuation dynamics. We give a short review of the consensus problem
of static disturbances is achieved. These control algorithms are
applied to decentralized frequency control of electrical power to position our theoretical results with respect to the literature.
systems. We show that the proposed algorithm can attenuate The coordination of autonomous agents based solely on local
step disturbances of power loads. We provide simulations of the interactions and decentralized control algorithms [7], has
proposed control algorithm on the IEEE 30 bus test system that applications in multi-vehicle control [8], formation-control
demonstrate its efficiency. [9], flocking [10], rendezvous [11] and distributed estimation
I. I NTRODUCTION [12], [13]. Consensus with integral action has been studied
in [14] for agents with single integrator dynamics. It was
Distributed control is in many large-scale systems the
shown that the proposed consensus protocol can attenuate
only feasible control strategy. To attenuate disturbances,
constant and some time-varying disturbances to a certain
proportional-integral (PI) controllers are often employed in
degree. The main theoretical contributions of this paper is
such plants. However, it is still an open problem when
the extension of consensus for agents with single integrator
distributed controllers can stabilize a plant in general [1].
dynamics under integral action, and consensus for agents with
One important and increasingly interesting class of sys-
damped double integrator dynamics under integral action. We
tems that require integral action to attenuate disturbances
study consensus algorithms both with and without absolute
are electric power systems. Due to disturbances in form
position measurements.
of load and generation changes, a proportional frequency
The rest of this paper is organized as follows. In section II we
controller is not able to reach the desired reference frequency
introduce the mathematical notation. In section III we analyze
asymptotically. To eliminate static errors, integrators are
consensus protocols with PI controllers for agents with single
employed. Due to the inherent difficulties with distributed
integrator dynamics. In section IV we propose two consensus
PI control, automatic frequency control of power systems
protocols for agents with double integrator dynamics, one
is typically carried out at two levels. In the inner control
without and one with absolute position measurements. We
loop the frequency is controlled with a proportional controller
also provide av example to support our theoretical results.
against a dynamic reference frequency. At the outer loop, the
In section V, based on the developed theory, we propose
reference frequency is controlled with a PI controller to elim-
a decentralized control algorithm for frequency control of
inate static errors [2]. While this control architecture often
electrical power systems, and compare the performance to
works satisfactorily today, future power system developments
traditional control algorithms. The paper ends by some con-
might render it unsuitable. For instance, large scale pene-
cluding remarks in section VI.
tration of renewable power generation increases generation
fluctuations, creating a need for fast and local disturbance II. P RELIMINARIES
attenuation. Decentralized control of power systems might
Let G be a graph. Denote by V = {1, . . . , n} the vertex
also provide efficient anti-islanding control and self-healing
set of G, and by E = {1, . . . , m} the edge set of G. Let
features, even when communication is unavailable [3], [4].
Ni be the set of neighboring nodes to i. Denote by B =
In this paper we propose a fully decentralized frequency
B(G) the vertex-edge adjacency matrix of G, and let L be the
control algorithm for electrical power systems. We model the
Laplacian matrix of G. In this paper we will only consider
power system by the swing equation, used in transient stabil-
static, undirected and connected graphs. For the application
ity [5] and fault detection [6]. We apply and further develop
of frequency control of power systems, this is a reasonable
the theory of distributed consensus to prove the stability of
assumption as long as there are no power line failures. For
the proposed algorithms. These results could be of interest
undirected graphs it holds that L = BB T . Let C denote
This work was supported in part by the European Commission by the the open left half complex plane, and C its closure. We
Hycon2 project, the Swedish Research Council (VR) and the Knut and Alice will denote the position of agent i as xi , and its velocity as
Wallenberg Foundation. The 3rd author is also affiliated with the Centre for
Autonomous Systems at KTH and is supported by the VR 2009-3948 grant. vi , and collect them into column vectors x = (x1 , . . . , xn )T ,
Corresponding author. E-mail: mandreas@kth.se v = (v1 , . . . , vn )T . We denote by cnm a vector or matrix
of dimension n m whose elements are all c. In denotes the negative real part we obtain
identity matrix of dimension n.
lim eAt = lim P eJt P 1
t
t
1 t 01(2n2)
III. C ONSENSUS BY DISTRIBUTED INTEGRAL ACTION 1
= P lim
0 1 01(2n2)
P
FOR SINGLE INTEGRATOR DYNAMICS WITH DAMPING t 0
0(2n2)1 0(2n2)1 eJ t

Consider agents with dynamics: 1 t 01(2n2)
= lim P 0 1 01(2n2) P 1
t
zi = xi (1) 0(2n2)1 0(2n2)1 0(2n2)(2n2)
 
X 1 1nn t1nn
(zi zj )+(xi xj ) (xi xi (0))+di (2)


xi = = lim
t n 0nn 1nn
jNi
Given an initial P position x(0) = x0 , we obtain
where R+ , R+ , R+ are fixed parameters, and limt xi (t) = n1 iV x0,i i V, i.e., the agents
di R is an unknown disturbance. converge to the average of their initial positions.
We now consider the case where = 0 and di 6= 0 i V.
Theorem 1. Under the dynamics (1)(1) the agents converge Define the output of the system
to a common value x for any constant disturbance di . If
di = 0 i V, the agents converge to x = n1 iV xi (0)
P   " # 
yz BT 0(n1)n z
for arbitrary vi (0), and for zi (0) = 0 f oralli V. If =
yx 0(n1)n BT x
absolute position measurements are not present, i.e. = 0, | {z }
we still have limt |xi (t)xj (t)| = 0 i, j V for any set ,C
of disturbances di and any , R+ . However the absolute
states are in general unbounded, i.e. limt |xi (t)| = and consider the linear coordinate change
h x = Su, T
i u = S x,
1 n1
i, j V. z = Sw, w = S z, where S = n 1
T S , and S is

Proof: First consider the case where = 0 and di = a matrix such that S is an orthonormal matrix. In the new
0 i V. By introducing z = [z1 , . . . , zn ]T we may rewrite coordinates, the system dynamics (3) become:
(1) (2) as: w = u
" #
  
z 0 In
 
z 0 01(n1)
= nn . (3) u = w
x L L x 0(n1)1 S T LS
| {z } " # " #
1
0 01(n1) 1
,A
+ u + n 1(n) d.
0(n1)1 S T LS ST
By elementary column operations we note that the character-
istic equation of A is given by 0 = det (s + )L + s2 In .

We note that the states u1 and w1 are both unobservable
By comparing the characteristic polynomial with the char- and uncontrollable. We thus omit these states to obtain a
minimal realization by defining the new
T coordinates u =
0
acteristic equation of L: 0 = det (L In ), with solutions T
u2 , . . . , un and w = w2 , . . . , wn , thus obtaining the
0

= i 0, we obtain the equation 0 = s2 + i s + i ..
This equation has two solutions s = 0 if i = 0, and system dynamics
solutions s C if i > 0. Since the above equation  0  " (n1)(n1) #  " #
has exactly two solutions for every i , it follows that the w 0 I(n1) w0 ST d
= + .
algebraic multiplicity of the eigenvalue 0 must be two. It is u0 S T LS S T LS u0 0(n1)1
well-known that for connected graphs G, 1 is the only zero-
eigenvalue of the Laplacian By Lemma 10 in [14], S T LS is invertible, so we may define
 L. By straightforward
T calcula-
tions we obtain that e1 = 11n 01n is an eigenvector
1  00   0  " 1 T #
1 T
T w w (S LS) S d
and e21 = 01n 11n is a generalized eigenvector of A = 0 .

u00 u 0(n1)1
corresponding to the eigenvalue 0. It can also
 be verified
that v1 = n1 11n 01n and v2 = n1 01n 11n

It is easily verified that the origin is the only equilibrium of
are a generalized left eigenvector and an eigenvector of A, the system dynamics, which in the new coordinates are given
respectively, corresponding to the eigenvalue 0, and that by
v1 e11 = 1, v2 e21 = 1 and v2 e11 = 0, v1 e21 = 0. If we let  00  " # 
P be an orthonormal matrix consisting of the normalized w 0(n1)(n1) I(n1) w00
00 = .
eigenvectors of A, we can chose the first columns of P to u T
S LS S LS u00
T

be e11 and e21 , and the first rows of P 1 to be v1 and v2 , | {z }

respectively. Since all other eigenvalues of A have strictly ,A00
By [15], the characteristic polynomial in of A00 is given by IV. C ONSENSUS BY DISTRIBUTED INTEGRAL ACTION FOR
det(2 I(n1) + ( + )S T LS). By comparing this polyno- DOUBLE INTEGRATOR DYNAMICS WITH DAMPING
mial with the characteristic polynomial det sI + B T LB ,


In this section we generalize the results of Section III to
which by lemma 10 in [14] has solutions si < 0, we know double integrator dynamics. Consider agents with dynamics:
that the eigenvalues of A00 must satisfy 2 + si + si = 0,
with solutions C . Thus A00 is Hurwitz. zi = xi (4)
Now consider the case where 6= 0 and di = 0 i V. xi = vi (5)
The dynamics can be written as:
X
vi (xi x0i )+di


vi = (zi zj ) + (xi xj )
jNi
     (6)
z 0 In z
= nn . where x0i = xi (0), R , R , R , R+
+ + +
x L L I x
| {z } and di R is an unknown scalar disturbance. The above
,A protocol does not require communication of the integral state
between the agents, as it suffices for each agent to measure
its neighbors states and integrate the relative differences.
By elementary column operations, the charac-
teristic polynomial of A may be written as Theorem 2. Under the dynamics (4)(6), the agents converge
det (s + )L + (s2 + s)In . By similar to a common value x for any constant disturbance di ,


0 =
arguments used in the previous parts of the proof, provided that < . P If di = 0 i V, the agents
A has a simple eigenvalue 0, converge to x = n1 iV xi (0) for arbitrary vi (0). If
T with the corresponding
eigenvector e 1 = 11n 0 1n and the left eigenvector absolute position measurements are not present, i.e., = 0,
v1 = n1 11n 01n , whereas all other eigenvalues have we still have limt |xi (t) xj (t)| = 0 i, j V for any
 

negative real part. We see that v1 e1 = 1, and hence it set of disturbances di . However the absolute states are in
follows that general unbounded, i.e., limt |xi (t)| = i, j V.
Also, the system is stable if and only if < .
Proof: The proof follows the same principle ideas as the
lim eAt = lim P eJt P 1 proof of Theorem 1. However, as we consider second-order
t
" t
0nn dynamics, the problem is inherently different to first-order
#  
1 01(2n1) 1 1 1nn
0nn dynamics. First consider the case where = 0. We first let
= P lim 0 P =
t 0(2n1)1 eJ t n 0nn
i = 0 i V. Let also di = 0 i V. By introducing the
state vector z = [z1 , . . . , zn ]T we may rewrite (4) (6) in
vector form as:
Given any initial position x(0) = x0 , it immediately follows
that limt x(t) = 0. Now let di 6= 0 and x(0) 6= 0 for at z 0nn In 0nn z
x = 0nn 0nn In x .
least one i V. Defining the output as
v L L In v
| {z }
  " #  ,A
yz BT 0(n1)n z
= , By elementary column operations it is easily shown that
yx 0(n1)n In x
the characteristic polynomial of A can be written as 0 =
det((+s)L+s2 (s+)I), where I is the identity matrix of
| {z }
,C
appropriate dimensions. Comparing the above equation with
the characteristic polynomial of L, we get that 0 = s3 +
the proof is analogue to the proof when di 6= 0 and x(0) = s2 + i s + i , where i is an eigenvalue of L. If i > 0,
0 i V, and is hence omitted. the above equation has all its solutions s C if and only
Finally, if di = 0 i V, the stationarity of x(t)
implies if < , and , , > 0 by the Routh-Hurwitz stability
criterion. Since G by assumption is connected, 1 = 0 and
limt 11nP Lz(t) Lx(t) x(t) + x(0) = 0, i > 0 i = 2, . . . , n. For 1 = 0, the above equation has
so nx = iV xi (0), which concludes the proof.

the solutions s = 0, s = .  By straightforwardcalculations
T
Note 1. Theorem 1 guarantees that the agents converge to  it can be shown that e 1
= 11n 01n 01n and e21 =
T1
a common state, albeit constant disturbances. If at least 01n 11n 01n are an eigenvector and a generalized
one absolute position measurement is available, the agents eigenvector, respectively, corresponding to the eigenvalue 0.
converge to a constant, bounded state. Since the agents Furthermore v1 = 21n 2 11n 01n 11n and v2 =

converge to the average of their initial states in the absence of 1
01n 11n 11n are a generalized left eigenvector
 
n
disturbances, and the system remains stable for any integral and a left eigenvector of A corresponding to he eigenvalue 0.
gain, there are no immediate performance degradations by Furthermore v1 e11 = 1, v2 e21 = 1 and v2 e11 = 0, v1 e21 = 0.
the introduction of integral action. Hence the first columns of P can be chosen as e11 and e21 ,
and the first rows of P 1 can be chosen to be v1 and v2 . that its (generalized) eigenvector will be distinct from e11 and
Since all other eigenvalues of A have strictly negative real e21 . Thus consensus is not reached.
part we obtain We now turn our attention to the case where 6= 0 and
di = 0 i V. The dynamics (4)(6) can then be written in
lim eAt = lim P eJt P 1
t t vector form as

z 0nn In 0nn z
1+t x = 0nn 0nn In x .
1 t1nn 2 1nn
1 nn 1 v L L I In v
= lim 0nn 1nn 1nn
.

t n | {z }
0nn 0nn 0nn ,A

Given any initial position

P x(0) = x10 ,P v(0) = v0 , we obtain By performing elementary column operations the character-
limt xi (t) = n1 iV x0,i + n iV v0,i i V. istic polynomial of A can be written as det(sI) det((s +
Now let us turn our attention to the case where where = 0 )I) det(( + s)L + (s3 + s2 + s)I). The first two factors
and di 6= 0 i V. We define the output of the system as give the solutions s = 0 and s = , respectively. By
comparing the third factor with the characteristic polynomial
yz BT 0(n1)n 0(n1)n z of L, it can be seen that the other eigenvalues s of A satisfy
yx = 0(n1)n BT 0(n1)n x ,

0 = s3 + s2 + ( + i )s + i , where i spec(L)
yv 0(n1)n 0(n1)n BT v Since G by assumption is connected, L has a single simple
eigenvalue 1 = 0, which gives the characteristic equation
| {z }
,C
0 = s(s2 + s + ), with one solution s = 0, and two
and consider the same linear coordinate change of z, x and solutions s C . As all other eigenvalues are strictly
v as applied to z and x in the proof of Theorem 1. In the positive, the solutions s corresponding to strictly positive i
new coordinates the system dynamics are satisfy s C iff i + < i . This is satisfied whenever
T
z 0 = x0 < . It can be verified that e1 = 11n 01n 01n
T
and v1 = n1 11n 11n 11n are a right and left

x0 = v 0
h i eigenvector of A, corresponding to the eigenvalue 0. Fur-
v 0 = 001(n1) 0(n1)1 S T LS z 0 thermore v1 e1 = 1. Since all other eigenvalues have strictly
" # "
1
# negative real part, we have
0 01(n1) 0 0 n 11(n)
(n1)1 x v + d.
0 S T LS ST 1
1nn 1nn 1nn
lim eAt = lim P eJt P 1 = 0nn 0nn 0nn .
We note that the states z10 , x01 and v10 are unobservable t t n
0nn 0nn 0nn
and uncontrollable. We thus omit these states to obtain a
minimal realization by
 0defining0 the new coordinates z 00 = Given any initial position x(0) = x0 , v(0) = v0 , we obtain
 0 0 T
 00 T
z2 , . . . , zn , x = x2 , . . . , xn and v = v2 , . . . , vn0
00
 0 T that limt x(t) = 0. Now when 6= 0 and di 6= 0 for at
we obtain the system dynamics least one i V, defining the output of the system as
T

00
z 0(n1)2 I(n1)2 0(n1)2
00
z S d T y z B 0(n1)n 0 (n1)n z
T
x00 = 00 y = 0 B 0 x

0(n1)2 0(n1)2 I(n1)2 x + 0 . x (n1)n (n1)n

v 00 S T LS S T LS I(n1)2 v 00 0 yv 0nn 0nn In v
| {z } | {z }
,A00 ,C

We now shift the state space by defining the proof is analogous to the proof when = 0 and di 6=
0 i V, and is hence omitted.
1
z (3) z 00 (S T
LS)1 T
S d Finally, if di = 0 i V, the stationarity of v(t)
(3) 00
implies:
x = x 0(n1)1 . 11n Lz Lx x + x(0) v = 0, so

limtP
(3) 00
v v 0(n1)1 nx = iV xi (0), which concludes the proof.
It is easily verified that the origin is the only equilib- Example 1. In this example we will consider the task of
rium of the system dynamics, and that the dynamics in reaching position-consensus for a multi-robot system. We
the new coordinates are also characterized by the matrix consider a distributed PI controller for mobile robots with
A00 . By a similar argument used when showing that A is velocity-dependent damping of the form of (4)(6). Let the
negative semi-definite, we may show that A00 is negative parameters be given by = 5, = 3 and = 0. The setup
semi-definite. But since S T LS is full-rank, A00 must also we will consider consists of a string of 5 mobile robots, whose
be full-rank, and hence A00 is Hurwitz. This implies that communication topology is a string graph. We consider the
limt |xi (t) xj (t)| = 0 i, j V even in the presence system with a constant disturbance d = [1, 0, 0, 0, 0], for
of disturbances di . It is also clear that whenever , = 0, = 1, and = 15. The initial conditions are given
at least one eigenvalue will have non-negative real part, and by x(0) = [5, 6, 8, 4, 5]T , v(0) = [0, 0, 0, 0, 0]T .
 =0 at bus i and ui is the mechanical input. kij = |Vi ||Vj |bij ,
where Vi is the voltage of bus i, and bij is the susceptance
5 of the edge (i, j), here referred to as line. By defining the
state vectors = i , . . . , n and = = 1 . . . , n , we

x(t)

0 may rewrite (7) in state-space form as

" #       
5 0nn In 0n1 0n1
= + + (8)
0 5 10 15 20 M Lk M D M pm Mu
t
where M = diag( m11 , . . . , m1n ), D = diag(d1 , . . . , dn ), Lk
=1 is the weighted Laplacian with edge weights kij , pm =
T T
, u = ui , . . . , un .
 m
p1 , . . . , pm

n
5
A. Centralized PI control
x(t)

0
Traditionally, automatic frequency control of power sys-
5 tems is carried out at two levels [2]. In the inner control
loop the frequency is controlled with a proportional controller
0 5 10 15 20
against a reference frequency. At the outer loop, the reference
t frequency is controlled with a PI controller to eliminate static
= 20 errors. We model the inner controller of bus i as
100
ui = ( i ) (9)
0 and the outer PI controller is assumed to be given by
x(t)

1X
100 z = ref i (10)
n
iV
0 2 4 6 8 10 1X
t = ref i + z (11)
n
iV
Figure 1. The first figure shows the state trajectories of (4)(6) when = 0,
the second figure shows the state trajectories when = 1, the third figure where we have assumed that the average frequency of the
shows the state trajectories when = 20. buses is measured by the central controller. In reality the
frequency is often measured at a specific bus. This will
typically lead to longer delays, since disturbances need to
By Theorem 2 stability is guaranteed if and only if < propagate through the system before control action can be
. In Figure 1 the state trajectories are shown for different taken. The centralized controller architecture is illustrated in
choices of . We observe that asymptotic consensus amongst figure 2.
the mobile robots is only reached when = 1. When =
0, consensus is not reached due to the static disturbance. 1
When = 1, the disturbance is attenuated, and asymptotic + u1
consensus is reached. However, as we increase to 20 = , ref C1 Bus 1

the system becomes unstable, in accordance with Theorem 2.
This is illustrated in the lower right plot of Figure 1.
Cc ..
V. F REQUENCY CONTROL OF POWER SYSTEMS .
n
+ un +
In this section we show that a similar protocol to the Cn Bus n
one proposed in section IV, can be employed for automatic +
frequency control in power systems. Let us consider a power
system modeled by a graph G = (V, E). Each node, here
referred to as a bus, is assumed to obey the linearized swing
equation [2] Figure 2. The figure illustrates the centralized control architecture.
X
mi i + di i = kij (i j ) + pm
i + ui (7)
jNi
Theorem 3. The power system described by (8) where ui
where i is the phase angle of bus i, mi and di are the inertia is given by (9)(11), is stable for > 0, > 0, > 0
and damping coefficient respectively, pm i is the power load satisfying < . Furthermore limt i (t) = ref .
 ref u1 1 , 1 buses. Even the slightest frequency measurement error will
C10 Bus 1 be integrated and prevent the frequencies from reaching con-
sensus [2]. With recent advances in phasor measurement unit
(PMU) technology however, phase measurements are likely
.. to be available to all generator buses [16]. By employing
.
optimal PMU placement, the number of PMU:s needed for
ref un n , n
Cn0 complete observability can be drastically reduced [17]. By
Bus n
integrating (13) we obtain zi = ref t i . This implies that
in order to accurately estimate the integral state zi , each
generator bus needs access only to accurate time and phase
Figure 3. The figure illustrates the decentralized control architecture. measurements, both provided by PMU:s. The decentralized
controller architecture is illustrated in figure 3.
Theorem 4. The power system described by (8) where ui is
Proof: As we are only interested in studying , we may given by (13)(14), is stable for any choice of > 0, > 0.
rewrite (8) with u given by (9)(11), as Furthermore limt i (t) = ref .

0 n 11n n 11n ref Proof: As we are only interested in studying , we may
ref rewrite (8) with ui is given by (13)(14), as
z = 0 0 I z + 1

n1 nn n
1n1 M L M DIn pm       " ref #
z 0 nn In z
= + (15)
pm
| {z }
,A
M L Inn M D M
(12)
| {z }
,A
We now consider the the matrix A0 defined as where the ith element of is i . Let m = mini 1/mi and
d = mini di . We can now write: M D = m dIn + D0 , where

0 n 11n n 11n
A0 , 0n1

0nn In
D0 is a diagonal matrix with non-negative entries. We now
1n1 M L In consider the the matrix A0 defined as
 
0nn In
By performing elementary column operations on A0 , A0 ,
M L Inn m dIn
we may write the characteristic polynomial of A0 as
s + s + s + det M L + (s + ) = 0. The first
3 2


By elementary row and column operations, the eigenvalues
of A0 are given by det (s2 + sm d)In + M L + Inn .


factor has solutions s C if < . Comparing the
second factor with the characteristic polynomial of M L, Comparing this with the characteristic polynomial of M L +
we have that s must satisfy s2 + s + ti = 0, where diag(), we conclude that s must satisfy s2 + sm d + ti = 0,
ti 0 is an eigenvalue of M L. Also this equation has where ti 0 is an eigenvalue of M L + diag(). By [18]
all solutions s C . Thus A must be Hurwitz. Finally, M L + diag() is positive definite, and hence the above
stationarity implies that the equilibrium solution must satisfy equation has all its solutions in C . It follows that also A is
= ref 1n1 . Hurwitz. Now consider the coordinate shift
 0    
z z z
B. Decentralized PI control = o
0 o
In this section we analyze a completely decentralized PI
controller, where each bus controls its own frequency based where z0 = (In + M L)1 (M D ref + pm ), 0 = ref 1n1 .
only on local information. Thus, no frequency measurements In the translated coordinates, 02n1 is the only equilibrium
need to be sent to a central controller, and there is no need to of the system. Hence limt i (t) = ref i V.
send control signals. This architecture might be favorable due C. Simulations
to security concerns when sending unencrypted frequency
The centralized and decentralized frequency control algo-
measurements and control signals over large areas. Another
rithms were tested on the IEEE 30 bus test system [19].The
benefit could be better performance when the tripping of one
line admittances were extracted from [19] and the voltages
or several power lines causes the network to be split up into
were assumed to be 132 kV for all buses. M and D were
two or more sub-networks, so called islanding. The controller
set to reasonable numerical values. The power system is
of node i is assumed to be given by
initially in an operational equilibrium, until the power load
zi = ( ref i ) (13) is increased by a step of 200 kW in the buses 2, 3 and
ui = ( ref
i ) + zi (14) 7. This will immediately result in decreased frequencies
at the extra load buses. The frequency controllers at the
The controller architecture is illustrated in Figure 3. This buses will then control the frequencies towards the desired
decentralized control architecture is not practically feasible frequency of ref = 50 Hz. In the centralized controller the
with only frequency measurements available at the generation parameters were set to = 0.8, = 0, = 0.04, while
 Centralized VI. D ISCUSSION AND C ONCLUSIONS
In this paper we have studied consensus protocols with
50
integral action for agents with double and single integrator
(t) [Hz]

dynamics. We have proved that the proposed consensus

49.95 protocols reach asymptotic consensus even in the presence
of constant disturbances. If we allow for absolute position
measurements, the proposed consensus protocol converges
49.9
0 20 40 60 80 100 asymptotically to a common state. In the absence of dis-
t [s] turbances, the proposed consensus protocols asymptotically
solve the initial average consensus problem. Furthermore
Centralized we have demonstrated that a similar consensus protocol can
be employed for decentralized frequency control of power
60 systems, with reasonable performance and possible benefits
P (t) [kW]

over centralized frequency control.

40
R EFERENCES
20
[1] M. Morari and E. Zafiriou. Robust Process Control. Prentice Hall,
0 Englewood Cliffs, 1989.
0 20 40 60 80 100 [2] J. Machowski, J.W. Bialek, and J.R. Bumby. Power System Dynamics:
Stability and Control. Wiley, 2008.
t [s] [3] N. Senroy, G.T. Heydt, and V. Vittal. Decision tree assisted controlled
islanding. Power Systems, IEEE Transactions on, 21(4):1790 1797,
Figure 4. The upper figure shows the bus frequencies with centralized nov. 2006.
frequency control, while the lower figure shows the control signals at all [4] B. Yang, V. Vittal, and G.T Heydt. Slow-coherency-based controlled
buses. islanding: A demonstration of the approach on the august 14, 2003
blackout scenario. Power Systems, IEEE Transactions on, 21(4):1840
Decentralized 1847, 2006.
[5] F. Doerfler and F. Bullo. on the critical coupling for Kuramoto oscilla-
tors. SIAM Journal on Applied Dynamical Systems, 10(3):10701099,
50 2011.
(t) [Hz]

[6] I Shames, A. M.H. Teixeira, H. Sandberg, and K. H. Johansson.

Distributed fault detection for interconnected second-order systems.
49.95 Automatica, 47(12):2757 2764, 2011.
[7] R. Olfati-Saber and R.M. Murray. Consensus problems in networks of
agents with switching topology and time-delays. IEEE Transactions
on Automatic Control, 49(9):1520 1533, sept. 2004.
49.9 [8] W. Ren and R.W. Beard. Distributed Consensus in Multi-vehicle
0 20 40 60 80 100
Cooperative Control. Springer, 2008.
t [s] [9] J.A. Fax and R.M. Murray. Information flow and cooperative control of
vehicle formations. IEEE Transactions Automatic Control, 49(9):1465
Decentralized 1476, sept. 2004.
[10] H.G. Tanner, A. Jadbabaie, and G.J. Pappas. Stable flocking of mobile
agents, part i: Fixed topology. In IEEE Conference on Decision and
60 Control, volume 2, pages 2010 2015 Vol.2, dec. 2003.
P (t) [kW]

[11] F. Bullo, J. Cortes, and S. Martnez. Distributed Control of Robotic

40 Networks. Applied Mathematics Series. Princeton University Press,
2009.
20 [12] L. Xiao, S. Boyd, and S. Lall. A scheme for robust distributed sensor
fusion based on average consensus. In International Symposium on
0 Information Processing in Sensor Networks, pages 63 70, april 2005.
0 20 40 60 80 100 [13] R. Carli, A. Chiuso, L. Schenato, and S. Zampieri. Distributed Kalman
filtering based on consensus strategies. IEEE Journal on Selected Areas
t [s] in Communications, 26(4):622 633, may 2008.
[14] R.A. Freeman, Peng Yang, and K.M. Lynch. Stability and conver-
Figure 5. The upper figure shows the bus frequencies with decentralized gence properties of dynamic average consensus estimators. In IEEE
frequency control, while the lower figure shows the control signals at all Conference on Decision and Control, pages 338 343, dec. 2006.
buses. [15] J. R. Silvester. Determinants of block matrices. The Mathematical
Gazette, 84(501):pp. 460467, 2000.
[16] A.G. Phadke. Synchronized phasor measurements in power systems.
in the decentralized control architecture the parameters were Computer Applications in Power, IEEE, 6(2):10 15, april 1993.
[17] R.F. Nuqui and A.G. Phadke. Phasor measurement unit placement
= 0.8, = 0.04. The step responses of the frequencies is techniques for complete and incomplete observability. Power Delivery,
plotted in figures 45. We note that if there is a centralized IEEE Transactions on, 20(4):2381 2388, oct. 2005.
PI-controller for the reference frequency, the generation is [18] A. Teixeira. Multi-agent systems with fault and security constraints.
Masters thesis, Royal Institute of Technology (KTH), June 2009.
increased uniformly among the generators. If however the [19] Power systems test case archive-30 bus power flow test case.
integral action is distributed amongst the generators, some available in: http://www.ee.washington.edu/research/pstca/pf30/pg-
generators will increase their generation more than others. tca30bus.htm.