-
1
A Distributed Frequency Regulation Architecture forIslanded
Inertia-Less AC Microgrids
Stanton T. Cady, Student Member, IEEE, Madi Zholbaryssov,
Student Member, IEEE,Alejandro D. Domı́nguez-Garcı́a, Member, IEEE,
and Christoforos N. Hadjicostis, Senior Member, IEEE
Abstract—We address the problem of frequency regulationin
islanded ac microgrids with no inertia, i.e., those
consistingentirely of generators interfaced through power
electronics. Thecontrol architecture we propose to achieve this is
designed todrive the average frequency error to zero while ensuring
thatthe frequency at every bus is equal and that the operatingpoint
that results is stable. We also introduce a
distributedimplementation of the proposed control architecture that
relieson a combination of three distributed algorithms. Two of
thealgorithms, which are well-established consensus-type
algorithms,allow the generators and loads to acquire global
informationneeded for making control decisions; the third
algorithm, whichwe propose herein, enables the generators to obtain
output valuesthat balance the total demand for load without
violating line flowconstraints. Collectively, these algorithms
eliminate the need fora centralized entity with complete knowledge
of the network,its topology, or the capabilities or properties of
the generatorsand loads therein. Moreover, the distributed
implementation wepropose relies on minimal measurements, requiring
only thatthe power injection at each bus be measured. To verify
ourproposed control architecture and the algorithms on which
itsdistributed implementation relies, we analytically show that
theresulting closed-loop system is stable and establish
convergenceof our proposed algorithm. We also illustrate the
features ofthe architecture using numerical simulations of three
test casesapplied to six- and 37-bus networks.
I. INTRODUCTION
Although the nascency of microgrids precludes a strict
def-inition, any collection of interconnected generators and
loadsthat is capable of islanded operation is generally considered
tobe one. It follows, then, that there are no formal restrictionson
the types of generators that may be present in a microgrid;however,
like the prototypical example of a neighborhoodcomprising homes
with rooftop-mounted photovoltaic (PV) ar-rays [1], microgrids are
typically envisaged to consist entirelyof generators that are
interfaced through power electronics.Consequently, without the
spinning mass inherent in traditionalsynchronous generators, this
class of microgrid has little tono effective inertia. Moreover,
irrespective of the presence ofinertia, the power demands of loads
in a microgrid can belarge relative to the output capabilities of
each generator.
Even though the properties mentioned above may notcharacterize
all types of microgrids, e.g., dc microgrids or
S. T. Cady, M. Zholbaryssov, and A. D. Domı́nguez-Garcı́a are
with theECE Department at the University of Illinois at
Urbana-Champaign, Urbana,IL 61801, USA. E-mail: {scady2, zholbar1,
aledan}@ILLINOIS.EDU.
C. N. Hadjicostis is with the ECE Department at the University
ofCyprus, Nicosia, Cyprus, and also with the ECE Department at the
Uni-versity of Illinois at Urbana-Champaign, Urbana, IL 61801, USA.
E-mail:[email protected].
microgrids with traditional generation (see, e.g., [2] and
thereferences therein), in this paper, we restrict considerationto
islanded ac microgrids with purely power electronics-interfaced
generators, i.e., those with no inertia, which wehenceforth refer
to simply as microgrids. Furthermore, whilethese properties, among
others, complicate the problem offrequency regulation, microgrids
are unencumbered by therequirements and well-established concepts
of their largercounterparts (see, e.g., [3]), making them amenable
to newcontrol paradigms. In particular, whereas generation
controlarchitectures for large power systems typically depend on
acentrally located decision-maker, numerous distributed
archi-tectures have been proposed for microgrids (see, e.g., [4],
[5]).Through a combination of computations performed by proces-sors
located at each bus and information exchanged betweenneighboring
processors, these distributed architectures, whichcan meet the same
objectives as their centralized counterparts,obviate the need for
an entity with complete information aboutthe number, type, or
capabilities of the generation units in thesystem. Additionally, by
eliminating the need for a centralizedprocessor and a communication
network connecting it to eachgenerator, these distributed
approaches can achieve highersystem-level efficiency, reliability,
and adaptability.
Regardless of the control paradigm employed, a
frequencyregulation architecture for microgrids must be designed
toensure that stable operation results and that the frequency
atevery bus in the system is equal to the desired reference
value.While numerous control schemes for microgrids have
beenproposed in the literature, designing a frequency
regulationarchitecture that achieves the aforementioned properties
hasreceived limited attention thus far. In fact, to the
authors’knowledge, no previous work has presented a scheme
forfrequency control in inertia-less ac microgrids for
whichstability and system-wide operation at the same frequency
areguaranteed. Next, we discuss some of the work that has
beenpresented in the literature on control for microgrids
includingfrequency regulation.
The authors in [6]–[10] propose centralized approaches
tomicrogrid control, with overall control schemes presented in[9]
and strategies for optimizing operation provided in [7].While
[8]–[10] propose control architectures for frequencyregulation of
microgrids with power electronics-interfacedgenerators, none of
these references provides analytical resultsguaranteeing both
stability and system-wide operation at thedesired frequency.
Distributed approaches to frequency controlhave been proposed in
[5], [11]–[15], as well as our ownprior work in [16]. The authors
in [11], [12], [14] proposed
-
2
control schemes for frequency regulation that they
analyticallyshowed to be stable; however, [11] relies on a
linearizednetwork model, [12] can only guarantee stability for
certaininitial conditions, and [14] does not guarantee that every
busoscillates at the same frequency. In our previous work in
[16],we proposed a distributed approach to frequency regulationin
microgrids, including those with generation resources in-terfaced
through power electronics. Although we were ableto experimentally
validate the proposed architecture, we onlydid so using a system
with synchronous generators, andcould not analytically guarantee
that our controller resultedin closed-loop stability or that the
frequency at every buswould be equal. Finally, the authors in [5],
[15] proposed acontrol architecture that they analytically showed
to result insystem-wide operation at the desired frequency, but
could onlyguarantee local stability.
Despite the lack of existing control architectures designedto
ensure stable operation at a common reference frequency,the authors
in [17] provide a condition for a broad classof coupled
oscillators, which can model the generators andloads in an
inertia-less microgrid, that is sufficient for meetingthe so-called
phase-cohesiveness requirement. In the contextof power flows in a
microgrid, satisfying this condition isequivalent to ensuring that
the angle difference between everypair of connected buses in steady
state will be strictly less thanπ2 . Moreover, if the condition for
phase cohesiveness is met,the natural frequencies of the
oscillators—equivalent to thepower injected at each bus in a
microgrid—and the couplingbetween them—equivalent to the
admittances of the branchesinterconnecting buses in a microgrid—are
such that the systemexhibits coherent behavior, i.e., the angle of
every oscillatorevolves at the same rate. Thus, when combined with
a schemefor regulating the average frequency error, ensuring
phasecohesiveness is sufficient for ensuring stability and
system-wide operation at a common frequency. While we providedsome
details for a control architecture that makes use of
thephase-cohesiveness condition in our earlier work in [18], wedid
not specifically address the problem of choosing feedbackgains that
ensure stability of the closed-loop system, nor didwe have a proof
for one of the main algorithms on which theoperation of the
architecture relies; we address both of theseand other issues in
this paper.
In the discussion that follows, we provide an overview of
thedistributed architecture we propose for frequency regulation
inislanded inertia-less ac microgrid and outline the
contributionspresented herein. As in [18], our proposed control
architectureis designed to take advantage of the results in [17] by
trackinggenerator set-points that, for some initial load demand,
areknown to result in phase-cohesive operation. Following one
ormore small perturbations to the power demanded by the loads,the
architecture iteratively adjusts the generator set-points todrive
the average frequency to some reference value while alsoensuring
that the operating point that results is phase cohesive.By
regulating the average frequency around an operating pointthat is
known to be phase cohesive, our controller ensuressmall-signal
stability of the closed-loop system while alsoguaranteeing that the
frequency at every bus is the same. Tohandle larger load
perturbations, we also provide a method
for triggering the recomputation of the generator set-points
tobe tracked based upon an estimate for the amount by whichthe
system has deviated from the original phase-cohesiveoperating
point. Additionally, we propose and provide a proofof convergence
for a distributed algorithm that enables thecomputation of
generator output values that collectively meetthe total load power
demand without violating generationor line flow limits. Using this
distributed algorithm andtwo consensus-type algorithms, we also
outline a distributedimplementation of our proposed control
architecture. Thesethree algorithms enable the acquisition of
global informationwith which processors located at each bus can
make decisionsto collectively achieve the system-level objectives
of ourproposed control architecture. Finally, we provide
analyticalcriteria for choosing gains that result in closed-loop
stabilityand demonstrate the operation of our proposed architecture
andits distributed implementation using numerical simulations.
II. PRELIMINARIES
We begin this section by outlining a model to represent
thephysical network of a microgrid. The model uses notions
fromgraph theory to represent the interconnections between busesin
the system as well as differential equations to represent
thedynamic behavior of the generators and loads that compriseit.
Next, we introduce another graph-theoretic model to rep-resent the
communication network on which the distributedimplementation of our
proposed control architecture relies.Finally, we formally define
the notion of phase cohesivenessand outline a sufficient condition
for achieving it; we thenpose the criterion for achieving phase
cohesiveness and othercontrol objectives as a feasibility
problem.
A. Physical Layer Model
We consider ac microgrids comprising loads and generatorsthat
can be represented by a first-order dynamical model; wefurther
restrict consideration to systems in which all gener-ation units
are interfaced through power electronics, i.e., weconsider systems
comprising generators with no inertia; non-dispatchable generators
are considered as loads injecting nega-tive power. Although
electrical lines in typical microgrids havenon-negligible losses,
we make the reasonable assumption thatthe resistance-to-reactance
ratio (commonly referred to as theR/X ratio) is homogeneous across
all lines, which allows usto use a linear transformation in [15] to
recover a losslessmodel. To facilitate the analytical discussion
presented herein,we assume that the generators have no internal
impedance,the power network is lossless, and the voltage magnitude
atevery bus is constant and unity. Although these assumptionsmay
lead to modeling inaccuracies under extreme loadingscenarios, the
control architecture we propose in Section III isdesigned for
operation under nominal conditions where suchmodels justifiably
approximate the behavior of generators andloads (see, e.g.,
[17]).
For an n-bus microgrid, let Gp = (Vp, Ep) be an undi-rected
simple graph representing the interconnections betweenbuses. The
vertex—or bus—set, Vp, is defined to be Vp :=
-
3
{1, 2, . . . , n} = V(g)p ∪V(`)p , where V(g)p and V(`)p denote
gener-ator and load bus sets, respectively. Without loss of
generality,we partition the bus set such that V(g)p := {1, 2, . . .
,m},V(`)p := {m + 1,m + 2, . . . , n}, and V(g)p ∩ V(`)p = ∅,i.e.,
each bus has only a generator or load attached, butnot both. The
edge—or branch—set, Ep, is defined to beEp ⊆ {{i, j} : i 6= j, i, j
∈ Vp}, where the edge {i, j} ∈ Epif buses i and j, i 6= j, are
connected electrically. Wedenote the set of buses to which each bus
i is connected byNp(i) := {j ∈ Vp : {i, j} ∈ Ep}, and denote the
number ofsuch buses by δp(i) = |Np(i)|. Finally, we assume that
noislands exist in the microgrid such that the graph Gp consistsof
a single connected component.
If we arbitrarily assign a direction to each edge e = {i, j}
∈Ep, the oriented incidence matrix, denoted by M ∈ Rn×|Ep|,is
defined as M := [mie] where
mie =
−1, if i is the sink node of edge e,1, if i is the source node
of edge e,0, otherwise.
(1)
Furthermore, we define the weighted Laplacian matrix of Gpas L =
Mdiag({Bij : {i, j} ∈ Ep})MT, where diag({Bij :{i, j} ∈ Ep}) is an
|Ep| × |Ep| diagonal matrix consisting ofsusceptances, Bij = Bji,
{i, j} ∈ Ep, between electricallyconnected buses i and j. Given
that the Laplacian is singular(due to its zero eigenvalue), we will
utilize its Moore-Penrosepseudo inverse, denoted by L† (see, e.g.,
[19] for details onhow to compute it).
At time t > 0, let the voltage angle at bus i ∈ Vp be
denotedby θi(t); similarly, let the set-point of generator i ∈
V(g)p bedenoted by ui(t), and let ui and ui denote lower and
upperbounds such that 0 ≤ ui ≤ ui(t) ≤ ui. Additionally, let`i(t)
:= `
0i + ∆`i(t) ≥ 0 denote the real power demand at
load bus i ∈ V(`)p where `0i ≥ 0 is the demand at t = 0,
and∆`i(t) is a load perturbation that occurs at some t > 0.
Foreach inverter we adopt a controllable voltage source behinda
reactance model (see, e.g., [15]), and write the microgriddynamics
as follows:
Didθi(t)
dt= Diω0 + ui(t)−
∑j∈Np(i)
Bij sin(θi(t)− θj(t)),
(2)
for i ∈ V(g)p , and
Didθi(t)
dt=Diω0 − (`0i + ∆`i(t))
−∑
j∈Np(i)
Bij sin(θi(t)− θj(t)), (3)
for i ∈ V(`)p , where Di > 0 is a time constant
associatedwith the dynamics of the generator or load at bus i ∈ Vp,
andω0 denotes the reference frequency. The model in (2) – (3)is
based on the standard decoupling approximation where thefrequency
and voltage control loops are decoupled so that thefrequency is
regulated to adjust the active power assumingconstant voltage
magnitudes, and the voltage magnitude isregulated to adjust the
reactive power [15].
B. Cyber Layer Model
In Section V, we will introduce a distributed implementationof
our proposed frequency regulation architecture that relieson
processors located at each bus. To realize this
distributedimplementation, we assume that each processor is capable
ofperforming low-complexity computations, e.g., addition
andmultiplication, based upon information obtained locally andfrom
the processors of neighboring buses. In this section,we introduce a
second graph-theoretic model to representthe communication network
over which the local processorscan exchange information with other
neighboring processors.While the connectedness of the graph used to
model thephysical layer is sufficient for the algorithms on which
ourdistributed implementation relies, we abstract the cyber
layerfrom the underlying physical layer to enable a more
generalcommunication modality.
To allow for the possibility of unidirectional transfers
ofinformation between neighboring processors, we represent thecyber
layer by a directed graph, which we denote by Gc =(Vc, Ec). The
vertex—or node—set, Vc, consists of processors,one for each bus in
the physical layer, i.e., Vc := {1, 2, . . . , n}.Although Vc and
Vp are defined identically, the physical andcyber layer vertex sets
need not be equal; however, since eachload and generator is
outfitted with a processor, there must bea one-to-one
correspondence between the vertex sets of thetwo layers, i.e.,
there must exist a bijection hc : Vp 7→ Vc.The communication graph
edge set, Ec, is defined to be Ec ⊆Vc × Vc, where the ordered pair
(i, j) ∈ Ec if node i canreceive information from node j.
We refer to the set of nodes from which node i can
receiveinformation as its in-neighborhood and denote it by N−c (i)
:={j ∈ Vc : (i, j) ∈ Ec}. Similarly, we refer to the set of nodesto
which node i can send information as its out-neighborhoodand denote
it by N+c (i) := {j ∈ Vc : (j, i) ∈ Ec}. Thenumber of nodes that
can receive information from node i,which we refer to as its
out-degree, is the cardinality of theout-neighborhood and we denote
it by δ+c (i) := |N+c (i)|.
For a pair of nodes i, j ∈ Vc, and for ν0, ν1, . . . , ντ ∈
Vcand e0, e1, . . . , eτ−1 ∈ Ec, we refer to the alternating
sequenceof nodes and edges i ≡ ν0, e0, ν1, e1, ν2, . . . , ντ−1,
eτ−1, ντ ≡j such that ei = (νl+1, νi) for l = 0, 1, . . . , τ−1 as
a directedpath of length τ between nodes i and j. Furthermore,
werefer to the minimum distance from i to j, i 6= j, as
theshortest-length directed path between the nodes, and denoteits
value by dc(i, j), with dc(i, j) = ∞ if no path exists.The diameter
of Gc, which we denote by ∆c, is defined tobe the longest shortest
path between any two nodes, i.e.,∆c := maxi,j∈Vc, i 6=j dc(i, j).
Finally, for the algorithmsintroduced in Section IV that rely on
Gc, we require thatthe graph modeling the communication network be
stronglyconnected, i.e., we require that the diameter of Gc be
finite.
C. Phase Cohesiveness Criterion
The control architecture we will propose in Section III
isdesigned to adjust the output of the generators in a microgridto
track pre-determined set-points while simultaneously elimi-nating
the frequency error that results from small perturbations
-
4
in load. These pre-determined set-points are chosen such
that,when applied, the system is stable around the resultant
operat-ing point, the average frequency equals some reference
value,and the frequency at every bus is the same. In the
discussionthat follows, we formalize the properties that
characterize suchan operating point, and introduce a feasibility
problem forchoosing generator set-points, the solution of which
meetsthese properties.
Given constant generator set-points, which we denote by u∗i ,i ∈
V(g)p , and the initial load power demands `0i , i ∈ V
(`)p , as
described in Section II-A, let θ∗ =[θ∗1 , θ
∗2 , . . . , θ
∗n
]Tdenote
a steady-state operating point of the system in (2) – (3).Then,
for some reference frequency, ω0, we would like todetermine the u∗i
’s such that, when applied to the generators,the operating point θ∗
exists and is characterized by thefollowing properties:
P1. the average frequency in the system equals the
reference,i.e.,
∑i∈Vp θ̇i(t)
n = ω0;P2. the frequency at every bus is equal, i.e.,
|θ̇i(t)−θ̇j(t)| = 0
for i, j ∈ Vp; andP3. the system is stable around the operating
point θ∗, i.e.,
−∇θ(t)h(θ(t))∣∣∣∣θ=θ∗
� 0, where θ :=[θ1(t), . . . , θn(t)
]Tand h(θ(t)) = [he(θ(t))] with he(θ(t)) := Bij sin(θi −θj) for
e = {i, j} ∈ Ep.
While we will show in Section III that meeting PropertyP1 can be
achieved by balancing generation and demand,i.e., choosing the u∗i
’s such that
∑i∈V(g)p
u∗i =∑i∈V(`)p
`0i ,ensuring that Properties P2 and P3 are met is more
difficult,especially when trying to do so in a distributed
fashion.However, if generation and demand are balanced and the
so-called phase-cohesiveness condition is met, i.e., |θ∗i − θ∗j | ≤
φfor {i, j} ∈ Ep and φ ∈ [0, π/2) [17], it can be shownthat the
resultant operating point, θ∗, exists and is char-acterized by
Properties P1 – P3. To that end, if we letu∗ =
[u∗1, u
∗2, . . . , u
∗m
]Tand `0 =
[`0m+1, `
0m+2, . . . , `
0n
]T,
then it follows from the results in [17] that, for a broad
classof network topologies, if∥∥∥∥MTL† [ u∗−`0
]∥∥∥∥∞≤ sin(φ), (4)
for some angle φ ∈ [0, π/2), the resulting operating point,
θ∗,exists and is phase cohesive. In subsequent developments,
werefer to such u∗i ’s as phase-cohesive set-points; although
thisterm is a slight abuse of nomenclature, we use it as
short-handto refer to set-points that result in phase-cohesive
operation,subject to the imposed load demands. Despite being
provenfor several topologies, including acyclic networks and
thosecomprising low-dimensional cycles, the condition in (4) isnot
sufficient for general networks. As a result, we
restrictconsideration to network topologies for which (4) holds
(werefer the reader to [17] and supporting information for
details).
Beyond finding phase-cohesive set-points that are balancedwith
demand, we must further restrict the problem to findingthose
set-points that lie within the individual bounds of thegenerators,
i.e., ui ≤ u∗i ≤ ui, i ∈ V
(g)p . Thus, the task
of finding generator set-points that meet all of the above-
described requirements can be summarized by the
followingfeasibility problem:
find u
subject to∑i∈V(g)p
ui =∑i∈V(`)p
`0i ,∥∥∥∥MTL† [ u−l0]∥∥∥∥∞≤ sin(φ),
ui ≤ ui ≤ ui, ∀i ∈ V(g)p .
(5)
III. FREQUENCY REGULATION ARCHITECTUREIn this section, we
propose a control architecture that is
designed to regulate the frequency in an inertia-less ac
micro-grid. We begin by providing an overview of the
architectureand its operation. Then, we formalize the control
scheme andoutline criteria for choosing gains that yield
closed-loop sta-bility. Finally, we describe a method for
determining when torecompute generator set-points to ensure phase
cohesiveness.
A. Overview
From a system-level perspective, the scheme we propose
forfrequency regulation is similar to a discrete-time
proportionalintegral controller. More specifically, after
determining set-points that satisfy (5), which we denote by u∗i , i
∈ V
(g)p , the
set-point of each generator i is incrementally adjusted awayfrom
u∗i over several discrete time intervals. Following oneor more
small perturbations in the power demanded by theloads, these
incremental changes serve to drive the resultingfrequency error in
the system to zero. However, in general,these incremental
adjustments move the system away from anoperating point that is
known to be phase cohesive. To counterthis adverse effect, our
control architecture includes a methodfor determining when to
recompute the u∗i ’s based upon anestimate for the amount by which
the system has deviatedfrom the phase-cohesive operating point.
Without loss of generality, if we transform the microgridmodel
in (2) – (3) to a reference frame rotating at the
referencefrequency, ω0, we see that an operating point at which
thefrequency of every bus equals ω0 is equivalent to one at
whichthe derivative of the voltage angle at each bus is zero,
i.e.,dθi(t)dt = 0, i ∈ Vp. [To keep notation simple, we use the
same
θ in the original and the rotating reference frames.] Let
∆ω(t)denote the average frequency error in this reference
frame,weighted by the time constants associated with the dynamicsof
the generators and loads, i.e.,
∆ω(t) :=
∑ni=1Di
dθi(t)dt∑n
i=1Di, (6)
for t > 0. Replacing the numerator in (6) with a summationof
all of the equations in (2) and (3), it follows that the valueof
the average frequency error at time t > 0 is given by
∆ω(t) =1∑n
i=1Di
∑i∈V(g)p
ui(t)−∑i∈V(`)p
`i(t)
. (7)As mentioned above, our frequency regulation
architecture
incrementally adjusts the generator set-points over several
-
5
discrete time intervals. We refer to the intervals during
whichthe controller is executed as rounds and index them byr = 0,
1, 2, . . . . To simplify notation, we reset the round indexevery
time the set-points to be tracked are computed, i.e., theu∗i ’s are
always determined immediately prior to round r = 0.We denote the
duration of the rounds by T0 and define thetime at the beginning of
each round r to be tr := rT0. Forour controller to eliminate the
frequency error that results fromchanges in load, we assume that a
sufficient number of roundshave elapsed between load perturbations.
As a result, for theanalysis of the control scheme we present next,
we assumethat the power demanded by the loads is constant, i.e., if
thepower demanded by load i ∈ V(`)p is perturbed by ∆`i(t) att =
t0, then `i(t) = `0i + ∆`i(t0) for t0 < t < t0 + r0T0 andr0
sufficiently large.
Let ui[r] := ui(t), tr ≤ t < tr+1, be the set-point
ofgenerator i ∈ V(g)p during round r, which is adjusted at
thebeginning of the round and held constant for the
remainingduration. Additionally, define D := 1∑n
i=1Di; then, given the
assumption that the power demanded by the loads is constant,it
follows from (7) that, for tr ≤ t < tr+1,
∆ω[r] := D
∑i∈V(g)p
ui[r]−∑i∈V(`)p
(`0i + ∆`i)
. (8)
B. Control Scheme
From (8), if we assume that ui[r] = u∗i , for i ∈ V(g)p and
r = 0, 1, 2, . . . i.e., the set-points of the generators are
constantand satisfy (5), it is clear that, following one or more
changesto the power demanded by the loads, the average
frequencyerror at round r is given by
∆ω[r] = D
∑i∈V(g)p
u∗i −∑i∈V(`)p
`0i −∑i∈V(`)p
∆`i
. (9)From (9), we see that, for constant generator set-points,
thevalue of the average frequency error is equal to the
additiveinverse of the sum of the load perturbations, weighted by
thesum of the generator and load time constants, i.e., ∆ω[r]
=−D
∑i∈V(`)p
∆`i. Thus, in order to drive the frequency errorto zero, it is
sufficient to adjust the ui[r]’s such that
limr→∞
∑i∈V(g)p
ui[r] =∑i∈V(`)p
`0i + ∆`i. (10)
Given previously determined phase-cohesive set-points, u∗i ,i ∈
V(g)p , and in order to satisfy (10), our control
architectureadjusts the set-point of generator i at round r
according to
ui[r] = u∗i + ∆ui[r], (11)
where ∆ui[r] denotes the incremental amount by which ui[r]is
adjusted away from u∗i . We define the incremental set-pointto be
∆ui[r] := αiei[r], where the value of ei[r] is updatedrecursively
as
ei[r + 1] = ei[r] + κi∆ω[r], (12)
for appropriately chosen gains αi and κi for i ∈ V(g)p , andwith
ei[0] = 0, i ∈ V(g)p . [Note that, in addition to resettingthe
round index, the value of ei, i ∈ V(g)p must also be resetto zero
when the u∗i ’s are recomputed.] Next, we establishcriteria for
choosing gains that ensure the closed-loop systemis stable and that
∆ω[r]→ 0 as r →∞.
C. Stability Analysis and Criteria for Choosing Gains
If we define α :=[α1, . . . , αm
]T, κ :=
[κ1, . . . , κm
]T, and
e[r] :=[e1[r], . . . , em[r]
]T, then, by stacking (8) with (12) for
i ∈ V(g)p , we can represent the closed-loop system by[∆ω[r +
1]e[r + 1]
]=
[β DαT
κ Im
]︸ ︷︷ ︸
:=Φ
[∆ω[r]e[r]
]+
[−10m
]D∑i∈V(`)p
∆`i,
(13)
where Φ ∈ R(m+1)×(m+1), β := D∑i∈V(g)p
αiκi, and Im and0m represent the m-dimensional identity matrix
and all-zerosvector, respectively.
To ensure that the closed-loop system in (13) is stable,the αi’s
and κi’s specified above must be chosen such thatthe spectral
radius of Φ lies on the boundary of or withinthe unit circle. More
specifically, for marginal stability, wemust choose the αi’s and
κi’s such that ρ(Φ) ≤ 1, i.e.,|λj | < 1 for j = 1, 2, . . . ,m +
1, where λj is the jtheigenvalue of Φ. Given such gains, we can
compute theasymptotic value that
[∆ω[r], e[r]
]Ttakes, which we denote
by[∆ωss, ess
]T. As r →∞, we have that
[∆ω[r], e[r]
]T=[
∆ω[r + 1], e[r + 1]]T
=[∆ωss, ess
]T, and (13) becomes[
1− β −DαT−κ 0m×m
] [∆ωss
ess
]=
[−10m
]D∑i∈V(`)p
∆`i, (14)
where 0m×m is an m×m matrix of all zeros. The
followingproposition, a proof of which is provided in Appendix
A,establishes the criteria for choosing gains such that the
closed-loop system is marginally stable and ∆ωss = 0.
Proposition 1: Consider the system in (13). If αi and κi fori ∈
V(g)p are chosen such that −2 < β = D
∑i∈V(g)p
αiκi < 0,then the system is marginally stable and the average
frequencyerror asymptotically approaches zero, i.e., ρ(Φ) ≤ 1
and∆ω[r]→ 0 as r →∞. Additionally, the value of the
averagefrequency error at any round r is given by
∆ω[r] = (1 + β)r−1(β − 1)D∑i∈V(`)p
∆`i. (15)
D. Estimating Deviation From Phase-Cohesive Operation
Recall the criterion for phase cohesiveness in (4), whichwas
defined for generator set-points and load demands u∗
and `0, respectively. More generally, for any set-points andload
demands, u =
[u1, . . . , um
]Tand ` =
[`m+1, . . . , `n
]T,
respectively, we can define the function
hp(u, `) :=
∥∥∥∥MTL† [ u−`]∥∥∥∥∞, (16)
-
6
where the operating point that results from the injections atthe
generators and loads is phase cohesive if the value that
thefunction hp(·, ·) takes is smaller than sin(φ) for φ ∈ [0,
π/2).
Given set-points found to result in phase-cohesive
operationsubject to the initial power demanded by the loads, u∗ and
`0,respectively, we define ∆u :=
[∆u1, . . . ,∆um
]Tand ∆` :=[
∆`m+1, . . . ,∆`n]T
to be vectors representing the amountby which the generators and
loads deviate from the initial set-points and demands,
respectively. Then, to ensure the systemremains phase cohesive as
it evolves away from the operatingpoint for which the u∗i ’s were
found, we must have h
p(u∗ +∆u, `0 + ∆`) ≤ sin(φ), which implies that hp(u∗ + ∆u, `0
+∆`) ≤ [sin(φ)−hp(u∗, `0)] for φ ∈ [0, π/2). In the discussionthat
follows, we describe a method for estimating the value thathp(·, ·)
takes as our control architecture adjusts the generatorset-points
in response to changes in load.
By inspecting the update rule in (11) and the values onwhich the
incremental set-point adjustment for each generatordepends, we see
that ∆ui[r] ∝
∑j∈V(`)p
∆`j . More specifi-cally, if we substitute the value of ∆ω[r]
from (15) into (12),we see that the value of ei[r] at any round r
for i ∈ V(g)p is
ei[r] = ei[0] + κi
r−1∑s=0
∆ω[s], (17)
= ei[0] + κiD
∑i∈V(`)p
∆`i
(r−1∑s=0
(1 + β)s
), (18)
and, given that ei[0] = 0, i ∈ V(g)p , we have that
theincremental set-point of generator i at round r is given by
∆ui[r] = αiκiD
∑j∈V(`)p
∆`j
(r−1∑s=0
(1 + β)s
). (19)
Thus, since the ∆ui[r]’s depend on the load perturbations, wecan
define a new function,
hl(`0 + ∆`) := hp(u∗ + ∆u(∆`), `0 + ∆`),
=
∥∥∥∥MTL† [u∗ + ∆u(∆`)−(`0 + ∆`)]∥∥∥∥∞, (20)
that takes the value of the load demands as input, and
returnsthe same value hp(·, ·) takes when evaluated at u = u∗ +
∆uand ` = `0 + ∆` for the closed-loop system. If we define
thegradient of hl(·) as
∇hl(`) =[∂hl(`)∂`m+1
, · · · , ∂hl(`)∂`n
], (21)
it follows that, for small perturbations in load around `0,
wecan estimate (20) as
hl(`0 + ∆`) ≈ hl(`0) +∇hl(`)∣∣∣∣`=`0
∆`,
= hl(`0) +∑i∈V(`)p
∂hl(`)
∂`i
∣∣∣∣`=`0
∆`i. (22)
By maintaining an estimate for the value of ∇hl(`)∣∣∣∣`=`0
∆`
as the amount of power demanded by the loads changes, the
point at which phase-cohesive set-points should be recomputedcan
be determined. For example, one strategy could be toinitiate a
recomputation of the u∗i ’s when
ξ(c) :=
∑i∈V(`)p
∂hl(`)
∂`i
∣∣∣∣`=`0
∆`i
c(23)
exceeds unity, where 0 < c <[sin(φ)− hl(`0)
]is a constant
that controls the frequency with which set-points are
computedand the margin of phase cohesiveness.
IV. DISTRIBUTED ALGORITHMSIn this section, we introduce three
distributed algorithms
that will be used as primitives for distributively
implementingthe frequency regulation architecture proposed in
Section III.We begin by providing an overview of two
consensus-typealgorithms, both of which rely on a communication
networkthat is described by the strongly connected directed
graphmodel Gc = (Vc, Ec) introduced in Section II-B. The
firstenables the computation of the maximum of locally
maintainednodal values in finite time; the second enables the nodes
toasymptotically determine a ratio of sums of values knownlocally
by each of the individual nodes; in order to convergeto desired
values, both algorithms require Gc to be stronglyconnected (see,
e.g., [20], [21]). The third algorithm, whichwe refer to as the
feasible flow algorithm, enables the com-putation of generator
outputs and the resultant branch powerflows that collectively meet
the total power demanded by theloads without violating any branch
flow limits; the feasibleflow algorithm requires a communication
network topology tomatch that of the physical layer which also
necessitates havingbidirectional communication links. We index the
iterationsover which all three algorithms update locally
maintainedvalues by k = 0, 1, 2, . . . .
A. Max ConsensusConsider the vector η :=
[η1, η2, . . . , ηn
]T, where the value
of ηi, i ∈ Vc, is known only by node i, and suppose the nodesare
interested in finding the maximum value among the ηi’s,which we
define to be µ := maxi∈Vc ηi = ‖η‖∞. Let µi[k]be an estimate for µ
maintained by node i ∈ Vc at iteration k.Then, as shown in [20], if
the nodes initialize their estimatesas µi[0] = ηi and update them
according to
µi[k + 1] = maxj∈N−c (i)∪{i}
µj [k] (24)
for each iteration k, it follows that, after a finite number
ofiterations bounded from above by the diameter of the graph,every
node can obtain the value of µ, i.e., for km ≥ ∆c,µi[km] = µ for i
∈ Vc.
If, in addition to iteratively updating the value of µi[k],
thenodes also maintain a second value, denoted by νi[k], andupdate
it at each k as
νi[k + 1] = argmax{µj : j∈N−c (i)∪{i}}
µj [k], (25)
it follows that νi[km] = argmax{ηj : j∈Vc}
ηj for i ∈ Vc and km ≥ ∆c,
i.e., every node can obtain the index of the node that has
the
-
7
maximum ηi at the same iteration that µ is acquired. [Notethat
in the case that two or more nodes have the maximizingvalue, e.g.,
ηj1 = ηj2 = µ, the value of νl[km + 1] can betaken to be the
largest index between j1 and j2.]
B. Ratio Consensus
Similar to the max-consensus algorithm described above,nodes
participating in the so-called ratio-consensus algorithmiteratively
update two state variables, which we denote by yi[k]and zi[k] for i
∈ Vc. As shown in [21], by updating the statevariables according
to
yi[k + 1] =∑
j∈N−c (i)∪{i}
1
δ+c (j) + 1yj [k], (26)
zi[k + 1] =∑
j∈N−c (i)∪{i}
1
δ+c (j) + 1zj [k], (27)
the ratio of the state variables, which we define to be γi[k]
:=yi[k]zi[k]
, converges to the same value ∀i ∈ Vc; more specifically,each of
the γi[k]’s asymptotically approaches
γ = limk→∞
γi[k] =
∑nj=1 yj [0]∑nj=1 zj [0]
, i ∈ Vc. (28)
By properly initializing the state variables, as we will showin
Section V, the ratio-consensus algorithm can enable eachgenerator
controller to obtain system-level information withwhich it can make
local decisions.
Remark 1: Although the discussion above implies that theprocess
used to obtain the value of γ must extend over aninfinite number of
iterations, it is possible to distributivelydetermine the iteration
at which γ is computed to within apre-specified bound of the
asymptotic value. As shown inour previous work in [22], by
maintaining and periodicallyreinitializing min- and max-consensus
algorithms as describedin Section IV-A (the formulation of an
algorithm for min-consensus follows analogously), the nodes can
compute anapproximation to the asymptotic value of γ after a
finitenumber of iterations, kr. To obtain the simulation
resultspresented in Section VI, we make use of this variant ofthe
ratio-consensus algorithm. However, when using suchapproximations
for frequency regulation, large approximationerrors should be
avoided because they might lead to largeerrors in the computation
of the set-points, which mightcause instability problems if the
resulting equilibrium pointis no longer phase-cohesive. However, by
choosing φ in (4)strictly smaller than π2 , we allow our controller
to have anadditional stability margin to provide resiliency against
smallapproximation errors. �
C. Feasible Flow Algorithm
We begin the following discussion by introducing the prob-lem of
determining individual generator outputs that collec-tively balance
the total power demanded by the loads in amicrogrid without
violating any generation or branch powerflow limits. After
formalizing the objectives of this problem,which we refer to as the
feasible flow problem, we propose analgorithm for distributively
solving it.
1) Feasible Flow Problem Statement: For each edge{i, j} ∈ Ep,
let fij denote the power flow between nodes iand j, where fij >
0 if the actual direction of the flow is frombus i to bus j. We
restrict each flow to be within lower andupper limits, f
ijand f ij , respectively, i.e., f ij ≤ fij ≤ f ij .
For each generator node i ∈ V(g)p , let gi denote its
output,with 0 ≤ g
i≤ gi ≤ gi, where gi and gi denote lower
and upper limits, respectively. [Note that, while related,
theset-point of generator i ∈ V(g)p and its respective limits,
i.e.,ui, ui, ui, defined in Section II-A are not necessarily
equivalentto gi, gi, gi.] Similarly, for each load node i ∈ V
(`)p , let `i ≥ 0
denote its power demand. Finally, let
bi :=
{−∑j∈Np(i) fij + gi, i ∈ V
(g)p ,
−∑j∈Np(i) fij − `i, i ∈ V
(`)p ,
(29)
be the flow balance at each node i ∈ Vp.Based upon the
definitions above, the feasible flow problem
is summarized as follows. Suppose that the power demandsat the
load buses, i.e., the `i’s, are known; then, the objectiveis to
assign values to each of the flows fij , {i, j} ∈ Ep andgenerator
outputs gi, i ∈ V(g)p , such that the flow balance ateach node is
zero, the total generator outputs are balancedwith the load power
demands, and all flow assignments andgenerator outputs are within
limits. More specifically, theobjective is to obtain a set of flows
{fij : {i, j} ∈ Ep} andgenerator outputs {gi : i ∈ V(g)p } such
that:
F1. fij≤ fij ≤ f ij for {i, j} ∈ Ep;
F2. gi≤ gi ≤ gi for i ∈ V
(g)p ; and
F3. bi = 0 for i ∈ Vp.
If we define f := [fe], f := [fe], f := [fe] ∈ R|Ep| for e ∈
Ep
and g := [gi], g := [gi], g := [gi] ∈ R|V(g)p |, for i ∈ V(g)p ,
the
feasible flow problem can be written in matrix form as: findf
and g such that Mf −
[gT,−`T
]T= 0|Vp|, f ≤ f ≤ f ,
and g ≤ g ≤ g, where 0|Vp| is the |Vp|-dimensional all
zerosvector.
2) Feasible Flow Algorithm: In the discussion that follows,we
introduce an algorithm that iteratively adjusts locallymaintained
estimates for flows and generator outputs such thatthe estimates
asymptotically approach values that satisfy theconstraints of the
feasible flow problem. Unlike the consensus-type algorithms
presented in Section IV-A and Section IV-B,which were defined in
terms of a directed graph, the commu-nication network on which this
algorithm relies must conformto the undirected graph modeling the
physical layer, Gp.
To support the distributed nature of our proposed algorithm,each
local processor maintains an estimate for the value of theflows
along edges connecting it to all of its neighbors.
Morespecifically, for each node i ∈ Vp, the estimate maintained byi
for the flow to each j ∈ Np(i) at iteration k = 0, 1, 2, . . .is
denoted by f (i)ji [k]. Additionally, for each generator nodei ∈
V(g)p , we denote an estimate for its output at iteration kby
gi[k]. Finally, based upon the flow and generator outputestimates,
the value of the flow balance at each node is
-
8
computed at each iteration as
bi[k] :=
{−∑j∈Np(i) f
(i)ij [k] + gi[k], i ∈ V
(g)p ,
−∑j∈Np(i) f
(i)ij [k]− `i, i ∈ V
(`)p .
(30)
The algorithm we propose for distributively solving thefeasible
flow problem is given by the following procedure inwhich the flow
and generator output estimates are initializedand iteratively
updated using a three-step process.
[Initialization] Each node initializes its flow estimates tobe
the average of its respective lower and upper limit, i.e.,f
(i)ij [0] =
12 (f ij + f ij), i ∈ Vp, j ∈ Np(i). Analogously, the
estimate for the output of each generator node is initialized
asgi[0] =
12 (gi + gi), i ∈ V
(g)p .
[Step 1] Node i adjusts its estimate for each flow in such away
that drives the flow balance estimate to zero, i.e., bi[k]→0 as k
→∞. More specifically, node i adjusts its estimate foreach flow
as
f̃(i)ij [k + 1] = f
(i)ij [k] +
bi[k]
wi, j ∈ Np(i), (31)
where wi := δp(i) + 1. If node i ∈ V(g)p , its output estimateis
adjusted in a similar fashion to the flows, i.e.,
ĝi[k + 1] = gi[k]−bi[k]
wi. (32)
[Step 2] Since each node updates its flow estimates duringStep 1
independently, two neighboring nodes may have differ-ent estimates
for the flow between them. Thus, by exchangingflow estimates with
neighboring nodes, each node updates itslocally maintained
estimates as
f̂(i)ij [k + 1] =
1
2
(f̃
(i)ij [k + 1]− f̃
(j)ji [k + 1]
). (33)
for j ∈ Np(i). Note that while the generator outputs
areanalogous to positive injections, only each respective
localprocessor maintains an estimate for its output, eliminating
theneed for an agreement step.
[Step 3] Finally, during the first two steps, each flowestimate
may have been adjusted in such a way that violateslimits. To ensure
the flow assignment to which the nodesconverge is feasible, any
flow estimate exceeding its upper orlower bound is clamped to be
within limits, i.e., for j ∈ Np(i),
f(i)ij [k + 1] =
f ij , if f̂
(i)ij [k + 1] > f ij ,
fij, if f̂ (i)ij [k + 1] < f ij ,
f̂(i)ij [k + 1], otherwise.
(34)
Similarly, the estimates for each generator output must
beclamped to be within limits:
gi[k + 1] =
gi, if ĝi[k + 1] > gi,gi, if ĝi[k + 1] < gi,
ĝij [k + 1], otherwise.(35)
Given that the nodes incident to any given edge clamp theirflow
estimates to the same limits during Step 3, it follows thatat the
beginning of each iteration, the estimates maintained byboth nodes
are additive inverses, i.e., f (i)ij [k] = −f
(j)ji [k] =:
fij [k], {i, j} ∈ Ep; thus, the progress of the algorithm in
Steps
Algorithm 1: Distributed feasible flow algorithm
Input: fe, fe, e ∈ Ep; gl, gl, l ∈ V
(g)p ; `j , j ∈ V(`)p
Output: f (i)∗ij {i, j} ∈ Ep; g∗l , l ∈ V(g)p
Each node i ∈ Vp separately does the following:begin
initializef
(i)ji [0] =
12 (f ji + f ji), j ∈ Np(i)
gi [0] =12 (gi + gi), i ∈ V
(g)p
foreach iteration, k = 0, 1, . . . , kf docompute
bi[k] ={−∑j∈Np(i) f
(i)ij [k] + b̃i[k] + gi[k], i ∈ V
(g)p
−∑j∈Np(i) f
(i)ij [k] + b̃i[k]− `i, i ∈ V
(`)p
transmitbi[k]/wi to j ∈ Np(i)
receivebj [k]/wj from j ∈ Np(i)
computef̂
(i)ij [k+ 1] = f
(i)ij [k]−
bj [k]2wj
+ bi[k]2wi , j ∈ Np(i)ĝi[k + 1] = gi[k]− 12
bi[k]wi
, if i ∈ V(g)pset
forj ∈ Np(i), f (i)ij [k + 1] =f ij , if f̂
(i)ij [k + 1] > f ij ,
fij, if f̂ (i)ij [k + 1] < f ij ,
f̂(i)ij [k + 1], otherwise,
gi[k + 1] =gi, if ĝi[k + 1] > gi,gi, if ĝi[k + 1] <
gi,
ĝi[k + 1], otherwise,if i ∈ V(g)p
For kf sufficiently large, set:f
(i)∗ij = f
(i)ij [kf ], j ∈ Np(i); g∗l = gl[kf ], l ∈ V
(g)p
1-3 can be summarized by the following iterations:
fij [k + 1] =
[fij [k] +
1
2
bi[k]
wi− 1
2
bj [k]
wj
]fijfij
, ∀{i, j} ∈ Ep,
(36)
gi[k + 1] =
[gi[k]−
1
2
bi[k]
wi
]gigi
, ∀i ∈ V(g)p ,
(37)
where [·]xx denotes the projection onto [x, x].A summary of the
feasible flow algorithm is given in Algo-
rithm 1. The following proposition establishes the convergenceof
the algorithm to a solution that satisfies the feasible flowproblem
(see Appendix A for a proof).
Proposition 2: Suppose, for given load power demands,`, a
solution to the feasible flow problem specified by the
-
9
r = 0 r = 1 r = 2 r = r0 r = 0
FeasibleFlow
MaxConsensus
RatioConsensus
kf
u∗, hl(`0)
kf
{f (i)∗ij (�vl), f(i)∗ij (−�vl)}
σkm
kr
km
∆hl(`)∆`i
σkm
kr
σkm
kr
σkm
kr
γg → α, κ ξ(c),∆ω[0]
ξ(c),∆ω[1]
ξ(c),∆ω[r0]
Compute phase-cohesive criterion-dependent values Execute
frequency regulation scheme
Fig. 1: Timeline of distributed implementation of frequency
regulation architecture.
objectives in F1 – F3 exists and let f∗e for e ∈ Ep and g∗l forl
∈ V(g)p denote the flows and generator outputs that satisfyit.
Then, Algorithm 1 is guaranteed to yield f∗ and g∗ such
that Mf∗ −[g∗
−`
]= 0|V|, and f ≤ f∗ ≤ f and g ≤ g∗ ≤ g,
where f∗ := [f∗e ] for e ∈ Ep and g∗ := [g∗i ] for i ∈ V(g)p ,
i.e.,
as k →∞, fe[k]→ f∗e , e ∈ Ep; gl[k]→ g∗l , l ∈ V(g)p .
From the summary in Algorithm 1, we can think of thefeasible
flow algorithm as a function that takes flow andgenerator output
limits and load demands as inputs and yieldsflows and generator
outputs that satisfy the feasible flowproblem, i.e., the algorithm
can be represented by
hf : (f, f , g, g, `) 7→ (f∗, g∗). (38)
Remark 2: As with ratio consensus, the algorithm wepropose to
solve the feasible flow problem must extend overan infinite number
of iterations. In reality, by stopping theiterative process after a
finite number of iterations, whichwe denote by kf , the worst-case
error between the flow andgenerator output estimates and their true
asymptotic valuescan be made arbitrarily small. Additionally,
similar to thefinite-time approach proposed in [22], the nodes can
use maxconsensus, re-initialized every ∆d iterations, to
periodicallycompute the value of �b[∆dk] := maxi∈Vp bi[∆dk]. Given
thatthe values of the flow and generator output estimates dependon
the flow balances, it may be possible to use this instanceof max
consensus to distributively determine the iteration atwhich all
values of f (i)ij [k], {i, j} ∈ Ep and gi[k], i ∈ V
(g)p are
within some bound of their true asymptotic values. �
V. DISTRIBUTED IMPLEMENTATION
As described in Section III, our proposed frequency reg-ulation
architecture requires global information to be im-plemented. In
particular, the following values on which thearchitecture relies
require system-level information to becomputed: the average
frequency error, ∆ω[r]; the value ofβ = D
∑i∈V(g)p
αiκi, for ensuring closed loop stability; the
phase-cohesive set-points, u∗i , i ∈ V(g)p ; and the value
of
∇hl(`)∣∣∣∣`=`0
∆` to determine ξ(c). In this section, we describe
how the three algorithms outlined in Section IV—the max-
andratio-consensus and feasible flow algorithms—can
collectivelyenable the distributed computation of all these values.
Beyondeliminating the need for global information, the distributed
im-plementation we propose does not rely on time-synchronized
measurements of the frequency or phase angle; instead, onlythe
injection at each bus is required.
We begin the following discussion by outlining how
ratioconsensus can be used to compute the value of ∆ω[r] at
eachround and to compute gains that ensure the closed loop systemis
stable. Then, we describe how the feasible flow algorithmcan be
used to compute the set-points that satisfy (5) and, whencombined
with the max- and ratio-consensus algorithms, canenable the
distributed computation of an estimate for the valueof ξ(c) as
defined in (23). Finally, we discuss the timeline overwhich each of
the necessary values is distributively computed.
A. Computing the Average Frequency Error
By inspection of (8), we see that the definition of the
averagefrequency error at each round r is a ratio of sums of
valuesknown by each local processor. In particular, if we
define
xi[r] =
{ui[r], if i ∈ V(g)p ,−`i, if i ∈ V(`)p ,
(39)
it is clear that ∆ω[r] =∑
i∈Vp xi[r]∑i∈Vp Di
. Thus, if we use an
instance of ratio consensus at each round r, where y(r)i [k]and
z(r)i [k] denote the states maintained by node i, which
areinitialized as y(r)i [0] = xi[r] and z
(r)i [0] = Di, it follows that
the average frequency error can be asymptotically computed,i.e.,
limk→∞
y(r)i [k]
z(r)i [k]
= ∆ω[r], i ∈ Vc.
B. Computing Stable Gains
From the analysis in Section III-C, it can be shown that, inthe
limit as r → ∞, if the generator set-points are updatedaccording to
the control scheme in (11) – (12), the steady-statevalue of
generator i’s set-point is given by
ussi = u∗i +
αiκi∑l∈V(g)p
αlκl
∑j∈V(`)p
∆`j , (40)
which implies that the product of gains αiκi affects the
amountby which generator i will adjust its set-point away fromu∗i
in order to meet the total incremental demand for load.Furthermore,
from (15), we see that if we choose the αi’sand κi’s such that β :=
D
∑i∈V(g)p
αiκi = −1, it followsthat the frequency error that results from
one or more changesin load can be eliminated after one round of our
proposedcontrol architecture. Thus, to ensure that β = −1, we
seethat the summation −
∑i∈Vp Di must be divided among the
-
10
generators, and that the specific choice of gains will
dictatethe proportion of the total incremental demand attributed
toeach generator.
Let ∆ui = ui − u∗i and ∆ui = ui − u∗i be the lower andupper
bounds on the amount by which generator i can adjustits set-point
away from u∗i without violating its output limits
and define γg :=−
∑i∈Vp Di−
∑i∈V(g)p
∆ui∑i∈V(g)p
∆ui−∆ui. Then, if we choose
αiκi = hgi (γg), where
hgi (γg) :=∆ui + γg(∆ui −∆ui),=ui − u∗i + γg(ui − ui),
i ∈ V(g)p , (41)
it is easy to see that∑i∈V(g)p
αiκi = −∑i∈Vp Di, and that
the summation −∑i∈Vp Di is divided among the generators
proportionally to their incremental set-point limits.Similar to
the so-called fair splitting allocation scheme in
[16], we can use ratio consensus to compute the value of γg
.More specifically, given the dependence of the value of γgon the
u∗i ’s, we can use an instance of ratio consensus eachtime the
phase-cohesive set-points are computed. We denotethe states
maintained by node i for each of these instancesby y(g)i [k] and
z
(g)i [k], and, if we initialize them as y
(g)i [0] =
−Di − ∆ui if i ∈ V(g)p , y
(g)i [0] = −Di if i ∈ V
(`)p , and
z(g)i [0] = ui − ui if i ∈ V
(g)p , z
(g)i [0] = 0 if i ∈ V
(`)p , it
follows from (28) that limk→∞y(g)i [k]
z(g)i [k]
= γg . By using this
value as the argument of the hgi (·) function defined in
(41),each node can compute the product of its gains as αiκi =hgi
(y
(g)i [k0]/z
(g)i [k0]), for sufficiently large k0. Since there are
no constraints on the individual αi’s and κi’s, we choose
gainssuch that αi = h
gi (y
gi [k0]/z
gi [k0]) and κi = 1. [Note that if
we use the approximate ratio consensus algorithm described
inRemark 1 to compute γg , the approximation error that resultswill
prevent the value of β from exactly equalling −1; thus, inreality,
it may take more than one round to drive the frequencyerror to
zero.]
Remark 3: Although using the function in (41) to assigngains
ensures that β = −1, it only guarantees that theincremental
set-point of every generator is within its respectiveincremental
limits for small enough collective load perturba-tions,
∑i∈V(`)p
∆`i. However, given that the control schemein (11) – (12) is
designed to regulate the frequency for smallperturbations away from
a pre-determined phase-cohesive op-erating point, we can assume
that the load perturbations aresufficiently small such that this
choice of gain ensures allincremental set-points are within limits
for the operating rangein which they are used.
An alternative function for choosing the gains that ensuresβ =
−1 and guarantees all incremental set-points are withinlimits
is:
h̃gi (γ̃g, γ̃d) := −γ̃d (∆ui + γ̃g(∆ui −∆ui)) , (42)
for i ∈ V(g)p , where γ̃g :=∑
i∈V(`)p∆`i−
∑i∈V(g)p
∆ui∑i∈V(g)p
∆ui−∆uiand
γ̃d :=
∑i∈Vp Di∑
j∈V(`)p∆`j
. If we assume that the total amount by
which the load deviates is within the collective incremental
ca-pacity of the generators, i.e.,
∑i∈V(g)p
∆ui ≤∑i∈V(`)p
∆`i ≤
Load Bus
Generator Bus1
4
5
2
3
6
Fig. 2: Physical layer model for six-bus microgrid.∑i∈V(g)p
∆ui, this choice of gains will divide the incrementaldemand
among the generators while also ensuring β = −1.Although this
alternative function has advantages comparedto hg(·), it requires
that each load knows the value of ∆`i inorder to compute γ̃d.
Additionally, while ratio consensus canbe used to compute γ̃d,
doing so requires another operationbefore the incremental
set-points can be determined. �
C. Computing Phase-Cohesive Set-Points
By inspection, we see that ensuring the phase
cohesivenesscriterion in (4) is satisfied is equivalent to ensuring
that allnetwork flows are within limits. More specifically, when
thevector of injections at the generator and load buses is
pre-multiplied by the matrix product MTL†, the elements of
theresultant vector are the network power flows, normalized bythe
susceptances of the branches connecting each pair of buses,i.e.,
the value of each element is equal to sin(θi(t) − θj(t))for {i, j}
∈ Ep. To ensure phase-cohesive operation results,every normalized
flow in the network must be less than sin(φ),i.e., MTL†
[uT,−`T
]T ≤ sin(φ)1|Ep|, where 1|Ep| is the |Ep|-dimensional all-ones
vector and φ ∈ [0, π/2). Thus, we seethat the problem of choosing
phase-cohesive set-points subjectto the power injections at the
load buses is analogous tochoosing generator outputs such that no
branch power flowexceeds its limit.
From the discussion above, it follows that, if a solution tothe
feasible flow problem exists for given load power demands,we can
use Algorithm 1 to find the generator set-pointsthat result in an
operating point that satisfies the feasibilityproblem in (5), i.e.,
the operating point is phase-cohesive, thetotal demand for load is
balanced by the collective generatorset-points, and all the
generator set-points are within limits.In order to enforce
appropriate limits for the feasible flowalgorithm that ensure phase
cohesiveness, we see that themaximum branch power flow on any given
line is upperbounded by its susceptance. Thus, if we enforce lower
andupper limits on the flow along each edge in the feasible
flowproblem equal to f
ij= −Bij sin(φ) and f ij = Bij sin(φ),
{i, j} ∈ Ep, respectively, we can find set-points that
satisfy(5) using Algorithm 1 as follows. Let u :=
[u1, . . . , um
]Tand u :=
[u1, . . . , um
]Tbe vectors representing the lower
and upper set-point limits of the generators and define B′
:=[Bij sin(φ) : {i, j} ∈ Ep] ∈ R|Ep|; then, using the
functionalrepresentation of Algorithm 1, i.e., hf (·), with
parameters(−B′, B′, u, u, `0), the generator outputs that result
satisfy (5),i.e., u∗ = g∗.
D. Computing Phase Cohesiveness Margin
In Section III-D, we showed that by monitoring the value ofξ(c)
as the amount of power demanded by the loads changes,
-
11
0 2 4 6 8−0.2
−0.15
−0.1
−0.05
0
0.05
time, t [s]
θ i(t
)[r
ad]
θ1(t) θ2(t) θ3(t)
θ4(t) θ5(t) θ6(t)
(a) Bus voltage angles.
0 2 4 6 8
0
−0.01
−0.02
−0.03
−0.04
−0.05
−0.06
time, t [s]
∆ω
(t)
[rad
/s]
(b) Weighted average frequency error.
0 2 4 6 80.75
1
1.25
1.5
time, t [s]
ui(t
)[p
u]
u1(t) u2(t) u3(t)
u∗1 u∗2 u
∗3
(c) Generator set-points.
Fig. 3: Case I simulation results with loads `4, `5, and `6
perturbed at t = 2 s, t = 4 s, and t = 6 s, respectively.
the nodes can estimate when to recompute phase-cohesivegenerator
set-points. Recall from (23) that the sensitivity ofthe function
hl(·) to small perturbations in load, evaluated at`0, ∇hl(`)
∣∣∣∣`=`0
, is required to compute ξ(c). By noting that
the output of the function hl(·) is the maximum normalizedbranch
flow subject to injections at the generators and loads,i.e., given
set-points u∗ + ∆u(∆`) and demands `0 + ∆`,∥∥∥∥MTL† [u∗ +
∆u(∆`)−(`0 + ∆`)
]∥∥∥∥∞
= max(i,j)∈Ep
sin(θi(t)− θj(t)),
(43)
we can use the following procedure, which combines thefeasible
flow algorithm and max consensus, to distributively
compute ∇hl(`)∣∣∣∣`=`0
.
From the functional representation of Algorithm 1 definedin
(38), we see that, in addition to the generator outputs g∗i , i
∈V(g)p , the set of flows {f (i)∗ij : {i, j} ∈ Ep} is also
computed.Thus, if we use the max-consensus algorithm, with the
value
initially known by node i given by ηi = maxj∈Np(i)|f(i)∗ij |
Bij sin(φ),
it follows that, without measurements of θi(t), i ∈ Vp,
asimplied by (43), we can compute the value of∥∥∥∥MTL† [g∗−`
]∥∥∥∥∞
= max(i,j)∈Ep
|f (i)∗ij |Bij sin(φ)
= maxl∈Vc
ηl,
subject to the parameters passed to the hf (·) function.
If we approximate the gradient of hl(·) by
∇hl(`) ≈[
∆hl(`)∆`m+1
, · · · , ∆hl(`)
∆`n
], (44)
we can compute it using 2|V(`)p |-instances of the feasible
flowand max-consensus algorithms. More specifically, given thatwe
are interested in determining the value of ξ(c) for theclosed-loop
system as it evolves away from the operatingpoint for which the u∗i
’s were computed, we can approximatethe sensitivity of hl(·) to
small changes in demand at loadi ∈ V(`)p , i.e., the value of
∆h
l(`)∆`i
∣∣∣∣`=`0
, subject to our proposed
control architecture, as follows. Define vi ∈ R|Vp| to be
avector with 1 in the ith coordinate and 0’s elsewhere; then,for
each i ∈ V(`)p , let `(�vi) := `0 + �vi be a vector of load
demands with the ith load perturbed by �. From (40), let
uj(�) := u∗j + �αjκj/
∑l∈V(g)p
αlκl (45)
be the set-point of generator j ∈ V(g)p given one such
loadperturbation, subject to the control scheme in (11) –
(12).Furthermore, let f∗e (�vi), e ∈ Ep be the flows that result
fromAlgorithm 1 with parameters (−B′, B′, u(�), u(�), `(�vi)),where
u(�) :=
[u1(�), . . . , um(�)
]T. Then, we can use an
instance of the max-consensus algorithm, where the valueknown by
node j is
ηj(�vi) = maxl∈Np(j)
|f (j)∗jl (�vi)|Bjl sin(φ)
, (46)
to compute the value of hl(`(�vi)) = max(j,l)∈Ep|f(j)∗jl
(�vi)|Bjl sin(φ)
=
maxj∈Vc ηj(�vi). If we use the central difference approxima-
tion to estimate the value of ∆hl(`)
∆`i
∣∣∣∣`=`0
, then it follows that
∆hl(`)
∆`i
∣∣∣∣`=`0
≈ hl(`(�vi))− hl(`(−�vi))
2�, (47)
where hl(`(−�vi)) is computed analogously to hl(`(�vi))
with`(−�vi) = `0 − �vi and uj(−�) = u∗j −
�αjκj∑l∈V(g)p
αlκl.
Given that the value of ξ(c) depends on an appropriatechoice of
c, which depends on the value of hl(`) for ` = `0,we can use the
process described above for computing theindividual hl(`(�vi))’s to
determine a value of c that is lessthan sin(φ) − hl(`0). By using
max consensus to computehl(`0), which can be combined with a
pre-determined localrule for choosing c, the node with the
maximizing flow can
also be determined, i.e., if ηi = maxj∈Np(i)|f(i)∗ij |
Bij sin(φ), the
node l = argmax{ηj : j∈Vc}
ηj is the one with the maximizing flow.
Then, by combining the appropriately chosen value of c withthe
approximation to the sensitivity of the function hl(·) foundusing
the process described above, the nodes can compute thevalue of ξ(c)
as follows. Let y(p)i [k] and z
(p)i [k] denote the
states maintained by node i, where y(p)i [0] =∆hl(`)
∆`i
∣∣∣∣`=`0
∆`i
if i ∈ V(`)p , and y(p)i [0] = 0 otherwise; and z(p)i [0] =
c
if i = argmax{ηj : j∈Vc}
ηj and z(p)i [0] = 0 otherwise. Then, by
-
12
0 2 4 6 80.18
0.19
0.2
0.21
0.22
time, t [s]
Max
imum
Net
wor
kFl
ows Estimated Actual
Fig. 4: Estimated vs. actual maximum flowfor Case I.
0 5 10 15 20 25 30 35 40 45 50
−1−0.5
0
0.5
1
iterations, k
b̂ i[k
]
b1[k] b2[k] b3[k] b4[k] b5[k] b6[k]
Fig. 5: Nodal balance estimates as feasible flow algorithm
evolves for initialcomputation of u∗ for Case I.
updating the states according to (26) – (27), the nodes
canasymptotically obtain
limk→∞
y(p)i [k]
z(p)i [k]
=
∑i∈V(`)p
∆hl(`)
∆`i
∣∣∣∣`=`0
∆`i
c.
E. Timeline
In the preceding discussion, we showed how each of thevalues
necessary to implement our proposed frequency reg-ulation
architecture can be acquired using a combination ofthe distributed
algorithms introduced in Section IV. Next, weprovide an overview of
the timeline over which our proposeddistributed implementation
operates, specifically discussingthe order in which the algorithms
must be executed to properlycompute the necessary values.
As discussed in Section III-B, at each round of our
proposedfrequency regulation architecture, in addition to the
weightedaverage frequency error, ∆ω[r], the output of each
generatori ∈ V(g)p is adjusted according to its phase-cohesive
set-point,u∗i , and gains, αi and κi, all of which depend on
globalinformation. Additionally, as discussed in Section III-D,
toensure the system remains phase cohesive for large changesin
load, the nodes monitor the value of ξ(c), the definition
of which depends on ∆hl(`)
∆`i
∣∣∣∣`=`0
, i ∈ V(`)p , as the system
evolves. Given that these quantities—u∗i , αi, κi, i ∈ V(g)p
and ∆hl(`)
∆`i
∣∣∣∣`=`0
, i ∈ V(`)p , which we collectively refer toas phase-cohesive
criterion-dependent values—must be knownto compute the generator
set-points and ξ(c), it is clear thatthey need to be determined
before our proposed frequencyregulation scheme can be executed.
Moreover, from (41), we
see that, in order to compute the ∆hl(`)
∆`i
∣∣∣∣`=`0
’s, the controller
gains, αi and κi, i ∈ V(g)p , must be known
beforehand.Similarly, from (45), we see that the u∗i ’s must be
known inorder to compute the αi’s and κi’s. Thus, it is clear that
thephase cohesive criterion-dependent values must be computedin a
specific order; we provide a detailed description of thisorder
next.
Prior to operation and immediately following the roundfor which
the value of ξ(c) is found to exceed unity, i.e.,before round r =
0, the nodes use the feasible flow algorithmto compute the
phase-cohesive set-points, u∗. The generator
nodes then use the u∗i ’s to determine their incremental
limits,∆ui and ∆ui, i ∈ V
(g)p and, together with the processors
located at the load buses, use ratio consensus to determinethe
value of γg . With the value of γg , or an approximationthereof,
each generator processor then determines its controllergains, αi
and κi, according to the function h
gi (·). Using
2|V(`)p |-instances of the feasible flow algorithm, the nodes
thendetermine the sets of flows {f (i)∗ij (�vl) : l ∈ V
(`)p , {i, j} ∈ Ep}
and {f (i)∗ij (−�vl) : l ∈ V(`)p , {i, j} ∈ Ep}, where f (i)∗ij
(�vl)
and f (i)∗ij (−�vl) are the flow assignments that result when
loadl ∈ V(`)p is perturbed by � and −�, respectively, and the
set-point of generator j ∈ V(g)p is lower- and upper-bounded byu(�)
and u(−�), respectively. From the sets of flows that result,the
nodes can use max consensus to determine the values ofhl(`(�vl))
and hl(`(−�vl)) for l ∈ V(`)p , with which the loadprocessors can
compute an estimate of ∆h
l(`)∆`l
∣∣∣∣`=`0
. After the
phase-cohesive set-points, generator gains, and load
sensitivityestimates are computed, the frequency regulation scheme
canbegin operation, with the nodes using separate instances ofratio
consensus to compute the values of ξ(c) and ∆ω[r] ateach round. An
overview of the order in which each value iscomputed is illustrated
by the timeline in Fig. 1.
VI. SIMULATION RESULTS
In this section, we present numerical simulation results
forthree test cases in which our proposed control architectureand
its distributed implementation are used. For each case,we
demonstrate closed-loop operation given a series of
smallperturbations to the loads and discuss several metrics
thatillustrate its effectiveness. We consider a six-bus ring
networkfor the first two test cases; the network consists of
threegenerators and three loads. For the third case, we utilize a
treenetwork consisting of 37-buses, 15 of which are generators,and
22 of which are loads. Generator, load, network, andsimulation
parameters for each of the cases can be foundin [23, Appendix B];
furthermore, in all cases, we makeuse of the finite-time
ratio-consensus algorithm mentioned inRemark 1.
A. Cases I and II: Six-Bus Ring Network
We consider a six-bus microgrid, the physical layer model
ofwhich is illustrated by the graph in Fig. 2. In Cases I and
II,
-
13
0 2 4 6 8
−0.15
−0.1
−0.05
0
0.05
time, t [s]
θ i(t
)[r
ad]
θ1(t) θ2(t) θ3(t)
θ4(t) θ5(t) θ6(t)
(a) Bus voltage angles.
0 2 4 6 8
0.05
0.04
0.03
0.02
0.01
0
−0.01−0.02−0.03−0.04
time, t [s]
∆ω
(t)
[rad
/s]
(b) Weighted average frequency error.
0 2 4 6 80.80.9
11.11.21.31.41.5
time, t [s]
ui(t
)[p
u]
u1(t) u2(t) u3(t)
u∗1 u∗2 u
∗3
(c) Generator set-points.
Fig. 6: Case II simulation results with loads `4, `6, and `5
perturbed at t = 2 s, t = 4 s, and t = 6 s, respectively.
the total power initially demanded by the loads is taken
tobe∑i∈V(`)p
`0i = 3.3 pu and at time t = 0−, prior to the
computation of u∗i ’s, the generators equally share the load
suchthat ui[0−] = 1.1 pu, i = 1, 2, 3. Using the method describedin
Section V-D, the sensitivity of the function hl(·) to changesin
load `6 is found to be negative, i.e., a decrease in the
powerdemanded by `6 will lead to an increase in the value of
hl(·).
1) Case I: For this case, we consider the closed-loopresponse of
the six-bus microgrid subject to an increase inpower demand by
loads `4, `5, and `6 at time t = 2, t = 4,and t = 6, respectively.
Each load demand is perturbed byan increase of at most 7.5% of its
original value, with `4perturbed by 5% such that ∆`4 = 0.0575 pu,
`5 perturbed by7.5% such that ∆`5 = 0.09375 pu, and `6 perturbed by
5%such that ∆`6 = 0.045 pu. Given these perturbations and thechoice
of c = 0.01, our proposed control architecture triggersa
recomputation of the u∗i ’s at time t ≈ 4 s, immediatelyafter load
`5 is perturbed. The response of the closed-loopsystem to the three
perturbations and the recomputation of u∗
is illustrated by the plots in Fig. 3. From Fig. 3a and Fig. 3b,
itcan be seen that, following each load perturbation, the
voltageangles stabilize and the weighted average frequency
errorthat results from the changes in load is quickly
eliminated.Additionally, from Fig. 3c, in which the generator
set-pointsand their respective values of u∗i are illustrated, it
can beseen that the generators increase their output according to
thecontrol scheme in (11) – (12) at times t = 2 s and t = 6 s
andthat new values of u∗ are computed and applied following
theperturbation to load `5 at t = 4 s.
To demonstrate the effectiveness of our proposed methodfor
estimating the value of the function hl(·) as the loads
areperturbed, the actual maximum normalized flow as computedat each
time step, i.e., max(i,j)∈Ep sin(θi(t) − θj(t)), and its
estimate, i.e., hl(`0)+∇hl(`)∣∣∣∣`=`0
, are shown in Fig. 4. From
the figure, we see that the estimated maximum flow closelytracks
the true value, with an error on the order of 0.001throughout the
simulation period. Additionally, we see that theestimate properly
predicts that an increase in `6 will lead to adecrease in the
maximum network flow. Finally, we illustratethe evolution of the
nodal flow balances maintained by eachnode for the first 50
iterations of the feasible flow algorithmas it is used to compute
the initial u∗i ’s in Fig. 8. From this
figure, it can be seen that the flow balances maintained by
allnodes quickly approach zero, with every node’s estimate
beingwithin 0.001 of zero within the first 50 iterations.
2) Case II: For this case, we again consider the
six-busmicrogrid, but, in order to demonstrate the negative
sensitivityof the function hl(·) to changes in demand at load `6,
weevaluate its closed-loop response to increases in demand atloads
`4 and `5, and a decrease in demand at load `6.Specifically, the
loads are perturbed as follows: `4 is increasedby 5% at t = 2 s
such that ∆`4 = 0.0575 pu, `6 is decreasedby 7.5% at t = 4 s such
that ∆`6 = −0.0675 pu, and `5increased by 4.8% at t = 6 s such that
∆`5 = 0.06 pu. Giventhese perturbations and the choice of c = 0.01,
our proposedcontrol architecture triggers a recomputation of the
u∗i ’s (andthe load sensitivities) at time t ≈ 6 s, immediately
after load`5 is perturbed.
The response of the closed-loop system to the three
pertur-bations and the recomputation of u∗ is illustrated by the
plotsin Fig. 6. As in Case I, Fig. 6a and Fig. 6b, illustrate
that,following each load perturbation, the voltage angles
stabilizeand the weighted average frequency error that results from
thechanges in load is quickly eliminated. Additionally, Fig.
6cillustrates the generator set-points and their respective u∗i
’sas the control architecture responds to the load
perturbations.The figure illustrates that the third load
perturbation, at timet = 6 s, increases the estimate for hl(·) such
that, given thevalue of c, a recomputation of the u∗i is
performed.
B. Case III: 37-Bus Tree Network
For this case, we consider the closed-loop response of the37-bus
tree network, the topology of which is illustrated inFig. 7.
Additionally, the total power initially demanded by theloads is
taken to be
∑i∈V(`)p
`0i = 28.92 pu and at time t = 0−,
prior to the computation of u∗i ’s, the generators equally
sharethe load such that ui[0−] = 1.928 pu, i = 1, . . . , 15.
To demonstrate the closed-loop response of our proposedfrequency
regulation architecture using the 37-bus network,we increase the
amount of power demanded by loads `20,`25, and `33 by 15% at t = 3
s, t = 5 s, and t = 7 s,respectively, such that ∆`20 = 0.2143 pu,
∆`25 = 0.2624 pu,and ∆`33 = 0.2195 pu. Given these load
perturbations andthe choice of c = 0.01, a recomputation of the u∗i
’s and the
-
14
Load Bus
Generator Bus
17
16
18
19
20
21 23
22
24
25
26
27
28
29 30
31
32
34
33
35 3637
1 2
3 4
5 6
7
8
9
10
11 12 13 14
15
Fig. 7: Graph-theoretic model of physical layer of 37-bus
microgrid.
0 1 2 3 4 5 6 7 8 90.34
0.36
0.38
0.4
time, t [s]
Max
imum
Net
wor
kFl
ows Estimated Actual
Fig. 8: Estimated vs. actual max flow for Case III.
sensitivities of hl(·) to changes in load is performed
shortlyafter load `33 is perturbed at t ≈ 7 s. The response ofthe
closed-loop system is illustrated by the plots in Fig.
9.Specifically, the evolution of the weighted average
frequencyerror is shown in Fig. 9a, from which it can be seen
thatthe frequency error that results from the load perturbations
isquickly eliminated. Additionally, the generator set-points
andrespective u∗i ’s (illustrated by the dotted traces) are shownin
Fig. 9b. From the figure, it can be seen that the amountby which
the generators deviate away from the originallycomputed u∗i ’s is
relatively small, and that the new u
∗i ’s are
computed at t ≈ 7 s after load `33 is perturbed.As in the
six-bus cases, we illustrate the evolution of
the estimated and computed values for the maximum flowthroughout
the simulation in Fig. 8. Similar to the smallernetwork cases, the
estimated value of hl(·) closely tracks thetrue computed value,
with errors on the order of 0.0001 forthe entire 9-second
duration.
VII. CONCLUDING REMARKS
In this paper, we introduced a control architecture suitablefor
regulating the frequency in an islanded ac microgrid withno
inertia. The approach we proposed is designed to regulatethe
average frequency subject to small load perturbations byadjusting
the output of the generators in the system around set-points that
are known to result in phase-cohesive operation. Tohandle larger
perturbations in the loads, we also proposed amethod that enables
the computation of an estimate for theamount by which the system
has deviated from the phase-cohesive operating point; the
controller monitors this estimateas the system evolves to determine
when the set-points to betracked should be recomputed. We also
proposed a distributedimplementation of our proposed architecture,
including theintroduction of an algorithm that enables processors
locatedat each bus to determine phase-cohesive set-points and
thenetwork flows that result. Finally, we demonstrated our
pro-posed control architecture and its distributed
implementationusing three case studies applied to two test
systems.
APPENDIX AANALYTICAL RESULTS
A. Proof for Proposition 1
In order to choose αi’s and κi’s such that ρ(Φ) ≤ 1, whereΦ is
defined in (13), we consider the characteristic equation of
Φ given by p(λ) = det(Φ− λIm+1). By taking advantage ofthe
structure of Φ − λIm+1 and using its cofactor expansionalong row 1,
p(λ) becomes
p(λ) = (β − λ)(1− λ)m +Dm∑j=1
(−1)jαj det(M1,j+1),
(48)
where M1,j+1 is the matrix Φ−λIm+1 with the first row and(j +
1)th column deleted. Using cofactor expansion along thefirst column
of M1,j+1, it can be shown that
det(M1,j+1) = (−1)j+1κj(1− λ)m−1. (49)
Thus, the characteristic equation becomes
p(λ) =(β − λ)(1− λ)m
+D(1− λ)m−1m∑j=1
(−1)j(−1)j+1αjκj
=(1− λ)m−1λ[λ− (β + 1)], (50)
from which it follows that λ1 = β + 1, λm = 0, and λj = 1,j = 3,
4, . . . ,m+ 1.
From the analysis above, it is clear that by
appropriatelychoosing gains such that −2 ≤ β ≤ 0, all of the
eigenvaluesof Φ will lie on the boundary of or within the unit
circle.However, for marginal stability, given that λ = 1 has
algebraicmultiplicity m− 1, we must show that Φ is nondefective,
i.e.,we must show that the geometric multiplicity of λ = 1 ism − 1.
If we consider the characteristic equation in (50) forλ = 1, we see
that since the rightmost m columns of thematrix Φ−Im+1 are linearly
dependent, the rank of Φ−Im+1is 2. Given the relationship between
the rank and nullity of amatrix, it is clear that the m− 1
eigenvectors associated withλ = 1 are linearly independent, which
implies that λ = 1 hasgeometric multiplicity m− 1.
Given that Φ is nondefective, we can write its Jordanform as Φ =
UJU−1 where J := diag({λj}) and U is anonsingular matrix composed
of columns corresponding to theeigenvectors of Φ given by
U =
[β −1 0 . . . 0κ κ w1 . . . wm−1
], (51)
where wj ∈ Rm, j = 1, 2, . . . ,m−1, are linearly independentand
orthogonal to α. It then follows that we can write the value
-
15
0 1 2 3 4 5 6 7 8 9
0.01
0
−0.01−0.02−0.03−0.04−0.05−0.06−0.07
time, t [s]
∆ω
(t)
[rad
/s]
(a) Weighted average frequency error.
0 1 2 3 4 5 6 7 8 9
1.4
1.6
1.8
2
2.2
2.4
2.6
time, t [s]
ui(t
),i
=1,...,1
5[p
u]
(b) Generator set-points.
Fig. 9: Case III simulation results with loads `20, `25, and `33
perturbed at t = 3 s, t = 5 s, and t = 7 s, respectively.
of[∆ω[r], e[r]
]Tin matrix form as[
∆ω[r]e[r]
]=UJrU−1
[∆ω[0]e[0]
]+
[r−1∑s=0
UJsU−1
] [−10m
]D∑i∈V(`)p
∆`i, (52)
where ∆ω[0] = D∑i∈V(`)p
∆`i and e[0] = 0m. Note that
(U(Jr −r−1∑s=1
Js))1 = [β(1 + β)r − β
r−1∑s=1
(1 + β)s, 0Tm],
(53)
where (U(Jr −∑r−1s=1 J
s))1 is the first row of U(Jr −∑r−1s=1 J
s). Then, by using (52) – (53), we obtain that
∆ω[r] =
(((1 + β)r −
r−1∑s=1
(1 + β)s)β(U−1)11 − 1
)∆ω[0],
(54)
where (U−1)ij is the (i, j)-th entry of the matrix U−1.
Bydirectly computing ∆ω[1] and solving the following equation:
∆ω[1] = (β − 1)∆ω[0] = (β(1 + β)(U−1)11 − 1)∆ω[0],(55)
we find that (U−1)11 = 11+β . Then, by using (54), weobtain that
the value of the average frequency error given thecontroller in
(11) – (12) at any round r is given by
∆ω[r] = (1 + β)r−1(β − 1)D∑i∈V(`)p
∆`i. (56)
From (56), we see that choosing gains such that β = −2
willresult in oscillatory behavior from the closed loop
system;similarly, choosing gains such that β = 0 will lead
toconstant frequency error, i.e., ∆ω[r] ≡ −D
∑i∈V(`)p
∆`i,r = 0, 1, 2, . . . . Thus, to ensure that ∆ω[r] → 0 as r →
∞,we must enforce strict inequalities on the value of β, i.e.,
werestrict to choosing gains such that −2 < β < 0.
B. Proof for Proposition 2
Consider the following quadratic optimization problem:
ming,f
12b
TQb
s.t. fij≤ fij ≤ f ij , {i, j} ∈ Ep
gi≤ gi ≤ gi, i ∈ V
(g)p ,
(57)
which is a quadratic programming problem, where Q is a diag-onal
weight matrix with entries Qii = 1wi =
1δp(i)+1
, ∀i ∈ Vp.To solve this problem, we define Dg = {g : gj ≤ gj ≤gj
, j ∈ V
(g)p } and Df = {f : f ij ≤ fij ≤ f ij , {i, j} ∈ Ep},
and apply the gradient projection method, the iterations ofwhich
are given by [24, Section 2.3]:
f [k + 1] =
[f [k]− s
2
(∂bTQb
∂f[k]
)T]+Df
, (58)
g[k + 1] =
[g[k]− s
2
(∂bTQb
∂u[k]
)T]+Dg
, (59)
where [·]+D denotes the projection onto the set D. The
iterationsin (58) – (59) can be simplified as follows:
f [k + 1] =
[f [k]− s
(∂b
∂f
)TQb[k]
]+Df
=[f [k] + sMTQb[k]
]+Df, (60)
g[k + 1] =
[g[k]− s
(∂b
∂g
)TQb[k]
]+Dg
= [u[k]− sQb[k]]+Dg . (61)
The iterations in (60) – (61) can be written as:
fij [k + 1] =
[fij [k] + s
bi[k]
wi− sbj [k]
wj
]fijfij
, ∀{i, j} ∈ Ep
gi[k + 1] =
[gi[k]− s
bi[k]
wi
]gigi
, ∀i ∈ V(g)p ,
which are identical to iterations (36) – (37) when s = 1/2.Now,
we show that, for 0 < s ≤ 0.8, iterations (60) – (61)converge
globally to a phase-cohesive solution. Define x[k] =
-
16
[gT[k], fT[k]
]Tand ∆x[k + 1] = x[k + 1]− x[k]. Then,
‖∆x[k + 1]‖ =∥∥[Ax[k]]+ − [Ax[k − 1]]+∥∥ , (62)
where [·]+ is the projection onto the positive orthant, A =[Im −
sQ sQMsMTQ I|Ep| − sMTQM
]. Since [·]+ is a non-expansive
mapping, we have that ‖[Ax[k]]+ − [Ax[k − 1]]+‖ ≤‖A∆x[k]‖ [24,
Proposition 2.1.3]; therefore,
‖∆x[k + 1]‖ ≤ ‖A∆x[k]‖ . (63)
Define F =[
Q QMMTQ MTQM
], which is a symmetric
and positive-semidefinite matrix, and let λmax(F ) denote
thelargest eigenvalue of F . Clearly, if 0 < s < 2λmax(F ) ,
then,−1 < λ(A) ≤ 1, where λ(A) is any eigenvalue of A. Next,we
find γ such that γ ≥ λmax(F ). By using the determinantformula for
the Schur complement of Q−λIm, where λ is aneigenvalue of F , we
have that
0 = det(F − λIm+|Ep|),= det(Q− λIm) det(MTQM − λI|Ep| −H),
(64)
where H = MTQ(Q − λIm)−1QM . Let λmax denote thelargest
eigenvalue of MTQM , and choose γ ≥ λmax + 0.5.Since Q ≤ 0.5Im, we
have that γIm − Q ≥ λmaxIm ≥2λmaxQ; thus,
MTQM − γI|Ep| −H ≤MTQM(1 +
0.5
λmax)− γI|Ep| ≤ 0,
which implies that λmax(F ) ≤ γ. The absolute sum of theentries
in each row corresponding to an edge {i, j} ∈ Ep ofMTQM can be
easily shown to be equal to δp(i)δp(i)+1 +
δp(j)δp(j)+1
.Therefore, by Gershgorin disc theorem (see [25, Theo-rem
6.1.1]), λmax ≤ max
{i,j}∈Ep
δp(i)δp(i)+1
+δp(j)δp(j)+1
. Define χ =
max{i,j}∈Ep
δp(i)δp(i)+1
+δp(j)δp(j)+1
. Then, we can choose γ = χ+ 0.5
and any 0 < s < 2γ . Note that χ < 2 and γ <
2.5,which implies that we can choose any 0 < s ≤ 0.8 tohave 1 ≥
λ(A) > −1. Since A is symmetric, we candecompose x[k] as
follows: x[k] = v[k] + c[k], whereAv[k] 6= v[k], Ac[k] = c[k] and
v[k] ⊥ c[k]. Define∆v[k+ 1] = v[k+ 1]− v[k] and ∆c[k+ 1] = c[k+ 1]−
c[k].Then, ‖A∆x[k]‖ ≤ σ2(A) ‖∆v[k]‖+‖∆c[k]‖, where σ2(A) isthe
second largest singular value of A. By using (63), we havethat
‖∆v[k + 1]‖+‖∆c[k + 1]‖ ≤ σ2(A) ‖∆v[k]‖+‖∆c[k]‖.Then, clearly,
‖∆v[k]‖ → 0 as k → ∞ since σ2(A) < 1.Suppose that v[k]→ v and
v[k] = v+ �[k] for some sequence�[k]. To show that ‖∆x[k]‖ → 0, it
suffices to show that‖∆c[k]‖ → 0. The iterations in (60) – (61) can
be written as:
v[k + 1] + c[k + 1] = [Ax[k]]+ = v[k] + c[k]−
sX[k]Fv[k],(65)
where X[k] is an (m+ |Ep|)×(m+ |Ep|) diagonal matrix with0 ≤
Xii[k] ≤ 1. Then, from (65), we have that
v + �[k + 1] + c[k + 1] =v + �[k] + c[k]− sX[k]F (v + �[k]),
or c[k+1] = �[k]+c[k]−sX[k]F (v+�[k])−�[k+1]. Finally,from the
last equation we obtain that
limk→∞
(∆c[k + 1] + sX[k]Fv) = limk→∞
(�[k]− sX[k]F�[k]
− �[k + 1]) = 0.
If limk→∞
sX[k]Fv = 0, then, limk→∞
∆c[k + 1] = 0, whichwould establish the convergence of the
algorithm. Suppose thatsX[k]Fv does not converge to 0. Since
Xi[k](Fv)i ≥ 0 orXi[k](Fv)i ≤ 0 for all k, then, without loss of
generality, wecan assume that there exists some i for which
Xi[k](Fv)i ≥ δ,∀k > N , for some δ > 0 and N . From (65), we
have that
xi[k + 1] = xi[k]− sXi[k](Fv)i − sXi[k](F�[k])i. (66)
Since �[k] → 0, (66) implies that xi[k] converges to itsupper
limit and Xi[k](Fv)i → 0, which is a contradiction.Therefore,
lim
k→∞sX[k]Fv = 0, lim
k→∞∆c[k] = 0, and x[k]
converges to a stationary point [24, Proposition 2.3.2], whichis
also a global minimum since the cost function 12b
TQb isconvex in f and g [24, Proposition 2.1.2].
REFERENCES
[1] R. Lasseter et al., “Integration of distributed energy
resources: TheCERTS microgrid concept,” Lawrence Berkeley National
Laboratory,Tech. Rep. LBNL-50829, Apr. 2002.
[2] J. Guerrero, J. Vasquez, J. Matas, L. de Vicuña, and M.
Castilla,“Hierarchical control of droop-controlled ac and dc
microgrids–a generalapproach toward standardization,” IEEE Trans.
Ind. Electron., vol. 58,no. 1, pp. 158 – 172, Jan. 2011.
[3] A. Wood and B. Wollenberg, Power Generation, Operation, and
Control.New York, NY: Wiley, 1996.
[4] M. Marwali, J.-W. Jung, and A. Keyhani, “Control of
distributedgeneration systems - part II: Load sharing control,”
IEEE Trans. PowerElectron., vol. 19, no. 6, pp. 1551–1561, Nov.
2004.
[5] J. W. Simpson-Porco, F. Dörfler, and F. Bullo,
“Synchronization andpower sharing for droop-controlled inverters in
islanded microgrids,”Automatica, vol. 49, no. 9, pp. 2603–2611,
2013.
[6] N. Hatziargyriou, H. Asano, R. Iravani, and C. Marnay,
“Microgrids,”IEEE Power and Energy Magazine, vol. 5, no. 4, pp.
78–94, July 2007.
[7] A. Tsikalakis and N. Hatziargyriou,