3794 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO.

5, SEPTEMBER 2016

Fast Local Voltage Control Under Limited Reactive

Power: Optimality and Stability Analysis
Hao Zhu, Member, IEEE, and Hao Jan Liu, Student Member, IEEE

Abstract—High penetration of distributed energy resources this centralized problem using information exchanges among
presents several challenges and opportunities for voltage reg- neighboring buses. To balance overall system VAR resources,
ulation in power distribution systems. A local reactive power the second-stage VAR control in [4] relies on a consensus av-
(VAR) control framework will be developed that can fast re-
spond to voltage mismatch and address the robustness issues
eraging protocol. Alternating-direction method-of-multipliers
of (de-)centralized approaches against communication delays (ADMM) has been advocated in [7], [8], while a subgradient
and noises. Using local bus voltage measurements, the proposed iterative solver has been developed by [9]. More recently, a
gradient-projection based schemes explicitly account for the VAR stochastic-approximation approach has been adopted in [10]
limit of every bus, and are proven convergent to a surrogate cen- to handle high system variability and measurement noises.
tralized problem with proper parameter choices. This optimality Nonetheless, all (de-)centralized approaches would require
result quantifies the capability of local VAR control without re-
quiring any real-time communications. The proposed framework
high-quality communication of the measurement and control
and analysis generalize earlier results on the droop VAR control signals, which is not yet a reality for almost all distribution
design, which may suffer from under-utilization of VAR resources systems. Since these optimization-based control methods are
in order to ensure stability. Numerical tests have demonstrated the designed in an open-loop fashion, potential communication
validity of our analytical results and the effectiveness of proposed delays or noises would challenge their optimality and stability
approaches implemented on realistic three-phase systems. for real-time implementations.
Index Terms—Voltage regulation, dynamic system stability, To tackle this, one can design VAR control strategies using
Volt/VAR control, power distribution systems, gradient-projection locally available information such as bus voltage magnitude
methods. measurements [1], [2], [11]. Since power system voltage is
more significantly affected by local VAR inputs compared to
those elsewhere ([12], Section 10.8), a local control framework
I. INTRODUCTION
would very fast and effectively respond to voltage deviation.

R ECENT technological advances in distribution systems

have led to growing interest in the voltage regulation
problem using distributed energy resources (DERs) of inverter
One challenge in local control design is the guaranteed dynamic
system stability. As shown by [13]–[15], local inverter VAR
control would potentially increase the number of tap changes
interfaces [1]. High variability of photovoltaic (PV) solar gen- triggered for on-load tap-changing (OLTC) transformers.
eration and abrupt load increase due to electric vehicle charging Hence, stability of local VAR control could be crucial to ensure
could result in unexpected voltage fluctuations, at time-scales that it will not offset, or even adversely affect OLTC trans-
much faster than the existing voltage control using on-load tap- formers and other voltage regulating devices. One popular local
changers or capacitor banks; see e.g., [1]–[4]. Further DER de- control is the piecewise linear droop design using instantaneous
ployment gives rise to unprecedented capability of extremely local bus voltage mismatch input, as advocated by the IEEE
fast voltage regulation through inverter VAR output control, in 1547.8 Standard [16]. However, as shown by [17] the droop
order to meet the 5% voltage deviation bounds per the ANSI slope has to be small enough to ensure system stability. This
Standard C84.1 [5]. condition is equivalent to enforcing high penalty on inverter
With the full system information available centrally, the VAR output, which would result in insufficient utilization of
VAR control problem can be cast as an optimal power flow VAR resources. A delayed droop scheme was developed in [3],
(OPF) one that minimizes system operational costs such as which could relax the droop stability condition. However, it
power losses and voltage violation [6]. Several distributed is not clear how to choose the delay parameter to balance the
optimization algorithms have also been proposed to solve stability and the convergence speed. Stability and optimality
analysis for similar integral control approaches have been
offered in [18], [19], but they assume unconstrained VAR
Manuscript received April 15, 2015; revised August 18, 2015, October 18,
2015; accepted November 22, 2015. Date of publication December 21, 2015; resources at every bus and thus fail to account for inverter
date of current version August 17, 2016. This work was supported in part by the rating limits in practice.
Illinois Center for a Smarter Electric Grid (ICSEG), in part by the Trustworthy The present paper offers a general framework for developing
Cyber Infrastructure for the Power Grid (TCIPG), and in part by the Cyber Re-
silient Energy Delivery Consortium (CREDC). Paper no. TPWRS-00512-2015. local VAR control strategies that explicitly account for the VAR
The authors are with the Department of Electrical & Computer Engineering, limits at every bus and potential VAR supplying penalty. This
University of Illinois, Urbana, IL 61801 USA (e-mail: [email protected]; problem turns out to be a box-constrained quadratic optimiza-
[email protected]).
Color versions of one or more of the figures in this paper are available online
tion problem, which motivates us to leverage the iterative gra-
at http://ieeexplore.ieee.org. dient-projection (GP) method ([20], Section 2.3). Interestingly,
Digital Object Identifier 10.1109/TPWRS.2015.2504419 the GP iterations naturally decouple into local updates requiring

0885-8950 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ZHU AND LIU: FAST LOCAL VOLTAGE CONTROL UNDER LIMITED REACTIVE POWER 3795

only bus voltage magnitude, shown to attain the optimal so-

lution to a surrogate centralized VAR control problem under
proper choice of parameters. Compared to the existing litera-
ture on (local) VAR control, our contribution is three-fold. First,
our proposed GP-based local control generalizes the existing
droop and delayed droop methods for limited VAR resources
in [3], [17]. Second, compared to these earlier approaches, our
Fig. 1. A radial distribution feeder with bus and line associated variables.
GP-based update can be stabilized under any choice of VAR
supplying penalty. Last but for least, motivated by the acceler-
ation from Newton's method, we have advocated to diagonally
Assuming the loss is negligible compared to line flow, a linear
scale the GP stepsize parameters based on the inverse Hessian
approximation of (1) can be constructed. The approximation
matrix. This proposed stepsize selection method can improve
error introduced is relatively small, at the order of 1% [17].
the system conditioning and convergence speed, while sacri-
Under relatively flat voltage profile, i.e., , we
ficing no optimality condition.
have . This leads to a small approxi-
The rest of the paper is organized as follows.
mation error at about 0.25% (1%) if there is a 5% (10%) devi-
Section II presents the linearized distribution system flow
ation in voltage magnitude approximation [17]. Under the two
model, which is further reformulated by introducing graph
assumptions, the linearized DistFlow (LinDistFlow) model can
matrices in Section III. The gradient-projection (GP) algorithm
be established for every , as
is introduced in Section IV to solve the limit-constrained
VAR optimization problem, which leads to equivalent local (2a)
voltage control updates. Stability conditions have been offered
in Section V for selecting GP stepsize parameters, which are
based on only system topology and line admittance. Numerical (2b)
tests on realistic (three-phase) feeder systems have been
performed for real-time implementations, which demonstrate (2c)
the effectiveness of the proposed scaled control design with
improved stability and voltage regulation performance. An The total injected reactive power , where
important issue of local VAR control is its interactions with denotes the reactive power contributed by e.g., PV inverters per
other voltage regulation devices, and we are actively pursuing bus , while corresponds to the load reactive power consump-
this direction to extend the present work. tion at bus . Using the bold symbols to denote their vector coun-
terparts, the VAR control problem becomes to solve for for
II. SYSTEM MODELING AND PROBLEM STATEMENT given and , such that
Consider a distribution network with buses collected in (3a)
the set , and line segments represented by
the set ; see Fig. 1 for a radial feeder il- (3b)
lustration. For tree-topology distribution networks, the number
(3c)
of line segments . Bus 0 denotes the point
of common coupling (PCC), usually at the distribution substa- where the substation voltage is fixed to be unit and is
tion and assumed to be of reference voltage. For every bus , let the voltage profile to be achieved. For example, to attain the
denote its voltage magnitude, and and denote the bus flat-voltage profile, we can set . Other choices also exist
real and reactive power injection, all in per unit (p.u.). For each depending on certain operational objectives such as better en-
line segment , let and denote its resistance and reac- ergy-efficiency as motivated by conservation voltage reduction
tance, and and the real and reactive power from bus to (CVR); see e.g., [22]. The bounds on could be due to the in-
, respectively. In addition, the subset denotes bus 's verters' apparent power limit and real power generation [1], or
neighboring buses that are further down from the feeder head. depend on the inverter power factor ratings.
The DistFlow equations [21] to model the distribution network
flow are given for every line as III. GRAPH-REPRESENTATION BASED VAR CONTROL
We will present our new LinDistFlow representation using
(1a) graph-based matrices, to better formulate the VAR control
problem. Let matrix of size denote the
graph incidence matrix for ; see e.g., ([23], pg. 6). Its
(1b)
-th column corresponds to one line segment , the
entries of which are all zero except for the -th and -th ones.
Specifically, set and when ;
i.e., bus is closer to the feeder head. Furthermore, let vector
(1c) denote the first row of which corresponds to bus 0,
while the rest of matrix denoted by of size ; i.e.,
where the nonlinear term represents the squared . Since the tree topology of ensures
line current magnitude, leading to the power loss terms in (1). that it is a connected graph, the rank of equal to
3796 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 5, SEPTEMBER 2016

= [23]. Therefore, the square matrix is of full rank , does not affect very much the power loss terms in (1a)–(1c).
and thus invertible. Hence, relatively constant loss terms can be even captured by
With the graph matrix notation and , all voltage (2c) as an operating-point related voltage profile.
in the LinDistFlow model become Remark 2 (Squared Voltage Profile): The other assumption
used for linearization relates to the squared voltage terms. How-
(4) ever, (10) can be generalized to include the original squared
voltage, instead of , as the input. To keep the squared voltage
where is an diagonal matrix with the -th diagonal difference of (1c), the graph-incidence based reformu-
entry equal to ; and similarly for which captures all 's. lation (2c) and (4) can be modified by substituting the voltage
Similarly, the power balance equations in (2a) and (2b) can be squared vector, with a scaling difference by a factor of 1/2.
respectively concatenated into This way, matrix becomes the linear sensitivity of (squared)
voltage profile due to change in . The exact form of this sen-
(5) sitivity matrix is given by [18], which is closely approximated
(6) by . Based on this modification, the VAR control problem (3)
and the ensuing analysis can be generalized accordingly to use
Solving for and and substituting them into (4) yield the squared voltage term for higher accuracy.
Remark 3 (Meshed Networks): Although the matrix LinDis-
(7) tFlow model in (10) has been derived for tree-topology net-
works, it can also be generalized to meshed networks. Since
or equivalently,
, it is the weighted graph Laplacian matrix.
This exactly coincides with the definition of the system Bbus
(8)
matrix in the popular fast-decoupled power flow (FDPF) model
where the two invertible matrices and for transmission network analysis ([24], Section 6.16). Hence,
. Since , the linear relation the model (10) and the ensuing algorithms also hold for general
from input to the output becomes distribution networks such as ring-topology systems.
Remarks 1–3 corroborate the validity of using the linearized
(9) model (10) for representing realistic distribution networks with
lossy lines, non-flat voltage profile, and even meshed topology.
where denotes the voltage profile This will be further demonstrated by numerical tests for even
under no additional VAR support. three-phase unbalanced cases in Section VI.
Proposition 1: Both and are positive definite (PD). Upon defining the voltage mismatch and rep-
Proof: Picking any non-zero vector of length and resenting by for notational convenience in the rest of the
defining , one can show that paper, the VAR control problem now becomes

(11a)

Since diagonals of are all positive resistance values, (11b)

holds for any non-zero . This way, we prove is
a PD matrix. Similarly, since all diagonals of are positive, where denotes the Euclidean norm operator. This is a
matrix is PD as well. box-constrained quadratic program (QP) and could be easily
Note that the matrix LinDistFlow model (8) has been estab- solved using off-the-shelf convex solvers. To account for the
lished by [17]. However, the linearity and positive definiteness cost of supplying VAR, a quadratic penalty term per bus will
were shown using a more complicated induction-based proof. be introduced, as motivated by the linear droop control design
Using the graph incidence matrix, the analysis here for the ma- [17]. The total cost of supplying is defined as
trix LinDistFlow model is much more straightforward. More-
over, the definitions based on suggest a very efficient way to (12)
form and , and even to model the approximation error in-
troduced by perturbed system information. Interestingly, the in-
verse of as denoted by is also a PD where the penalty coefficient and the diagonal matrix
matrix, and it is actually the power network Bbus matrix used in . Clearly, the case of no penalty on VAR
the fast decoupled power flow (FDPF) analysis; see e.g., ([24], support as in (11) can be included by setting . The reason
Section 6.16). The matrix LinDistFlow model (9) is equivalent for penalizing the VAR supply with a positive is three-fold.
to First, it will be shown in Section V.A that the popular linear
droop control design naturally comes from a positive which
(10) corresponds to the inverse of the droop slope. Second, since
larger reactive power flow could result in higher line current
Remark 1 (Power Loss Terms): Albeit an approximation of level and hence power losses, it is possible to prevent unneces-
(1), the linear relation in (10) is very meaningful as it could be sary VAR loss and decrease the overall network power losses by
considered as the sensitivity of to around any operating discouraging the net VAR injection at some locations. Last but
point. As shown by [8], [18], change of reactive power injection not least, as detailed soon in Section V, a larger tends to better
ZHU AND LIU: FAST LOCAL VOLTAGE CONTROL UNDER LIMITED REACTIVE POWER 3797

and faster stabilize the system. Thus, it would decrease the oc- . By scaling each bus with a different stepsize ,
currence and level of abrupt changes in system dynamics, and the most general GP form is
has the potential to prevent local control actions from adversely
affecting the operations of OLTC and other voltage regulating
devices. (18)
To allow for local control schemes, it turns out that the voltage
mismatch norm in (11) needs to be weighted by the PD matrix where the diagonal matrix . Clearly, the
, leading to the following surrogate VAR control problem original GP iteration (17) is a special case of (18) by setting
.
(13) Interestingly, the GP iteration (18) can be easily implemented
by setting the VAR input at all buses to be the instantaneous
. Since always holds, it is feasible to use the
where the objective
latest GP iterate as the network reactive power input by setting
at every iteration . Under this setting, the gradient
(14) direction for any becomes (cf. (9))
This surrogate problem is still convex as is PD. Because of (19)
the box constraint, the weighted error norm with would
attain a different solution compared to the original unweighted Hence, the -th entry of the gradient does not depend
problem (11). However, if every bus has unlimited VAR capa- on the full vector , but only local information on its own bus
bility (i.e., unconstrained case), the optimal solutions to both voltage and VAR input . Hence, the GP iteration (18)
error norm objectives coincide at can be completely decoupled into local updates, as given by
if . This implies that under abundant VAR resources, the
optimal solution to the weighted (13) has the potential to closely
approximate the minimum of the (11). As detailed soon, the sur- (20)
rogate problem (13) will facilitate the development of fast local
control schemes. where denotes the projection at bus to the interval ,
which is again a local computation. The proposed local VAR
IV. GRADIENT-PROJECTION METHOD control design relying on (20) is essentially equivalent to the
centralized GP solver for (13).
This section will introduce our proposed VAR control frame-
Proposition 2: Under constant , the fixed-point of the iter-
work by solving the constrained optimization problem (13). The
ative update (18), or equivalently its local counterpart (20), will
key of solving (13) lies in the separable structure of the con-
achieve the optimum to the VAR control problem (13).
straint. To project a vector to the set , one can threshold
Proof: First, existence and uniqueness of the optimum
it on each coordinate. This motivates one to adopt the gradient-
follows from the strong convexity and none-empty constraint
projection (GP) method, a generic optimization solver for con-
of (13). Furthermore, the first-order optimality condition ([20],
strained problems; see e.g., ([20], Section 2.3). The GP method
Prop. 2.1.2) for (13) implies that holds
extends the iterative first-order gradient methods to solving con-
. Thus, the necessary condition for to be optimum
strained optimization problems like (13). Upon forming the gra-
boils down to
dient direction of (14) by using as

(15) (21)
the simplest GP iteration finds the feasible direction by pro-
jecting the gradient update, as or equivalently,
(16)

where the operator thresholds any input to the constraint set (22)
, and is the stepsize. Clearly, the GP iteration (16)
boils down to the steepest descent method if the optimization would hold for any PD diagonal matrix . This is exactly the
problem is unconstrained. condition that a fixed point to (18) would satisfy.
A more general GP form takes the weighted linear combina- Remark 4 (Features of Local Control): The GP-based VAR
tion of the last iterate and the projection, given by control (20) requires each bus to measure its local voltage mag-
nitude, which can be implemented with minimal hardware up-
(17) dates. The computational requirement is very minimal since
(20) involves only scalar operations. As detailed soon, the GP
where the weighting parameter . This choice of stepsize parameters need to be chosen judiciously to ensure the
ensures that would always satisfy the box con- dynamic stability. However, they can be determined off-line
straint for any iteration , as long as the last iterate . using solely system topology and line admittance information.
Hence, every GP iterate is feasible to (13) as long as Without the network information, it is also possible to develop
3798 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 5, SEPTEMBER 2016

some parameter tuning schemes that adaptively react to poten-

tial local oscillation. In addition, even though the analytical re-
sults rely on the linearized power flow model, the local update
(20) itself can be easily implemented in any three-phase feeders
with lossy and coupling effects. Last but not least, the local VAR
control can also be updated in an asynchronous fashion with var-
ious update rates under minimal centralized coordination. This
is of particular interest for the “plug-and-play” functionality in
microgrid design. Fig. 2. Droop VAR control curve.
Albeit its simple design and robust features, the local VAR
update could suffer from performance degradation due to the
surrogate error norm objective (14). As mentioned earlier, is positive definite (cf. (24)), is a singular value of
this objective would yield the same optimum solution under , implying the scaling coefficient of (23)
unlimited VAR resources. For the case of limited but abundant
VAR resources, it is expected that the GP-based local update
could closely approximate the desired optimum of the
original problem (11). Otherwise, it is possible to quantify the
exact performance degradation from numerical simulations Assuming , having for every
with given feeder information as shown in Section VI. ensures that (23) is a contraction mapping for every . Accord-
ingly, the scaled error norm will go to 0 in the limit;
and similarly for since is PD. Hence, to ensure
V. STABLE LOCAL VAR CONTROL the stability of (20) the largest eigenvalue of needs to be less
than 2.
We will analyze the dynamic stability issue of the proposed
Under the linearized model (10), Proposition 3 coincides with
GP-based VAR control in this section. With highly variable re-
the classical dynamic system stability conditions on the eigen-
newable generation and elastic load, the stability of the local
values of the Jacobian matrix used by [3], [4]. However, our
VAR control will be crucial to cope with the system dynamics
proof based on contracting error norm could handle the projec-
at a very fast time-scale. This is closely related to the static-case
tion operator under limited VAR resources.
analysis on the convergence conditions of (20). This is equiva-
lent to having the iterative error mismatch when A. Droop VAR Control
. To this end, diagonally scale all iterates with
As advocated by the EPRI smart-inverter initiative [25] and
, and define and
IEEE 1547.8 standard [16], the droop control scheme is to scale
. Since projection is a nonexpansive mapping,
the inverter VAR output based on the instantaneous bus voltage
the scaled error norm
mismatch. For given , the GP-based local control (20) actually
generalizes this scheme by setting and ,
yielding
(25)

Such a linear droop control design has a negative slope of

and no deadband, as illustrated by Fig. 2. Based on (12), the
larger is, the more costly it is to provide VAR at bus . Ac-
cordingly, the droop slope would be smaller in magnitude to
make it less sensitive to the local voltage mismatch. As pointed
(23) out by [3], this local control scheme is prone to instability, while
[17] has proven that (25) is stable if is PD. Propo-
where the second equality holds by substituting (22). Denoting sition 3 also generalizes this result, since the droop condition
matrix , one can establish that its -th leads to
eigenvalue
(26)
(24) which is equivalent to the condition that is PD be-
cause both matrices are PD.
where the last inequality holds because , and are all If has very large eigenvalues, the local droop slope needs
PD. A sufficient stability condition is to ensure the non-negative to be very small and thus it is less sensitive to the voltage mis-
error is contracting at every iteration, which leads match. A smaller droop slope could affect the algorithmic con-
to the following proposition. vergence rate. In addition, this would lead to a large and
Proposition 3: The local update (20) is guaranteed to be higher penalty on VAR supply, which is not necessarily desir-
stable if 's largest eigenvalue . able. As pointed out by [18], matrix becomes much worse
Proof: By definition the matrix Euclidean norm conditioned if the size of network grows up or the main feeder
equals to the largest singular value of . Because matrix line gets longer. Hence, it is expected that the droop control
ZHU AND LIU: FAST LOCAL VOLTAGE CONTROL UNDER LIMITED REACTIVE POWER 3799

strategy would be more likely to be unstable for large-scale dis- To better understand the effect of , let us assume there are
tribution systems. abundant VAR resources, and thus the projection in (20) is never
active. This is exactly the scenario where instability of the droop
B. Scaled VAR Control control would emerge, as argued at the end of Section V.A.
To address the stability concerns of droop control, we Under this assumption, a closer look at the mismatch error in
will consider other designs of for local VAR control. The (23) yields
Newton's second-order method has been a very popular ap-
(30)
proach to accelerate the convergence of iterative solvers, by
scaling the gradient with the inverse Hessian of the objective; The effective Jacobian now becomes for the local con-
see e.g., ([20], Section 1.3). However, such a scaling method trol update (20). Hence, the sufficient stability condition for (20)
would be problematic when it comes to constrained problem is updated as . As , the delayed VAR
solvers with a projection operator. It can be easily shown that control is more likely to be stable. This analysis corroborates
the fixed-point of the Newton's update generally does not attain that the stability of droop control can be improved by the de-
the optimum solution of a constrained problem that it intends layed design, as pointed out by [3].
to solve ([20], Section 2.4). Hence, we will set for the
iteration (18), and scale it using the inverse of the diagonals of VI. NUMERICAL TESTS
Hessian matrix; i.e., This section presents numerical test results to demonstrate the
effectiveness of the proposed local control methods, for single-
(27)
and three-phase feeder systems. Both static and dynamic sce-
for some . The stability condition of Proposition 3 now narios on the system loading and generation will be considered.
becomes To better compare various algorithms, the desired voltage mag-
nitude is set to be unit in p.u. at every bus , along with
(28) fixed at 1. Each bus is equipped with a certain amount of
PV panels, which are able to offer flexible VAR supply to the
Compared to the droop control, the proposed scaled design feeder via effective inverter design. For the dynamic simulation
can stabilize the system dynamics for any matrix . This offers scenarios, the VAR limits are updated at every time slot
better flexibility for choosing the VAR supply penalty. In addi- based on the given inverter ratings and the instantaneous real
tion, the Hessian based diagonal scaling in (27) helps improve power generated.
the matrix conditioning and would speed up the convergence All numerical tests use the open source simulator OpenDSS
rate. More numerical simulations will be given in Section VI to [26] to solve for the actual power flow, instead of the approxi-
demonstrate this improvement. mate solution using (9). In addition, the actual bus voltage mag-
nitude, instead of the one obtained by the LinDistFlow model,
C. Delayed VAR Control is used for updating the VAR control outputs and numerical per-
With a non-unit stepsize in (20), the most general formance comparisons.
GP update takes the weighted average between the last iterate A. Single-Phase 16-Bus Radial Feeder
and the projection result. This coincides with the delayed
droop control method developed in [3], in the form of A 12 kV radial distribution feeder of 16 buses is first consid-
ered; i.e., the network in Fig. 1 with . Each line segment
(29) has the same impedance of . For the static
case, each bus has a constant load of kVA, and
where can be chosen using either the droop or the scaled abundant VAR resources of kVA. To include
control design. The work in [3] proposes this practical solution the VAR supply penalty, is chosen to be 0.2 at every bus for
to address the instability issues of droop control, along a very the proposed scaled and delayed schemes. The (delayed) droop
general stability condition. The latter can only be used to check control will used based on a linear droop curve of no
a specific distribution system with all case information given, deadband as in Fig. 2, since the voltage limits are set at [0.95,
but does not provide the exact bounds on the stepsize based on 1.05] and VAR limits at kVA.
the graph based matrices as in Proposition 3. 1) Static Scenario: Fig. 3 plots the iterative voltage mis-
Convergence of the GP method could also depend on the match error norm for the local VAR control methods.
choice of , which is not reflected by Proposition 3. As in The centralized solution corresponds to the optimum solution
([20], Section 2.3), the GP method in the form of (18) is con- to (13) with . To provide the benchmark perfor-
vergent as long as is kept constant and is selected using mance under the original unweighted objective, the matrix in
the limited minimization rule or the Armijo rule. The gist of (14) is substituted by , where the scalar is the average of
both rules is to choose adaptively to ensure sufficient de- 's eigenvalues.
cent in the objective value at every iteration. Numerical tests As for the stepsize choice, is chosen to be 0.3 for the scaled
performed in [3] demonstrate that a small and constant choice and delayed control methods, based on the stability conditions
such as , would lead to stable VAR control empiri- in Proposition 3. The stepsize is set to be constant 0.3 at
cally, even if the corresponding droop control is unstable. This every iteration for both delayed schemes. Choices of stepsize
suggests that a small could contribute to diminish the error will be discussed soon in more detail. As depicted by Fig. 3, the
norm in (23) as well. droop control fails to converge to a fixed operating condition,
3800 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 5, SEPTEMBER 2016

Fig. 4. Iterative voltage mismatch error performance for the scaled VAR con-
Fig. 3. Voltage mismatch error versus iteration index for various VAR control trol method with different values.
methods under the static system setting.

which oscillates between two operating points. This coincides

with the earlier discussion that a larger penalty coefficient due
to a steeper linear droop slope would lead to instability. By de-
laying the droop VAR update with a small , it is possible to
stabilize the droop control since the effective Jacobian's eigen-
values will become smaller. Since the VAR provision penalty
is higher for the delayed droop control, its voltage mismatch
error will be larger than the scaled and delayed ones. The scaled
control method converges very fast to the centralized solution,
same for the delayed one. The best voltage mismatch perfor-
mance for local control strategies under the surrogate VAR ob-
jective is around 0.055, as compared to the benchmark perfor-
mance around 0.031. This speaks for the competitiveness of the
proposed totally local methods, which require minimal coordi-
nation in selecting stepsize. Fig. 5. Iterative voltage mismatch error performance for the delayed VAR con-
The effect of stepsize choices on VAR updates is further in- trol method with different constant values.
vestigated, which shows a trade-off between the stability and
convergence rate. Under the same static setting, Fig. 4 plots the
voltage mismatch performance for the scaled control method stabilize the system. Nonetheless, it does not seem that the de-
using various values. Clearly, the larger is, the faster the up- layed scheme with the best would converge faster than the
date converges. However, this could also lead to potential oscil- scaled one. Therefore, the numerical tests suggest that the pro-
lation in the error performance, demonstrating potential insta- posed scaled VAR control offers very competitive convergence
bility under fast dynamics. The upper bound of is calculated to rate by tuning up the stepsize, which should be advocated for
be around 0.63 according to Proposition 3, which coincides with practical implementations.
the observation that the error oscillation would happen when The average algorithmic run time has been collected using
. The stepsize of seems to be a very good choice MATLAB® R2014a software on a typical Windows 8 com-
for the 16-bus feeder, while the analysis on the conditioning puter with a 2.6 GHz CPU. The computational time for all local
number of the Jacobian matrix also suggests that this would lead (droop, scaled, or delayed) control schemes is around 1–2 mi-
to a very good convergence rate. Hence, the stability results of croseconds per node per iteration. This numerically corroborate
Section V are very helpful to select a good value if the full the fast computation feature in Remark 4.
feeder information is available. Otherwise, it is also possible to 2) Dynamic Tests: Numerical tests under dynamic system
select the stepsize on-the-fly, by adaptively reducing the value operating conditions have also been performed. Dynamic sce-
for each inverter based on its local voltage oscillation intensity. narios include PV generation drops due to cloud coverage, and
Similar analysis has been performed for the delayed con- suddenly increasing load during large appliance start-up. To
trol scheme with different constant values, as plotted in model these, a daily residential load profile and a solar PV gen-
Fig. 5. All the error curves are based on fixing as in eration profile are used, both at minute resolution as shown by
Fig. 4. The value of is set to be constant for every iteration, Fig. 6. They are taken from a real dataset in [27], which was col-
varying from 0.01 to 0.9. Under this setting, Fig. 5 shows that lected at a residential home in the central Illinois region during
a smaller would lead to slower convergence rate. It is true a particular day in Summer 2010. Note that the profile data was
that under large values, a small would be very helpful to collected on a Friday, when house residents are more likely to
ZHU AND LIU: FAST LOCAL VOLTAGE CONTROL UNDER LIMITED REACTIVE POWER 3801

Fig. 7. Daily bus voltage profile under no VAR support.

Fig. 6. Sample daily residential load and solar PV generation profiles.

stay late in the night. This explains why the midnight load is
observed to be slightly higher than the house base load in the
early morning or afternoon. The installed solar panels have 3
kW peak capacity. The dynamic tests construct the load at every
bus to consist of 18 residential homes. Each home has the same
PV generation profile as in Fig. 6, where the inverter apparent
power limit is 5% higher than the peak capacity of 3 kW.
Fig. 7 illustrates the daily feeder voltage profile without any
VAR support. Peak voltage is observed to happen during noon
time when the load is minimal and the solar generation is at
its peak. Meanwhile, under-voltage violation (below 0.95 p.u.)
has been experienced at the end of feeder, during the evening
time with increasing power demand and decreasing PV gener-
ation. Because of its stability issue, the droop method is not
suitable for dynamic implementation. In addition, the delayed Fig. 8. Daily voltage mismatch error at every minute for the 16-bus case.
method has been omitted since it achieves the same steady-state
performance as the scaled method but at a slower convergence
rate. Hence, the proposed scaled VAR control scheme is imple-
mented here for tackling the dynamic under- and over-voltage
issues, along with the delayed droop method. The parameter and
stepsize settings follow from the static tests, while the droop
slope is time-varying based on the instantaneous VAR limits
computed from the PV generation at every minute. Both local
control schemes update every 5 seconds while the load and PV
generation stay constant within a minute. The voltage mismatch
comparison is illustrated in Fig. 8, which shows that additional
VAR control outperforms the case with no VAR support. More-
over, because of the constant setting, the scaled control
method is more effective in maintaining a flat voltage profile
than the delayed droop design, especially for the evening hours
from 18:00 to 22:00.

B. Modified 16-Bus Meshed Network

Fig. 9. Voltage mismatch error versus iteration index for various VAR control
To corroborate Remark 3 on the applicability to meshed net- methods for the modified 16-bus meshed network.
works, we have modified the 16-bus radial feeder by adding
two additional lines, one connecting nodes 12–14 and one con-
necting 13–15. All other system settings follow from the static whose performance slightly degrades from the benchmark so-
scenario in Section VI.A. Fig. 9 plots the voltage mismatch lution to the original unweighted problem. This test verifies the
error per iteration for all control schemes, which shows similar applicability of the proposed framework to networks of general
comparison results as in Fig. 3. Because the penalty coefficient topology.
is not sufficiently large, the droop control has shown
to be oscillating as well, which has been stabilized using the de- C. IEEE 123-Bus Test Feeder
layed strategy. The scaled control scheme converges very fast to To demonstrate the effectiveness of the proposed schemes for
the centralized optimum solution to the weighted problem (13), practical three-phase feeders, we have also implemented them
3802 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 5, SEPTEMBER 2016

Fig. 10. IEEE 123-bus test feeder with solar PV panel locations.

Fig. 11. Three-phase 123-bus feeder voltage profile.

on the IEEE 123-bus test case [28]. Fig. 10 shows the one-line
diagram for this distribution feeder case. In order to show the ef-
fects of inverter VAR control, the four three-phase voltage reg-
ulators are taken out of the 123-bus feeder system. In addition,
the case load information is used to determine the number of res-
idential homes connected to each load bus, while each home's
load demand and PV generation profile are same as Fig. 6. Lo-
cations of the load buses that are equipped with solar panels are
shown in Fig. 10. The feeder voltage profile for this three-phase
system under no VAR support is plotted in Fig. 11 for one time
instance, demonstrating that the system is unbalanced.
Figs. 12(a)–(c) plot the daily voltage mismatch error at all
three phases for the 123-bus case. Similar to Fig. 8 for the
16-bus case, only the scaled and delayed droop control methods
are compared to the case of no VAR support. For the scaled
control, is set to be constant at 0.01 while the delayed droop
slope is again time-varying based on the instantaneous inverter
VAR limits. Both local VAR control methods improve the
voltage support performance over the no VAR support scenario.
Because of its constant minimal VAR penalty, the scaled VAR
control scheme slightly outperforms the delayed droop method.
However, the difference between voltage mismatch error Fig. 12. Daily voltage mismatch error at (a) phase a; (b) phase b; and (c) phase
among all three scenarios is less significant compared to the c for the 123-bus case.
16-bus system in Fig. 8. This less significant voltage regulation
performance is because the 123-bus case has a lower level of
PV penetration (30%) compared to the 16-bus case (100%). voltage mismatch error by half from the baseline case of no
Accordingly, the inverter VAR output would be less effective VAR support.
in attaining the ideal constant voltage profile. Nonetheless, the
scaled VAR control becomes more effective during the evening VII. CONCLUSIONS AND FUTURE WORK
hours at higher voltage violation. As shown by the zoom-in This paper presents a general framework for developing local
view in Figs. 12(a)–(c), the scaled scheme almost reduces the VAR control methods with high penetration of distributed VAR
ZHU AND LIU: FAST LOCAL VOLTAGE CONTROL UNDER LIMITED REACTIVE POWER 3803

resources. By linearizing the distribution network power flow [15] G. Valverde and T. Van Cutsem, “Model predictive control of voltages
model, the VAR control problem is cast as one to minimize in active distribution networks,” IEEE Trans. Smart Grid, vol. 4, no. 4,
pp. 2152–2161, 2013.
the voltage mismatch error. Using the graph matrix represen- [16] IEEE Draft Recommended Practice for Establishing Methods and Pro-
tation, we formulate a weighted error minimization problem cedures That Provide Supplemental Support for Implementation Strate-
under box constraints that represent VAR limits at every bus. gies for Expanded Use of IEEE Standard 1547, IEEE P1547.8/D8, July
2014, Jun. 2014, pp. 1–176.
The gradient-projection (GP) method is evoked for solving this [17] M. Farivar, L. Chen, and S. Low, “Equilibrium and dynamics of local
constrained problem, which naturally decouples into local VAR voltage control in distribution systems,” in Proc. IEEE 52nd Conf. De-
updates requiring only the instantaneous bus voltage magnitude cision and Control (CDC), Dec. 2013, pp. 4329–4334.
information. The GP-based VAR control framework general- [18] B. Zhang, A. Dominguez-Garcia, and D. Tse, “A local control approach
to voltage regulation in distribution networks,” in Proc. 2013 North
izes existing droop and delayed droop control methods, while Amer. Power Symp. (NAPS), Sep. 2013, pp. 1–6.
allows for stability analysis to tune up the parameters based on [19] N. Li, G. Qu, and M. Dahleh, “Real-time decentralized voltage control
the network Bbus matrix. Numerical tests have been performed in distribution networks,” in Proc. 52nd Annual Allerton Conf. Com-
munication, Control, and Computing, Sep. 2014, pp. 582–588.
on single- and three-phase systems using the exact ac power [20] D. P. Bertsekas, Nonlinear Programming, 2nd ed. Belmont, MA,
flow model, which corroborate the analytical results on the per- USA: Athena Scientific, 1999.
formance guarantees of proposed methods for realistic system [21] M. Baran and F. Wu, “Optimal capacitor placement on radial distribu-
tion systems,” IEEE Trans. Power Del., vol. 4, no. 1, pp. 725–734, Jan.
implementations. 1989.
The future research plan for this work includes to investigate [22] H. J. Liu, R. Macwan, N. Alexander, and H. Zhu, “A methodology
the impact of potential asynchronous control updates among dif- to analyze conservation voltage reduction performance using field test
ferent buses, due to the potential lack of coordination [29]. We data,” in Proc. IEEE Int. Conf. Smart Grid Communications, Nov.
2014, pp. 529–534.
are also actively investigating the interactions between inverter [23] D. B. West, Introduction to Graph Theory, 2nd ed. Upper Saddle
VAR control and other voltage regulation devices, as well as River, NJ, USA: Prentice Hall, 2001.
pursuing a distributed VAR control framework which has the [24] A. J. Wood, B. F. Wollenberg, and G. B. Sheble, Power Generation,
Operation, and Control, 3rd ed. New York, NY, USA: Wiley, 2013.
potential to achieve the original unweighted error objective. [25] Common Functions for Smart Inverters, Electric Power Research In-
stitute (EPRI), Smart Inverter Working Group Recommendations, Feb.
2014 [Online]. Available: http://www.epri.com/abstracts/Pages/Pro-
REFERENCES ductAbstract.aspx?ProductId=000000000001026809
[26] OpenDSS Simulation Tool [Online]. Available: http://smartgrid.epri.
[1] K. Turitsyn, P. Sulc, S. Backhaus, and M. Chertkov, “Options for con- com/SimulationTool.aspx
trol of reactive power by distributed photovoltaic generators,” Proc. [27] “Individual household electric power consumption dataset,” UCI Ma-
IEEE, vol. 99, no. 6, pp. 1063–1073, Jun. 2011. chine Learning Repository [Online]. Available: http://archive.ics.uci.
[2] P. Carvalho, P. F. Correia, and L. Ferreira, “Distributed reactive power edu/ml/
generation control for voltage rise mitigation in distribution networks,” [28] IEEE PES Distribution Test Feeders, Sep. 2010 [Online]. Available:
IEEE Trans. Power Syst., vol. 23, no. 2, pp. 766–772, May 2008. http://www.ewh.ieee.org/soc/pes/dsacom/testfeeders.html
[3] P. Jahangiri and D. Aliprantis, “Distributed Volt/VAR control by pv [29] H. Zhu and N. Li, “Asynchronous local voltage control in power distri-
inverters,” IEEE Trans. Power Syst., vol. 28, no. 3, pp. 3429–3439, bution networks,” in Proc. 41st IEEE Int. Conf. Acoustics, Speech and
Aug. 2013. Signal Processing (ICASSP), Mar. 2016, submitted for publication.
[4] B. Robbins, C. Hadjicostis, and A. Dominguez-Garcia, “A two-stage
distributed architecture for voltage control in power distribution sys-
tems,” IEEE Trans. Power Syst., vol. 28, no. 2, pp. 1470–1482, May
2013.
[5] American National Standard for Electric Power Systems and Equip-
ment- Voltage Ratings (60 Hertz), ANSI C84.1-2011 Std. Hao Zhu (M’12) received her B.S. from Tsinghua
[6] M. Farivar, R. Neal, C. Clarke, and S. Low, “Optimal inverter var con- University, Beijing, China in 2006, and the M.Sc.
trol in distribution systems with high pv penetration,” in Proc. IEEE and Ph.D. degrees from the University of Minnesota,
Power and Energy Soc. General Meeting, Jul. 2012, pp. 1–7. Minneapolis in 2009 and 2012, respectively.
[7] E. Dall'Anese, H. Zhu, and G. Giannakis, “Distributed optimal power She is currently an Assistant Professor of ECE
flow for smart microgrids,” IEEE Trans. Smart Grid, vol. 4, no. 3, pp. at the University of Illinois, Urbana-Champaign
1464–1475, Sep. 2013. (UIUC). She worked as a postdoctoral research
[8] P. Sulc, S. Backhaus, and M. Chertkov, “Optimal distributed control associate on power system validation with the
of reactive power via the alternating direction method of multipliers,” Information Trust Institute at UIUC from 2012 to
IEEE Trans. Energy Convers., vol. 29, no. 4, pp. 968–977, Dec. 2014. 2013. Her current research interests include power
[9] B. Zhang, A. Lam, A. Dominguez-Garcia, and D. Tse, “An optimal and system monitoring and operations, dynamics and
distributed method for voltage regulation in power distribution sys- stability, distribution systems, and energy data analytics.
tems,” IEEE Trans. Power Syst., vol. 30, no. 4, pp. 1714–1726, Jul. Dr. Zhu is currently the IEEE Signal Processing Society's Representative on
2015. the IEEE TRANSACTIONS ON SMART GRID, and a steering committee member
[10] V. Kekatos, G. Wang, A. Conejo, and G. Giannakis, “Stochastic reac- for IEEE SMART GRID.
tive power management in microgrids with renewables,” IEEE Trans.
Power Syst., vol. 30, no. 6, pp. 3386–3395, Nov. 2015.
[11] H.-G. Yeh, D. Gayme, and S. Low, “Adaptive var control for distribu-
tion circuits with photovoltaic generators,” IEEE Trans. Power Syst.,
vol. 27, no. 3, pp. 1656–1663, Aug. 2012. Hao Jan (Max) Liu (S’12) was born in Taipei,
[12] A. R. Bergen and V. Vittal, Power System Analysis, 2nd ed. Upper Taiwan. He received his B.S. from Missouri Univer-
Saddle River, NJ, USA: Prentice Hall, 2000. sity of Science and Technology, Rolla, MO in 2011.
[13] Q. Zhou and J. Bialek, “Generation curtailment to manage voltage con- He is currently a Ph.D. candidate of ECE at the
straints in distribution networks,” IET Gener., Transm., Distrib., vol. University of Illinois, Urbana-Champaign (UIUC).
1, no. 3, pp. 492–498, 2007. His current research interests cover the areas of op-
[14] M. Kraiczy, M. Braun, G. Wirth, S. Schmidt, and J. Brantl, “Interfer- timization for power distribution networks and smart
ences between local voltage control strategies of a HV/MV transformer grid technology, including Volt/VAR and conserva-
and distributed generators,” in Proc. 28th Eur. PV Solar Energy Conf. tion voltage reduction studies.
Exhib., Sep. 2013.