This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON SMART GRID

A Fully Distributed Reactive Power Optimization

and Control Method for Active
Distribution Networks
Weiye Zheng, Wenchuan Wu, Senior Member, IEEE, Boming Zhang, Fellow, IEEE,
Hongbin Sun, Senior Member, IEEE, and Yibing Liu, Student Member, IEEE

AbstractThis paper presents a fully distributed reactive

power optimization algorithm that can obtain the global optimum
solution of nonconvex problems for distribution networks (DNs)
without requiring a central coordinator. Second-order conic
relaxation is used to achieve exact convexification. A fully distributed second-order cone programming solver (D-SOCP) is
formulated corresponding to the given division of areas based
on the alternating direction method of multipliers (ADMM)
algorithm, which is greatly simplified by exploiting the structure of active DNs. The problem is solved for each area with
very little interchange of boundary information between neighboring areas. D-SOCP is extended by using a varying penalty
parameter to improve convergence. A proof of its convergence
is also given. The effectiveness of the method is demonstrated
via numerical simulations using the IEEE 69-bus, 123-bus DNs,
and a real 1066-bus distribution system.
Index TermsActive distribution networks (ADNs), alternating direction method of multipliers (ADMM), distributed reactive
power optimization, second-order conic (SoC) relaxation.

N a

Ea
Ga
Mj , Mj
O
Xa
Variables
PGi , QGi
QCi
Pij , Qij
Vi , vi

N OMENCLATURE
Indices and Sets
i, j, k
a, b
t
u
N
E
ET
G
Na

Indices of buses, from 1 to N.

Indices of areas, from 1 to R.
Iteration index for distributed solvers
(inner loop).
Regulation index for voltage regulators
(outer loop).
Set of buses in the entire system.
Set of edges in the entire system.
Set of edges of a spanning tree selected from
the graph {N, E}.
Set of generators and reactive power (VAR)
compensators in the entire system.
Set of buses in area a.

Manuscript received April 23, 2014; revised October 11, 2014 and
December 8, 2014; accepted January 22, 2015. This work was supported
in part by the National Key Basic Research Program of China under
Grant 2013CB228205, in part by the National Science Foundation of
China under Grant 51477083, and in part by New Century Excellent Talents
in University under Grant NCET-11-0281. Paper no. TSG-00340-2014.
The authors are with the State Key Laboratory of Power Systems,
Department of Electrical Engineering, Tsinghua University, Beijing 100084,
China (e-mail: wuwench@tsinghua.edu.cn).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSG.2015.2396493

Extended set of buses in area a comprising N a and the other end of the tie-lines in
neighboring areas.
Set of edges in area a.
Set of generators in area a.
Set of areas that contain bus j and its
cardinality.
Set of buses shared by different areas.
Feasible region of state vector xa in area a.

Iij , lij
V
Pa,Gi , Qa,Gi
Pa,i , Qa,i
Qa,Ci
Pa,ij , Qa,ij
Va,i , va,i
Ia,ij , la,ij

Injected active and reactive powers of generators at bus i.

Total reactive power from compensation
devices at bus i.
Active and reactive powers at the sending end
of branch ij.
Voltage magnitude at bus i and the square
thereof.
Current magnitude at branch ij and the square
thereof.
Voltage adjustment due to voltage regulation.
Injected active and reactive powers of generators at bus i in area a.
Total injected active and reactive powers at
bus i in area a.
Total reactive power from compensation
devices at bus i in area a.
Active and reactive powers at the sending end
of branch ij in area a.
Voltage magnitude at bus i in area a and the
square thereof.
Current magnitude of branch ij in area a and
the square thereof.

Vectors
x
sj
xa
xa,j
a,j
rt

Global state vector for the entire system.

Global state vector at bus j.
Local state vector in area a.
Local copy of global state vector sj in area a.
Local Lagrangian multiplier vector corresponding to bus j in area a.
Global primal residual vector for whole system in the tth iteration.

c 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
1949-3053 
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2

IEEE TRANSACTIONS ON SMART GRID

dt
rta
dta

Global dual residual vector for whole system

in the tth iteration.
Local primal residual vector for area a in the
tth iteration.
Local dual residual vector for area a in the tth
iteration.

Functions
fa (), ga (),
Objective function, inequality constraints and
ha ()
equality constraints in area a.
L()
Global augmented Lagrangian function.
Local augmented Lagrangian function in
La ()
area a.
Parameters
t
PDj , QDj
rij , xij
gj , bj

Vref
Vc max , Vc min
Vn max , Vn min
t
at
uta,j
t ,
P ta,j , Q
a,j
t ,
P ta,ij , Q
a,ij
t
v ta,j , la,ij

P Gi

Gi
QGi , Q
V i , V i
P ij , Pij
ij , Q
Q
ij
Iij

Border residual in the tth iteration, as defined

in (26).
Active and reactive power load demands at
bus j.
Resistance and reactance of branch ij.
Aggregated shunt conductance and susceptance of equivalent parameters for capacitor
banks, lines and transformers on bus j.
Voltage magnitude at the point of common
coupling (PCC).
Highest and lowest voltages across all buses.
Maximum and minimum allowable voltages
across all buses.
Global penalty parameter in the tth iteration of fully distributed second-order cone
programming (D-SOCP) solver.
Local penalty parameter in area a in the tth
iteration of varying penalty parameter (VPP).
Local auxiliary vector for bus j in area a in
the tth iteration.
Detailed entries in auxiliary vector uta,j .

Predefined tolerance for convergence.

Upper bound of injected active powers of
generators at bus i.
Lower and upper bound of injected reactive
powers of generators at bus i.
Upper bound and lower bound of voltage
magnitude at bus i.
Upper bound and lower bound of active power
at the sending end of branch ij.
Upper bound and lower bound of reactive
power at the sending end of branch ij.
Upper bound of current magnitude of
branch ij.
I. I NTRODUCTION

HE OPTIMIZATION of distribution networks (DNs)

has conventionally been solved in a centralized manner
using a distribution management system (DMS). However,

in future smart grid systems, an increasing number of

distributed generators (DGs) comprising wind power and photovoltaic (PV) generation units are expected to be integrated
into the DN, forming active DNs (ADNs) [1][3]. Because of
the distributed nature of the active components of the network,
centralized algorithms for ADNs are expected to encounter
significant technical challenges.
The volume of data that arises as a result of large-scale
integration of DGs is expected to lead to a communication
bottleneck in the near future because the central coordinator
must collect and process all of this information. Additionally,
the size of the problem makes it very computationally expensive for the centralized DMS to solve [4]. Furthermore, the
regional distribution system may belong to various owners, and
it is thus problematic to implement a centralized optimization
because of privacy and security concerns [5].
Another important issue is robustness in terms of the cyberphysical system. The central coordinator becomes a vulnerable
target for both cyber and physical attacks. If the central coordinator fails for some reason, the entire system will fail [5].
One of the key operating tasks of ADNs is to maintain the
network parameters, including the bus voltages and branch
power flows, within predefined limits. However, a high penetration of DGs may result in overvoltage at ADNs due to
reverse power flow [6]. Therefore, reactive power optimization, which is closely related to optimal power flow (OPF)
problems, is used to minimize network power losses and
prevent voltage violations by dispatching reactive power control devices [7]. Considering the significant challenges posed
by a centralized DMS, a fully distributed reactive power
optimization algorithm is preferred.
Unfortunately, the reactive power optimization problem
is essentially nonconvex and NP-hard. Furthermore, directly
applying distributed algorithms to nonconvex problems does
not have guaranteed convergence. Therefore, various convexification or linearization techniques have been proposed to
reduce the complexity of the problem. Because the original nonconvex problem stems from the nonlinear power flow
equations, simplification techniques have mainly focused on
the power flow constraints. The dc power flow approximation linearizes the constraints and makes the problem easier
to solve [8], [9]; however, accuracy is not guaranteed because
this may not be a good approximation for ADNs. Semi-definite
relaxation may be used to solve a convex Lagrangian dual
problem instead of the primal nonconvex one [10]; this method
works in radial networks, but it involves additional large-scale
variables and cannot be extended to applications in meshed
networks [11]. Second-order conic (SoC) relaxation is based
on a branch flow model and is a promising convexification
method. It was first proposed by Jabr [12] and was further
developed to deal with radial networks [13] and extended to
meshed networks [14].
If the nonconvex problem is transformed into a convex
problem, it is relatively easy to solve in either a centralized
or a distributed manner. The alternating direction method of
multipliers (ADMM) is a promising distributed algorithm [15]
that has been applied to state estimation (SE) [16] and OPF
problems [17] in power systems. Kekatos and Giannakis [16]

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ZHENG et al.: FULLY DISTRIBUTED REACTIVE POWER OPTIMIZATION AND CONTROL METHOD FOR ADNs

described a distributed algorithm for robust SE; however,

the solutions deviate from the optimum of the original
problem due to the poor dc power flow approximation.
Sun et al. [17] described an ac OPF solution in a fully distributed manner; however, only one generator was considered.
A distributed semi-definite programming (SDP) solver for
OPF problem with a bus injection model in microgrids has
been reported [18]; however, the positive semidefiniteness of
a global matrix reflects the coupling of all buses in the systems,
which is very hard to decompose.
Simulation results have been reported [18] that reveal
an interesting and universal property, namely that the performance of ADMM relies on the value of the penalty
parameter. Self-adaptive tuning techniques have been used to
deal with monotone variational inequality problems [19], [20].
Furthermore, varying the penalty parameter to deal with the
general optimization problem has been reported [15]; however,
there was no proof of convergence and cannot directly be used
in distributed manner.
In this paper, we use the SoC relaxation technique described
in [13] and propose a D-SOCP for reactive power optimization problems in ADNs based on ADMM, in which we deploy
a VPP tuning technique that does not require any central coordinator to accelerate convergence. The major contributions are
as follows.
1) Most pioneering works apply ADMM to nonconvex OPF
problems directly; however, convexity is a prerequisite
for convergence of the ADMM algorithm. Using a few
additional auxiliary variables, SoC relaxation can be
used to transform the nonconvex ADN problem into
a convex second oder optimization program (SOCP)
model exactly for radial networks.
2) ADMM itself does not guarantee a fully distributed
solution. In this paper, an ADMM-based D-SOCP for
reactive power optimization problem is achieved by
eliminating global variables and global updates. Besides,
a revised iterative procedure of ADMM is proposed by
considering the features of DNs. Solving SOCP in a distributed fashion is more practical than its distributed
SDP counterpart for two reasons: first, the conic constraints only reflect couplings among neighboring buses
rather than those among all buses in the system, which
means that SOCP is easier for exact decomposition; second, the distributed SDP counterpart only applies to
radial networks, whereas D-SOCP can handle both radial
and meshed networks. Furthermore, SDP relaxation may
lead to unrealistic results in some cases [21]. Note that
mixed variables (branch powers and the voltage at buses)
for SoC relaxation of the branch flow equations also
make the distributed modeling more challenging.
3) The performance of the ADMM-based distributed algorithms relies on the choice of the penalty parameter.
VPP tuning techniques in existing literatures cannot
work without central coordination [19], [20]. Therefore,
a fully distributed tuning technique is proposed to accelerate convergence of the D-SOCP solver as a refinement.
The remainder of this paper is organized as follows.
Section II describes the original nonconvex reactive power

Fig. 1.

Schematic of the distflow branch equations.

Fig. 2.

Scheme of the reactive power optimization with voltage regulation.

optimization problem for ADNs in detail. Section III details

the SoC relaxation technique. In Section IV, we propose
an ADMM-based D-SOCP solver, and in Section V, we
extend D-SOCP by using a VPP to improve performance with
arbitrary initial values for penalty parameters. Numerical simulations are described in Section VI, and concluding remarks
are given in Section VII. The Appendix provides derivations
of the proposed distributed algorithms.

II. N ONCONVEX ADN O PTIMIZATION P ROBLEM

Consider an ADN with N buses, in which the bus set is
represented by N := {1, . . . N}, where 1 denotes the PCC. The
branches are denoted as the edge set E := {(i, j)} N N,
and set of generators and VAR compensators is represented
by G. {N, E} should be regarded as a directed graph, and the
following reference orientation is adopted. Select any spanning tree {N, ET } of the graph rooted from the PCC; all
tree branches, e.g., ET , are in the direction away from the
root, and the link branches, e.g., E\ET , may be in an arbitrary direction. We denote a branch as i j if it points
from bus i to bus j.
As shown in Fig. 1, a radial network was used as an example
to illustrate the Distflow branch equations [22], [23].
For DN, especially in China, the transformers in substation
are critical facilities to regulate voltage profiles by adjusting
their tap-changers. A reactive power optimization scheme with
voltage regulation for ADNs can be formulated as shown in
Fig. 2, which comprises two steps including reactive power
optimization in feeders and voltage regulation in substations.
These two steps can be executed either separately or coordinately. The technical details will be explained in the following
sections.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4

IEEE TRANSACTIONS ON SMART GRID

A. Reactive Power Optimization Model for Feeders

1) Objective Function: We aim to minimize the active
power loss of the network, that is

PGi
(Model 1)
min
QGi ,QCi

subject to

iG

Equations (1)(5)

where QCi is the total reactive power from the compensation

devices, such as static VAR compensators or capacitor banks,
at bus i, and PGi , QGi are the injected active and reactive
powers of the DGs (or other generators) at bus i.
2) DistFlow Branch Equality Constraints: We have



(Pij )2 + (Qij )2 r + P g V 2 = 

P
ij
ij
j
j j
i: ij
k: jk Pjk
(Vi )2

2
2


P + Q
Qij ( ij )(V )(2 ij ) xij + Qj bj Vj2 = k: jk Qjk , j N
i: ij
i

2 2


V 2 = V 2 2
rij Pij + xij Qij  + r2 + x2 Pij + Qij
j

ij

ij

Vi2

(1)
and

PGj PDj , j G
j N\G
PDj ,

QGj + QCj QDj , j G
Qj =
j N\G
QDj ,
Pj =

(2)

where rij , xij denote the resistance and reactance of branch ij,
respectively; Pij , Qij are the active and reactive powers, respectively, at the sending end of each branch ij; Pj , Qj are the
injected active and reactive powers at bus j; PDj , QDj are the
injected active and reactive powers at the load; and Vj denotes
the voltage magnitude at bus j. For brevity, G denotes the
set of both generators and VAR compensation devices. Lines
are modeled as -equivalent components, and transformers
are modeled as series components with transmission parameters that depend on the connection type [24]. Connections
to earth have been modeled as aggregated shunt conductance
gj and susceptance bj of equivalent parameters for capacitor banks and transformers on bus j. Since transformers have
been included, the method can be also applied to distribution
systems with multivoltage levels.
3) Reference Voltage of the PCC: Let
V1 = Vref

(3)

where Vref is a voltage magnitude controlled by the substation side.

4) Security Constraints: We have
V i Vi V i ,
 
Iij Iij
Pij Pij P ij,
Q Q Q
ij
ij

i N
(i, j) E

(4)

ij

where |Iij | denotes the current magnitude of branch ij; Iij is

the upper limit of branch current magnitude for branch ij;
ij represent the lower and upper bounds of
and Pij , P ij , Qij , Q
Pij and Qij , respectively; V i and V i are the lower and upper
bounds of the voltage magnitude at bus i, respectively.

Fig. 3.

Equivalent network of a substation.

5) Operating Constraints of DG and Compensation

Devices: Suppose that the output of a DG follows the signal
from maximum power point tracking. Therefore, all DGs generate the maximum allowable active power, and the reactive
power can be controllable and optimized.
We have

PGi = P Gi
Gi ,
Q QGi Q
i N
(5)
Gi

QCi QCi QCi

where P Gi is the upper bound of the injected active powers of
Gi are the lower and upper bounds
the DG at bus i, QGi and Q
of the injected reactive powers of the DG at bus i, respectively,
Ci are the lower and upper bounds of the reactive
and QCi , Q
power from the compensation devices at bus i, respectively.
B. Voltage Regulation in Substation
As shown in Fig. 1, the voltage regulation is implemented
by transformer in substation. Based on the reactive power
optimization results from the feeders, we can obtain the highest and lowest voltages over buses in the uth regulation
(i.e., outer loop), denoted Vcumax := sup{Vi |i N} and
Vcumin := inf{Vi |i N}; however, according to the predefined
limit and requirement for voltage quality, the maximum and
minimum allowable voltages over all buses can be calculated
using Vn max := sup{V i |i N} and Vc max := sup{V i |i N}.
We define the voltage gap as
Vu
+ Vcumin
Vn max + Vn min
c max
.
(6)
2
2
Fig. 3 shows the equivalent network of a substation, where
Z = Z T /k (k 1)//(jXC ), Z = Z T /k + Z eq and where k
is the transformer ratio of the voltage regulator, and Z T is the
impedance of the voltage regulator. Z eq , V0 are the Thevenin
equivalent parameters, and XC denotes the shunt reactance at
bus 1.
The voltage at the PCC can be regulated according to the
following expression:
V u =

V1u+1 := V1u + V u

(7)

and the active and reactive powers supplied by the substation

satisfy


 
V 0 V 1
V 1
V 1
V 1
P = Re
(8)
Z
Z

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ZHENG et al.: FULLY DISTRIBUTED REACTIVE POWER OPTIMIZATION AND CONTROL METHOD FOR ADNs

Equation (12) can be rewritten in a standard SoC form as



 2Pij 


 2Qij  lij + vi
(13)


 lij vi 
2
Fig. 4. Information exchange between voltage regulation in substation and
reactive power optimization on feeders in the uth iteration.

thus making it possible to obtain the global optimum of the

nonconvex Model 1 by solving the following SOCP problem:

(Model 3) min
PGi
(14)
QGi ,QCi

and


Q = Im

V 0 V 1
Z

V 1

V 1
Z

subject to

V 1

(9)

iG





Pij lij rij + Pj gj Vj2 =
Pjk
i: ij

respectively.
The information exchange between voltage regulation in
substation and reactive power optimization on feeders in the
uth iteration is figured out in Fig. 4. With the regulation
described by (6) and (7), the range of voltage magnitudes of
all buses is expected to lie in the middle of the operational
interval, which ensures a wide margin for voltage magnitude
deviations for security concerns.

k: jk



Qij lij xij + Qj bj Vj2 =
Qjk j N

i: ij

vj = vi 2 rij Pij + xij Qij

k: jk


+ rij2 + xij2 lij ,

and (2) (5), (10), and (13).

(15)

Note that Model 3 is almost identical to Model 2 except

for the constraint in (13), which is relaxed from the constraint
in (11) in Model 2. This SOCP relaxation is exact for radial
networks [13].

III. SOCP R ELAXATION

Note that Model 1 is nonconvex and hard to solve.
Furthermore, distributed optimization algorithms based on
dual ascent typically do not guarantee convergence when
applied to general nonconvex problems. Thus, here we use
SoC relaxation to transform it into a convex problem [13],
which allows for an efficient computational solution as well as
subsequent distributed algorithms with a proof of convergence.
SOCP problems are typically solved using convex
programming [26], and a global optimal solution can be found
using interior point methods in polynomial time [27].
Model 1 can be rearranged by introducing two new variables
Vi and Iij , which are defined as follows:
vi = Vi2 ,

i N

(10)

IV. F ULLY D ISTRIBUTED SOCP S OLVER

Suppose the DN is divided into R areas. Let N a , Ea , Ga
represent the bus set, edge set, and generator/VAR compensator set in area a, respectively. Also, let sj :=
{Pj , Qj , Pij , Qij , vj , lij |i : i j} be a global state vector
corresponding to bus j. For any area a, considering the interconnections with other areas, the bus set N a is extended to the
set N a , which comprises both N a and the other end of tie lines
among as neighboring areas. The vectorxa := {xa,j |j N a } is
then introduced to describe the local state variables in area a
with the vector xa,j , which represents a local copy of sj in
area a. Model 3 can now be restated as follows:
(Model 4)

min

R


fa (x)

a=1

and

2
2
Pij + Qij
2
lij = Iij =
,
(Vi )2

subject to
(i, j) E.

(11)

As a result, Model 1 can be restated as follows:


(Model 2)
min
PGi
QGi ,QCi

subject to

iG

Equations (15), (2)(5), and (10)(11)

Proposition 1 (SoC Relaxation): The following relaxation

of the equality given in (11) is exact:

2
2
Pij + Qij
lij
,
(i, j) E
(12)
(Vi )2
if the following two conditions are satisfied.
1) The objective function increases with lij .
2) There are no upper bounds for the loads.
Proof: Refer to [13] and [14].

ga (x) 0
ha (x) = 0,
x x x
)T

a = 1, 2, . . . , R
(16)

where x := (x1 , . . . , xR is a global state vector that collects

all copies in each area. For simplicity, all inequality constraints
in area a are denoted by the vector function ga (x), which
is nonnegative, whereas the equality constraints are denoted
by the vector function ha (x), which is equal to 0, and fa (x)
represents the objective function corresponding to area a.
Let the set Mj := {a| j N a } contain all indexes of the
areas that contain bus j with cardinality denoted as Mj := |Mj |.
Clearly, the global entry sj should be duplicated in all local
states xa if area a belongs to the set Mj . Also note that Mj > 1
means that more than one area overlap at bus j, and let the
overlapping bus set O := { j|Mj > 1} contain all such buses.
With the example shown in Fig. 5, suppose that the bus set
supervised by area 1 is N 1 = {0 3, 18 21, 22 24}, as
shown by the red solid curve in the figure. However, when

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6

IEEE TRANSACTIONS ON SMART GRID

where R+ is the penalty parameter for the ADMM

algorithm.
Note that the global Lagrangian can be decoupled spatially,
that is
R

   

 
(21)
L {xa }, {sj }, {a,j } =
La xa , sj , a,j
a=1

Fig. 5. Schematic of the area decoupling scheme. The colored solid lines
cover the set of buses in each area, and the dotted rectangles denote the
extended set of buses in each area.

considering the interconnection with other areas, e.g., area 2,

the bus set of area 1 must be extended to include the border
of area 2, e.g., bus 4, so that the bus set of area 1 is extended
to N 1 = {0 3, 18 21, 22 24, 4}, as shown by the pink
area in Fig. 5. Similarly, the bus set of area 2 was initially
N 2 = {4 7, 25 26}; however, it was later extended to
N 2 = {4 7, 25 26, 3, 8}, which includes the border of its
neighboring areas. We then have N 1 N 2 = {3, 4}; it follows
that buses 3 and 4 are shared by areas 1 and 2. Therefore,
M3 = {1, 2}, M3 = 2, and O := {3, 4, 7, 8}.
For any bus j that is shared by two neighboring
areas, e.g., a and b, the local copies of state variables,
e.g., xa,j and xb,j should be all driven to the same value as
the global state variable sj . For the system shown in Fig. 5,
we have x1,3 = s3 , x2,3 = s3 . However, such constraints are
interesting only for bus j, which belongs to the overlapping
set O, as regional coupling exists only in this case. Model 4
may, therefore, be transformed equivalently to Model 5, which
can be expressed in terms of the local variable xa and global
variable sj , that is
(Model 5) min

R


fa (xa )

(17)

a=1

ga (xa ) 0
ha (xa ) = 0
xa xa x a
xa,j = sj ,

a = 1, 2, . . . , R
(18)
a, j O N a .

(19)

The novelty here comes from solving Model 5 in a fully

distributed manner using ADMM [28]. Let a,j denote the
Lagrangian multiplier vector corresponding to the equality
constraints in (19), and suppose that the constraints in (18)
define a feasible set X a for area a; the augmented Lagrangian
of Model 5 can now be written as follows:

R
 


  

fa (xa ) +
Ta,j xa,j sj
L {xa }, {sj }, a,j =
a=1

jON a




2
+ xa,j sj 2 , xa X a , a
2
(20)

with a local Lagrangian for each area a defined as

   
La xa , sj , a,j
 




xa,j sj 2
:= fa (xa ) +
Ta,j xa,j sj +
2
2
jON a

xa X a .

(22)

Let t denote the iteration index. The standard ADMM

iterations consist of the following three steps:
    
xa , stj , ta,j
(23)
=
arg
min
L
xt+1
a
a
xa X a



   t 
st+1
= arg min La xt+1
(24)
a , sj , a,j
j
{sj |jO}

and


t+1
t+1
t
t+1
,
=

s
a,j
a,j
a,j
j

j O N a . (25)

The termination criteria may be determined by considering

the primal residual column vector reflecting the primal feasibility during the tth iteration. This can be defined as rt , which
describes {|xta,j stj ||a, j ON a } for all areas in a column.
The dual residual column vector reflecting dual feasibility can
a
be defined as dt , which describes {|stj st1
j ||a, j O N }
for any overlapping bus j [15]. Given a feasibility tolerance
> 0, a reasonable criterion for termination is that the primal
and dual residuals must be small, i.e., their maximum element
must be less than a predetermined tolerance , where
 t 
 r 
t

(26)
=
 dt  < .

However, directly applying ADMM does not lead to a fully

distributed algorithm. Although the update of local primal and
dual optimization vectors in (23) and (25) can be executed by
all areas in parallel, (24) updates the global vector sj for any
overlapping bus j, which necessitates that the control center
store the global data and execute a global update.
The major adaption here is to implement fully distributed
optimization by eliminating the global vectors {sj } and the
global update in (24) from the standard ADMM iterations.
The following two lemmas are useful for our discussion.
Lemma 1 (Principle of Area Division): Each area is considered as a node of a macroscopic graph. Assume that the
macroscopic graph is a tree with the PCC as the root. Then,
regardless of the initial division of the radial distribution here
denoted as {N0a |a}, each bus set N a can be set up with a minor
modification of N0a , and we can guarantee that, with the partition {N a |a}, any overlapping bus will be shared by only two
areas, that is

2, j O
Mj =
(27)
1, j N\O.
Proof: See the Appendix.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ZHENG et al.: FULLY DISTRIBUTED REACTIVE POWER OPTIMIZATION AND CONTROL METHOD FOR ADNs

Lemma 2 (Complementary Lagrangian Multipliers): From

Lemma 1, any overlapping bus j belongs to a unique pair of
neighboring areas a and b. The Lagrangian multipliers updated
by (25) then satisfy ta,j + tb,j = 0 for each iteration t.
Proof: See the Appendix.
Proposition 2 (D-SOCP Solver): A local auxiliary vector
is introduced for any area a corresponding to Lagrangian
multipliers, that is

t+1
t
ut+1
a,j = ua,j + xb,j

b = Mj \a.

xta,j

jON a
+ xtb,j

j O N a
(30)

To be more specific, for the reactive power optimization

problem, we have xa := {xa,j |j N a } = {Pa,j , Qa,j , Pa,ij ,
Qa,ij , va,j , la,ij |j N a }, where variables with subscript
a are local variables that belong to area a. Accordingly,
t , P t ,
uta,j can be expressed element-wise as uta,j = (P ta,j , Q
a,j
a,ij
t
t
t
T

Q , v , l ) , and (29) is equivalent to solving the followa,ij

a,j

{Qa,Gj ,Qa,Cj |jN a } iGa

 

jON a

Pa,j P ta,j

lim t = 0

2 
2

ta,ij
+ Pa,ij P ta,ij + Qa,ij Q
2 
2 

t
t

(31)
+ va,ij v a,ij + la,ij la,ij

k:jk



Qa,ij la,ij xij + Qa,j bj vj =
Qa,jk

i: ij

j N a

k: jk

va,j = va,i 2 rij Pa,ij + xij Qa,ij +


Pa,Gj PDj , j Ga
Pa,j =
PDj ,
j N a \Ga

Qa,Gj + Qa,Cj QDj , j Ga
Qa,j =
QDj ,
j N a \Ga
a
V1 = Vref ,
if 1 N
2
2
V va,i V i ,
i N a
i
2

lij Iij
Pij Pa,ij P ij ,
(i, j) Ea

Q Qa,ij Q
ij
ij
rij2

+ xij2

lim

subject to




Pa,ij la,ij rij + Pa,j gj vj =
Pa,jk

xa ,

(38)
a = 1, 2, . . . , R

(39)

and

2

+ Qa,j Qta,j

i: ij

(37)

For initialization, the state variables xa can be set to arbitrary values x0a and auxiliary variables u0a,j = x0a,j + x0b,j /2.
The termination criteria can be determined by considering
the primal residual vector in the tth iteration rt , which collects {|xta,j xtb,j /2|} in each column, and the dual residual
t1
vector dt , which collects all {|(xta,j + xtb,j ) (xt1
a,j + xb,j )/2|}.
Proof: See the Appendix.
Proposition 3 (Proof of Convergence): If the penalty parameter is bounded, e.g., < +, then the iteration described
by (29) and (30) in Proposition 2 will approach both feasibility
and optimality, that is

lim xt
t a
2

(36)

a,ij

ing subproblem for each area a:


min
Pa,Gi

Schematic of the fully distributed algorithm.

Pa,Gi = P Gi
Gi ,
Q Qa,Gi Q
i N a
Gi
Ci
Q Qa,Ci Q
 Ci

 2Pa,ij



 2Qa,ij
 la,ij + va,i ,
(i, j) Ea .


 la,ij va,i 

ta,j

,
a, j O N a .
(28)

By exploiting the feature of radial DNs stated in Lemma 1

and the problem structure stated in Lemma 2, the ADMM
algorithm described by (23)(25) can be significantly simplified for any area a as follows:
2
 


= arg mina fa (xa ) +
xt+1
xa,j uta,j  (29)
a
xa
2
2
uta,j := stj

Fig. 6.

(32)
la,ij ,
(33)

(34)

(35)

R



fa xta = p

(40)

a=1

where xa and p are the optimal solutions and optimum of

centralized Model 5, respectively.
Proof: Since the adaption in Proposition 2 is a variable substitution and equivalent transformation, the convergence proof
of standard ADMM remains valid (see [15, Appendix A]).
The fully distributed algorithm using (29) and (30) can
be implemented in parallel manner. The local subproblems
described by (31)(37) can be solved for all areas simultaneously, and then the local auxiliary variable ua,j in (30) can
be updated using the latest boundary data. For illustration,
Fig. 6 shows the distributed optimization process for two areas.
V. D ISTRIBUTED P ENALTY PARAMETER T UNING
From Proposition 3, convergence of the D-SOCP algorithm
is guaranteed in theory. However, in practice, if the penalty
parameter is not appropriately chosen, convergence may be
slow. An optional extension of ADMM-based algorithms is
to vary the penalty parameter t depending on the (real-time)
conditions during each iteration t (see [15, Sec. 3.4.1]). It can
be very difficult to provide a proof of convergence for the
tuning scheme given in [15, Sec. 3.4.1] because may vary
in an unbounded manner, so Proposition 3 no longer applies.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
8

IEEE TRANSACTIONS ON SMART GRID

A self-adaptive tuning technique was reported [20] for solving monotonic variational inequality problems. However, this
kind of tuning technique requires a central coordinator to
gather the primal and dual residuals from all areas, and it
cannot be used in distributed manner.
Here, we propose a fully distributed penalty parameter tuning technique and it is used to improve the performance of
ADMM-based D-SOCP solver.
The penalty parameter tuning scheme is as follows:
t
 
 
a

if rta 2 < dta 2 and t tmax

1+ ,
 t
 
d  < rt  and t tmax
at+1 = (1 + )at ,
if
a
a 2
2

t
a ,
otherwise
(41)
where (0, 1) and > 0 are parameters and tmax is
the maximum allowable time for tuning. Typical choices are
= 0.1 and = 1.
Proposition 4 (Proof of Convergence): The iterations
described by (29) and (30) in Proposition 2 with the VPP
in (41) are guaranteed to converge.
Proof: Because at becomes fixed after a finite period
of adaptation, at a0 (1 + )tmax < + is bounded.
Proposition 3 can be applied, and then Proposition 4 can be
deduced readily.
The ADMM algorithm will converge very slowly if any one
of the residual terms rt or dt converges slowly [20]. Because
of this, the main principle of the above tuning method is to
balance the speed with which the primal residual rt and dual
residual dt converge to zero.
When applying this penalty parameter tuning technique
to distributed optimization problems, the physical interpretation is relatively straightforward. As seen from the augmented Lagrangian in (22), a larger t penalizes violations
of primal feasibility more harshly and reduces rt , which
means that it is easier for different areas to reach agreement in the shared data at the borders. The drawback is
that both converge to the optimal solution more slowly.
In contrast, a smaller t and a larger rt provide greater
optimality within an area, but they make it more difficult for the shared data between different areas to reach
agreement.
VI. N UMERICAL S IMULATIONS
The D-SOCP in Proposition 2 and D-SOCP with the VPP
in (41) were tested using the following two networks.
1) An IEEE 69-bus distribution system derived from a portion of the practical Pacific Gas and Electric Company
distribution system as shown in Fig. 7.
2) An IEEE 123-bus distribution system with a meshed
structure that was partitioned into five areas, as shown
in Fig. 8.
Suppose that each area is operated by a different company.
Detailed information on these systems is given in [29][31].
The base power was 1.0 MW, and the following calculations were all based on per-unit data. The security constraints were given by voltage boundaries of V = 0.9 and
V = 1.042 [6]. For both systems, the maximum active and

Fig. 7. IEEE 69-bus distribution system divided into four areas. The dashed
lines indicate the bus set corresponding to each area.

Fig. 8. Meshed IEEE 123-bus distribution system divided into five areas.
The dashed lines indicate the bus set corresponding to each area.

reactive powers that could be supplied by substation were

both set to 6.0 MW. The line transmission limit was set
ij = 6.0 MW. The state variable xa was initialized
to P ij = Q
using a flat profile.
For the IEEE 69-bus system, ten DGs were located at buses
G = {19, 20, 27, 54, 66, 34, 38, 3, 51, 48}, and the maximum
active power they could supply was set to P Gi = 300 kW.
Ci =
Gi = Q
The differences in reactive powers were set to Q
300 kvar and QGi = QCi = 0. For the meshed IEEE 123-bus
system, four DGs were located at buses {7, 12, 33, 47}, and
six VAR compensators were located at buses {121, 67, 76,
94, 115, 108}. The maximum active power they could supply
was, P Gi = 150 kW and the differences in reactive powers
Gi = Q
Ci = 150 kvar and Q = Q = 0.
were set to Q
Gi
Ci
The algorithms were implemented using MATLAB on
a machine with an Intel Core i5-3210M 2.50-GHz processor
and 8 GB of RAM.
Both IEEE 69-bus and 123-bus systems were derived from
practical distribution systems. The radial IEEE 69-bus system
corresponds to parallel region partitioning, whereas the IEEE
123-bus represents a challenging scenario with a large number
of buses and a loop.
In the following, Sections VI-AVI-D focus on convergence, accuracy and sensitivity analyses of reactive power
optimization on the feeder side (inner loop) without considering voltage regulation, whereas Section VI-E investigates the
effects of reactive power optimization with voltage regulation
(an inner loop with an outer loop).

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ZHENG et al.: FULLY DISTRIBUTED REACTIVE POWER OPTIMIZATION AND CONTROL METHOD FOR ADNs

(a)

(b)

(c)

(d)

(e)

(f)

Fig. 9. Convergence of D-SOCP for reactive power optimization. (a) Active power loss, (b) reactive powers schedule, and (c) border residual with the 69-bus
system. (d) Active power loss, (e) reactive powers schedule, and (f) border residual with the 123-bus system. In (a) and (d), the optimal active power losses
determined using the original nonconvex centralized Model 1 (CEN) with the IPOPT solver [32] are shown by the dashed lines.

(a)

(b)

Fig. 10. Border residual as a function of the iteration number with different
penalty parameters with D-SOCP. (a) IEEE 69-bus system. (b) IEEE 123-bus
system.

A. Convergence of the D-SOCP

Fig. 9 shows convergence of the real and reactive power
using the D-SOCP described in (29) and (30). The penalty
parameter was set to = 1.0 here. Fig. 9(a) and (d) shows the
convergence of the objective function. Note that the optimum
active power loss is reached in both systems. The convergence
of optimal reactive outputs of DGs and VAR compensators,
i.e., QGi and QCi respectively, is shown in Fig. 9(b) and (e).
Fig. 9(c) and (f) shows the residual t as defined in (26),
which exhibits piecewise linear convergence. A small value
of t corresponds to minor border gaps, which quantify the
agreement between different areas and the stability of local
solution. The residual drops rapidly at the beginning until the
proposed D-SOCP converges to modest accuracy (103 ), and
then decreases at a slower rate. Fortunately, however, an accuracy of 103 is often sufficient for practical applications
of ADNs.

(a)

(b)

Fig. 11. Convergence of the border residual with the VPP. (a) IEEE 69-bus
system. (b) IEEE 123-bus system.

i.e., 1 100, D-SOCP performs well; however, when

= 0.01, the residual in IEEE 69-bus system did not decrease
significantly until around the 100-th iteration, whereas the
residual in IEEE 123-bus system decreased to around 101
after 300 iterations. The border residual converged more
slowly with a small value of , which may result in slower
convergence in practice.
Setting empirically may not be suitable for all problems
for DNs because in practice, this may require large historical
datasets. Even if there is sufficient time to build large historical
datasets and an empirical rule can be determined, this rule may
fit the historical data, but not the real-time application if there
are significant changes to the network (which we may expect,
as there are often drastic and unpredictable changes in the
loads or even the structure of DNs). Hence, a tuning method
is preferable to accelerate convergence, even when is not
set appropriately.

B. Sensitivity to the Penalty Parameter

C. Effects of VPP and the Accuracy of Distributed

Algorithms

Fig. 10 shows a sensitivity analysis of penalty parameter

for both systems, with the penalty parameter initialized in
the range 0.01 100. The performance of D-SOCP is
dependent on the value of ; when is sufficiently large,

The D-SOCP method with a VPP, as described in equations (29), (30), and (41), is expected to improve the performance of D-SOCP. The effect of the VPP is shown in
Fig. 11. When the initial value of penalty parameter was

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
10

IEEE TRANSACTIONS ON SMART GRID

TABLE I
C OMPARISON OF O PTIMAL S OLUTIONS AND ACTIVE P OWER
L OSSES A MONG T HREE A LGORITHMS

set to 0.01, convergence was accelerated through the use

of VPP. Without VPP, when was initialized to 0.01, the
border residual obtained using D-SOCP was too large in the
IEEE 123-bus system, even after 300 iterations; however, this
large residual was not observed when VPP was used. When
was initialized to a relatively large value, the D-SOCP and
VPP methods exhibited similar performance (not shown here
for brevity).
The accuracy of D-SOCP and VPP in both test systems is
listed in Table I, where the optimal solutions and active power
losses obtained by the CEN, D-SOCP, and D-SOCP with VPP
are compared. When = 1.0, D-SOCP reached the optimal
solutions in both systems. However, when 0 = 0.01, for the
radial IEEE 69-bus system, although D-SOCP still converged
to the optimal solution, but convergence was slow, as shown in
Fig. 11; similarly, when 0 = 0.01, VPP can get more accurate
results with less iterations than D-SOCP for the meshed IEEE
123-bus system.
Note that the accuracy with the meshed network was
slightly inferior compared with that for the radial network
because the SOCP relaxation may not be strictly exact for
meshed networks without the aid of a phase shifter [13].
Fortunately, in practice, DNs typically operate radially, which
means that both D-SOCP and VPP have potential applications
for DNs.
D. Sensitivity to Partitioning
To analyze the sensitivity to partitioning, D-SOCP was
tested for three types of partitioning of the IEEE 69-bus
system. Detailed area divisions are listed in Table II. The original division is denoted as partition 1; partition 2 represents the
case whereby different bus sets are chosen for the same zones;
partition 3 represents a more challenging case of fragmentary
division in a larger number of areas.
The convergence of the three partitioning schemes is shown
in Fig. 12. Partition 2 exhibited the same number of areas
as partition 1, but with significant changes of bus sets in
areas 1 and 3 as well as minor changes of bus sets in
areas 2 and 4. However, the convergence characteristics did
not differ significantly, and it follows that choosing different
bus sets for the same number of areas does not significantly
affect the optimization process.
Partition 3 retained the same bus sets in areas 2 and 4 as
partition 1 did, with other buses further divided into another

TABLE II
A REA D IVISION OF THE IEEE 69-B US D ISTRIBUTION N ETWORK

Fig. 12.

Convergence of D-SOCP for the three types of partitioning.

five areas. In this case, D-SOCP converged in a slightly slower

manner in comparison to partitions 1 and 2 due to the larger
number of areas.
E. Effect of Voltage Regulation
We can further improve the quality of voltage using voltage
regulation, and we can maintain the range of voltage magnitudes of all buses located at the center of operational interval,
which ensures a wide margin of voltage magnitudes. Both systems with the original partitioning were tested, and the results
are shown in Fig. 13.
For the IEEE 69-bus system, although the active power loss
could be minimized via reactive power optimization, the voltage magnitude nearly approached the upper bound, and the
margin was very narrow without voltage regulation, as shown
in Fig. 13(a). In this case, an overvoltage was likely to occur
in the presence of unexpected system disturbance. However,
with voltage regulation at the PCC, this risk can be mitigated,
as shown in Fig. 13(b).
For the IEEE 123-bus system, no risk of voltage violation
was observed in the base case. To test the effect of voltage
regulation, the load demand was raised to 150% here and
the voltage magnitude approached the lower bound. With two
times of voltage regulation, the narrow margin of voltage had
been improved significantly.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ZHENG et al.: FULLY DISTRIBUTED REACTIVE POWER OPTIMIZATION AND CONTROL METHOD FOR ADNs

11

(a)

(b)

(c)

(d)

(e)

(f)

Fig. 13. (a) Voltage magnitude without regulation, (b) voltage magnitude with regulation, and (c) voltage adjustment defined in (6) for the IEEE 69-bus
system. (d) Voltage magnitude without regulation, (e) voltage magnitude with regulation, and (f) voltage adjustment defined in (6) for the IEEE 123-bus
system.

TABLE III
E FFECT OF D-SOCP IN A R EAL S YSTEM

Fig. 14.
system.

Convergence of border residual of D-SOCP in a real 1066-bus

Fig. 16.

Fig. 15.

Convergence of objective of D-SOCP in a real 1066-bus system.

F. Numerical Test on Real 1066-Bus Distribution System

To stress further the issue of practical application, D-SOCP
has been tested on a real distribution system in Shandong
province of China. It consists of 1066 buses, six VAR
compensation devices, four capacitors, and 51 transformers.
Additionally, 44 PV power plants are added in the real-network

Schematic showing area adjustment.

for demonstration. The network is unbalanced three phase. The

detailed parameters of the system are available online [33].
To demonstrate the effectiveness of D-SOCP, the results
have been compared to those obtained by original nonconvex
Model 1.
As shown in Figs. 14 and 15, both the border residual
and the objectives of D-SOCP were able to converge steadily.
Compared to CEN, the relative error of optimum obtained
by D-SOCP was 104 . This convergence is guaranteed theoretically by Proposition 3. Table III shows that D-SOCP
can achieve the same solution as the original centralized
model. Detailed optimal reactive power schedule obtained by
D-SOCP and voltages of all buses in the system are available
online [33].

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
12

IEEE TRANSACTIONS ON SMART GRID

VII. C ONCLUSION

Substituting (45) into (28) yields

We have described a method to solve nonconvex reactive power optimization problems using a D-SOCP. Based on
variable splitting and ADMM, a powerful fully distributed
algorithm was derived by eliminating global variables and
global updates. A fully distributed penalty parameter tuning
technique (VPP) was shown to accelerate D-SOCP and to
handle worst-case scenarios. A proof of its convergence was
presented. Simulated data indicated that the proposed distributed method with VPP can converge reliably and achieve
accurate solution.

uta,j =

xta,j + xtb,j

ta,j

and (46) can be written as

, j O N a , b = Mj \a

xta,j + xtb,j

uta,j , j O N a , b = Mj \a. (47)

2
Using (45), the multiplier updates in (25) can eliminate
global vectors, that is


t+1
t+1
t
j O N a
=

+
0.5
x
+
x
t+1
a,j
a,j
a,j
b,j ,
ta,j =

b = Mj \a.
A PPENDIX
Proof of Lemma 1: For any given {N0a |a}, let N a = N0a ,
we apply the construction method of N a in Section IV.
If Mj = 2, j O, then Lemma 1 holds.
Otherwise, for every bus j where Mj > 2, without loss of
generality, suppose that Mj = {a, b, c}, as shown in Fig. 15.
We select a tie-line j k between areas a and b and apply
the following adjustment: N a = N0a {k}, N b = N0b \{k}. We
then have Mj := Mj 1, as shown in Fig. 16. This can be
repeated until Mj = 2.
Proof of Lemma 2: From Lemma 1, for any overlapping bus
j O, Mj = 2, and therefore bus j belongs to a unique pair
of neighboring areas a and b = Mj \a. Adding the updates of
ta,j and tb,j yields
stj

xta,j + xtb,j

2
b = Mj \a.

t1
t1
a,j + b,j

ta,j + tb,j
2

, j O N a
(42)

However, the optimization of (24) for areas a and b can be

expressed analytically and summed, that is
stj =

xta,j + xtb,j
2

t1
t1
a,j + b,j

, j O N a , b = Mj \a.
(43)

Hence, by comparing (42) and (43), Lemma 2 is

proved [15].
Proof of Proposition 2: By completing the squares, (23) is
equivalent to

  T 
t+1
ta,j
xa,j stj
xa = arg min fa (xa ) +
xa Xa

jON a

2 


t

s
xa,j
j
2
2


2
 
ta,j 


t
 .
xa,j sj


+

= arg min fa (xa ) +

xa

Xa

jON a

(44)
Substituting (28) into (44) yields the expression in (29).
From Lemma 2, the global vectors in (43) can be expressed
as local vectors, that is
stj =

xta,j + xtb,j
2

j O N a ,

b = Mj \a.

(45)

(46)

(48)

Substituting (47) into (48) yields (30). Hence, the claim of

the proposition follows readily.
R EFERENCES
[1] C. DAdamo, S. Jupe, and C. Abbey, Global survey on planning
and operation of active distribution networksUpdate of CIGRE
C6.11 working group activities, in Proc. 20th Int. Conf. Exhibit. Elect.
Distrib. (CIRED), Prague, Czech Republic, 2009, pp. 14.
[2] T. J. T. Hashim, A. Mohamed, and H. Shareef, A review on voltage
control methods for active distribution networks, Prz. Elektrotechniczn.,
vol. 88, pp. 304312, Jun. 2012.
[3] S. Chowdhury, S. P. Chowdhury, and P. Crossley, Microgrids and Active
Distribution Networks. London, U.K.: Inst. Eng. Technol., 2009.
[4] S. Kaisler, F. Armour, and J. A. Espinosa, Introduction to big data:
Challenges, opportunities, and realities minitrack, in Proc. 47th Hawaii
Int. Conf. Syst. Sci. (HICSS), Waikoloa, HI, USA, 2014, p. 728.
[5] S. M. Amin, Smart grid security, privacy, and resilient architectures:
Opportunities and challenges, in Proc. IEEE Power Energy Soc. Gen.
Meeting, San Diego, CA, USA, 2012, pp. 12.
[6] W. Yang, Z. Peng, L. Wenyuan, X. Weidong, and A. Abdollahi, Online
overvoltage prevention control of photovoltaic generators in microgrids,
IEEE Trans. Smart Grid, vol. 3, no. 4, pp. 20712078, Dec. 2012.
[7] L. Franchi, M. Innorta, P. Marannino, and C. Sabelli, Evaluation of
economy and/or security oriented objective functions for reactive power
scheduling in large scale systems, IEEE Trans. Power App. Syst.,
vol. PAS-102, no. 10, pp. 34813488, Oct. 1983.
[8] O. Alsac, J. Bright, M. Prais, and B. Stott, Further developments in
LP-based optimal power flow, IEEE Trans. Power Syst., vol. 5, no. 3,
pp. 697711, Aug. 1990.
[9] B. Stott, J. Jardim, and O. Alsac, DC power flow revisited, IEEE Trans.
Power Syst., vol. 24, no. 3, pp. 12901300, Aug. 2009.
[10] J. Lavaei and S. H. Low, Zero duality gap in optimal power flow
problem, IEEE Trans. Power Syst., vol. 27, no. 1, pp. 92107,
Feb. 2012.
[11] B. C. Lesieutre, D. K. Molzahn, A. R. Borden, and C. L. DeMarco,
Examining the limits of the application of semidefinite programming
to power flow problems, in Proc. 49th Annu. Allerton Conf. Commun.
Control Comput. (Allerton), Monticello, IL, USA, 2011, pp. 14921499.
[12] R. A. Jabr, Radial distribution load flow using conic programming,
IEEE Trans. Power Syst., vol. 21, no. 3, pp. 14581459, Aug. 2006.
[13] M. Farivar and S. H. Low, Branch flow model: Relaxations and
convexificationPart I, IEEE Trans. Power Syst., vol. 28, no. 3,
pp. 25542564, Aug. 2013.
[14] M. Farivar and S. H. Low, Branch flow model: Relaxations and
convexificationPart II, IEEE Trans. Power Syst., vol. 28, no. 3,
pp. 25652572, Aug. 2013.
[15] S. Boyd, N. Parikh, and E. Chu, Distributed optimization and statistical learning via the alternating direction method of multipliers, Found.
Trends Mach. Learn., vol. 3, no. 1, pp. 1122, 2011.
[16] V. Kekatos and G. B. Giannakis, Distributed robust power system state
estimation, IEEE Trans. Power Syst., vol. 28, no. 2, pp. 16171626,
May 2013.
[17] A. X. Sun, D. T. Phan, and S. Ghosh, Fully decentralized AC optimal power flow algorithms, in Proc. IEEE Power Energy Soc. Gen.
Meeting (PES), Vancouver, BC, Canada, 2013, pp. 15.
[18] E. DallAnese, Z. Hao, and G. B. Giannakis, Distributed optimal power
flow for smart microgrids, IEEE Trans. Smart Grid, vol. 4, no. 3,
pp. 14641475, Sep. 2013.

This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
ZHENG et al.: FULLY DISTRIBUTED REACTIVE POWER OPTIMIZATION AND CONTROL METHOD FOR ADNs

[19] B. S. He, H. Yang, and S. L. Wang, Alternating direction method with

self-adaptive penalty parameters for monotone variational inequalities,
J. Optim. Theory Appl., vol. 106, pp. 337356, Aug. 2000.
[20] S. L. Wang and L. Z. Liao, Decomposition method with a variable
parameter for a class of monotone variational inequality problems,
J. Optim. Theory Appl., vol. 109, pp. 415429, May 2001.
[21] S. Bose, S. H. Low, T. Teeraratkul, and B. Hassibi, Equivalent relaxations of optimal power flow, IEEE Trans. Autom. Control, to be
published.
[22] M. E. Baran and F. F. Wu, Optimal capacitor placement on radial distribution systems, IEEE Trans. Power Del., vol. 4, no. 1, pp. 725734,
Jan. 1989.
[23] M. E. Baran and F. F. Wu, Network reconfiguration in distribution
systems for loss reduction and load balancing, IEEE Trans. Power Del.,
vol. 4, no. 2, pp. 14011407, Apr. 1989.
[24] W. H. Kersting, Distribution System Modeling and Analysis (Electric
Power Generation, Transmission, and Distribution), 3rd ed. Boca Raton,
FL, USA: CRC Press, 2012, pp. 158.
[25] S. Paudyal, C. A. Canizares, and K. Bhattacharya, Optimal operation of
distribution feeders in smart grids, IEEE Trans. Ind. Electron., vol. 58,
no. 10, pp. 44954503, Oct. 2011.
[26] F. Alizadeh and D. Goldfarb, Second-order cone programming,
Math. Program., vol. 95, no. 1, pp. 351, 2003.
[27] M. S. Lobo, L. Vandenberghe, S. Boyd, and H. Lebret, Applications
of second-order cone programming, Linear Algebra Appl., vol. 284,
nos. 13, pp. 193228, 1998.
[28] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed
Computation:
Numerical
Methods.
Belmont,
MA,
USA:
Athena Scientific, 1997.
[29] (Aug. 5, 2013). Distribution Test Feeders. [Online]. Available:
http://ewh.ieee.org/soc/ pes/dsacom/testfeeders/
[30] B.
Venkatesh,
S.
Chandramohan,
N.
Kayalvizhi,
and
R. P. Kumudini Devi, Optimal reconfiguration of radial distribuion
system using artificial intelligence methods, in Proc. IEEE Toronto
Int. Conf. Sci. Technol. Human. (TIC-STH), Toronto, ON, USA, 2009,
pp. 660665.
[31] W. H. Kersting, Radial distribution test feeders, in Proc. IEEE
Power Eng. Soc. Winter Meeting, vol. 2. Columbus, OH, USA, 2001,
pp. 908912.
[32] (Jul.
08,
2013).
IPOPT.
[Online].
Available:
http://www.coin-or.org/Ipopt/
[33] (Apr. 1, 2014). [Online]. Available: https://drive.google.com/file/d/
0BzVXltCTbTuyZU1aY0MzVUZGTGc/view?usp=sharing

Weiye Zheng received the B.S. degree in electrical

engineering from Tsinghua University, Beijing,
China, in 2013, where he is currently pursuing the
Ph.D. degree.
His current research interests include smart grids
operation and control, and renewable energy management in active distribution networks.

13

Wenchuan Wu (SM14) received the B.S., M.S.,

and Ph.D. degrees from the Department of Electrical
Engineering, Tsinghua University, Beijing, China,
in 1996, 1998, and 2003, respectively.
He is currently a Full Professor with the
Department of Electrical Engineering, Tsinghua
University. His current research interests include
energy management system, and active distribution
system operation and control.

Boming Zhang (SM95F10) received the

Ph.D. degree in electrical engineering from
Tsinghua University, Beijing, China, in 1985.
Since 1985, he has been with the Department of
Electrical Engineering, Tsinghua University, where
he was promoted to Professor in 1993. His current
research interests include power system analysis
and control, especially in the energy management
system (EMS) advanced applications in the Electric
Power Control Center (EPCC). He has published
over 300 academic papers and implemented over
70 EMS/distributer training systems in China.
Prof. Zhang is currently a Steering Member of the CIGRE China State
Committee and the International Workshop of the EPCC.

Hongbin Sun (SM12) received the double B.S. degrees and the Ph.D.
degree in electric engineering from the Department of Electrical Engineering,
Tsinghua University, Beijing, China, in 1992 and 1997, respectively.
He is currently a Professor with the Department of Electrical Engineering,
Tsinghua University. His current research interests include energy management system and distribution management system.

Yibing Liu (S14) received the M.S. degree in electric engineering from the
Department of Electrical Engineering, Tsinghua University, Beijing, China,
in 2014.
He is currently a Lecturer with the Department of Hydro and Electrical
Engineering, Qinghai University, Xining, China. His current research interests
include renewable energy management in active distribution networks.