This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
1

Distributed Optimal Conservation Voltage

Reduction in Integrated Primary-Secondary
Distribution Systems
Qianzhi Zhang, Student Member, IEEE, Yifei Guo, Member, IEEE, Zhaoyu Wang, Senior Member, IEEE,
and Fankun Bu, Student Member, IEEE

Abstract—This paper proposes an asychronous distributed Parameters

leader-follower control method to achieve conservation voltage An , Bn Topology matrices for boundary system.
reduction (CVR) in three-phase unbalanced distribution systems NS Total number of secondary networks.
by optimally scheduling smart inverters of distributed energy
resources (DERs). One feature of the proposed method is to con- NeS Setting number of secondary networks for par-
sider integrated primary-secondary distribution networks and tial barrier.
voltage dependent loads. To ease the computational complexity k1p , k2p , k3p Constant-impedance (Z), constant-current (I)
introduced by the large number of secondary networks, we parti- and constant-power (P) coefficients for active
tion a system into distributed leader-follower control zones based ZIP loads.
on the network connectivity. To address the non-convexity from
the nonlinear power flow and load models, a feedback-based lin- k1q , k2q , k3q Constant-impedance (Z), constant-current (I)
ear approximation using instantaneous power and voltage mea- and constant-power (P) coefficients for reactive
surements is proposed. This enables the online implementation ZIP loads.
of the proposed method to achieve fast tracking of system varia- pLi,φ,t , qi,φ,t
L
Real and reactive load multipliers.
tions led by DERs. Another feature of the proposed method is the pgi,φ,t Three-phase real power injections by the smart
asynchronous implementations of the leader-follower controllers,
which makes it compatible with non-uniform update rates and ro- inverter.
cap
bust against communication delays and failures. Numerical tests qi,φ,t Three-phase reactive power capacity of smart
are performed on a real distribution feeder in Midwest U. S. to inverters.
validate the effectiveness and robustness of the proposed method. m
Sij,φ,t Three-phase apparent power measurements
feedback from the system.
Index Terms—Alternating direction method of multipliers scap
i,φ,t Power capacity of smart inverters.
(ADMM), asynchronous update, conservation voltage reduc- T Time length for termination.
tion (CVR), feedback-based linear approximation, integrated
v min , v max Minimum and maximum limits for squared
primary-secondary distribution networks.
nodal voltage magnitude.
m
vij,φ,t Three-phase voltage measurements feedback
N OMENCLATURE from the system.
Sets and Indices zij , rij , xij Matrices of the line impedance, resistance and
B Set of boundary buses. reactance.
Cj Set of children buses of bus j. τn Setting iteration for boundary delay.
E Set of branches. µ, τ dec , τ inc Parameters for updating penalty factor.
N Set of buses. N = P ∪ B ∪ S. Variables
Mt , N t Sets of follower controllers in asynchronous Lρ Augmented Lagrangian.
communication. Pij,φ,t Three-phase real power flows.
P Set of primary network buses. pi,φ,t , qi,φ,t Three-phase active and reactive power bus in-
S Set of secondary network buses. jections.
X Set of variables for primary network. pZIP , q ZIP
i,φ,t i,φ,t Three-phase real and reactive ZIP loads.
Zn Set of variables for secondary networks. Qij,φ,t Three-phase reactive power flows.
g
k Index of iteration. qi,φ,t Three-phase reactive power injections by the
n Index of secondary network. smart inverter.
t Index of time instant. rnk , skn Primal and dual residuals.
φ Index of three-phase φa , φb , φc . Sij,φ,t Three-phase apparent power flow.
vi,φ,t Squared of three-phase voltage magnitude.
This work was supported in part by the U.S. Department of Energy Wind √
Energy Technologies Office under Grant DE-EE0008956, and in part by the v̄i,φ,t Estimation of the nonlinear term vi,φ,t .
National Science Foundation under ECCS 1929975 (Corresponding author: x, zn Compact variables of primary and secondary
Zhaoyu Wang). networks.
The authors are with the Department of Electrical and Computer
Engineering, Iowa State University, Ames, IA 50011 USA (e-mail: xB,n , zB,n Compact variables for boundary of primary
[email protected]; [email protected]; [email protected]; [email protected]).

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
2

network and coupling secondary network. makes the optimization programs non-convex and NP hard.
λn Lagrange multipliers. Though heuristic algorithms (e.g. differential evolution algo-
ρk Iterative varying penalty coefficient for con- rithm [13]) or general nonlinear programming solvers (e.g.
straint violation. fmincon) can solve these problems, it often suffers the sub-
εpij,φ , εqij,φ Active and reactive power loss nonlinear terms. optimality without proven optimal gaps. Other studies have
εvi,φ Voltage drop nonlinear term. directly dropped nonlinear terms (e.g. LinDistFlow) [12] or
used first-order Taylor expansion at a fixed point, to reduce
I. I NTRODUCTION the computational complexity [14]. However, such offline lin-
ONSERVATION voltage reduction (CVR) is to lower the ear approximation methods may bring non-negligible errors to
C voltage for peak load shaving and long-term energy sav-
ings, while maintaining the voltage at end users within the
power flow and bus voltage computation, thus, hindering the
CVR performance. In addition, voltage-dependent load mod-
bound of set by American National Standards Institute (ANSI) els must be used when studying CVR because the nature of
[1], [2]. CVR is that load is sensitive to voltage. Therefore, the non-
Conventionally, CVR is implemented by rule-based or linear ZIP or exponential load models further complicate the
heuristic voltage controls at primary feeders by legacy regu- VVO-CVR problem.
lating devices, such as on-load tap-changers, capacitor banks, (3) Solution Algorithms: The VVO-CVR can be directly
step-voltage regulators, in slow timescales [3], [4]. The in- solved by centralized solvers, which naturally requires global
creasing integration of distributed energy resources (DERs), communication, monitoring, data collection and computation.
e.g., residential solar photovoltaics (PV), in secondary net- Centralized solvers may be computationally expensive and less
works challenges conventional methods; but in turn, it also reliable for large systems, which is particularly true for a dis-
provides new voltage/var regulation capabilities by injecting tribution system with a number of secondary networks. The
or absorbing reactive power. The interactions between CVR information privacy of customers is another concern for cen-
and widespread DERs have been explored in [5]–[7]. It is tralized control. To this end, some studies have developed dis-
demonstrated that DERs can flatten voltage profiles along tributed algorithms to solve VVO-CVR based on distribution
feeders to allow deeper voltage reduction. In addition, the optimization methods, such as alternating direction method of
fast and flexible reactive power capabilities of four-quadrant multipliers (ADMM) [12] and primal-dual gradient algorithms
smart inverters enable implementing CVR in fast timescales. [15]. In [12] and [16], the ADMM is applied to solve VVO-
To achieve system-wide optimal performance, voltage/var op- CVR in a three-phase unbalanced distribution network. In [14]
timization based CVR (VVO-CVR), which can be cast into an and [17], to provide a fully distributed solution, the convex-
optimal power flow program, has spurred a substantial body of ified voltage regulation model is solved by ADMM. In [18],
research. In [8], a linear least-squares problem is formulated different loading and PV penetration levels are tested for op-
for optimizing the CVR objective with a linearly approximated timal reactive power control in large-scale distribution sys-
relation between voltages changes and actions of voltage reg- tems. In [19] and [20], ADMM is implemented for solving
ulating devices. The integration of optimal CVR and demand the optimal reactive power dispatch problem of PV inverters.
response is considered in [9] to maximize the energy effi- In [21], optimal coordinated voltage control is achieved by
ciency. Voltage optimization algorithm is developed in [10] to ADMM for multiple distribution network clusters. Note that
implementing CVR by reactive power control of aggregated the distributed control algorithms in existing works inherently
inverters. In [11], a convex optimization problem is formulated require synchronous update, which implies that the computa-
with network decomposition to optimally regulate voltages in tion efficiency depends on the slowest agent. They are sig-
a decentralized manner. In [12], the large-scale VVO-CVR nificantly affected by the differences in processing speed and
problem is divided into a number of small-scale optimization communication delays, which may deteriorate the control per-
problems using a distributed framework with only local in- formance [22]–[24]. For example, the synchronous distributed
formation exchange, which coordinates multiple bus agents to algorithms may lose the fast-tracking capabilities for large sys-
obtain a solution for the original centralized problem. While tems.
the previous works have contributed valuable insights to VVO- To address these challenges, this paper proposes a leader-
CVR, there are problems remaining open, summarized as fol- follower distributed algorithm based on asynchronous-ADMM
lows: (async-ADMM) [25] to solve the VVO-CVR problem and
(1) Integrated Primary-Secondary Distribution Networks: A enable online implementation with feedback-based linear ap-
practical distribution system is composed of medium-voltage proximation, where the primary network corresponds to the
(MV) primary networks and low-voltage (LV) secondary net- leader control and each secondary network corresponds to a
works, where most loads and residential DERs are connected follower control. The contributions of this paper are threefold:
to secondary networks. However, previous studies have fo- • Mapping primary-secondary distribution system to
cused on primary networks while simplifying secondary net- ADMM-based leader-follower control framework: To bet-
work by using aggregate models to reduce computational bur- ter model DERs’ impacts and improve the grid-edge volt-
den. The grid-edge voltage regulation in distribution networks age regulation performance, we consider an integrated
has not been well addressed. primary-secondary distribution system with detailed mod-
(2) Power Flow Models: Some VVO-CVR studies have used eling of secondary networks. To solve the VVO-CVR
full AC power flow models; however, the nonlinear nature problem in a distributed way, we first split the primary

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
3

and secondary networks from modeling perspective, then Distributed leader-follower control
framework (cyber layer)
introduce coupling constraints at boundary nodes, finally
- Information exchange Follower controller #Ns
map the primary and secondary networks into leader and - Measurements
Follower controller #2 ...
follower controllers in ADMM distributed framework. - Dispatchings
Leader controller
• Online feedback-based linear approximation method for Follower controller #1
power flow and ZIP load: We propose an online
feedback-based linear approximation method, where the
instantaneous power and voltage measurements are used Secondary network #1
Primary network
as system feedback in each iteration of ADMM to lin-
Secondary network #2
earize the nonlinear terms of power flow calculation for
both power flow and ZIP load models, which can signifi- Secondary network #Ns
cantly reduce the computational complexity and lineariza- ...
- Boundary
tion errors by instantaneously tracking system variations. - Load node
• Asynchronous implementation of ADMM: We develop an Integrated primary-secondary distribution system - DER
(physical layer)
asynchronous counterpart of conventional ADMM-based
distributed control algorithms, which is robust against
non-uniform update rates and communication delays, Fig. 1. Overall framework of the proposed distributed CVR with online and
making it suitable for real-world applications. asynchronous implementations.
The remainder of the paper is organized as follows: Section
II presents the overall framework of the proposed method.
Section III describes a centralized VVO-CVR in an integrated III. O PTIMAL CVR IN I NTEGRATED
primary-secondary distribution system. Section IV proposes P RIMARY-S ECONDARY D ISTRIBUTION S YSTEMS
the distributed algorithm with online and asynchronous im-
A. Modeling Integrated Primary-Secondary Distribution Net-
plementation. Simulation results and conclusions are given in
works
Section V and Section VI, respectively.
A real distribution system consists of substation transform-
ers, MV primary networks, service transformers, and LV sec-
II. OVERVIEW OF THE P ROPOSED F RAMEWORK ondary networks. Here, we consider a three-phase radial dis-
tribution system with N buses denoted by set N and N − 1
The general framework of the proposed distributed CVR
branches denoted by set E. The buses in primary network
with online and asynchronous implementations is shown in
and secondary networks are denoted by sets P and S, re-
Fig. 1. A VVO-CVR framework that dispatches smart inverters
spectively. The three-phase φa , φb , φc are simplified as φ.
is developed for unbalanced three-phase distribution systems.
The time instance is represented by t. For each bus i ∈ N ,
The integration of primary-secondary networks with detailed
pZIP ZIP
i,φ,t , qi,φ,t ∈ R
3×1
are the vector of three-phase real and re-
secondary network models will be taken into account for better
voltage regulation at grid-edge. Inspired by the physical struc-
active ZIP loads at time t; pgi,φ,t , qi,φ,tg
∈ 3×1 are the vector R
of three-phase real and reactive power injections by the smart
ture of the distribution systems shown in Fig. 1, the primary
network corresponds to the leader controller and each sec-
inverter at time t; vi,φ,t := Vi,φ,t Vi,φ,t ∈ 3×1 represents R
the vector of three-phase squared voltage magnitude at time t.
ondary system corresponds to a follower controller. We then
Cj denotes the set of children buses. For any branch (i, j) ∈ E,
develop a distributed solution algorithm via ADMM frame-
C
zij = rij +ixij ∈ 3×3 are matrices of the three-phase branch
work to solve the VVO-CVR problem in a leader-follower
distributed fashion, where the leader and followers controllers
resistance and reactance; Sij,φ,t = Pij,φ,t + iQij,φ,t ∈ 3×1 C
denote the vector of three-phase real and reactive power flow
only exchange aggregate power and voltage magnitude infor-
from buses i to j at time t.
mation at boundaries. Note that, we specially address the asyn-
chronous counterpart of the distributed solver to achieve robust Most of the loads and DERs are connected to secondary net-
and fast solutions while guaranteeing the convergence. works, the power flows through the service transformers can
be equivalently considered as the power injections pi,φ,t , qi,φ,t
The nonlinear power flow and ZIP load models make the
at the boundary bus i ∈ B (i.e., LV side bus of service trans-
proposed problem nonconvex. To handle this issue, we pro-
former), where B ⊆ N denotes the boundary bus set and let
pose to leverage voltage and line flow measurements as feed-
bus i0 be the copy of bus i at time t. Accordingly, the phys-
back to linearize these nonlinear models and make the program
ical coupling of active power, reactive power and voltage at
tractable. This feedback-based linear approximation method
the boundary bus i are expressed as,
will be embedded within the distribution solution algorithm
and combined with the online implementation of the dis- X
pi,φ,t + Pi0 j,φ,t = 0, ∀i ∈ B (1)
tributed algorithm, where the reactive power outputs of smart
j∈Ni
inverters will be updated at each iteration by solving a time- X
varying convex optimization program in a leader-follower dis- qi,φ,t + Qi0 j,φ,t = 0, ∀i ∈ B (2)
tributed fashion. In this way, we transform the conventional of- j∈Ni
fline VVO-CVR to be an online feedback-based control model. vi,φ,t − vi0 ,φ,t = 0, ∀i ∈ B. (3)

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
4

B. VVO-Based CVR power point tracking for active power control. Note that we
The aim of CVR is to reduce the total power consumption of focus on proposing an online distributed VVO-CVR to opti-
the entire system while maintaining a feasible voltage profile mally dispatch the smart inverters in fast timescale. However,
across primary and secondary networks. Therefore, the VVO- the conventional voltage regulation devices, such as on-load
CVR program can be formulated as follows, tap changer (OLTC) and capacitor banks (CBs), have slow
X X reaction speed and limited number of switching operation,
min Re{S0j,φ,t } (4a) which cannot handle the fast changes in system states caused
j:0→j φ∈{a,b,c} by loads and renewable energy resources in modern distribu-
s.t. (1)-(3) tion systems. Thus, they should be controlled in a rather slow
X timescale instead of together with smart inverters, which is
Pij,φ,t = Pjk,φ,t − pgj,φ,t + pZIP p
j,φ,t + εij,φ,t (4b)
out of the scope of this paper. But it should be highlighted
k:j→k
X that, the operation of OLTC and CBs can be controlled by the
g
Qij,φ,t = Qjk,φ,t − qj,φ,t ZIP
+ qj,φ,t + εqij,φ,t (4c) leader controller, of which the impact can be taken into ac-
k:j→k count in the fast timescale control of smart inverters. In this
Qij,φ,t + εvi,φ,t


vj,φ,t = vi,φ,t − 2 r̄ij Pij,φ,t + x̄ij way, the coordination among them can be easily achieved.
(4d) The power flow model (4b)–(4d) includes non-linear terms
p p √ p
εpij,φ , εqij,φ and εvi,φ . In the unbalanced three-phase branch flow
pZIP L
i,φ,t = pi,φ,t ki,1 · vi,φ,t + ki,2 · vi,φ,t + ki,3 (4e)
ZIP L q q √ q
model, these nonlinear terms render the program non-convex
qi,φ,t = qi,φ,t ki,1 · vi,φ,t + ki,2 · vi,φ,t + ki,3 (4f) that is hard to solver. However, simply dropping these nonlin-
v min ≤ vi,φ,t ≤ v max , ∀i ∈ N (4g) ear terms may cause non-negligible modeling errors that de-
cap
−qi,φ,t g
≤ qi,φ,t cap
≤ qi,φ,t , ∀i ∈ G. (4h) teriorates the voltage regulation performance. Similarly, when
calculating active/reactive ZIP loads in constraints (4e) and
In objective (4a), the Re{S0j,φ,t } denotes the three-phase √
(4f), the nonlinear part vi,φ,t also introduces non-convexity.
active power supplied from the substation of the feeders at To make the problem tractable, we propose to estimate the
time t. For any branch (i, j) ∈ E, the unbalanced three-phase nonlinear terms with instantaneous voltage and line flow mea-
branch flow model can be represented by constraints (4b)– surements, which can be referred to as a feedback-based linear
(4d). Here, the and denote the element-wise multiplica- approximation method. Such approximate models of power
tion and division. If the network is not too severely unbalanced flow and ZIP load are integrated with the online implementa-
[14], then the voltage magnitudes between the phases are sim- tion of the distributed solver, which will be detailed in Section
ilar and relative phase unbalance α is small. The unbalanced IV-C.
three-phase resistance matrix r̄ij and reactance matrix x̄ij can
be referred to [12]. The active and reactive ZIP loads pZIP i,φ,t
ZIP
and qi,φ,t are calculated in constraints (4e) and (4f), where C. Reformulating VVO-CVR for Distributed Solution by Split-
pLi,φ,t , qi,φ,t
L
R
∈ 3×1 are the vectors of three-phase active
p p
ting Primary and Secondary Networks
and reactive load multipliers on bus i, respectively. ki,1 , ki,2 , We first compactly define the decision vector x :=
p q q q T
ki,3 and ki,1 , ki,2 , ki,3 are constant-impedance (Z), constant- [pi,φ,t , qi,φ,t , vi,φ,t ] , i ∈ P for primary network and zn :=
T
current (I) and constant-power (P) coefficients for active and [Pi0 j,φ,t , Qi0 j,φ,t , vi0 ,φ,t ] , i ∈ S for nth secondary network,
reactive ZIP loads on bus i. Our work is proposing a dis- that consist of all the active/reactive branch flows and squared
tributed CVR model based on static optimal power flow prob- bus voltage magnitudes belonging to the primary network
lem, which focuses on system level optimization. The dynamic and nth secondary network, respectively. Accordingly, the
model, such as induction motor, is not included in the scope boundary variables xB,n and zB,n (sub-vectors of x and
of our work. In constraint (4g), the (squared) bus voltage mag- zn , respectively) regarding nth secondary network (suppose
nitude limits are set to the bus voltage v min and v max , which bus i is the boundary bus) can be compactly represented
T
are typically [0.952 , 1.052 ] p.u., respectively. The nodal volt- by: xB,n := [pi,φ,t , qi,φ,t , vi,φ,t ] , i ∈ B and zB,n :=
hP iT
age constraint (4g) is applied to all nodes in the distribution P 0 ,
P
Q 0 , v 0 , i ∈ B, respectively.
j∈Ci i j,φ,t j∈Ci i j,φ,t i ,φ,t
system, including primary network and secondary networks.
By decomposing the constraints into primary network, sec-
In constraint (4h), the reactive power output of smart in-
ondary networks and boundary systems, the VVO-CVR prob-
verter is limited by the available reactive power of smart in-
cap lem in (1)–(3) and (4) can be compactly reformulated as,
verters qi,φ,t . Based on the capacity of the smart inverter scap
i,φ,t
and the active power output of smart inverter pgi,φ,t , we can min f (x) (5a)
x,zn ,∀n
calculate the available capacity for reactive power generation
cap s.t. x ∈ X := {x|(4b)–(4g)} (5b)
of smart inverters qi,φ,t . According to the requirement for re-
active power capability of the DERs in IEEE 1547-2018 Stan- zn ∈ Zn := {zn |(4b)–(4h)} , ∀n (5c)
dard [26], the DERs shall provide voltage regulation capabil- An xB,n + Bn zB,n = 0 ⇐⇒ {(1)–(3)} , ∀n (5d)
ity by injecting reactive power or absorbing reactive power.
Therefore, we assume there are enough reactive power capa- where constraint sets (5d) is defined for boundary system. The
bility for DER inverters in our proposed VVO-CVR problem. An = I9 and Bn = blkdiag(I6 , −I3 ) for three-phase sec-
We also assume the DER system operates with the maximum ondary networks and An = I3 and Bn = blkdiag(I2 , −I1 )

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
5

for single-phase secondary networks, where Im denotes the Follower Controller #1 τ 1

m × m identity matrix. Follower Controller #2 τ 2

Follower Controller #3 τ
3

Follower Controller #4 τ
4

IV. P ROPOSED D ISTRIBUTED S OLUTION A LGORITHM FOR Follower Controller #5 τ 5

A SYNCHRONOUS AND O NLINE I MPLEMENTATIONS 0 1 2 3 4 5 6

Leader Controller Clock

A. Standard Distributed Solution Algorithm via ADMM

The augmented Lagrangian of the compact VVO-based Fig. 2. An example of leader-follower async-ADMM framework.
CVR (5) is shown as,

NS
X B. Asynchronous Implementation
Lρ = f (x) + λn (An xB,n + Bn zB,n )
When implementing sync-ADMM to solve the VVO-CVR
n=1
NS in above formulations (7)–(9), the leader controller of the pri-
X ρk 2 mary network has to wait till all the follower controllers of
+ kAn xB,n + Bn zB,n k2 (6)
n=1
2 the secondary networks finish updating their variables zn to
receive the latest boundary variables zB,n and proceed. Thus,
where the λn is the vector of the Lagrange multipliers for the the sync-ADMM is not ideal for optimally dispatching smart
primary network (leader controller) and the coupling nth sec- inverters in a fast timescale and robust for communication de-
ondary network (follower controller), k denotes the iteration lay. To alleviate this problem, an async-ADMM method [25]
index, and ρk > 0 is the iterative varying penalty coefficient is implemented, where the leader controller only needs to re-
for constraint violation. ceive the updates from a minimum number of N eS ≥ 1 follower
The ADMM solves the problem (5) by alternatingly min- controllers, and NS can be much smaller than the total num-
e
imizing the augmented Lagrangian (6) over x, zn and λn . It ber of follower controllers NS . This relaxation is the so called
consists of the following steps: (i) By (7), the leader con- partial barrier. Here a small number of N eS based on partial
troller first updates the variables x associated with primary barrier means that the update frequencies of the slow follower
system, where the update boundary variables xk+1 B,n will be controllers can be much less than those faster follower con-
sent to each corresponding follower controller. (ii) By (8), the trollers. To ensure sufficient freshness of all the updates, we
follower controllers update the variables zn associated with also require a bounded delay, i.e., the n-th follower controller
each secondary system by. Since each distributed follower con- must communicate with the leader controller and receive the
troller only solves the problem in terms of the local variables results from the leader controller for updating local variables
in secondary systems so that this step can be performed in at least once every τn ≥ 1 iterations. Consequently, the up-
k+1
parallel. The updated boundary variables zB,n will be sent to date in every follower controller can be at most τn iterations
the leader controller. (iii) As in (9), each follower controller later than the leader’s clock. An example of the asynchronous
is also responsible for updating the variables λn by xk+1B,n and update is given in Fig. 2, where the partial barrier N eS = 2. In
k+1 this example, the leader controller receives the updates from
zB,n . The newly updated variables λk+1 n will be sent to the
leader controller. follower controller 1 at clock time two; the leader controller
receives the updates from follower controllers 2 and 5 at clock
Ns
X time three; the leader controller receives the updates from fol-
xk+1 ← arg min f (x) + λkn (An xB,n + Bn k
zB,n ) lower controllers 3 and 4 at clock time six. Meanwhile, the
x∈X n=1
leader controller has already preserved the update of follower
Ns
X ρk k 2 controller 1 for five iterations and follower controllers 2 and
+ An xB,n + Bn zB,n , (7)
2 2 5 for four iterations.
n=1
The convergence rate of this async-ADMM is in the or-
der of O(NS τn /2T N eS ) [25]. The T is the total time length
znk+1 ← arg min λkn An xk+1

for termination. This convergence rate can be intuitively ex-
B,n + Bn zB,n
zn ∈Zn plained by different value of NS , NeS and τn : (i) If the number
k 2 of secondary networks in the system, NS , is large, more iter-
ρ
+ An xk+1
B,n + Bn zB,n , (8) ations k in the async-ADMM are needed for convergence. It
2 2
is because each follower controller’s update is less informa-
tive with a smaller data subset. (ii) If there is a large number
λk+1
n ← λkn + ρk (An xk+1
B,n + Bn
k+1
zB,n ), (9) NeS of secondary networks exchanging information with the
primary network in the async-ADMM, the number of itera-
where the sync-ADMM necessitates the use of a global clock tions k required for convergence is reduced. This is because
k for both leader controller and follower controllers. The con- the primary network can collect more information from the
vergence and optimality analyses of this conventional sync- secondary networks in each iteration. (iii) If a large τn ex-
ADMM can be found in [27]. ists, due to the very infrequent information exchange between

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
6

Algorithm 1 Online and Asynchronous Implementations of by a symbol t in the distributed algorithm. Specifically, the in-
Distributed VVO-CVR stantaneous power and voltage measurements at time t − 1 are
1: Initialization: Set t = 0 and choose x(0), zn (0), n = used as the system feedback to estimate the nonlinear terms
1, . . . , NS . of power flow and ZIP load models at time t. In this paper,
2: repeat we assume a widespread coverage of meters throughout the
3: t ← t + 1. network. The leader and follower controllers have access to
4: If leader controller receives the newly updated zB,n the instantaneous measurements of line flow and voltage. 1
and λn from some follower controller n, then Mt ← Thus, the nonlinear terms εpij,φ,t , εqij,φ,t and εvi,φ,t in (4b)–(4d)
Mt−1 ∪ {n}. at time t can be estimated as constants with the system feed-
t t et ← λt , n ∈ Mt and zet
5: Let zeB,n ← zB,n ,λ n n B,n ← back measurements from previous time t − 1 as,
t−1 et t−1 t
zeB,n , λn ← λn , n ∈
e /M.
6: if |Mt | ≥ N eS then εpij,φ,t =Re{(Sij,φ,t−1
m m
vi,φ,t−1m
)
(vi,φ,t−1 m
− vj,φ,t−1 )},
7: Update xt+1 by (7) using zeB,nt
. (13)
8:
t+1
Send xB,n to follower controller n ∈ Mt . εqij,φ,t m m m m
=Im{(Sij,φ,t−1 vi,φ,t−1 ) (vi,φ,t−1 − vj,φ,t−1 )},
9: Reset Mt ← ∅. (14)
end if ∗ ∗
εvi,φ,t m m
 
10: = zij ((Sij,φ,t−1 ) (vi,φ,t−1 ) )
11: for every n ∈ N t do  ∗ m m

zij (Sij,φ,t−1 vi,φ,t−1 ) , (15)
12: Update znt+1 by (8).
13: Update λt+1
n by (9). m
where the Sij,φ,t−1 R
∈ 3×1 , vi,φ,t−1
m
∈ 3×1 and vj,φ,t−1
m
∈R
14:
15: end for
t+1
Send zB,n and λt+1
n to leader controller. R 3×1
are the instantaneous three-phase apparent power and
voltage measurements feedback from the system at time t − 1.
16: for every n ∈ / N t do Similarly, to handle the non-convexity due to the nonlinear
17: Let zn ← znt and λt+1
t+1
← λtn . √
n part vi,φ,t in active/reactive ZIP loads, we use the first-order
18: end for Talyor expansion to linearize it around the instantaneous volt-
19: Update ρt by (10)–(12). m
age measurements vi,φ,t−1 as,
20: Update reactive power output of inverters as per znt+1 .
21: Update the nonlinear terms εpij,φ,t , εqij,φ,t and εvi,φ,t by m
v̄i,φ,t =vi,φ,t−1
(13)–(15) with measurements feedback from the system. 1 m
22: Update the estimation of the nonlinear term v̄i,φ,t in + (vi,φ,t−1 )−1 m
(vi,φ,t − vi,φ,t−1 m
vi,φ,t−1 ),
2
ZIP loads (16)–(18) with measurements feedback from the (16)
system.
23: until t terminates.
√
R
where v̄i,φ,t ∈ 3×1 is the estimation of the nonlinear term
vi,φ,t . Therefore, the active and reactive ZIP loads in (4e)
and (4f) are re-written as follows,
the leader controller and follower controllers, a larger number p p p
of iteration k is needed for convergence. To further improve pZIP L
i,φ,t ' pi,φ,t (ki,1 · vi,φ,t + ki,2 · v̄i,φ,t + ki,3 ), (17)
ZIP L q q q
the convergence performance and capture fast system varia- qi,φ,t ' qi,φ,t (ki,1 · vi,φ,t + ki,2 · v̄i,φ,t + ki,3 ). (18)
tion of the async-ADMM, as well as make the performance
less dependent on the initial choice, we implement an iterative In this way, the above feedback-based linear approxima-
varying penalty update [27] as follows, tion method with online system measurements can make the
 inc k sub-problems of leader and follower controllers convex and
 τ ρ , if krk k2 > µksk k2 can be efficiently solved. Due to the distributed solution algo-
k+1
ρ := ρk /τ dec , if ksk k2 > µkrk k2 (10) rithm, the original large-scale centralized VVO-CVR problem
ρk , otherwise

is decomposed to several sub-problems for leader controller
where µ > 1, τ dec > 1 and τ inc > 1 are the updating param- of primary network and follower controllers of secondary net-
eters. The primal and dual residuals rnk and skn are calculated works, implying better a scalability. This is exactly an inherent
as, advantage of distributed optimization techniques. The detailed
procedure of the online async-ADMM is shown in Algorithm
rnk = An xkB,n + Bn zB,n
k
, ∀n (11) 1. The Mt denotes the set of follower controllers whose lo-

k+1
 cal updates have arrived at leader controller at iteration t and
skn = ρk ATn Bn zB,n k
− zB,n , ∀n. (12) N t denotes set of follower controllers that receives the newly
updated xB,n at iteration t. During the iteration, if the nth fol-
C. Online Implementation lower controller n ∈ / N t , which does not update the variable
at iteration t, then the values of xB,n , zB,n and λn and xB,n
To accurately track the fast variations of renewable genera- remain unchanged until the newly updated values come.
tion and load demand for better CVR performance, we address
the online implementation of the proposed distributed algo- 1 If line flow measurements are not available, one can approximately esti-
rithm. In this context, we directly represent the iteration index mate them through the linearized power flow model.

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
7

V. C ASE S TUDIES S
S
S S
A. Simulation Setup
S
S
A real-world distribution feeder located in Midwest U.S. S
S
S
[28] in Fig. 3 is used to illustrate our proposed scheme. This S S
S S
real feeder is shared by our utility partner, which consists of S S
one primary network and forty-four secondary networks. The S S
S S
primary network is denoted by overhead lines (blue) and un- S
S S
derground lines (red), and the secondary network is denoted S
S S S
by a circled capital letter S. Each secondary network includes S S
S S S
S
a service transformer, a secondary circuit with multiple cus- S S
S
S S
tomers and DERs. We have two reasons for choosing this real S S S
S
distribution feeder as the test system: (i) The real distribution S S
grid model [28] is an integrated primary-secondary distribu-
tion, which can be used to verify our proposed distributed CVR S
model. While most of the IEEE standard distribution systems,
such as IEEE 13-bus system and IEEE 123-bus system, only
have primary network. (ii) Customers in the real distribution
grid model [28] are equipped with smart meters, which can
help us to achieve the proposed online feedback-based linear
approximation method.
The time-series multiplier of load demand and solar power
with 1-minute time resolution are shown in Fig. 4. In the
Overhead Line
case study, PV smart inverters are installed in the secondary Underground Cable
networks and the total capacity of PV can serve 30% load. Substation S Secondary Network
The base voltages in the primary distribution network and the
secondary networks are 13.8 kV and 0.208 kV, respectively. Fig. 3. A real primary-secondary distribution feeder in Midwest U.S. [28],
consisting one MV primary network and forty-four LV secondary networks.
The base power value is 100 kVA. The selected parameters
for simulations are summarized in Table I, where the choice
of hyper-parameters depends on cross-validation. In general, 1
Load
a bad choice of hyper-parameter will affect the convergence PV
speed and the results. For example, a very large value of the 0.8
initial penalty factor ρ may lead to a sub-optimal solution,
Multipler

while a too small value of ρ will cause a slow convergence 0.6

speed. The choice of updating factor µ has the similar im-
pacts on convergence speed and results. In Table I, the ZIP 0.4
coefficients of active and reactive loads follow [29].
0.2
We develop a simulation framework in MATLAB R2019b,
which integrates YALMIP Toolbox with IBM ILOG CPLEX
0
12.9 solver for optimization, and the Open Distribution System 00:00 04:00 08:00 12:00 16:00 20:00 24:00
Simulator (OpenDSS) for power flow analysis. The OpenDSS Time [hr]
can be controlled from MATLAB through a component ob-
ject model interface, allowing us to carry out the feedback-
based linear approximation, performing power flow calcula- Fig. 4. Time-series multipliers of load demand and PV power.
tions, and retrieving the feedback results. In this section, we
present the convergence analysis to show the impact of asyn-
chronous update on convergence speed. We also demonstrate problem is solved by the proposed distributed method with
the effectiveness of our proposed method through numerical asynchronous updates. The performance testing for different
evaluations on several benchmarks to study load consumption numbers of secondary networks (follower controllers) in the
reduction through CVR implementation: (i) The base case is asynchronous distributed algorithm will be presented, where
generated by setting the unity-power factor control mode for the secondary networks are random selected in each iteration
all PV inverters where no additional reactive power support is to imitate the possible communication failure or delay in the
considered. (ii) The VVO-CVR problem is solved by a cen- practical cases. For example, if the number of secondary net-
tralized solver, where the nonlinear terms εpij , εqij and εvij in works (follower controllers) is set to be 20 in the asynchronous
power flow equations are neglected. (iii) The VVO-CVR prob- implementation, it will have 20 follower controllers to update
lem is solved by the proposed distributed method, which re- and communicate with the leader controller in each iteration.
quires globally synchronous updates between the leader con- The rest of follower controllers, which are not selected, will
troller and all the follower controllers. (iv) The VVO-CVR remain unchanged in this iteration.

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
8

TABLE I 2
S ELECTED PARAMETERS No failure
0 Failure during 30th -50 th iteration
Description Notion Value

log(residual)
Initial penalty factor ρ 0.05 -2 Failure of leader controller
Updating factor µ 10
-4
Increasing/Decreasing factor τ inc ,τ dec 5,5
Active load ZIP Coefficients k1p , k2p , k3p 0.96,−1.17,1.21
-6
Reactive load ZIP Coefficients k1q , k2q , k3q 6.28,−10.16,4.88
-8
5 -10
10 20 30 40 50 60 70 80 90 100
Iteration
0

Fig. 6. Convergence speed of the proposed distributed method by considering

log(residual)

the potential failure of the primary network (leader controller).

-5

The distributed leader-follower methods may suffer from the

-10 reliability issues when considering the potential failure of the
leader controller. To show the impacts of the potential failure
async (1/44 follower)
async (10/44 followers)
of the primary network (leader controller), the convergence
-15
async (20/44 followers) speeds of normal communication and communication failure
async (30/44 followers) of primary network (leader controller) are compared. In this
sync (44/44 followers)
case, we assume that the primary network (leader controller)
50 100 150 200 250 300 350 400 450 500 could have communication failure by not updating its own sub-
Iteration problem and communicating with secondary networks (fol-
lower controllers) during 30th to 50th iteration, then recover
the communication at 51st iteration. In Fig. 6, it can be ob-
Fig. 5. Convergence speed of the proposed distributed method with syn- served that the overall convergence speed is still acceptable
chronous and asynchronous implementation.
even the primary network (leader controller) fails to update
and communicate for 20 iterations. Therefore, our proposed
B. Convergence Analysis distributed algorithm is still efficient for certain level of com-
munication failure of primary network (leader controller).
The logarithm values of the norm of primal residuals (11)
with synchronous and different asynchronous communication
settings are illustrated in Fig. 5, which can be considered as C. Effect of Online Feedback-Based Approximation
one indicator of the convergence speed for the synchronous To show the effect of online feedback measurements, we
and asynchronous updates with different numbers of secondary solve the VVO-based CVR problem at a fixed point (at 19:00)
networks (follower controllers). It can be observed that, if there with different control strategies in centralized and distributed
is no communication failure or delay, the proposed distributed manners. The iterative objective function values (the active
algorithm with the standard ADMM can achieve the best con- power flow through substation) are recorded in Fig. 7. Even
vergence speed; the asynchronous implementation with 20 or though the difference of the objective solutions between the
30 activated secondary networks (follower controllers) can still centralized solver (blue dashed line) and the proposed dis-
guarantee the convergence with an acceptable speed; while the tributed method (red line) is about 0.26% after nearly 50 iter-
performance of convergence with 10 or even less secondary ation, the proposed method can still achieve a better result than
networks (follower controllers) are not as good as other cases. the centralized method. It is because the proposed distributed
Hence, there is a trade-off between the work stress/need on method can use measurements feedback from the system to
communication system and the convergence performance. The approximate the nonlinear terms successively, while the cen-
principle of partial barrier is balancing the trade-off between tralized method neglects the nonlinear terms.
the work stress/need on communication system and the per- To show the effect of approximation of the nonlinear part
√
formance of convergence. In our case, the threshold of the vi,φ,t in (16), we calculate the difference between the ac-
number of secondary networks (follower controllers) is 20 to curate ZIP load and the approximate ZIP load with a given
maintain the calculation accuracy. Here, the acceptable speed time series voltage (1-minute time resolution). The accurate
can be quantified as: if the primal residuals is lower than 10−3 ZIP load at time t is calculated based on the original ZIP
within 30 iterations, then we consider the convergence speed load model (4e)–(4f) with the instantaneous voltage at time t.
is acceptable. Keep in mind that the thresholds may vary in While the approximate ZIP load is estimated based on (16)–
different cases, which should be adjusted accordingly. (18) with the voltage measurement of previous time t − 1. In

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
9

11.56
Objective function [p.u.]

Voltage [p.u.]
11.55 0.98

11.54
w/o secondary
w/ secondary
11.53 0.95
w/o feedback Grid-edge voltage 0.9337 p.u. violates the lower
limit by 1.3%
w/ feedback
11.52 0 100 200 300 400 500
10 20 30 40 50 Node
Iteration

Fig. 9. Nodal voltage profiles with and without the secondary networks.
Fig. 7. Objection function values under a fixed-point test.

10-5 in the optimal CVR. However, the grid-edge voltage within

1 one secondary network is 0.9377 p.u., which violates the volt-
age lower limit 0.95 p.u. by 1.3%, if we only consider the
0.5 primary network and aggregate secondary networks as nodal
injections.
Diff.

0
E. Reactive Power output of Smart Inverters
In this test case, there are forty-four secondary networks,
-0.5
and each secondary network are installed with two smart in-
verters, one in the middle and one in the end of the secondary
-1 network. Note that the optimal position and sizing of inverter
00:00 04:00 08:00 12:00 16:00 20:00 24:00 are not included in the scope of this work. To show the reac-
Time [hr] tive power of inverters in a clear way, we select two inverters
as examples with different reactive power behaviors. As shown
Fig. 8. Difference between the accurate ZIP load and the approximate ZIP
in Fig. 10, the inverter 1 (blue curve) is installed in the end of
load. the secondary network, where the reactive power injections are
always required to maintain the voltage above the lower volt-
age limit; while the inverter 2 (red dashed curve) is installed
Fig. 8, it can be observed that if the voltage difference be- in the middle of the secondary network, where the reactive
tween t and t − 1 is not large, then the differences between power injection and absorbing are both required to maintain
the accurate ZIP load and approximate ZIP load are ranging the voltage within predefined voltage limits. Therefore, the
from −10−5 to 10−5 , which is acceptable. reactive power output of inverter will be affected by the in-
stallation positions. In our case, it is possible that the reactive
D. Grid-Edge Voltage Profile power outputs of inverter reach its capacity. For example, be-
cause the inverter 1 is installed in the end of a long secondary
In real distribution system, most loads and residential DERs
feeder, our proposed optimal CVR determines inverter 1 to in-
are connected to secondary networks. If the secondary net-
ject enough reactive powers, which satisfy both the reactive
works are simplified by using aggregate models in primary
power capacity constraints and voltage limit constraints.
network, it will hinder the performance of grid-edge voltage
regulation. To show the importance of considering detailed
models of secondary networks in CVR implementation, two F. Comparison Between Different Control Strategies
cases are presented: we solve the optimal CVR with and with- To show the time-series simulation, the VVO-CVR is
out considering detailed secondary network models, then input performed in a daily operation of the integrated primary-
the optimal reactive power dispatch results of smart inverters secondary distribution grid (with 1-minute time resolution)
in the distribution system to evaluate the CVR performance. with different control strategies in centralized and distributed
If the secondary networks are not considered in the optimal manners, respectively. Note that the online implementation of
CVR, the optimal reactive power setting at each primary node the async-ADMM method is used here, where the nonlinear
has to be proportionally distributed to PV inverters in the sec- terms of the network and load models are approximated with
ondary networks. The primary and secondary nodal voltage the power and voltage measurements feedback from the sys-
profiles of the two cases are presented in Fig. 9, respectively. tem with the last-minute dispatch. Existing studies [30], [31]
It can be observed that the grid-edge voltages can be well reg- have been conducted based on smart meters with 1-minute
ulated if both primary and secondary networks are considered time resolution. Therefore, the online implementation of the

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
10

0.02 TABLE II
E NERGY C ONSUMPTION R ESULTS W ITH D IFFERENT C ONTROL
Reactive power [p.u.]

S TRATEGIES
0.01
Energy (kWh) Reduction (%)
Base case (w/o control) 262,167.4 -
0
CCVR 227,269.9 13.3%
DSCVR 226,339.5 13.6%
-0.01 Inv 1 DACVR (20 followers) 227,325.1 13.2%
Inv 2
-0.02
00:00 04:00 08:00 12:00 16:00 20:00 24:00
The numerical comparisons of total energy consumption
Time [hr]
over one day and the energy reduction are presented in Table
II among the base case, CCVR, and distributed sync. CVR
Fig. 10. Reactive power output of two smart inverters as examples. (DSCVR) and DACVR with 20 followers. Compared to the
base case, the VVO-based CVR method can achieve the en-
450 ergy reduction around 13.2% to 13.6%. In theory, the differ-
CCVR ences between CCVR, DSCVR and DACVR shall be small,
400 DACVR (20 followers) because they are solving the similar VVO-CVR problems. The
Base case (w/o control) reasons why they do not have the exact same solution are: (i)
Substation Powers [kW]

350
Because of the missing nonlinear terms in power flow calcula-
300 tions, CCVR cannot obtain the accurate solution; (ii) DACVR
obtain the solution by receiving updates from limited num-
250 ber of secondary networks (follower controllers). Based on
200 the comparison between CCVR and DSCVR and DACVR, it
can be seen that the total energy consumption results from the
150 CCVR, DSCVR and DACVR are very similar, and DSCVR
yields slightly better results than other two cases. This is be-
100
cause DSCVR has the online power and voltage feedback
50 measurements from the system to accurately approximate the
nonlinear terms of the power flow calculations and ZIP load
00:00 04:00 08:00 12:00 16:00 20:00 24:00
models. While the nonlinear terms εpij,φ,t , εqij,φ,t and εvi,φ,t are
Time [hr]
neglected in CCVR, this offline linear approximation method
may bring inaccurate power flow and bus voltage computa-
Fig. 11. Substation feed-in active power with different control strategies. tions, consequently, hindering the CVR performance. The en-
ergy reduction of DACVR is also slightly less than DSCVR,
because DACVR only receives updates from limited number of
async-ADMM method can by achieved by using 1-minute time follower controllers, while DSCVR can receive updates from
resolution measurements sent by smart meters. We also as- all follower controllers. It is concluded that DACVR can still
sume that the change of the system is not that large within obtain a good energy reduction performance with updates from
1-minute, so that the measurements from the last-minute can limited number of follower controllers. Compared to CCVR,
still be used to approximate the nonlinear term for the next the advantages of the proposed DSCVR and DACVR can be
minute. summarized as follows: (i) The CCVR is disadvantageous on
The active power supplies from the substation of the base scalability, because CCVR must solve a large-scale VVO-CVR
case (without control), centralized CVR (CCVR) and dis- problem. With increasing size of decision models, the com-
tributed async. CVR (DACVR) with 20 secondary networks putation burden of CCVR increases extensively. While the
(follower controllers) are shown in Fig. 11. As can be ob- the proposed DSCVR and DACVR decompose the large-scale
served, the proposed method can effectively reduce the power problem into multiple small-scale sub problems, therefore, the
supply from substation, especially during the peak load pe- computation burden is reduced. (ii) In the proposed DSCVR
riod, e.g., 16:00–20:00. To verify the online performance of the and DACVR, the data privacy and ownership of customers are
proposed distributed method, we compare the time-series so- respected, including local consumption measurement data and
lutions of the CCVR (green curve) with DACVR with 20 fol- cost functions. However, CCVR requires the system-wide col-
lowers (purple dotted curve). It can be seen that, the DACVR lection of data, and a costly communication infrastructure to
with 20 followers can provide a similar control performance enable information passing between a control center and reg-
to CCVR. Therefore, when there are at least 20 follower con- ulation devices. (iii) Moreover, the CCVR are susceptible to
trollers updating and communicating with leader controller in single point of failure. While DACVR is resilient against agent
the asynchronous implementation, a good control performance communication failure or limited communication.
can be achieved. In Fig. 12, the 1440-minute time-varying voltage profiles

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
11

have validated the robustness and effectiveness of the pro-

posed method. According to the case studies, we have shown
that: (1) With a reasonable setting of asynchronous update,
the proposed async-ADMM method is able to guarantee the
convergence with acceptable speed. (2) Compared to using ag-
gregate models of secondary networks, the grid-edge voltages
can be better regulated with detailed secondary network mod-
els in the proposed CVR implementation. (3) With the online
feedback-based linear approximation, the proposed VVO-CVR
can achieve good performance of energy/voltage reductions
while maintaining voltage level in predefined ranges.

(a) Base case (w/o control) R EFERENCES

[1] American National Standard For Electric Power Systems and
Equipment-Voltage Ratings (60 Hertz). American National Standards
Institute C84.1, 2016.
[2] Z. Wang and J. Wang, “Review on implementation and assessment of
conservation voltage reduction,” IEEE Trans. Power Syst., vol. 29, no. 3,
pp. 1306–1315, May 2014.
[3] D. Kirshner, “Implementation of conservation voltage reduction at com-
monwealth edison,” IEEE Trans. Power Syst., vol. 5, no. 4, pp. 1178–
1182, Nov. 1990.
[4] Z. Wang, M. Begovic, and J. Wang, “Analysis of conservation voltage
reduction effects based on multistage SVR and stochastic process,” IEEE
Trans. Smart Grid, vol. 5, no. 1, pp. 431–439, Jan. 2014.
[5] J. Wang, C. Chen, and X. Lu, “Guidelines for implementing ad-
vanced distribution management systems-requirements for DMS inte-
gration with DERMS and microgrids,” Argonne National Lab.(ANL),
Argonne, IL (United States), Tech. Rep., 2015.
(b) DACVR with 20 followers. [6] Q. Shi, W. Feng, Q. Zhang, X. Wang, and F. Li, “Overvoltage mitigation
through volt-var control of distributed PV systems,” in Proc. IEEE Power
Fig. 12. Voltage profiles with different control strategies (each line represents and Energy Society Tranmission and Distribution (T&D), 2020, pp. 1–5.
a phase-wise voltage magnitude of a bus). [7] F. Ding, A. Nagarajan, S. Chakraborty, M. Baggu, A. Nguyen,
S. Walinga, M. McCarty, and F. Bell, “Photovoltaic impact assess-
ment of smart inverter volt-var control on distribution system conserva-
tion voltage reduction and power quality,” National Renewable Energy
of the base case and DACVR with 20 followers are com- Lab.(NREL), Golden, CO (United States), Tech. Rep., 2016.
pared. Each line represents a phase-wise voltage magnitude [8] T. V. Dao, S. Chaitusaney, and H. T. N. Nguyen, “Linear least-squares
of a bus. As shown in Fig. 12 (a), where there is no reac- method for conservation voltage reduction in distribution systems with
photovoltaic inverters,” IEEE Trans. Smart Grid, vol. 8, no. 3, pp. 1252–
tive power control in the base case, there are voltage viola- 1263, Mar. 2017.
tions of the lower limit 0.95 p.u., during the heavy-load peri- [9] M. S. Hossan and B. Chowdhury, “Integrated CVR and demand response
ods, e.g., 16:00–20:00. On the other hand, when the CVR is framework for advanced distribution management systems,” IEEE Trans.
Sustain. Energy, vol. 11, no. 1, pp. 534–544, Feb. 2020.
implemented with optimal reactive power control, the system [10] F. Ding and M. Baggu, “Coordinated use of smart inverters with legacy
achieves maximum voltage reduction while maintains voltage voltage regulating devices in distribution systems with high distributed
levels with the predefined range [0.95,1.05] p.u., as shown in pv penetration — increase CVR energy savings,” IEEE Trans. Smart
Grid, pp. 1–1, To be published 2018.
Fig. 12 (b). [11] C. Feng, Z. Li, M. Shahidehpour, F. Wen, W. Liu, and X. Wang, “Decen-
tralized short-term voltage control in active power distribution systems,”
IEEE Trans. Smart Grid, vol. 9, no. 5, pp. 4566–4576, Sept. 2018.
VI. C ONCLUSION [12] Q. Zhang, K. Dehghanpour, and Z. Wang, “Distributed CVR in unbal-
To better regulate voltages at the grid-edge while imple- anced distribution systems with PV penetration,” IEEE Trans. Smart
Grid, vol. 10, no. 5, pp. 5308–5319, Sept. 2019.
menting CVR in distribution system, a distributed VVO- [13] B. Zhou, D. Xu, K. W. Chan, C. Li, Y. Cao, and S. Bu, “A two-stage
CVR algorithm is developed to optimally coordinate the framework for multiobjective energy management in distribution net-
smart inverters in unbalanced three-phase integrated primary- works with a high penetration of wind energy,” Energy, vol. 135, pp.
754–766, Sept. 2017.
secondary distribution systems. In order to handle the non- [14] B. A. Robbins and A. D. Domı́nguez-Garcı́a, “Optimal reactive power
convexity of power flow and ZIP load models, a feedback- dispatch for voltage regulation in unbalanced distribution systems,” IEEE
based linear approximation method has been proposed to suc- Trans. Power Syst., vol. 31, no. 4, pp. 2903–2913, July 2016.
[15] G. Qu and N. Li, “Optimal distributed feedback voltage control under
cessively estimate the nonlinear terms in these models. An limited reactive power,” IEEE Trans. Power Syst., vol. 35, no. 1, pp.
ADMM-based distributed framework is established to solve 315–331, Jan. 2020.
the optimal CVR problem in a leader-follower distributed fash- [16] R. R. Jha, A. Dubey, T. Hong, and D. Zhao, “Distributed algorithm for
volt - var optimization in unbalanced distribution system,” in 2020 IEEE
ion, where the primary system corresponds to the leader con- Power Energy Society Innovative Smart Grid Technologies Conference
troller and each secondary system corresponds to a follower (ISGT), May 2020, pp. 1–5.
controller. We further address its asynchronous implementa- [17] W. Zheng, W. Wu, B. Zhang, H. Sun, and Y. Liu, “A fully distributed
reactive power optimization and control method for active distribution
tion with a frozen strategy that allows asynchronous updates. networks,” IEEE Trans. Smart Grid, vol. 7, no. 2, pp. 1021–1033, Mar.
Simulation results on a real Midwest U.S. distribution feeder 2016.

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2021.3088010, IEEE
Transactions on Smart Grid
12

[18] P. Šulc, S. Backhaus, and M. Chertkov, “Optimal distributed control of Zhaoyu Wang (S’13–M’15–SM’20) is the Harpole-
reactive power via the alternating direction method of multipliers,” IEEE Pentair Assistant Professor with Iowa State Univer-
Trans. Energy Convers., vol. 29, no. 4, pp. 968–977, Dec. 2014. sity. He received the B.S. and M.S. degrees in elec-
[19] Y. Wang, T. Zhao, C. Ju, Y. Xu, and P. Wang, “Two-level distributed trical engineering from Shanghai Jiaotong Univer-
volt/var control using aggregated pv inverters in distribution networks,” sity, and the M.S. and Ph.D. degrees in electrical
IEEE Trans. Power Deliv., vol. 35, no. 4, pp. 1844–1855, Aug. 2020. and computer engineering from Georgia Institute of
[20] T. Xu and W. Wu, “Accelerated ADMM-based fully distributed inverter- Technology. His research interests include optimiza-
based volt/var control strategy for active distribution networks,” IEEE tion and data analytics in power distribution systems
Trans. Industr. Inform., vol. 16, no. 12, pp. 7532–7543, Aug. 2020. and microgrids. He is the Principal Investigator for
[21] Y. Chai, L. Guo, C. Wang, Z. Zhao, X. Du, and J. Pan, “Network parti- a multitude of projects focused on these topics and
tion and voltage coordination control for distribution networks with high funded by the National Science Foundation, the De-
penetration of distributed PV units,” IEEE Trans. Power Syst., vol. 33, partment of Energy, National Laboratories, PSERC, and Iowa Economic De-
no. 3, pp. 3396–3407, May 2018. velopment Authority. Dr. Wang is the Chair of IEEE Power and Energy Soci-
[22] S. Magnússon, G. Qu, and N. Li, “Distributed optimal voltage control ety (PES) PSOPE Award Subcommittee, Co-Vice Chair of PES Distribution
with asynchronous and delayed communication,” IEEE Trans. Smart System Operation and Planning Subcommittee, and Vice Chair of PES Task
Grid, vol. 11, no. 4, pp. 3469–3482, July 2020. Force on Advances in Natural Disaster Mitigation Methods. He is an editor
[23] H. J. Liu, W. Shi, and H. Zhu, “Hybrid voltage control in distribution of IEEE Transactions on Power Systems, IEEE Transactions on Smart Grid,
networks under limited communication rates,” IEEE Trans. Smart Grid, IEEE Open Access Journal of Power and Energy, IEEE Power Engineering
vol. 10, no. 3, pp. 2416–2427, May 2019. Letters, and IET Smart Grid. Dr. Wang was the recipient of the National Sci-
[24] Q. Zhang and M. Sahraei-Ardakani, “Impacts of communication limits ence Foundation (NSF) CAREER Award, the IEEE PES Outstanding Young
on convergence of distributed DCOPF with flexible transmission,” in Engineer Award, and the Harpole-Pentair Young Faculty Award Endowment.
Proc. North American Power Symposium (NAPS), 2017, pp. 1–6.
[25] R. Zhang and J. Kwok, “Asynchronous distributed ADMM for consensus
optimization,” in Proc. International conference on machine learning,
2014, pp. 1701–1709.
[26] “IEEE standard for interconnection and interoperability of distributed en-
ergy resources with associated electric power systems interfaces,” IEEE
Std 1547-2018 (Revision of IEEE Std 1547-2003), pp. 1–138, 2018.
[27] S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein et al., “Distributed
optimization and statistical learning via the alternating direction method
of multipliers,” Found. Trends Mach. Learn., vol. 3, no. 1, pp. 1–122,
2011.
[28] F. Bu, Y. Yuan, Z. Wang, K. Dehghanpour, and A. Kimber, “A time-
series distribution test system based on real utility data,” in Proc. North
American Power Symposium (NAPS), 2019, pp. 1–6.
[29] K. Schneider, J. Fuller, F. Tuffner, and R. Singh, “Evaluation of conser-
vation voltage reduction (CVR) on a national level,” Pacific Northwest
Nat. Laboratory (PNNL) Tech. Rep, Sept. 2010.
[30] Y. Wang, Q. Chen, T. Hong, and C. Kang, “Review of smart meter data
analytics: Applications, methodologies, and challenges,” IEEE Trans.
Smart Grid, vol. 10, no. 3, pp. 3125–3148, 2019.
[31] R. Granell, C. J. Axon, and D. C. H. Wallom, “Impacts of raw data tem-
poral resolution using selected clustering methods on residential elec-
tricity load profiles,” IEEE Trans. Power Syst., vol. 30, no. 6, pp. 3217–
3224, 2015.
Fankun Bu (S’18) received the B.S. and M.S. de-
grees from North China Electric Power University,
Baoding, China, in 2008 and 2013, respectively.
From 2008 to 2010, he worked as a commissioning
engineer for NARI Technology Co., Ltd., Nanjing,
Qianzhi Zhang (S’17) is currently pursuing his China. From 2013 to 2017, he worked as an electri-
Ph.D. in the Department of Electrical and Computer cal engineer for State Grid Corporation of China at
Engineering, Iowa State University, Ames, IA. He Jiangsu, Nanjing, China. He is currently pursuing his
received his M.S. in electrical and computer engi- Ph.D. in the Department of Electrical and Computer
neering from Arizona State University in 2015. From Engineering, Iowa State University, Ames, IA. His
2015 to 2016, he worked as a research engineer with research interests include distribution system mod-
Huadian Electric Power Research Institute. His re- eling, smart meter data analytics, renewable energy integration, and power
search interests include the applications of machine system relaying.
learning and advanced optimization techniques in
power system operation and control.

Yifei Guo (M’19) received the B.E. and Ph. D. de-

grees in electrical engineering from Shandong Uni-
versity, Jinan, China, in 2014 and 2019, respectively.
Currently, he is a Postdoctoral Research Associate
with the Department of Electrical and Computer En-
gineering, Iowa State University, Ames, IA, USA.
He was a visiting student with the Department of
Electrical Engineering, Technical University of Den-
mark, Lyngby, Denmark, in 2017–2018.
His research interests include voltage/var control,
renewable energy integration, wind farm control, dis-
tribution system optimization and control, and power system protection.

1949-3053 (c) 2021 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.
Authorized licensed use limited to: Tencent. Downloaded on June 30,2021 at 09:32:05 UTC from IEEE Xplore. Restrictions apply.