1

Hierarchical Hybrid Control Architecture for Volt/Var

Management of Smart Power Distribution Grids
Ahmad R. Malekpour, Member, IEEE, Anuradha M. Annaswamy, Fellow, IEEE, and Jalpa Shah, Member, IEEE
 introduces several problems associated with power quality such
Abstract—This paper presents an efficient, actively coordinated, as voltage rise [1], line overloading, reverse flow [4], voltage
hierarchical control architecture for Volt/Var management of fluctuation [5], harmonics, and flickers [6]. The legacy devices
distribution grids in the presence of high penetration of such as LTC, VR, and CB alone do not suffice in ensuring
distributed generators (DGs). The architecture includes three satisfactory VVC, and necessitate addition of several layers of
layers with specific operational goals associated with each layer defense. The first such layer is the use of smart inverters, which
based on data resolution, communication network, and control
can provide a suite of advanced functionalities such as reactive
device response time. The top layer carries out central
optimization and seeks to minimize power losses via optimal daily power support to improve grid voltage stability and control.
scheduling of slow response legacy devices such as load tap However, smart inverters alone do not suffice, as they are not
changers (LTCs), voltage regulators (VRs), and capacitor banks equipped for fast responses to ramping behavior of solar PV.
(CBs), and DGs. The middle layer ensures voltage regulation using Moreover, transient cloud cover can rapidly reduce solar power
a distributed approach for reactive power provisioning by fast generation, cause temporary voltage drop and tap changer
response DGs. Finally the bottom layer carries out local decision cycling. Given the current growth of DG in the U.S. [1] as well
making that enlists edge intelligence to operate LTCs, VRs and as the projected growth of renewable DGs such as wind and
CBs to cope with fast and real-time changes in DGs and loads. The solar resources [4] and [5], in addition to the use of legacy
top layer generates overall references for the lower layers
devices and smart inverters, a well-coordinated control of DG
optimized over a 24-hour horizon, which are then suitably
corrected at faster time-scales (at 15-minute intervals in the will play a significant role in mitigating the issues and
middle layer and 2-sec interval at the bottom layer) to address providing volt/var support. This paper proposes a novel
voltage deviations at multiple time-scales. Appropriate hierarchical-hybrid networked control architecture for VVC in
coordination is introduced between each layer so as to meet distribution grids.
combined goals of optimization and accommodation of load and Over the years, various methods have been proposed for
generation uncertainties in real-time. The proposed method is VVC, which we broadly classify into centralized, distributed,
validated using a modified IEEE 34 bus test feeder and at 80% local, and hybrid methods. The approaches adopted for the
penetration of DGs. centralized class rely on: optimal power flow (OPF) techniques
Index Terms— Power distribution systems, power quality, applied to a snapshot [7]-[10], or a time-series simulation [11]-
reactive power control, renewable energy, smart grid, solar power. [13] using mixed integer linear programming (MILP) [7],
I. INTRODUCTION sequential convex programming (SCP)[6], interior point
method [9], quasi-Newton method [10], gradient descent
O ne of the fundamental operating requirements of power
distribution system is volt/var control (VVC) where its
primary purpose is to improve voltage profiles along the
methods [11], evolutionary algorithms [12], [13] model
predictive control [14], and nonlinear programming (NLP) [15].
In [7], mixed integer linear programming (MILP) is used to
feeder and reduce real power losses under all loading formulate the volt/var optimization including transformer load
conditions. Traditionally, utilities use load tap-changer (LTC) tap changers (LTC), switchable capacitors, and reactive power
transformer, step voltage regulator (SVR) and capacitors for of DGs. In [8], three-phase OPF using sequential convex
reactive power and voltage control in distribution systems as programming (SCP) is introduced to solve the volt/var control
shown in Figure 1. LTCs, located in distribution substations, are problem where the goal was to minimize reactive power of DG
the most common devices to regulate voltage. LTCs are tap- units while satisfying operational constraints. In [9], primal-
changing autotransformers designed to regulate voltage if it dual interior point method for three phase electrical system
violates the pre-set limits. Typically, SVRs are located optimization was used, in which the phase unbalances were
downstream of the substation. They are also tap-changing taken into account. In [10], OpenDSS for three-phase load flow
autotransformers designed to regulate voltage. Capacitors are analysis and quasi-Newton method were used to perform OPF
common reactive power compensators in distribution systems in unbalanced networks. In [11], an optimal dispatch of DGs
and can be found in both substation and distribution feeders. and loss minimization via switching shunt capacitors were
LTC and SVR are designed to change positions a few times a formulated using gradient descent methods. In [12], Genetic
day to regulate the voltage with respect to variations in the load. Algorithm (GA) was used to minimize the power losses and
High penetration of solar on distribution grid can introduce regulate the voltage by finding the optimal scheduling of LTC
significant challenges in carrying out efficient VVC, as it

This work was supported by Eaton Corporation. Annaswamy are with Department of Mechanical Engineering, Massachusetts
Ahmad R. Malekpour was with with Department of Mechanical Institute of Technology, Cambridge, MA 02139, USA ([email protected])
Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, Jalpa Shah is with Eaton Corporation Inc., Eden Prairie, MN 55344, USA
USA ([email protected]). Curently, he is the Manager of Grid Strategy and ([email protected])
Analytics at ComEd.([email protected]);. Anuradha
 2

and step voltage regulators (SVRs). Probabilistic volt/var involving the sensitivity theory in conjunction with the
optimization was formulated in [13] considering the uncertainty Lyapunov function. In [23], an automatic distributed voltage
associated with renewable DGs via applying Particle Swarm control algorithm based on sensitivity approach is used to
Optimization (PSO). In [14], model predictive control was used control the node voltages regulating the reactive power injected
to regulate distribution network voltages in the presence of high by DGs. A sensitivity-based decentralized local active and
penetration of DG. In [15], nonlinear programming (NLP) was reactive power control of DGs to overcome voltage and thermal
proposed to formulate OPF including LTC and capacitor issues near the point of connection was presented in [24]. The
operation. In [16], multi-agent systems (MAS) were adopted to interaction of DG and LTC was not modeled and explored. In
solve VVC problems. In all of the above papers which consist [25], a distributed local reactive power control was introduced
of centralized decision-making, it is assumed that problem to mitigate voltage rise due to DG integration in distribution
information, cost parameters, objectives, and decision authority systems with occasional communication with the distribution
are given to an entity such as a distribution system operator system operator. The approach increased feeder losses as well
(DSO). And the DSO achieves the optimal solution using a as stress in tap changing of LTC. Local approaches impose no
centralized OPF method. However, with high proliferation of or minimum communication burden between control devices.
DGs in medium- and large-scale power distribution system, Therefore, they are very simple and the least expensive to
solving centralized OPF needs significant investment in meters, implement. However, they are inherently locally optimal due to
communications, and control system infrastructure. the lack of full system information.
The second class of consists of distributed approaches, where The fourth method is hybrid in nature, and combines one or
the entire distribution network is split into sub-networks with more of the above methods. For example, combined local and
production and consumption capabilities. An optimization remote voltage and reactive power control to minimize
problem is formulated for each sub-network in order to manage distribution system losses in the presence of induction DG
the resources and make decision autonomously through machines was presented in [26] where LTC and substation
bidirectional power exchange with its adjacent subnetworks. capacitors were adjusted remotely using a dynamic
Several approaches have been proposed to develop distributed programming approach. The approach however was not
algorithms for optimal power flow (OPF) and/or controlling successful in curtailing the voltage fluctuations due to
DGs. Dual-decomposition distributed algorithm with gradient renewable integration and also resulted in increased OLTC and
ascent was proposed in [16] for voltage regulation, where it was cap switching during the study. Combined central and local
shown to solve a convex relaxation of power flow equations. In active and reactive power control of PV inverters was presented
[17], [18], two decentralized optimization frameworks were in [27] where a local controller was incorporated into
developed by leveraging the dual-ascent and alternating centralized optimization to reduce the search space and
direction method of multipliers (ADMM) to solve a computation complexity. In [28], a predefined control scheme
semidefinite programming relaxation of OPF. The goal was to is suggested to coordinate LTC, capacitors and DGs in which
determine the active and reactive power set points of DGs. The distribution grid is divided into zones with individual voltage
approach partitioned distribution network into clusters and regulation and reactive support schemes. The approach does not
optimization was carried out under the assumption that address the fact that if DGs adjust reactive power output before
customer to utility, customer to control center, and control OLTC, VR, capacitor, the number of OLTC/VR/Capacitor
center to control center communications in each cluster were operation can be reduced. Also, operational scenarios for longer
available. In [19], two distributed approaches, dual-ascent and period were not studied to see if voltage regulator device setting
consensus-based ADMM algorithms, were proposed to control needs to be changed daily/weekly/monthly.
reactive power from PV inverters. The approaches require that The method that we propose in this paper is hybrid in nature,
each node has its own set of constraints, its own objective and combines the three approaches mentioned above including
function and neighboring node to node message passing a centralized, a distributed and a local decision-making
capability. Hierarchical optimization for DG control without component. A hierarchical structure is imposed on this
coordination formulation between the levels were presented in combination so that the best feature of each of these three
[20], [21]. These approaches grouped under the second class components is retained by the overall hierarchical-hybrid
can reduce the computational burden and communication needs network (HHN) of decision-making. The main advantage of the
by sharing the optimization task among multiple control centers proposed HHN is its ability to accommodate a large number of
within each network and enables large-scale implementation. DGS. One of the main challenges of the ongoing grid
However, it requires advanced metering infrastructure (AMI) modernization, especially in the context of distribution grids, is
as well as multiple control centers within the distribution grid the increased presence of DGs. As DGs introduce distributed
to perform the optimization which makes it expensive for field reactive power injection, the spatio-temporal management of
implementation. Also there may not be a feasible solution that voltage fluctuations and therefore volt/var control becomes
meets all of the constraints and solves the objective functions highly nontrivial. We will show that the proposed HHN can
chosen at various levels. effectively carry out VVC.
The third class can be categorized as local methods. These The top layer of the proposed HHN is a centralized optimizer.
approaches are based on local measurements and limited This provides an overall optimization framework and helps
communication between voltage control devices. For example, design the HHN so as to minimize the losses in the distribution
in [22], reactive power provisioning of PV inverters was locally grid. The second feature of the proposed HHN is its use of a
optimized to reduce the power losses in the network. The distributed layer based on the concept of consensus. Consensus
control design was based on an optimization procedure is the problem of finding an agreement among autonomous
 3

entities such as people, autonomous software agents or

computers in a network. Assuming that each of these entities
$i$ has a state-variable $x_i$, the idea is for all states to arrive
at a consensus of having the same value and to accomplish this
goal while communicating directly with only a subset of the
other agents. Consensus problems have been studied in
computer science [14], physics [15], operations research [16]
and control theory [17] in a very successful manner, especially
when the underlying network is large and subject to
uncertainties. The third layer of the HHN is edge-intelligence
and is enabled to through smart local devices.
The common utility practice for voltage regulation in LV
network is local control of slow-response LTC, VR, and CB,
i.e. edge control of the grid, based on engineering judgments
and archived information [2]. For fast-response PV inverter,
several grid edge reactive power provisioning methods have
been suggested, including 𝑐𝑜𝑠𝜑(𝑃), 𝑄(𝑉), and 𝑄(𝑉)/𝑃(𝑉)
strategies [8], [9]. In the 𝑐𝑜𝑠𝜑(𝑃) strategy, inverter feed-in Fig. 1. Hierarchical architecture
power is monitored; once the power exceeds a predetermined
limit, the PV inverter starts to absorb reactive power at a A. Centralized Optimization Formulation
constant rate. In the 𝑄(𝑉) strategy, a droop characteristic based Let 𝜑𝑏 indicates the set of totally 𝑁 nodes in distribution grid
on inverter bus voltage is used to calculate PV inverter reactive and 𝜑𝐷𝐺 , 𝜑𝐶 , and 𝜑𝑉𝑅 denote the set of nodes in 𝜑𝑏 that include
power. We employ such a droop-strategy through the use of DGs, cap banks, and LTC/VR, respectively. Mathematical
smart inverters in the third and final layer of the HHN. optimization for the layer one is formulated as
𝐻 𝑁
II. APPROACH min 𝐹 = ∑ ∑ 𝑟𝑖𝑗 𝑙𝑖𝑗 [𝐾] (1)
The proposed hierarchical hybrid approach includes three 𝑞𝑖𝐺,𝑆𝑖 ,𝑇𝑖𝑗
𝑘=1 𝑖=1
control layers where each layer operates in a different time- s.t.
scale and interacts with LTC, VRs, CBs, and DGs in physical
Distribution power flow and voltage constraints, ∀𝑖, 𝑗 ∈ 𝜑𝑏 :
layer. Fig. 1 shows the control and physical layers as well as the
information flow and voltage measurements signals for the ∑ 𝑃𝑚𝑖 [𝐾] = 𝑃𝑖𝑗 [𝐾] + 𝑝𝑖𝐷 [𝐾] − 𝑝𝑖𝐺 [𝐾] (2)
reliable operation of the grid. 𝑚,𝑖,𝑗∈𝜑𝑏
Layer one receives a day-ahead 15 minute load and ∑ 𝑄𝑚𝑖 [𝐾] = 𝑄𝑖𝑗 [𝐾] + 𝑞𝑖𝐷 [𝐾] − 𝑞𝑖𝐺 [𝐾] (3)
generation forecasted data and calculates the on/off status of
𝑚,𝑖,𝑗∈𝜑𝑏
cap banks, tap operation of LTC/VRs, and reactive power
𝑉𝑗 [𝐾] = 𝑉𝑖 [𝐾] − (𝑟𝑖𝑗 𝑃𝑖𝑗 [𝐾] + 𝑥𝑖𝑗 𝑄𝑖𝑗 [𝐾])/𝑉1 (4)
provisioning from DGs for the next 24 hours. As shown in Fig. 2 2
1, these setting will be communicated to layers - two and three. 𝑙𝑖𝑗 [𝐾] = ((𝑃𝑖𝑗 [𝐾]) + (𝑄𝑖𝑗 [𝐾]) ) /𝑉1 (5)
In layer two, each DG receives the setting from the top layer, 𝑉𝑖 ≤ 𝑉𝑖 [𝐾] ≤ 𝑉𝑖 (6)
measures the voltage at its terminal voltage, and determines the 2 2 2
required reactive power for local voltage regulation based on (𝑃𝑖𝑗 [𝐾]) + (𝑄𝑖𝑗 [𝐾]) ≤ (𝑆𝑖𝑗 ) (7)
droop-control method. If the voltage is higher or lower than Constraints on DG operation, ∀𝑖 ∈ 𝜑𝐷𝐸𝑅 :
predefined upper/lower critical voltages, DG requests for 𝑝𝑖 < 𝑝𝑖𝐺 [K] < 𝑝𝑖 (8)
reactive power from its neighboring DGs that have additional 2
capacity. Each DG calculates its share of contribution to meet (𝑝𝑖𝐺 [𝐾])2 + (𝑞𝑖𝐺 [𝐾])2 ≤ (𝑆𝑖𝐷𝐸𝑅 ) (9)
𝐺 −1 𝐺 𝐺 [K] 𝐺 −1 𝐺
the requested reactive power via a distributed algorithm that −𝑝𝑖 tan(cos 𝑃𝐹𝑖 ) < 𝑞𝑖 < 𝑝𝑖 tan(cos 𝑃𝐹𝑖 ) (10)
requires communication network (e.g., Wi-Fi, ZigBee, or Constraints on cap operation, ∀𝑖 ∈ 𝜑𝐶 :
power line communication) to exchange information among 𝑞𝑖𝐺 [𝐾] = 𝑆𝑖 [𝐾]𝑄𝑖𝐶 (11)
neighboring DGs. Layer three performs local decision-making 𝐻
by enabling intelligent edge devices (i.e. LTCs, VRs and CBs) ∑|𝑆𝑖 [𝐾 + 1] − 𝑆𝑖 [𝐾]| < 𝑆𝑚𝑎𝑥 (12)
to measure the voltage, analyze the data and then quickly act 𝐾=1
locally and autonomously to cope with fast and real-time 𝑆𝑖 ∈ {0,1} (13)
changes in DGs and loads. Constraints on LTC/VR operation, ∀𝑖, 𝑗 ∈ 𝜑𝑉𝑅 :
𝑇𝑖𝑗 [𝐾]𝑉𝑗 [𝐾] = 𝑉𝑖 [𝐾] − (𝑟𝑖 𝑃𝑖𝑗 [𝐾] + 𝑥𝑖 𝑄𝑖𝑗 [𝐾])/𝑉1 (14)
𝐻

∑|𝑇𝑖𝑗 [𝐾 + 1] − 𝑇𝑖𝑗 [𝐾]| < 𝑇𝑚𝑎𝑥 (15)

𝐾=1
𝑇𝑖𝑗 < 𝑇𝑖𝑗 [𝐾] < 𝑇𝑖𝑗 (16)
𝑇𝑖𝑗 ∈ {0.95,096, … ,1.05} (17)
 4

The decision variables of problem (1) are 𝐱 = [𝑇𝑖𝑗 , 𝑆𝑖 , 𝑞𝑖𝐺 ] 𝑇𝑖𝑗 𝑉𝑗 = ∑ 𝑤ℎ (24)
and state variables 𝐲 = [𝑃𝑖𝑗 , 𝑄𝑖𝑗 , 𝑉𝑖 ]. The objective function (1) ℎ
is the total power losses over the study horizon 𝐻. In the above Equation (24) is still bilinear due to the product of continuous
formulation, (2)-(4) are are the linear form of the DistFlow and binary variables. In this paper, McCormick inequalities [7]
equations which have been extensively verified and used in the are deployed to linearize the binary bilinear terms as follows
literature [29]–[31], (5) is the power loss in line segment 𝑖𝑗 at 𝑤ℎ ≥ 𝜃ℎ 𝑉𝛼ℎ (25)
time instant 𝐾, and (6) and (7) are bus voltage magnitude and 𝑤ℎ ≥ 𝜃ℎ (𝑉𝑗 + 𝑉𝛼ℎ − 𝑉) (26)
line thermal limits. Constraint (8) is the DG’s active power 𝑤ℎ ≤ 𝜃ℎ 𝑉𝛼ℎ (27)
generation limit. Constraint (9) imposes the inverter rating curve
𝑤ℎ ≤ 𝜃ℎ (𝑉𝑗 + 𝑉𝛼ℎ − 𝑉) (28)
limit on DG generation capacity while constraint (10) reinforce
upper and lower bounds on the reactive power provisioning of where 𝑉 and 𝑉 are max and min ranges of voltage. In (25)-(28),
DGs (specified by a given power factor (𝑃𝐹)). Constraint (11) when 𝛼ℎ = 0, (25) and (27) imply that 𝑤ℎ = 0 and constraints
represents the zero or full capacity reactive power injection by (26) and (28) become redundant. Similarly, when 𝛼ℎ = 1 , (26)
cap bank at node 𝑖 and time instant 𝐾. Constraint (12) limits the and (28) imply that 𝑤ℎ = 𝜃ℎ 𝑉𝑗 and (25) and (27) are redundant
number of cap bank switching over the study horizon 𝐻. The constraints. Therefore, (24) can be reformulated as
on/off status of the cap banks is stablished by (13). Equation ∑ 𝑤ℎ = 𝜃𝑠 𝑉𝑗 , 𝛼𝑠 = 1 . (29)
(14) calculates the voltage at LTC/VR buses. Constraint (15)
ℎ
limits the number of LTC/VR tap operation over the study Using the above formulation, constraint (20) can be replaced
horizon 𝐻. Constraint (16) represents the maximum and by linear equations (21), (22), (25)-(28).
minimum LTC/VR tap steps, and constraint (17) stablishes the
B. Distributed Control Formulation
discrete status of LTC and VRs.
1) DG Communication Network Model: Consider a grid in
Note that the model (1)–(17) is nonlinear due to the products
of integer variable (𝑇𝑖𝑗 ) and continous variable (𝑉𝑗 ) in (14). In which n nodes have DGs and can only interact with each other
through local communication. The exchange of information
order to handle the nonlinearities, a fictitious bus 𝑚 is
introduced where the tap-changing transformer between nodes between DGs (active nodes) is represented by a weighted graph
𝑖𝑗 can be represented as an admittance 𝑟𝑖𝑗 +𝑗𝑥𝑖𝑗 in series with an 𝒢(𝒱, 𝐸, 𝒜) where the set of active nodes (vertices) 𝒱 =
ideal transformer with the tap ratio 1: 𝑇𝑖𝑗 as shown in Fig. 2. {1,2, … , 𝑛} are communicating through a set of undirected links
(edges) 𝐸 ⊆ 𝒱×𝒱\diag(𝒱). The nonnegative matrix 𝒜 =
[𝑎𝑖𝑗 ] (with the 𝑎𝑖𝑗 entry at the 𝑖th row, 𝑗th column) is a
weighted adjacency matrix that matches exactly the set of links
in the communication graph. 𝑎𝑖𝑗 is positive if there is a
Fig. 2. Tap-changing transformer equivalent model communication link between nodes 𝑖, 𝑗, i.e. (𝑖, 𝑗) ∈ 𝐸, and is
Since the transformer model in Fig. 2 is ideal (i.e. lossless zero otherwise. In order to follow the literature, we assume that
transformer), (14) can be reformulated as: the communication graph does not include any self-loops.
𝑃𝑖𝑗 [𝐾] + 𝑗𝑄𝑖𝑗 [𝐾] = 𝑃𝑖𝑎 [𝐾] + 𝑗𝑄𝑖𝑎 [𝐾] (18) Nodes that nodes that can send information to node 𝑗 are
𝑉𝑎 [𝐾] = 𝑉𝑖 [𝐾] − (𝑟𝑖 𝑃𝑖𝑎 [𝐾] + 𝑥𝑖 𝑄𝑖𝑎 [𝐾])/𝑉1 (19) defined as in-neighbors of node 𝑗 and denoted by 𝒩𝑗 − =
𝑇𝑖𝑗 [𝐾]𝑉𝑗 [𝐾] = 𝑉𝑎 [𝐾] (20) {𝑖 ∈ 𝒱|(𝑗, 𝑖) ∈ 𝐸}. Similarly, nodes that receive information
The reformulated transformer model is still nonlinear and from node 𝑗 are represented as out-neighbors of node 𝑗 and
non-convex due to bilinear term 𝑇𝑖𝑗 𝑉𝑗 in (20) where 𝑉𝑗 and 𝑇𝑖𝑗 denoted by 𝒩𝑗 + = {𝑖 ∈ 𝒱|(𝑖, 𝑗) ∈ 𝐸}.
are continuous and discrete variables, respectively. In order to In order to implement the proposed distributed control approach
reduce the nonlinearity of the above formulation, some and quarantee the convergence, we make the following
constraints can be reformulated. Let 𝜃ℎ , ℎ = 1,2 … ,11, denotes assumptions on the network communication graph 𝒢 [32]-[35]:
the set of states of transformer tap ratios that 𝑇𝑖𝑗 can take. 1) Strongly connectivity: There exists a path between any two
Referring to (17), 𝜃ℎ can take values from {0.95,096, … ,1.05}. vertices, i.e. ∀𝑖, 𝑗 ∈ 𝒱, 𝑖 ≠ 𝑗, there is a sequence of nodes 𝜈𝑖 =
Corresponding to 𝜃ℎ , a special ordered set of type 1 (SOS1) 𝜈𝑙1 , 𝜈𝑙2 , 𝜈𝑙3 , … , 𝜈𝑙𝐵 = 𝜈𝑗 (𝐵 ≥ 2) such that (𝐼, 𝐼 + 1) ∈ 𝐸 for all
variable is introduced as
𝐼 = 1,2, … , 𝐵 − 1.
∑ 𝛼ℎ = 1 (21) 2) Nondegeneracy: There exists a constant 𝛽 > 0 such that
ℎ 𝑎𝑖𝑖 ≥ 𝛽, and 𝑎𝑖𝑗 , for , 𝑖 ≠ 𝑗, satisfies 𝑎𝑖𝑗 ∈ {0} ∪ [𝛽, 1].
where 𝛼ℎ is a set of binary variables of which one can be 1 and
all others are zero. Therefore, the discrete variable 𝑇𝑖𝑗 can be 3) Column stochastic adjacency matrix:
represented as ∑ 𝑎𝑖𝑗 = 1 , ∀𝑗 ∈ 𝒱 (30)
𝑇𝑖𝑗 = ∑ 𝛼ℎ 𝜃ℎ (22) 𝑖∈{𝑗}∪{𝒩𝑗− }

𝑚 In order to fulfil the nondegeneracy and column stochastic

Moreover, a new set of variable 𝑤𝑚 is introduced as follows charachteristics, we define matrix 𝒜 = [𝑎𝑖𝑗 ] with the following
𝑤ℎ = 𝛼ℎ 𝜃ℎ 𝑉𝑗 (23) charachteristic:
Hence, 𝑇𝑖𝑗 𝑉𝑗 can be replaced by the following constraint
 5

2/[(𝒩𝑖 + + 𝒩𝑗 + ) + 𝜀] , 𝑖𝑓 𝑖 ∈ 𝒩𝑗 ∑𝑛𝑗=1 𝑞𝑗 𝑚𝑎𝑥

Γ𝑗 = 𝑞 . (43)
+ + ∑𝑛𝑗=1 𝑞𝑗𝑚𝑎𝑥 𝑗
𝑎𝑖𝑗 = 1 − ∑ 2/[𝒩𝑖 + 𝒩𝑗 + 𝜀] , 𝑖𝑓 𝑖=𝑗 (31)
𝑗∈𝒩𝑖
Meanwhile, DG 𝑗 computes
𝑥𝑗 [𝑡]
{0 , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 γ𝑗 [𝑡] = (44)
2) Distributed Control Algorithm: 𝑦𝑗 [𝑡]
In order to mitigate stress on grid assests and improve the and adjusts the reactive power to be provided at the equilibrium
system performance, a distributed control algorithm is proposed point as
that coordinates DGs and enables them to collectively Γ𝑗 , 𝑖𝑓 lim γ𝑗 [𝑡] ≤ 1
𝑒𝑞 𝑘→∞
𝑞𝑗 = { 𝑚𝑎𝑥 . (45)
participate in voltage control. At time instant 𝑡 the DG at bus 𝑗 𝑞𝑗 , 𝑖𝑓 lim γ𝑗 [𝑡] > 1
𝑘→∞
monitors its terminal voltage 𝑉𝑗 [𝑡] to determine if it is higher or From (43), it can be observed that summation over the initial
lower than the predefined upper and lower critical voltages 𝑉𝑐𝑟𝑈 reactive power request of all DGs ∑𝑛𝑗=1 𝑞𝑗 is equal to the
and 𝑉𝑐𝑟𝐿 . If so, the DG estimates the amount of reactive power reactive power shared among all DGs at the equilibrium
required to push the voltage back within (𝑉𝑐𝑟𝑈 , 𝑉𝑐𝑟𝐿 ) using the ∑𝑛𝑗=1 Γ𝑗 . In (44), DG 𝑗 monitors γ𝑗 value. If lim γ𝑗 [𝑡] ≤ 1 at the
𝑗𝑗 𝑘→∞
sensitivity of bus voltage magnitude to the reactive power (𝑆𝑉𝑄 ) equilibrium point, i.e., ∑𝑛𝑗=1 𝑞𝑗 ≤ ∑𝑛𝑗=1 𝑞𝑗𝑚𝑎𝑥 , DG capacity will
as not violated. If lim γ𝑗 [𝑡] > 1, the calculated Γ𝑗 by DG 𝑗 will be
𝑗𝑗 𝑘→∞
(𝑉𝑗 [𝑡]−𝑉𝑐𝑟𝑈 )⁄𝑆𝑉𝑄 , 𝑖𝑓 𝑉𝑗 [𝑡] ≥ 𝑉𝑐𝑟𝑈 larger than its capacity, so DG 𝑗 fixes its contribution to be
𝑞𝑗 [𝑡] = { 𝐿 (32) 𝑒𝑞
𝑗𝑗
(𝑉𝑐𝑟 − 𝑉𝑗 [𝑡])⁄𝑆𝑉𝑄 , 𝑖𝑓 𝑉𝑐𝑟𝐿 ≤ 𝑉𝑗 [𝑡] 𝑞𝑗 = 𝑞𝑗𝑚𝑎𝑥 .
Each DG requests for reactive power from its neighboring 3) Convergence: Convergence of the overall method can be
DGs that have additional capacity by initiating a peer-to-peer analyzed through the weight matrix 𝒜. Since the graph of 𝒜 is
information exchange. strongly connected, 𝒜 is irreducible. Moreover, 𝒜 is column
stochastic with non-negative entries. Therefore, 𝒜 is primitive,
To this end, two sets of auxiliary variables 𝑥𝑗 and 𝑦𝑗 are
i.e., there is a 𝑡 such that all the entries of 𝒜𝑡 are positive. From
introduced that each DG communicates with its neighboring the Perron Frobenius Lemma [3] for non-negative matrices
DG. 𝑥𝑗 represents the average reactive power and 𝑦𝑗 denotes 1
states that lim 𝒜𝑡 = 11𝑇 where 1 is the vector of all ones
𝑛
maximum reactive power of DG 𝑗. 𝑘→∞
[35]. The speed of convergence is determined by matrix 𝒜
The initial values for 𝑥𝑗 , 𝑦𝑗 are set as follows
which is carefully formed in (31). The faster convergence
𝑥𝑗 [0] = 𝑞𝑗 (33) achieved with lower values of 𝜆2 , the second largest eigenvalue
𝑚𝑎𝑥
𝑦𝑗 [0] = 𝑞𝑗 (34) of 𝒜. In general, a more connected communication graph yiels
where 𝑞𝑗 is calculated in (32). At time step 𝑡, each DG 𝑗 updates faster convergence speed [36]. In practice, a terminating
its information states 𝑥𝑗 [𝑡] and 𝑦𝑗 [𝑡] to 𝑥𝑗 [𝑡 + 1] and 𝑦𝑗 [𝑡 + 1] criterion is defined to quantify the number of iterations and
reach the equilibrium as ‖𝑥[𝑡] − 𝑥‖∞ ≤ 𝜎, ‖𝑦[𝑡] − 𝑦‖∞ ≤ 𝜎
via a weighted linear combination of its own value and the
[23].
available information received by its neighbors via
In this paper, we have used a 10 ms iteration step time to
communication as follows execute each iteration of the distributed algorithm which have
𝑥𝑗 [𝑡 + 1] = 𝑎𝑗𝑗 𝑥𝑗 [𝑡] + ∑ 𝑎𝑗𝑖 𝑥𝑖 [𝑡] (35) been verified and used in the literature [36].
𝑖∈{𝒩𝑗−} C. Edge Intelligence Formulation
One of the grid-edge Volt/Var providers is LTC/VR device
𝑦𝑗 [𝑡 + 1] = 𝑎𝑗𝑗 𝑦𝑗 [𝑡] + ∑ 𝑎𝑗𝑖 𝑦𝑖 [𝑡] (36)
that its basic function is to monitor voltage and maintain it
𝑖∈{𝒩𝑗− }
within a preset range. Most modern LTC/VR controls offer
𝑇 𝑇
Let 𝑥[𝑡] = [𝑥1 [𝑡], … , 𝑥𝑛 [𝑡]] and 𝑦[𝑡] = [𝑦1 [𝑡], … , 𝑦𝑛 [𝑡]] , several modes of operation, such as sequential, time integrating,
(35) and (36) can be written in a compact form as and voltage averaging [37]. In this paper, the time integrating
𝑥[𝑡 + 1] = 𝒜𝑥[𝑡] (37) control is applied as shown in Fig. 3. Autonomosly controlled
LTC/VRs use set voltage (𝑉𝑠𝑒𝑡 ), bandwidth (𝐵𝑊), and time
𝑦[𝑡 + 1] = 𝒜𝑦[𝑡] (38)
delay (𝑇𝑑 ) to regulate the voltage at the secondary side of the
or can be represented in terms of the initial vectors as
transformer. 𝑉𝑠𝑒𝑡 is the voltage level (in 120V base) and 𝐵𝑊 is
𝑥[𝑡 + 1] = 𝒜𝑡 𝑥[0] (39) 1
the acceptable voltage range around set voltage (𝑉𝑠𝑒𝑡 ± 𝐵𝑊)
𝑦[𝑡 + 1] = 𝒜𝑡 𝑦[0] (40) 2
At every time step, DG 𝑗 calculates to which the control will regulate. 𝑇𝑑 is the number of seconds
𝑥𝑗 [𝑡] the control waits from the start of an out-of-band condition,
Γ𝑗 [𝑡] = 𝑦 [0] (41) before initiating a tap change. In order to prevent unnecessary
𝑦𝑗 [𝑡] 𝑗 LTC/VR operations due to temporary voltage fluctuations, 𝑇𝑑
If 𝑡 is sufficiently large, Γ𝑗 [𝑘] will asymptotically converge is typical set to 30 through 90 sec.
to
∑𝑛𝑗=1 𝑥𝑗 [0]
Γ𝑗 = 𝑛 𝑦 [0] (42)
∑𝑗=1 𝑦𝑗 [0] 𝑗
or equivalently
 6

Fig. 4. (a) Coordination between layers one and two (b) Coordination
Fig. 3. Control principle of autonomously controlled LTC/VR between layers one and three
Fig. (a) shows the coordination between layer one and two in
The control initiates a counter 𝑇𝑡 which counts up when the which each DG measures the voltage at the point of
1
measured voltage is out of 𝑉𝑠𝑒𝑡 ± 𝐵𝑊 and counts down to zero interconnection and uses its own local intelligence to calculate
2
when the voltage is in bound. the rective power injection/absorption required (𝑞𝑗𝐿 [𝑡]) at time
1
max(0, 𝑇𝑡 − 1) , 𝑖𝑓 |𝑉𝑗 [𝑡] − 2𝐵𝑊| ≤ 𝑉𝑠𝑒𝑡 𝑘. Meanwhile each DG receives the reactive power scheduling
𝑇𝑡 = { (46) from layer two (𝑞𝑗𝑇 [𝑘]) and determines the final reactive power
𝑇𝑡 + 1 , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 𝑓
Once 𝑇𝑡 becomes greater than 𝑇𝑑 , a control pulse is sent to (𝑞𝑗 [𝑡]) to be injected/absorbed. This is summarized in (51)
the LTC/VR mechanism in order to move the tap position up or below:
down. 𝑞𝑗𝑇 [𝑘]𝑠𝑖𝑔𝑛(𝑉𝑗 [𝑡] − 𝑉) , 𝑖𝑓 𝑉𝑗 [𝑡] ≥ 𝑉
1
𝑇𝑎𝑝𝐿 − 1 , 𝑖𝑓 𝑉𝑠𝑒𝑡 + 2𝐵𝑊 < 𝑉𝑗 [𝑡] , 𝑇𝑡 > 𝑇𝑑 𝑓
𝑇𝑎𝑝𝐿 = { (47) 𝑞𝑗 [𝑡] = { 𝑞𝑗𝐿 [𝑘] , 𝑖𝑓 𝑉 ≤ 𝑉𝑗 [𝑡] ≤ 𝑉 (51)
1
𝑇𝑎𝑝𝐿 + 1 , 𝑖𝑓 𝑉𝑠𝑒𝑡 − 2𝐵𝑊 > 𝑉𝑗 [𝑡] , 𝑇𝑡 > 𝑇𝑑 𝑞𝑗𝑇 [𝑘]𝑠𝑖𝑔𝑛(𝑉𝑗 [𝑡]
− 𝑉) , 𝑖𝑓 𝑉𝑗 [𝑡] ≤ 𝑉
Using the same analogy for the cap bank operation, the As shown in Fig. (b), each switching device (LTC/VR and
control method initiates a time delay that counts up when the cap bank) measures the local voltage and uses its own
measured voltage is out of standard regulatory bound and intelligence to calculate a tap operation and/or cap switching
counts down all the way down to zero when the measured (𝑇𝑎𝑝𝐿𝑗 [𝑘]) while receiving a tapping/switching schedule from
voltage is in bound.
layer one (𝑇𝑎𝑝𝑇𝑗 [𝑘]). In order to coordinate between local edge
max(0, 𝑇𝑡 − 1) , 𝑖𝑓 𝑉 ≤ 𝑉𝑗 [𝑡] ≤ 𝑉
𝑇𝑡 = { (48) and global tapping/switching decisions for each switching
𝑇𝑡 + 1 , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 device at node 𝑗, a coordination rate 𝜆𝑗 is introduced. The 𝜆𝑗
If the voltage remains out of bound for the duration of the value close to 1 represents a situation in which the LTC/VR or
time delay setting, an appropriate cap switching is activated. cap bank is more likely follows the scheduling signal from layer
0 , 𝑖𝑓 𝑉𝑗 [𝑡] ≥ 𝑉 , 𝑇𝑡 > 𝑇𝑑 one. As 𝜆𝑗 approaches 0, the LTC/VR or cap bank is more likely
𝑆𝑊𝐿 = { (49)
1 , 𝑖𝑓 𝑉𝑗 [𝑡] ≤ 𝑉 , 𝑇𝑡 > 𝑇𝑑 follows the tapping/switching decitions that was made based on
DGs are assumed to follow a droop control mechanism [9] as edge intelligence. The mathematical formulation is presented as
𝑗
𝑞𝑗 [𝑡]+𝑞𝑗 𝑈 𝑈 𝑇𝑎𝑝𝑇 [𝑘] , 𝑖𝑓 𝜆𝑗 = 1
𝑈 (𝑉𝑗 [𝑡]−𝑉𝑐𝑟 ), 𝑖𝑓 𝑉𝑗 [𝑡] ≥ 𝑉𝑐𝑟 𝑗
𝑞𝑗 [𝑡] = 𝑞𝑗 [𝑡] −
1.05−𝑉𝑐𝑟
{ 𝑞𝑗[𝑡]−𝑞𝑗 (50) 𝑇𝑎𝑝𝑓 [𝑘] = {𝑟𝑜𝑢𝑛𝑑(𝜆𝑗 𝑇𝑎𝑝𝑇𝑗 [𝑘] + (1
𝑗
− 𝜆𝑗 )𝑇𝑎𝑝𝐿 [𝑘]) , 𝑖𝑓 𝜆𝑗 ∈ (0,1)
𝑗
(𝑉 𝐿 − 𝑉𝑗 [𝑡]), 𝑖𝑓 𝑉𝑐𝑟𝐿 ≤ 𝑉𝑗 [𝑡] 𝑇𝑎𝑝𝐿 [𝑘] , 𝑖𝑓 𝜆𝑗 = 0
𝑉 𝐿 −0.95 𝑐𝑟
𝑐𝑟
where 𝑞𝑗 and 𝑞𝑗 are maximum and minimum DG inverter (52)
III. SIMULATION RESULTS
reactive power capacities at node 𝑗.
A. Network description
D. Coordination Between Layers
A modified IEEE 34 bus distribution system is used to
In order to achieve system-wide voltage control and mitigate
investigate performance of the proposed approach as shown in
the adverse impact of high penetration of DG while optimally
Fig. 5. The network line and load data can be found in [36].
use voltage control assets, a coordination between control layes
DGs are connected to buses #11, #20, #29 with the
is required as shown in Fig. 4.
corresponding capacity of 500 kW, 500 kW, and 450 kW,
respectively. Two cap banks of 150 kVar are located in buses
#25, #27 with corresponding time delay of 30 and 25 seconds.
 7

Home load data were extracted from the eGauge website centralized+edge methods for sunny days. Max voltage is
[39], which provides load data with up to 1-minute resolution. observed via edge intelligence method (at bus 20) while min
Typical home data is shown in Fig. 6. Load reactive power is voltage is realized for centralized and hybrid methods (at bus
defined in proportion to the real load connected at the same bus 29). Inherently designed for local decision making, the edge
with a power factor of 0.9 lagging. The 1-second resolution PV intelligence method is not able to(a)efficiently use voltage control
generation is obtained from a station near Hawaii’s Honolulu assets since its not using the LTC. Note that LTC is the main
International Airport on the island of Oahu as available in [40]. voltage control device that is located at the infeed to the feeder
Global Horizontal Irradiance is shown in Fig. 6, which includes and its tap operation impacts the whole feeder voltage. As a
both the clear sky and transient cloud movement periods. The result, losses are the most via (b) edge intelligence method as
assumption was made that the nodes are geographically close shown in Table II. Compared to edge intelligence method,
in the network such that outputs of PV units follow the same Central+Edge approach performs better in terms of power
generation pattern. LTC/VR and Cap bank data are shown in losses, however, it’s performance is worse than centralized
table I. method. It’s mainly because in (c) sunny days there is no fast
variation in solar irradiance, therefore, implementing 15 minute
tap changer and cap switching signals generated from layer one
(centralized method only) result in better volt/var control
management scheme than that of (d) Central+Edge approach in
which local decision making at the edge of the device
deterioriates the signals received from layer one. Hybrid
method performs very close to centralized method but
marginally better via scheduling (e) of VR1 and VR2 with
comparable number of tap operation. Moreover, minimal losses
are realized with less stress on DGs for reactive power
provisioning by adding distributed communication among DGs
Fig. 5. Modified IEEE 34node test feeder. for volt/var control as (f) shown in Table 2.

(g)

Fig. 6. Typical home data (left) and global horizontal irradiance (right)

TABLE I: LTC/VR DATA

𝑉𝑠𝑒𝑡 (p.u.) 𝐵𝑊 (p.u.) 𝑇𝑑 (s) Initial Tap Position
LTC 1.025 0.01 60 0
VR1 1 0.02 50 3
VR2 1 0.03 40 0

B. Results
The proposed hybrid control architecture was compared to
the centralized, edge intelligence and centralized+edge
methods. We examine four scenarios: Case 1) sunny day, Case
2) intermittent cloudy day, Case 3) intermittent cloudy day with
forecast error, Case 4) distributed DG; i.e. 8 DGs instead of 3
DGs shown in Fig. 5, but at the same level of penetration (80%).
Case 1) Figs. 7(a)-(g) show whole-day simulations of the
voltage at bus 29, the tap changer operation, and reactive power
provisioning of DGs for case 1 under all control strategies. For
quantitative comparison, the investigated control methods are
also summarized in Table I. Performance metrics include total
power losses, number of LTC/VR tap operations and cap bank
switchings, and max and min voltages observed throughout the Fig. 7. (a) Voltage at node 29, (b)-(d) Reactive power provisioning from DGs
day. Referring to Table I, it can be inferred that application of 1 to 3, (e) LTC tap operation, (f) VR1 tap operation, (g) VR2 tap operation.
centralized method results in 3.8382 MW power losses, which
shows that the centralized method performs relatively better in
terms of the power losses as compared to edge intelligence and
 8

TABLE II: Daily power losses, number of LTC/VR tap operations, cap bank impact of solar forecast error is much higher on the centralized
switchings, max, min voltage, and Lambda values for case 1 method as the power losses increased the most while the worst
Ploss Number of tap voltage regulation is exhibited (min voltage 0.8720 p.u.). This
Operations
Max, Min
𝜆𝑗
(MW)
LTC VR1
Voltage (p.u.)
VR2
is mainly because the day-ahead scheduling in layer is altered
Cent. 3.8382 6 6 2 1.0313,0.9162 -- by the forecast errors as the difference between 1-hour-ahead
Edge 4.5819 0 8 6 1.0405,0.9268 -- and day-ahead forecasts is increased over time. Also, such 15-
Cent+Edge 4.1667 6 6 4 1.0323,0.9268 1,1,0.1 min scheduling decisions have been used to manage solar
Hierarchical-Hybrid 3.8311 6 7 3 1.0313,0.9162 1,0.9,0.5
variability which is in the timescale of seconds to minutes. This
Case 2) The hybrid control approach has also been emphasizes that consideration of these errors is needed to
implemented for an intermittent cloudy day and the result were capture the true impact of the forecasts on power system
compared to centralized, edge intelligence and operations. In contrast, edge intelligence method is not
centralized+edge methods as shown in Table III. impacted by solar forecast errors since real data are used in this
approach (edge intelligence method is using a data-driven
TABLE III: Daily power losses, number of LTC/VR tap operations, cap
bank switchings, max, min voltage, and Lambda values for case 2 approach to perform decision making).
Number of tap Application of centralized+edge method under forecast error
Ploss Max, Min
changing Voltage (p.u.)
Lambda results in lower losses compared to centralized and edge
(MW)
LTC VR1 VR2
intelligence methods. This confirms the ability of edge devices
Cent. 3.6502 5 6 5 1.0313,0.9162 --
Edge 3.8573 0 8 4 1.0412,0.9268 -- in improving the accuracy of analytic applications. In contrast
Cent+Edge 3.61 5 8 3 1.0313,0.9268 1,0.9,0 to centralized, edge intelligence and centralized+edge methods,
Hierarchical-Hybrid 3.5780 5 8 3 1.0313,0.9268 1,0.9,0 all DGs in the proposed hybrid approach contribute via reactive
In case 2, where solar irradiance changes fast and frequently power support in order to achieve minimal losses and improve
due to intermittent clouds passing over the grid, centralized voltage regulation. Moreover, the proposed hybrid approach is
method exhibits the worst voltage regulation (min voltage the case in which losses are reduced, which shows the ability of
0.9162 p.u.) while edge intelligence method performs much the hybrid approach in dealing with forecast error and
better than the sunny day in terms of reducing the power losses. uncertainties in input data for smart grid analytics applications.
In particular, percentage of difference between power losses Switch status of capacitors didn’t change for cases 1-3.
incurred from edge intelligence and centralized methods IV. CONCLUSIONS AND FUTURE WORK
reduced from 19.37% to 5.65% from case 1 to case 2. The edge This paper presents an efficient, actively coordinated,
intelligence value is more realized in Central+Edge approach hierarchical control architecture for Volt/Var management of
where edge devices process and analyse data independently and distribution grids in the presence of high penetration of DGs.
improve the performance by reducing the power losses even The architecture includes three layers with specific operational
more than centralized method as shown in Table III. The goals associated with each layer based on data resolution,
proposed hybrid approach presented in this paper further communication network, and control device response time. The
reduces losses to 3.587 MW by efficiently scheduling of top layer carries out central optimization and seeks to minimize
LTC/VRs and allowing DGs to communicate with each other power losses via optimal daily scheduling of slow response
which results in maximum reactive power provision from DGs. legacy devices such as load tap changers (LTCs), voltage
Case 3) The study in case 2 was on the basis that layer one regulators (VRs), and capacitor banks (CBs), and DGs. The
receives a perfect forecast of solar generation for day ahead middle layer ensures voltage regulation using a distributed
LTC/VR and cab banks scheduling. However, in practice there approach for reactive power provisioning by fast response DGs.
is an error associated with solar irradiance forecast that if not Finally the bottom layer carries out local decision making that
considered, it can result in substantial economic losses and enlists edge intelligence to operate LTCs, VRs and CBs to cope
power system reliability issues. In case 3, 15-min forecasted with fast and real-time changes in DGs and loads. The
data with errors was used for the whole day simulation. Mainly, simulations studies carried out on a modified IEEE-34 bus
the maximum average percentage error (MAPE) reported for shows the advantage of the hierarchical architecture.
hour ahead forecast is 11% and it increases to 15% for the day
ahead forecast [41]. REFERENCES
TABLE IV: Daily power losses, number of LTC/VR tap operations, cap [1] E. Demirok, P. C. Gonzalez, K. H. B. Frederiksen, D. Sera, P. Rodriguez,
bank switchings, max, min voltage, and Lambda values for case 3 and R. Teodorescu, "Local Reactive Power Control Methods for
Number of tap Overvoltage Prevention of Distributed Solar Inverters in Low-Voltage
Ploss
(MW) changing
Max, Min
Voltage (p.u.) 𝜆𝑗 Grids," IEEE J. Photovoltaics, vol. 1, pp. 174-182, Oct 2011.
LTC VR1 VR2 [2] R. Caldon, M. Coppa, R. Sgarbossa, and R. Turri, "A simplified algorithm
Cent. 3.80 6 4 6 1.0313,0.8720 -- for OLTC control in active distribution MV networks," in AEIT Annual
Edge 3.8573 0 8 4 1.0412,0.9268 -- Conference, 2013, 2013, pp. 1-6.
Cent+Edge 3.65 6 5 6 1.0424,0.9268 1,0.9,0.1 [3] [ref1b] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge, U.K.:
Hierarchical-Hybrid 3.57 6 5 6 1.0444,0.9268 1,0.9,0 Cambridge Univ. Press, 1987
[4] C.-S. Chen, C.-T. Tsai, C.-H. Lin, W.-L. Hsieh, and T.-T. Ku, “Loading
Comparing the results in Tables III and IV, it is clear that the balance of distribution feeders with loop power controllers considering
losses are increased for the centralized, edge intelligence and photovoltaic generation,” IEEE Trans. Power Syst., vol. 26, no. 3, pp.
1762–1768, Aug. 2011.
centralized+edge methods due to the solar forecast error. In [5] A. R. Malekpour, A. Pahwa, S. Das, “Inverter-based var Control in Low
particular, 4%, and 1.1% increase were observed for the Voltage Distribution Systems with Rooftop Solar PV,” in Proc. 2013
centralized, and centralized+edge methods, respectively. The IEEE 45th North American Power Symposium, pp. 1-5.
 9

[6] N.-K. C. Nair and L. Jing, “Power quality analysis for building integrated [29] R. Tonkoski and L. A. C. Lopes, "Voltage Regulation in Radial
PV and micro wind turbine in new zealand,” Energy and Buildings, vol. Distribution Feeders with High Penetration of Photovoltaic," in Energy
58, pp. 302–309, 2013. 2030 Conference, 2008. ENERGY 2008. IEEE, 2008, pp. 1-7.
[7] A. Borghetti, '"Using mixed integer programming for the volt/var [30] R. Aghatehrani and R. Kavasseri, "Reactive Power Management of a
optimization in distribution feeders," Electr.Power Syst.Res., vol. 98, 5, DFIG Wind System in Microgrids Based on Voltage Sensitivity
2013, pp. 39-50. Analysis," IEEE Trans.Sust. Energy, vol. 2, pp. 451-458, Oct 2011.
[8] S. Deshmukh, B. Natarajan and A. Pahwa, '"Voltage/VAR Control in [31] R. Walling, Z. Gao, "Eliminating Voltage Variation Due to Distribution-
Distribution Networks via Reactive Power Injection Through Distributed Connected Renewable Generation," 2011 DistribuTECH Conference and
Generators," IEEE Transactions on Smart Grid, vol. 3, no. 3, 2012, pp. Expo, San Diego, Feb. , 2011.
1226-1234. [32] A. Nedic and A. Ozdaglar, “Distributed subgradient methods for
[9] L.R. Araujo, D.R.R. Penido, S. Carneiro and J.L.R. Pereira, '"A Three- multiagent optimization,” IEEE Trans. Autom. Control, vol. 54, no. 1, pp.
Phase Optimal Power-Flow Algorithm to Mitigate Voltage Unbalance," 48–61, Jan. 2009.
IEEE Trans.Power Del., vol. 28, no. 4, 2013, pp. 2394-2402. [33] A. Nedic, A. Ozdaglar, and P. A. Parrilo, “Constrained consensus and
[10] S. Bruno, S. Lamonaca, G. Rotondo, U. Stecchi and M. La Scala, optimization in multi-agent networks,” IEEE Trans. Autom. Control, vol.
'"Unbalanced Three-Phase Optimal Power Flow for Smart Grids," IEEE 55, no. 4, pp. 922–938, Apr. 2010.
Transactions on Industrial Electronics, vol. 58, no. 10, 2011, pp. 4504-
[34] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and cooperation
4513.
in networked multi-agent systems,” Proc. IEEE, vol. 95, no. 1, pp. 215–
[11] Y. Zhu and K. Tomsovic, '"Optimal distribution power flow for systems
233, Jan. 2007.
with distributed energy resources," International Journal of Electrical
[35] A. Olshevsky and J. N. Tsitsiklis, “Convergence speed in distributed
Power & Energy Systems, vol. 29, no. 3, 3, 2007, pp. 260-267.
consensus and averaging,” SIAM J. Control Optim., vol. 48, no. 1, pp.
[12] T. Senjyu, Y. Miyazato, A. Yona, N. Urasaki and T. Funabashi, '"Optimal
33–55, 2009.
Distribution Voltage Control and Coordination With Distributed
[36] T. F. Wu, C. L. Kuo, K. H. Sun, and H. C. Hsieh, "Combined Unipolar
Generation," IEEE Trans.Power Del., vol. 23, no. 2, 2008, pp. 1236-1242.
and Bipolar PWM for Current Distortion Improvement During Power
[13] A.R. Malekpour and T. Niknam, '"A probabilistic multi-objective daily
Compensation," IEEE Trans. Power Elect., pp. 1702-1709, Apr 2014.
Volt/Var control at distribution networks including renewable energy
[37] Voltage Regulators: Cooper Power System's VR-32 Regulator and CL-
sources," Energy, vol. 36, no. 5, 5, 2011, pp. 3477-3488.
2A Control Installation, Operation and Maintenance Instructions Parts
[14] G. Valverde and T. Van Cutsem, '"Model Predictive Control of Voltages
Replacement Information
in Active Distribution Networks," IEEE Transactions on Smart Grid, vol.
http://www.cooperindustries.com/content/dam/public/powersystems/res
4, no. 4, 2013, pp. 2152-2161.
ources/library/225_VoltageRegulators/S225105.pdf
[15] S. Paudyal, C.A. Canizares and K. Bhattacharya, '"Optimal Operation of
[38] N. Mwakabuta, A. Sekar, Comparative Study of the IEEE 34 Node Test
Distribution Feeders in Smart Grids," IEEE Trans.Ind.Electron., vol. 58,
Feeder under Practical Simplifications, 39th North American Power
no. 10, 2011, pp. 4495-4503.
Symposium (NAPS 2007), 484-491, 2007.
[16] B. Zhang, A. Lam, A. Dom ́ınguez-Garc ́ıa, and D. Tse, “Optimal
[39] [Online]. Available:http://egauge360.egaug.es/."
distributed voltage regulation in power distribution networks,”arXiv
[40] [Online]. Available:http://www.nrel.gov/midc/oahu_archive/.
preprint arXiv:1204.5226, 2012.
[41] J. Zhang, A. Florita, B.-M. Hodge, S. Lu, H. F. Hamann, V.
[17] E. Dall'Anese, E. DallAnese, S.V. Dhople, B.B. Johnson and G.B.
Banunarayanan, and A. M. Brockway, “ A suite of metrics for assessing
Giannakis, '"Decentralized Optimal Dispatch of Photovoltaic Inverters in
the performance of solar power forecasting,” Sol. Energy 111, 157–175
Residential Distribution Systems," IEEE Trans.Energy Convers., vol. 29,
(2015).
no. 4, 2014, pp. 957-967.
[18] E. Dall'Anese, G.B. Hao Zhu and Giannakis, '"Distributed Optimal Power BIOGRAPHIES
Flow for Smart Microgrids," IEEE Transactions on Smart Grid, vol. 4, no.
3, 2012, pp. 1464-1475. Dr. Ahmad Reza Malekpour (M’16) received the Ph.D. degree in electrical
[19] P. Šulc, P. Sulc, S. Backhaus and M. Chertkov, '"Optimal Distributed engineering from Kansas State University, Manhattan, in 2016. He was a
Control of Reactive Power Via the Alternating Direction Method of Postdoctoral Associate within Active-Adaptive Control Laboratory,
Multipliers," IEEE Trans.Energy Convers., vol. 29, no. 4, 2013, pp. 968- Massachusetts Institute of Technology (MIT), Cambridge. Curently, he is the
977. Manager of Grid Strategy and Analytics at ComEd. His research interests
[20] T. Alquthami and A.P. Meliopoulos, '"Hierarchical optimization and include renewable procurement strategies, energy management of microgrids
control of a distribution system," 2013 North American Power and distributed energy resources, stochastic and distributed optimization in
Symposium (NAPS), pp. 1-6. power and energy system.
[21] A.P. Meliopoulos, G.J. Cokkinides, R. Huang and E. Farantatos,
'"Integrated Smart Grid Hierarchical Control," 2012 45th Hawaii Dr. Anuradha Annaswamy received her Ph.D. in Electrical Engineering from
International Conference on System Sciences, pp. 1967-1976. Yale University in 1985. She has been a member of the faculty at Yale, Boston
[22] A. Cagnano, E. De Tuglie, M. Liserre and R.A. Mastromauro, '"Online University, and MIT where currently she is the director of the Active-Adaptive
Optimal Reactive Power Control Strategy of PV Inverters," IEEE Control Laboratory and a Senior Research Scientist in the Department of
Trans.Ind.Electron., vol. 58, no. 10, 2011, pp. 4549-4558. Mechanical Engineering. Her research interests pertain to adaptive control
[23] M. Brenna, E. De Berardinis, L. Delli Carpini, F. Foiadelli, P. Paulon, P. theory and applications to aerospace, automotive, and propulsion systems,
Petroni, G. Sapienza, G. Scrosati and D. Zaninelli, '"Automatic cyber physical systems science, and CPS applications to Smart Grids, Smart
Distributed Voltage Control Algorithm in Smart Grids Applications," Cities, and Smart Infrastructures.
IEEE Trans. Smart Grid, vol. 4, no. 2, 2013, pp. 877-885.
[24] T. Sansawatt, L.F. Ochoa and G.P. Harrison, '"Smart Decentralized
Control of DG for Voltage and Thermal Constraint Management," IEEE Dr. Jalpa Shah received her Ph.D. in Electrical Engineering from University
Trans.Power Syst., vol. 27, no. 3, 2012, pp. 1637-1645. of Minnesota, twincities in 2011. She is currently a Specialist Engineer with
[25] P.M.S. Carvalho, P.F. Correia and L.A.F. Ferreira, '"Distributed Reactive Corporate Research and Techngoloy , Eaton Corporation leading technology
Power Generation Control for Voltage Rise Mitigation in Distribution development for intelligent power grid controls. She has priorly worked with
Networks," IEEE Trans.Power Syst., vol. 23, no. 2, 2008, pp. 766-772. John Deere and Rockwell Automation in the domain of industrial motor control
[26] F.A. Viawan and D. Karlsson, '"Combined Local and Remote Voltage and and automation and on/off road vehicle electrification. Her research interests
Reactive Power Control in the Presence of Induction Machine Distributed include advanced controls for power electronics and electrical power systems.
Generation," IEEE Trans.Power Syst., vol. 22, no. 4, 2007, pp. 2003-
2012. .
[27] S. Weckx, C. Gonzalez and J. Driesen, '"Combined Central and Local
Active and Reactive Power Control of PV Inverters," IEEE Transactions
on Sustainable Energy, vol. 5, no. 3, 2014, pp. 776-784.
[28] J. Barr, and R. Majumder, Integration of Distributed Generation in the
Volt/VAR Management System for Active Distribution Networks, IEEE
Trans. on Smart Grid, vol. 6, no. 2, 2014, pp. 576-586.