Electrical Power and Energy Systems 167 (2025) 110632

Contents lists available at ScienceDirect

International Journal of Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

A cloud-edge collaborative optimization control strategy for voltage in

distribution networks with PV stations
Guoqing Li a , Wei Wang a , Dan Pang b, Zhipeng Wang b , Weixian Tan a , Zhenhao Wang a,*,
Jinming Ge a
a
Key Laboratory of Modern Power System Simulation and Control & Renewable Energy Technology, Northeast Electric Power University, Jilin 132012, China
b
State Grid Changchun Power Supply Company, Changchun 130021, China

A R T I C L E I N F O A B S T R A C T

Keywords: With the continuous expansion of the power system scale and the continuous development of the power network,
Cloud edge synergy the traditional power system management and optimization methods face many challenges. In order to meet the
Reactive-voltage sensitivity requirements of voltage optimization and adjustment, the optimization problem is divided into cloud front
Modularity function
precomputation and edge computing device cooperative optimization computation with the framework of cloud-
Mixed-integer second-order cone planning
edge cooperation. The cloud front-end precomputation uses an improved reactive-voltage sensitivity based on an
improved modularity function to partition the power system on a 15 min basis and stores the results in the cloud
data memory. The voltage threshold device detects the node voltage overrun and triggers the collaborative
optimization computation of the edge computing devices, which sends a command to the cloud to call the
partitioning result of this time period, and the cloud sends the result to each edge computing device, which
determines the area it is responsible for, and adjusts the voltage overrun partitioning by using the mixed-integer
second-order conic planning, and ultimately realizes the optimization strategy within the minute-level zone.
Since the voltage adjustment is a fine-grained optimization of the local area, it is highly flexible and targeted.
Moreover, using the cloud-edge collaboration technology, the intelligent management and optimization of the
power system is finally realized. Case analysis and comparative verification show that the method proposed in
this paper is accurate and highly efficient.

1. Introduction The traditional centralized control method is to regulate the voltage

control equipment in the system (e.g., reactive power compensation
With the increasing scale and complexity of power systems, distrib­ equipment, transformer taps, etc.) by real-time monitoring of the
uted photovoltaic (DP), energy storage system (ESS), static Var voltage level at each node of the power system and according to the pre-
compensator (SVC), group switching capacitor (SC) and other coordi­ set control strategy and algorithm [7–9], so that the voltage meets the
native control devices are connected [1,2], which makes the traditional operational requirements. However, with the increasing scale and
power system management and optimization methods face many chal­ complexity of the power system, the traditional centralized control of
lenges, such as voltage overrun, power balance and so on. switching real-time collection of commands and simultaneous processing of data
capacitor (SC) and other coordinative control devices [1,2], which limits its application in practice. Literature [10] proposed a centralized
makes the traditional power system management and optimization voltage regulation method for a high percentage of distributed photo­
methods face many challenges, such as voltage overruns, power voltaics, which coordinates the devices using a mixed integer nonlinear
imbalance and other increasingly serious problems [3,4]. In today’s programming model, finds the optimal set using the NSGA-II algorithm,
digital and intelligent background, the application of digital means for and proposes a fast decision-making algorithm to find the optimal so­
power system optimization has become a trend, and the cloud-side lution in the optimal set. Literature [11] proposed a two-stage optimi­
cooperative approach has gradually received public attention and zation method based on meta-heuristic optimization algorithm (MOA) to
become a research hotspot [5,6]. regulate the voltage level by controlling the charging and discharging

* Corresponding author at: Key Laboratory of Modern Power System Simulation and Control & Renewable Energy Technology, Northeast Electric Power University,
Jilin 132012, China.
E-mail address: [email protected] (Z. Wang).

https://doi.org/10.1016/j.ijepes.2025.110632
Received 31 May 2024; Received in revised form 5 November 2024; Accepted 25 March 2025
Available online 29 March 2025
0142-0615/© 2025 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

rate of the battery for the operation of energy storage systems such as regional multi-energy system scheduling process.
batteries. Literature [12] proposes a Deep Reinforcement Learning In summary, at present, the voltage regulation of the distribution
(DRL)-based scheduling scheme for energy storage systems to mitigate network is faced with multiple challenges such as immature intelligent
system voltage deviations in unbalanced low-voltage distribution net­ regulation technology, high volatility of renewable energy, aging
works. The ESS-based voltage regulation problem is formulated as a equipment, and insufficient capacity, which is difficult to ensure the
multi-stage quadratic stochastic programming that ultimately mini­ quality of the system and voltage, and may also affect the stability of the
mizes the total daily voltage regulation cost. These centralized optimi­ system.
zation methods generally suffer from disadvantages such as dependence Therefore, this paper proposes a cloud-based front-end pre­
on control centers, communication delays, and high complexity and cost computation strategy and an edge voltage optimization control strategy
[13–16]. based on cloud-edge synergy, following the ideas of reactive-voltage
In view of the above problems, many studies have proposed grid partitioning control and Mixed-integer Second Order Cone Program­
partition voltage optimization control methods [17,18], that is, based on ming (MISOCP). Firstly, the cloud-based 15 min precomputation parti­
the original grid structure of the voltage optimization problem is divided tioning model is proposed to establish reactive-voltage sensitivity
into a number of sub-area optimization problems, the method has an partitioning indexes, combined with the modularity function partition­
important role for the optimization of the control space of the dimen­ ing standard, to partition the power system to achieve high coupling
sionality reduction [19–21]. Literature [22] synthesizes the distribution within the region and weak coupling between the domains, which
network zonal distributed control with active reactive power compre­ effectively improves the operation and management efficiency of the
hensive consideration, and proposes a distributed active regulation power system; secondly, the minute-level intra-area optimization strat­
strategy, firstly, based on the reasonable zoning of the distribution egy based on the edge computing equipment is proposed to view the
network, and secondly, after the ADMM to minimize the active network voltage optimization problem as a second-order cone optimization
loss and voltage deviation as the objective function for optimization problem, and real-time optimization computation is carried out by the
adjustment. Literature [23] proposes a regional voltage control method edge computing devices of the power system to achieve fast response
for distributed PV system distribution network based on network par­ and optimization adjustment to the local area of the power system,
titioning, which achieves network partitioning and voltage control which improves the flexibility of the system; finally, the effectiveness of
through an improved modularity function with a greedy network par­ the proposed cloud-edge cooperative voltage optimization method is
titioning optimization algorithm. In real-time control each region is only verified in the improved IEEE33 node arithmetic system.
responsible for controlling its own voltage can achieve similar results
with centralized control, and has more obvious real-time performance. 2. Cloud-side cooperative voltage optimization control method
Literature [24] divides the distribution system into various zones with for photovoltaic-containing power plants
respect to the type of users, followed by reactive power/voltage opti­
mization in each zone using an optimal current modeling method based In recent years, with the rapid development of digitalization and
on banded relaxation semi-deterministic planning. This method is intelligent technology, the emerging cloud-edge cooperative technology
characterized by fast convergence and is non-iterative, hence no penalty transfers the original centralized computing to each cloud device and
parameters are required. Literature [25] utilizes improved active and edge computing device [30], which provides a new type of information
reactive voltage sensitivity combined electrical distance metrics, com­ processing solution for the power system, which can integrate various
bined with AP algorithm for partitioning the system, and ADMM for resources, process data more efficiently, and support the high-speed
global voltage optimal control on different time scales using active development of the power system. The cloud-edge cooperative voltage
network loss as the objective function. Literature [26] proposes a dual optimization method proposed in this paper combines the cloud with
time-scale distributed voltage control scheme based on Model Predictive edge computing devices, i.e., the cloud is responsible for 15 min front-
Control (MPC) and Alternating Direction Multiplier Method (ADMM) for end reactive power-voltage sensitivity partitioning model training, and
distributed generator unbalanced distribution networks. The optimiza­ the edge computing devices retrieve the cloud partitioning data for
tion problem is distributed to various regions of the distribution network minute-level voltage optimization calculations within the zone.
by incorporating ADMM without the need of a central controller. Different from the traditional centralized control architecture, the cloud
With the continuous development of distribution Internet of Things no longer directly receives the state and action information of each
(IoT) technology, the trend of power grid is gradually evolving from fine device and terminal, but is collected and uploaded to the cloud by the
control to the direction of distribution security, reliability and effi­ edge computing device, as shown in Fig. 1.
ciency, how to use sensors, communication devices, data acquisition and Fig. 1 clearly shows the functions of the cloud and the computational
processing technology, connecting the nodes of the electric power sys­ scope of the edge computing devices. Due to the participation of new
tem to achieve real-time data collection, transmission and analysis has energy sources, its optimized dispatch cycle has strict requirements.
become a current hot spot. Literature [27] proposed a network recon­ Therefore, the establishment of the cloud partition model is set to the 15
figuration strategy based on cloud-side collaboration in response to the min level, i.e., the power system voltage, power, equipment status, ac­
traditional centralized distribution network reconfiguration problems tions and other information stored in the edge devices are called up
such as long computation time and heavy solving task. A differential every 15 min at equal intervals. The collected information is utilized by
hybrid model with an event-triggered strategy is established, and a the cloud processor to calculate the reactive power-voltage sensitivity,
graph convolutional neural network with a multilayer perceptron is and the improved modularity function is used for partitioning, and
used in this model to select the optimal solution. Literature [28] pro­ finally the results are stored in the cloud data memory. This step is called
poses an integrated electricity-gas cloud-edge cooperative distributed 15 min cloud front pre-computation.
scheduling structure, which rationally divides the integrated energy On the edge computing side, the voltage threshold device monitors in
system scheduling problem into multiple sub-problems, sets up cloud real time information such as the grid operation state within the range it
computing centers in different energy networks, and sets up edge is responsible for, triggers an action condition of the edge computing
computing units in the energy centers, so as to solve the cost and carbon device when voltage overrun occurs, and immediately sends a demand
emission problems of regional integrated energy systems. Literature to the superior cloud processor to invoke the 15 min partitioning result,
[29] proposed a cloud-edge cooperative scheduling framework for and the cloud processor sends down the data stored in the cloud data
regional multi-energy systems. The regional multi-energy system is memory to each edge computing device, dividing the power system into
divided into five service layers, and corresponding computing servers a number of subregions, each of which region is responsible for at least
are configured in different scheduling service layers to optimize the one edge computing device, and according to the subregion it is in, the

2
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

Fig. 1. Cloud edge collaborative optimization mode.

edge computing device chooses whether to perform MISOCP optimiza­ cloud processor for further analysis and processing.
tion or not. The MISOCP optimization is adjusted according to mini­ 2) The cloud processor performs reactive-voltage sensitivity calcu­
mizing the network loss considering the reactive power constraints of lation SQU,ij based on the collected 15 min power system network
the PV plants, SVCs and SCs, the active power constraints of the PV operation information, and improves the sensitivity index based on
plants and ESSs, the voltage and current constraints, and so on. This step electrical distance.
is called minute-level reactive power-voltage-network loss optimization 3) For the improved sensitivity metrics, the cloud processor parti­
adjustment. tions them by the improved modularity function partitioning strategy
Thus, in this paper, voltage optimization is performed according to and uploads the results to the cloud memory for the edge computing
the cloud-side cooperative two-tier network architecture. devices to call after voltage crossing.
The cloud front-end precomputation proposed in this paper is a 4) Under the normal operation state, the cloud front-end pre­
centralized training model for power system regions from the cloud. The computation step is performed every 15 min.
cloud front-end precomputation is based on the power system model, Eventually, the 15 min cloud front pre-computation is realized. The
which calls the data of each node every 15 min and utilizes the parti­ specific logic is shown in Fig. 2.Ensure that you return to the ‘Els-body-
tioning method to partition the system and sends the results to the edge text’ style, the style that you will mainly be using for large blocks of text,
computing devices in time to clarify the area that each edge device is when you have completed your bulleted list.
responsible for. Fig. 2 demonstrates the 15 min cloud front pre-calculation logic,
The cloud front pre-computation steps are as follows: which is an important preparatory part of the cloud-side cooperative
1) During the normal operation of the power system, the edge voltage optimization method for power systems. In this process, it is first
computing device collects key information of the power system network necessary to collect various data of the power system, carry out pre-
every 15 min, including: network topology, load data, generator data, processing calculations, divide the nodes into regions, and divide the
branch circuit data, voltage level, operation status, etc. After these data results. Every 15 min, the calculation results are updated to ensure the
are collected and organized, they are packaged and uploaded to the real-time and accuracy of the system status. When there is a call demand

3
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

to the terminals in the sub-area. Through minute-level reactive power-

voltage-network loss optimization and adjustment, combined with 15
min − level cloud-based pre-calculation, the optimal control of the
power system is ultimately realized, which greatly improves the oper­
ating efficiency of the power system.

3. Cloud-based partitioning strategies

The distribution network partitioning problem in the cloud is a

complex network partitioning problem characterized by the fact that the
distances between nodes can be measured not only by traditional simi­
larity definitions such as Euclidean distances, but also by using electrical
distance definitions or modularity functions with actual physical
meaning. Many scholars have attempted to implement partitioning of
distribution networks using improved modularity function community
discovery algorithms or customized electrical distances combined with
clustering algorithms. In this paper, we propose a zoning strategy based
on 15 min cloud computing by combining voltage sensitivity and
modularity function.

3.1. Cloud processor-based reactive-voltage sensitivity calculation

The cloud processor accesses the grid information collected by the

edge computing device every 15 min. Among these information, the
voltage magnitude information is more coupled with reactive power and
less coupled with active power, therefore, the node information in the
grid information is divided into a set of reactive nodes {G} that can
Fig. 2. 15 min cloud pre calculation logic.
provide reactive power and load nodes {L}. The control model of the
system containing the above nodes is:
[ ] [ ][ ]
from the edge computing device, the latest results are sent down to the ΔQG HGG NGL ΔUG
= (1)
edge computing device to support real-time minute-level reactive ΔQL JLG LLL ΔUL
power-voltage-network loss optimization adjustment. Through 15 min
Where: HGG、NGL、JLG、LLL are the elements in the control matrix,
cloud front pre-calculation, the operation and management efficiency of
ΔQG is the reactive power variation at the reactive node, ΔQL is the
the power system can be effectively improved.
reactive power variation at the load node, ΔUG is the reactive node
When voltage overrun occurs in the power system, the edge
voltage magnitude variation and ΔUL is the load node voltage variation.
computing devices perform reactive power regulation of PV plants, SVCs
If the node is both reactive and load node, equation (1) can be
and SCs, and active power regulation of PV plants and ESSs in this sub-
changed to:
area based on the results of the cloud-based pre-computation to alleviate
the voltage overrun problem in the grid and reduce the network loss. [ΔQG ] = [ HGG LLL ][ΔUL ] (2)
The steps of co-optimization calculation by edge computing devices
The injected reactive power at the load node is assumed to be con­
are as follows:
stant, i.e., ΔQL = 0. By eliminating the reactive node voltage variation
(1) The voltage threshold device detects the voltage overrun signal,
ΔUG from Eq. (1), the sensitivity expression between the unit voltage
triggers the action conditions of the edge device, and sends the partition
magnitude at node i of the load and the reactive power variation at node
adjustment demand for this time period to the cloud processor.
j of the reactive j can be obtained as:
(2) The cloud processor fetches the partitioning result of this time
period in the cloud memory and sends it down to each edge computing ΔUL,i JLG
SQU,ij = = (3)
device. ΔQG,j JLG NGL − LLL HGG
(3) Each edge computing device determines the nodes to be
controlled by this device according to the partitioning results sent down When node i is both reactive and load, according to the expression of
from the cloud. Eq (1) in the text, it can be obtained:
(4) The edge computing device uses the preset voltage optimization ΔQG = HGG ⋅ΔUG + NGL ⋅ΔUL (4)
control model to carry out optimization solving, and the specific solving
results include the control instruction values of PV power plant, ESS, ΔQL = JLG ⋅ΔUG + LLL ⋅ΔUL (5)
SVC, SC and so on. And the solution results are sent down to each
JLG and NGL can be obtained from the following
measurement and control terminal.
Ultimately, it realizes minute-level reactive power-voltage-network ΔQG − NGL ⋅ΔUL
HGG = (6)
loss optimization adjustment. Its specific logic is shown in Fig. 3. ΔUG
Fig. 3 demonstrates the minute-level reactive-voltage-network loss
optimization adjustment, which is the final optimization link in the ΔQL − LLL ⋅ΔUL
JLG = (7)
cloud-edge collaborative voltage optimization and control method, in ΔUG
which the cloud-front pre-calculation results are called by the edge From equation (3), the node voltage amplitude variation can be
computing device, and the optimization algorithm is utilized to adjust expressed as:
the system to minimize the active network loss as the objective function,
and the adjustment process meets, for example, the voltage constraints ΔU = SQU,ij ΔQ (8)
and other requirements. Finally, the optimization results are sent down The traditional reactive-voltage sensitivity lacks the systematic

4
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

Fig. 3. Minute level reactive power voltage network loss optimization adjustment logic.

consideration of correlation and connectivity among nodes, therefore, the partitioning problem of complex networks [31], and the modularity
the introduction of the concept of electrical distance in the cloud-based function can measure the node partitioning status in a high-speed and
reactive-voltage sensitivity concept proposed in this paper improves the effective way. Applying the modularity function to cloud pre-
reactive-voltage sensitivity, which enables the cloud processor to more computation can provide accurate guidance for power system parti­
accurately assess the degree of coupling among nodes and improves the tioning. The module degree function formula is defined as follows:
accuracy and effectiveness of the reactive-voltage sensitivity. The elec­ ⎧
0, Partitions i and j are not connected
trical distance can be expressed as: ⎪
⎨
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ Q= 1 ∑ [ ki kj
] (11)
dij = R2ij + X2ij (9)
⎪
⎩ A ij − δ(i, j), Partitions i and j are connected
2m ij 2m

Where: dij denotes the electrical distance between node i and node j, Where: Aij is the connectivity line between node i and node j. If node i
and Rij and Xij denote the resistance and reactance between node i and is connected to node j, then Aij = 1, otherwise Aij = 0. Based on the
node j, respectively. characteristics of Aij, it is known that the connectivity matrix of the
The reactive-voltage sensitivity is improved by using the electrical unphased graph is a symmetric array along the main diagonal, i.e., Aij =
distance as a weighting factor and the improved reactive-voltage Aji. If node i is in the same partition as node j, the δ-function is 1,
sensitivity can be expressed as: otherwise δ(i,j) = 0.
ki denotes the sum of all weights connected to node i:
SʹQU,ij = SQU,ij × dij (10)
∑
ki = Aij (12)
By introducing the concept of electrical distance, the improved i
reactive-voltage sensitivity index more accurately responds to the in­
fluence between nodes, improves the accuracy of the sensitivity calcu­ m denotes the sum of all weights in the network.
lation, is closer to the operating state of the actual power system, and ∑∑
m= Aij (13)
better responds to the dynamic characteristics of the system by i j
combining with 15 min electrical feature information extraction.
Based on the improved sensitivity analysis above, the module degree
3.2. A Reactive-Voltage sensitivity partitioning framework based on function is improved as follows:
improved modularity functions 1 ∑ n ∑ n
ki kj
QVQ = [wij − ]δ(i, j) (14)
2mQ i=1 j=1 2mQ
Newman et al. introduced the modularity function in order to solve

5
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

Where: wij is the improved reactive-voltage sensitivity value, i.e., wij on cloud computing makes full use of cloud computing resources and
= Aij. ki then denotes the sum of all the improved reactive-voltage provides a new partitioning idea for the power system. By front-end
∑
sensitivity values connected to the node, i.e.,:ki = nj=1 wij precomputing reactive-voltage sensitivity combined with the
The specific steps to divide areas according to QVQ are (see Fig. 4).: improved modular degree function, the system nodes can be divided into
1) Obtain the data of the distribution network, treat each node as a different partitions. This partitioning strategy can better guide the
cluster, and calculate the modularity value according to (14); management of the power system and contribute to the realization of the
2) Initial node i starts to randomly select node j for combination to goal of smart grid.
obtain a new cluster, calculate modularity (14), and calculate modu­
larity increment. If the modularity increment is positive, node i and node 4. In-zone voltage regulation strategy based on edge computing
j are regarded as the same cluster; devices
3) Consider the current cluster as a new cluster, combine with other
clusters, and repeat step 2). The voltage adjustment strategy in the zone based on edge
4) When no node can be integrated and the modularity function computing devices is based on the aforementioned cloud-based zoning
reaches the maximum, the clustering process stops, and the clustering is strategy. With the increasing scale and complexity of the power system,
the optimal clustering result. the traditional power system management has been facing great chal­
The modularity function is calculated using the network topology, lenges. In the context of new energy access, this section firstly analyzes
and the coupling degree between different nodes can only be charac­ the optimal control model of PV power plant, which takes the reactive
terized according to the network topology, and the nodes with different power compensation capacity of PV power plant into consideration,
coupling degrees can be divided into optimal clusters. However, if there which is of great significance for maintaining the stable operation of
is a high proportion of distributed photovoltaic access in the distribution power system and improving the power quality of power grid. Secondly,
network, it is unreasonable to divide the cluster only according to the a minute-level optimal control model for edge computing devices is
topology of the distribution network. Considering the high proportion of presented, which takes minimizing active network loss as the objective
distributed photovoltaic reactive power involved in voltage regulation function and considers Distflow constraints, node voltage and line cur­
and the dispersion of photovoltaic installation location. After the rent constraints, PV output constraints, SC constraints, SVC constraints,
partition step is completed, constraints are added to judge the rationality and ESS constraints. Finally, voltage regulation is performed by the
of the partition result. cooperative optimization computation of edge computing devices
Through the improved modular degree function partitioning strat­ described above.
egy, the power system area can be formally decoupled in the cloud
processor, so that the partition has the condition of in situ voltage 4.1. Optimal control model for photovoltaic power plants
control, realizing the characteristics of strong coupling within the
partition and weak coupling between the intervals; and the partitions The large-scale access of new energy sources to the power system
are strictly connected to each other, and there is no isolated area or causes the power system to face a series of challenges, such as: unstable
node, which is conducive to the improvement of the stability and reli­ power supply, voltage fluctuation and other problems. However,
ability of the system. photovoltaic power plant, as a representative of new energy, is different
The improved modular degree function partitioning strategy based from the conventional power source, and the power output situation of
photovoltaic power plant is not correlated with the grid load, which is

Fig. 4. The specific steps to divide areas.

6
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

shown in Fig. 5. The active voltage regulation function possessed by PV pled partitioned grid, the optimization model is established as shown
power plants provides new possibilities for power system voltage opti­ below:
mization. However, the participation of PV power plants in the active ∑
voltage regulation of the grid should not only take into account the in­ min lij rij (17)
i,j∈Ωline
tensity of light affecting the PV output situation, but also need to take
into account the size of the load. Therefore, this paper introduces the Where: Ωline is the set of lines; lij is the squared value of the current
active output rate of PV power plants and the reactive output rate of PV flowing through line i-j; rij is the resistance of line i-j.
power plants to more accurately assess the ability of PV power plants to 2 Constraints for edge computing devices.
regulate the grid voltage so as to maintain the stable operation of the The constraints include Distflow current constraints, node voltage
power system. and line current constraints, PV plant output constraints, SC constraints,
1 PV power plant active output rate. SVC constraints and ESS constraints.
Define the ratio of the active output of the PV plant to the rated (1) Distflow tide constraints.
power of the PV plant at the current moment as its active output rate, In order to enable the edge computing device to better handle integer
and the expression is shown below: variables and to improve the efficiency and reliability of the solution of
α = PPV PV
(15) the optimization problem, the new variables vi and lij are utilized instead
OUT /SΣ
of the squared quantities (Vi)2 and (Iij)2.
∑
Where: PPV OUT is the active output of the PV plant, and SPV is the
vi = (Vi )2 (18)
rated power of the PV plant.
2 Reactive power output rate of PV plant. ( )2
lij = Iij (19)
Define the ratio of the reactive power output of the PV plant to the
rated power of the PV plant at the current moment as its reactive power The original tidal current constraints are transformed into tidal
output rate, and the expression is shown below: current constraints based on second-order cone planning, and the
constraint equations can be expressed as follows:
β = QPV PV
OUT /SΣ (16)
⎧ ∑ ∑
⎪ Pij = Pjk + rij lij + Pj
Where: QPV OUT is the reactive power output of the PV plant, and the ⎪
⎪
⎪ ij∈Ωline k∈(j,k)
⎪
further can be obtained as α2 + β2 ≤ 1.
⎪
⎪ ∑ ∑
⎪
⎪
The I region in Fig. 5 is the inductive reactive power emitted from the
⎨ Qij = Qjk + xij lij + Qj
ij∈Ωline k∈(j,k) (20)
PV power plant, and the optimization solution in this region is conducive ⎪
⎪
⎪
to the enhancement of the system low voltage; the II region is the
⎪
⎪
⎪
⎪ Pj = Pj + Pj + Pdch,j − Pch,j − Pj
G PV ESS ESS load
⎪
⎪
capacitive reactive power emitted from the PV power plant, and the ⎩
Qj = QGj + QPV j + Qj + Qj − Qload
SC SVC
j
optimization solution in this region is conducive to the recovery of the
system over-voltage. By fully utilizing the active voltage regulation ( ) ( )
capability of PV power plant, it can quickly respond and regulate the vj = vi − 2 rij Pij + xij Qij + rij2 + x2ij lij (21)
grid voltage when it fluctuates, thus reducing the instability of the grid
voltage. P2ij + Q2ij = lij vi (22)
⃦ ⃦
⃦ 2Pij ⃦
⃦ ⃦
4.2. Collaborative optimization control model based on edge computing ⃦ 2Qij ⃦ ⩽lij + vi , ∀i, j ∈ Ωline (23)
⃦ ⃦
devices ⃦ lij − vi ⃦
2

Edge devices are playing an increasingly important role as a new where Pij and Qij are the active and reactive power flowing through the
intelligent tool. For the Mixed-integer Nonconvex Programming (MINP) line connected to node i-j, respectively; Pj and Qj are the active and
problem in optimal control in power systems, edge computing devices reactive power input to node j, respectively; rij and xij are the resistance
solve this challenge by converting it into a mixed-integer second-order and reactance of line i-j, respectively; PG j and QG j are the active and
cone planning problem. In this section, the MISOCP model based on reactive power input to the generator at node j, respectively; Ωline is the
edge computing devices will be developed, which consists of two parts: set of nodes; PPV j and QPV j are the active and reactive power of the PV
objective function and constraints. input at node j, respectively; QSC j is the reactive power of the group-
1 Objective function of edge computing device. switching capacitor at node j, respectively; PESS dch,j and PESS ch,j
With the objective of minimizing the total power loss in the decou­ are the discharging and charging power of the energy storage system at
node j, respectively; QSVC j is the reactive power of the SVC at node j;
and Pload j and
Qload j are the load at node j, respectively. active power and reactive
power.
(2) Node voltage and line current constraints
⎧
⎨ U2 ⩽vj ⩽U2
j,min j,max
(24)
⎩ 0⩽lj ⩽Iij,max
2

Where: Uj.max and Uj,min are the maximum and minimum values of
the voltage at node j, respectively; Iij,max is the maximum value of the
current at line i-j.
(3) Photovoltaic power plant output constraints
⃒ ⃒ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( )2 ( )2
⃒ PV ⃒
⃒Qj ⃒⩽ SPV
j − PPV j (25)
Fig. 5. Power output diagram of photovoltaic power station.

7
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

Where: the SPV j is the total PV power at node j. Finally, the optimized result is sent to the corresponding equipment
(4) SC constraints terminal to realize the minute level voltage optimization adjustment.
⎧ The implementation process of the method is shown in the Fig. 6.
⎨ QSC = kSC ΔQSC
j j j
(26)
⎩ 0⩽kSC SC
j ⩽Kj,max
6. Calculus analysis

Where: ΔQSC j is the compensated reactive power per group of group 6.1. Introduction to the algorithm
switching capacitors at node j; kSC j is the number of group switching
capacitor actions at node j, and KSC j,max is the maximum number of In order to verify the effectiveness and accuracy of the proposed
group switching capacitor actions at node j. cloud-side cooperative voltage optimization control method with PV
(5) SVC constraints power plants, the improved IEEE33 node arithmetic example[32] is
selected for analysis in this paper.
QSVC SVC
j,min ⩽Qj ⩽QSVC (27)
j,max Using MATLAB R2021a to calculate the improved module degree
Where: QSVC j,min, QSVC j,max are the lower and upper limits of the partitioning strategy based on cloud. Edge optimization using MATLAB
SVC output reactive power at node j, respectively. combined with YALMIP to develop voltage optimization program.
(6) ESS constraints. CPLEX 12.6 for simulation solution. The simulation test environment
ESS constraints mainly include charging and discharging power adopted in this paper is: Intel Core i7-10700 @2.9 GHz, ram 32G.
constraints and battery storage capacity constraints. In the improved IEEE33 node example, there are 9 PV plants, 1 set of
⎧ SC device, 2 sets of SVC device and 1 set of ESS device. The PV plants
⎪
⎪
⎪ 0⩽PESS
ch,j ⩽Ach,j Pch,j,max
ESS ESS numbered PV1-PV9 are located at nodes 5, 8, 11, 15, 18, 21, 25, 29, and
⎨
33, respectively, and the capacity of each PV plant is shown in Table 1.
0⩽Pdch,j ⩽Adch,j Pdch,j,max
ESS ESS ESS
(28)
⎪
⎪ The SC unit is installed at node 7, which is divided into five gears, with a
⎪
⎩ A + A ⩽1
ESS
ch,j
ESS
dch,j single set of 100 kVar. The SVC units are installed at nodes 13 and 27,
both with a capacity of 300 kVar. The ESS unit is installed at node 31,
⎧
1 ESS with a capacity of 750 kW. The maximum and minimum safe voltages in
⎨ EESS
⎪ = EESS
j,0 + ηPch,j Δt −
ESS
Pdch,j Δt
j
η (29) the calculations are 1.05p.u. and 0.95p.u. respectively. The root node is
⎪
⎩ node 1. The improved IEEE33 node example connection diagram is
0.18EESS ESS ESS
j,max ⩽Ej ⩽0.8Ej,max
shown in Fig. 7.
There are certain requirements for the setup of edge computing de­
Where: AESS ch,j and AESS dch,j are Boolean variables denoting the
vices: 1) Edge computing devices can be set up in the central node of the
charging and discharging states of the energy storage device at node j,
network topology; 2) If a node has a large load and requires more data
respectively; PESS ch,j,max and PESS dch,j,max are the maximum power
processing and optimization control, edge computing devices need to be
limit of charging and discharging of the energy storage device at node j,
installed in these nodes; 3) Edge computing devices are installed in the
respectively; Δt is the discharging time; η is the charging and discharging
nodes with better communication performance to ensure the reliable
efficiency, which is taken to be 0.9; EESS j,0 is the initial state of charge
transmission of information. If there is a lack of edge computing devices,
of the energy storage device; EESS j,max is the maximum power that can
neighboring edge computing devices are required to be responsible for
be stored by the energy storage device at node j.
two or more partitions. Therefore, the improved IEEE33 node edge
According to the above edge computing device-based voltage
computing devices are installed at nodes 2, 3, 6, 10, 12, 15, and 29.
adjustment strategy in the zone, the voltage threshold device monitors
the voltage status in real time, and when the voltage is detected to be out
of limits, the edge computing device responds immediately and initiates 6.2. Example results
the adjustment mechanism. First, the edge computing device requests to
fetch the cloud-based front-end pre-computation partitioning results for (1) Cloud front pre-computed partitioning results.
this time period, and obtains the required power system network oper­ At a certain moment, the load of each node surges to 1.2 times of the
ation information. Second, the cloud front-end precomputation sends baseline load, and voltage overruns occur, and the voltage situation is
the partitioning results to each edge computing device, and the edge shown as the blue line in Fig. 8, which adopts the voltage optimization
computing device performs MISOCP optimization for the partition it is control method of this paper. The edge computing device calls the cloud
responsible for. Finally, the optimized results are sent down to the front precomputation result at the beginning of this time segment.
corresponding device terminals to achieve minute-level voltage opti­ The cloud front precomputation at the beginning of this time
mization adjustment. This process makes full use of the computing segment is performed by the cloud processor to synthesize the network
power of the edge computing devices and improves the speed of power key information collected by the edge computing device, and the par­
system fault recovery. titioning is performed in accordance with the reactive-voltage sensi­
tivity partitioning strategy of the improved modularity function. When
5. The proposed cloud-edge collaborative optimization control the number of clusters is 4, the improved modularity function achieves
strategy the maximum value ρ = 0.144, so the optimal number of clusters is 4.
The partitioning result is shown in Fig. 9, which is divided into four
Based on the voltage adjustment strategy of the edge computing regions, A1, A2, A3 and A4. The included nodes are shown in Table 2.
device, the voltage threshold device monitors the voltage status in real Node 1 is the root node and does not participate in grid partitioning.
time. When the voltage exceeds the limit, the edge computing device From the results, the cluster partitioning results are correlated with the
will respond immediately and start the adjustment mechanism. geographic attributes of the load nodes, this is because the improved
First, the edge computing device requests the cloud pre-calculation modularity function contains the concept of electrical distance which is
partition results at this time, and obtains the required power system related to geographic attributes, the sensitivity between nodes that are
network operation information. directly connected is high, and the sensitivity between nodes that are not
Secondly, the cloud pre-calculation delivers the partition results to directly connected is low, so that the partitioning results are correlated
each edge computing device, and the edge computing device performs with the geographic attributes.A1 is regulated by the edge computing
MISOCP optimization for the responsible partition. device of node 2 in the area, A2 is regulated by the node 6′s edge
computing device, A3 by the edge computing device at node 10, and A4

8
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

Fig. 6. The proposed cloud-edge collaborative optimization control strategy.

∑
Table 1 γ = avg( SʹQU,ij ) (30)
Connection nodes and rated capacity of photovoltaic power stations. i,j∈Ck

PV Connection node Rated capacity /kW If 0 < γ < 0.2, it is considered that the internal coupling is low.
PV1 5 500 If 0.2 < γ < 1, it is considered that the coupling inside the partition is
PV2 8 500 medium.
PV3 11 500
If 1 > γ, the partition is considered to be highly coupled.
PV4 15 500
PV5 18 750 The traditional concept of coupling degree usually uses a standard­
PV6 21 500 ized approach, limiting the index to between [0,1]. However, the
PV7 25 750 improved reactive voltage sensitivity, combined with the concept of
PV8 29 750 electrical distance, amplifies the actual coupling between nodes within
PV9 33 750
the partition. Therefore, the coupling degree in the absolute region is not
a normalized value, and its index will be greater than 1. This method is
by the edge computing device at node 12. closer to the actual physical characteristics of the system and helps to
The cloud processor sends the A1, A2, A3, and A4 partitioning results better evaluate the synergies between the nodes of the system.
to each of the four edge computing devices, ensuring that each device is In the example analysis, the improved IEEE33 node example is
responsible for 1 partition. These partitioning results are based on the divided into four partitions, namely A1, A2, A3 and A4. The absolute
topology of the power system combined with an improved modularity coupling degrees calculated for the four zones are 1.31, 1.09, 1.16 and
function to provide the basis for subsequent voltage optimization. 2.42, respectively. The absolute coupling degree γ of the four zones is
Because the traditional concept of coupling degree in the area can greater than 1, indicating that the node correlation in the zone is strong,
not accurately reflect the improved system characteristics. In this paper, and the control demand must be highly coordinated to meet the voltage
the reactive power − voltage sensitivity is improved. Therefore, the regulation demand. Therefore, it is reasonable to use the method pro­
concept of absolute regional coupling degree is proposed, and its posed in this paper to partition.
expression is as follows: In A4 partition, the edge computing device on node 15 does not work
because the edge computing device on node 12 in A4 partition performs
the voltage optimization control task in A4 partition. Partition A4 has

9
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

Fig. 7. Schematic diagram of improved IEEE33 node calculation example.

compliance with the requirements, and the lowest point of voltage still
appears at node 18 at 0.961p.u.. Before adjustment, the active network
loss is 203 kW, and after adjustment, it is reduced to 70 kW. After
adjustment, the number of SC input groups is 2, and the total reactive
power provided is 200 kVar; the reactive power provided by the SVC
connected at node 27 is 300 kVar; and the active power provided by the
ESS connected at node 31 is 513 kW. The active power rate and reactive
power rate of PV power stations connected at nodes 5, 14, and 29 are
shown in Table 3. rate are shown in Table 3.
The active output rates of the three PV plants described in Table 3 are
90.96 %, 80.60 % and 66.67 %, respectively, and the reactive output

Table 2
Partition Results.
Fig. 8. Voltage distribution of system nodes before and after adjustment.
Partition number The zone contains nodes

multiple edge computing devices and redundancy. When a device on A1 2,3,4,5,19,20,21,22,23,24,25

node 12 fails or fails to perform a task due to other reasons, the device on A2 6,7,8,9,26,27,28,29,30,31,32,33
A3 10,11
node 15 automatically takes over the task. This improves system fault
A4 12,13,14,15,16,17,18
tolerance and reliability.
(2) Edge computing device intra-zone voltage adjustment results.
The edge computing devices are optimized and calculated using the
Table 3
edge computing device-based intra-zone voltage adjustment strategy
Active and reactive power output rates of photovoltaic power stations.
proposed in this paper, and the final results are shown as the red line in
Fig. 8. Photovoltaic plant number Connection node α Eq(15) β Eq(16)
Before adjustment, the system node voltages operated between PV2 8 90.96 % 30.64 %
0.894p.u. and 1.00p.u., with the lowest point of voltage occurring at PV4 15 80.60 % 8.40 %
node 18 at 0.894p.u.. After adjustment, the voltages are all in PV8 29 66.67 % 46.67 %

Fig. 9. Cloud pre calculation partition results.

10
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

rates are 30.64 %, 8.4 % and 46.67 %, respectively, which are all within
their rated capacities.
According to the cloud-side cooperative voltage optimization
method proposed in this paper, only the control capacity of three PV
power plants is utilized because the system is already in a stable oper­
ation state after the regulation of these three PV power plants, and the
operation status of each power equipment is within the normal range.
Therefore, there is no need for extra power output from the redundant
PV plants to regulate the grid voltage.
(3) 24 h results considering the fluctuation of output and load of
distributed energy.
Choose the day with the highest light intensity in July. The daily
residential and commercial load curves contained in the distribution
network are in line with the residential, commercial and industrial load
demand in July. The total load of the system and the change of photo­
voltaic output 24 h a day are shown in the Fig. 10.
After the optimization of the method proposed in this paper, the
results of the whole day are shown in Fig. 11. Fig. 11. The results of node voltage results and increase in 24 h.
It can be clearly seen from the above data that the control method
proposed in this paper can effectively improve the voltage level of the
Table 4
distribution network in 24 h, so that the voltage of each node does not Comparison of time required and active power loss between traditional opti­
exceed the limit value. The effectiveness and applicability of this method mization methods and optimization methods in this paper.
are further illustrated.
Method Time System active network
(4) Comparison of algorithms. required loss
In order to verify the feasibility of the proposed cloud-edge cooper­
Voltage Optimization Adjustment 3.629 s 72 kW
ative optimization method, two voltage optimization adjustment
Method 1
methods are added for comparative analysis. Voltage Optimization Adjustment 1.863 s 72 kW
Voltage optimization adjustment method 1: Using the partitioning Method 2
method and MISOCP method in this paper, firstly, the voltage threshold Optimization methods for this paper 1.624 s 70 kW
device detects the voltage overrun and triggers the edge computing
device to fetch the data from each terminal; secondly, the edge
computing device. First, the voltage threshold detection device detects
computing device uploads the data to the cloud processor, which per­
the voltage overrun and triggers the edge computing device to fetch the
forms the system partitioning process and sends the results down to the
data from each terminal; second, the edge computing device performs
edge computing device; finally, the edge computing device in each
the voltage optimization calculation of this power system using the
partition performs the MISOCP optimization computation and sends the
MISOCP method; finally, the edge computing device sends the results
results down to each device for execution. This method all performs the
down to each device for execution. Voltage optimization for improved
computation after the failure and is not pre-partitioned. Voltage opti­
IEEE33 node arithmetic.
mization for improved IEEE33 node arithmetic.
The voltage of each node in comparison of the voltage optimization
The voltage optimization tuning method 1 has the same results for
adjustment method 2 and the optimization method obtained in this
each node voltage as obtained by the optimization method in this paper,
paper is shown in Fig. 12.
but the time required to run is different. As shown in Table 4. This
From Fig. 12, the lowest point of voltage in the voltage optimization
method is only limited to the local scope of the cloud processor and the
and adjustment method 2 appears at node 33 with a voltage of 0.951p.u.
edge computing device, and cannot fully utilize the advantages of the
The lowest point of voltage in the optimization method of this paper also
cloud and the cloud memory. In contrast, the cloud-edge collaborative
appears at node 33 with a voltage of 0.961p.u.
voltage optimization method proposed in this paper takes advantage of
Voltage optimization and adjustment method 2 optimization method
the synergy between the cloud processor and the edge computing de­
has a larger fluctuation rate of node voltage enhancement than the cloud
vice, and is able to achieve local voltage fine-tuning, making full use of
edge cooperative optimization method proposed in this paper, and the
the cloud’s front-end pre-computing resources and the real-time per­
voltage gap between the nodes is larger. This is due to the fact that the
formance of the edge computing device, and thus improving the voltage
cloud edge cooperative optimization method in this paper is targeted
stability more effectively.
Voltage Optimization Adjustment Method 2: The traditional opti­
mization method, i.e., MISOCP method, is fully utilized by the edge

Fig. 10. The total load of the system and the change of photovoltaic output 24 Fig. 12. Voltage optimization adjustment method 2 optimization method and
h a day. this optimization method system node voltage distribution.

11
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

and can be selectively optimized using compensation equipment based

on regions with large voltage landings. In contrast, the voltage optimi­
zation and adjustment method 2 optimization method is optimized for
the entire grid and does not optimize for specific regions, which is not
flexible and selective.
Voltage optimization adjustment method 1 optimization method
takes the longest time to adjust, which is 3.629 s; voltage optimization
adjustment method 2 optimization method takes 1.863 s to adjust, and
the active network loss is 72 kW; the optimization method in this paper
takes 1.624 s to adjust, and the active network loss is 70 kW, as shown in
Table 4. The calculation speed of the method proposed in this paper is
faster for the following reasons:
(1) The use of edge computing equipment to collect network infor­
mation, upload it to the cloud processor, and store it in the cloud data Fig. 14. Voltage comparison between nodes in a 69-nodes system.
memory after partitioning by the processor, so that the edge computing
equipment can quickly obtain the corresponding partitioning results of verification. To further demonstrate and validate the feasibility of the
this 15 min section when voltage overrun occurs, without the need to proposed method on a large-scale system, this paper perform simulation
carry out a repetitive calculation process; on a 69-nodes system[33].
(2) After the occurrence of voltage overrun, the cloud front pre- The corresponding results of the voltage regulation method in this
calculated data is called up, so that the edge computing device carries paper are obtained by using a 69-nodes system:
out optimized calculation in the partition to which it belongs, which It can be seen that the proposed method is still applicable in the 69-
avoids the process of collecting and sending all the calculation tasks to nodes system, which further demonstrates the effectiveness of the
all the edge computing devices, and effectively shortens the data method.
transmission delay and calculation time;
(3) The edge computing device chooses mixed integer second-order 7. Conclusion
cone optimization, which has faster convergence speed and computa­
tional efficiency compared with other nonlinear planning, and can In this paper, an optimization and control strategy based on cloud-
obtain the optimal solution set in a shorter time. edge collaboration is proposed, and this strategy is divided into a
The method proposed in this paper has lower active network losses cloud front precomputation strategy and an edge voltage optimization
for the following reasons: and control strategy, aiming to improve the restoration efficiency of the
(1) Through the cloud-side cooperative voltage optimization power system. In the cloud front precomputation strategy, the ability of
method, we are able to adjust the power system more accurately, opti­ the edge computing device to collect grid information is utilized to
mize the power allocation at each node, and reduce the active network transmit data to the cloud, and the cloud processor is utilized to establish
loss; an 15 min cloud front precomputation model. In the edge voltage
(2) The partitioning strategy used by the cloud processor is able to optimization and control strategy, the action conditions are triggered by
divide the system into several mutually independent zones, which helps the voltage threshold device, which calls the cloud precomputation
in local voltage adjustment and further reduces the active network loss model stored in the cloud for the current time period, i.e., the partition
of the system; model for the current time period, and sends it to each edge computing
(3) The mixed integer second-order cone optimization method used device, which is responsible for calculating and optimizing the voltage
for the intra-zone voltage adjustment strategy of the edge computing of the nodes in the region, and ultimately realizes the optimization
device, which takes minimizing the active network loss as the objective strategy within the zone at the minute level. The method proposed in
function, is able to find the optimal solution quickly. this paper is able to carry out fine optimization for the local area of the
It can be seen that using the cloud-edge cooperative voltage opti­ power system, with high flexibility and relevance, using cloud-edge
mization method in this paper can greatly improve the recovery speed cooperative technology to realize intelligent management and optimi­
after voltage faults and reduce the active network losses. zation of the power system, and provide strong support for the security
(5) Test in a 69-nodes system (see Figs. 13 and 14). and stability of the power system.
The applicability of the proposed algorithm truly requires further However, the method proposed in this paper also has some

Fig. 13. Topology of a 69-nodes system and partition results using the method presented in this paper.

12
G. Li et al. International Journal of Electrical Power and Energy Systems 167 (2025) 110632

drawbacks. It relies on the performance and reliability of the edge [8] Shen X, Shahidehpour M, Zhu S, Han Y, Zheng J. Multistage planning of active
distribution networks considering the cooptimization of operation strategies. IEEE
computing equipment, and once the equipment fails, it may affect the
Trans Smart Grid 2018;9(2):1425–33.
stability and reliability of the final optimization results. Secondly, due to [9] Song Y, Zheng Y, Liu T, Lei S, Hill DJ. A new formulation of distribution network
the complexity and uncertainty of the power system, the method pro­ reconfiguration for reducing the voltage volatility induced by distributed
posed in this paper may not be applicable to all operating conditions, generation. IEEE Trans Power System 2020;35(1):496–507.
[10] Ma W, Wang W, Chen Z, Ruonan Hu. A centralized voltage regulation method for
resulting in the inability to operate in some extreme cases. The cloud- distribution networks containing high penetrations of photovoltaic power. Int J
based cooperative optimization strategy for extreme cases will be Electr Power Energy Syst 2021;129:106852.
further investigated in the future. Consideration will also be given to [11] Alrashidi M, Rahman S. A bi-level optimization method for voltage control in
distribution networks using batteries and smart inverters with high wind and
distribution network with interconnected switches or flexible multi- photovoltaic penetrations. Int J Electr Power Energy Syst 2023;151:109217.
state switches, which will be a hybrid AC-DC system. [12] Wang S, Liang Du, Fan X, Huang Q. Deep reinforcement scheduling of energy
storage systems for real-time voltage regulation in unbalanced LV networks with
high PV penetration. IEEE Trans Sustain Energy 2021;12(4):2342–52.
CRediT authorship contribution statement [13] Cheng Y, Zhang N, Zhang B, Kang C, Xi W, Feng M. Low-carbon operation of
multiple energy systems based on energy-carbon integrated prices. IEEE Trans
Guoqing Li: Writing – review & editing, Supervision, Conceptuali­ Smart Grid 2020;11(2):1307–18.
[14] Werth A, Kitamura N, Tanaka K. Conceptual study for open energy systems:
zation. Wei Wang: Writing – review & editing, Writing – original draft, distributed energy network using interconnected dc nanogrids. IEEE Trans Smart
Data curation, Conceptualization. Dan Pang: Validation, Software, Re­ Grid 2015;6(4):1621–30.
sources, Investigation, Funding acquisition. Zhipeng Wang: Validation, [15] Ouammi A, Dagdougui H, Dessaint L, Sacile R. Coordinated model predictive-based
power flows control in a cooperative network of smart microgrids. IEEE Trans
Software, Resources, Methodology, Investigation, Funding acquisition.
Smart Grid 2015;6(5):2233–44.
Weixian Tan: Writing – original draft, Data curation. Zhenhao Wang: [16] Fathi M, Bevrani H. Statistical cooperative power dispatching in interconnected
Writing – review & editing, Project administration, Funding acquisition, microgrids. IEEE Trans Sustain Energy 2013;4(3):586–93.
Conceptualization. Jinming Ge: Writing – review & editing, Writing – [17] Sun Y, Zhang B, Ge L, et al. Day-ahead optimization schedule for gas-electric
integrated energy system based on second-order cone programming. CSEE J Power
original draft, Software. Energy Syst 2020;6(1):142–51.
[18] Zhang C, Xu Y, Dong Z, et al. Three-stage robust inverter-based voltage/var control
Declaration of competing interest for distribution networks with high-level PV. IEEE Trans Smart Grid 2017;10(1):
782–93.
[19] Wang Z, Chen B, Wang J, Begovic MM, Chen C. Coordinated energy management
The authors declare that they have no known competing financial of networked microgrids in distribution systems. IEEE Trans Smart Grid 2015;6(1):
interests or personal relationships that could have appeared to influence 45–53.
[20] Dall’Anese E, Zhu H, Giannakis GB. Distributed optimal power flow for smart
the work reported in this paper. microgrids. IEEE Trans Smart Grid 2013;4(3):1464–75.
[21] Tang Z, Hill DJ, Liu T. Fast distributed reactive power control for voltage
Acknowledgements regulation in distribution networks. IEEE Trans Power Syst 2018;34(1):802–5.
[22] Lingfeng KOU, Ming WU, Yang LI, Xiaoyun QU, Hui XIE, Bo GAO, et al.
Optimization and control method of distributed active and reactive power in active
This work is supported by Science and Technology Project of SGCC. distribution network. Proc CSEE 2020;40(6):1856–63.
(No.SGJLCC00KJJS2203595). [23] Zhao Bo, Zhicheng Xu, Chen Xu, Wang C, Lin F. Network partition-based zonal
voltage control for distribution networks with distributed PV systems. IEEE Trans
Smart Grid 2018;9(5):4087–98.
Data availability [24] Yilmaz M, ElShatshat R. Zone-oriented 2-stage distributed voltage control
algorithm for active distribution networks. Elec Power Sys Res 2023;217:109127.
The authors do not have permission to share data. [25] Meng L, Yang X, Zhu J, Wang X, Meng X. Network partition and distributed voltage
coordination control strategy of active distribution network system considering
photovoltaic uncertainty. Appl Energy 2024;362:122846.
References [26] Li Z, Qiuwei Wu, Chen J, Huang S, Shen F. Double-time-scale distributed voltage
control for unbalanced distribution networks based on MPC and ADMM. Int J Elect
[1] Tian Liu, Xiaoqi Tan, Bo Sun, Yuan Wu, Danny HK. Tsang. Energy management of Power Energy Sys 2023;145:108665.
cooperative microgrids: A distributed optimization approach. International Journal [27] Yue D, He Z, Dou C. Cloud-edge collaboration-based distribution network
of Electrical Power & Energy Systems, 2018, 96: 335-346. reconfiguration for voltage preventive control. IEEE Trans Ind Info 2023;19(12):
[2] Zheng Y, Song Y, Hill DJ, et al. Online distributed MPC-based optimal scheduling 11542–52.
for EV charging stations in distribution systems. IEEE Trans Ind Inf 2018;15(2): [28] Zhu Xu, Yang J, Zhan X, Sun Y, Zhang Y. Cloud-edge collaborative distributed
638–49. optimal dispatching strategy for an electric-gas integrated energy system
[3] Liang H, Lizhong Xu, Zhao Y, Zhang Si, Chen C, Tang J, et al. Day-ahead joint considering carbon emission reductions. Int J Electr Power Energy Syst 2022;143:
scheduling of multiple park-level integrated energy systems considering coupling 108458.
uncertainty of electricity-carbon-gas prices. Int J Electr Power Energy Syst 2024; [29] Liu S, Teng Y, Cheng SQ, NingWei Xu, Sun P, Zhang K, et al. A cloud-edge
158. cooperative scheduling model and its optimization method for regional multi-
[4] Attar M, Homaee O, Falaghi H, et al. A novel strategy for optimal placement of energy systems. Front Energy Res 2024;12.
locally controlled voltage regulators in traditional distribution systems. Int J Electr [30] Raak F, Susuki Y, Hikihara T. Data-driven partitioning of power networks via
Power Energy Syst 2018;96:11–22. koopman mode analysis. IEEE Trans Power Syst 2016;51(6):31–5.
[5] Shen X, Shahidehpour M, Han Y, Zhu S, Zheng J. Expansion planning of active [31] Girvan M, Newman ME. Community structure in social and biological networks.
distribution networks with centralized and distributed energy storage systems. Proceedings of the National Academy of Science of the United States of America
IEEE Trans Sustain Energy 2017;8(1):126–34. 2002;99(12):7821–6.
[6] Ji H, Wang C, Li P, et al. A centralized-based method to determine the local voltage [32] Baran ME, Wu FF. Network reconfiguration in distribution systems for loss
control strategies of distributed generator operation in active distribution reduction and load balancing. IEEE Trans Power Delivery 1989;4(2):1401–7.
networks. Appl Energy 2018;228:2024–36. [33] D. Das, Optimal placement of capacitors in radial distribution system using a
[7] Sun H, et al. Review of challenges and research opportunities for voltage control in Fuzzy-GA method, International Journal of Electrical Power & Energy Systems,
smart grids. IEEE Trans Power System 2019;34(4):2790–801. Volume 30, Issues 6–7, July–September 2008, Pages 361-367.

13