This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
Pinning-based Hierarchical and Distributed
Cooperative Control for AC Microgrid Clusters
Xiangyu Wu, Member, IEEE, Yin Xu, Senior Member, IEEE, Jinghan He, Fellow, IEEE,
Xiaojun Wang, Member, IEEE, Juan C. Vasquez, Senior Member, IEEE, and Josep M. Guerrero, Fellow, IEEE
Abstract—With the large-scale application of microgrids
(MGs), interconnecting nearby MGs to form a microgrid cluster
(MGC) enables a higher utilization of renewable sources. This
paper presents a pinning-based hierarchical and distributed
cooperative control strategy for AC MGC, which includes DG-
layer, MG-layer and MGC-layer controls. The DG-layer control
regulates the local voltage/current of each DG unit. The MG-layer
control is performed for each individual MG, managing DG units
in a cooperative manner through several sparse communication
networks. By representing each MG as an MG agent, the MGC-
layer control coordinates MGs based on a higher level peer-to-
peer communication network among MG agents. The interaction
between MG-layer and MGC-layer is established by pinning some
DG units of each MG to communicate with the MG agent.
Compared with the existing literature, the contributions of this
paper are (i) simultaneously realizing multiple control objectives
in the MGC system level including frequency/voltage regulation
and active/reactive power sharing, (ii) presenting a systematic
approach to construct a unified small-signal dynamic model of the
MGC system, and (iii) performing a detailed small-signal stability
analysis to evaluate the system dynamic performance. Time-
domain simulation and experiments are carried out to validate
the effectiveness of the proposed methods.
Index Terms—distributed cooperative control, dynamic model,
hierarchical control, microgrid cluster, small-signal stability.
I. I NTRODUCTION
M ICROGRID (MG) is a viable technical option to sys-
tematically organize the distributed generation (DG)
units and facilitate their local consumption [1]–[3]. With the
large-scale development of MGs, a number of adjacent MGs
could be connected together to form a microgrid cluster (MGC)
in a certain region [4], [5]. Due to the power support capability
among different MGs, the MGC can further promote the
utilization of renewable sources and also enhance the system
resiliency after major disasters [6]–[8].
A desirable control system is of vital importance for an MGC
to ensure its stable and efficient operation. First, compared with
a single MG, the MGC not only needs to consider the control
This work was supported in part by the National Key R&D Program of
China (2018YFB0905200), in part by the National Natural Science Foundation
of China (51807005), in part by the China Postdoctoral Science Foundation
(2017M620600). The work of J. M. Guerrero was supported by the Villum
Investigator Grant from The Villum Foundation under Grant 999730. (Corre-
sponding author: Yin Xu)
Xiangyu Wu, Yin Xu, Jinghan He and Xiaojun Wang are with the School
of Electrical and Engineering, Beijing Jiaotong University, Beijing, 100044,
China (email: wuxiangyu@bjtu.edu.cn; xuyin@bjtu.edu.cn; jhhe@bjtu.edu.cn;
xjwang1@bjtu.edu.cn).
Juan C. Vasquez and Josep M. Guerrero are with the Department of
Energy Technology, Aalborg University, 9220 Aalborg East, Denmark (email:
juq@et.aau.dk; joz@et.aau.dk).
0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
1
of DG units inside each individual MG, but also ought to
realize the coordination control among different MGs. Second,
multiple control objectives both in the individual MG level and
in the whole system level need to be realized for an MGC.
Therefore, considering that each MG can be regarded as a
control area, the MGC control is a kind of multi-area, multi-
layer and multi-objective control problem, which is extremely
challenging.
One strategy to realize the MGC control is the centralized
control method, which requires a central control unit and
bidirectional communication links between the central control
unit and each control device. In [9]–[12], a system central
controller is employed for coordinating each DG unit in the
MGC to realize the system frequency/voltage/power control
objectives. However, the centralized method suffers from the
problem of “single point of failure” for the central control
unit, which significantly decrease the system reliability and
scalability [13].
In contrast, the distributed control method does not rely on
the central control unit. It is based on a sparse communication
network in which each control device requires its own informa-
tion and that of its neighbours to perform control actions [14].
The distributed control method has been widely studied in a
single MG [13], [15]–[18]. For the application of distributed
control in the MGC, relevant works are summarized as follows.
References [19], [20] propose a distributed control method
to realize the control of system frequency, DG units output
voltages or power sharing among DG units in the MGC.
However, all the DG units from different MGs are in the
same communication network. This is similar to control “a
large MG” and thus it breaks the autonomy characteristic of
each MG as an area, which may result in problems of privacy
protection, slow algorithm convergence speed and difficulties in
the MG plug-and-play [21], [22]. To overcome this drawback,
references [23], [24] introduce an MG agent for each MG.
The MG agents communicate with each other to achieve the
coordination among different MGs. For the DG units, only DG
units from the same MG have communications. This method
can reserve the autonomy characteristic of MGs, however,
the control methods in [23], [24] are designed for DC MGC
rather than AC MGC. For AC MGC, reference [21] introduces
an interface controller to mimic the droop characteristic for
each MG, while the newly introduced controller increases the
implementation complexity. In addition, reference [21] aims to
share the output power from the point of common coupling
(PCC) of each MG in proportion to the capacities of MGs.
However, it will result in no power support among MGs when
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
all the loads are inside MGs. In [22], [25], the concept of MG
agent [22] or DG-head [25] (similar to MG agent) is introduced
for AC MGC and corresponding distributed control methods
are proposed, which can realize power support among MGs
regardless of load positions. However, references [22], [25]
have the following limitations.
• The active power sharing among DG units is in proportion
to their capacities. However, an economic operation of
the MGC system, e.g, minimization of generation costs,
is usually more desirable from the perspective of overall
system operation, which cannot be realized in [22], [25].
• In [22], [25], the average value of DG unit output volt-
ages is controlled at the set value. However, actually
the voltages of critical buses in the system should be
controlled to ensure continuous operation of sensitive
loads and facilitate the synchronization of MGC system
to the main grid or the synchronization of sub-MG to the
MGC system.
• The validation of control methods in [22], [25] is based
on time-domain simulations rather than experiments.
Moreover, stability performance is another important con-
cern for the reliable operation of an MGC system. References
[22], [25] provide the MGC system stability analysis results
based on the Lyapunov-based method. However, the Lyapunov-
based method cannot provide the information of damping and
oscillation frequency. In addition, it suffers from the problem
of conservativeness for the stability analysis result, which may
result in undesirable errors. In contrast, the eigenvalue-based
small-signal stability analysis can provide many insights on
the system dynamics. A small-signal dynamic model can be
used for analyzing the system small-signal stability and tuning
control parameters. Although the small-signal dynamic model
of a single MG is well established [26]–[30], to the best of our
knowledge, a small-signal dynamic model for the AC MGC
system and the corresponding detailed stability analysis have
less been studied before. Compared with a single MG, the
modeling of an MGC system is more complicated, since it
needs to consider the interconnection of MGs and the multi-
area/multi-layer controllers.
Motivated by the aforementioned limitations, this paper
presents a pinning-based hierarchical control strategy for is-
landed AC MGC based on the distributed cooperative control
theory. The control strategy consists of three layers, i.e., DG-
layer control, MG-layer control to manage DG units inside
each individual MG, and MGC-layer control to coordinate
MGs. The major contributions of this paper are three-folds.
• Compared with the existing MGC control methods, the
proposed method can simultaneously (i) realize an eco-
nomic operation of the system by minimizing the overall
generation cost, (ii) provide the voltage control capability
for the system critical bus and the point of common
coupling (PCC) of each MG, (iii) restore the system
frequency to the desired (rated) values, and (iv) achieve
an accurate reactive power sharing among DG units.
• A systematic approach to modeling the MGC system is
presented. The constructed small-signal dynamic model
considers the modules of DGs and MGs with the proposed
0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
2
three-layer distributed controllers, lossy-line networks and
loads. All the modules are combined to be a unified model.
• A detailed small-signal stability analysis for the MGC
system is carried out. The relation between the dominant
oscillation modes and key state variables/controllers is
established. General guidelines for tuning control param-
eters are also provided.
The remainder of this paper is organized as follows. Section
II presents the three-layer distributed control framework of
MGC system. Section III elaborates the proposed control
method. Section IV develops the small-signal dynamic model
of MGC system. Numerical results including stability analy-
sis and time-domain simulation are presented in Section V.
Section VI provides the experimental results. Section VII
concludes the paper.
II. T HREE - LAYER D ISTRIBUTED C ONTROL F RAMEWORK
OF MGC
A. Control Framework
Assume that there are M MGs in the MGC system, labeled
as M G1 , . . . , M Gk , . . . , M GM . There are Nk DG units in
M Gk , k = 1, 2, . . . , M . Fig.1 shows the three-layer distributed
control framework of MGC in this paper, which includes the
DG-layer, MG-layer and MGC-layer.
• DG-layer: The DG-layer control is responsible for regu-
lating the local voltage and current of each DG unit in
a decentralized manner, i.e., there are no communications
among DG units in this layer. The droop-based control [1]
has been identified as an effective decentralized control
approach for the DG-layer.
• MG-layer: The MG-layer control is performed for each
MG individually, realizing the coordination of DG units
inside MG in a distributed manner. Each MG has a dis-
tributed sparse communication network (lower communi-
cation network) among DG units to achieve a cooperative
control. The MG-layer control of each MG is usually
used for allocating internal DG units’ power and tracking
the MG voltage and frequency references dictated by the
MGC-layer.
• MGC-layer: In the MGC-layer, each MG is represented
as an MG agent. Each MG agent communicates with
its neighbors to form a distributed upper communication
network. In each MG, at least one DG unit is selected
as the pinned DG unit, which has a bidirectional com-
munication link (pinning link) to the corresponding MG
agent. Through the pinning links, the MG agent collects
the needed information from the pinned DG units and
also sends references to the pinned DG units, realizing
the interaction between MGC-layer and MG-layer. The
MGC-layer control should coordinate each MG to share
active/reactie power among them and also regulate the
system frequency and voltage in a distributed manner.
B. Double-layer Cyber Network Modeling
The lower communication networks in the MG-layer can be
modeled as M directed graphs (digraphs) G1 , . . . , Gk , . . . , GM
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
3

MG1 MGM
MGC-layer Agent Agent
control
(Distributed)
Upper communication MG2 MG3
Agent Agent
network
Pinning link

DG3 DG3 DG3 DG3

MG-layer DG NM
DG1 DG N1 DG1 DG N2 DG1 DG N3 DG1
control
DG2 DG2 DG2 DG2
(Distributed) MG1 MG2 MG3 MGM

Lower communication networks

LCL Rest of
DG-layer filter microgrid
control
(Decentralized) PWM DG
controller

Fig. 1. Three-layer distributed cooperative control framework of MGC

corresponding to each MG. For DGi in M Gk , the set of its
neighbours is denoted as Nik , i = 1, 2, . . . , Nk . The associated
adjacency matrix of Gk is defined as Ak = [akij ]Nk ×Nk , where
akij is the weight of the communication link from DGj to DGi .
The upper communication network among MGs in the
MGC-layer can be modeled as a digraph G. e The set of
k algorithm in which xref is assigned to the pinned nodes to
neighbours of M Gk is denoted as H , k = 1, 2, . . . , M . The
associated adjacency matrix of G e is defined as Ae = [ãkl ]M ×M ,
where ãkl is the weight of the communication link from M Gj
to M Gi .
To model the pinning links between upper and lower cyber
networks, we introduce the leader-adjacency matrix Bpin k
= C OOPERATIVE C ONTROL M ETHOD
k k k k
[g1 , g2 , . . . , gNk ] for each M Gk . The pinning gain gi = 1 if
DGi is the pinned DG unit; otherwise gik = 0.
More details about the graph theory can be referred to [13].
C. Basic Principle of Distributed Cooperative Control Theory
In this paper, the distributed cooperative control theory [14]
is employed to design controllers in the MG-layer and MGC-
layer. Based on the sparse communication network modeled
by a digraph, its basic principle is illustrated as follows.
Assume that each node i has a scalar information state xi .
Node i needs information of its own and that of its neighbours
j (j ∈ Ni ) to update its state, where Ni is the set of its
neighbours. Based on the continuous consensus algorithm, the
update rules can be categorized as synchronization problem
(1) and tracking problem (2), where aij is the element of the
associated adjacency matrix, and gi is the pinning gain.
X active and reactive power droop coefficients, output active and
ẋi (t) = − aij (xi (t) − xj (t)), (1)
j∈Ni
X
ẋi (t) = − aij (xi (t) − xj (t)) − gi (xi (t) − xref ). (2)
j∈Ni
If the communication network has a spanning tree [14], then
(i) for (1), the states xi of all the nodes will synchronize
to a consensus value, and (ii) for (2), xi of all the nodes
will track the reference value xref . In sum, equation (1)
is a synchronization-based algorithm in which the consensus
value is not prescribed, while equation (2) is a tracking-based
realize the reference tracking. It should be noted that xref
could also have dynamics.
III. P INNING - BASED H IERARCHICAL AND D ISTRIBUTED
Based on the three-layer control framework in Section II,
this section presents detailed control strategies for each layer.
A. DG-layer control
The droop-based control is adopted in this layer for each DG
unit, which includes the power controller, voltage controller
and current controller [1]. The power controller is responsible
for obtaining the frequency and voltage reference of DG units
based on the power droop characteristics, given by
∗
ωki = ωn − mki (Pki − Pki ), (3)
∗
|voki | = V0α − nki Qki , (4)
where the subscript “ki” corresponds to DGi in M Gk (i.e.,
∗
DGki ), ωki and voki are the frequency and output voltage refer-
ence of DGki , ωn is the system rated frequency (2π×50rad/s),
∗
V0α is the no-load voltage, and mki , nki , Pki , Qki , Pki are the
reactive power, and rated active power of DGki , respectively.
Then, the voltage and current controllers enable a closed-loop
regulation of DG unit output voltage voki to its reference value
∗
voki .

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
4

The traditional active power droop controller (3) shares the 1) MG-layer frequency controller: The frequency controller
active power among DG units based on the ratios of their in the MG-layer is responsible for realizing objective (MG-i),
capacities. However, an economic operation of the system given by
cannot be guaranteed with this kind of power sharing scheme.
ωki = ωn − ked ηki (Pki ) + Ωki , (11)
Instead of (3), we adopt an optimal active power droop control  
method in this paper, which can share the active power among dΩki X
DG units such that the system overall generation cost is = −cωki  akij (ωki − ωkj ) + gik (ωki − ωrefk )
dt k
minimized. j∈N
Xi
The generation cost function of DGki is denoted as −cP ki akij (ked ηki (Pki ) − ked ηkj (Pkj )), (12)
GCki (Pki ). Thus, the overall generation cost minimization j∈Nik
problem, i.e., economic dispatch (ED) problem, in M Gk can
be formulated as Equation (11) is transformed from (9) with an additional shift
variable Ωki . In (12), Ωki is determined based on a combination
N
Pk
min GCki (Pki ) of tracking problem (2) and synchronization problem (1),
i=1
(5) where cωki and cP ki are positive control gains, and ωrefk is
N
Pk the frequency reference value for M Gk from the MGC-layer.
s.t. Pki = PLoad ωrefk is sent to the pinned DG units through pinning links. The
i=1
dynamic process of (12) results in ωki of all DG units in M Gk
where PLoad is the total load power in M Gk . The Lagrange
converging to ωrefk and the realization of (8).
multiplier method can be used to solve the optimization prob-
Remark 1: The transients of ωrefk of different MGs could
lem (5). The Lagrange function Lag is constructed as
be different to adjust the power flow of MG. However, they
Nk Nk
!
X X will be synchronized to a common value in the steady state by
Lag = GCki (Pki ) + η PLoad − Pki , (6) the MGC-layer controller as illustrated in Section III.C.
i=1 i=1
2) MG-layer voltage controller: The voltage controller in
where η is the Lagrange multiplier. According to the optimality the MG-layer is responsible for realizing objective (MG-ii),
conditions, we have given by
∂Lag ∂GCki (Pki ) ∗
= − η = 0, i = 1, 2, . . . , Nk . (7) vodki = Vn − nki Qki + λki −hki , (13a)
∂Pki ∂Pki
| {z }
Vf ki
∂GCki (Pki ) ∗
∂Pki is the generation cost increment value (GCIV) of voqki = 0, (13b)
DGki , which can denoted as ηki (Pki ). Then from (7), we have ∗ ∗
where vodki and voqki are the d-axis and q-axis components
ηk1 (Pk1 ) = ηk2 (Pk2 ) = . . . = ηkNk (PkNk ). (8) ∗
of voki . Equation (13a) is transformed from (4) with two addi-
Equation (8) is the equal increment principle (EIP) of economic tional shift variables λki and hki . Note that Vn − nki Qki + λki
∗
dispatch problem to minimize the overall generation cost. is denoted as an intermediate voltage variable Vf ki , and voqki
Thereafter, the optimal active power droop equation can be is set to zero.
designed as λki is determined based on the tracking problem (2), given
ωki = ωn − ked ηki (Pki ), (9) by
 
where ked is a positive scalar coefficient. In the steady state, dλki X
ωki of all the DG units must be equal [26], and thus ηki (Pki ) = −cvki  akij (Vf ki − Vf kj ) + gik (Vf ki − Vf refk ) ,
dt k j∈Ni
are also equal from (9), i.e., equation (8) is satisfied. Therefore,
(14)
the system overall generation cost can be minimized with (9).
where cvki is a positive control gain. Vf refk is the intermediate
In sum, (9) and (4) are the active/reactive power droop
voltage reference for M Gk from the MGC-layer, which is sent
controllers used in this paper for the DG-layer control.
to the pinned DG units through pinning links. The dynamic
B. MG-layer control process of (14) will result in Vf ki of all DG units in M Gk
converging to Vf refk .
For M Gk , the objectives of the MG-layer control are
hki is determined based on the synchronization problem (1),
• (MG-i): enabling each DG unit’s frequency track the refer-
given by
ence value ωrefk from the MGC-layer, while maintaining
the EIP (8) achieved in the DG-layer control; dhki X
= cQki akij (nki Qki − nkj Qkj ), (15)
• (MG-ii): enabling each DG unit’s intermediate voltage dt k j∈Ni
Vf k (defined later) track the reference value Vf refk from
the MGC-layer, while realizing an accurate reactive power where cQki is a positive control gain. The dynamic process of
sharing among DG units based on their capacities, i.e., (15) will impose nk1 Qk1 = nk2 Qk2 = . . . = nkNk QkNk in
the steady state.
nk1 Qk1 = nk2 Qk2 = . . . = nkNk QkNk . (10) Remark 2: Since the consensus regulation of DG output
Based on the lower communication network Gk , the frequen- voltages and an accurate reactive power sharing is contradictory
cy and voltage controllers of M Gk can be designed according [15], this paper preferentially chooses to realize an accurate
to the distributed cooperative control theory. reactive power sharing. Therefore, in (14), the intermediate

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
5

voltage Vf ki , rather than the actual output voltage, is regulated 2) MGC-layer voltage controller: The voltage controller in
to the consensus value Vf refk . the MGC-layer is responsible for realizing objective (MGC-ii),
given by
∗
C. MGC-layer control VPCCk = V0β + Tk −Hk , (20)
| {z }
Vsk
The MGC-layer control determines ωrefk and Vf refk for each
MG. The objectives of this layer are ∗
where VPCCk is the reference value for the PCC voltage VPCCk
• (MGC-i): restoring the system frequency to the predefined of M Gk , V0β is a voltage constant (normally set around rated
∗
value ωsys while ensuring an equal generation cost incre- voltage), Tk and Hk are two regulation variables. Note that
ment value (GCIV) among pinned DG units of all MGs; Vn + Tk is denoted as Vsk .
• (MGC-ii): restoring the system critical bus voltage Vcri Tk is determined by
∗
to the predefined value Vcri , providing the voltage control
capability for the PCC voltage of each MG, and realizing
 
an accurate reactive power sharing among pinned DG dTk X
= −cvk  ãkl (Vsk − Vsl ) + g̃k (Vsk − Vsref )(21)
units of all MGs. dt k l∈H
∗ ∗
Without loss of generality, and ωsys Vcri
are chosen as the rated
frequency and voltage values in this paper.
∗
Based on the upper communication network G e among MG Vsref = Vn + kPS (Vcri − Vcri )
Z
agents, the MGC-layer controllers can also be designed based ∗
+kIS (Vcri − Vcri )dt, (22)
on the distributed cooperative control theory.
To facilitate the controller design, we firstly define two
new variables ηk,pin and Qk,pin . From Fig.1, each MG agent In (21), cvk is a positive control gain. Equation (21) is a
collects information from the pinned DG units through pinning tracking problem, which results in Vsk of all MGs converging
k to Vsref . From (22), Vsref is obtained from a PI controller
 Denote F as the set of pinned DG units in M Gk , and
links.
F k  is the cardinality of F k . Then, ηk,pin and Qk,pin are such that the system critical bus voltage Vcri can recover to its
∗
defined as reference Vcri . kPS and kIS are the proportional and integral
gains of the PI controller.
1 X
ηk,pin = k ηki (Pki ), k = 1, 2, . . . , M, (16) Hk is determined by
|F |
i∈F k
1 X dHk X
Qk,pin = k nki Qki , k = 1, 2, . . . , M. (17) = cQk ãkl (Qk,pin − Ql,pin ), (23)
|F | dt k
k i∈F l∈H

From (16) and (17), ηk,pin and Qk,pin are the average values In (23), cQk is a positive control gain. Equation (23) is a
of ηki (Pki ) and nki (Qki ) of all the pinned DG units in M Gk , synchronization problem, which results in Qk,pin of all MGs
respectively. being equal with each other.
Moreover, at least one MG agent needs to be pinned to Finally, Vf refk is obtained from a PI controller, which can
∗ ∗
receive the system reference value, e.g., ωsys . The pinning gain make VPCCk track its reference VPCCk , i.e.,
g̃k = 1 if M Gk agent is pinned; otherwise, g̃k = 0.
∗
1) MGC-layer frequency controller: The frequency con- Vf refk = V0γ + kP (VPCCk − VPCCk )
troller in the MGC-layer is responsible for realizing objective
Z
∗
(MGC-i), given by +kI (VPCCk − VPCCk )dt, (24)

ωrefk = ωn + Φk , (18) where kP and kI are the proportional and integral gains of
the PI controller, and V0γ is a voltage constant (normally set
  around rated voltage).
dΦk X
∗  Remark 3: For each MG, without loss of generality, the PCC
= −cωk  ãkl (ωrefk − ωrefl ) + g̃k (ωrefk − ωsys )
dt k voltage VPCCk is selected to be controlled to its reference.
l∈H
X One can also choose another node voltage in the MG to be
−cP k ãkl (ked ηk,pin − ked ηl,pin ). (19)
controlled.
l∈Hk
Remark 4: Since (i) ηk,pin of all MGs are equal with each
From (18), ωrefk is the sum of the rated frequency ωn and other, and (ii) the MG-layer control can ensure (8), it can be
a regulation variable Φk . From (19), Φk is determined based deduced that the GCIVs of all pinned DG units are equal and
on a combination of tracking problem (2) and synchronization furthermore the GCIVs of all DG units in the MGC are equal.
problem (1), where cωk and cP k are positive control gains. Therefore, an economic operation of the whole MGC system
The dynamic process of (19) will result in ωrefk of all MGs can be realized. Similarly, it can also be deduced that nki Qki
∗
converging to ωsys and ηk,pin of all MGs being equal with each of all DG units in the MGC are equal, which results in an
other. accurate reactive power sharing in the whole system.

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
6

ηki(Pki)
ω *
sys
nkiQki
... DGki DG-layer control
MG1 ... MGk DGk1
Vcri* ωrefk Ωki ωki voltage &
MGC-layer MG-layer Droop current
Upper cyber
network Controller
Lower cyber
network
Controller λki Controller controller
(11)-(15) *
(18)-(24)
Vfrefk hki
(4),(9) voki
MG2 G MGM DGk2 G
...k DGknk PWM
...

PCCk DGki
Vsys VPCCk ioki voki
Line Lc Lf
Line
CBk Cf
(closed) The rest Local
The rest of Elecrical network
of MGk load
the MGC of MGk

System critical bus

Fig. 2. A block diagram of the pinning-based hierarchical and distributed cooperative control for MGC

D. Summary tributed cooperative control. A general modeling procedure in

Fig.2 shows a block diagram of the proposed pinning-based
hierarchical and distributed cooperative control for MGC. Note
that for simplicity, only the most important control details
are presented in Fig.2. For the physical network, DGki is
connected to the PCC bus of M Gk (PCCk ) through a LCL
filter, local load bus and feeder line. Then, M Gk is connected
to the system critical bus and the rest of MGC through a
circuit breaker (CBk ) and feeder line. The voltages of the
system critical bus and PCCk are measured and fed back to
the MGC-layer controller. The general control implementation
procedures in Fig.2 are summarized in the following.
• MGC-layer control: The system frequency and critical bus
∗
voltage reference ωsys and Vcri∗
are sent to the pinned
MG agents. Then, the MGC-layer controllers (18)-(24) are
executed based on the information exchange among MG
agents in the upper communication network G. e Finally,
the frequency and voltage references ωrefk and Vf refk for
M Gk are obtained.
• MG-layer control: ωrefk and Vf refk are sent to the pinned
DG units in M Gk . The pinned DG units also send their
information of ηki (Pki ) and nki Qki to the corresponding
MG agent. Then, the MG-layer controllers (11)-(15) are
executed based on the information exchange among DG
units in the lower communication work Gk . Finally, the
regulation variables Ωki , λki and hki are obtained.
• DG-layer control: Ωki , λki and hki are sent to the droop
controller (4) and (9). Then, the output frequency and
∗
voltage references ωki and voki for DGki are obtained
and then sent to the voltage and current controller. Finally,
the PWM signals are generated to control the inverter.
IV. S MALL -S IGNAL DYNAMIC M ODELING OF MGC
This section presents the small-signal dynamic model of
MGC system with the proposed pinning-based three-layer dis-
this section is (i) modeling individual DG unit, (ii) combining
DG unit models to obtain each individual MG model, (iii)
combining the models of all MGs, (iv) modeling the MGC-
layer controller, (v) considering the correlation between MG-
layer and MGC-layer and then combining all the models to
obtain the complete MGC system model. Detailed step-by-
step modeling process and the details of parameter matrices
are provided in Appendix B.
A. Individual MG model
Without loss of generality, the local dq reference frame of
DG1 in M G1 , i.e., DG11 , is selected as the global common
DQ-frame. Thus, we have ωg = ω11 , where ωg is the
frequency of the global DQ-frame. Then, δki is defined as
the angle between the local dq-frame of DGki and the global
DQ-frame, i.e.,
δ̇ki = ωki − ωg . (25)
This paper mainly focuses on the dynamics of power droop
controller, MG-layer controller and MGC-layer controller,
while the fast dynamics of the voltage/current controllers and
LC filters are not the focus of this paper. Thus, they are omitted
in the model by assuming
∗ ∗
vodki = vodki , voqki = voqki , (26)
where vodki and voqki are d-axis and q-axis components of the
output voltage voki of DGki , Fig.2.
Then, in light of (26), the small-signal dynamic model of
DGki can be obtained by linearizing (11)-(15) and rearranging
the linearized results, i.e.,
∆Ẋinvki = Ainvki ∆Xinvki + Binvki ∆ioDQki + Cinvki ∆ωg
X
+ Finvkij ∆Xinvkj + Hinvki ∆Vf refk
j∈Nik
+Rinvki ∆ωrefk , (27)

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
7

where Ainvki , Binvki , Cinvki , Finvkij , Hinvki and Rinvki are the where YM = (YA − YB YD−1 YC )−1 . Then, from (31) and (33),
parameter matrices and their details are provided in Appendix we have
B. Finvkij represents the correlations among DG units in M Gk
due to their information exchange in the lower communication ∆ioDQ = WINV ∆XINV , WINV = YM LINV . (34)
network. ioDQki is the deviation of the DGki output current
ioki , Fig.2, in the global DQ-frame. There are 6 state variables To describe Vf refk in (29), we focus on the linearized model
of DGki , given by of (24) in the following. Denote ψk as the state variable of (24),
T
i.e.,
∆Xinvki = [∆δki , ∆Pki , ∆Qki , ∆Ωki , ∆λki , ∆hki ] , (28)
ψ˙k = V ∗ − VPCCk ,
PCCk (35)
where the superscript “T ” represents the transpose of a matrix. q
Combination of all the individual DG unit models (28) in 2
where VPCCk = VPCCDk 2
+ VPCCQk . By linearizing (24) and
M Gk provides the small-signal dynamic model of M Gk , given (35), we have
by
∆Vf refk = −kP APCCk ∆VPCCDQk + kI ∆ψk
∆ẊINVk = AINVk ∆XINVk + BINVk ∆ioDQk + CINVk ∆ωg
∗
+kP ∆VPCCk , (36)
+HINVk ∆Vf refk + RINVk ∆ωrefk , (29)
∗
∆ψ̇k = −APCCk ∆VPCCDQk + ∆VPCCk , (37)
T T T
where ∆XINVk = [∆Xinvk1 , ∆Xinvk2 , . . . , ∆XinvkN k
]T
and ∆ioDQk = [∆ioDQk1 , ∆ioDQk2 , . . . , ∆ioDQkNk ]T .
T T T
where ∆VPCCDQk = [∆VPCCDk , ∆VPCCQk ]T , and APCCk is
AINVk , BINVk , CINVk , HINVk and RINVk are parameter the parameter matrix. Since ∆VPCCDQk is part of ∆vN LDQ ,
matrices. based on (32), it can be expressed in terms of ∆ioDQ as
In addition, by linearizing (13) and (26) and then combining
the linearized results of all DG units in M Gk , we have ∆VPCCDQk = ZPCCk ∆ioDQ , (38)
∆voDQk = LINVk ∆XINVk , (30) where ZPCCk can be obtained from Yaug .
where ∆voDQk is the deviation of voki of all the DG units in To describe ∆ωg in (29), we linearize (11) for DG11 . Thus,
M Gk , and LINVk is the parameter matrix. ∆ωg can be expressed as

∆ωg = DINV ∆XINV , (39)

B. Combination of all individual MG’s models
By combining (30) of all the MGs, we have where DINV is the parameter matrix.
Then, the combination model of all the individual MGs
∆voDQ = LINV ∆XINV , (31) can be formulated by (i) substituting (36),(38),(39) into the
individual MG model (29), (ii) combining the models of all the
where LINV is the parameter matrix. ∆voDQ is the com-
individual MGs, and (iii) substituting (34) into the combined
bination of ∆voDQk of all the MGs, i.e., ∆voDQ =
T T T results, i.e.,
[∆voDQ1 , ∆voDQ2 , . . . , ∆voDQM ]T . ∆XINV is the com-
bination of ∆XINVk of all the MGs, i.e., ∆XINV =
T T T ∆ẊINV = JINV ∆XINV + kI HINV ∆ψ
[∆XINV1 , ∆XINV2 , . . . , ∆XINVM ]T .
∗
For the network and load models, we consider the steady- +kP HINV ∆VPCC + RINV ∆ωref , (40)
state algebraic equations of the series RL lossy feeder lines and
RL-type constant impedance loads. Therefore, their models are where
described by the node voltage equations I = Y U with Y being
the node admittance matrix. The linearized form of I = Y U JINV = AINV + CINV DINV + BINV WINV
in this paper is given by −kP NINV WINV . (41)
    
∆ioDQ YA YB ∆voDQ
= , (32) In (40) and (41), AINV , BINV , CINV , HINV , RINV and NINV
0 YC YD ∆vN LDQ
| {z } are parameter matrices, ∆ψ = [∆ψ1 , ∆ψ2 , . . . , ∆ψM ]T ,
∗ ∗ ∗ ∗
:=Yaug ∆VPCC = [∆VPCC1 , ∆VPCC2 , . . . , ∆VPCCM ]T , and ∆ωref =
T
[∆ωref1 , ∆ωref2 , . . . , ∆ωrefM ] .
where Yaug can be obtained from Y , ∆ioDQ is the combination
of ∆ioDQk of all the MGs, and ∆vN LDQ is the deviation of all In addition, the state-space equation of ∆ψ can be for-
the voltages of the network and load nodes. Note that in (32) mulated by (i) substituting (38) into (37), (ii) combining the
the injected currents corresponding to the network and load substituted results of all MGs, and (iii) substituting (34) into
nodes are zero because the constant impedance type of loads the combined results, i.e.,
is considered. A contractive form of (32) can be obtained by
reducing the zero components of the injected currents, i.e., ∆ψ˙ = −TPCC WINV ∆XINV + ∆VPCC
∗
, (42)

∆ioDQ = YM ∆voDQ , (33) where TPCC is the parameter matrix.

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
8

MG1 10
1 Lc PCC1
DG11 Line4 19

Load1 Load2 System critical bus

23
2 Lc 11
DG12 MG3
Line5 PCC3
16
CB1 Line1 Lc 7
12
21 Line10 DG31
Load3
Line3
3 Lc Load6
Line6 Load7
DG13 CB3
17 Lc 8
DG32
Load9 Line11
18
MG2 PCC2 Load8
4 Lc 13 Line7 20
DG21 Load11 9
Line12 DG33
Lc
5 Lc 14 Line8 Line2
DG22
Load4 CB2
6 Lc 15 CB Grid
DG23
Load10 22

Load5 Line9 0.38kV/10kV

Fig. 3. A schematic diagram of the studied MGC system

q
C. Modeling of MGC-layer controller 2
where Vcri = VcriD 2 . By linearizing (22) and (46), we
+ VcriQ
From (20), we have have
∆Vsref = −kPS Acri ∆VcriDQ + kIS ∆ϕ, (47)
Vsk = V0β + Tk . (43)
∆ϕ̇ = −Acri ∆VcriDQ , (48)
For the MGC-layer controller of M Gk agent, linearizing (18), where ∆VcriDQ = [∆VcriD , ∆VcriQ ]T , and Acri is the parame-
(19), (21), (23) and (43) and rearranging the linearized results ter matrix. Since ∆VcriDQ is part of ∆vN LDQ , based on (32),
yield it can be expressed as
X
∆ṠMGk = AMGCk ∆SMGk + FMGCkl ∆SMGl ∆VcriDQ = Zcri ∆ioDQ , (49)
l∈Hk
where Zcri can be obtained from Yaug .
+HMGCk ∆Vsref + EMGCk ∆XINV (44)
Then, substituting (34), (47) and (49) into (45) yields
where ∆SMGk = [∆Φk , ∆Tk , ∆Hk ]T , AMGCk , FMGCkl , ∆ṠMG = AMGC ∆SMG + JMGC ∆XINV
HMGCk and EMGCk are parameter matrices. FMGCkl repre- +kIS HMGC ∆ϕ, (50)
sents the correlations among MG agents due to their informa-
tion exchange in the upper communication network. EMGCk where
represents the correlations from pinned DG units to MG agents JMGC = EMGC − kPS HMGC Acri Zcri WINV . (51)
by the pinning links.
By combining (44) of all the MG agents, we have Moreover, substituting (34) and (49) into (48) yields
∆ϕ̇ = −Acri Zcri WINV ∆XINV . (52)
∆ṠMG = AMGC ∆SMG + HMGC ∆Vsref
+EMGC ∆XINV (45) D. Complete MGC system model
T T T To further formulate the models of (40) and (42), we need
where ∆SMG = [∆SMG1 , ∆SMG2 , . . . , ∆SMGM ]T , AMGC , ∗
to express ∆VPCC and ∆ωref in terms of ∆SMG , which can
HMGC and EMGC are parameter matrices.
be done by linearizing (18) and (20) and then combining the
To describe ∆Vsref in (45), we focus on the linearized model
linearized results of all MGs, given by
of (22) in the following. Denote ϕ as the state variable of (22),
i.e., ∆ωref = Acou ∆SMG , (53)
∗ ∗
ϕ̇ = Vcri − Vcri , (46) ∆VPCC = Bcou ∆SMG , (54)

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
9

DG13 TABLE II
G1 G ENERATION COST PARAMETERS

MG1 αki βki γki

DG12 DG11 Agent (* , V * )
sys cri DG11 DG12 DG13 (MG1 ) 0.094 1.22 51
DG33
Pinning G3
G
DG21 DG22 DG23 (MG2 ) 0.078 3.41 31
link DG31 DG32 DG33 (MG3 ) 0.105 2.53 78
DG23
G2 MG2 MG3 DG31 DG32
Agent Agent
TABLE III
DG22 DG21 PARAMETERS OF DG- LAYER CONTROL

nki Pmax Qmax

Parameters ked
Fig. 4. Double-layer distributed communication network of the studied MGC (kV/kVar×10−3 ) (kW) (kVar)
system
DG11 DG12 DG13 0.1 0.78 60 20
DG21 DG22 DG23 0.1 0.52 50 30
TABLE I
DG31 DG32 DG33 0.1 0.78 60 20
E LECTRICAL PARAMETERS OF THE MGC SYSTEM

Type Electrical parameters

TABLE IV
Line1=0.008Ω+0.03mH, Line2=0.005Ω+0.02mH, PARAMETERS OF MG- LAYER AND MGC- LAYER CONTROL
Line Line3=0.014Ω+0.07mH, Line4,8,10,11=0.15Ω+0.15mH,
Line5,9=0.11Ω+0.35mH, Line6,7,12=0.11Ω+0.2mH Parameters MG-layer MGC-layer
Load1,5,8=15kW+7.5kVar, Load2=40kW+15kVar, cωki /cωk 560 200
Load3,6=12kW+5kVar, Load4,7=50kW+20kVar, cP ki /cP k 95 35
Load
Load9=75kW+22.5kVar, Load10=2kW+5kVar, cvki /cvk 20 5
Load11=20kW+10kVar cQki /cQk 60 13
kP / 25
kI / 200
where Acou and Bcou are parameter matrices. kPS / 5
Finally, the complete MGC system model can be formulated kIS / 67
by (i) substituting (53) and (54) into (40) and (42), and (ii)
combining the substituted results and (50) and (52), i.e.,
∆Ẋsys = Asys ∆Xsys , (55) Fig.4 shows the upper and lower communication networks
of the studied MGC system, where the arrowed and dotted
T T T
where ∆Xsys = [∆XINV , ∆ψ , ∆SMG , ∆ϕ]T , and Asys is the lines denote the communication links. M G1 is the pinned
∗ ∗
state matrix of the MGC system. MG agent which receives the references of ωsys and Vcri .
The system small-signal stability can be evaluated based on DG11 , DG21 and DG31 are the pinned DG units of each MG.
analyzing the eigenvalues and eigen-structure of Asys , which The corresponding associated adjacency matrix and leader-
will be presented in Section V.B. adjacency matrix parameters are provided in Appendix A.
The generation cost function GCki (Pki ) in this paper is
V. N UMERICAL S TUDY described by the quadratic function [31], i.e.,
Based on a test MGC system, this section provides the 2
GCki (Pki ) = αki Pki + βki Pki + γki , (56)
system stability analysis (Section V.B) and time-domain sim-
ulation results in the PSCAD/EMTDC platform (Section V.C) where αki , βki and γki are cost parameters. Table II shows the
to validate the effectiveness of the proposed methods. cost parameters of each DG unit [31].
Table III and IV provide the parameters of the DG-layer,
A. Study System MG-layer, and MGC-layer controls, where Pmax and Qmax
are the active/reactive power capacities of DG units.
Fig.3 shows a schematic diagram of the studied 23-bus three-
phase MGC system, which is a 0.38kV and 50Hz system
including 3 MGs, 9 DG units, 12 feeder lines and 11 loads. B. Eigen-analysis Results
The MGC is connected to the utility grid through a circuit 1) Eigenvalue Spectrum: Fig.5 compares the low-frequency
breaker (CB) and 10kV/0.38kV transformer. The CB is open, eigenvalue spectrum of the MGC system with and without
i.e, the system is operated in the islanded mode. CB1 − CB3 the MGC-layer control in service. In Case 1 (cyan squares),
are the circuit breakers of each MG. They are closed. Bus 23 is there are only DG&MG-layer controls for the MGC system.
the system critical bus, which is the common bus connecting In Case 2 (green stars), the MGC-layer control is added. Fig.5
to all MGs. Buses 19, 20 and 21 are the PCC bus of each indicates that the MGC-layer control (i) significantly changes
MG. Each feeder line is modeled by a series RL branch and the shaping of eigenvalues on the complex plane, and (ii)
each load is represented by a series RL impedance. Lc is introduces some new oscillation modes to the system. The
the coupling inductance. The system electrical parameters are dominant oscillation modes of Case 2 are labeled as modes
shown in Table I. 1-7 in Fig.5. Table V shows their corresponding eigenvalues

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
10

0.6 Mode 1 1.4

DG11 Mode 5
0.5 DG12 1.2
DG13 1
0.4
DG21
0.3 DG22 0.8
DG23 0.6
0.2 DG31
DG32 0.4
0.1 DG33 0.2
0 MG1 0
MG2
Δδki ΔP MG3 Δδki ΔP
ki ΔQ ki ΔQ
kiΔΩ
ki Δλki
Δhki Δψk Δφ ki ΔΩ
kiΔλ
ki Δhki
ΔΦk ΔTk Δψk ΔΦ
ΔHk Δφ k ΔT
k ΔH
k Δφ

0.7 Mode 3 DG11 0.8 Mode 7

0.6 DG12
DG13 0.6
0.5 DG21
0.4 DG22
DG23 0.4
0.3
DG31
0.2 DG32 0.2
0.1
DG33
MG1
0 MG2 0
Δδki ΔP MG3 Δδki ΔP
ki ΔQ
kiΔΩ Δφ ki ΔQ
kiΔΩ
ki Δλki ki Δλki
Δhki Δψk Δhki Δψk
ΔΦk ΔTk ΔΦk ΔTk
ΔHk Δφ ΔHk Δφ

Fig. 6. Participation factors of (a) mode 1, (b) mode 3, (c) mode 5, and (d) mode 7, which describe the association degrees between state variables and
modes. The variables in the horizontal axis correspond to the state variables. States ∆δki , ∆Pki , ∆Qki , ∆Ωki , ∆λki , ∆hki , which belong to DG units
(DG11 –DG33 ), mainly correspond to the DG-layer and MG-layer controls. States ∆ψk , ∆Φk , ∆Tk , ∆Hk , which belong to MG agents (M G1 –M G3 ), and
∆ϕ mainly correspond to the MGC-layer control. The vertical axis shows the participation factor value.

TABLE V
D OMINANT OSCILLATORY MODES WITH MGC- LAYER CONTROL

Case1: DG&MG-layer controls eigenvalue damping

Case2: DG&MG&MGC-layer controls mode 1 -27.17±j105.20 0.250
150 mode 2
mode 2 -24.42±j106.51 0.223
mode 1
mode 3 -5.33±j40.08 0.132
100 mode 5 mode 4 -4.57±j37.45 0.121
Imaginary (rad/s)

mode 5 -46.80±j65.43 0.582

mode 3 mode 4 mode 6 -41.37±j60.82 0.562
50 mode 6
mode 7 mode 7 -4.53±j8.51 0.470
0

-50 and dampings. The frequency range of modes 1-7 varies from
8.51rad/s to 106.55rad/s, which indicates a “wide-frequency-
-100 band oscillation”. In addition, modes 1-4 are with relatively
-150 low dampings, which may lead to more oscillatory responses
-60 -50 -40 -30 -20 -10 0 10 compared with Case 1.
Real (1/s) 2) Participation Factor Analysis: Participation factor is the
multiplication of the corresponding element in the right and left
Fig. 5. Low-frequency eigenvalue spectrum of the studied MGC system with eigenvectors of the state matrix. It can be used for evaluating
and without the MGC-layer control the association degree between state variables and modes. In
this paper, we perform the participation factor analysis based
on the state matrix Asys in (55).

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
11

150 150 • Modes 1 and 3 are “interaction modes” because the

modes 1,2
100 100 state variables of both DG&MG-layer and MGC-layer
Imaginary (rad/s)

Imaginary (rad/s)
contribute to modes 1 and 3. From Table V, modes 1
50 50
and 3 are with relatively low dampings. Therefore, it can
0 0 be deduced that the control layer interaction may result
-50 -50 in low-damping modes.
-100 -100
• Mode 5 is a kind of “DG&MG-layer mode” since the
-150
state variables of DG&MG-layer mainly contribute to it.
-60 -40 -20 0 20 -150 • Mode 7 is a kind of “MGC-layer mode” since the state
-60 -40 -20 0 20
Real (1/s) Real (1/s)
(a) cωk varies from 30 to 200 (b) cPk varies from 35 to 300
variables of MGC-layer mainly contribute to it. From
Table V, the oscillation frequency of mode 7 is the lowest.
120 200
This is because the response speed of the MGC-layer
Imaginary (rad/s)

80 control should be slow enough to ensure a time-scale

Imaginary (rad/s)

100
40 separation between MGC-layer and MG-layer.
mode 7
• States ∆Qki , ∆hki and ∆Hk , which are associated with
0 0
the reactive power controllers (15) and (23), are generally
-40 the states with the highest contribution to modes 1,3,5,7.
-100
-80 Therefore, controllers (15) and (23) are the key controllers
-120 -200 which affect the system dynamics most significantly.
-60 -40 -20 0 -60 -40 -20 0 20
Real (1/s) Real (1/s) 3) Sensitivity Analysis–Impact of MGC-layer Control Pa-
(c) cvk varies from 5 to 40 (d) cQk varies from 5 to 40 rameters on System Stability: Fig.7 shows the traces of the
dominant oscillatory modes 1-7 as a function of the MGC-
Fig. 7. Traces of dominant oscillatory modes when (a) cωk varies from 30
to 200, (b) cP k varies from 35 to 300, (c) cvk varies from 5 to 40, and layer control parameters cωk , cP k , cvk and cQk , where the
(d) cQk varies from 5 to 40. Blue circles denote eigenvalues with beginning blue circles denote eigenvalues with beginning parameters, and
parameters, and purple diamonds denote eigenvalues with ending parameters. purple diamonds denote eigenvalues with ending parameters.
Corresponding to Fig.7, Table VI summarizes the variation
TABLE VI trends of the mainly affected modes when increasing the MGC-
VARIATION TRENDS OF DOMINANT OSCILLATORY MODES WHEN
INCREASING MGC- LAYER CONTROL PARAMETERS
layer control parameters. Note that (i) the results of increasing
kP and kPS are shown in Table VI while not in Fig.7 for
Parameter
Mainly Modes variation simplicity, and (ii) parameters kI and kIS have little impact on
affected modes when increasing parameters modes 1-7 and thus their results are not presented. Fig.7 and
cωk modes 5,6 move towards stable region Table VI indicate that
cP k modes 1,2 move towards unstable region
move towards stable region first,
• The system may lose stability with variations of con-
cvk mode 7 trol parameters, i.e., there are eigenvalues located in the
then splits at the real axis
modes 1,2 move towards unstable region right half plane with improper parameters. Therefore, the
cQk control parameters should be carefully selected to ensure
modes 3,4 move towards stable region
kPS mode 7 move towards stable region sufficient stability.
modes 1,2 move towards unstable region • The system dynamics are relatively complex. One param-
kP
modes 3,4 move towards stable region eter may affect different modes, and one mode may be
affected by different parameters, e.g., modes 1 and 2 are
affected by cP k , cQk and kP .
Fig.6 shows the participation factor analysis results for • In particular, Fig.7(d) indicates that increasing cQk has
modes 1, 3, 5 and 7. Note that the results of modes an opposite effect on the variation trends of modes 1,2
2, 4, 6 are similar to those of modes 1, 3, 5 and thus and modes 3,4. Therefore, the selection of cQk should
not presented for simplicity. The variables in the horizon- consider a tradeoff between the dynamic performance of
tal axis of Fig.6 correspond to the state variables. States modes 1,2 and modes 3,4. In addition, cQk is the control
∆δki , ∆Pki , ∆Qki , ∆Ωki , ∆λki , ∆hki , which belong to DG parameter of controller (23), which is a key controller
units (DG11 –DG33 ), mainly correspond to the DG-layer and from the previous analysis. Hence, the selection of cQk is
MG-layer controls. States ∆ψk , ∆Φk , ∆Tk , ∆Hk , which be- of vital importance.
long to MG agents (M G1 –M G3 ), and ∆ϕ mainly correspond In sum, the system dynamics is relatively complicated due
to the MGC-layer control. The vertical axis shows the par- to multiple control layers and control parameters. The above
ticipation factor value. A higher value indicates a stronger analysis can facilitate a better understanding of the system
association between the state variable and the analyzed mode. dynamics and stability characteristics, and also provide a
Moreover, to better understand Fig.6, we take ∆δki as an general guideline for the control parameter tuning. The control
example. For the position of ∆δki in Fig.6, the participation parameter selection can also be formulated as an optimization
factor values of ∆δki of different DG units are stacked together. problem to achieve largest stability margin and best dynamic
The results of Fig.6 are summarized as follows. performance by optimizing control parameters. This research

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
12

DG&MG-layer DG&MG&MGC-layer TABLE VII

DG-layer control controls controls E LECTRICAL PARAMETERS OF EXPERIMENTAL MGC SYSTEM
50.4
Type Electrical parameters
Frequency (Hz)

50 Filter inductance: 1.8mH

Inverter
Filter capacitance: 27µF
49.6
Line1=1.8mH, Line2=1.8mH
49.2 Line Line3=1.8mH, Line4=1.8mH
Line5=1.9Ω+2.5mH, Line6=1.6Ω+2.1mH
48.8
2 8 14 20 26 Load1=92Ω, Load2=153.3Ω,
Load
Time(s) Load3=39.8Ω+j32.6Ω, Load4=120.2Ω+j37.1Ω
(a) System frequency (Hz)
1.005
t=17s, ωrefk is regulated by (18) and (19), and the system
Voltage (p.u.)

1
frequency is maintained at 50Hz.
0.995 • Fig.8(b) indicates that the system critical bus voltage

0.99
exhibits deviations during the DG-layer and MG-layer
controls, while it can be restored to the rated value 1p.u.
0.985 by the MGC-layer controller.
2 8 14 20 26
Time(s) • Fig.8(c) indicates that during the stage of DG-layer con-
(b) System critical bus voltage Vcri (p.u.) trol, the GCIVs (ηki ) of all DG units are equal due
10.5 to the regulating characteristic of droop equation (9).
GCIV ηki ($/kW)

After the MG-layer control is activated, the neighbour

9.5
communications in the lower communication networks
8.5 result in a consensus of GCIVs for each individual MG,
DG21 DG31
while the consensus values for different MGs could be
7.5 DG11
DG12 DG22 DG32 different. After the MGC-layer control is employed, it can
DG13 DG23 DG33
6.5 realize a coordination among different MGs with the help
2 8 14 20 26
Time(s)
of pinning links and the upper communication network,
(c) DG generation cost increment value ηki and thus the GCIVs of all DG units become equal again,
indicating an economic operation of the whole system.
25
• Fig.8(d) indicates that under the DG-layer control the
Reactive power

20 reactive power sharing is not accurate. After t=8s, the

(kVar)

15 MG-layer control can impose an accurate reactive power

10
DG11
5 DG12
DG13
0
2 8 14
Time(s)
(d) DG output reactive powers Qki (kVar)
Fig. 8. Simulation results of (a) system frequency, (b) system critical bus
voltage Vcri , (c) DG generation cost increment value ηki , and (d) DG output
reactive powers Qki . The system is initially operated with the DG-layer
control. At t=8s the MG-layer control is activated, and at t=17s the MGC-
layer control is employed.
will be included in our future work.
C. Time-domain Simulation Results
The MGC system is initially operated with the DG-layer
control. At t=8s, the MG-layer control is activated. At t=17s,
the MGC-layer control is employed. Fig.8 shows the corre-
sponding simulation results.
• Fig.8(a) indicates that the DG-layer control results in a
frequency deviation, while the system frequency is re-
stored to the rated value 50Hz after applying the MG-layer
control. This is because in our study ωrefk of each MG is
set to be the rated frequency by default if without MGC-
layer control. After the MGC-layer control is employed at
sharing inside each individual MG. After t=17s, an accu-
DG21 DG31 rate reactive power sharing for the whole MGC system is
DG22 DG32
DG23 DG33 realized by the MGC-layer control. Note that the reactive
20 26 power is shared based on nki shown in Table III.
In sum, the control objectives can be well realized with the
proposed pinning-based hierarchical and distributed coopera-
tive control method.
VI. E XPERIMENTAL R ESULTS
This section presents experimental results to validate the
practical implementation feasibility of the proposed control
method. Three experimental studies are carried out. Study 1
validates the basic control performance. Study 2 and 3 evaluate
the control performance under conditions of load changes and
communication link failures, respectively.
A. Experimental Setup
Fig.9 shows the hardware setup for the experimental MGC
system in the laboratory. The experimental setup includes a
dSPACE DS1006 platform, four Danfoss inverters, RL branch
lines, resistive and inductive loads, switches to control the
relays and a control desk. The voltage and current are measured
through LEM measurement boards and then processed by
ADC boards. The signals after ADC boards are fed back to

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
13

TABLE IX
PARAMETERS OF DG- LAYER CONTROL FOR EXPERIMENTAL MGC SYSTEM
dSPACE
DS1006
Danfoss nki Pmax Qmax
Parameters ked
Inverters (V/Var×10−3 ) (kW) (kVar)
Resistive
Load DG11 2 6.479 1.8 1.2
Inductive Resistive DG12 2 8.639 1.35 0.9
Load Load DG21 2 6.479 1.8 1.2
DG22 2 8.639 1.35 0.9

TABLE X
Control Lines PARAMETERS OF MG- LAYER AND MGC- LAYER CONTROLS FOR
Desk EXPERIMENTAL MGC SYSTEM

Switch Parameters MG-layer MGC-layer

cωki /cωk 400 80
cP ki /cP k 400 80
cvki /cvk 150 30
Fig. 9. Experimental setup in the laboratory.
cQki /cQk 20 2
kP / 1.2
MG1 System critical bus kI / 42
PCC1
7 kPS / 0.3
1 Line1 5
PCC2 MG2
DG11 6
kIS / 10
Line3 3
DG21
Load1 Line5
CB1 Line6
2 Load2
measurements. The communication between dSPACE and PC
CB2
DG12 4 is based on Ethernet cables.
Line2
DG22 Fig.10 shows the physical configuration of the experimental
Line4
Load3 MGC system, which includes 2 MGs, 4 DG units, 6 feeder
Load4 lines and 4 loads. The system rated frequency is 50Hz and
rated line-to-line RMS voltage is 215V. The circuit breakers
Fig. 10. Physical configuration of the experimental MGC system. CB1 and CB2 are closed. Bus 7 is the system critical bus.
Buses 5 and 6 are the PCC buses for M G1 and M G2 . Fig.11
G1 MG1 shows the double-layer distributed communication network for
Agent the experimental MGC system. In the upper cyber network G,
(sys
*
, Vcri* )
e
DG12 DG11 ∗ ∗
M G1 agent receives the references ωsys and Vcri . In the lower
cyber networks G1 and G2 , DG11 and DG21 are the pinned
Pinning
link G DG units of M G1 and M G2 , respectively.
The system electrical parameters, generation cost parame-
G2 ters, DG-layer, MG-layer and MGC-layer control parameters
MG2
DG22 DG21 Agent are shown in Table VII–Table X. The ratio of reactive power
capacities among DG units is Qmax11 : Qmax12 : Qmax21 :
Qmax22 = 4 : 3 : 4 : 3.
Fig. 11. Topology of the double-layer distributed communication network for
the experimental MGC system
B. Study 1: Basic Control Performance Validation
TABLE VIII The DG-layer control is engaged initially. At t=t1 , the MG-
G ENERATION COST PARAMETERS FOR EXPERIMENTAL MGC SYSTEM
layer control is activated, and at t=t2 , the MGC-layer control
αki βki γki is introduced. Fig.12 shows the corresponding experimental
DG11 0.108 0.23 0.7 results. Fig.12(a) indicates that the system frequency can be
DG12 0.127 0.15 0.9 restored and then maintained at the rated value 50Hz by
DG21 0.108 0.23 0.7 the MG-layer and MGC-layer controls after t=t1 . Fig.12(b)
DG22 0.127 0.15 0.9
indicates that the MGC-layer control can restore the system
critical bus voltage to the rated value 1p.u. after t=t2 . Fig.12(c)
indicates that (i) the GCIVs (ηki ) of all DG units are equal at
the dSPACE DS1006 platform, in which the control strategy the stage of DG layer control, (ii) after t=t1 , ηki inside each
is programmed and then executed in real-time. Finally, the individual MG are equal, i.e., η11 = η12 and η21 = η22 , due
dSPACE generates PWM signals to switch the inverters with to the regulating effect of MG-layer control, and (iii) after
a switching frequency of 10kHz. The control desk installed t=t2 , the MGC-layer control imposes equal ηki of all DG
in a PC is responsible for monitoring the system states and units again, realizing a minimum generation cost of the whole

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
14

DG-layer DG&MG-layer DG&MG&MGC- Load4 on t3 Load4 off t4 Load4 on

control t1 controls t2 layer controls 50.2
50.2

Frequency (Hz)
50.1
Frequency (Hz)

50.0
50.0
49.7
49.9
49.5
49.8
49.2
50 55 60 65 70 75 80
10 15 20 25 30 35 40 Time (s)
Time (s) (a) System frequency (Hz)
(a) System frequency (Hz)
t3 t4
t1 t2
1.05

Voltage (p.u.)
1.00
Voltage (p.u.)

1.00
0.95

0.90
0.95
50 55 60 65 70 75 80
10 15 20 25 30 35 40
Time (s)
Time (s)
(b) System critical bus voltage Vcri (p.u.) (b) System critical bus voltage Vcri (p.u.)
t1 t2 t3 t4
0.32
η12
GCIV ηki ($/kW)

0.31
GCIV ηki ($/kW)

0.30 η11 0.30

η22
0.29
η21 0.28
DG11 DG12 DG11 DG12
0.28 DG21 DG22 DG21 DG22
10 15 20 25 30 35 40 50 55 60 65 70 75 80
Time (s) Time (s)
(c) DG generation cost increment value ηki (c) DG generation cost increment value ηki
t1 t2 t3 t4
0.30 0.23
Q11
Reactive power
Reactive power

0.20 Q11
Q12
(kVar)
(kVar)

0.20
Q21 Q21
0.15
Q12
0.10 DG11 DG12 DG11 DG12
Q22 DG21 Q22 DG21
0.04 DG22 0.10 DG22
10 15 20 25 30 35 40 50 55 60 65 70 75 80
Time (s) Time (s)
(d) DG output reactive powers Qki (kVar) (d) DG output reactive powers Qki (kVar)

Fig. 12. Experimental results for Study 1–basic control performance evalua-
tion: (a) system frequency, (b) system critical bus voltage, (c) DG generation
cost increment value ηki , (d) DG output reactive powers Qki . The DG-layer
control is engaged initially. At t=t1 , the MG-layer control is activated, and at
t=t2 , the MGC-layer control is introduced.
Fig. 13. Experimental results for Study 2–load change performance evaluation:
(a) system frequency, (b) system critical bus voltage, (c) DG generation cost
increment value ηki , (d) DG output reactive powers Qki . The system is
operated under the three-layer controls. At t=t3 , Load 4 is switched off. At
t=t4 , Load 4 is switched on again.

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
15

system. Fig.12(d) indicates that (i) the reactive power sharing

is not accurate before t=t1 , (ii) an accurate reactive power
sharing inside each individual MG, i.e., Q11 : Q12 = 4 : 3
t5 t6 t7 t8
and Q21 : Q22 = 4 : 3, is realized after applying the MG-
50.2 Load3 on Load3 off Load3 on layer control, and (iii) after t=t2 , the MGC-layer control can
Link MG2 to achieve an accurate reactive power sharing across the whole
Frequency (Hz)

50.1 MG1 fails

system, i.e., Q11 : Q12 : Q21 : Q22 = 4 : 3 : 4 : 3.
50.0

49.9 Link DG22 to C. Study 2: Load Change Performance Evaluation

DG21 fails
The system is operated under the three-layer controls. At
49.8 t=t3 , Load 4 is switched off. At t=t4 , Load 4 is switched
90 100 110 120 130 on again. Fig.13 indicates that after Load 4 switched off and
Time (s)
switched on, the proposed three-layer control can always (i)
(a) System frequency (Hz)
maintain the system frequency at 50Hz, Fig.13(a), (ii) maintain
t5 t6 t7 t8
1.10 the system critical bus voltage at 1p.u., Fig.13(b), (iii) impose
equal ηki of all DG units, Fig.13(c), and (iv) realize an accurate
Voltage (p.u.)

Link MG2 to
1.05 MG1 fails
reactive power sharing Q11 : Q12 : Q21 : Q22 = 4 : 3 : 4 : 3
across the whole system, Fig.13(d).
1.00

0.95 Link DG22 to D. Study 3: Communication Link Failure Resiliency

DG21 fails
In Study 3, Load 4 is always disconnected from the system,
0.90 while Loads 1-3 are connected to the system. Under this
90 100 110 120 130
Time (s)
circumstance, the active and reactive power of Load3 are
(b) System critical bus voltage Vcri (p.u.) 46.4% and 100% of the whole active power and reactive
t5 t6 t7 t8 power load, respectively. The system is operated under the
three-layer controls. At t=t5 , a communication link failure in
GCIV ηki ($/kW)

0.30 the lower communication network happens, i.e., the link from
0.28 DG22 to DG21 fails. Subsequently, at t=t6 , Load 3 is switched
off. Then, at t=t7 , a communication link failure in the upper
Link DG22 to Link MG2 to
0.26 DG21 fails MG1 fails communication network happens, i.e., the link from M G2
agent to M G1 agent fails. Finally, at t=t8 , Load 3 is switched
0.24 DG11 DG12 on again. Fig.14 shows that the control objectives in terms
DG21 DG22 of (a) system frequency, (b) system critical bus voltage, (c)
0.22
90 100 110 120 130 GCIV and (d) reactive power sharing can still be realized after
Time (s) communication link failures. This is because the remaining
(c) DG generation cost increment value ηki communication network still contains a spanning tree, and thus
t5 t6 t7 t8
the convergence condition of algorithms (1) and (2) is still
satisfied.
Reactive power

0.20
Remark 5: Switching on and off Load 3 corresponds to a
(kVar)

large load disturbance to the system. The results of Fig.14

0.10 Link MG2 to
show that (i) no significant overshoots are observed, and (ii)
Link DG22 to MG1 fails
DG21 fails the inverters are not tripped. Therefore, the robustness of
0 DG11 DG12 our control method against large load disturbances can be
DG21 DG22 validated.
-0.08
90 100 110 120 130
Time (s) VII. C ONCLUSIONS
(d) DG output reactive powers Qki (kVar) This paper presents a pinning-based hierarchical and dis-
tributed cooperative control strategy for AC MGC. The pro-
Fig. 14. Experimental results for Study 3–communication link failure resilien- posed method fully reserves the autonomy characteristic of
cy: (a) system frequency, (b) system critical bus voltage, (c) DG generation
cost increment value ηki , (d) DG output reactive powers Qki . The system each individual MG while realizing their coordination through
is operated under the three-layer controls. At t=t5 , the communication link pinning links and communications among MG agents in the
from DG22 to DG21 fails. At t=t6 , Load 3 is switched off. At t=t7 , the upper cyber network.
communication link from M G2 agent to M G1 agent fails. At t=t8 , Load 3
is switched on again. With the proposed method, multiple control objectives can
be achieved, including (i) minimization of the system overall
generation cost, (ii) restoration of system frequency and critical
bus voltage, (iii) voltage control capability for the PCC bus of

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
each MG, and (iv) an accurate reactive power sharing among
DG units in the whole system.
In addition, this paper also develops a unified small-signal
dynamic model for AC MGC to analyze the system small-
signal stability. The model captures the details of the three-
layer controllers of DGs and MGs, and also considers the
circuit features of networks and loads. Eigen-analysis results
reveal that (i) the dominant oscillation modes are associated
with different control layers and the interaction among control
layers may lead to low-damping modes, and (ii) the relation
between dominant oscillatory modes and control parameters
exhibits complex multivariate interactions.
Finally, the performance of the proposed method is evaluated
and validated by time-domain simulation and experimental
results.
R EFERENCES
[1] J. M. Guerrero, J. C. Vasquez, J. Matas, D. Vicuna, L. Garcı́a, and
M. Castilla, “Hierarchical control of droop-controlled AC and DC
microgrids–A general approach toward standardization,” IEEE Transac-
tions on Industrial Electronics, vol. 58, no. 1, pp. 158–172, 2011.
[2] Y. W. Li and C.-N. Kao, “An accurate power control strategy for power-
electronics-interfaced distributed generation units operating in a low-
voltage multibus microgrid,” IEEE Transactions on Power Electronics,
vol. 24, no. 12, pp. 2977–2988, 2009.
[3] H. Jia, Q. Xiao, and J. He, “An improved grid current and DC capacitor
voltage balancing method for three-terminal hybrid AC/DC microgrid,”
IEEE Transactions on Smart Grid, vol. 10, no. 6, pp. 5876–5888, 2019.
[4] Y. Han, K. Zhang, H. Li, E. A. A. Coelho, and J. M. Guerrero, “MAS-
based distributed coordinated control and optimization in microgrid and
microgrid clusters: A comprehensive overview,” IEEE Transactions on
Power Electronics, vol. 33, no. 8, pp. 6488–6508, 2018.
[5] E. Pashajavid, A. Ghosh, and F. Zare, “A multimode supervisory control
scheme for coupling remote droop-regulated microgrids,” IEEE Trans-
actions on Smart Grid, vol. 9, no. 5, pp. 5381–5392, 2018.
[6] Y. Xu, C.-C. Liu, K. P. Schneider, F. K. Tuffner, and D. T. Ton, “Mi-
crogrids for service restoration to critical load in a resilient distribution
system,” IEEE Transactions on Smart Grid, vol. 9, no. 1, pp. 426–437,
2018.
[7] Z. Wang, C. Shen, Y. Xu, F. Liu, X. Wu, and C.-C. Liu, “Risk-limiting
load restoration for resilience enhancement with intermittent energy
resources,” IEEE Transactions on Smart Grid, vol. 10, no. 3, pp. 2507–
2522, 2019.
[8] Y. Wang, Y. Xu, J. He, C.-C. Liu, K. P. Schneider, M. Hong, and D. T.
Ton, “Coordinating multiple sources for service restoration to enhance
resilience of distribution systems,” IEEE Transactions on Smart Grid,
2019, early access article online.
[9] Z. Xu, P. Yang, Y. Zhang, Z. Zeng, C. Zheng, and J. Peng, “Control
devices development of multi-microgrids based on hierarchical structure,”
IET Generation, Transmission & Distribution, vol. 10, no. 16, pp. 4249–
4256, 2016.
[10] T. John and S. P. Lam, “Voltage and frequency control during microgrid
islanding in a multi-area multi-microgrid system,” IET Generation,
Transmission & Distribution, vol. 11, no. 6, pp. 1502–1512, 2017.
[11] R. Zamora and A. K. Srivastava, “Multi-layer architecture for voltage
and frequency control in networked microgrids,” IEEE Transactions on
Smart Grid, vol. 9, no. 3, pp. 2076–2085, 2018.
[12] N. J. Gil and J. P. Lopes, “Hierarchical frequency control scheme for
islanded multi-microgrids operation,” in 2007 IEEE Lausanne Power
Tech. IEEE, 2007, pp. 473–478.
[13] A. Bidram, A. Davoudi, F. L. Lewis, and Z. Qu, “Secondary control
of microgrids based on distributed cooperative control of multi-agent
systems,” IET Generation, Transmission & Distribution, vol. 7, no. 8,
pp. 822–831, 2013.
[14] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and coopera-
tion in networked multi-agent systems,” Proceedings of the IEEE, vol. 95,
no. 1, pp. 215–233, 2007.
[15] J. W. Simpson-Porco, Q. Shafiee, F. Dörfler, J. C. Vasquez, J. M.
Guerrero, and F. Bullo, “Secondary frequency and voltage control of
islanded microgrids via distributed averaging,” IEEE Transactions on
Industrial Electronics, vol. 62, no. 11, pp. 7025–7038, 2015.
0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
16
[16] V. Nasirian, Q. Shafiee, J. Guerrero, F. Lewis, and A. Davoudi, “Droop-
free distributed control for AC microgrids,” IEEE Transactions on Power
Electronics, vol. 31, no. 2, pp. 1600–1617, 2016.
[17] V. Nasirian, S. Moayedi, A. Davoudi, and F. L. Lewis, “Distributed
cooperative control of DC microgrids,” IEEE Transactions on Power
Electronics, vol. 30, no. 4, pp. 2288–2303, 2015.
[18] Q. Shafiee, J. M. Guerrero, and J. C. Vasquez, “Distributed secondary
control for islanded microgrids–A novel approach,” IEEE Transactions
on Power Electronics, vol. 29, no. 2, pp. 1018–1031, 2014.
[19] D. O. Amoateng, M. Al Hosani, M. S. Elmoursi, K. Turitsyn, and
J. L. Kirtley, “Adaptive voltage and frequency control of islanded multi-
microgrids,” IEEE Transactions on Power Systems, vol. 33, no. 4, pp.
4454–4465, 2018.
[20] M. S. Golsorkhi, D. J. Hill, and H. R. Karshenas, “Distributed voltage
control and power management of networked microgrids,” IEEE Journal
of Emerging and Selected Topics in Power Electronics, vol. 6, no. 4, pp.
1892–1902, 2018.
[21] X. Wu, Y. Xu, X. Wu, J. He, J. M. Guerrero, C.-C. Liu, K. P. Schneider,
and D. T. Ton, “A two-layer distributed cooperative control method for is-
landed networked microgrid systems,” IEEE Transactions on Smart Grid,
2019, early access article online, DOI: 10.1109/TSG.2019.2928330.
[22] X. Lu, J. Lai, X. Yu, Y. Wang, and J. M. Guerrero, “Distributed
coordination of islanded microgrid clusters using a two-layer intermittent
communication network,” IEEE Transactions on Industrial Informatics,
vol. 14, no. 9, pp. 3956–3969, 2018.
[23] S. Moayedi and A. Davoudi, “Distributed tertiary control of DC micro-
grid clusters,” IEEE Transactions on Power Electronics, vol. 31, no. 2,
pp. 1717–1733, 2016.
[24] Q. Shafiee, T. Dragičević, J. C. Vasquez, and J. M. Guerrero, “Hierar-
chical control for multiple DC-microgrids clusters,” IEEE Transactions
on Energy Conversion, vol. 29, no. 4, pp. 922–933, 2014.
[25] J. Lai, X. Lu, X. Yu, and A. Monti, “Cluster-oriented distributed
cooperative control for multiple AC microgrids,” IEEE Transactions on
Industrial Informatics, vol. 15, no. 11, pp. 5906–5918, 2019.
[26] N. Pogaku, M. Prodanović, and T. C. Green, “Modeling, analysis and
testing of autonomous operation of an inverter-based microgrid,” IEEE
Transactions on Power Electronics, vol. 22, no. 2, pp. 613–625, 2007.
[27] X. Wu, C. Shen, and R. Iravani, “A distributed, cooperative frequency
and voltage control for microgrids,” IEEE Transactions on Smart Grid,
vol. 9, no. 4, pp. 2764–2776, 2018.
[28] X. Wu and C. Shen, “Distributed optimal control for stability en-
hancement of microgrids with multiple distributed generators,” IEEE
Transactions on Power Systems, vol. 32, no. 5, pp. 4045–4059, 2017.
[29] N. Bottrell, M. Prodanovic, and T. C. Green, “Dynamic stability of a
microgrid with an active load,” IEEE Transactions on Power Electronics,
vol. 28, no. 11, pp. 5107–5119, 2013.
[30] H. Liang, B. J. Choi, W. Zhuang, and X. Shen, “Stability enhancement
of decentralized inverter control through wireless communications in
microgrids,” IEEE Transactions on Smart Grid, vol. 4, no. 1, pp. 321–
331, 2013.
[31] G. Chen and E. Feng, “Distributed secondary control and optimal power
sharing in microgrids,” IEEE/CAA Journal of Automatica Sinica, vol. 2,
no. 3, pp. 304–312, 2015.
A PPENDIX A
PARAMETERS OF THE D OUBLE - LAYER C YBER N ETWORK
The associated adjacency matrix and leader-adjacency ma-
trix parameters for Fig.4 are given by
   
0 1 1 0 1 1
Ae =  1 0 1  A1 =  1 0 1
1 1 0 1 1 0
   
0 1 1 0 1 1
A2 =  1 0 1  A3 =  1 0 1
1 1 0 1 1 0
1 2 3
 
Bpin = Bpin = Bpin = 1 0 0 (57)
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
17

A PPENDIX B q̃ki = vodki ioqki − voqki iodki . (63b)

D ETAILED S MALL - SIGNAL DYNAMIC M ODELING P ROCESS
IN S ECTION IV The average active power Pki and reactive power Qki can be
obtained based on 1st-order low-pass filters, i.e.,
This appendix provides details of the small-signal dynamic
ωc
modeling process and parameter matrices in Section IV. Denote Pki = p̃ki , (64a)
N as the s + ωc
PMnumber of all the DG units in the MGC system, i.e., ωc
N = k=1 Nk . The generation cost function GCki (Pki ) is
Qki = q̃ki , (64b)
s + ωc
assumed to be the quadratic function (56). Therefore, we have
ηki (Pki ) = 2αki Pki + βki . where ωc is the cut-off frequency of the 1st-order low-pass
ωc
Linearizing (11) yields filter s+ω c
and it is set as 100rad/s for the numerical and
  experimental studies in this paper.
∆δki By linearizing (25), (63), (64) and substituting (58) into the
∆ωki = Cpw  ∆Pki  + ∆Ωki , linearized results, we have
∆Qki    
  ∆δ̇ki ∆δki
Cpw = 0 −2ked αki 0 . (58)  ∆Ṗki  = AP  ∆Pki  + DP ∆Ωki + BP 1 [∆vodqki ]
In light of (58), linearizing (12) yields ∆Q̇ki ∆Qki
 
∆δki X +BP 2 [∆iodqki ] + CP ∆ωg
∆Ω̇ki = Af  ∆Pki  + Bf ∆Ωki + cωki akij ∆Ωkj
   
0 −2ked αki 0 −1
∆Qki j∈Nik AP =  0 −ωc 0  CP =  0 
0 0 −ωc 0
X
−2 (cωki − cP ki ) akij ked αkj ∆Pkj + cωki gik ∆ωrefk    
j∈Nik 0 0 1
    BP 1 =  ωc Iodki ωc Ioqki  DP =  0 
Af = 0 2 (cωki − cP ki )
X
akij + cωki gik  ked αki 0 −ωc Ioqki ωc Iodki 0
 
j∈Nik 0 0
  BP 2 =  ωc Vodki ωc Voqki  (65)
X ωc Voqki −ωc Vodki
Bf = −cωki  akij + gik  . (59)
j∈Nik where Vodki , Voqki , Iodki , Ioqki are the equilibrium point values
of vodki , voqki , iodki , ioqki for linearization.
From (13a), we have
Recall (26). Linearizing (13) yields
Vf ki = Vn − nki Qki + λki . (60)  
∆δki
Linearizing (14), (60) and rearranging the linearized results ∆vodqki = CP V  ∆Pki  + DP V 1 ∆λki + DP V 2 ∆hki
yield ∆Qki
   
0 0 −nki
 
∆δki 1
X CP V = DP V 1 =
∆λ̇ki = Ag  ∆Pki  + Bg ∆λki − cvki akij nkj ∆Qkj 0 0 0 0
∆Qki
 
j∈Ni k
−1
DP V 2 = . (66)
0
X
k k
+cvki aij ∆λkj + cvki gi ∆Vf refk
j∈Nik The DG unit output voltage vodqki and current iodqki are
expressed in the local dq reference frame of DGki . To integrate
" ! #
akij + gik
P
Ag = 0 0 cvki nki DGki into the whole system model, vodqki and iodqki need to
j∈Nik be transformed to the form in the global DQ-frame, i.e., voDQki
 
X and ioDQki . The transformation is given by
Bg = −cvki  akij + gik  . (61)
∆voDQki = Ts ∆vodqki + Tc ∆δki (67)
j∈Nik
−1 −1
∆iodqki = Ts ∆ioDQki + TV ∆δki (68)
By linearizing (15), we have  
cos δ0ki − sin δ0ki

∆δki
 Ts =
X sin δ0ki cos δ0ki
∆ḣki = Ah  ∆Pki  − cQki akij nkj ∆Qkj  
−Vodki sin δ0ki − Voqki cos δ0ki
∆Qki j∈Nik Tc =
 P k  Vodki cos δ0ki − Voqki sin δ0ki
Ah = 0 0 cQki aij nkj .  
(62) −IoDki sin δ0ki + IoQki cos δ0ki
k j∈Ni TV−1 = , (69)
−IoDki cos δ0ki − IoQki sin δ0ki
Based on the instantaneous power theory, the instantaneous
where δ0ki is the equilibrium point value of δki .
output active power p̃ki and reactive power q̃ki of DGki are
Then, the small-signal dynamic model of DGki , i.e., (27),
expressed as
can be obtained by (i) substituting (66) and (68) into (65), and
p̃ki = vodki iodki + voqki ioqki , (63a) (ii) combining the substituted results and (59), (61) and (62).

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
18

Details of the parameter matrices in (27) are provided in (80). Then, LINVk and LINV in (30) and (31) can be expressed as
Moreover, from (58), we have
 
∆ωki = Dinvki ∆Xinvki Linvk1 0 ... 0
   0 L invk2 ... 0 
Dinvki = Cpw 1 0 0 1×6 (70) LINVk = .
 
.. .. ..
 ..

∆ωg = DINV1 ∆XINV1 . . . 
DINV1
 
= Dinv11 0 . . . 0 1×6N (71) 0 0 . . . LinvkNk 2Nk ×6Nk
1  
LINV1 0 ... 0
 0 LINV2 ... 0 
Thus, DINV in (39) is LINV = .

.. .. ..

(74)
 ..

. . . 
  0 0 . . . LINVM 2N ×6N
DINV = DINV1 0 . . . 0 1×6N (72)

Details of the parameter matrices in the small-signal dynam- h i

VPCCQk0
ic model of M Gk , i.e., (29), are provided in (81). In (36) and (37), APCCk = VVPCCDk0
PCCk0 V PCCk0
, where
Substituting (66) into (67) yields VPCCDk0 , VPCCQk0 and VPCCk0 are the equilibrium point
values of VPCCDk , VPCCQk and VPCCk .

∆voDQki = Linvki ∆Xinvki Details of the parameter matrices in the combination model
h i of all individual MGs (40) and (41) and the state-space
Linvki = Ts CP V + Tec 0 Ts DP V 1 Ts DP V 2 equation of ∆ψ (42) are provided in (82).
2×6
 
Tec = Tc 0 0 (73) For the MGC-layer controller of M Gk agent, linearizing

AP + BP 1 CP V + BP 2 TV−1 0 0 DP BP 1 DP V 1 BP 1 DP V 2
   
 Af Bf 0 0 
Ainvki = 
 Ag 0 Bg 0 
Ah 0 0 0 6×6
−1
       
BP 2 Ts CP 0 0
 0   0   0   cωki g k 
Binvki =  Cinvki =   Hinvki =   Rinvki =  i 
 0   0   cvki gik   0 
0 6×2
0 6×1 0 6×1
0 6×1
 
0 0 0 0 0 0
 0 −2(cωki − cP ki )akij ked αkj 0 c ωki ak
ij 0 0 
Finvkij =  (80)
0 0 −cvki akij nkj 0 cvki akij 0 
0 0 −cQki akij nkj 0 0 0 6×6

AINVk = GINVk + FINVk

   
Ainvk1 0 ... 0 Finvk11
Finvk12 . . . Finvk1Nk
 0 A invk2 . . . 0   Finvk21
Finvk22 . . . Finvk2Nk 
GINVk =  FINVk = 
   
.. .. . .. ..  .. .... .. 
 . . .   . . . . 
0 0 . . . AinvkNk FinvkNk 1 FinvkNk 2 . . . FinvkNk Nk
 6Nk ×6Nk    6Nk ×6Nk

Binvk1 0 ... 0 Cinvk1 Hinvk1
 0 B invk2 . . . 0   Cinvk2   Hinvk2 
BINVk =  CINVk =  HINVk =
     
.. .. . . ..  ..  .. 
 . . . .   .   . 
0 0 . . . BinvkNk 6Nk ×2Nk
CinvkNk 6Nk ×1
HinvkNk 6Nk ×1
 
Rinvk1
 Rinvk2 
RINVk =  (81)
 
.. 
 . 
RinvkNk 6Nk ×1

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
19

   
AINV1 0 ... 0 BINV1 0 ... 0
 0 AINV2 ... 0   0 B INV2 ... 0 
AINV = BINV =
   
.. .. .. ..  .. .. .. .. 
 . . . .   . . . . 
0 0 . . . AINVM 0 0 . . . BINVM
 6N ×6N  6N ×2N
HINV1 0 ... 0 RINV1 0 ... 0
 0 H INV2 ... 0   0 R INV2 ... 0 
HINV = RINV =
   
.. .. .. ..  .. .. .. .. 
 . . . .   . . . . 
0 0 . . . HINVM 6N ×M
0 0 . . . RINVM 6N ×M
     
CINV1 HINV1 APCC1 ZPCC1 APCC1 ZPCC1
 CINV2   HINV2 APCC2 ZPCC2   APCC2 ZPCC2 
CINV =  NINV =  TPCC =  (82)
     
..  ..  .. 
 .   .   . 
CINVM 6N ×1
HINV1M APCCM ZPCCM 6N ×2N
APCCM ZPCCM M ×2N

AMGC = GMGC + FMGC

   
AMGC1 0 ... 0 HMGC1
 0 AMGC2 . . . 0   HMGC2 
GMGC =  H =
   
.. .. .. ..  MGC  .. 
 . . . .   . 
0 0 . . . AMGCM 3M ×3M HMGCM 3M ×1
   
FMGC11 FMGC12 . . . FMGC1M EMGC1
 FMGC21 FMGC22 . . . FMGC2M   EMGC2 
FMGC =  E = (83)
   
.. .. .. .
..  MGC  .. 
 . . .   . 
FMGCM 1 FMGCM 2 . . . FMGCM M 3M ×3M
EMGCM 3M ×6N

 
JINV kI HINV RINV Acou + kP HINV Bcou 0
 −TPCC WINV 0 Bcou 0 
Asys =  (84)
 JMGC 0 AMGC kIS HMGC 
−Acri Zcri WINV 0 0 0 (6N +4M +1)×(6N +4M +1)

(18) and (19) yields Linearizing (23) yields

 
X X 1 X
∆Φ̇k = Bcf ∆Φk + cωk ãkl ∆Φl ∆Ḣk = cQk ãkl  k nki ∆Qki 
|F |
l∈Hk l∈Hk i∈F k
   
X 1 X X 1 X
−cP k ãkl  k 2αki ∆Pki  −cQk ãkl  l nli ∆Qli  (77)
|F | |F |
l∈Hk i∈F k k l∈H l i∈F
 
X 1 X Then, (44) can be obtained by combining (75), (76) and (77).
+cP k ãkl  l 2αli ∆Pli 
|F | Details of AMGCk , FMGCkl and HMGCk are given by
l∈Hk i∈F l
   
Bcf 0 0 0
 
X
Bcf = −cωk  ãkl + g̃k  (75) AMGCk =  0 Bcg 0  HMGCk =  cvk g̃k 
l∈Hk 0 0 0 3×3 0 3×1
 
cωk ãkl 0 0
Linearizing (21) and (43) yields FMGCkl =  0 cvk ãkl 0  (78)
0 0 0 3×3
X
∆Ṫk = Bcg ∆Tk + cvk ãkl ∆Tl + cvk g̃k ∆Vsref EMGCk is a 3 × 6N sparse matrix. For the first row of
l∈Hk EMGCk , the elements corresponding to ∆Pki of the pinned
DG units of M Gk are −cP k l∈Hk ãkl 2α
P ki
, and the elements
 
X |F k |
Bcg = −cvk  ãkl + g̃k  (76) corresponding to ∆Pli of the pinned DG units of M Gl are
l∈Hk cP k ãkl 2α li
|F l |
. The second row of EMGCk are all zeros. For the

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPEL.2020.2972321, IEEE
Transactions on Power Electronics
20

third row of EMGCk , the elements corresponding to ∆Qki of (54) are given by
P nki
the pinned DG units of M Gk are cQk l∈Hk ãkl |F k | , and the
 
Ayy 0 ... 0
elements corresponding to ∆Qli of the pinned DG units of  0 Ayy
nli ... 0 
M Gl are −cQk ãkl |F l| .
 
Acou =  .  Ayy = 1 0 0
 
.. .. ..
 .. . . . 
0 0 . . . Ayy
Then, details of the parameter matrices in (45) are provided  
in (83). Byy 0 ... 0
 0 Byy ... 0   
Bcou =  .  Byy = 0 1 −1 . (79)
 
.. .. .
 .. . ..
hi
VcriQ0
In (47) and (48), Acri = VVcriD0 cri0 Vcri0
, where VcriD0 , . 
VcriQ0 and Vcri0 are the equilibrium point values of VcriD , VcriQ 0 0 . . . Byy
and Vcri . Finally, the detail of Asys in the complete MGC system
model (55) is provided in (84).
Details of Acou ∈ RM ×3M and Bcou ∈ RM ×3M in (53) and

0885-8993 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 04:30:00 UTC from IEEE Xplore. Restrictions apply.