SPECIAL SECTION ON KEY ENABLING

TECHNOLOGIES FOR PROSUMER ENERGY MANAGEMENT

Received December 24, 2020, accepted January 9, 2021, date of publication January 13, 2021, date of current version January 22, 2021.
Digital Object Identifier 10.1109/ACCESS.2021.3051464

Research on the Bi-Level Optimization Model of

Distribution Network Based on Distributed
Cooperative Control
YANCHUN XU1 , JIN ZHANG 1, PING WANG1 , AND MI LU2 , (Senior Member, IEEE)
1 Hubei Provincial Key Laboratory for Operation and Control of Cascaded Hydropower Station, China Three Gorges University, Yichang 443002, China
2 Department of Electrical & Computer Engineering, Texas A&M University, College Station, TX 77840, USA

Corresponding author: Jin Zhang ([email protected])

This work was supported in part by the National Natural Science Foundation of China under Grant 51707102, and in part by the Key
Project of Science and Technology Research Plan of Education Department of Hubei under Grant D20201203.

ABSTRACT The uncertainty of high-permeability renewable energy output brings about great challenges
to the operation control in distribution network. The traditional control mode has lower reliability, higher
calculation and greater communication pressure. It is more difficult to realize the technical requirements of
‘‘plug and play’’ of power components. Therefore, a bi-level coordinated control model is proposed for the
operation and energy management optimization of active distribution network with multi-microgrid in this
paper. The consensus algorithm which exists in the multi-agent system is applied in the upper-level model
and the capacity utilization ratio of the different microgrids is set as the consistent variable because the ratio
of resistance and reactance is usually larger in distribution network and the interests in the different microgrid
entities should be distributed fairly. Moreover, in the lower-level model, the lowest operating cost is the final
goal and the model predictive control rolling optimization is utilized to reasonably allocate the output of the
equipments in the microgrid according to the changing trend of renewable energy. Thereby the tracking of
the upper-level control commands is realized. Finally, the simulations are conducted in the IEEE-33 nodes
distribution system connected to a multi-microgrid to verify the feasibility and effectiveness of the proposed
method, in which a reference for the optimal operation of higher penetration renewable energy to distribution
network is provided.

INDEX TERMS Bi-level model, consensus algorithm, distributed control, high-permeability, model predic-
tive control.

I. INTRODUCTION the consensus algorithm is the interaction between one con-

With the continuous expansion of power system scale and troller and another neighboring controller in the same sys-
the large-scale access of renewable energy with strong tem through the communication link. The distributed agent
randomness and volatility, the energy management and oper- makes decisions based on the obtained information and its
ation control in power system are facing more severe chal- own state in order to achieve the same state for themselves
lenges [1]. The traditional dispatch and operation methods finally [4]. Distributed collaborative control has been exten-
for distribution network cannot meet the needs of the soci- sively researched and applied in many engineering fields,
ety [2]. Compared with the traditional centralized control such as automatic generation control [5], unmanned aerial
and the decentralized control, the distributed control system vehicle (UAV) formation flight [6], convergent control of
has better control robustness, communication flexibility and multi-agent [7].
system scalability [3] and it can realize ‘‘plug and play’’ Reference [8] shows the influence on the convergence of
of power components, which is flexible and convenient. the consensus algorithm because of the transmission delay
In this paper, the basic idea of distributed control using of the communication link between multi-agents, thus the
feasibility of the consensus algorithm in the operation con-
The associate editor coordinating the review of this manuscript and trol of the distribution network is proved. Reference [9]
approving it for publication was Huai-zhi Wang . shows the influence on the convergence results because of

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
11798 VOLUME 9, 2021
Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

error consistency variable propagation in the different micro-

grid individuals for improving their own economic benefits.
Meanwhile, the lines of distribution network in the medium
and low voltage level has a larger resistance/reactance ratio.
Therefore, the nodal voltage is more sensitive to the injection
of active power. This is the reason that the method should
be adopted for the regulation voltage of active power [10]. FIGURE 1. Centralized and decentralized control model. (a) centralized
control model. (b) decentralized control model.
Reference [11] proposes that the microgrid is regarded as an
effective unit for integrating distributed generator (DG) and
the fluctuation of renewable energy output is reduced to a cer- the full utilization of historical data as the reference data lacks
tain extent by taking advantage of the time-complementary robustness.
characteristics of wind and solar output. Thus, the microgrid It can be found that the bi-level control model is used in
doesn’t have to face large number of DGs with distributed some papers. It has obvious advantages in maintaining the
access directly. It provides the idea for the integration opera- stable operation or the lowest operating cost compared with
tions of a higher proportional renewable energy. the traditional single level model. However, there are fewer
The reasonable distribution of the output for each device in papers utilizing the bi-level model to simultaneously consider
the microgrid needs to be combined with the changing trend the two issues of voltage stability and operating economy in
of the renewable energy output and the forecast accuracy distribution network.
of renewable energy output is negative correlation with the In this paper, a bi-level coordinated control model is pro-
forecast lead time. The data in reference [12] show that the posed for the operation and energy management optimization
error distribution interval in the forecast 24 hours ahead is in distribution system with multiple microgrids. By using the
5.45 times that in the forecast 1 hour ahead if the 95% confi- bi-level control model, the lowest operating cost of the micro-
dence interval is taken as an example. The model predictive grid will be realized under the premise of safe operation.
control (MPC) can perform rolling optimization based on The distributed control method is adopted in the upper-level
the latest predictive data to make reasonable control of DG model and the capacity utilization ratio of the microgrid is
for the uncertainty of the high-penetration renewable energy used as the consistency variable. The total output power of
output after it is connected to the distribution network [13]. microgrid can be calculated separately in each microgrid to
According to the forecast data of the load and renewable maintain the voltage stability in distribution network. At the
energy, two adjacent moments are put together for real-time same time, the day-ahead forecast data of renewable energy is
power generation control optimization decision to meet the employed in the lower-level model in order to make the day-
constraints of safety operation in distribution network [14]. ahead scheduling plan. The latest forecast data in the day is
However, the centralized control method requires the dispatch combined with MPC to optimize the operation status of the
center to master global information and it is easier to cause device in microgrid and avoid large fluctuations in the high
its poor reliability. If the incremental ratio of consumption proportion of renewable energy, which saves economic costs
for distributed generator is set as a consistent variable, the and brings about higher robustness.
cooperative control method is adopted in the consensus algo-
rithm to improve the control reliability and realize the optimal II. CONTROL STRATEGY OF MULTI-MICROGRID SYSTEM
economic dispatch in distribution network [15]. However, the A. DISTRIBUTED COLLABORATIVE CONTROL
voltage stability is not considered in reference [15]. The traditional control methods for distributed powers in
Hierarchical control can decompose the overall control distribution network are mainly centralized and distributed as
goal into several sub-goals in order to reduce the control shown in Fig. 1. The fundamental difference between the two
difficulty and improve the effects of control. Reference [16] control methods is whether there is a central controller in the
proposes multiple-step greedy policy based on consensus system. In the centralized control model, all DGs are directly
Q-learning strategy and it improves the adaptability of the managed by the central controller, and there is no communi-
consensus algorithm in a dynamic random environment. But cation connection between two adjoining DGs. The central
it doesn’t consider the economic costs. The scenario model- controller performs the centralized calculations according
ing of the uncertainty for the load and renewable energy are to the information uploaded by each DG, and delivers the
built based on the establishment of a bi-level control model commands to each DG, which causes the centralized control
in reference [17]. The expectation of the historical output model lack of reliability [18]. However, in the decentralized
scenario for renewable energy is regarded as the day-ahead control model, each DG performs update control based on its
dispatch value to achieve the lowest operating cost. Moreover, own information, so the coordination is poor.
the bi-level model can be converted to a single-layer model Nowadays, a novel distributed control model is utilized
for solution by the Karush–Kuhn–Tucker (KKT) condition and there is a communication link between adjacent DGs.
and constraint relaxation technology. Besides, the voltage sta- Meanwhile the DGs are controlled by interacting necessary
bility is ignored in this method and the low solution efficiency information so that the operating status of each DG tends to
is not suitable for multi-microgrid connection. Meanwhile, be consistent for achieving some specific control goals shown

VOLUME 9, 2021 11799

 Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

node voltage Vcp,i (i = 1,. . . , n) is controlled, and this voltage

can be measured in real time by the measuring element in the
microgrid i. Based on the consensus algorithm, the update
formula of capacity utilization ratio in the microgrid j can be
calculated as formula (2)
FIGURE 2. Distributed control model. N
ref
X
ϕj (k + 1) = wj,m ϕm (k) + wi,j ω1ϕi (k) (2)
in Fig. 2. The distributed control model has better economy m=1
and scalability. where
The establishment of communication links in distributed ϕj (k+1) represents the updated capacity utilization ratio of
control model can be explained by knowledge of graph the- microgrid j after the kth iteration;
ory. The topology of a multi-agent network is represented wj,m represents the weight of microgrid j received micro-
by G = {V , ε, A}. The distributed topology point set is grid m information;.
represented by V = {v1 ,. . . , vn }. The directed edge set of ω represents iteration step, its value range is 0-1. Generally
graph G is represented by ε ⊆ V × V ; the adjacency matrix speaking, the larger the iteration step, the faster the conver-
of graph G is represented by A(G) = {aij }n×n . gence speed. However, the worse the convergence accuracy
In a simple graph, the adjacency matrix consists of 0 or 1 is worse [26];
and its diagonal element is 0 [19]. The degree of the node vi , ref
1ϕi (k) represents the reference value of the capacity
is di . Thus D(G) equals diag(d1 ,. . . , dn ). The laplace matrix utilization ratio in the kth iteration for the microgrid i.
of G is L(G) and the matrix L(G) equals D(G)-A(G). The ref
In order to get the value of 1ϕi (k), the voltage control
elements of the matrix L(G) is lij shown in formula (1). performance evaluation function is defined as shown in for-

lii =
X
aij mula (3)
i6=j (1) 1
fv,i = (VCP,i − 1)2 (3)
2

lij = −aij
ref
For the undirected graphs, L(G) is a symmetric positive It also can define 1ϕi as the negative value of the partial
semi-definite matrix. The eigenvalue of L(G) contains a lot of derivative for the goal function fv,i with respect to ϕi , as shown
information of the graph G. The second smallest eigenvalue in formula (4)
of L(G) is the ‘‘algebraic connectivity’’ of the graph G, which ref ∂fv,i
1ϕi (k) = − (4)
is a quantitative indicator of the convergence in the consensus ∂ϕi ϕi (k)
algorithm [20].
According to formula (3), we can deduce formula (5)
B. ACTIVE POWER ALLOCATION BASED ON CONSENSUS ∂fv,i
ALGORITHM = VCP,i − 1 (5)
∂VCP,i
The outputs of wind turbine and photovoltaic have the com-
ref
plementary characteristics. In order to reduce the impact The calculation method of 1ϕi is shown as formula (6)
of the randomness and the fluctuation of renewable energy ∂fv,i ∂fv,i ∂VCP,i ∂Pi
ref
output on distribution network, the microgrid system with 1ϕi =− =− × ×
∂ϕi ∂VCP,i ∂Pi ∂ϕi
wind-and-solar-storage complement is usually connected to
∂VCP,i
the distribution network [21]. While the investors of micro- = −(VCP,i − 1) × × Si × cos ϕi (6)
grid are different and the actual operating conditions do ∂Pi
not support the full-load output of the power source, the where Si represents the capacity of the microgrid i; cosϕi
power generation is distributed proportionally according to represents the power factor of the microgrid i; Pi represents
the connected capacity, which is also a relatively simple and the active power output of the microgrid i.
fair benefit allocation [22]. Meanwhile, the renewable energy  X
 1P i = P Gi − PLi − U i Uj (Gij cos θij + Bij sin θij ) = 0
generator needn’t to work in the maximum power point track- 

j∈i
ing (MPPT) mode. Wind turbine can change the output active ∂Pi
1Pi = PGi − PLi − Ui − Ui2 Gii = 0

power by changing the pitch angle, rotor speed and other 
∂Vi
measures [23] and photovoltaic can increase device voltage to
(7)
reduce output power [24], [25]. The utilization ratio of micro-
grid capacity is selected as a consistent variable. It can protect According to the network power flow (7), Eq. (8) could be
the privacy of information transmission between the different derived as follows.
microgrids. In this study, the sum of the output of renewable ∂VCP,i VCP,i
energy and the capacity of fuel generators in the microgrid are = 2 (8)
∂Pi VCP,i Gii + Pi
defined as the capacity of the microgrid. While n microgrids
are connected to distribution network, It can assume that the where Gii represents the self-admittance of the node i.

11800 VOLUME 9, 2021

Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

FIGURE 4. Communication network digraph. MG is the abbreviation of

microgrid.

A. UPPER-LEVEL DISTRIBUTION NETWORK MODEL

The upper-level model is responsible for maintaining the
voltage stability in distribution network. With the help of
FIGURE 3. Bi-level control model.
the communication connection of the distributed agent in
different microgrids [29], the distributed agent can adjust the
The weight of information received between microgrids is output power of microgrid according to the controlled nodal
only related to the topology of the communication network, voltage while the loads fluctuate.
which is determined by the degree of node [27]. However, The microgrid in which the nodal voltage needs to be
the different importance of the information transmitted from controlled is selected as the ‘‘leader’’. The ‘‘leader’’ should
different microgrids isn’t considered in the weight setting. have the function of real-time voltage measurement at the
It limits the convergence speed of the consistency variables. point of common coupling (PCC). Generally its capacity is
Because the reference value of microgrid capacity uti- the largest in all microgrids.
lization ratio is determined by the microgrid i in which the According to the requirements of the distributed control
voltage of the connection point can be measured in each method, the communication connections need to be estab-
iteration of calculation, the information transmitted by the lished among the distributed agents in different microgrids.
microgrid i is more important. The weight factor z is proposed What is more, the minimum number of communication links
to increase the weight of the communication link connected should meet the requirements that communication topology
to the microgrid i. The wight z needs to meet the following is a connected graph.
formula (9) in order to reduce the volatility of iteration. As for the microgrids shown in Fig. 4, MG1 is chosen
as the ‘‘leader’’, which means MG1 sets the reference value
n
X of the capacity utilization ratio in each microgrid according
zi,j lij = 0, i = 1, 2, · · · , n (9)
to the voltage of its PCC. In this condition, the information
j=1
transmitted by MG3 is less useful information for MG2 and
III. A HIERARCHICAL COLLABORATIVE OPTIMIZATION MG4. According to the previous discussion about the setting
MODEL FOR DISTRIBUTION NETWORK of weight setting, the weights of the communication line from
The influencing factors of the voltage stability in distribution MG1 to MG2 and from MG1 to MG4 should be increased
network with high permeability DG mainly come from the and the weights from MG3 to MG2 and from MG3 to MG4
uncertainty of the loads and the renewable energy output [28]. should be reduced.
A bi-level optimization model for the distribution network is The calculation of the elements in the improved communi-
established in this study, which is aimed at the voltage stabil- cation link weight matrix W (k) is shown as follows.
ity and operation economy in distribution network as shown zi,j lij
in Fig. 3. The upper-level model is made up of distribution wi,j = n (10)
P
network and distributed agent, the lower-level is made up of zi,j lij
j=1
distributed agent and microgrid.
The connection between the bi-level optimization model The matrix form is written for the iterative process of
is established by the distributed agent, and each distributed microgrid capacity utilization ratio as follows.
agent has a communication link with other distributed agents.
ϕ(k + 1) = W (k)ϕ(k) + ωϕ ref (k) (11)
In the upper-level control, the distributed agents are responsi-
ble for calculating the power output of the respective micro- where k represents the number of iterations;
grid based on the distribution network voltage, and in the ω represents the iteration step;
lower-level control, the distributed agents are responsible for ϕ = [ϕ1 , ϕ2 , ϕ3 , ϕ4 ]T represents the matrix of microgrid
distributing the power output of each device in the microgrid capacity utilization ratio matrix;
ref
based on the principle of lowest economic cost according to ϕ ref = [1ϕ1 ,0,0,0]T represents the reference matrix of
the results of the upper-level control. microgrid capacity utilization ratio.

VOLUME 9, 2021 11801

 Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

The calculation pressure gets relief in the distributed con- The cost of renewable energy output is expressed by the
trol because the calculation amount of the central controller cost of abandoning wind and solar and the calculation is
is distributed to each microgrid. Under this condition that shown as follows.
the communication topology diagram is connected, the dif- fore
ferent distributed agents can reach convergence, even if some CRES (PRES (t)) = γ (PRES (t) − PRES (t)) (15)
communications are interrupted. It extremely improves the
where PRES (t) represents the power output of wind turbine
reliability of the system.
and photovoltaic at the time t;
Meanwhile, the new electrical device to work only needs fore
PRES (t) represents the forecast power output of renewable
to establish the communication connection with any device
energy at the time t;
that is already running on the network. It does not need
γ represents the penalty term of abandoning wind and
to upload the information to the central controller and the
solar.
amount of information exchange is small, which reduces the
requirements of the communication network and provides
2) RESTRICTIONS
convenience for the DG’s ‘‘plug and play’’.
The internal constraints in the microgrid can be divided into
B. LOWER-LEVEL MICROGRID MODEL equality constraints and inequality constraints.
The distributed agent in the lower-level model needs to Active power balance constraint is shown as formula (16).
track the output instructions issued by the upper-level control The distributed agent in the microgrid distributes the output of
model and the output of each device in the microgrid is the internal DG and the energy storage device after the output
optimally distributed to achieve the lowest economic cost at instruction from the upper control is received.
the same time. PRES (t) + Pfuel (t) + PESS (t) − Pload (t) = Ptotal (t) (16)
1) OBJECTIVE FUNCTION where Ptotal (t) represents power output command issued by
The cost of the microgrid includes the cost of abandoning upper control at the time t;
wind and solar, the cost of energy storage devices, the cost of Pload (t) represents the internal load consumption power of
fuel generators as the output instructions issued by the upper internal load in the microgrid at the time t;
level is tracked. The specific description is as follows. PRES (t), Pfuel (t), PESS (t) represent the power output of
T
X renewable energy, fuel generators and energy storage devices
minC = [CESS (PESS (t)) + Cfuel (Pfuel (t)) in the microgrid at the time t respectively.
t=1 DG operation constraints is shown as formula (17). The
+CRES (PRES (t))] (12) types of DG in the microgrid include renewable energy gen-
erators (wind turbines and photovoltaics) and fuel generators.
where CESS represents the cost of energy storage device; For the former, the output should not exceed its predicted
Cfuel represents the cost of fuel generator; output. Of course, the latter should not exceed its maximum
CRES represents the cost of wind and solar energy. power output [31].
The operating cost of the energy storage device includes
fore

the cost of the operation and maintenance, depreciation cost 0 ≤ Pwind (t) ≤ Pwind (t)

of the energy storage unit, excluding the charging cost, which fore
0 ≤ Ppv (t) ≤ Ppv (t) (17)
can be equivalent to the quadratic function of the charge and 
0 ≤ Pfuel (t) ≤ Pfuel,max

discharge power [30]. The quadratic function is shown as
follows. fore fore
where Pwind (t) and Ppv (t) represents the predicted power
1
CESS (PESS (t)) = αP2ESS (t) (13) output of wind turbines and photovoltaics at the time t respec-
2 tively; Pfuel,max represents maximum power output of fuel
where PESS (t) represents the power output of the energy generator.
storage device at the time t; The change of the output power for the fuel generator
α represents the cost coefficient of energy storage. should meet the limit of the ramping power, which can be
The operating cost of a fuel generator is related to the shown as follows.
active power of the generator output and the calculation of
max max
the operating cost for a fuel generator is shown as follows. −1Pd_
fuel 1T ≤ Pfuel (t + 1)−Pfuel (t) ≤ 1Pr_
fuel 1T (18)
Cfuel (Pfuel (t)) = aP2fuel (t) + bPfuel (t) + c (14) max max
where 1Pd_ fuel (t) and 1Pr_
fuel (t) represents the maximum
where Pfuel (t) represents the power output of the fuel gener- downhill power and maximum uphill power for the fuel gen-
ator at the time t; erator respectively.
a, b, c represents the quadratic coefficient, the monomial Operational constraints of energy storage devices are
coefficient, the constant term coefficient of the fuel generator shown as formula (19) and (20). The operation constraints
cost respectively. of energy storage device include the constraints of maximum

11802 VOLUME 9, 2021

Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

charge and discharge power and the constraints of state of function as follows.
charge (SOC). T
" #
X CESS (PESS (t)) + Cfuel (Pfuel (t))
minC = fore
( +CRES (PRES (t)) + λ PESS (t) − PESS (t)
max max
−Pc_
ESS ≤ PESS (t) ≤ Pd_
ESS
t=1
(19) (21)
SOCmin ≤ SOC(t) ≤ SOCmax
SOC(t + 1) where λ represents the power penalty term of energy stor-

PESS (t)1T age device charge and discharge, the larger the value of λ,
SOC(t)(1 − σ ) −
 PESS (t) > 0 the closer the optimization result to the global optimization
= ηd Q (20)
P (t)1T ηc result, otherwise the closer to the local optimization result;
SOC(t)(1 − σ ) − ESS
 PESS (t) < 0
Q P fore ESS(t) represents the charging and discharging power
of the energy storage device at the time t in the day-ahead
max max scheduling plan.
where Pc_
ESS and Pd_ESS represents the maximum charging At last, the constraint term of SOC is added to the con-
power and the maximum discharging power of energy storage straint condition as shown in formula (22).
device respectively;
σ represents the self-discharge rate of energy storage SOC(t) ≤ β · SOC(t)fore (22)
device;
ηc and ηd represents energy storage device charging effi- where β represents the constraint term of SOC;
ciency and discharging efficiency of energy storage device SOC(t)fore represents the value of SOC at the time t in the
respectively; day-ahead scheduling plan.
Q represents the capacity of energy storage device; The day-ahead scheduling plan can be calculated by using
SOCmin and SOCmax represents the limit of SOC in energy the forecast data of the load and renewable energy in the
storage device respectively. previous 1h class according to formula (12)-(20). The smaller
the term λ, the higher the local economy. However, SOC will
reach the extreme value easily. Similarly, the larger the term
IV. SOLUTION OF THE BI-LEVEL MODEL
λ is, the larger the proportion of global economy is. But the
A. CHOICE OF THE BI-LEVEL MODEL SOLVING METHOD
energy storage device cannot play the role of peak shaving
The voltage stability and operating economy in distribu- and valley filling. By adding the setting of β, the amount
tion network can be guaranteed separately by establishing of charge and discharge at the time of per unit is limited,
the bi-level control model. For the voltage stability, the which could avoid the fluctuation of SOC frequently and
distributed agent obtains the update of the correct capacity extend the service life of energy storage devices. Meanwhile,
utilization ratio in the microgrid through the consensus algo- the full-day charging and discharging for the energy storage
rithm in the upper-level control. In the lower-level model, the device considers the global economy and the local economy.
distributed agent allocates the output of renewable energy, The variability and randomness of generator output are
fuel generators and energy storage devices in the microgrid greater in distribution network with high penetration rate DG
according to the principle of lowest economic cost under the compared with the traditional distribution network containing
command of the upper-level control, which belongs to the renewable energy [32]. Although wind-solar hybrid systems
quadratic programming problem and can be solved efficiently can reduce the fluctuations of power output, traditional gen-
by commercial solver such as Cplex or Gurobi. erators may not be able to adjust the power output of the
microgrid smoothly due to power ramping and output power
B. BI-LEVEL MODEL SOLUTION BASED ON MODEL restrictions. Thus, the dispatch commands cannot be executed
PREDICTIVE CONTROL rightly. MPC is the algorithm which is model-based closed-
In the control of the lower-level model, the small fluctua- loop optimization control with rolling optimization in a finite
tions of power output for renewable energy can be smoothly time domain [13]. Its core consists of three parts: predictive
adjusted by the energy storage device, but the output of the model, rolling optimization and feedback correction. The
fuel generator is required to supplement for the short-term ability of the microgrid to execute the commands from the
severe fluctuations. The output of the energy storage device upper-level output can be improved by the joining of MPC
should refer to the day-ahead scheduling plan and reserve a and the operating pressure for the traditional generators can
certain adjustment margin. Thus, it could avoid over-charge be reduced.
and over-discharge in a short time caused by pursuing the The forecast data of renewable energy is updated in
lowest cost. every 15 minutes generally. The forecast error of renewable
Therefore, several improvements should be made based on energy shows a decreasing trend with the shortening of the
formula (12) in the actual control process. Firstly, the upper forecast time scale and the update of forecast data [33]. There-
limit of wind turbine and photovoltaic output is changed to fore, 1 hour and 15min are chosen as the prediction horizon
the latest forecast data. Then the power penalty term of the and as the control horizon respectively to carry out rolling
energy storage charge and discharge is added in the objective optimization control for the internal devices in the microgrid.

VOLUME 9, 2021 11803

 Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

FIGURE 6. Structure of IEEE-33.

FIGURE 5. Schematic diagram of model predictive control.

The fluctuation characteristics of renewable energy is consid-

ered so as to eliminate the deviation of the daily output plan
from the actual operation because of the large forecast error.
The working diagram of MPC is shown in Fig. 5.
The output command issued by the distributed agent
in the upper-level control can be regarded as feed-
back correction in MPC, which changes Ptotal (t) in for-
mula (16) according to the real-time voltage in distribution
network.
The rolling optimization is realized through the MPC in
the lower-level model to convert the control problem into an
FIGURE 7. Microgrid composition.
optimization problem for solution. Meanwhile, the optimal
control sequence in the current control horizon is obtained,
which can make the fuel generator or energy storage device
prepare in advance for the processing of renewable energy
volatility in the future. Therefore, the impact caused by DG
power output and climbing speed limitations is avoided on
executing the plan of power output in the microgrid. It ensures
that the obtained sequence of optimal control has better sta-
bility and robustness.

V. CASE ANALYSIS
A. CASE SETTINGS
In this study, the improved IEEE-33 system is taken for anal-
ysis. The upper limit of the voltage in distribution network
is 1.05p.u. and the lower limit is 0.95p.u. The microgrids are
connected to the nodes 11, 19, 26, 28 respectively. The power
factor of each microgrid is 0.9. The node 11 is selected as
the node which the voltage is controlled. The communica- FIGURE 8. Renewable energy and distribution network load data. (a) the
tion link between distributed agents in different microgrids data of wind turbine (blue line: predict value; orange line: true line).
(b) the data of photovoltaic (blue line: predict value; orange line: true
is represented in the green dotted line as shown in Fig. 6. line). (c) the data of renewable energy (blue line: predict value; orange
The composition of each microgrid is shown in Fig. 7. The line: true line). (d) the data of load (blue line: predict value; orange line:
true line).
day-ahead forecast and actual data of the power output for the
load and renewable energy in distribution network are shown
in Fig. 8.
It can be seen from Fig. 8(a), Fig. 8(b), Fig. 8(c) of the total power output for photovoltaic and wind tur-
that the power output of wind turbine and photovoltaic bine is 0.3108. Wind-solar hybrid system can reduce the
have complementary characteristics in time. According to volatility of renewable energy output, which is conducive
the actual output data of wind turbines and photovoltaics, to achieving high penetration rate access of renewable
it can be calculated that the dispersion coefficient of the energy.
power output for wind turbine and photovoltaic are 0.4490 The parameters of the devices in microgrid are shown in
and 1.2416 respectively, while the dispersion coefficient TABLE 1.

11804 VOLUME 9, 2021

Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

TABLE 1. Device parameters in microgrids.

FIGURE 10. Voltage of the controlled node.

FIGURE 9. Capacity utilization ratio in the microgrid.

B. SIMULATION ANALYSIS OF UPPER-LEVEL MODEL

The improved weight matrix of the communication link in the FIGURE 11. Voltage of the controlled node.
microgrid W (k) is set as follows.
 
0.5 0.25 0 0.25
 0.45 0.5 0.05 0 
W (k) =  0

0.05 0.45 0.5 
0.25 0 0.25 0.5
The load data and the forecast data of renewable energy
at 10:00 in distribution network are taken for simulations.
The microgrid connected to the node 11 has the function of
real-time voltage measurement. Therefore, the node 11 is set
as the node in which the voltage is controlled. The initial
capacity utilization ratio for each microgrid increases from
10%. The change of the capacity utilization ratio and the
controlled nodal voltage during the iterations are shown in
Fig. 9 and Fig. 10 respectively. FIGURE 12. Microgrid capacity utilization ratio.
It can be seen from Fig. 9 and Fig. 10 that the capacity
utilization ratio of the different microgrid finally converges to
the same value, and the controlled nodal voltage is stable to In the operation of actual distribution network, most of
1p.u. through information exchange between the distributed the conditions are the load fluctuations and all microgrids
agents in the different microgrid. The value of microgrid connected to the distribution network cannot start with a
capacity utilization ratio is lower after stable convergence lower capacity utilization ratio at the same time. According
because the microgrid capacity refers to the sum of the pre- to Fig. 11 and Fig. 12, the microgrid capacity utilization ratio
dicted power output for renewable energy and the maximum could reach a new balance with a little change and the voltage
power output for the fuel generators in this research. of the controlled node can restore to 1p.u. quickly while
The load 100kW is added in the 7th node in the distribution the load fluctuates in distribution network. The simulations
network while the iteration is 100 times. The change of micro- verify that the proposed consensus algorithm can calculate
grid capacity utilization ratio and the voltage changes of the the output of each microgrid quickly and keep the voltage to
controlled node are shown in Fig. 11 and Fig.12 respectively. be 1p.u. in the controlled node.

VOLUME 9, 2021 11805

 Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

FIGURE 15. Total output in each microgrid.

FIGURE 13. The weight of communication link is same.

Meanwhile the number of times that the microgrid capac-

ity utilization ratio converges to the same value increases
from 59 times to 180 times while the unidirectional fail-
ure occurs in the communication link between the micro-
grid 28 and the microgrid 11.

C. SIMULATION ANALYSIS OF LOWER-LEVEL MODEL

According to formula (12)-(20), each distributed agent calcu-
lates the output of planned power in the four microgrids based
on the day-ahead forecast data of the load and renewable
energy in the upper-level model and the results are shown in
Fig. 15.
Moreover, the distributed agent in each microgrid dis-
FIGURE 14. The communication link between the microgrid 28 and the tributes the power output according to the output plan and
microgrid 11 fails. the results are shown in Fig. 16. The setting of power output
for the energy storage device needs to refer to the data of the
energy storage device in the day-ahead scheduling plan for
The data transmitted by the microgrid which is connected the economic cost in microgrid operation throughout the day.
to the controlled voltage node to the other microgrid is more Thus, the calculation of power distribution for each device
accurate, because the reference value of capacity utilization in the microgrid should be based on formula (21) and (22).
ratio is calculated by the distributed agent in this microgrid. The penalty term λ is set to 10 and the constraint term β is
The weight of the communication link between the microgrid set to 0.5. It’s can be seen that the energy storge device is
connected to the node 11 and the other microgrids is increased charging while the fuel generator is discharging at 22:00 in
in the setting of the weight matrix W (k) of communication Fig. 16. This is because the wind turbine output power rises
link. The weights of the communication links between the significantly at 22:00, and the fuel generator cannot drop to
different microgrids are set to be the same in reference [34] the calculated output power due to the maximum downhill
and the changes of the capacity utilization ratio in each power restrictions. It needs to rely on energy storage devices
microgrid are shown in Fig. 13 under the same conditions as to absorb the excess power.
Fig. 9. According to Fig. 8(c), the power output of renewable
According to the consensus algorithm, the capacity utiliza- energy has dropped sharply at 14:00. This phenomenon will
tion ratio still will reach a same value if the communication become more evident with the increase of renewable energy
topology is also a connected graph. The changes in the capac- penetration rate. As described in Fig. 17, the microgrid isn’t
ity utilization ratio for the different microgrids are shown in able to execute the commands of the upper-level power output
Fig. 14 while the communication connection between some owing to the restrictions on the operation of fuel generators
microgrids fails. and energy storage devices.
According to Fig. 9 and Fig. 13, the convergence speed The microgrid can make preparation in advance according
is significantly increased in the improved consensus algo- to the forecast trend of the output for renewable energy
rithm. The iteration times drop from 78 times to 59 times, through the rolling optimization in the MPC model. Thereby,
which shows that the improved consensus algorithm is more the operating pressure of fuel generators and energy storage
suitable for real-time adjustment of the microgrid output. devices get relief. The adverse effects on control from the
According to Fig. 14, the convergence still can be achieved. randomness of power output for renewable energy becomes

11806 VOLUME 9, 2021

Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

FIGURE 16. Arrangements for the day-ahead dispatch of each microgrid. (a) the dispatch of microgrid 11(blue bar: battery; red bar: renewable energy;
yellow bar: fuel generator). (b) the dispatch of microgrid 19(blue bar: battery; red bar: renewable energy; yellow bar: fuel generator). (c) the dispatch of
microgrid 26. (d)the dispatch of microgrid 28(blue bar: battery; red bar: renewable energy; yellow bar: fuel generator).

FIGURE 17. Microgrid 11 active power shortage. FIGURE 18. The output of each device in the microgrid 11.

weaken. The results of the power output control in the micro- renewable energy and prolong the service life for the fuel
grid 11 through the MPC rolling optimization are shown in generator.
Fig. 18. The actual power output and the planned power output
It is shown that the fuel generator can make a judgment in for the renewable energy, energy storage device and fuel
advance to change its power output according to the future generator in the four microgrids are shown in Fig. 19.
trend of renewable energy by adding MPC rolling optimiza- It can be seen from Fig. 19 that the power outputs of
tion. It also can lessen the impact of microgrid executing of the energy storage device and the fuel generator follow the
upper-level commands due to the limitation of the climbing day-ahead scheduling arrangement to a certain extent and the
power of the fuel generator especially at 14:00. The power economics of the whole day is considered. Meanwhile, the
output of the energy storage device can be flexibly adjusted MPC rolling optimization reduces the impact of the random-
around 18:00, which can make up the small fluctuations of ness of the power output of renewable energy on the total

VOLUME 9, 2021 11807

 Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

FIGURE 19. The actual output of each device in the microgrid and the planned output. (a) the data of microgrid 11(blue dotted line, blue line, red
dotted line, red line, green dotted line, green line are the output of battery(day-ahead), battery, renewable energy(day-ahead), renewable energy,
fuel generator(day-ahead) and fuel generator respectively). (b) the data of microgrid 19(blue dotted line, blue line, red dotted line, red line, green
dotted line, green line are the output of battery(day-ahead), battery, renewable energy(day-ahead), renewable energy, fuel generator(day-ahead)
and fuel generator respectively). (c) the data of microgrid 26(blue dotted line, blue line, red dotted line, red line, green dotted line, green line are
the output of battery(day-ahead), battery, renewable energy(day-ahead), renewable energy, fuel generator(day-ahead) and fuel generator
respectively). (d) the data of microgrid 28(blue dotted line, blue line, red dotted line, red line, green dotted line, green line are the output of
battery(day-ahead), battery, renewable energy(day-ahead), renewable energy, fuel generator(day-ahead) and fuel generator respectively).

FIGURE 20. 24h distribution of voltage at each node before optimization FIGURE 21. 24h distribution of voltage at each node after the
of distribution network. optimization in distribution network.

voltage is shown in Fig. 21 after the bi-level model is adopted

power output in the microgrid. Fuel generators can make for the optimization.
preparation in advance before large fluctuation of renewable According to Fig. 20, the access of high-permeability DG
energy happens. It is conducive to the safe and stable opera- causes the out-of-limit voltage in distribution network. How-
tion in the microgrid system. ever, the proposed bi-level model of control optimization can
effectively reduce voltage fluctuations and improve voltage
D. COMPARISON OF THE EFFECT OF BI-LEVEL stability as shown in Fig. 21. It proves the effectiveness of
OPTIMIZATION MODEL the bi-level model of control optimization.
The 24h distribution of the voltage at each node in the dis-
tribution network is shown in Fig. 20 without the control VI. CONCLUSION
of power output for the wind turbines and photovoltaic in The method of DG cooperative control is introduced into the
the distribution network. The 24h distribution of each node operation control for multiple microgrids and a novel bi-level

11808 VOLUME 9, 2021

Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

model of control optimization is proposed. It is applied to the [13] S. Vazquez, J. Rodriguez, M. Rivera, L. G. Franquelo, and M. Norambuena,
distribution network containing high-penetration renewable ‘‘Model predictive control for power converters and drives: Advances and
trends,’’ IEEE Trans. Ind. Electron., vol. 64, no. 2, pp. 935–947, Feb. 2017.
energy. Combined with the simulation results, it can be found [14] H. Zhang, T. Xu, X. Ma, M. Wang, H. Yang, and T. Chen, ‘‘Novel online
that the convergence speed of the microgrid capacity utiliza- real-time scheduling decision-making approach with adjacent time-point
tion ratio can be accelerated in the upper-level control, and the rolling,’’ Power Syst. Technol., vol. 44, no. 9, pp. 3355–3360, 2020.
[15] Z. Zhang and M.-Y. Chow, ‘‘Convergence analysis of the incremental cost
target of voltage control can be achieved faster by increasing consensus algorithm under different communication network topologies
the weight of the communication link in the microgrid which in a smart grid,’’ IEEE Trans. Power Syst., vol. 27, no. 4, pp. 1761–1768,
Nov. 2012.
the nodal voltage is controlled and the effective informa- [16] L. Xi, L. Zhang, Y. Xu, S. Wang, and C. Yang, ‘‘Automatic generation
tion is transmitted. The collaborative control based on the control based on multiple-step greedy attribute and multiple-level alloca-
consensus algorithm can still achieve the control goal while tion strategy,’’ CSEE J. Power Energy Syst., early access, Nov. 20, 2020,
doi: 10.17775/CSEEJPES.2020.02650.
the communication link fails or the load fluctuates. Thus the [17] Z. Huang, Y. Zhang, F. Zheng, J. Lin, X. An, and H. Shi, ‘‘Day-
stronger robustness is verified. In the lower-level control, the ahead and real-time energy management method for active distribution
distributed agents calculate the best distribution for power networks based on coordinated optimization of different stakeholders,’’
Power Syst. Technol., early access, Sep. 9, 2020, doi: 10.13335/j.1000-
output that all-day economy is considered according to the 3673.pst.2020.0410.
output command and the microgrid cost function. The MPC [18] H. Yang, L. Zhu, F. Gao, J. Fan, and H. Wen, ‘‘Measurement and analysis of
rolling optimization is adopted to reduce the impact of high passive intermodulation induced by additional impedance in loose contact
coaxial connector,’’ IEEE Trans. Electromagn. Compat., vol. 61, no. 6,
proportion of the output randomness of renewable energy on pp. 1876–1883, Dec. 2019.
power output of the microgrid. It also decreases the difficulty [19] W. Ren and R. W. Beard, ‘‘Consensus seeking in multiagent systems
under dynamically changing interaction topologies,’’ IEEE Trans. Autom.
of microgrid regulation. The foreseeability is brought for the Control, vol. 50, no. 5, pp. 655–661, May 2005.
fuel generator operation plan. Compared with the distribu- [20] Y. Xu and Z. Li, ‘‘Distributed optimal resource management based on the
tion network without control, the bi-level optimization model consensus algorithm in a microgrid,’’ IEEE Trans. Ind. Electron., vol. 62,
no. 4, pp. 2584–2592, Apr. 2015.
can reduce voltage fluctuations and improve voltage stability [21] Y. Li, Z. Xu, L. Xiong, G. Song, J. Zhang, D. Qi, and H. Yang,
effectively. ‘‘A cascading power sharing control for microgrid embedded with wind
and solar generation,’’ Renew. Energy, vol. 132, pp. 846–860, Mar. 2019.
REFERENCES [22] S. Baros and M. D. Ilic, ‘‘Distributed torque control of deloaded wind
DFIGs for wind farm power output regulation,’’ IEEE Trans. Power Syst.,
[1] L. Xi, J. Wu, Y. Xu, and H. Sun, ‘‘Automatic generation control
vol. 32, no. 6, pp. 4590–4599, Nov. 2017.
based on multiple neural networks with actor-critic strategy,’’ IEEE
[23] R. Sitharthan, M. Karthikeyan, D. S. Sundar, and S. Rajasekaran, ‘‘Adap-
Trans. Neural Netw. Learn. Syst., early access, Jul. 14, 2020, doi:
tive hybrid intelligent MPPT controller to approximate effectual wind
10.1109/TNNLS.2020.3006080.
speed and optimal rotor speed of variable speed wind turbine,’’ ISA Trans.,
[2] H. Xin, R. Zhao, L. Zhang, Z. Wang, K. P. Wong, and W. Wei,
vol. 96, pp. 479–489, Jan. 2020.
‘‘A decentralized hierarchical control structure and self-optimizing control [24] H. Xin, Z. Qu, J. Seuss, and A. Maknouninejad, ‘‘A self-organizing strategy
strategy for F-P type DGs in islanded microgrids,’’ IEEE Trans. Smart for power flow control of photovoltaic generators in a distribution net-
Grid, vol. 7, no. 1, pp. 3–5, Jan. 2016. work,’’ IEEE Trans. Power Syst., vol. 26, no. 3, pp. 1462–1473, Aug. 2011.
[3] L. Xi, J. Chen, Y. Huang, Y. Xu, L. Liu, Y. Zhou, and Y. Li, ‘‘Smart [25] M. N. Kabir, Y. Mishra, G. Ledwich, Z. Xu, and R. C. Bansal, ‘‘Improv-
generation control based on multi-agent reinforcement learning with the ing voltage profile of residential distribution systems using rooftop PVs
idea of the time tunnel,’’ Energy, vol. 153, pp. 977–987, Jun. 2018. and battery energy storage systems,’’ Appl. Energy, vol. 134, no. 1,
[4] Y. Xu and Z.-G. Wu, ‘‘Distributed adaptive event-triggered fault-tolerant pp. 290–300, Dec. 2014.
synchronization for multiagent systems,’’ IEEE Trans. Ind. Electron., [26] B. Zhang, A. Y. S. Lam, A. D. Dominguez-Garcia, and D. Tse, ‘‘An optimal
vol. 68, no. 2, pp. 1537–1547, Feb. 2021. and distributed method for voltage regulation in power distribution sys-
[5] L. Xi, L. Yu, Y. Xu, S. Wang, and X. Chen, ‘‘A novel multi-agent DDQN- tems,’’ IEEE Trans. Power Syst., vol. 30, no. 4, pp. 1714–1726, Jul. 2015.
AD method-based distributed strategy for automatic generation control of [27] B. Xie, C. Qi, H. Ben, and W. Yu, ‘‘The applications of graph theory in
integrated energy systems,’’ IEEE Trans. Sustain. Energy, vol. 11, no. 4, electric network,’’ in Proc. Int. Conf. Sens., Diag., Prognostics, Control
pp. 2417–2426, Oct. 2020. (SDPC), Aug. 2019, pp. 780–784.
[6] Y. Kuriki and T. Namerikawa, ‘‘Formation control with collision avoidance [28] H. Yang and H. Wen, ‘‘TDR prediction method for PIM distortion in loose
for a multi-UAV system using decentralized MPC and consensus-based contact coaxial connectors,’’ IEEE Trans. Instrum. Meas., vol. 68, no. 12,
control,’’ in Proc. Eur. Control Conf. (ECC), Jul. 2015, pp. 3079–3084. pp. 4689–4693, Dec. 2019.
[7] S. Zhang and Y. Cao, ‘‘Consensus in networked multi-robot systems via [29] H. Wen, H. Yang, H. Kuang, X. Qin, and G. Cai, ‘‘Global threshold predic-
local state feedback robust control,’’ Int. J. Adv. Robotic Syst., vol. 16, no. 6, tion of multicarrier multipactor with time distribution and material coeffi-
Nov. 2019, Art. no. 172988141989354. cients,’’ IEEE Trans. Electromagn. Compat., vol. 60, no. 5, pp. 1163–1170,
[8] G. Chen and Z. Zhao, ‘‘Delay effects on consensus-based distributed Oct. 2018.
economic dispatch algorithm in microgrid,’’ IEEE Trans. Power Syst., [30] C. Wu, S. Lin, C. Xia, and L. Guan, ‘‘Distributed optimal dispatch of
vol. 33, no. 1, pp. 602–612, Jan. 2018. microgrid cluster based on model predictive control,’’ Power Syst. Technol.,
[9] H. J. LeBlanc, H. Zhang, X. Koutsoukos, and S. Sundaram, ‘‘Resilient vol. 44, no. 2, pp. 530–538, 2020.
asymptotic consensus in robust networks,’’ IEEE J. Sel. Areas Commun., [31] D. Xu, Q. Wu, B. Zhou, C. Li, L. Bai, and S. Huang, ‘‘Distributed
vol. 31, no. 4, pp. 766–781, Apr. 2013. multi-energy operation of coupled electricity, heating, and natural gas
[10] M. Brenna, E. De Berardinis, L. Delli Carpini, F. Foiadelli, networks,’’ IEEE Trans. Sustain. Energy, vol. 11, no. 4, pp. 2457–2469,
P. Paulon, P. Petroni, G. Sapienza, G. Scrosati, and D. Zaninelli, Oct. 2020.
‘‘Automatic distributed voltage control algorithm in smart grids [32] Y. Xu, Y. Gao, Z. Li, and M. Lu, ‘‘Detection and classification of power
applications,’’ IEEE Trans. Smart Grid, vol. 4, no. 2, pp. 877–885, quality disturbances in distribution network based on VMD and DFA,’’
Jun. 2013. CSEE J. Power Energy Syst., vol. 6, no. 1, pp. 122–130, 2019.
[11] L. Xu, X. Ruan, C. Mao, B. Zhang, and Y. Luo, ‘‘An improved optimal [33] H. Wang, Y. Liu, B. Zhou, C. Li, G. Cao, N. Voropai, and
sizing method for Wind-Solar-Battery hybrid power system,’’ IEEE Trans. E. Barakhtenko, ‘‘Taxonomy research of artificial intelligence for deter-
Sustain. Energy, vol. 4, no. 3, pp. 774–785, Jul. 2013. ministic solar power forecasting,’’ Energy Convers. Manage., vol. 214,
[12] Y. Wang, S. Zhao, Y. Xu, and J. Yin, ‘‘Rolling dispatch of Jun. 2020, Art. no. 112909.
wind/thermal/storage system based on chance constrained goal [34] J. Le, Q. Zhou, C. Wang, and X. Li, ‘‘Research on voltage and power
programming,’’ Power Syst. Technol., vol. 41, no. 1, pp. 187–193, optimal control strategy of distribution network based on distributed col-
2017. laborative principle,’’ Proc. CSEE, vol. 40, no. 4, p. 1249, 2020.

VOLUME 9, 2021 11809

 Y. Xu et al.: Research on the Bi-Level Optimization Model of Distribution Network Based on Distributed Cooperative Control

YANCHUN XU received the Ph.D. degree from PING WANG received the bachelor’s degree
the Department of Electrical Engineering, Harbin in electrical engineering from Northeast Electric
Institute of Technology (HIT), Harbin, China, Power University, Jilin, China, in 2002.
in 2010. He worked with China Three Gorges Univer-
She worked with China Three Gorges Uni- sity, Yichang, China. His research interests include
versity, Yichang, China. Her research interests power system relay, and distribution network oper-
include power quality detection with distributed ation and control.
generation, harmonics detection of power sys-
tems, as well as matric converter applied in power
systems.

JIN ZHANG received the bachelor’s degree in MI LU (Senior Member, IEEE) received the M.S.
electrical engineering from Shandong Agricultural and Ph.D. degrees from Rice University, in 1984
University, Tai’an, China, in 2019. He is currently and 1987, respectively, both in electrical and com-
pursuing the M.S. degree with the College of Elec- puter engineering.
trical Engineering and New Energy, China Three She is currently a Professor with Texas A&M
Gorges University, Yichang, China. University. Her research interests include digital
His research interests include voltage stability system design and microprogrammed control of
analysis, and distribution network operation and digital systems. She is a Senior Member of IEEE
control. Computer Society.

11810 VOLUME 9, 2021