Electrical Power and Energy Systems 73 (2015) 8996

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

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

Improved average consensus algorithm based distributed cost

optimization for loading shedding of autonomous microgrids
Wei Liu a, Wei Gu a,, Yinliang Xu b, Shuai Xue a, Ming Chen a, Bo Zhao c, Miao Fan d
a

Southeast University, School of Electrical Engineering, Nanjing, China

Sun Yat-sen University, SYSU-CMU Joint Institute of Engineering, SYSU-CMU Shunde International Joint Research Institute, Guangzhou, China
c
Zhejiang Electric Power Test and Research Institute, Hangzhou, China
d
Siemens Industry, Inc., Schenectady, NY, USA
b

a r t i c l e

i n f o

Article history:
Received 13 July 2014
Received in revised form 25 March 2015
Accepted 18 April 2015
Available online 14 May 2015
Keywords:
Distributed cost optimization (DCO)
Load shedding (LS)
Improved average consensus algorithm
(IACA)
Global information
Synchronization processing

a b s t r a c t
This paper addresses a new distributed cost optimization (DCO) method for load shedding (LS) of an
islanded microgrid considering cost. A two-layer improved average consensus algorithm (IACA) of
multi-agent system (MAS) is proposed, and the consensus characteristic of which is analyzed in detail.
With the global information discovered in the rst layer of the IACA, the DCO of LS can be solved by using
the synchronization processing of the IACA in the second layer. PSCAD/EMTDC-based simulation models
are built to study the value settings of consensus constants and the performances of the proposed DCO
method. Simulation results veried the convergence improvement of the IACA and the effectiveness of
the proposed DCO.
2015 Elsevier Ltd. All rights reserved.

Introduction
There will be more microgrids in future power systems, including grid-connected microgrids and islanded microgrids [1,2]. An
islanded microgrid is more likely to occur a disturbance because
of the randomness and uncertainty of renewable distributed generators (DGs) and the low-inertia of the microgrid [3]. Thus a sudden change in generation or load poses a challenge for stabilizing
the frequency of an islanded microgrid. Furthermore, a large disturbance could result in a rapid frequency drop or even microgrid
blackout [4]. The best way to maintain the stability of the frequency response is to execute a load shedding (LS) scheme [5,6].
Accordingly, many studies focus on adaptive and intelligent LS
schemes to avoid the instability caused by shedding an improper
amount of load. On one hand, the magnitude of the active power
unbalance estimation method based on rate of change of frequency
(ROCOF) and system inertia constant was proposed in [710], by
using the estimated value the amount of load shedding can be
determined and an accurate and quick LS can be implemented in
a central control manner. On the other hand, a multi-agent system
(MAS) based LS is introduced to obtain adaptive LS, the control
modes of the MAS-based LS are centralized or distributed with
Corresponding author. Tel.: +86 13814005169; fax: +86 25 87796196.
E-mail address: wgu@seu.edu.cn (W. Gu).
http://dx.doi.org/10.1016/j.ijepes.2015.04.006
0142-0615/ 2015 Elsevier Ltd. All rights reserved.

coordination [11,12]. The centralized scheme requires a central

controller, which easily suffers from a failure to handle the huge
amount of data. Moreover, taking the uncertainty of intermittent
renewable energy resources into consideration, a generation uctuation may result in unintentional structure changes, which will
further increase the burden on centralized strategy. Distributed
coordination theory is realized based on consensus and cooperation theory in networked MAS, and the average consensus algorithm such as Metropolis method and improved Metropolis
method were applied to discover global information, thus a local
decision-making can be made considering the global information
[13,14]. On this basis, many distributed cooperative schemes are
proposed, in [15], the authors proposed a consensus theory based
multi-agent coordination scheme for information discovery in
microgrids via wireless networks, the Metropolis method and pairwise average was applied accordingly. In [16], the authors proposed a novel model of a distributed demand side management
mechanism that allows agents, by adapting the deferment of their
loads based on grid prices, to coordinate in a distributed manner. In
[17], the authors proposed an improved average-consensus theorem based global information discovery method and a distributed
multi-agent-based LS algorithm, which can make efcient LS decision based on discovered global information. However, the distributed scheme with a leaderfollower structure, which uses a
leader agent to gather information from all of the other local agents

90

W. Liu et al. / Electrical Power and Energy Systems 73 (2015) 8996

Nomenclature
LS
DCO
MAS
IACA
MC
DG

load shedding
distributed cost optimization
multi-agent system
improved average consensus algorithm
marginal cost
distributed generator

and sum it to obtain the global information, raises similar concerns

about system performance and reliability when possible malfunctions and attacks occur at the leader agent like central controllers.
In [18], the authors proposed a two layer consensus based fully distributed LS scheme without leader agent, advantages of a fully distributed scheme include the ability to survive uncertain
disturbances and fully distributed data updating, which leads to
efcient information sharing and eventually a faster decision-making process and operation.
Recently, the cost prioritized schemes in generation side for a
rural or islanded microgrid attracted special attention, in [19],
two cost-prioritized droop schemes for distributed generators
(DGs) were proposed; in [20], a distributed subgradient-based
coordination of multiple renewable generators considering different generation cost was proposed in a microgrid. With regard to
the LS schemes, a cost prioritized based optimal LS in customer
side also deserves more attention. Typically, the different important grades (non-vital/semi-vital/vital) [8] and the different stages
of load (three stages of load) [10] were taken into accounts during
the LS scheme, the cost of LS and the willing of customers to shed
load are also needed to be considered during the implementation
of the distributed LS [21,22] in customer side. The proposals of this
study are inspired from the multi-stage LS scheme [10] and the distributed cooperative algorithm [13,14,17,18], and the main objectives are as follows: (1) to implement a practical LS considering
the cost of LS and the customers willing to shed load; and (2) to
implement a fully distributed LS without leader or virtual leader.
This study proposed a two-layer improved average consensus
algorithm (IACA) based DCO method for LS of an islanded microgrid to overcome the mentioned shortcomings caused by leader
agents. It can implement an adaptive LS considering the cost and
marginal cost of LS and the willing of customers to LS in customer
side. Several related problems have been addressed in this study,
including the convergence improvement of consensus method,
the DCO for LS considering cost and marginal cost, and the adaptation of DCO during plug-and-play operations. Representative cases
are evaluated to testify the improvement of IACA and the effectiveness of the proposed DCO method.
The rest of the paper is organized as follows: Section Two-layer
IACA based DCO for LS introduces on the proposed two-layered
IACA based DCO of LS; Section Case studies illustrates and simulates the proposed DCO in different cases; the conclusions are duly
drawn in Section Conclusion.
Two-layer IACA based DCO for LS
In this study, the two-layer IACA of MAS is proposed to implement the distributed LS of an islanded microgrid. Compared with
a centralized scheme which coordinates its local agents by central
control agent, the proposed distributed scheme coordinates its distributed agents based on IACA based global information sharing
and the synchronization processing of IACA. For the MAS in this
study, every agent knows its corresponding power outputs and
load conditions, but does not have direct access to the global

DER
MT
WT
PV
MH

distributed energy resource

micro turbine
wind turbine
photovoltaic
mini hydro

information since it can only communicate with its immediate

neighboring agents. Thus, the main challenge with the design of
DCO and distributed LS is to discover the global information and
realize optimization through information exchange between distributed agents. According to the two-layer IACA strategy, in the
rst layer, the local agents monitor the power disturbances
throughout the microgrid and share the global information to each
agent. While in the second layer, the amount of load to be shed is
optimized in a distributed manner based on the IACA, considering
the cost of each load. Through the global information shared in the
rst layer, every agent in the second layer can make decision
locally according to the global information without the coordination of a leader. Meanwhile the DCO can be implemented by using
the synchronization processing of IACA in a fully distributed manner in the second layer.
Fig. 1 illustrates the owchart of the two-layer IACA based DCO
for LS of an islanded microgrid. The corresponding steps of the proposed DCO are described as follows:
Step 1: Firstly global information (the total active power deciency PDEF) is monitored by the corresponding faulty agent
when an active power deciency or shortage occurs in an
islanded microgrid.
Step 2: In this step, through the IACA based distributed global
information sharing, the total active power deciency PDEF of
the islanded microgrid is shared among MAS when the average
consensus is reached in the rst layer IACA.
Step 3: Ranked in the second layer of IACA, each agent evaluates
its corresponding cost and marginal cost of load shedding
locally according to the cost and marginal cost functions.
Step 4: The DCO of LS is solved by using the synchronization
processing of the IACA, and the optimal results will be reached
when the marginal cost of each load agent converges to a common value asymptotically.
Step 5: Lastly based on the results calculated by DCO in Step 4,
the distributed optimal LS can be implemented for the islanded
microgrid.
The rst layer: IACA based global information sharing
The average consensus algorithm which relies on local information of agents are used to guarantee the global information be
shared in a distributed way. To adapt for the changes of the communication topologies, we have studied ACA method [14] and
NNCA method in [18] respectively. In this study an IACA is chosen
to achieve the global information in MAS with communication
constraints and improve the convergence rate. The IACA based global information sharing and the corresponding Lyapunov stability
proof is described in Fig. 2.
In Fig. 2, pi 2 R denotes the state variable of agent i; i = 1, 2, . . .,
k

k1

n, j = 1, 2, . . ., n; k is the discrete-time index; pi and pi

are the
information shared by agent i at iteration k and k + 1 respectively,
k

pi is the information of agent j; P[k] is information sharing matrix;

wij is the coefcient for information exchanged between

91

W. Liu et al. / Electrical Power and Energy Systems 73 (2015) 8996

Start
Fsys < FLS

A1
A2

Local prediction for

the total power deficiency

An

A3
Global Power Sharing in the
first layer of IACA
First Layer IACA
Local evaluation for
the cost of load shedding

An-1

A2(2)

A1(2)

Distributed cost optimization in

the second layer of IACA
Second Layer IACA

An(2)
A3(2)
An-1(2)

Distributed optimal load

shedding
The first layer agent

The second layer agent

End
Communication Networks

3
4

Uncertain Links

Fig. 1. Flowchart of the distributed optimization method using two-layer IACA.

Fig. 2. IACA based information sharing.

neighboring agents i and j, if agents i and j are connected through a

communication line wij 0, otherwise wij = 0, W is the updating
matrix; and Ni is the indexes of agents that are connected to agent
i, ni and nj are respectively the numbers of agents in the neighborhood of agent i and j; h is the convergence constant which affects
the convergence rate of IACA.

To obtain a stable algorithm, the updating matrix and the convergence constant are needed to be properly designed. In this
study, the IACA is proposed to improve the convergence rate of
consensus, accordingly different value settings of convergence constants are compared. By the verication proof in Fig. 2, when
1 < h < 0, DL 6 0 is yielded, which implies that the stability of the

92

W. Liu et al. / Electrical Power and Energy Systems 73 (2015) 8996

proposed consensus under a communication topology can be

reached asymptotically. Note that the communication topology
must be a strong connected topology [11,12], each agent on the
communication topology cannot be isolated from the communication topology, at least one communication link must be connected
to exchange information with its neighboring agents. Depending
on the proposed method, an agent can adjust its coefcients locally
to adapt the changes of system conguration, such as plug-andplay operations of switching topologies. When a power disturbance occurs in an island microgrid, the global information (such
as n, the total magnitude of active power imbalance, and the
marginal cost of load) can be calculated and shared among the
distributed agents by IACA.

The second layer: DCO for LS

With the global information discovered in the rst layer, the
DCO is implemented in the second layer. Firstly assuming that
the load unit has a quadratic cost function [21,22] as described
in Fig. 3, the objective of the DCO is to minimize the total cost of
LS operation. The cost and marginal cost functions and the IACA
based synchronization processing of the DCO are all illustrated in
Fig. 3.
In Fig. 3, Ci and CMCi are the cost function and the marginal cost
function of the ith load respectively; K1 and K2 are constants; CTOT
is the total cost of load shedding; PLSi is the amount of load shedding of the ith load, PLS is the total amount of load shedding; PG
and PL are the total power generation and load; li is the load shedding willing variable which describing the willing to shed load.
This form of the cost function suggests that the customer with
the lowest willing have the highest marginal cost and hence the
lowest marginal benet. By following the synchronization processing of the IACA described in Fig. 3, the optimal results of DCO can
be solved when all agents converge to a common marginal cost
asymptotically.

Case studies
Microgrid architecture
To test the effectiveness of proposed approach, a PSCAD/EMTDC
based simulation system is developed. The conguration of simulated autonomous microgrid and its corresponding communication
topology are shown in Fig. 4. Note that an agent can only communicate with its immediate neighboring agents in the communication topology. The MAS-based microgrid is modeled by the
PSCAD/EMTDC and Matlab interface technologies. More specically, the MAS are simulated in Matlab, while the microgrid is
operated in PSCAD, and the user-dened interface (UDI) models
in PSCAD are dened to associate these two platforms together.
Through these interface technologies, the IACA based global information sharing and the DCO can be implemented and veried
[14,18].
The types and parameters of all lines are shown in Table 1. In
Table 1, OL is overhead line, SC is customer service line; R1 and
X1 are the positive sequence resistance and reactance, R0 and X0
are the zero sequence resistance and reactance respectively.
All the data of DERs and loads are shown in Tables 2 and 3
[23,24].
Case A: overload scenario
At the beginning of the simulation case, the microgrid works in
grid-connected mode and all DERs work in PQ mode. At t = 2 s, the
microgrid islands from the main grid, and MT operates from PQ to
V/f control mode to maintain the power balance of the islanded
microgrid. When t = 5 s, the autonomous microgrid occurs an overload accident, as a result the power balance between supply and
demand does not match at the moment and the frequency starts
to uctuate, and the LS is implemented based on the proposed
DCO when the frequency of the simulated microgrid is lower than
49.5 Hz. The simulation results are shown in Fig. 5.

The second layer

Fig. 3. Distributed cost optimization of LS.

93

W. Liu et al. / Electrical Power and Energy Systems 73 (2015) 8996

Fig. 4. Microgrid architecture.

Table 1
Lines types and parameters.
Type

R1

X1

R0

X0

OL-4  120 mm2

OL-3  70mm2 + 54.6 mm2
SC-4  6 mm2
SC-4  16 mm2
SC-4  25 mm2
SC-3  50 mm2 + 35 mm2

0.284
0.497
3.690
1.380
0.871
0.822

0.284
0.497
3.690
1.380
0.871
0.822

1.136
2.387
13.64
5.52
3.48
2.04

0.417
0.447
0.472
0.418
0.409
0.421

(1) Improvement of the global information sharing: Firstly the

proposed IACA based information sharing is used to discover
the total active power imbalance in the rst layer. As mentioned in Fig. 2, the convergence constant h has great inuence on the convergence rate of IACA. Thus in this case,
four representative values of the convergence constant h

are selected and compared to verify the improvement of

the value settings of h. As can be seen from Fig. 5(a) that,
the IACA reach consensus within fewest iterations when setting h = 0.8, and the global information can be discovered
with least communication time delay. With the global information discovered by IACA, the DCO in the second layer can
be implemented in a fully distributed manner without a
leader.
(2) DCO of LS: Secondly to implement the DCO in the second
layer, the cost and marginal cost of each load is calculated
by each distributed agent locally as described in Fig. 3. The
values of willing variables li of load shedding are shown
in Table 3. For the sake of simplicity, assuming that k1 and
k2 are 0.5 and 1 in this study respectively, note that these
assumptions do not affect the fundamental concepts to be
considered here. Observed from Fig. 5(b) that, the marginal
Table 3
Data of loads.

Table 2
Data of DERs.

Load

DER

Capacity/kVA

Pref/kW

Control mode

MT
WT
PV
MH

100
40
30
50

100
15
10
50

PQ-V/f
PQ
MPPT
PQ

Load
Load
Load
Load
Load

1
2
3
4
5

Active power/kW

Reactive power/kVar

li

5.13
22.50
54.72
5.13
22.50

2.48
10.90
15.96
2.48
10.90

0.1
0.4
0.8
0.2
0.5

94

W. Liu et al. / Electrical Power and Energy Systems 73 (2015) 8996

(a)

(b)

(c)
Fig. 5. Control performances of the proposed method in Case A.

costs of loads are asymptotically converged to a common

value after the synchronization processing of IACA based
DCO.

(3) Distributed optimal LS: Lastly depending on the optimization results from DCO, the amounts of load shedding for each
load can be derived, and the LS strategy is carried out to

W. Liu et al. / Electrical Power and Energy Systems 73 (2015) 8996

95

(a)

(b)

(c)
Fig. 6. Control performances of the proposed method in Case B.

achieve a cooperative frequency recovery. The frequency of

autonomous microgrid recovers to the rated value after the
distributed cost optimal LS, as can be observed from Fig. 5(c).
Case B: overload scenario with plug out operation
In this case, the PV and load 4 disconnects from the autonomous
microgrid and its corresponding agent 4 plugs out accordingly, as a
result the communication topology of the simulation microgrid
changes. When t = 5 s, the islanded microgrid occurs an overload,

and LS is needed to be implemented when the frequency of microgrid drops down lower than 49.5 Hz. The simulation results are
shown in Fig. 6.
Firstly the updating matrix W of IACA updates according to the
updating method described in Fig. 2. By using the proposed IACA
updating method, only the neighboring agents of agent 4 needs
to update during the plug out operation. The global information
is shared as in Case A, and four representative values of the convergence constant h are compared to nd a better setting of h.
Observed from Fig. 6(a) that, the IACA reach consensus within

96

W. Liu et al. / Electrical Power and Energy Systems 73 (2015) 8996

fewest iterations when setting h = 0.9, the global information can

be discovered with least communication time delay.
Observed from Fig. 6(b) that, the marginal costs of loads are
asymptotically converged to a new common value after the synchronization processing of DCO, thus the amounts of load to be
shed for the remaining four loads can be derived from the optimization results of DCO.
Lastly observed from Fig. 6(c), the frequency of autonomous
microgrid can recover to the rated value after the implementation
of the distributed cost optimal LS.

Conclusion
This study proposed and veried a two-layered IACA based DCO
method for LS of an islanded microgrid. Several related problems,
including the convergence improvement of the consensus method,
the DCO of load shedding considering marginal cost, and the adaptation of plug-and-play operations, have been studied in detail.
Two representative cases are simulated to illustrate the advantages
of the proposed method: (1) it is a new fully DCO by using the synchronization processing of IACA; (2) it is an accurate optimal load
shedding which takes cost and marginal cost into considerations;
and (3) it can meet the requirements for plug-and-play operations
of DGs.
Acknowledgments
This work was supported in part by the National High
Technology Research and Development Program of China (863
Program Grant No. 2011AA05A107), the National Science
Foundation of China (Grant No. 51277027), the Natural Science
Foundation of Jiangsu Province of China (Grant No.
SBK201122387) and the State Grid Corporation of China.
References
[1] Gu W, Wu Z, Bo R, Liu W, Zhou G, Chen W, et al. Modeling, planning and
optimal energy management of combined cooling, heating and power
microgrid: a review. Int J Electr Power Energy Syst 2014;54(1):2637.
[2] Pogaku N, Prodanovic M, Green TC. Modeling, analysis and testing of
autonomous operation of an inverter-based microgrid. IEEE Trans Power
Electron 2007;22(2):61325.
[3] Peas Lopes JA, Moreira CL, Madureira AG. Dening control strategies for
microgrids islanded operation. IEEE Trans Power Syst 2006;21(2):91624.

[4] Kim JY, Jeon JH, Kim SK, et al. Cooperative control strategy of energy storage
system and microsources for stabilizing the microgrid during islanded
operation. IEEE Trans Power Electron 2010;25(12):303748.
[5] Lukas S, Ignacio E, Luis R. A method for the design of UFLS schemes of small
isolated power systems. IEEE Trans Power Syst 2012;27(2):9518.
[6] Giroletti M, Farina M, Scattolini R. A hybrid frequency/power based method for
industrial load shedding. Int J Electr Power Energy Syst 2012;35(1):194200.
[7] Karimi M et al. A new approach of under-frequency load shedding technique in
an islanded distribution network. Int Rev Electr Eng 2012;7(2):415562.
[8] Karimi M, Mohamad H, Mokhlis H, et al. Under-frequency load shedding
scheme for islanded distribution network connected with mini hydro. Int J
Electr Power Energy Syst 2012;42(1):12738.
[9] Terzija VV. Adaptive underfrequency load shedding based on the magnitude of
the disturbance estimation. IEEE Trans Power Syst 2006;21(3):12606.
[10] Gu W, Liu W, Shen C, Wu Z. Multi-stage underfrequency load shedding for
islanded microgrid with equivalent inertia constant analysis. Int J Electr Power
Energy Syst 2013;46(2):369.
[11] Olfati-Saber R, Fax JA, Murray RM. Consensus and cooperation in networked
multi-agent systems. Proc IEEE 2007;95(1):21533.
[12] Zhang Z, Chow MY. Convergence analysis of the incremental cost consensus
algorithm under different communication network topologies in a smart grid.
IEEE Trans Power Syst 2011;6(7):313441.
[13] Xu Y, Liu W. Novel multiagent based load restoration algorithm for microgrids.
IEEE Trans Smart Grid 2011;2(1):15261.
[14] Liu W, Gu W, Sheng W, Meng X, Wu Z, Chen W. Decentralized multi-agent
system based cooperative frequency control for autonomous microgrid with
communication constraints. IEEE Trans Sustain Energy 2014;5(2):44656.
[15] Liang H, Abdrabou A. Multi-agent coordination in microgrids via wireless
networks
<http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.261.
2611>.
[16] Ramchurn SD, Vytelingum P, Rogers A, Jennings N. Agent-based control for
decentralised demand side management in the smart grid. In: Proceedings of
the 10th international conference on autonomous agents and multiagent
systems, international foundation for autonomous agents and multiagent
systems, AAMAS 11, Richland, SC, USA, vol. 1, 2011; p. 512.
[17] Xu Y, Liu W, Gong J. Stable multi-agent based load shedding algorithm for
power systems. IEEE Trans Power Syst 2011;26(4):200614.
[18] Gu W, Liu W, Zhu J, Zhao B, Wu Z, Luo Z, Yu J. Adaptive decentralized underfrequency load shedding for islanded smart distribution networks. IEEE Trans
Sustain Energy 2014;5(3):88695.
[19] Nutkani IU, Poh CL, Peng W, Blaabjerg F. Cost-prioritized droop schemes for
autonomous AC microgrids. IEEE Trans Power Electron 2015;30(2):110919.
[20] Xu Y, Zhang W, Liu W, et al. Distributed subgradient-based coordination of
multiple renewable generators in a microgrid. IEEE Trans Power Syst
2014;29(1):2333.
[21] Fahrioglu M, Alvarado Fernando L. Designing incentive compatible contracts
for effective demand management. IEEE Trans Power Syst 2000;15(4):
125560.
[22] Fahrioglu M, Alvarado Fernando L. Using utility information to calibrate
customer demand management behavior models. IEEE Trans Power Syst
2001;16(2):31722.
[23] Papathanassiou S, Hatziargyriou N, Strunz K. A benchmark low voltage
microgrid network. CIGRE Symposium 2005, Athens; 2005. p. 15.
[24] Rudion K, Styczynski ZA, Hatziargyriou N, et al. Development of benchmarks
for low and medium voltage distribution networks with high penetration of
dispersed generation. In: Proceedings of the international symposium on
modern electric power systems, Wroclaw; 2006. p. 11521.