IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO.

3, SEPTEMBER 2012

1525

A Cooperative Multiagent Framework for

Self-Healing Mechanisms in Distribution Systems
Aboelsood Zidan, Student Member, IEEE, and Ehab F. El-Saadany, Senior Member, IEEE

AbstractBecause of societys full dependence on electricity and

high cost of system outages, one important goal is to increase the
reliability of the power system, which means that a salient attractive feature of smart grid is its self-healing ability. Smart grids will
develop and enhance the automation of distribution by operating
in a distributed manner through new digital technologies such as
monitoring, automatic control, two-way communication, and data
management. In this work, the smart grid concept and technologies
have been applied to construct a self-healing framework for use in
smart distribution systems. The proposed multiagent system is designed to locate and isolate faults, then decide and implement the
switching operations to restore the out-of-service loads. The proposed control structure has two layers: zone and feeder. The function of zone agents in the first layer is monitoring, making simple
calculations, and implementing control actions. Feeder agents in
the second layer are assigned to negotiation. The constraints include voltage limits, line current limits, and radial topology. Load
variation has been taken into consideration to avoid the need for
further reconfigurations during the restoration period. The results
of the simulation conducted using the new framework demonstrate
the effectiveness of the proposed control structure.
Index TermsDistributed control, distribution system, fault location, multiagent, service restoration, smart grid.

I. INTRODUCTION

HE AUTOMATION of distribution provides increased

efficiency in day-to-day operations and improves the
restoration process, which includes responding rapidly to
outages, reducing operator error, and shortening the duration
of outages [1], [2]. Detecting and isolating fault locations and
restoring service are among the important functions of distribution automation (DA) and form the cornerstone of strategies
for developing smart grids [3]. Service restoration is defined as
finding suitable backup feeders and laterals to transfer the loads
in out-of-service areas using operational criteria through a series of switching operations [4]. Restoration is achieved through
the switching operation of sectionalizes (normally closed) and
tie switches (normally opened). Different restoration methods
thus entail different configurations, which may affect service
quality. In addition, because the task of restoration is usually

Manuscript received November 24, 2011; revised March 12, 2012; accepted
April 29, 2012. Date of publication May 25, 2012; date of current version August 20, 2012. Paper no. TSG-00653-2011.
The authors are with the Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, ON N2L 3G1 Canada (e-mail:
aazidan@uwaterloo.ca; ehab@uwaterloo.ca).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSG.2012.2198247

performed under emergency conditions, time constraints can

add to the complexity of the problem [4], [5].
Currently, many distribution systems are not fully automated.
Therefore, in most cases, the corrective actions that are required
for the fault location, isolation and service restoration (FLISR)
function are performed manually by human operators. This
process may require more time than the self-healing feature of a
smart grid, which can detect faults and reconfigure the network
automatically. Therefore, one vital pillar of smart grids is
self-healing, which can be described as the quality of a system
that enables it, when subjected to a fault, to automatically and
intelligently perform corrective actions to restore the system
to the best possible state, thus enabling it to perform basic
functions without violating any system constraints. The ability
to quickly and flexibly reconfigure the network to restore the
loads previously deenergized by the fault, therefore, represents
the key component of the self-healing function. This automated
FLISR or self-healing function will: minimize the workload for
the field human operators, provide almost immediate restoration for the customers, and improve reliability and robustness
of distribution networks. The approaches currently used to
solve this problem (heuristics, expert system, meta-heuristic,
and mathematical programming) operate in a centralized way
so that a central optimization solver must read all of the system
data and then process them in order to obtain the solution [6].
The main advantage of centralized control is that it may
provide the best solution, especially for small-scale systems.
However, several challenges might impede the use of centralized control in future distribution networks: the large number
of small DG units and the increase in the level of uncertainty
due to renewable resources, electric vehicles, and variable
demand. Smart grids therefore represent a promising solution
for tackling these challenges through the development and
enhancement of distribution automation, the distributed operation of functions, and the reduction or elimination of human
intervention through the enhanced deployment of information,
two-way communication technologies, and data management.
Multiagent systems (MAS) are composed of multiple interacting computing elements, known as agents. These agents react
to changes in the environment and are capable of acting (making
decisions) in order to achieve specified goals [7]. A multiagent system can be considered the plate form of distributed processing, parallel operation, and autonomous solving. It can also
be much faster for solving discrete and nonlinear problems [8].
In the last few years, some research studies have been directed at
implementing the self-healing feature of a smart grid in distribution systems [9][15]. In [9], the application of a MAS for planning the service restoration of a distribution system was pro-

1949-3053/$31.00 2012 IEEE

1526

IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012

posed; however, the authors did not consider load shedding, load
priorities, or the obtaining of extra available capacity through
load transfers from the main backup feeders to their neighbors.
Their work was limited to service restoration and did not include the use of agents for the detection and isolation of fault
locations. In [10][15], a multiagent framework was proposed
for the self-healing of a distribution system, but, no studies have
presented a well-defined control structure, which should include
the following:
1) an appropriate control structure and operation mechanism
to be implemented in each agent;
2) appropriate coordination and communication among
agents in a distributed fashion;
3) a general procedure for a restoration algorithm that is effective for any distribution network (i.e., without systembased rules);
4) consideration of load variation, load priorities and operational constraints;
5) a proper simulation model for verifying the effectiveness
of the proposed control structure.
The work presented in this paper has the goal of including
consideration of the above points, which have not yet been
well defined in previous studies, in the design of an appropriate distributed control system for self-healing distribution
feeders. The remainder of this paper is organized as follows.
Section II presents the formulation of the restoration problem
and a discussion of related practical issues. In Sections III
and IV, the structure of the proposed multiagent control is
described. The operating mechanism of each control agent has
been defined according to the concept of intelligent agents, and
the decision makers of the control agents are coordinated using
expert-based Foundation for Intelligent Physical Agents (FIPA)
communication protocols. Section V discusses important issues
related to the practical implementation of the proposed algorithm. Section VI presents the results of a simulation conducted
in order to validate the proposed algorithm, and Section VII
provides conclusions and summarizes the main contributions
of this research.
II. RESTORATION PROBLEM IN DISTRIBUTION SYSTEMS
A. Problem Formulation
The service restoration problem can be formulated as a multiobjective multiconstraint optimization problem [16]. The proposed objectives and constraints considered are as follows:
a) Objective Functions:
1) maximizing the load restored with consideration of load
priority
(1)
: the number of energized buses after service
where
restoration; : load at the th bus; : status of the load at
the th bus (i.e., equals 1 for a restored load and 0 for an
unrestored one); and : priority or importance level of the
load at the th bus;

2) minimizing the number of switching operations and thus

reducing the time required for and the operational cost of
the restoration [4]

(2)
where : the total number of switches; : status of the th
switch in the restored network (i.e., equals 1 for a closed
one and 0 for an opened one); and
: status of the th
switch immediately after the fault has been isolated;
3) minimizing losses during the restoration period
(3)
: total number of branches; : current in the th
where
branch; and : resistance of the th branch.
b) Constraints:
1) radial network structure should be maintained;
2) bus voltages at all buses should be kept within limits
(4)
where : voltage at the th bus, and
: maximum
and minimum acceptable bus voltages (in the case study for
this research, they are 1.05 p.u. and 0.9 p.u., respectively);
3) all branch currents should be kept within their limits
(5)
where : current in branch j, and
current.

: maximum line

B. Practical Issues Related to the Service Restoration Problem

This section highlights some of the important aspects related
to the operational practices of the restoration problem, which
were used in extracting the rules for designing the expert-based
decision makers for the control agents described in Section IV.
The related issues are as follows:
For locating and isolating the fault and coordinating the
protection device, it is important to maintain the radial network topology during the service restoration process [16].
In the case of limited available capacity for restoration or
partial restoration, the restoration should begin with the
highest priority customers first (i.e., hospitals, industrial
loads, etc.) as formulated in (1) [16].
Load variation during the restoration period should be
considered during the building of the restoration plan.
Using the rated or the daily peak load may lead to the
limited restoration of some loads. Using the prefault load
may cause overloading in some backup feeders [5]. Therefore, in order not to violate the operational constraints
(i.e., voltage and current limits) and to benefit from the
available capacity for restoration (to restore as much load
as possible), the peak load over the restoration period (i.e.,
the average duration for fault repair) should be used to
build the restoration plan [5], [17]. This point is aligned
with the second objective in (2) because considering load

ZIDAN AND EL-SAADANY: A COOPERATIVE MULTIAGENT FRAMEWORK FOR SELF-HEALING MECHANISMS IN DISTRIBUTION SYSTEMS

1527

Fig. 1. Levels of backup feeders for the restoration process.

variation will prevent the need for further reconfigurations

during the future restoration hours, thus reducing the
number of switching operations.
Restoration is accomplished by transferring loads in the
out-of-service area to their neighboring backup feeders
through the on-off control of tie-sectionalizing switches.
Because the time required for the restoration process
depends on the number of switching operations [4], [16],
that number should be kept as low as possible, as formulated in (2). For example, the typical operating time
of remote-controlled and manually controlled switches
are 50 s and 12001500 s, respectively [6]. Reducing
the number of switching operations also reduces the
possibility of switching surges, the risk of outages, and
the number of transient disturbances in the system due
to multiple switching operations. The operational cost
of the switching operations is also lowered (i.e., the cost
of maintenance and dispatching technicians for nonautomated systems as well as the costs associated with the
reduced lifetime of the switches).
If the restoration plan provides a network configuration
with a topology closer to that of the prefault configuration,
it will be easier to return to the normal configuration after
the fault is cleared.
Based on the last two points, if the capacity of the backup
feeder is sufficient, the out-of-service area can be restored
as a single group through one switching operation using
one backup feeder (i.e., group restoration) [18][21].
This option minimizes the response time by reducing
the number of switching operations to one. If no backup
feeder has the capacity required for group restoration, the
out-of-service loads can be restored as multigroups (i.e.,
zone restoration). A zone indicates a segment of a distribution feeder that is bounded by two or more switches. Zone
restoration is accomplished by finding a plan to restore as
many zones as possible through suitable paths [20], [21].

If any zones are still unrestored after zone restoration,

loads can be transferred from the main backup feeders
to their neighbors in order to secure what might be a
sufficient margin for the main backup feeders to restore
the remainder of the unrestored loads, as shown in Fig. 1.
If any zones are still unrestored after the load transfer,
shedding of the least priority loads can enable the remainder of the unrestored loads to be restored with the
remaining limited capacity.
For practical purposes, the loss reduction calculated according to (3) is definitely to be used in normal operating
conditions and is not appropriate for service restoration
during an emergency situation [16], [20], [21].
III. THE PROPOSED MAS CONTROL STRUCTURE
A. Structure of the Proposed MAS
A cooperative MAS control structure is proposed for the implementation of the self-healing operation. The control agents
possess the intelligent ability to use communication and negotiation in order to determine the current and predicted states of
the system and then set their actuators and switches in a way
that will achieve their own objectives as closely as possible
while satisfying any constraints. Each distribution feeder is divided into segments (zones) based on the location of protective
devices. Thus, each section bounded by two or more switches
is a zone. To help in the restoration process when a fault occurs, feeders of the distribution networks are connected by tie
switches that are normally open. The proposed control structure
therefore has two layers: zone and feeder. The function of the
zone agents in the first layer is monitoring, making simple calculations, and implementing control actions. The feeder agents in
the second layer are assigned to negotiation. Each feeder agent
communicates only with its neighbored feeder agents which are
connected with them by tie switches. Fig. 2 shows a depiction of

1528

IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012

Fig. 2. Concept of the distributed control structure.

the concept of the intelligent control agents and their two-way

communication for the proposed control structure.
B. Coordination Among Control Agents via Communication
The Foundation for Intelligent Physical Agents (FIPA) was
originally formed in 1996 to specify software standards for
agent-based systems [22]. FIPA developed an agent communication language (ACL) for agent communications [22], [23],
which consists of 20 ACL message types (e.g., informing,
requesting, and composite speech acts). These messages have
been used in this work as a means of coordinating the proposed
intelligent control agents in order to achieve self-healing.
IV. THE PROPOSED OPERATION MECHANISM AND
COORDINATION BETWEEN CONTROL AGENTS
Fig. 3 shows the objectives and the two-way communication
among the control agents. As shown in the figure, the objectives
common to all agents are the maximization of the loads restored
and the minimization of the number of switching operations.
Each agent also has an additional objective. For example, the
initiator feeder agent (i.e., the feeder that has been subjected to
a fault at one of its components and hence has at least one out-ofservice zone) will locate and isolate the faulty zone before beginning the restoration process. The responder feeder agents
(i.e., level-1 or immediate-neighbor backup feeders of the initiator) and the subcontractor feeder agents (i.e., level-2 or immediate- neighbor backup feeders of the responder feeders) will
provide whatever capacity they have to assist with the restoration without violating their operational constraints. This section describes the operating mechanism of each agent and the
mechanism for their coordination using two-way communication during both stages of the self-healing operation (i.e., the first
stage, which is the detection and isolation of the fault location,
and the second stage, which is the restoration of the out-of-service load).

A. Fault Location Detection and Isolation Algorithm (Stage 1)

Once a permanent fault occurs in a distribution feeder, the
feeder circuit breaker is tripped in real-time operation. The
fault location detection and isolation algorithm is then applied
in order to locate and isolate the faulty section from both directions. As soon as the faulty section is isolated, the upstream
out-of-service loads are restored through the closing of the
feeder circuit breaker. A restoration algorithm is applied to
restore the downstream out-of-service loads. When the faulty
section is repaired, the reverse switching sequence is applied
so that the distribution system is returned to its normal configuration. Due to the radial topology of the distribution feeders,
the occurrence of a fault in a distribution feeder affects only
its sections [i.e., sections between the substation and the faulty
section as well as the downstream sections, when distributed
generation units (DGs) are present]. Therefore, only the control
agents of the feeder that has the faulty section will participate
at this stage. Due to the voltage potential difference, the normal
power flows being from the source to the grid. However, the
introduction of DG units may change the direction of power
flows from unidirectional to bidirectional. When a fault occurs somewhere in the distribution system, the power flow
magnitude and direction change. Fault current flows from the
substation and DG units to the lowest potential point at the fault
location. Therefore, when a fault occurs in one of the zones
between the substation and other zones that involve DG units,
the following two conditions apply [13]:
1) the fault is fed by both the substation and the DG units
in the downstream zones. The current in both boundary
breakers of this zone will thus flow into the zone;
2) the current in at least one of its breakers will exceed its
limit.
The former condition means that there is no fault outside
this zone. The latter condition always applies because the
former one can be implemented under normal conditions (i.e.,
a reverse power flow due to a high generation level produced

ZIDAN AND EL-SAADANY: A COOPERATIVE MULTIAGENT FRAMEWORK FOR SELF-HEALING MECHANISMS IN DISTRIBUTION SYSTEMS

1529

Fig. 3. Coordination via two-way communication among the control agents.

TABLE I
LOGIC CIRCUIT FOR ZONE BINARY SIGNAL AND FAULT CURRENT FLOW

from DG units in the downstream zones). On the other hand,

when a fault occurs in a zone that has no downstream zones
containing DGs, its entrance breaker current will exceed its
limit.
Based on these conditions for fault occurrence and on the proposed control structure in Fig. 3, the fault location detection and
isolation algorithm for a single fault at a time can be described
as follows:
1) Monitoring devices using direction and over-current relays
provide two signals to indicate a change in the status of

the current flow. One signal indicates that the magnitude

of the current exceeds its limit, and the other indicates the
direction of the current.
2) Zone agents utilize these signals in a logic circuit to generate simplified binary status signals (0 or 1), as shown in
Table I.
3) They then send these binary signals to their feeder agent
through inform messages.
4) The feeder agent determines which zone is the faulty zone,
as shown in Table I:

1530

IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012

if it receives a binary signal with a value of one from a

type 1 zone, it sends a request message to this zone agent
asking it to open its boundary breakers;
if it receives a binary signal with a value of one from a
type 2 zone, it sends a request message to this zone agent
asking it to open its boundary breakers;
if it receives a binary signal with a value of one from
more than one type 2 zones, it sends a request message
to the last zone agent (i.e., the zone at the feeder end
side) asking it to open its boundary breakers;
if it receives a binary signal with a value of zero, no
action is taken.

the number of switching operations as much as possible. Fig. 4

shows the overall procedure for the proposed agents. The remainder of this section describes the operating mechanism of
each agent during the restoration process.
1) Initiator Control Agent Operating Mechanism: After the
fault is isolated, the downstream zones are isolated. The affected
zones communicate with their feeder agent (initiator), as shown
in Figs. 2 and 3, in order to build a restoration plan. The details
of the proposed architecture are as follows:
1. Each zone agent in the out-of-service area sends a request
message to the initiator, including its load demand and
priority

B. Service Restoration Algorithm (Stage 2)

After the faulty section is isolated, the upstream out-of-service loads are restored through the closing of the feeder circuit
breaker. The restoration algorithm is applied in order to restore
the downstream out-of-service loads. The restoration problem
is a combinatorial problem of large scale due to many combinations of possible switching operations that increase with
the increase of networks dimensions. For this reason, most of
the published work didnt apply global optimization algorithms
(i.e., metaheuristic and mathematical programming) to solve the
restoration problem. Expert-based algorithms are preferred over
global optimization algorithms to determine restoration plans
[20] for two main reasons: i) while global optimization algorithms may arrive at high quality solutions, they take an inappropriately long time to reach a solution, hence, slow specially for realistic distribution networks [16], [20]. Furthermore,
a well-designed expert-based algorithm may arrive at equally
good solution and the benefit from the computational burden of
global techniques may not always be realized [20]. ii) Global
optimization algorithms propose the final configurations only.
In contrast, expert-based algorithms provide a feasible sequence
of switch operations to guarantee that implementing the restoration plan will not cause unwanted effects on the network [20].
Therefore, in this work, an expert-based decision-making algorithm has been used to govern the control agents. The rules
have been extracted from the practical issues related to the service restoration problem, as presented in Section II of this paper.
The restoration plan has two primary objectives: 1) to provide
as much service as possible to the customers with consideration
of their priorities (1); and 2) to be implemented as fast as possible that can be translated into minimization of the number of
switching operations to be performed (2). Hence, based on the
problem objectives, the proposed algorithm implements a multistep procedure while avoiding visits to infeasible solutions such
that: 1) the outage loads are restored as a single group if there is
supporting feeder that has enough capacity; 2) if group restoration is not valid, zone restoration is implemented by restoring as
many zones as possible through suitable paths; 3) load transfer
from the main backup feeders to their neighbors is applied to
restore the remainder of the unrestored loads; and 4) load shedding of the least priority loads is applied to restore the remainder
of the unrestored loads with the remaining limited capacity.
In this way, the proposed algorithm builds its restoration plan
with: reenergizing as much as possible of the out-of-service
loads with the consideration of their priorities and minimizing

(6)

2.

3.

4.

5.

: weighting factor and load demand of cuswhere

tomer in zone , respectively, and : total number of
customers in zone .
The initiator control agent starts negotiations using a contract net protocol (CNP) [23] by sending call for proposal
(CFP) messages to the responder feeder agents.
After the responder feeder agents reply with their proposal messages, which contain their available remaining
capacity (ARC), the initiator agent sends these two input
items to its decision maker. The input consists of the
load demands and priorities from the out-of-service zone
agents, and the ARCs from the responder agents.
The decision maker component in the initiator agent uses
expert-based rules that have been extracted from the practical issues explained in Section II along with the input it
has received in order to determine its output.
The initiator agent compares the maximum ARC with the
total demand from the out-of-service zones
(7)

where : load demand of zone

: the total number of
out-of-service zones; and
: the total number of responder agents.
6. If (7) is satisfied, the initiator decides to initiate group
restoration by restoring all out-of-service zones through
one switching operation.
7. The actions of the initiator agent are therefore to send an
accept-proposal message to the responder agent that has
the highest ARC and to send a request message to its zone
agent that is the neighbor of the selected backup feeder
asking it to close its tie switch for the completion of the
restoration process.
8. If (7) is not satisfied, the initiator decides to initiate zone
restoration by building a zone/switch relationship table
[9]. The initiator identifies all possible out-of-service zone
combinations
as follows:
(8)
where

: a collection of

adjacent zones bounded by

ZIDAN AND EL-SAADANY: A COOPERATIVE MULTIAGENT FRAMEWORK FOR SELF-HEALING MECHANISMS IN DISTRIBUTION SYSTEMS

1531

Fig. 4. Overall self-healing procedure using the proposed agents.

two or more switches, at least one of which is a tie switch.

After the elements in
are identified, the initiator builds
its zone/switch relationship table. The information in this
table is: 1) the zones in each combination
; 2) the
bounded switches for each combination
; 3) the load
demand
in each combination
according to (9); and
4) the equivalent priority index for each combination
according to (10).
(9)

(10)
9. Based on the zone/switch relationship table and the ARCs
communicated from the responder agents, the initiator
agent searches for possible combinations of zone restoration. It compares the ARCs with the elements of
,
which are listed in descending order based on their priority indices.
10. After checking for the available restoration possibilities,
the initiator agent actions are: 1) to send accept-proposal

1532

messages to those responders that will be used in the

restorations; 2) to place tie switches between the feeder
agents that have accepted proposals and the selected combinations for the restoration in a switch-to-be-closed
list (SCL); 3) to place the bounded sectionalizing
switches of the selected combinations for restoration
in a switch-to-be-opened list (SOL) in order to satisfy the
radial constraint; and 4) to update the zone/switch relation
table.
11. It then checks to determine whether the table is empty (i.e.,
whether all zones have been restored).
12. If the table is empty, the initiator sends request messages
to the appropriate zone agents asking them to open their
sectionalizing switches that are included in the SOL list in
order to partition the outage area and then to close their tie
switches that are included in the SCL list.
13. If the table is not empty, the initiator sends request messages to the responder feeder agents that are neighbors of
the remaining unrestored zone combinations. This request
prompts these responders to start negotiations with their
neighbors (subcontractors) to find load transfers that can
provide additional ARC. The request message includes the
load demand required for the remaining unrestored zone
combinations.
14. After these responders reply with their ARC, the initiator
agent repeats steps 911.
15. If the table is empty, the initiator sends request messages
to the appropriate zone agents to open their sectionalizing
switches included in the final SOL list and then to close
their tie switches included in the final SCL list.
16. If the table is not empty, the initiator determines the necessity for load shedding of the lowest priority load (i.e.,
the lowest priority zone index) in the remaining unrestored
zone combinations. It then checks to determine whether all
zones have been restored, as in step 9.
17. The initiator agent repeats step 16 until the zone/switch relationship table becomes empty, when it then executes step
15 in order to implement the switching actions required for
the completion of the restoration process.
Fig. 5 shows the overall procedure for the operation of the proposed initiator control agent.
2) Level-1 Backup Feeder (Responder) Operating Mechanism: The available remaining capacity (ARC) of each supporting feeder represents the maximum load that can be supplied by this feeder without violation of its current and voltage
constraints. During the restoration period, higher currents will
appear only in branches between the root node (i.e., substation)
and the correspondent node at the tie switch with the affected
feeder (i.e., restoration path - - as shown in Fig. 6). Hence, the
maximum additional load, which will not lead to overloading in
the supporting feeder can be obtained by considering the current
constraint at zones of this restoration path only [24]. In addition,
due to the radial topology of distribution networks, as the load
point comes far away from the root node, its voltage drop increases, and it becomes maximum at points located at the end
of the feeder (i.e., point as shown in Fig. 6). Hence, the maximum additional load, which will not lead to unpermitted low
voltages in the supporting feeder can be obtained by consid-

IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012

ering the voltage constraint at zones including these points only

[25]. The operating mechanism of each level-1 backup feeder
(responder) agent will be as follows:
1. after the responder agent receives a CFP message from the
initiator, it starts to build its proposal (ARC);
2. it sends query messages to its appropriate zone agents
about their spare capacities and bus voltage values;
3. each zone agent replies by sending an inform message that
includes the spare capacity
of its branch and/or the
bus voltage magnitude of its bus:
(11)
where
: represents the available capacity of each
zone before it becomes overloaded and before its protection device operates;
: upper bound current
in branch ; and
: magnitude of the current flow in
branch ;
4. if any zone has more than one branch, it sends the minimum
spare capacity of its branches, and a zone with more than
one bus sends the lower voltage magnitude of its buses;
5. after the responder agent receives these replies, it calculates its ARC as follows [24]:
(12)
(13)
(14)
where : zones along the restoration path (i.e., path - in Fig. 6);
: the lowest bus voltage magnitude of
the values received from zone agents;
: minimum allowable voltage magnitude in the network (i.e., 0.9 p.u.);
: series impedance of the path between the substation and the node closest to node
on the restoration
path (i.e., part of the restoration path between root node
and node in Fig. 6). This impedance could be determined by carrying out offline simulation if the forecasted
load is available (i.e., to determine which point among the
points located at the end of the feeder will have minimum
voltage, hence, its
will be used). Another option is to
determine
for those possible points in advance and
based on the received minimum voltage value, the appropriate impedance is used. In this paper the first option was
used to determine this
: maximum spare capacity
of the restoration path without overloading (current limit
constraint);
: maximum spare capacity of the restoration path to avoid under-voltage at any node (voltage limit
constraint); and
: maximum spare capacity of the
restoration path without violating operating constraints.
6. If this spare capacity
from this supporting feeder
will be used to restore an out-of-service load
at voltage
(15)

ZIDAN AND EL-SAADANY: A COOPERATIVE MULTIAGENT FRAMEWORK FOR SELF-HEALING MECHANISMS IN DISTRIBUTION SYSTEMS

1533

Fig. 5. Procedure for the proposed initiator control agents.

to include the voltage limit

feeder will thus be

pu, the ARC for this

(16)

7.
8.

9.

10.
Fig. 6. Restoration path and

for supporting feeder.

This ARC guarantees that voltage limits will not be violated for the restored zones.
The responder sends to the initiator agent a propose message that includes this ARC.
If the responder receives an accept-proposal message from
the initiator, it replies to the initiator by sending an inform
message to indicate that it is committed to the completion
of the task.
If the responder receives from the initiator a request message for additional ARC through load transfer, it begins
negotiations by sending CFP messages to its neighboring
feeders, if available (i.e., level-2 backup feeders or subcontractor agents).
This load transfer from a level-1 backup feeder to a level-2
backup feeder would involve the responder securing

1534

IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012

Fig. 7. Procedure for the proposed responder control agent.

a margin that could enable it to restore the remaining

out-of-service zone combinations. The best amount of the
transferred load (TL) is then

load of remaining unrestored zone combinations

remaining ARC of this level-1 backup feeder

(17)

11. Due to the discrete nature and possibly the limited ARC
of the level-2 backup feeders, the TL cannot be exactly
the same as what is required. The responder thus selects
its zones to be transferred to the level-2 backup feeder as
follows:

Transferred zone

ARC of level
(18)

12. After the responder determines the zones to be transferred,

it sends a propose message to the initiator with its new
ARC.
13. If the responder receives an accept-proposal message from
the initiator, it sends a confirm message to the subcontractor agent and request messages to the appropriate zone
agents asking them to open the bounded sectionalizing
switches for the selected zones to be transferred and to
close the tie switch to complete the load transfer to the subcontractor.
Fig. 7 shows the overall procedure for the proposed responder
control agent.
3) Level-2 Backup Feeders (Subcontractors) Operating
Mechanism: The operating mechanism of the subcontractor
agents is similar to steps 17 in the mechanism of the responder
agents, and subcontractor agents also negotiate with responder
agents.

V. PRACTICAL ISSUES RELATED TO THE IMPLEMENTATION OF

THE PROPOSED SELF-HEALING ALGORITHM
In this section, issues that are related to the practical implementation of the two-way communication control algorithm are
presented.
A. Implementation of Two-Way Communication
Two-way communication is the key technology in the implementation of the smart grid concept. The appropriate design for
the physical, data, and network communication layers is still a
topic of intense debate. Several types of communication technologies can be used (e.g., fiber optics, wireless, and wire line).
The harsh and complex power system environment entails significant challenges with respect to the reliability of wireless
communications for smart-grid applications [26]. Power line
communication (PLC) may be considered a good candidate for
smart-grid applications because it is the only technology that has
a deployment cost that can be considered comparable to that of
wireless systems since the lines already exist. A power grid will
therefore be the information source and the information delivery
system [27].
B. Communication Link Failure Issue
For MAS-based self-healing, communication is essential for
the collection of enough information to enable good decisions
[10]. Thus, if communication between two agents fails for any
reason, the system should be able to continue the execution of
the algorithm. A time-out procedure is simply the expectation
of executing specific tasks within a specific maximum predetermined time interval and can be introduced for communication between agents. When the time expected for the receipt of
a message from another agent is exceeded, the receiving agent
will modify the execution of its algorithm in order to tackle the
problem. For example, in this proposed algorithm, communication among agents is limited to two levels: the first level between each zone agent and its feeder agent, and the second level

ZIDAN AND EL-SAADANY: A COOPERATIVE MULTIAGENT FRAMEWORK FOR SELF-HEALING MECHANISMS IN DISTRIBUTION SYSTEMS

1535

Fig. 8. Current implementation of the procedure with data flow between the MAS and the simulated distribution network.

between neighboring feeder agents. Because of their high numbers in practical systems, to reduce the amount of communication and hence the delay time during the negotiation process,
the algorithm does not allow zone agents to negotiate with one
another. As a backup algorithm, if one zone agent were to lose
its ability to communicate with its feeder agent, it can send its
message to its neighboring zone agent (i.e., rerouting the communication path). In consequence, the latter agent will forward
this message to their feeder agent and then forward its reply to
the former zone agent.
VI. CASE STUDIES
The proposed MAS-based self-healing algorithm was tested
on a distribution system having two substations, four feeders,
and 70 nodes [28]. The following factors were taken into
consideration:
1. Renewable DG units, such as those generating power from
wind and solar energy, are characterized as fluctuating
power sources due to changes in wind speed and solar
irradiance. Therefore, at each time interval (depending on
the accuracy of the forecasting), a zone agent that includes
this type of DG sends its feeder agent an inform message
containing the power predicted to be generated during this
time interval.
2. The typical hourly load patterns of residential, commercial,
and industrial customers are used [29] to include the priorities of the loads and to estimate accurately the loading of
zones along each feeder as well as the ARC of the supporting feeders.
3. Based on the policies of each utility regarding the number
of switching operations allowed and the forecasting accuracy with respect to renewable DG units and load demand,
the service restoration algorithm can provide an overall
plan for the whole restoration period or multiple plans (e.g.,
one plan for each time interval).

4. As mentioned in Section II of this paper, the peak load and

the lowest generated power from renewable-energy-based
DGs over each restoration plan period are used to build this
restoration plan.
5. The islanded operation of DGs is not included in this work.
Utilities have installed intelligent electronic devices (IEDs)
such as modern protective relays and recloser controls in order
to provide local measurement and control capabilities. IEDs
provide the required input and output signals for the control
system, such as switch open/close indication, live/dead voltage
indication, fault current indication, load current, and open/close
commands. The data required for the simulation were obtained
from load flow calculations in Matlab software. In this work,
the algorithm for each control agent was implemented in a Java
Agent Developing Framework (JADE), which provides the required communication and platform services, and the distribution system was modeled with Matlab. Fig. 8 shows the simulation sequence whereby the forecasted load demand and DG
generation are fed into the distribution network. This step is followed by the sending of the generated data to the MAS in JADE.
The MAS executes the algorithm, and the switching actions are
sent back to the distribution network and held until the end of
the plan time interval.
For the sake of verification and comparison, a comparison
among three scenarios (1st proposed scenario with consideration of variant load, 2nd scenario based on prefault load, and
3rd scenario based on peak load) to show the effect of load variation on restoration plans is presented. The first scenario implements the proposed algorithm so that [17], [30][32]: 1) the
peak load at each out-of-service zone over the restoration time
interval is determined; 2) for the candidate supporting feeders,
the network simulator runs hourly power flow over the restoration time interval; 3) in this way, the branch current flow and
voltage for each zone per any candidate supporting feeder are
estimated (i.e., the highest branch current and lowest voltage);
4) if all zones in a candidate supporting feeder have similar

1536

IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012

Fig. 9. Four-feeder test system with zone numbers.

load patterns, one power flow for the peaking hour will be required; and 5) these estimated currents and voltages are sent to
responder agents so that they can calculate their ARC, as explained in Section IV.
The second scenario builds its restoration plan based on the
prefault values, and then responds to any overloading through
post-restoration load management using load transfers and
shedding. The third scenario builds its restoration plan based
on the maximum known daily values of loads.
A. Simulation Study Without DG
Fig. 9 shows the system with zone numbers. Feeders F1 and
F2 are residential feeders, feeder F3 is an industrial feeder, and
feeder F4 is a commercial feeder. A fault is assumed to have

occurred at Z1 in feeder F4. Two outage periods were selected:

1) during the lightest loading period of the system (i.e., from 1
AM to 6 AM); and 2) during the highest loading period of the
system (i.e., from 5 PM to 10 PM).
1) Case 1: Fault Occurred During the Lightest Loading
Period:
a) First Scenario: Based on the proposed fault location
and isolation algorithm, the Z1 agent (ZAG1) sends an inform
message to its feeder agent (FAG4) that includes a binary signal
with a value of one. FAG4 then sends a request message to
ZAG1 to open its sectionalizing switch s1 for fault isolation,
and FCB4 will also open in response to the fault. After the
fault has been isolated, the out-of-service zones (zones 210)
start the restoration algorithm by sending request messages to

ZIDAN AND EL-SAADANY: A COOPERATIVE MULTIAGENT FRAMEWORK FOR SELF-HEALING MECHANISMS IN DISTRIBUTION SYSTEMS

1537

TABLE II
OPEN SWITCHES, LOSS, MINIMUM VOLTAGE, SWITCHING OPERATIONS, AND LOAD SHED

TABLE III
OPEN SWITCHES, LOSS, MINIMUM VOLTAGE, SWITCHING OPERATIONS, AND LOAD SHED

their FAG4. The contents of the request messages are load demand and priority index. FAG4 (initiator) starts the negotiation
process by sending CFP messages to its neighbors FAG1, FAG2,
and FAG3 (responders). FAG1, FAG2, and FAG3 send query
messages to their ZAGs to obtain their ARCs. These ZAGs
then send inform messages to their FAGs. The contents of the
inform messages are spare capacity and voltage magnitudes.
Each responder FAG calculates its ARC using these data and
sends a propose message to the initiator. FAG4 evaluates these
proposals by comparing the proposed ARCs with its required
demand. In this case, the available ARC from FAG3 is sufficient compared to that required; FAG4 thus determines to use
group restoration by closing only s4. FAG4 sends accept-proposal message to FAG3. FAG3 then reply with inform message
to indicate that it is committed to completing the task. FAG4
places switch s4 on the switch-to-be-closed list (SCL). In consequence, FAG4 sends a request to ZAG4 to close its tie switch
s4 for restoration.
b) Second Scenario: The sequence is the same as in the
first scenario. Based on the prefault data, group restoration is
implemented by closing s4. There is no constraint violation will
occur till the end of the restoration period.
c) Third Scenario: Based on the peak load data, zone
restoration and load transfer are implemented. FAG3 transfers
its load (L50) to FAG1 by opening switch s13 and then closing
switch s14. Hence, it will enable FAG4 to restore all outage
zones. FAG4 will implement zone restoration by restoring
zones (27) from FAG3, zones (89) from FAG2, and zone 10
from FAG1. The switching operations to be implemented are:
opening switches s8 and s11 to divide the outage area into three
groups; then closing switches s4, s10, and s12 to complete the

restoration process. There is no constraint violation will occur

till the end of the restoration period.
Table II shows the open switches, losses, minimum voltage,
switching operations, and the load shed during the restoration
plan [Base: 11 KV, 1 MVA].
The second scenario (prefault) has the same restoration
plan as the first scenario. Based on these results, the third
scenario provides higher voltage and lower losses because
it has been built based on the peak load. However, the third
scenario required a higher number of switching operations
(seven switching operations compared to the first scenario that
required one switching operation only), leading to the conclusion that the first scenario is a good alternative to the third one,
thereby aligning more closely with operational practices.
2) Case 2: Fault Occurred During the Highest Loading
Period: Table III shows the open switches, losses, minimum
voltage, switching operations, and the load shed during the
restoration plan for the three scenarios.
Both of the first and the second scenarios required the same
switching actions (i.e., opening s8, s11, and closing s4, s10,
s12). In the first scenario, these switching actions were implemented at the beginning of the restoration period. In the second
scenario, the switching actions had to be implemented in two
stages: opening s8 and closing s4, s12 at the beginning of the
restoration period and then opening s11 and closing s10 at 6
PM in order to relieve the overloading of feeder 1. This process
causes customers at zones Z8 and Z9 to be interrupted again
during the switching actions at 6 PM. In addition, as shown in
Table III, the first scenario provides a higher voltage and lower
losses during the restoration period. Also, the third scenario required a higher number of switching operations (7 switching op-

1538

IEEE TRANSACTIONS ON SMART GRID, VOL. 3, NO. 3, SEPTEMBER 2012

Fig. 10. Load and DG power profiles over 24 hours.

TABLE IV
OPEN SWITCHES, LOSS, MINIMUM VOLTAGE,
SWITCHING OPERATIONS, AND LOAD SHED

erations compared to the first scenario that required 5 switching

operations).
B. Simulation Study With DG
For the sake of comparison with cases 1 and 2, the outage
periods and fault location were selected to be the same as in the
previous cases. Three wind-based DG units, each with a rating
of 0.5 MW were inserted at load points L13, L27, and L40 as
shown in Fig. 9. Fig. 10 shows the load and the total generation
from these DG units over 24 hours. The sequence among agents
is the same as in the previous cases. Table IV shows the open
switches, losses, minimum voltage, switching operations, and
the load shed during the restoration plan.
The second case (i.e., fault occurred during the highest
loading period from 5 PM to 10 PM) shows the effectiveness
of DG units in supporting the ARC for responders and then in
reducing the number of switching actions. For example, in case
without DG, 5 switching operations were required, and in case
with DG, only 3 were needed.

From these case studies and based on the results shown in

Tables II, III, and IV it appears that the proposed algorithm has
the following advantages:
it can locate and isolate the faulty component;
during the mean time to repair the fault, it provides a
restoration plan that includes the following:
1) consideration of variable demand and generation;
2) consideration of load priorities;
3) satisfaction of operational constraints (voltages and
currents do not violate their limits);
4) provision of the minimum number of switching operations to reduce operational costs and to avoid further switching that causes interruption to some customers, as occurred in the second scenario, as shown
in Table III;
5) benefits of supporting power derived from DG units
It can be concluded that:
Scenario 1 is more effective than scenario 2 because the
former results in a restoration plan that may involve a lower
number of switching operations and that does not require
further reconfigurations and customer interruptions.
Scenario 1 is more effective than scenario 3 because the
former results in a restoration plan that involves a lower
number of switching operations and it avoids unnecessary
load shedding that may be required by scenario 3. Therefore, scenario 1 is aligning with the operational practices.
VII. CONCLUSION
In this paper, a proposed distributed control structure for the
self-healing of smart distribution systems has been presented.

ZIDAN AND EL-SAADANY: A COOPERATIVE MULTIAGENT FRAMEWORK FOR SELF-HEALING MECHANISMS IN DISTRIBUTION SYSTEMS

The proposed control structure consists of two main types of

controllers: zone and feeder. The operating mechanism of each
controller has been designed based on the concept of a multiagent system. An expert-based decision maker has been proposed for each agent in order to achieve its objectives and satisfy
its constraints. Coordination via two-way communication protocols has been proposed as a means of achieving effective cooperation among agents and of implementing both the detection
and isolation of a fault location and the restoration objectives.
For the purposes of verification of the proposed algorithm, a
simulation model was developed. The results show that cooperation among agents through two-way communication provides
a good solution for fault location and isolation and for building
an effective restoration plan. This proposed algorithm and the
implementation of the procedures for the agents could enable
the concept of a self-healing smart grid to be employed in future distribution systems.
REFERENCES
[1] R. Prez-Guerrero, G. Heydt, N. Jack, B. Keel, and A. Castelhano, Optimal restoration of distribution systems using dynamic programming,
IEEE Trans. Power Del., vol. 23, no. 3, pp. 15891596, Jul. 2008.
[2] S. Kazemi, M. Fotuhi-Firuzabad, M. Sanaye-Pasand, and M. Lehtonen,
Impacts of automatic control systems of loop restoration scheme on
the distribution system reliability, IET Gener. Transm. Distrib., vol.
3, no. 10, pp. 891902, 2009.
[3] X. Mamo, S. Mallet, T. Coste, and S. Grenard, Distribution automation: The cornerstone for smart grid development strategy, in Proc.
IEEE Power Energy Soc. General Meeting, Calgary, AB, Canada,
2009, pp. 16.
[4] W. Chen, Quantitative decision-making model for distribution system
restoration, IEEE Trans. Power Syst., vol. 25, no. 1, pp. 313321, Feb.
2010.
[5] M. Tsai, Development of an object-oriented service restoration expert
system with load variations, IEEE Trans. Power Syst., vol. 23, no. 1,
pp. 219225, Feb. 2008.
[6] T. Nagata and H. Sasaki, An efficient algorithm for distribution network restoration, in Proc. IEEE Power Eng. Soc. Summer Meeting,
Vancouver, BC, Canada, 2001, vol. 1, pp. 5459.
[7] P. Li, B. Song, W. Wang, and T. Wang, Multi-agent approach for service restoration of microgrid, in Proc. 5th IEEE Conf. Indust. Electron. Appl., Taichung, Taiwan, 2010, pp. 962966.
[8] J. Wu, T. Lee, C. Lu, and S. Su, An autonomous decision approach for
fault allocation and service restoration in electrical distribution systems
by multi agent system, in Proc. 9th Int. Conf. Hybrid Intell. Syst.,
Shenyang, China, Aug. 2009, pp. 8994.
[9] M. Tsai and Y. Pan, Application of BDI-based intelligent multi-agent
systems for distribution system service restoration planning, Euro.
Trans. Electr. Power, pp. 17831801, 2011.
[10] X. Yinliang and L. Wenxin, Novel multiagent based load restoration
algorithm for microgrids, IEEE Trans. Smart Grid, vol. 2, no. 1, pp.
152161, Mar. 2011.
[11] D. Ye, M. Zhang, and D. Sutanto, A hybrid multiagent framework
with Q-learning for power grid systems restoration, IEEE Trans.
Power Syst., vol. 26, no. 4, pp. 18, Nov. 2011.
[12] L. Liu, K. P. Logan, D. A. Cartes, and S. K. Srivastava, Fault detection, diagnostics, and prognostics: Software agent solutions, IEEE
Trans. Vehic. Technol., vol. 56, no. 4, pp. 16131622, Jul. 2007.
[13] T. Kato, H. Kanamori, Y. Suzuoki, and T. Funabashi, Multi-agent
based control and protection of power distributed systemProtection
scheme with simplified information utilization, in Proc. 13th Int.
Conf. Intell. Syst. Appl. Power Syst., Arlington, VA, Nov. 2005, pp.
4954.
[14] T. Nagata and H. Sasaki, A multi-agent approach to power system
restoration, IEEE Trans. Power Syst., vol. 17, no. 2, pp. 457462, May
2002.
[15] J. M. Solanki, S. Khushalani, and N. Schulz, A multi-agent solution
to distribution systems restoration, IEEE Trans. Power Syst., vol. 22,
no. 3, pp. 10261034, Aug. 2007.
[16] Y. Kumar, B. Das, and J. Sharma, Multiobjective, multiconstraint service restoration of electric power distribution system with priority customers, IEEE Trans. Power Del., vol. 23, no. 1, pp. 261270, Jan.
2008.

1539

[17] V. Donde, Z. Wang, F. Yang, and J. Stoupis, Short-term load forecasting based capacity check for automated power restoration of electric distribution networks, in Proc. 2010 IEEE PES Trans. Distr. Conf.
Exposition, New Orleans, LA, Apr. 2010, pp. 18.
[18] A. Zidan and E. F. El-Saadany, Service restoration in balanced and
unbalanced distribution systems with high DG penetration, in Proc.
IEEE Power Energy Soc. General Meeting, Detroit, MI, July 2011, pp.
18.
[19] W. Lin and H. Chin, A new approach for distribution feeder reconfiguration for loss reduction and service restoration, IEEE Trans. Power
Del., vol. 13, no. 3, pp. 870875, Jul. 1998.
[20] K. Miu, H. Chiang, and R. McNulty, Multi-tier service restoration
through network reconfiguration and capacitor control for large-scale
radial distribution networks, IEEE Trans. Power Syst., vol. 15, no. 3,
pp. 10011007, Aug. 2000.
[21] C. Liu, S. Lee, and S. Venkata, An expert system operational aid for
restoration and loss reduction of distribution systems, IEEE Trans.
Power Syst., vol. 3, no. 2, pp. 619626, May 1988.
[22] FIPA, [Online] [Online]. Available: http://www.fipa.org
[23] G. Weiss, Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence. Cambridge, MA: The MIT Press, 2000.
[24] R. Ciric and D. Popovic, Multi-objective distribution network restoration using heuristic approach and mix integer programming method,
Electr. Power Energy Syst., pp. 497505, 2000.
[25] K. Aoki, T. Satoh, M. Itoh, H. Kuwabara, and M. Kanezashi, Voltage
drop constrained restoration of supply by switch operation in distribution systems, IEEE Trans. Power Del., vol. 3, no. 3, pp. 12671274,
July 1988.
[26] V. Gungor, B. Lu, and G. Hancke, Opportunities and challenges of
wireless sensor networks in smart grid, IEEE Trans. Indust. Electron.,
vol. 57, no. 10, pp. 35573564, Oct. 2010.
[27] S. Galli, A. Scaglione, and Z. Wang, For the grid and through the grid:
The role of power line communications in the smart grid, Proc. IEEE,
vol. 99, no. 6, pp. 9981027, Jun. 2011.
[28] D. Das, Reconfiguration of distribution system using fuzzy multi-objective approach, Electr. Power, Energy Syst., vol. 28, pp. 331338,
2006.
[29] E. Lopez, H. Opazo, L. Garcia, and P. Bastard, Online reconfiguration
considering variability demand: Applications to real networks, IEEE
Trans. Power Syst., vol. 19, no. 1, pp. 549553, Feb. 2004.
[30] M.-Y. Huang, C.-S. Chen, and C.-H. Lin, Innovative service restoration of distribution systems by considering short-term load forecasting
of service zones, Electr. Power Energy Syst., vol. 27, no. 56, pp.
417427, Jun.Jul. 2005.
[31] K. Aoki, K. Nara, M. Itoh, T. Satoh, and H. Kuwabara, A new algorithm for service restoration in distribution systems, IEEE Trans.
Power Del., vol. 4, no. 3, pp. 18321839, Jul. 1989.
[32] S. Devi, D. P. S. Gupta, and S. Sargunaraj, Optimal restoration of
supply following a fault on large distribution systems, in Proc. Int.
Conf. Adv. Power Syst. Contr., Operation Management, Hong Kong,
Nov. 1991, vol. 2, pp. 508513.
Aboelsood Zidan (S11) was born in Sohag, Egypt,
in 1982. He received the B.Sc. and M.Sc. degrees
from Assiut University, Assiut, Egypt, in 2004 and
2007, respectively, both in electrical engineering. He
is currently working toward the Ph.D. degree in electrical and computer engineering at the University of
Waterloo, Waterloo, ON, Canada.
His research interests include distribution automation, renewable DG, distribution system planning,
and smart grids.

Ehab F. El-Saadany (M01SM05) was born in

Cairo, Egypt, in 1964. He received the B.Sc. and
M.Sc. degrees in electrical engineering from Ain
Shams University, Cairo, Egypt, in 1986 and 1990,
respectively, and the Ph.D. degree in electrical engineering from the University of Waterloo, Waterloo,
ON, Canada, in 1998.
He is currently a Professor in the Department of
Electrical and Computer Engineering, University
of Waterloo. His research interests are distribution
system control and operation, power quality, distributed generation, power electronics, digital signal processing applications
for power systems, and mechatronics.