McArthur, S. D. J. and Davidson, E. M. and Catterson, V. M. and Dimeas, A. L. and Hatziargyriou, N. D.
and Ponce, F.A. and Funabashi, T. (2007) Multi-agent systems for power engineering applications -
part 1: concepts, approaches and technical challenges. IEEE Transactions on Power Systems, 22 (4).
pp. 1743-1752. ISSN 0885-8950
http://strathprints.strath.ac.uk/
26472/
Strathprints is designed to allow users to access the research output of the University
of
Strathclyde. Copyright © and Moral Rights for the papers on this site are retained by the
individual authors and/or other copyright owners. You may not engage in further
distribution of the material for any profitmaking activities or any commercial gain. You
may freely distribute both the url (http://strathprints.strath.ac.uk
) and the content of this
paper for research or study, educational, or not-for-profit purposes without prior
permission or charge. You may freely distribute the url (http://strathprints.strath.ac.uk)
of the Strathprints website.
Any correspondence concerning this service should be sent to The
Strathprints Administrator: eprints@cis.strath.ac.uk
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 4, NOVEMBER 2007 1
Multi-Agent Systems for Power Engineering
Applications—Part 1: Concepts, Approaches, and
Technical Challenges
S. D. J. McArthur, E. M. Davidson, V. M. Catterson, A. L. Dimeas, N. D. Hatziargyriou, F. Ponci, T. Funabashi
Abstract—This is the first part of a 2-part paper that has
arisen from the work of the IEEE Power Engineering Society’s
Multi-Agent Systems (MAS) Working Group.
Part 1 of the paper examines the potential value of MAS
technology to the power industry. In terms of contribution, it
describes fundamental concepts and approaches within the field
of multi-agent systems that are appropriate to power engineering
applications. As well as presenting a comprehensive review of the
meaningful power engineering applications for which MAS are
being investigated, it also defines the technical issues which must
be addressed in order to accelerate and facilitate the uptake of
the technology within the power and energy sector.
Part 2 of the paper explores the decisions inherent in engi-
neering multi-agent systems for applications in the power and
energy sector and offers guidance and recommendations on how
MAS can be designed and implemented.
Index Terms—Multi-agent systems
I. INTRODUCTION
F
OR over a decade the proposed use of multi-agent sys-
tems (MAS) to address challenges in power engineering
has been reported in IEEE transactions and conference papers.
MAS technology is now being developed for a range of
applications including diagnostics [1], condition monitoring
[2], power system restoration [3], market simulation [4], [5],
network control [6], [7] and automation [8]. Moreover, the
technology is maturing to the point where the first multi-agent
systems are now being migrated from the laboratory to the
utility, allowing industry to gain experience in the use of MAS
and also to evaluate their effectiveness [1].
Nevertheless, despite a growing awareness of the technol-
ogy, some fundamental questions often arise from other re-
searchers and, in particular, industrial partners when discussing
multi-agent systems and their role in power engineering. These
are: what benefits are offered by multi-agent systems? What
differentiates them from existing systems and approaches? To
what sort of problem can they be applied?
If and when MAS technology is deemed appropriate for a
particular power engineering application, then other questions
naturally follow: how should multi-agent systems be designed?
S. D. J. McArthur, E. M. Davidson, and V. M. Catterson are with the
Institute for Energy and Environment, University of Strathlcyde, Glasgow,
UK (email: s.mcarthur@eee.strath.ac.uk).
A. L. Dimeas and N. D. Hatziargyriou are with the Power Division of the
Electrical and Computer Engineering Department of the National Technical
University of Athens, Greece.
F. Ponci is with the Electrical Systems Department, University of South
Carolina, USA.
T. Funabashi is with the Meidensha Corporation, Japan.
How should multi-agent systems be implemented? Are there
any special considerations for the application of MAS in power
engineering?
The IEEE Power Engineering Society’s (PES) Intelligent
System Subcommittee (within the PSACE Committee) has
formed a Working Group to investigate these questions about
the use of multi-agent systems. Its first remit is to define
the drivers for and benefits gained by the use of multi-agent
systems in the field of power engineering. As MAS are a rela-
tively new technology, a number of technical challenges need
to be overcome if they are to be used effectively. The Working
Group’s second remit is to identify and disseminate details of
those challenges. Its third and final remit is to provide technical
leadership in terms of recommendation and guidance on the
appropriate use of the standards, design methodologies and
implementation approaches which are currently available.
This paper reports on the research of the Working Group. It
begins by describing key concepts and approaches associated
with multi-agent systems. As a result of research and discus-
sions by the Multi-Agent Systems Working Group, definitions
of MAS terminology and concepts have been tailored for use
by the power engineering community.
The engineering drivers behind the use of MAS and the
benefits they may offer are presented. The recent increase
in activities in this area has led to some inappropriate uses
of the technology; hence it considers the principal problems
which can be tackled by MAS. Comparisons with existing
technologies, such as web services, grid computing and intel-
ligent systems techniques are drawn to illustrate how MAS
differ.
Additionally, this part of the paper (part 1) presents a
comprehensive review of the power engineering applications
for which MAS technology is being investigated, and outlines
the key technical issues and research challenges which the
authors believe need to be addressed if MAS technology is to
be deployed within the power industry.
The uptake of multi-agent systems has increased over the
last few years, in terms of number of research projects. How-
ever, it is essential at this stage of maturity of research into the
application of MAS that appropriate standards and guidance
are available for those developing multi-agent systems in the
power engineering community; these are discussed in the
companion Part 2 paper.
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 4, NOVEMBER 2007 2
II. CONCEPTS: TERMINOLOGY AND DEFINITIONS
In order to explore the potential benefits of MAS to power
engineering and the areas where their application may be
justified, the basic concepts and approaches associated with
multi-agent systems need to be understood. This leads us to a
basic but essential, and unfortunately difficult, question: what
is an agent?
A. The definition of Agency
The computer science community has produced myriad
definitions for what an agent is [9]–[13]. The fact that so many
different definitions exist, testifies to the difficulty in defining
the notion of agency. A comparison of these definitions and
discussion of their relative merits and weaknesses, from a
computer science perspective, can be found in [14].
While all the definitions referenced above differ, they all
share a basic set of concepts: the notion of an agent, its
environment, and the property of autonomy. Wooldridge’s
basic definition of an agent [13] echoes that of Russell and
Norvig [9] and Maes [10]. According to Wooldridge an agent
is merely “a software (or hardware) entity that is situated
in some environment and is able to autonomously react to
changes in that environment.”
The environment is simply everything external to the agent.
In order to be situated in an environment, at least part of
the environment must be observable to, or alterable by the
agent. The environment may be physical (e.g. the power
system), and therefore observable through sensors, or it may
be the computing environment (e.g. data sources, computing
resources, and other agents), observable through system calls,
program invocation, and messaging. An agent may alter the
environment by taking some action: either physically (such as
closing a normally-open point to reconfigure a network), or
otherwise (e.g. storing diagnostic information in a database
for others to access).
The separation of agent from environment means that agents
are inherently distributable. Placing copies of the same agent
in different environments will not affect the reasoning abilities
of each agent nor the goals it was designed to achieve; rather,
the specific actions taken by each may differ due to different
observations from the two environments. This means that an
agent can operate usefully in any environment which supports
the tasks the agent intends to perform.
Under Wooldridge’s definition, an entity situated in an
environment is only an agent if it can act autonomously in
response to environmental changes. Autonomy is a somewhat
elusive term, used in all definitions of agency, but rarely
defined. The loosest definition of autonomy says that an agent
“exercises control over its own actions” [14], meaning that it
can initiate or schedule certain actions for execution. Russell
and Norvig go further, by requiring the scheduling of actions
to be in response to some change in the environment, and not
simply the result of the agent’s in-built knowledge [9]. This
requirement for environmental change is in agreement with
Wooldridge, and makes intuitive sense; can an agent really
be considered autonomous if it takes action at times prede-
fined by the agent designer, regardless of external changes in
circumstance? Autonomy is therefore the ability to schedule
action based on environmental observations.
From an engineering perspective this definition is problem-
atic: it does not clearly distinguish agents from a number of
existing software and hardware systems. Arguably, under the
definition above some existing systems could be classed as
agents. For example, a protection relay could be considered
as an agent. It is situated in its environment, i.e. the power
system. It reacts to changes in it environment, i.e. changes to
voltage or/and current. It also exhibits a degree of autonomy.
Similar arguments can be made for software systems such as
Unix daemons and virus checkers.
Renaming existing systems or new systems built using
existing technologies as “agents” offers nothing new and
no concrete engineering benefit. While Russell and Norvig
[9] argue that “The notion of an agent is meant to be a
tool for analyzing systems, not an absolute characterization
that divides the world into agents and non-agents”, being
able to distinguish agent systems from existing systems is
important. There is a need to know how agents and multi-agent
systems differ from existing systems and system engineering
approaches. Moreover, it is the potential advantages gained
through these differences that interest us as power engineers
and that have motivated the exploration of the application of
MAS to power engineering problems.
B. Definition of an Intelligent Agent
In order to help differentiate MAS from existing systems the
authors have adopted the definition of agency as proposed by
Wooldridge [13]. Wooldridge extends the concept of an agent,
given above, to that of an intelligent agent by extending the
definition of autonomy to flexible autonomy. An agent which
displays flexible autonomy, i.e. an intelligent agent, has the
following three characteristics:
• Reactivity: an intelligent agent is able to react to changes
in its environment in a timely fashion, and takes some
action based on those changes and the function it is
designed to achieve.
• Pro-activeness: intelligent agents exhibit goal directed
behavior. Goal directed behavior connotes that an agent
will dynamically change its behavior in order to achieve
its goals. For example, if an agent loses communication
with another agent whose services it requires to fulfill its
goals, it will search for another agent that provides the
same services. Wooldridge describes this pro-activeness
as an agent’s ability to “take the initiative”.
• Social ability: intelligent agents are able to interact with
other intelligent agents. Social ability connotes more than
the simple passing of data between different software and
hardware entities, something many traditional systems do.
It connotes the ability to negotiate and interact in a coop-
erative manner. That ability is normally underpinned by
an agent communication language (ACL), which allows
agents to converse rather than simply pass data.
While an agent, in terms of our earlier definition, and
many existing systems display the characteristic of reactiv-
ity, in order to be classed as an intelligent agent under
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 4, NOVEMBER 2007 3
Wooldridge’s definition, an agent must also have some form
of pro-activeness and some form of social ability. It is the
goal-directed behavior of individual agents and the ability to
flexibly communicate and interact that set intelligent agents
apart.
Not only do the characteristics of reactivity, pro-activeness
and social ability help us distinguish agents from traditional
hardware and software systems, it is from these characteristics,
as shall be discussed in the following sections, that many of
their benefits are derived.
C. The definition of a Multi-Agent System
A multi-agent system is simply a system comprising two or
more agents or intelligent agents. It is important to recognize
that there is no overall system goal, simply the local goals of
each separate agent. The system designer’s intentions for the
system can only be realized by including multiple intelligent
agents, with local goals corresponding to sub-parts of that
intention.
Depending on the definition of agency adhered to, agents
in a multi-agent system may or may not have the ability
to communicate directly with each other. However, under
Wooldridge’s definitions, intelligent agents must have social
ability and therefore must be capable of communication with
each other.
For the sake of this paper the authors have focused on
MAS where this communication is supported. This clearly
differentiates the type of MAS discussed in this paper from
other types of systems.
III. THE POTENTIAL BENEFITS OF MAS TECHNOLOGY
AND DRIVERS FOR ITS USE IN POWER ENGINEERING
APPLICATIONS
To answer the question of how (and why) MAS may be
applied in power engineering requires an understanding of the
basic ways MAS can be exploited. In this paper the authors
have called these “approaches”.
To date MAS have a tendency to be exploited in two
ways: as an approach to building flexible and extensible
hardware/software systems; and as a modeling approach.
A. MAS as an approach to the construction of robust, flexible,
and extensible systems
There are many power engineering application areas for
which flexible and extensible solutions are beneficial.
Flexibility connotes the ability to respond correctly to
dynamic situations, and support for replication in varied situ-
ations (environments). This sounds very similar to autonomy
and therefore intelligent agents should automatically be flex-
ible; but if autonomy is the ability of an agent to schedule
its own actions, flexibility relates to having a number of
possible actions from which to select the most appropriate.
Some specific examples of flexible behavior would be correct
handling of different formats of one type of data (such as
temperatures in Centigrade or Fahrenheit); or the ability to
construct a new plan if a particular control action fails; or a
system that can be deployed on any feeder, which senses the
connection of distributed generation and changes protection
settings accordingly.
Extensibility connotes the ability to easily add new func-
tionality to a system, augmenting or upgrading any existing
functionality. For example, a condition monitoring system
may gain a new type of sensor, and require a new data
analysis algorithm. A state-estimator system may be upgraded
to use a faster load-flow calculation algorithm. For distribution
networks, a distributed network control and management sys-
tem responsible for voltage control may be extended to also
automate restoration and the management of distributed gen-
eration. Importantly, a truly extensible system will allow new
functionality to be added without the need to re-implement the
existing functionality.
Across many applications in power engineering there is also
a requirement for fault tolerance and graceful degradation:
should part of the system fail for whatever reason, the system
should still be able to meet its design objective or, if that is not
possible, it should accomplish what it can without interfering
with other systems.
MAS can provide a way of building such systems. Indeed,
the ability of MAS to be flexible, extensible, and fault tolerant
is often part of the justification for their use. However, in
order for that justification to be valid, the way in which MAS
provide flexibility, extensibility, and fault tolerance needs to be
understood. The properties of agents and MAS that produce
these qualities are examined below.
1) Benefits of autonomy and agent encapsulation: An agent
encapsulates a particular task or set of functionality, in a
similar way to modular or object-oriented programming. This
means that the benefits of standard interfaces and information-
hiding are also available with agent programming through the
use of messaging with a standard agent communication lan-
guage, but there is also the additional capability of autonomous
action.
Recall that autonomous action means each agent is able
to schedule its own activity in order to achieve its goals. In
a modular programming situation, external modules can call
functions which the module has no choice but to execute. With
agent programming, external agents can only send messages
requesting the agent take some action: the autonomous agent
can decide whether to fulfill the request, the priority of the
task, and if other actions should also be scheduled. This can be
useful in situations when an agent is receiving many requests
and cannot fulfill them all within a reasonable timescale, such
as with multiple requests for a processing-intensive task like
a load-flow calculation.
The autonomy of each agent and the messaging interface
are what contribute most to flexible and extensible systems.
Because agents are not directly linked to others, it is easy to
take one out of operation or add a new one while the others are
running. Any agents interacting with the stopped one can use
the standard service location facilities to locate another agent
that performs the same task, and by this mechanism new agents
can be included within the system. The agent framework
provides the functionality for messaging and service location,
meaning that new agent integration and communications are
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 4, NOVEMBER 2007 4
handled without effort from the system designer.
This allows systems to be extensible: extra functionality can
be added simply by deploying new agents, which use service
location to find others to communicate with; and parts of
systems can be upgraded by deploying a replacement agent
and removing the obsolete one. Flexibility also follows: the
appropriate mix of agents can be deployed to fit the details
of individual situations, and flexible handling of messages
between agents allows the system to self-configure. Finally,
legacy systems can be incorporated within the system simply
by wrapping legacy functionality in a layer of agent messag-
ing.
2) Benefits of open MAS architectures: An open agent ar-
chitecture places no restrictions on the programming language
or origin of agents joining the system, and allows flexible
communication between any agents. This is achievable through
adherence to messaging standards: the separation of an agent
from its environment means that the messaging language an
agent understands is important for inter-agent communication,
rather than the programming language in which it was imple-
mented.
An example of a set of standards for an open architecture
is that defined by the Foundation for Intelligent Physical
Agents (FIPA) [15]. The FIPA Agent Management Reference
Model covers the “framework within which FIPA agents
exist”, defining standards for creating, locating, removing, and
communicating with agents. This is more generally called
the agent platform, and is simply one part of an agent’s
environment. One requirement of an open agent architecture
is that the platform places no restrictions on the creation and
messaging of agents, while a second is that some mechanism
must be available for locating particular agents or agents
offering particular services within the platform. Under the
FIPA model, this is achieved through a separate agent called
the Directory Facilitator: an agent which manages a searchable
list of services offered by other agents within the platform.
Early agent systems tended to be closed architectures, as one
set of agents would be deployed every time the system was
run, with all communication explicitly defined by the system
creator. An example is the ARCHON system for distribution
network management, originally built to integrate four legacy
systems [16]. Such an architecture is said to be closed because
new agents cannot be added to the community: even if a new
agent is created and run, other agents have no way of locating
it and communicating with it. A closed architecture removes
the possibility of an extensible or flexible system, severely
limiting the benefits of using agents.
How to specifically design an open agent architecture is
discussed in detail in Part 2 of this paper.
3) Platform for distributed systems: An agent is distinct
from its environment, meaning that it can be placed in different
environments and still have the same goals and abilities.
However, the environment impacts upon which actions an
agent takes and in what order, as the agent autonomously
schedules action in response to sensor inputs and messages.
For this reason an agent is inherently distributable, having
no fixed ties to its environment. In practice, distribution of
agents across a network is supported by the agent platform: the
platform is run on every computer that will host an agent, and
the agents are deployed within the platform as usual. To agents
within one platform, there is no difference between agents on
the same computer and agents on a different computer, as
the instances of the platform running on separate machines
seamlessly connect and appear as a single instance.
This means that the same set of agents can be deployed
on one computer, and alternatively on multiple networked
computers, without modifying or changing the agent code.
4) Fault tolerance: Building redundancy into systems is
one of the standard engineering approaches to gaining fault
tolerance. Building redundancy into MAS simply involves
providing more than one agent with a given set of abilities.
If an agent needs the services of a second agent in order to
fulfill its goals, and the second agent fails, the agent can pro-
actively seek an alternative agent (perhaps using the Directory
Facilitator) to provide the services it requires.
This redundancy may be provided by simple duplication
of each agent, possibly with distribution of duplicates across
different computers. This would provide a tolerance to physical
faults, such as the loss of a network connection, or damage
to a computer. Tolerance to programming-related faults would
require a more design-intensive solution: rather than simply
running two copies of a single agent, the same functionality
would be coded differently in two agents. Various applications
and operating environments will have differing requirements
for levels of robustness and fault tolerance, and so the approach
taken must be application-specific.
However, the flexibility offered by an open architecture of
agents with good social ability easily leads to the design of a
fault tolerant system.
B. Multi-agent systems as a modeling approach
Multi-agent systems are more than a systems integration
method, they also provide a modeling approach. By offering
a way of viewing the world, an agent system can intuitively
represent a real-world situation of interacting entities, and give
a way of testing how complex behaviors may emerge.
Natural representation of the world has previously been
given as an advantage of object-oriented (OO) systems design,
where entities in a system are modeled as objects. This has
recently found favor with the power engineering community
in standards such as the Common Information Model (CIM)
[17] and IEC 61850 [18]. The main benefit of the object
approach is data-encapsulation: the particular data structures
used to hold attributes of an object are hidden from external
objects, but are indirectly accessible through method calls and
standard interfaces. Agent-based design adds another level of
abstraction to this: not only are internal data structures hidden,
but the “methods” (actions) an agent can perform are also
hidden, yet indirectly accessible through standard messaging
interfaces.
This is a very natural way of modeling actors in some
systems such as markets: in a real market actors have attributes
(such as desired price and lowest price for a seller) and
possible actions (e.g. start auction, accept bid) which other
actors cannot manipulate directly. Indirect access is available
