IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO.

3, AUGUST 2005 1447

Operation of a Multiagent System

for Microgrid Control
Aris L. Dimeas, Student Member, IEEE, and Nikos D. Hatziargyriou, Senior Member, IEEE

AbstractThis paper presents the operation of a multiagent

system (MAS) for the control of a Microgrid. The approach
presented utilizes the advantages of using the MAS technology
for controlling a Microgrid and a classical distributed algorithm
based on the symmetrical assignment problem for the optimal
energy exchange between the production units of the Microgrid
and the local loads, as well the main grid.
Index TermsAuction algorithm, distributed generation, energy
market, microgrids, multi agent system, symmetrical assignment
problem.

I. INTRODUCTION

A Microgrid is a possible future energy system paradigm [1],

formed by the interconnection of small, modular genera-
tion (micro-turbines, fuel cells, PV, etc.), together with storage
devices (flywheels, energy capacitors and batteries) and control-
lable loads at low voltage distribution systems. Such systems Fig. 1. Control levels of the microgrid environment.
can be operated interconnected to the power grid, or islanded,
if disconnected from the grid. The operation of micro-sources Local Controllers (LC), which could be either Micro
in the network introduces considerable complexity in the oper- Source Controllers or Load Controllers.
ation of an LV grid, but at the same time, it can provide distinct
The DNO is responsible for the technical operation in a medium
benefits to the overall system performance, if managed and co-
and low voltage area, in which more than one Microgrids may
ordinated efficiently.
exist. In addition, one or more MOs are responsible for the
Significant research is currently carried out regarding opera-
Market Operation of this area. These two entities do not belong
tion and control of Microgrids [2], [3]. In this paper, a control
to the Microgrid, but they are the delegates of the grid. The DNO
system is proposed, aimed at the following:
refers to the operational functions of the system and the MO to
1) optimal use of local distributed resources; the Market functions. It should be noted that, despite the au-
2) feeding of local loads; tonomous operation of the Microgrid, it should ideally appear
3) operation simplicity. as a controlled, intelligent unit in coordination with the DNO.
It should be noted that the optimal operation of the production This is important in view of a possible increased number of such
units and the controllable loads does not refer exclusively to systems in the future.
their market participation, but also to their technical operation, The main interface between the DNS/MO and the Microgrid
since it is not possible to have constantly specialized personnel is the Microgrid Central Controller (MGCC). The MGCC is
for monitoring at this system level. the main responsible for the optimization of the Microgrid oper-
The paper presents a distributed control approach for Mi- ation, or alternatively, it simply coordinates the local controllers,
crogrids. These systems operate inside the wider distribution which assume the main responsibility for this optimization.
system and, for that reason, the overall environment is intro- The lower level of control consists of the LC. The LCs con-
duced first. trol the Distributed Energy Resources (DER), production and
Three control levels are distinguished, as presented in Fig. 1. storage units, and some of the local loads. Depending on the
Distribution Network Operator (DNO) and Market Oper- mode of operation, they have certain level of intelligence, in
ator (MO) at the level of the Medium Voltage. order to take decisions locally. Of course, in any type of opera-
Microgrid Central Controller (MGCC). tion there are certain decisions that can be taken only locally. For
example for voltage control, the LCs do not need the coordina-
Manuscript received November 8, 2004; revised March 14, 2005. This work tion of the MGCC and all necessary calculations are performed
was supported by the European Union research project Microgrids under Con- locally.
tract ENK 5-CT-2002-00610). Paper no. TPWRS-005902004. There are several levels of decentralization that can be pos-
The authors are with the National Technical University of Athens, Athens, GR
15773, Greece (e-mail: [email protected]; nh @power.ece.ntua.gr). sibly applied, ranging from centralized control to a fully de-
Digital Object Identifier 10.1109/TPWRS.2005.852060 centralized approach. According to the fully decentralized ap-
0885-8950/$20.00 2005 IEEE
1448 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 3, AUGUST 2005

proach, adopted in this paper, the main responsibility is given

to the DER controllers, which compete to maximize their pro-
duction in order to satisfy the demand and probably provide the
maximum possible export to the grid taking into account current
market prices. Furthermore, LCs should take all the appropriate
decisions to ensure safe and smooth operation of the DER they
are controlling.
The organization of a controlled, intelligent entity (whole
Microgrid) formed by several less intelligent entities (local Fig. 2. Actions sequence for the market operation in the time domain.
controllers) can be based on a multiagent system (MAS). This
was preliminary shown in a previous paper [4] of the authors.
The multiagent technology is the evolution of the classical
distributed technology with some specific characteristics that
provide new capabilities in controlling complex systems [9].
This technology has already been proposed in the operation of
large power systems, as well for the control of single machines
[5][8].
The use of distributed control in a Microgrid provides effec-
tive solutions for a number of specific operational problems. For
example, local loads and production or storage units may have
different owners and several decisions should be taken locally
and independently. Microgrids operating in a market require that Fig. 3. Power flows and bids in the microgrid.
the controller of each unit participating in the market, has a cer-
tain degree of intelligence. Moreover, the local DER not only
the local loads and the decisions of the MGCC, but performs
sell power to the network, but may have also other tasks, like:
optimal resource allocation. The DER units within the Micro-
producing heat for local installations; grid adjust their set points, after negotiation with the other units
feeding local critical loads and preserving enough energy based on the grid Market prices, their operational cost and the
or fuel to supply them for some time; local load demands. It is also assumed that there is no limit to
load-shedding capabilities; the power that can be sold or bought from the grid, since the
black start the microgrid; capacity of the grid is almost infinite compared to the local de-
ensure a seamless transition from grid connected to island mands.
mode and vice versa; The overall procedure is the following (Fig. 2).
voltage control.
The general organization of the paper is as follows: In Sec-
1. The Market Operator (MO) announces the
tion II the Market Operation of the Microgrid is presented. In
prices for selling (SP) or buying (BP) energy
to the Microgrid. Normally it is SP > BP.
Section III the Multi Agent Technology is shortly introduced
together with some of its benefits for the control of a Microgrid.
2. The local loads announce their demands for
Section IV provides the mathematical background and the al-
the next 15 minutes and an initial price (DP)
for the kWh. It is DP > BP and DP < SP.
gorithm that ensure the optimal market operation of the system,
as well some preliminary results in Section V. In Section VI,
3. The production units accept or decline the
the method is adapted to the practical problem of the Microgrid
load offer according to an Auction Price (AP)
Operation and Section VII provides a technical description of its
which is the result of the algorithm presented
MAS implementation. Section VIII describes the MAS installed
in Section IV (Fig. 3).
in the laboratory Microgrid of the National Technical University
4. The negotiation continues for a specific
of Athens. Section IX provides a comparison between the cen-
time (5 min).
tralized and decentralized approach and Section X concludes.
5. After the end of the negotiation time, all
the units have adjusted their set points. If
II. MICROGRID MARKET OPERATION there is no production unit within the Micro-
In order to analyze the control system presented in the next grid to satisfy the load demand, the power is
sections, it is necessary to describe the market model in which bought from the grid. In addition, the grid
the Microgrid is assumed to operate. It should be noted that the can be considered as a load too, so the pro-
system presented does not aim at the optimal participation of duction or storage units can sell energy to
the Microgrid in the market, but at its optimal internal opera- the grid.
tion. According to the model adopted, optimal market operation
is a task of the MGCC. The MGCC will make all the necessary For this market model, it is assumed that the distributed con-
negotiations with the Market Operator in order to achieve the trol system is an entity that tries to maximize the overall internal
best energy prices. The algorithm does not alter the demand of benefit, i.e., minimize the operational cost of the Microgrid.
DIMEAS AND HATZIARGYRIOU: OPERATION OF A MAS FOR MICROGRID CONTROL 1449

It should be noted that the proposed control system with that the agent uses the resources, skills and services character-
minor modifications can be easily adapted to other market izes its behavior. As a consequence, it is obvious that the be-
models, for example, it can deal with longer term base ne- havior of every agent is formed by its goals. An agent that con-
gotiations for base load predictions. The horizon of the time trols a battery system and its goal is to supply uninterruptible
scheduling is adaptable to the needs of the market model. power to a load will have different behavior from a similar bat-
tery, whose primary goal is to maximize profits by bidding in
the energy market.
III. MAS THEORETICAL BACKGROUND

Although there is no strict definition about what an agent is, IV. MATHEMATICAL BACKGROUND
the literature [9], [10] provides some basic characteristics: The core of the algorithm applied is based on an auction al-
The first characteristic is that an agent can be a physical en- gorithm for the solution of the symmetric assignment problem.
tity that acts in the environment or a virtual one, i.e., with no This method provides maximization of the internal benefit of
physical existence. In our case the physical entity is the agent the system. The symmetric assignment problem is formulated
that controls directly a Microturbine and a virtual one a piece of as follows.
software that makes bids to the energy market or stores data in Consider persons and objects that should be matched.
a database. There is a benefit for matching person with object . In the
An agent is capable of acting in the environment, i.e., the presented application, the benefit for each person is his revenues
agent changes its environment with its actions. A diesel gener- for obtaining object , i.e., an agreement for producing a certain
ator by altering its production changes the setpoints of the other amount of energy. The main target is to assign the persons to
local units, changes the voltage level of the adjacent buses and objects and to maximize the total benefit
in a more global point of view changes the security level of the
system [the stability of the system in case of a short circuit for
example].
(1)
Agents communicate with each other and this could be re-
garded as part of their capability for acting in the environment.
As an example, lets consider a system that includes a wind gen- The price is an algorithmic variable that is formed by the bids
erator and a battery system: the battery system uses power from of all persons and so expresses the global desire. The prices of all
the wind turbine to charge and to discharge it, in times of no objects form the price vector. These prices should not be con-
wind. In order to achieve this operation optimally, the two agents fused with the market prices. Furthermore, the difference be-
have to exchange many messages. This is a type of action be- tween the benefit and the price is the actual value of an object
cause the environment is altered in a different way by this com- for a specific person. The actual value for a specific object is
munication, rather than if the two agents were acting without different for two persons, since it is related to the benefit. At the
any kind of coordination. beginning of the iterations, the price vector is zero and so the
Agents have a certain level of autonomy, which means that actual value is equal to the benefit, although variations of the
they can take decisions without a central controller or com- proposed methods use initial non- zero values for faster conver-
mander. To achieve this, they are driven by a set of tendencies. gence.
For a battery system a tendency could be: charge the batteries In order to clarify the above terms, we consider an example
when the price for the kWh is low and the state of charge is low, with two objects and two persons that belong to a larger set
too. Thus, the MAS decides when to start charging based on of persons and objects. The first person has a benefit vector
its own rules and goals and not by an external command. In ad- , the second one has benefits
dition, the autonomy of every agent is related to the resources . Taking into account only the benefits, the first person
that it possesses and uses. These resources could be the avail- has higher benefit for the first object and the second person for
able fuel for a diesel generator. the second object. If we assume a price vector for
Another significant characteristic of the agents is that they the two objects, the actual values for the two persons are {9, 1}
have partial or none at all representation of the environment. and {6,2}.Both players desire the first object more, since both
For example, in a power system the agent of a generator knows have greater actual value for it than for the second, however the
only the voltage level of its own bus and, maybe, it can estimate second person has greater benefit for the second object, than for
what is happening in certain specific buses. However, the agent the first. It can be said that the benefit represents local informa-
does not know what is happening in the whole system. This is tion for each person and the price vector global information for
the core of the MAS technology, since the goal is to control the whole system. The price for an object increases until at most
a very complicated system with minimum data exchange and one person wants it. Increasing the price of an object is an indi-
minimum computational demands. cation that there is another person that desires this object, too.
Finally, another significant characteristic is that an agent has a The auction algorithm used calculates the price vector , in
certain behavior and tends to satisfy certain objectives using its order to satisfy the -complementary slackness condition sug-
resources, skills and services. An example of these skills could gested in [11], [12]. The steps are described next.
be the ability to produce or store power and an example for the At the beginning of each iteration, the -complementary
services could be the ability to sell power in a market. The way slackness condition is checked for all pairs of the assign-
1450 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 3, AUGUST 2005

ment. The is the object that person wants to be assigned TABLE I

to. So the formulation of this condition is RESULTS OF THE AUCTION ALGORITHM FOR THE FIRST FOUR ITERATIONS

(2)

is the set of objects that can be matched with person .This

inequality has two parts: is the actual value of object
for person , as described before. The right part refers to the
object that gives maximum value to person minus . is a
timal solution, as long as there is one. The maximum number of
positive scalar, added in the bid of each object, in order to avoid
iterations is
possible infinite iterations in case two or more objects provide
maximum benefit to the same person, as explained later.
If all persons are assigned to objects, the algorithm termi- (8)
nates. Otherwise, a nonempty subset of persons that are unas-
signed is formed. Similarly, the nonempty subset is formed and the algorithm terminates in finite number of iterations if
by the available objects. The following two steps are performed
only for persons that belong to . (9)
The first step is the bidding phase, where each person finds
an object which provides maximal value; this is
The above algorithm is further explained by considering three
persons and three objects. All bidders have zero benefit for the
(3) third object and the benefit for the rest of the objects is a constant
. So the benefits for the objects are for ,2,3
Following this, the person computes a bidding increment and and the other benefits . The initial prices
are considered zero. For this example, the first four iterations of
the algorithm are presented in Table I.
(4)
The second column of the table called Prices shows the
price of the three objects at the beginning of each iteration. The
is the best object value
column called Pairs shows the pairs (persons, object) that are
assigned at the end of the iteration. The fourth column called
(5) Bidder/ Object shows which person was the bidder at the end
of the negotiation and for which object a bid is made. Since
and the second best object value we have only three objects and three persons, there can be only
one bidder in each iteration. The last column shows the bidding
increment at the end of the iteration.
(6)
In the first iteration, all persons desire object 1, so there are
bids for this object. It should be mentioned that, in case the bids
According to the previous equations, the bidding increment is are equal, some selections are random or based on rules. The
based on the two best objects for every person. The price of an initial selection could be the second object or two persons could
object rises, if there are two or more bids for it and the price bid for the first object and the other for the second. The selected
increment is the larger bidding increment between the bids. It is strategy does not affect the convergence of the algorithm ac-
obvious that, if the scalar and the benefits for the first and cording to [11]. According to the benefits, all the persons in this
the second best object are the same, then and this leads iteration have the same increment bid, the value of the bid is
the algorithm to infinite iterations. The scalar ensures that the and the winning bidder is the last person. In the second iteration
minimum increment for the bids is . the bid increment for all persons is because
The next phase is the assignment phase, where each object
selected as best object by the nonempty subset of persons (10)
in , determines the highest bidder
In this example, the need of is clearly illustrated, because oth-
(7) erwise no price would increase. All persons desire object 1 or
2, since the benefit is for both of them. In the last itera-
Object raises its prices by the highest bidding increment tion the prices of 1 or 2 object will be greater than , object 3
, and gets assigned to the highest bidder . The receives a bid and this is the end of the algorithm.

person that was assigned to at the beginning of the iteration,

V. PRELIMINARY RESULTS
if any, becomes unassigned.
The algorithm iterates until all persons have an object as- In order to explore the performance of the algorithm, a simple
signed. It is proven [11] that the algorithm converges to the op- program was developed in a single processor machine using
DIMEAS AND HATZIARGYRIOU: OPERATION OF A MAS FOR MICROGRID CONTROL 1451

TABLE II
RESULTS OF THE AUCTION ALGORITHM FOR VARIOUS SIZES OF THE PROBLEM

random values for the benefits of the objects. The primary con-
cern was to discover the average and the maximum number of Fig. 4. Blocks of energy that form the assignment problem.
iterations needed to converge to solution. This is critical for this
application, since in the distributed control system, there is sig- As explained in Section V, it is desirable to limit the algorithm
nificant delay in each iteration. During the development and execution to 100 iterations at each negotiation, i.e., a number of
testing of the Java program, it was concluded that it is not effec- 20 to 30 blocks should be selected including the additional vir-
tive to have more than 100 iterations, since otherwise negotia- tual load. According to the problem formulation of Section IV,
tions would last several minutes or hours. It should be noted that the persons correspond to the blocks of the Available Power
in the single machine application, there are no communication and the objects to the demand blocks. The agent market oper-
delays. In a multiagent application, built over a local network or ation based on the described model is illustrated in Fig. 5. The
the Internet, the communication delay should be seriously con- Production Unit Agents control the DER, the Load Unit Agents
sidered in the development. In Table II, results from the program represent the loads and the Grid Agent generates Market Player
are presented. Agents. The Market Player Agents are virtual agents and their
The conclusion from Table II is that no more than 30 objects task is to accomplish the negotiation. There are two types of
(and 30 persons) should be used. This is enough for our appli- Market Player Agents: The Seller and the Buyer. The Buyer is
cation and detailed justification for this selection will be given the object in the assignment problem and the Seller is the person.
in the next section. Furthermore, the use of proper value for the Every Market Player Agent represents a single block of energy.
factor ensures that there will be no more than 150 iterations.
Similar to the local loads, the virtual load is represented by
This means that .
Market Player Agents that are created from the Grid Agent.
According to the proposed market model, each producer has
the ability to sell all the production to the grid and, similarly,
VI. MICROGRID MARKET OPERATION EXPRESSED AS
every load can buy energy from the grid. For this reason the
SYMMETRICAL ASSIGNMENT PROBLEM
Grid Agent finds the number of pairs of Market Player Agents
In order to describe the Microgrid market operation, two (Sellers and Buyers) and creates extra sellers and buyers. The
types of physical agents and one type of virtual agent are number of the agents is equal to Market Agents that are created
introduced. The two physical agents are the Production Unit from the Production Units Agents and the Load Unit Agents. In
Agent and the Load Unit Agent. These two agents are physical, this way, buying or selling energy from the grid is determined
because they directly control a production or storage unit and a by the algorithm.
load panel, respectively. The third type is the Grid Agent. This A major issue in Microgrids operation is the estimation of the
agent is virtual, because it cannot control the grid in any way upper limits for the demand or the available power of each DER
and just announces the prices for selling or buying energy. All for the next time interval. This should be done for each partic-
other agents, introduced later in this section, are virtual and ipant, separately. It should be noted, that although forecasting
their operation concerns the auction algorithm only. techniques are well advanced for larger interconnected systems
Let us consider that there are production units with total and typically hourly resolution times, there is little experience
capacity and Loads with total capacity . The symmetrical in forecasting with a high temporal resolution (e.g., min)
assignment problem requires that . In order to overcome with a horizon of 34 h for very small loads, like the loads in a
the problem of surplus or deficient local production, a virtual Microgrid. This happens because the load fluctuation in a small
load with proper price is added, as shown in Fig. 4. Similarly, system is very high and has no specific pattern.
virtual production can be added. The virtual load or production In this application the upper limit is defined by two methods
corresponds to the extra energy that is bought or sold to the grid. depending on the type of load or DER. The first method is to
As mentioned before, it is assumed that the grid can offer or consider that the upper limit is the nominal capacity. For units
receive infinite energy. like a Diesel Generator or a water heater this is quite realistic. On
In order to apply the algorithm for the solution of the sym- the contrary, for units like Photovoltaic Panels, Wind generators
metric assignment problem, the load should be divided into or lights, the persistent method is used, i.e., it is assumed that the
equal blocks, similar to the available production. Blocks that average energy production or demand for the next 15 min will
belong to the same load have equal benefits, since the system be the same as the current one.
will provide all the necessary power for the whole load or none. It should be noted that other functionalities of the Micro-
For example, if we consider a water heater that demands 500 grid (like security check, battery management, voltage control,
Wh for the next 15 min, the system should provide 500 Wh or etc.) can be included in this operation. For example, the offered
nothing. power of battery bank production could be reduced in order to
1452 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 3, AUGUST 2005

Fig. 6. Three types of agents for the proposed MAS.

Fig. 5. Virtual market player agents that are created for the need of the
negotiation.

maintain the state of charge status and keep certain amount of

energy to serve the system, in case of a grid black out.

VII. MAS IMPLEMENTATION

The Microgrid is a micrograph of a large power system, en-
capsulating a large amount of technical parameters and details.
Centralized control is possible in two ways: the first option is to
create a very complex model with full description of all its ele-
ments and, as a consequence, requires a communication system
capable to exchange large amount of information, as well as
complex algorithms. The second option is to create a simplified Fig. 7. Part of the proposed ontology for the MAS.
description of the system and obtain results that would probably
be suboptimal. The MAS technology allows creating a model control. Thus, all software modules of the control system per-
of the system, as detailed as possible. In this model, every agent ceive the system in the same way and this is important in order
uses the exact piece of information it needs, leaving the tech- to develop a high level communication system, where the agents
nical details for the agents that are below it in the organization exchange knowledge. In order to understand the importance
chart. It is therefore essential to have a formal way of describing of having a common perception of the environment, consider
information and giving to each agent the information it needs. the concept energy which has different meaning in quantum
In our application, the system model has three types of agents, physics, food, or electric systems. Furthermore, the object ori-
as shown in Fig. 6. ented nature of the ontology and the data abstraction, support the
Control Agents: they control directly the physical units of development of a distributed control system, since each agent
the system. handles only the necessary (or allowable) part of information
Management Agents: they manage the Microgrid and and knowledge. Fig. 7 presents a small simplified part of the
take decisions regarding the state of it. ontology concerning the loads, which shows the advantages of
Ancillary Agents: they perform ancillary services, like this architecture. The main class Ont_load provides general de-
communication tasks or data storage. tails about the whole load bank and the inherited classes have
The development of the application is in conformity with the details for each sub load, like market values or the address in the
standards proposed by the international Foundation of Intelli- Programmable Logic Controller (PLC) that controls the switch
gent Physical Agents (FIPA) [13]. This organization aims to of the specific sub load. The load unit agent needs to know and
standardize the development of such systems, especially in the handle only the information of the ont_load class and nothing
area of communication between the agents and the organization more. Furthermore, the Market Player agent needs to know the
of the MAS. For our implementation, the Java Agent DEvelop- information of the ont_market class. This architecture simplifies
ment Framework (Jade) 3.0 platform is selected [14]. Jade is a extremely much the algorithmic development of each function
Java-based tool for developing MAS systems. of the system.
One main feature of FIPA compliant MAS platforms is using The second main task is to support high level communication,
the Agent Communication Language (ACL) [15] with high- so that the agents not only exchange simple values, but also
level ontologies. It is a general conclusion that high-level com- knowledge, commands, beliefs or procedures that have to be
munication is one significant element [9], [10] to develop in- followed. For example, the agent that controls a load participates
telligence inside a society, no matter if this society is a human in the market by sending a request message to all the production
one or a MAS. For the system presented here, an ontology was units agents declaring the amount of energy that it needs. In
developed according to the needs of the system. This ontology ACL, this message has the following parts.
supports two main tasks. 1) Message Type: in our case is a request.
The first task is to provide an adequate description of the 2) Receiver: All the agents that produce energy and partici-
system, so that it includes all the necessary information for its pate in the market.
DIMEAS AND HATZIARGYRIOU: OPERATION OF A MAS FOR MICROGRID CONTROL 1453

3) Language: The agents should use the same language in

order to parse the messages correctly.
4) Ontology: The agents share the same ontology in order to
understand the content of the message.
5) Content: in our case is an object of the ontology called
energy.
The agents that participate in the market communicate by
sending proposals and accepting or rejecting proposals and this
is how the auction algorithm was implemented.
Using ontology and conversation based communication,
MAS can relatively easily deal with other functionalities of the
system, provided that the agents understand a larger ontology
and have the ability (intelligence) to handle a larger variety of Fig. 8. Laboratory microgrid in the Power System Laboratory of the National
messages. For example, if they receive a message that a black Technical University of Athens.
out occurred, they understand that it is normal to measure zero
voltage and that they should stop their participation in the
VIII. MAS IMPLEMENTATION AT A LABORATORY MICROGRID
market. After that, they should follow some predefined actions,
in order to restore the system in isolated mode and feed at least
some critical loads. In a similar way, the functionalities men- In Fig. 8, the Microgrid installed in the Power System Labo-
tioned in the introduction can be implemented in the control ratory of the National Technical University of Athens is shown.
software. This is a modular system, comprising a PV generator as the
The final issue is how agent behaviors are implemented in primary source of power. The micro sources are interfaced to
the MAS software. The JADE platform provides an excellent the 1-phase AC bus via DC/AC PWM inverters. A battery bank
library for implementing behaviors by using a basic object is included, interfaced to the AC system via a bi-directional
called Behavior. Furthermore the platform provides a number PWM voltage source converter. The Microgrid is connected to
of classes that extend the basic object Behavior, such as: the local LV grid [17]. Furthermore, a panel with controllable
loads is available. This Microgrid is the test field of the MAS
1) SimpleBehavior; application.
2) OneShotBehavior; In this section a general description of the part of the MAS
3) CyclicBehavior; that controls the laboratory system is presented. This part of the
4) CompositeBehavior; MAS incorporates the following agents.
5) SequentialBehavior.
MGCC: a Microgrid Central Controller Agent that an-
The JADE provides as well methods for manipulating these be- nounces the beginning and the end of the negotiation pe-
haviors. Detailed description of this functionality can be found riod for the market operation.
in [16]. Using these classes, next to auxiliary behaviors for mes- PV: a Production Unit Agent that is dedicated to represent
sage handling and monitoring, three main behaviors have been the photovoltaic panel.
introduced: Battery unit: a Production Unit Agent that is dedicated
1) Market Operation; to represent the battery panel. This agent can sell or buy
2) System Stand By; energy from the market depending on the state of charge
3) System Start Up. of the batteries.
Power system: a Grid Agent that represents the grid. It
The first behavior refers to market participation, the second does should be mentioned that the grid is considered as a pro-
nothing (stand by mode) and the third performs an elementary duction unit or consumption unit with infinite capabilities.
system start up with predefined critical loads. Load unit: a Load Unit Agent that is specialized in rep-
The implementation of each behavior is not the same in every resenting the loads of the system.
agent, since the main type of agents (Production Unit Agents, Power seller agents: Market Player Agents that bid
Load Units Agents, and Market Player Agents) have different directly in the market and represent the agents that
roles in the system. Furthermore, the Market Player Agents do sell power to the system. According to the assignment
not have the System Start Up behavior, since they are virtual problem, these agents are the persons.
agents created for the market operation only. One critical point Power buyer agent: Market Player Agents that partici-
is that the first and the last behaviors are incompatible and, in pate in the market and represent the agents that buy power
order to avoid dangerous situations, the MGCC coordinates the from the system. According to the assignment problem,
transition between the behaviors by sending proper messages. these agents correspond to objects.
The simple but effective solution is to send all agents to stand by
An operational cycle of the system is the following.
mode and then to select the new behavior. A detailed description
of how these behaviors are implemented is beyond the scope of 1) The MGCC announces the beginning of the market pe-
this paper. riod.
1454 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 20, NO. 3, AUGUST 2005

2) The power sellers estimate their actual production capa-

bilities and generate the proper number of power-seller
market agents.
3) The loads estimate their actual demands and generate the
proper number of power buyer market agents.
4) The power system agent, depending on the number of the
seller and buyer agents, generates sufficient number of
agents, in order to make the system symmetrical. Next,
it calculates the total number of the sellers (or buyers)
and generates an extra set of sellers and buyers. This
allows a production unit to sell directly to the grid or gives
the ability to a load to buy directly from the grid.
5) The MGCC announces the start of the new negotiation
cycle.
6) The energy market sellers and buyers agents bid in the
market, according to the algorithm described in Sec-
tion IV.
7) The MGCC announces the end of the current negotiation
8) The sellers and buyers agents announce to their parent
agent their assignment and then terminate themselves.
9) The MGCC announces the end of the energy market pe-
riod.
10)The MGCC announces the new set points
Fig. 9. Study case network of a microgrid.
It is important to consider the required negotiation time ac-
cording to the auction algorithm. Every negotiation cycle needs
a maximum of eight messages between each pair of seller and 2) Control Actions, like seamless transition to isolated mode
buyer. There is a maximum delay of approximately 200 ms for or reconnection to the main grid
the receipt of a message, i.e., each cycle needs approximately In the first category, each agent can calculate the amount of
2.5 s. This explains why the maximum number of iterations power needed for the specific operation and place a bid in the
should be less than 100; otherwise each negotiation could last auction algorithm
for hours. In the second category, coordination algorithms should be in-
A total number of 30 agents can therefore be implemented, cluded in the agents. An example is the reconnection to the main
which means that the power production or load units can pro- grid, where the system has to ensure that the main grid has ac-
duce 15 agents (considering that the power system unit gener- ceptable voltage and frequency and, after that, it may proceed
ates an extra number of agents equal to the initial). If every en- to all the necessary switching actions.
ergy block is 250 Wh, the system can handle 3750 Wh. In the
application presented, the block size is fixed, but it can be easily IX. COMPARISON OF DECENTRALIZED VERSUS
adjusted, according to the total demand or available production. CENTRALIZED APPROACH
A critical point of the application concerns communication,
It is interesting to compare the decentralized approach pro-
not only between the agents, but also between agents and the
posed in this paper with a more centralized one, where the Mi-
various devices. Several technologies have been used, like
crogrid Central Controller decides about the setpoints of each
eXtentible Markup Language (XML), XML remote procedure
unit [19]. The main difference between the two approaches lies
call (XMLRPC), andObject linking and embedding for Process
in the amount of information that is processed in each case. If
Control (OPC). An optimized usage of these technologies
the MGCC had available and could process all the information
would minimize the delay up to 50 or 60 ms, in order to allow
of the LCs, then its solution would be at least as good, as
the deployment of more than 50 agents. This number is ade-
the one provided by the decentralized control. This is because
quate for the control of a larger and more realistic Microgrid,
each LC does not have direct access to the information of its
like the one presented in Fig. 9 [18]. This network includes 6
neighbor controller, although MAS technology allows to ask for
DER units, including Photovoltaic panels, Fuel Cell, Batteries,
it. In practice however, it is very difficult for the MGCC to have
Wind Generators and CHP based on Microturbines.
access to all available information. For example, it is very com-
It should be noted that, due to space limitations, the approach plicated for the MGCC to know and handle the temperature of
presented here does not deal with other functionalities that the the battery of a specific storage unit. Similar conclusions can
control mechanism of a Microgrid should have. Although not be found in the literature [20]. It is very hard to implement at a
included in the presented implementation, two main categories reasonable cost a centralized system that can bid in the market
of functionalities can be distinguished: every hour and, at the same time, has the ability to shut down a
1) Energy Related Functionalities, like reserve power han- specific load or change a specific generator set point, in case of
dling. unstable operation within the next 300 ms.
DIMEAS AND HATZIARGYRIOU: OPERATION OF A MAS FOR MICROGRID CONTROL 1455

A related issue concerns data communication infrastructure. [6] J. A. Hossack, J. Menal, S. D. McArthur, and J. R. McDonald, A mul-
The proposed method only needs a simple local network and tiagent architecture for protection engineering diagnostic assistance, in
Proc. Power Engineering Society General Meeting, vol. 18, May 2003,
the information exchange is limited to the essential data only. pp. 639647.
A more centralized approach requires a significant data flow [7] T. Nagata and H. Sasaki, A multi-agent approach to power system
toward a single central point, in order to achieve similar re- restoration, in Proc. ISAP 2003, Lemnos, Greece, Sep. 2003.
[8] T. Hiyama, M. Kouno, H. Ono, and K. Furukawa, Multi-agent based
sults. The problem becomes extremely difficult and expensive wide area stabilization control of electric power systems, in Proc. ISAP
to solve, if real time functionalities are required, such as on line 2003, Lemnos, Greece.
security assessment. [9] J. Ferber, Multi-Agent Systems. An Introduction to Distributed Intelli-
gence. Reading, MA: Addison-Wesley, 1999.
Another issue is the openness of the system. Adopting a de- [10] J. M. Bradshaw, Software Agents. Cambridge, MA: MIT Press, 1997.
centralized approach, allows every manufacturer of DER units [11] D. P. Bertsekas and D. A. Castanon, A forward/reverse auction algo-
or loads to embed a programmable agent in the controller of his rithm for asymmetric assignments problems, Comput. Optimiz. and Ap-
plic., vol. 1, pp. 277297, 1992.
equipment, according to some rules. This would provide the re- [12] D. P. Bertsekas, Auction algorithms for network flow problems: A tuto-
quired plug and play capability of future DER units and loads. rial introduction, Comput. Optimiz. and Applic., vol. 1, pp. 766, 1992.
On the contrary, in a centralized system the installation of any [13] Foundation of Intelligent Physical Agents (FIPA) [Online]. Available:
http://www.fipa.org
new component would require extra programming of the central [14] Java Agent DEvelopment Framework [Online]. Available:
controller. http://jade.tilab.com/
[15] FIPA, FIPA 97 Specification Part 2: Agent Communication Language
(1997). [Online]. Available: http://www.fipa.org/
X. CONCLUSIONS [16] JADE PROGRAMMERS GUIDE [Online]. Available:
http://jade.tilab.com/doc/index.html
A Microgrid, although small in size, has complex operations. [17] S. Papathanassiou, D. Georgakis, N. Hatziargyriou, A. Engler, and
In this paper, application of MAS for the control of a Microgrid C. Hardt, Operation of a prototype microgrid system based on
is presented. The MAS approach was selected as a tool, not only micro-sources equipped with fast-acting power electronics interfaces,
in Proc. 31th PESC 2004, Aachen, Germany, Jun. 2004.
to provide intelligence for the needs of complex tasks, but also [18] S. Papathanassiou, N. Hatziargyriou, and K. Strunz, A benchmark LV
to facilitate management in the design of the algorithm. In this microgrid for steady state and transient analysis, in Proc. CIGRE Symp.
paper, emphasis is placed on the internal operation of the Micro- Power Systems with Dispersed Generation , Athens, Greece, Apr. 2005.
[19] A. Tsikalakis and N. D. Hatziagyriou, Economic scheduling functions
grid and its participation in the Energy Market. The main idea of a microgrid participating in energy markets, in Proc. CIGRE Symp.
of the algorithm presented, is that every DER or controllable Power Systems with Dispersed Generation, Athens, Greece, Apr. 2005.
load decides what is best for it, taking into account the overall [20] F. Ygge and H. Akkerman, Decentralized markets versus central con-
trol: A comparative study, J. Artif. Intell. Res., vol. 11, pp. 301333,
benefit through the auction algorithm. Of course, a MAS does Oct. 1999.
not aim exclusively to market participation, but also needs other
functionalities. Thus, the proposed architecture is considered as
the first step of a more integrated control mechanism. Practical
aspects from the implementation of the method on a small lab-
oratory Microgrid and projection to more realistic LV grids are Aris L. Dimeas (S04) was born in Athens, Greece, in 1977. He received the
discussed. General conclusions regarding the benefits from this diploma in electrical and computer engineering from the National Technical
University of Athens (NTUA), Athens, Greece. He is currently pursuing the
decentralized control approach compared to centralized control Ph.D. degree in the Electrical and Computer Engineering Department of NTUA.
are finally presented. His research interests include dispersed generation, artificial intelligence
techniques in power systems, and computer applications in liberalized energy
markets.
REFERENCES Mr. Dimeas is member of the Technical Chamber of Greece and student
[1] EU Research Project Microgrids [Online]. Available: http://micro- member of IEEE.
grids.power.ece.ntua.gr
[2] Microgrids: reliable power in a small package [Online]. Available:
http://www.lbl.gov/Science-Articles/Archive/EETD-microgrids.html
[3] R. Lasseter, A. Akhil, C. Marnay, J. Stephens, J. Dagle, R. Guttromson,
A. Meliopoulos, R. Yinger, and J. Eto, White Paper on Integration of
Distributed Energy Resources. The CERTS MicroGrid Concept, Con- Nikos D. Hatziargyriou (S80M82SM90) was born in Athens, Greece. He
sortium for Electric Reliability Technology Solutions (CERTS), CA, received the diploma in electrical and mechanical engineering from the National
Tech. Rep. LBNL-50 829, 2002. Technical University of Athens (NTUA), Athens, Greece, and the M.Sc. and
[4] A. Dimeas and N. D. Hatziagyriou, A Multiagent system for Micro- Ph.D. degrees from UMIST, Manchester, U.K.
grids, in Proc. IEEE PES General Meeting, Denver, CO, Jun. 2004, He is currently a Professor in the Power Division of the Electrical and Com-
PESGM2004001 244. puter Engineering Department of NTUA. His research interests include dis-
[5] Z. Ming, R. Jianwen, L. Gengyin, and X. Xianghai, A multi-agent based persed generation, artificial intelligence techniques in power systems, modeling
dispatching operation instructing system in electric power systems, in and digital techniques for power system analysis and control.
Proc. Power Engineering Society General Meeting 2003, vol. 1, Jul. Dr. Hatziargyriou is a member of CIGRE SCC6 and of the Technical Chamber
1317, 2003, pp. 436440. of Greece.