You are on page 1of 7

Ant Colony Optimization for Microgrid

Multi-Objective Power Management


C.M. Colson, Student Member, IEEE, M.H. Nehrir, Senior Member, IEEE, C. Wang, Senior Member, IEEE

Abstract Steadily increasing needs for electrical power,


progress in power deregulation, tight construction constraints on
new high voltage lines for long distance power transmission, and
global environmental concerns have created increased interest in
alternative energy (AE) generation. Hybrid combination of AE
sources can significantly improve their reliability and better
deliver power to customer loads without reliance on centralized
electricity production. It is expected that alternative energy
distributed generation (AEDG) microgrids that capitalize on
diverse energy sources, are controlled in a decentralized way, and
reduce the burden on the utility grid by generating power close to
the consumer will penetrate the existing grid-infrastructure in
the near future.
This paper presents a framework for an
intelligent supervisory controller that utilizes ant colony
optimization (ACO) methods for AEDG microgrid dispatch
control. The novelty of this work is the application of ACO to the
rapid microgrid power management problem given complex
constraints
and
objectives
including:
environmental,
fuel/resource availability, and economic considerations. Given
the compound nature of the multi-objective, multi-constraint
energy management problem for integrated AEDG systems, this
paper develops a constraint satisfaction problem (CSP) algorithm
capable of finding Pareto optimal dispatch solutions. Microgrid
power management control is not an easy problem, but its
development is critical for widespread AEDG system
implementation.
Index Terms Ant colony optimization, constraint satisfaction,
microgrid, power management.

I. INTRODUCTION

ODAY, the overwhelming majority of American homes,


businesses, and factories rely on a steady supply of
electrical power.
American electricity consumption is
projected to grow from 4,065 billion kilowatt-hours in 2006 to
5,619 billion kilowatt-hours in 2030, with growth rates of
2.2% per year for commercial customers, 1.5% per year for
industrial customers, and 0.8% per year for residential
customers [1]. This rising demand for electricity, an aging
existing infrastructure, and an uncertain energy future pose
significant risk to the complex and critical U.S. power system.
In 2007, the North American Electric Reliability Corporation
(NERC) cited critical demands for renewable energy and

C.M. Colson (email: christopher.colson@myportal.montana.edu) and M.H.


Nehrir (email: hnehrir@ece.montana.edu) are with the Electrical and
Computer Engineering Department, Montana State University, Bozeman, MT
59717 USA.
C. Wang is with the Division of Engineering Technology, Wayne State
University, Detroit, MI 48202 USA (e-mail: cwang@wayne.edu).

978-1-4244-3811-2/09/$25.00 2009 IEEE

storage resources, in addition to the monumental need for bulk


transmission modernization as key long-term emerging
issues [2]. The efforts to prepare for the energy future will
likely include measures to modernize the electrical grid,
enhance the quality and reliability of the energy supply,
diversify how the nation sources its immense hunger for
electricity, and address the looming crisis of environmental
impact of emissions from energy consumption. As an
example of the enormous scale faced by industry
infrastructure maintenance and upgrade, the Edison Electric
Institute (EEI) projected a $37 billion expenditure on its
member company transmission system assets from 2007 to
2010 alone [3].
Given the state of the nations electrical generation
capacity, it is impossible to discuss the future of electric power
in the United States without addressing the nations primary
reliance on only three sources (coal, nuclear power, and
natural gas) for nearly 90% of the nations electricity [1]. This
huge proportion of American electrical generation is derived
in a centralized manner. As demand continues to rise,
increased pressure will be placed on existing central power
plants, transmission assets, and distribution systems. This
steadily escalating need for electrical power, progress in
power deregulation, tight construction constraints on new high
voltage lines for long distance power transmission and
reliance on central generation places the nations energy future
in a difficult predicament. Recognizing these challenges for
central generation, along with growing global environmental
concerns, interest has increased in alternative energy (AE)
generation. Wind, photovoltaic, fuel cell, microturbine, and
other AE devices are attractive primarily because of low or
zero emissions, high efficiency, scalable application, and/or
adaptability to remote implementation. It is believed that
these desirable characteristics can be capitalized upon through
hybrid combination of AE sources, networked within a
microgrid framework to significantly improve their reliability
and better deliver power close to customer loads. Although
ideally adaptable to islanded off-grid applications, it is
expected that alternative energy distributed generation
(AEDG) microgrids that capitalize on diverse energy sources,
are controlled in a decentralized way, and reduce the burden
on the utility grid will penetrate the existing infrastructure
network in the near future [4]. In fact, the DOEs Vision
2020 statement set forth mid-term goals of achieving upwards
of 20% of new generation capacity added by 2010 in the form
of distributed energy resources [5].
Although there is little consensus on a standard definition,
for the purpose of this paper, a microgrid is specified as a

small (typically several MW or less in scale) power system


that has three primary components: distributed generators,
autonomous load centers, and the ability to operate
interconnected with or islanded from the larger utility
electrical grid. Interest in the microgrid concept, at the
distribution level, with multiple AEDG sources has been
increasing worldwide [6]-[9], mainly because the scalable
generation sources connected to microgrids can be coordinated
and controlled in a decentralized way. This operational mode
allows diverse AEDG sources to provide their full benefits
while reducing the coordination and control burden on the
utility grid. Further, the desire to become less dependent on
fossil fuel-based resources indicates the need for innovation in
ways to complement vast centralized non-renewable energy
power generation with renewable energy sources so that
energy utilization for the entire system is stable and
sustainable to the maximum extent. The primary goal of
microgrid architectures is to significantly improve energy
production and delivery for load customers, while facilitating
a more stable electrical infrastructure with a measurable
reduction in environmental emissions.
This paper seeks to address the power management
framework critical for making progress towards the goal of
properly integrating multiple AEDGs into a microgrid. The
conventional implementation of up to a medium-sized
generator onto a distribution system may not have a significant
impact on power quality at the feeder level. Additionally,
with only one generator operating, the power management
problem reduces to a trivial one. However, coupling multiple
AEDGs and dispatchable loads, within a microgrid framework
that is capable of utility interconnection, is a new concept and
creates a daunting challenge for safe and efficient operation.
This paper lays out an artificially intelligent power
management supervisory algorithm that can be used to
optimize AEDG dispatch based on local load demand and
additional realistic constraints such as: emissions concerns,
fuel availability and cost, weather conditions, the spot-market
price of electricity, and local history information. To this end,
the artificially intelligent supervisor should find a rapid Pareto
optimum to what is by definition a multi-objective
combinational problem. In short, this work addresses the
complicated engineering problem of how multiple AEDGs and
dispatchable loads can be successfully controlled on a
microgrid in a near-optimal manner using artificial
intelligence concepts.
In Section II, the multi-objective, multi-constraint microgrid
power management problem is discussed.
The ACO
algorithm and its application to microgrid constraint
satisfaction is discussed in Section III.
Conclusions and
planned future work are given in Section IV.
II. MULTI-OBJECTIVE CONFIGURATION

maximizations or minimizations. However, in most


cases, a global maximum of any particular individual objective
function may not be a satisfactory solution for the remaining
objectives [10]. Because of this, we need to alter our concept
of optimality for such problems. With similarity to economic
systems, a Pareto optimum can be reached where the solution
represents a state of satisfaction for one objective that
cannot be raised further without lowering another objectives
satisfaction. Multiple formal mathematical techniques exist
to derive Pareto solutions, but they can be computationally
expensive [11]. The power management of a microgrid
clearly fits into the broad field of multi-objective optimization
problems of interest today and currently, such problems have
not been proven to be solvable in strongly polynomial-time.
Therefore, evolutionary algorithms and heuristic techniques
such as ant colony optimization (ACO) may yield excellent
near-optimum results in short periods of time, which is critical
for good power management and dispatch [12].
The formulation of the multi-objective, multi-constraint
microgrid power management problem begins by identifying
the attributes of the microgrid. These include the microgrid
capability to:
- Meet consumer load demand within given voltage and
frequency standards.
- Reduce or defer distribution and transmission capacity
additions.
- Reduce line losses, provide VAR support, and
stimulate power market opportunities for trading
ancillary services to the utility grid.
- Facilitate customization as load composition, diversity,
and profile vary (example: urban vs. rural).
- Allow the integration of stable and adaptable control
system behavior for seamless plug-and-play
implementation and sensible load control.
These attributes lead to the definition of the primary
constraints:
- AEDG resource availability (i.e., solar insolation, wind
energy, etc.)
- Microgrid bus voltage levels and stability factors
- Real/Reactive power demands from microgrid loads
- Status of interconnection
The primary objectives for the microgrid power
management problem are:
- Minimize economic factors (i.e. fuel costs, operation
and maintenance, etc.)
- Minimize environmental impact
- Optimally dispatch dump/shedable loads
- Maximize delivery to utility grid (including ancillary
services, reserve margin, etc.)
Based on this objective and constraint formulation of the
microgrid power management problem, and ACO algorithm is
developed to achieve rapid solutions, as explained in the next
section.

For typical multi-objective optimization problems, the goal


is to seek an optimal solution for numerous competing
objectives, simultaneously. In such problems, the satisfaction
of the objective functions becomes a combination of vector
2

III. ANT COLONY OPTIMIZATION


A. Framework
The Ant Colony Optimization (ACO) algorithm is a natureinspired metaheuristic optimization method proposed by
Dorigo, et al. [12], for solving NP-hard combinatorial
optimization problems such as the benchmark traveling
salesman problem (TSP), scheduling problems, subset
problems, and a host of others. This artificial intelligence (AI)
technique is of particular interest because it has been shown to
develop Pareto-optimal solutions with short time complexity
[13]. Additionally, the ACO algorithm has been shown to
outperform other general purpose optimization algorithms
including genetic algorithms (GA) when applied to a number
of benchmark combinatorial optimization problems, although
this claim is widely interpretational [12].
Within the ACO algorithm, colonies of artificial ants
cooperate to find solutions to discrete optimization problems.
Problems in ACO are generally represented with a
construction graph, where each node in the graph corresponds
to a component of the solution. The goal is to find the
solution with the minimum cost or distance path, which
ultimately represents the best answer. Each ant walks on
this graph and incrementally builds a solution. Modeled after
real ants, the behavior of the artificial ants is governed by two
primary factors: stigmergic tendency and random exploration.
Stigmergy, or indirect communication facilitated by the
environment, is accomplished in nature through the deposit of
chemical pheromones. The pheromones reinforce good
solutions and guide the search. For the ACO algorithm, a
simulated ant uses artificial pheromones at each step in the
construction graph, along with other problem specific
heuristics, to randomly select the next component in the
solution.
The algorithmic framework for ACO depends on the
construction graph representation, is typically simplified as:
G = (N,A,C).
For the graph G, the set of nodes
N = {d1,d2, . . . ,dn} represents where path decisions must be
made; the set of arcs A ={lij} link the nodes i to j; and
(optionally) C = {cij} is the set of costs associated with arcs A.
The elements of sets N and A are typically constrained. In setcovering or constraint satisfaction variants of the
combinational optimization problem, the order of the solution
sequence is not important. Fig. 1 shows an example of a
construction graph that the ACO algorithm would attempt to
solve. Nodes d1 through d3 are shown in Fig. 1, as well as
how additional nodes (dn) would be added to the construction
graph. A complete path on the graph passes through each
node once, contains a set of arcs (represented as dashed lines
in Fig.2), and is called a solution (s). The minimum cost path
is called the optimal solution (s*) and is represented by the
darkened arcs in Fig. 2 as the shortest complete path.

Fig. 1. Typical ACO construction graph framework.

Fig. 2. Optimal solution (s*) representation.

In the constraint satisfaction problem (CSP), the node and


arc representation on the construction graph is slightly
modified from the general case. The graph is defined by a
triple (X,D,C), where X is a set of variables, D is a set of
domains, and C is a set of constraints. Each variable xi in X
can be assigned a value from its domain Di in D. The
constraints in C restrict the set of values that these variables
can take simultaneously.
A complete solution, s, is
constructed by incrementally selecting a value for each
variable in X. The goal in a CSP is to find a solution that
satisfies all the constraints in C [12].
For the microgrid power management problem, the CSP
construction graph is easily defined. Each generator, storage
system, and/or dispatchable load is represented by a variable.
It is noted that the microgrid could have a population of
generators and loads that are a mix of dispatchable and nondispatchable systems, and this fact simply modifies the
number and type of variables defined. The domain of each
variable represents the quantized operating points at which the
generator or load could be commanded to operate.

Fig. 3: CSP construction graph.

The construction graph for a CSP contains a node for every


possible variable assignment. For example, Fig. 3 is the
construction graph for the variables x1, x2, and x3. The graph
is fully connected except for nodes of the same variable (e.g.
x1,d1 cannot be connected to x1,d2). This prescribed constraint
enforces the stipulation that only one instantiation for each
variable is allowed in a solution set. Based on the microgrid
power management formulation, given in section II, this is
analogous to selecting one operating point for an individual
generator (it is invalid to select two or more operating points
for the same generator). A generator operating point could be
its power, voltage, and/or frequency setpoints. The ants walk
the graph until a value has been selected for each variable,
resulting in dispatch positions for each generator attached to
the microgrid. The inclusion of every possible assignment
typically results in a very large solution space, and the
computation of transition probabilities is significantly more
computationally expensive than for other applications, such as
the TSP. A method for reducing this complexity is suggested
later in the section discussing data representation. The nonfully connected construction graph for the power management
problem with n-generators (excluding controllable loads at this
time) and m-by-p-domains (in V and f) is displayed in Fig. 4.
It should be noted that all nodal connections are not indicated
on the construction graph for clarity.

Fig. 4: Construction graph for power management CSP problem (some


connections omitted for clarity).

B. ACO Algorithm Development


Ant tour construction is based on a probabilistic function.
At each construction step, an ant must choose the next node
to visit in pursuit of a complete solution. The probability of
choosing node j from node i, is a combination of pheromone
weighting and random exploration, expressed by:

p =
k
ij

ij nij

[ ] [ n ]

lNik

il

(1)

il

where: ij is the pheromone on arc ij, is the pheromone


weighting factor, nij is the heuristic value on arc ij, is the
heuristic weighting factor, and Ni is the feasible neighborhood
of options for the ant, k, to traverse from node i. In addition
to global heuristics that help improve the simulated ants
capabilities, pheromones associated with the construction
graph are deposited on the arc chosen and globally
evaporate over time. Pheromone decay on the construction
graph arcs are accomplished by:

ij f = (1 ) ij o

(2)

where: ijo is the current step pheromone value on arc ij, ijf is
the updated pheromone value on arc ij, and is the decay
factor. This combined effect allows for the artificial
stigmergic effect which leads to better solution convergence.
Numerous modifications of the ACO algorithm exist that
allow customization for particular problems. Primarily, the
level to which the best solution found by the ants is exploited
4

for pheromone updating and the frequency of pheromone reset


are design trade-offs that must be considered when applying
ACO algorithms. Typically, the number of ants per colony
and the number of colonies is predetermined by the designer.
This is especially crucial in the elitist-based, max-min Ant
System (MMAS) system [12] because pheromone update and
scaling does not occur until an entire colony of ants has
selected its best colony solution. That information is passed
on to the next colony by changing the pheromone matrix that
the new colony operates on. There are some practical and
theoretical tradeoffs that are incorporated into the choice of
how many colony ants and how many colonies are applied.
Clearly, as the number of colonies increases, more best-so-far
updates of the pheromone matrix are passed on, presumably
leading to better solutions. However, increasing the number
of colonies becomes more computationally expensive, and
because the pheromones on good solutions are becoming
stronger relative to weak solutions, the level of exploration by
subsequent colony ants is diminished. By increasing the
number of ants within a colony, initially there is great
exploration, but it is computationally expensive and unless the
and values, in equation (1), are weighted properly, the ant
behavior approximates random search.
The primary algorithm structure is described by the
following pseudocode:
procedure Ants (input # of ants per colony, input # of colonies)
Derive parameters
Initialize pheromone and heuristic matrices
for ( number of colonies )
Find_Best_Tour_of_Colony
Determine_Fitness_of_Best_Solution
Update_Pheremone_Matrix
Update_Heuristic_Matrix
end
Calculate_Dispatch_Quantities
end

Find best tour of colony: Based on selected constraints,


each ant in a colony constructs its solution set by selecting a
generator value from each category according to (1).
Determine fitness of best solution: After the ants construct
their solutions, the best solution is determined by comparing
solutions to a performance measure. This information is
retained in memory for later comparison, updating the
pheromone and heuristic matrices, and ultimately determining
the solution dispatch quantities.
Update pheromone matrix: The pheromone matrix is
updated based on the random proportional rule [12].
Update heuristic matrix: Ants are only allowed to select
generator values that meet the selected constraints. Values
that violate constraints (infeasible solutions) are given a
heuristic value of zero. The heuristic matrix is updated to
reflect infeasible solutions.
Calculate dispatch quantities: Based on the solution
determined by the ants, quantities such as a generators
voltage magnitude and angle are calculated.
C. Data Structure
There are numerous ways to represent the constraints and
objectives for the microgrid optimization problem. For
simplicitys sake, a matrix data structure for constraints and
objectives is suggested and shown in Fig. 5. Each generator
(Gx) has a corresponding vector of values (Nx), the length of
which is determined by the resolution of the functions that
represent power produced or consumed, environmental cost,
monetary
cost,
fuel/resource
availability,
etc.
Correspondingly, the heuristic and pheromone matrices map
directly to the constraint and objective dimensions, as will be
explained later.

Derive parameters: For the purpose of the microgrid power


management CSP, generator operating characteristics are
defined and classified. A matrix is used to represent the
operating characteristics and constraints associated with each
generator.
Initialize pheromone and heuristic matrices:
ACO
algorithms such as MMAS utilize limits on the amount of
pheromone. For the microgrid power management CSP, the
pheromones are normalized (upper limit of 1and lower limit of
0). The pheromone matrix is initialized to the upper limit. By
initializing the pheromone values and utilizing a small
pheromone evaporation parameter (such as 0.02), this results
in an initial search phase that is very explorative [12].
Significant initial exploration is beneficial to the problem at
hand because it allows rapid investigation of a broader scope
of the construction graph and is not as hampered by local
minima/maxima.

Fig. 5: Data structure for ACO implementation in the microgrid case.

A key implementation feature unique to the microgrid


optimization problem is for this data structure scheme to allow
the algorithm to update infeasible nodes according to time.
For example, as resources vary with time (e.g., blowing wind
or varying sun insolation intensity), the resource levels
ultimately determine the operating ranges of each generator
(i.e. power output from a photovoltaic panel). To monitor
these generator parameter changes, the resource constraint
matrix is correspondingly updated for nodes that are not

available to be chosen for dispatch. This is shown in Fig. 6;


the dark nodes indicating unavailability and the light nodes
indicating availability. In this implementation, infeasible
solutions (dark nodes) are given a heuristic value of zero.
Using this framework, it should be possible to have the ACO
algorithm operate in real-time, taking in sensory data
corresponding to resource information, and modify the
constraint and objective matrices as ants are seeking their
solutions. This feature will allow the algorithm to respond
dynamically to changes in microgrid capability.

Fig. 6: Data structure for ACO time-varying resource constraint matrix.

The data structure of the pheromone and heuristic matrices


are significantly larger in size than the constraint and objective
matrices described above. This is primarily due to the
geometric expansion required to fully connect each node with
every other node. Subsequently, in the case of n-generators
with a one dimensional domain of size m, the pheromone and
heuristic matrices are of size nxmxT. As shown in Fig. 7, the
additional dimension, represented by T, corresponds to the
to arcs, from the generator/node cell. For example, if the
current ant position is generator #2 at node #4, then matrixsheet T8 represents the totality of arc travel options available
from that node.

dependence on computationally expensive methods [14].


Most ant colony metaheuristic applications hav
ve not
explicitly focused on constraint satisfaction problems (CSPs)
(
[15]-[16, et.al.]. This is a curious result for a few reeasons.
First, it presents the obvious question as why ant algo
orithms
have not been more hotly pursued to solve CSPs. This may
m be
because, although ant algorithms, such as MMAS and
d ACO,
can find good, near-optimal solutions to CSPs, they cannot
prove that a satisfactory Pareto optimum does not exist
e
if
conditions demonstrate a lack of an optimum [12]. This
T
is
undesirable for many computer science and engin
neering
applications. Despite this drawback, ant algorithms seeem to
be ideally suited to a class of CSPs that represent prractical
application problems of interest to engineers. Additionally,
for practical applications speed is a critical aspect of
intelligent systems that will operate in real-time or neaar realtime applications, such as for microgrid power manag
gement.
Clearly, for engineers seeking practical implementations, the
drawback ACO exhibits with regards to provability for
f the
existence of an optimum (or no optimum) is irrelevant. What
n that is
the engineer desires is a rapid convergence to a solution
presumably near optimal. Ant algorithms can achiev
ve this,
and possibly because less is known about their theo
oretical
expression, this is why they have been of less interest to
researchers.
In most engineering applications to date, the ACO
algorithm has been applied to off-line problem solving
g [17][20]. This involves using computational resources to solve
s
a
problem, not in real-time, and have those solutions used
d later.
In one recent example of a purely engineering application for
ant algorithms [21], the ant metaheuristic framew
work is
applied to a relatively straightforward engineering design
problem, storm water network design. The optimal storm
water network solutions, solved off-line, can be applied to
construction design. The framework presented in thiss paper
has sought to build upon such examples and app
ply ant
metaheuristics to a real-time CSP application. The ultimate
purpose of applying ACO to the microgrid power manag
gement
problem is to capitalize on ACOs two most po
owerful
properties: rapid convergence and inherent stigmergic memory
m
that seem ideally suited for real-time applications.
IV. CONCLUSION AND FUTURE WORK

Fig. 7: Data structure for ACO heuristic and pheromone matrices.

D. ACO Engineering Applications


It should be noted that both the constraint and objective
functions in the microgrid power management problem have
been shown to be non-linear, non-homogenous, and
sometimes time-varying in the desired variables. While it is
possible to linearize these functions, it is not desirable to do
so. This exhibits the need for good, rapidly converging
techniques for determining near optimal solutions without the

The power management of microgrid generation is a very


complicated problem. In this paper, a framework haas been
developed for utilizing the promising ACO heuristic tecchnique
for implementation in a microgrid power manag
gement
controller. The formulation details for the power manag
gement
problem were discussed, as well as the structure of thee ACO
algorithm as applied to the microgrid scenario.
This paper has specifically addressed a smalll, but
fundamental portion of the challenge to coordinate and control
c
multiple AEDG sources connected to a microgrid. The design
of an intelligent decision-making strategy, based on
n ACO
heuristics, is an important step toward autonomous operation
of microgrids that achieve energy sustainability while
minimizing operational costs and environmental emiissions.
6

Follow-on publications and future work will detail results


from simulation studies using this ACO technique for
microgrid power management. Constraints including real and
reactive power dispatch, time-varying resource and load
matrices, and a discussion regarding performance factors will
be discussed. Techniques for addressing solution stagnation
and pheromone resetting results will be included.

[17] P. Korosec and J. Silc, The distributed multilevel ant-stigmergy


algorithm used at the electric-motor design Engineering Applications of
Artificial Intelligence, Vol. 21, No. 6, September 2008.
[18] M. Lopez-Ibanez, T.D. Prasad, B. Paechter, Ant colony optimization for
optimal control of pumps in water distribution networks Journal of Water
Resources Planning and Management, Vol. 134, No. 4, July/August 2008.

ACKNOWLEDGMENT

[19] S.K. Chaharsooghi and A.H. Meimand Kermani, An effective ant


colony optimization algorithm (ACO) for multi-objective resource allocation
problem (MORAP), Applied Mathematics and Computation, Vol. 200, No. 1,
Jun 15, 2008.

The authors would like to thank Kristie Simpson of


Montana State University/Dynojet Research for her research
assistance and contributions to this work.

[20]A. Ahuja, S. Das, and A. Pahwa. An AIS-ACO Hybrid Approach for


Multi-Objective Distribution System Reconfiguration, IEEE Transactions on
Power Systems, Vol. 22, No. 3, August 2007.

REFERENCES
[1] Energy Information Administration - Office of Integrated Analysis and
Forecasting, Annual Energy Outlook 2006, U.S. Department of Energy,
DOE/EIA-0383, 2006.

[21] M.H. Afshar, Partially constrained ant colony optimization algorithm for
the solution of constrained optimization problems: Application to storm water
network design, Advances in Water Resources, Vol. 30, 2007.

[2] North American Electric Reliability Corporation (NERC), NERC 2007


Long-Term Reliability Assessment, Princeton, NJ, October 2007.
[3] Edison Electric Institute, Transmission Projects: At A Glance,
Washington, D.C., January 2008.
[4] N. Hatziargyriou, H. Asano, R. Iravani, and C. Marnay, Microgrids,
IEEE Power & Energy Magazine, Vol. 5, Issue 4, July/August 2007.
[5] Strategic plan for distributed energy resources, US Department of
Energy,
September
2000.
Available
at:
http://www.eere.energy.gov/de/pdfs/de_program_plan.pdf.
[6] R. Lasseter, A. Abbas, C. Marnay, J. Stevens, J. Dagle, R. Guttromson, A.
Sakis Meliopoulos, R. Yinger, and J. Eto, Integration of distributed energy
resources: The CERTS microgrid concept, California Energy Commission,
P500-03-089F, Oct. 2003.
[7] Y. Fujioka, H. Maejima, S. Nakamura, Y. Kojima, M. Okudera, and S.
Uesaka, Regional power grid with renewable energy resources: A
demonstrative project in Hachinohe, CIGRE 2006, Paris, August 2006.
[8] B. Bucholz, et al, PE Magazine, July/August 2007.
[9] European Commission (Apr. 2006), European SmartGrids technology
platform: Vision and strategy for Europes electricity networks of the future.
Available at: http://ec.europa.eu/research/energy/pdf/smartgrids_en.pdf.
[10] Saul Gass, Linear Programming: Methods and Applications - 5th Edition,
Dover Publications, Inc. 2003, pgs. 224-250.
[11] A. Osyczka, Multicriteria optimization for engineering design, Design
Optimization edited by J. Gero, Academic Press Inc., 1985.
[12] M. Dorigo and T. Stutzle, Ant Colony Optimization. MIT Press, 2004.
[13] B. Fox, W. Xiang, H.P. Lee, Industrial applications of the ant colony
optimization algorithm, International Journal of Advanced Manufacturing
Technology, Vol. 31, No. 7-8, January 2007.
[14] K.Y. Lee and M.A. El-Sharkawi, Modern Heuristic Optimization
Techniques Theory and Application to Power Systems, IEEE Press,
Piscataway, NJ, 2008.
[15] Christine Solnon, Ants Can Solve Constraint Satisfaction Problems,
IEEE Transactions on Evolutionary Computation, Vol. 6, No. 4, August 2002.
[16] Finbarr Tarrant & Derek Bridge, When Ants Attack: Ant Algorithms for
Constraint Satisfaction Problems, Artificial Intelligence Review, Vol. 24,
2005.