This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON SMART GRID 1

Dynamic Data Driven Adaptive Simulation

Framework for Automated Control
in Microgrids
Aristotelis E. Thanos, Student Member, IEEE, Mehrad Bastani, Student Member, IEEE,
Nurcin Celik, Member, IEEE, and Chun-Hung Chen, Senior Member, IEEE

Abstract—In this paper, we introduce a novel dynamic for human operators to make the necessary changes (some-
data driven adaptive simulation framework for the operation times needed in a fraction of a second) to quarantine problems
and control of microgrids (MGs) that significantly accelerates in the network. When the dust finally settled, over 50 million
the real-time computation of the resource allocation, and controls
decisions to optimize the operational cost, energy surety, as well as people were without power, and the majority were without
emissions per MW. The proposed framework includes a database power for the adjacent two days or more. It is estimated
receiving input from electrical and environmental sensors, a fault that this particular outage cost the North American economy
detection algorithm that discovers liabilities and potential haz- $6 billion.
ards within the MG, an agent-based simulation of the MG system, Microgrids (MGs) are autonomous electricity environments
an optimal computing budget allocation-based control selection
algorithm that uses the agent-based simulation to decide the best that function within a larger electrical system. They are able
control design of the MG, and a multiobjective algorithm for to react to a crisis (or advise a human operator on how
optimizing the decisions of the MG given the best control design. to react) inside the necessary time frame in order to con-
For validating our framework, we use the structure of a realistic tain local problems when necessary. They have potential as
MG that is simulated using real-historical data. The experiments a mechanism to increase reliability and efficiency while also
reveal that the proposed framework significantly reduces the
computational burden of a considerably complex multiobjective offering new services which are unavailable under a traditional
problem. grid system [1]. These new services may include dynamic
pricing [2] and distributed energy resource management [3].
Index Terms—Agent-based simulation, autonomous con-
trol, microgrids (MGs), multiobjective optimization, real-time A typical MG can cater to many different types of consumers
simulation. including customers with traditional demands, customers with
sensitive loads, and customers who have micro-sources con-
sisting of photovoltaic (PV) panels, wind turbines, fuel cells,
I. I NTRODUCTION micro-turbines, diesel generators, or battery storages [4].
ODERN society has come to depend on the easy In this paper, we propose a real-time dynamic data driven
M and reliable flow of electricity. When a major power
outage occurs, damages, and the inability to conduct nor-
adaptive simulation (DDDAS) framework for MGs. The
DDDAS paradigm entails the ability to dynamically incor-
mal life operations cause billions of dollars in losses per porate additional data into an executing application, and in
day, in addition to costing human lives. America’s great reverse, the ability of an application to dynamically steer the
wake-up call about the severe cost of insufficient invest- measurement process [5], [6]. It allows the application to be
ment and oversight in its electrical infrastructure came on able to guide data collection and determine when it is most
August 14, 2003 when a widespread power outage hit many beneficial to gather future data.
parts of the Northeastern and Midwestern United States as The DDDAS is a promising and evolving new technol-
well as the Canadian province of Ontario. The U.S. electrical ogy with, in recent years, direct engineering and science
system is so dynamic and interconnected that it is impossible applications [7]. Carnahan and Reynolds [8] used a DDDAS
framework and focus on identifying the crucial information
Manuscript received January 22, 2015; revised May 30, 2015; accepted that will determine the flexible points in which the simula-
July 20, 2015. This work was supported by the Air Force Office of Scientific tion should automatically be adapted to the new situation.
Research under Award FA9550-13-1-0105. Paper no. TSG-00073-2015. Park et al. [9] developed a new method of data reduction
(Corresponding author: Nurcin Celik.)
A. E. Thanos, M. Bastani, and N. Celik are with the Department to enhance data transmission for sensors used in structural
of Industrial Engineering, University of Miami, Coral Gables, health monitoring. Khaleghi et al. [10] proposed a DDDAS
FL 33143 USA (e-mail: a.thanosfilis@umiami.edu; m.bastani@umiami.edu; framework for effective surveillance and crowd control using
celik@umiami.edu).
C.-H. Chen is with the Department of Systems Engineering and Operations unmanned aerial and ground vehicles. Blasch et al. [11]
Research, George Mason University, Fairfax, VA 22030 USA (e-mail: explored the concept of an adaptive fine-tuning of secure
cchen9@gmu.edu). communication trust analysis that aims to achieve a balance
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. between standard static solutions versus dynamic data driven
Digital Object Identifier 10.1109/TSG.2015.2464709 solutions. Finally, Thanos et al. [12] defined an MG as
1949-3053 c 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON SMART GRID

a system of systems and apply the DDDAS paradigm to evaluation module is embedded in the model to swiftly evalu-
select the best simulation fidelity for a military MG while ate the designs. The simulation model communicates with the
Thanos et al. [13] and Celik et al. [14] applied a DDDAS OCBA-based control design selection algorithm that makes
framework to power network systems to address the economic real-time decisions on which designs need to be simulated
load dispatch problem [15]. and on how many replications. The simulation is used to
Many fields deal with prediction without a full dataset. calculate the performance of various control designs (in an
A DDDAS assists in filling in the gaps of a complex iterative manner) and, subsequently, the OCBA-based con-
(or even real world) model, while also improving itself trol design selection algorithm ranks these control designs
semi-autonomously [16]. Each DDDAS, however, must be based on their calculated performance. Once the best control
constructed differently depending on the requirements of the design is selected, the problem is solved using the proposed
current experiment. In our presented framework, autonomous multiobjective optimization algorithm which is described
control is enabled in an MG system via varying modeling in Section II-C.
fidelities that can operate at multiple levels and can accom- The connection amongst the four major components of our
modate uncertainties and various complexities, in a compu- framework is demonstrated in Fig. 2.
tationally efficient manner. Here, as the set of the potential
control designs is finite, optimal computing budget alloca-
tion (OCBA) is embedded for the selection of the most appro- A. Algorithm I: Fault Detection and Isolation
priate fidelity of the model with a high probability of correct FDI is a crucial element of numerous operations manage-
selection. ment automation systems. Fault detection is defined as the
identification of a situation that a problem has occurred, even if
the origin cause is not yet known. Fault diagnosis, also known
II. A RCHITECTURE OF THE P ROPOSED S YSTEM as fault isolation, on the other hand, is responsible for locating
In this paper, we present a novel DDDAS framework for the origin causes of problems, allowing the system to take the
the operation and control of MGs that speeds up the real-time appropriate corrective actions. Automated fault detection and
computation of the resource allocation and significantly con- diagnosis prominently relies on the input of sensors or calcu-
trol decisions while optimizing the operational cost, energy lated performances of the system [17]. In application systems,
surety, and emissions. As illustrated in Fig. 1, our frame- the most common faults encountered are sensor failures, and
work includes: 1) a database receiving data from electrical and as a result, a major focus in FDI is to identify these problems
environmental sensors; 2) a fault detection algorithm; 3) an and address the challenge of distinguishing them from process
agent-based simulation of the MG system that includes sepa- problems [18].
rate modules for energy storage and energy generation from FDI using equation-based analytical models involves and
solar, wind, and diesel sources; 4) an OCBA-based control necessitates a static set of state variables and measurements.
design selection algorithm; and 5) a multiobjective optimiza- In matrix-based techniques particularly, every time that the
tion algorithm for optimizing the decisions of the real MG. FDI analysis is performed, all state variables are recalcu-
A high level description of the framework follows. lated, updated, and stored. This process does not scale well to
The real system is equipped with sensors that feed the large systems where an immense collection of state variables
database of our framework with electrical and environmen- and sensors exists. Other approaches, such as the general-
tal data. More specifically, the database is collecting wind ized discriminant analysis (GDA) or the kernel GDA [19],
speed, solar irradiance, temperature, voltage, current, and fre- decrease the computational burden by only transmitting note-
quency data from the real MG. The collected data is crucial worthy new information, but even then, every state variable
for the framework since the environmental data is responsible in the system still has to be defined and preserve space for
for the forecasting of the energy generation from renewable storage.
sources and the electrical data is responsible for the proper On the other hand, event-driven systems do not require
operation of the MG and the smooth transition from the nor- a predetermined storage size to represent the system’s state.
mal operation mode to the islanded mode. These data are A specific event might occur several times, and events that
then used by the fault detection and isolation (FDI) algorithm do not occur necessitate no storage. Events contain some
to discover liabilities and potential hazards within the MG. attributes including but not limited to an event category, an
After the FDI algorithm scans the system, its results are associated object that might reveal the root cause of the prob-
provided to the agent-based simulation model. The agent- lem, and a time stamp that is tied with the exact time that
based simulation of the MG system imitates the real system the problem is observed. When the system is monitored in
under several different control designs. The simulation model the highest fidelity, recent events are often stored in the rapid
approaches the MG as a system of systems and each agent access memory (RAM) for brisk retrieval.
simulates a sub-system within the overarching system. The The possible events that may be reported in our model are
goal of the simulation is to evaluate the performance of dif- listed in Table I. When a measured variable gets a suspicious
ferent control designs by isolating the most appropriate regions value then the corresponding event is created and a possible
in the system and by finding an optimal resource allocation fault is associated with it. If there is a sensor error, the sys-
that minimizes the operational cost while also maximizing tem stops acquiring data from this source and, concurrently,
the energy surety of the system. To this end, a performance increases the details of the MG control design.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

THANOS et al.: DDDAS FRAMEWORK FOR AUTOMATED CONTROL IN MGs 3

Fig. 1. DDDAS framework for automated control in MGs.

TABLE I
L IST OF E VENTS IN FDI A LGORITHM

Fig. 2. Connection of the components of the proposed framework.

B. Algorithm II: OCBA-Based Control Design Selection

This algorithm addresses the design selection problem from
a finite set of control design alternatives with stochastic
constraints in MG simulations for the allocation of the simula-
tion replications. The real MG has several segregation points
which are used to isolate specific areas or even individual
buildings. Our framework automatically controls sets of these
points and isolates regions accordingly when needed. For
example, as shown in Fig. 3 for a group of 22 buildings, we
may have a design where the buildings are partitioned into two
subgroups, each controlled by a segregation point (design A),
or four subgroups and segregation points (design B), or 22 sub-
groups and segregation points (design C), etc. These different
control designs are predetermined to fit the real system that
is simulated. The goal of this algorithm is to select the
most appropriate design and provide it to the multiobjective Three demand categories, namely, critical, priority, and non-
optimization algorithm as an input. Using this design, the mul- critical are used to prioritize buildings within the MG. Critical
tiobjective optimization algorithm provides a Pareto solution demand areas are comprised of required areas for the adminis-
while saving significantly from the computational time. tration and community control; priority demand areas cover the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON SMART GRID

Fig. 3. Different designs for a group of buildings.

buildings in which typical day-to-day functions take place; and

noncritical demand buildings include the recreation, housing,
and shopping facilities. There are also three different genera-
tion sources within the considered MG: 1) diesel generators;
2) PV arrays; and 3) wind turbines.
For the proposed OCBA-based control design selection
algorithm, the simulation performance of the system is for-
mulated as a single maximization objective function as
shown in (1). This single objective function combines the aver-
age percentage of satisfaction for different load types and total
operational cost of the MG during the simulation time
cmax − c
f (X) = α · Pcrit + β · Ppr + γ · Pncrit + δ · (1)
cmax
where X is the design for which the performance is evaluated,
Pcrit , Ppr , and Pncrit are the percentages of energy surety of
the critical, priority, and noncritical loads, respectively, cmax
is the maximum cost calculated so far between all replications
and designs, c is the cost of the current replication for design
X and α, β, γ , and δ are the coefficients of the percentages of
energy surety/cost and correspond to the priority that is given
to the different objectives.
We assign simulation replications to design alternatives
utilizing the aforementioned performance function and the
OCBA technique [20]–[25]. As the simulation proceeds, we
compute the mean and variance of the MGs performance using
the data which are collected up to that stage. The algorithm
is presented in Fig. 4. Fig. 4. Operations of the isolated MG control design algorithm.
In the OCBA-based control design selection algorithm, the
best design (b) may change in each iteration (l). As the simu-
lation proceeds by increasing the number of iterations, the best the environmental factors by minimizing the emissions of
design in most cases converges to the optimal design as l goes carbon oxides (CO2 ), oxides of nitrogen (NOx ), and sul-
to infinity (the probability of correct selection approaches very fur oxides (SO2 ). The notation of the problem is provided
close to 1). In order to conduct OCBA appropriately, the ini- in Table II.
tial number of replications (n0 ), and one-time increment () The mathematical formulation of the objectives of the prob-
should not be too small to avoid a poor estimation of the mean lem is given in (2)–(4). Z1 , Z2 , and Z3 represent the operational
and variance. cost of the MG, the amount of emissions and the energy surety,
respectively
C. Algorithm III: Multiobjective Optimization
The multiobjective optimization algorithm aims to deter- 
k  
mine the decision variables of the MG real system with MinZ1 = fc uz + vz Pz + wz P2z (2)
accuracy. Thus, given a set of segregation points, demand z=1
in each building, and renewable generation, the multiobjec- k 3
tive optimization algorithm defines the region that will be MinZ2 = ay ezy Pz (3)
disconnected as well as the generation sources that will be z=1 y=1
initiated to satisfy the demand. The goal of this algorithm n
 
is to maximize energy surety while at the same time min- MaxZ3 = t1 cri xi di + t2 pri xi di + t3 1 − cri − pri xi di . (4)
imize the total operational cost. Our model also considers i=1
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

THANOS et al.: DDDAS FRAMEWORK FOR AUTOMATED CONTROL IN MGs 5

TABLE II
S UMMARY OF N OTATION AND F ORMULATION OF THE P ROBLEM

Fig. 5. Segregation points and solution space complexity.

It must be noted here that when the control of the isola-

tion mechanism within the MG becomes less intensive, the
algorithm may end up with decisions that are suboptimal sac-
rificing from energy surety, cost, and emissions. To this end, it
is crucial to effectively select the fidelity level for the control
mechanism in order to speed up the computational time while
also producing near-optimal decisions for the system.
Our multiobjective optimization algorithm utilizes the
∈-constraint method which has multiple advantages over the
The constraints of the problems are weighting method in cases where there exist multiple objec-
tives as follows.

k 
n 1) It can produce unsupported efficient solutions for mul-
pW + pS + Pz = di xi (5) tiobjective integer programming problems.
z=1 i=1 2) There is no need to scale the objective functions in order
Pmin < Pz < Pmax ∀z = 1, . . . , k (6) to obtain a weighted sum.
z  z
3) The total number of efficient solutions can be controlled
i∈sj xi
gj = ∀j = 1, . . . , m (7) by adjusting the number of grid points [26]–[28].
lj The appropriate ∈ for each constraint can be acquired by
xi ∈ {0, 1} ∀i = 1, . . . , n (8) a parametrical variation of the right-hand-side part of each
gj ∈ {0, 1} ∀j = 1, . . . , m. (9) constraint. In order to prevent redundant or dominated solu-
tions, a range for each objective is calculated using the payoff
In our framework, each segregation point is associated with table. The payoff table is created using the results from the
a specific set of buildings depending on the topology of individual optimization of each objective function using the
the MG. When the fidelity level of the control design becomes other objectives as constraints. After the calculation of the pay-
higher, so does the number of segregation points. Therefore, off table, the ranges of the values of the objectives can be
the number of buildings that is associated with each point divided in an arbitrary number of intervals.
becomes smaller. In particular, in the highest possible fidelity,
each segregation point controls a single building. As each seg- D. Agent-Based Simulation
regation point is represented in the problem’s mathematical Recent advances in simulation and vast growth in computing
formulation with a binary variable, when the fidelity level is power have made it possible to optimize complex stochastic
higher, the multiobjective optimization algorithm has to deal systems that were previously intractable [29]. One of the most
with a much more complex problem. This is because the solu- promising ways to simulate the behavior of an MG is using
tion space of the problem is exponential to the number of the an agent-based simulation model. Agent-based simulation is
segregation points. Hence, if there are N segregation points, a pioneering technique for modeling systems that contains
the solution space contains O(2N ) different solutions. As illus- several autonomous components that interact with each other.
trated in Fig. 5, if a small group of three buildings is controlled Agent-based modeling is a novel approach to the simulation
by a single segregation point, then the solution space contains of power network systems due to the multiple independent
only two potential solutions. However, if each building is con- agents (power generation resources) that interact with and feed
trolled by a unique segregation point, then the solution space into a larger network that can realistically represent the struc-
contains eight potential solutions. ture of an electricity distribution system. This larger network
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON SMART GRID

learns from their ever-increasing data-collection and adapts where PW is the power output, AW is the total swept area,
the overall system behavior to better suit the needs of its and η is the efficiency of the wind turbine generator.
customers. In the simulation model, for each second, the simulation
In our case, when creating the MG simulation model, agents triggers an event and calculates the power output of the
are designed for the loads, wind turbines, PV arrays, and diesel wind farm using the wind speed measurements and the given
generators. Our simulation model is built using a Java-based parameters of the wind turbines.
software that brings together system dynamics, discrete event, 4) Diesel Generator Agent: Diesel backup generators offer
and agent-based methods within a single model development an uninterruptible source of electricity when all other resources
environment. Details of the major agents are described below. of electricity fail. In this paper, the diesel generator agent acts
1) Demand Agent: For each building on the MG area, as a stand-by system that is activated automatically whenever
a demand agent is created. The agent reads its parameters from the system detects a lack of power. When a power loss occurs,
the database’s corresponding tables. To this end, it reads the the MG controller signals a number of generators depending
type of demand (critical, priority, noncritical), the type of load on the estimated amount of electricity needed. When power
(residential, commercial, or industrial), and the peak demand is restored or renewables are again able to meet the demand,
of the building. By using a power factor plot for the load, the the controller automatically transfers the electrical load back
current demand is calculated per second. More specifically, to the utility or the renewables. The MG then sends a signal
following the values of the power factor in the schedule, the to the generators to power down, return to standby mode, and
current demand is calculated by also adding a Gaussian noise. wait for the next power outage.
This agent contains an event that is triggered each second and
calculates the new demand and total power. III. E XPERIMENTS AND R ESULTS
2) Solar Agent: A PV cell transforms the solar irradiance
In order to demonstrate the validity of the proposed frame-
into electricity. In this paper, the solar generator agent is
work, a realistic case study is designed and simulated using
designed to simulate each solar farm in the system each con-
real data obtained from a military base. The proposed algo-
sisting of several PV cells. To this end, the agent utilizes the
rithms, described in Section III were then applied to it in order
area and efficiency of each cell along with the total number
to show the computational time reduction that the framework
of cells. It then uses the embedded functions to calculate the
offers.
power output. Each second, the simulation triggers an event
and calculates the power output for the solar farm using the
solar radiation measurements. A. Case Study Microgrid Characteristics
Equation (10) allows for the calculation of the power out- The considered simulated system is a military base in the
put from PV system with an area A(m2 ) when the total United States. The average annual demand for the simulated
solar irradiance of Ir((kWh)/m2 ) is incident on the PV MG is 130.6 GWh. In addition, as stated before, the buildings
surface [30], [31] within the MG are prioritized into critical, priority, and non-
critical demand categories. The average annualized demand
P = Ir · η · A (10)
for each demand type of the MG is 31.7 GWh of critical
where η denoted the system efficiency and ηm the module loads, 40.1 GWh of priority loads, and 58.8 GWh of non-
efficiency and are defined as critical loads. The data collected for this paper includes real
data collected over a period of two years by the National
η = ηm · ηpc · Pf (11) Renewable Energy Laboratory from sensors located approxi-
ηm = ηr · (1 − β(Tc − Tr )) (12) mately 60 miles from Panama City, Florida for solar irradiance
and wind speed. This case study is limited to 44 acres of usable
where ηr is the manufacturer’s reported efficiency, ηpc is the
land for a solar farm. The simulation makes use of SunPower
power conditioning efficiency, Pf is the packing factor, β is the
E19/425 solar panels with a rated efficiency of 19.7%. Due
array temperature coefficient, Tr is the reference temperature
to the high cost of installation and because of the renewable
for the cell efficiency, and Tc is the current temperature.
energy credits available to the commercial companies willing
3) Wind Generator Agent: Wind power is increasing in
to invest in solar power, this particular agent utilizes a power
popularity primarily due to the fact that it does not produce
purchase agreement between the MG and the commercial util-
greenhouse gas emissions while it uses a relatively insignif-
ity company (2.23 cents per kWh). Moreover, in this case
icant amount of space. Wind turbines are responsible for
study, the simulation uses the GE Class I 1.5-77 wind turbines
transforming wind energy into electricity. In the proposed sim-
(80 m tall with 37 m long blades). A two turbine system costs
ulation model, a wind generator agent is designed to compute
$9.6 million with an estimated $25 000 in operating costs.
the power output of the wind farms over short periods of time
The considered MG is designed to have five different con-
(on the order of seconds). The agent takes into consideration
trol fidelities with 1, 3, 25, 104, and 186 segregation points,
the real-time measurements of the wind speed, the number of
respectively, starting from the least detailed fidelity control
wind turbines, their efficiency, and their rated power output to
design. The weights for the critical, priority, and noncritical
calculate the overall power output. Equation (13) is used for
demands in the multiobjective optimization model have been
the actual power output of a wind turbine [31]
set to 6, 3, and 1, respectively, according to the maintenance
P = PW · AW · η (13) manager of the considered military base.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

THANOS et al.: DDDAS FRAMEWORK FOR AUTOMATED CONTROL IN MGs 7

Fig. 6. Comparison of OCBA, PTV, and EA in a five-fidelity MG. Fig. 8. Comparison of OCBA, PTV, and EA in a 125-fidelity MG.

hand, using PTV and EA algorithms results in the same P{CS}

after 2250 and 2500 replications. These results illustrate that
using the proposed algorithm is approximately four times
faster than PTV and EA.
3) MG Structure With 125 Fidelities: Fig. 8 shows that the
OCBA-based control design selection algorithm reaches the
Fig. 7. Comparison of OCBA, PTV, and EA in a 25-fidelity MG. 94.5% and 98.45% probability of correct selection after 4500
and 6500 replications, respectively. Conversely, using EA and
PTV algorithms only resulted in 93% probability of correct
selection after 36 500 and 40 500 replications, respectively.
B. Validation of the Components of the Framework
To validate the effectiveness of the proposed multiobjec-
In order to validate the performance of the proposed frame- tive optimization algorithm, we compare the quality of the
work, we first test the performance of its crucial components solutions by solving the problem of (2)–(9) within the con-
in isolation (before they are all embedded into our pro- sidered MG when the control fidelity is 3 and 5, with 25 and
posed decision framework). To this end, we first validate the 186 segregation points, respectively.
OCBA-based control design selection algorithm by testing the The objectives considered in the multiobjective optimiza-
total number of replications needed from the simulation to tion algorithm are to maximize the total energy surety
reach the desired probability of correct selection of the best by giving priority to the critical and priority loads, min-
design in the considered case study MG. We compare the total imize the operational cost, and minimize the amount of
number of replications acquired by the proposed algorithm emissions by allocating the generation resources and con-
against those obtained from traditional approaches, including trolling the segregation points. All solutions were obtained
equal allocation (EA) and proportional to variance (PTV) algo- using general algebraic modeling system (GAMS) with
rithms, with three different predetermined control design sets. solving constraint integer programs solver version 2.1.1.
The first, second, and third control design sets includes 5, 25, The experiments were conducted on a computer with 3.40 GHz
and 125 different designs (i.e., fidelities), respectively. In EA, processor and 8.00 GB memory (RAM).
the replications of the simulation are equally allocated among In order to use the ∈-constraint method in the optimization
all of the design candidates [20]. PTV can be regarded as problem described in Section III-C, the optimization prob-
a sequential version of the two-stage procedure of Rinott [32] lem was first solved for each objective function separately by
where the computing budget is allocated proportionally to the considering the rest of the objective functions as constraints.
estimated sample variances. The details of the experiments and The results obtained from that procedure led to a minimum
the results obtained from the proposed approach in compar- and maximum value for each objective function. These max-
ison with those obtained from the traditional approaches are imum and minimum values were then used for the creation
provided in Figs. 6–8. of ten intervals for both cost and emissions objective func-
1) MG Structure With Five Fidelities: The results of this tions. The minimum and maximum of every interval represents
experiment show that the OCBA-based control design selec- values of ∈i for the objective functions. In the last step,
tion algorithm reaches a probability of correct selection of the objective function with the highest priority (i.e., energy
98.5% after 35 replications while the PTV and EA algo- surety objective) was considered as the only objective func-
rithms need 55 and 65 replications to reach the same P{CS} tion of the problem, and the cost/emission objectives were
(see Fig. 6). Therefore, the OCBA-based control design selec- added as constraints allowing them to take values only within
tion algorithm performs approximately 60% and 85% better the intervals described above. Since for both cost and emis-
than PTV and EA algorithms, respectively. sions objectives ten intervals were created, their combination
2) MG Structure With 25 Fidelities: The results shown in resulted in 100 different optimization problems and, respec-
Fig. 7 reveal that 500 and 750 replications of the OCBA-based tively, in 100 distinct Pareto solutions. However, some of the
control design selection algorithm lead to a 97.9% and 99.8% obtained solutions were nonfeasible while some others were
probability of correct selection, respectively. On the other not in the Pareto frontier. In Fig. 9, we illustrate the obtained
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON SMART GRID

TABLE III
C OMPUTATIONAL T IME C OMPARISON FOR THE F RAMEWORK VALIDATION

Fig. 9. Set of Pareto solutions for fidelity 5 (left) and fidelity 3 (right).

Pareto frontier solutions solving the problem in fidelity 5 (left)

and in fidelity 3 (right), respectively.
The large (and bold) and small dots represent the
Pareto solutions, and their projections on the axis-planes,
respectively. The x-, y-, and z-axis represent the operational
cost of the MG, the emissions, and the amount of satisfied
demand, respectively.
The challenge in comparing the solutions here is that the
solutions in each case correspond to a different optimization
model, while both models are using the same input data and
are optimizing the same objectives. While the highest fidelity
model includes substantially more binary variables and hence,
acquires more solutions in the Pareto frontier, considering the
fact that we have to choose only one solution (for the MG
operation and control) and that the energy surety is the most
important objective, we should focus on the solutions that
maximize the z-axis (energy surety). We can see that in both
cases of fidelity 3 and 5, we acquire several Pareto solutions,
but when the control of the segregation points is limited, we
acquire less Pareto solutions. This is expected since in the
second case the number of the binary decision variables is
significantly limited. However, for the Pareto solutions that
are acquired in the space region where the most important
objective is optimized (z-axis for energy surety objective),
in both cases, the ∈-constraint method leads to the same
efficiency.
Fig. 10. Comparison of computational time for solving the problem.

C. Validation of the Entire Framework

Our proposed framework is applied to the considered MG problem without considering any smart control design selec-
for a period of 24 h starting at 12 A . M . for validation purposes. tion that would reduce the complexity. In Table III, the time
The MG under investigation suffers from a blackout and is in seconds is shown for both of these considered cases. Even
isolated using only its own generation resources to satisfy the though the proposed framework utilizes the simulation and
demand. the OCBA-based control design selection algorithm, the total
The main goal of the framework is to significantly speed up time of solving the framework is significantly reduced. This
the computational time of solving the aforementioned problem. occurs since with the proposed framework the complexity of
To this end, the simulation model is used for the selection of the multiobjective optimization problem substantially drops.
the most appropriate control/design mechanism, and then, the As discussed in Section II-C, the solution space is exponen-
problem is solved with the multiobjective optimization algo- tial to the number of segregation points. To this end, with the
rithm of the framework using the best selected control design proposed framework the number of segregation points is much
as an input. smaller, and thus, the multiobjective optimization algorithm
In order to validate the framework, we compare the com- solves the problem much faster.
putational time needed for solving the problem using the The total computational time needed to solve the prob-
proposed framework to the computational time needed for lem using the proposed framework in comparison to solve
solving the problem as a plain multiobjective optimization it as a plain multiobjective optimization problem is shown in
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

THANOS et al.: DDDAS FRAMEWORK FOR AUTOMATED CONTROL IN MGs 9

Fig. 10. Here, the dashed black line illustrates the amount of [5] F. Darema, “Dynamic data driven applications systems: A new paradigm
time it takes to solve the problem as a plain optimization prob- for application simulations and measurements,” in Computational
Science—ICCS (LNCS 3038). Berlin, Germany: Springer-Verlag,
lem. This time also corresponds to the computational time of Jan. 2004, pp. 662–669.
solving the considered problem using our framework in the [6] S. Karaman, R. G. Sanfelice, and E. Frazzoli, “Optimal control of
maximum possible control fidelity. The solid black line is the mixed logical dynamical systems with linear temporal logic specifi-
cations,” in Proc. IEEE Conf. Decis. Control, Cancún, Mexico, 2008,
computational time it takes to solve the problem using our pp. 2117–2122.
proposed DDDAS framework, and this time includes the time [7] Y. Ouyang, J. E. Zhang, and S. M. Luo, “Dynamic data driven applica-
required for the simulation to acquire the best possible con- tion system: Recent development and future perspective,” Ecol. Model.,
vol. 204, pp. 1–8, May 2007.
trol design. The time required to solve the considered problem [8] J. C. Carnahan and P. F. Reynolds, “Requirements for DDDAS
using our proposed framework is significantly less than that flexible point support,” in Proc. WSC, Monterey, CA, USA, 2006,
of solving the problem as a plain optimization problem. The pp. 2101–2108.
[9] C. Park, J. Tang, and Y. Ding, “Aggressive data reduction for damage
95% confidence interval reduction in the computational time detection in structural health monitoring,” Struct. Health Monit., vol. 9,
is 50.38% ± 11.09%; where in some cases even a reduction pp. 59–74, Jan. 2010.
of up to 70% is observed. [10] A. Khaleghi et al., “A DDDAMS-based planning and control framework
for surveillance and crowd control via UAVs and UGVs,” Expert Syst.
Appl., vol. 40, pp. 7168–7183, Dec. 2013.
IV. C ONCLUSION [11] E. Blasch, Y. Al-Nashif, and S. Hariri, “Static versus dynamic data infor-
mation fusion analysis using DDDAS for cyber security trust,” in
MGs are pioneering systems that offer managerial inde- Proc. ICCS, Macau, China, 2014, pp. 1299–1313.
pendence, operational independence, emergent behavior, geo- [12] A. E. Thanos, D. Moore, X. Shi, and N. Celik, “System of systems mod-
graphic distribution, and evolutionary development. This paper eling and simulation for microgrids using DDDAMS,” in A Handbook
for Modeling and Simulation for System of Systems. Hoboken, NJ, USA:
reveals that DDDAS and agent-based simulation are promis- Wiley, 2014.
ing methods to model such systems as they provide means [13] A. E. Thanos, X. Shi, J. P. Saenz, and N. Celik, “A DDDAMS frame-
to find the most efficient method for optimizing their con- work for real-time load dispatching in power networks,” in Proc. WSC,
Washington, DC, USA, 2013, pp. 1893–1904.
trol mechanism in order to enhance their operation and attain [14] N. Celik, A. E. Thanos, and J. P. Saenz, “DDDAMS-based dispatch
several benefits. Our framework includes a database respon- control in power networks,” in Proc. ICCS, Barcelona, Spain, 2013,
sible for collecting data from electrical and environmental pp. 1899–1908.
[15] J. P. Sáenz, N. Celik, H. Xi, Y. Son, and S. Asfour, “Two-stage
sensors; a fault detection/isolation algorithm that identifies economic and environmental load dispatching framework using parti-
liabilities as well as potential threats within the MG; an cle filtering,” Int. J. Elect. Power Energy Syst., vol. 48, pp. 93–110,
agent-based simulation of the MG system that includes mod- Jun. 2013.
[16] D. Allaire and K. Willcox, “A mathematical and computational frame-
ules for energy generation from renewable and fuel sources, work for multifidelity design and analysis with computer models,”
and an heuristic algorithm for fast optimization in terms of Int. J. Uncertain. Quantif., vol. 4, pp. 1–20, Jan. 2014.
cost and demand satisfaction; an OCBA-based control design [17] H. Jiang, J. J. Zhang, W. Gao, and Z. Wu, “Fault detection, identifi-
cation, and location in smart grid based on data-driven computational
selection algorithm that uses the agent-based simulation to methods,” IEEE Trans. Smart Grid, vol. 5, no. 6, pp. 2947–2956,
select the best control design of the MG; and a multiob- Nov. 2014.
jective algorithm for optimizing the decisions of the MG [18] Z. Zhang, S. Gong, A. D. Dimitrovski, and H. Li, “Time synchronization
given the best control design. The proposed work is quite attack in smart grid: Impact and analysis,” IEEE Trans. Smart Grid,
vol. 4, no. 1, pp. 87–98, Mar. 2013.
promising in determining the optimal level of detail to ana- [19] X. Zhang, S. Chen, Y. Zhu, and W. Yan, “Fault detection and diagnosis
lyze a system while minimizing the simulation time and for steam turbine based on kernel GDA,” in Proc. ICMIC, Shanghai,
required computational resources involved therein. The frame- China, 2011, pp. 58–62.
[20] C. H. Chen and L. H. Lee, Stochastic Simulation Optimization: An
work is validated using a case study MG, and the results Optimal Computing Budget Allocation, vol. 1. Hackensack, NJ, USA:
show that the computational time is significantly reduced World Scientific, 2010.
(50.3% on average) while the quality of the solutions is not [21] W. Chen, S. Gao, C. H. Chen, and L. Shi, “An optimal sample allocation
strategy for partition-based random search,” IEEE Trans. Autom. Sci.
compromised. Eng., vol. 11, no. 1, pp. 177–186, Jan. 2014.
[22] B. W. Hsieh, C. H. Chen, and S. C. Chang, “Efficient simulation-based
composition of scheduling policies by integrating ordinal optimization
ACKNOWLEDGMENT with design of experiment,” IEEE Trans. Autom. Sci. Eng., vol. 4, no. 4,
The authors would like to thank R. Belleville and T. Watley pp. 553–568, Oct. 2007.
[23] S. Yan, E. Zhou, and C. H. Chen, “Efficient selection of a set of good
for their valuable feedback. enough designs with complexity preference,” IEEE Trans. Autom. Sci.
Eng., vol. 9, no. 3, pp. 596–606, Jul. 2012.
[24] M. Bastani, A. E. Thanos, N. Celik, and C. H. Chen, “Efficient design
R EFERENCES selection in microgrid simulations,” in Proc. WSC, Savanah, GA, USA,
[1] K. Moslehi and R. Kumar, “A reliability perspective of the smart grid,” 2014, pp. 2762–2773.
IEEE Trans. Smart Grid, vol. 1, no. 1, pp. 57–64, Jun. 2010. [25] J. Xu et al., “MO2TOS: Multi-fidelity optimization with ordinal transfor-
[2] D. T. Nguyen and L. B. Le, “Optimal bidding strategy for microgrids mation and optimal sampling,” Asia-Pac. J. Oper. Res., to be published.
considering renewable energy and building thermal dynamics,” IEEE [26] G. Mavrotas, “Effective implementation of the ε-constraint method
Trans. Smart Grid, vol. 5, no. 4, pp. 1608–1620, Jul. 2014. in multi-objective mathematical programming problems,” Appl. Math.
[3] C. Ahn and H. Peng, “Decentralized voltage control to minimize distri- Comput., vol. 213, pp. 455–465, Jul. 2009.
bution power loss of microgrids,” IEEE Trans. Smart Grid, vol. 4, no. 3, [27] R. E. Steuer, Multiple Criteria Optimization: Theory, Computation and
pp. 1297–1304, Sep. 2013. Application. Hoboken, NJ, USA: Wiley, 1986, p. 546.
[4] F. A. Mohamed and H. N. Koivo, “System modelling and online opti- [28] A. E. Thanos, N. Celik, and J. P. Saenz, “An evolutionary sequential
mal management of microgrid with battery storage,” in Proc. ICREPQ, sampling algorithm for multi-objective optimization,” Asia-Pac. J. Oper.
Seville, Spain, 2007, pp. 1–5. Res., 2015.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON SMART GRID

[29] J. Xu, E. Huang, C. H. Chen, and L. H. Lee, “Simulation optimization: Nurcin Celik (M’14) received the Ph.D. degree
A review and exploration in the new era of cloud computing and big in systems and industrial engineering from the
data,” Asia-Pac. J. Oper. Res., vol. 32, no. 3, Jun. 2015, Art. ID 1550019. University of Arizona, Tucson, AZ, USA, in 2010.
[30] J. Zhao, E. Mazhari, N. Celik, and Y. J. Son, “Hybrid agent-based simu- She is an Assistant Professor with the Department
lation for policy evaluation of solar power generation systems,” Simulat. of Industrial Engineering, University of Miami,
Model. Pract. Theory, vol. 19, pp. 2189–2205, Nov. 2011. Coral Gables, FL, USA. Her current research inter-
[31] J. P. Sáenz, N. Celik, S. Asfour, and Y. J. Son, “Electric utility ests include the areas of integrated modeling and
resource planning using continuous-discrete modular simulation and decision making for large-scale and complex and
optimization (CoDiMoSO),” Comput. Ind. Eng., vol. 63, pp. 671–694, dynamic systems.
Nov. 2012. Prof. Celik was a recipient of the Eliahu I. Jury
[32] Y. Rinott, “On two-stage selection procedures and related probability Early Career Research Award in 2014, the Air
inequalities,” Commun. Stat., vol. 7, no. 8, pp. 799–811, 1978. Force Scientific Research Young Investigator Research Award in 2013 and
the International Association for Management of Technology Outstanding
Research Project Award in 2011.

Aristotelis E. Thanos (S’15) received the B.S.

and M.S. degrees in applied mathematics from the
National Technical University of Athens, Athens,
Greece, in 2010, and the M.S. degree in com-
puter science from the University of Athens,
Athens, in 2012. He is currently pursuing the
Ph.D. degree with the Department of Industrial
Engineering, University of Miami, Coral Gables,
FL, USA.
Dr. Thanos was a recipient of the University
of Miami Outstanding Graduate Research Assistant
Award in 2014-2015, and the Computer and Information Systems Division
Best Student Paper at the Industrial and Systems Engineering Research
Conference in 2015.

Chun-Hung Chen (S’91–M’94–SM’01) received

the Ph.D. degree in engineering sciences from
Harvard University, Cambridge, MA, USA, in 1994.
He is a Professor with the Department of
Systems Engineering and Operations Research,
Mehrad Bastani (S’15) received the B.Sc. degree George Mason University, Fairfax, VA, USA. He
in industrial engineering from the Sharif University was sponsored by National Science Foundation,
of Technology, Tehran, Iran, in 2011, and the M.Sc. National Institutes of Health, Department of Energy,
degree in industrial engineering from the University National Aeronautics and Space Administration,
of Tehran, Tehran, in 2013. He is currently pursuing Muscular Dystrophy Association, Air Force, and
the Ph.D. degree with the Department of Industrial Federal Aviation Administration, and has researched
Engineering, University of Miami, Coral Gables, the development of efficient methodologies for stochastic simulation optimiza-
FL, USA. tion and its applications.
Dr. Bastani was a recipient of the Computer and Prof. Chen was a recipient of several awards including the National
Information Systems Division Best Student Paper Thousand Talents Award in 2011, the Best Automation Paper Award from
at the Industrial and Systems Engineering Research the IEEE International Conference on Robotics and Automation in 2003, and
Conference in 2015. the Eliahu I. Jury Award from Harvard University, in 1994.