IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 32, NO.

3, MAY 2017 2305

Distributed Generation Monitoring for Hierarchical

Control Applications in Smart Microgrids
Younes Seyedi, Student Member, IEEE, Houshang Karimi, Senior Member, IEEE,
and Santiago Grijalva, Senior Member, IEEE

Abstract—Hierarchical control/protection applications in smart control, a microgrid central controller (MGCC) is responsible
microgrids require knowledge of real-time status of distributed for reliable and optimal operation of microgrid [5]–[7]. In gen-
generation (DG) systems. Lack or failure of communications with eral, different types of DG units may exist in microgrids and
the microgrid central controller (MGCC) can significantly under-
mine performance of such applications since the MGCC cannot their operating conditions may change over time. Centralized
determine the number of operational energy sources. To overcome power sharing, dispatching, and frequency/voltage restoration
these challenges, the MGCC needs a secondary mechanism in can be carried out if there are grid-supporting DG units that
order to track presence or absence of DG systems. This paper can bidirectionally communicate with the MGCC. In addition
proposes a new monitoring approach that empowers the MGCC to grid-supporting DG systems, there are some grid-feeding DG
to estimate the number of operational DG systems and thus deter-
mine the total generation capacity of the microgrid. A parameter units which are locally controlled by a maximum power point
estimator is developed to extract an autoregressive model for the tracking (MPPT) controller. Generally, DG units that rely on re-
synchrophasors of current symmetrical components (CSC) of the newable energy sources are grid-feeding, where the MPPT con-
main point of common coupling (PCC). The extracted model is troller determines the reference values for active powers without
used by an adaptive algorithm that identifies abrupt changes in DG communicating with the MGCC. Since grid-feeding DG units
by evaluating the norm of forward prediction error. The proposed
approach uses real-time synchrophasor data to dynamically contribute to power flow, their presence or absence can affect
update the criterion for event detection and is very robust against exchange of power between microgrid and the host grid (in the
abrupt load changes. The performance is verified using extensive secondary and tertiary control levels.) It is known that, presence
simulations of the IEEE 13-Bus benchmark with four photovoltaic or absence of DG units can also affect fault currents that flow
(PV) units. through the feeders and the point of common coupling (PCC)
Index Terms—Distributed generation (DG), hierarchical [8]. Therefore, intelligent electronic devices (IEDs) that protect
control, microgrid, monitoring, smart grid, synchrophasor. the grid should change their protection settings in order to adapt
to new fault characteristics in microgrid [9], [10].
I. INTRODUCTION In practice, lack or failure of communication systems is in-
MART microgrids and active distribution networks (ADNs) evitable for both types of DG systems. Under these circum-
S are key elements of a more intelligent power grid. Such
networks are inherently exposed to diverse operating conditions
stances, disconnection or connection of DG units cannot be
identified, and thus invalid centralized control and protection
[1] due to prevalence of distributed generation (DG) as well functionalities may be carried out. Hence, operation and pro-
as distributed energy storage systems. To achieve the optimal tection of entire microgrid become liable to disturbances and
operating conditions in smart grids, intelligent control archi- faults. To overcome the aforementioned challenges, the MGCC
tectures and data-centric tools are under development [2], [3]. needs a secondary mechanism in order to track presence or ab-
In fact, smart microgrids can effectively harness DG and DS sence of DG units based on information received from the PCC.
units to improve the grid performance in terms of power quality, By virtue of such a secondary mechanism, the MGCC can up-
reliability, efficiency, cost, and sustainability. date the pickup level of IEDs and also estimate the total rated
Hierarchical control structures employ distributed controllers power generated by operational DG units. Therefore, reliabil-
and communication systems in order to control microgrids in ity of microgrid control is not diminished by lack or failure of
three different levels [4]. In the secondary level of hierarchical communications between the MGCC and DG units.
This paper proposes a secondary monitoring approach which
empowers the MGCC to track the number of operational DG
Manuscript received March 8, 2016; revised July 16, 2016; accepted Septem-
ber 12, 2016. Date of publication September 15, 2016; date of current version units over time. DG monitoring is accomplished by detecting
April 17, 2017. Paper no. TPWRS-00372-2016. the following two events: 1) Disconnection event: when a DG
Y. Seyedi and H. Karimi are with the Department of Electrical unit is suddenly removed from the microgrid, for example, due
Engineering, Polytechnique Montreal, Montreal, QC H3T 1J4, Canada (e-mail:
[email protected]; [email protected]). to a false islanding, failure of local controller, or a local fault.
S. Grijalva is with the Georgia Institute of Technology, Atlanta, GA 30332 2) Connection event: when a DG unit becomes operational, for
USA (e-mail: [email protected]). example, after successful transition to the grid-connected mode.
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. We first show that synchrophasors of current symmetrical
Digital Object Identifier 10.1109/TPWRS.2016.2610322 components (CSC) at the PCC can be modeled as complex

0885-8950 © 2016 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.
2306 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 32, NO. 3, MAY 2017

Gaussian random processes. The basic premise is that, the over- II. A STOCHASTIC ANALYSIS OF THE PHASORS
all flow of power through the PCC encompasses different loads, OF CURRENT SYMMETRICAL COMPONENTS
generation, storage, and renewable sources. A block-based Assume that the instantaneous voltage at the main PCC and
parameter estimator is developed which sequentially extracts an
the current flowing into the main PCC and are given by
autoregressive model based on the least squares criterion. The
⎛ ⎞ ⎛ ⎞
extracted model is then used by a monitoring algorithm which Va sin(ωt + θa ) Aa sin(ωt + δa )
decouples abrupt changes in generation from changes in con- ⎜ ⎟ ⎜ ⎟
vabc = ⎝ Vb sin(ωt + θb ) ⎠ , iabc = ⎝ Ab sin(ωt + δb ) ⎠ .
sumption by evaluating the norm of prediction error. Moreover,
the performance of monitoring algorithm is investigated by as- Vc sin(ωt + θc ) Ac sin(ωt + δc )
sessment of probability of non-detection and probability of false Generally, the PCC voltage in the above equation may repre-
detection. sent an unbalanced three-phase signal. For the sake of brevity,
The concept of DG monitoring is considered by several works the dependency of amplitudes and phase angles on the time
as a solution for reliable operation of active distribution net- are dropped from the notations. The three-phase current sig-
works in smart grids [11]–[14]. The works presented in [11] and nal can be decomposed into its symmetrical components [23] as
[12] deal with a novel monitoring approach based on state esti- Iabc = Ip + In + Iz , where the subscripts p, n, and z denote the
mation and geometric tests to identify the operating conditions positive, negative, and zero sequence components, respectively.
of DG systems. The aforementioned works, which aim to iden- For the rest of our analysis, the variable x ∈ {p, n, z} represents
tify active DG sources under unbalanced conditions, are based either of these components. The phasors of the x-sequences of
on two types of measurements: real-time data provided by a current signal at the main PCC are
supervisory control and data acquisition (SCADA) system, and
pseudo-measurements based on load forecast. The authors in Ix = Ax exp(jδx ) = Ix,r + jIx,i . (1)
[13] point out that real-time monitoring using communications
between controllers, switches and smart meters can improve the In real-world microgrids, a large number of time-varying
operational stability of microgrids. An optimal sensor selection loads/sources are connected to each phase within the network.
technique is presented in [14] which uses principal component Under such circumstances, according to the central limit the-
analysis for condition monitoring of DG units in wind farms. orem [24], the algebraic sum of the demanded and generated
As a matter of fact, previous works do not address the reliability powers at any time instant follows a Gaussian distribution. Sev-
and efficiency of DG monitoring under the impact of sudden eral studies have shown that Gaussian distributions can effec-
load changes, measurement noise, and renewable sources tively describe the randomness of flow of power in smart grids
fluctuations. [20], [25]–[28]. Assume that the aggregated active and reactive
The main contributions of this paper are summarized as fol- powers from all the sources and loads within the microgrid are
lows: 1) Using offline data with predetermined detection criteria denoted by vectors p = [Pa Pb Pc ]T , and q = [Qa Qb Qc ]T ,
is not efficient under dynamic and time-varying load changes. In respectively. We define these vectors as
view of this fact, we first develop an online parameter estimator ⎡ ⎤ ⎡ ⎤
Aa Va cos(θa − δa ) Aa Va sin(θa − δa )
which adaptively extracts a realistic model for the synchropha- ⎢ ⎥ ⎢ ⎥
p  ⎣ Ab Vb cos(θb − δb ) ⎦ , q  ⎣ Ab Vb sin(θb − δb ) ⎦ .
sors of CSC. Previous works related to optimal filtering and state
estimation [15]-[20] deal with the positive sequence component Ac Vc cos(θc − δc ) Ac Vc sin(θc − δc )
of voltage at several buses (multi-node single-component). (2)
This allows them to assume that the state transition matrix is The aggregated active and reactive powers are characterized
a priori known, and the process noise is always zero-mean. by their mean vectors and covariance matrices, i.e., p ∼
However, our approach is based on all (positive, negative, and N (p̄, CP ), q ∼ N (q̄, CQ ) with
zero) sequence components of current flowing into the PCC
p̄ = E{p} = [μP ,a μP ,b μP ,c ]T (3)
(single-node multi-component). 2) Our proposed approach
requires one phasor measurement unit (PMU) [21], [22] to q̄ = E{q} = [μQ ,a μQ ,b μQ ,c ]T (4)
report synchrophasors of CSC to the MGCC via a reliable data  
communication link. Therefore, the complexity and communi- CP = E (p − p̄)(p − p̄)T (5)
cations overhead are not affected by the topology of microgrid.  
CQ = E (q − q̄)(q − q̄)T (6)
Moreover, real-time data are used to dynamically determine
the criterion for event detection. 3) Unlike previous methods In the presence of three-phase loads, transformer connec-
[11]–[14], the proposed approach uses a statistical metric to as- tions, and non-transposed lines, the flowing power at each
sess how reliably the MGCC detects disconnection/connection phase becomes interrelated with the power at the other phases.
events under noisy PMU data and random fluctuations of Therefore, CP and CQ are not necessarily diagonal matrices.
renewable energy. It turns out that, the proposed monitoring ap- Moreover, the active and reactive powers absorbed by loads are
proach has excellent robustness against abrupt load changes and generally correlated. Hence, the following cross-covariance ma-
data noise. trix is required to properly describe the random power flow at
SEYEDI et al.: DISTRIBUTED GENERATION MONITORING FOR HIERARCHICAL CONTROL APPLICATIONS IN SMART MICROGRIDS 2307

the PCC:
CP Q = E{(p − p̄)(q − q̄)T } (7)
Now, for a set of arbitrary angles Φ = (ϕ1 , ϕ2 , ϕ3 ), we define
the following trigonometric vector functions:
cz (Φ)  [ cosϕ1 cosϕ2 cosϕ3 ]T
cp (Φ)  [ cosϕ1 cos(ϕ2 + 2π/3) cos(ϕ3 − 2π/3)]T
cn (Φ)  [ cosϕ1 cos(ϕ2 − 2π/3) cos(ϕ3 + 2π/3)]T
sz (Φ)  [ sinϕ1 sinϕ2 sinϕ3 ]T
sp (Φ)  [ sinϕ1 sin(ϕ2 + 2π/3) sin(ϕ3 − 2π/3)]T
sn (Φ)  [ sinϕ1 sin(ϕ2 − 2π/3) sin(ϕ3 + 2π/3)]T Fig. 1. The concept of secondary monitoring and control by the MGCC in the
hierarchical control structure.
The x-sequences of current signal in the phasor domain are
determined by
1 Ix becomes a complex Gaussian random variable. By taking
Ix,r = [Aa Ab Ac ]cx (δa , δb , δc ) (8)
3 expectation of (14) and (15), the mean values are found as
1 μx,r = uTx,c p̄ + uTx,s q̄ (16)
Ix,i = [Aa Ab Ac ]sx (δa , δb , δc ) (9)
3
Based on (2) and after some mathematical manipulations, it can μx,i = uTx,s p̄ − uTx,c q̄ (17)
be shown that and the variances of real and imaginary parts are given by
 T  
1 1 1 2
σx,r = E (Ix,r − μx,r )(Ix,r − μx,r )T ,
cx (δa , δb , δc ) =
Aa Va Ab Vb Ac Vc
2
 
  σx,i = E (Ix,i − μx,i )(Ix,i − μx,i )T .
 p  cx (θa , θb , θc ) + q  sx (θa , θb , θc ) , (10)
Using (14) and (15), it can be shown that the variances are
and written as:
 T 2
1 1 1 σx,r = uTx,c CP ux,c + uTx,s CQ ux,s
sx (δa , δb , δc ) =
Aa Va Ab Vb Ac Vc
  + uTx,c CP Q ux,s + uTx,s CTP Q ux,c (18)
 p  sx (θa , θb , θc ) − q  cx (θa , θb , θc ) , (11) 2
σx,i = uTx,s CP ux,s + uTx,c CQ ux,c
where  stands for the element-wise product. By substituting − uTx,s CP Q ux,c − uTx,c CTP Q ux,s (19)
(10) into (8), (11) into (9), the following expressions are obtained
  Since the powers absorbed by loads or injected by DG units
Ix,r = 1T3 ux,c  p + ux,s  q vary over the time, the phasors of CSC flowing into the PCC
  are represented by time series. The time series is essentially
Ix,i = 1T3 ux,s  p − ux,c  q
comprised of Gaussian random processes.
where 13 is a column vector having three elements equal to 1,
and III. SECONDARY MONITORING OF DISTRIBUTED GENERATION
 T
1 1 1 1 The concept of secondary monitoring and control in smart mi-
ux,c =  cx (θa , θb , θc ), (12)
3 Va Vb Vc crogrids is illustrated in Fig. 1. The underlying notion is that the
 T time series of CSC conveys valuable information about the real-
1 1 1 1 time behavior of DG units in the microgrid. The synchrophasors
ux,s =  sx (θa , θb , θc ). (13)
3 Va Vb Vc of CSC flowing into the PCC are extracted by a PMU and then
The above equations can be further simplified as: reported to the MGCC via a data link. At the MGCC, first a pa-
rameter estimator identifies an autoregressive model for the time
Ix,r = uTx,c p + uTx,s q (14) series of phasors of CSC based on the received synchrophasor
packets. A monitoring algorithm uses the identified model to
Ix,i = uTx,s p − uTx,c q (15)
track disconnection or connection of DG units over the time.
According to the above expressions, the real and imaginary Upon detection of a change in DG, a secondary control appli-
parts of Ix are linear functions of the aggregated power vectors cation sends the control commands and feedback signals via
defined in (2). Since p and q have a Gaussian distribution, control links.
2308 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 32, NO. 3, MAY 2017

A. Autoregressive Model of the Phasors of CSC

Algorithm 1: The Estimation Dataset.
The stochastic analysis presented in Section II points out that 1) Initialize: k = 0:
the synchrophasors of CSC flowing into the PCC are correlated Begin with S(k) = {y0 , y0 , ..., y0 }, where y0 =
Gaussian processes with non-zero mean. Assume that the syn-   
N elements
chrophasor data packets carry the real and imaginary parts of [Aa cos(δa ) Aa sin(δa ) 0 0 0 0]T is the initial data
Ix , along with their measurement time tags. The MGCC models vector.
the time series of the synchrophasors by a vector autoregressive 2) Update: ∀k ≥ 1:
process: 1) Obtain the data vector y(k) by reading the output
of the data link at the time instant t = kTs .
x(k + 1) = Ax(k) + w(k), k = 0, 1, 2, ... (20)
2) Construct a temporary dataset S̃ by removing the
where vector y(l|k − 1) from S(k − 1), where y(l|k − 1)
∈ S(k − 1), 1 ≤ l ≤ N is the data vector whose
 T
x(k)  Ip,r (k) Ip,i (k) In ,r (k) In ,i (k) Iz ,r (k) Iz ,i (k) measurement time tag is the earliest.
3) Construct the complete dataset by appending y(k)
is the vector of actual phasors of the x-sequences at time index k. to the temporary dataset: S(k) = {y(k), S̃}.
A ∈ R6×6 is the transition matrix and w(k) ∈ R6 represents the
vector of the process noise. Assume that e(k) ∈ R6 represents
the vector of data noise which depends on measurement noise B. Online Parameter Estimation
and harmonics distortion, and is independent of the process First, the MGCC extracts an autoregressive model for the
noise. The data vector received by the MGCC is thus written as: time series of synchrophasors of CSC. To this aim, the time-
varying parameters A, Q, and m must be estimated based on
y(k) = x(k) + e(k). (21) the data vectors delivered by the data link. The online param-
eter estimation can be achieved by employing a block-based
The process noise has a Gaussian probability density function
parameter estimator and the measurement time tags reported
(PDF) with the following parameters:
by the PMU. To this aim, Algorithm 1 constructs a dataset,
m = E{w(k)}, (22) S  {y1 , y2 , ..., yN }, which is updated every Ts seconds. Once
  the estimation dataset is obtained, the vectors in S(k) are cen-
Q = E (w(k) − m)(w(k) − m)T . (23) tered according to

1 
The data noise has also a Gaussian PDF but with zero mean and N

the covariance matrix y i = yi − yi , yi ∈ S(k). (24)

N i=1
R = E{e(k)e(k)T }. The estimation of the transition matrix is such that the prediction
errors based on the centered data become as small as possible
Since e(k) depends mainly on the accuracy of the PMU in
for i = 2, ..., N . The following minimization problem is thus
measuring the synchrophasors, R is typically a priori known
constructed:
matrix. Based on the stochastic analysis presented in Section II,
1 
the following points are highlighted: N
2
1) The process noise is not necessarily zero-mean, i.e., Â = arg min yi − Θyi−1
Θ N − 1 i=2
m = 0. Therefore, the elements of x(k) are not gener-
ally zero-mean random processes. Moreover, the process By solving the above problem, the least-squares estimate of A
noises of different sequence components may not be in- is written as
dependent, i.e., Q is not necessarily a diagonal matrix. N −1
Strictly speaking, the value of Q depends on the covari- N
yi yi−1  yi−1 yi−1
T T

ances CP , CQ , and CP Q . Â(k) = (25)

i=2
N − 1 i=2 N − 1
2) If the microgrid is stable (e.g., by using decentralized
control methods [29]), then the transition matrix for x(k) For estimation of the mean vector of the process noise, it
is stable, i.e., max |λi | ≤ 1 where λi ∈ SP(A). Hence, the follows from (20) that
i
mean vector and the covariance matrix converge to fixed
values. 
i−1
x(i + 1) = Ai x(1) + Aj w(i − j). (26)
In practical microgrids, there are other factors that affect j =0
the time series of synchrophasors at the PCC. In particular,
renewable energy variations, imperfect voltage regulators and Note that e(k) ∼ N (0, R). Thus, the equality E{y(k)} =
tap changer actions result in voltage variability. Under such E{x(k)} always holds. Since A is stable (as discussed in the
circumstances, statistical analysis must be done for performance previous subsection), if the autoregressive process is stationary,
evaluation of the MGCC in terms of DG monitoring. the expected value of x(k) converges to a fixed point. Therefore,
SEYEDI et al.: DISTRIBUTED GENERATION MONITORING FOR HIERARCHICAL CONTROL APPLICATIONS IN SMART MICROGRIDS 2309

the parameter m is estimated by

Algorithm 2: Secondary Monitoring Algorithm.
∞ −1 N
  yi For a given Γ > 0, N > 0 and ∀k > N :
m̂(k) = Â (k)
i
. (27) 1) Obtain S(k) according to Algorithm 1.
N
i=0 i=1 2) Extract the PCC’s parameters Â(k) and m̂(k), based
Based on (20) and (21), it can be easily shown that the fol- on S(k).
lowing relationship holds: 3) Calculate the forward prediction error as:
y(i) − Ay(i − 1) = w(i − 1) + e(i) − Ae(i − 1). (28) ε(k) = y(k) − Â(k − 1)y(u|k − 1) − m̂(k − 1), (30)

Note that w(i − 1), e(i − 1), and e(i) are independent random where y(u|k − 1) ∈ S(k − 1), 1 ≤ u ≤ N is the data
vectors, and consequently, the estimate of the process noise vector whose measurement time tag is the latest.
covariance matrix is given by: 4) Calculate the backward mean and standard deviation
of norms as:
Q̂(k) = −R − Â(k)RÂT (k) k −1
 ε(i)
N 
  T μ(k) = , (31)
1 N
+ yi − Â(k)yi−1 yi − Â(k)yi−1 i=k −N
N −1 !
i=2 " k −1
 "  ε(i) 2
1 N   ζ(k) = # − μ(k)2 . (32)
− (y i − Â(k)yi−1 i=k −N
N
(N − 1)2 i=2
5) Calculate the norm of the positive-sequence vector as:
N 
 T
× (yi − Â(k)yi−1 (29) η(k) = y(k)  [1 1 0 0 0 0]T (33)
i=2
6) If | ε(k) − μ(k)| > Γζ(k), then proceed to step 7,
It should be noted that, the estimation dataset has a fixed size otherwise go to step 1 with time index k + 1.
equal to N , however, it is updated every Ts seconds. Hence, 7) If η(k) > η(k − 1) & (k − 1) ≥ 1, set
the recent N data vectors are always used for online parame- (k) = (k − 1) − 1.
ter estimation. This parameter estimator is adaptive because it If η(k) < η(k − 1) & (k − 1) < LD G , set
accounts for time-varying properties of the PCC’s currents in a (k) = (k − 1) + 1.
realistic manner. Otherwise, go to step 1 with time index k + 1.

C. Secondary Monitoring Algorithm

Algorithm 2 describes the procedure of DG monitoring based
on the autoregressive model identified by the proposed param- D. Performance Assessment
eter estimator. The variable (k) is the operational number of The performance of high-level control/protection applica-
microgrid which is an estimate of the number of DG units that tions that depend on DG monitoring is directly related to the
are operational (connected to the microgrid). LD G is the to- accuracy of the estimated operational number. Imperfect data
tal number of DG units deployed within the microgrid. The links and harsh power system environments intensify noise in
operational number of microgrid always satisfies inequality the synchrophasors data which in turn degrades the parameter
0 ≤ (k) ≤ LD G . It should be noted that the MGCC which estimation and model-based monitoring. Moreover, DG units
is synchronized with the PMU invokes Algorithm 2 at the time can impose voltage/current deviations due to fluctuations of
instants t = kTs , ∀k > N . their renewable energy sources. Such random fluctuations are
The DG monitoring based on Algorithm 2 relies on two pa- adversary in the sense that they increase uncertainty in the main
rameters: the size of estimation dataset N , and the deviation PCC’s dynamic behavior and may degrade decision making and
coefficient Γ. In a stochastic microgrid setting, abrupt changes monitoring.
in status of DG units result in significant deviations of the one- Statistically, the aforementioned phenomena increase the
step ahead prediction error. Based on this premise, Algorithm probability of false detection and probability of non-detection,
2 decouples connection/disconnection of DG units from sud- denoted by pf d and pn d , respectively. pf d shows the probability
den load changes by comparing the norm of prediction error that the MGCC declares a false disconnection/connection event.
with its backward average values. Afterward, the MGCC dis- pn d is the likelihood that an event of disconnection/connection
tinguishes between the connection and disconnection events by is not detected by the MGCC. Both pf d and pn d affect the esti-
assessing the norm of the positive-sequence component. Note mated operational number of the microgrid over the time. For a
that Algorithm 2 uses real-time data to dynamically determine fixed N , the optimal value of the deviation coefficient Γ yields
the criterion for event detection. This dramatically increases ro- the maximum performance as:
bustness of the monitoring process against abrupt load changes.
The optimal choice of the deviation coefficient is explained in
the next subsection. Γ∗ = arg max (1 − pf d )(1 − pn d ). (34)
Γ
2310 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 32, NO. 3, MAY 2017

Basically, the statistical quantity (1 − pf d )(1 − pn d ) shows the

efficiency of the MGCC in terms of DG monitoring. When this
quantity is equal to unity, the MGCC is able to track presence
or absence of DG units without any false event.
It should be highlighted that solving Eq. (34) to reach the
highest efficiency requires prior knowledge of load variability
patterns. This knowledge can be obtained by two classes of
methods: a) offline load characterization, and b) online load
characterization. Offline methods analyze empirical ac-
tive/reactive power profiles at different buses and are performed
at the microgrid design and planning stage. Online methods
estimate active/reactive power profiles based on collected in-
formation from smart meters [30]. Online load characterization
methods can be easily implemented as part of the MGCC in the
control structure depicted in Fig. 1.
Basically, the MGCC evaluates the variability patterns by
processing the smart meter data in the microgrid. This is ac-
Fig. 2. The single-phase circuit diagram of the test microgrid and the locations
complished by extracting the mean and the standard deviation of DG units.
of the loads. Based on these statistics and the knowledge of the
rated power of DG units, a grid-search method finds the solution
for Eq. (34). Then, the MGCC periodically updates the param-
are statistically independent of each other. Two unbalanced load
eters N and Γ, which are optimal according to the variability
profiles with high and low degree of variability are given in
pattern of the loads.
Table I. With the low variability profile, the mean and the stan-
Apart from data noise and fluctuations in renewable energy,
dard deviation of active power demanded by a local load are
bad PMU data may affect the performance of the DG monitor-
60 kW and 3.46 kW, respectively. With the high variability pro-
ing algorithm. In general, loss of time synchronization, PMU
file, the mean value of active power is 60 kW but the standard de-
failures and communications errors result in bad data. When
viation increases to 14.43 kW. In particular, the high variability
a single PMU is deployed at the main PCC, an elaborate data
profile in the unbalanced scenario is used to investigate the abil-
validation algorithm, such as the one developed in [31], is able
ity of the proposed monitoring approach to discriminate between
to detect outliers in the synchrophasor data stream and replace
significant load changes and DG connection/disconnection.
them with correct data. Moreover, availability of measurement
diversity through installation of more PMUs can enhance the
quality of data conditioning [32]. Therefore, the MGCC can A. Monte Carlo Simulations
prevent bad PMU data from entering the monitoring process by The aim of Monte Carlo load flow simulations is to numer-
adding such data validation modules to the proposed structure. ically validate the stochastic analysis of the synchrophasors of
The augmented monitoring structure mitigates the probability CSC presented in Section II. In order to observe the actual dis-
of false detection and improves the reliability of secondary con- tribution of CSC, DG units do not inject any power and the
trol/protection applications. synchrophasor data do not incur any measurement noise. The
statistically balanced and unbalanced (with high variability pro-
IV. SIMULATION RESULTS file) scenarios are simulated with per-phase distributions given
by Table I. For the main PCC, the cumulative distribution func-
The IEEE 13-Bus radial distribution network [33] with the tions (CDFs) of the real part of the phasor of the negative-
single-phase diagram depicted in Fig. 2 is adopted as the test sequence component are depicted in Fig. 3. The solid curves
microgrid. The microgrid circuit and the MGCC’s algorithms show the simulated CDF and the dashed curves show the theo-
are implemented in MATLAB/Simulink. Bus 1 is the main PCC retical CDF of a Gaussian process. For the balanced scenario, the
where the microgrid is connected to the host grid. The host grid is Gaussian CDF has parameters μn ,r = 0, σn ,r = 2.37, and for
characterized by the short circuit power 300 MVA and the short the unbalanced scenario with the high variability profile, the the-
circuit ratio SCR = 7. The nominal voltage of the network is oretical CDF is plotted with parameters μn ,r = 8.9, σn ,r = 1.9.
15 kV and the system frequency is 60 Hz. The test microgrid has The excellent match between the simulated and theoretical
four DG units (LD G = 4 ) driven by photovoltaic (PV) energy CDFs confirms that the synchrophasors of CSC are Gaussian
sources. The DG units have identical rated power and operate random processes as discussed in Section II.
under unity power factor in the current control mode.
The distributions of active and reactive powers demanded by
the PQ loads are given in Table I, where U[α, β] stands for B. Time Domain Simulations
the uniform PDF over the interval [α, β], inclusive of the lim- In the time domain simulations, a PQ load is randomly se-
its. In all simulations, the PDFs of demanded powers are the lected every Ts = 1 sec, and its demanded powers change ac-
same among all buses. Moreover, the loads at different buses cording to the unbalanced scenarios in Table I. Voltage/current
SEYEDI et al.: DISTRIBUTED GENERATION MONITORING FOR HIERARCHICAL CONTROL APPLICATIONS IN SMART MICROGRIDS 2311

TABLE I
THE PER-PHASE DISTRIBUTIONS OF THE PQ LOADS IN DIFFERENT SCENARIOS

Statistically Balanced Statistically Unbalanced (High Variability Profile) Statistically Unbalanced (Low Variability Profile)

Active power (kW) Reactive power (kVAr) Active power (kW) Reactive power (kVAr) Active power (kW) Reactive power (kVAr)

Phase a U[0, 20] U[0, 5] U[20, 40] U[0, 20] U[28, 32] U[8, 12]
Phase b U[0, 20] U[0, 5] U[10, 30] U[0, 10] U[18, 22] U[3, 7]
Phase c U[0, 20] U[0, 5] U[5, 15] U[0, 5] U[8, 12] U[1, 5]

Fig. 3. The simulated (dashed curves) and theoretical (solid curves) CDFs of
In , r (k) at the main PCC.
Fig. 4. The per-phase active power injected by DG units in the period of one
day: (a) Sunny day (b) Cloudy day.
fluctuations due to solar energy variability are taken into ac-
count by using realistic solar power data measured in the state
of Georgia [34]. Two daily scenarios are adopted with differ-
ent spatio-temporal properties of solar energy: sunny day with
clear sky (high degree of correlation between DG units), and
cloudy day with scattered clouds in sky (low degree of cor-
relation between DG units). After a level scaling, the actual
powers generated by the DG units in the two daily scenarios
are shown in Fig. 4(a) and (b). It is evident that, the cloudy
day results in the highest degree of voltage/current variability
and thus corresponds to the worst-case scenario. A sequence of
connection/disconnection events are intentionally generated as
follows: DG 3 is disconnected at t = 150 min, and reconnects
at t = 300 min, DG 4 is disconnected from the network at t =
250 min, and then reconnects with the network at t = 350 min.
Figs. 5–7 show the time domain results for the sunny day in
the unbalanced scenarios. Fig. 5(a) illustrates abrupt peaks in
the norm of forward prediction error caused by changes in the Fig. 5. DG monitoring over the time in the unbalanced scenario with the low
variability profile: R = 10 −2 I, N = 120, Γ = 4.
status of DGs 3 and 4 under the low variability profile. In this
scenario, the standard deviation of local loads is 3.46 kW which
is much smaller than the generation capacity of DG units in the scenario, the standard deviation of local loads is 14.43 kW
sunny day. The monitoring algorithm is able to properly update which is quite comparable to the rated powers of DG units.
the operational number of the microgrid over the time, as shown However, the monitoring algorithm is able to disregard sud-
in Fig. 5(b). den load changes and prevents spurious detections as indicated
The time domain results with the high variability profile are in Fig. 6(b). To elucidate the results, connection/disconnection
shown in Figs. 6 and 7. In contrast with Figs. 5(a) and 6(a) ex- events are detectable as long as the rated power of DG units is
hibits major deviations which are caused by significant load comparable to their local loads and the standard deviation of the
changes under the high variability profile. Note that in this loads is not dominating over DG capacities.
2312 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 32, NO. 3, MAY 2017

Fig. 8. The probability of false detection vs. the size of estimation dataset:
Fig. 6. DG monitoring over the time in the unbalanced scenario with the high Unbalanced scenario with the high variability profile, R = 10−2 I.
variability profile: R = 10 −4 I, N = 120, Γ = 4.

Fig. 9. The probability of non-detection vs. the size of estimation dataset:

Fig. 7. DG monitoring over the time in the unbalanced scenario with the high Unbalanced scenario with the high variability profile, R = 10−3 I.
variability profile: R = 10 −2 I, N = 120, Γ = 4.

increasing the deviation coefficient. The probability of a false

The impact of data noise combined with huge load changes
detection in the cloudy day is slightly higher which means that
under the high variability profile is inspected in Fig. 7. It can be
the estimated operational number is more liable to error under
observed that, a false detection occurs at t = 190 min. However,
high DG variability and severe data noise.
the disconnection events at t = 150 min and t = 250 min and
Fig. 9 illustrates the probability of non-detection for differ-
the connection events at t = 300 min have been detected. The
ent values of N . For the sunny day, if the size of estimation
connection event at t = 350 min is actually detected but it is dis-
dataset is large enough, then the accuracy of the parameter
regarded since the operational number of the microgrid cannot
estimator is very good and thus the MGCC detects most of
exceed the total number of DG units. It can be inferred that, the
disconnection/connection events. It is concluded that, the size
data noise is detrimental to the secondary DG monitoring only
of estimation dataset directly affects the estimation error in the
if the standard deviation of loads is comparable to the injected
PCC’s parameters. On the other hand, for the cloudy day, a
powers by DG units.
bigger estimation dataset can slightly decrease the probability
of non-detection. This implies that, substantial fluctuations in
C. Performance Simulations
the renewable energy can obscure the disconnection/connection
Extensive simulations are conducted in order to numerically events during the monitoring process.
evaluate the performance of the model-based monitoring algo- The efficiency of the monitoring algorithm as a function of
rithm presented in Section III. Fig. 8 shows the probability that the deviation coefficient is demonstrated in Figs. 10 and 11. The
the MGCC detects at least one false event during each day. results shown in Fig. 10 reveal the impact of load variability
It is evident that, the size of estimation dataset should be large on the performance of the MGCC in terms of event detection.
enough to ensure reliable performance of the MGCC. Moreover, When the standard deviation of local loads is smaller than the
the monitoring algorithm becomes robust against data noise by DG capacities, the optimal value of Γ yields the maximum
SEYEDI et al.: DISTRIBUTED GENERATION MONITORING FOR HIERARCHICAL CONTROL APPLICATIONS IN SMART MICROGRIDS 2313

results presented in Fig. 11 indicate that the optimal value of the

deviation coefficient is very similar for both sunny and cloudy
days.

V. CONCLUSION
In this work, we have developed a new autoregressive model-
based approach for secondary monitoring of DG units in
smart microgrids. By implementing the proposed approach, the
MGCC can estimate the number of operational DG units at any
time instant. A stochastic analysis shows that the synchropha-
sors of CSC flowing through the main PCC can be modeled as
Gaussian random processes. A block-based parameter estimator
has been developed which adaptively extracts the autoregressive
model of the synchrophasors of CSC based on the least squares
Fig. 10. The efficiency of the monitoring algorithm as a function of the devi-
criterion. The extracted model is employed by the monitoring al-
ation coefficient: Sunny day, R = 10−2 I, N = 180 gorithm which identifies disconnection/connection of DG units
belonging to the microgrid. Simulation results confirm that the
monitoring algorithm dynamically adapts its criterion for event
detection and thus reliably tracks the operational number of
microgrid. By employing distributed energy storage systems in
smart microgrids, severe renewable energy fluctuations can be
alleviated and the reliability of the DG monitoring process is
significantly improved.

REFERENCES
[1] S. M. Ashabani and Y. A. I. Mohamed, “New family of microgrid control
and management strategies in smart distribution grids—Analysis, compar-
ison and testing,” IEEE Trans. Power Syst., vol. 29, no. 5, pp. 2257–2269,
Sep. 2014.
[2] A. P. Sakis Meliopoulos et al., “Smart grid technologies for autonomous
operation and control,” IEEE Trans. Smart Grid, vol. 2, no. 1, pp. 1–9,
Mar. 2011.
[3] M. Kezunovic, L. Xie, and S. Grijalva, “The role of big data in improving
power system operation and protection,” in Proc. IREP Symp.-Bulk Power
Syst. Dyn. Control - IX, 2013, pp. 1–9.
Fig. 11. The efficiency of the monitoring algorithm as a function of the [4] J. M. Guerrero et al.,“Advanced control architectures for intelligent
deviation coefficient: Unbalanced scenario with the high variability profile, microgrids-Part I: Decentralized and hierarchical control,” IEEE Trans.
R = 10 −3 I, N = 180. Ind. Electron., vol. 60, no. 4, pp. 1254–1270, Apr. 2013.
[5] D. E. Olivares et al., “Trends in microgrid control,” IEEE Trans. Smart
Grid, vol. 5, no. 4, pp. 1905–1919, Jul. 2014.
[6] J. M. Guerrero, “Hierarchical control of droop-controlled AC and DC
efficiency of 100%, i.e., the MGCC is able to detect every microgrids—A general approach toward standardization,” IEEE Trans.
connection/disconnection event without any false event. By Ind. Electron., vol. 58, no. 1, pp. 158–171, Jan. 2011.
[7] O. Palizban and K. Kauhaniemi, “Secondary control in AC microgrids
increasing the standard deviation of loads, the performance challenges and solutions,” in Proc. Int. Conf. Smart Cities Green ICT
decreases and the maximum efficiency reduces to 93.4%. This Syst., May 2015, pp. 1–6.
implies that with the high variability profile, significant load [8] G. Antonova et al., “Distributed generation and its impact on power grids
and microgrids protection,” in Proc. Protective Relay Eng., Apr. 2012,
changes combined with data noise may obscure some of the pp. 152–161.
DG connection/disconnection events. [9] W. K. A. Najy, H. H. Zeineldin, and W. L. Woon, “Optimal protection
The results shown in Fig. 11 reveal the impact of renewable coordination for microgrids with grid-connected and islanded capability,”
IEEE Trans. Ind. Electron., vol. 60, no. 4, pp. 1668–1677, Apr. 2013.
energy fluctuations on the efficiency of DG monitoring. Since [10] Y. Zheng, D. Wang, and Z. Zhou, “Hierarchical protection control system
the variance of data noise is reduced by 10 dB, the maximum of smart substations,” J. Modern Power Syst. Clean Energy, vol. 2, no. 3,
efficiency in the sunny day with the high variability profile pp. 282–288, Sep. 2014.
[11] A. Souza, E. M. Loureno, and A. S. Costa, “Real-time monitoring of
arises to 97%. This observation substantiates robustness and distributed generation through state estimation and geometrically-based
reliability of the proposed monitoring approach under signif- tests,” in Proc. Bulk Power Syst. Dyn. Control (iREP)-VIII (iREP),
icant load changes and trivial renewable energy fluctuations. Aug. 2010, pp. 1–8.
[12] D. Issicaba, A. S. Costa, and J. L. Colombo, “Real-time monitoring of
The maximum efficiency in the cloudy day is 29% less than points of common coupling in distribution systems through state estima-
the maximum efficiency in the sunny day. Distributed energy tion and geometric tests,” IEEE Trans. Smart Grid, vol. 7, no. 1, pp. 9–18,
storage systems can be used to mitigate non-trivial renewable Jan. 2016.
[13] R. Jingding, C. Yanbo, and Z. Lihua, “Discussion on monitoring scheme
energy fluctuations [35], and restore the performance in cloudy of distributed generation and micro-grid system,” in Proc. Int. Conf. Power
days with high variability profiles. Moreover, the numerical Electron. Syst. Appl., Jun. 2011, pp. 1–6.
2314 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 32, NO. 3, MAY 2017

[14] Y. Wang, X. Ma, and J. Malcolm, “Enhancing condition monitoring of [34] NREL, “Solar power data for integration studies datasets.” (2006).
distributed generation systems through optimal sensor selection,” in Proc. [Online]. Available: http://www.nrel.gov/
Annu. Conf. IEEE Ind. Electron. Soc., Nov. 2013, pp. 7610–7616. [35] M. J. E. Alam, K. M. Muttaqi, and D. Sutanto, “Mitigation of rooftop solar
[15] K. R. Shih and S. J. Huang, “Application of a robust algorithm for dynamic PV impacts and evening peak support by managing available capacity of
state estimation of a power system,” IEEE Trans. Power Syst., vol. 17, distributed energy storage systems,” IEEE Trans. Power Syst., vol. 28,
no. 1, pp. 141–147, Feb, 2002. no. 4, pp. 3874–3884, Nov. 2013.
[16] Z. Huang, K. Schneider, and J. Nieplocha, “Feasibility studies of applying
Kalman filter techniques to power system dynamic state estimation,” in Younes Seyedi (S’12) received the B.Sc. degree from
Proc. 8th Int. Power Eng. Conf., Dec. 2007, pp. 376–382. the University of Tehran, Tehran, Iran, in 2010, and
[17] J. Liu et al., “Trade-offs in PMU deployment for state estimation in active the M.Sc. degree from Tehran Polytechnic, Tehran,
distribution grids,” IEEE Trans. Smart Grid, vol. 3, no. 2, pp. 915–924, Iran, in 2012, both in electrical engineering. He
Jun. 2012. is currently working toward the Ph.D. degree in
[18] S. Sarri et al., “State estimation of active distribution networks: Com- the Department of Electrical Engineering, Polytech-
parison between WLS and iterated Kalman-filter algorithm integrating nique Montreal, Montreal, QC, Canada. His recent
PMUs,” in Proc. 3rd IEEE PES Innov. Smart Grid Technol. Eur. Conf., research is focused on data-driven approaches for
Oct. 2012, pp. 1–8. power grid monitoring and control, intelligent pro-
[19] J. Zhang et al., “A two-stage Kalman filter approach for robust and real- tection applications for smart grids, and synchropha-
time power system state estimation,” IEEE Trans. Sustain. Energy, vol. 5, sor networks. He is also a Research Assistant at
no. 2, pp. 629–636, Apr. 2014. Polytechnique Montreal.
[20] L. Zanni et al., “Probabilistic assessment of the process-noise covariance
matrix of discrete Kalman filter state estimation of active distribution
networks,” in Proc. Int. Conf. Probabilistic Methods Appl. Power Syst.,
Jul. 2014, pp. 1–6. Houshang Karimi (S’03–M’07–SM’12) received
[21] IEEE Standard for Synchrophasor Measurements for Power Systems, Std. the B.Sc. and M.Sc. degrees from Isfahan University
C37.118.1 2011, Dec. 28, 2011. of Technology, Isfahan, Iran, in 1994 and 2000, re-
[22] J. D. L. Ree, V. Centeno, J. S. Thorp, and A. G. Phadke, “Synchronized spectively, and the Ph.D. degree from the University
phasor measurement applications in power systems,” IEEE Trans. Smart of Toronto, Toronto, ON, Canada, in 2007, all in elec-
Grid, vol. 1, no. 1, pp. 20–27, Apr. 2010. trical engineering. He was a Visiting Researcher and
[23] G. C. Paap, “Symmetrical components in the time domain and their appli- a Postdoctoral Fellow in the Department of Electrical
cation to power network calculations,” IEEE Trans. Power Syst., vol. 15, and Computer Engineering, University of Toronto,
no. 2, pp. 522–528, May 2000. from 2001 to 2003 and from 2007 to 2008, respec-
[24] A. Papoulis and S. U. Pillai, Probability, Random Variables, and Stochastic tively. He was with the Department of Electrical En-
Processes, 4th ed. New York, NY, USA: McGraw-Hill, 2002. gineering, Sharif University of Technology, Tehran,
[25] M. He and J. Zhang,“A dependency graph approach for fault detection Iran, from 2009 to 2012. From June 2012 to January 2013, he was a Visiting
and localization towards secure smart grid,” IEEE Trans. Smart Grid, Researcher in the ePower Lab, Department of Electrical and Computer Engi-
vol. 2, no. 2, pp. 342–351, Jun. 2011. neering, Queens University, Kingston, ON, Canada. He joined the Department
[26] G. Valverde, A. T. Saric, and V. Terzija, “Stochastic monitoring of distri- of Electrical Engineering, Polytechnique Montreal, Montreal, QC, Canada, in
bution networks including correlated input variables,” IEEE Trans. Power 2013, where he is currently an Assistant Professor. His research interests include
Syst., vol. 28, no. 1, pp. 246–255, Feb. 2013. control systems, distributed generations, and microgrid control.
[27] R. Singh, B. C. Pal, and R. A. Jabr, “Statistical representation of distri-
bution system loads using Gaussian mixture model,” IEEE Trans. Power
Syst., vol. 25, no. 1, pp. 29–37, Feb. 2010. Santiago Grijalva (M’02–SM’07) received the
[28] A. Bracale et al.,“A Bayesian-based approach for a short-term steady- M.Sc. and Ph.D. degrees in electrical and computer
state forecast of a smart grid, ” IEEE Trans. Smart Grid, vol. 4, no. 4, engineering from the University of Illinois at Urbana-
pp. 1760–177, Dec. 2013. Champaign, Champaign, IL, USA, in 1999 and
[29] J. Rocabert et al., “Control of power converters in AC microgrids,” IEEE 2002, respectively. He is the Georgia Power Distin-
Trans. Power Electron., vol. 27, no. 11, pp. 4734–4749, Nov. 2012. guished Professor of Electrical and Computer Engi-
[30] E. J. Palacios-Garca et al., “Smart metering system for microgrids,” in neering and the Director of the Advanced Computa-
Proc. 41st Annu. Conf. IEEE Ind. Electron. Soc., Nov. 2015, pp. 003289– tional Electricity Systems (ACES) Laboratory, Geor-
003294. gia Institute of Technology, Atlanta, GA, USA. His
[31] K. D. Jones, A. Pal, and J. S. Thorp, “Methodology for performing syn- research interest is on decentralized power system
chrophasor data conditioning and validation,” IEEE Trans. Power Syst., control, power system informatics and economics,
vol. 30, no. 3, pp. 1121–1130, May 2015. and future sustainable energy systems. He has been a Principal Investigator
[32] M. B. Do Coutto Filho, J. C. Stacchini de Souza, and M. A. Ribeiro for various research projects under the Department of Energy, ARPA-E, EPRI,
Guimaraens, “Enhanced bad data processing by phasor-aided state es- PSERC, NSF, and other industry and Government sponsors. From 2002 to 2009,
timation,” IEEE Trans. Power Syst., vol. 29, no. 5, pp. 2200–2209, he was with PowerWorld Corporation as a Software Architect and Consultant.
Sep. 2014. From 2013 to 2014, he was on assignment to the National Renewable Energy
[33] W. H. Kersting, “Radial distribution test feeders,” IEEE Trans. Power Laboratory (NREL) as a founding Director of the Power System Engineering
Syst., vol. 6, no. 3, pp. 975–985, Aug. 1991. Center (PSEC).