Inertia estimation of equivalent areas by a
PMU-based approach following perturbations
G. R. Moraes, A. Berizzi, V. Ilea, G. D’Antona
Department of Energy
Politecnico di Milano
Milan, Italy
Abstract—This paper investigates the problem of inertia es-
timation of power system equivalent areas by a PMU-based
approach. Assuming measurements available on the borders of
the area, a reduced model is built based on the data acquired
during a perturbation. An iterative method is proposed to
calculate the parameters of an equivalent generator, and in
particular its inertia. The dynamic behavior of the equivalents
are compared with the behavior of the Center of Inertia (COI) of
each area, and the equivalent inertia is estimated using a Least
Squares approach. Results obtained for a 11-bus, 4 generators
test-system are presented and discussed.
Index Terms—Inertia estimation, dynamic equivalent, model
reduction, phasor measurement units.
I. I NTRODUCTION
The behavior of the system during the transient period
shortly after the occurrence of a perturbation in the grid is
governed by the inertia of each machine in operation: it is the
so-called inertial response. The increasing penetration of RES-
based intermittent fonts mainly connected to the power system
by means of converters, and the consequent disconnection
of traditional synchronous generation, are more and more
reducing the overall power system inertia, thus increasing
the probability of large frequency transients following large
perturbations. For that reason, estimating inertia in real-time
is becoming more and more important for the overall security
assessment function.
Recently, the installation of Phasor Measurement Units
(PMUs) has brought the possibility of acquiring measurements
in real-time in a precise and synchronized way in various
points of the grid. Many studies concerning on-line inertia
estimation were developed in the past few years, normally
divided in small-perturbation [1], [2], large-perturbation [3]–
[6] and steady-state [7], [8] approaches. The former approach
is based on analysis of oscillations, but some techniques can
be applied for large-perturbation analysis too. The second
approach is based on detection of perturbations on the system,
and is the field of interest of this paper. The latter is based on
probabilistic analysis of historical data.
The post-detection methods are mainly based on solving
the swing equation to estimate the inertia of the machine.
The works [3], [4] propose to restate the swing equation in
terms of only variables measured by PMUs, solving it for
selected data according to the moment of detection of the
perturbation. The approaches are simple enough to be easily
implemented and present robustness to false detections. The
978-1-5386-5186-5/18/$31.00 © 2018 IEEE
works [5], [6] instead, proposed Extended Kalman Filters with
recurrent estimation, that depend on model calibration and
interpolation, but presented robust results independent on the
perturbations tested. Although the cited methods work well,
a known limitation is the need of monitoring individually
all the generators connected on the grid, which makes these
approaches unfeasible in real power systems nowadays.
Reference [1] introduced an Interarea Model Estimation
(IME) method to build dynamical equivalents based on mea-
surements of interarea oscillations, acquired from PMUs at
the boundaries. The areas are determined following coherency
criteria, and the parameters can be estimated solving equations
of bus voltage and bus frequency oscillations as function of
their location in the transmission path.
However, one may be interested on the dynamical behavior
of an area monitored by a PMU independent of the coherency
of the internal machines, especially for frequency stability
issues. This paper proposes an approach to determine dynam-
ical equivalents considering the measurements of the transient
period following a perturbation. Not dealing with oscillations,
the coherency and modal decomposition are not considered
to build the equivalents. An IME technique similar to [1] is
proposed, with the goal of determining the equivalent inertia
of the considered areas, taking into account also resistances of
transmission lines (neglected in other studies) by means of a
novel iterative strategy; the proposed method requires at least
three measurement points, that could also be reduced to two,
and it is based on the determination of equivalent reactances
first and on the estimation of inertias subsequently, by means
of a combination of Least Squares Estimation and iterative
procedures.
Numerical results obtained with a two-area 11-bus test
system are presented to illustrate the proposed methodology.
The occurrence of a perturbation inside each one of the areas
is studied in two different cases, and the dynamic behavior of
the equivalent built is compared to the dynamic behavior of
the center of inertia of each area.
The remaining of this paper is organized as follows: Section
2 presents the methodology to build the dynamic equivalents
and estimate the inertia of the area from the measurements
acquired. Section 3 presents numerical results and analysis of
the method performance. Main conclusions are presented in
Section 4.
 II. M ETHODOLOGY imag(ZiE2 ) imag(ZTot )
µAi = arctan − arctan (6)
real(ZiE2 ) real(ZTot )
Given a large power system, and a couple of buses 1 and 2,
where the system has to be reduced, the goal of this paper is to imag(ZiE1 ) imag(ZTot )
µBi = arctan − arctan (7)
find out the parameters of the two equivalent generators GA1 real(ZiE1 ) real(ZTot )
and GA2 of Figure 1. Each area is represented by a voltage ZTot = X1 + X2 + ZLT (8)
source E1 and E2 behind its own equivalent reactance X1 and
X2 . The two areas are interconnected by an interconnection Step 3: Consider its variation ∆Vi with respect to a voltage
line with impedance ZLT . Measurements are assumed to be angle change linearized about an equilibrium point 0
acquired at buses 1 and 2 by PMUs.
E1 (t)E2 (t)ΨAi ΨBi sin ($i (t))
∆Vi = [∆δ2 − ∆δ1 ] (9)
GA1 1 GA2 |V0i |
2
ZLT
Step 4: Define the normalized voltage Vni as
x1 x2
E2
E1
Vni (t) , ∆Vi (t)V0i (10)
where it is worth noticing that both ∆Vi (t) and Vni (t) are time
dependent. As the PMUs will provide a number of samples
within the considered time interval, the normalized voltages
can be computed based on the PMU output. The approach
for this step is different from the one proposed in [1], as it
Fig. 1. Equivalent system considers more than one point in the equations.
Step 5: Let us assume that three PMUs are installed at
The procedure to determine the equivalent parameters of points 1, 2 and 3 (at some point in the interconnection line).
each area starts with the estimation of equivalent reactances Point 3 can be either a physical bus (where a real PMU is
and then moves to the estimation of inertias HA1 and HA2 . installed) or a virtual bus, where data can be computed based
First, a determination of reactances in the assumption of loss- on the data coming from the PMUs installed at 1 and a 2, [1].
less system takes place as it is done in [1]; losses are eventually Write the following expressions:
considered and the updated reactances so determined are used Vn3 (t) ΨA3 ΨB3 sin ($3 (t))
to compute the equivalent inertias by means of the swing = (11)
Vn1 (t) ΨA1 ΨB1 sin ($1 (t))
equation and the use of least squares method. At this point
this paper differs totally on the approach of [1], that proposes Vn2 (t) ΨA2 ΨB2 sin ($2 (t))
= (12)
modal analysis together with an extrapolation technique. Vn1 (t) ΨA1 ΨB1 sin ($1 (t))
The methodology can be summarized in the following steps: This expressions cannot be solved in this step because δ1 (t)
Step 1: The voltage Vi of each measurement point i, can and δ2 (t) (that appear inside the functions $i (t)) are not
be expressed in terms of the electrical distance to the voltage directly measured.
sources E1 and E2 : Step 6: Consider for the moment the resistance of the trans-
Vi (t) = ZiE2 E1 (t) + ZiE1 E2 (t) (1) mission line negligible (the relevant results are characterized
by the subscript I): RT L = 0.
where ZiE2 is the electrical distance of point i to the voltage Taking one of the generators as a reference (e.g. δ1 = 0),
source E2 and ZiE1 is the electrical distance of point i to the expressions 11 and 12 can be simplified to be dependent only
voltage source E1 . Both ZiE1 and ZiE2 are functions of the on system parameters:
unknowns X1 , X2 and known ZLT .
Step 2: Take the magnitude of Vi . Vn3I (t) X3E2 X3E1
= (13)
q Vn1I (t) X1E2 X1E1
Vi (t) = E1 (t)2 Ψ2Ai + E2 (t)2 Ψ2Bi + Vn2I (t) X2E2 X2E1
= (14)
2E1 (t)E2 (t)ΨAi ΨBi cos($i (t)) (2) Vn1I (t) X1E2 X1E1
Call the right hand side of equations 13 and 14 as
where E1 and δ1 are the magnitude and angle of the voltage
E1 , while E2 and δ2 are the magnitude and angle of the X3E2 X3E1
αI = (15)
voltage E2 and the other parameters are defined below. X1E2 X1E1
|ZiE2 | X3E2 X3E1
ΨAi = (3) βI = (16)
|ZTot | X1E2 X1E1
|ZiE1 | where the subscript I denotes the lossless case.
ΨBi = (4) Rewrite expressions 13 and 14 as:
|ZTot |
$i (t) = µAi − µBi + δ1 (t) − δ2 (t) (5) Vn3I = αI Vn1I (17)
 Vn2I = βI Vn1I (18)
where Vn1I , Vn2I and Vn1I are vectors consisting of a limited
number of samples of the variables Vn1 (t),Vn2 (t) and Vn3 (t),
respectively. In the present work, a limited time-window of 2s
around the moment the perturbation takes place is selected.
The over-determined systems (17) and (18) can be solved
for αI and βI using the linear Least Squares method:
α̂I = (Vn1I T Vn1I )−1 Vn1I T Vn3I (19)
β̂I = (Vn1I T Vn1I )−1 Vn1I T Vn2I (20)
With the solutions of (19) and (20), equations (15) and (16)
can be solved for X1I and X2I , the internal reactances of the
equivalent machines G1 and G2 for the lossless case.
Step 7: At this point, in order to take into account the
resistive feature of the interconnection line, which is one of
the original contributions of the present paper, the estimated
X1I and X2I can be combined with the measurement coming
from the PMUs to compute the estimate of internal voltages
E1I (t) = E1I (t)∠δ1I (t) and E2I (t) = E1I (t)∠δ2I (t) through
Ohm’s law.
Step 8: The values of the internal voltages newly computed
make it possible to come back to expressions (11) and (12).
As δ1 (t) and δ2 (t) are not directly measured, the angles δ1I (t)
and δ2I (t) obtained in Step 5 are used in this step as an
approximation.
The values of the internal voltages newly computed make
it possible to determine an improved linearization of voltage
magnitudes at the installation point of PMUs and to pro-
vide updated expressions for the updated normalized voltages
Vn1II , Vn2II and Vn3II :
Vn3II ΨA3 ΨB3 sin (µA3 − µB3 − δ1I − δ2I )
= (21)
Vn1II ΨA1 ΨB1 sin (µA1 − µB1 − δ1I − δ2I )
Vn2II ΨA2 ΨB2 sin (µA2 − µB2 − δ1I − δ2I )
= (22)
Vn1II ΨA1 ΨB1 sin (µA1 − µB1 − δ1I − δ2I )
where the underline denotes vector and the subscript II
denotes the present step, considering RT L 6= 0.
Calling αII the right hand side of equation (21), and βII the
right hand side of equation (22), both underdetermined sytems
can be solved using least squares method to obtain X1II and
X2II , similar to what was done in Step 6.
At this point, again, the updated internal voltages EiII (t)
can be computed by Ohm’s law, using the newly estimated
X1II and X2II .
Step 9: Test if XjI − XjII is lower than a tolerance
(10−3 p.u. used), where j = 1, 2 denotes the Area of the
calculated equivalent reactance. If not, set XjI = XjII and
go back to Step 7. If yes, the iterative process of reactance
estimation converged. Proceed to Step 10.
Step 10: In this step, the area equivalent inertia is estimated.
HAj d2 δi (t)
= ∆PAj (t), p.u. (23)
πf 0 dt2
where HAj is the inertia constant and ∆PAj is the power
imbalance of Area j (j = 1, 2), f 0 is the nominal frequency
of the system. Due to the slow behavior of the moving power
in comparison to the electromagnetic power generated by syn-
chronous machines, ∆Pj (t) ≈ Pint st
− Pintj (t), where Pintj
j
is the power exiting Area j through the interconnection line
st
and Pint is its value in steady state (before the perturbation).
j
Equation 23 can be solved through Least Squares method
for HAj . In the present work, a limited time-window of 2s
around the moment the perturbation takes place is selected.
How to detect the perturbation and how to select the best
moment to apply the time-window is out of the scope of this
paper.
III. R ESULTS
Consider the test system represented in Figure 2, divided in
two different areas for the present study. The parameters of the
system are specified in [11]. The simulation was handled in
PowerFactory 2018 with the detailed Model 2.2 for generators
[12] and constant impedance model for loads. Governor, PSS
and AVRs were modeled using benchmarking parameters
and the standards provided by DigSilent. The methodology
proposed in this paper was applied in Matlab software.
Fig. 2. 11-bus test-system, [11]
Area 1 is defined by the boundary bus 5, and contain only
one generating machine. Area 2 is defined by the boundary
bus 6, composed by three non-coherent generating units.
Supposing PMUs installed in those boundary buses, the data
acquired from the simulator are highlighted in Figure 3, that is
a snipping of Figure 2. The voltage at bus 5 is represented as
V5 ∠θ5 and the current injection exiting Area 1 is I56 ∠φ56 .
Similar for the other area, V6 ∠θ6 is the voltage at bus 6
and and I65 ∠φ65 is the current injection exiting Area 2. The
sampling time used in the simulations was 20ms.
Case 1 is based on a simulated step increase of 20%
in the load of bus 9 (see Figure 2). In this case, Area 1
oscillates against Area 2. However, G2 oscillates against G3
and G4 locally in Area 2. The idea is to estimate the inertia
of Area 1 (that is equal to the inertia of G1) and Area 2
(that is an equivalent including G2, G3 and G4). For the
data considered, the process converged in 8 iterations, and
the internal reactances estimated were X1 = 0.034p.u. and
X2 = 0.185p.u. With X1 and X2 , the internal voltages
E1 (t)∠δ1 (t) and E2 (t)∠δ2 (t) are calculated.
 LS_IME_cr2
1.0005
f1

1 f COI
1

0.9995

Frequency (p.u.)
0.999

0.9985

0.998

0.9975

Fig. 3. Snipping of the 11-bus test system 0.997

17 18 19 20 21 22 23 24 25 26 27
Time(s)

Fig. 4. Frequency - Area 1

The electrical frequencies related to each equivalent gener-
ator are calculated according to
1 dδj (t) LS_IME_cr2
fj (t) = + fref (t) (24) 1.0005
2πf0 dt f2
1 f
where fj refers to the equivalent machine of Area j = 1, 2, COI
2

f0 is the nominal frequency of the system, δj is the internal 0.9995

angle of the voltage Ej and fref is the frequency of the
Frequency (p.u.)

0.999
reference machine. After calculating fj (t) with equation (24),
a median filter is applied to reduce the peaks caused by the 0.9985
finite difference method.
The electrical frequency of each equivalent machine is 0.998

compared with the mean frequency (fCOI ) of each area, that

0.9975
is a hypothetical concept associated to the Center of Inertia
(COI) of a system, which is a rotational analogy of the center 0.997
of mass of an object [11], [13]. It is an important figure
0.9965
of merit for comparison because following a perturbation, 17 18 19 20 21 22 23 24 25 26 27
the frequency decay of all generators tends to approach the Time(s)
calculated mean rate of the system. The mean frequency of an
area j is determined by the dynamic behavior of each of the n Fig. 5. Frequency - Area 2
machines contained in that Area, as defined in the following
expression
Pn Generator 2 than to the hypothetical COI of the system, such
k=1 Hk SbGk fGk
fCOIj (t) = P n (25) that the influence of the frequency of the generator 2 (fG2 )
i=1 Hk SbGk brings the estimated f2 closer to its behavior. This can be
where Hk is the inertia, SbGk is the nominal power and fGk is seen in Figure 6, where fG2 , fG3 and fG4 are the frequencies
the electrical frequency of each Generator Gk present in Area of each generator of area 2 following the diagram of the test
j. system (Figure 1).
The comparisons between the electrical frequencies and the However, it worths to observe in Figure 6 that the influence
mean frequencies of each area are presented in Figures 4 and of fG3 and fG4 is also present in f2 . In between t = 20s and
5, where f1 and f2 are the frequencies related to the equivalent t = 21s it can be seen that f2 presents a peak, following the
machines of Area 1 and Area 2, and fCOI1 and fCOI2 are the behavior of fG2 , but right after it decreases like fG3 and fG4
mean frequencies of Area 1 and 2, respectively. instead, such that the equivalent inertia related to f2 takes into
It can be seen that the estimated frequency of area 1 is very account the contribution of all the machines of area 2.
close to the mean frequency of that area, what is expected The estimated inertia results for Case 1 can be seen in
once this area has only one generator. The differences are due Table I, where Hest is the estimated inertia and HCOI is the
to inaccuracies of the methods used. About area 2, instead, the calculated inertia related to the mean frequency.
behavior of f2 is slightly different from fCOI2 . This happens It can be seen that the inertia estimated for Area 1 is very
because the measurement point is electrically closer to the close to the inertia of the COI of this area (that is equal to
 1.0005
Considering the mechanical power as known, the method
1 is retested and the estimation obtained for Area 2 is Hest =
19.03, a value 1% greater than the HCOI .
0.9995
A second experiment (named Case 2) is performed includ-
0.999 ing a new load of 150MW and 50MVAR in Area 1, and
Frequency (p.u.)

0.9985
a step increase of 20% in the active power consumption is
simulated in this area, instead of the previous perturbation. For
0.998 this case, the reactances estimated were X1 = 223.0201p.u.
0.9975 and X2 = 0.1249p.u. The increase of load in Area 1 was
f2
f
reflected direclty on the estimated X1 , much higher than in
0.997 G2
f G3 Case 1, while the estimations of X2 were not so different
0.9965 f G4 in between each other. This happens because the presence
of three generating units in Area 2 partially compensates the
0.996
17 18 19 20 21 22 23 24 25 26 27 perturbation locally, such that the load increase has not a big
Time(s) impact seen from the equivalent point of view. In Area 1,
instead, the impact of a local load increase is much bigger
Fig. 6. Frequencies - Area 2
because the area is limited to only one generating unit. The
estimated reactance not only represents the real impedances of
TABLE I the area but also has the effect of imposing to the estimated
I NERTIA ESTIMATIONS FOR C ASE 1 internal voltage the impact of the perturbation viewed from
Area 1 Area 2 the interconnection path.
Hest (s) 6.2504 -0.7719 Table II present the inertia estimations obtained for the
HCOI (s) 6.5000 18.8500 named Case 2.

TABLE II
I NERTIA ESTIMATIONS FOR C ASE 2
the inertia of G1). However, the method failed to estimate
the inertia for Area 2. This happens because the assumption Area 1 Area 2
for the mechanical power assumed in Step 8 of the method Hest (s) -22.2999 18.7216
HCOI (s) 6.5000 18.8500
doesn’t hold for this area, due to the fact that the perturbation
is located inside. The perturbation alters the moving power
such that it oscillates together with the loads of the area, as The first line of the table shows that the method was
shown in Figure 7. The legend P eA2 stands for the active accurate to estimate the equivalent inertia of Area 2, but
power exiting Area 2 and P mA2 represents the moving power struggled again when trying to estimate the inertia of the area
related to the area. that contains the perturbation (for the same reasons presented
in Case 1). When the moving power of Area 1 is considered
as known, the estimations improve to Hest = 7.179, such that
-0.75
Pe
a next step for the research is to estimate the moving power
A2

-0.8 PmA2 of the area that contains the perturbation.

Figure 8 present the inertia estimated in time for both areas,
-0.85 using a sliding window of 100 samples, considering Pm1
known. In the legend, H1 and H2 refer to the estimated inertia
Power (p.u.)

-0.9 of the equivalent machines of area 1 and 2, respectively.

The method didn’t present high sensitivity for the size of the
-0.95 time window, such that the results were similar to every length
tested (1.5s to 3s). It can be seen that the inertias estimated for
-1
both areas presented a peak in the first miliseconds after the
perturbation, and later presented a nearly constant behavior.
-1.05
The peaks are due to the use of finite difference method to
-1.1
calculate the first and second derivative of the equivalent rotor
0 5 10 15 20 25 30 35 40 45 50 angle. The use of filtering reduced those peaks, but they can
time(s)
still be observed in both resulting signals. Alternatives to the
Fig. 7. Moving power and active power - Area 2 finite difference method together with filtering will be studied
in the future to improve the results.
 [6] K. Kalsi et al. Calibrating multi-machine power system parameters
50 with the extended kalman filter. IEEE Power and Energy Soc. General
H1 Meeting (PES), 2011.
45
H2 [7] N. Petra, C. G. Petra, Z. Zhang, E. M. Constantinescu, and M. Anistescu.
40 A bayesian approach for parameter estimation with uncertainty for
dynamic power systems. IEEE Transactions on Power Systems, 2016.
35 [8] X. Cao, B. Stephen, I. F. Abdulhadi, C. D. Booth, and G. Burt. Switching
markov gaussian models for dynamic power system inertia estimation.
30 IEEE Transactions on Power Systems, 2016.
Inertia(s)

[9] J. H. Chow. Power System Coherency and Model Reduction. New York:
25
Springer, 2013.
20 [10] Y. Wehbe, L. Fan and Z. Miao. Least Squares Based Estimation of Syn-
chronous Generator States and Parameters with Phasor Measurement
15 Units. IEEE, 2012.
[11] P. Kundur. Power System Stability and Control. McGraw-Hill.,
10
[12] IEEE. Guide for Synchronous Generator Modelling Practices and Ap-
5 plications in Power System Stability Analyses. Tech. Rep. 1110-2002,
IEEE, 2002.
0 [13] H. Golpı̂ra and A. Messina. A Center-of-Gravity-based Approach to
19.5 20 20.5 21 Estimate Slow Power and Frequency Variations. IEEE Transactions on
Time (s) Power Systems, 2018.
[14] F. Milano and A. Ortega. Frequency Divider. IEEE Transactions on
Fig. 8. Inertias estimated - Case 2 Power Systems, 2017.

IV. C ONCLUSIONS
A PMU-based method for building dynamic equivalents and
estimating the equivalent inertia following a perturbation was
presented. The method is based on the approach proposed in
[9] for study of oscillations, and adapted in a novel iterative
approach to consider losses without increasing the number of
measurement points needed.
The method was tested with data produced by a 11-bus test
system simulated in PowerFactory, considering detailed model
2.2 (IEEE standard) for the generators and considering the
presence of frequency control. Results for two test cases were
presented, variating the location of the perturbation considered.
The results were accurate and practical when estimating the
inertia of the areas that doesn’t include the perturbation inside.
For the area that contains a perturbation, the method works but
requires the knowledge of the equivalent moving power of the
area in time, which estimation is a possible topic of interest
for a future work.
V. ACKNOWLEDGMENTS
The authors gratefully acknowledge the financial and tech-
nical support from the Italian TSO, Terna.
R EFERENCES
[1] J. H. Chow, A. Chakrabortty, L. Vanfretti, and M. Arcak. Estimation of
power system inertia constant and capacity of spinning-reserve support
generators using measured frequency transients. IEEE Transactions on
Power Systems, 2008.
[2] L. Fan, Z. Miao and Wehbe Y. Application of Dynamic State and Pa-
rameter Estimation Techniques on Real-World Data. IEEE Transactions
on Smart Grid, 2013.
[3] P. Wall and V. Terzija. Simultaneous Estimation of the Time of Distur-
bance and Inertia in Power Systems. IEEE, 2014.
[4] P. M. Ashton, G. A. Taylor, A. M. Carter, M. E. Bradley, and W. Hung.
Application of phasor measurement units to estimate power system
inertial frequency response. IEEE Power and Energy Soc. General
Meeting, 2013.
[5] H. Zhenyu, D. Pengwei, D. N. Kosterev, and Y. Bo. Application
of extended kalman filter techniques for dynamic model parameter
calibration. IEEE Power and Energy Soc. General Meeting (PES), 2009.