This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

Online Estimation of Power System Inertia

Constant Under Normal Operating Conditions
Fanhong Zeng1, Student Member, IEEE, Junbo Zhang1, Senior Member, IEEE, Ge Chen1,
Student Member, IEEE, Zikun Wu1, Siwei Huang1 and Yingqi Liang2, Member, IEEE
1
School of Electric Power, South China University of Technology, GuangZhou 510641, China
2
Department of Electrical and Computer Engineering, National University of Singapore, Singapore
Corresponding author: Junbo Zhang (e-mail: [email protected]).

This work was supported in part by the National Natural Science Foundation of China under Grant 51761145106, and in part by the
Natural Science Foundation of Guangdong Province, China, under Grant 2018B030306041.

ABSTRACT An online estimation method for the power system inertia constant under normal operating
conditions is proposed. First of all, a dynamic model relating the active power to the bus frequency at the
generation node is identified in the frequency domain using ambient data measured with the phasor
measurement units (PMUs). Then, the inertia constant at the generation node is extracted from the unit step
response of the identified model in the time domain using the swing equation. Finally, with the sliding
window method and the exponential smoothing method, the estimated inertia constant is updated in real-
time. Compared to the conventional methods using large disturbance data or field test data, the proposed
method can estimate the inertia constant under normal operating conditions, and therefore, can provide the
tracking trajectory of the power system inertia constant in real-time. The effectiveness of the proposed
method is validated in the IEEE 39-bus system. The results show that the relative error of the identified
inertia constant is below 5% and the identified inertia constant can be updated within 1s.

INDEX TERMS Inertia constant estimation, ambient signals, subspace identification, step response, sliding
window, exponential smoothing

I. INTRODUCTION controls and many of the protection relays [1]. When the
In the point view of physics, the inertia of a power system is penetration of RESs becomes higher, the situation is different.
its capability to resist energy fluctuations caused by external For instance, the system frequency may drop badly when the
disturbances, which, in conventional power systems, are system suffers an active power deficiency, and consequently,
supplied mainly by the kinetic energy stored in the rotating the protection and control devices such as under-frequency
mass of the synchronous generators and quantified by the load shedding or disconnection of generators may be
inertia constant [1]. triggered reluctantly [9]. Since there is no mature method to
With the increasing share of the power electronic inverter track the power system inertia in real-time, to avoid the
interfaced renewable energy sources (RESs), some changes malfunction of the protection relays and stability controllers,
have taken place in power system inertia. Many synchronous the transmission system operators (TSOs) must adopt more
generators are displaced by RESs, thus leading to a persistent conservative operational schemes to ensure the stable
decrease in the conventionally available inertia resources [2]- operation of the power system, which inflates the power
[4]. Meanwhile, with different control strategies and system operational costs.
parameters, various inertia suppliers such as virtual inertia Given the issues mentioned above, the need for online
control and energy storage systems are employed to improve estimation of power system inertia (constant) is highlighted
the power system inertia [5]-[7]. Therefore, the power system and has drawn consistent attention in recent years. According
inertia, as well as its quantification, not only decrease but to the measurement data types, the existing estimation
also become time-varying [8]. methods can be categorized into two groups: 1) disturbance
In conventional power systems, the inertia constant is data methods, and 2) non-disturbance data methods.
steady over the long term, and therefore, is chosen as a For most of the estimation methods based on disturbance
fundamental reference to the design of frequency stability measurements, the inertia (constant) is estimated using the

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

swing equation, utilizing post-event data that record the In relation to the existing work on inertia estimation, the
transient active power and frequency characteristics of the contributions of this study are clarified in Fig. 1 and
generators [10]-[21]. For instance, a procedure for estimating summarized as follows: 1) it provides a more stable and
the total inertia of the Great Britain power system was precise solution to identify power system inertia constant
proposed in [13], which calculated the total inertia for the using ambient signals; 2) it realizes the estimation of inertia
whole system by summing all the estimated regional inertias. constant at different hierarchies (individual generator, area
An online algorithm to estimate the system inertia after a and the whole system) under normal operating conditions; 3)
disturbance based on sliding windows of active power and it realizes the real-time online tracking of inertia constant in
frequency derivative measurements was proposed in [14]. In the time scales of seconds under normal operating conditions,
[17], an inertia estimation method based on the extended which can timely provide important information for stable
Kalman filter was proposed, which needs to assume the time operation of the power system.
of disturbance. Based on the transient and steady state
characteristics of the frequency response after a disturbance, Inertia Estimation
a method to estimate the equivalent inertia and damping
constant simultaneously was proposed in [18]. In [19], an
approach for online inertia estimation in the power system Disturbance Data Non-disturbance Data

network with solar photovoltaic source was proposed using

the synchronized measurements from PMUs. In [20], based Swing Equation
Electromechanical
Probing Signal Ambient Signal
Wave Theory
on the measured frequency response for any arbitrary
disturbance in the system, a method to estimate the available
inertia in an islanded microgrid was proposed. In addition, Online Inertia Online Inertia Estimation Online Inertia
Estimation at Different Hierarchies Tracking
some other disturbance data methods have recently been
developed. In [21], an approach based on electromechanical Existing Work
More Stable and
wave theory was presented to identify the change of power Precise Method Contribution of the Study
system inertia distribution. Though both the offline and
online methods based on recorded disturbances can estimate FIGURE 1. Contributions of the study to inertia estimation

the inertia accurately, they cannot achieve continuous inertia

estimation as the natural transient events are deficient, and The remainder of the paper is organized as follows: section
the transient field experiments are expensive. II provides the theoretical fundamentals about power system
Compared to the disturbance data methods, estimators inertia; section III introduces the proposed inertia constant
based on non-disturbance measurements are quite limited. In online estimation method; section IV presents simulation
[22], a statistical model-based real-time inertia estimation results in the IEEE 39-bus system to verify the feasibility of
method was proposed, in which the model was trained to the proposed method; section V concludes the paper.
learn the features that relate the steady-state average
II. THEORETICAL FUNDAMENTALS
frequency variations and the system inertia. In [23], ambient
frequency and active power data were employed to estimate
A. INERTIAL RESPONSE AND INERTIA CONSTANT
the effective inertia of a power system, where a combined
After a disturbance, the frequency response of a traditional
model of inertial response and primary control was identified,
power system will sequentially go through three stages: the
and then the inertia was extracted from the impulse response
inertial response, the primary response and the secondary
of the model. However, the inertia tracking trajectory could
response [13]. During the inertial response, the rotating
not be provided in real-time in [23]. In our previous work, a
kinetic energy stored in the synchronous generators is
closed-loop identification method was proposed for the
released spontaneously to maintain the power balance, thus
power system equivalent inertia constant online estimation,
reducing the rate of change of frequency (RoCoF).
which could achieve a precise estimation result, but it
The rotational kinetic energy of a synchronous generator is
required the injection of an additional probing signal [24].
usually normalized to an inertia constant, which is defined as
In this paper, we propose a more robust online estimation
the ratio of the stored rotational kinetic energy at the rated
method for the power system inertia constant under normal
rotating speed to the rated capacity of the synchronous
operating conditions. Compared to the conventional methods
generator [1], namely:
based on transient test, probing and etc., the proposed method
Ei
can achieve a precise inertia constant estimation without any Hi = (1)
disturbance event and probing injection, and can provide the Si
tracking trajectory of the inertia constant in real-time. To our where Hi is the inertia constant of generator i, Ei is the
knowledge, it is the first work to track the inertia constant rotational kinetic energy of generator i and Si is the rated
under normal operating conditions. capacity of generator i. Physically, Hi represents the time
duration to supply energy for the demand that equals to the

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

rated capacity of the generator, without any additional respectively; fj denotes the aggregated frequency of the area
mechanical input. or the system (in p.u.); Hj denotes the equivalent inertia
In a power-electronics-dominated power system, the constant of the area or the system; and Dj denotes the total
converter-interfaced RESs are initially inertia-free and the damping coefficient of the area or the system. The
system frequency response does not have clear three-stages. theoretical value of Hj can be calculated using (2).
However, virtual inertial response can be obtained with Under normal operating conditions, all the variables in (3)
control of electrical converters, then converter-interfaced vary around the steady-state operating point. Therefore,
RESs can provide equivalent inertia for the power system. formula (3) can be written as the incremental formulation
Besides, the dynamics of inertial response provided by around the steady-state operating point as follows:
converter-interfaced RESs can be described by a first-order 1
differential equation [25]-[26], which is similar to the swing f r ,i = (Pm,i − Pe,i − Di f r ,i ) (5)
2H i
equation of the synchronous generator. Namely, the inertial
response of converter-interfaced RESs is mathematically Assuming ΔPm,i to be zero and taking the Laplace
equivalent to that of the synchronous generators [27]. transform on both sides of (5), we can reformulate the
Therefore, when estimating the equivalent inertia constant swing equation as a first-order transfer function:
of the converter-interfaced RESs, it is reasonable to regard f r ,i(s) 1
Gi(s) =  − (6)
them as the equivalent synchronous generators and the Pe,i(s) 2H i s + Di
proposed estimation method can be employed.
Further, in a multimachine power system, if we consider where Δfr,i represents the rotor electrical frequency
the other inertia contributors as the equivalent synchronous deviation of generator i, ΔPe,i represents the electrical power
generators, then the equivalent inertia constant of the entire deviation of generator i, and Gi(s) (s is the Laplace operator)
power system can be calculated as follows: is the transfer function from ΔPe,i to Δfr,i, which is
characterized by Hi and Di. Besides, the transfer function
1 N
H sys =
Ssys
 H i Si (2) from ΔPe,j to Δfj is similar to that in (6), which is
i =1 characterized by Hj and Dj.
where Hsys is the equivalent inertia constant of the system,
Ssys is the rated capacity of the system and N is the number III. METHODOLOGY
of generators. The methodology proposed in this paper is for power
system inertia constant real-time online estimation under
B. SWING EQUATION normal operating conditions, which is shown in Fig.2. The
The dynamics between active power and frequency of a key procedures include signal selection and preprocessing,
synchronous generator in a short time frame after a power system identification, inertia constant extraction and inertia
mismatch can be modeled by the swing equation. For constant tracking. All the key procedures can be realized
synchronous generator i, considering the damping effects, automatically with low computational burden, and their
the swing equation can be written as [1] details are expanded in the following subsections.
1
f r ,i = ( P − Pe,i − Di f r ,i ) (3) A. SIGNAL SELECTION AND PREPROCESSING
2H i m,i According to section Ⅱ-B, the dynamic model of generator i
where Pm,i and Pe,i are the mechanical power (in p.u.) and can be identified using electrical power Pe,i as the input and
the electrical power (in p.u.) of generator i, respectively; fr,i rotor electrical frequency fr,i as the output. However, Pe,i and
is the rotor electrical frequency (in p.u.) of generator i; and fr,i in the real power system are difficult to measure. As a
Di is the damping coefficient of generator i. Physically, substitute, Pe,i and fr,i can be approximated by the active
when suffering a power mismatch, the RoCoF of the power output Pi and the frequency fi at the generator
synchronous generator is constrained by the inertia constant connection bus, respectively. In practice, Pi and fi can be
and the damping coefficient in a short time frame, thus the measured by PMUs installed at the generator connection
frequency of the generator cannot change suddenly and the bus. Therefore, we can identify the dynamic model of
frequency stability can be improved. generator i using active power output Pi as the input and
As an approximation, the dynamic behavior of a certain bus frequency fi as the output. Namely, the following
area or the system can be represented as an equivalent equation holds:
synchronous generator j, which leads to the aggregated fi(s) 1
swing equation as follows: Gi(s)   − (7)
Pi(s) 2H i s + Di
1
fj = ( P − Pe,j − D j f j ) (4) Additionally, it is possible to identify the dynamic model
2 H j m, j of a certain area (or the system) using the total active power
where Pm,j and Pe,j denotes the total mechanical power and output and the aggregated frequency of the area (or the
the total electrical power of the area or the system (in p.u.), system) as input and output, respectively. It would be

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

Bus H
P
G Power system

Individual/
Equivalent
t
P f PMU Time-varying inertia

Disturb the
Convert into per model with a
unit values unit step signal Sliding
N4SID window
Remove mean Calculate the method with
values initial slope exponential
Model cross-
validation smoothing
Prefilter Extract the inertia
Downsample constant using
swing equation
Signal System Inertia Constant Inertia Constant
Selection and Identification
Preprocessing Extraction Tracking

FIGURE 2. Flow chart of the proposed methodology

simple to obtain the total active power output by summing frequency components that can impair the inertia constant
up all active power outputs in the area (or the system). To estimation. Finally, the signals are downsampled to the
aggregate the frequency, however, the center of the inertia range of 5-10 Hz to avoid the numerical problems when
frequency is commonly used, which is an average running the identification algorithm.
frequency weighted by the inertia of each node and cannot
be measured directly [23]. B. SYSTEM IDENTIFICATION
Here we propose a simplified aggregated frequency to Generally, the power system is nonlinear. However, under
represent the center of inertia frequency. The frequency of normal operating conditions, the disturbance to the power
the area (or the system) is evaluated by a weighted average system is small, so the nonlinear power system can be
of the measured frequencies as follows: approximated by a linear state space model around the
steady-state operating point. An nth other multi-input-multi-
 i =1 wi fi
i=Nj

fj = (8) output state space model can be described as follows:

 i =1 wi
i=N j
xk +1 = Axk + Buk + ωk
(9)
where fi denotes the frequency at the generator connection yk = Cxk + Duk + vk
bus, and Nj denotes the number of online generators in the
where xk∈Rn is the state vector; uk ∈Rm is the input vector;
area (or the system). Here, the weights wi are determined
yk ∈Rl is the output vector; and A∈Rn×n, B∈Rn×m, C∈Rl×n,
based on the following consideration: the larger the inertia
and D ∈ Rl×m are system matrixes to be identified.
constant of generator i, the smaller the variations of the bus
Additionally, ωk and νk denote random sequences of process
frequency fi under normal operating conditions. Therefore,
noises and measurement noises, respectively.
the weights wi are defined as the inverse of the variance of
Subspace identification methods are effective algorithms
bus frequency fi, namely, wi = 1/var(fi).
to identify the state space model with ambient data and can
Under normal operating conditions, the ambient data
be implemented in different ways [30]. As a widely used
measured by PMUs are commonly polluted with noise, so
algorithm in subspace identification [28], the N4SID
signal preprocessing is a necessary procedure to improve
(Numerical algorithm for Subspace State Space System
identification efficiency and accuracy before running an
IDentification) algorithm is used in this paper. The main
identification algorithm [28], [29]. First, all the signals are
feature of the N4SID algorithm is to calculate matrix Γk
converted into per unit values by dividing their base values.
through oblique projection as follows:
Then, all the signals are detrended by removing their mean
values and prefiltered using a noncausal Butterworth low-  U p  
pass filter. As the typical values for Hi are in the range of 2- Γk = Y f |    / U f (10)
  Y p  
10 s [1], a noncausal Butterworth low-pass filter with a 0.5 
Hz cut-off frequency is suitable to attenuate the higher

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

where / denotes the oblique projection. Then, singular value According to section Ⅲ-B, the real power system can be
decomposition (SVD) is applied on Γk to determine the approximated by a linear model and the model is reliable if
order of the identified model. Specifically, the order of the the model cross-validation performs well. In other words,
identified model is equal to the number of the dominant the identified model can be regarded as the real power
singular values of matrix Γk. The SVD can be partitioned system model to some extent. In this perspective, when a
into the following form: certain disturbance is applied to the identified model, the
S 0 V1T  disturbance source can be regarded as the active power
W1 Γ kW2 = U1 U 2   1   T   U1 S1V1 (11)
T
deviation ΔPi and the corresponding response can be
 0 S2  V2  regarded as the bus frequency deviation Δfi.
where W1 and W2 are the identity weighing matrixes. In If we disturb Gi(s) with an unit step signal, namely, ΔPi =
(11), the insignificant singular values are neglected by -ε(t) (ε(t) is the unit step function), then Δfi can be
removing S2 as the dominant singular values determine the expressed as (13) in Laplace domain.
main dynamics of the system. Finally, the system matrixes 1 1
fi(s) = (13)
A, B, C and D can be obtained by solving the linear 2H i s + Di s
equations. Readers can refer to [28] for the details of the
The equation can be solved directly and be written as
N4SID algorithm.
follows in the time domain:
Generally, the order of the real system is rather high as it Di
−
contains many complicated control systems. However, a 1 t
(1 − e 2 H i )
fi(t ) = (14)
model with a lower order is identified in this part. Though Di
the order of the identified model is lower than the order of
Then the slope of the unit step response at t = 0 can be
the real power system, it is accurate enough to capture the
calculated as:
dynamics of the inertial response. The N4SID algorithm
1
can search the best order for the identified model fi t = 0 = (15)
automatically after setting a range of orders from nmin to 2H i
nmax. Empirically, nmin can be set as 1 and nmax can be set as According to (15), the inertia constant Hi is determined
10 for inertia constant estimation. by the initial RoCoF fi t = 0 , namely the initial slope of
After running the identification algorithm, the reliability
of the identified model should be verified by model cross- the unit step response of Gi(s). Therefore, we can disturb
validation. The model cross-validation can be performed by the identified model with a unit step signal and calculate the
comparing the validated output yˆ(t ) and the original output initial slope of the corresponding response as fi t = 0 .The
y(t). To evaluate the reliability of the identified model
problem turns into how to estimate fi t =0
after the
quantitatively, the fitting ratio (FR) between the validated
output yˆ(t ) and original output y(t) is defined as follows: disturbance as accurate as possible. As the inertial response
 
N
( yˆ(t ) − y(t ))2 
activates immediately after the disturbance and lasts a short
FR =  1 − t =1   100% (12) time, we recommend calculating fi t = 0 with a 500 ms
 
N
y(t )2 
 t =1  sample-by-sample sliding window1 over 1-2 s period [13]
where N is the number of samples. from the unit step response of the identified model. For
power systems of various sizes, the data length of the unit
C. INERTIA CONSTANT EXTRACTION step response for the fi t = 0 calculation can be adjusted
Theoretically, the state space model includes the inertia
constant but as an implicit value, so further analysis should properly according to the dynamics of the systems. During
be employed for the identified model. A direct way is to each sliding window, fi t = 0 is estimated as the slope of a
extract the inertia constant from the model itself in the linear fit to the response. The maximum slope is then taken
frequency domain after some transformation. First, the state
to represent the fi t = 0 during the inertial response period
space model can be transformed into a transfer function
from ΔPi to Δfi. Then, the transfer function is reduced to a following the disturbance.
first order inertia function the same as (7) and finally, the Note that, the discrete time identified model should be
inertia constant Hi is estimated together with Di. However, converted into continuous time model before evaluating
large error may be introduced during the order reduction step response. Then, the continuous time model should be
process, leading to inaccurate estimation of the inertia checked for stability in s-domain, namely, the real part of
constant. For this reason, we propose a method to extract poles should be less than zero. The functions d2c, step,
the inertia constant from the step response of the identified polyfit in MATLAB can be used in this part for inertia
model in the time domain. constant extraction.
The proposed RoCoF calculation method is also
applicable for the event data of real power systems. In other

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

words, the inertia constant can be estimated from the event IV. CASE STUDY
data based on the proposed RoCoF calculation method and
the swing equation. In this case, the event data can be A. SYSTEM BACKGROUND
treated as a complementary of ambient data. The proposed method was tested in the IEEE 39-bus system,
which is a simplified model of high voltage transmission
D. INERTIA CONSTANT TRACKING system in the northeast of the USA (New England area).
In this part, sliding window2 method and exponential The system consists of 10 generators, 39 buses, 19 loads, 34
smoothing method are used to update the estimated results lines and 12 transformers. The rated frequency is 60 Hz and
in real-time, thus realizing the online tracking of the inertia the main voltage level is 345 kV. Generator 1 is the
constant. For the sliding window method, a fixed size time equivalent generator for the external power grid, and
window is used to estimate the inertia constant, and the generator 2 is the balance generator. Automatic voltage
window is gradually updated for the next estimation. The regulators and governors are used for generator 2-10. The
performance of the sliding window method is dependent on time domain simulations are carried out using Digsilent/
both the choice of sliding window length and the estimation Powerfactory software. The single line diagram of the test
refresh rate. system is shown in Fig. 3.
Generally, the estimation accuracy increases with sliding
10 8
window length. However, no estimates are available during 30 37
the first sliding window, and the time delay of the first 26 28 29
25
estimation result can be reduced with a shorter sliding 2 Area 4
27
window length. The estimation refresh rate determines how 1
38
3 18 17 9
fast a new estimation should be done. Though a faster
39 21
refresh rate makes the estimator more responsive for real- 1
15 16
time estimation as well as tracks change of inertia constant
Area 1 Area 2 Area 3
better, it increases computational burden, thus leading to a 7
longer execution time. 4 14 24 36
Though the inertia constant extracted from the identified 23
5 13
models can be updated in real-time, there may be a few low 9 6 19
FR models that introduce inaccurate estimates, leading to 12 22
7
large fluctuations of the inertia constant. To smooth the 11 20
estimates from the low FR models, the exponential 10 33 35
smoothing method was used, whose original form is 8 31 32 34
2 3 5 4 6
defined as:
 t = t + (1 −  ) t −1 (16) FIGURE 3. Single line diagram of the IEEE 39-bus system[31]
where νt and νt-1 are the smoothing value at time t and time
t-1, respectively; θt is the actual value at time t; and α is the B. ONLINE INERTIA CONSTANT ESTIMATION
smoothing constant, ranging from 0 to 1. In this part, the inertia constant estimation accuracy of the
In addition, the inertia constant extracted from the proposed method is validated in the widely used IEEE 39-
unstable identified model is far from the real value, which bus system mentioned above. All 19 loads in the 39-bus
should be detected and removed. In this paper, we replace system are injected with Gaussian noise filtered by a low-
the abnormal estimates from the unstable identified model pass filter with a cut-off frequency of 5 Hz, thus simulating
at time t with the estimates at time t-1. the small random load fluctuations and other related small
Combining the methods mentioned above, the inertia variations of real power systems during normal operating
constant can be updated as follows: conditions. Then, the inertia constant of all the individual
generators as well as the equivalent inertia constant of both

t ht + (1 − t )H t −1 ,if stable
k k
Ht =  (17) the areas and the whole system is estimated using the

 H t −1 ,if unstable methods introduced in section Ⅲ-A to C.
1) Online Inertia Constant Estimation of The Individual
where Ht and Ht-1 are the smoothing value of inertia
Generator: The experiment of a single estimation can be
constant at time t and time t-1, respectively; ht is the inertia
carried out with the following procedures:
constant extracted from the identified model at time t; μt is
Step 1: The bus frequency fi at the generator connection
the fitting coefficient at time t, which is equal to FR divided
bus and the active power output Pi are measured by PMUs.
by 100; and k is the exponent, which can be set from 20 to
The length of the measurements is 200 s and the sampling
80 to get a good smoothing effect.
1,2
The sliding windows in these two parts are different. The former is rate is 100 Hz. The first 100 s signals are used for model
for RoCoF calculation, and the latter is for inertia constant tracking. identification, and the last 100 s signals are used for model
cross-validation.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

Step 2: Data preprocessing is employed to the signals. Step 4: The inertia constant of the generator is extracted
The bus frequency fi and the active power output Pi are from the identified model using the method proposed in
converted into per unit values by dividing their base values section Ⅲ-C. The identified model is disturbed with a unit
60 Hz and 100 MW, respectively; the trend is removed; the step signal; the unit step response of the identified model is
high-frequency components are filtered using a noncausal collected; the initial RoCoF is calculated from the unit step
low-pass filter with a cut-off frequency of 0.5 Hz; and the response of the identified model; and the inertia constant is
sampling rate is decreased from 100 Hz to 5 Hz to avoid extracted using (15).
numerical unstable problems. Fig. 6 shows the unit step response of the identified
Fig. 4 shows the preprocessed active power and bus model of generator 2 as well as the calculation of initial
frequency of generator 2. The fluctuations of both the RoCoF during inertial response, where the first 0.5 s of the
active power and the bus frequency under normal operating result is enlarged. Table Ⅰ shows the detailed results of the
conditions is rather small. The preprocessed measurements inertia constant estimation with a rated apparent power (Sb)
of other generators are not provided here since they are of 100 MW. In Table Ⅰ, the estimated inertia constant (Hest)
similar to that in Fig. 4. is very close to the real inertia constant (Href), and the
Step 3: With the processed active power Pi and bus largest relative error for the estimated inertia constant (Hest)
frequency fi, the state space model with Pi as input and fi as is below 5%, thus the feasibility of the proposed method to
output is identified using N4SID. Then the model cross- estimate inertia constant of individual generator is verified.
validation is carried out to evaluate the reliability of the
identified model.
Fig. 5 shows the model cross-validation result of
generator 2. The model cross-validation results of other
generators are not provided here since they are similar to
that in Fig. 5. It indicates a reliable model is identified as
the validated output fits the original output well.

FIGURE 6. Calculation of the initial RoCoF during inertial response

TABLE I
ESTIMATED INERTIA CONSTANT OF ALL GENERATORS
Generator Bus Order Href Hest Error (%)
G01 39 5 500 506.06 1.21
G02 31 7 30.3 30.96 2.18
G03 32 7 35.8 36.42 1.73
G04 33 6 28.6 29.34 2.59
G05 34 7 26.0 25.40 -2.31
FIGURE 4. Preprocessed measurements of generator 2 G06 35 5 34.8 34.99 0.55
G07 36 7 26.4 27.49 4.13
G08 37 5 24.3 23.61 -2.84
G09 38 7 34.5 35.93 4.14
G10 30 5 42 43.56 3.71

Generally, the results from a single estimation are

uncertain. To evaluate the accuracy of the proposed method
and to further reduce the relative error by averaging the
results from multiple estimations, the sliding window
method is used to obtain various estimated results from one
measurement period. In our simulations, a 10-minute
measurement with a window length of 200 s and a refresh
rate of 1 s is used; in total, 401 inertia constant estimations
are carried out. For each estimation, steps 1-4 are employed
sequentially; then, all the estimated results are obtained.
FIGURE 5. Model cross-validation result of generator 2

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

TABLE Ⅱ
STATISTICAL ANALYSIS RESULTS OF THE ESTIMATED INERTIA CONSTANT
Generator Href Havg ρH Havg CI-95% ρH CI-95% Error (%)
G01 500 510.01 29.61 498.73-521.28 20.50-40.06 2.00
G02 30.3 30.87 2.86 29.86-31.89 2.30-3.78 1.88
G03 35.8 36.24 3.25 35.67-36.81 2.89-3.71 1.23
G04 28.6 29.10 2.29 28.56-29.65 1.96-2.74 1.75
G05 26.0 26.54 1.52 25.97-27.10 1.21-2.04 2.08
G06 34.8 35.47 2.89 33.87-37.07 2.12-4.56 1.93
G07 26.4 26.90 1.99 26.00-27.81 1.52-2.87 1.89
G08 24.3 24.40 1.56 24.16-24.63 1.41-1.74 0.41
G09 34.5 35.02 2.13 33.77-35.08 1.75-2.70 1.51
G10 42 42.05 3.34 41.27-42.82 2.87-3.98 0.12

After all of the estimations are completed, the results from is estimated by summing the active power outputs and
the identified model with an FR above 95% are selected for aggregating the bus frequencies.
statistical analysis, which means the point estimation and To evaluate the accuracy of the developed method and to
the 95% confidence interval (CI) estimation are carried out. further reduce the relative error, the sliding window method
Fig. 7 shows the normal distribution behavior of the is used to obtain various estimates from one measurement
estimated inertia constant (Hest) of generator 2. It can see period. In our simulations, a 10-minute measurement with a
that 106 inertia constant estimates are selected for statistical window length of 200 s and a refresh rate of 1 s is used; in
analysis. The distribution of the estimated inertia constant total, 401 inertia constant estimations are carried out. In
from other generators is not provided here due to the each estimation, steps 1 to 4 mentioned above are employed
limited space. Table Ⅱ shows the statistical analysis results sequentially. After all of the estimations are completed, the
of the estimated inertia constant. In Table Ⅱ, we can see results from the identified model with an FR above 95% are
that the average of the estimated inertia constant (Havg) is selected for statistical analysis.
nearly the same as the real inertia constant (Href), and all the Fig. 8 shows the normal distribution behavior of the
real inertia constant (Href) values lie in the 95% CI. The estimated inertia constant (Hest) of area 2. It can see that
variance of the estimated inertia constant (ρH) is rather 258 inertia constant estimates are selected for statistical
small, meaning that the estimated values (Hest) fluctuate analysis. The distribution of the estimates from other areas
around the real values (Href) in a small range. Further, the is not shown here because of space considerations. Table Ⅲ
relative error of all the estimated inertia constants is shows the statistical analysis results for the estimations. In
approximately 2%, which is less than the results of a single Table Ⅲ, we can find out that the average of the estimated
estimation, indicating sufficiently good performance of the inertia constant (Havg) is nearly the same as the real inertia
proposed method. constant (Href), and all of the real inertia constant (Href)
values are in 95% CI. The variance of the estimated inertia
constant (ρH) is small, indicating the estimated values (Hest)
narrowly fluctuate around the real values (Href). Further, the
relative errors of all the estimated values are below 5%,
thus verifying the effectiveness of the proposed method for
estimating the equivalent inertia constant of the area or the
whole system.

FIGURE 7. Distribution of the estimated inertia constant of generator 2

2) Online Equivalent Inertia Constant Estimation of the
Area and the Whole System: Considering the coherence of
the generators after a large disturbance, the 39-bus system
can be divided into 4 areas, as shown in Fig. 2 [32]. Then,
the equivalent inertia constant of the area and the whole
system is estimated following steps 1 to 4. Note that the
equivalent inertia constant of the area and the whole system FIGURE 8. Distribution of the estimated inertia constant of area 2

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

TABLE Ⅲ
STATISTICAL ANALYSIS RESULTS OF THE ESTIMATED INERTIA CONSTANT
Area/System Href Havg ρH Havg CI-95% ρH CI-95% Error (%)
Area 1 500 510.01 29.61 498.73-521.28 20.50-40.06 2.00
Area 2 66.1 64.81 5.96 63.57-66.31 5.49-6.53 -1.95
Area 3 115.8 118.24 6.87 115.77-120.80 5.49-9.18 2.15
Area 4 100.8 104.43 2.29 103.25-105.63 1.70-3.48 3.60
System 782.7 757.94 32.96 724.57-791.30 30.82-35.42 -2.94

3) Comparison with the Disturbance Data Method: The aggregated by all the generators of a certain area. The
disturbance data method in [13] is carried out to compare inertia constant of the synchronous generator of the 39-bus
with the method proposed in this paper. system is changed to simulate the time-varying equivalent
Step 1: The active power of Load 15 is increased from inertia constant caused by the switching of the generators in
320 MW to 1820MW suddenly, resulting in an active the area. Considering a measurement length of 120 min, the
power deficiency in the system. In this case, the active inertia constant of all generators remains at H1 during the
power imbalance ΔP is equal to 15 (in p.u.). first 40 min, then changes to H2 and remains there from 40
Step 2: The transient frequency following this event is to 80 min. The inertia constant stays at H3 during the last 40
measured. A five-order low pass Butterworth filter with a min. H1, H2 and H3 can be regarded as the inertia constant
cutoff frequency of 0.5 Hz is used to isolate the dominant vectors here that represent the inertia constant of all
system inertial response. The processed transient frequency generators in different periods. Table Ⅳ shows the inertia
is shown in Fig. 9. constant of all generators in different periods.
Step 3: The RoCoF is calculated using a 500-ms sample-
by-sample sliding window, over a 2 s period following the TABLE Ⅳ
event. The maximum value is then taken to represent the INERTIA CONSTANT OF ALL GENERATORS IN DIFFERENT PERIODS
RoCoF following the event before the primary frequency Generator Bus
H1 H2 H3
response starts to take effect. (0-40 min) (40-80 min) (80-120 min)
G01 39 500 600 400
Step 4: With the method in [13], the equivalent inertia G02 31 30.3 28 35
constant of the whole system is estimated as 809.98 s. The G03 32 35.8 32 40
estimate value is close to the real value of 782.7 s and the G04 33 28.6 24 36
relative error is 3.49%. G05 34 26.0 30 33
Besides, the relative error of the estimated inertia G06 35 34.8 40 44
G07 36 26.4 21 31.5
constant from the disturbance data method and that from
G08 37 24.3 31.5 21
the method proposed in this paper are both below 5%, thus G09 38 34.5 38 42
verifying the effectiveness of the proposed method further. G10 30 42 36 48

All the loads in IEEE 39-bus system are modeled to be

random, the same as section Ⅳ-B. Then, the inertia
constant of all generators in the 39-bus system is tracked
using the methods introduced in section Ⅲ-A to D. The
active power output Pi and bus frequency fi with a 120-min
length are measured by PMUs. The sliding window method
with a window length of 100 s and a refresh rate of 1 s is
employed. For each sliding window, steps 2-4 in section
Ⅳ-B-1) are employed sequentially and then, the estimates
are smoothed using the exponential smoothing method in
(17). Once the measurements from the sliding window are
obtained, the inertia constant will be estimated immediately.
The execution time of a single estimation is less than 1 s,
which means that the inertia constant can be updated before
FIGURE 9. Preprocessed transient frequency measurement new samples are completely collected, leading to real-time
online tracking of the inertia constant.
C. ONLINE INERTIA CONSTANT TRACKING Inertia constant tracking trajectories of generator 7 and
In this part, the inertia constant online tracking capability of generator 8 are shown in Fig. 10 and Fig. 11, respectively.
the proposed method is validated in the IEEE 39-bus The inertia constant tracking trajectories of other generators
system. Here, each synchronous generator of the 39-bus are not provided here for the limited space. Both Fig. 10
system is regarded as an equivalent generator, which is and Fig. 11 show that the inertia constant tracking

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

trajectories fluctuate around the reference trajectories in a For various power systems, the sliding window length
small range, meaning that the inertia constant can be can be adjusted properly until an acceptable dynamic model
tracked accurately in real-time using the proposed method. can be identified and the faster refresh rated can be
determined to track the inertia constant better using a more
powerful computer. All the analyses above confirm that the
proposed method is suitable for power system inertia
constant real-time online tracking.

V. CONCLUSION
This paper proposed an online estimation method for
power system inertia constant under normal operating
conditions. First, the dynamic model between active power
output and bus frequency measured by PMUs was
identified using the subspace identification method. Then,
FIGURE 10. Inertia constant tracking trajectory of generator 7 the inertia constant was extracted from the unit step
response of the identified model in the time domain. Finally,
the sliding window method and the exponential smoothing
method were used to update the inertia constant in real-time.
The proposed method was tested in the IEEE 39-bus
system. The results confirmed that different hierarchical
power system inertia constant (individual generator, area
and the whole system) could be estimated with high
accuracy using ambient data, which could overcome the
disadvantages of the disturbance data methods. Moreover,
by using the proposed method, the inertia constant could be
FIGURE 11. Inertia constant tracking trajectory of generator 8 updated on a time scale of seconds, and the inertia constant
tracking trajectories could also be provided in real-time.
According to section Ⅲ-D, the execution time of the
procedures depends on the choice of sliding window length REFERENCES
[1] P. Kundur, Power System Stability and Control. New York, NY,
and estimation refresh rate. Table V shows the execution USA: McGraw-Hill, 1994, pp.128-136.
time for different settings of sliding window length and [2] V. Gevorgian, Y. Zhang and E. Ela, "Investigating the impacts of
refresh rate with a 120-min measurement. The simulations wind generation participation in interconnection frequency
response," IEEE Trans. Sustain. Energy, vol. 6, no. 3, pp. 1004-1012,
are carried out in a regular office laptop with Inter(R) July 2015.
Core(TM) i7-7700HQ and CPU at 2.80 GHz. [3] Eftekharnejad S, Vittal V, Heydt G T, B Keel, J Loehr. “Impact of
increased penetration of photovoltaic generation on power systems,”
TABLE V IEEE Trans. Power Syst., vol. 28, no. 1, pp. 893-901, May 2013.
EXECUTION TIME FOR DIFFERENT SETTINGS [4] Adrees A, Papadopoulos P N, Milanovic J V. “A framework to
Sliding Window Refresh Rate Execution assess the effect of reduction in inertia on system frequency
Setting response,” 2016 IEEE Power and Energy Society General Meeting,
Length (s) (s) Time(s)
1 100 1 2217.55 pp. 1-5.
2 100 2 1075.64 [5] H. Shao et al., "Equivalent Modeling and Comprehensive Evaluation
3 100 4 542.35 of Inertia Emulation Control Strategy for DFIG Wind Turbine
Generator," IEEE Access, vol. 7, pp. 64798-64811, 2019.
4 100 5 447.59
[6] Arani M F M, El-Saadany E F. “Implementing virtual inertia in
DFIG-based wind power generation,” IEEE Trans. Power Syst., vol.
The results of the first four settings show that the 28, no. 2, pp. 1373-1384, May 2013.
execution time is inversely proportional to the refresh rate [7] A. Ghafouri, J. Milimonfared, and G. B. Gharehpetian. “Coordinated
control of distributed energy resources and conventional power
as fewer estimations are done with a slower refresh rate. plants for frequency control of power systems,” IEEE Trans. Smart
The execution time of the fifth setting is comparable to the Grid, vol. 6, no. 1, pp. 104-114, Jan. 2015.
execution time of the first setting, indicating that a larger [8] Ulbig A, Borsche T S, Andersson G. “Impact of low rotational
inertia on power system stability and operation,” IFAC Proc.
sliding window length will increase the computational Volumes, Vol. 47, no. 3, pp. 7290-7279, Dec. 2014.
burden. Additionally, processing one measurement period [9] Spahic E, Varma D, Beck G, Kuhn G, Hild V. “Impact of reduced
takes approximately 6.22-31.36% of its length using the system inertia on stable power system operation and an overview of
possible solutions,” 2016 IEEE Power and Energy Society General
laptop introduced above. On average, it takes 3.73-18.81 s Meeting, pp. 1-5.
to analyze 1 min of the measurements. In addition, a sliding [10] T. Inoue, H. Taniguchi, Y. Ikeguchi and K. Yoshida, "Estimation of
window length of 100-200 s with a refresh rate of 1-2 s is power system inertia constant and capacity of spinning-reserve
responsive enough to obtain reliably accurate results in our support generators using measured frequency transients," IEEE
Trans. Power Syst., vol. 12, no. 1, pp. 136-143, Feb. 1997.
simulations.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

[11] D. P. Chassin, Z. Huang, M. K. Donnelly, C. Hassler, E. Ramirez stochastic subspace identification," IEEE Trans. Power Syst., vol. 29,
and C. Ray, "Estimation of WECC system inertia using observed no. 1, pp. 349-358, Jan. 2014.
frequency transients," IEEE Trans. Power Syst., vol. 20, no. 2, pp. [31] Canizares C, Fernandes T, Geraldi E, et al. “Benchmark models for
1190-1192, May 2005. the analysis and control of small-signal oscillatory dynamics in
[12] P. Du and Y. Makarov, "Using disturbance data to monitor primary power systems,” IEEE Trans. Power Syst., vol. 32, no. 1, pp. 715-
frequency response for power system interconnections," IEEE Trans. 722, Jan. 2017. 1943.
Power Syst., vol. 29, no. 3, pp. 1431-1432, May 2014. [32] N. Ma and D. Wang, "Extracting Spatial-Temporal Characteristics of
[13] Ashton P M, Saunders C S, Taylor G A, Carter A M, Bradley M E. Frequency Dynamic in Large-Scale Power Grids," IEEE Trans.
“Inertia estimation of the GB power system using synchrophasor Power Syst., vol. 34, no. 4, pp. 2654-2662, July 2019.
measurements,” IEEE Trans. Power Syst., vol. 30, no.2, pp. 701-709,
Mar. 2015.
[14] P. Wall and V. Terzija, "Simultaneous estimation of the time of
disturbance and inertia in power systems," IEEE Trans. Power Del.,
Fanhong Zeng (S’19) was born in Guangdong,
vol. 29, no. 4, pp. 2018-2031, Aug. 2014.
China, on Mar. 25, 1994. He received his B.Eng.
[15] D. Zografos and M. Ghandhari, "Estimation of power system degree in electrical engineering from the School
inertia," 2016 IEEE Power and Energy Society General Meeting
of Electric Power, South China University of
(PESGM), Boston, MA, 2016, pp. 1-5.
Technology, in 2017. He is currently working
[16] Y. Bian, H. Wyman-Pain, F. Li, R. Bhakar, S. Mishra and N. P. toward his M. Eng. degree in the same department.
Padhy, "Demand side contributions for system inertia in the GB
His research interests include power system
power system," IEEE Trans. Power Syst., vol. 33, no. 4, pp. 3521-
stability and control, system identification and
3530, July 2018. power system inertia estimation.
[17] D. del Giudice and S. Grillo, "Analysis of the Sensitivity of the
Extended Kalman Filter Based Inertia Estimation Method to the
Assumed Time of Disturbance," 2018 IEEE International
Conference on Environment and Electrical Engineering and 2018
IEEE Industrial and Commercial Power Systems Europe (EEEIC /
I&CPS Europe), Palermo, 2018, pp. 1-6.
Junbo Zhang (S’10-M’14-SM’18) received his
[18] A. Darbandsari, A. Maroufkhani and T. Amraee, "The estimation of
B.S. and Ph. D. degree in Electrical Engineering
inertia and load damping constants using phasor measurement data,"
from Tsinghua University in 2008 and 2013,
2017 Smart Grid Conference (SGC), Tehran, 2017, pp. 1-7
respectively. He visited the Hongkong
[19] R. K. Panda, A. Mohapatra and S. C. Srivastava, "Online Estimation
Polytechnic University and Stanford University
of System Inertia in a Power Network utilizing Synchrophasor
from 2009 to 2010, and from 2018 to 2019,
Measurements," IEEE Trans. Power Syst., Early Access.
respectively. He is currently a professor at the
[20] C. Phurailatpam, Z. H. Rather, B. Bahrani and S. Doolla,
School of Electric Power Engineering, South
"Measurement Based Estimation of Inertia in AC Microgrids," in
China University of Technology. His research
IEEE Trans. Sustainable Energy, Early Access.
areas include artificial intelligent and hyper-
[21] You S, Liu Y, Kou G, et al. “Non-invasive identification of inertia
automatics in complex system operation and decision making, knowledge-
distribution change in high renewable systems using distribution
based expert systems with distributed cloud computing, power system
level PMU,” IEEE Trans. Power Syst., vol. 33, no. 1, pp. 1110-1112,
simulation with high performance parallel computing, and power system
Jan. 2018.
operation and control.
[22] X. Cao, Bv f. Stephen, I. F. Abdulhadi, C. D. Booth and G. M. Burt,
"Switching Markov Gaussian models for dynamic power system
inertia estimation," IEEE Trans. Power Syst., vol. 31, no. 5, pp.
3394-3403, Sept. 2016.
[23] K. Tuttelberg, J. Kilter, D. Wilson and K. Uhlen, "Estimation of
Ge Chen (S’20) was born in Guangdong, China,
power system inertia from ambient wide area measurements," IEEE
on Nov. 14, 1995. He received the bachelor’s
Trans. Power Syst., vol. 33, no. 6, pp. 7249-7257, Nov. 2018.
degree in electrical engineering from the School
[24] J. Zhang and H. Xu, "Online identification of power system
of Electric Power Engineering, South China
equivalent inertia constant," IEEE Trans. Ind. Electron., vol. 64, no.
University of Technology, in 2018. He is
10, pp. 8098-8107, Oct. 2017.
currently pursuing his M. Eng. Degree in
[25] D'Arco, S.; Suul, J.A., "Equivalence of Virtual Synchronous
electrical engineering in the same department. His
Machines and Frequency-Droops for Converter-Based MicroGrids,"
research interests include power system dispatch,
in Smart Grid, IEEE Transactions on, vol.5, no.1, pp.394-395, Jan.
and application of artificial intelligence in smart
2014.
grids.
[26] Y. Li, Z. Xu and K. P. Wong, "Advanced Control Strategies of
PMSG-Based Wind Turbines for System Inertia Support," in IEEE
Transactions on Power Systems, vol. 32, no. 4, pp. 3027-3037, July
2017.
[27] L. Xiong et al., "Static Synchronous Generator Model: A New
Perspective to Investigate Dynamic Characteristics and Stability
Issues of Grid-Tied PWM Inverter," in IEEE Transactions on Power
Zikun Wu was born in Guangdong, China, on
Electronics, vol. 31, no. 9, pp. 6264-6280, Sept. 2016.
Feb. 27, 1996. He received his B.Eng. degree in
[28] Ning Zhou, J. W. Pierre and J. F. Hauer, "Initial results in power
electrical engineering from the School of Electric
system identification from injected probing signals using a subspace
Power, South China University of Technology, in
method," IEEE Trans. Power Syst., vol. 21, no. 3, pp. 1296-1302,
2019. He is currently pursuing his M. Eng. degree
Aug. 2006.
in electrical engineering in the same department.
[29] J. Zhang and H. Xu, "Microperturbation method for power system
His research interests include power system
online model identification," IEEE Trans. Ind. Informat., vol. 12, no.
modeling, simulation and analysis, and software
3, pp. 1055-1063, June 2016.
architecture in power systems.
[30] S. A. Nezam Sarmadi and V. Venkatasubramanian,
"Electromechanical mode estimation using recursive adaptive

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2997728, IEEE Access

Siwei Huang was born in ShanXi, China, on Oct.

08, 1994. She received her B.Eng. degree in
electrical engineering from the Hehai University
in 2017. She is currently pursuing her M. Eng.
degree in electrical engineering in South China
University of Technology. Her research interests
include system identification and power system
inertia estimation, and application of artificial
intelligence algorithms in power system.

Yingqi Liang (S’18-M’18) received her B.Eng.

degree in Electrical Engineering from South China
University of Technology and is pursuing her
Ph.D. degree in Electrical and Computer
Engineering at National University of Singapore.
Her research interest lies in the learning,
optimization, and control of networked cyber-
physical systems, with particular applications to
power systems.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.