IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 59, NO.

4, JULY/AUGUST 2023 4893

Machine Learning Assisted Inertia Estimation

Using Ambient Measurements
Mingjian Tuo , Student Member, IEEE, and Xingpeng Li , Senior Member, IEEE

Abstract—With the increasing penetration of converter-based results in under frequency load shedding as well as tripping
renewable resources, different types of dynamics have been intro- of generator protection devices; the failure of successive units
duced to the power system. Due to the complexity and high order would furthermore cause cascading outages [3].
of the modern power system, mathematical model-based inertia
estimation method becomes more difficult. This paper proposes Inertia estimation can ensure the accountability and reliability
two novel machine learning assisted inertia estimation methods of inertia response through implementation of frequency control
based on long-recurrent convolutional neural (LRCN) network and ancillary services [4], [5]. In large-scale deregulated intercon-
graph convolutional neural (GCN) network respectively. Informa- nection power systems, inertia information is only available
tive features are extracted from ambient measurements collected within operators’ own territories. Thus, system-wide inertia esti-
through phasor measurement units (PMU). Spatial structure with
high dimensional features and graphical information are then in- mation is important for operators to provide frequency regulation
corporated to improve the accuracy of the inertia estimation. Case services. Traditionally, system frequency response is analyzed
studies are conducted on the IEEE 24-bus system. The proposed by looking at the collective performance of all generators using
LRCN and GCN based inertia estimation models achieve an ac- a system equivalent model. Based on event measurements and
curacy of 97.34% and 98.15% respectively. Furthermore, the pro- mathematical model, the system inertia could be estimated
posed zero generation injection bus based optimal PMU placement
(ZGIB-OPP) has been proved to be able to maximize the system by the number and size of actively connected synchronous
observability, which subsequently improves the performance of all units.
proposed inertia estimation models. Reference [6] proposed an inertia estimation approach which
Index Terms—Ambient synchrophasor data, graph neural
divides the system into multiple subareas and estimates iner-
network, inertia estimation, low inertia power grid, phasor tia of each subarea separately, but the approximation made in
measurement units. mathematical model introduces additional errors. The Electric
Reliability Council of Texas (ERCOT) uses a real-time suffi-
ciency monitoring tool to monitor inertia based on the operating
I. INTRODUCTION plans submitted by the generation resources [7]. As previous
ENEWABLE energy sources (RES) are replacing tradi- inertia estimation methods are based on mathematical models,
R tional synchronous generators with the primary goal of
carbon dioxide emission reduction and environmental benefits
the ability of these methods is dependent on factors such as the
size of disturbance, accuracy of frequency measurement and lo-
[1]. Increasing penetration of inverter-based resources such cation of measurement point relative to in-feed loss [8]. They are
as wind power, solar photovoltaics (PV) and energy storage difficult to implement in inertia monitoring due to randomness of
systems (ESS), has degraded the system inertia during this system events. Therefore, inertia estimation using ambient wide
transition and introduced different dynamics into traditional area measurements are considered as a more accurate method
power systems [2]. Traditionally, power system inertia plays an that can reflect the system-wide status.
important role in regulating rate of change of frequency (RoCoF) Inertia estimation method based on mathematical model is
and frequency excursion after a disturbance. Insufficient system highly dependent on accuracy of measurements from phasor
inertia would lead to dramatical change in frequency and further measurement units (PMUs) or equivalent devices. However,
modern power systems are connected to different devices which
provide frequency regulation service, and inertia constant esti-
Manuscript received 16 December 2022; revised 15 March 2023; accepted 15
April 2023. Date of publication 25 April 2023; date of current version 19 July mation purely based on synchronous generators is inaccurate [9].
2023. Paper 2022-PSEC-1390.R1, presented at the 2022 IEEE Industry Applica- Moreover, RES and other inverter-based sources are traditionally
tions Society Annual Meeting, Detroit, MI, USA, Oct. 09–14, and approved for considered passive in terms of inertial response. The variability
publication in the IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS
by the Power Systems Engineering Committee of the IEEE Industry Applica- nature of RES also imports uncertainties into the system inertial
tions Society [DOI: 10.1109/IAS54023.2022.9940112]. Long-Term Recurrent response as well as system inertia constant [10]. Recent study
Convolutional Network-based Inertia Estimation using Ambient Measurements. in [11] shows that control schemes emulating synchronous
(Corresponding author: Xingpeng Li.)
The authors are with the Department of Electrical and Computer Engineer- machine response can be used to contribute system inertia.
ing, University of Houston, Houston, TX 77204 USA (e-mail: [email protected]; Therefore, the swing equation-based models may not be able
[email protected]). to capture the entire characteristics. In addition, nonlinearities
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TIA.2023.3269732. in system frequency response such as deadbands and saturations
Digital Object Identifier 10.1109/TIA.2023.3269732 cannot be taken into considerations either. Thus, the estimated

0093-9994 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Indian Institute of Technology Patna. Downloaded on June 05,2024 at 04:30:24 UTC from IEEE Xplore. Restrictions apply.
4894 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 59, NO. 4, JULY/AUGUST 2023

value based on mathematical model may suffer inaccuracy in the simulation setup, and the results and analysis are presented
various conditions. in Section V. Section VI presents the concluding remarks and
The invention and development of PMU based wide area future work.
measurements systems (WAMS) enable the application of data-
driven techniques in power system analysis [12]. A neural II. SYSTEM FREQUENCY DYNAMICS
network-based inertia estimation technique is proposed in [13],
The frequency of the power system is one of the most im-
which utilizes inter-area model information as neural network
portant metrics that indicate the system stability. Traditionally,
inputs and estimates the inertia constant as an output of the
the frequency is treated as unique of the whole power system,
network. However, this approach only estimates the inertia
which is derived from the system equivalent model extended
constant for large systems with only traditional synchronous
from one-machine swing equation.
generation. A convolutional neural network (CNN) based model
The inertia constant of a generator is a parameter describing
is proposed in [14], which estimates the system inertia through
the ability of synchronous generator in counteracting the fre-
frequency response and RoCoF data. Graphs are a kind of data
quency excursion due to power imbalance occurring in power
structure which models a set of objects (nodes) and their rela-
systems. The energy stored in large rotating generator and some
tionships (edges) [15]. Recent advances in deep neural network
industrial motors gives them the tendency to remain rotating.
(DNN) offer an opportunity to integrate graph topology into a
The rotational energy Ei in the rotor of the machine at nominal
neural network, creating a graph neural network (GNN) model
speed is defined by the following formula:
[16]. Power system can be represented as a graph with high
dimensional features and interdependency among buses. This 1
Ei = Ji ωi2 (1)
perspective may offer a better state of the art machine learning 2
for power systems analysis. where Ji is the moment of inertia of the shaft in kg·m2 s and ωi
To bridge aforementioned gaps, we propose two model-free is the nominal rotational speed. The inertia constant Hi is then
ambient measurements-based machine learning approaches in given in seconds, which can be expressed as:
this paper to dynamically estimate the system inertia constant.
The major contributions of this work are as follows, Ji ωi2
Hi = (2)
1) First, although data driven approaches have been investi- 2SBi
gated for system inertia estimation in previous work [13],
where SBi is the generator rated power in MVA. When multiple
the topological information and high dimensional features
generators connected to the power system, dynamics of these
haven’t been studied thoroughly. To tackle this issue, a
generators’ rotors are directly coupled with the grid electrical
long-term recurrent convolutional network (LRCN) based
dynamics. Thereby the power system could be represented by a
algorithm are proposed to efficiently process temporal
single equivalent model of inertia. The total power system inertia
measurements, and a graph convolutional neural networks
Esys is then considered as the summation of the kinetic energy
(GCN) assisted method model is used to efficiently iden-
stored in all dispatched generators synchronized with the power
tify spatial data. The proposed models have been com-
system. It can be shown in the form of either the stored kinetic
pared with other state of art methods such as CNN and
energy or inertia constants as follows.
DNN models.
2) Secondly, previous inertia estimation methods only uti- N
 N

1
lize frequency and RoCoF data derived from single fre- Esys = Ji ωi2 = Hi S B i (3)
i=1
2 i=1
quency measurement, results may suffer high errors when
non-monotonic frequency deviation occurs. We consider The inertia constant of the power system in seconds is given
measuring heterogeneous responses from multiple nodal by the equation below,
PMUs, and the extracted features are then reformulated N
into the form of graph structure for GCN training. Hi S B i
Hsys = i = 1 (4)
3) Thirdly, a wrapper feature selection algorithm is used SB
to optimize the feature combination set for inertia esti- where SB is the total rated power of the whole system.
mation. The proposed ambient measurements-based al- The simplified system equivalent model is based on the ex-
gorithms are examined under multiple noise conditions, tension of one-machine swing equation. For a single machine,
which demonstrates the proposed methods can improve the dynamic of its rotor can be described in (5) with M =
the estimation accuracy as well as estimator robustness. 2H denoting the normalized inertia constant and D denoting
4) Last, a zero generation injection bus based optimal PMU damping constant respectively.
placement (ZGIB-OPP) method is proposed in this paper
dΔω
to maximize the observability of WASM, which further- ΔPm − ΔPe = M + DΔω (5)
more improves the performance of all inertia estimation dt
models. where ΔPm is the total change in mechanical power and ΔPe is
The remainder of this paper is organized as follows. In the total change in electric power of the power system. dΔω/dt
Section II, the frequency dynamics of power systems are de- is commonly known as RoCoF.
scribed. Section III details the proposed inertia estimation algo- The dynamics between power and frequency during a short
rithms using LRCN and GCN techniques. Section IV describes period of time following a disturbance can be modeled by the
Authorized licensed use limited to: Indian Institute of Technology Patna. Downloaded on June 05,2024 at 04:30:24 UTC from IEEE Xplore. Restrictions apply.
TUO AND LI: MACHINE LEARNING ASSISTED INERTIA ESTIMATION USING AMBIENT MEASUREMENTS 4895

Fig. 1. Generator transfer function model.

swing equation. As an approximation, the equation commonly

used for system equivalent model is expressed as follows,
dΔω ΔPm − ΔPe
= ωn (6) Fig. 2. Wide area measurement system.
dt 2HSB
If the size of disturbance and RoCoF are accurately measured
in advance, then the total inertia of the system can be estimated
based on the equivalent expression. However, such approxi- III. INERTIA ESTIMATION
mation fails to capture the entire characteristics, nonlinearities
in system frequency response and controls are not taken into A. Wide Area Monitoring System
consideration. Synchronized measurement technology makes it possible to
The power system with frequency control loops is shown in sample analogue voltage and current wave data in synchronism
Fig. 1. For primary frequency control, once the power mismatch with a global positioning system (GPS) clock, and record the
event has occurred and frequency has started to drop from corresponding frequency related data from widely distributed
nominal value, the deviation is fed into closed control droop locations.
where turbine-governor counteracts the power mismatch [17]. PMUs are widely used for modern power systems. Fig. 2
In this paper, we only focus on the short period following the shows the topology of wide area monitoring system (WAMS).
disturbance. Thus, the model can be simplified further when Measurements from PMUs are obtained from widely distributed
we analyze only primary frequency control, i.e., model the locations, and synchronized with respect to a GPS clock. Syn-
dynamics before any secondary control gets involved. chrophasor technologies allow direct measurement of frequency
As most inertia estimation approaches rely on event transient and bus voltages. With the development of PMU based WAMS,
measurement of collective system model following recorded the accuracy of measurements improves significantly.
disturbances, study in [18] have found that the approximation Most PMUs can calculate up to 30 to 60 samples per cycle with
may introduce high error due to inertia heterogeneity, and thus the GPS time stamp provided by hardware that has an accuracy
cause issues in system operations. of millisecond or higher [12], reporting rates of 10 - 240 samples
Different from focusing on the collective performance of the per second are allowed. In this paper, the sampling rate of PMU
power system equivalent model, the frequency response expe- is set to 200 per second.
rienced by each bus could be very distinct. Therefore, dynamic
model is preferred in modern power system analysis. Using the
topological information and system parameters, when multiple
generators connected in a bus, equivalent (5) can be extended B. System Perturbation Using Probing Signal
and applied to all buses to describe the oscillatory behavior of Low level probing signal method has been conventionally
each individual bus, used for generator dynamic studies. A modified form of de-
trended fluctuation analysis has been introduced to determine
mi θ̈i + di θ̇i = pin,i − pe,i (7)
the event suitability for probing signal method [14]. With
where mi and di denote the inertia coefficient and damping PMUs installed throughout the system providing highly accurate
ratio for node i respectively, while pin,i and pe,i refer to the measurements and test improvement such as microperturbation
power input and electrical power output, respectively. If bus method (MPM), the probing signal method can provide effec-
i is asynchronous machine, then mi is the generator’s inertia, tive approach for system inertia identification without affecting
and di would be primary control [19]. If bus i connected to system stability [21].
a inverter-based resources, di is the droop control coefficient, A sample probing signal, fed to the system with an amplitude
and mi accounts for power measurement constant, a control of PE , and corresponding PMU measurements are shown in
gain [11], [20]. Accordingly, the ambient measurements of Fig. 3. With varying system inertia and probing signal amplitude,
nodes from the system would provide more information of each a number of ambient measurements of Δω, Δω̇ and v can then
subarea and thus improve the accuracy of estimation model. be collected.

Authorized licensed use limited to: Indian Institute of Technology Patna. Downloaded on June 05,2024 at 04:30:24 UTC from IEEE Xplore. Restrictions apply.
4896 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 59, NO. 4, JULY/AUGUST 2023

Fig. 6. General architecture of proposed LRCN model.

data. An LSTM cell typically compromises three gates: input,

Fig. 3. A sample of probing signal, ambient measurements for PE = 0.001
p.u. forget and output gates [24] as shown in Fig. 5.
The fundamental equations of LSTM network can be repre-
sented as follows:
ht = (1 − zt ) ∗ ht−1 + zt ∗ ht (8)
zt = σ (Wf [ht−1 , xt ] + bf ) (9)
where, xt is the network input; ht is the output state of the neuron
from LSTM network; ht−1 is the previous state of the neuron; zt
computes the necessary information and removes the irrelevant
Fig. 4. Illustration of convolutional neural network architecture [22], [23]. data; σ is the sigmoid function; Wf is the weight and bf is the
bias.
Since LRCN leverages the strength of rapid progress in CNN
and has the ability to capture the dependencies in a sequence,
it has been successfully used in computer vision, image pro-
cessing, and other fields in signals and time-series analysis [25].
The architecture of the proposed LRCN model is illustrated in
Fig. 6 and can be trained to estimate system inertia from ambient
measurements obtained from the PMU.
The proposed LRCN model first processes the measurements
input with 1-D-CNN layers, whose outputs are then fed into the
LSTM recurrent sequence model; and the fully connected layer
finally produces the estimated inertia constant. The samples in
training set are defined in batches which will be propagated
through the networks. One epoch of training is completed when
Fig. 5. Illustration of an LSTM cell.
all the training samples have been passed forward and backward
once. The number of iterations is defined as the number of
passes, and each pass uses the same batch size that is the number
C. Inertia Estimation Using LRCN of samples. At each training iteration, the LRCN model input
Motivation of CNNs roots in the history of neural networks size is b × c, and the output will be a column vector of size b
for graph data processing, recurrent neural networks (RNN) are with inertia estimates for corresponding input in the time period.
utilized on graphs and cycles. Study in [22] has shown that While the dimension of c is determined by the set of features
CNNs have the ability to extract spatial features and compose and feature sampling rate.
them to construct expressive representations. An example of a The mean squared error (MSE) measures the average squared
convolutional neural network is shown in Fig. 4 [22], [23]. difference between actual and predicted outputs. The goal of
Long short-term memory (LSTM) is an extended frame of training is to minimize MSE via back propagation which will
RNN which can exhibit temporal behavior of time-series input provides best estimator [26]. Comparing to root mean squared

Authorized licensed use limited to: Indian Institute of Technology Patna. Downloaded on June 05,2024 at 04:30:24 UTC from IEEE Xplore. Restrictions apply.
TUO AND LI: MACHINE LEARNING ASSISTED INERTIA ESTIMATION USING AMBIENT MEASUREMENTS 4897

error (RMSE) and mean absolute value, MSE is more sensitive

to outliers in absolute terms. Such selection helps us to train
the model with no outlier predictions with huge errors The fully
connected network used in this model includes one flatten layer
and two hidden layers. MSE is defined as:
n
1 
M SE = (yi − ỹi )2 (10)
n i=1

where n is the total number of training samples, yi is the actual

value of ith output, and ỹi is the estimated value corresponding
to the ith output. Similarly, the weight update equation via back
propagation is expressed as:
∂EM SE
wt+1 = wt − α (11)
∂wt
where wt is the weight for current iteration, wt+1 is the updated
weight for next iteration, α is the learning rate, and EM SE is
the MSE obtained from expression (11).

D. Graph Convolutional Neural Networks Fig. 7. Example of message passing mechanism.

Geometric deep learning is a recent emerging field. Tra-

ditional CNNs have limitations in processing graphical data
which have explicit topological graph correlation embedded
E. Optimal PMU Placement
[27]. Recent advancement of CNN results in the rediscovery of
GNNs. GCN has been developed by extending the convolution The cost of PMUs depends on the count of measurement
operation onto graphs and in general onto non-Euclidean spaces. channels available [30]. Due to the high cost of having a PMU
Previous studies in [28], [29] have proved that GCN provides at each node, various techniques are used to solve the optimal
state-of-arts performance in graph analysis tasks. PMU placement (OPP) problem for observability of the system
Power system is an interconnected network of generators for static and dynamic state estimators [31].
and loads. The graph structure of the power system consists of Given a system with N buses, traditionally OPP is a binary
nodes (buses) and edges (branches). The branches in the power optimization problem, the objective functions considered in OPP
system are undirected, such graphs provide information on buses problems in previous studies are mainly for minimization of the
and their connections. The convolution operator in propagation number of PMUs (15), maximization of observability (16), or
module is used to aggregate information from neighbors. Con- both as a multi-objective function.
sidering G = (V, E) as an undirected graph representing a
power system, where V ∈ RN denotes its nodes and E ∈ RK N

denotes its edges. Let A ∈ RN ×N be the adjacency matrix of G, Min F1 = xi (15)
we can define a renormalization trick as: i=1

V = D̃− 2 ÃD̃− 2 N

1 1
(12)
Max F2 = oi (16)
where à = A + IN represents an adjacency matrix with self- i=1
connections. Typically, the element at (i, j) of the adjacency
matric A is defined as follows: where xi , i = 12, . . . , N is elements of the PMU installation

1; if Vi , Vj ∈ V, (Vi , Vj ) ∈ E indication row vector X. For the base case, oi denotes the observ-
Aij = (13)
0; if Vi , Vj ∈ V, (Vi , Vj ) ∈
/E ability of each bus. In this work, we focus on the maximization
of observability of bus considering resources on neighboring
where (Vi , Vj ) denotes the branches from i to j. The diagonal
 connected buses. With limited PMU resources. The complete
degree matrix D̃ for G is defined as D̃ii = j Ãij . topological observability can be expressed as follows,
The graph convolutional layer is defined as follows,
 
F l (X, A) = σ V F (l−1) (X, A) Wkl + bl (14) O = A·X (17)
O≥u (18)
where F l is the convolutional activations and bl is the bias matrix
at the l-th layer; F 0 = X is the input matrix. Fig. 7 demonstrates
the message passing mechanism in forward propagation, a target where u is a row vector N × 1 consisting of binary variables,
node (bus 8) receiving information from its neighboring nodes. representing that the monitored bus is observed by PMU.

Authorized licensed use limited to: Indian Institute of Technology Patna. Downloaded on June 05,2024 at 04:30:24 UTC from IEEE Xplore. Restrictions apply.
4898 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 59, NO. 4, JULY/AUGUST 2023

IV. SIMULATION SETUP as follows.

A. Overview |{y ∈ T D | |y − ỹ| ≤ μ }|
ACC = (19)
The IEEE 24-bus system [32] was used for the experiment to | T D|
collect the training data. The system has 24 buses (17 buses with
where ACC denotes the proportion of the correctly predicted
loads), 38 branches, and 38 generators. The system inertia M
values with μ tolerance. T D denotes the validation dataset, and
typically ranges from 3s to 8s. Hence, to ensure the practicality of
μ is the predetermined threshold.
the proposed model, the measurements snapshots were collected
for 11 different values of M from 3s to 8s with an increment of
0.5s following daily dispatches considering the RES penetration. D. PMU Formulations
Similarly, probing signals with 100 different values of PE from Constraints of PMU locations are not considered in the base
0.001 p.u. to 0.01 p.u. with an increment 0.0001 p.u. were used. case. With a resampling rate of 200Hz, measurements on each
The modeling and simulation of the power system, along with PMU node gives 200 data points at a sampling frame of 1s. In
data collection, were conducted in MATLAB/Simulink 2019b. this paper, ambient measurements of Δω, Δω̇ and v on generator
The data pre-processing was conducted in both MATLAB and buses are assumed as available measurements.
Python. The proposed LRCN and GCN based models were Availability of PMU data are affected by several realistic
developed in Python using Keras and PYG. factors, which have been studied in [8]. Given limited resources,
the objective function (16) of OPP is applied in this work to
maximize the system observations.
B. Data Preprocessing Zero injection bus effect is also added. Traditionally, zero
The initial data analyzed in this study were acquired from injection bus (ZIB) means no load or generator is connected to
PMUs with a sampling rate larger than 200 Hz; nodal measure- it. Since we are more interested in the statement of generators,
ments of Δω and Δω̇ are obtained. By using only one second the impact of zero generation injection bus (ZGIB) is considered
sampling frame for normalization, the real-time applicability in this work where generator is the only factor. The topological
of this method is maintained. Similarly, following the same observability constraint of each of the bus connected to a ZGIB
pattern we obtained nodal voltage measurement v. Since the is updated by introducing virtual connections to every other non-
training data come from the ambient measurements of all PMUs ZGIB bus connected to that same ZGIB.
may suffer different sampling rate which is larger than 200 N N
 
Hz, without dimension reduction process the original training õi = Aij xj + wij xj ∀i = 12, . . . , N (20)
data would increase the complexity of the model and may also j j
lead to overfitting. In addition, there are bad data in the raw
measurements which would introduce high error in analysis where wij is an auxiliary binary variable: 1, if buses i and j are
results. Therefore, the bad data points are first discarded from both connected to the same ZIB; otherwise 0.
raw measurements, and then we downsample the measurements
to 200 Hz for next step [14]. Additionally, Gaussian noise signal E. Hyperparameters Selection
is added to the constituent tonic to mimic the noisy measure-
ments. Different signal-to-noise ratios (SNR) are investigated For the convolution layers in the LRCN model, the channel
in this paper. The data are collected between multiple sessions, number are set p = 10 and q = 20, and kernels with sizes r
all the measurements are normalized by employing min/max = s = 3. Rectified linear unit (ReLU) is used as the activation
normalization between [0, 1]. function. The memory unit value l of LSTM layer is set as 32.
The training was operated in batches of 32 data points.
The GCN model consists of one GNN layer and two hidden
C. Feature Selection fully connected layer. The dimension of input features is defined
based on the number of available PMUs which gives n × d,
To find the best time frame of data extraction, different time where d is the nodal features of GCN layer. The fully connected
windows of the ambient measurements are determined: (1) the layers are set as 64 and 128 respectively.
time frame is first chosen from 0s to 1s following the perturba- An MSE based dynamic learning rate strategy is used for
tion, where initial RoCoF is included; (2) the second time frame the training. Learning rate schedule is applied in the training
is from 0.5s to 1.5s after the signal infeed. With Δω and Δω̇ as process by reducing the learning rate accordingly, and the factor
basic features combination, the coefficient of determination and by which the learning rate will be reduced is set 0.5 and the
validation accuracy are used as evaluating metrics. patience value is set 50 epochs.
In order to determine the optimal combination of features for
inertia estimation model training, a wrapper feature selection
method is utilized: (i) the proposed LRCN model is used as the V. RESULT ANALYSIS
inertia estimator, (ii) accuracy score is used as the evaluation A total of 1100 samples were collected, the entire dataset is
metric, (iii) greedy forward selection as the subset selection first divided into two subsets: 880 samples (80%) for training
policy. The specific metrics for feature evaluations is expressed and 220 samples (20%) for validation.

Authorized licensed use limited to: Indian Institute of Technology Patna. Downloaded on June 05,2024 at 04:30:24 UTC from IEEE Xplore. Restrictions apply.
TUO AND LI: MACHINE LEARNING ASSISTED INERTIA ESTIMATION USING AMBIENT MEASUREMENTS 4899

TABLE I
PERFORMANCE COMPARISON FOR DIFFERENT MODELS

TABLE II
COMPARISON OF DIFFERENT FEATURES SETS FOR LRCN MODEL

A. Time Frame Selection

To investigate the best time period for feature extraction, all
PMUs are considered available in base case. Measurement Δω̇ is
selected as training feature combination; training data extracted
from two periods are then fed into the aforementioned models,
including the proposed LRCN and GCN models.
Performances of different inertia estimation models are sum- Fig. 8. Absolute error of prediction with the proposed LRCN model using
marized in Table I. As it can be observed, the DNN model features extracted from different time periods.
has the lowest validation accuracy for both scenarios. CNN
based model has a relatively higher validation accuracy, showing
advantage in processing spatial data. The proposed LRCN model
has a validation accuracy of 96.89% with a tolerance of 0.5s for
features extracted from period 0.0 - 1.0s period, while it is only
82.57% for the use of features extracted from period 0.5 - 1.5s.
GCN based model has the highest validation accuracy in both
scenarios at 98.15% and 83.19% respectively. On the whole,
the features extracted from the time frame following the dis-
turbance contain prominent inertial response information, and
accordingly have a positive impact on the overall performance
of inertia constant estimation model.
Fig. 8 compares the distribution of absolute prediction error
for the LRCN model with features extracted from 0.5s - 1.5s and
0.0s - 1.0s respectively. Using features extracted from period 0.0
- 1.0s, the coefficient of determination of the proposed LRCN
model is 0.9625, which is higher the use of features extracted Fig. 9. The learning curve of the proposed LRCN model: MSE losses versus
from period 0.5 - 1.5s at 0.7619. the number of epochs.

B. Analysis of Model Performance

feature combinations. Thus, Δω + Δω̇ is selected as the optimal
Optimal combination of features extracted from 0.5s - combination for model training.
1.5s period is first selected through greedy forward selection. Figs. 9 and 10 depict the evolution of MSE losses on the
Table II compares the results with LRCN model as performance training and validation datasets over the training process of the
estimator. It can be observed that only considering Δω mea- proposed models. As it can be seen, for both LRCN and GCN
surement as input feature provides a validation accuracy of cases, MSE decreases as the number of epochs increases. In
80.30%. Combination of Δω and Δω̇ has a highest validation terms of MSE, the validation loss of LRCN model drops faster
accuracy at 97.34% with 0.5s tolerance, which outperforms other and reaches minimal value at 0.025. It should be noted that a

Authorized licensed use limited to: Indian Institute of Technology Patna. Downloaded on June 05,2024 at 04:30:24 UTC from IEEE Xplore. Restrictions apply.
4900 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 59, NO. 4, JULY/AUGUST 2023

TABLE IV
COMPARISON OF DIFFERENT MODELS WITH SNR AT 45DB

Fig. 10. The learning curve of the proposed GCN model: MSE losses versus.
the number of epochs.

TABLE III
COMPARISON OF DIFFERENT MODELS WITH OPTIMAL FEATURE COMBINATION

Fig. 11. Prediction results of the benchmark CNN model with SNR at 45dB.

sudden drop is observed in GCN model training, and the GCN noise power has been recognized by most of the researchers [33].
based model has a lower validation loss at 0.020. Usually, lowpass Butterworth filter could be applied to isolate the
The proposed LRCN and GCN based approaches are then system inertial frequency response from the measured transient
compared with benchmark algorithm [14] in Table III, Δω noise [8]. And bad data conditioning, missing data prediction
and Δω̇ are used as primary features to train the model. Both and statistical filter design could be also applied to data with non-
algorithms are employed to train the inertia constant estimation Gaussin distributions. However, the ideal condition cannot be
model. The results show that the CNN model has a validation guaranteed in all scenarios. The proposed models are compared
accuracy of 95.18% with 0.5s tolerance, which is higher than with the benchmark models under high noise conditions. Study
the validation accuracy of DNN model at 93.45%. These results in [33] has shown that a SNR of 45dB is considered as a good
reflect that CNN based model has the better capability to process approximation of noise power under realistic conditions.
spatial data comparing to traditional DNN model. Nevertheless, Table IV shows the inertia estimation accuracy of all models
the proposed LRCN model has a validation accuracy of 97.34% with combination of Δω and Δω̇ as training features. After
with 0.5s tolerance, and GCN model has a validation accuracy adding additional Gaussian noise signal with a SNR of 45dB
at 98.15%. Additionally, the coefficient of determinations of to the ambient measurements, the overall MSE increases for
LRCN model and GCN model are 0.9725 and 0.9826 respec- both models, while the validation accuracy reduces accordingly.
tively, which are higher than the benchmark CNN and DNN Understandably, a significant reduction in validation accuracy
models. An explanation could be that the proposed LRCN and can be observed in both cases.
GCN models are more efficient algorithms to identify critical The scatter points of CNN model, the proposed LRCN and
temporal information and graphical information embedded in GCN models when SNR of 45dB is applied are depicted in
the collected power system data. Combining Table III and Figs. 11 –13. Results show that only considering measurements
Figs. 9–10, we can observe that GCN has the highest valida- of frequency related data may suffer high error when noise is
tion accuracy and the lowest MSE mean absolute percentage applied.
error (MAPE), indicating that GCN has a better performance in The method described in this research uses a wrapper feature
processing graphical data. selection process and then. Results of different feature combi-
nations are listed in Tables V and VI. It can be observed that the
proposed LRCN model has a validation accuracy of 93.25% with
C. Impact of Low SNR
0.5s tolerance under the condition of SNR at 45dB. With voltage
PMU data always has noise associated with the measured data. measurements added, the validation accuracy is improved to
The Gaussian distribution of the noise with varying degree of 93.87% with 0.5s tolerance.

Authorized licensed use limited to: Indian Institute of Technology Patna. Downloaded on June 05,2024 at 04:30:24 UTC from IEEE Xplore. Restrictions apply.
TUO AND LI: MACHINE LEARNING ASSISTED INERTIA ESTIMATION USING AMBIENT MEASUREMENTS 4901

TABLE VI
COMPARISON OF DIFFERENT FEATURES SETS FOR GCN MODEL WHEN
CONSIDERING SNR AT 45 DB

TABLE VII
OPTIMAL LOCATIONS OF PMUS GIVEN LIMITED RESOURCES

Fig. 12. Prediction results of the proposed LRCN model with SNR at 45dB.

Fig. 14. Distribution of absolute values for random PMU locations.

Fig. 13. Prediction results of the proposed GCN model with SNR at 45dB.
Table VII listed the results of PMU location considering the
proposed ZGIB-OPP method. Given the total number of PMUs
TABLE V limited by two, the proposed OPP method considering ZGIB
COMPARISON OF DIFFERENT FEATURES SETS FOR LRCN MODEL WHEN suggests locating PMUs at buses 2 and 16 for the best inertia esti-
CONSIDERING SNR AT 45 DB
mation performance. When the total number of PMUs increases,
the proposed ZGIB-OPP adds more buses into the optimal set,
indicating the consistency in optimal PMUs locations.
We apply the proposed ZGIB-OPP to IEEE 24-bus system,
and the system wide measurements of limited channels are then
obtained for model training. Figs. 14 and 15 compare the distri-
bution of prediction absolute error for different PMUs settings
using LRCN model. Comparing to the random PMU location
In summary, (i) under a low noise condition, measurements case, the proposed ZGIB-OPP improves the performance of
of Δω and Δω̇ are the optimal set of features suitable for inertia LRCN model, resulting in more samples with absolute error
estimation; (ii) with SNR of 45dB added, the performance of landed within 0.1-0.2s.
benchmark model decreases significantly, while the proposed Results listed in Table VIII also show that with limited PMU
LRCN model based on optimal features combination of Δω, channels, as well as combination of Δω and Δω̇ as input feature,
Δω̇ and v shows higher robustness and better performance. the validation accuracy of all models drops accordingly. For
DNN and CNN model, the validation accuracy drops to 90.00%
and 90.45% respectively. Validation accuracy of LRCN and
D. Optimal PMU Placement GCN model drops slightly when there are limited channels or
The application of proposed ZGIB-OPP approach to IEEE under random PMU locations. With ZGIB-OPP applied, GCN
24-bus system is carried, and the corresponding impact on inertia based model has the highest accuracy of 95.89%. It should be
estimation models are investigated in this section. noted that applying ZGIB-OPP improves the performances of all

Authorized licensed use limited to: Indian Institute of Technology Patna. Downloaded on June 05,2024 at 04:30:24 UTC from IEEE Xplore. Restrictions apply.
4902 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 59, NO. 4, JULY/AUGUST 2023

[2] M. Tuo and X. Li, “Dynamic estimation of power system inertia distri-
bution using synchrophasor measurements,” in Proc. 52nd North Amer.
Power Symp., 2021, pp. 1–6.
[3] F. Milano, F. Dörfler, G. Hug, D. J. Hill, and G. Verbic, “Foundations
and challenges of low-inertia systems,” in Proc. Power Syst. Computation
Conf., 2018, pp. 1–25.
[4] S. You et al., “Comparative assessment of tactics to improve primary
frequency response without curtailing solar output in high photovoltaic
interconnection grids,” IEEE Trans. Sustain. Energy, vol. 10, no. 2,
pp. 718–728, Apr. 2019.
[5] H. Gu, R. Yan, and T. K. Saha, “Minimum synchronous inertia requirement
of renewable power systems,” IEEE Trans. Power Syst., vol. 33, no. 2,
pp. 1533–1543, Jun. 2017.
[6] Y. Cui, S. You, and Y. Liu, “Ambient synchrophasor measurement based
system inertia estimation,” in Proc. IEEE Power Eng. Soc. Gen. Meeting,
2020, pp. 1–5.
[7] ERCOT, “Inertia: Basic concepts and impacts on the ERCOT
grid,” ERCOT, Tech. Rep., Apr. 2018. [Online]. Available: https:
//www.ercot.com/files/docs/2018/04/04/Inertia_Basic_Concepts_
Impacts_On_ERCOT_v0.pdf
Fig. 15. Distribution of absolute values for ZGIB-OPP. [8] P. M. Ashton, C. S. Saunders, G. A. Taylor, A. M. Carter, and M. E.
Bradley, “Inertia estimation of the GB power system using synchrophasor
measurements,” IEEE Trans. Power Syst., vol. 30, no. 2, pp. 701–709,
TABLE VIII Mar. 2015.
COMPARISON OF DIFFERENT MODELS WITH FIVE PMUS LIMIT [9] F. Milano and Á. Ortega, “A method for evaluating frequency regulation
in an electrical grid – Part I: Theory,” IEEE Trans. Power Syst., vol. 36,
no. 1, pp. 183–193, Jan. 2021.
[10] Nordic TSOs, “Challenges and opportunities for the nordic power system,”
Aug. 2016. [Online]. Available: https://www.fingrid.fi/globalassets/
dokumentit/fi/yhtio/tki-toiminta/report-challenges-and-opportunities-
for-the-nordic-power-system.pdf
[11] B. K. Poolla, S. Bolognani, and F. Dörfler, “Optimal placement of virtual
inertia in power grids,” IEEE Trans. Autom. Control, vol. 62, no. 12,
pp. 6209–6220, Dec. 2017.
[12] R. B. Sharma and G. M. Dhole, “Wide area measurement technology in
power systems,” Procedia Technol., vol. 25, pp. 718–725, 2016.
models, indicating that the overall observability of the WAMS [13] A. Schmitt and B. Lee, “Steady-state inertia estimation using a neural
is improved through the proposed ZGIB-OPP algorithm. network approach with modal information,” in Proc. IEEE Power Energy
Soc. Gen. Meeting, 2017, pp. 1–5.
[14] A. Poudyal, R. Fourney, R. Tonkoski, T. M. Hansen, U. Tamrakar, and
VI. CONCLUSION R. D. Trevizan, “Convolutional neural network-based inertia estimation
using local frequency measurements,” in Proc. 52nd North Amer. Power
Neural networks have been applied for inertia estimation as Symp., 2021, pp. 1–6.
extensive amounts of data can be obtained through power system [15] F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini,
“The graph neural network model,” IEEE Trans. Neural Netw., vol. 20,
digital equipment and advanced measuring infrastructures such no. 1, pp. 61–80, Jan. 2009.
as PMUs. In this paper, LRCN and GCN based learning algo- [16] T. Pham and X. Li, “Reduced optimal power flow using graph
rithms are proposed to estimate system inertia constant. System neural network,” in Proc. 54th North Amer. Power Symp., 2022,
pp. 1–6.
wide ambient measurements based on WAMS are used as candi- [17] U. Markovic, Z. Chu, P. Aristidou, and G. Hug, “LQR-based adaptive
date features for model training, and a wrapper feature selection virtual synchronous machine for power systems with high inverter pen-
is also used to optimize the feature combination. Considering the etration,” IEEE Trans. Sustain. Energy, vol. 10, no. 3, pp. 1501–1512,
Jul. 2019.
limitation on PMU settings, an ZGIB-OPP method is proposed [18] A. Adrees, J. V. Milanović, and P. Mancarella, “Effect of iner-
to maximize the observability of the WAMS given limited PMU tia heterogeneity on frequency dynamics of low-inertia power sys-
resources. Results indicate that the proposed LRCN and GCN tems,” IET Gener., Transmiss. Distrib., vol. 13, no. 14, pp. 2951–2958,
Jul. 2019.
models have better performances than the benchmark DNN and [19] E. Carlini, F. Del Pizzo, G. Giannuzzi, D. Lauria, F. Mottola, and C.
CNN models. The proposed LRCN model and GCN model Pisani, “Online analysis and prediction of the inertia in power systems
also show high robustness under conditions with higher noises. with renewable power generation based on a minimum variance harmonic
finite impulse response filter,” Int. J. Elect. Power Energy Syst., vol. 131,
The proposed ZGIB-OPP method has been proved to be capa- 2021, Art. no. 107042.
ble of improving the performance of all implemented models. [20] A. Fernández-Guillamón, E. Gómez-Lázaro, E. Muljadi, and Á. Molina-
Considering that the IEEE 24-bus system model used in this García, “Power systems with high renewable energy sources: A review
of inertia and frequency control strategies over time,” Renewable Sustain.
research has a mix generation of both synchronous generators Energy Rev., vol. 115, 2019, Art. no. 109369.
and inverter-based resources, the proposed approach can also be [21] J. Zhang and H. Xu, “Online identification of power system equivalent in-
applied to estimate inertia constant in realistic conditions. ertia constant,” IEEE Trans. Ind. Electron., vol. 64, no. 10, pp. 8098–8107,
Oct. 2017.
[22] V. Sridharan, M. Tuo, and X. Li, “Wholesale electricity price forecast-
REFERENCES ing using integrated long-term recurrent convolutional network model,”
Energies, vol. 15, no. 20, 2022, Art. no. 7606.
[1] M. Tuo and X. Li, “Long-term recurrent convolutional network-based [23] S. Albeiwi and A. Mahmood, “A framework for designing the architectures
inertia estimation using ambient measurements,” in Proc. IEEE Ind. Appl. of deep convolutional neural networks,” Entropy, vol. 19, no. 6, May 2017,
Soc. Annu. Meeting, 2022, pp. 1–6. Art. no. 242.

Authorized licensed use limited to: Indian Institute of Technology Patna. Downloaded on June 05,2024 at 04:30:24 UTC from IEEE Xplore. Restrictions apply.
TUO AND LI: MACHINE LEARNING ASSISTED INERTIA ESTIMATION USING AMBIENT MEASUREMENTS 4903

[24] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Mingjian Tuo (Student Member, IEEE) received the B.E. degree in electrical
Computation, vol. 9, no. 8, pp. 1735–1780, Nov. 1997. engineering and automation from the Huazhong University of Science and
[25] J. Donahue et al., “Long-term recurrent convolutional networks for visual Technology, Wuhan, China, and the M.Sc. degree in computer application
recognition and description,” IEEE Trans. Pattern Anal. Mach. Intell., technology from the National Engineering Research Center for E-Learning,
vol. 39, no. 4, pp. 677–691, Apr. 2017. Central China Normal University, Wuhan, in 2015 and 2018, respectively. He
[26] A. Chugh, “MAE, MSE, RMSE, coefficient of determination, is currently working toward the Ph.D. degree in electrical engineering with the
adjusted R squared - which metric is better?”, Dec. 2020. [Online]. University of Houston, Houston, TX, USA. His research interests include power
Available: https://medium.com/analytics-vidhya/mae-mse-rmse- systems operations and optimization, power system stability and control, deep
coefficient-of-determination-adjusted-r-squared-which-metric-is- learning for data analysis in power system, and system dynamic modeling.
better-cd0326a5697e
[27] P. E. Pope, S. Kolouri, M. Rostami, C. E. Martin, and H. Hoffmann,
“Explainability methods for graph convolutional neural networks,” in
Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2019, pp. 10772–10781.
[28] F. Ebrahimzadeh, M. Adeen, and F. Milano, “On the impact of topology on
power system transient and frequency stability,” in Proc. IEEE Int. Conf.
Environ. Elect. Eng. IEEE Ind. Commercial Power Syst. Europe, 2019,
pp. 1–5.
[29] H. J. T.Weckesser, M. Glavic, and J. Østergaard, “An improved on-line Xingpeng Li (Senior Member, IEEE) received the B.S. degree in electrical
contingency screening for power system transient stability assessment,” engineering from Shandong University, Jinan, China, in 2010, the M.S. degree
Electric Power Compon. Syst., vol. 45, no. 8, pp. 852–863, May 2017. in electrical engineering from Zhejiang University, Hangzhou, China, in 2013,
[30] T. Johnson and T. Moger, “A critical review of methods for optimal the second M.S. degree in industrial engineering and the Ph.D. degree in
placement of phasor measurement units,” Int. Trans. Elect. Energy Syst., electrical engineering from Arizona State University, Tempe, AZ, USA, in
vol. 31, no. 3, 2021, Art. no. e12698. 2016 and 2017, respectively, and the third M.S. degree in computer science
[31] N. H. Abbasy and H. M. Ismail, “A unified approach for the optimal PMU from the Georgia Institute of Technology, Atlanta, GA, USA, in 2023. He is
location for power system state estimation,” IEEE Trans. Power Syst., currently an Assistant Professor with the Department of Electrical and Computer
vol. 24, no. 2, pp. 806–813, May 2009. Engineering, University of Houston, Houston, TX, USA. He previously worked
[32] M. Tuo, A. Venkatesh Ramesh, and X. Li, “Benefits and cyber- for ISO New England, Holyoke, MA, USA, and PJM Interconnection, Audubon,
vulnerability of demand response system in real-time grid operations,” in PA, USA. Before joining the University of Houston, he was a Senior Application
Proc. IEEE Int. Conf. Commun., Control, Comput. Technol. Smart Grids, Engineer for ABB (now is Hitachi Energy), San Jose, CA, USA. His research
2020, pp. 1–6. interests include power system operations, control and planning, applications
[33] M. Brown, M. Biswal, S. Brahma, S. J. Ranade, and H. Cao, “Character- of optimization and machine learning in power systems, energy markets, and
izing and quantifying noise in PMU data,” in Proc. IEEE Power Energy microgrids.
Soc. Gen. Meeting, 2016, pp. 1–5.

Authorized licensed use limited to: Indian Institute of Technology Patna. Downloaded on June 05,2024 at 04:30:24 UTC from IEEE Xplore. Restrictions apply.