Received: 11 February 2023 Revised: 3 August 2023 Accepted: 28 August 2023 IET Generation, Transmission & Distribution

DOI: 10.1049/gtd2.12976

ORIGINAL RESEARCH

A robust fusion bus frequency estimation method to improve

frequency oscillation damping in power systems

Ali Farahani1 Amir H. Abolmasoumi1,2 Lamine Mili3 Mohammad Bayat1,2

1
Department of Electrical Engineering Arak Abstract
University, Arak, Iran
This paper proposes a new robust method for accurately and reliably estimating remote
2
Research Institute of Renewable Energy, Arak bus frequencies in a power system. Two different measurement sources for the remote bus
University, Arak, Iran
frequency are considered, that is, the ideal Frequency Divider (FD) and the Synchronous
3
Department of Electrical and Computer Reference Frame Phase Locked Loops (SRF-PLLs). Each measurement signal encoun-
Engineering Virginia Polytechnic Institute and State
University, Falls Church, VA, USA ters different uncertainties and data quality issues. In this paper, both data sources are
employed and fused together to better estimate the remote bus frequencies. To this end,
Correspondence the model structure of the power system is selected and an Unscented Kalman Filter (UKF)
Amir H. Abolmasoumi, Research Institute of is utilized together with a fusion covariance intersection method to enhance the accuracy
Renewable Energy, Arak University, Arak,
38156-8-8349, Iran.
of the estimated bus frequency. Since the fusion estimation method fails in case of qual-
Email: [email protected], ity issues on both channels, a robust Generalized Maximum-likelihood UKF (GM-UKF)
[email protected] using a novel outlier detection criteria is developed. The impact of the resulting robust
fusion filter on the estimation of the remote bus frequencies and on the performance of
WAPSS, which makes use of the estimated frequencies as the feedback signal, is examined
via simulations. The results demonstrate the excellent performance and reliability of the
proposed method in dealing with noise filtering and outlier suppression while ensuring a
high statistical efficiency.

1 INTRODUCTION measuring frequencies in one area and acting on generators in

another area [4, 7]. In other words, Wide-Area Power System
The damping of frequency oscillations is of essential impor- Stabilizers (WAPSSs) that take the frequency of a bus in differ-
tance for guaranteeing the stability of wide-area power systems ent area as the measurement feedback are the right option [4,
with inter-connected synchronous generators. Dominant low 8–10].
frequency oscillations include local modes in the range of 1–2 The essential step in all design methods of a WAPSS is to
Hz and, more importantly, inter-area modes in the range of 0.1– determine the optimal bus on which the feedback frequency
1 Hz [1, 2]. If not damped efficiently, these modes may result in should be measured and the generator at another area where
line and generator outages, power system splitting into islands, the WAPSS should be placed. A large number of studies in the
and cascading failures leading to large-scale blackouts [3]. The literature deal with the placement of the meters and WAPSSs
common solution for the stabilization of the power systems is to to achieve the highest system observability and controllability
deploy local Power System Stabilizers (PSSs) [2, 4] and Flexible based on a sensitivity analysis [11], a geometric-based approach
AC Transmission System (FACTS) devices [5, 6]. As for the local [12], and wide-area damping controllers [9]. There are also sev-
PSSs, they use the local frequency measurements as feedback eral WAPSS design schemes proposed in the literature; see for
signals sent to the voltage automatic regulators to stabilize the example [7, 8] and [13]. The control scheme being considered
inter-area electromechanical oscillations [2]. However, the local consists of remote metered values taken from Wide-Area Mea-
stabilizers are shown not the best option to achieve this goal. surement Systems (WAMSs), using Phasor Measurement Units
Indeed, inter-area oscillations, which are observable at buses (PMUs), and transmitted to a WAPSS [14]. However, providing
belonging to different areas, should be damped by stabilizers a high quality frequency measurement from a remote bus to the

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is
properly cited.
© 2023 The Authors. IET Generation, Transmission & Distribution published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology.

IET Gener. Transm. Distrib. 2023;1–16. wileyonlinelibrary.com/iet-gtd 1

2 FARAHANI ET AL.

WAPSS is problematic because of the issues regarding the data gross errors, communication losses, and injected false values.
quality such as uncertainties, noises and deviant data, termed As for the measurement errors, they are shown to have some
outliers. The latter may be due to imperfect measuring meth- influence on the frequency regulators, which may be signif-
ods or devices, sampling and processing issues, communication icant for large errors [27]. Huang et al . [28, 29] categorize
problems or even cyber-attacks, to name a few [15]. the PMU quality issues such as time synchronization accuracy,
Regarding the frequency estimation, it is a sub-topic of the data losses, and data latency and discuss each one separately.
overall power system state estimation problem. There are a For more on the PMU measurement quality issues, see [30,
number of power system dynamic state estimation methods 31] and the references therein for further details. Regarding
making use of the PMU measurements proposed in the liter- the PLLs, they may be unable to guarantee low noise and
ature; see for example [16, 17] and the references therein. In high-speed response at the same time while suffering from
[18], a switched dynamic state estimation is utilized to address numerical integration errors [4, 32]. Furthermore, their per-
the problem of denial-of-service cyber attacks in power systems. formance may deteriorate as a result of a step change in the
Two subsystems are employed and the estimation outcomes speed and the presence of harmonics and unbalanced power
are combined. Huang et al. [19] propose a decentralized UKF systems [23, 32].
method based on the analog voltage and current measure- It is known that for large disturbances, the FD calculations
ments installed at transformer buses. Here, the magnitude and provide more reliable data on remote bus frequencies compared
the frequency of the voltages and currents are obtained based to the SRF-PLLs. However, there are many scenarios where the
on statistical signal processing techniques without the need SRF-PLLs measurement data contain usable data that must not
of PMUs. Another method to estimate the frequency from be thrown away. Moreover, the FD needs the power system
measurements is presented in [20] using an adaptive detection parameter values, which may not be available with a high pre-
methodology, which is shown to be robust to the changes in the cision. As a result, there are cases where the FD calculations are
noise statistics and the step changes in the voltage and current also unreliable to some extent. Furthermore, as discussed, some
amplitudes and phases. In [21], a frequency estimator is designed data sets transmitted over both the FD and the SRF-PLLs chan-
based on the measurements by the Hilbert-transform PLLs nels may suffer from quality issues. These weaknesses motivate
together with a complex mean squared estimation method. us to employ both measurement sources at the same time to
Instead of using bus frequencies measured from locally installed improve bus frequency estimation. In this paper, we investigate
devices such as instrument transformers or PLLs, Milano and the fusion of two measurement feedback signals taken from the
Ortega [22] propose the FD formula to obtain the remote bus FD calculations and the SRF-PLLs to obtain more accurate and
frequencies given the angular velocity of the generators since reliable estimates from of the frequency at the remote bus and,
the PMUs are usually installed at the generator buses. How- as a result, to enhance the overall efficiency of the WAPSS. The
ever, the FD relies on the PMU-based rotor speed signals, which current paper may be seen as the extension of the results in [23]
may be noisy and biased. Also the FD makes use of the sus- into the case where SRF-PLLs measurements are available. To
ceptance matrix of the power system, whose parameters may perform the fusion, we first make use of a covariance intersec-
be subjected to uncertainties. Zhao et al. [23] advocate the use tion method in combination with the UKF method, which we
of a GM-UKF aimed at robustly estimating the rotor speed, call fusion UKF. The latter takes the measurements from the
and thereby the remote bus frequency via the FD, while achiev- FD and the SRF-PLLs and calculates the dynamically weighted
ing lower communication load by transmitting only the rotor sum of these measurements to construct the feedback signal for
angle and speed signals to the FD calculation center. How- the WAPSS. Since two measurement signals come from differ-
ever, the locally estimated frequencies have to be transmitted ent sources, they may suffer from different data quality issues.
through the communication channel to the FD center where the Consequently, the fusion filter is used to preserve the quality of
load bus frequencies are calculated. Obviously, this may intro- the output feedback signal in presence of uncertainties, noises,
duce new errors to the transmitted signals, including missing and outliers in one of the channels. The fusion will be shown to
data, various noises, and gross errors as a result of cyber-attacks be the most effective method when one channel is contaminated
or channel interference. On the other hand, different types of with large errors.
PLLs have been recently utilized to provide the bus frequencies To robustify the fusion filter against outliers or thick-tailed
[24]. As one of the best performing variants, the synchronous non-Gaussian noises, which may occur simultaneously at
reference frame PLLs (SRF-PLLs) [25] are well-functioning both measurements channels, we suggest to utilize the robust
high-bandwidth algorithms that are capable of providing a fast GM-UKF recently presented in [33–37] as a powerful power
and accurate phase-angle estimation. In [4], the dynamic per- system dynamic state estimator. The GM-UKF makes use
formance of the WAPSSs using either an ideal FD-calculated of the concept of projection statistics to weight the data
frequency or SRF-PLLs-based measurements are compared and based on their inlier-outlier behaviour. As an extension to the
the effect of the time-delays on the performance of the WAPSS fusion estimation, the outlier detection step in the GM-UKF
are investigated. is modified. Specifically, the organization of the data vector
There are several quality issues of the values provided by local set for which the projection statistics and the corresponding
bus frequency meters such as SRF-PLLs or the FD-calculated outlier weights that are calculated are modified to achieve a
ones achieved from PMUs installed at generator buses. How- fusion GM-UKF. It is verified that in addition to the general
ever, as indicated in [23, 26], PMUs may be corrupted by advantages of the fusion method, it also enhances the outlier
FARAHANI ET AL. 3

detection capability of the robust estimation process presented

in [23] and accordingly improves the performance of the robust
frequency estimation by the GM-UKF. In other words, the
fusion GM-UKF proposed in this paper is more robust than
the GM-UKF presented in [23]. We also show that especial
attention should be paid to choosing the appropriate dynamic FIGURE 1 The block diagram of a power system stabilizer.
model of the power system in terms of its relationship with
system observability and its communication load. The latter
is determined by the number of metered signals (for example 2.1 Wide area power system stabilizer
electrical power, rotor speed etc.) that are sent through the com-
munication channels to the estimation center. This work is the The block diagram of the PSS is depicted in Figure 1. As
generalization of the initial ideas in [38, 39] by the authors of the observed, the PSS takes Vsi j = Δ𝜔 as the feedback and applies
present paper. the voltage Vso as an added signal to the Automatic Voltage Reg-
The proposed fusion UKF and fusion GM-UKF methods ulator (AVR) reference input to make change on the generator
are simulated on two cases. The two-area four-machine power exciter on which the WAPSS is installed. The PSS is composed
system [4] and the New England 39-bus 10-machine power of a washout filter followed by the lead-lag blocks. In Figure 1,
system NEW England Case System [9]. The results show the GWAPSS represents the WAPSS gain, TW is the time constant of
improvement of the bus frequency estimation, which leads the washout filter, and Tn1 , Tn2 , Td 1 , Td 2 are the four time con-
to better frequency oscillation damping by the PSSs in com- stants that belong to each of four stabilizing blocks. The results
parison with the case where the conventional FD or PLLs from [40–42] show that the inter-area oscillations are better
measurements are individually used to construct the feed- observable when using the difference between the rotor speeds
back signals. It is also demonstrated that the designed fusion of some machines and the frequencies on some load buses in
GM-UKF is robust to gross errors among the measure- other areas of the power system. Assuming that the power sys-
ment data, which enhances the performance of the proposed tem stabilizer is installed at Bus j , we get Vsi = Δ𝜔i j = 𝜔Bi −
method. 𝜔G j . Here, Vsi = Δ𝜔i j is the feedback error signal that is fed
The next sections of the paper are organized in the follow- to the PSS controller. Note that the selection of Buses i and j
ing manner: Section 2 presents preliminaries to power system are based on some controllability/observability analysis based
stabilization and WAPSS structure, and explain how to obtain on, for example, the geometric approach [4]. Calculating the
the frequency at the target point via the SRF-PLLs and the FD feedback signal Vsi calls for two elements to be known, namely
methods. Section 3 discusses the architecture of the proposed 𝜔Bi and 𝜔G j . Therefore, the generator rotor speed and the bus
state estimation method. Section 4 proposes the GM-UKF frequency are essential components in the construction of the
to robustify the fusion filter. Section 5 provides some simu- feedback signal. Accordingly, improving the estimation of the
lations carried out on the New England 10-machine system bus frequency contributes to improve the performance of the
and analyzes the performance of the proposed state estima- frequency oscillation control. As mentioned before, the output
tion method in various scenarios. Finally Section 6 concludes of the PSS block Vso is an added control voltage to the refer-
the paper. ence voltage of the local AVR,(Vre f 0 ), and enables the controller
to damp the electromechanical oscillations through the excita-
tion controller. Finally, the voltage reference is then calculated
2 PRELIMINARIES TO WIDE-AREA as Vre f = Vre f 0 + Vso.
POWER SYSTEM STABILIZER AND
FREQUENCY MEASUREMENTS
2.2 Obtaining the value of the bus frequency
The algebraic differential state-space equations that governs a
power system are expressed as Two methods aimed at obtaining the value of the frequency on
the remote buses are considered. The first method consists in
( )
ẋ = ̃f x, y, u , measuring the frequency by using the SRF-PLLs as a frequency
( ) (1) measurement device installed at a local bus. The PLLs devices
g x, y, u = 0, are primary aimed at synchronizing the power electronic con-
verters to the electrical grid. A byproduct is the estimation of the
where ̃f and g are the vector functions of the power system deviation of the frequency from its base value [32]. The input is
state vector x, the dependent variables contained in y, and the the three-phase bus voltage. The voltages are transformed to
control input u, which here is the voltage applied to the exciter the d − q reference frame and the q component is controlled to
of the target generator as the wide area power system stabi- track zero value, which guarantees that the converter is complete
lizer. In state estimation, some signals such as the electrical synchrony with the AC power system frequency at the connec-
or the mechanical power are assumed to be known, result- tion point and utilizes vd (t ) as the phase reference. Note that
ing in a reduced dimension of x and in a larger number of if the locked angle, 𝜃, is the same as the angle of the system
input variables. voltage vabc (t ), then the Park transformation requires that vq (t )
4 FARAHANI ET AL.

Another method for calculating the frequency of the local

target bus makes use of the FD formula proposed by [22].
According to FD method, the bus frequencies can be calculated
from the known generator rotor speeds as follows:

Δ𝝎B = DΔ𝝎G ,

D = −(BBB + BB0 )−1 BBG , (5)

Δ𝜔Bi,FD = di Δ𝝎G .

Here the vectors Δ𝝎B and Δ𝝎G are the deviations of the bus
frequencies and generator speeds from their base values; the
matrices BBB , BB0 , and BBG are the susceptance matrices of the
FIGURE 2 The structure of the SRF-PLL: (a) the overall idea of
power transmission system, and di is the i-th row of the matrix
SRF-PLL, and (b) the frequency estimation by SRF-PLL.
D. The FD method has been investigated from the viewpoint
of numerical stability and complexity in [23] and [43].

is equal to zero and vd (t ) is constant. Figure 2a represents the

idea of SRF-PLLs, where Vabc denotes the three-phase voltages 3 DYNAMIC STATE ESTIMATION
at the bus of connection. Also Vd and Vq denote the dq volt- USING THE FUSION UKF METHOD
ages. Accordingly, [Tdq ] is the transfer matrix to the dq reference
frame and 𝜃Bi is the bus angle. As shown in Figure 2, a Pro- As discussed in Section 2, two methods can be used to cal-
portional Integral (PI) controller and an integrator are used as culate the value of the selected load bus frequency, that is,
the Loop Frequency (LF) and the Voltage Controlled Oscillator the FD calculations and the SRF-PLLs measurements. Each of
(VCO), respectively. The frequency of the local bus is obtained these methods are exposed to various types of uncertainties.
via 𝜔Bi = 𝜔Base + Δ𝜔Bi , where 𝜔Base denotes the base frequency The SRF-PLLs measurements may be contaminated by noises
of the power system. Since the measurement of the signal and and outliers. Besides, the transmission of such measurement
the application of the Park transformation is not instantaneous, through the communication channel may add even more uncer-
it is modeled by a time delay. Moreover, as seen in Figure 2, tainties. On the other hand, the FD method also needs accurate
it is possible to extract an estimation of the frequency deviation values of the generator speeds at its inputs. However, these val-
from the SRF-PLLs. Here, it is assumed that the frequency mea- ues should be first estimated by local decentralized estimators
surement by the SRF-PLLs is precise. However, there are some using PMUs. Now, these PMUs suffer from additive noise and
reported drawbacks such as the tradeoff between the noise and biases. To address this problem, Zhao et al. [23] propose to use
the response speed, the numerical integration errors, and the a robust FD based on a robust GM-UKF [33–36]. While this
vulnerability to step changes in the speed [4, 32]. There may method locally deals with observation outliers among the mea-
also additional vulnerabilities to harmonics and power system surements as well as structural and innovation outliers in the
unbalance [23, 32]. By defining assumed dynamic model to avoid the deviation of the frequency
estimation, it is vulnerable to gross-errors when the signals are
̃ x2 = transmitted to the FD center.
x1 = 𝜃, 𝜀𝜃 dt , (2)
∫

the SRF-PLLs will be described by the following dynamic state- 3.1 Structure of the dynamic state
space equations: estimation for WAPSS

ẋ1 = −kP x1 + ki x2 + k p u, In this paper we construct an estimation structure based on

(3) a data fusion filter to improve the feedback signal using both
ẋ2 = −x1 + u. the FD and the SRF-PLL methods. Note that this method is
robust, especially when only one of the measurements suffers
The frequency value is then calculated as from communication failure or strong noise at a certain time.
The overall structure of the wide area control of the power sys-
yPLL = 𝜔Bi,PLL = −kP x1 + ki x2 + k p u + 𝜔Base . (4) tem using the data fusion is represented in Figure 3. As seen, the
Local Frequency Estimators (LFEs) estimate the rotor speed
The measurements are sent to the WAPSS or a processing at each generator bus. Afterwards, each local estimator sends
center through a communication network, which may impose the estimated rotor speed of the corresponding generator to the
further quality issues such as different noises, missing data, and FD calculation unit to achieve the frequency of the target Bus
gross errors. i. Such calculated frequency may be contaminated by various
FARAHANI ET AL. 5

nonlinear dynamical model expressed as

xk+1 = f (xk , uk ),
(6)
yk = h(xk ) + vk = H k xk + vk .

First, it is assumed that the measurement noises vk on SRF-

PLL measurements and also on signals received from the
generator units (say the local unit rotor speeds) include only
Gaussian noises as a result of being transmitted through the
communication network. Later in Sections 3.4 and 3.5, this
assumption is relaxed and it is assumed that there are outliers
among measurements. It can be interpreted as non-Gaussian
noise contamination. For example, the measurements could be
contaminated with a thick-tailed noise such as Cauchy. The GM-
UKF addresses this problem. The exact statistical parameters of
outliers are not necessary to be known.
For the given discrete-time state space the fusion UKF calcu-
lates the estimates of the states in two steps which are referred
to as the prediction and update steps. The process is carried out
in a recursive manner. The UKF processes a set of 2n sample
FIGURE 3 The overall structure of wide area control via WAPSS using
the fusion filter to mix the frequency data from FD and SRF-PLL.
points, called sigma points. In the beginning of each step, sigma
points are obtained as
(√ )
X ik−1|k−1 = x̂ k−1|k−1 ± nPk−1|k−1 , i = 1, … , n , (7)
noises and gross errors as a result of processing and transmis- i
sion noises or missing data. On the other hand, the frequency
where Pk−1|k−1 is the error covariance of the previous step and
measurements in Bus i can be performed by the SRF-PLLs. In
√
general, the uncertainties on frequency measurements received nPk−|k−1 is the square-root which may come from Cholesky
at the feedback center from the SRF-PLLs are different from decomposition. Here, n is the order of the system. Sigma
those of the FD-calculated frequencies. The data fusion filter points propagate through the evolution equation to produce the
combines two data signals by utilizing filtering methods depend- prediction as
ing on the considered power system dynamic model. If a small
signal dynamic model is considered, the fusion Kalman filter is ( )
X ik|k−1 = f X ik−1|k−1 . (8)
suggested. Now, if a nonlinear model is considered, the UKF
may be used as the basis estimator. However, when there are
observation, innovation, or structural outliers either among the The predicted state vector and the predicted covariance matrix
received measurement data or within the assumed power system of estimations are given by
dynamic model, the GM-UKF should be applied.
The selection of the dynamic model for the fusion filter is of ( )
∑
2n
critical importance. Utilizing a highly detailed dynamic model x̂ k|k−1 = 𝛾i X ik|k−1 = f X ik−1|k−1 , uk ,
with too many states weakens the system observability, resulting i=1
in a poor performance of the frequency estimation. Alternately,
employing simple models with limited number of states while ∑
2n ( )( )T
considering mechanical or electrical power on each generator Pk|k−1 = 𝛾i X ik|k−1 − x̂ k|k−1 X ik|k−1 − x̂ k|k−1 + Qk ,
as known inputs will enhance the system observability and, as i=1
a result, will improve the performance of the frequency esti-
mation. The model selection will be discussed more in the Y k|k−1 = H k X k|k−1 ,
next sections.
∑
2n
ŷ k|k−1 = 𝛾i Y i,k|k−1 ,
3.2 Fusion unscented Kalman filter i=1
(9)
Data fusion techniques merge the measurement data from two
or more sensors to obtain more accurate outcomes [44]. there where 𝛼i = 1∕2n is the equal weight attributed to for the sigma
are plenty of research works on measurement fusion methods points. After the prediction completed, the estimated states and
as applied to dynamic systems [45]. Consider the power system covariance are updated step then is carried out to calculated the
6 FARAHANI ET AL.

as follows: 3.3 Power system dynamical model,

observability and communication load
∑
2n ( )( )T
Py = 𝛾i Y ik|k−1 − ŷ k|k−1 Y ik|k−1 − ŷ k|k−1 , Since the fusion state estimation technique is a wide-area estima-
i=1
tion algorithm, communication load should be considered in the
filter design. When the number of signals transferred through
∑
2n ( )( )T the communication channel increases, there may be some
Pxy = 𝛾i X ik|k−1 − x̂ k|k−1 Y i,k|k−1 − ŷ k|k−1 , decrease of the reliability of the frequency estimation as a result
i=1
of the network delays, signal corruptions, and cyber-attacks.
(10) Therefore, estimation methods involving lower required bur-
Kk = Pxy P−1
y , den on the communication system are of interest. The UKF as
applied to the FD calculations can utilize simple models of the
( ) power system, provided that the filter has access to some known
x̂ k|k = x̂ k|k−1 + K k yk − ŷ k|k−1 , signals. Let us consider the classical generator model given by

𝛿̇ = 𝜔 − 𝜔0 ,
Pk|k = Pk|k−1 − K k Py K k T , (14)
(2H ∕𝜔0 )𝜔̇ = Pm − Pe − D p (𝜔 − 𝜔0 ),
where in (9) and (10) yk = [ yPLL,k yG 1 ,k … yGn ,k ]T is vec-
in which 𝛿 denotes the rotor angle of the corresponding gener-
tor of measurements, here the PLL frequency measure- ator , 𝜔 denotes the angular rotor speed, H denotes the inertia
ment yPLL,k and measured values of the generator rotor constant, D p denotes the friction/damping ratio, 𝜔0 denotes the
angular velocity values [yG 1 ,k … yGn ,k ]T . The matrix H k = base rotor speed, and Pm , Pe denote mechanical and active elec-
diag(H PLL, k , I n ) determines the dependency of the mea- trical powers, respectively. Here, the mechanical and electrical
surements on the states. Moreover, matrices Qk and Rk power values of each generator at every time step should be sent
are the covariance of the process and observation noises, to the fusion center by each unit. Note that using (14) provides
respectively. The FD formula utilizes the estimated gen- the acceptable level of observability to estimate the frequency
erator rotor speeds to calculate the i − th bus frequency of the target load bus. Instead of using (14), which requires
as dispatching the mechanical and electrical active power values
through the communication network, one may consider a set
[ ]T
ŷ FD,k = 𝜔̂ Bi,FD,k = H FD ŷ G,k = H FD ŷG 1,k … ŷGn,k , (11) of more detailed dynamical models such as the two-axis model
or the flux decay model, which account for the dynamics of the
where H FD is the same as d i given by (5). Let the estimated fre- automatic voltage regulator and/or of the turbine governor.
Figure 4 shows the structure of the UKF-FD with the appli-
quency of the i − th bus be calculated as ỹ k = [̂yPLL,k ŷFD,k ]T . cation of three models. Figure 4a depicts the usage of the simple
Thus, the error covariance reduces to swing equation for each generator. In this case the electrical and
[ ] mechanical powers are assumed as known inputs by the filter.
( ) 𝛼k 0 Since the UKF-FD estimator is located in the fusion center,
̃ k x̂ k|k =
S k = cov ỹ k − H , (12)
0 𝛽k the instantaneous values of the electrical and mechanical power
will be transmitted to the center. The observability here simply
where the error covariance matrix S k is diagonal because the means whether the frequency estimation in the center is suc-
cessful or not. Since the model is simple with only two state
measurement errors of FD and PLL are assumed to be indepen- variables and the inputs, which are known by the estimator, are
dent and H̃ k = diag(H PLL, , H FD ). Finally, the fusion signal is
k transmitted to the center, the system observability is high and
obtained as the weighted average of the estimates by the FD and the bus frequency estimation is successful. Note that it is com-
PLL as: mon in the literature to consider the mechanical power to be
𝛼k 𝛽k fixed and known during the transient process [16, 23, 36]. If
𝜔̂ fusion,k = ŷ + ŷ . (13) Pmi , i = 1, .., N for all the generators are fixed and known by
𝛼k + 𝛽k FD,k 𝛼k + 𝛽k PLL,k
the filter, then the transmission of Pe is sufficient and model1
is the best model for the frequency estimation. In that case,
Equation (13) implies that which means that the estimate with
the model selection can be skipped. There are cases where Pmi
lower error is more weighted in calculation of the fusion esti- is time-varying or not a priori known by the filter. To address
mates of the bus frequency. This method was first introduced in this problem, one option is to include the governor dynamics in
[46] and has been applied to a number of state estimation prob- the model used by the filter rather than to directly send the Pm
lems (see [47] and [48]). The next section discusses the tradeoff value to the center. This has been shown in Figure 4b. In that
between the observability of the model states and the necessary case, the governor state variables can also be estimated, which
communication load. is an additional benefit. However, as the number of the state
FARAHANI ET AL. 7

FIGURE 4 Three power system dynamical models and the required input signals for UKF-FD: (a) simple swing equation based on FD; (b) model with
inclusion of the governor states; (c) complete model with inclusion of the electrical and the mechanical subsystems.

TABLE 1 Comparison of the dynamic order, observability and the of its statistical assumptions and its accuracy. It turns out that
communication load for three generator models used for frequency estimation the variance weighting may not be sufficient when there are
from FD measurements.
severe deviations from the statistical assumptions. Indeed, as
Model 1 Model 2 Model 3 one of the most critical data quality issues, the measurements
Filter inputs Pm , Pe Pe E fd , Id , Iq or calculations sent from the FD or the SRF-PLL may be
contaminated by outliers. Outliers are deviant data point among
States 𝜔, 𝛿 𝜔, 𝛿, xtg 𝜔,𝛿,𝜓kq ,Ed′ ,𝜓kd ,Eq′ ,xtg
measurements, which may occur in different stages of the
Observability High Medium Low
calculation, transmission, and processing of the data. This is
Communication load Medium Low Medium likely to occur when there is an overall lack of bandwidth, or
partial packet dropouts, or a severe cyber-attack, or a random
concurrence of outliers among the measurements.
To deal with outliers in our fusion state estimator, the sam-
variables increases and the known values and measurements (Pe ples will be first processed via an outlier detection technique.
and 𝜔G ) remain the same, the estimation procedure is of lower There are several techniques presented in the literature to detect
accuracy since the system observability decreases. As a rule of and deal with outliers in power system state estimation. Among
thumb, the higher the number of the model state variables are them, Projection Statistics (PS) have been shown to be robust
the lower the system observability will be if the number of the and fast to calculate (See [49] and [50]). Using PS, different
measured/observed signals is kept constant. robust Kalman filters have been proposed in the literature (See
It is also possible to adopt a more complex generator model [16-19] and [23]). For a sample set {z1 , … , zm }, PS [49] are
for the filter by including both the mechanical and electrical calculated as
sub-systems as shown in Figure 4c. As observed, the num- | T ( )|
|z v − med j zT v |
ber of the state variables of a generator is seven and the | i j |
PSi = max‖v‖=1 | |
variables, Id ,Iq and E fd , of each generator plus the estimated [ ( )] , (15)
rotor speed of each generator should be transmitted to the fil- 1.4826medk zk v − med j zTj v
T

tering center. This results in a low system observability and,

thereby, to an unsuccessful state estimation. A comparison of for i = 1, m. PS make use of the sample median as the estimator
the required communication load and the observability level of location and the Median-Absolute-Deviation (MAD) as the
using each model can be seen in Table 1, where 𝜔, 𝛿, 𝜓kq , estimator of scale, which are calculated in the direction originat-
Ed′ , 𝜓kd , Eq′ are states of generator and xtg is turbine governor ing from the coordinate-wise medians and passing through each
state vector. data points. It is shown that the squared projection statistics,
PS 2 , follow a 𝜒n2 distribution when the data points follow the
Gaussian distribution [50]. In this paper, we propose to calculate
3.4 Dealing with outliers the PS of the data points Z̃ k defined as

Taking advantage of the measurement redundancy, the UKF Z̃ k = [ Zk−1 Zk ], (16)

fusion filter is able to handle noises and uncertainties in the
communication network in an effective way. This is because where
it makes use of the SRF-PLL measurement and the FD cal- [ ]T
culations as two independent data sets and combines them Zk = yPLL,k − ŷPLL,k yG ,k − ŷG ,k . (17)
to provide accurate frequency estimation at the i-th bus.
However, it makes use the fusion weighting based on the It should be noted that the selection of the data points
estimated covariance matrix and thereby, relies on the validity is paramount for outlier detection and weighting. Here, the
8 FARAHANI ET AL.

fusion framework provides the information on each genera- function given by

tor rotor speed estimation error along with the PLL errors.
{
By integrating all the innovation terms of the generator rotor 0.5ri 2 −𝜆 < ri < 𝜆,
speed estimation and of the PLL bus frequency estimation, 𝜌 (ri ) = (22)
the fusion technique constructs the Z matrix. More elements 𝜆|ri | − o.5𝜆2 ri < −𝜆 or ri > 𝜆,
in Z lead to larger redundancy, resulting in a higher reliabil-
ity of the PS. In other words, the outliers are being detected where the parameter 𝜆 is the threshold between the quadratic
with higher probability by incorporating multiple sources of and the linear component and is selected to be around 1.5 [50].
data. Also the innovations Zk−1 are used to get more redun- Putting the derivative of the objective function equal to zero
dancy in Step k. Then, by applying the PS, the weights are results in
calculated as [ ]−1
( j +1 )
( ) x̂ k|k = T Tk Q( j ) T k T k Q( j ) z̃ k , (23)
2
wi = min 1, (d ∕PSi ) , (18)
where Q = diag(q(ri )), q(ri ) = 𝜑(ri )∕ri and 𝜑(.) is the deriva-
where d is chosen to be 1.5 in order to achieve high statistical tive of function 𝜌(.) with respect to ri . Also j represents the
efficiency under Gaussian noises. j +1 j
iteration counter. The iterations stop when ∥ x̂ k|k − x̂ k|k ∥≤
𝜀, for example 𝜀 = 0.01. Gandhi and Mili [33] derived the
influence function of this robust filter, which is shown to be
3.5 Generalized maximum-likelihood bounded, and thereby its covariance matrix. After obtaining the
unscented Kalman filter estimation by the GM-UKF using the proposed weight calcu-
lation method, one can robustly estimate frequency on i − th
Based on the weights calculated by using The PS, the bus by combining the estimations obtained by two measure-
robust GM-UKF algorithm are applied to achieve a reliable ment sources, that is, the FD-calculated and the PLL-calculated
estimation of the load bus frequency. Note that the GM- measurements given by (13).
UKF utilizes the same steps of prediction as those of the The whole implementation procedure of the modified fusion
UKF. In order to update the predictions, we first obtain GM-UKF is depicted in Figure 5. Based on this figure, the
the uncorrelated representation of the regression problem local filter estimators (LFEs) transmit the needed signals of each
as follows. The UKF can be represented in a batch-mode unit (unit rotor speeds, measured powers etc) to the estima-
regression as tion center. The transmitted signals depends on the model in
[ ] [ ] [ ] use by the central estimation algorithm. The model selection is
zk Hk vk discussed in detail in Section 3.3 of the manuscript. For gener-
= xk + , (19)
x̂ k|k−1 I ek ality, here, the vector of transmitted signals from the generator
units is referred to as yG ,k . The frequency measurements by the
SRF-PLL are also transmitted to the estimation center. There-
in which ek is the prediction error and ṽ k = [ vk ek ]T is the
fore, we have two sets of measurement. Besides, the dynamic
total error on the measurement and the prediction, which has
evolution models for such signals are known for the central esti-
a covariance matrix given by P̃ k = diag(Rk , Pk|k−1 ), which will
mators (Dynamic model of generators and SRF-PLL dynamic
be factorized as P̃ = M k M Tk . Multiplying (19) by M −1
k leads to model). Using these known models and the received sets of
[ ] measurements the estimated vector of generator speeds ŷ G ,k are
Δ zk calculated and is given to FD formula to calculate the frequency
z̃ k = M −1 = T k xk + ṽ k , (20)
k
x̂ k|k−1 estimation at Bus i. The same is done for the SRF-PLL measure-
ments. The estimates are finally combined using an covariance
where T k = M −1 T
k [ H k I ] and ṽ k has the identity covariance intersection method. Next, we show through simulations that
I n+ny , indicating that the errors are uncorrelated. To solve for the robust state estimation by the GM-UKF improves when the
the robust Kalman gain, we make use of a Shweppe-type GM- fusion of two measurements is employed.
estimator that minimizes an objective function defined as

∑
m+n 4 SIMULATION RESULTS
Jk = wi2 𝜌 (ri ),
i=1 (21) Let us firstly assess the performance of the proposed robust
fusion filter when it is applied to the four machine test case
ri = (z̃ i − aiT x)∕(sw
̂ i ), [2] provided with a WAPSS. As depicted in Figure 6, each
of two areas include two synchronous generators connected
where aiT is the i-th row of matrix T k , s = to step-up transformers and are related through two parallel
1.4826bm median|zi − aiT x|
̂ is a robust estimation of the scale 20-kV tie lines. To do the simulations, we make use of the power
and parameter bm is a corrective coefficient. Also 𝜌(.) is a Huber system toolbox [51]. Similar to [4], a WAPSS on Generator 3 is
FARAHANI ET AL. 9

FIGURE 5 Implementation procedure of the fusion GM-UKF method for estimating the frequency at Bus i.

1 consists of only the swing equations. If the electrical and

mechanical power values are transmitted to the fusion filter, the
UKF-FD enjoys a high level of observability despite a heavy
communication load. In Model 2, the turbine-governor dynam-
ics are taken into account by the UKF in addition to the rotor
speed and rotor angle dynamics. As a result, the model is of
higher order with a lower system observability. However, there
FIGURE 6 Four-machine two-area power system.
is no need to transmit the mechanical power to the fusion
center, which halves the communication load. The most com-
plex Model 3 considers the dynamics of both the electrical
considered, taking the feedback from the remote local Bus 10. and the mechanical subsystems. Using this model, the only sig-
It should be noted that we make use of the same WAPSS setup nal that is known by the UKF-FD is the excitation voltage
and parameters used in [4, 9]. In fact, the mission here is not E fd , which is sent to the fusion center. The system observ-
the WAPSS design but to provide a feedback signal of higher ability is low, inducing the divergence of the UKF-FD in some
quality for the WAPSS. It is also assumed that the SRF-PLL is cases.
available at the same bus. A 3-phase fault is assumed to occur The performance of the UKF-FD to estimate the frequency
on the line connecting Bus 3 to 101 at time t = 2.1 s. The fault at the load Bus 10 is assessed using the three described mod-
is removed at time t = 2.15 s from Bus 3 and at t = 2.20 s from els. Figures 7 and 8 display the results for disturbances applied
101. In Section 4.4, other types of faults and load shedding to the generator shaft torque and for the three-phase fault,
are also considered. To do the bus frequency estimation, local respectively. While being more accurate, a system model with
estimations of the generators’ rotor speeds together with some a higher number of states generally reduces the observability of
input signals such as the generators’ mechanical or electrical the system. However, when using simpler models, some signals
powers (based on the selected model) are transmitted to the such as the generator mechanical or electrical powers may be
fusion center, where the measurements from bus metering transmitted as known inputs to the control center, yielding a
devices such as SRF-PLL are also received. At that center, the higher observability with improved model accuracy. Therefore,
local estimates go through UKF filters, as we call here UKF-FD a tradeoff between the needed number of known/transmitted
and UKF-PLL. The results are combined using the covariance signals and system’s observability/accuracy leads us to choose
weighting method. Before we evaluate the performance of Model 2.
the fusion filter in the frequency stabilization, we first study
the effect of using different dynamic generator models on the
performance of the UKF-FD filter. 4.2 Frequency estimation results based on
fusion filtering

4.1 Frequency estimation results using By applying the covariance intersection method, the fusion fil-
different generator models ter combines two signals, namely one from the SRF-PLL and
the other one from the FD. Under the Gaussian assumption of
As we discussed previously, the selection of the model complex- the noises, the output of the filter is equal to a mixture of two
ity level for the UKF-FD has a great impact on the observability estimations, with higher weight placed on the one with smaller
and the required communication load. In Table 1, three models error. Although the FD calculations are more accurate than the
are specified, which have different levels of complexity. Model PLL measurements, especially when large disturbances occur,
10 FARAHANI ET AL.

FIGURE 7 The effect of the generator dynamic model selection on the frequency estimation at Bus 10 using the UKF-FD for the torque disturbance of
magnitude 0.5 p.u. applied on the generator shaft from t = 5 s to t = 5.5 s; (a) frequency estimation at the local Bus 10; (b) instantaneous frequency mean squared
errors; and (c) total frequency cumulative mean squared errors.

FIGURE 8 The effect of the generator dynamic model selection on the frequency estimation at Bus 10 using the UKF-FD for a three-phase fault occurred on
the line between Bus 3 and 101 at t = 2.1 s and cleared at t = 2.2 s: (a) frequency estimation of the local Bus 10; (b) instantaneous frequency mean squared errors;
and (c) total frequency cumulative mean squared errors.

the data received at the control center are also contaminated by the total cumulative angular frequency deviation of the gener-
several types of errors, such as communication and processing ators from the base value for the three methods of UKF-FD,
errors, to name a few. Furthermore, the FD relies on the suscep- UKF-PLL and UKF-Fusion from t = 10 s to t = 14 s.
tance matrix of the grid, which may include some uncertainties.
Therefore, there may be situations where the communication
or the process noises or other uncertainties greatly influence 4.3 Robust frequency estimation using the
only one data channel while lightly impacting the other channel. fusion GM-UKF
In that case, the fusion process puts higher weight on the data
channel that has smaller error variance, yielding more accurate First, recall that the suggested fusion filter is designed based
final estimation results. Therefore, the fusion of the estimates on the UKF and the variance weighting method relying on the
improves the estimated load bus frequency, which is the feed- assumption of Gaussian noise. Although the fusion UKF can
back signal used by the WAPSS controller, and thereby enhances deal with some limited deviations from that assumption, it will
the quality of the frequency damping. To examine the perfor- be shown that the Fusion-UKF is vulnerable to noise with thick-
mance of UKF fusion method, we change the noise power on tailed distributions such as the Cauchy distribution. It is also
both channels alternately, that is, the SRF-PLL data is becoming observed that this filter is not robust to outliers among the mea-
noisy between t = 10 s and t = 12 s and the FD measure- surements, especially when the latter are on both measurement
ments faces a high intensity Gaussian noise between t = 12 s channels. Indeed, the outlying data points violate the Gaussian
and t = 14 s. Figure 9 shows that the fusion estimation provides assumption and disrupt the frequency estimations provided by
estimates closer to the measurement with the lowest noise. Also the fusion filter. The poor performance of the fusion method in
the positive performance of the fusion method on the damp- dealing with outliers is shown in Figure 10, where the FD/SRF-
ing of the frequency on Generator 3 is observed. As seen, the PLL data are considered to be contaminated with observation
fusion of the signals improves the feedback signal, resulting in a outliers of magnitude 0.005 p.u, from t = 6s to t = 8s. It is
better damping of the oscillations as compared with each indi- also shown that outliers can affect the frequency stabilization
vidual measurement method. Figure 9c shows a comparison of performed by the WAPSS.
FARAHANI ET AL. 11

FIGURE 9 Bus frequency estimation using different fusion filters and their effect on the stabilization of the electromechanical oscillations: (a) frequency
estimations of the load Bus 10; (b) the effect of the fusion filters on damping the frequency oscillations of G 3; and (c) total frequency cumulative mean squared
errors calculated from t = 10 s to t = 14 s.

FIGURE 10 Effect of outliers on Fusion-UKF: (a) frequency estimations of the load Bus 10 when only FD measurements are contaminated with outliers of
magnitude 0.005 p.u. from t = 6 s to t = 8 s (Case 1); (b) frequency estimations of the load Bus 10 when only SRF-PLL measurements are contaminated with
outliers of magnitude 0.005 p.u. from t = 6 s to t = 8 s (Case 2); (c) frequency estimations of the load Bus 10 when both FD and SRF-PLL measurements are
contaminated with outliers of magnitude 0.005 p.u. from t = 6 s to t = 8 s (Case 3); and (d) total frequency cumulative mean squared errors.

FIGURE 11 Estimation of the load bus frequency by the Fusion-UKF and the fusion GM-UKF in the presence of outliers among SRF-PLL data from t = 3 s
to t = 3.49 s and among FD data from t = 3.5 s to t = 4 s using the Fusion-UKF and the fusion GM-UKF: (a) Performance of fusion UKF methods in the
presence of outliers, (b) comparison of fusion UKF and fusion GM-UKF in stabilizing the frequency of Generator G 3 in the presence of outliers.

Let us now demonstrate that using the fusion GM-UKF 0.1 p.u. from t = 3 s to t = 3.49 s and from t = 3.51 s to t = 4
results in a robust estimation method. Recall that the fusion s, respectively. As seen in Figure 11, the GM-UKF can provide
GM-UKF is a combination of the GM-UKF developed in a reliable and accurate estimation, resulting in better damping of
[33–37] and the covariance intersection fusion method while the frequency oscillations.
making use of the weights calculated by the projection statistics. Let us now analyze the ability of the Fusion-UKF and the
The frequency estimation by the fusion GM-UKF is compared fusion GM-UKF to deal with a noise following a thick tailed
to that of the fusion UKF and the results are displayed in probability distribution. Figure 12 displays the instantaneous
Figure 11. It is assumed that the data on the SRF-PLL and the mean squared errors of the state estimation methods consid-
FD channels are contaminated by outliers with a magnitude of ering that the measurement noise follows four distributions,
12 FARAHANI ET AL.

FIGURE 12 Performance of the fusion estimation methods against noise following non-Gaussian probability distributions: (a) Laplace noise distribution with
mean 0 and standard deviation 0.0003 p.u.; (b) Student-t distribution with two degrees of freedom and with mean 0 and standard deviation 0.0006 p.u.; (c) Cauchy
distribution with the location parameter a = 0 and the scale parameter b = 0.0003; and (d) frequency total cumulative mean squared errors for three noise
probability distributions.

FIGURE 13 Comparison of the frequency estimation using the fusion GM-UKF, the GM-UKF-FD, and the GM-UKF-PLL. The noise on the FD follows the
Cauchy distribution with the location parameter a = 0 and the scale parameter b = 0.0003 while the noise on the PLL follows the Gaussian distribution: (a) mean
squared errors of the frequency estimation; (b) total frequency cumulative mean squared errors.

namely, (1) a Gaussian distribution, (2) a Laplace distribution, estimation methods when they estimate the frequency on the
(3) a Student’s t-distribution, and (4) a Cauchy distribution. The local Bus 10. The overall outcome of this comparison is that
results show that the Fusion-UKF performs satisfactory for the fusion approach has increased the efficiency and robust-
a noise following the Gaussian and the Laplace distribution ness of the frequency estimation method even more than the
while the fusion GM-UKF performs better for a noise follow- GM-UKF-FD presented in [23]. The first reason is the sta-
ing thicker tailed distributions such as the Student-t distribution tistical enhancement that the covariance intersection approach
with a low degree of freedom and the Cauchy distribution. provides to the frequency estimation. The other and more
Now, let us compare the Fusion GM-UKF, which makes use important reason is the higher outlier detection ability provided
of both measurements, with the GM-UKF presented in [23]. We by the projection statistics to the redundant SRF-PLL measure-
refer to the latter as the GM-UKF-FD. For a more comprehen- ments. In fact, even if the SRF-PLL measurements are not quite
sive study, we also consider the so-called GM-UKF-PLL, which precise, they still can play a role in increasing the probability of
only processes SRF-PLL measurements. Note that in [23], only the outlier detection technique.
the FD calculations are employed. Figure 13 compares the per- Let us now analyze the impact of the delays on the received
formances of the fusion GM-UKF, the GM-UKF-FD, and the FD and SRF-PLL signals. Figure 14 presents the state estima-
GM-UKF-PLL. For a better assessment of the Fusion GM- tion results of the bus frequency when two signals are not of the
UKF effect, we assume that the noise on the FD follows the same time benchmark. As noticed, the fusion GM-UKF enjoys
Cauchy distribution with a location parameter of a = 0 and a the highest estimation performance when there are delays on
scale parameter of b = 0.0003 and that the noise on the PLL the FD measurements. It can also handle missing measure-
follows the Gaussian distribution with the previously assumed ments. To show that, let us assume that the measurements on
parameters. Figures 13a and 13b display the mean squared both channels are missed between t = 3 s and t = 4 s and the
errors and the total cumulative mean squared errors of the three last received measurements in t = 2.99 s are used as the only
FARAHANI ET AL. 13

FIGURE 14 Performance of the fusion estimation methods against delay differences between the SRF-PLL and the FD signals: (a) comparison of the
performance of the UKF-FD, the UKF-PLL, and the Fusion-UKF for delay=50 ms; (b) effect of the time delay on the damping of the frequency oscillations of G 3;
(c) mean squared errors of the frequency estimation methods for different delays.

FIGURE 15 Performance of the fusion estimation of the frequency estimation at Bus 10 and its effect on improving frequency oscillation damping of
Generator G3 by the WAPSS with missing data between t = 3 s to t = 4 s: (a) frequency estimation at Bus 10; (b) frequency oscillation damping of Generator G3;
and (c) total mean squared errors of the oscillation damping estimation.

available information in the next time interval. Figure 15 dis- 4.5 Performance of the fusion GM-UKF
plays the results on the performance of the fusion GM-UKF frequency estimation in new-england test
in this situation. As seen, it provides the best performance as system
compared to the other methods.
We carry out a study on frequency estimation within the 39-Bus
New-England power system, whose data are provided in [9].
4.4 fusion GM-UKF estimation for other Zhang and Bose [9] study inter-area oscillations in that system
faults/load shedding scenarios and recommend to install a WAPSS on Generator G 7. The lat-
ter takes the frequency at Bus 16 as its feedback signal. As in the
As another important evaluation of the Fusion GM-UKF esti- previous example, it is assumed that a SRF-PLL is located at Bus
mation and its effect on the frequency oscillation damping, we 16, which calculates the angular frequency at that bus and sends
apply other types of disturbances on the line between Buses 3 it to the fusion processing center. Also, the angular frequencies
and 101 at time t = 2.1 s. These disturbances are: (1) a line- of all the generators are simultaneously estimated in a decen-
to-ground fault; (2) a two-line-to-ground fault; (3) a line-to-line tralized manner and are sent to the fusion center. In the fusion
fault; and (4) a loss-of-line leading to a setpoint change. We also center, after carrying out a centralized estimation and fusion,
apply a fifth disturbance that consists in a 20-percent loss of the high-quality estimate of frequency at Bus 16 is applied to
loads on Bus 4 starting from t = 2.1 s. The results are displayed WAPSS installed on Generator G 7. A 3-phase fault is assumed
in Figure 16. Here, outliers are included among the SRF-PLL to happen on the line between Buses 1 and 39 at t = 2.1 s. The
data from t = 3 s to t = 3.49 s and among the FD data from fault is then removed at t = 2.15 s from Bus 1and at t = 2.20 s
t = 3.5 s to t = 4 s. It is observed that the fusion method is from 39.
able to provide the feedback signal of a higher quality for the To show the effectiveness of the fusion GM-UKF method,
WAPSS, leading to an improved frequency oscillation damping. it is assumed that the FD data is contaminated by observation
14 FARAHANI ET AL.

FIGURE 16 The ability of the fusion GM-UKF to improve the frequency oscillation damping by the WAPSS against various disturbances at t = 2.1s, which
includes: (a) a line-to-ground fault on Line 3-101; (b) a two-line-to-ground fault on the same line; (c) a line-to-line fault; (d) a line outage leading to a setpoint change;
and (e) a 20 percent loss-of-load at Bus 4.

FIGURE 17 Estimation performed at Bus 16 in the New England test system by the fusion filters when there are outliers among the FD data from t = 3 s to
t = 4 s using the Fusion-UKF and the fusion GM-UKF; (a) the impact of outliers on the frequency estimations at Bus 16; (b) the impact of outliers on the damping
of the frequency oscillations of Generator G 7; (c) comparison of the MSEs of the Fusion-UKF and of the fusion GM-UKF for all the generators.

outliers of 0.1 p.u. from t = 3s to t = 4s. Figure 17a displays filtering method is a limitation especially for very large networks
the results of the angular frequency estimation at Bus 16. The calling for fast processors. This can be regarded as the price
effect of the robust frequency estimation on the WAPSS fre- to pay to have higher robustness against non-Gaussian noises
quency stabilization is shown in Figure 17b. As seen, the fusion and outliers. Note that the proposed fusion UKF has reason-
GM-UKF provides an accurate estimation at the time where able computational speed and may be used most of times when
the outlier is applied, but the Fusion-UKF dos not since it no outliers exist.
tracks the outlier. This leads to better stabilization when using
the fusion GM-UKF estimations as the feedback signal rather
than using the Fusion-UKF. Figure 17c displays the MSEs. As 5 CONCLUSIONS
for Table 2, it provides the simulation times of all the tested
algorithms. As seen from Table 2 and the previously presented To construct a high quality feedback signal for the WAPSS,
figures, the fusion GM-UKF algorithm provides lower estima- the Fusion-UKF is adapted based on the time dependent
tion error with higher computing time. The run-time of the weighting on the signals provided by the SRF-PLL and the FD
FARAHANI ET AL. 15

TABLE 2 Comparison of the simulation times (in seconds) for all the 6. Mohanty, A.K., Barik, A.K.: Power system stability improvement using
tested algorithms. facts devices. Int. J. Modern Eng. Res. 1(2), 666–672 (2011)
7. Chow, J.H., Sanchez-Gasca, J.J., Ren, H., Wang, S.: Power system damping
Algorithms Two-area Case New England case
controller design-using multiple input signals. IEEE Control Syst. Mag.
GM-UKF-Fusion 39 98 20(4), 82–90 (2000)
8. Aboul-Ela, M.E., Sallam, A.A., McCalley, J.D., Fouad, A.A.: Damping con-
GM-UKF-FD 37 95
troller design for power system oscillations using global signals. IEEE
GM-UKF-PLL 32 89 Trans. Power Syst. 11(2), 767–773 (1996)
UKF-Fusion 29 76 9. Zhang, Y., Bose, A.: Design of wide-area damping controllers for interarea
oscillations. IEEE Trans. Power Syst. 23(3), 1136–1143 (2008). https://
UKF-FD 27 74 doi.org/10.1109/TPWRS.2008.926718
UKF-PLL 24 72 10. Shakarami, M.R., Faraji Davoudkhani, I.: Wide-area power system
stabilizer design based on grey wolf optimization algorithm con-
sidering the time delay. Electr. Power Syst. Res. 133, 149–159
(2016)
11. Chung, C.Y., Wang, K.W., Tse, C.T., Bian, X.Y., David, A.K.: Probabilis-
tic eigenvalue sensitivity analysis and pss design in multimachine systems.
calculations. Furthermore, to address the problem of outliers IEEE Trans. Power Syst. 18(4), 1439–1445 (2003)
among the data, a robust fusion GM-UKF has been devel- 12. Zhang, P., Messina, A.R., Coorick, A., Cory, B.J.: Selection of locations
oped with a modified calculation of the weights based on the and input signals for multiple svc damping controllers in large scale power
projection statistics. The results show that the latter is able to systems. In: IEEE Power Engineering Society 1999 Winter Meeting (Cat.
No.99CH36233), vol. 1, pp. 667–670. IEEE, Piscataway (1999)
improve the performance of the frequency damping. Moreover, 13. Heniche, A., Kamwa, I.: Assessment of two methods to select wide-area
it enhances the quality of the feedback data when there are signals for power system damping control. IEEE Trans. Power Syst. 23(2),
outliers among the data provided by the SRF-PLL and the FD. 572–581 (2008)
14. Song, F.F., Bi, T.S., Yang, Q.X.: Study on wide area measurement system
AUTHOR CONTRIBUTIONS based transient stability control for power system. In: 2005 International
Power Engineering Conference, vol. 2, pp. 757–760. IEEE, Piscataway
Ali Farahani: Investigation, software, writing - original (2005)
draft. Amir Abolmasoumi: Conceptualization, formal analy- 15. Zhao, J., Zhang, G., La Scala, M.: A two-stage robust power system state
sis, methodology, resources, software, supervision, writing - estimation method with unknown measurement noise. In: 2016 IEEE
review and editing. Lamine Mili: Conceptualization, supervi- Power and Energy Society General Meeting (PESGM), pp. 1–5. IEEE,
sion, writing - review and editing. Mohammad Bayat: Resources, Piscataway (2016)
16. Zhou, N., Meng, D., Huang, Z., Welch, G.: Dynamic state estimation of a
supervision, validation. synchronous machine using pmu data: A comparative study. IEEE Trans.
Smart Grid 6(1), 450–460 (2014)
CONFLICT OF INTEREST STATEMENT 17. Zhao, J., Gómez-Expósito, A., Netto, M., Mili, L., Abur, A., Terzija, V.,
The authors have declared no conflict of interest. Kamwa, I., Pal, B., Singh, A. K., Qi, J., et al.: Power system dynamic
state estimation: Motivations, definitions, methodologies, and future work.
IEEE Trans. Power Syst. 34(4), 3188–3198 (2019)
FUNDING INFORMATION 18. Chen, J., Dou, C., Xiao, L., Wang, Z.: Fusion state estimation for power
There is no funding to declare regarding this manuscript. systems under dos attacks: A switched system approach. IEEE Trans. Syst.
Man Cybern. Syst. 49(8), 1679–1687 (2019)
DATA AVAILABILITY STATEMENT 19. Singh, A.K., Pal, B.C.: Decentralized robust dynamic state estimation in
Data available on request from the corresponding author. power systems using instrument transformers. IEEE Trans. Signal Process.
66(6), 1541–1550 (2018)
20. Mir, A.S., Singh, A.K., Senroy, N.: Robust observer based methodology for
ORCID frequency and rate of change of frequency estimation in power systems.
Amir H. Abolmasoumi https://orcid.org/0000-0001-9739- IEEE Trans. Power Syst. 36(6), 5385–5395 (2021)
1340 21. Chen, X., Henry, M., Duncan, S.R.: An enhanced algorithm for fre-
Mohammad Bayat https://orcid.org/0000-0003-1465-0015 quency estimation in power systems. In: 2018 UKACC 12th International
Conference on Control (CONTROL), pp. 140–145. IEEE, Piscataway
(2018)
REFERENCES 22. Milano, F., Ortega, A.: Frequency divider. IEEE Trans. Power Syst. 32(2),
1. Klein, M., Rogers, G.J., Kundur, P.: A fundamental study of inter-area 1493–1501 (2017)
oscillations in power systems. IEEE Trans. Power Syst. 6(3), 914–921 23. Zhao, J., Mili, L., Milano, F.: Robust frequency divider for power system
(1991) online monitoring and control. IEEE Trans. Power Syst. 33(4), 4414–4423
2. Kundur, P.: Power System Stability and Control. In: EPRI Power System (2018)
Engineering Series. McGraw-Hill, New York (1994) 24. Ortega, A., Milano, F.: Comparison of different pll implementations for
3. Rogers, G.: Power system oscillations. In: Power Electronics and Power frequency estimation and control. In: 2018 18th International Conference
Systems. Springer, New York (2012) on Harmonics and Quality of Power (ICHQP), pp. 1–6. IEEE, Piscataway
4. Tzounas, G., Liu, M., Murad, M.A.A., Milano, F.: Impact of realistic bus (2018)
frequency measurements on wide-area power system stabilizers. In: 2019 25. Freijedo, F.D., Doval-Gandoy, J., Lopez, O., Acha, E.: Tuning of phase-
IEEE Milan PowerTech, pp. 1–6. IEEE, Piscataway (2019) locked loops for power converters under distorted utility conditions. IEEE
5. Abolmasoumi, A.H., Moradi, M.: Nonlinear ts fuzzy stabilizer design for Trans. Ind. Appl. 45(6), 2039–2047 (2009)
power systems including random loads and static synchronous compen- 26. Zhao, J., Zhang, G., La-Scala, M., Wang, Z.: Enhanced robustness of state
sator. international transactions on electrical energy systems. IEEE Trans. estimator to bad data processing through multi-innovation analysis. IEEE
Power Syst. 28(4), 1559–1564 (2018) Trans. Ind. Inf. 13(4), 1610–1619 (2016)
16 FARAHANI ET AL.

27. Zhao, C., Topcu, U., Low, S.H.: Optimal load control via frequency mea- 40. Aboul-Ela, M.E., Sallam, A., McCalley, J.D., Fouad, A.: Damping controller
surement and neighborhood area communication. IEEE Trans. Power design for power system oscillations using global signals. IEEE Trans.
Syst. 28(4), 3576–3587 (2013) Power Syst. 11(2), 767–773 (1996)
28. Huang, C., Li, F., Zhan, L., Xu, Y., Hu, Q., Zhou, D., Liu, Y.: Data qual- 41. Yao, W., Jiang, L., Wu, Q., Wen, J., Cheng, S.: Delay-dependent stabil-
ity issues for synchrophasor applications part ii: problem formulation and ity analysis of the power system with a wide-area damping controller
potential solutions. J. Mod. Power Syst. Clean Energy 4(3), 353–361 (2016) embedded. IEEE Trans. Power Syst. 26(1), 233–240 (2010)
29. Huang, C., Li, F., Zhou, D., Guo, J., Pan, Z., Liu, Y., Liu, Y.: Data quality 42. Wang, Y., Yemula, P., Bose, A.: Decentralized communication and control
issues for synchrophasor applications part i: A review. J. Mod. Power Syst. systems for power system operation. IEEE Trans. Smart Grid 6(2), 885–
Clean Energy 4(3), 342–352 (2016) 893 (2014)
30. Liu, Y., Gracia, J.R., Ewing, P.D., Zhao, J., Tan, J., Wu, L., Zhan, L.: Impact 43. Ortega, A., Milano, F.: Impact of frequency estimation for vsc-based
of measurement error on synchrophasor applications. Oak Ridge National devices with primary frequency control. In: 2017 IEEE PES Innovative
Lab. (ORNL), Oak Ridge, TN (2015) Smart Grid Technologies Conference Europe (ISGT-Europe), pp. 1–6.
31. Sundararajan, A., Khan, T., Moghadasi, A., Sarwat, A.I.: Survey on syn- IEEE, Piscataway (2017)
chrophasor data quality and cybersecurity challenges, and evaluation of 44. Liggins, M., Hall, D., Llinas, J.: Handbook of Multi Sensor Data
their interdependencies. J. Mod. Power Syst. Clean Energy 7(3), 449–467 Fusion: Theory and Practice. In: Electrical Engineering & Applied Signal
(2019) Processing Series, 2nd ed. CRC Press, Boca Raton, FL (2017)
32. Ortega, A., Milano, F.: Comparison of different pll implementations for 45. Raol, J.R.: Multi-Sensor Data Fusion with MATLAB. CRC Press, Boca
frequency estimation and control. In: 2018 18th International Conference Raton, FL (2009)
on Harmonics and Quality of Power (ICHQP), pp. 1–6. IEEE, Piscataway 46. Uhlmann, J.K.: Dynamic Map Building and Localization: New Theoretical
(2018) Foundations. University of Oxford, Oxford (1995)
33. Gandhi, M.A., Mili, L.: Robust kalman filter based on a generalized 47. Chen, L., Arambel, P.O., Mehra, R.K.: Estimation under unknown cor-
maximum-likelihood-type estimator. IEEE Trans. Signal Process. 58(5), relation: Covariance intersection revisited. IEEE Trans. Autom. Control
2509–2520 (2010) 47(11), 1879–1882 (2002)
34. Zhao, J., Mili, L.: A robust generalized-maximum likelihood inscented 48. Julier, S.J., Uhlmann, J.K.: Using covariance intersection for slam. Rob.
kalman filter for power system dynamic state estimation. IEEE J. Sel. Top. Auton. Syst. 55(1), 3–20 (2007)
Signal Process. 12(4), 578–592 (2018) 49. Donoho, D.L., Gasko, M.: Breakdown properties of location estimates
35. Zhao, J., Mili, L.: Robust unscented kalman filter for power system dynamic based on halfspace depth and projected outlyingness. Ann. Statist. 20(4),
state estimation with unknown noise statistics. IEEE Trans. Smart Grid 1803–1827 (1992). https://doi.org/10.1214/aos/1176348890
10(2), 1215–1224 (2019) 50. Mili, L., Cheniae, M.G., Vichare, N.S., Rousseeuw, P.J.: Robust state estima-
36. Zhao, J., Netto, M., Mili, L.: A robust iterated extended kalman filter for tion based on projection statistics [of power systems]. IEEE Trans. Power
power system dynamic state estimation. IEEE Trans. Power Syst. 32(4), Syst. 11(2), 1118–1127 (1996)
3205–3216 (2017). https://doi.org/10.1109/TPWRS.2016.2628344 51. Chow, J.H., Cheung, K.W.: A toolbox for power system dynamics and con-
37. Ortega, A., Milano, F.: Comparison of bus frequency estimators for power trol engineering education and research. IEEE Trans. Power Syst. 7(4),
system transient stability analysis. 2016 IEEE International Conference 1559–1564 (1992)
on Power System Technology (POWERCON), pp. 1–6. IEEE, Piscataway
(2016)
38. Farahani, A., Abolmasoumi, A.H., Bayat, M., Mili, L.: A fast outlier-robust
fusion estimator for local bus frequency estimation in power systems. How to cite this article: Farahani, A., Abolmasoumi,
In: 2020 10th Smart Grid Conference (SGC), pp. 1–6. IEEE, Piscataway A.H., Mili, L., Bayat, M.: A robust fusion bus frequency
(2020) estimation method to improve frequency oscillation
39. Farahani, A., Abolmasoumi, A.H., Bayat, M.: Fusion estimation of local
bus frequency for robust wide area power system stabilizer. In: 2021 7th
damping in power systems. IET Gener. Transm. Distrib.
International Conference on Control, Instrumentation and Automation 1–16 (2023). https://doi.org/10.1049/gtd2.12976
(ICCIA), pp. 1–5. IEEE, Piscataway (2021)