578 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 12, NO.

4, AUGUST 2018

A Robust Generalized-Maximum Likelihood

Unscented Kalman Filter for Power System
Dynamic State Estimation
Junbo Zhao , Member, IEEE, and Lamine Mili , Life Fellow, IEEE

Abstract—This paper develops a new robust generalized However, these functions are difficult to be achieved without
maximum-likelihood-type unscented Kalman filter (GM-UKF) the development of a fast and robust Dynamic State Estimator
that is able to suppress observation and innovation outliers while (DSE). Indeed, the dynamic state estimates of the synchronous
filtering out non-Gaussian process and measurement noise. Be-
cause the errors of the real and reactive power measurements cal- machines can be utilized by various devices to enhance small
culated using phasor measurement units (PMUs) follow long-tailed signal stability [1] and to initiate generation outages and load
probability distributions, the conventional UKF provides strongly shedding during transient instabilities, such as power system
biased state estimates since it relies on the weighted least squares stabilizers, automatic voltage regulators, and under-frequency
estimator. By contrast, the state estimates and residuals of our GM- relays [2], [3].
UKF are proved to be roughly Gaussian, allowing the sigma points
to reliably approximate the mean and the covariance matrices of To date, a variety of dynamic state estimators have been pro-
the predicted and corrected state vectors. To develop our GM-UKF, posed in the literature; they are based on the Extended Kalman
we first derive a batch-mode regression form by processing the Filter (EKF) [4], [5], the Iterated EKF (IEKF) [6], [7], the un-
predictions and observations simultaneously, where the statistical scented Kalman filter (UKF) [8]–[10], to cite a few. However, all
linearization approach is used. We show that the set of equations these methods suffer from several important shortcomings, pre-
so derived are equivalent to those of the unscented transforma-
tion. Then, a robust GM-estimator that minimizes a convex Huber cluding them from being adopted by power utilities for power
cost function while using weights calculated via projection statis- system real-time applications. To be specific, they cannot han-
tics (PSs) is proposed. The PSs are applied to a two-dimensional dle i) non-Gaussian process and observation noise of the sys-
matrix that consists of a serially correlated predicted state and tem nonlinear dynamic models, ii) innovation, observation and
innovation vectors to detect observation and innovation outliers.
structural outliers and iii) cyber attacks.
These outliers are suppressed by the GM-estimator using the iter-
atively reweighted least squares algorithm. Finally, the asymptotic There are several reasons for these shortcomings. Firstly,
error covariance matrix of the GM-UKF state estimates is derived the current DSE approaches assume that both the process and
from the total influence function. Extensive simulation results car- the observation noise of the system nonlinear dynamic models
ried out on IEEE New England 39-bus 10-machine test system are Gaussian. However, two recent investigations conducted by
verify the effectiveness and robustness of the proposed method.
PNNL [11], [12] revealed that the PMU measurement errors of
Index Terms—Dynamic state estimation, robust estimation, un- the voltage and current magnitudes obey non-Gaussian probabil-
scented Kalman filter, non-Gaussian noise, total influence function, ity distributions. Evidently, the presence of non-Gaussian noise
outliers, cyber attacks, power system dynamics. calls for new research and development in robust power system
I. INTRODUCTION DSE based on robust statistics. Secondly, three types of outliers
associated with a given dynamical system model have been de-
A. Motivation fined by Gandhi and Mili [13], namely observation outliers,
HE widespread deployment of Phasor Measurement Units which affect the metered values; innovation outliers, which cor-
T (PMUs) on power transmission grids has made possible the
real-time monitoring and control of power system dynamics.
rupt the predicted state estimates; and structural outliers, which
affect the system dynamic states and the observation functions.
Observation outliers may result from large biases in PMU mea-
Manuscript received September 7, 2017; revised February 16, 2018; accepted surements due to infrequent calibration, or instrument failures,
March 20, 2018. Date of publication April 16, 2018; date of current version or impulsive communication noise [14], [15]. As for innovation
June 27, 2018. This work was supported in part by the U.S. National Science outliers, they may occur in several different ways. For example,
Foundation under Grant ECCS-1711191 and in part by the Advanced Grid
Modeling program of the U.S. Department of Energy (DOE) Office of Elec- some of the generator models may not be well calibrated, re-
tricity Delivery & Energy Reliability. The guest editor coordinating the review sulting in highly inaccurate model outputs that are inconsistent
of this manuscript and approving it for publication was Dr. Javier Contreras. with the measurements. This was precisely the case in the 1996
(Corresponding author: Junbo Zhao.)
The authors are with the Bradley Department of Electrical and Computer En- blackout, where the model being used predicted system stability
gineering, Virginia Polytechnic Institute and State University, Northern Virginia while in reality the system was undergoing numerous cascading
Center, Falls Church, VA 22043 USA (e-mail:,[email protected]; [email protected]). failures, which resulted in a rapid system collapse that occurred
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. within minutes [16], [17]. Innovation outliers may also be in-
Digital Object Identifier 10.1109/JSTSP.2018.2827261 duced by the approximations in the state prediction model or by
1932-4553 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 ZHAO AND MILI: ROBUST GM-UKF FOR POWER SYSTEM DYNAMIC STATE ESTIMATION 579

a system process impulsive noise [18]. By contrast, structural r A theorem is provided, which shows that the state estimates
outliers are induced by wrong circuit breaker statuses or gross provided by our GM-UKF are asymptotically Gaussian
errors in the model parameters of the transmission lines, or of the even when the system process and measurement noise obey
automatic voltage regulators, or of the synchronous machines. thick-tailed distributions; this is precisely the case when
In [19], it is reported that wrong estimates of the parameters using PMU measurements [12]. As a result, the means and
of the synchronous machine models may result from the use the covariance matrices of the state estimates can be rep-
of erroneous metered values. It turns out that the conventional resented well by sigma points. Furthermore, this theorem
filters, namely the EKF, the EnKF, the UKF, and the Particle guarantees a good performance of our GM-UKF for any
Filter (PF) are not robust to any type of outliers. For instance, distributions.
it is demonstrated in [20] that the performances of the EnKF, r Our GM-UKF is shown to be able to handle very strong
EKF, UKF and PF are significantly degraded in the presence of system nonlinearity while other alternatives fail to con-
observation outliers. In addition, since the EKF, UKF and EnKF verge or yield large biased state estimates.
are derived based on Gaussian assumption [21], they may yield r The robustness our GM-UKF to outliers and cyber attacks
significantly biased estimates when non-Gaussian noise occurs. is investigated; it is shown that it can handle at least 25%
To address observation outliers, Rouhani and Abur [10] devel- of incorrect PMU measurements and predicted state vari-
oped a robust UKF-based DSE using the Least-Absolute-Value ables.
(LAV) estimator. However, the authors do not address the vul- r It is demonstrated both theoretically and numerically that
nerability of the DSE to innovation outliers. In [7], a robust IEKF our GM-UKF is able to achieve high statistical efficiency
was proposed to handle observation and innovation outliers, but for a wide-range of system process and observation noise
it may suffer from divergence problems if the nonlinearity of probability distributions.
the system model is strong. In addition, both [7], [10] do not The rest of the paper is organized as follows. Section II
address the non-Gaussianity of the measurement and system presents the problem formulation. Section III develops the the-
process noise. In signal processing and communication areas, ory of the proposed GM-UKF and Section IV presents the ap-
Huber estimator-based robust UKF methods have been advo- plication of GM-UKF to power system. Section V shows and
cated to suppress observation outliers and bound the influence analyzes the simulation results. Finally Section VI concludes
of non-Gaussian observation noise [22]–[25]. However, these the paper.
methods are unable to handle innovations outliers, which can
occur quite often in power systems due to controller parame- II. PROBLEM FORMULATION
ter errors and model deficiency. Furthermore, they achieve very
low statistical efficiency or may diverge when both process and A. Nonlinear Discrete-Time Dynamical System Model
observation noises are non-Gaussian. A discrete-time state space representation of a general non-
linear dynamical system is expressed as
B. Contributions and Paper Organization xk = f (xk −1 , uk ) + wk , (1)
To address the aforementioned challenges, this paper resorts
z k = h (xk , uk ) + v k , (2)
to robust statistics and develops a robust Generalized Maximum-
Likelihood-type UKF (GM-UKF) for power system dynamic where xk ∈ Rn ×1 and z k ∈ Rm ×1 are the state vector and the
state estimation. Note that in [26], we have shown some prelim- measurement/observation vector at time sample k, respectively;
inary results of our robust UKF, for instance that it is able to f and h are vector-valued nonlinear functions; wk and v k are
handle unknown Gaussian and non-Gaussian noise, and obser- the system process and observation noise, respectively; they are
vation outliers. However, no theoretical proofs are provided to assumed to be independent and identically distributed with zero
quantify its statistical efficiency and its breakdown point. This mean and covariance matrices Qk and Rk , respectively; uk is
paper bridges these gaps by providing the corresponding theo- the system input vector.
retical justifications. Specifically, it includes the following new
contributions:
r The statistical linearization approach is shown to yield the B. Dynamic State Estimation Using UKF
same solution as the sigma-points-based unscented trans- The main idea underlying the UKF is the application of a de-
formation; this allows us to derive a redundancy linear terministic sampling technique known as the unscented transfor-
batch-mode regression form by processing the predictions mation, which allows us, under the Gaussian noise assumption,
and observations simultaneously. Note that this overdeter- to choose a set of sample points, termed sigma points, that
mined system of equations results in the data redundancy have the same mean and covariance matrix as those of the a
needed for the detection and suppression of the innovation priori state vector [27]. These sigma points are then propagated
and observation outliers. through the non-linear functions f and h, yielding an estimation
r The application of the weighted least squares (WLS) esti- of the a posteriori state statistics by using the Kalman filter ap-
mator to the linear batch-mode regression form is proved proach, i.e., the sample mean and the sample covariance matrix.
to yield the same solution as the traditional UKF, making Consequently, no calculation of Jacobian matrices is required,
this filter vulnerable to non-Gaussian process and mea- which can be by itself a difficult task to achieve in some cases
surement noise and any type of outliers. or computationally costly.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 580 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 12, NO. 4, AUGUST 2018

To be specific, given a state estimate at time step k– the measurement errors on nodal voltage and line current an-
1, xk −1|k −1 ∈ Rn ×1 , having a covariance matrix given by gles, which are roughly Gaussian, the measurement errors on
P k −1|k −1 , its statistics are captured by 2n weighted sigma points
xx
both nodal voltage and line current magnitudes obey a bimodal
defined as Gaussian mixture distribution. As for the measurement errors of
  line real and reactive powers calculated from voltage and cur-
k −1|k −1 ±
χik −1 |k −1 = x nP xx
k −1|k −1 , (3) rent phasors, they follow a thick tailed distribution that may be
i
1
approximated by either the Laplace or the Cauchy distribution.
with weights wi = 2n , i = 1, . . . , 2n. Then, each sigma point For more details about how to derive these error statistics, please
is propagated through the nonlinear system process model (1), refer to [12]. As a result, the sigma points may not capture the
yielding a set of transformed samples expressed as complete statistics of the state vector, resulting in poor or even
 
diverging state estimates. Furthermore, since the UKF lacks sta-
χik |k −1 = f χik −1 |k −1 . (4)
tistical robustness, it is sensitive to any type of outliers, including
Next, the predicted sample mean and sample covariance matrix observation, innovation and structural outliers. When dealing
of the state vector are calculated by with a general nonlinear dynamical system, three types of out-
liers, namely observation, innovation and structural outliers, are
2n
 defined in [7]. In power system DSE, observation outliers re-
k |k −1 =
x wi χik |k −1 , (5) fer to the phase biases and gross errors in PMU measurements
i=1
[15]; innovation outliers may be induced by incorrect generator
2n
 parameter values, failure of brushless exciter rotating diodes,
P xx
k |k −1 = k |k −1 )(χik |k −1 − x
wi (χik |k −1 − x k |k −1 )T + Qk . or impulsive system process noise; and structural outliers may
i=1 be caused by transmission parameter errors or topology errors.
(6) In the following section, we will propose a robust GM-UKF
Finally, the measurement updating is performed and the filtered that is able to suppress observation and innovation outliers and
k |k with the covariance matrix P xx
state x k |k are calculated by
to filter out various types of thick-tailed observation noises.
Note that the problem of the identification and suppression of
 −1
K k = P xz P zz
, (7) structural outliers is outside the scope of this paper since it re-
k |k −1 k |k −1
quires a different formulation; it will be addressed in a future
 
k |k = x
x k |k −1 + K k z k − zk |k −1 , (8) paper.

P xx
k |k = P k |k −1 − K k P k |k −1 K k ,
xx zz T
(9) III. THEORY OF THE PROPOSED GM-UKF

where zk |k −1 = 2n i=1 wi z k |k −1 is the predicted measurement

i The proposed GM-UKF is based on the following assump-
tions, which are satisfied by many engineering systems, includ-
vector and z k |k −1 = h(χk |k −1 ); the self and cross-covariance
i i
ing power systems:
matrices, P zk |kz −1 and P xz
k |k −1
, are respectively calculated by r Assumption 1: The measurement redundancy is sufficient
2n
 to make the state of the dynamical system observable. Re-
P zk |kz −1 = wi (z ik |k −1 − zk |k −1 )(z ik |k −1 − zk |k −1 )T + Rk , call that this condition is a prerequisite for state estimation.
i=1
r Assumption 2: The majority of the metered values are re-
(10) liable. In other words, the percentage of bad data among
2n the measurements does not exceed 25 percent, which is a

P k |k −1 =
xz
k |k −1 )(z ik |k −1 − zk |k −1 )T . (11)
wi (χik |k −1 − x sizable percentage that is rarely exceeded in practice.
r Assumption 3: The dynamical system model is sufficiently
i=1
accurate. This assumption is acceptable since typically
C. Problem Statement such a model is derived from physical laws and its pa-
rameters are known with reasonable precision.
If the system process and measurement noise obey a Gaussian
Our filter consists of four major steps, namely a batch-mode
k −1|k −1 , will follow
probability distribution, the filtered state, x
regression form step, a robust pre-whitening step, a robust re-
a Gaussian distribution as well. In that case, the sample mean
gression state estimation step, and a robust error covariance
and the sample covariance matrix of x k −1|k −1 will be captured
matrix updating step. They are described next.
by the sigma points and the UKF will produce reliable state
estimates. However, the Gaussianity assumption may not hold
A. Derivation of the Batch-Mode Regression Form
true in practice. This is precisely the case in power systems;
for instances, impulsive process noise may occur due to sys- In this section, we first show the equivalence of statistical lin-
tem model inaccuracy at a certain time window and the PMU earization and the unscented transformation using sigma points.
measurement noise may not follow a Gaussian distribution as We then derive the proposed batch-mode regression form. The
shown in Fig. 1. This figure displays histograms and parametric former claim is presented in the following theorem:
probability density estimates of PMU errors on nodal voltage Theorem 1: Given the state estimate vector x k −1|k −1 and its
magnitudes and angles, line current magnitudes and angles, and associated covariance matrix P xx k −1|k −1 , statistical linear regres-
line real and reactive powers. It is observed that except for sion [28] applied to an arbitrary nonlinear function g(x) yields
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 ZHAO AND MILI: ROBUST GM-UKF FOR POWER SYSTEM DYNAMIC STATE ESTIMATION 581

Fig. 1. Error distributions of the phasor voltage angle and magnitude, the phasor current angle and magnitude, the real and reactive power using field PMU data
provided by Pacific Northwest National Laboratory.

results that are equivalent to those of the unscented transforma- applied to the nonlinear observation equation, yielding
tion using the sigma points generated according to (3).  
z k = H k xk − x k |k −1 + zk |k −1 + ν k + v k , (12)
Proof: See Appendix A. 
By applying a statistical linearization to the nonlinear system where H k = (P xz k |k −1
)T (P xx
k |k −1
)−1 , which is no longer
process model, we obtain the predicted state vector x k |k −1 along a Jacobian matrix. Here, the covariance of the statisti-
with its covariance matrix P xx k |k −1 . We define 
x k |k −1 = xk − cal linearization error term is Rk = E ν k ν Tk = P zk |kz −1 −
δ k , where xk is the true state vector; δ k is the prediction error; −1 xz
(P xz )T (P xx
k |k −1 ) P k |k −1, where P k |k −1 and P k |k −1 are two
zz xz
and E δ k δ Tk = P xxk |k −1 . Then, statistical linearization can be
k |k −1
covariance matrices that are calculated by following the same

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 582 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 12, NO. 4, AUGUST 2018

steps as those of the UKF. By processing the predictions and the to suppress observation outliers, they fail to handle innovation
observations simultaneously, we get the following batch-mode outliers, which can happen quite frequently in power system due
regression form: to controller parameter errors and model deficiency. Last but not
   the least, they cannot handle the cases when both process and
zk + H k x k |k −1 − zk |k −1 Hk ν k + vk
= xk + observation noise are non-Gaussian.
k |k −1
x I −δ k
(13) B. Robust Prewhitening
which can be rewritten in a compact form as Before carrying out a robust regression, we uncorrelate the
k xk + ek .
zk = H (14) state prediction errors of the batch-mode regression form. This
is achieved by pre-multiplying S −1 k on both sides of (14),
The error covariance matrix is given by yielding
 
Σk |k −1 0 S −1 −1  −1
W k = E ek ek = T
= S k S Tk , (15) k z k = S k H k xk + S k ek , (20)
0 P xx
k |k −1
which can be put into the compact form given by
T
where Σk |k −1 = E[(ν k + v k )(ν k + v k ) ] = Rk + Rk ; I is y k = C k xk + ξ k , (21)
an identity matrix; S k is calculated by the Cholesky decompo-
sition technique. where E[ξ k ξ k T ] = I. However, if outliers occur, the applica-
Theorem 2: The weighted least squares estimator of the tion of S −1
k will corrupt the prewhitening [13]. To overcome this
batch-mode regression form (14) yields an estimated state vector problem, we first detect the outliers and calculate the weights
k |k and its associated covariance matrix P xx
x k |k
that are equiva- using the projection statistics (PS) [13], [29]. Those weights
lent to those of the UKF. will be incorporated in the objective function that is defined in
Proof: It is well-known that the state vector estimate based the proposed GM-estimator shown in Section III-C. Now, we
on the regression model given by (14) using the weighted least describe the procedures used to calculate the weights. We ap-
squares estimator is expressed as ply the PS to a 2-dimensional matrix Z k that contains serially
 −1 correlated samples of the innovations and of the predicted state
xk |k = H T W −1 H k T W −1 z k ,
H (16)
k k k k variables. Note that the innovation vector is defined as the dif-
ference between the observations and their associated predicted
with the covariance matrix P xx T W −1 H
= (H k )−1 . By ap-
k |k k k values at the previous step. Formally, we have
plying an algebraic substitution and using the matrix inversion  
lemma, we get z k −1 − h(xk −1|k −2 ) z k − h( xk |k −1 )
 Zk = , (22)
 −1 −1 k −1|k −2
x k |k −1
x
−1
P xx
k |k
= H T
k R k H k + P xx
k |k −1
where z k −1 − h( xk −1|k −2 ) and z k − h( xk |k −1 ) are the inno-
 −1 vation vectors while x  k |k −1 are the predicted state
 k −1|k −2 and x
= P xx
k |k−1
−P xx
k |k−1
H Tk H k P xx
k |k−1
H Tk +Rk H k P xx
k |k−1 vectors at time instants k–1 and k, respectively. We may also
apply the PS to higher dimensional samples, but we found that
= (I − K k H k ) P xx = P xx − K k P zk |kz −1 K Tk ,
k |k −1 k |k −1 2 dimensions are enough to identify outliers. Note that the PS
(17) values of the predictions and of the innovations are separately
where the gain matrix is expressed as calculated because the values taken by the former and the latter
are centered around different points.
K k = P xx
k |k −1
H Tk (H k P xx
k |k −1
H Tk +Rk )−1 = P xz
k |k −1
(P zk |kz −1 )−1 The PS of the jth row vector, lTj , of the predictions (respec-
(18) tively the innovations) in Z k is defined as the maximum of
Thus, we can conclude that the estimation error covariance is the standardized projections of all the lTj ’s on every direction
identical to that of the UKF in (9). By applying similar substi-  that originates from the coordinatewise medians of the pre-
tutions and using the matrix inversion lemma, we can also show dictions (respectively the innovations) and that passes through
that the estimated state vector is given by every data point, and where the standardized projections are
  based on the sample median and the median-absolute-deviation
x k |k −1 + K k z k − zk |k −1 ,
k |k = x (19) [29]. The implementation of the PS to detect outliers in matrix
which completes the proof.  Z k is displayed in Fig. 2, while its mathematical expression is
Remark: In the literature, a few Huber estimator-based ro- defined as [29]:
T  
bust UKF methods have been proposed and applied to various l  − medi lT  
applications in signal processing, target tracking, to name a P Sj = max
j
 i
 T  , (23)
=1 1.4826 medκ lT κ  − medi li 

few [22]–[25]. However, in their developed regression mod-
els, ν k that compensates higher order Taylor-series expansion for i, j, κ = 1, 2, . . . , m + n.
error terms is neglected. As a consequence, the estimation re- Once the PS values are calculated, they are compared to a
sults are biased. In addition, although these estimators are able statistical threshold to identify outliers. Extensive Monte Carlo

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 ZHAO AND MILI: ROBUST GM-UKF FOR POWER SYSTEM DYNAMIC STATE ESTIMATION 583

Fig. 2. Application of the projection statistics to the matrix Zk for detecting outliers in a two-dimensional dataset that consists of the innovation vectors and the
predicted state vectors.

Fig. 3. Q-Q plots of the sample quantiles of the PS vs. the corresponding quantiles of the χ 22 and χ 24 distributions, where (a) and (b) represent Q-Q plots of PS
with Gaussian and Laplace noise, respectively.

simulations and Q-Q plots reveal that the probability distribu- where i is calculated by (24); rS i = ri /si is the standard-
tions of the PS applied to Z k , whose data points obey bivari- ized residual; ri = yi − cTi x  is the residual, where cTi is the ith
ate Gaussian and Laplace probability distributions, follow chi- row vector of the matrix C k ; s = 1.4826 · bm · mediani |ri | is
square distributions with degree of freedom 2 and 4, respectively the robust scale estimate; bm is a correction factor to achieve
(See Fig. 3). This investigation allows us to apply statistical tests unbiasedness for a finite sample of size m + n at a given prob-
to the PS and to flag all the data points that satisfy PSi > d as ability distribution [29]; ρ(·) is the nonlinear function of rS i. In
outliers. The latter are downweighted via this paper, the convex Huber-ρ function [30] is adopted, that is
  
1 2
i = min 1, d2 /P Si2 , (24) rS , for |rS i | < λ
ρ (rS i ) = 2 i , (26)
where the parameter d is set equal to 1.5 to yield good statistical λ |rS i | − λ2 /2 elsewhere
efficiency at different distributions without increasing too much where the parameter λ between the quadratic and the linear
the bias induced by outliers. As an example, when the noise is segment of ρ(·) is typically chosen between 1.5 to 3 in the
assumed to be Laplacian, the PS obeys a chi-square distribution literature.
with 4 degrees of freedom. In that case, we can choose the To minimize (25), one takes its partial derivative with respect
statistical detection threshold as χ24,0.975 at a significance level to xk and sets it equal to zero, yielding
of 97.5%.

m +n
∂J (xk ) i c i
= − ψ (rS i ) = 0, (27)
C. Robust Filtering and Solution ∂xk i=1
s
To suppress the outliers and filter out thick-tailed non- where ψ (rS i ) = ∂ρ (rS i )/∂rS i is the so-called ψ-function. By
Gaussian measurement noise, we develop a robust GM- dividing and multiplying the standardized residual rS i to both
estimator that minimizes the following objective function: sides of (27) and putting it in a matrix form, we get

m +n  (y k − C k xk ) = 0,
C Tk Q (28)
J (xk ) = i2 ρ (rS i ) , (25)
i=1  = diag(q (rS )) and q (rS ) = ψ (rS )/rS .
where Q i i i i

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 584 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 12, NO. 4, AUGUST 2018

By using the IRLS algorithm [13], [31], the state vector cor- Proof: See Appendix C. 
rection at the j iteration is calculated through Now, we are in a position to derive the covariance matrix
 −1 P xx
k |k from (30). First, let us prove the following theorem:
Δ
(j +1)
xk |k = C Tk Q  (j ) C k  (j ) y k ,
C Tk Q (29) Theorem 4: The sample variance of the robust scale esti-
mator s of the GM-standardized residuals tends to one as the
(j +1) (j +1) (j )
where Δ
xk |k k |k
=x k |k . The algorithm converges
−x number of observation tends to infinity.
when Δ
xk |k
(j +1)
∞ ≤ 10−2 . Proof: See Appendix D. 
Finally, by virtue of Theorems 3 and 4, the asymptotic error
covariance matrix of our GM-UKF at time sample k is updated
D. Asymptotic Error Covariance Matrix of the GM-UKF
by
State Estimates
Upon convergence of the iterative algorithm, the error co- k |k = E[IF (x; Φ, T ) · IF (x; Φ, T ) ]
P xx T

variance matrix P xx k |k is updated so that the state prediction

EΦ ψ 2 (rS i )  T −1  T  −1
for the next step can be performed. To this end, consider the = 2 Ck Ck C k Q C k C Tk C k
-contamination model G = (1 − ) Φ + Δr , where G and Φ {EΦ [ψ  (rS i )]}
are the contaminated and the true cumulative probability dis- (32)
tribution function of the residual, respectively; Δr is the point  
mass to model outliers. Inspired by Fernholtz [32] and Hampel where Q = diag i2 .
et al. [33], the error covariance matrix is updated based on the It should be noted that thanks to the Gaussian normality of
following theorem: our GM-UKF state estimates and residuals, the bulk of the data,
Theorem 3: Let T (·) be the functional form of the GM- including forecasted state variables and received observations,
estimator with a bounded ψ(·) function and Φα be the empirical obey Gaussian distribution. As a result, our GM-UKF always
cumulative probability distribution function, then achieves a high statistical efficiency for a broad range of noise
probability distributions. This will be demonstrated on the IEEE
√ d
α(T (Φα ) − T (Φ)) → N (0, P xxk |k ), (30) 10-machine system in Section V-G.
Remark: To achieve robustness to outliers or model devia-
d
where α = m + n; → means convergence in probability distri- tions, an estimator has to sacrifice some estimation efficiency.
bution; P xxk |k = E[IF (x; Φ, T ) · IF (x; Φ, T ) ] with the total
T Under ideal conditions, including perfect linear system model,
influence function IF (x; Φ, T ) evaluated at Φ. known Gaussian system process and measurement noise and
Proof: See Appendix B.  reliable data without any outliers, the Kalman filter is the maxi-
Discussion: The UKF is able to provide good results only mum likelihood estimator, yielding 100% estimation efficiency.
when the process and observation noises obey a Gaussian dis- Under these assumptions, our proposed GM-UKF achieves at
tribution [27]. Indeed, in that case the filtered state vector x k |k least 95% statistical efficiency as demonstrated in Section V-G.
is Gaussian and the mean and covariance matrix of x k |k can However, if any of these assumptions is violated, the UKF
be accurately estimated by the sample mean and the sample achieves quite poor estimation efficiency and may provide
covariance matrix of the sigma points. However, this property strongly biased state estimates. By contrast, our GM-UKF is
no longer holds true if the Gaussianity assumption of the noises able to withstand these deviations and yield a slightly decreased
is violated. In that case, the state estimate vector x k |k obtained estimation efficiency. Since the aforementioned assumptions
from the UKF is significantly biased due to the lack of statistical are usually violated in practice, our GM-UKF is guaranteed
robustness of the filter to thick-tailed non-Gaussian noise. By to provide reliable state estimates and to yield better efficiency
contrast, our GM-UKF guarantees the asymptotic Gaussianity than the UKF under the foregoing conditions.
of xk |k for thick-tailed non-Gaussian noises and yields reliable
state estimates with good statistical efficiency. It is worth point- IV. APPLICATION TO POWER SYSTEM
ing out that the robustness of the Huber M-estimator-based UKF
The developed robust GM-UKF is a generic technique for
to outliers and to non-Gaussian noises reported in [22]–[24] has
online monitoring of many dynamical cyber-physical systems,
not been confirmed by the simulations carried out on the IEE
including smart grids, autonomous vehicles, aircraft tracking,
New-England 39-bus 10-machine system. In fact, it is found
GPS tracking and navigation, radar systems, to name a few. In
that this estimator provides strongly biased state estimates for
this paper, we take power system dynamic state estimation as
that system as shown in Section V.
an illustrative example to demonstrate the capabilities of our
Let us now derive the total influence function, IF (x; Φ, T ),
GM-UKF for suppressing observation and innovation outliers
of our GM-UKF at the cumulative probability distribution Φ.
while being able to filter out non-Gaussian noise.
Corollary 3.1. The total influence function of the GM-UKF
For an electric power system, its discrete-time state space
defined by (27) using the regression model (21) is expressed as
representation can be expressed as in (33) and (34) as follows:
  −1
1  T 
IF (x; Φ, T ) = ψ (rS i ) CC T ( Φ ) dΦ Cψ (rS i ) . xk = f (xk −1 , uk ) + wk , (33)
s
(31) z k = h (xk , uk ) + v k , (34)

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 ZHAO AND MILI: ROBUST GM-UKF FOR POWER SYSTEM DYNAMIC STATE ESTIMATION 585

where the state vector xk contains the rotor angle, rotor speed, vector given by xk = [δ ω Ed Eq Ef d VF VR TM PS V ]. The
the d- and q- axis state variables of the synchronous generator, relationships given by (35)–(43) and (44)–(46) are represented
the exciter, the voltage regulator, and the governor. z k is the in compact forms by the vector-valued function f (·), which
measurement vector that contains a collection of voltage pha- relates xk to xk −1 , and by h(·), respectively. The system input
sors, current phasors, real and reactive power flows and power vector is denoted by uk = [Vr ef PC ]T . The measurement vector
injections. The noise wk and v k , which are not necessary to be z k contains a collection of voltage phasor measurements V ∠θ
Gaussian noise, are assumed to be white and independent of each and real and reactive power injections Pe and Qe . They are
other. uk represents the input vector; h(·) is the vector-valued measured by the PMUs.
measurement function while f (·) is the vector-valued function Remark: There exist two ways of implementing a DSE to
that relates xk to xk −1 . In this paper, the detailed 9th order estimate power system dynamic state variables: centralized es-
two-axis model with IEEE-DC1A exciter and TGOV1 turbine- timation vs. decentralized estimation. The centralized DSE ap-
governor are implemented and tested. It can be represented by proach requires accurate knowledge of each system component
the following differential and algebraic equations [34]: and all real-time wide-area PMU measurements. Although the
Differential equations: estimation results are helpful for global control applications,
dEq it requires large computing times that may not be compatible

Tdo = −Eq − (Xd − Xd ) Id + Ef d , (35) with real-time applications in large-scale interconnected power
dt
systems. By contrast, the decentralized DSE approach is im-
dE    plemented at each local synchronous generator, and therefore,
Tq o d = −Ed − Xq − Xq Iq , (36)
dt is fast to execute. The proposed GM-UKF can be implemented
dδ in both ways but only the decentralized GM-UKF is used for
= ω − ωs , (37) demonstration in this paper.
dt
2H dω
= TM − Pe − D (ω − ωs ) , (38) V. NUMERICAL RESULTS
ωs dt
dEf d The performances of our GM-UKF to handle thick-tailed non-
TE = − (KE + SE (Ef d )) Ef d + VR , (39) Gaussian noise along with innovation and observation outliers
dt
are assessed on the IEEE 39-bus system. The UKF, the robust
dVF KF KF
TF = −VF + VR − (KE + SE (Ef d )) Ef d , Huber-UKF [22]–[24] and the GM-IEKF [7] are included for
dt TE TE comparisons. The time-domain simulation results are used to
(40) generate a collection of samples of the nodal voltage magni-
dVR tudes and phase angles as well as of the real and reactive power
TA = −VR + KA (Vr ef − VF − V ) , (41) injections at the terminal buses of all the generators. A sam-
dt
dTM pling rate of 50 samples/second is assumed. A synthetic noise is
TC H = −TM + PS V , (42) added to the true values following the probability distributions
dt
  displayed in Fig. 1. Specifically, a zero mean Gaussian noise
dPS V 1 ω is assumed for the voltage angles, a bimodal Gaussian mixture
TS V = −PS V + PC − −1 , (43)
dt R D ωs distribution is assumed for the noise of the voltage magnitudes,
Algebraic equations: and either a Laplace or a Cauchy distribution is assumed for the
noise of the real and reactive power measurements. The two-
Vd = V sin (δ − θ) , Vq = V cos (δ − θ) , (44) axis generator model is assumed and tested, whose parameters
Eq − Vq Vd − Ed are taken from [35]. A disturbance is applied at t = 0.5 s by
Id =  , Iq = , (45) opening the transmission line between Buses 15 and 16. The
Xd Xq
maximal number of iterations allowed for the IRLS algorithm
Pe = Vd Id + Vq Iq , Qe = −Vd Iq + Vq Id , (46) is 20. The parameters λ and d are set to 1.5. The convergence
 tolerance threshold of the IRLS algorithm is 0.01. For the state
where Tdo , Tq o ,
TE , TF , TA , TC H and TS V are time constants,
initialization, the steady-state values with 10% errors are used.
in seconds; KE , KF and KA are controller gains; Vr ef and PC
Due to space limitation, not all the 9 state variables of each gen-
are known control inputs; Eq , Ed , Ef d , VF , VR , TM and PS V
erator are shown; instead estimated values of the rotor angle and
are the q-axis and d-axis transient voltages, field voltage, scaled
speed, the field voltage and the mechanical power of Generator
output of the stabilizing transformer and scaled output of the
5 are utilized for illustration purposes. The mean absolute error
amplifier, synchronous machine mechanical torque and steam
(MAE) is utilized to evaluate the overall performance of each
valve position, respectively; Xd , Xd , Xq and Xq are generator
method.
parameters; V and θ are the terminal bus voltage magnitude and
phase angle, respectively; Pe and Qe are the active and reactive
electrical power outputs; Id and Iq are the d and q axis currents, A. Case 1: Thick-Tailed Non-Gaussian Measurement Noise
Without Outliers
respectively.
By applying time discretization to (35)–(46) using the 4th- In this section, we first evaluate the performance of the GM-
order Ruger-Kutta method, we get (33)–(34), yielding the state IEKF, the UKF, the Huber-UKF and the GM-UKF under normal

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 586 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 12, NO. 4, AUGUST 2018

Fig. 5. Tracking performance of the GM-IEKF, UKF, the Huber-UKF and

GM-UKF when the variances of the process and measurement noise change
over time; (a) the estimated rotor angle and speed, field voltage and mechanical
Fig. 4. Case 1: Tracking performance of the GM-IEKF, UKF, the Huber-UKF power of Generator 5 are used for illustration purposes; (b) the mean absolute
and GM-UKF without outliers; (a) the estimated rotor angle and speed, field error of each of the four filters.
voltage and mechanical power of Generator 5 are used for illustration purposes;
(b) mean absolute error of each of the four filters.

commonly done in most of the literature, but to obtain accurate

conditions. Specifically, a zero mean Gaussian noise with co- dynamic state variables of the governor for controls and stability
variance matrices of 10−4 I with the appropriate dimensions analysis.
are assumed for the process noise and the noise of voltage and In practical power systems, the process and measurement
current angles; the noise of the voltage magnitudes follows a noise may change over time, yielding unknown noise statistics.
bimodal Gaussian mixture with zero mean, variances of 10−4 To further validate the performance of the proposed GM-UKF
and 10−3 and weights of 0.9 and 0.1, respectively; a Laplace under these conditions, we consider the following scenario: the
noise with zero mean and scale 0.02 is added to the real and variance of the system process noise is time varying and follows
reactive power injections measurements. The test results are a Gaussian distribution with mean 10−2 and standard deviation
displayed in Fig. 4. It is observed that the UKF is not able to 2 × 10−3 ; the variances of the Gaussian mixture model are as-
cope with a Laplacian noise even in the absence of outliers. sumed to be time varying as well; in particular, their variances
By contrast, the GM-IEKF and the Huber-UKF can filter out are assumed to follow Gaussian distributions with means 10−4
such noise to a certain degree. However, their statistical effi- and 10−3 , and standard deviations 4 × 10−5 and 3 × 10−4 , re-
ciencies are much lower than our GM-UKF. In particular, the spectively; the scales of the Laplace distributions are assumed
GM-IEKF poorly estimates the field voltage and the rotor speed to follow Gaussian distributions with means 0.25 and 0.3, and
while the Huber-UKF tracks the rotor angle and the field voltage standard deviations 2 × 10−3 and 3 × 10−3 , respectively. The
poorly. By observing Fig. 4, it is interesting to note that the tur- test results are shown in Fig. 5. It is observed from the figure
bine mechanical power is changing during the transient process. that due to the time-varying process and observation noise, the
According to the CIGRE report [36], it can significantly vary estimates of UKF and Huber-UKF show a lot of oscillations and
when control features such as fast valving or special protection impulses. In addition, UKF, Huber-UKF and GM-IEKF obtain
schemes are used to limit the output of the steam driven gen- increased estimation errors. By contrast, our proposed GM-UKF
erator during transients. Consequently, it is of vital importance is able to filter out time varying process and measurement noise,
to not assume it to be fixed at a constant steady-state value as yielding much better results than the other alternatives.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 ZHAO AND MILI: ROBUST GM-UKF FOR POWER SYSTEM DYNAMIC STATE ESTIMATION 587

Fig. 6. Tracking performance of the GM-IEKF, UKF, the Huber-UKF and

GM-UKF when all state variables are initialized with zeros.

We also test the case when all the state variables are initial-
ized to zeros to show the convergence speed of the proposed
GM-UKF. It is worth pointing out that this is an extreme case
because for practical power systems, we can always infer a rea-
sonable initialization for the state variables from the power flow
solutions. The test results are displayed in Fig. 6. Note that
due to the incorrect initialization and the non-Gaussian noise,
both the GM-IEKF and the Huber-UKF diverge and their re-
sults are not shown. In fact, the UKF diverges as well since
it does not estimate the rotor angle correctly. By contrast, our
proposed GM-UKF is able to track the true state variables al-
though it needs some time to converge due to incorrect state
initialization. Fig. 7. Case 2: Tracking performance of the GM-IEKF, the UKF, the Huber-
UKF and the GM-UKF in the presence of observation outliers from t = 4 s to
t = 6 s, where the real and reactive power measurements of Generator 5 are
B. Case 2: Thick-Tailed Non-Gaussian Measurement Noise corrupted with 20% errors; (a) the estimated rotor angle and speed, field voltage
With Observation Outliers and mechanical power of Generator 5 are used for illustration purposes; (b)
mean absolute error of each of the four filters.
The settings are the same as those of Case 1 except for the
presence of observation outliers from t = 4 s to t = 6 s. The latter
are simulated by adding 20% errors to the real and the reactive
are simulated by adding 20% errors to the predicted rotor angle
power measurements of Generator 5. The results are presented in
of Generator 5. This innovation outlier is induced by a gross
Fig. 7. From this figure, we observe that the UKF is not robust to
parameter value in the model. The comparison results are shown
observation outliers since it yields significantly biased results.
in Fig. 8. As expected, due to the non-robustness of the UKF,
Although the GM-IEKF and Huber-UKF can handle them, it
it is unable to handle innovation outliers. Furthermore, Huber-
produces increased biases on the estimates at the time when
UKF fails to suppress that as well since it cannot distinguish
observation outliers occur. By contrast, the GM-UKF suppresses
outliers from observations and predicted state variables. On the
the outliers and produces much less bias than the GM-IEKF and
other hand, the GM-IEKF can handle them, but it produces
Huber-UKF. Note that the Gaussianity of the GM-estimator
larger biases compared with Case 2. This can be explained by
used in the estimation step of the GM-UKF allows that method
the fact that the model errors will not only affect the predicted
to filter out thick-tailed noise while its statistical robustness
state vector but also will produce a smearing effect throughout
enables it to suppress the outliers, hence achieving very good
the Jacobian matrix. As a result, it downweights several good
estimates.
measurements. By contrast, our GM-UKF is capable of handling
both observation and innovation outliers, yielding comparable
C. Case 3: Thick-Tailed Non-Gaussian Measurement Noise performances. Because only the sigma points associated with
With Innovation Outliers
the model errors will be affected and downweighted, this filter
The settings are the same as those of Case 1 except for the obtains better estimates than the GM-IEKF in presence of model
presence of innovation outliers from t = 4 s to t = 6 s. They errors.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 588 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 12, NO. 4, AUGUST 2018

Fig. 8. Case 3: Tracking performance of the GM-IEKF, the UKF, the Huber-
UKF and the GM-UKF in the presence of innovation outliers from t = 4 s to Fig. 9. Case 4: Tracking performance of the GM-IEKF, the UKF, the Huber-
t = 6 s, where the predicted rotor angle of Generator 5 is corrupted with 20% UKF and the GM-UKF in the presence of Cauchy power measurement noise.
errors; (a) the estimated rotor angle and speed, field voltage and mechanical The estimated rotor angle and speed, field voltage and mechanical power of
power of Generator 5 are used for illustration purposes; (b) mean absolute error Generator 5 are used for illustration purposes; (b) mean absolute error of each
of each of the four filters. of the four filters. Since the UKF and Huber-UKF diverge, their results are not
shown in the figure.

D. Case 4: Long Tail Cauchy Measurement Noise system states. This is not the case for our GM-UKF since it
As discussed before, the real and the reactive power measure- achieves a comparable performance and tracks the system state
ment noises may obey a Cauchy distribution as well, which is a at the very beginning of the transient process.
very thick tailed distribution with no moments being defined. To
test the capability of our GM-UKF to handle that case, we as-
sume that the simulation settings are the same as those of Case 1 E. Case 5: Robustness to Strong System Nonlinearity
except that the Cauchy noises with zero median and a scale of Practical systems may be heavily loaded, resulting in strong
0.005 is added to the real and the reactive power injection mea- nonlinear dynamics. To illustrate the capability of our GM-UKF
surements. The obtained test results are displayed in Fig. 9. to handle that case, we assume that the load at Bus 7 is increased
Note that in presence of Cauchy measurement noise, the UKF from 233.8 MW to 1500 MW to stress the system before switch-
and Huber-UKF have a non-positive definite covariance matrix, ing Line 15-16 while the other simulation settings are the same
resulting in their divergences. By contrast, thanks to the Gaus- as those of Case 1. Note that the steady-state maximum load-
sian normality and robustness of our GM-UKF, the total influ- ability at Bus 7 is around 2000 MW. After the line switching,
ence function-based covariance matrix updating approach can the system operates under even greater stressed conditions. The
always guarantee its positive-definiteness. On the other hand, test results are displayed in Fig. 10. It is observed from these
compared with the results obtained when using Laplacian power two figures that the GM-IEKF and UKF fail to converge at the
measurement noises, the GM-IEKF produces larger biases of the very beginning of the transient process while the Huber-UKF
state estimates and takes much longer time to approach the true diverges latter on. By contrast, our GM-UKF is able to handle

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 ZHAO AND MILI: ROBUST GM-UKF FOR POWER SYSTEM DYNAMIC STATE ESTIMATION 589

F. Breakdown Point of the GM-UKF to Cyber Attacks

To investigate the breakdown point of the GM-UKF to cyber
attacks, which is defined as the maximum number of outliers
that the filter can handle without yielding unreliable estimates,
we carry out extensive simulations on the IEEE 39-bus test
system using the concept of finite sample breakdown in nonlin-
ear regression introduced by Stromberg and Ruppert [37]. By
replacing a varying percentage of observations by outliers in
the vector y k , it is observed that the GM-UKF can handle at
least 25% of corrupted observations. It is worth noting that the
breakdown point of the GM-estimator in nonlinear regression
is still unknown. This problem will be investigated as a future
work. Another interesting problem is the determination of the
maximum breakdown point that any regression estimator may
have in structured nonlinear regression such as power system
state estimation problems; this will be an interesting extension
of the results proved in Mili and Coakley [38] in the linear
case.

G. Discussions About the Statistical Efficiency of the

GM-UKF
Let us now show that our proposed GM-UKF with λ = 1.5
has a high efficiency under various probability distributions of
the measurement noise as depicted in Fig. 1. This can be done
by analyzing the derived asymptotic variance matrix shown in
(32). Under Gaussian noise assumed for all the measurements
in the system, our filter exhibits a high statistical efficiency of
96.4%. Under the Laplace distribution for the real and reac-
Fig. 10. Case 5: Tracking performance of the GM-IEKF, the UKF, the Huber- tive power measurement noises while the measurement noises
UKF and the GM-UKF in the presence of strong system nonlinearity. (a) The
estimated rotor angle and speed, field voltage and mechanical power of Gener- of the voltage magnitudes and phase angles respectively obey
ator 5 are used for illustration purposes; (b) mean absolute error of each of the a Gaussian mixture and the Gaussian distribution as indicated
four filters. in Fig. 1, our filter exhibits an average efficiency of 96.4%
for 500 samples (10 s). These high efficiencies do not come
as a surprise since for λ tending to infinity or to zero, the GM-
estimator becomes a maximum likelihood estimator at the Gaus-
this scenario while achieving excellent tracking performance. sian or at the Laplace distribution, respectively. Finally, under
The underlying reasons are as follows:
r Under stressed system operating conditions, the first-order the Cauchy distribution of the real and reactive power measure-
ment noises, our GM-UKF has an average efficiency of 83.3%
Taylor series expansion used in the EKF and the GM- for 500 samples. A lower efficiency is expected here because
IEKF is too approximate and is unable to account for strong the Cauchy distribution has much thicker tails than the two oth-
system nonlinearities. As a result, these two filters produce ers. Recall that the Cauchy tails are so thick that no moment
very large approximation errors and eventually diverge;
r Thanks to the sigma-points-based unscented transforma- exists. Note that when the estimation error covariance matrix
is updated, all the outliers with respect to the Gaussian dis-
tion and its approximation accuracy up to at least third- tribution, which include the measurements associated with the
order Taylor series expansion, the UKF, Huber-UKF and tails of the Laplace or the Cauchy distribution or other thick-
the GM-UKF are able to handle strong system nonlinear- tailed distributions, will be heavily downweighted through the
ities. However, due to the accumulative estimation error matrix Q , yielding bounded biases and variances in the state
induced by the non-Gaussian measurement error, the esti- estimates.
mation error covariance matrices of the UKF and Huber-
UKF are close to non-positive definitiveness. Therefore,
they produce large estimation errors and finally diverge. H. Computational Efficiency
By contrast, our GM-UKF leverages the strength of the To validate the applicability of the proposed GM-UKF to on-
unscented transformation to handle system nonlinearities line estimation with a PMU sampling rate of 30 or 60 samples per
while the robustness of the GM-estimator allows it to filter second, its computational efficiency is analyzed and compared
out thick-tailed non-Gaussian measurement noise, yielding to that of the UKF, the Huber-UKF and the GM-IEKF in Cases
good estimation results. 1–5. The test is performed on a PC with Intel Core i5, 2.50 GHz,

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 590 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 12, NO. 4, AUGUST 2018

TABLE I where ζ i = γ i − (Aχi + b). By taking the derivative of the

AVERAGE COMPUTING TIMES OF THE FOUR DSE METHODS FOR EVERY PMU
SAMPLE, WHERE NA REPRESENTS NOT APPLICABLE objective function with respect to A and b and let them equal to
zero, respectively, we obtain


b = y − Ax, (48)
 = P T P −1 ,
A (49)
xy xx

where x = 2n i=1 wi χi ; y =
2n
i=1 wi g (χi ) =
2n
i=1 wi γ i ;
2n T 2n
P xx = i=1 wi (χi − x)(χi − x) ; P xy = i=1 wi (χi −
8 GB of RAM. The average computing time of each method for
x)(γ i − y)T . Then, the estimation error covariance matrix is
every PMU sample is displayed in Table I. We observe from this
calculated as
table that the UKF has the best computational efficiency, exhibit-
ing computing times much lower than the PMU sampling period,
which are 33.3 ms and 16.7 ms for 30 sample/s and 60 samples/s, 2n

wi ζi ζi
T
respectively. Although the execution times of the GM-UKF are a P ζζ =
little bit longer, they are still smaller than the PMU sampling pe- i=1
riod, demonstrating its ability to track system real-time dynamic 2n
   T
states. =  (χi −x) γ i −y − A
wi γ i −y− A  (χi − x)
i=1

VI. CONCLUSION = P y y − AP T = P y y − P T P −1 P xy ,
 xx A (50)
xy xx
In this paper, the robust GM-UKF is proposed to handle
non-Gaussian noise and outliers. We show first that the UKF
estimates the state vector via a weighted least squares esti- where P y y = 2n T
i=1 wi (γ i − y)(γ i − y) . Now, by taking the
mator under the Gaussianity assumption of the system pro- expectation and the outer product of the statistical linearized
cess or measurement noises; consequently, it yields strongly model, respectively, we obtain the posterior statistics given by
biased state estimates when the noises follow non-Gaussian
probability distributions, which is precisely the case when pro- 2n 2n
cessing PMU measurements. By contrast, the state estimates  
 +
y = Ax  =
wi χi − Ax wi χi , (51)
and residuals of our GM-UKF are proved to be asymptoti-
i=1 i=1
cally Gaussian, allowing the sigma points to reliably approx-
imate the mean and the covariance matrices of the predicted  xx A
P y y = AP T + P ζ ζ
and corrected state vectors. Furthermore, by relying on the 2n

projection statistics and the GM-estimator, the proposed GM- = P Txy P −1
xx P xy + wi (γ i − y)(γ i − y)T
UKF is able to suppress observation and innovation outliers i=1
while exhibiting high statistical efficiency of the state esti-
mates. In addition, we derive the expression of the asymp- − P Txy P −1
xx P xy
totic error covariance matrix of the GM-UKF state estimates 2n

from the total influence function of the GM-estimator. Numer- = wi (γ i − y)(γ i − y)T , (52)
ical results carried out on the IEEE 39-bus 10-machine system i=1
demonstrate the effectiveness and robustness of the proposed
method.
which are the same expressions as those obtained by apply-
ing the unscented transformation to the nonlinear function
APPENDIX A y = g(x), which completes the proof. 
PROOF OF THEOREM 1
Proof: Consider a nonlinear function y = g(x) evaluated in
2n points, i.e., (χi , γ i ), where γ i = g (χi ) for i = 1, . . . , 2n. APPENDIX B
Assuming that the nonlinear function is statistically linearized PROOF OF THEOREM 3
as y = Ax + b + ζ, the objective is to find A  and 
b so that the
Proof: By taking a von Mises expansion of the functional
point-wise linearization error ζ i is minimized, i.e.,
form of the estimator T with respect to Φ, we get
  2n

 
A, b = arg min wi ζ i T ζ i , (47)
i=1 T (Φα ) = T (Φ) + T  (Φα − Φ) + Rem(Φα − Φ), (53)

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 ZHAO AND MILI: ROBUST GM-UKF FOR POWER SYSTEM DYNAMIC STATE ESTIMATION 591

which
√ can be reorganized into the following form by multiplying APPENDIX D
α on both sides of the equality: PROOF OF THEOREM 4
√ Proof: By virtue of the law of large numbers, the distri-
α (T (Φα ) − T (Φ))
√ √ bution of the residuals tends to the Gaussian distribution, i.e.,
= αT  (Φα − Φ) + αRem (Φα − Φ) (54) Φ ∼ N (μ, σ 2 ). Since the median absolute deviation (MAD =
 s
√ √ 1.4826·b m ) is a consistent estimator for the standard deviation σ
= α IF (x; Φ, T )d(Φα − Φ) + αRem(Φα − Φ) of a Gaussian distribution, we get
 
(55) 1 s
 = P |X − μ| ≤
√ √ 2 1.4826 · bm
= α IF (x; Φ, T )dΦα + αRem(Φα − Φ) (56)  
|X − μ| s
=P ≤
1 
α
√ σ 1.4826 · bm · σ
=√ IF (xi ; Φ, T ) + αRem (Φα − Φ) , (57)  
α i=1 s
= 2Φ − 1. (63)
1.4826 · bm · σ
where the definition of the influence function is applied to yield  
(38) to (39).
 By virtues of Fisher consistency at the distribution Therefore, we obtain s/σ = 1.4826 · bm · Φ−1 34 = bm → 1
Φ, that is, IF (x; Φ, T )dΦ = 0, (39) reduces to (40). Finally, as m tends to infinity, where Φ is the cumulative proba-
by using the property of the empirical cumulative probability bility function of the standard Gaussian distribution. On the
distribution function, we have other hand, from (21) and E[ξ k ξ k T ] = I, the residuals can be
 shown to follow the standard Gaussian distribution. Therefore,
1
α
IF (x; Φ, T )dΦα = IF (xi ; Φ, T ), (58) EF [s2 ] = s2 → 1. 
α i=1
ACKNOWLEDGMENT
yielding (40) to (41).
Following the work of Fernholz [32], we can show that The authors would like to thank Dr. Z. Huang from PNNL
√ p
for providing us with real PMU data for analyzing the statistical
αRem (Φα − Φ) → 0, (59) probability distributions of the PMU measurement errors.
p
where → means probability convergence. Therefore, by apply- REFERENCES
ing the central limit theorem and Slutsky’s lemma to (57), it
[1] Z. Wang and Q. Huang, “A closed normal form solution under near-
follows that resonant modal interaction in power systems,” IEEE Trans. Power Syst.,
√ d vol. 32, no. 6, pp. 4570–4578, Nov. 2017.
α(T (Φα ) − T (Φ)) → N (0, P xx
k |k ), (60) [2] I. Kamwa, R. Grondin, and Y. Hebert, “Wide-area measurement based
stabilizing control of large power systems—A decentralized/hierarchical
k |k = E[IF (x; Φ, T ) · IF (x; Φ, T ) ].
where P xx T
 approach,” IEEE Trans. Power Syst., vol. 16, no. 1, pp. 136–153, Feb. 2001.
[3] H. Ni, G. T. Heydt, and L. Mili, “Power system stability using robust wide
area control,” IEEE Trans. Power Syst., vol. 17, no. 4, pp. 1123–1131,
APPENDIX C Nov. 2002.
PROOF OF COROLLARY 3.2 [4] E. Ghahremani and I. Kamwa, “Dynamic state estimation in power system
by applying the extended Kalman filter with unknown inputs to phasor
Proof: In our previous work [7], the total influence function measurements,” IEEE Trans. Power Syst., vol. 26, no. 4, pp. 2556–2566,
Nov. 2011.
of a GM-estimator based on a nonlinear regression model given [5] Z. Huang, K. Schneider, and J. Nieplocha, “Feasibility studies of applying
by y = ϕ(x) + ξ is expressed as Kalman filter techniques to power system dynamic state estimation,” in
Proc. 8th Int. Power Eng. Conf., Singapore, Dec. 2007, pp. 376–382.
IF (x; Φ, T ) [6] L. Fan and Y. Wehbe, “Extended Kalman filtering based real-time dynamic
state and parameter estimation using PMU data,” Elect. Power Syst. Res.,
    −1 vol. 103, pp. 168–177, 2013.
ψ (r S i ) ∂ϕ (x) ∂ϕ (x)T 
= −ψ (r S i ) D T (Φ) dΦ [7] J. B. Zhao, M. Netto, and L. Mili, “A robust iterated extended Kalman
filter for power system dynamic state estimation,” IEEE Trans. Power
s ∂x ∂xT
Syst., vol. 32, no. 4, pp. 3205–3216, 2017.
[8] A. K. Singh and B. C. Pal, “Decentralized dynamic state estimation in
∂ϕ (x)
· ψ (r S i ) , (61) power systems using unscented transformation,” IEEE Trans. Power Syst.,
∂x vol. 29, no. 2, pp. 794–804, 2014.
2
[9] S. Wang, W. Gao, and A. P. S. Meliopoulos, “An alternative method for
where D = ∂∂ xϕ(x)
i ∂ xj
is the Hessian matrix of ϕ(x). Since we power system dynamic state estimation based on unscented transform,”
IEEE Trans. Power Syst., vol. 27, no. 2, pp. 942–950, 2012.
have ϕ(x) = Cx for the GM-UKF, (61) reduces to [10] A. Rouhani and A. Abur, “Linear phasor estimator assisted dynamic state
  −1 estimation,” IEEE Trans. Smart Grid, vol. 9, no. 1, pp. 211–219, Jan. 2018.
1  T  [11] Z. Huang et al., “Capturing real-time power system dynamics: Opportu-
IF (x; Φ, T ) = ψ (rS i ) CC T ( Φ ) dΦ Cψ (rS i ) . nities and challenges,” in Proc. IEEE Power Eng. Soc. General Meeting,
s Jul. 2015, pp. 1–5.
(62) [12] S. Wang, J. B. Zhao, Z. Huang, and R. Diao, “Assessing Gaussian assump-
tion of PMU measurement error using field data,” IEEE Trans. Power Del.,
 to be published.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.
 592 IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, VOL. 12, NO. 4, AUGUST 2018

[13] M. Gandhi and L. Mili, “Robust Kalman filter based on a general- [35] IEEE PES TF on Benchmark System for Stability Controls, “Benchmark
ized maximum-likelihood-type estimator,” IEEE Trans. Signal Process., systems for small-signal stability analysis and control,” IEEE PES, Pis-
vol. 58, no. 5, pp. 2509–2520, May 2010. cataway, NJ, USA, Tech. Rep. PES-TR18, Aug. 2015.
[14] J. B. Zhao, S. Wang, L. Mili, R. Huang, and Z. Huang, “A robust state es- [36] “Modeling of gas turbines and steam turbines in combined cycle power
timation framework considering measurement correlations and imperfect plants,” CIGRE Rep., Paris, France, Task Force C4.02.25, 2003.
synchronization,” IEEE Trans. Power Syst., to be published. [37] A. J. Stromberg and D. Ruppert, “Breakdown in nonlinear regression,” J.
[15] K. D. Jones, A. Pal, and J. S. Thorp, “Methodology for performing syn- Amer. Statist. Assoc., vol. 87, no. 420, pp. 991–997, 1992.
chrophasor data conditioning and validation,” IEEE Trans. Power Syst., [38] L. Mili and C. W. Coakley, “Robust estimation in structured linear regres-
vol. 30, no. 3, pp. 1121–1130, May 2015. sion,” Ann. Statis., vol. 24, no. 6, pp. 2593–2607, 1996.
[16] D. N. Kosterev, C. W. Taylor, and W. A. Mittelstadt, “Model validation
for the August 10, 1996 WSCC system outage,” IEEE Trans. Power Syst.,
vol. 14, no. 3, pp. 967–979, Aug. 1999.
[17] Z. Huang, P. Du, D. Kosterev, and S. Yang, “Generator dynamic model
validation and parameter calibration using phasor measurements at the
point of connection,” IEEE Trans. Power Syst., vol. 28, no. 2, pp. 1939– Junbo Zhao (M’18) received the Bachelor’s degree
1949, May 2013. in electrical engineering from Southwest Jiaotong
[18] J. B. Zhao and L. Mili, “Power system robust decentralized dynamic state University, Chengdu, China, in 2012, and the Ph.D.
estimation based on multiple hypothesis testing,” IEEE Trans. Power Syst., degree from the Bradley Department of Electrical and
to be published. DOI: 10.1109/TPWRS.2017.2785344. Computer Engineering Virginia Polytechnic Institute
[19] M. A. M. Ariff, B. C. Pal, and A. K. Singh, “Estimating dynamic model and State University, in 2018. He did the summer
parameters for adaptive protection and control in power system,” IEEE internship at Pacific Northwest National Laboratory
Trans. Power Syst., vol. 30, no. 2, pp. 829–839, Mar. 2015. from May to August, 2017. He is now the Chair of
[20] N. Zhou, D. Meng, Z. Huang, and G. Welch, “Dynamic state estimation IEEE Task Force on Power System Dynamic State
of a synchronous machine using PMU data: A comparative study,” IEEE and Parameter Estimation, the secretary of the IEEE
Trans. Smart Grid., vol. 6, no. 1, pp. 450–460, Jan. 2015. Working Group on State Estimation Algorithms and
[21] G. Evensen, “Sequential data assimilation with a nonlinear quasi- the IEEE Task Force on Synchrophasor Applications in Power System Opera-
geostrophic model using Monte Carlo methods to forecast error statistics,” tion and Control. He has authored 2 book chapters, authored or coauthored more
J. Geophys. Res., vol. 99, no. 5, pp. 143–162, May 1994. than 40 peer-reviewed journal and conference papers, and 9 Chinese patents.
[22] X. Wang, N. Cui, and J. Guo, “Huber-based unscented filtering and its His research interests are power system real-time monitoring, operations,
application to vision-based relative navigation,” IET Radar, Sonar Navig., and cyber security that include power system state estimation, power system
vol. 4, no. 1, pp. 134–141, 2010. dynamics and stability, static and dynamic load modeling, power system cy-
[23] L. Y. Chang et al., “Huber-based novel robust unscented Kalman filter,” ber attacks and countermeasures, big data analytics, and robust statistics with
IET Sci., Meas. Technol., vol. 6, no. 6, pp. 502–509, 2012. applications in the smart grid.
[24] L. B. Chang and K. Li, “Unified form for the robust Gaussian information
filtering based on M-estimate,” IEEE Signal Process. Lett., vol. 24, no. 4,
pp. 412–416, Apr. 2017.
[25] W. Li, S. Sun, Y. Jia, and J. Du, “Robust unscented Kalman filter with
adaptation of process and measurement noise covariances,” Digital Signal
Process., vol. 48, pp. 93–103, 2016.
[26] J. B. Zhao and L. Mili, “Robust unscented Kalman filter for power system Lamine Mili (LF’17) received the Diploma in elec-
dynamic state estimation with unknown noise statistics,” IEEE Trans. trical engineering from the Swiss Federal Institute of
Smart Grid, to be published. DOI: 10.1109/TSG.2017.2761452. Technology, Lausanne, Switzerland, in 1976, and the
[27] S. Julier, J. Uhlmann, and H. F. Durrant-Whyte, “A new method for the Ph.D. degree from the University of Liège, Liège,
nonlinear transformation of means and covariances in filters and estima- Belgium, in 1987. He is currently a Professor with
tors,” IEEE Trans. Autom. Control, vol. 45, no. 3, pp. 477–482, Mar. 2000. the Department of Electrical and Computer Engineer-
[28] T. Lefebvre, H. Bruyninckx, and J. De Schutter, “Comment on ‘a new ing, Virginia Tech, Northern Virginia Center, Falls
method for the nonlinear transformation of means and covariances in filters Church, VA, USA. He has five years of industrial ex-
and estimators’,” IEEE Trans. Autom. Control, vol. 47, no. 8, pp. 1406– perience with the Tunisian electric utility, STEG. At
1408, Aug. 2002. STEG, he worked in the planning department from
[29] L. Mili, M. Cheniae, N. Vichare, and P. Rousseeuw, “Robust state esti- 1976 to 1979 and then at the Test and Meter Labo-
mation based on projection statistics,” IEEE Trans. Power Syst., vol. 11, ratory from 1979 to 1981. He was a Visiting Professor with the Swiss Federal
no. 2, pp. 1118–1127, May 1996. Institute of Technology, Lausanne, the Grenoble Institute of Technology, the
[30] P. J. Huber, Robust Statistics. New York, NY, USA: Wiley, 1981. École Supérieure D’électricité, France and the École Polytechnique de Tunisie
[31] L. Thomas and L. Mili, “A robust GM-estimator for the automated detec- in Tunisia, and did consulting work for the French Power Transmission com-
tion of external defects on barked hardwood logs and stems,” IEEE Trans. pany, RTE.
Signal Process., vol. 55, no. 7, pp. 3568–3576, Jul. 2007. His research has focused on power system planning for enhanced resiliency
[32] L. Fernholz, Von Mises Calculus for Statistical Functionals (Lecture Notes and sustainability, risk management of complex systems to catastrophic failures,
in Statistics 19). New York, NY, USA: Springer-Verlag, 1983. robust estimation and control, nonlinear dynamics, and bifurcation theory. He
[33] F. R. Hampel, E. M. Ronchetti, P. J. Rousseeuw, and W. A. Stahel, Robust is the co-founder and Co-Editor of the International Journal of Critical Infras-
Statistics: The Approach Based on Influence Functions. New York, NY, tructure. He is the Chairman of the IEEE Working Group on State Estimation
USA: Wiley, 1986. Algorithms. He was a recipient of several awards including the US National
[34] P. W. Sauer and M. A. Pai, Power System Dynamics and Stability. Engle- Science Foundation (NSF) Research Initiation Award and the NSF Young In-
wood Cliffs, NJ, USA: Prentice-Hall, 1998. vestigation Award.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY ROORKEE. Downloaded on December 30,2024 at 12:26:55 UTC from IEEE Xplore. Restrictions apply.