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

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2018.2870327, IEEE
Transactions on Smart Grid

IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2018 1

A Decentralized H-infinity Unscented Kalman Filter

for Dynamic State Estimation Against Uncertainties
Junbo Zhao, Member, IEEE, Lamine Mili, Life Fellow, IEEE

Abstract—The widely used traditional Kalman filter-type Xd , Xq d-axis and q-axis transient reactance in p.u.
power system dynamic state estimator (DSE) is unable to address V Terminal bus voltage magnitude in p.u.
the unknown and non-Gaussian system process and measure- θ Terminal bus voltage angle in rad
ment noise as well as dynamical model uncertainties. To this
end, this paper proposes a decentralized H-infinity unscented I Terminal bus current magnitude in p.u.
Kalman filter that leverages the strength of the H-infinity criteria φ Terminal bus current angle in rad
developed in robust control for handling system uncertainties f Vector-valued differential equations
with the advantage of the UKF for addressing strong model h Vector-valued measurement equations
nonlinearities. Specifically, the statistical linerization approach is w, v System process and measurement noise vec-
used to derive a linear-like batch-mode regression model similar
to the linear Kalman filter. This allows us to resort to the linear tors, respectively
H-infinity Kalman filter framework for the development of the u System input vector
proposed H-infinity UKF in the Krein space. An analytical form Q, R System process and measurement noise co-
is also derived to tune the parameter of the H-infinity criterion. variance matrices, respectively
Two decoupled models are presented to enable the decentralized χ Sigma points
implementation of the H-infinity UKF using the local PMU
measurements. Extensive simulation results carried out on the z Measurement vector
IEEE 39-bus system show that the proposed H-infinity UKF is F,H Statistical linerization matrices of the state
able to bound the influences of various types of measurement and prediction and the measurement models, re-
model uncertainties while obtaining accurate state estimates. spectively
Index Terms—Dynamic state estimation, decentralized estima- γ Tuning parameter
tion, model uncertainties, unscented Kalman filter, non-Gaussian
w Weights of the sigma points
noise, H-infinity filter, extended Kalman filter, robustness.
n Number of state variables
N OMENCLATURE
ω, ωs Rotor speed and its base value in rad/s I. I NTRODUCTION
δ Rotor angle in rad
Ed , Eq
Ef d
d-axis and q-axis transient voltages in p.u.
Generator excitation field voltage in p.u.
R EAL-time dynamic states of synchronous generators are
of vital importance for power system monitoring, fault
location, protection and control [1]–[4]. With the wide-area
VF , VR Scaled output of the stabilizing transformer deployment of phasor measurement units (PMUs), the devel-
and the amplifier in p.u. opment of dynamic state estimator (DSE) becomes feasible.
TM , PSV Mechanic torque and steam valve position in To date, various Kalman filter-type DSEs have been pro-
p.u. posed in the literature, including the extended Kalman filter
 
Tdo , Tqo Generator stator winding time constants in s (EKF) and the unscented Kalman filter, among others [ 5]–[14].
TE , TF , TA Excitor time constants in s They can be grouped into centralized and decentralized ones.
TCH , TSV Governor time constants in s The centralized methods are based on a global DSE, which
KE , KF , KA Excitor control gains processes information collected from all the system devices
Vref Reference voltage in p.u. and the real-time PMU measurements [5]–[9]. Since all the
H Generator inertia in p.u. measurements are communicated to the control center, a good
D Generator damping constant in p.u. global measurement redundancy is achieved, which in turn
PC Control input of the governor in p.u. enhances the capability of the DSE to handle measurement
RD Governor speed regulation constant in p.u. losses and outliers. However, for large-scale interconnected
Id , Iq d-axis and q-axis stator currents in p.u. power systems, a global DSE may be incompatible with real-
Vd , Vq d-axis and q-axis stator voltage magnitudes time applications due to prohibitively large executing times.
in p.u. By contrast, decentralized methods are implemented at each
Xd , X q d-axis and q-axis synchronous reactance in local synchronous generator, and therefore are independent of
p.u. system size. Consequently, they are suitable for large-scale
This work is supported by the U.S. National Science Foundation under power system online monitoring. In [10], [11], a decentralized
Grant ECCS-1711191 and the Advanced Grid Modeling program of the U.S. DSE based on EKF is proposed to simultaneously estimate
Department of Energy Office of Electricity Delivery & Energy Reliability. the states and the parameters of a classical generator model.
Junbo Zhao and Lamine Mili are with the Bradley Department of Electrical
and Computer Engineering, Virginia Polytechnic Institute and State University, In [12], a decentralized DSE based on UKF is advocated
Falls Church, VA 22043, USA (e-mail: [email protected], [email protected]). to overcome the weaknesses of the EKF in handling strong

1949-3053 (c) 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2018.2870327, IEEE
Transactions on Smart Grid

IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2018 2

nonlinearity of the dynamical model. In [ 13], a fourth-order • It derives a linear like batch-regression form that pre-
generator model is considered and the mechanical torque and serves all the properties of the UKF when casted into the
field voltage of the exciter are taken as system inputs when H-infinity framework. We then turn to the Krein space
implementing the decentralized EKF-based DSE. Anagnostou Kalman filter framework for the development of the H-
and Pal [14] improved that approach by proposing a derivative infinity UKF; it is interesting that this is the first attempt
free UKF-based DSE, where the unknown mechanical torque to derive nonlinear H-infinity UKF in the Krein space;
and field voltage of the exciter are estimated along with the • It leverages two ways for the decentralized implementa-
generator state variables. [12]–[14] showed that their methods tion using local PMU measurements; their effectiveness
are sensitive to generator parameter uncertainties and large are compared;
measurement noise. However, they did not propose the related • It has better robustness to unknown process and measure-
mitigation approaches. ment noise and uncertain generator model and controller
It turns out that the aforementioned decentralized DSEs parameter values as compared to other approaches, in-
do not address the cases of unknown or thick-tailed process cluding the H-infinity EKF [22], the traditional decen-
and measurement noises as well as model uncertainties. In tralized UKF [12] and the robust EKF [7].
practice, due to unknown disturbances and changes of system The remainder of the paper is organized as follows. Section
inputs, the assumed dynamical model may not properly predict II presents the dynamical model of synchronous generator and
the dynamical behavior of a power system, yielding unknown the problem statement. Section III describes the proposed de-
process noise. The PMU measurement noise actually follows centralized H-infinity UKF and Section IV shows and analyzes
a non-Gaussian distribution as validated in [15], [16]. Due the simulation results. Finally, Section V concludes the paper.
to magnetic saturation, changing temperature of the machine
windings and aging process, the generator transient reactance
II. P ROBLEM F ORMULATION
and other parameters can be time-varying, causing large model
uncertainties. Under these conditions, the performances of A. Synchronous Generator Modeling
Kalman filter-type DSEs may significantly degrade, resulting In this paper, the two-axis generator model with the IEEE-
in unreliable state estimates [17]. Note that the robustness of DC1A exciter and the TGOV1 turbine-governor is considered.
a DSE to model and measurement uncertainties is mandatory Note that this is one of the recommended models by IEEE for
to justify its values for practical applications. transient stability analysis [24]. It consists of the following
To handle a variety of system uncertainties, we resort to continuous time differential and algebraic equations (DAEs)
robust control theory and develop a decentralized H-infinity [25]:
UKF. Our choice of an H-infinity filter is motivated by the fact Differential equations:
that it has a proved performance in a variety of engineering
applications, such as designing controllers to address measure- 
dEq
Tdo = −Eq − (Xd − Xd ) Id + Ef d , (1)
ment delays, unknown inputs and unknown noise statistics, dt
among others [17]–[20]. Unlike the Kalman filter that mini-
 dEd
  
mizes the mean-square error, the H-infinity filter minimizes the Tqo = −Ed − Xq − Xq Iq , (2)
dt
maximum variance of the state estimates [17]. To overcome the
dδ
fact that the H-infinity filter is designed for linear dynamical = ω − ωs , (3)
system and thus may not be effective for power systems, dt
first-order Taylor series approximation-based H-infinity EKF 2H dω
is proposed [21], [22]. However, a power system may exhibit = TM − Pe − D (ω − ωs ) , (4)
ωs dt
strong nonlinearity on the dynamical model when it is heavily
dEf d
loaded or subject to large disturbances. As a result, first- TE = − (KE + SE (Ef d )) Ef d + VR , (5)
dt
order Taylor series approximation will induce large estimation
errors, which may cause a divergence of the EKF and the H- dVF KF KF
TF = −VF + VR − (KE + SE (Ef d )) Ef d , (6)
infinity EKF [23]. dt TE TE
The proposed decentralized H-infinity UKF has the follow- dVR
ing salient features: TA = −VR + KA (Vref − VF − V ) , (7)
dt
• It stems from the conversion of the traditional nonlinear
dTM
UKF into an equivalent linear like batch-mode regression TCH = −TM + PSV , (8)
form through the use of the statistical linerization ap- dt
 
proach. Note that the latter does not involve any Jacobian dPSV 1 ω
TSV = −PSV + PC − −1 , (9)
matrix while being equivalent to the unscented trans- dt RD ωs
formation; therefore, it is able to handle strong system
Algebraic equations:
nonlinearities. By contrast, the first-order Taylor series
approximation based H-infinity EKF involves cumber- Vd = V sin (δ − θ) , Vq = V cos (δ − θ) , (10)
some linerization through the calculation of Jacobian
matrices, which results in significant biased results in Eq − Vq Vd − Ed
Id =  , Iq = . (11)
presence of strong nonlinearity; Xd Xq

1949-3053 (c) 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2018.2870327, IEEE
Transactions on Smart Grid

IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2018 3

The discrete-time state space representation of the above so, a synchronous generator can be modeled as an isolated en-
DAEs can be expressed as tity from a DSE prospective, without modeling the remaining
part of the system to which it is interconnected. Depending
xk = f (xk−1 , uk ) + wk , (12) on the way of treating the voltage and current phasors, two
zk = h (xk ) + vk , (13) model decoupling approaches can be implemented. They are
presented below as Approach A1 and Approach A2. Note that
where xk is the state vector that includes generator state the i-th generator is used as an example and the q-axis is
variables ω, δ, Ed and Eq , exciter state variables Ef d , VF assumed to lead d-axis.
and VR , and governor state variables T M and PSV ; f (·) Approach A1: Consider a synchronous generator where the
is the nonlinear state transition function corresponding to stator current phasor I i ∠φi is supposed to be a known input,
the equations (1)-(9); uk is the system input vector; z k is i.e., ui = [Ii φi ]T , and terminal bus voltage phasor V i ∠θi
the measurement vector, including metered generator terminal and frequency f i as outputs. Consequently, we have I di =
π π
voltage or current phasors and frequency; h(·) is the vector- real[Ii ej(φi −δi + 2 ) ], Iqi = Imag[Ii ej(φi −δi + 2 ) ], Vdi = Edi

+
  
valued measurement function corresponding to the equations Iqi Xqi , and Vqi = Eqi − Idi Xdi . Furthermore, we have z i =
(10)-(11); wk and vk are system process and measurement [z1 z2 z3 ]T and their expressions can be shown as follows:
noises with covariance matrices Q k and Rk , respectively. 
z1 = Vi = Vdi2 + Vqi2 + vV i , (14)
B. Problem Statement
Vdi
To estimate the state variables of the synchronous generator z2 = θi = δi − arctan( ) + vθi , (15)
Vqi
using (12)-(13), several decentralized EKF and UKF-based
z3 = fi = f0 (Δωi + 1) + vf i , (16)
approaches have been proposed in [ 10]–[14]. These Kalman
filter-type DSEs work well if the following conditions are where vi = [vV i vθi vf i ]T is the measurement noise with zero
satisfied [17]: i) the system process and observation noises are mean and covariance matrix R ik .
assumed to have zero mean and zero correlation at each time Approach A2: Consider a synchronous generator where
instant; ii) the covariance matrices Q k and Rk are known; the terminal bus voltage phasor, V i ∠θi , is supposed to be
iii) the discrete-time nonlinear state-space model is accurate. known input, i.e., u i = [Vi θi ]T , and current phasor I i ∠φi
However, these assumptions may be violated for practical and fi are considered as outputs. Consequently, we have,

systems. For instance, the disturbances and the system inputs Vdi = Vi sin(δi − θi ), Vqi = Vi cos(δi − θi ), Idi = (Eqi −
  
may be unknown under some operating conditions of the Vqi )/Xdi , and Iqi = (−Edi +Vdi )/Xqi . Furthermore, we have
system; the synchronization process may induce additional zi = [z1 z2 z3 ]T and
noise to PMU measurements, yielding unknown or even non- 
Gaussian measurement noise; the synchronous generator tran- z1 = Ii = Idi 2 + I2 + v , (17)
qi Ii
sient reactance can vary [26] due to the magnetic saturation of
z2 = φi = arg(Iqi − jIdi ) + δi + vφi , (18)
the iron core of the generator and to a changing temperature of
the machine windings; due to the aging process, the addition z3 = fi = f0 (Δωi + 1) + vf i , (19)
of new digital protective relaying packages, replacement of where vi = [vIi vφi vf i ]T is the measurement noise with zero
the old rotating pilot exciters and voltage regulators, and mean and covariance matrix R ik . It should be emphasized that
the controller failures, to name a few, controller parameters in both approaches, the rotor speed deviation Δω i can be either
can change from time to time [27], [28]. As a result, their obtained from the PMUs or calculated using the approach
performances can be significantly degraded. To bound the shown in [13]. The measurement or approximation error is
influence of these uncertainties on the state estimates while characterized by v f i and treated as one of the uncertainties.
handling strong nonlinear dynamical system model, we de-
Consequently, the discrete time state-space model of the ith
velop a decentralized H-infinity UKF as described next. generator at the time sample k is derived as
III. T HE P ROPOSED D ECENTRALIZED H- INFINITY UKF  
xik = fi xik−1 , uik + wki , (20)
This section presents two model decoupling approaches of  
zki = hi xik + vki , (21)
a synchronous generator that enable the decentralized imple-
mentation of a DSE. Statistical linearization approach is used where wki is the system process noise; the input vector u ik
to convert the traditional nonlinear UKF into an equivalent lin- that is composed of either voltage or current phasor contains
ear batch-mode regression form. The latter allows us to resort noise. As the model is nonlinear, the system process noise is
to the linear H-infinity filter framework for the development not additive and its covariance matrix Q ik becomes unknown.
of an H-infinity UKF in the Krein space. Remark: To implement the decentralized DSE, the local
system observability should be satisfied. In other words, the
A. Discrete-Time State Space Model for Decentralized DSE local system that contains more than one generator can be
Evidently, the dynamics of a synchronous generator can be observable by PMUs. If a generator terminal bus is not
accurately captured if one pair of the voltage or current phasors equipped with PMU, a PMU-based linear state estimator [8]
provided by PMUs are taken as inputs while the other pair of for that local system should be performed first. Then, the
measurements are considered as model outputs [29]. By doing estimated measurements at the generator terminal buses can be

1949-3053 (c) 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2018.2870327, IEEE
Transactions on Smart Grid

IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2018 4

2n

obtained and the decentralized DSE is executed using either

Pk|k−1
zz
= wj (zk|k−1
j
− zk|k−1 )(zk|k−1
j
− zk|k−1 )T + Rk ,
approach A1 or A2. Note that the local phasor data center
j=1
(PDC) is in charge of communicating and processing the (27)
PMU measurements. As the local system that contains the where zk|k−1
j
= h(χjk|k−1 ).
interested generators is usually of small-scale, the concern of In the work [23], [31], the unscented transformation has
communication issue is unlikely to happen. been shown to be equivalent to the statistical linerization. As
B. Derivation of the H-infinity UKF a result, the nonlinear regression model can be converted to its
equivalent linear regression form. Motivated by this property,
The model decoupling approaches allow us to estimate
we apply statistical linerization to the nonlinear state transition
the states of a generator through (20) and (21) using only
and measurement functions around x k−1|k−1 and x k|k−1 ,
local PMU measurements. As elaborated before, the derived
respectively, yielding 
discrete-time state-space model is subject to uncertainties,
xk = Fk xk−1 − x k−1|k−1 + x k|k−1 + ek + wk (28)
which will affect significantly the performance of the tradi-
tional UKF. To overcome this difficulty, we develop an H-  
infinity UKF that bounds the influence of the uncertainties on k|k−1 + zk|k−1 + εk + vk
zk = Hk xk − x (29)
−1
the state estimates, and thereby provides accurate generator where Fk = (Pk|k−1 ) (Pk−1|k−1 ) ;
xχ T xx
state estimates. The proposed H-infinity UKF consists of
2n
two key steps, namely the derivation of an equivalent linear xχ
Pk|k−1 = k−1|k−1 )(χjk|k−1 − x
wj (χjk−1|k−1 − x k|k−1 )T ,
regression form of the UKF and of the state filtering using the j=1
H-infinity criterion. For the sake of simplicity, in the sequel (30)
the index i associated with the ith synchronous generator is and ek is the statistical linearization error term on
neglected. the nonlinear state transition function; it is normally
1) Derivation of the Equivalent Linear Regression Form of distributed with zero mean and covariance matrix
the Nonlinear UKF: The main idea of the UKF is to rely Lk = Pk|k−1
xx
− (Pk|k−1
xχ
)T (Pk−1|k−1
xx
)−1 Pk|k−1
xχ
;
−1
on the unscented transformation for nonlinear state transition Hk = (Pk|k−1 ) (Pk|k−1 ) ;
xz T xx

and measurement prediction by strategically selecting sigma 2n

points within the Kalman filter framework. The use of the Pk|k−1
xz
= k|k−1 )(zk|k−1
wj (χjk|k−1 − x j
− zk|k−1 )T ,
sigma points for statistical information propagation avoids the i=1
linearization of the nonlinear state transition and measurement (31)
functions, and thus yields accurate state estimates [30]. and εk is the statistical linearization error term on the
Given the state estimate with mean x  k−1|k−1 ∈ Rn×1 and nonlinear measurement function; it is normally distributed
covariance matrix P k−1|k−1 at time step k-1, 2n weighted
xx with zero mean and covariance matrix Σ k = Pk|k−1 zz
−
−1 xz
sigma points are generated through (Pk|k−1 ) (Pk|k−1 ) Pk|k−1 . Then, the nonlinear dynamical
xz T xx

 model expressed by (20) and (21) can be converted into the

χjk−1|k−1 = xk−1|k−1 ± xx
nPk−1|k−1 , (22) following equivalent linear-like form:
j

where the weights wj = 1/2n, j = 1, ..., 2n. These sigma xk = Fk xk−1 + x k−1|k−1 + ek + wk ,
k|k−1 − Fk x (32)
points are propagated through the nonlinear system process  k|k−1 + εk + vk .
zk = Hk xk + zk|k−1 − Hk x (33)
model (20) to obtain the following transformed sigma points
 Note that Fk and Hk are no longer the Jacobian matrices.
χjk|k−1 = f χjk−1|k−1 . (23) In addition, by applying the linear Kalman filter to ( 32) and
(33) will yield the same results as the traditional UKF. This
k|k−1 and its covariance
Then, the predicted state vector x
is a significant result that allows us to resort to the linear
matrix are approximated by the weighted sample mean and
H-infinity Kalman filter framework for the development of our
sample covariance matrix of the transformed sigma points,
H-infinity UKF in the Krein space.
respectively, yielding
2n

2) Derivation of the H-infinity UKF: We aim to design a
k|k−1 =
x wj χjk|k−1 , (24) filter that guarantees a finite upper bound on the estimation
j=1 error [17]. If the metered sequence values are provided up to
2n

time N , the foregoing objective can be achieved by satisfying
Pk|k−1
xx
= k|k−1 )(χjk|k−1 − x
wj (χk|k−1 − x
j
k|k−1 )T +Qk .
the following inequality:
j=1
N  2
(25) xk − x
k|k P −1
k=0
Next, the sigma points are updated to capture the information sup  2 N 
k|k
 ≤ γ2,
of system process noise through  {x0 ,vk ,wk } x0 − x
0|0  + k=0 wk Q−1 + vk 2R−1
2
−1
P0|0 k k
k|k−1 ±
χjk|k−1 = x xx
nPk|k−1 . (26)
j where x0 and P0|0 are the initial state vector and its covariance
matrix, respectively; γ is a positive scalar parameter that
Then, the predicted measurement vector is given by zk|k−1 =

2n bounds the uncertainties.
wj h(χjk|k−1 ) and its associated error covariance matrix is It is worth pointing out that the solution of the objective
j=1 function can be shown to be equivalent to that of the Krein

1949-3053 (c) 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2018.2870327, IEEE
Transactions on Smart Grid

IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2018 5

space Kalman filter [32]. The first step is to convert the Remark: the system model is always with uncertainties and
suboptimal H-infinity filtering problem to an indefinite form the measurement noise statistics may change from time to
so that the Krein space Kalman filter can be used. Formally, time, yielding unknown and time-varying Q and R. Indeed,
we get the input vector that is composed of either voltage or current
phasor contains noise. As the model is nonlinear, the system
 2
N
N

J∞ = x 0 − x 
0|0 P −1 + 2
wk Q−1 +
2
vk R−1 process noise is not additive and its covariance matrix Q
k k
0|0
k=0 k=0 becomes unknown and time-varying. It is thus difficult to

N tune them for the traditional UKF in a real-time manner. By
− γ −2 xk − x k|k 2P −1 contrast, our proposed H-infinity UKF adopts the min-max
k|k
k=0 criteria to bound these uncertainties and does not require the

N −1 online tuning of them. Note that the non-additive noise can be

2
= x0 − x 0|0 2P −1 + wk Q−1 regarded as a source of model uncertainty and suppressed by
0|0 k
k=0 the H-infinity UKF.

N T −1 
zk − h(xk ) Rk 0 zk − h(xk ) C. Algorithm Implementation and Discussions
+
xk − x k|k 0 −γ 2 I xk − x k|k
k=0 The detailed step by step implementation of the proposed
(34) approach is summarized as follows:
• Step 1: Model synchronous generators with given param-
By substituting the derived linear-like regression form ( 32)
and (33) into (34) and following the steps of the Krein space eter values;
• Step 2: Obtain the discrete time state-space form repre-
Kalman filter [32], we can derive our H-infinity UKF as
follows: sentation of synchronous generator and derive the decen-
  tralized DSE implementation model using either (14)-(16)
k|k = x
x k|k−1 + K k zk − zk|k−1 , (35) or (17)-(19);
K k = Pk|k−1
xx
HkT (Hk Pk|k−1xx
HkT + Rk )−1 , (36) • Step 3: Generate sigma points and obtain the linear-like
regression model (32)-(33) using statistical linearization
= Pk|k−1
xz
(Pk|k−1
zz
)−1 (37)
approach;
−1 • Step 4: Perform H-infinity UKF using equations (35)-
Pk|k = (I − Pk|k−1 [Hk I]Re,k [Hk I] )Pk|k−1 ,
xx xx T T T xx
(38)
(41), where the parameter γ 2 is tuned using (44);
−1
= Pk|k−1 − [Pk|k−1 Pk|k−1 ]Re,k [Pk|k−1 Pk|k−1 ] ,
xx xz xx xz xx T
• Step 5: the steps 1-4 are recursively implemented for the
(39) next time instant.
 
T T
Rk + Hk Pk|k−1 Hk
xx T
(Pk|k−1 Hk )
xx Thanks to the combination of H-infinity criterion with
Re,k = 2 ; (40) UKF, the proposed H-infinity UKF is able to bound various
Pk|k−1 Hk
xx T
−γ I + Pk|k−1xx
  sources of uncertainties according to robust control theory,
Rk + Pk|k−1zz
[Pk|k−1
xz
]T such as model and measurement uncertainties. Note that how
= , (41)
Pk|k−1
xz
−γ 2 I + Pk|k−1
xx
to effectively address model and measurement uncertainties is
still in its infancy and very limited work has been done [ 22].
and where I is an identity matrix. This paper provides an attempt to handle that. Furthermore,
To show the relationship between the tuning parameter the proposed H-infinity UKF is implemented in a decentralized
γ and the estimation error covariance matrix, we apply the manner and thus is very fast to execute. The state filtering steps
matrix inversion lemma to (38), yielding of H-infinity UKF are similar to the traditional UKF while the
−1 added computational burdens are due to the inversion of larger
xx −1 Rk 0
(Pk|k ) = (Pk|k−1 xx
)−1 + [HkT I] 2 [HkT I]T scale matrices and the tuning of parameter γ 2 . However, due
0 −γ I
T −1 to the decentralized implementation, the number of state vari-
= (Pk|k−1 ) + Hk Rk Hk − γ −2 I.
xx −1
(42)
ables of each machine is small. Thus, the inversion of matrices
Thus, when γ tends to infinity, the H-infinity UKF will reduce are fast to calculate. The computational efficiency assessment
to the traditional UKF. Therefore, γ can be seen as a tuning results shown in Section IV validate these observations. On the
parameter to balance the tradeoff between the H-infinity and other hand, it is known that H-infinity criterion is unable to
the minimum mean-square error performance. To guarantee address outliers and cyber attacks while the robust estimators
the positive-definiteness of the error covariance matrix P k|k xx developed based on robust statistics criterion can suppress
T −1
i.e., (Pk|k ) + Hk Rk Hk − γ I  0, γ must be tuned to them [19], [21]. A combination of H-infinity criterion with
xx −1 −2

satisfy: other statistical robust estimators seems to be a good solution.

How to achieve that is out of the scope of this paper and will
−1
γ 2  (Pk|k−1 + HkT R−1 k Hk )
−1
, (43) be regarded as our future work.
which is equivalent to IV. N UMERICAL R ESULTS
−1
γ 2 ≥ max{eig(Pk|k−1 + HkT R−1
k Hk )
−1
}, (44) We carry out extensive simulations on the IEEE 39-bus
system to demonstrate the performance of the proposed
where eig(A)−1 denotes the eigenvalue of matrix A −1 . DSE subject to various types of uncertainties. The two-axis

1949-3053 (c) 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2018.2870327, IEEE
Transactions on Smart Grid

IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2018 6

3
× 10-3 • HDUKFI: decentralized H-infinity UKF with measured
DUKFV terminal current phasor as model inputs while treating the
DUKFI
2.5
HDUKFI voltage phasor and frequency as outputs/measurements.
Root Mean Squared Error [pu]

2
0.045
DUKFV
DUKFI
0.04
1.5 HDUKFI

Root Mean Squared Error [pu]

0.035

1
0.03

0.025
0.5
0.02

0 0.015
0 1 2 3 4 5 6 7 8 9 10
time [s] 0.01

Fig. 1: Root-mean-squared error of the DUKFV, the DUKFI 0.005

and the HDUKFI with unknown gaussian process and mea- 0

surement noises. They are assumed to be normally distributed 0 1 2 3 4 5 6 7 8 9 10

with zero means and covariance matrices Q = 10 −5 I9×9 and

time [s]

R = 10−4 I3×3 , respectively. Fig. 2: RMSE of the DUKFV, the DUKFI and the HDUKFI in
the presence of non-Gaussian process and measurement noises.
model provided with the IEEE-DC1A exciter and the TGOV1 They are assumed to follow a Gaussian mixture model, where
turbine-governor is assumed for each generator, where the 5% of the data are contaminated with covariance matrices Q =
parameter values can be found in [33]. Here, the transient 10−5 I9×9 and R = 10−5 I3×3 while the left 95% data are with
stability simulations are performed to generate measurements true covariance matrices Q = 10 −6 I9×9 and R = 10−6 I3×3 ,
and true state variables using the Matlab-based software PST respectively.
[34] with some modifications. The 4th order Ruger-Kutta ap-
proach is adopted with an integration step of t=1/120 s to solve
differential and algebraic equations. The simulations consist of
the following steps: Line 15-16 is tripped at t=0.5s to simulate 0.09
DUKFV
a system disturbance; the voltage phasor, current phasor and 0.08
DUKFI
HDUKFI
frequency at each generator’s terminal bus are corrupted by
Root Mean Squared Error [pu]

0.07
additive noise to simulate realistic PMU measurements; the
scan rate of the PMU measurements is 1/60s; note that if the 0.06

integration step of the dynamic simulation is different from 0.05

the PMU sampling rate, interpolation technique [35] is used 0.04

to coordinate the simulations and real-time measurements. Of
0.03
course, this would induce some errors. But, they have negligi-
ble effects on the DSE as demonstrated in [35]. On the other 0.02

hand, these errors can be regarded as model uncertainties and 0.01

will be bounded by our proposed H-infinity UKF. Furthermore,
0
the computing speed of our approach is smaller than the typical 0 1 2 3 4 5 6 7 8 9 10
time [s]
PMU scan rate, i.e., 30-120 samples/s, which will be shown by
simulation results. If the PMU sampling rate is 120 samples/s, Fig. 3: RMSE of the DUKFV, the DUKFI and the HDUKFI
no interpolation technique is required; the root-mean-squared when process and measurement noises follow a Gaussian
error (RMSE) of all estimated generator state variables is used mixture model, where 10% of the data are contaminated with
as the overall performance index while the estimated states covariance matrices Q = 5 × 10 −5 I9×9 and R = 10−4 I3×3
of Generator 5 are taken for illustration. Three approaches while the left 90% data are with true covariance matrices
are considered and compared, namely DUKFV, DUKFI and Q = 10−6 I9×9 and R = 10−6 I3×3 , respectively.
HDUKFI. They are defined as follows:
• DUKFV: decentralized UKF with measured terminal volt-
age phasor as model inputs while treating the current A. Case 1: Unknown Gaussian Process and Measurement
phasor and frequency as outputs/measurements; this is Noises
the existing approach proposed in [12]; In practical power systems, the process and measurement
• DUKFI: decentralized UKF with measured terminal cur- noises can vary under different system operating conditions,
rent phasor as model inputs while treating the voltage yielding deviations from nominal values. In such cases, they
phasor and frequency as outputs/measurements; this is are unknown to the operator. In this section, the process and
an extension of [12] presented in this paper; the measurement noises are assumed to be normally distributed

1949-3053 (c) 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2018.2870327, IEEE
Transactions on Smart Grid

IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2018 7

90
True Value
85
and DUKFI increase significantly, which demonstrates their
DUKFV 80
80 DUKFI
HDUKFI 75
sensitivities to non-Gaussian noise. By contrast, the increased
estimation error of our HDUKFI remains small thanks to the
δ 5 [degree]

70 70

60
65
8 8.5 9 9.5 10
H-infinity criterion. For example, by looking at the estimated
rotor angle estimates, we find that the proposed HDUKFI has
50
much higher accuracy than other two approaches, i.e., about
1.0006
40
0 1 2 3 4 5 6 7 8 9 10
5 degrees at some times, which is a lot and can significantly
1.0004
affect the intended applications.
ω [pu]

1.0002
C. Case 3: Robustness to Dynamical Model Uncertainties
5

0.9998
Due to aging processes, variations of the machine tempera-
0 1 2 3 4 5 6 7 8 9 10
1.0005 ture during its operation, the effect of saturation on generator
1.0004 inductances, to cite a few, the reactance and transient reactance
1.0003
of a synchronous generator can change with time. In our
ω5 [pu]

1.0002

1.0001
simulations, we assume that after the disturbances, both the d-
1
and q-axis transient reactance of the generators deviate from
0.9999
0 1 2 3 4 5 6 7 8 9 10
the nominal values by a percentage of 10% error. The latter
is simulated by a Gaussian random variable with zero mean
Fig. 4: Estimated states of Generator 5 by each method in the and specified error as standard deviation. Fig. 5 displays the
presence of unknown non-Gaussian process and measurement estimated state variables of Generator 5 by each method. It is
noises. This is associated with the results shown in Fig. 3. observed that both the DUKFV and the DUKFI are sensitive to
dynamical model uncertainties and provide significantly biased
with zero means and covariance matrices Q = 10 −5 I9×9 and estimation results; the reader is referred to the estimated rotor
R = 10−4 I3×3 , respectively. Note that the true covariance angle and mechanical power of Generator 5 as examples. By
matrices are Q = 10−6 I9×9 and R = 10−6 I3×3 . Simulation contrast, our HDUKFI bounds the uncertainties and provides
results of the DUKFV, the DUKFI and the HDUKFI are reasonably good state estimates. Further explanations of these
displayed in Fig. 1. It is found that DUKFI achieves slightly conclusions can be stated as follows:
better results than the DUKFV in presence of unknown process • In theory, both EKF and UKF assume that the dynamical
and measurement noise. By contrast, the proposed HDUKFI model is accurate and the system noise statistics, and
relies on the H-infinity criterion to obtain a better balance system inputs are known. Then, they can produce good
between model and measurements, yielding better results. estimation results. If these assumptions are violated, the
state estimates can be significantly biased [17], [21]. Our
B. Case 2: Unknown Non-Gaussian Process and Measure- results are consistent with these theoretical conclusions;
ment Noise • According to the work [36]–[39] in terms of generator
Due to communication channel noise, the measurement model validation and calibration, it is found that only a
noise associated with PMUs usually deviates from the Gaus- few key parameters play the major role in the responses
sian assumption [16]. In addition, since the PMU measure- of the generator. In other words, changes among these
ments are taken as system inputs, the system process noise is parameters can affect the generator response significantly
no longer Gaussian. To assess the sensitivity of each method while the changes on other parameters induce negligible
to unknown non-Gaussian process and measurement noise, difference. Furthermore, the sensitivity of parameters can
we assume they follow a Gaussian mixture model with two change in presence of different system disturbances. This
mixture components, where 5% of the data are contaminated has been validated and emphasized in NERC standard
with covariance matrices Q = 10 −5 I9×9 and R = 10−5 I3×3 [38] when performing parameter calibrations. Last but
while the left 95% data are with the true covariance matrices not the least, both generator and controller parameters
Q = 10−6 I9×9 and R = 10−6 I3×3 . The RMSE of each can have high sensitivities. Following [36]–[39], we carry
method is displayed in Fig. 2. It can be observed that the out trajectory sensitivity analysis to Generator 5 with and
DUKFV is very sensitive to non-Gaussian noises and provides without parameter uncertainties on generator d- and q
the worst estimation results. Although the DUKFI performs axis transient reactance, exciter gain K A and governor
slightly better than the DUKFV, its statistical efficiency is TSV , respectively. Due to space limitation, the detailed
much lower than our HDUKFI. results are not shown here. The results show that the
When a higher percentage of data deviates from the Gaus- former two groups of parameters play the major role
sian assumption, say 10% data is contaminated with covari- in the responses of the generator. This demonstrates
ance matrices Q = 5 × 10−5 I9×9 and R = 10−4 I3×3 while why uncertainties on generator d- and q axis transient
the left 90% data are with the true covariance matrices Q = reactance can significantly affect EKF and UKF.
10−6 I9×9 and R = 10−6 I3×3 , the RMSE of each method and In practice, the controller parameters can change from time
the estimated states of generator 5 are shown in Figs. 3-4. It to time [27], [28] due to aging process, inappropriate tuning
is observed that with the increased degree of deviation from and controller failures, to name a few. On the other hand, other
the Gaussian assumption, the estimation errors of the DUKFV parameters can vary in the model at the same time. To simulate

1949-3053 (c) 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2018.2870327, IEEE
Transactions on Smart Grid

IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2018 8

90 1.0005
true value
DUKFV 1.0004
80
DUKFI
1.0003

δ 5 [degree]
HDUKFI

ω [pu]
70
1.0002

5
60
1.0001

50 1

40 0.9999
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
0.65
1.2

0.6
1.1
E'd5 [pu]

E'q5 [pu]
0.55
1
0.5

0.9
0.45

0.4 0.8
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
1.95 1.95

1.9 1.9
[pu]

[pu]
1.85 1.85
fd5

R5
V
E

1.8 1.8

1.75 1.75
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
× 10-3
5 5.2

5.15

0 5.1

TM 5 [pu]
[pu]

5.05
F5
V

-5 5

4.95

-10 4.9
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
time [s] time [s]

Fig. 5: Estimated states of Generator 5 by each method with model uncertainties. Here, it is assumed that after the disturbance
is applied, both the d- and q-axis transient reactance of Generator 5 deviates from the nominal values by a percentage of 10%.

80 1.0006
true value
DUKFV
70 DUKFI 1.0004
δ 5 [degree]

HDUKFI
ω5 [pu]

60 1.0002

50 1

40 0.9998
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
0.65 1.1

1.05
0.6
[pu]
E'd5 [pu]

1
0.55
q5

0.95
E'

0.5
0.9

0.45 0.85
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
2.2 2.5

2
2
[pu]
Efd5 [pu]

1.8
R5
V

1.5
1.6

1.4 1
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
0.02
5.1

0.01 5.08
[pu]

TM [pu]

0 5.06
F5

-0.01 5.04
V

-0.02 5.02

-0.03 5
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
time [s] time [s]

Fig. 6: Estimated states of Generator 5 by each method with model uncertainties. Here, it is assumed that the d- and q axis
steady-state and transient reactances and the excitor gain K A of the generator 5 are all subject to 5% uncertainties after the
disturbance.

a more realistic scenario, we assume that both the d- and q-axis Fig. 6. It can be concluded that when more parameters of the
steady-state and transient reactances and the excitor gain K A generator are subject to uncertainties, both the DUKFV and
of the generator 5 are all subject to 5% uncertainties after the the DUKFI obtain larger estimation errors. By contrast, the
disturbance. The estimated states of generator 5 are shown in proposed HDUKFI is still able to bound these uncertainties

1949-3053 (c) 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2018.2870327, IEEE
Transactions on Smart Grid

IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2018 9

TABLE I: Average Computing Times of The Three Methods

0.012
GM-IEKF At Each PMU Scan
HEKF
HDUKFI
0.01 Cases DUKFV DUKFI HDUKFI
Root Mean Squared Error [pu]

Case 1 0.36ms 0.31ms 1.15ms

0.008
Case 2 0.38ms 0.35ms 1.20ms
Case 3 0.39ms 0.37ms 1.22ms
0.006
the proposed HDUKFI achieves much better capability of
handling the model uncertainties than the first-order Taylor
0.004
series expansion-based H-infinity EKF and the GM-IEKF.
Thanks to the min-max criterion, the H-infinity EKF is able
0.002
to balance the trade-off between model and measurements,
yielding better robustness than the GM-IEKF in terms of
0
0 1 2 3 4 5 6 7 8 9 10 handling model uncertainties.
time [s]
D. Assessment of Computational Efficiency
Fig. 7: RMSE of the GM-IEKF, the HEKF and the HDUKFI To assess the computational efficiency of each method, the
with model uncertainties, where after the disturbance is ap- average computing times of the DUKFV, the DUKFI and the
plied, both d- and q-axis transient reactance of Generator 5 HDUKFI for Cases 1-3 at each PMU scan are presented in
deviate from the nominal value by a percentage of 10%. Table. I. All the tests are performed on a PC with Intel Core
i5, 2.50 GHz, 8GB of RAM. From the table, we find that
0.014
the DUKFV and the DUKFI have comparable computational
GM-IEKF
HEKF
efficiency in all cases while the proposed HDUKFI spends a
0.012 HDUKFI little bit more time. Their computing times are much smaller
than the PMU scan rate, which is 16.7ms for 60 samples/s (or
Root Mean Squared Error [pu]

0.01
8.3ms for 120 samples/s). This is because all three approaches
0.008
are implemented in a decentralized manner and thus are very
fast to execute. Furthermore, the reason why the HDUKFI
0.006 needs slightly more times is that it involves the inversion of
larger scale matrices and the tuning of parameter γ 2 . But
0.004
thanks to the decentralized implementation, the number of
0.002
state variables of each machine is small and therefore the
inversion of matrices are fast to calculate. In summary, the
0
0 1 2 3 4 5 6 7 8 9 10
HDUKFI is able to keep up with the PMU scan rate and
time [s] suitable for large-scale power system online applications.
Fig. 8: RMSE of the GM-IEKF, the HEKF and the HDUKFI E. Practical Values of the Proposed HDUKFI
with model uncertainties, where the d- and q-axis steady- For practical power systems, the estimated dynamic state
state and transient reactances and the excitor gain K A of variables can be used for many different applications, such
the generator 5 are all subject to 5% uncertainties after the as development of local and wide-area oscillation damping
disturbance. controllers, design of adaptive out-of-step protection scheme,
to name a few [3], [4], [39]. Note that many dynamic state
with small increase of estimation errors. For example, by variables are not measured by PMUs, such as field voltage,
comparing with Fig. 5, it can be observed that the estimation generator rotor angle, mechanical power and so on. With the
errors of rotor angle, field voltage and mechanical power DSE, all of them can be estimated online and used for further
by the DUKFV and the DUKFI are increased significantly. power grid monitoring, control and protection.
However, those errors for the proposed HDUKFI are still much In principle, the more accurate the estimated state variables
smaller. Furthermore, it is interesting that both DUKFV and are, the more effective the developed damping controllers,
DUKFI are able to get quite reasonable rotor speed estimation adaptive out-of-step protection scheme, etc, will be. For exam-
even in presence of model uncertainties while producing poor ple, with poorly estimated rotor speed and angle, the decen-
results for other state variables. By contrast, our proposed tralized damping controller may be ineffective and sometimes
HDUKFI outperforms all other approaches, yielding accurate may induce adverse affects [4]. In addition, they may lead
state estimates in different conditions. to incorrect conclusions of the rotor angle stability and the
To further demonstrate the advantages of the proposed failure of the out-of-step protection scheme [ 3]. Our proposed
HDUKFI over other competitive robust methods, such as the H-infinity UKF outperforms other approaches significantly
GM-IEKF [7] and the H-infinity EKF (HEKF) [22], the former in the presence of various sources of model uncertainties,
two scenarios are repeated and the corresponding results are demonstrating its potentials for wide-area applications. The
shown in Figs. 7-8. Due to space limitation, only the RMSE power system applications based on the estimated dynamic
results are shown. It is clear from those two figures that state variables are beyond the scope of the paper. We will
with the combination of UKF and the H-infinity criterion, investigate that in our future work.

1949-3053 (c) 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2018.2870327, IEEE
Transactions on Smart Grid

IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2018 10

V. C ONCLUSIONS AND F UTURE W ORK [13] E. Ghahremani, I. Kamwa, “Local and wide-area PMU-based decentral-
ized dynamic state estimation in multi-machine power systems,” IEEE
In this paper, a decentralized H-infinity UKF is proposed to Trans. Power Syst., vol. 31, no. 1, pp. 547–562, 2016.
address various sources of uncertainties. It has the following [14] G. Anagnostou, B. C. Pal, “Derivative-free Kalman filtering based
salient features: approaches to dynamic state estimation for power systems with unknown
inputs,” IEEE Trans. Power Syst., vol. 33, no. 1, pp. 116-130, 2018.
• the use of statistical linerization approach allows us to [15] Z. Huang, N. Zhou, R. Diao, S. Wang, S. Elbert, D. Meng, S. Lu, “Cap-
convert the traditional nonlinear UKF into an equivalent turing real-time power system dynamics: opportunities and challenges,”
Proc. IEEE Power Eng. Soc. General Meeting, July 2015.
linear like batch-mode regression form. The latter is fur- [16] S. Wang, J. B. Zhao, Z. Huang, R. Diao, “Assessing Gaussian assump-
ther leveraged to develop the H-infinity UKF by resorting tion of PMU measurement error using field data,” IEEE Trans. Power
to the linear H-infinity Kalman filter framework in the Delivery, 2017, in press.
[17] D. Simon, Optimal State Estimation. Hoboken, NJ: Wiley, 2006.
Krein space; [18] K. Zhou, J. C. Doyle, K. Glover, Robust and Optimal Control, Prentice-
• two model decoupling approaches are presented to enable Hall, NewJersey, 1996.
the decentralized estimation of generator state variables [19] M. Gandhi, L. Mili, “Robust Kalman filter based on a generalized
maximum-likelihood-type estimator, IEEE Trans. Signal Processing, vol.
through local PMU measurements. As a result, no com- 58, no. 5, pp. 2509–2520, 2010.
plicated communications and other system models are [20] W. Li, Y. Jia, “H-infinity filtering for a class of nonlinear discrete-time
needed, enhancing the capability of the H-infinity UKF systems based on unscented transform,” Signal Processing, vol. 90, no.
12, pp. 3301-3307, 2010.
for large-scale power system online applications; [21] G. A. Einicke, L. B. White, “Robust extended Kalman filtering,” IEEE
• comparison results carried out on the IEEE 39-bus system Trans. Signal Processing, vol. 47, no. 9, pp. 2596–2599, 1999.
show that the proposed H-infinity UKF has much better [22] J. B. Zhao, “Dynamic state estimation with model uncertainties using
H-infinity extended Kalman filter,” IEEE Trans. Power Syst., vol. 33,
robustness than other alternatives in terms of bounding no. 1, pp. 1099–1100, 2018.
various types of uncertainties, including the unknown [23] J. B. Zhao, L. Mili, “A robust generalized-maximum likelihood un-
Gaussian and non-Gaussian process and measurement scented Kalman filter for power system dynamic state estimation,” IEEE
Journal of Selected Topics in Signal Processing, vol. 12, no. 4, pp. 578-
noise, and the dynamical model uncertainties. 592, 2018.
It is known that H-infinity criterion is unable to address [24] IEEE Std 1110-2002, IEEE Guide for Synchronous Generator Modeling
outliers while the robust estimators developed based on robust Practices and Applications in Power System Stability Analyses, IEEE
Power and Engineering Society.
statistics criterion can suppress them [19], [21]. In the future [25] P. Sauer, M. A. Pai. Power system dynamics and stability. Urbana, 1997.
work, we will combine the H-infinity criterion with other [26] E. Kyriakides, et. al, “Online estimation of synchronous generator pa-
statistical robust estimators to develop a hybrid robust DSE, rameters using a damper current observer and a graphic user interface,”
IEEE Trans. Energy Conversion, vol. 19, no. 3, pp.499-507, 2004.
which is able to bound model and measurement uncertainties [27] J. Chow, et al, “Generator and exciter parameter estimation of Fort
while being robust to outliers and cyber attacks. Patrick Henry hydro unit 1,” IEEE Trans. Energy Conversion, vol. 14,
R EFERENCES no. 4, pp. 923-929, 1999.
[28] D. Kosterev, “Hydro turbine-governor model validation in Pacific North-
[1] Y. Liu, A. S. Meliopoulos, Z. Tan, L. Sun, R. Fan, “Dynamic state west,” IEEE Trans. Power Syst., vol. 19, no. 2, pp. 1144-1149, 2004.
estimation-based fault locating on transmission lines,” IET Generation, [29] J. B. Zhao, L. Mili, “Power system robust decentralized dynamic state
Transmission & Distribution, vol. 11, no. 17, pp. 4184-4192, 2017. estimation based on multiple hypothesis testing,” IEEE Trans. Power
[2] Y. Liu, A. P. Meliopoulos, R. Fan, L. Sun, Z. Tan, “Dynamic state Syst., vol. 33, no. 4, pp. 4553-4562, 2018.
estimation based protection on series compensated transmission lines,” [30] S. Julier, J. K. Uhlmann, “Unscented filtering and nonlinear estimation,”
IEEE Trans. Power Delivery, vol. 32, no. 5, pp. 2199-2209, 2017. Proceedings of the IEEE, vol. 92, no. 3, pp. 401–422, 2004.
[3] A. P. S. Meliopoulos, G. J. Cokkinides, et. al, “Dynamic state estimation- [31] J. B. Zhao, L. Mili, “Robust unscented Kalman filter for power system
based protection: status and promise,” IEEE Trans. Power Delivery, vol. dynamic state estimation with unknown noise statistics,” IEEE Trans.
32, no. 1, pp. 320-330, 2017. Smart Grid, 2017.
[4] A. K. Singh, B. C Pal, “Decentralized control of oscillatory dynamics [32] B. Hassibi, Ali H. Sayed, T. Kailath, “Linear estimation in Krein spaces-
in power systems using an extended LQR,” IEEE Trans. Power Syst., Part II: applications,” IEEE Trans. Automatic Control, vol. 41, no. 1, pp.
vol. 31, no. 3, pp. 1715-1728, 2016. 34-49, 1996.
[5] E. Ghahremani, I. Kamwa, “Dynamic state estimation in power system [33] IEEE PES TF on Benchmark System for Stability Controls, “Benchmark
by applying the extended Kalman filter with unknown inputs to phasor systems for small-signal stability analysis and control,” Aug. 2015.
measurements,” IEEE Trans. Power Syst., vol. 26, no. 4, pp. 2556–2566, [34] J. H. Chow, K. W. Cheung, “A toolbox for power system dynamics and
2011. control engineering education and research,” IEEE Trans. Power Syst.,
[6] J. B. Zhao, L. Mili, “Robust dynamic state estimator to outliers and vol. 7, no. 4, pp. 1559-1564, 1992.
cyber attacks,” Proc. IEEE Power Energy General Meeting, 2017, pp. [35] S. Akhlaghi, N. Zhou, Z. Huang, “A multi-step adaptive interpolation
1-5. approach to mitigating the impact of nonlinearity on dynamic state
[7] J. B. Zhao, M. Netto, L. Mili, “A robust iterated extended Kalman filter estimation,” IEEE Trans. Smart Grid, vol. 9, no. 4, pp. 3102-3111, 2018.
for power system dynamic state estimation,” IEEE Trans. Power Syst., [36] S. Benchluch, J. Chow, “A trajectory sensitivity method for the iden-
vol. 32, no. 3, pp. 2484-2486, 2017. tification of nonlinear excitation system models,” IEEE Trans. Energy
[8] A. Rouhani, A. Abur, “Linear phasor estimator assisted dynamic state Conv., vol. 8, no. 2, pp. 159-164, 1993.
estimation,” IEEE Trans. Smart Grid, vol. 9, no. 1, pp. 211– 219, 2018. [37] I. A. Hiskens, “Nonlinear dynamic model evaluation from disturbance
[9] S. Wang, W. Gao, A. P. S. Meliopoulos, “An alternative method for measurements,” IEEE Trans. Power Syst., vol. 16, no. 4, pp. 702-710,
power system dynamic state estimation based on unscented transform,” 2001.
IEEE Trans. Power Syst., vol. 27, no. 2, pp. 942–950, 2012. [38] Reliability Guideline of Power Plant Dynamic Model Verification using
[10] Z. Huang, P. Du, D. Losterev, B. Yang, “Application of extended Kalman PMUs, North American Electric Reliability Corporation, 2016.
filter techniques for dynamic model parameter calibration,” Proc. IEEE [39] Z. Huang, P. Du, D. Kosterev, S. Yang, “Generator dynamic model
Power Energy General Meeting, 2009, pp. 1–6. validation and parameter calibration using phasor measurements at the
[11] L. Fan and Y. Wehbe, “Extended Kalman filtering based real-time point of connection,” IEEE Trans. Power Syst., vol. 28, no. 2, pp. 1939–
dynamic state and parameter estimation using PMU data,” Elect. Power 1949, 2013.
Syst. Res., vol. 103, pp. 168–177, 2013.
[12] A. K. Singh, B. C. Pal, “Decentralized dynamic state estimation in power
systems using unscented transformation,” IEEE Trans. Power Syst., vol.
29, no. 2, pp. 794–804, 2014.

1949-3053 (c) 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2018.2870327, IEEE
Transactions on Smart Grid

IEEE TRANSACTIONS ON SMART GRID, VOL. , NO. , 2018 11

Junbo Zhao (M’18) is an assistant professor (re-

search) at Virginia Tech. He received the Ph.D.
degree from Bradley Department of Electrical and
Computer Engineering at Virginia Tech in 2018.
He did the summer internship at Pacific Northwest
National Laboratory from May-August, 2017. He
is now the chair of IEEE Task Force on Power
System Dynamic State and Parameter Estimation,
the secretary of the IEEE Working Group on State
Estimation Algorithms and the IEEE Task Force
on Synchrophasor Applications in Power System
Operation and Control.
He has written 2 book chapters, published more than 40 peer-reviewed
journal and conference papers. His research interests lie in power system real-
time monitoring, operations and cyber security that include state estimation,
dynamics and stability, cyber attacks and countermeasures, big data analytics
and robust statistics with applications in the smart grid. He serves as the
Associate Editor of International Journal of Electrical Power & Energy
Systems and IET Generation, Transmission & Distribution.

Lamine Mili (LF’17) received the Electrical Engi-

neering Diploma from the Swiss Federal Institute
of Technology, Lausanne, in 1976, and the Ph.D.
degree from the University of Liège, Belgium, in
1987. He is a Professor of Electrical and Computer
Engineering, Virginia Tech, Blacksburg. He has five
years of industrial experience with the Tunisian
electric utility, STEG. At STEG, he worked in the
planning department from 1976 to 1979 and then at
the Test and Meter Laboratory from 1979 till 1981.
He was a Visiting Professor with the Swiss Federal
Institute of Technology in Lausanne, the Grenoble Institute of Technology,
the École Supérieure D’électricité in France and the École Polytechnique de
Tunisie in Tunisia, and did consulting work for the French Power Transmission
company, RTE.
His research has focused on power system planning for enhanced resiliency
and sustainability, risk management of complex systems to catastrophic
failures, robust estimation and control, nonlinear dynamics, and bifurcation
theory. He is the co-founder and co-editor of the International Journal of
Critical Infrastructure. He is the chairman of the IEEE Working Group on
State Estimation Algorithms. He is a recipient of several awards including
the US National Science Foundation (NSF) Research Initiation Award and
the NSF Young Investigation Award.

1949-3053 (c) 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.