Model-Free MLE Estimation for Online Rotor
Angle Stability Assessment with PMU Data
Shaopan Wei, Student Member, IEEE, Ming Yang, Member, IEEE, Junjian Qi, Member, IEEE,
Jianhui Wang, Senior Member, IEEE, Shiying Ma, and Xueshan Han
Abstract—Recent research has demonstrated that the rotor
angle stability of a power system can be assessed by identifying
the sign of the system’s maximal Lyapunov exponent (MLE).
A positive (negative) MLE implies unstable (stable) rotor angle
dynamics. However, because the MLE may fluctuate between
positive and negative values for a long time after a severe
disturbance, determining the system stability is difficult when
observing a positive or negative MLE without knowing its further
fluctuation trend. In this paper, a new approach for online
rotor angle stability assessment is proposed to address this
problem. The MLE is estimated by a recursive least squares-
based method from real-time rotor angle measurements, and two
critical parameters, the Theiler window, and the MLE estimation
initial time step are carefully chosen to make sure that the
calculated MLE curves present distinct features for different
stability conditions. By using the proposed stability assessment
criteria, the developed approach can provide a timely and reliable
assessment of rotor angle stability. Extensive tests on the New
England 39-bus system and the Northeast Power Coordinating
Council 140-bus system verify the effectiveness of the proposed
approach.
Index Terms—Lyapunov exponent, model-free, online stabil-
ity assessment, phasor measurement unit, rotor angle stability,
Theiler window.
I. I NTRODUCTION
T RANSIENT rotor angle stability refers to the ability
of synchronous generators of an interconnected power
system to remain in synchronism after a severe disturbance [1].
With the development of synchrophasor technologies, utilities
are now able to track rotor angle deviations and take actions
to respond to emergency events. However, since the dynamics
This work was supported by the National Basic Research Program of China
(973 Program) under Grant 2013CB228205, the State Grid Corporation of
China under Grant XT71-15-056, and the National Science Foundation of
China under Grants 51007047 and 51477091.
M. Yang (Corresponding Author) is with the Key Laboratory of Power Sys-
tem Intelligent Dispatch and Control, Shandong University, Jinan, Shandong
250061, China. He was a visiting scholar with Argonne National Laboratory,
Argonne, IL 60439, USA (e-mail: myang@sdu.edu.cn).
S. Wei and X. Han are with the Key Laboratory of Power System Intelligent
Dispatch and Control, Shandong University, Jinan, Shandong 250061, China
(e-mail: spw sdu@sina.com; xshan@sdu.edu.cn).
J. Qi is with the Department of Electrical and Computer Engineering,
University of Central Florida, Orlando, FL 32816, USA (e-mail: Jun-
jian.Qi@ucf.edu).
J. Wang is with the Department of Electrical Engineering at Southern
Methodist University, Dallas, TX 75275, USA and the Energy Systems
Division at Argonne National Laboratory, Argonne, IL 60439, USA (email:
jianhui.wang@ieee.org).
S. Ma is with Institute of Electric Power System, China Electric
Power Research Institute, Haidian District, Beijing 100192, China (e-mail:
mashiy@epri.sgcc.com.cn).
'LJLWDO2EMHFW,GHQWL¿HU73:56
‹,(((3HUVRQDOXVHLVSHUPLWWHGEXWUHSXEOLFDWLRQUHGLVWULEXWLRQUHTXLUHV,(((SHUPLVVLRQ
6HH KWWSZZZLHHHRUJSXEOLFDWLRQV VWDQGDUGVSXEOLFDWLRQVULJKWVLQGH[KWPO IRU PRUH LQIRUPDWLRQ
of power systems are complex, the online assessment of rotor
angle stability is still very challenging [2], [3].
In [4], an adaptive out-of-step relay was proposed for the
Florida–Georgia system. The equal area criterion was applied
to change the settings of the protection system based on phasor
measurement unit (PMU) measurements. In [5], the dynamics
of the power transfer paths were monitored based on the
energy functions of the two-machine equivalent system, and
the PMU data were used to identify the parameters of the
energy functions. In [6], PMU measurements were used as
inputs for estimating the differential/algebraic equation model
to predict post-fault dynamics. In [7], an online dynamic
security assessment scheme was proposed based on self-
adaptive decision trees, where PMU data were used for online
identification of the critical attributes of the system. In [8],
rotor angle stability was estimated by using artificial neural
networks and the measured voltage and current phasors were
used as inputs of the offline-trained estimation model. In [9], a
systematic scheme for building fuzzy rule-based classifiers for
a fast stability assessment was proposed. By testing on a large
and highly diversified database, it was demonstrated that the
analysis of post-fault short-term PMU data can extract useful
features satisfying the requirements of stability assessment.
Lyapunov exponents (LEs) characterize the separation rate
of infinitesimally close trajectories, and they are important
indices for quantifying the stability of dynamic systems. If
the system’s maximal Lyapunov exponent (MLE) is positive,
the system is unstable, and vice versa. LEs were first applied
to power system stability analysis in [10], in which it was
verified that LEs can predict the out-of-step conditions of
power systems. In [2], a model-based MLE method was pro-
posed for online prediction of rotor angle stability from PMU
measurements. The work built solid analytical foundations for
LE-based rotor angle stability assessment. In [11], the LEs
were calculated with dynamic component and network models
to identify coherent groups of generators.
Although model-based MLE estimation approaches have
made significant progresses on online rotor angle stability
assessment, they are usually computationally expensive, es-
pecially when applied to large power systems. On the other
hand, since the 1980s, the computation of LE from time series
has been widely studied [12]–[15]. Based on these theoretical
studies, two model-free MLE estimation approaches have been
proposed for transient voltage stability assessment [16] and
rotor angle stability assessment [3], in which the MLEs are
estimated by using only PMU measurements.
Model-free LE-based stability assessment approaches are at-

tractive, because they can eliminate model errors and simplify
the calculation. However, when applying these approaches,
a time window has to be prespecified for MLE observation.
The window size is crucial for obtaining reliable and timely
assessment results, i.e., a very small window size will lead
to unreliable assessment results while a too large window
size will lead to accurate but untimely results. Determining
the window size in advance is difficult because the estimated
MLEs may fluctuate between positive and negative values
for quite a long time after disturbances and the window size
should change with different fault scenarios.
In this paper, an LE-based model-free rotor angle stability
assessment approach is proposed. The MLEs are estimated
by a recursive least squares (RLS) based method from real-
time rotor angle measurements. When two critical parameters
are chosen according to the characteristics of the relative
rotor angles of the selected generator pairs, the calculated
MLE curves will present distinct features for different sta-
bility conditions, based on which the stability criteria are
correspondingly designed to capture the MLE features and
perform online assessment of rotor angle stability. Compared
with the existing approaches, the proposed approach does not
need a prespecified time window to identify the sign of MLEs.
Instead, the proposed approach can always make a reliable and
timely assessment as soon as the crucial features are observed
from the estimated MLE curves.
The remainder of this paper is organized as follows. Section
II introduces the theoretical basis of model-free MLE estima-
tion. Section III proposes a rotor angle stability assessment
approach, and discusses the parameter selection principles and
stability criteria. In Section IV, simulation results on the New
England 39-bus system and the Northeast Power Coordinating
Council (NPCC) 140-bus system are presented to validate the
effectiveness of the proposed approach. Section V discusses
the impacts of the noise and sampling rate of PMUs. Finally,
conclusions are drawn in Section VI.
II. MLE E STIMATION F ROM T IME S ERIES
LEs can reflect the exponential divergence or convergence
of neighboring trajectories in the state space of a dynamic
system [17]. An N -dimensional dynamic system, i.e., there
are N state variables in the dynamic equation ẋ = f (x), has
N LEs, and the largest one is defined as the MLE of the
system. The MLE is a useful indicator of system stability: A
positive MLE indicates unstable system dynamics whereas a
negative MLE indicates asymptotically stable dynamics [2].
The MLE can be estimated by using Jacobian matrix-based
(model-based) methods [2], [17] or direct model-free methods
[3], [16], [17]. Compared with the Jacobian matrix-based
methods, direct methods are more suitable for online stability
assessment mainly because they do not need the repeated
computing of the Jacobian matrix or even the dynamic model
of the system.
According to Oseledec’s multiplicative ergodic theorem
[17]–[19], for a reference point X 0 and its neighboring point
X m(0) chosen from the state space of a nonlinear dynamic
system, the distance between the trajectories emerging from
2
X 0 and X m(0) , i.e., the original trajectory and the neigh-
boring trajectory, will have three different growth phases as
illustrated in Fig. 1. In Phase I, the difference vector between
the states of the trajectories gradually converges towards
the most expanding direction, and the distance between the
trajectories will exhibit fluctuations. In Phase II, the distance
experiences an exponential growth characterized by the MLE,
which corresponds to a linear segment in the semi-logarithmic
plot. Finally, in Phase III, the separation of the trajectories is
saturated and the distance converges to a constant value.
Fig. 1. Logarithmic distance of states on different trajectories.
As proved by [17] and [19]–[21], the MLE can be estimated
from the mean logarithmic separation rate of the trajectories
in Phase II as
 
1 d (m(n), n, k)
λk ≈ log
kΔt d (m(n), n, 0)
 
1 ||X m(n)+k − X n+k ||
= log ,
kΔt ||X m(n) − X n ||
X n , X m(n) , X n+k , and X m(n)+k ∈ Phase II, (1)
where λk is the estimated MLE, k is the lagged time steps
for the MLE estimation, Δt is the time duration for each
time step, X n and X m(n) are the initial points for the
MLE estimation on the original and neighboring trajectories,
respectively, X n+k is the kth point behind X n on the original
trajectory, X m(n)+k is the kth point behind X m(n) on the
neighboring trajectory, d (m(n), n, 0) is the Euclidean distance
between the MLE estimation initial points, d (m(n), n, k) is
the Euclidean distance between the kth points behind the MLE
estimation initial points, and ||A − B|| denotes the Euclidean
distance between points A and B.
It should be noted that the original and the neighboring
trajectories are usually from the same observed time series
with different initial points, as illustrated in Fig. 2. To make
sure that the trajectories are temporally separated and hence
can be seen as different trajectories, the trajectory’s initial
points should satisfy |m(0) − 0| > w, where w is called the
Theiler window [17] and should be determined according to
the characteristics of the system. In Fig. 2, X 0 and X m(0) are
the initial points of the original and neighboring trajectories
generated from the time series, respectively. Here, 0 and m(0)
are the time labels of the observed time series, satisfying m(0)
− 0 = w. Furthermore, the MLE should be estimated from
the linear segment of the logarithmic distance curve (Phase
II in Fig. 1). Correspondingly, X n and X m(n) are the initial
points for the MLE estimation on the original and neighboring

trajectories, ensuring that the MLE estimation is executed
within the right segment. Here, n and m(n) are the time labels
satisfying m(n) − n = m(0) − 0 = w.
III. LE-BASED ROTOR A NGLE S TABILITY A SSESSMENT
Here, we propose a rotor angle stability assessment ap-
proach based on MLE estimation. The approach is data-driven
and can perform online stability assessment using only the
rotor angle and the rotor speed of the generators [2], [3]. Rotor
angle and rotor speed may not be directly provided by the
PMUs, but can be estimated by using various dynamic state
estimation algorithms, which can prove real-time and accurate
rotor angle and rotor speed estimation results with respect to
the PMU measurements [22]–[30].
A. State Variable Selection
Even a moderate-sized power system may still have hun-
dreds of state variables. Due to both the calculation in-
tractability and the insufficiency of measurements, using all
these variables to form the state space for MLE estimation is
impractical. It is more feasible to reconstruct the power system
dynamics with only a small number of state variables.
According to Takens’ theorem [31]–[37], the dynamics of
a nonlinear system can be reconstructed by the observations
of a single state variable, the essential mechanism of which is
that the evolution of any component in a nonlinear system
is affected by the changes of other related components of
the system, and hence the information of all these related
components is inherently included in the development of
the single component. Although Takens’ theorem emerged
from investigations of chaotic systems, they are applicable
to general nonlinear physical systems [37]. The reconstructed
state vector can be expressed as
 
Θt = θt , θt−τ , θt−2τ , · · · , θt−(N −1)τ , (2)
where N is the embedding dimension, τ is the lag time, and
θt , θt−τ , θt−2τ , · · · , θt−(N −1)τ are the observations of the
observed variable at the corresponding time steps.
Fig. 2. Trajectories and the effect of the Theiler window.
3
In the reconstructed state space, the Euclidean distance in
(1) becomes
d(m(n), n, i) = ||Θm(n)+i − Θn+i ||, i ∈ {0, k}, (3)
where Θn+i and Θm(n)+i are the observations on the recon-
structed trajectory.
Moreover, another simplification can be made based on
[17] and [21]. According to the statistical properties of the
local divergence rates of the nearby trajectories, this universal
simplification rule guarantees that considering only the first
component of the reconstructed state vector to estimate the
MLE is reasonable. This is essentially because all components
will grow exponentially at the rate of the MLE. Therefore,
Θn+i and Θm(n)+i in (3) can be replaced by their first
components, and the distance in (1) can be calculated by
d(m(n), n, i) = |θm(n)+i − θn+i |, i ∈ {0, k}. (4)
In this paper, the relative rotor angle of the severely
disturbed generator pair (SDGP) is selected as the observed
variable, because these SDGPs are, in general, responsible for
the system dynamics after considerable disturbances [38]. An
SDGP should be composed of a severely disturbed generator
and the least disturbed generator to reflect the dynamics of the
severely disturbed generators [39]. In particular, the SDGPs
can be identified as follows.
1) Obtain the rotor speed of all generators at the fault
clearing moment, ωtc ,n , n = 1, 2, . . . , NG , where NG is
the number of generators. The rotor speed of generators
can be estimated by dynamic state estimation based on
PMU measurements of the voltage phasors and current
pahsors at the terminal bus of some generators that are
determined by the optimal PMU placement method in
[27].
2) Obtain the maximal absolute value of the rotor speed,
i.e., ωt∗c = max |ωtc ,n |. Define generator g as one
n=1,2,...,NG
of the severely disturbed generators, if |ωtc ,g | /ωt∗c > σ,
where σ is a predetermined threshold. In this work, σ
is chosen as 0.7, as in [40].
3) Find the least disturbed generator with the minimal
absolute value of the rotor speed.
4) Form an SDGP by combing one of the severely disturbed
generators and the least disturbed generator. Iterate over
the severely disturbed generators and form all SDGPs.
For each SDGP, it can be viewed as a grid-connected non-
linear subsystem [41]. Since Takens’ theorem is universal for
nonlinear systems, it is applicable for the SDGP subsystems.
It should be noted that when applying Takens’ theorem, two
premises should be satisfied to ensure the performance of the
reconstruction: (a) The time series used for the reconstruction
should be able to reflect the dynamics of the system [35];
(b) The embedding dimension, i.e., N , should be sufficiently
large, and the lag time, i.e., τ , should be properly selected
[34].
When analyzing the dynamics of the SDGPs, the premises
of Takens’ theorem can be satisfied, because:
 4

1) For each SDGP, the relative rotor angle of the gener- where X k is the coefficient matrix and Y k is the observation
ator pair can be regarded as the state variable of the vector. According to the aforementioned definitions, X k and
subsystem [41], whose trajectory can directly reflect the Y k can be expressed as
⎡ ⎤
dynamics of the subsystem, so the premise (a) can be m(n)Δt 1
satisfied. ⎢ (m(n) + 1)Δt 1 ⎥
Xk = ⎣ ,
··· ⎦
(8)
···
2) In this research, the objective of the reconstruction is (m(n) + k)Δt 1
to calculate the MLE. According to the simplification in
[17] and [21], the selection of N and τ can be ignored in ⎡ ⎤
L(m(n))
the proposed approach (although there will be an optimal ⎢ L(m(n) + 1) ⎥
Y k =⎣ ⎦. (9)
N and τ ) since the selection of N and τ will not affect ···
the first element of the reconstructed vector. L(m(n) + k)
In fact, the relative rotor angle of any generator pair can be
Moreover, to avoid the repeated calculation of the inverse
regarded as the state variable of the generator pair subsystem,
matrix in (7), a recursive estimation algorithm is applied [43],
reflecting the dynamics of the subsystem. In the paper, the
which can be formulated as
relative rotor angles of the SDGPs are selected as the observed
variables to identify the system stability because: Ê k+1 = Ê k + Gk+1 yk+1 − x
k+1 Ê k ,
1) The relative rotor angle of an SDGP can be observed to P k xk+1
identify the stability of the SDGP. Gk+1 = ,
1 + x
k+1 P k xk+1
2) The stability of the SDGPs can be used to identify the P k+1 = P k − Gk+1 x
k+1 P k , (10)
stability of the whole system [38].

The proposed approach is not based on the assumption that where xk+1 is [(m(n) + k + 1)Δt 1] , yk+1 is the new
each generator is equipped with a PMU. Instead, the rotor observation L(m(n)+k +1), Ê k and Ê k+1 are the estimation
angle and the rotor speed of the generators can be obtained results before and after obtaining the new observation, P k and
from the measurements of the installed PMUs, which can P k+1 are the covariance matrices, and Gk+1 is the gain vector.
ensure the observability of the system, by using dynamic state The algorithm includes the following three steps:
estimation algorithms [27], [42]. 1) With the first two groups of data (k = 1 in (5)), set the
initial values of Ê 1 and P 1 to be (X 
1 X 1)
−1
X1Y1
 −1
B. RLS-Based MLE Estimation and (X 1 X 1 ) , respectively.
The MLE can be estimated by calculating the slope of 2) Obtain a new observation of the logarithmic distance
the logarithmic distance curve in Phase II. Considering the between the trajectories, and then sequentially calculate
influences of measurement errors and nonlinear fluctuations, Gk+1 , Ê k+1 , and P k+1 according to (10).
we adopt the least squares algorithm to estimate the MLE. 3) Set k = k + 1 and return to step 2.
For the linear segment of the logarithmic distance curve,
Compared with the normal least squares method, the RLS
the dominant development trend of the distance between the
method has a considerably lighter calculation burden, and can
original and neighboring trajectories can be well extracted by
improve the computational efficiency significantly. However, it
using the least squares estimation, which is very helpful for
should be clarified that the normal least squares and the RLS
performing fast and reliable stability assessment.
will have the same estimation results on the same waveform.
From (1) and Fig. 2, we see that the MLE is estimated
starting from time step m(n). From this time step, k + 1
sequential logarithmic distances can be obtained as C. Parameter Setting
Because the Theiler window w and the MLE estimation
L(m(n) + i) = log (d (m(n), n, i))
  initial time step m(n) determine the shape of the estimated
= log |θm(n)+i − θn+i | , i = 0, 1, . . . , k. (5) MLE curve, they are crucial for a quick and reliable rotor angle
Since the k + 1 logarithmic distance samples are located stability assessment. Here, we discuss the way of choosing
in Phase II of Fig. 1, the corresponding MLE (the slope of these parameters according to the rotor angle swing features.
the logarithmic distance curve in Phase II) can be estimated (1) Theiler window selection
by using the linear regression model that describes the linear
increase of the logarithmic distance with respect to time, as The Theiler window determines the temporal separation be-
tween the initial points θ0 and θm(0) . If the Theiler window is
L(m(n) + i) = λk · (m(n) + i)Δt + Ck + ξk , too small, the separation of the trajectories will be insufficient,
i = 0, 1, . . . , k, (6) causing interferences and fluctuations of the calculated MLE
where λk is the MLE to be estimated, Ck is the constant term, trajectory. On the other hand, a Theiler window that is too
and ξk is the residual term. large will result in unnecessary waiting time and delay the
The solution of the least squares estimation is stability assessment. Thus, choosing a proper Theiler window
 is crucial for performing quick and reliable rotor angle stability
λk
Ê k = = (X k Xk)
−1
Xk Y k, (7) assessment. However, in [3], it is implied that the Theiler win-
Ck dow is always set to one sampling step, which cannot always

ensure that the trajectories are fully separated. Therefore, the
relationship between the original and neighboring trajectories
cannot be quickly identified to estimate their convergence or
divergence trend, i.e., the stability condition of the system. On
the contrary, in the proposed approach, the Theiler window is
automatically determined according to the rotor angle swing
features, by checking the distance between the trajectories.
This approach can always ensure that the trajectories have
been sufficiently separated.
The post-fault rotor angles of the SDGPs have significant
swing patterns [44]–[46]. According to the features of the
relative rotor speed of the SDGPs, six distinct swing patterns
can be identified as shown in Fig. 3, and their theoretical
foundations are provided in Appendix. Different w are chosen
for different patterns as follows.
• Pattern I: In this pattern, the relative rotor speed in-
creases after fault clearing. No decelerating area exists
for the SDGP and the system will lose stability during
the first swing. Since the original and the neighboring
trajectories separate rapidly, w is set to 1 (the smallest
positive integer) to minimize the estimation waiting time.
• Pattern II: In this pattern, the relative rotor speed de-
creases after fault clearing. However, since the decel-
erating area is relatively small, the relative rotor speed
increases again after a short time period, and the in-
creasing trend continues until the system loses stability.
The key feature of this pattern is that the relative rotor
Fig. 3. Relative rotor speed curves of different patterns.
5
speed at the fault clearing moment, denoted by v0 in
Fig. 3, appears again after the initial decrease. To achieve
obvious separation between the trajectories, w is set to
be the time step lags of the reappearance of v0 .
• Pattern III: In this pattern, the relative rotor speed first
decreases to −v0 , and then exhibits periodic oscillations.
The decelerating area is large enough to reduce the
relative rotor speed to zero, and the system stability
depends on the damping characteristics of the post-fault
equilibrium point. In this case, w is set to be the time
step lags of the first appearance of −v0 .
• Pattern IV: In this pattern, the relative rotor speed first
decreases to some value greater than −v0 , and then
oscillates periodically. The key feature of this pattern
is that the relative rotor speed v0 and −v0 cannot be
observed after fault clearing. This pattern is a special case
of Pattern III, and w is set to be the time step lags of the
first appearance of the local minimal relative rotor speed
after fault clearing.
• Pattern V: This pattern is similar to Pattern III, and
usually appears after very quick fault clearing. The key
feature of this pattern is that the relative rotor speed shows
decelerated growth immediately after fault clearing, and
the relative rotor speed −v0 can be observed after that.
w is set in the same way as in Pattern III.
• Pattern VI: This pattern is a special case of Pattern
V. The key feature of this pattern is that the relative
rotor speed shows decelerated growth immediately after
fault clearing, and the relative rotor speed −v0 cannot be
observed during the oscillations. w is set in the same way
as in Pattern IV.
(2) MLE estimation initial time step selection
When calculating the MLE, the fluctuation segment of the
logarithmic distance curve (Phase I in Fig. 1) should be
skipped over to ensure the accuracy of the calculated MLE.
Otherwise, the dynamic characteristics of rotor angles after a
disturbance will be submerged by ruleless MLE fluctuations,
and the overall development trend of the MLE (positive or
negative as the time approaches infinity) cannot be quickly
identified from the initial variation features of the calculated
MLE trajectory. That is, the MLE estimation initial time step
m(n) should ensure that the slope is estimated for Phase II
of the logarithmic growth of the distance. In this phase, the
original and neighboring trajectories have been sufficiently
separated, and thus the logarithmic distance curve has shown
a clear development trend [17].
In fact, the distance between the rotor angle trajectories
has a close relationship with the relative rotor speed of the
SDGP. According to the definition of the Theiler window,
the distance between the corresponding points of the original
and the neighboring trajectories at any time step j can be

reformulated as
dj = d(m(0), 0, j) = θm(0)+j − θ0+j
⎛ ⎞
j+w
 j+w
 will increase immediately after the MLE estimation initial time
⎝
≈ θ0+j + ⎠
vt ·Δt −θ0+j = vt ·Δt , (11)
t=j+1 t=j+1
where vt is the relative rotor speed at the relevant time step.
It is revealed in (11) that the distance dj is equal to the area
enclosed by the time axis and the relative rotor speed curve
within the Theiler window, as indicated by the shaded regions
in Fig. 3. Therefore, m(n) can be determined as follows.
• Patterns I–II: With the selected w, it is seen in Fig. 3 that
the distance dj of these patterns monotonically increases
after fault clearing, indicating the logarithmic growth of
the distance in Phase II as soon as the fault is cleared.
Therefore, in these two patterns, m(n) is set to be w to
minimize the estimation waiting time.
• Patterns III–VI: For these patterns the distance will
exhibit periodic fluctuations after fault clearing. To ensure
that the trajectories have been sufficiently separated and
to capture the main trend of the fluctuations, m(n) is set
to be w + j ∗ , where j ∗ is the time step when dj reaches
its first local maximum after fault clearing. In fact, since
j ∗ = n = m(n) − w (see Fig. 2 and 3), j ∗ also indicates
the MLE estimation initial point of the original trajectory.
D. Rotor Angle Stability Assessment Criteria
When the MLE curve of an SDGP is estimated, the fol-
lowing criteria can be used to determine the stability of the
SDGP.
• Criterion I: If the MLE of the SDGP increases at the
beginning, the SDGP is unstable.
• Criterion II: If the MLE decreases at the beginning,
it will have oscillations. If the first peak point of the
oscillation is positive, the SDGP is unstable.
• Criterion III: If the MLE decreases at the beginning
and the first peak point of the oscillation is negative, the
SDGP is stable.
• Criterion IV: If all SDGPs are stable, the system is
stable; otherwise, the system is unstable.
Fig. 4. Typical MLE curves for Criterion I–III.
6
Typical MLE curves corresponding to these criteria are
shown in Fig. 4. If the condition in Criterion I is satisfied,
Pattern I or II in Fig. 3 will happen, for which the distance
step and the trend will last until the SDGP loses stability.
Criteria II and III correspond to Patterns III–VI. In these
patterns, the relative rotor angle will oscillate after fault
clearing. However, when the selected parameters are used,
the logarithmic distance will exhibit a significant development
trend as illustrated in Fig. 5. If the relative rotor speed
has an undamped oscillation, the SDGP is unstable and the
logarithmic distance curve will fluctuate periodically as shown
in Fig. 5(a). Since the MLE is the average slope of the
logarithmic distance curve from the MLE estimation initial
time step, the MLE will first decrease and then increase to a
positive peak value, as in Fig. 4. By contrast, if the relative
rotor speed has a damped oscillation, the SDGP is stable and
the logarithmic distance curve will be similar to Fig. 5(b).
The MLE will first decrease and then increase to a negative
peak value, as in Fig. 4. It should be noted that the MLE
is identical to the slope of the curve shown in Fig. 5, and
from the first cycle of the logarithmic distance variation, i.e.,
from the initial point to the first peak point, the variation
trend of the logarithmic distance (increase or decrease) can
be readily determined. Thus, in Criterion II and III, the first
peak point of the MLE curve should be considered for the
stability assessment. And the extensive test results in Section
IV have shown that the first peak point of the MLE curve is
effective on indicating the stability of the system.
Fig. 5. The logarithmic distance curves for Pattern III–VI.
Because SDGPs are responsible for the system dynamics
after disturbances [38], Criterion IV is obviously established
for the angle stability assessment of the whole system.
The proposed approach does not need a prespecified ob-
serving window to identify the sign of the calculated MLE
for the stability assessment, and thus can provide faster and
more reliable stability assessment results compared with the
benchmark approaches.
E. Rotor Angle Stability Assessment Procedure
The proposed online rotor angle stability assessment pro-
cedure is shown in Fig. 6, which includes the measurement
data preparation module, the parameter setting module, the
MLE estimation module, and the stability assessment module.
Specifically:

1) When a fault is detected, the measurement data prepa-
ration module will be immediately activated to identify
the SDGPs and collect the corresponding relative rotor
angle and rotor speed measurements in real time.
2) The relative rotor speed variation patterns are identified
online according to the rules described in Section III.C,
and the parameters are assigned accordingly.
3) The MLE sequences are calculated by using the RLS-
based algorithm presented in Section III.B.
4) Finally, the stability condition is assessed by using
the criteria provided in Section III.D according to the
features of the estimated MLE curves.
IV. C ASE S TUDIES
The proposed approach is tested on the New England 39-
bus system and the NPCC 140-bus system. Simulations are
performed with the Power System Analysis Toolbox (PSAT)
[47] in MATLAB. And all tests are carried out on a 3.2-GHz
Intel(R) Core(TM) i7-4790S based desktop.
A. New England 39-Bus System
The New England 39-bus system has 10 generators and
46 branches. The parameters can be found in [48]. Unless
otherwise specified, all generators in the tests are described by
the fourth-order transient model with Type I turbine governor
(TG), Type II automatic voltage regulator (AVR), and Type
II power system stabilizer (PSS) (see PSAT documentations).
Fig. 6. Flow chart of the assessment procedure.
7
All loads are described by the ZIP model and the ratios of
the constant impedance, constant current, and constant power
loads are 0.4, 0.5, and 0.1, respectively. The sampling rate
of the PMU measurements used for MLE estimation is 120
samples/s [3].
To verify the effectiveness of the proposed approach, a
three-phase-to-ground fault is applied at bus 2 at t = 1 s, and
the fault is cleared by opening line 2–3 at tc = 1.243 s and
tc = 1.244 s for Scenarios I and II, respectively. According
to the time-domain simulation, the system is stable under
Scenario I and unstable under Scenario II.
By using the method outlined in Section III.A, the generator
pairs 38–39 and 37–39 are identified as the SDGPs for both
scenarios. Figures 7–8 show the relative rotor angles and the
estimated MLEs of the SDGPs.
From Fig. 7(a), it is seen that the relative rotor angles of the
SDGPs tend to be stable after a long period of oscillations. In
Fig. 7(b), the MLEs of the SDGPs decrease immediately from
the initial points of the MLE estimation. The first peak points
of the oscillations are both less than 0. Therefore, according
to Criteria III and IV, the system is stable under this scenario.
In contrast, under Scenario II, the MLE of the generator
pair 38–39 increases immediately from the initial point of the
MLE estimation, as shown in Fig. 8(b). Therefore, according
to Criteria I and IV, the system is unstable under this scenario.
For the two scenarios, the system stability can be assessed,
respectively, within 2.82 s and 1.40 s after fault clearing,
demonstrating that the proposed approach can detect the insta-
bility early. In fact, according to the proposed approach, the
first-swing instability can be identified very quickly, because
the instability can be detected at the beginning of the MLE
curve. On the other hand, for the multiswing stability or
Fig. 7. Simulation results of Scenario I.
Fig. 8. Simulation results of Scenario II.

instability, the assessment time is slightly longer since the
approach needs to check the first peak point of the MLE curve.
In industrial applications, a predetermined relative rotor
angle value, i.e., π rad, is usually set as the threshold for
determining rotor angle stability [49]. Although this pragmatic
criterion is easy to execute, the relative rotor angle correspond-
ing to rotor angle instability may change significantly with
changes in topologies, parameters, and operating conditions. In
all tests, eight times of misclassification exist if π is used as the
stability assessment threshold, and the proportion is about 1%.
Moreover, if more critically stable scenarios are involved, more
misclassification cases will be observed. In [50], it is explicitly
stated that setting a fixed threshold of the relative rotor angle
for stability assessment is dangerous. It may produce wrong
stability assessment results, e.g., misclassify a stable scenario
as an unstable scenario. For instance, under Scenario I, the
relative rotor angle between generators 38 and 39 can reach
up to 3.207 rad while the system is still stable.
It should also be noted that foundational differences exist
between the proposed approach and the approach proposed
in [3]. In the proposed approach, the system stability can be
explicitly determined at the latest when the first peak point
of the MLE curve is observed. In contrast, for the approach
proposed in [3], the MLE curve must be observed for a
long period to ensure the sign of the MLE. Unfortunately,
in the approach proposed in [3], predetermining the observing
window size to provide reliable and timely assessment results
is actually difficult (for instance, see Figs. 1–2 in [3]). From
this perspective, the proposed approach is more time-efficient
and more reliable.
Furthermore, a new three-phase-to-ground fault is applied
at bus 26 at t = 1 s, and the fault is cleared by opening line
25-26 at tc = 1.08 s, tc = 1.147 s, tc = 1.148 s, and tc =
1.16 s, respectively. According to the time-domain simulation,
the system is critically stable when the fault is cleared at tc =
1.147 s, and the system is critically unstable when the fault is
cleared at tc = 1.148 s. In the tests, the generator pair 38–39
is identified as the SDGP. For comparison, the relative rotor
angle curves or the estimated MLE curves are plotted in the
same figure, i.e., Fig. 9(a) shows the relative rotor angles and
Fig. 9(b) shows the estimated MLE curves.
Fig. 9. Simulation results of the fault at bus 26 for different tc .
From Fig. 9(a), it can be seen that the relative rotor angle
curves of the SDGP change gradually from stable to unstable
8
in the four cases. In Fig. 9(b), the stability of all cases can
be successfully identified from the estimated MLE curves
according to the proposed criteria. Meanwhile, for the unstable
cases, it can be found that the more serious the fault is,
the shorter the time required for the stability assessment is.
Moreover, the time required to identify an unstable case is
much shorter than the time required to identify a stable case.
To further verify the accuracy of the proposed approach,
extensive tests were performed on the New England 39-
bus system. Specifically, a three-phase-to-ground fault was
created for each bus (except the generator buses) at t = 1
s, and was cleared at 1.08 s, 1.16 s, 1.24 s, and 1.32 s,
respectively. The test results show that the proposed approach
successfully determines the rotor angle stability in all the 224
tests. The occurrence frequencies of the fault patterns and the
success rate of the proposed stability assessment approach are
summarized in Table I.
TABLE I
S UMMARY OF THE T ESTS CONDUCTED ON THE N EW E NGLAND 39-B US
S YSTEM
Occurrence Times of the Swing Patterns
tc /s
I II III IV V VI
1.08 N/A N/A 4 43 6 3
1.16 N/A 6 10 36 3 1
1.24 4 13 9 27 2 1
1.32 38 12 2 4 N/A N/A
Success Rate 100% 100% 100% 100% 100% 100%
It is seen in Table I that Patterns I and II that correspond to
first-swing instability usually occur with longer fault clearing
times. In contrast, Patterns V and IV usually occur with a
shorter fault clearing time, and the system will have a good
chance to maintain stability in these patterns. Among the tests,
the first-swing instability identification time ranges from 1.2 s
to 1.5 s, and the multiswing stability assessment time ranges
from 2.2 s to 2.5 s, further confirming the efficiency of the
proposed approach. Meanwhile, for the simulations of the
39-bus system, the time consumption of the SDGP selection
ranges from 25 μs to 50 μs, and this process is executed only
once after detecting a credible fault. For each sampling step,
the time needed for obtaining the logarithmic distance ranges
from 4 μs to 8 μs, and the MLE calculation time ranges
from 30 μs to 40 μs. Thus, the longest calculation time for
each sampling step will be shorter than 0.1 ms (including
the time consumption of SDGP selection), fully satisfying the
requirements of the online assessment.
Because the system has sufficient damping to ensure the
stability of the post-fault equilibrium points, there are no mul-
tiswing instability cases. To generate a multiswing instability
case, all the PSSs are removed and the parameters of the AVRs
are tuned. In the modified system, a three-phase-to-ground
fault is applied at bus 28 at t = 1 s, and is cleared by opening
line 27–28 at tc = 1.12 s. In this test, only the generator pair
38–39 is identified as the SDGP, whose relative rotor angle
curve and MLE curve are shown in Fig. 10.
In Fig. 10(b), the MLE decreases immediately from the
initial point of the MLE estimation, and the following peak
MLE is positive. According to Criteria II and IV, the MLE

Fig. 10. Simulation results of the three-phase-to-ground fault at bus 28.
variation pattern indicates that the system is multiswing un-
stable, which is verified by the relative rotor angle curve. In
this case, the stability assessment time is 2.36 s, which is
considerably shorter than the time required for directly looking
at the relative rotor angle curve.
A similar phenomenon occurs in Fig. 11 and Fig. 12. In
Fig. 11, a three-phase-to-ground fault is applied at bus 15 at
t = 1 s, and is cleared by opening line 15-16 at tc = 1.07 s,
in which the generator pair 37–39 is identified as the SDGP.
In Fig. 12, a three-phase-to-ground fault is applied at bus 21
at t = 1 s, and is cleared by opening line 21-22 at tc = 1.07 s,
in which the generator pair 38–39 is identified as the SDGP.
Fig. 11. Simulation results of the three-phase-to-ground fault at bus 15.
Fig. 12. Simulation results of the three-phase-to-ground fault at bus 21.
During the tests shown in Fig. 11 and Fig. 12, the multi-
swing instability can be identified timely and accurately by
using the proposed approach, and the analysis of the tests is
similar to that of the test shown in Fig. 10.
9
B. NPCC 140-Bus System
The proposed approach is also tested on the NPCC 140-bus
system [51]. The settings are the same as those for the New
England 39-bus system.
A three-phase-to-ground fault is first applied at bus 35 at
t = 0.1 s, and is cleared by opening line 34–35 at tc = 0.307 s
(Scenario III, stable) and tc = 0.308 s (Scenario IV, unstable),
respectively. Generator pairs 1–48 and 2–48 are identified as
the SDGPs for both cases. The relative rotor angle curves and
MLE curves under these two scenarios are shown in Figs.
13–14, respectively.
In Fig. 13, it is seen that the curves of Scenario III
are similar to those of Scenario I. In this case, the system
is determined as stable because of the same reason as for
Scenario I. As shown in Fig. 14, under Scenario IV, the MLE
curves of both SDGPs exhibit unstable features, indicating that
both generator pairs will lose stability and hence the system
will be unstable.
Furthermore, a new three-phase-to-ground fault is applied
at bus 12 at t = 0.1 s, and the fault is cleared by opening line
12-13 at tc = 0.24 s, tc = 0.267 s, tc = 0.268 s, and tc =
0.32 s, respectively. According to the time-domain simulation,
the system is critically stable when the fault is cleared at tc =
0.267 s, and the system is critically unstable when the fault is
cleared at tc = 0.268 s. In the above four cases, the generator
pairs 1–48 and 6–48 are identified as the SDGPs. Fig. 15
shows the relative rotor angle curves and the estimated MLE
curves of the SDGPs. As shown in Fig. 15(b), in all cases
the stability of the SDGP 1-48 can be successfully identified
according to the proposed criteria, within 2.07 s, 2.42 s, 1.54
s, and 0.38 s after fault clearing, respectively. Meanwhile, as
Fig. 13. Simulation results of Scenario III.
Fig. 14. Simulation results of Scenario IV.
 10

shown in Fig. 15(d), the stability of the generator pair 6–48 In the proposed method, the calculated MLE can capture
can be successfully identified within 2.29 s, 2.47 s, 1.76 s, and the divergence/convergence features of the relative rotor angle
0.64 s after fault clearing, respectively. Thus, in the four cases, trajectories of the SDGPs accurately and timely. Meanwhile,
the time consumptions of the stability assessment of the whole the identified severely disturbed generators (crucial generators)
system are 2.29 s, 2.47 s, 1.54 s, and 0.38 s, respectively. are usually the major control objects in emergency rotor
Extensive tests are also executed on the NPCC 140-bus angle stability control [33]. Therefore, the MLE estimation
system. A three-phase-to-ground fault is applied at each bus results can provide valuable indications on when and where to
(except the generator buses) at t = 0.1 s, and is cleared at perform the emergency control, making the proposed method
0.18 s, 0.26 s, 0.32 s, and 0.40 s, respectively. Table II lists helpful for designing automatically triggered counteractions to
the occurrence frequencies of the fault patterns and the success prevent the crucial generators from falling out-of-step.
rate of the proposed stability assessment approach during the Compared with IEEAC or SIME [52], the proposed stability
tests. assessment approach is purely data-driven, without involving
any system model or parameter; hence, it can avoid incorrect
TABLE II assessment results caused by parameter errors. Meanwhile, in
S UMMARY OF THE T ESTS ON THE NPCC 140-B US S YSTEM
the proposed approach, the preprocesses of IEEAC or SIME,
Occurrence Times of the Swing Patterns
tc /s e.g., the system equivalent, simplification, and homology units
I II III IV V VI
0.18 N/A N/A 6 166 2 5
identification, are not required, and thus the misleading as-
0.26 N/A 10 26 143 N/A N/A sessment results caused by improper preprocesses can also
0.32 8 57 38 76 N/A N/A be avoided. Moreover, in the proposed approach, the MLE
0.40 60 12 49 58 N/A N/A can be easily estimated by using the RLS, only involving
Success Rate 100% 100% 100% 100% 100% 100% arithmetic operations. Therefore, the proposed approach has
a considerably lighter computation burden compared with
According to the test results, the proposed approach can IEEAC or SIME, which needs to map the trajectories of all
accurately determine system stability in all 716 tests. The first- units to the trajectory of the aggregated single machine and
swing instability identification time ranges from 1.1 s to 1.7 s, involves complex numerical integration operations. Therefore,
and the multiswing stability assessment time ranges from 1.8 s the proposed approach is more suitable for online application
to 2.4 s. Meanwhile, for the simulations of the 140-bus system, in bulk power systems and can provide timely assessment
the time consumption of the SDGP selection ranges from 30 results.
μs to 60 μs. The calculation time for obtaining the logarithmic
distance at each sampling step ranges from 4 μs to 10 μs. In V. D ISCUSSIONS
addition, the MLE calculation time ranges from 30 μs to 50
μs. Thus, the longest calculation time for each sampling step A. Discussions on Noise and Sampling Rate
is within 0.12 ms, which illustrates that the proposed approach Obviously, the existence of noise in the data will impact the
runs fast enough to meet the online calculation requirements performance of any measurement-based method. Nevertheless,
for this larger system. owing to the effect of dynamic state estimation, the noise in
the raw data obtained from PMUs can be perfectly filtered,
as has been fully verified in the corresponding papers [22]–
[30]. Moreover, state estimation errors must exist no matter
which kind of estimation algorithm is applied. Despite this,
the dominant development trend of state variables, e.g., rotor
angles, can at least be truly reflected in the estimation results.
Under this circumstance, the proposed approach can still work
well as the development trend of rotor angles is the only
concern of the approach. Besides, with the development of
dynamic estimation algorithms, the reliability of the proposed
approach can also be improved.
The sampling rate is another important aspect that should be
considered when designing a data-driven analysis algorithm.
However, the sampling rate of PMUs will not affect the effec-
tiveness of the proposed approach significantly, because the
approach only focuses on the dominant development trend of
the relative rotor angle of SDGP. As long as the sampling cycle
is significantly shorter than the swing cycle of the relative
rotor angle, the swing features can be truthfully reflected by
the calculated MLE curve. In practice, the sampling rate of
a commercially available PMU ranges from 30 samples per
second to 120 samples per second, and the dynamic state
Fig. 15. Simulation results of the fault at bus 12 for different tc . estimation can provide the estimation results in real time

(shorter than a sampling cycle [29]), and can sufficiently
satisfy the requirement of our approach. For instance, Fig.
16 shows the estimated MLE curves under different sampling
rates for the critically stable and unstable cases of Fig. 9. From
Fig. 16, it can be seen that the sampling rate does not affect
the effectiveness of the proposed approach. Furthermore, under
both sampling rates (30 samples per second and 120 samples
per second), the success rates of the stability assessment for
the 39-bus system and the 140-bus system are always 100%.
Fig. 16. MLE curves of θ(38,39) under different sampling rates.
B. Discussions on the Universal Parameter Assignment Rule
Although the algorithm based on the six patterns works
quite well in the extensive tests, the categorization of the six
patterns is not easy to be strictly proved in theory. Thus,
it’s necessary to provide a universal parameter assignment
rule that has a more solid theoretical foundation in order to
further improve the understanding of the proposed approach.
Fortunately, such a universal parameter assignment rule, which
does not depend on the post-fault patterns, does exist.
As having been described, the Theiler window should be
large enough to ensure that the original and neighboring
trajectories obtained from the same time series have been fully
separated. Meanwhile, according to the post-fault dynamics of
power systems, the relative rotor angle (as well as the relative
rotor speed) of a generator pair will experience a cyclical
swing after fault clearing. Thus, it can be imagined that the
half swing circle of the relative rotor angle (or speed) can be a
good candidate for the Theiler window w in theory, since the
observed points on the original and neighboring trajectories
are always approximately opposite for a half swing circle lag.
In practice, the half swing cycle can be estimated quickly
by various methods, e.g., the FFT-based frequency estimation
approach in [53] or the simple frequency calculation approach
in [54], for the SDGPs. It should be noted that although the
half swing cycle is a good value for the Theiler window, the
Theiler window selection allows a certain level of deviation
as long as the trajectories can be separated sufficiently.
Correspondingly, the MLE estimation initial time step
m(n), which ensures that the MLE estimation is performed
in Phase II of the logarithmic distance curve, can be selected
as w + j ∗ , i.e., the time step of the observed point on the
neighboring trajectory when the separation distance between
11
the rotor angle trajectories reaches the first local maximum.
Here, a special case is the first-swing instability. In this case,
the separation distance will increase monotonically, and no
local maximum logarithmic distance can be found. Therefore,
if a monotonic increase trend of the logarithmic distance is
detected, m(n) can be set to w to shorten the estimation
waiting time. The stability criteria described in Section III.D
are still valid when the aforementioned parameter assignment
rule is applied to the SDGPs.
This universal parameter assignment rule is strict in theory
since it does not assume that the post-fault dynamics of an
SDGP are similar to those of the single-machine infinite bus
system. Nevertheless, in most test cases, the dynamics of
SDGPs are more or less similar to the single-machine infinite
bus system [44]–[46], and the stability can be successfully
assessed according to the approach based on the six patterns.
In fact, it can be observed that the parameter assignment rule
respecting the six patterns is an instance of the universal rule
under the situation that the SDGPs have similar post-fault
dynamics as the single-machine infinite bus system, in which
the MLE calculation waiting time can be shorten significantly
since the value of w is smaller and the online identification of
the half swing cycle can be avoided.
VI. C ONCLUSION
In this paper, a model-free approach for online rotor angle
stability assessment was proposed based on the MLE. By using
the proposed MLE estimation algorithm, parameter setting
rules, and the stability criteria, the approach can identify the
system stability condition online by using PMU measurements.
The approach does not need a predetermined observing win-
dow to identify the sign of the MLE, and can provide reliable
and timely assessment results by analyzing the features of
the estimated MLE curve. To verify the performance of the
proposed approach, extensive tests were performed on the New
England 39-bus system and the NPCC 140-bus system. The
proposed approach could successfully determine the system
stability conditions in all 945 tests. Moreover, among all the
tests, the first-swing stability could be assessed within 1.7 s
and the multiswing stability could be assessed within 2.5 s.
In transient stability assessment, the stability margin is an
important indicator for power system dispatch and control.
Unfortunately, MLE-based stability assessment approaches
cannot be directly used for estimating the stability margin.
To address this issue will be our next research focus.
A PPENDIX
T HEORETICAL F OUNDATIONS OF THE S WING PATTERNS
First, for a single-machine infinite bus system, the post-
fault variation scenarios of the relative rotor speed can be
categorized into the six patterns straightforwardly.
Fig. 17 shows the power–angle diagrams for three network
conditions of the system: (i) pre-fault, (ii) during fault, and (iii)
post-fault. In the figure, δ0 is the rotor angle corresponding
to the pre-fault equilibrium point a, δk is the rotor angle
corresponding to the post-fault equilibrium point k, δc is the
fault clearing angle, δcm is the critical fault clearing angle, δb is

the rotor angle corresponding to the post-fault operating point
h (it is an unstable equilibrium point), A1 is the accelerating
area, and A2 is the decelerating area.
Fig. 17. Power–angle diagrams.
From the equal area criterion and the rotor-motion equations
[55], it is easily known that the six patterns are all the possible
post-fault patterns for the single-machine infinite bus system.
For different fault clearing times, the corresponding patterns
are listed below.
• Pattern I: δc ≥ δb . As shown in the figure, in this pattern,
no decelerating area exists after fault clearing, and the
accelerating power, i.e., Pm − Pe3 , will keep increasing,
making the relative rotor speed increase continuously.
• Pattern II: δcm < δc < δb . In this pattern, a small
decelerating area exists; hence, the relative rotor speed
decreases after fault clearing. However, since the decel-
erating area is not large enough, the relative rotor speed
increases again after a short while, and the accelerating
trend continues until the system loses stability.
• Pattern III and IV: δk ≤ δc ≤ δcm . In these two
patterns, the relative rotor speed will decrease after fault
clearing. As the decelerating area is sufficiently large, the
relative rotor speed will decrease from v0 (the relative
rotor speed at the fault clearing moment) to 0, and then
reversely increase to −v0 (Pattern III) or some value
between −v0 and 0 (Pattern IV), depending on the co-
actions of the inverse driving torque and the damping
torque. In these two patterns, the rotor speed will exhibit
periodic oscillations, and the system stability depends on
the dynamic characteristics of the post-fault equilibrium
point [55].
• Pattern V and VI: δ0 < δc < δk . In these two patterns,
a small accelerating area exists after fault clearing, but
the accelerating power, i.e., Pm − Pe3 , keeps decreasing,
making the relative rotor speed show decelerated growth
after fault clearing. Subsequently, when the rotor angle
becomes greater than δk , the relative rotor speed will
begin to decrease since Pe3 becomes greater than Pm .
Then, as in Pattern III and Pattern IV, the relative rotor
speed will decrease to 0, and reversely increase to −v0
(Pattern V) or some value between −v0 and 0 (Pattern
VI). In these two patterns, the rotor speed will also
exhibit periodic oscillations, and the system stability also
12
depends on the dynamic characteristics of the post-fault
equilibrium point.
Second, in an interconnected bulk system, for a given
disturbance, it is reasonable to assume that only one or a
few machines are severely disturbed. These generators are
responsible for the system stability under the disturbance,
while the rest of the machines are relatively insensitive to
the disturbance [38]. The dynamic behavior of the severely
disturbed generator can be reflected by the SDGP, which is
composed of the severely disturbed generator and the least
disturbed generator. Thus, the stability of the whole system
can be assessed by analyzing the post-fault behavior of relative
rotor angles of the SDGPs [39].
Moreover, since the behavior of the least disturbed generator
is similar to that of the infinite bus, i.e., the generator will
approximately maintain the pre-fault operating condition, the
dynamic characteristics of the severely disturbed generator
with respect to the least disturbed generator will be similar to
those of the single generator with respect to the infinite bus in
the single-machine infinite bus system [44]–[46]. Therefore,
the post-fault relative rotor speed variation patterns for the
single-machine infinite bus system are also suitable for the
SDGPs.
R EFERENCES
[1] P. Kundur, J. Paserba, V. Ajjarapu, G. Andersson, A. Bose et al.,
“Definition and classification of power system stability,” IEEE Trans.
Power Syst., vol. 19, no. 3, pp. 1387–1401, Aug. 2004.
[2] J. Yan, C. C. Liu, and U. Vaidya, “PMU-based monitoring of rotor angle
dynamics,” IEEE Trans. Power Syst., vol. 26, no. 4, pp. 2125–2133, Mar.
2011.
[3] S. Dasgupta, M. Paramasivam, U. Vaidya, and V. Ajjarapu, “PMU-based
model-free approach for real-time rotor angle monitoring,” IEEE Trans.
Power Syst., vol. 30, no. 5, pp. 2818–2819, Sep. 2015.
[4] V. Centeno, A. Phadke, A. Edris, J. Benton, M. Gaudi, and G. Michel,
“An adaptive out-of-step relay,” IEEE Trans. Power Delivery, vol. 12,
no. 1, pp. 61–71, Jan. 1997.
[5] J. H. Chow, A. Chakrabortty, M. Arcak, B. Bhargava, and A. Salazar,
“Synchronized phasor data based energy function analysis of dominant
power transfer paths in large power systems,” IEEE Trans. Power Syst.,
vol. 22, no. 2, pp. 727–734, May 2007.
[6] C. W. Liu and J. S. Thorp, “New methods for computing power system
dynamic response for real-time transient stability prediction,” IEEE
Trans. Circuits Syst. I, vol. 47, no. 3, pp. 324–337, Mar. 2000.
[7] K. Sun, S. Likhate, V. Vittal et al., “An online dynamic security
assessment scheme using phasor measurements and decision trees,”
IEEE Trans. Power Syst., vol. 22, no. 4, pp. 1935–1943, Nov. 2007.
[8] A. Del Angel, P. Geurts, D. Ernst, M. Glavic, and L. Wehenkel,
“Estimation of rotor angles of synchronous machines using artificial
neural networks and local PMU-based quantities,” Neurocomputing,
vol. 70, no. 16, pp. 2668–2678, Oct. 2007.
[9] I. Kamwa, S. Samantaray, and G. Joos, “Development of rule-based
classifiers for rapid stability assessment of wide-area post-disturbance
records,” IEEE Trans. Power Syst., vol. 24, no. 1, pp. 258–270, Feb.
2009.
[10] C. W. Liu, J. S. Thorp, J. Lu et al., “Detection of transiently chaotic
swings in power systems using real-time phasor measurements,” IEEE
Trans. Power Syst., vol. 9, no. 3, pp. 1285–1292, Aug. 1994.
[11] S. K. Khaitan and J. D. McCalley, “Vantage: A lyapunov exponents
based technique for identification of coherent groups of generators in
power systems,” Electr. Pow. Syst. Res., vol. 105, pp. 33–38, Dec. 2013.
[12] A. Wolf, J. B. Swift, H. L. Swinney, and J. A. Vastano, “Determining
lyapunov exponents from a time series,” PHYSICA D, vol. 16, no. 3,
pp. 285–317, Jul. 1985.
[13] J. P. Eckmann and D. Ruelle, “Ergodic theory of chaos and strange
attractors,” REV MOD PHYS, vol. 57, no. 3, pp. 617–656, Jul. 1985.

[14] J. P. Eckmann, S. O. Kamphorst, D. Ruelle, and S. Ciliberto, “Liapunov
exponents from time series,” PHYS REV A, vol. 34, no. 6, pp. 4971–
4979, Dec. 1986.
[15] J. P. Eckmann and D. Ruelle, “Fundamental limitations for estimating
dimensions and lyapunov exponents in dynamical systems,” PHYSICA
D, vol. 56, no. 2–3, pp. 185–187, May. 1992.
[16] S. Dasgupta, M. Paramasivam, U. Vaidya, and V. Ajjarapu, “Real-time
monitoring of short-term voltage stability using PMU data,” IEEE Trans.
Power Syst., vol. 28, no. 4, pp. 3702–3711, Nov. 2013.
[17] C. Skokos, G. A. Gottwald, and J. Laskar, Chaos Detection and
Predictability. Heidelberg, Germany: Springer, 2016.
[18] V. I. Oseledec, “A multiplicative ergodic theorem. lyapunov characteris-
tic numbers for dynamical systems,” Trans. Moscow Math. Soc, vol. 19,
no. 2, pp. 197–231, Dec. 1968.
[19] M. T. Rosenstein, J. J. Collins, and C. J. De Luca, “A practical method
for calculating largest lyapunov exponents from small data sets,” Physica
D, vol. 65, no. 1-2, pp. 117–134, May 1993.
[20] J. Gao and Z. Zheng, “Local exponential divergence plot and optimal
embedding of a chaotic time series,” Phys. Lett. A, vol. 181, no. 2, pp.
153–158, Oct. 1993.
[21] H. Kantz, “A robust method to estimate the maximal lyapunov exponent
of a time series,” Phys. Lett. A, vol. 185, no. 1, pp. 77–87, Jan. 1994.
[22] Z. Huang, K. Schneider, and J. Nieplocha, “Feasibility studies of apply-
ing kalman filter techniques to power system dynamic state estimation,”
in 2007 International Power Engineering Conference (IPEC 2007).
IEEE, Dec. 2007, pp. 376–382.
[23] E. Ghahremani and I. Kamwa, “Dynamic state estimation in power
system by applying the extended kalman filter with unknown inputs
to phasor measurements,” IEEE Trans. Power Syst., vol. 26, no. 4, pp.
2556–2566, Nov. 2011.
[24] K. Sun, J. Qi, and W. Kang, “Power system observability and dynamic
state estimation for stability monitoring using synchrophasor measure-
ments,” Control Eng. Pract., vol. 53, pp. 160–172, Aug. 2016.
[25] E. Ghahremani and I. Kamwa, “Local and wide-area pmu-based de-
centralized dynamic state estimation in multi-machine power systems,”
IEEE Trans. Power Syst., vol. 31, no. 1, pp. 547–562, Jan. 2016.
[26] J. Qi, K. Sun, J. Wang, and H. Liu, “Dynamic state estimation for
multi-machine power system by unscented Kalman filter with enhanced
numerical stability,” IEEE Trans. Smart Grid, to be published.
[27] J. Qi, K. Sun, and W. Kang, “Optimal PMU placement for power system
dynamic state estimation by using empirical observability gramian,”
IEEE Trans. Power Syst., vol. 30, no. 4, pp. 2041–2054, Jul. 2015.
[28] M. A. Ariff, B. C. Pal, and A. K. Singh, “Estimating dynamic model
parameters for adaptive protection and control in power system,” IEEE
Trans. Power Syst., vol. 30, no. 2, pp. 829–839, Mar. 2015.
[29] N. Zhou, D. Meng, Z. Huang, and G. Welch, “Dynamic state estimation
of a synchronous machine using PMU data: A comparative study,” IEEE
Trans. Smart Grid, vol. 6, no. 1, pp. 450–460, Jan. 2015.
[30] A. F. Taha, J. Qi., J. Wang, and J. H. Panchal, “Risk mitigation for
dynamic state estimation against cyber attacks and unknown inputs,”
IEEE Trans. Smart Grid, to be published.
[31] F. Takens, “Detecting strange attractors in turbulence,” in Dynamical
systems and turbulence, Warwick 1980. Heidelberg, Germany: Springer,
1981, pp. 366–381.
[32] N. H. Packard, J. P. Crutchfield, J. D. Farmer, and R. S. Shaw, “Geometry
from a time series,” Phys. Rev. Lett., vol. 45, no. 9, pp. 712–716, Nov.
1980.
[33] P. Bhui and N. Senroy, “Real time prediction and control of transient
stability using transient energy function,” IEEE Trans. Power Syst.,
vol. 32, no. 2, pp. 923–934, Mar. 2017.
[34] H. Kantz and T. Schreiber, Nonlinear time series analysis. Cambridge
university press, 2004.
[35] E. R. Deyle and G. Sugihara, “Generalized theorems for nonlinear state
space reconstruction,” PLoS One, vol. 6, no. 3, p. e18295, Mar. 2011.
[36] M. P. McCullough, Phase space reconstruction: Methods in applied
economics and econometrics. Washington State University, 2008.
[37] A. Röbel, “Synthesizing natural sounds using dynamic models of sound
attractors,” COMPUT MUSIC J, vol. 25, no. 2, pp. 46–61, Mar. 2001.
[38] M. Haque and A. Rahim, “Determination of first swing stability limit
of multimachine power systems through taylor series expansions,” IEE
Proc.-C, vol. 136, no. 6, pp. 373–380, Nov. 1989.
[39] M. Yin, C. Chung, K. Wong, Y. Xue, and Y. Zou, “An improved iterative
method for assessment of multi-swing transient stability limit,” IEEE
Trans. Power Syst., vol. 26, no. 4, pp. 2023–2030, Nov. 2011.
[40] M. Haque, “Further developments of the equal-area criterion for mul-
timachine power systems,” Electr. Pow. Syst. Res., vol. 33, no. 3, pp.
175–183, Jun. 1995.
13
[41] L. Fang and Y. Ji-Lai, “Transient stability analysis with equal area
criterion directly used to a non-equivalent generator pair,” in 2009
International Conference on Power Engineering, Energy and Electrical
Drives. IEEE, 2009, pp. 386–389.
[42] J. Qi, K. Sun, and W. Kang, “Adaptive optimal pmu placement based on
empirical observability gramian,” IFAC-PapersOnLine, vol. 49, no. 18,
pp. 482–487, Aug. 2016.
[43] P. C. Young, Recursive estimation and time-series analysis: An introduc-
tion for the student and practitioner. Heidelberg, Germany: Springer,
2011.
[44] B. Spalding, H. Yee, and D. Goudie, “Coherency recognition for stability
studies using singular points,” IEEE Trans. Power App. Syst., vol. 96,
no. 4, pp. 1368–1375, Jul. 1977.
[45] M. Takahashi, K. Matsuzawa, M. Sato, K. Omata, R. Tsukui, T. Naka-
mura, and S. Mizuguchi, “Fast generation shedding equipment based
on the observation of swings of generators,” IEEE Trans. Power Syst.,
vol. 3, no. 2, pp. 439–446, May 1988.
[46] G. G. Karady, A. A. Daoud, and M. A. Mohamed, “On-line stability
enhancement using multi-agent technique,” in Proc. IEEE Power Eng.
Soc. Winter Meeting, 2002, vol. 2, pp. 893–899.
[47] F. Milano, “An open source power system analysis toolbox,” IEEE Trans.
Power Syst., vol. 20, no. 3, pp. 1199–1206, Aug. 2005.
[48] A. Pai, Energy function analysis for power system stability. Boston,
MA, USA: Kluwer Academic Publishers, 1989.
[49] Y. Xu, Z. Y. Dong, J. H. Zhao, P. Zhang, and K. P. Wong, “A reliable
intelligent system for real-time dynamic security assessment of power
systems,” IEEE Trans. Power Syst., vol. 27, no. 3, pp. 1253–1263, Aug.
2012.
[50] Y. Xue, V. C. Thierry, and M. Ribbens-Pavella, “A simple direct method
for fast stability assessment of large power systems,” IEEE Trans. Power
Syst., vol. 3, no. 2, pp. 400–412, May. 1988.
[51] J. H. Chow, R. Galarza, P. Accari, and W. Price, “Inertial and slow
coherency aggregation algorithms for power system dynamic model
reduction,” IEEE Trans. Power Syst., vol. 10, no. 2, pp. 680–685, May
1995.
[52] M. Pavella, D. Ernst, and D. Ruiz-Vega, Transient Stability of Power
Systems: A Unified Approach to Assessment and Control. Kluwer
Academic Publishers, 2000.
[53] H. Wen, S. Guo, Z. Teng, F. Li, and Y. Yang, “Frequency estimation
of distorted and noisy signals in power systems by fft-based approach,”
IEEE Trans. Power Syst., vol. 29, no. 2, pp. 765–774, Mar. 2014.
[54] B. Wang, X. Su, and K. Sun, “Properties of the frequency–amplitude
curve,” IEEE Trans. Power Syst., vol. 32, no. 1, pp. 826–827, Jan. 2017.
[55] P. Kundur, Power System Stability and Control. New York: McGraw-
hill, 1994.
Shaopan Wei (S’16) received the B.S. degree in
the School of Electrical Engineering, Hebei Univer-
sity of Technology, Tianjin, China, in 2015. He is
currently pursuing the M.S. degree in the School of
Electrical Engineering, Shandong University, Jinan,
China. His research interest is stability analysis of
power systems.
 14

Ming Yang (M’09) received the B.E. and Ph.D
degrees in electrical engineering from Shandong
University, Jinan, China, in 2003 and 2009, respec-
tively. He was an exchange Ph.D. student at Energy
System Research Center, The University of Texas at
Arlington, Arlington, TX, USA, from October 2006
to October 2007. He conducted postdoctoral research
at School of Mathematics of Shandong University
from July 2009 to July 2011. From November 2015
to October 2016 he was a Visiting Scholar at the
Energy Systems Division, Argonne National Labo-
ratory, Argonne, IL, USA.
Currently, he is an Associate Professor at Shandong University and an
associate editor of IEEE Transactions on Industry Applications. His research
interests are power system optimal operation and control.
Xueshan Han received the B.E. and M.S. degrees in
electrical engineering from Northeast Dianli Univer-
sity, Jilin, China, in 1984 and 1989, respectively. He
received the Ph.D degree in electrical engineering
from Harbin Institute of Technology, Harbin, China,
in 1994. He conducted postdoctoral research at
Nanyang Technological University, Singapore, from
Oct. 1998 to Oct. 2000. He is now a full professor
of electrical engineering at Shandong University.
In recent 15 years, he leads a research group at
Shandong University to develop the dispatch theory
of modern power systems, and has published more than 200 academic papers.
His research interests include power system operation and control, power
system reliability assessment, and power economics.

Junjian Qi (S’12–M’13) received the B.E. degree

in electrical engineering from Shandong University,
Jinan, China, in 2008, and the Ph.D. degree in elec-
trical engineering from Tsinghua University, Beijing,
China, in 2013. He was a Visiting Scholar with
Iowa State University, Ames, IA, USA, in 2012, a
Research Associate with the Department of EECS,
University of Tennessee, Knoxville, TN, USA, from
2013 to 2015, and a Postdoctoral Appointee with
the Energy Systems Division, Argonne National
Laboratory, Argonne, IL, USA, from 2015 to 2017.
He is currently an Assistant Professor with the Department of Electrical and
Computer Engineering, University of Central Florida, Orlando, FL, USA.
He is the Secretary of the IEEE Task Force on Voltage Control for Smart
Grids. His research interests include cascading blackouts, power system
dynamics and stability, state estimation, synchrophasors, voltage control, and
cybersecurity.

Jianhui Wang (M’07–SM’12) received the Ph.D.

degree in electrical engineering from Illinois Insti-
tute of Technology, Chicago, IL, USA, in 2007.
Presently, he is the Section Lead for Advanced
Power Grid Modeling at the Energy Systems Divi-
sion at Argonne National Laboratory, Argonne, IL,
USA.
Dr. Wang is the secretary of the IEEE Power
& Energy Society (PES) Power System Operations
Committee. He is an associate editor of Journal of
Energy Engineering and an editorial board member
of Applied Energy. He is also an affiliate professor at Auburn University
and an adjunct professor at University of Notre Dame. He has held visiting
positions in Europe, Australia and Hong Kong including a VELUX Visiting
Professorship at the Technical University of Denmark (DTU). Dr. Wang is
the Editor-in-Chief of the IEEE Transactions on Smart Grid and an IEEE
PES Distinguished Lecturer. He is also the recipient of the IEEE PES Power
System Operation Committee Prize Paper Award in 2015.

Shiying Ma received the B.E. in electrical engineer-

ing from Xi’an Jiaotong University, Xi’an, China,
in 1990. He received the M.S. and Ph.D degrees
in Institute of Electric Power System from China
Electric Power Research Institute, Beijing, China, in
1996 and 2013, respectively. He is now the deputy
director of Institute of Electric Power System, China
Electric Power Research Institute. He is also a
professor of engineering, and has published more
than 60 academic papers.