Académique Documents
Professionnel Documents
Culture Documents
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2018.2884655, IEEE
Transactions on Power Systems
1
0885-8950 (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/TPWRS.2018.2884655, IEEE
Transactions on Power Systems
2
Numerical results are presented in Section IV. Final comments The previous approach is especially well suited for the
appear in Section V and Section VI shows the used references. analysis of large sparse power system models. In the following
subsection, this previously described approach will be applied
II. A FREQUENCY DOMAIN INTERPOLATION FRAMEWORK in detail within the Loewner matrix method framework
In this section, the generalized realization framework is following the works from [22] and [26].
introduced. Also, the solution to this problem towards the B. A standard Loewner matrix method
Loewner matrix method is proposed. An important feature of In this section, the construction of the models form (1) to (3)
this framework relays on its applicability to both the reduction is proposed using the Loewner matrix method. For the sake of
of linear models and the extraction of linear models from brevity, only a short overview of the standard Loewner matrix
measured frequency data. framework is presented [26]. Mainly, it is focused on the main
A. The generalized realization problem steps of the algorithm in [25].
If the input-output direct relation is neglected the The application of the Loewner matrix method requires
generalized realization problem consists in constructing the frequency data of the form:
state space representations Σ:=(E,A,B,C) from the array of
data giving by the set { ( sk,H(sk) ) : {sk}i∊ ; {H(sk)}i ∊ m,p; s , H s ,1 k N
k k (4)
H(x)=C(xE‒A)-1B , ∀x∊{sk} }, where m and p are the number
of outputs and inputs in the model respectively. Such where each element of the set {sk}i∊ is a single frequency at
realizations have the form: which some frequency response from an unknown model
HE(s)∊ m,p is measured. In case the model is already known it
Ex Ax Bu can be said that {sk} is a frequency at which the transfer
(1)
y Cx function of the model H(s)∊ m,p is sampled. N is the number
of elements of {sk}. The elements of the set (4) are appended
with the complex conjugates at negative frequencies
where x∊ n,u∊ p and y∊ m are the state, input and output
vectors respectively and {E,A}∊ n, n, B∊ n,p and C∊ m,n are
real matrices of proper dimensions. s , H s s , H s s , H s
k k k
H
k l l
(5)
There exists different software which can provide frequency
responses form a wide range of non-explicit models. Also, the where 1≤ l ≤ 2N, and the symbol (.)H denotes the Hermitian
frequency response of a particular transfer function can be transpose. Then, following the matrix format tangential
computed if an explicit linear model is known, or measured interpolation approach, the resulting data in (5) are partitioned
from application of low-level multi-sine probing inputs and into two sets, one containing the odd frequency samples along
frequency-domain processing of resulting synchrophasor with their complex conjugates, and the other containing the
recorded signals [28]- [29]. In either case, the first step for the even samples along with their complex conjugates:
extraction of a linear model in the generalized realization
framework will consist of obtaining the frequency response of
the desired part of the model to be captured. Then after the
s , H s , H , , H
l l k k k k (6)
construction of (1) a projection of the state vector into a
reduced basis of the form x =Txr will lead to a reduced model Considering 2N=N1+N2, with N1 and N2 the number of odd
of the form: and even frequency samples respectively. Then it is possible
to build the set of matrices
Exr A r xr Br u
(2) H j H i
y Cr x r L j ,i
j i
where xr∊ r, Ē=T-1ET, Ā=T-1AT, =T-1B, and T-1T=I∊ r,r, j H j i H i
with r < n. If the pseudo inverse of matrix Ē exists, (2) takes M j ,i (7)
the form: j i
xr A r xr B r u
F j H j , W i H i
(3)
y Cr x r These matrices are of dimensions {L,M}∊ N¹m,N²p , F∊ N¹m,p
and W∊ m, N²p . They are called the Loewner matrix (L), the
where Ar= Ē-1Ā and Br= Ē-1 . The quality of this model can shifted Loewner matrix (M), and the input (F) and output (W)
be measured in terms of its transfer function of the form matrices associated with μk, and υk, respectively. Of particular
Hr(s)= Cr(sI‒Ar)-1Br as an approximation of the transfer interest, the Loewner matrices form a linear subspace in the
function of the original model H(s)= C(sE‒A)-1B by the use space of all complex N1m x N2p matrices. Equivalent real
of the H-∞ norm of the difference between both of them given matrices can be computed regarding a similarity
by: ||H(s) ‒ Hr(s)||∞. transformation
0885-8950 (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/TPWRS.2018.2884655, IEEE
Transactions on Power Systems
3
Using this notion, assume now that a realization for the data In the following section, two different approaches to the
in (6) that solves the problem in IIA is given by: extraction of the frequency data given in (4) are proposed. The
first approach is suitable when a linearized model of a large
LRe x L M Re x L FReu sparse system is available. The second one is proposed when
(10)
y WRe x L such a model is not available, but the inputs of the underlying
black-box macro-model can be manipulated (empirically or by
with xL∊ N²p. Following strong theoretical concepts a reduced
simulation) with certain freedom using probing signals.
basis is obtained using a singular value decomposition (SVD)
of the regular part of the Loewner matrix pencil, (xLRe‒MRe),
x∊{μk}∪{υk}. The wrong choice of x could lead to ill- III. MODEL ORDER REDUCTION AND SYSTEM IDENTIFICATION
conditioned matrices; to avoid such thing the extraction of the In this section, a general framework for model order
energy of the Loewner matrix is proposed for the construction reduction and system identification of black-box macromodels
of a reduced and consistent model. The singular value based on the Loewner matrix technique is proposed. In both
decomposition LRe = ΦΣΨH, where Σ is a diagonal matrix cases, the resulting models are suitable for control design.
containing the singular values, σ, while Φ and Ψ are Note that while the passivity of Loewner matrix based
orthonormal matrices is computed. To select the order r of the macromodels has been studied in previous work such as in
reduced model an energy criterion is given by: [26], we do not address it in this paper. The main reason is that
passivity is important primarily in the context of transient
i network simulations which is not relevant to this work.
r
E i 1
(11)
A. Order reduction in large sparse linear models
q
j 1 j
One of the applications of the Loewner interpolation
method proposed in this work is the extraction of ROMs from
can be used, where q is the number of non-zero singular
values of LRe and ε is the percentage of the energy to be large sparse linear system models. For such purposes consider
retained (e.g., 99.9%). The model is then extracted by an a model with the form (1) is provided, then (4) is given by:
oblique projection using the two orthonormal matrices Φr and
Ψr, which are constructed by taking the first r columns of Φ, s , H s j , C jE A
k k k
1
B
(14)
and Ψ, respectively. First, the product of the state equation in
(10) and matrix ΦrH is obtained. Then the projection xL = Ψrxr
is used to finally get a reduced order model (ROM) with a where H(sk)∊ m,p, and jωk∊ . Now consider that the
form given by: frequency domain responses from linear transfer functions can
be given as:
xr Er 1A r x r Er 1B r u (12)
H jk C jE A B
1
(15)
y Cr x r
CX B jk
H
where Er=Φr LReΨr, Ar=ΦrHMReΨr, Br=ΦrHFRe, and finally
Cr=-WReΨr . Notice that the matrix -1
Er always exists because By adequately expressing the matrix product with respect to
the singular value decomposition of LRe matrix is used in the the term given by XB(jωk) =( jωk E-A)-1B, it can be seen that:
construction of Φr and Ψr.
The selection of r is key for the performance of the ROMs H jk CXB jk (16)
since a bad choice of this parameter tends to lead to unstable
0885-8950 (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/TPWRS.2018.2884655, IEEE
Transactions on Power Systems
4
m,p
with XB(jωk)∊ .Now notice that this implies a relation A key point from the extraction of the frequency response is
given by: the existence of a reliable method for phase shift
measurement. By measuring the phases in both signals (20)
jE A XB jk B (17) and (21) the relative phase shift between them (θq,i – ϕ0) can
be detected even if the method itself generates a phase error at
the signals proving signal (e.g., ϕ0 ≠ 0). Band-limited multi-
The previous form allows fast solutions for XB in terms of a sine signal probing associated with empirical transfer function
suitable sparse LU factorization as performed by references estimation is also an effective mean to improve phase and
[5], [8]. In the latter references, using sparse representations of magnitude accuracy [29].
models of the form (1) helps to improve the efficiency of the The measured magnitude and phase can be collected in the i-
extraction of the controllability and observability subspaces. th column of the empirical array HE(jωk)∊ m,p for each jωk as
In this case, it similarly helps to enhance the extraction of a complex number given by
linear model frequency responses. Finally, the set of data used
in the Loewner method is given as follows:
q ,i j q ,i 0
Re H E jk q ,i Re
e
s , H s j , CX j
k k k B k
(18) (22)
q ,i j q ,i 0
Once (18) is obtained, the proposed Loewner matrix method
Im H E jk q ,i Im
e
introduced in section IIB can be applied to extract ROMs.
Finally, the Loewner matrix technique can be applied at the set
B. System Identification using frequency response
of data given by:
When explicit models of power system electromechanical
oscillations are not available or large-scale models lead to
linearized versions with critical numerical errors the Loewner s , H s j , H j .
k k k E k (23)
matrix method can be used as an identification tool to obtain
small-signal models. Consider a transient stability model of Again the advantage of this classical approach to measure
the form frequency responses relies on the freedom for selecting any
x f x, u desirable frequency from any frequency range ωk ∊{χk}.
(19) This feature is key because with a larger input-output
y g x, u
behavior the dimensions of the set of Loewner matrices (7)
can still provide a feasible computation of the SVD process if
where again {y,g}∊ m,{x,f}∊ n, u∊ p are the output and the amount of interpolation data sk decreases.
input vectors, respectively. Furthermore, the duration of time-domain simulations for
To obtain a set of tangential data to be interpolated by the each signal might be significant because of the presence of ŷq,i.
Loewner matrix method, consider a large amount of probing However, decreasing the number of needed frequency
test signals of the form:
response components, results in a substantial reduction of the
computational effort required to build the data set.
u ei sin k t 0 (20) Other well-known approaches like the use of multi-
sinusoidal probing test signals [29]-[31], and the application of
such that ei∊ p is the i-th column of a proper identity matrix, numerical transforms, could be applied to enhance the
ωk corresponds to a frequency of a selected range ωk ∊{χk} , efficiency of the frequency response extraction but are not
the phase of reference is set to zero ϕ0 = 0, and α must be of a addressed on this document.
high enough level to provide signal responses of sufficient Drawing on the presented approaches a novel technique for
magnitude, e.g., above numerical or background noise floor both, model order reduction and model identification based on
for actual tests, but sufficiently weak to avoid any given the Loewner matrix method is introduced. The advantages of
nonlinear effect on (19). this method will be explored in the following section.
Assume now that the q-th output of the model by the action
of the i-th test input yq,i will be given by: IV. NUMERICAL RESULTS
0885-8950 (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/TPWRS.2018.2884655, IEEE
Transactions on Power Systems
5
0885-8950 (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/TPWRS.2018.2884655, IEEE
Transactions on Power Systems
6
0885-8950 (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/TPWRS.2018.2884655, IEEE
Transactions on Power Systems
7
The first parts of all the signals were cut off from the The Loewner matrices were built based on the approach in
records to capture only the steady-state behavior, considering (7). For reference, the reported Loewner matrix had
that all the transient dynamics with damping higher than dimensions LRe∊ 12480,3200.
1.75% had already died away. Giving these parameters and using the criterion in (13) set
The magnitude gain and the phase shift measured at the to a tolerance corresponding with the value ε =1 x 10-4, the
output signals associated with an input at a giving frequency order of the obtained adjusted model was found to be r =117
ωk correspond to a sample of the linear transfer function of the (for reference the non-linear system has 118 states).
model at that specific frequency. Figs. 5-6 depict frequency domain comparison and a map of
To compute (22) the fast Fourier transform was applied to the eigenvalues of the linearized version of the system versus
all input and output signals. It should be stressed that, even if a the identified Loewner model respectively. Notice that an
Hann window was used, the ratio (βq,i/α) and the phase attempt to capture the same number of inputs and outputs in a
difference (θq,i -ϕ0) are able to extract the required information time domain based realization technique will raise enormous
for the construction of (23). computational issues.
From the set (23), considering the Nyquist frequency, all The associated high dimensions of equivalent Hankel based
samples corresponding to frequencies above 20 Hz were models would generate out-of-memory errors (in Matlab),
ignored resulting in N=64 frequency points. solved by reducing the number of time domain samples, at the
Fig. 4 depicts the match between the measurements and the cost of losing frequency domain resolution by undersampling
frequency response of a linearized model obtained using the time-domain responses.
„svm_mgen.m‟ script for the same given input-output pair, and
at the same frequency points. 3) Test no.2: Comparison against other techniques
To compare against other identification methods described
in the power systems literature a new test was performed.
Using a similar procedure as before new frequency responses
were extracted using the reference voltage of the six modeled
exciters as inputs and the same outputs as the previous case.
Two main methods were taken as baselines; the first one is
the eigensystem realization algorithm (ERA) [17], [38]-[39],
which is a reliable classical technique that has been applied
successfully for the same problem presented in this work.
Two different versions of ERA were implemented, the
most basic version reported in reference [39] which uses
impulse responses (A), and a version that can be implemented
in practical cases [17],[36]. The later uses responses to pulse
Fig. 5. H-∞ norm of the linearized and identified model by Loewner matrix inputs with specified short duration instead of impulse
method.
responses for model extraction (B).
The second method is a modified version of the matrix
fitting technique [19], [38], this method was chosen as a
reference because the same data used by Loewner matrix
method can be used for the extraction of a model. It is also the
most popular method within the frequency-domain network
equivalent research group [40], whose main goal is developing
simulation tools for large-scale power system electromagnetic
transient studies.
For the first case of the ERA method, namely ERAA, the
procedure was implemented as it appears in reference [39].
The use of a unit impulse may introduce nonlinear behavior in
the system response; to extract linear responses, the magnitude
Fig. 6. Comparison between the linearized model eigenvalues and the
of the prove impulse (Kronecker delta) was then scaled to
identified model eigenvalues. a=0.01 pu according to the results of several superposition
2) Test no. 1: Extraction of MIMO linear models tests.
For this first test, the frequency response measurements As the linear model of the system is available, the general
derived in the previous section were used. The Loewner gain of the system was computed, and a correction factor for
matrix method was applied considering 50 inputs and 195 the range of interest was applied to the identified model. Since
outputs corresponding to all generator mechanical power this method is the only method using an excitation close to an
references as inputs, and all generator speed deviations and actual impulse response, it was used as ERA benchmark
voltage magnitude response signals at all 145 buses in the reference.
network as outputs.
0885-8950 (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/TPWRS.2018.2884655, IEEE
Transactions on Power Systems
8
For the second version of ERA, namely ERAB, time domain Also, notice a mismatch in the ERA cases for lower
solutions of the model were obtained for a probing signal frequencies. Fig. 8 and Table III provide an insight of the
corresponding to a 0.01 pu amplitude pulse of a duration of captured eigenvalues for the three techniques.
0.2 seconds. In both cases, these probing inputs were injected
at the six voltage reference of the modeled exciters. A linearity
test was also performed to asses that only linear behavior was
excited by the signals. A total of 14 seconds for each
simulation was retained using an integration step of 0.05
seconds. These settings were adopted for both cases, ERAA
and ERAB to simulate enough time span to reach the steady-
state conditions, while obtaining Hankel matrices whose
dimensions remain small enough to allow SVD to fit into
memory limits.
For the Loewner matrix method, all the proving signals and
parameters were set to be the same values as in the previous
tests no 1. The only differences were the selected inputs and
outputs, which in this case are the same as specified above in
ERA methods.
For model identification using matrix fitting method, the
code related to [40] was modified to work with non-
symmetrical non-square transfer functions H(s) instead of the
typical admittance matrix for which the original code was
developed. The same frequency points (N=64) used in the
Loewner case were chosen for the matrix fitting.
TABLE II
SOME CHARACTERISTICS OF THE DIFFERENT METHODS
matrix Loewner Fig. 7. a) H-∞ norm of the linearized and identified model by Loewner matrix
method ERAA ERAB method (teal) the ERA method (black & blue) and Matrix fitting method (red).
fitting matrix
b) Error in percentage.
type of time time frequency frequency
method domain domain domain domain
probing { { ( ) ( )
input** ( )
dimensions
of a basis Hkl∊ 27300,840
Hkl∊ 27300,840
X∊ 48010
LRe∊ 12484,384
matrix*
Fig. 8. Comparison of eigenvalues between the linearized model and the
identified models Loewner (teal), ERA (black & blue) and Matrix fitting
*Implies the SVD of the Hankel matrices (Hkl) and the Loewner matrix
(red).
(LRe); and the mean square solution of the vector X. ** α=0.001, a=0.01. ***
Given by the percentage of the energy of a basis matrix. +For the extraction of Observe that time-domain-solution based methods such as
the frequency response. ERAB, a practical version of ERA, requires fewer simulations.
(Table II). However, the match obtained in the linear modes
The selected order of approximation for matrix fitting was for the proposed version is higher. Notice also, for this case,
set to r =40 to be consistent with the results for the Loewner with the same information the Matrix fitting method obtained
case. The iterations were set to it =5. The same weight wht = 1 less accuracy in the identification of the corresponding linear
was applied to all frequencies. Table II depicts a comparison modes.
of the signals and other characteristics of the different 4) Test no.3: Extraction of equivalents for a given frequency
methods, such as the number of signals per input or band
dimensions of the bases for the extraction of linear models. In this final test, the frequency range of interest was set to
Fig. 7 shows for reference the comparison between the 0-0.1 Hz, in the aim of capturing all frequency modes in the
linearized models and the three identification H-∞ norms. As range of the common low-frequency oscillation mode (global
shown, more iteration is needed in the Matrix fitting identified system-wide mode) which is a challenging task for currently
system in order to obtain a better match between the norms. available identification techniques.
0885-8950 (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/TPWRS.2018.2884655, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 9
TABLE III
SLOWEST IDENTIFIED EIGENVALUES AT THE DIFFERENT MODELS
Loewner matrix
linearized model ERAA[39]* ERAB [38]
matrix fitting [40]
-0.4934± j0.3437 -0.4934± j0.3438 -0.4890 ± j0.3263 - -
-0.2971± j0.3763 -0.2971± j0.3763 -0.2968± j0.3769 - -0.3059± j0.3845
-0.3351± j0.4554 -0.3349± j0.4554 -0.3338± j0.4550 -0.3330± j0.4538 -0.3506± j0.4139
-0.6079 ± j0.5233 -0.6106± j0.5216 -0.6256± j0.5269 -0.6056± j0.4852 -
-0.4194 ± j0.5307 -0.4194 ± j0.5308 -0.4175± j0.5318 - -0.4104± j0. 5094
-0.6653 ± j0.7137 -0.6646 ± j0.7148 -0.6805± j0.6908 -0.6885± j0.7386 -0.6218± j0. 7314
*Benchmark for reference
methodology is that of the stability analysis of large power
Using the same procedure for the extraction of the frequency system models subject to severe topology changes.
response of the system and considering the same inputs and Hence, its application for control purposes, tuning, and
outputs again as in test no. 2, a new set {χk} was defined using validation, in a wide geographical area over a wide frequency
only 4 logarithmically spaced frequency points between the range is straightforward. Extracting frequency responses at
values of 10-1 and 1 rad/sec. A Loewner method based state- different frequency points enables a single frequency control
space model was constructed to capture 99.99% of the energy model covering inter-area, local modes and other frequency
in the singular values of the Loewner matrix. The order of the ranges of interest; with high accuracy in each band.
obtained adjusted model was found to be r =15. Fig. 9 shows Furthermore, the arbitrary selection of input injection points
the match between the modes of the actual linearized model makes the previous concept suitable for tuning all kinds of
and the obtained Loewner based model for this test. modulation based controls such as HVDC, renewable energy
or smart loads controls.
Finally, giving the ability of the method to extract models
with a considerable number of inputs and outputs, independent
system operators can better locate oscillations sources for
different topologies.
V. CONCLUSIONS
In this paper, we developed and tested a new model
reduction and linear black-box system identification scheme
designed for capturing electromechanical dynamics of large
power systems, based on Loewner matrix interpolation. The
new approach can handle large non-square multi-input multi-
output systems, using computed or measured frequency
Fig. 9. Comparison between the exact linearized model and Loewner based response data. Its application to electromechanical models was
model eigenvalues corresponding to four logarithmically spaced frequency evaluated extensively on large-scale multi-machine examples
points in the range 0.1-1 rad/sec.
TABLE IV and then compared with and shown to be superior to the
5 LARGEST RESIDUES FOR THE MODE @ 0 .52 RAD/SEC (WORST MATCH) current state of the art techniques, especially the popular
original captured
input output matrix fitting method when both are provided with the same
residue residue
frequency response data. In particular, the proposed method
1 0.2229∟30.79° 0.2200∟30.37° VREF5 ||V|| BUS #110
2 0.1975∟43.21° 0.1856∟41.59° VREF5 ||V|| BUS #93 was able to reduce or identify the large systems and extract the
3 0.1901∟38.11° 0.1794∟36.59° VREF5 ||V|| BUS #33 target natural modes with high fidelity using a relatively small
4 0.1901∟38.11° 0.1794∟36.59° VREF5 ||V|| BUS #35 number of frequency data samples.
5 0.1893∟38.11° 0.1787∟36.59° VREF5 ||V|| BUS #34
Even though presented results focused on the extraction of
Table IV shows the captured residues corresponding to the the modes of a particularly challenging frequency range
mode with the worst match for the frequency range of interest. related to system frequency-response control, the user can
Notice that the Loewner matrix associated to this case is of select any frequency band of interest in the proposed method
dimensions LRe∊ 780,24 proving that, provided that accurate quite arbitrarily to retain desirable information. The previous
frequency response data is available, the Loewner matrix makes the technique suitable for different intended purposes,
method potentially results in an adequate identification of the based on selective, logarithmically spaced, frequency response
dynamics of interest for control proposes with lower data.
computational effort than other comparable techniques. The drawback of the presented approach when dealing with
5) Discussion system identification is the need for effective strategies to
Loewner matrix method allows the extraction of a linear compute or measure frequency responses of the power system
model from a set of data from multiple frequency ranges at the under consideration. Another related issue is the lack of an a
same time. Linear models extracted by this method can priori strategy for selecting the number of needed frequency
provide insight into the system stability characteristics where domain samples achieved a given level of model accuracy.
conventional small signal analysis is not feasible. A particular These critical aspects of the method are left out as open
scenario of interest envisaged for the application of this research topics for now.
0885-8950 (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/TPWRS.2018.2884655, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 10
0885-8950 (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.