Electrical Power and Energy Systems: Hamed Shakouri G., Hamid Reza Radmanesh
Electrical Power and Energy Systems: Hamed Shakouri G., Hamid Reza Radmanesh
Identification of a continuous time nonlinear state space model for the external
power system dynamic equivalent by neural networks
Hamed Shakouri G. a, Hamid Reza Radmanesh b,*
a
Industrial Engineering Faculty, University of Tehran, Tehran, Iran
b
Engineering Faculty, Shahed University, Tehran, Iran
a r t i c l e i n f o a b s t r a c t
Article history: Based on the concept of the external power system dynamic equivalent obtained for the study system, in
Received 14 June 2008 this paper a reduced-order artificial neural network is proposed, to construct a model for the external
Received in revised form 5 March 2009 part. The mastermind behind the proposed method is to identify the external part as a dynamic–algebraic
Accepted 16 March 2009
ANN, and this separation between dynamic equations in the state space and algebraic equations is useful
to solve the prediction problem. Moreover, using similarity transformations, the state space model can be
simplified, such that all the nonlinearities are embedded in the algebraic part. Since usually the study sys-
Keywords:
tem equations are available in the continuous time domain, the external part is converted to the contin-
Dynamic equivalent
Artificial neural networks
uous time domain by a novel method. To obtain this model, the system should be excited first by a sort of
Nonlinear identification random disturbances, and then data measured on the boundary nodes is used to identify the model. Iden-
Numerical differentiation tification process is accomplished by training the proposed network which can be used to predict behav-
ior of the external system with a high degree of accuracy. Such an equivalent has wide applications for
dynamic stability studies.
Ó 2009 Elsevier Ltd. All rights reserved.
1. Introduction become serious rival of the old conventional methods (e.g. the
coherency [15,16] and the modal [17] approaches). The equivalent
Modeling, as the first step of any system analysis, is an impor- estimation methods have spread, because it can be estimated
tant task in scientific studies. Usually, the model is obtained either based on data measured only on the boundary nodes between
using physical laws or through an identification process [1,2]. the study system and the external system [18–20]. This way, with-
Therefore, a variety of linear and nonlinear models are developed out any need to information from the external system, estimation
to get in use for power system studies, and many computer pack- process tries to estimate a reduced order linear model which is re-
ages are implemented as well [3–5]. placed for the external part. Evidently, estimation methods can be
A set of low order linear models is still used in some dynamic used, in presence of perfect data of the network as well to compute
studies [6–8] as well as many nonlinear models [9]. However, for the equivalent by simulation and/or model order reduction.
vastly interconnected networks, such as the European and the Mid- Sophisticated techniques have become interesting subject to
dle East networks, even low order models lead to large system researchers for solving identification problems since 90s. For
matrices causing memory managing and/or system analysis diffi- example, to obtain dynamic equivalent of an external subsystem,
culties. Therefore, in company with system expansion, utilizing dy- an optimization problem has been solved by Genetic Algorithm
namic equivalents has been spread manifestly. A review through (GA) likewise Levenberg–Marguardt algorithm [21]. Neural net-
different former approaches on dynamic equivalents can be found works is the most prevalent method between these techniques be-
in [10]. cause of its high inherent ability for modeling nonlinear systems,
On the other hand, due to the lack of complete system data, including power system dynamic equivalents [22–27].
and/or frequently variations of the parameters while system oper- Some researchers have modeled the external part as a nonlinear
ations, the importance of estimation methods is revealed notice- ARX (NARX) [22,23]. Since simultaneous solution of the algebraic
ably. Especially, on-line model correction aids for employing and dynamic equations in the state space model is more conve-
adaptive controllers, power system stabilizers (PSS) or transient nient, it would be better rearranged the NARX model in a state
stability assessment [11–14]. The capability of such methods have space form to avoid complexity of solution caused by nonlinear
terms and transformation. To model external part as a state space
* Corresponding author.
model, [26,27] introduce two ANNs. A ‘‘bottleneck” ANN, is used by
E-mail addresses: [email protected] (H. Shakouri G.), [email protected] the former to extract ‘‘states” of the reduced-order equivalent; and
(H.R. Radmanesh). a ‘‘recurrent” ANN is embedded in an ordinary differential
0142-0615/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijepes.2009.03.016
H. Shakouri G., H.R. Radmanesh / Electrical Power and Energy Systems 31 (2009) 334–344 335
As an introductory tool, we should first mention the numerical ci;i ¼ ejðdi p=2Þ ; i ¼ 1; 2; . . . ; m ð12Þ
differentiation method used in this paper introduced by Li [28].
where m is number of the machines. Hence, the direct and the
A (2n + 1)-point centered difference formula of order 2n to approx-
quadrature axis voltage and current relationships are defined by:
imate the first derivative of a function, f(x), at the middle point x0 is
determined by:
V ¼ CðV d þ jV q Þ ð13Þ
0 1X n
2n I ¼ CðId þ jIq Þ ð14Þ
f ðx0 Þ ¼ d2nþ1;0;j f ðxj Þ þ O2n;0 ðh Þ ð1Þ
h j¼n
Now, the matrices A, B, C, and D are defined as diagonal matrices
where
with the diagonal elements defined as follows:
ð1Þjþ1 ðn!Þ2
d2nþ1;0;j ¼ ; j ¼ 1; 2; . . . ; n ð2Þ Rs
jðn jÞ!ðn þ jÞ! Ai;i ¼ ð15Þ
R2s þ X 0d X 0q
and
X 0q
d2nþ1;0;0 ¼ 0: Bi;i ¼ ð16Þ
R2s þ X 0d X 0q
Thus, the symmetric 7-point approximation for the first derivative X 0d
can be written as C i;i ¼ ð17Þ
R2s þ X 0d X 0q
1 6 Rs
f 0 ðx0 Þ ¼ ðf3 þ 9f 2 45f 1 þ 45f 1 9f 2 þ f3 Þ þ Oðh Þ ð3Þ Di;i ¼ ð18Þ
60
R2s þ X 0d X 0q
Both the backward and the forward difference formulas are avail-
able in [28]. However, it should be mentioned that usually centered In addition, two matrices namely G0 and B0 are defined by:
difference formula has a better performance than the other
formulas. b red C
G0 ¼ Re½C1 Y ð19Þ
0 1 b
B ¼ Im½C Y red C ð20Þ
3. The power system model
This way, Vd and Vq can be obtained using the regular matrix meth-
3.1. Dynamic equations ods given below.
" #1 " #
To simulate a multi-machine power system, the state equations Vd G0 þ A B0 þ B AE0d þ BE0q
¼ ð21Þ
are derived from [29]. Each generator can be represented by the Vq B0 þ C G0 þ D CE0d þ DE0q
following two-axis model:
0 Once Vd and Vq are known, Id and Iq can be calculated using (22) and
dEqi (23):
T 0doi ¼ E0qi ðX di X 0di ÞIdi þ Efdi ð4Þ
dt
dE
0 Id ¼ G0 V d B0 V q ð22Þ
T 0qoi di ¼ E0di þ ðX qi X 0qi ÞIdi ð5Þ Iq ¼ B0 V d þ G0 V q ð23Þ
dt
336 H. Shakouri G., H.R. Radmanesh / Electrical Power and Energy Systems 31 (2009) 334–344
4. The concept of dynamic equivalent Conversely, when the frontier nodes cannot be assumed as infi-
nite buses, both the magnitude and phase of voltage on the frontier
To understand the concept of dynamic equivalent, it is relevant node involve dynamical terms. The voltage at the frontier nodes
to start our study with the simplest case: the frontier (boundary) can be considered as a function of injection currents and the exter-
nodes are infinite buses (Fig. 1). Magnitude and phase of voltages nal part state variables (See Fig. 2).
on the frontier nodes ðV fn Þ are constant without dynamics. The notion explained above can be achieved analytically as well.
Differential-algebraic equations (DAEs) in part III can be repre-
sented symbolically as follows [29]:
Fig. 3. The whole system equations (green box in the left hand: dynamic equations, blue box in the right hand: algebraic equations).
H. Shakouri G., H.R. Radmanesh / Electrical Power and Energy Systems 31 (2009) 334–344 337
to the model, and both magnitude and phase of the frontier nodes nodes. For each set we can choose two of four waveforms as the in-
voltages ðV fn Þ are outputs for the algebraic equations of the exter- put set and two remaining waveforms as the output set. Eqs. (31)–
nal part. There are other options for choosing input–output data (33) are in polar form and it is possible to be rewritten in (D, Q)
set. Let us consider the following sets of variables: components form as well.
Since integration of the frequencies to obtain the angles hides
V fn \V fn Ifn \Ifn ð31Þ the fluctuations caused by disturbances, there should be a refer-
ence to compute differences between the angles. To do so, we need
V fn \V fn Sfn \Sfn ð32Þ
to choose one of the generator power angle deviations as a refer-
Ifn \Ifn Sfn \Sfn ð33Þ ence for angle deviations. It is apparent in Fig. 4 that the power an-
gles (simulation results for WSCC test system) seem float and these
where j.j means the magnitude, \(.) denotes the phase angle, and Sfn float waveforms do not provide appropriate data for identification
is the complex power measurements in the interconnecting frontier process. Instead, the waveforms in Fig. 5 are obtained by choosing
power angle deviation 3 as a reference for the angle deviations.
Now these waveforms appear as persistently exciting [1] and
appropriate data for identification process. To carry out this trans-
formation, we should convert xS to xn in (6) where n is the num-
ber of chosen generators as the reference for power angles. Hence,
for generator n, (6) is eliminated and the degree of the differential
equations reduces.
Fig. 6. The initial structure of the reduced-order dynamic–algebraic equivalent (DEA) The number of the neurons of each layer is given under the number of each layer of the
NN blocks. These are used for the case study defined in Section 7.
338 H. Shakouri G., H.R. Radmanesh / Electrical Power and Energy Systems 31 (2009) 334–344
Fig. 7. Nonlinear layers in the paths from the inputs to the outputs of three realizations of DAEs.
Fig. 8. The simplified structure of the reduced-order dynamic–algebraic equivalent (DEA) The number of the neurons of each layer is given under the number of each layer of
the NN blocks. These are used for the case study defined in Section 7.
algebraic equations group ‘‘G2”. In Fig. 7a, G1 paths through two layer: the algebraic equations (thin blue line). In Fig. 7b nonlinear-
nonlinear layers: the dynamic equations and the algebraic equa- ity is concentrated only in the algebraic equations. Hence, both G1
tions (bold green line); but G2 paths through only one nonlinear and G2 face to one nonlinear layer (thin blue line). Similarly, in
H. Shakouri G., H.R. Radmanesh / Electrical Power and Energy Systems 31 (2009) 334–344 339
Fig. 7c, G1 faces to one nonlinear layer: the dynamic equations ber of nonlinear layers. The simplified version of the structure
(thin blue line), but G2 does not faces to any nonlinear layer shown in Fig. 6 is illustrated in Fig. 8. Since the bias connection
(dashed red line). From Fig. 7, it is apparent that the structure (c) for layer 1 is useless, it is pruned from the ANN.
cannot be a perfect structure, because there is no nonlinear term
affecting the external inputs to the algebraic part. It is worthy to 6. The continuous time state space model
mention that the structure (b) is more efficient than (a), because
each nonlinear unit in the algebraic part is shared between the The state space model, illustrated in Fig. 8, is a discontinuous
two groups of the inputs (inputs to the dynamic part and inputs (discrete) time model, but the study system equations are usually
to the algebraic part) in their paths to the outputs. Furthermore, given in continuous time domain. A discontinuous time state space
complexity of the structure (b) is less than (a), due to the less num- model can be represented as in Fig. 9. Hence, for the ith state var-
iable we can write:
xi ðk þ 1Þ ¼ fi ðxi ðkÞ; u
i ðkÞÞ ð34Þ
where xi ðkÞ and u
i ðkÞ are called effective state variables and effec-
tive inputs at t = k to construct the ith state variable at t = k + 1.
Eq. (34) can be written in continuous time domain as
x_ i ðtÞ ¼ F i ðxi ðtÞ; u
i ðtÞÞ ð35Þ
A relatively accurate approximate of the first derivative of xi(t),
i.e. ^
x_ i ðtÞ, can be determined by (3). Therefore, we can train the neu-
ral network in Fig. 10, the target of which is ^ x_ i ðtÞ; i = 1, 2, . . . , ns
where ns is the number of state variables considered for the equiv-
alent system. The effective equivalent states, xi ðtÞ, which indeed
Fig. 9. The discontinuous (discrete) time state space model. are artificial states and so dimensionless are:
xi ðtÞ ¼ X ext ðtÞ ð36Þ
and the inputs are:
i ðtÞ ¼ Vt 3
u It3 \Vt3 \It 3 ð37Þ
The disturbances take place on mechanical torque and reference signals, which are illustrated in Fig. 12, and the corresponding re-
voltage of the study system generator. sponse, i.e. the magnitude and phase of the voltage and current in
It should be mentioned that the method can be applied on any node 3 are plotted in Fig. 13. The sampling period is chosen to be
larger system with more than one generator in the study system 25 ms.
and/or more than one frontier nodes. However, the vast number The measurements are used to train the equivalent ANN.
of calculations needed for training the ANN causes limits to the This should be done according to Figs. 6 and 8, for both the initial
on-line applications of the method. and simplified structures, respectively. The number of the neu-
rons of each layer is given under the number of each layer of
8. Application the NN blocks. To perform order reduction, number of the state
variables for the external part is assumed to be 5; this will be
The test system is simulated by simultaneously exciting both equal to the number of the neurons in layer 2. The size of each
the mechanical torque and the reference voltage of generator 3. layer is indicated in Figs. 6 and 8 for the initial and simplified
These disturbances are designed as BPR (binary pseudo random) structures.
Fig. 12. The mechanical torque and reference voltage waveforms (pu: per-units).
Fig. 13. Magnitude (pu: per-units) and phase (radians) of the voltage and the current in node 3.
H. Shakouri G., H.R. Radmanesh / Electrical Power and Energy Systems 31 (2009) 334–344 341
Before that the measurements can be used for identification, and the simplified structures. It is obvious that although the sim-
since the scales of deviations in the signals are different, data plified structure involves less complexity, it is more accurate, be-
should be normalized. This is done by mapping the samples of each cause in some cases nonlinear layer(s) between delay(s) can have
waveform (WK) according to the following rule: a preventive effect for the performance of training of the ANNs.
The external part state space variables for the simplified structure
W K W OP related to the validation data are plotted in Fig. 16. Fig. 17 com-
W KMapped ¼ ð38Þ
MaxðjW W OP jÞ pares the measured magnitude of voltage for node 3 with the cor-
responding output of the ANN for validation data. The figure is
where WOP is the operating point for each waveform, and is the vec- zoomed in for the time interval [54, 56] to show accuracy of the
tor of data waveform. This normalization makes the waveform sig- equivalent model output. This is illustrated in Fig. 18. Similar plots
nals dimensionless. for the magnitude of the current waveform are illustrated in Figs.
Training and validation results illustrated in Figs. 14 and 15, 19 and 20.
which are the mean square errors (MSE) of the model outputs, im- Now, the discontinuous time model should be transformed to the
ply accurate prediction of the output waveforms by both the initial continuous time domain according to Fig. 10. The transformation
Fig. 14. Training and validation performance (MSE) for the initial structure. Fig. 15. Training and validation performance (MSE) for the simplified structure.
Fig. 16. The external part state space variables for the simplified structure.
342 H. Shakouri G., H.R. Radmanesh / Electrical Power and Energy Systems 31 (2009) 334–344
Fig. 17. Comparison between the measured magnitude of the normalized voltage Fig. 20. Comparison between the original magnitude of mapped-current for node 3
for node 3 and the corresponding output of the ANN for the validation data. and related output of the ANN for validation data in the time interval [54, 56].
Fig. 18. Comparison between the original magnitude of mapped-voltage for node 3 Fig. 21. Comparison between 7-point approximation and ANN estimated of the first
and related output of the ANN for validation data in the time interval [54, 56]. state space variable derivative for the validation data.
Fig. 22. Comparison between 7-point approximation and ANN estimated of the first
Fig. 19. Comparison between the original magnitude of mapped-current for node 3
state space variable derivative for the validation data in the time interval [54, 56].
and related output of the ANN for the validation data.
9. Comparative analysis
should be done for all state variables. Figs. 21 and 22 show the
results for the top state variable in Fig. 16. The ANN has two hidden Now that the proposed method is fully described, it would be
layers with 20 neurons in each layer. helpful if the place of the method among all methods introduced
H. Shakouri G., H.R. Radmanesh / Electrical Power and Energy Systems 31 (2009) 334–344 343
in the literature of power system dynamic equivalents is clarified structure. This method is indeed based on nonparametric system
[6]. There are various methods to obtain a proper dynamic equiv- identification. Therefore, obtained equivalent is not limited by
alent for a large power system. To have a comparative survey, parameters; this is considered as an advantage. However, identi-
these methods are first categorized in this section. Power systems fied system is dependent on the study system, i.e. any change in
dynamic equivalents also may be classified due to different views. parameters of the study system, causes some changes in the equiv-
One may classify into: Empirical simplification; Eigenvalue analy- alent identified for the external part.
sis; Frequency domain analysis; Optimization methods and Ran- The method proposed in this paper is indeed a composition of
dom process analysis. However, from the view point of this the two recent approaches. Although it is a black box model and
paper, principally there are two kinds of equivalent making meth- identification process is nonparametric, but its nonparametric
ods: Direct Methods and Estimation Methods. structure is derived from the complete system structure either
for the dynamic part or for the algebraic part.
9.1. Direct methods
[16] Eduardo J.S. Pires de Souza, Identification of coherent generators considering [24] Sowa Paweł, Azmy Ahmed M, Erlich István. Dynamic equivalents for
the electrical proximity for drastic dynamic equivalents. Int J Electr Power calculation of power system restoration. Wisla, 7–9 czerwca 2004.
Energ Syst 2007. [25] Rahim AHMA, Al-Ramadham AJ. Dynamic equivalent of external power system
[17] Erling, I.; Kasztel, Z.; Schegner, P., Enhanced modal based technique for and its parameter estimation through artificial neural networks. Elsevier
construction of power system dynamic equivalents. In: 14th PSCC. Sevilla; Sience Ltd.; 2002.
June 2002. [26] Stankovic Aleksandar M, Saric Andrija T, Milosevic Mirjana. Identification of
[18] Ibrahim MAH, Mostafa OM, El-Abiad AH. Dynamic equivalent using operating nonparametric dynamic power system equivalents with artificial neural
data and stochastic modeling. IEEE PAS 1976;95(5):1713–22. networks. IEEE Trans Power Syst 2003;18(4).
[19] Alvarado FL. Real time external equivalents for static security analysis. IEEE [27] Stankovic Aleksandar M, Saric Andrija T. Transient power system analysis with
TPAS 1979;98:505–7. measurement-based gray box and hybrid dynamic equivalents. IEEE Trans
[20] Yu Yn. Power system dynamics. Academic Press; 1983. Power Syst 2004;19(1).
[21] Ramirez Arredondo Juan M, Garcia Valle Rodrigo. An optimal power system [28] Li Jianping. General explicit difference formulas for numerical differentiation. J
model order reduction technique. Electr Power Energ Syst 2004;26:493–500. Comput Appl Math 2005;183:29–52.
[22] Enrico De Tuglie, Lorenzo Guida, Francesco Torelli, Dario Lucarella, Massimo [29] Sauer Peter W, Pai MA. Power system dynamics and stability. Upper Saddle
Pozzi, Giuliano Vimercati. Identification of dynamic voltage–current power River: Prentice Hall; 1998.
system equivalents through artificial neural networks. Bulk power system [30] Douglas Hines Paul. A power systems capstone design project and associated
dynamics and control – VI. Italy: Cortina d’Ampezzo; 22–27 August 2004. simulation software designed to meet the changing needs of the electrical
[23] Azmy Ahmed M, Erlich István. Identification of dynamic equivalents for power industry and engineering accreditation requirements. Master of Science
distribution power networks using recurrent anns. 0-7803-8718-X/04/ in Electrical Engineering, University of Washington; 2001.
$20.00Ó2004 IEEE.