You are on page 1of 6

Abstract - The paper addresses practical aspects of integration

of classical and artificial neural network (ANN)-based dynamic


equivalents for transient analysis of large-scale interconnected
power systems. Following standard model reduction procedures
in power systems, we divide the system in two parts one re-
tained in detail, and the remainder that is to be replaced with a
simpler (equivalent) model. Our equivalent is based on two
ANNs, and is derived from measurements at boundary points. In
this paper we focus on a complete two-way interaction between
the retained portion of the system that is modeled with explicit
differential-algebraic equations (DAEs) and the ANN-based
equivalent. The key issue here is the consistency between dynamic
variables produced by the two different mechanisms at boundary
points, and a two-way communication between the corresponding
software modules. The overall model is denoted as hybrid, and the
paper proposes a fast two-way iterative interaction algorithm. We
demonstrate the feasibility of implementing our approach in stan-
dard industrial software by integrating our ANN-based equiva-
lent with the Matlab Power System Toolbox. We illustrate capa-
bilities of our approach to transient analysis on a benchmark
multi-machine example, derived from the New England/New
York power system.
Keywords - Power systems, Reduced-order systems, Neural net-
works, Nonlinear dynamic models.
I. INTRODUCTION
OWER systems have undergone profound changes in re-
cent years. This process is primarily driven by a desire to
increase efficiency, with concomitant reduction in energy
costs, energy losses and component size. The ongoing de-
regulation of the power utility industry adds a new dimension
to the main issues in modeling and control of power systems.
Power systems of the future are likely to be more stressed than
in the past, while the participating energy providers will be
less inclined to cooperate in overall power system control.
Future utilities are thus likely to increasingly rely on dynami-
cal equivalents in parts of interconnected power system and on
communication and computer technology to maintain a stable
system operation. Such models will have to be generated and

The authors are with the Department of Electrical and Computer Engi-
neering, Northeastern University, Boston, USA (e-mail: {astankov, asaric
@ece.neu.edu).
This work has been supported by the National Science Foundation under
grant ECS98-20977.
validated mostly from locally available prior information and
from real-time measurements.
Dynamic models of interconnected power systems are very
large, with sizes well into thousands of state variables. Existing
equivalencing methods tend to require a complete database
about the interconnected system [1]-[5] (with the exception of
[1]). This means that reduced models are sometimes based on
unreliable input data. The accuracy of the reduced model is
validated by comparing the waveforms at the point of coupling
between the retained and reduced portion of the system. When
there is a need to improve the equivalent, the tuning task may
prove difficult, as there is little indication as to which parame-
ters should be adjusted [6], [7]. The industrial experience is
also that equivalent derived for one class of events is typically
not very successful for other classes [3], [4].
An alternative (introduced in [1], and extended to large sig-
nal ANN-based models in [8], [9]), is to directly derive the
dynamic equivalent from measurements at points connecting
the retained subsystem with the remainder that will be reduced.
Our ANN-based approach is non-parametric in the sense that
we do not postulate any fixed dynamic model in advance.
Equivalencing procedure consists of two conceptual steps. At
the first stage, one ANN (of a special, bottleneck type) is
used to extract estimates about states of the reduced-order
equivalent; at the second stage, another ANN (recurrent
type) is embedded in an ordinary differential equation (ODE)
solver structure, and trained to approximate the right-hand
side of a continuous-time system, whose states were extracted
at the first step; for details, see [8]. ANN-based equivalent can
be used together with a classical equivalent in two ways: 1) it
is possible to hard-wire part of equivalent using any of the
classical model structures (gray box formulation); and 2) it
is possible to use both classes of equivalents simultaneously
with a retained system model [9].
In this paper we generalize the training procedure for the
two ANNs used in our equivalent. In the case of the bottleneck
network that extracts the state information, we use both meas-
ured quantities and variables calculated from them using the
retained system model. In the case of the recurrent network
that extracts state dynamics (active and reactive power in this
case), we add information about other quantities at interface
buses (voltage magnitude and angle). This redundancy will
prove helpful in deriving a two-way interaction procedure
An Integrative Approach to Transient Power
System Analysis with Standard and
ANN-Based Dynamic Models
Aleksandar M. Stankovi and Andrija T. Sari
P
0-7803-7967-5/03/$17.00 2003 IEEE
Paper accepted for presentation at 2003 IEEE Bologna Power Tech Conference, June 23th-26th, Bologna, Italy
within the hybrid model (cf. one-way interaction in [9]).
The rest of the paper is organized as follows the general-
ized problem formulation is given in Section II, followed in
Section III by the generalization of ANN training sets. The
two-way interaction between classical and ANN-based
equivalents in transient analysis is treated in Section IV. The
application of the proposed procedure on example hybrid
equivalenced 68-bus and 16-generator New England/New
York power system is outlined in Section V, followed by con-
clusions in Section VI.
II. GENERALIZED PROBLEM FORMULATION
Consider a power system divided to two different modeled
parts: 1) classical retained subsystem (including possibly areas
equivalenced in a conventional way), described by system of
DAEs, and 2) ANN-based reduced subsystem, modeled with a
set of transformation functions between vector of interface
variables and the state vector and an ODE integrator structure
that assures the continuous-time nature of the equivalent (we
select the trapezoidal recursive integrator here). Retained sub-
system is assumed to be exactly known, while the ANN-based
reduced subsystem is unknown, or only partially known. The
subsystems are connected via boundary nodes, equipped with
on-line measurements (say as part of the SCADA system). We
also assume that the measurements are accurate and that rec-
ords are of sufficient duration for ANNs training.
The standard DAEs formulation for problem of transient
analysis in classical modeled subsystem (subscript 1) is [10]:
( )
1 1 1 1
, , f =
&
G G
X X V I ; (1a)
( )
1 1 1
, g =
G G
I X V ; (1b)
1 11 1 1
1
i L
iL L LL L

=


V I Y Y
V I Y Y
, (1c)
where:
1
X is state vector;
1 G
V (
1 G
I ) is complex vector of
generator voltages (currents);
1
V (
L
V ) and
1 i
I (
iL
I ) are
vector of node voltages and bus injected currents in subsystem
(boundary) nodes, respectively (determined in part from avail-
able measurements and partially calculated from (1)); Y is
subsystem (index 1) and boundary nodes (index L) network
admittance matrix.
The minimal necessary number of measurements per
boundary node (electrical port) in formulation (1) is two. For
example, active and reactive power flows measurements lead
to a PQ-formulation (similar to standard load flow classifica-
tion of nodes) for the node, while active power flow and node
voltage measurements lead to a PV-formulation. Additional
(redundant) measurements can be used for model parameter
tuning [6], [7] and for detection of measurement errors. Simi-
larly, additional information about the reduced subsystem can
be used in structures that combine physical and signal-based
models, such as gray box models [9].
ANN-based reduced subsystem (subscript 2) is described
as:
2 2

( )
L L
= X j - mapping part; (2a)
( ) 2 2

L L
= X g - demapping part; (2b)
( ) 2 1 2 2 2

( ) ( ) ( ), ( )
2

n n n n
t
t t t t
+
D

= +

X X X y
( ) 2 2 1 1

( ), ( )

n n
t t
+ +

y X , (2c)
where:
2

X is reduced-order state vector;


L
is set of inter-
face variables in the boundary nodes;
2

L
j and
2

L
g are map-
ping and demapping transformations linking interface vari-
ables with states, respectively;
2
y is the integration function;
is parameter vector describing conditions for prediction of
state vector (Section III).
III. GENERALIZED ANN TRAINING SETS
The synthesis of the ANN-based equivalent depends on: 1)
the selection of interface variable set in the boundary nodes
(
L
), and 2) available knowledge about physical structure and
parameter database for the retained subsystem. The initial
training requires sufficiently long measurement records. When
the equivalent is later used on-line, we are interested in how
different the operating conditions are from those in the training
set. If the discrepancy is deemed substantial, then additional
training is needed. Criteria for similarity of training and op-
eration interface variables from database include: 1) Mean
square error between interface variables at the input and output
layer of trained bottleneck ANN; 2) Minimal mean square de-
viation between the operation interface variables and the
closest training pattern in the database.
Even when a minimal measurement set at boundary nodes is
available, the training of the bottleneck ANN can be per-
formed with a complete (redundant) interface variable set, as
we can use the model of subsystem 1 to reconstruct the non-
measured variables (we assume that measurement set is com-
posed with active and reactive power flow in interconnections,
as in Section II):
T T
L Lm Lc Lm Lm Lc Lc
= =

P Q V q , (3)
where indices m and c signify measured and calculated val-
ues, respectively. The symbol
Lm
P (
Lm
Q ) denotes vector of
measured active (reactive) power flows, while
Lc
V (
Lc
q ) de-
notes vector of calculated voltage magnitude (angle) in bound-
ary nodes. This interface variable set induces the following
structure of mapping transformation function (same holds for
demapping function
2

L
g ):
2 2 2 2 2 2 2

T T
L Lm Lc LP LQ LV Lq
j j j j
= =

j j j . (4)
The generalization of the training set also helps in improv-
ing the effectiveness of the bottleneck and recurrent ANN, and
quality of solution obtained by ANN-based equivalent. With
our particular selection of an implicit ODE integrator structure
(2c), we want to improve its numerical stability [9]. An im-
plicit integrator has advantages in systems with widely differ-
ent time scales (stiff DAEs), such as power system. Because
of that, we preferred implicit ODE integrator application in our
equivalencing procedure. From generalized equation (2c) we
see that the integration transformation function
2
y depends
both the state vector
2

X and conditions described by parame-


ter vector . Given our choice of the interface variable set
(3), a minimal parameter vector is determined by calcu-
lated interface variables, or:
T
Lc Lc Lc
= =

V q . (5)
Our knowledge about physical structure of reduced subsys-
tem allows us to distinguish two characteristic cases. If the
dynamic model (including parameters) for that subsystem is
very uncertain, ANN-based equivalent will be formed on the
basis of a black box model [8]. In this case, the state vector in
the reduced subsystem does not have physical nature. When
the dynamic model is partly known, we can apply a gray box
procedure for reduced-order state extraction [9]. In this way
the physical interpretability of the state variables is largely
retained (typically, in addition to improved accuracy).
Interaction between classical (retained) and ANN-based
(reduced) subsystems in off-line (training) phase is shown on
Fig. 1.
Fig. 1. One-way interaction between classical and ANN-based reduced sub-
systems in off-line (training) phase.
IV. INTERACTIONS BETWEEN RETAINED AND ANN-BASED
REDUCED SUBSYSTEMS IN OPERATION PHASE
Two main aspects determine the type of interaction: 1)
availability of initial (on-line) interface measurement set, and
2) phase of transient analysis (initial or prediction). The re-
sulting interaction can be one-way (when the initial input is
available and analysis in the initial phase), or two-way (un-
available initial input or prediction phase of the analysis).
A. One-way interaction.
This interaction is typically of short duration (see Figs.
4-6), and is similar to the training phase in Fig. 1. The main
difference is that the duration of calculation is limited only to
the initial input segment. Additionally, bottleneck ANN in-
putoutput transformation function (mapping+demapping
part)
( )
2 2
( )
L L L L
= g j serves as the criterion to evaluate
the quality of bottleneck ANN training; in this case, the recur-
rent ANN is not used.
B. Two-way interaction.
When the initial (minimal) measurement set of interface
variables is not available, the corresponding data segment can
be synthesized by extrapolation (for example, by using a third
ANN [8]). We will illustrate the subsequent two-way interac-
tion on a more complicated case of transient analysis in the
prediction phase (see Fig. 2).
Calculation for time instant
1 n
t
+
is initialized by the calcu-
lation for the ANN-based reduced subsystem, since it contains
information about the operating condition (encoded in ANNs
during training). For the classical subsystem, we have to spec-
ify in advance two interface variables per boundary node.
Following steps are then performed in the prediction phase (we
omit the time instant in the sequel to reduce the notational
clutter):
Step 1: Calculation of initial prediction of state vector
2

X
from recurrent ANN (2c), with constraints on demap-
ping transformation function for parameter vector
(5) on the output from bottleneck ANN (2b):
T
ANN ANN ANN
Lc Lc Lc

= =

V q
2 2
2 2
2 2

( )

( )

( )

LV
Lc
Lq
g
g

= =


g
X
X
X
, (6)
Step 2: Calculation of other initial interface variables (speci-
fied as measured in the classically modeled subsys-
tem, index m in eq. (3)) on the output from demap-
ping part of bottleneck ANN (2b):
T
ANN ANN ANN
Lm Lm Lm

=

P Q
2 2
2 2
2 2

( )

( )

( )

LP
Lm
LQ
g
g

= =


g
X
X
X
. (7)
Step 3: Calculation of prediction state vector in retained sub-
system (
1
X ) and subvector of calculated interface
variables
T
cl cl cl
Lc Lc Lc

=

V q for interface variable
subvector
T
ANN ANN ANN
Lm Lm Lm

= =

P Q from Step 2.
Step 4: Verification of the convergence criterion:
( )
2
cl ANN
Lc Lc
e - . (8)
If criterion (8) is satisfied, transient analysis is per-
forms for next time prediction (Step 1). Otherwise,
calculation continues with Step 5.
Minimal measurement
set (training duration)
Classical
(retained)
subsystem
ANN-based
(reduced)
subsystem
Other calcu-
lated inter-
face variable
Off-line (training) phase
Step 5: Calculation of corrected prediction of state vector
'
2

X
from recurrent ANN (2c), with constant parameter
vector
T
cl cl cl
Lc Lc Lc

= =

V q , calculated in Step 3.
Step 6: Calculation of corrected parameter vector
' ' ANN
Lc
= (5) on the output from demapping part
of bottleneck ANN (2b), for corrected prediction of
state vector
'
2

X from Step 5.
Step 7: Calculation of other corrected interface variables
(
' ANN
Lm
) - this step is same as Step 2.
Fig. 2. Two-way interaction between classical and ANN-based reduced sub-
systems in calculation of transient analysis (on-line (prediction) phase).
In the case of the interaction for initial data segment algo-
rithm is simplified, as only the bottleneck ANN is used. Note
that the algorithm can be divided into the initial prediction
phase (Step 1-Step 4) and correction phase (Steps 5, 6, 7, 3
and 4, or closed loop in Fig. 2). While it is possible to train the
recurrent ANN with the extended parameter vector
T
L Lm Lc
= =

, the resulting ANN would be sig-
nificantly larger and much harder to train. If such ANN were
used, Step 2 (Step 7) would be removed from the algorithm,
since a solution for all interface variables would be obtained
directly from the recurrent ANN (Step 1 and Step 6).
V. APPLICATION
The effectiveness of our two-way interactive transient
analysis procedure will be evaluated on the example of New
England/New York power system from Fig. 3 [11]. The system
comprises 68 buses and 19 generators. The full model is char-
acterized with a 144-dimensional state vector. The system can
be naturally divided in two parts connected with three lines.
Points for active and reactive power flow measurements in
interconnections on Fig. 3 is shown as . The left part of the
system, described by a 42-dimensional state vector, is retained,
while the right portion (shaded on Fig. 3), originally described
by a 102-dimensional state vector, is equivalenced. The
equivalent is sought for the case of a three-phase fault (started
at 0.5s) on generator G15 in the retained part. For ANN-
training purposes, measurements of active and reactive power
flows in interconnections (simulated by ETMSP [12]) are
used, corresponding to fault clearing times in the range
0.05-0.5s. ANNs were trained with the Levendberg-Marquardt
back-propagation training algorithm [13]. Bottleneck ANN
was trained using the gray box methodology [9]. For this pur-
pose, the right part of the system in Fig. 3 is equivalenced by
software package DYNRED [14]. After reduction, reduced-
order subsystem comprises only two equivalent generators
with generator speed and generator absolute angle as state
variables (thus assuming the classical model of equivalent syn-
chronous generator) and a 46-dimensional reduced overall
system state vector. Since the parameters and waveforms at
equivalent generators are very similar, for initialization (gray
box) of bottleneck ANN only two states (or two neurons in
middle, bottleneck layer) from one generator are used.
The size of the measurement training subvector
Lm
is 6
(active and reactive power flows in interconnections 1-2,
1-27 and 9-8, Fig. 3), while the dimension of the calculated
training subvector
Lc
is 4 (voltage magnitude and angle in
measurements nodes 1 and 9, Fig. 3).
Other basic data for ANNs training are:
- Number of training measurement responses: 5.
- Length of training waveforms for bottleneck ANN: 0.5-9s.
- Duration of training waveforms for recurrent ANN: 1-9s.
- Dimension of training pattern for fault (post fault) bottle-
neck ANN: 129 (1375).
- Dimension of training pattern for recurrent ANN: 190.
- Number of neurons in fault (post fault) bottleneck ANN:
10; 10; 2; 10; 10 (10; 10; 2; 25; 10).
- Number of neurons in recurrent ANN: 14; 20; 20; 2.
Our integrated dynamic ANN-based model reduction and
transient analysis software package is realized in Matlab envi-
ronment, with the help of the Power System Toolbox [11].
In Figs. 4-7 we show responses obtained by applying the
generalized training set to the ANN equivalent. Figure 4 dis-
plays three traces of generator speed the first obtained solely
from the bottleneck ANN, the second by combining the bottle-
neck and recurrent ANN, and the third from a classical
equivalencing procedure (DYNRED, also used for the initiali-
Initial prediction
2

X , with
constraint on
ANN
Lc
=
(recurrent and demapping
part of bottleneck ANN)
Initial (corrected) predic-
tion
ANN
Lm
(demapping part
of bottleneck ANN)
Convergence
criterion
satisfied?
Next time
prediction
Yes
No
Corrected prediction
' ' ANN
Lc
= (demapping
part of bottleneck ANN)
Corrected prediction
'
2

X ,
with
cl
Lc
=
(recurrent ANN)
Predictions
1
X and
cl
Lc

in the classical (retained)


subsystem
On-line (prediction) phase
zation of the ANN gray box). In Figs. 5 and 6 we show active
and reactive power flows in interconnection line 9-8, together
with waveforms obtained in the original system. Recalling that
the dimension of the state vector of the ANN equivalent is just
2 (as contrasted with the 102-dimensional original state), we
conclude that the overall response quality is satisfactory.
Fig. 3. New England/New York test power system [11].
Fig. 4. Equivalent generator speed in reduced subsystem.
Fig. 5. Active power flow in interconnection 9-8.
1
G1
53
G8
60
26 25
28 29
61
G9
24
27
30
9
8
3
18
17
2
4 14
15
21
16
57
19
20
G5
56
G4
58
G6
22
59
G7
23
13
11
12
5
6
7
54
G2
55
G3
10
65
G13
37
G14
66
40 48
47
41
G10
62
31
G11
63
32
33
38 46
G15
67
42
49
36
64
G12
34
35
43
68
G16
52
50
51 45
44
39
ANN-based reduced-order
subsystem
Retained subsystem
Interconnection and
measurement points
0 1 2 3 4 5 6 7 8 9
0.991
0.992
0.993
0.994
0.995
0.996
0.997
0.998
0.999
1
1.001
G
e
n
e
r
a
t
o
r

s
p
e
e
d

(
p
u
)
Time (sec)
Initial (DYNRED)
Gray box (bottleneck ANN)
Gray box (recurrent ANN)
0 1 2 3 4 5 6 7 8 9
-2
-1.5
-1
-0.5
0
0.5
A
c
t
i
v
e

p
o
w
e
r

(
p
u
)
Time (sec)
Original system
Gray box (bottleneck ANN)
Gray box (recurrent ANN)
The proposed interactive two-way transient analysis proce-
dure will be illustrated on the example of the active power
flow in interconnection 9-8 (Fig. 5), and the results are shown
in Fig. 7 (including the original and a classically reduced sys-
tem). They suggest that both the classical and the ANN-based
equivalent successfully capture the system dynamics, espe-
cially right after the fault. Note, however, that the input data
requirements are quite different for the two methods.
Fig. 6. Reactive power flow in interconnection 9-8.
Fig. 7. Active power flow in interconnection 9-8.
VI. CONCLUSIONS
The paper addresses practical aspects of integration of clas-
sical and artificial neural network (ANN)-based dynamic
equivalents for transient analysis of large-scale power systems.
The results shown in the paper demonstrate that a complete
two-way on-line communication between the two models leads
to encouraging results for a challenging multi-machine bench-
mark power system.
VII. REFERENCES
[1] W. W. Price, D. N. Ewart, E. M. Gulachenski and R. F. Silva, Dynamic
equivalents from on-line measurements, IEEE Trans. on Power Appa-
ratus and Systems, PAS-94, pp. 1349-1357, Jul./Aug. 1975.
[2] R. Podmore, Identification of coherent generator dynamic equivalents,
IEEE Trans. on Power Appar. and Systems, vol. PAS-97, pp. 1344-
1354, Jul./Aug. 1978.
[3] L. Wang, M. Klein, S. Yirga and P. Kundur, Dynamic reduction of
large power systems for stability studies, IEEE Trans. on Power Sys-
tems, vol. 12, pp. 889-895, May 1997.
[4] W. W. Price, A. W. Hargave, B. J. Hurysz, J. H. Chow and P. M.
Hirsch, Large-scale system testing of a power system dynamic equiva-
lencing program, IEEE Trans. on Power Systems, vol. 13, pp. 768-774,
Aug. 1998.
[5] X. Lei, D. Povh and O. Ruhle, Industrial approaches for dynamic
equivalents of large power systems, in Proceedings of the 2002 IEEE
Power Engineering Society Winter Meeting, New York, USA.
[6] I. A. Hiskens, Nonlinear dynamic model evaluation from disturbance
measurements, IEEE Trans. on Power Systems, vol. 16, pp. 702-710,
Nov. 2001.
[7] M. Burth, G. C. Verghese and M. Vlez-Reyes, Subset selection for
improved parameter estimation in on-line identification of a synchro-
nous generator, IEEE Trans. on Power Systems, vol. 14, pp. 218-225,
Feb. 1999.
[8] A. M. Stankovi, A. T. Sari and M. Miloevi, Identification of non-
parametric dynamic power system equivalents with artificial neural
networks, Submitted for publication.
[9] A. M. Stankovi and A. T. Sari, Transient power system analysis with
measurement-based gray box and hybrid dynamic equivalents, Sub-
mitted for publication.
[10] P. W. Sauer and M. A. Pai, Power system dynamics and stability, Pren-
tice Hall, 1998.
[11] G. Rogers, Power system oscillations, Kluwer Academic Publishers,
2000.
[12] * * * Extended transient-midterm stability program (ETMSP), Users
Manual, Vol. 1-6, Version 3.1, EPRI TR-102004-V2R1, Final Report,
May 1994.
[13] * * * Neural network toolbox, for use with MATLAB

, Users Guide,
Version 4, The Math Works Inc., 2001.
[14] * * * Dynamic reduction program (DYNRED), Users Manual, Version
1.1, EPRI TR-102234 Project 2447-1, Final Report, Oct. 1993.
VIII. BIOGRAPHIES
Aleksandar M. Stankovi (1960) obtained the Dipl. Ing. degree from the
University of Belgrade, Yugoslavia in 1982, the M.S. degree from the same
institution in 1986, and the Ph.D. degree from Massachusetts Institute of
Technology in 1993, all in electrical engineering. He has been with the De-
partment of Electrical and Computer Engineering at Northeastern University,
Boston since 1993, presently as a Professor. He serves as an Associate Editor
for IEEE Transactions on Power Systems and he served IEEE Transactions on
Control System Technology in the same capacity from 1997 to 2001.
Andrija T. Sari (1962) received the B.Sc. degree from the University of
Belgrade, Yugoslavia, in 1988, and the M.Sc. and Ph.D. degrees in Electrical
Engineering in 1992 and 1997, respectively, from the same university. He is
an Associate Professor of Electrical Engineering at aak College of Engi-
neering, University of Kragujevac, Yugoslavia. Presently, he is a Visiting
Professor with the Department of Electrical and Computer Engineering,
Northeastern University, Boston, USA. His main areas of interest are power
system control, analysis, planning and optimization, as well as application of
artificial intelligence methods in these areas.
0 1 2 3 4 5 6 7 8 9
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
R
e
a
c
t
i
v
e

p
o
w
e
r

(
p
u
)
Time (sec)
Original system
Gray box (bottleneck ANN)
Gray box (recurrent ANN)
0 1 2 3 4 5 6 7 8 9
-2
-1.5
-1
-0.5
0
0.5
A
c
t
i
v
e

p
o
w
e
r

f
l
o
w

(
p
u
)
Time (sec)
Original system
DYNRED equivalent
ANN equivalent

You might also like