Load modeling for wide area power system

Ping Ju

, Chuan Qin, Feng Wu, Huiling Xie, Yan Ning

College of Energy and Electrical Engineering, Hohai University, Nanjing 210098, China
a r t i c l e i n f o
Article history:
Received 3 March 2010
Received in revised form 12 October 2010
Accepted 9 December 2010
Available online 24 February 2011
Keywords:
Power systems
Load modeling
Power system identication
a b s t r a c t
The electrical loads have signicant impacts on the dynamic performances of the wide area power sys-
tem, and the dynamic behaviors of the power system can not be reproduced by the simulation system,
unless the electrical load was modeled accurately. In the previous load modeling, the dynamics of the
electrical load itself is of concern. Using such load models, the satised dynamic behaviors of the wide
area power system can not be obtained by simulation. In this paper, a system-wide load modeling strat-
egy is proposed. The electrical loads are rstly classied into a few categories using the component based
load modeling method. According to the problems studied, the output variables of the power system are
selected, and the objective function is constructed using the data from Wide Area Measurement System
(WAMS). The trajectory sensitivities of the load model parameters with respect to the output variables
are analyzed. Based on which, the key parameters playing important roles on dynamics of the power sys-
tem are identied and included in further parameter estimation. The load models for the wide area power
system are built simultaneously, and the system-wide load modeling method is implemented on the IEEE
39-bus system to evaluate its effectiveness.
2011 Elsevier Ltd. All rights reserved.
1. Introduction
It has been well recognized that the electrical loads have signif-
icant impacts on the dynamic performances of the wide area
power system, and the accurate load models are very important
for the power system dynamic simulation and analysis [16]. Since
the characteristics of the electrical load are random and changing,
the load modeling is a difcult task [7]. In the past few decades,
extensive studies have been carried out in the eld of electrical
load modeling. Two load modeling methods, namely component
based method and measurement based method, are commonly
used in the previous studies. To represent the electrical load char-
acteristics more accurately, the measurement based method re-
cently attracted more and more interests from power engineers
and researchers [812]. In the measurement based method, the dy-
namic responses of the electrical load under a system disturbance
are measured, and a set of parameters for the load model are ob-
tained by the optimization strategy with the objective to minimize
the difference between the simulated dynamic responses of the
load model under the system disturbance and those measured.
Using the measurement based load modeling method, the model
for electrical loads connected at the load transformer, where the
measurement equipment is installed, can be constructed, and the
dynamics of the electrical load itself is concerned. Since there are
hundreds of such electrical loads in the modern power system, it
is difcult and uneconomical to established model for each electri-
cal load, respectively. Hence, the system-wide load modeling has
to be developed.
The initial work about the system-wide dynamic load modeling
in WSCC was carried out in [13,14]. The simulation system was
constructed to reproduce the dynamics of the outage occurred in
western North America on August 10, 1996, and to verify the sys-
tem dynamic models in [13]. It was found that if the load models
were not properly set, the simulation system cannot show the sim-
ilar damping as that of the actual system. It should be mentioned
that numerous simulations were performed to search the suitable
load model parameters. It takes quite long time to repeat many
times simulations, and it was not guaranteed to obtain the satis-
ed result. In [14], the impact of the electrical load, which was
modeled by the composite load model, on the damping of the
power system was studied. The sensitivity of the parameters of
the composite load model was analyzed, and it is concluded that
the percentage of motor modeled at the load bus has the largest
sensitivity. The damping of the power system decreases with the
increase of the percentage of the motor in the total load. However,
as it was demonstrated in the paper, this work was a start in the
system-wide load modeling, and the load modeling method for
the wide area power system needs to be investigated further.
To build the system-wide load models, the measurements valid
among the wide area power system are necessary. With the devel-
opment of the wide area measurement system (WAMS), the sys-
tem-wide measurements, which synchronized by the GPS, can be
obtained. WAMS has been applied in power system monitoring,
0142-0615/$ - see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijepes.2010.12.030

Corresponding author. Tel.: +86 25 83787268; fax: +86 25 83735375.

E-mail address: pju@hhu.edu.cn (P. Ju).
Electrical Power and Energy Systems 33 (2011) 909917
Contents lists available at ScienceDirect
Electrical Power and Energy Systems
j our nal homepage: www. el sevi er . com/ l ocat e/ i j epes
stability assessment, system protection, model validation, fault
location, etc. [1521]. In [22], the WAMS was applied to build
the model for a load, and the accuracy of the load model can be im-
proved using the measurement of WAMS. While, it should be
pointed out that in this work, the load was represented by a static
load model, and loads in the power system were also modeled sep-
arately. Hence, the application of the WAMS in the system-wide
load modeling, where the dynamic composite load model is em-
ployed, has not been fully investigated.
In this paper, a system-wide load modeling method, which
combines the component based load modeling method and the
measurement based load modeling method, is proposed, and the
load models for the wide area power system are built simulta-
neously. The proposed load modeling method is implemented in
the IEEE 39-bus system, and its effectiveness is evaluated. The pa-
per is arranged as follows. In Section 2, the composite load model is
presented. The trajectory sensitivities of the parameters of the
composite load model are analyzed in Section 3. In Section 4, the
procedure of the system-wide load modeling method is intro-
duced, and then in Section 5, the WAMS based system-wide load
modeling is implemented on the IEEE 39-bus system. Conclusions
are drawn in Section 6.
2. Composite load model
The composite load model is the most attractive load model to
power engineers and researchers due to its clear physical meaning
and high proportions of induction motors in the power load, and it
has been widely used in the power system dynamic analysis.
Hence, the composite load model is employed in this work [23].
The composite load model is normally represented by an equiva-
lent induction motor in parallel with a static load. The dynamic
model of the induction motor may vary from one set of state vari-
ables to another, and the model introduced in the power system
analysis software package (PSASP), by China EPRI, is given by
ds
dt

1
T
jL
T
M
T
E
1
de
0
L
dt

1
T
0
d0L
e
0
L
jX X
0
I jT
0
d0L
e
0
L
sx
0
_
2
and
T
M
K
L
1 s
b
; T
E
R
e
e
0
L
I

_ _
e
0
tR
I
R
e
0
tI
I
I
_ _
X
0
X
s

X
r
X
m
X
r
X
m
; X X
s
X
m
; T
0
d0L
X
m
X
r
=R
r
where s is the rotor slip; e
0
L
is the voltage phasor behind the transient
reactance; e
0
tR
and e
0
tI
are the real part and imaginary part of the volt-
age phasor behindthe transient reactance, respectively; I is the stator
current phasor; I

is the conjugate of the stator current phasor I; I

R
and I
I
are the real part and imaginary part of the stator current pha-
sor, respectively; x
0
is the synchronous rotating speed; T
M
and T
E
are
the mechanical and electrical torque, respectively; X
0
is the transient
reactance; X is the stator reactance; R
r
is the rotor resistance; X
s
is the
stator leakage reactance; X
r
is the rotor leakage reactance; X
m
is the
magnetizing reactance; b is the mechanical load exponent; T
0
d0L
is
the transient rotor time constant; T
jL
is the inertia time constant,
and K
L
is the interim coefcients related to initial states.
The static load is represented by a polynomial function as
follows:
P
s
P
s0
A
P
V=V
0

2
B
P
V=V
0
C
P

Q
s
Q
s0
A
q
V=V
0

2
B
q
V=V
0
C
q

3
where P
s
and Q
s
are the active power and reactive power of the sta-
tic load, respectively; V is the terminal voltage of the load; A
p
and B
p
are the voltage dependent coefcients of the active power; A
q
and B
q
are the voltage dependent coefcients of the reactive power; C
p
and
C
q
are the voltage independent coefcients of the active power and
the reactive power, respectively, and V
0
is the pre-fault terminal
voltage of the load.
Eqs. (1)(3) describe the composite load model. If the static load
is represented by the constant impedance load, the independent
parameters of the composite load model are
h g
MP
; K
L
; X
s
; R
r
; X
r
; X
m
; b; T
jL
; T
0
dOL
:
where g
MP
P
M0
=P
L0
is the percentage of the motor in the load; P
M0
is the pre-fault active power of the induction motor, and P
L0
is the
pre-fault active power of the load.
3. Sensitivity analysis of composite load model
3.1. Trajectory sensitivity
The trajectory sensitivity of the composite load model can be
dened as
@y
i
h; k
@h
j
lim
Dh
j
!0

y
i
h
1
; . . . ; h
j
Dh
j
; . . . ; h
m
; k y
i
h
1
; . . . ; h
j
; . . . ; h
m
; k
Dh
j
4
where y
i
is the trajectory of the ith output variable; h is the vector of
the parameters of the composite load model; h
j
is the jth parameter;
k is the sequence of the sampling and D is the deviation of the
parameter.
The normalized trajectory sensitivity is given by
@y
i
h; k=y
i0

@h
j
=h
j0

y
i
h
1
; . . . ; h
j
Dh
j
; . . . ; h
m
; k y
i
h
1
; . . . ; h
j
; . . . ; h
m
; k=y
i0
Dh
j
=h
j0
5
For the convenience of comparison, the average sensitivity can
be calculated as
A
ij

1
K

K
k1
@y
i
h; k=y
i0

@h
j
=h
j0

6
where A
ij
is the average sensitivity of the jth parameter with respect
to the trajectory of ith output variable; y
i0
is the initial value of the
output variable.
3.2. Relation between trajectory sensitivity and parameter estimation
The trajectory sensitivity demonstrates the impact of the varia-
tion of the load parameter on that of the output variables trajec-
tory under a system disturbance. If the trajectory sensitivity of a
parameter is larger than those of the other parameters, the param-
eter plays a more important role in the dynamics of the output var-
iable, in other words, the parameter can be estimated easily using
the dynamics of the output variable. In contrast, if the trajectory
sensitivity of a parameter is very small, e.g. close to zero, the
parameter is unidentiable using the dynamics of the output var-
iable chosen. On the other hand, if a couple of parameters have an
unknown relationship with each other, they are dependent with
each other and unidentiable as well [24]. These unidentiable
parameters can also be detected using trajectory sensitivity analy-
sis. Assuming that parameters h
i
and h
i+1
are coupling with each
other, the output of the power system can be written as
910 P. Ju et al. / Electrical Power and Energy Systems 33 (2011) 909917
y f h
1
; h
2
; . . . ; uh
i
; h
i1
; . . . ; h
n
7
the sensitivities of the h
i
and h
i+1
to the output y can be analyzed as
follows:
@y
@h
i

dy
du
@u
@h
i
@y
@h
i1

dy
du
@u
@h
i1
_

_
8
then, we have
@y
@h
i1

@y
@h
i
@u=@h
i1
@u=@h
i
_ _
9
It should be pointed out that oy/oh
i+1
and oy/oh
i
in (9) vary with
time, while o//oh
i+1
and o//oh
i
are constant. Hence, oy/oh
i+1
and oy/
oh
i
reach zero at the same time. In other words, the trajectory sen-
sitivities of these two coupled unidentiable parameters are in
phase with each other or in reverse.
4. System-wide load modeling for power system
A system-wide load modeling method combining the compo-
nent based load modeling method and the measurement based
load modeling method is proposed, and it aims to make the simu-
lation system demonstrate similar dynamic performances to those
measured. It should be mentioned that during the system-wide
load modeling, it was assumed that there are no structural differ-
ences between the model and the actual simulated load. The pro-
posed system-wide load modeling method can consider one time
point, and only cover one event. To build the load models for the
wide area power system, a system-wide oscillation has to be ob-
served under a serious disturbance, which happens very rarely.
Most of the disturbance could only excite the dynamics of the local
electrical loads. Installing the load modeling equipments in the
substations, the characteristic of the load fed by the substation
can be obtained using the measurement based load modeling
method, which is a node-wide method. With the data observed
at different time and different location, the database of load model
for the power system can be constructed [11]. The load models are
different with the time, while they are similar during similar per-
iod, such as 79 pm in summer. With years accumulation, a few
load models can be obtained during this period. Based on which,
the parameters of the load model during this period can be aggre-
gated. However, this load model was only validated locally, since
the load models were built using the node-wide load modeling
method. Once an event of a system-wide oscillation was recorded
during this period, the proposed system-wide load modeling meth-
od can be implemented to verify the load models built by the node-
wide load modeling method, and the load model can be validated
and improved. The procedure and implementation of the load
modeling method proposed are described as follows.
4.1. Classifying the electrical load
There are a lot of electrical loads in the power system, and it is
impossible to build load model for each electrical load, respec-
tively. A compromise solution is that the electrical loads are classi-
ed into a few categories. The electrical loads in each category have
similar dynamic characteristic, and can be represented by the load
model with the same parameters. In this situation, the number of
the load models needed to be built is reduced signicantly. Com-
pared to the previous simulation system with all the electrical
loads represented by identical load model, the simulation accuracy
of the simulation system with the proposed system-wide load
model can be improved greatly.
The characteristics of the electrical loads in the power grid are
rstly investigated using the component based load modeling
method. The components of the electrical load can be commonly
divided into two categories, namely industrial load and non-
industrial load. Generally, the motor percentage in the industrial
load is much larger than that in the non-industrial load. Hence,
based on the percentage of the industrial load, the electrical loads
can be classied into the following three categories, load with large
portion of industrial load, load with medium portion of industrial
load and load with low portion of industrial load.
4.2. Selecting key parameters to be estimated
The number of the parameters to be estimated also has a signif-
icant impact on the accuracy of the parameter estimation. Apply-
ing the classication of the electrical load in Section 4.1, the
number of load models to be built has been reduced, and the num-
ber of parameters to be estimated is subsequently decreased. Even
though only a few load model to be built, tens of parameters need
to be estimated, since each composite load model contains nine
parameters. If all these parameters are estimated simultaneously,
the optimization strategy will have the problem of convergence,
and fail to nd the parameters for the load models. Hence, the
number of the parameters to be estimated has to be reduced
further.
Performing the trajectory sensitivity analysis, it can be found
that only a few the parameters, such as g
MP
, K
L
, have signicant im-
pact on the dynamic of the power system. It is important to include
these parameters in the estimation algorithm. While, the other
parameters playing limited roles in the dynamic of the power sys-
tem can be set as the default values. In this situation, the efciency
of the parameter estimation can be improved.
4.3. Dening the objective function
The objective of the load modeling is to search the best param-
eter vector h

, which minimizes the error index E.

E

min
hh

2S
Eh 10
The error index can be dened as follows:
E

N
k1
Y
m
k Y
c
k
2
11
or E

N
k1
jY
m
k Y
c
kj 12
or E

N
k1
Y
m
k Y
c
k
2

_
13
where Y
m
and Y
c
are the measured output and the model output,
respectively.
In the traditional load modeling using measurement based
method, the active power and reactive power absorbed by the load
are usually chosen as the model output. Using such objective func-
tions, the dynamic performance of the electrical load itself is of
concern, and the best load model output tting with those mea-
sured can be obtained, while the best dynamic performance of
the power system may not be guaranteed. Hence, the output vari-
ables in the objective function for the system-wide load modeling
have to be chosen as those variables representing the dynamics of
the power grid rather than the dynamics of the electrical load
itself.
The choice of the output variables of the system-wide objective
function depends on the topic researched or the critical issue of the
power grid. For example, with the development of the modern
P. Ju et al. / Electrical Power and Energy Systems 33 (2011) 909917 911
power systems, the very low frequency oscillation happens more
frequently after a system disturbance, and the oscillation mode
are usually the inter-area oscillation, where a group of generators
oscillate against another group. The dynamics of the rotor angle
difference between the centers of the inertia (COI) of these two
groups is very important for the analysis of the oscillating charac-
teristics. Hence, when the inter-area oscillation is studied, the rotor
angle difference between the COI of the two groups of generators
can be chosen as the output variable for the system-wide load
model. The system key variables to be observed are also depend
on fault locations.
To calculate the COI, the phasor of the internal voltage of the
generator is necessary, while it can not be measured directly. Based
on the principle of the synchronous generator, it can be concluded
that the internal voltage keeps still with respect to the rotor. If a
gear wheel and a sensor are installed on the rotor and the stator
of the synchronous generator, respectively, the position of the ro-
tor can be measured. Correcting the phasor difference between
the rotor position and terminal voltage at the no-load condition,
the phasor of the internal voltage can be obtained. While, the mea-
surement of rotor angle difference between different generators is
very difcult for the traditional RTU system, because the genera-
tors locate at different places. Fortunately, with the development
of the wide area measurement techniques, the rotor angle of each
generator can be synchronized by the phasor measurement unit
using GPS. Using the synchronized information, the dynamics of
the rotor angle difference between COIs can be obtained.
4.4. Estimating the parameters
Ant Colony Search algorithm (ACS) is employed in estimating
the key parameters selected in Section 4.2. This algorithm simu-
lates the procedure of the real ant to nd the shortest route be-
tween the ants nest and the source of food by sharing the
information and cooperating among the colony. The ACS has been
employed successfully in load modeling [11,25] and economic dis-
patch of the electrical power [26]. The procedure of parameter esti-
mation using ACS is briey introduced as follows, while the details
can be found in [27,28].
Step 1: Initialization
The initial group of ants is generated randomly in the problem
space, and each ant represents a vector of the parameters of the
load models and is a potential solution for the problem.
Step 2: Evaluation
Evaluate the tness values of the initial colony of ants. The
objective function given in (10) is used to evaluate the tness of
an ant, so as to nd the best ant to minimize the error between
the dynamics measured and the dynamics simulated.
Step 3: Stopping criteria
If the stopping criteria are satised, then stop; else continue to
next step. Here, the stopping criteria may be the maximal number
of iterations, or the tness value of a ant in the colony smaller than
a specied positive value. In this paper, the maximal number of
iterations and the specied positive value are chosen as 20,000
and 10
7
, respectively.
Step 4: Re-distribution of the ant colony
Using the trial distribution function method, the ant colony can
be re-distributed. The number of the ants in each interval is
changed, and the ants in each interval move according to the num-
ber of the ants in the neighboring intervals. Ants move towards the
area with high probability in order to nd optimum solution. After
moving, go to step 2.
5. Implementation on the IEEE 39-bus system
5.1. System description
The proposed system-wide load modeling method is imple-
mented on the IEEE 39-bus system, whose conguration is de-
picted in Fig. 1. During the load modeling, the assumption was
made that the models of the other components in the system have
been validated. As shown in Fig. 1, the IEEE 39-system can be di-
vided into three areas, namely Area A, Area B and Area C.
The characteristics of the electrical load depend on their com-
ponents. Practically, the electrical load in the same area may have
different characteristics, while the characteristics of electrical
loads locating in the different areas may be similar. Hence, the
electrical loads are classied into the following categories in this
paper, and the electrical loads in the same category are modeled
with the identical load model.
Category 1: Electrical loads connected at the load buses no. 3, 4,
15, 16, 18, 21, 23 and 25.
Category 2: Electrical loads connected at the load buses no. 7,
20, 26, 27 and 28.
Category 3: Electrical loads connected at the load buses no. 8, 24
29 and 39.
A three-phase short circuit was applied in the middle of the
transmission line between Bus 26 and Bus 29, and was cleared
after 0.06 s. Under the disturbance, the dynamic simulations
were performed using the software PSASP with the default mod-
els. The default parameters of the load models are listed in the
Appendix. The simulated dynamic responses of the wide area
power system using PSASP are assumed to be obtained by
WAMS. Hence, the data used in this paper for load model identi-
cation is noise free, and the impact of the noise on the load
model parameter estimation needs to be further investigated in
the practical application.
5.2. Trajectory sensitivity analysis
The sensitivities of the composite load model parameters,
h = [g
MP
, K
L
, X
s
, R
r
, X
r
, X
m
, b, T
jL
, T
0
dOL
], for the three categories with
respect to the trajectory of the rotor angle differences between
the COIs of area A and area B, d
BA
, were calculated. d
BA
is dened
as follows:
d
BA
d
B
d
A
d
A

i2A
d
i
T
Ji
T
JA
; T
JA

i2A
T
Ji
d
B

i2B
d
i
T
Ji
T
JB
; T
JB

i2B
T
Ji
14
where d
A
and d
B
are the rotor angle of the COIs of area A and area B,
respectively; T
JA
and T
JB
are the inertia of the COIs of area A and area
B, respectively; d
i
is the rotor angle of the ith generator; T
Ji
is the
inertia of the ith generator. Since the parameters of X
0
and X have
signicant impact on the dynamics of the electrical load, the sensi-
tivities of these two parameters were also calculated using (15) and
(16).
912 P. Ju et al. / Electrical Power and Energy Systems 33 (2011) 909917
where
DX
0
X
s
DX
s

X
r
DX
r
X
m
DX
m

X
s
DX
s
X
m
DX
m
X
s

X
r
X
m
X
s
X
m
_ _
;
DX DX
s
DX
m
The average sensitivities of the parameters, g
MP
, K
L
, X
0
, X
s
and X,
which have more signicant impacts on the trajectory of d
BA
than
the other parameters, are listed in Table 1. It can be seen that the
parameters of g
MP
and K
L
have the highest sensitivities. The loads
in the category 3 have the most signicant effect on the dynamics
of the d
BA
, while the load in the category 1 have the smallest effect
on the dynamics of the d
BA
. The dynamics of the sensitivities are
also shown in Fig. 2. It can be concluded that g
MP
, K
L
and X
s
are
independent with each other, and identiable. While, since the
sensitivity of X
0
are in phase with X, X and X
0
are unidentiable
using the dynamics of d
BA
. Additionally, X and X
0
are the parameters
derived using X
s
, X
r
and X
m
. Hence, only the nine parameters, i.e.
g
MPi
, K
Li
, and X
si
(i = 1, 2, 3), are included in the further step of
parameter estimation, while the other parameter are set as the de-
fault values.
The sensitivities of the composite load model parameters with
respect to the trajectory of the power owing between Bus 28
and Bus 29 as P
2829
and Q
2829
, and the power owing between
Bus 2 and Bus 3 as P
23
and Q
23
were also analyzed. The average
sensitivity of g
MP
, K
L
and X
s
with respect to P
2829
, Q
2829
, P
23
and Q
23
are listed in Tables 2 and 3, respectively, and the dynam-
ics of these sensitivities with respective to P
23
are also depicted in
Fig. 3.
Comparing Tables 13, it can be concluded that the sensitivities
of the load model parameters with respect to d
BA
, P
23
and Q
23
are
much higher than those with respect to P
2829
and Q
2829
, because
the transmission line between bus 28 and 29 locates quite near the
fault location, and its power ow is more sensitive to the fault.
Hence, the dynamics of the angle difference between COIs and
the power owing through the transmission line 23 can be cho-
sen as the output variable for the system-wide load modeling.
Area A
Area B
Area C
Fig. 1. Conguration of the IEEE 39-bus system.
Table 1
Average sensitivities of parameters of composite load models with respect to d
BA
.
g
MP
K
L
X
s
X
0
X
Sensitivities (10
3
) for Category 1 192.3 187.5 16.2 19.9 10.4
Sensitivities (10
3
) for Category 2 576.5 298.3 83.3 131.9 63.4
Sensitivities (10
3
) for Category 3 828.7 204.9 34.1 44.1 18.0
@y
i
h; k=y
i0

@ X
0
=X
0
0
_
y
i
h
1
; . . . ; X
s
DX
s
; X
r
DX
r
; X
m
DX
m
; . . . ; h
m
; k y
i
h
1
; . . . ; X
s
; X
r
; X
m
; . . . ; h
m
; k=y
i0
DX
0
=X
0
0
15
@y
i
h; k=y
i0

@X=X
0

y
i
h
1
; . . . ; X
s
DX
s
; X
m
DX
m
; . . . ; h
m
; k y
i
h
1
; . . . ; X
s
; X
m
; . . . ; h
m
; k=y
i0
DX=X
0
16
P. Ju et al. / Electrical Power and Energy Systems 33 (2011) 909917 913
Fig. 2. Trajectory sensitivity of the parameters of the load model with respect to d
BA
.
Table 2
Average sensitivities of parameters of composite load models with respect to P
2829
,
Q
2829
.
Active power Reactive power
g
MP
K
L
X
s
g
MP
K
L
X
s
Sensitivities (10
3
) for Category
1
14.2 10.9 5.8 10.4 8.4 5.3
Sensitivities (10
3
) for Category
2
31.2 34.6 14.7 21.1 21.5 11.2
Sensitivities (10
3
) for Category
3
30.0 19.1 9.4 17.7 14.2 7.8
Table 3
Average sensitivities of parameters of composite load models with respect to P
23
,
Q
23
.
Active Power Reactive power
g
MP
K
L
X
s
g
MP
K
L
X
s
Sensitivities (10
3
) for
Category 1
98.5 77.8 25.7 25.5 18.9 5.9
Sensitivities (10
3
) for
Category 2
109.9 120.2 49.6 34.1 39.2 13.3
Sensitivities (10
3
) for
Category 3
84.1 62.2 32.9 26.8 20.5 7.5
914 P. Ju et al. / Electrical Power and Energy Systems 33 (2011) 909917
5.3. Parameter estimated with different error indices
The parameters of the load models are estimated using ACS. The
upper and lower bounds of the parameters are listed in Appendix
B, which is the searching space of the ants. With the following error
indices
Index 1: E

N
k1
d
BA m
k d
BA c
k
2
.
Index 2: E

N
k1
jd
BA m
k d
BA c
kj.
The estimated parameters are listed in Table 4. From Table 4, it
can be seen that using these two error indices, the estimated
parameters are all satised. The parameters estimated using Index
1 are a little better than those parameters estimated using Index 2.
5.4. Parameter estimated with different output variables
The parameters are also estimated using the index 1 with differ-
ent output variables, such as d
BA
, and P
23
and Q
23
. The indices can
be written as Index 1 and Index 3, respectively.
Index 3: E

N
k1
P
23 m
k P
23 c
k
2

N
k1
Q
23 m
k Q
23 c
k
2
.
The parameter estimation results are listed in Table 5. Since the
sensitivities of the load parameters to d
BA
are higher than those to
Fig. 3. Trajectory sensitivity of the parameters of the load model with respect to P
23
.
0 5 10 15 20 25 30 35 40 45
6
8
10
12
14
16
18
20
22
t(s)

3
8
Fig. 4. Dynamics of d
38
with system-wide load model and the accurate model.
P. Ju et al. / Electrical Power and Energy Systems 33 (2011) 909917 915
P
23
and Q
23
, the parameters estimated using index 1 is better than
those using index 3. The dynamics of the rotor angle difference be-
tween the generator connected to bus 38 and COI of area A, d
38
, ob-
tained with accurate model and that simulated with model
estimated with index 1 are shown in Fig. 4. It can be seen that
the dynamics of d
38
simulated with estimated model using Index
1 is quite close to that obtained with accurate model. The dynamics
of the output of the system using the system-wide load model ts
those dynamics measured very well.
5.5. Comparing to the traditional node-wide load modeling method
The load models built using the proposed system-wide load
modeling method are also compared with those using the tradi-
tional node-wide load modeling method. During the load model-
ing using the node-wide method, one load model in each category
was constructed, namely load no. 3 in Category 1, load no. 28 in
Category 2 and load no. 8 in Category 3. Three sets of load model
parameters were obtained, and they were applied to the other
load models in the same categories, respectively. The parameters
estimated using different load modeling methods are listed in Ta-
ble 6, and the dynamics of d
38
obtained using the accurate model,
model built using system-wide load modeling method and the
model built using node-wide load modeling method are shown
in Fig. 5.
From Table 6, it can be seen that all the parameters except K
L1
and X
s2
estimated using the system-wide load modeling method
are better than those estimated using the node-wide load model-
ing method. From Fig. 5, it can also be seen that the dynamics of
the d
38
simulated using the model built using these two method
both t that simulated using the accurate model well, and the
dynamics of d
38
simulated using system-wide load model is a little
closer to the accurate dynamics. Hence, using the system-wide
load modeling method, the accuracy of the load model built can
be improved.
Table 5
Load parameters estimated with different output variables.
Parameters True
value
d
BA
P
23
and Q
23
Estimated
value
Error
%
Estimated
value
Error %
g
MP1
0.40 0.4173 4.33 0.4152 3.80
K
L1
0.40 0.4267 6.68 0.4289 7.23
X
s1
0.18 0.1811 0.61 0.2061 14.50
g
MP2
0.50 0.4991 1.80 0.5049 0.98
K
L2
0.60 0.6045 0.75 0.6887 14.78
X
s2
0.16 0.1856 16.00 0.1338 16.37
g
MP3
0.40 0.4239 5.98 0.4186 4.65
K
L3
0.50 0.5070 1.40 0.4293 14.14
X
s3
0.18 0.1809 0.50 0.2059 14.39
Table 4
Load parameters estimated with different error indices.
Parameters True
value
Index 1 = 0.5959637 Index 2 = 0.0927247
Estimated
value
Error
%
Estimated
value
Error
%
g
MP1
0.40 0.4173 4.33 0.4212 5.30
K
L1
0.40 0.4267 6.68 0.4270 6.75
X
s1
0.18 0.1811 0.61 0.1810 0.56
g
MP2
0.50 0.4991 1.80 0.5018 0.36
K
L2
0.60 0.6045 0.75 0.6026 0.43
X
s2
0.16 0.1856 16.00 0.1857 16.06
g
MP3
0.40 0.4239 5.98 0.4242 6.05
K
L3
0.50 0.5070 1.40 0.5065 1.30
X
s3
0.18 0.1809 0.50 0.1808 0.44
Table 6
Load parameters estimated using different load modeling method.
Parameters True
value
System-wide method Node-wide method
Estimated
value
Error
%
Estimated
value
Error %
g
MP1
0.40 0.4173 4.33 0.3658 8.55
K
L1
0.40 0.4267 6.68 0.3835 4.13
X
s1
0.18 0.1811 0.61 0.1692 6.00
g
MP2
0.50 0.4991 1.80 0.4787 4.26
K
L2
0.60 0.6045 0.75 0.4989 16.85
X
s2
0.16 0.1856 16.00 0.1800 12.50
g
MP3
0.40 0.4239 5.98 0.3591 10.23
K
L3
0.50 0.5070 1.40 0.4187 16.26
X
s3
0.18 0.1809 0.50 0.1507 16.28
30 35 40 45
12.2
12.4
12.6
12.8
13
13.2
13.4
13.6
13.8
14
14.2
t(s)

3
8
System wide load model
Node wide load model
Accurate load model
Fig. 5. Dynamics of d
38
with different models.
916 P. Ju et al. / Electrical Power and Energy Systems 33 (2011) 909917
6. Conclusion
In this paper, a WAMS based system-wide load modeling meth-
od has been proposed. Compared with the traditional node-wide
load modeling method, the proposed system-wide load modeling
method combining the component based load modeling method
and measurement based load modeling method, and the load mod-
els for the wide area power system can be built simultaneously.
The accuracy of the load model can be improved using the pro-
posed system-wide load modeling method. Simulating with
system-wide load models, the best dynamics performances corre-
lating to those recorded have been obtained. During the load mod-
eling, the trajectory sensitivities of the parameters of the
composite model with respect to the output of the wide area
power system have also been analyzed. The key parameters which
have important impact on the dynamics of the wide area power
system have been detected, and the relationship between the tra-
jectory sensitivity and the parameter identiability has also been
established.
Acknowledgments
This paper was sponsored by the Scientic Funds for Outstand-
ing Young Scientists of China under Contract No. 50725723. The
authors are also grateful to Energy and Electrical Engineering Col-
lege at Hohai University, China.
Appendix A. Load parameters of the composite load models
Parameters of the load models in Category 1:
g
MP
0:4; K
L
0:4; X
s
0:18; R
r
0:02; X
r
0:12; X
m
3:5; b 2:0; T
jL
2:0 s; s
0
0:0116:
Parameters of the load models in Category 2:
g
MP
0:5; K
L
0:6; X
s
0:18; R
r
0:02; X
r
0:12; X
m
3:5; b 2:0; T
jL
2:0 s; s
0
0:0116:
Parameters of the load models in Category 3:
g
MP
0:4; K
L
0:5; X
s
0:18; R
r
0:02; X
r
0:12; X
m
3:5; b 2:0; T
jL
2:0 s; s
0
0:0116:
Appendix B. Lower and upper bounds for estimated parameters
g
MP1
: 0:32; 0:48; K
L1
: 0:32; 0:48; X
S1
: 0:144; 0:216;
g
MP2
: 0:40; 0:60; K
L2
: 0:48; 0:72; X
S2
: 0:128; 0:192;
g
MP3
: 0:32; 0:48; K
L3
: 0:40; 0:60; X
S3
: 0:144; 0:216:
References
[1] Hiskens IA, Alseddiqui J. Sensitivity, approximation, and uncertainty in power
system dynamic simulation. IEEE Trans Power Syst 2006;21(4):180820.
[2] Banejade M, Ledwich G. Quantication of damping contribution from loads.
IEE Proc Gener Transm Distrib 2005;152(3):42934.
[3] Gordon M. Impact of load behaviour on transient stability and power transfer
limitations. In: Power and Energy Society General Meeting, 2009. PES 09. IEEE,
2630 July 2009, Calgary, Alberta, Canada.
[4] Pereira RMM, Pereira AJC, Ferreira CMM, Barbosa FPM. Dynamic voltage
stability assessment of an electric power network using composite load models.
In: Proceedings of the 43rd international universities power engineering
conference, UPEC 2008, September 1, 2008September 4, 2008, Padova,
Italy.
[5] Kao WS. The effects of load models on unstable low-frequency oscillation
damping in Taipower system experience w/wo power system stabilizers. IEEE
Trans PWRS 2001;16(3):46372.
[6] Ellis A, Kosterev D, Meklin A. Dynamic load models: where are we? Proc IEEE
Power Eng Soc Transm Distrib Conf 2006:13204.
[7] Li YH, Chiang HD, Choi BK, Chen YT, Huang DH, et al. Load models for modeling
dynamic behaviors of reactive loads: evaluation and comparison. Int J Electr
Power Energy Syst 2008;30(9):497503.
[8] Choi BK, Chiang HD, Li YH, Li H, Chen YT, Huang DH, et al. Measurement-based
dynamic load models: derivation comparison and validation. IEEE Trans Power
Syst 2006;21(3):127683.
[9] Han D, Ma J, He RM, Dong ZY. A real application of measurement-based load
modeling in large-scale power grids and its validation. IEEE Trans Power Syst
2009;24(4):175664.
[10] He RM, Ma J, Hill DJ. Composite load modeling via measurement approach.
IEEE Trans Power Syst 2006;21(2):66372.
[11] Ju P, Wu F, Shao ZY, Zhang XP, Fu HJ, Zhang PF, et al. Composite load models
based on eld measurements and their applications in dynamic analysis. IET
Gener Transm Distrib 2007;1(5):72430.
[12] Ranade SJ, Ellis A, Mechenbier J. The development of power system load
models from measurements. Proc IEEE Power Eng Soc Transm Distrib Conf
2001;1:2016.
[13] Kosterev DN, Taylor CW, Mittelstadt WA. Model validation for the August 10,
1996 WSCC system outage. IEEE Trans Power Syst 1999;14(3):96779.
[14] Pereira L, Kosterev D, Undrill DDJ, Zhu W. An interim dynamic induction motor
model for stability studies in the WSCC. IEEE Trans Power Syst 2002;17(4):
110815.
[15] Kamwa I, Grondin R. PMU conguration for system dynamic performance
measurement in large multiarea power systems. IEEE Trans Power Syst
2002;17(2):38594.
[16] Karlsson D, Hemmingsson M, Lindahl S. Wide area system monitoring and
control. IEEE Power Energy Mag 2004;2(5):6876.
[17] Warichet JS, Maun JC. Considerations about synchrophasors measurement in
dynamic system conditions. Int J Electr Power Energy Syst 2009;31(9):45264.
[18] Bertsch J, Carnal C, Karlsson D, McDaniel J, Vu K. Wide-area protection and
power system utilization. Proc IEEE 2005;93(5):9971003.
[19] Yu CS, Liu CW. Self-correction two-machine equivalent model for stability
control of FACT system using real-time phasor measurements. IEE Proc: Gener
Transm Distrib 2002;149(4):38996.
[20] Wang C, Jia QQ, Li XB, Dou CX. Fault location using synchronized sequence
measurements. Int J Electr Power Energy Syst 2008;30(2):1349.
[21] Hauer JF, Mittelstadt WA, Martin KE, Burns JW, Lee H, Pierre JW, et al. Use of
the WECC WAMS in wide-area probing tests for validation of system
performance and modeling. IEEE Trans Power Syst 2009;24(1):2507.
[22] Zhang J, Wen JY, Cheng SJ, Dong ZY. Realization of the WAMS based power
system aggregate load area model. In: IEEE power and energy society 2008
general meeting: conversion and delivery of electrical energy in the 21st
century, PES, July 20, 2008July 24, 2008, Pittsburgh, PA, United States.
[23] Kundur P. Power system stability and control. New York: McGraw-Hill, Inc.;
1994. pp. 279306.
[24] Ju P, Handschin E. Identiability of load models. IEE Proc Gener Transm
Distrb 1997;144(1):459.
[25] Chen Q, Ju P, Shi KQ, Tang Y, Shao ZY, Yang WY. Parameter estimation
and comparison of the load models with considering distribution network
directly or indirectly. In: International journal of electrical power and energy
systems. <http://www.elsevier.com/wps/nd/journaldescription.cws_home/
30432/description#description>.
[26] Pothiya S, Ngamroo I, Kongprawechnon W. Ant colony optimisation for
economic dispatch problem with non-smooth cost functions. Int J Electr Power
Energy Syst 2010;32(5):47887.
[27] Vlachogiannis JG, HatziargyriouND, Lee KY. Ant colony system-based algorithm
for constrained load ow problem. IEEE Trans PWRS 2005;20(3):12419.
[28] Wang L, Wu QD. Ant system algorithm based system identication. ACTA
Automat Sin 2003;29(1):1029.
P. Ju et al. / Electrical Power and Energy Systems 33 (2011) 909917 917