Representative static load models for transient

stability analysis: development and examination

Y. Li, H.-D. Chiang, B.-K. Choi, Y.-T. Chen, D.-H. Huang and M.G. Lauby

Abstract: Load representation has a significant impact on power system analysis and control
results. Currently, static load models are still popular in power industry for transient stability analy-
sis. Dynamic load models are recommended in both industry and academia for possible improve-
ment in analysis accuracy. The accuracy of using static (nonlinear) load models with suitably
identified parameters for transient stability analysis is examined. Numerical studies conducted
using on-line measurement data indicate that static load models, as compared with induction
motor load models, are acceptable for modelling real power behaviours during disturbances and
hence are adequate for transient stability analysis. Using multiple sets of on-line measurements,
representative model parameters of five static nonlinear load models are derived and their perform-
ance in modelling dynamic behaviours of both real and reactive powers is compared. A method is
presented to determine a representative set of parameters of static (nonlinear) load models for each
loading condition.

1 Introduction improvement of accuracy. The incorporation of dynamic

load models, however, increases the dimension of system
Load modelling has a significant impact on power system representation for power system dynamic simulation.
analysis and control results. Adequate representation of This paper examines the accuracy of modelling real
loads may result in considerable savings in system power behaviours by static load models and the adequacy
reinforcement costs and increase the operation flexibility, of using static load models for transient stability analysis.
whereas inaccurate load models may lead to a power Numerical studies conducted in this paper using on-line
system being operated in modes that result in system col- measurement data indicate that static load models, as com-
lapse or separation [1]. Because of its importance, the pared with induction motor load models, are acceptable for
subject of load modelling has drawn significant research modelling real power behaviours during disturbance and
and development efforts from both academia and power hence are adequate for transient stability analysis. The
industry. associated parameters of five static load models, PSS/E
There are two classes of approaches for developing load static model [5], ZIP model [6], exponential model [6],
models: component-based approach and measurement- EPRI model [7] and the composite ZIP-exponential model
based approach. The component-based approach builds a [7], are derived using a total of 11 sets of measurement
load model using detailed information about all the individ- data. Among the five static load model structures, the expo-
ual components at each load bus. The measurement-based nential model offers reliable and consistent results. The per-
approach involves placing measurement systems at the formance of two static load models and two dynamic load
load buses for which load models are to be developed. models in modelling real power behaviours is compared.
The measurement-based approach has the advantage of Compared with static load models, dynamic load models
direct measurement of actual load behaviours. To take under study slightly improve the results in modelling real
advantage of the attractive features offered by the power behaviours at the expense of increasing the system
measurement-based approach, an on-line transient data dimension. A method is presented to determine a represen-
recording system was developed at the Taiwan Power tative set of parameters of static load models for each
Company to investigate the actual load behaviours during loading condition.
system disturbances [2 4].
At present, static load models are still commonly used in 2 Preliminaries
power industry for transient stability analysis. Dynamic
load models, which can capture dynamic responses of the Nine sets of self-acting monitoring systems are currently
loads to disturbances, are recommended for possible installed at primary substations and distribution substations
of the Taiwan power system. When a (natural) system dis-
# The Institution of Engineering and Technology 2007 turbance occurs, the monitoring system at relevant substa-
doi:10.1049/iet-gtd:20060044 tion is triggered automatically to record the three-phase
Paper first received 1st February and in revised form 2nd August 2006 currents and voltages of the substation (i.e. the load bus)
Y. Li, H.-D. Chiang and B.-K. Choi are with School of Electrical and Computer and store the data on a local computer. Currently, there
Engineering, Cornell University, Ithaca, NY 14853, USA are 18 predefined trigger types such as voltage level viola-
Y.-T. Chen and D.-H. Huang are with the System Planning Department, Taiwan tion, frequency limit violation and so on. The discrete
Power Company, Taipei, Republic of China Fourier transformation technique is used to transform the
M.G. Lauby is with the EPRI International, EPRI, Palo Alto, CA 94303, USA recorded data into phasor form. The voltage and current
E-mail: [email protected] phasors at each substation are then used to compute real
422 IET Gener. Transm. Distrib., 2007, 1, (3), pp. 422 431
and reactive power at the substation. The recorder system In this paper, five representative static models were selected
records signals that correspond to voltage variations from for study.
0.5 to 1.2 pu, current variations from 0.1 to 1.0 pu and
frequency variations from 55 to 65 Hz. The sampling rate
of the data recorder system can be flexibly adjusted and 3.1 PSS/E static model
the sampling rate of measurement data in our studies was
3840 Hz per channel (16 channels in parallel) with all In this model, real and reactive power part of the load is
channels being sampled at the same frequency. represented by (1) and (2) which are both voltage and
Input and output data for the study models should be frequency dependent
properly defined for load model developments. For develop-
n1
n2
n3 !
ing both static (nonlinear) and dynamic load models, input V V V
data is defined as voltage magnitude, frequency deviation PL P0 a1 a2 a3 1 a7 Df
V0 V0 V0
whereas output data is defined as real and reactive power
of the load. The real and reactive powers are calculated 1
from the measurement using the positive sequence

n4
n5
n6 !
method. A set of measurement (three-phase voltage and V V V
current) and its corresponding computed real and reactive QL Q0 a4 a5 a6 1 a8 Df
V0 V0 V0
power are shown in Fig. 1. Although it is recognised that
there is still an ongoing issue over the definition of reactive 2
power for unbalanced cases, we use the same definition as
the balanced case.
where a1 a8 , n1 n6 are the parameters to be estimated. And
a1 a2 a3 1, a4 a5 a6 1 when V V0.

3 Static load model structures 3.2 ZIP model

There are a considerable number of static load models pro- ZIP Model is a static load model that represents the power
posed in the literature for transient stability analysis; relationship to voltage magnitude as a polynomial equation
however, several of them have similar model structures. multiplied by frequency dependent term, usually in the

Fig. 1 Real and reactive power calculation using positive sequence method
a and b Positive sequence voltage and current
c Computed real power
d Computed reactive power

IET Gener. Transm. Distrib., Vol. 1, No. 3, May 2007 423

following form 3.5 ZIP-exponential model

"

1 # This model is a composite model of ZIP model and expo-

V 2 V nential model
PL P0 p1 p2 p3 1 Kpf Df 3
V0 V0
2
"
 # P V

1 1  Kpi Kpc Kp1 Kp2
V 2 V P0 V0
QL Q0 q 1 q2 q3 1 Kqf Df 4

npv1
V0 V0 V V
Kpi Kpc Kp1
V0 V0

npv2
where p1 and q1 are the constant impedance load par- V
 1 Kpf 1 Df Kp2 1 Kpf 2 Df 9
ameters, p2 and q2 are the constant current load parameters, V0
p3 and q3 are the constant power load parameters and Kpf
2
Q V
and Kqf are the frequency sensitivity parameters. And 1  Kqi Kqc Kq1 Kq2
p1 p2 p3 1.0, q1 q2 q3 1.0 when V V0. Q0 V0


nqv1
V V
Kqi Kqc Kq1
V0 V0
3.3 Exponential model
nqv2
V
 1 Kqf 1 Df Kq2 1 kqf 2 Df 10
Real and reactive power during restoration reveals exponen- V0
tial characteristics. This observation motivates the con-
sideration of exponential load model expressed in the where Kpi and Kqi are the constant current parts of the total
following form load, Kpc and Kqc the constant power parts of the total load
and Kp1 , Kpf1 , npv1, Kp2 , Kpf2, npv2, Kq1 , Kqf1 , nqv1, Kq2 ,

Kqf2 and nqv2 the parameters of exponential part of the
V KpV total load.
PL P0 1 Kpf Df 5
V0
4 Parameter estimation

KqV
V
QL Q0 1 Kqf Df 6 Once a model structure is specified, load modelling
V0 becomes a parameter estimation problem. Measurement-
based approach for developing static load models can then
where KpV is the voltage-dependent parameter of real be described by the following steps.
power and KqV the voltage-dependent parameter of reactive
power. Step 1: Obtain a set of input output data derived from a set
of measurement data taken at a load bus of the power
system under study.
Step 2: Select a static load model structure.
3.4 EPRI model Step 3: Estimate its parameters using a suitable method and
estimation criterion.
This model is used for dynamic studies in established stab- Step 4: Evaluate the derived static load model using the esti-
ility programs (e.g. EPRIs LOADSYN and ETMSP mation criterion.
packages). Step 5: If the criterion is not met, take remedial actions; for
example, try another estimation method, or try another
(
Kpv1
Kpv2 ) model structure and repeat step 3; otherwise, the estimated
V V parameter values with the associated static model structure
PL P0 Pa1 1 Kpf 1 Df
1  Pa1
V0 V0 constitute a static load model.

7 Our proposed identification process of static load models

(
Kqv1 h is illustrated in Fig. 2. We next describe some compu-
V i
Q 
0 tational schemes behind these steps.
QL P0 Qa1 1 Kq f 1 Df  Qa1
V0 P0 The parameter estimation task is formulated as a non-
) linear least squares problem

Kqv2
V
 1 Kq f 2 Df
8 1X n
V0 min 1P p min 12 p
p p 2 k1 Pk

where Pa1 is the frequency dependent fraction of real load, 1X n

min 1Q p min 12 p 11
Qa1 the reactive load coefficient of uncompensated reactive p p 2 k1 Qk
load to real power load, Kpv1 and Kpv2 are the voltage expo-
nents for frequency-dependent and frequency-independent s:t: gp 0
real power load, Kqv1 and Kqv2 the voltage exponents hp  0
for the uncompensated and compensated reactive power
load, Kpf1 and Kpf2 the frequency sensitivity coefficients where p is a vector of model parameters to be estimated and
for real and uncompensated reactive power load and n the total number of sampling points used for estimation.
Kqf2 is the frequency sensitivity coefficient for reactive 1Pk( p) and 1Qk(p) are errors between the measured value
compensation. and the model output value of real and reactive powers at
424 IET Gener. Transm. Distrib., Vol. 1, No. 3, May 2007
 Obtain measuremant
We split them into two parts: one part containing S 2 1
data form power system sets to be used for parameter estimation and the other part
is used for model validation. This procedure is repeated
for each run. After all S runs are finished, the validation
Prepare input and output data errors are then averaged. It has been shown that the general-
ity and validity of load models can be improved with the
cross-validation technique.
Select a static load
model structure
5 Static load model identification

Estimate parameters of the model

In this section, the proposed load model identification
scheme is applied to the five static model structures to
derive suitable parameter values for each model structure.
Evaluate the model The issues of estimation accuracy and model complexity
are examined to evaluate the estimation performance of
each static model.
No Eleven sets of measurement data belong to certain
Criterion is met? Take remedy actions
loading conditions are selected from the data recording
Yes system installed at an industrial substation, Panchiao substa-
Accept the load model structure tion, of Taiwan power system. On the basis of the recording
and parameters time of each measurement data set, we classified the loading
conditions into four categories: summer peak (SP), summer
Fig. 2 Measurement-based static load model identification medium (SM), summer light (SL) and winter light (WL).
scheme Table 1 summarises 11 sets of measurements at Panchiao
substation.
the kth sampling point. g(p) and h(p) represent equality and These sets of measurements were selected from a larger
inequality constraints for p. number of measurements, the range of voltage magnitudes
To solve the above parameter estimation problem, the variation lies between 5.2% and 16.5% with respect to its
well-known Newton-like method is applied. Since the use pre-disturbance steady-state value. The variation of real
of Newtons method to obtain the parameters can suffer power is in the range of 10.0 32.5% whereas the variation
from ill-conditioned Hessian matrix and can be unsatisfac- of reactive power is generally large (more than 60%).
tory in many instances [8], quasi-Newton methods are We first present the parameter estimation results of the
used in this paper. Among the quasi-Newton methods, a five static load models using measurement set 7 in which
modified algorithm of Broyden, Fletcher, Goldfarb and a large voltage swing occurred. Three-phase voltage and
Shanno (BFGS) is considered to be reliable and well current measurement of set 7 are shown in Fig. 3. In
accepted [8, 9]. It should, however, be noted that the appli- Section 6, measurement data obtained in SM loading con-
cation of any gradient-based optimisation method can dition, SM1 and SM2, are chosen to compare the perform-
potentially lead to a local optimum other than a global ance of static load models and dynamic load models in
optimal solution [10]. One alternative is the application of modelling real power behaviours during disturbance. In
a global optimisation technique to solve problem (11). Section 7, four measurement sets under SL loading con-
After estimation of model parameters, the model dition, SL1 SL4, and four measurement sets under WL
response is simulated with measured input for the validation loading condition, WL1 WL4, are used to derive represen-
of the developed load model. The residual calculation tative load model parameters for different loading
is used for evaluating and validating the accuracy of the conditions.
developed model by the following popular indices Table 2 lists the estimated value of parameters that mini-
mises the error function for the five static models. The
 P 1=2
1=n nk1 12Pk p obtained value may be a local optimal solution of (11).
sP  P 1=2  100%;
1=n nk1 P2k
Table 1: Eleven measurements sets at Panchiao
 P 1=2 substation
1=n nk1 12Qk p
sQ  P 1=2  100% 12 No. Time Loading Voltage Real power
1=n nk1 Q2k
Month Day Hour condition variation, % variation, %
where Pk , Qk are the measured values of real and reactive
1 8 2 00:24 SL1 5.6 10.4
power at the kth sampling point, respectively. If both sP
and sQ are less than desired values, the developed load 2 8 2 00:39 SL2 5.5 10.0
model is then acceptable; otherwise, remedial actions 3 8 2 01:43 SL3 5.5 10.6
should be taken, such as try another estimation algorithm 4 8 2 02:39 SL4 5.2 10.5
for accurate parameter values or try another model 5 8 9 11:34 SM1 9.1 26.4
structure. 6 6 18 11:03 SM2 11.4 12.4
It is desirable to validate the model on a different set of
7 6 19 14:38 SP1 16.5 25.8
measurement data (called validation data), since parameter
variance error could not be detected from the data used for 8 1 11 04:27 WL1 9.0 21.0
parameter estimation (called training data). When multiple 9 1 11 06:06 WL2 9.1 21.8
sets of measurement data are available, we can use the 10 1 10 03:08 WL3 12.6 32.5
cross validation technique for model validation. Suppose 11 1 6 06:51 WL4 9.8 25.0
that there are S(2) sets of measurement data available.
IET Gener. Transm. Distrib., Vol. 1, No. 3, May 2007 425
Fig. 3 Voltage and current profile of a substation during SP

For PSS/E model, the requirements that each parameter is also very satisfactory, 0.75%. It has been found that the
cannot be negative and that the summation of parameters relative errors in modelling reactive power behaviours are
is equal to 1.0 are imposed during the parameter estimation also large but less than that of PSS/E model.
procedure. For ZIP model, the requirements that all par- The parameter number of exponential model is the least
ameters except q2 must be non-negative and that the sum- among five static load models under study, only two. The
mation of parameters is equal to 1.0 are imposed during accuracy of modelling real power is close to other static
the parameter estimation procedure. load models, 0.76%. Moreover, we observe that the expo-
We observe from the estimation results that there are nential load model gives only single parameter lying
seven parameters to be estimated for PSS/E model. The within a reasonable range; hence, the issue of local
relative error in modelling real power behaviour is very sat- optimal solution does not exist in this case.
isfactory, only 0.74%. In contrast, the relative error in mod- The number of parameters of EPRI model to be estimated
elling reactive power behaviours is much larger and is four. The relative error of modelling real power is similar
unsatisfactory. This indicates that the PSS/E load model to other static models, 0.74%, whereas the relative error of
is inadequate for describing the reactive power behaviour modelling reactive power is smaller than other static
during disturbance. models, 12.61%. The ZIP-exponential model has eight par-
For ZIP model structure, the number of parameters is ameters and the relative error in modelling real power is
four. The relative error in modelling real power behaviours 0.75%, whereas the relative error of modelling reactive
power is similar to other static models, 19.51%.
In the above analysis, the amount of data used for the par-
Table 2: Estimation results of static models ameter estimation corresponds to the window length 2 s,
shown in Fig. 3. Pre-disturbance period is 0 0.5 s transient
Parameters Relative period is 0.5 1.5 s and post-disturbance is 1.5 2 s. If this
error, % window length is varied, the parameter estimation results
are shown in Table 3, taking exponential model as an
PSS/E
example.
[a1 a3 , n1 n3 , a7] [0.334, 0.334, 0.332, 0.759, 0.759, 0.74 In Table 3, the first row gives parameters and relative
0.759, 0.338] errors when window length is 2 s, from 0 to 2 s. The second
[a4 a6 , n4 n6 , a8] [0.333, 0.333, 0.333, 4.547, 4.547, 20.64 row gives parameters and relative errors when window
4.547, 6.371] length is 4 s, from 0 to 4 s. A long post-disturbance period
is covered. The third row gives parameters and relative
ZIP
errors when window length is 0.5 s, from 1 to 1.5 s, in
[ p1 p3 , Kpf] [0.189, 0.42, 0.391, 0.3398] 0.75 which a partial transient period is included.
[q1 q3 , Kqf] [2.0, 21.0, 0.0, 3.355] 16.81
exponential
[Kpv , Kpf] [0.772, 0.331] 0.76
Table 3: Estimation results of exponential model with
[Kqv , Kqf] [4.522, 6.479] 20.63 different window length
EPRI
Window Parameter Relative
[Pa1 , Kpv1 , Kpf1 , Kpv2] [0.015, 21.134, 18.753, 0.795] 0.74
length, s error, %
[Qa1 , Kqv1 , Kqf1 , Kqf2 , Kqv2] [1.023, 212.118, 2.814, 12.61
2.771, 212.732] 2 (0 2) [Kpv , Kpf] 0.772, 0.331 0.76
ZIP-exponential [Kqv , Kqf] 4.522, 6.479 20.63
[Kpi , Kpc , Kp1 , Kp2 , npv1, Kpf 1 , npv 2, Kpf 2] [0.6545, 0, 0.75 4 (0 4) [Kpv , Kpf] 0.765, 0.315 0.69
0.2941, 0.2999, 1.0, 0.5647, 1.0, 0.5711] [Kqv , Kqf] 4.243, 6.065 15.73
[Kqi , Kqc , Kq1 , Kq2 , nqv1, Kqf 1 , nqv 2, Kqf 2] [0, 21, 1, 19.51 0.5 (11.5) [Kpv , Kpf] 0.776, 0.378 1.10
1, 2.5562, 2.8596, 2.5553, 2.8618] [Kqv , Kqf] 10.059, 17.501 25.44

426 IET Gener. Transm. Distrib., Vol. 1, No. 3, May 2007

 From the estimation results, we observe that when the Real Power Modeling - ZIP Model
Relative Error (%) - 0.79
window length is less than the length of transient period, 9
derived parameters of real power are similar to whereas
derived parameters of reactive power are much different 8.8
from those when window length is 2 s. Relative error of
real power increases from 0.76% to 1.10% and that of reac- 8.6
tive power increases from 20.63% to 25.44%. It indicates

Real power in p.u.

that the window should cover the entire transient period 8.4
for accurate load model parameters.
Another observation is that when the windows cover a 8.2
long post-disturbance period, the derived model parameters
are similar, whereas the relative errors decrease with the 8
size of window. For instance, the relative error of real
power decreases from 0.76% to 0.69% and that of reactive 7.8 Measured P
power decreases from 20.63% to 15.73%. According to Modeled P
(12), one can observe that the summation of measured 7.6
power increases whereas the summation of difference 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
time in seconds
between measured power and modelled power is almost
same when a longer post-disturbance is covered. Therefore
Fig. 4 Modelled real power behaviour by ZIP model: corre-
relative errors decrease when window covers longer sponding relative error is 0.79%
post-disturbance period.
Overall, all the static load models under study are
acceptable for modelling real power behaviours during behaviours. Fig. 4 shows the measured real power and
disturbance (the relative errors are , 1%). For modelling modelled real power behaviours of ZIP model. And Fig. 5
reactive power behaviours, static load models are shows the measured real power and modelled real power
inadequate. Among the five static load models, we notice behaviours of exponential model.
that the exponential load model is very promising since it
gives reliable and consistently accurate solution and only 6.2 Induction motor load models
has two parameters to be estimated. Therefore from the
viewpoint of modelling real power behaviours during Induction motor load is an important component of power
disturbance, the static load models with suitably identified system dynamic loads. A first-order linear model for small
parameters are acceptable. Since transient stability analysis induction motor was derived in Lesieutre et al. [11]
is mainly concerned with real power behaviours of loads

dDPd dDV
during and after a disturbance, our study supports that DPd Tpp Kpv DV Tpv
appropriate static load models with suitably identified dt dt 13
parameters can be used for transient stability analysis. DQd Kqv DV Kqp DPd

6 Modelling dynamic behaviours of real power The total number of parameters to be estimated is five:
p [Tpp , Tpv , Kpv , Kqv , Kqp]T.
One key goal in this study is to compare the performance of A third-order linear induction motor load model was
static load models and dynamic load models in modelling developed in Chiang et al. [12]. The total number
real power behaviours during disturbance. In this section, of parameters to be estimated is nine: p [Rs , Rr , Xs , Xm ,
the performance of ZIP model, exponential model, first- Xr , H, T0 , vr0 , b]T.
order induction motor load model and third-order induction The parameter values and the relative errors of first-order
motor load model in modelling real power behaviours is and third-order induction motor load models in modelling
compared. Two sets of measurement data obtained in SM dynamic behaviours of real power are given in Table 5.
loading condition are chosen: one set of data is used to esti-
mate model parameters and simulate real power behaviours Real Power Modeling - Exponential Model
during disturbance and the other set of data is used to 9
Relative Error (%) - 0.70

validate the performance of the derived load models.

8.8
6.1 Static models
8.6
Table 4 lists the parameter values and the relative error of
Real Power in p.u.

ZIP and exponential models in modelling real power 8.4

Table 4: Modelling results of ZIP and exponential 8.2

models
8
Parameters Relative
error, % 7.8
Measured P
Modeled P
ZIP 7.6
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
[ p1 p3 , Kpf] [0.1452, 0.6908, 0.1688, 0.2217] 0.79 time in seconds
exponential
[Kpv , Kpf] [0.859, 0.2074] 0.70 Fig. 5 Modelled real power behaviour by exponential model:
corresponding relative error is 0.70%

IET Gener. Transm. Distrib., Vol. 1, No. 3, May 2007 427

Table 5: Modelling results of first-order and third-order Real Power Modeling - 3rd-order Induction Motor
Relative Error (%) - 0.56
induction motor load model 9

Parameters Relative 8.8

error %
8.6
first-order

Real Power in p.u.

[Tpp , Tpv , Kpv , Kqv , Kqp] [0.336, 0.1198, 6.8152, 0.62 8.4
211.7404, 1.4405]
third-order 8.2
[Rs , Rr , Xs , Xm , Xr , H, T0 , vr0 , b [0.0126, 0.56
0.00184, 0.00102, 0.346, 0.0252, 0.885, 15.56, 8

364.23, 12.176]
7.8 Measured P
Modeled P

Figs. 6 and 7 display the measured and modelled real power 7.6
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
behaviours of first-order and third-order induction motor time in seconds
load models. The relative error of first-order induction
motor load model is 0.62% and of third-order induction Fig. 7 Modelled real power behaviour by third-order induction
motor load model is 0.56%. motor model: corresponding relative error is 0.56%
A comparison of relative errors in modelling real power
by different load models is shown in Fig. 8. It is noticed Relative Errors of Modeling Real Power (%)

that in modelling dynamic behaviours of real power, 0.85

0.8
third-order induction motor load model gives the most accu-
0.75
rate result, first-order induction motor load model and expo- 0.7
nential model take the second and third places and ZIP 0.65
model gives a very slightly worse result. However, the 0.6
difference of the relative errors between third-order induc- 0.55

tion motor load model and ZIP model is ,0.25%. More 0.5
ZIP Exponential 1st-order 3rd-order
cases will be given in the next section to confirm this
observation. Fig. 8 Comparison of modelling real power by different load
How to model dynamic behaviours of real power during a models
disturbance is a long-standing issue. This issue includes the
appropriateness of using a static nonlinear load model for
modelling real power behaviours during a disturbance. only the increase in the dimension of differential equations
Although the adoption of a dynamic load model such as is considered. It clearly shows the extent that dynamic load
the third-order induction motor load model may slightly models increase the dimension of system representation.
improve the accuracy of modelling real power, it should The more complicated a dynamic load model is, the more
be recognised that a dynamic load model will increase the system dimension it increases.
dimension of system representation for power system Overall, induction motor load models slightly increase
dynamic analysis and simulation. For example, Taiwan results in modelling real power behaviours, as compared
power system is a medium-sized isolated power system. with ZIP and exponential models, at the expense of the
The number of load buses is around 500. Applying different increase of system dimension. From a viewpoint of captur-
load models for system dynamic simulation, the percentages ing dynamic behaviours of real power during disturbance,
of system dimension increase are shown in Table 6. Here, static nonlinear load models with suitably identified par-
ameters are acceptable and hence can be used for transient
stability analysis.
Real Power Modeling - 1st-order Induction Motor
Relative Error (%) - 0.62
9
7 Representative parameter values
8.9

Since a set of parameters is obtained based on a given indi-

8.7
vidual measurement data set, the question of how to deter-
8.5
mine a representative set of parameters for a specific
loading condition arises when multiple measurement data
Real Power in p.u.

8.3
sets are available pertained to the loading condition. It is

8.1 Table 6: Percentages of system dimension increase for

different load models
7.9
Load model System dimension
7.7 Measured P increase
Modeled P
7.5 ZIP 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
time in seconds exponential 0
first-order induction motor 12.5%
Fig. 6 Modelled real power behaviour by first-order induction third-order induction motor 37.5%
motor model: corresponding relative error is 0.62%

428 IET Gener. Transm. Distrib., Vol. 1, No. 3, May 2007

hence desirable to develop a set of representative par- Hence, a set of representative parameters for a specific
ameters for a load model structure that collectively can be loading condition has been obtained. The effectiveness
used for a loading condition under study. To this end, a of the representative parameters is verified using different
method is presented in this section to derive a representative measurement sets taken at the loading condition.
set of load parameters for each loading condition. Fig. 9 show the real and reactive power behaviours of
When multiple measurement data sets are available (e.g. exponential model of WL1 WL4 measurement data sets.
m measurement data sets), a representative set of parameters Again, it shows that static load models are adequate in
of a certain loading condition can be developed by minimis- modelling real power behaviours (relative error ranges
ing the following error function that is a weighted sum of from 0.7% to 1.1%) and can be used for transient stability
individual mismatch error of m measurement data sets analysis.
X
m
p arg min wi 1i p 13
p
i1 8 Conclusions
where wi is a weighting factor for ith error function In this paper, the adequacy of using static load models
1i ( p) . 1i( p) can be calculated according to (11). This for- for transient stability analysis is examined. Model par-
mulation is a weighted nonlinear least squares problem and ameters of five static nonlinear load models are derived
by introducing weighting factor wi , the importance of each with multiple sets of on-line measurements, and their
measurement data set can be emphasized. For simplicity, wi performances in modelling real and reactive power beha-
is fixed to unity in this study. viours are compared. From the numerical studies, we have
Table 7 gives representative sets of parameters of expo- the following.
nential model for SL and WL loading conditions, respect-
ively. The second column gives representative parameter 1. Static nonlinear load models can be developed to
values of different loading conditions. And the relative mis- approximate the dynamic behaviours of real power.
match errors and average error are listed in the third to the Hence, static load models can be suitable for transient stab-
seventh column when applying the representative parameters ility analysis; this development of static load models,
to four measurement data sets of each loading condition. however, requires accurate parameter estimation based on
In Table 1, voltage dips of different measurement sets are multiple measurement data.
given. Under SL loading condition, SL1 has the largest 2. For modelling real power behaviour, all five static
voltage dip whereas SL4 has the smallest voltage dip. models under study give similar relative errors. The expo-
Under WL loading condition, WL3 has the largest voltage nential load model is very promising since it gives reliable
dip whereas WL1 has the smallest voltage dip. However, and consistently accurate solutions and only has two
the load model with the representative parameters does not parameters to be estimated.
give the smallest relative error under SL1 or WL3 and the 3. In modelling real power behaviours, third-order induc-
largest error under SL4 or WL1. It appears that there is no tion motor load model gives the most accurate result,
direct correlation between voltage dip and relative errors. first-order induction motor load model and exponential
The representative real and reactive power models for SL model take the second and third places and the ZIP model
and WL are shown in (14) and (15). The average errors in give a very slightly worse result. However, the difference
modelling real power behaviours are 0.55% and 0.90%. of the relative errors between third-order induction motor
The average errors in modelling reactive power behaviours load model and ZIP model is very small, less than 0.25%.
are 8.68% and 9.32%. 4. In modelling real power behaviours, induction motor

0:836 load model, does not give much more accurate results
V than ZIP model and exponential model. Since transient stab-
SL: PL P0 1 0:386Df ;
V0 ility analysis is mainly concerned with real power beha-

9:4 viours of load during and after a disturbance, appropriate
V static load models are adequate for transient stability
QL Q0 1 7:468Df 14
V0 analysis.

1:174 5. Adopting dynamic load models in transient stability
V
WL: PL P0 1 0:419Df ; analysis increases system dimension and can significantly
V0 require computing effort for power system dynamic

11:953 simulation. This may limit the application of dynamic
V
QL Q0 1 3:094Df 15 load models in situations where computation speed is
V0 critical.

Table 7: Representative sets of parameters of exponential model for two loading conditions

A set of parameters Relative error in real power, % 1avg (p) (%)

[Kp v ,Kp f] S/WL1 S/WL2 S/WL3 S/WL4

SL [0.836, 0.386] 0.59 0.55 0.53 0.55 0.55

WL [1.174, 0.419] 0.95 0.70 1.07 0.86 0.90

A set of parameters Relative error in reactive power, % 1avg (p) (%)

[Kqv ,Kqf] S/WL1 S/WL2 S/WL3 S/WL4
SL [9.400, 7.468] 8.47 9.88 9.77 6.58 8.68
WL [11.953, 3.094] 10.83 10.63 8.83 6.99 9.32

IET Gener. Transm. Distrib., Vol. 1, No. 3, May 2007 429

 Real Power Modeling - Exponential Model Reactive Power Modeling - Exponential Model
Relative Error (%) - 0.95 Relative Error (%) - 10.83
3.9 0.7

3.8 0.6

3.7 0.5

Reactive Power in p.u.

3.6 0.4

Real Power in p.u.

3.5 0.3

3.4 0.2

3.3 0.1

3.2 0

3.1 Measured P -0.1 Measured Q

Modeled P Modeled Q
3 -0.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
time in seconds time in seconds
a
Real Power Modeling - Exponential Model Reactive Power Modeling - Exponential Model
Relative Error (%) - 0.70 Relative Error (%) - 10.63
3.9 0.7

3.8 0.6

3.7 0.5

3.6 0.4

Reactive Power in p.u.

Real Power in p.u.

3.5 0.3

3.4 0.2

3.3 0.1

3.2 0

3.1 -0.1

3 Measured P -0.2 Measured Q

Modeled P Modeled Q
-0.3
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
time in seconds time in seconds
b
Real Power Modeling - Exponential Model Reactive Power Modeling - Exponential Model
Relative Error (%) - 1.07 Relative Error (%) - 8.83
4.5 1.2
4.4
1
4.2
0.8
Reactive Power in p.u.

4
Real Power in p.u.

0.6

3.8
0.4

3.6
0.2

3.4 0

3.2 Measured P -0.2 Measured Q

Modeled P Modeled Q
3 -0.4
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
time in seconds time in second
c
Real Power Modeling - Exponential Model Reactive Power Modeling - Exponential Model
Relative Error (%) - 0.86 Relative Error (%) - 6.99
4.8 1.2

4.6 1

4.4 0.8
Reactive Power in p.u.
Real Power in p.u.

4.2 0.6

4 0.4

3.8 0.2

3.6 Measured P 0
Measured Q
Modeled P Modeled Q
3.4 -0.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
time in seconds time in seconds
d

Fig. 9 Modelled real and reactive power behaviours of exponential model of WL1 WL4 measurement data sets
a WL1: relative errors of modelled real and reactive powers are 0.95% and 10.83%, respectively
b WL2: relative errors of modelled real and reactive powers are 0.70% and 10.63%, respectively
c WL3: relative errors of modelled real and reactive powers are 1.07% and 8.83%, respectively
d WL4: relative errors of modelled real and reactive powers are 0.86% and 6.99%, respectively

430 IET Gener. Transm. Distrib., Vol. 1, No. 3, May 2007

6. From a practical viewpoint, it is desirable to develop a 4 Lin, C.-J., Chen, Y.-T., Chiang, H.-D., and Wang, J.-C.: Dynamic
representative set of parameters for a load model structure load models in power systems using the measurement approach,
IEEE Trans. Power Syst., 1993, 8, (1), pp. 309 315
that can capture the load behaviour for a specific loading 5 Tseng, K.-H., Kao, W.-S., and Lin, J.-R.: Load model effects on
condition. To this end, a method is presented in this paper distance relay settings, IEEE Trans. Power Deliv., 2003, 18, (4),
to derive a representative set of load parameters for a pp. 11401146
static load model structure for a loading condition. 6 Kundur, P.: Power system stability and control (McGraw-Hill, Inc.,
New York, 1994)
7. Load models are derived for each substation using the 7 IEEE Task Force on Load Representation for Dynamic
measurement-based approach. This decoupled approach PerformanceBibliography on load models for power flow and
is applicable to any power system, medium-size or large dynamic performance simulation, IEEE Trans. Power Syst., 1995,
scale, or to more complex systems. 10, (1), pp. 523538
8 Dias, L.G., and El-Hawary, M.E.: Nonlinear parameter estimation
experiments for static load modeling in electric power systems, IEE
Proc., 1989, 136, (2), pp. 68 77
9 References 9 Fletcher, R.: Practical methods of optimization (John Wiley & Sons,
New York, 1987, 2nd edn.)
1 CIGRE Task Force 38.02.05: Load modeling and dynamics, Electra, 10 Knyazkin, V., Canizares, C.A., and Soder, L.H.: On the parameter
1990, (130), pp. 122141 estimation and modeling of aggregate power system loads, IEEE
2 Chiou, C.-Y., Huang, C.-H., Chiang, H.-D., and Yuan, J.-L.: Trans. Power Syst., 2004, 19, (2), pp. 10231031
Development of a micro-processor-based transient data recording 11 Lesieutre, B.C., Sauer, P.W., and Pai, M.A.: Development
system for load behavior analysis, IEEE Trans. Power Syst., 1993, and comparative study of induction machine based dynamic P, Q
8, (1), pp. 1622 load models, IEEE Trans. Power Syst., 1995, 10, (1), pp. 182191
3 Wang, J.-C., Chiang, H.-D., Chang, C.-L., and Liu, A.-H.: 12 Chiang, H.-D., Wang, J.-C., Huang, C.-T., Chen, Y.-T., and Huang,
Development of a frequency-dependent composite load model C.-H.: Development of a dynamic ZIP-motor load model from
using the measurement approach, IEEE Trans. Power Syst., 1994, on-line field measurements, Int. J. Electr. Power Energy Syst.,
9, (3), pp. 1546 1556 1997, 19, (7), pp. 459 468

IET Gener. Transm. Distrib., Vol. 1, No. 3, May 2007 431