Sei sulla pagina 1di 6

Power System Dynamic Load Identification and Stability

S. Z. Zhu*

Z. Y. Dong** K. P. Won<** Z. H. Wang*

* Tsinghua University, Beijing, P. R. China


** National University of Singapore, Singapore

*** T h e University of Western Australia, Western Australia


ABSTRACT- Proper power system load models are playing more
and more important roles in power system stability analysis in

todays stressed power systems. Different load models may


cause large difference in stability analytical results.
Measurement based load modeling gives closer look at the real
power system loads and their dynamic characteristics. In this
paper, Genetic Algorithms and Evolutionary Programming
based system identification is used to locate the best available
parameters for proper load modeling based on real field test data
taken in North China.

- Static/Dynamic Load Modeling, System


Identification, Load Model Identification, Stability
Analysis, and Genetic Algorithms.

KEYWORDS

I. INTRODUCTION
Power system modeling is important for system stability
analysis. Traditionally, generation and system modeling
have been studied and various models have been given.
However, it is not enough for stability study without a
proper load model, especially in modem stressed power
grids. Different load models can have significant impact
on the overall stability analytical results [l] and [2]. Due
to some uncertainties in load modeling, it is regarded as a
difficult task to represent proper load models. Such
uncertainties include large number of load components,
the accessibility of certain loads due to different
ownership and difficulties in on line measurement for
various voltage and frequency ranges [3]. Besides the
traditional static model, there are many specific load
models for various kinds of loads and systems [ 11, [2], [3],
[41, [51, and [61.
There are two approaches in load modeling, (i) the
component based approach, which models the load on the
basis of familiarity with static and dynamic behavior of
all the individual loads and load components of a
particular load bus; and (ii) the measurement based
approach, which uses system identification to estimate a
proper model and its parameters. The component based
approach encounters difficulties when surveying the
individual characteristics of a large Varity of components,
which is possible even for a small area modeled as a load
bus. Moreover, it is even difficult to combine models of
the same type, which may operate under different voltage
levels [7]. Measurement based approach requires
measurement at various load buses to determine the
appropriate model structure and parameters. It comes

0-7803-6338-8/00/$10.00(~)2000IEEE

from direct measurement of real load behavior and may


give better modeling results over the component base one.
In this paper, we use the measurement-based approach to
identify the load models and their parameters aimed at
providing the most appropriate models for system
stability analysis. Genetic Algorithms (GAS) based search
and optimization approach is employed to achieve the
load model identification purpose. Measurement data
from field tests are used to give identification example
and show the effectiveness of the proposed algorithm.
11. POWER
SYSTEMLOADMODELING

The fact that loads are a complex, stochastic, and time


varying combination of many different devices may
inform power engineers and analysts that it is insensible,
probably even impossible to model all the devices
connected to the realistic power systems. In out context,
the term load refers to the equivalent representation of
the aggregate effect of many individual load devices, and
includes sub-transmission and distribution networks that
connects consumers to the transmission network and
generation.
and
the
the

-P

-Q

Figure 1. Composite dynamic load model


It is a composition of a static load (Ls), a generic dynamic
recovery load (LG) [5] and an aggregate induction motor
load ( L I ~ )[SI. The static load represents all the static
parts of the load. The generic recovery load is used to
take account of the effects of all down stream on-load tap
changer (OLTC) actions and thermo-statically controlled
heating loads. An aggregate induction motor is used to
represent all down stream compressors and other rotating

13

loads. Given wl, w 2 and w3 as the weighting factor of


each component, the composite load is represented as
L = wlLs + w2LG+ w3LIM,with C w = 1 (2.1)
The values of wl, w and w vary for different load buses
depending on their load compositions. For a load with
high concentration of industrial components, for example,
a larger value of w may be assigned.

or a polynomial function, such as a quadratic function;


Recovery time constants Tp and T, range from 60s to
150s. The values for a; and p, will take the same values
as a and p i n the static component, while & a 2.0,h;.2.5
[61.
I

1.1,

Here we describe the loads and their aggregated


characteristics that significantly present voltage
sensibility. Details of the load modeling are as follows,
(1) The static load is modeled as an exponential function
of voltage V

(Cf

($f

(2.2)
where Po, Qo are static load consumptions at the rated
voltage Vo. The indices a and p are the parameters
chosen to best represent the voltage dependence of the
aggregate load, and normally have a range of a = 0.5
1.8, p =1.5
6 according to Kundur [8]. Xu et al.
Proposed a = 0.31 1.50, p = 2.22 4.18 based on the
field test in [8]. The general trend is that, high
concentration of residential load exhibits larger a and
smaller p; while high concentration of commercial /
industrial loads exhibits smaller a and larger p [8] and [9].
For example, their values can be chosen as a = 0.8 1.5,
p = 2.0 4.0 for each bus depending on the load
composition. However these static load models neglect
the critical important dynamic behavior exhibited by
many loads.

Pd = PO

Qd

= QO

(2) A number of generic dynamic load models have been


proposed recently in the literature for voltage stability
studies, see - [SI, [6], [lo] and [SI. A first-order dynamic
recovery model proposed by Hill and Karlsson in [SI and
[6] will be used to illustrate the impact of load modeling
on system stability. This model captures the load
restoration characteristics with an exponential recovery
process. Figure 2. shows the typical power response of
aggregate loads to a voltage step and its exponential
approximation. Some examples of this response are
provided in [ I l l . Mathematically, this model can be
expressed in state space form as,
xp =P,(V)-P,
(2.3)

xq = Qs (U- Qd

(2.4)

+Q,W

(2.6)
where Pd and Qd are the load real and reactive powers,
xp and xy are the corresponding load states, Tp and Tq are
the load recovery time constants. Qs,
P, and Qt, Pl are the
steady state and transient load characteristics respectively.
Normally they are expressed as a function of node
voltage V,either exponential as
Qd

=txq

8, = QO

14

kp Q,
,

(3) A steady-state equivalent circuit of Induction Motor


as Figure 3 shows can be used [8]. When we neglect the
stator transients, the aggregate IM is represented by its
first order model,
(2.9)
s = [T, (s) - T, (3, U1
Where s is the motor slip; H is moment of inertia; T, and
T,,, are the electromagnetic and mechanical torques
respectively; T, a P&, V) in per unit if neglecting effects
of R, and T,,, is assumed to be constant. Parameters for IM
can be taken from [8].

&

TfS

Figure 3. IM steady-state equivalent circuit


These models will be used in load modeling and study of
modeling impact on system stability in the following
sections.
111. GA/EP & LOADIDENTIFICATION

(2.5)

+p,(V)

Pd = * x p

Figure 2. Typical power response of aggregate loads to


a voltage step

= QO

(gp

(2.7)

In this section, we propose the algorithm for load


identification. System identification and EP fitness
function for the specific purpose of load modeling
identification. [ 121
Genetic Algorithms (GAS) are heuristic algorithms,
which can locate the global optimal solution. The GA
optimization mechanism is developed from the concept of
natural evolution, where the strongest individuals survive
and the weaker ones die off during the evolution process.
Part of the work is to develop an effective modification of

a genetic algorithm to optimally determine the load


model parameters with system identification algorithms.

weight(i) = c w ,

;l'..l.J
'c
/=I

For better GA performance, adaptively adjusted mutation


probability can be used [13], [14], [15], [16], and [17]. In
this paper two parallel processes are used for mutation
probability control, as follows,
(4. I )
P,(i,y) = P," *exp(-+i/nl)*exp(-y/n2)
where P, is the mutation probability, which takes the
initial value of P,", +i is the fitness of the i-th individual,
y is the generation number, nl and n2 are adjusting factors
controlling the decreasing rate of P, taking into
consideration of fitness and generation number
respectively.
In our algorithm, the production mutation occurs when a
Gauss-distribution random vector is added into the parent
generation. The basic algorithm adopted in this paper is
as follows,

1. Problem Formulation: The solution X of the


optimization problem is represented by a ddimensional vector X = [x,, x2, ... ,xd 1, and uJ< xJ <
4, where xJis the j-th dimension of the vector; uJ,bJ,
are the lower and upper limits of xi. $), i=1,2,. .....,
n is i-th individual string in one generation, n is the
size of solutions.
2. Solution Initialization: The initial solution is
produced by a function on a random basis
considering the limits of each individuals by:
q ( I ) = rund()*(bJ-uJ)+uJ
(4.2)
Where V i = 1,2, ..., n, V j = 1,2, ..., d, and rand() is
function which can produce a random vector.
3. Fitness Function: For each solution $), its fitness
function is calculated by using function:
f;= F'(X(')), which is the objective function; and
e, = E(X(')),which is the identification error function.
in
4. Mutation Operation: Each new individual 9)
the current generation is reproduced by mutation.
The mutation operation is based on adding a
uniform probabilistic vector of last generation to the
current individual as given in Equation (4.2)

(4.4)

Where ,U',, is a random vector between (0, l), e, is


the advantage.
6. Selection: According to the value of weight, all
individual (2n) are arranged in sequence. The first n
individuals selected as the next generation.
7. Return to 4 until the convergence condition is
satisfied.
Iv. POWER SYSTEM MODEL ANALYSIS
In this section, we test out algorithm with some real field
test data to identify load models for further stability
analysis. The data is from the field measurement from
Tong Liao Power Plant and the'neighboring area of the
North East China grid. The one line diagram of the test
system is given in Figure 4. Tong Liao Power Plant
locates in the eastern part of Inner Mongolia. The
electricity is transmitted to North East China Grid via
three 220KV transmission lines over a distance of more
than 200 km.
17s

jdw

tongliao

rhuangliao

tcngbian

176, 177

Figure 4. Power System Around Tong Liao Power Plant

WJ

- x(')./ +N(o,PJ --+'J)>


-

emax

V j = 1,2,-. -,d
Where

N(O,P,

z , ) is the

(4.3)
mutation

emax

operation vector, and the

PJl e+

Z J is the

emax
determination variable based on the value of
mutation probability Z.,
5. Tournament: The competence of tournament is
calculated by each individual's tournament penalty
weighting factor, weighf(i).This factor is calculated
by comparing with other randomly selected
individuals.

1V.I. Load Model Identification


We take the population size n = 50, the number of each
individual competing with others m = 30. The Elite
percentage, which decides the percentage of reserved
individuals in each generation be lo%, so Elite = n*10%
= 5 ; mutation probability z = 0.001, and scale factor B =
e-O.OS*step-num
, where step-num is the total number of
iterations. The limits of variables to be identified are
given in Table 1.
Table 1. Variables range of solution vector

Lowerlimit I 0

-1

I -3

10

1 -1

-30

15

These limits are used only during the initial population is


being produced. Now we give some detailed case studies
from field test and computer identification simulation.
Example 1. Static load model
The active power P and reactive power Q are identified
respectively. The identification process terminates after
100 generations. The results are shown in Table 2.
Table 2. The identification results of static load model

parameter z = 0.001, and fl = e-0~08*step-num


. In our
algorithm, we consider two aspects of modeling: (1)
order of model, A4 is often set to be 1 or 2, and (2)
linearity of the model, i is Boolean variable, and i = 0
stands for linear model, i = 1 for nonlinear model.
We used dynamic load model to fit the test data with
iteration number of 3000. Both first-order linear model
and second-order nonlinear model are identified. Figure 6
shows the simulated results of active and reactive powers.
Identification error and modeling parameters are given in
Table 3.

Active Dower
Table 3. The identification results of the dynamic load
Error
2.42 16307e-002
2.42 1666 1e-002
2.42 16854e-002
2.42 16894e-002
2.42 17 187e-002
x3
-13.95077
-14.16945
-14.14562
-1426829
-1420455
Error
6.856216.8737617MKl3
6.8765852dH3
6.877633W3
6 . m 3
x3
-522094
-528481
-533203
-52'7143
-539731

Reactive Dower
A

Therefore, the identification gives the static load model as:


P = 0.454084V1.405687
Q = 0.179435V3.206189
In our identification process, PO,al, Qo, bl are very closed
after only one identification, and all solution individuals
converged close to their final solution point very quickly.
Since the frequency f changed very small in field tests,
the impact of frequency variation is ignored and the static
load model is used. The identified and measured real data
of the reactive and reactive powers are shown as Figure 5 .

-Y

x2
14.05219
1427952

X1

0.68979
0.68653
~~

0.68764
0.68925
0.69oOo

x4
-7.87665
-822913
-8.05477
-820608
-829263

1425461
1437090
1430551
X5

~~~~

XI

0.95424
0.95633
0.95445

9.16634
9.47256
921439
930574
9.42222
x2
535434
5.40915

5.46509
539985

0.95560

0.95516
x4
4.m
-5.73658
-5.75931
-5.97893
-5.98805

m
X$

82027
7.88834
7.79500
8.031%
8.08046

0.48

0.46
0.45

0.W

1.00

1.01

1.02

1.I3

Figure 5. The curves of static loads between identified


and real measured data
Example 2. Dynamic load model
Let the population size n = 200, the number of each
individual competing with others m = 80. One generation
is left by IO%, i.e., Elite = n*10% = 20, mutation

16

i n
n

qn

rn

cn

cm

inn

r?n

Figure 6. The comparison curves of reactive power:


identified and real measured
Based from the results obtained, we can conclude-that,

1.The results are satisfactory for the requirement of


system security analysis from field test. We also
performed Least Square identification, and it can be
seen that our G A E P based algorithm gives better
performance over the LS approach. The algorithm is
robust with different orders of linear model.
2. The initial generation of solutions can be produced by
random models or combining some consideration of
the actual plant to be identified.

-Constant

Eq' Model and Constant resistance

...... Classic Eq" Model and Constant resistance

..

1 4

IV.11. Load Modeling and System Stability


Based on the load identification results, simulation of the
power grid is carried out. The results are given in Figures
7 - 10. In the simulation, it is assumed that all 4
generators in Tong Liao Power Plant have the same
-Constant

10

t(s)

0
0

10

Figure 9. Power plant angle stability after Dianju


trrnsmiwinn line nn lnnd disrnnnertinn.

Eq' Model and Constant resistance

......Classic Eq" Model and Constant resistance

Online Eq" Model and Exponential function

-Constant
Eq' Model and Constant resistance
......Classic Eq" Model and Constant resistance
Online Eq" Model and Exponentialfunction

2.44

I
1P

Ac
tiv

0.4

3.5

, t~s)

0.2
0

10
1.51

Figure 7. P at # 1 Diantong transmission line after a %phase


dinrnnnertinn

generation level, the network operation condition is


during later peak hours, and system loads are taken the
classical loading levels under normal operation
conditions. The system faults include: (i) disconnection
of the three phase main transmission line, (ii)
disconnection of the three phase main transmission line at
-Constant Eq' Modeland Constant resistance
......Classic Eq" Model and Constant resistance
Online Eq" Model and Exponentialfunction

.
0

.
2

.
6

.,

tfs).

10

Figure 10. #2 Generator active power behavior after Dianju line


nn-lnad 3 nhrne dismnnertinn.

In Figures 7 - 10, the solid lines in each figure stand for


constant Eq' model, dashed lines stand for Eq" classical
model, and the short dashed lines stand for performances
following field test for the Eq" model. The load models
described before. As we can see, the simulation results
fits very closely to the real field tests. This demonstrated
These stability simulations under various system
operation and loading conditions are very useful for
future operation planning of Tong Liao Power Plant.
Algorithms used here can certainly be applied to other
systems to investigate the system load modeling and
stability conditions in a very practical and less risky way.

05

t(s)
0

V. CONCLUSIONS

10

Figure 8. Active power transients of Dianling line after 3 phase


dinrnnnrrtinn nf niantnne line.

the terminal out of the power plant, (iii) main


transmission line two phase short circuit at Ju Feng
terminal side.

Genetic Algorithms and Evolutionary Programming


based identification is used in the paper to identify the
power system load parameters based on data from field
measurement. Several load models are used to simulate
the identification process. Improvements over the basic
genetic algorithms are proposed including considerations

17

on mutation probability control, fitness formulation and a


progressive concept for search and optimization. Both
theoretical analytical simulation and field tests are carried
to validate the effective of the algorithm for load models
and their impact on system stability. It can be seen from
these simulations and tests the algorithms proposed in the
paper gives satisfactory results of identification for
further stability analysis. Further researches are being
carried out to develop a more comprehensive general load
model suitable for measurement based model
identification and system stability analysis.

REFERENCES
Lin, C.-J., et al., Dynamic Load Models in Power
Systems Using The Meansurement Approach, IEEE
Trans. Power Systems, Vol. 8, No. 1, 1993, pp.309315.

[9] Xu, W., E. Vaahedi, Y. Mansour and J. Tamby


(1997). Voltage stability load parameter
determination from field test on b.c. hydros system.
IEEE Trans. on Power Systems 12(3), 1290-1297.
[lO]Xu, W. and Y. Mansour (1994). Voltage stability
analysis using generic dynamic load models. IEEE
Trans. on Power Systems 9(1), 479-493.
[ 111IEEEPES Power System Stability Subcommittee
Special Publication, Voltage Stability Assessment
Procedures and Guides, 1999

[12]Zhu, S. Z., D. Shen, Y. H. Zhen, Q. Ai, L. Li and S .


M. Shahidehpour, The Study of Genetic Algorithm
for Load Modeling, Proc. International Power
Engineering Conference, Singapore 1999, pp.787792.

Mauricio, W. and A. Semlyen, Effect of Load


Characteristics on The Dynamic Stability of Power
Systems, IEEE Trans. Power Apparatus and
Systems, Vol. 91, 1972, pp.3303-3309.

[ 131Goldberg, D. E., Genetic Algorithms in Search,


Optimization, and Machine Learning, AddisonWesley Publishing Co. Inc., 1989.

IEEE Task Force, Load Representation for


Dynamic Performance Analysis, IEEE Trans.
Power Systems, Vol. 8, No. 2, 1993, pp. 472-482.

[14] Wong, K. P., and Y. W. Wong, Genetic and Genetic


/ Simulated-Annealing Approaches to Economic
Dispatch, IEE Proc. C. 1994, 141, (5), pp. 685 -692.

Hill, D. J. and I. A. Hiskens, Modeling, Stability


and Control of Voltage Behavior in Power Supply
Systems, Invited paper, IV Symposium of Specialists
in Electric Operational and Expansion Planning,
Foz do Iguacu, Brazil, May 1994.

[15]Ma, J. T., and L. L. Lai, Improved Genetic


Algorithm for Reactive Power Planning, Proc. 12th
Power Systems Computation Conference, Dresden,
Sweden, August 19-23, 1996, pp.499-505.

Hill, D. J. (1993). Nonlinear dynamic load models


with recovery for voltage stability studies. IEEE
Trans. on Power Systems 8(1), 166-176.
Karlsson, D. and D. J. Hill (1994). Modeling and
identification of nonlinear dynamic loads in power
systems. IEEE Trans. on Power Systems 9 (l), 157166.
Milanovic, J. V., The Influence of Loa& on Power
System Electromechanical Oscillations, PhD Thesis,
University of Newcastle, Australia, 1996.

18

[SI Kundur, P. (1994). Power system stability and


control.

[16] Wong, K.P., and J. Yuryevich, EvolutionaryProgramming-Based Algorithm for EnvironmentallyConstrained Economic Dispatch IEEE Trans.
Power Systems, Vol. 13, No. 2, May 1998, pp. 301306.
[17] Wong, K.P., and Y. W. Wong, Floating-Point
Number-Coding Method for Genetic Algorithms,
Proc. ANZIIS-93, Perth, Western Australia, 1-3
December 1993, pp. 512-516.

Potrebbero piacerti anche