Parameter Estimation of a Complex Load Model

Using Phasor Measurements

Siming Guo, Student Member, IEEE, and Thomas J. Overbye, Fellow, IEEE
Department of Electrical and Computer Engineering
University of Illinois at Urbana-Champaign
Champaign, Illinois, 61820
Email: sguo6@illinois.edu and overbye@illinois.edu

AbstractPower system operators use dynamic models of the

power grid to determine if the transient responses of transmission
devices are within tolerable limits. The focus of this work is on
the parameter estimation of a complex load model by using measurements from digital fault recorders or phasor measurement
units, and the goal is to create a method which will produce
more accurate models. Two methods are presented in this paper:
Compare and Resimulate, and Simulate then Calculate. In a
validation study, the former achieved a mean accuracy of 0.65%
with a 42 minute run time, while the latter had a lower accuracy
of 4.13%, but required only 3.2 seconds of computation time. It
achieved this by performing the PSS/E simulations beforehand,
and then calculating a linear representation of the nonlinear
simulation process. However, both algorithms have weaknesses:
Compare and Resimulate may be vulnerable to measurement
error or noise, and Simulate then Calculate is dependent on the
parameter distribution chosen for the simulations.

I. I NTRODUCTION
Simulations are a useful tool in power systems because they
allow operators and researchers to perform analysis or failure
testing of the electricity grid for planning purposes. More
importantly, simulations can be performed quickly and without
damaging actual components. Simulations require accurate
models of all components of the power system: generators,
transmission components, and loads. Historically, generator
and transmission modeling have received the most attention,
but due to increasing stresses on the grid from various sources,
load modeling has also become important [1]. For example, [2]
showed that in a simulation of a particular inter-area transfer,
the tie-line stability limit found can vary by as much as 50%
depending on the load model. In the absence of an accurate and
verified load model, an operator in such a situation may have to
be overly conservative, thus under-utilizing their resources and
increasing costs [3]. A reliable algorithm for the estimation of
load model parameters is the goal of this work.
II. L OAD M ODELING
Load modeling in theory is similar to generator modeling.
However, the key difference is scale: in the USA, there
are approximately 5800 power plants with a total of 18000
generators [4], whereas there are over 131 million homes
[5]. Additionally, many of the generators are standard types
(e.g. salient rotor synchronous machine), while each home
represents a diverse range of devices from digital electronics
to appliances. Thus, we can model each of the generators
978-1-4577-1683-6/12/$26.00 2012 IEEE

individually, but for the loads, we wish to create an aggregate

model (of a city for instance) that has far fewer parameters
than the addition of all the individual loads.
A. Aggregate Load Modeling Challenges
Aggregate load modeling poses two special challenges,
which are not faced during the modeling of other components
of the power system, such as transformers or generators.
Firstly, while most transmission components can be tested
in isolation before they are placed in service (e.g. to find
generator curves), loads can not. Thus, we need to develop
a method which can create a model of the load while it is online. The second challenge of load modeling is that aggregated
loads are inherently complicated and highly nonlinear [6].
This is because an aggregate load may contain many diverse
components, such as transformers built into devices, power
electronics, industrial motors, and constant impedance loads
such as a radiator. Even if each of these loads themselves
are linear, or even piecewise linearizable, aggregating them
together will most likely not preserve this property.
B. General Load Model Classifications
There are different types of aggregate load models, and
different methods by which to estimate their parameters. This
paper focuses on dynamic, time domain load modeling using
the measurement approach.
1) Dynamic vs Static: A static model is one which is
accurate for steady-state operating conditions. However, often
operators are more concerned with transient stability, and
wish to know if the system remains stable under a particular
disturbance, such as a fault. Disturbances cause a system
to transition from an initial steady state, through a transient
region, to a second operating point. Only a dynamic model can
capture the response during the transient region in addition to
the initial and final steady states. This is important because
even if a system is stable at two operating points, A and B, it
can go unstable during the transition from A to B.
2) Time Domain vs Frequency Domain: Load modeling is
a special application of the general field called system identification, which is a classical signal processing problem. In
system identification, we attempt to determine a mathematical
model of a dynamic system, using observations of responses
to disturbances, coupled with any prior knowledge we may

Fig. 1.

PSS/E CLOD complex load model [7].

have of the system [8]. There are two general methods for
system identification: time domain methods and frequency
domain methods. For power system applications, time domain
is preferred [9], so this paper will focus only on time domain
methods.
3) Measurement vs Component-Based Approach: There are
two time domain approaches currently used for load modeling:
the component-based approach and the measurement-based
approach [9]. The component-based approach determines an
aggregate load model at a bus from a priori knowledge of the
type and quantity of loads connected to that bus, by simply
adding them together. On the other hand, the only knowledge
we have in a measurement-based approach is the model
structurethe actual parameters in the model are calculated
to make the behavior of the model mimic measurements taken
on the grid.

aggregation of smaller models, they will naturally have more

degrees of freedom. Secondly, a complex load model takes
advantage of the knowledge we do have about the load.
For example, the CLOD model takes advantage of the fact
that we know a large portion of the loads on the grid are
either induction motors or discharge lighting systems [14],
by incorporating typical motor and lighting models into the
CLOD model as distinct elements [7]. The remainder of the
load is modeled more generically: as either constant MVA, or
constant current real power (KP = 1 in Fig. 1) and constant
impedance reactive power (PI/QZ). Transformer saturation and
winding losses can be included, are usually insignificant [14].
In the CLOD model, the internal parameters of each of the
basic load types are already defined. What we wish to calculate
is the percentage allocation for each of the load categories.
III. L OAD M ODEL PARAMETER E STIMATION

C. CLOD Load Model

There are many load models which are typically found in
power system simulation packages. Basic models include ZIP,
exponential, and induction machine models, which have all
commonly been used in the past for dynamic studies [10].
However, each of these models alone is relatively simple,
and aside from machine models, they have relatively few
parameters. This limits their potential accuracy, since [10]
found a close correlation between the number of parameters
(i.e. degrees of freedom) in a model and the accuracy of the
model. One way to remedy this is to use more complicated
models, for example by representing the load as set of general
polynomial, exponential, and/or differential equations [1113].
This paper will look at a type of model called a complex
load model, where the model is a collection of two or more
of the basic models listed at the beginning of this section.
Specifically, we will focus on the CLOD complex load model,
as defined in PSS/E, which is shown in Fig. 1. Complex
load models have two advantages. First, since they are an

A. General Approach
In order to determine the dynamic load model, we study
the response of the power system during disturbances. These
disturbances can include generator outages, load steps, and
line faults. The type of measurements used can include voltage magnitudes, angles, line flows, and frequencies, but will
inevitably be tied to the type of disturbance considered. For
example, it is well known that the coupling of real power to
voltage magnitude is much weaker than to voltage angle [15],
so for a real load step, the angle measurements are preferred
for the analysis. For this paper, we study the effect of faults
on bus voltage magnitudes. For a given fault on the system,
the approach taken is similar to that described in [13]:
P
1) Guess the load model parameters, p, with i pi = 1.
2) Simulate the fault in PSS/E using the load model in
Step 1 to generate voltage waveforms at each bus.
Concatenate the waveforms from all the buses, forming
v p [n].

Fig. 2.

37 bus case from [15] used for validation (fault bus highlighted).

3) Compare v p [n] with the actual waveform measured by

sensors on the grid, v real [n] .
4) Repeat Steps 1 to 3 with successively better guesses of
p.
Mathematically, this can be restated as the following least
squares minimization problem:
p = argmin {||v p [n] v meas [n]||2 }
p
X
s.t.
pi = 1

(1)

A perfect parameter estimation algorithm would calculate

a p equal to the real parameters, preal . However, since preal
is unknown, the best realistic algorithm would calculate p
such that v p [n] = v fict [n]. Herein lies a crucial assumption:
there must be a one-to-one mapping between the parameters,
p, and the measurements, v, for both the simulation and the
real power grid. That is, the fault and the simulation must be:
1) Deterministic: A given p should always produce the
same v[n].
2) Injective: Two different sets of parameters p1 and p2
should not produce the same v[n].

For the real power grid, neither determinism nor injectivity

can be guaranteed, and it is not something we can control.
For the simulation, determinism is a given as long as we do
not use stochastic methods such as Monte Carlo. Injectivity
of the simulation, while not guaranteed, can be controlled. As
an extremely simplistic example, if the simulation is linear or
monotonic, then different inputs of p will result in different
outputs v[n]. When the simulation (or a component, such
as the load model) is highly nonlinear and non-monotonic,
injectivity of the simulation also can not be guaranteed. The
implications of this will be discussed in further detail in
Section IV-A3.
B. Validation
In order to validate our algorithm, we require a set of
measurements for which we know the real load parameters. We
can then use the difference between the estimated parameters
and the real parameters to judge the algorithms efficacy.
However, such data does not exist. Therefore, we need to
generate a set of fictional voltage data from a set of fictional
parameters, which we will label v fict [n] and pfict , respectively,
and pretend this is the real load model on the system. For
this paper, the power system used was the 37 bus case provided
in [15], shown in Fig. 2. The disturbance simulated was a

PARAMETER
Large motor
Small motor
Discharge lighting
Transformer saturation
Constant MVA
PI/QZ
Transformer R and X

VALUE
24%
19%
30%
Neglected
9%
18%
Neglected

TABLE I
PARAMETERS DESIGNATED AS THE FICTIONAL LOAD MODEL .

three phase balanced fault at bus Demar69 (highlighted), with

fault impedance Z = j11012 per unit and cleared after
0.5 seconds. The fictional load model parameters arbitrarily
assigned are listed in Table I. The five bold ones are the
parameters the algorithm will solve for.
IV. PARAMETER E STIMATION A LGORITHMS
This paper will investigate two algorithms for calculating
p : Compare and Resimulate, and Simulate then Calculate. For
each, the algorithm will be described, followed by an analysis
of its performance on the parameter estimation problem, and
ending with a comparison of the two methods as well a note
on obstacles to be overcome.
A. Algorithm 1: Compare and Resimulate
1) Algorithm: In Compare and Resimulate (CaR), we solve
(1) using a general nonlinear optimization routine. For this
work, the MATLAB function fminsearch was used, which uses
a Nelder-Mead simplex algorithm [16]. This method was used
because it is a derivative-free algorithm, to minimize the risk of
divergence given the nonlinear nature of the objective function.
2) Performance: The starting guess for fminsearch was set
to 20% for each of the five parameters to be estimated. Plotted
in Fig. 3 are the values of p at each iteration of the algorithm,
and the horizontal lines are the values of pfict . Fig 3 shows the
algorithm converging to within a tolerance of ||pi pi1 ||2 <
1104 in 188 function evaluations. Table II lists the final
solution, and the absolute error |p pfict |. The runtime of
CaR for this validation case study was 42 minutes.
PARAMETER
Large motor
Small motor
Discharge lighting
Constant MVA
PI/QZ

CAR
S OLUTION
23.98%
19.00%
29.06%
8.31%
19.60%
M EAN E RROR

Fig. 3.

Convergence of the CaR algorithm on pfict .

3) Discussion: From Table II, it would seem that this

algorithm performs well. However, since this validation was
done using simulated data, one aspect not yet considered
is measurement error or noise, which would be present if
real sensors were used. Fig. 4 shows the results of 426
separate simulations, each with a different set of parameters p,
compared against the v fict [n] generated from pfict . The X-axis
is parameter distance, as defined in (2), and the Y-axis is the
residual, as defined in (3).
Residual = ||p pfict ||2
Parameter distance =

N
X

(v[n] v fict )2

(2)
(3)

n=1

Ideally, we would like to see that as the parameter distance increased (i.e. p increasingly different from pfict ), the difference

A BSOLUTE
E RROR
0.021%
0.00089%
0.94%
0.69%
1.60%
0.65%

TABLE II
C A R SOLUTION AND ERROR .
Fig. 4. The residual of 426 simulations as a function of the parameter
distance.

in the simulated waveform would also increase monotonically.

While this general trend is seen in Fig. 4, it is clear that a
monotonic increase does not exist. In other words, two vastly
different sets of parameters may produce waveforms that are
nearly identical. With the addition of measurement error or
noise into v fict , the injectivity condition in Section III-A may
no longer hold, resulting in the algorithm possibly converging
to a grossly incorrect p .
B. Algorithm 2: Simulate then Calculate
1) Algorithm: One of the issues of the CaR algorithm
is that the objective function contains a PSS/E simulation.
A new simulation must be performed for each of the 188
function evaluations, for the new value of p at that iteration,
thus leading to the 42 minute computation time. In Simulate
then Calculate (StC), we create a proxy which emulates the
simulation, but has a much shorter computation time. It uses a
combination of least squares fitting and matrix multiplication,
both of which are fast calculations. The proxy works as
follows:
1) Define S sets of parameters with a chosen distribution,
and with S >> 5, creating P RS5 . Also define P A
as the augmented matrix [ 1 | P ].
2) Simulate the fault for each of the S parameter sets,
creating waveforms V [n] RSN , where N is the
number of samples in the simulation period.
6N
3) Solve P A C[n] = V [n] for C
in the least
 R
1

squares sense: C = P A T P A
P AT V
4) Solve for p = argminp {||v p [n] v meas [n]||2 } =
argminp {||[ 1 | p ] C[n] v meas [n]||2 }.
The matrix C generated in Step 3 is the proxy, since it maps
parameters to waveforms. We augment a column of ones to
the the parameter matrix P in order to provide one additional
degree of freedom, since we are attempting to replace a
nonlinear process (a simulation) with a linear one (the matrix
C). Additionally, the reader may note that Step 4 is now just a
linear least squares problem,
 which hasthe analytical solution
1

p = (v meas C 1 ) C 2-6 T C 2-6 C 2-6 T

, where C 1 is the
first row of C, and C 2-6 are the second through sixth rows.
2) Performance: This algorithm replaces the simulation
process with C[n], so the computation should be much faster
PARAMETER
Large motor
Small motor
Discharge lighting
Constant MVA
PI/QZ

CAR
S OLUTION
24.44%
16.97%
25.60%
5.11%
27.88%
M EAN E RROR

A BSOLUTE
E RROR
0.44%
2.03%
4.40%
3.89%
9.88%
4.13%

TABLE III
S T C SOLUTION AND ERROR FOR A BOUNDED UNIFORM RANDOM
DISTRIBUTION .

since the simulation in the function evaluation is replaced by

the matrix multiplication p C. Indeed, we do see a large
reduction in computation time: the StC algorithm ran in 3.2
seconds. However, due to this linearization, the accuracy of
the algorithm suffered, with an increase in average error from
0.65% to 4.13%, shown in Table III
3) Discussion: Like the CaR algorithm, this algorithm also
has a vulnerability which is not immediately apparent from
the results. In Step 1 in the StC algorithm, we choose sets of
parameters to be simulated, based on a chosen distribution. For
Table III, 126 sets of parameters were created based on a uniform random distribution bounded by [0.05, 0.5] (essentially a
Monte Carlo simulation). That is, for each of the parameters
in p, a random value
P between 5% and 50% was chosen, while
still adhering to i pi = 1. An alternative would be to use
an unbounded uniform non-random distribution, where each
parameter in p is allowed to take on one of the values in the
set {0, 0.2, 0.4, 0.6, 0.8, 1}, and all 126 possible combinations
of parameters are considered (akin to an exhaustive search).
The parameter estimation result obtained is shown in Table IV,
where we can see that three of the five parameter estimates
are grossly inaccurate.
PARAMETER
Large motor
Small motor
Discharge lighting
Constant MVA
PI/QZ

CAR
S OLUTION
20.49%
16.26%
2.98%
11.95%
78.17%
M EAN E RROR

A BSOLUTE
E RROR
3.51%
2.74%
32.98%
20.95%
60.17%
24.07%

TABLE IV
S T C SOLUTION AND ERROR FOR AN UNBOUNDED UNIFORM
NON - RANDOM DISTRIBUTION .

One may note that, while the solution algorithm ran in 3.2
seconds, we had to perform 126 simulations beforehand as an
input into the algorithm, which would have taken about the
same amount of time as the CaR algorithm. However, these
126 simulations can be saved, and never need to be repeated.
They can be reused if, at a later time, we make improvements
to the algorithm and wish to recalculate the load parameters.
Alternatively, if we wish to improve the parameter estimation
by using a larger set of simulations, we can simply add on to
the simulations we already performed.
V. C ONCLUSION AND N EXT S TEPS
In this paper, we investigated the performance of two
algorithms for complex load model parameter estimation. The
first, Compare and Resimulate, solves for the load model
using a generic nonlinear minimization routine. It has high
accuracy but suffers from long run times, and is vulnerable to
measurement error and noise. The second, Simulate then Calculate, emulates the time-consuming simulation process using
a simple matrix multiplication, which reduces the computation

time significantly, but also decreases the accuracy of the

solution. Additionally, it is quite sensitive to the distribution
we choose for the parameter set P .
The next step will be to analyze the weaknesses of the
algorithms in more detail. For the CaR algorithm, we need
to determine the amount of noise v meas that it can tolerate (or
equivalently, the signal to noise ratio required). For the StC
algorithm, we need to find a suitable parameter distribution (or
be able to automatically calculate one), since inability to do
so will severely limit its potential. Finally, we wish to test the
algorithms on real transient data, which will involve additional
challenges in signal processing.

[12]

[13]

[14]

[15]
ACKNOWLEDGMENT
The authors would like to thank the Illinois Department of
Commerce and Economic Opportunity for funding this work
through the Illinois Center for a Smarter Electric Grid Project.
R EFERENCES
[1] F. John Meyer and Kwang Y. Lee. Improved Dynamic
Load Model for Power System Stability Studies. IEEE
Transactions on Power Apparatus and Systems, PAS
101(9):33033309, 1982.
[2] Maurice Kent, Wayne Schmus, Francis McCrackin, and
Luther Wheeler. Dynamic Modeling of Loads in Stability
Studies. IEEE Transactions on Power Apparatus and
Systems, PAS-88(5):756763, May 1969.
[3] Richard Bravo, Robert Yinger, Dave Chassin, Henry
Huang, Ning Lu, Ian Hiskens, and Giri Venkataramanan.
Load Modeling Transmission Research. Technical report,
CIEE, 2010.
[4] U S Energy Information Administration. Frequently
Asked Questions. Online document, 2011.
[5] U S Census Bureau. State & County QuickFacts. Online
document, 2011.
[6] Richard Bravo, Robert Yinger, Dave Chassin, Henry
Huang, Ning Lu, Ian Hiskens, and Giri Venkataramanan.
Load Modeling Transmission Research Appendix H Load Monitoring. Technical report, CIEE, 2010.
[7] Siemens Inc. PSS/E 31.0 Program Operation Manual
Volume II, 2007.
[8] K.J. Keesman. System Identification: An Introduction,
volume 2. Springer, 2011.
[9] Jin-cheng Wang and Hsiao-dong Chiang. Development
of a frequency-dependent composite load model using
the measurement approach. IEEE Transactions on Power
Systems, 9(3):15461556, 1994.
[10] Y. Li, H. Chiang, B. Choi, Y. Chen, D. Huang, and
M. Lauby. Load models for modeling dynamic behaviors
of reactive loads: Evaluation and comparison. International Journal of Electrical Power & Energy Systems,
30(9):497503, November 2008.
[11] C.-J. Lin, A.Y.-T. Chen, C.-Y. Chiou, C.-H. Huang, H.D. Chiang, J.-C. Wang, and L. Fekih-Ahmed. Dynamic
load models in power systems using the measurement ap-

[16]

proach. IEEE Transactions on Power Systems, 8(1):309

315, 1993.
Shanshan Liu. Dynamic-Data Driven Real-Time Identification for Electric Power Systems. PhD thesis, University
of Illinois at Urbana-Champaign, 2009.
H. Renmu, M. Jin, and D.J. Hill. Composite Load Modeling via Measurement Approach. IEEE Transactions on
Power Systems, 21(2):663672, May 2006.
J.A. De Leon and Bud Kehrli. The Modeling Requirements for Short-Term Voltage Stability Studies. 2006
IEEE PES Power Systems Conference and Exposition,
pages 582588, 2006.
J D Glover Et al. Power System Analysis and Design.
Thomson Learning, fourth edition, 2008.
MathWorks. Unconstrained nonlinear optimization algorithms. Online document, 2011.