www.ietdl.

org
Published in IET Generation, Transmission & Distribution
Received on 19th September 2008
Revised on 22nd December 2008
doi: 10.1049/iet-gtd.2008.0485

ISSN 1751-8687

Choice of estimator for distribution system

state estimation
R. Singh1 B.C. Pal1 R.A. Jabr 2
1

Department of Electrical and Electronic Engineering, Imperial College, London, UK

Electrical Computer and Communication Engineering Department, Notre Dame University, Zouk Mosbeh, Lebanon
E-mail: [email protected]
2

Abstract: In this study, a statistical framework is introduced to assess the suitability of various state estimation
(SE) methodologies for the purpose of distribution system state estimation (DSSE). The existing algorithms
adopted in the transmission system SE are recongured for the distribution system. The performance of three
SE algorithms has been examined and discussed in standard 12-bus and 95-bus UK-GDS network models.

Nomenclature
m, n
xt , x

number of measurements and state variables

true state and estimated state vectors, respectively
(n  1)

Px , P x

numerically computed and estimated

covariance matrices, respectively (n  n)

E[.]
e

expectation operator
normalised state error squared variable

measurement vector (m  1)

h(x)
szi

expectation of measurement vector (m  1)

standard deviation of the ith measurement

Rz
ez

measurement error covariance matrix (m  m)

measurement error vector (m  1)

ri

normalised residual of ith measurement

error

Introduction

Deregulation of power system and the introduction of

distributed generation (DG) to distribution networks has
challenged the operational philosophy of the distribution
systems. The passive nature of the network can only
accommodate restricted amount of DG capacity. This means
that signicant network reinforcement will be necessary to
accommodate DG and load growth in the future. An
alternative would be to change the approach to network
operation such as introducing control to distribution network
666
& The Institution of Engineering and Technology 2009

operation. A range of technology innovations is needed to

change the way distribution systems operate. The innovations
must pin down on new architecture for distribution network
control centre with performance critical software functions
like state estimation (SE), optimal power ow (OPF) and
network-specic sensor placement and integration.
In transmission systems, SE is a fairly routine task and a host
of established methodologies exist [1]. These cannot simply be
transferred to distribution systems because the planning, design
and operation philosophy of distribution networks are different
from those in the transmission networks. The distribution
network topology and characteristics are different and most
importantly the amount of available network measurements is
very limited. The SE methodologies adopted in transmission
systems start showing their limitations when exposed to the
specics of distribution networks [2].
Furthermore, the potential benets of using SE technologies
in distribution network control have not been explored mainly
because of the absence of adequate network measurements
and also the lack of rigorous methodology and tools that
could be applied on restricted measurements. The
development of new distribution system state estimation
(DSSE) is a challenging task as the tools to evaluate the
quality of SE must consider a number of issues relating to
measurement types, locations and numbers.
Methodologies on which such tools could be built are not
available at present. However, some interesting research has
IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

www.ietdl.org
been done in DSSE [3 10]. Lu et al. [3] propose a threephase DSSE algorithm. The algorithm uses a currentbased formulation of the weighted least-squares (WLS)
method in which the power measurements, current
measurements and voltage measurements are converted to
their equivalent currents, and the Jacobian terms are
constant and equal to the admittance matrix elements. The
observability analysis of the proposed distribution system is
also discussed. Lin and Teng [4] have proposed a new fast
decoupled state estimator with equality constraints. The
proposed method is based on the equivalent current
measurement in rectangular coordinates. Baran and Kelley
[5] have introduced a computationally efcient algorithm
based on branch currents as state variables. The method is
demonstrated to work well in radial and weakly meshed
systems. This concept is further rened by Wang and
Schulz [6] and they have presented a revised branch
current-based DSSE algorithm. In this algorithm, the load
estimated at every node from an automated metre reading
system is used as a pseudo measurement. Li [7] has
presented a distribution system state estimator based on
WLS approach and three-phase modelling techniques. Li
has also demonstrated the impact of the measurement
placement and measurement accuracy on the estimated
results. A rule-based approach for measurement placement
is presented by Baran et al. [8]. Ghosh et al. [9] have
presented an alternative approach to DSSE using a
probabilistic extension of the radial load ow algorithm
treating the real measurements as solution constraints. The
algorithm that accounts for non-normally distributed loads,
incorporates the concept of load diversity and can interact
with a load allocation routine. The eld results are
discussed in [10].

techniques along with the details of their state error

covariance matrices are discussed in Section 3. The efcacy
of the algorithms is examined on standard test systems and
discussed in Section 4.

The DSSE literature is either based on the probabilistic

load ow or direct adaptation of transmission system SE
algorithms (particularly WLS). The issue of measurement
inadequacy is addressed through pseudo measurements that
are stochastic in nature. However, the performance of the
SE algorithms under the stochastic behaviour of pseudo
measurements is not addressed in the DSSE literature.

where, P x , denotes the estimated state error covariance matrix.

The work presented in this paper investigates the existing

transmission system SE techniques and algorithms and
assesses their suitability to the DSSE problem. The selected
algorithms are tested on the 12-bus and 95-bus UK-GDS
network models against some statistical measures like bias,
consistency and overall quality of the estimates. Unlike many
other distribution systems, the UK distribution network is
fairly balanced and that has prompted us to go with a singlephase approach, although the method is generic.
Furthermore, the statistical measures utilised in this paper
mainly depend on the probability distribution of the
measurements and not on the line model of the network.
Following this introduction, a theoretical framework for
the statistical measures is established in Section 2. The
consistency and the quality of the estimates utilise the
asymptotic state error covariance matrix. The various SE
IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

Statistical measures

In distribution systems, measurements are predominantly

of pseudo type, which are statistical in nature, so the
performance of a state estimator should be based on some
statistical measures. Various statistical measures such as
bias, consistency and quality have been adopted for
assessing the effectiveness of SE in other technology areas
such as target tracking [11]. We explore these for the
DSSE applications. Briey we describe the statistical
measures as follows.

2.1 Bias
A state estimator is said to be unbiased if the expected value
of error in the state estimate is zero. Mathematically an
unbiased estimator can be dened as
E[(xt  x^ )] 0

(1)

2.2 Consistency
If the error in an estimate statistically corresponds to the
corresponding covariance matrix then the estimate (and
hence the technique generating this estimate) is said to be
consistent. One measure of consistency is the normalised
state error squared variable
1

e (xt  x^ )T P^ x (xt  x^ )

(2)

For the estimator to be consistent e should be within its

condence bounds, which can be obtained from the error
statistics.

2.2.1 Choice of condence regions: In the univariate

case when the estimation error is represented by a normal
distribution with zero mean and known variance, one can
use the tables of normal distribution to compute the
condence intervals. However, in the multivariate case when
the estimation error is represented by a normal distribution
with zero mean vector and known covariance matrix, such
condence intervals are difcult to compute because tables
are available only for the bivariate case. Alternatively, one
could setup limits for each component on the basis of
distribution, but this procedure has the disadvantages that
the choice of limit is somewhat arbitrary and in some cases
leads to tests that may be poor against some alternatives.
Moreover, such limits are difcult to compute. The
procedure given below, which is based on x2-statistics, can
be easily computed and applied in the multivariate case.
Furthermore, it can be theoretically justied based on the
following lemma. The proof the lemma can be found in [12].
667

& The Institution of Engineering and Technology 2009

www.ietdl.org
trace of the error covariance matrix can be written as

Lemma 1: If an n-component vector v is distributed

according to N (0, T ) (non-singular), then vT T 1 v is
distributed according to x2-distribution with n degrees of
freedom.

2.2.2 x2-statistics: It can be shown that if the errors

in measurements are normally distributed, the SE error

corresponding to these measurements will be normally
distributed with zero mean vector and covariance matrix
given by E[(xt  x^ )(xt  x^ )T ]. By utilising this fact and
Lemma 1, the normalised squared error e (2) should follow
a x2-distribution with n degrees of freedom for a consistent
estimator, where n is the number of states. In other words
for the estimator to be consistent, e should lie within its
condence bounds that can be obtained from the standard
x2-table for a chosen condence level a. Lower and upper
bounds for this condence level can be given by
x2n ((1  a)=2) and x2n ((1 a)=2), respectively. In statistics,
a 95% condence interval is considered to be adequate.

Various algorithms have been suggested for transmission system

SE [1]. All these algorithms work well in transmission systems
because there is high redundancy in the measurements.
However, in distribution systems, because of sparsity of
measurements, there is less or no redundancy in the
measurements. Hence, when these algorithms are exposed to
distribution systems they start showing their limitations. For
example, in transmission systems, weighted least absolute
value estimator (WLAV) eliminates bad data out of redundant
measurements, but in distribution systems it fails to work
because it treats every pseudo measurement as bad data and
there is no redundancy to eliminate these pseudo measurements.

2.2.3 x2-test over Monte Carlo simulations: In

practice, statistical tests are performed using a number of
Monte Carlo simulations. Consider the system has n
number of states and M is the number of Monte Carlo
simulations, then the normalised squared error follows a
x2-distribution with Mn degrees of freedom. Mathematically
E[e]

x2Mn (a)
M

(3)

For large number of Monte Carlo runs x2Mn (a) Mn, which
results in
E[e] n

(4)

Hence the mean of e should approach to the number of states

with the increase in the number of simulations.

2.3 Quality
Quality of an estimate is inversely related to its variance. For
the multivariate case, the square root of the determinant of
the error covariance matrix measures the volume of 1 2 s
ellipsoid and is used here to quantify the total variance of
an estimate. Hence, the quality of the estimate can be
dened as
!

1
Qdet log p
det (P x )

(5)

Sometimes, in large networks, it becomes difcult to

compute the determinant of the error covariance matrix
numerically because of precision limits of the solver. In this
situation, an alternate way to dene the quality is to use the
trace of the error covariance matrix. However, this ignores
the off-diagonal information. The quality as function of the
668
& The Institution of Engineering and Technology 2009


Qtrace log

1
tr(P x )


(6)

State estimation techniques

This section briey explains the most common SE

techniques to examine their suitability for the DSSE problem
under stochastic behaviour of the pseudo measurements and
limited or no redundancy. All these techniques use the
following measurement model.

3.1 Measurement model

(7)

z h(x) ez

where, ez  N (0, Rz ) is zero mean Gaussian noise with error

covariance matrix Rz ( diag{s2z1 , s2z2 , . . . , s2zm }). We dene
the normalised residual of ith measurement ri as
ri

zi  hi (x)
szi

(8)

where, ri  N (0, 1). The class of estimators discussed in this

section are based on maximum likelihood theory. They
rely on a priori knowledge of the distribution of the
measurement error (Gaussian in this case, with zero mean
and known covariance). A generalised estimation problem
seeks to minimise the following objective
J

m
X

r(ri )

(9)

i1

The different estimators can be characterised based on the

choice of the r function.

3.2 Weighted least squares estimation

WLS is a quadratic form of the maximum likelihood
estimation problem. The WLS problem can be stated as
the minimisation of the following objective function
1
[z  h(x)]T R1
z [z  h(x)]
2

(10)

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

www.ietdl.org
the estimator. An expression for a is given by [14]

The above objective takes the form given in (9) for

1
r(ri ) ri2
2

(11)

An estimate of state was obtained iteratively using the

Newton method according to
xk ))1 H T (^xk )R1
xk )]
x^ k1 x^ k (H T (^xk )R1
z H (^
z [z  h(^
(12)
where

H (^xk )

@h(x)
@x


(13)

[z  h(x)]jj1
jjR1=2
z

The numerical computation of a for various estimators

is given in the Appendix. Table 1 summarises the
various estimators used in this paper. Table 1 indicates a
typical value of a for SHGM considering avi 1:5. Since
the IRLS method is used for SHGM, the value of
a changes during the estimation process depending upon
the weight vi .

Case study

The algorithms discussed in the previous section were applied

on a 12-bus radial distribution network model and on a part
of the UK generic distribution system model (95-bus UKGDS). Figs. 1 and 2 show the schematic of the test
systems. Network and load data for these networks can be
found in [16] and [17], respectively.

4.1 State variables

r(ri ) jri j

(15)

Existing techniques use linear programming or interior point

methods to solve this problem. In this paper we have used a
primal dual interior point method [13].

This estimator combines both WLS and WLAV estimators.

The r function for SHGM estimator is given by
if

jri j  avi

The bus voltage magnitudes and angles were considered as

state variables except at the reference bus (Bus #1) for
which the bus angle was assumed to be zero. Hence, the
number of states to be evaluated was 23 and 189 for the
12-bus test system and the UK-GDS, respectively.

4.2 Measurements

3.4 Schweppe Huber generalised M

(SHGM) estimator

>
: av jr j  1 a2 v2
i i
i
2

(18)

(14)

which is equivalent to (9) when

r(ri )

(E[c0 (r)])2

where c(r) @r(r)=@r and c0 (r) @c(r)=@r.

WLAV estimator is based on the minimisation of the L1

norm of weighted measurement residual, and can be
expressed as

1 2
r
2i

E[c2 (r)]

x^xk

3.3 Weighted least absolute value

estimator

8
>
<

It was assumed that the errors associated with the

measurements are independent identically distributed
(i.i.d.). Three types of measurements were taken into
consideration. The telemetered measurements were utilised
as real measurements. Zero injections with a very low

(16)

otherwise

Table 1 State estimators: summary

Solution for x Asymptotic error covariance P x

The performance of this estimator highly depends upon the

weight factor vi and tuning parameter a. In this paper, the
solution to this problem was obtained using iteratively reweighted least squares (IRLS) method [1]. The parameter
a 1.5 was used in simulations.

3.4.1 State error covariance matrix: An estimate of

Newton

(HT (^x)R1
x))1
z H(^

WLAV

PDIP

SHGM

IRLS

p T
x)R1
x))1
z H(^
2 (H (^
1.037(HT (^x)R1
x))1 a
z H(^

WLS

a 1.5, vi 1

the asymptotic covariance matrix at convergence can be

expressed as [14, 15]
x))1
P^ x a(H T (^x)R1
z H (^

(17)

where x^ limk x^ k . The value of a depends on the choice of

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

Figure 1 Twelve-bus test system

669

& The Institution of Engineering and Technology 2009

www.ietdl.org

Figure 2 UK-GDS: 95-bus test system

variance (1028) were modelled as virtual measurements.

Loads were modelled as pseudo measurements. Various
scenarios considering the errors in real measurements as 1
and 3%, whereas 20 and 50% in pseudo measurements
were examined. The range of error in pseudo measurements
was chosen on the basis of errors in load estimates of
various classes of customers, like industrial, domestic and
commercial. The loads of the industrial customers can be
estimated more accurately than the domestic and
commercial, thus they have less error. On the other hand,
loads of domestic customers are difcult to estimate, hence
they have large error. The error in commercial load
estimates lies between the two. It was also taken into
consideration that with this choice of range, the maximum
demand limits at various buses are not violated and the
condition of linear approximation is valid. The mean value
for these measurements was obtained using distribution
system load ow. Table 2 summarises the measurements
and their redundancy level for the two test network models.

4.3 Measurement variance

A +3s deviation around the mean covers more than 99.7%
area of the Gaussian curve. Hence, for a given % of maximum
error about mean mzi , the standard deviation of error was
computed as follows

szi

mzi  %error
3  100

(19)

The square of standard deviation gives the variance of the

measurement.

4.4 Simulation results

The performance of the estimators was evaluated for the
following cases:
Case 1: Error in real measurement 1% and pseudo
measurement 20%

Table 2 Measurements used in study

Test system
twelve-bus
UK-GDS (a) limited
redundancy
UK-GDS (b)
increased
redundancy

Real measurements (mr)

Virtual & pseudo

measurements (mp)

Redundency
(mr mp)/n

3(V1 , P12 , Q12 )

22 (loads only no zero

injections)

25
1:09
23

5(V1 , P12 , Q12 , P185 , Q185 )

188 (loads and zero

injections)

193
1:02
189

21(V1 , V18 , V19 , V20 , V21 , V95 , P12 , Q12 ,

P185 , Q185 , P1819 , Q1819 , P8295 , Q8295
P1517 , Q1517 , P3435 , Q3435 , d19 , d20 , d21 )

188 (loads and zero

injections)

670
& The Institution of Engineering and Technology 2009

209
1:11
189

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

www.ietdl.org
Case 2: Error in real measurement 1% and pseudo
measurement 50%
Case 3: Error in real measurement 3% and pseudo
measurement 20%
Case 4: Error in real measurement 3% and pseudo
measurement 50%

4.4.1 Twelve-bus system: In the 12-bus test system, the

voltage magnitude measurement at bus #1 and power ow
measurement in line #12 were considered as real
measurement. Fig. 3 shows the variation of the expected value
of the normalised state error squared with various Monte
Carlo steps for the 12-bus distribution system. It is clear from
the gure that as the number of Monte Carlo steps increases,
the expected value of normalised state error square variable
approaches the number of states, which agrees with (4). Also
after 400 Monte Carlo steps, the error in E[e ] is within 1% of
the number of states. Hence, we chose 400 Monte Carlo steps
for the simulations. A larger number of Monte Carlo steps
gives slightly better results but it increases the computation time.
Fig. 4 shows the error plots with the number of simulations
for the three estimators. The plots shown are for the worst
case scenario (Case 4), that is, the error associated with real
measurements is 3% and that with pseudo measurements is
50%. The estimation errors for all the states are displayed
in Fig. 4, however, they are indistinguishable because of the
overlaps. It is evident from the gure that the error varies
about zero mean. This indicates that all the three
estimators are unbiased. It was also found that for all other
cases, the three estimators were unbiased.
Figs. 58 show the consistency plots for the estimators for
cases 14. A 95% condence level was used to dene the
condence bounds. It was found that WLS shows consistent
results in all test cases. On the other hand, WLAV is

Figure 4 Twelve-bus system estimation error plot for all

state variables: error in true measurements 3%, error in
pseudo measurements 50%
inconsistent in all the cases. It is interesting to note that
SHGM is inconsistent for small errors in pseudo
measurements and consistent for large errors in pseudo
measurements. The reason is that the measurement set
considered for study is predominantly comprised of the
pseudo measurements, and large error in pseudo
measurements increases the measurement variance (19). Also
the computation of variance in (19) is based on the maximum
error. This results in low normalised residual (jri j) for pseudo
measurements. Owing to this fact the normalised residual
becomes less than the cutoff value avi (16), and the estimator
behaves like WLS. However, this is not always true.
Whenever the normalised residual exceeds the cutoff value,
the estimator becomes inconsistent. It will be shown that for
the 95-bus UK-GDS system, SHGM becomes inconsistent
for these cases of large errors too.
Table 3 shows the performance summary of the 12-bus test
system. Two types of qualities are shown. As expected, the

Figure 3 Variation of E[e ] with different Monte Carlo steps

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

671

& The Institution of Engineering and Technology 2009

www.ietdl.org

Figure 5 Twelve-bus system consistency plot: error in true

measurements 1%, error in pseudo measurements 20%

Figure 6 Twelve-bus system consistency plot: error in true

measurements 1%, error in pseudo measurements 50%

quality of the estimates decreases with the increase in the

error in measurements. This decrease is signicant with the
increase in error in the real measurements as compared to
the pseudo measurements.

WLS was found to be consistent in all the test cases.

Hence, the consistency plots of WLS in all four test cases
are displayed in Fig. 9. The consistency plot for SHGMis
also shown in Fig. 10 for the test Case 2. It is clear from
Fig. 10 that the SHGM which was consistent in Case 2 in
the 12-bus system no longer remains consistent in larger
systems.

4.4.2 Ninety-ve-bus UK-GDS: The performance of

the estimators was also evaluated on the 95-bus test system
model for all the test cases analysed in the 12-bus test
system. It was observed that in the 95-bus test system also,
400 Monte Carlo steps are sufcient to bring down the
error in E[e ] within 1% of the number of states. The
following two cases were considered:
(a) Limited redundancy
In this case, the real measurements were considered to
be available at the main substation. Hence, the voltage
magnitude measurement at bus #1 and power ow
measurements in lines #1 2 and #1 85 were taken as real
measurements. It was observed that in the 95-bus test
system all the estimators were unbiased. However, only
672
& The Institution of Engineering and Technology 2009

(b) Increased redundancy

In this case, the redundancy was increased by placing
the measurements at DG locations rst and then
measurements were placed at optimal locations. The
optimality criterion and details of the measurement
placement appear in [18]. Furthermore, the phasor
measurements were also deployed at optimally selected
buses. The real measurement set in this study consists of
the following measurements:
1. Voltage measurements at buses #1, #18, #19, #20, #21
and #95
IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

www.ietdl.org

Figure 7 Twelve-bus system consistency plot: error in true

measurements 3%, error in pseudo measurements 20%

2. Line ow measurements in lines #1 2, #1 85, #82 95,

#18 19, #15 17 and #34 35
3. Phasor measurements at buses #19, #20 and #21

Figure 8 Twelve-bus system consistency plot: error in true

measurements 3%, error in pseudo measurements 50%
The consistency plots for WLS and SHGM with increased
redundancy are shown in Figs. 11 and 12, respectively. The
WLS shows the consistent performance whereas the
SHGM shows the inconsistency in all the simulated cases.

Table 3 Twelve-bus system performance summary

Estimator

Real 1%, pseudo 20%

Real 1%, pseudo 50%

Bias

Consistent/E[e ]

Quality tr/det

Bias

Consistent/E[e ]

Quality tr/det

WLS

unbiased

consistent/23.13

4.08/199.24

unbiased

consistent/23.25

3.7/179.01

WLAV

unbiased

inconsisten/136.7

3.17/191.46

unbiased

inconsistent/80.67

2.26/173.98

SHGM

unbiased

inconsistent/1033.1

3.86/193.1

unbiased

consistent/26.37

3.7/178.28

Real 3%, pseudo 20%

Real 3%, pseudo 50%

WLS

unbiased

consistent/23.85

1.89/198.4

unbiased

consistent/22.97

1.81/178.02

WLAV

unbiased

inconsistent/130.91

1.68/190.53

unbiased

inconsistent/72.81

1.45/173.07

SHGM

unbiased

inconsistent/953.8

1.86/192.37

unbiased

consistent/24.1

1.80/177.63

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

673

& The Institution of Engineering and Technology 2009

www.ietdl.org

Figure 10 Ninety-ve-bus system consistency plot with

limited redundancy: error in true measurements 1%,
error in pseudo measurements 50%

pseudo measurements. The value of E[e ] in the case of

SHGM does not converge to the number of states (i.e.
189), which numerically veries its inconsistency.
It is also important to note that with limited redundancy the
trace qualities dened in (6) are close for both WLS and
SHGM in Case 2 and Case 4. This gives the impression
that SHGM should be consistent for these cases. Since trace
captures the diagonal information of the error covariance
matrix, it can be attributed that inconsistency in SHGM is
mainly due to off-diagonal elements. In case of increased
redundancy there is signicant difference in the qualities of
WLS and SHGM in all the test cases. The quality of WLS
is better than the quality of SHGM.
In all the simulated cases only WLS satises the three
statistical criteria (bias, consistency and quality) under the
assumption of normal distribution of measurement errors.
It can be concluded that the WLS is a suitable solver for
the DSSE problem.

4.5 Comments on error distribution and

choice of solver
Figure 9 Ninety-ve-bus system consistency plot with
limited redundancy: WLS shows consistency in all test cases
A very high degree of inconsistency was observed in WLAV,
which is difcult to show graphically.
The performance summaries for both cases are shown in
Tables 4(a) and (b). In both the cases, the quality dened
in (5) gives numerical instability in computations, hence it
does not appear in the tables. Furthermore, the quality for
WLAV estimator is inconsistent and shows negative values.
This is because of very high variance of state estimates that
are unacceptable for SE. In WLS and SHGM, as expected
the qualities decrease with increase in errors in real and
674
& The Institution of Engineering and Technology 2009

The statistical criteria discussed in this paper depend on the

characteristics of the distribution of measurement errors. The
results presented are based on the assumption that the
measurement errors are normally distributed. Under this
assumption, the WLS satises the statistical criteria and
hence was found to be the suitable solver for the SE.
However, this may not be true if the measurement errors
are not normally distributed. For instance if the errors
follow the Laplace distribution [19], the WLAV estimator
gives better performance than WLS and SHGM. The
reason for this is that the WLAV is consistent with the
Laplace distribution and maximisation of log-likelihood of
the Laplace density function results in the WLAV
formulation. Hence, depending on the distribution of the
errors, the corresponding statistical criterion discussed in
Subsection 2.2 can be modied in order to identify the
consistent solver for that distribution.
IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

www.ietdl.org

Figure 11 Ninety-ve-bus system consistency plot with

increased redundancy: WLS shows consistency in all the
test cases

Figure 12 Ninety-ve-bus system consistency plot with

increased redundancy: SHGM shows inconsistency in all
the test cases

In reality, different probabilistic load distributions exist in the

distribution networks and no standard distribution can t all of
them. Furthermore, the large size of the distribution network
having various probability distributions at different buses
makes accommodating them in a single state estimator

impractical. A more practical approach is to model the actual

probability distributions as a mixture of several Gaussian
distributions (Fig. 13) and apply the WLS state estimator
which is consistent with the normal distribution. This
requires the modelling of the distribution of errors through

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

675

& The Institution of Engineering and Technology 2009

www.ietdl.org
Table 4 Ninety-ve-bus UK-GDS performance summary
Estimator

Real 1%, pseudo 20%

Bias

Real 1%, pseudo 50%

Consistent/E[e ]

Quality tr/det

Bias

Consistent/E[e ]

Quality tr/det

6.63/

unbiased

consistent/188.16

6.24/

252/

unbiased

inconsistent/1

(a) Limited redundancy

WLS

unbiased

consistent/190.02

WLAV

unbiased

inconsistent/1

SHGM

unbiased

inconsistent/3.06  10

6.46/

unbiased

Real 3%, pseudo 20%

WLS

unbiased

consistent/189.84

WLAV

unbiased

Inconsistent/1

SHGM

unbiased

Inconsistent/2.89  10

244.42/
5

inconsistent/2.53  10

6.16/

Real 3%, pseudo 50%

4.61/

unbiased

consistent/190.23

245.18/

unbiased

inconsistent/1

4.75/ 2

unbiased

4.41/
241.12/
5

inConsistent/2.27  10

4.4/

(b) Increased redundancy

Real 1%, pseudo 20%
WLS

unbiased

consistent/190

WLAV

unbiased

inconsistent/1

SHGM

unbiased

inconsistent/1.65  10

Real 1%, pseudo 50%

8.86/

unbiased

consistent/188.3

255.65/

unbiased

inconsistent/1

6.70/

unbiased

Real 3%, pseudo 20%

WLS

unbiased

consistent/188.65

WLAV

unbiased

inconsistent/1

SHGM

unbiased

inconsistent/6.58  10

Mc
X
k1

wk N (mk , s2k )(x) and

Mc
X

wk 1

inconsistent/1.82  10

6.85/

unbiased

consistent/189.23

243.73/

unbiased

inconsistent/1

4.32/

(20)

k1

unbiased

6.35/

Figure 13 Gaussian mixture approximation of the density

inconsistent/1.0  10

6.75/
249.88/
10

4.28/

The expectation maximisation algorithm [2022] is used to

obtain the parameters (wk , mk , s2k ) of the GMM.
In transmission systems, all the estimators work well
because of very high redundancy and thus the statistical
measures to evaluate the performance are not required. For
example, a highly erroneous measurement is treated as a
bad data by the WLAV estimator and a redundant
measurement is always available to replace this. But in
distribution systems, the measurements are mainly the
pseudo measurements with very limited redundancy. Since
pseudo measurements are derived from the historical load
proles and customer behaviour, they are highly erroneous.
This is why the statistical framework is required to identify
the suitable solver for the DSSE.

676
& The Institution of Engineering and Technology 2009

263.74/
9

Real 3%, pseudo 50%

Gaussian mixture model (GMM) [2022]. As shown in

Fig. 13, the GMM represents an arbitrary distribution as a
weighted combination of several Gaussian components.
Mathematically, a GMM having Mc mixture components
with mean and variance of kth component as mk and s2k can
be written as

f (x)

8.75/

Conclusion

The performance evaluation of SE techniques shows that

the existing solution methodology of WLAV and SHGM
cannot be applied to the distribution systems. In order to
obtain the consistent and good quality estimate, signicant
modications are required in these algorithms. WLS gives
consistent and better quality performance when applied to
distribution systems. Hence, WLS is found to be a suitable
solver for the DSSE problem.
IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

www.ietdl.org
The WLS works well if the noise characteristics are
known. In the absence of this knowledge either the WLS
needs to be modied or a new class of algorithms need to
be introduced. Furthermore, with growing interest in the
distribution automation, new DSSE techniques are
expected to be introduced in the future. However, any
modication in existing techniques or introduction of new
algorithms should qualify some statistical criteria because of
limited number of measurements. This paper highlights
some important statistical criteria against which a SE
algorithm should be tested to assess its suitability to DSSE.

[11] BLACKMAN S.S. : Multiple target tracking with radar

applications (Canton St. Norwood: Artech House, Inc.,
1986)
[12] ANDERSON T.W.: An introduction to multivariate
statistical analysis (Wiely, Seventh printing, 1966)
[13] JABR R.A.: Primal dual interior point approach to compute
the L1 solution of the state estimation problem, IEE Proc.Gener. Transm. Distrib., 2005, 152, (3), pp. 313320
[14]

Acknowledgment

The authors would like to thank Peter D. Lang of EDF Energy

Networks for his valuable suggestions and discussions.

Robust statistics (Wiely, 1981)

[15] CELIK M.K., LIU W.-H.E. : An incremental measurement

placement algorithm for state estimation, IEEE Trans.
Power Syst., 1995, 10, (3), pp. 1698 1703
[16] DAS D., NAGI H.S., KOTHARI D.P.: Novel method for solving
radial distribution networks, IEE Proc.-Gener. Transm. and
Distrib., 1994, 141, (4), pp. 291 298

References

[1] ABUR A., EXPOSITO A.G.: Power system state estimation:

theory and implementation (Marcel Dekker, Inc., 2004)
[2] SHAFIU A., JENKINS T.V. ,
networks, CIRED, 2005

HUBER P.J.:

STRBAC G. :

Control of active

[3] LU C.N., TENG J.H., LIU W.-H.E.: Distribution system state

estimation, IEEE Trans. Power Syst., 1995, 10, (1),
pp. 229 240
[4] LIN W.-M., TENG J.-H.: Distribution fast decoupled state
estimation by measurement pairing, IEE Proc.-Gener.
Transm. Distrib., 1996, 143, (1), pp. 43 48
[5] BARAN M.E., KELLEY A.W.: A branch current based state
estimation method for distribution systems, IEEE Trans.
Power Syst., 1995, 10, (1), pp. 483 491
[6] WANG H., SCHULZ N.N.: A revised branch current based
distribution system state estimation algorithm and meter
placement impact, IEEE Trans. Power Syst., 2004, 19, (1),
pp. 207 213
[7] LI K.: State estimation for power distribution system
and measurement impacts, IEEE Trans. Power Syst., 1996,
11, (2), pp. 911 916
[8] BARAN M.E., ZHU J.X., KELLEY A.W.: Meter placement for real
time monitoring of distribution feeders, IEEE Trans. Power
Syst., 1996, 11, (1), pp. 332 337
[9] GHOSH A.K., LUBKEMAN D.L. , DOWNEY M.J., JONES R.H. :
Distribution circuit state estimation using a probabilistic
approach, IEEE Trans. Power Syst., 1997, 12, (1), pp. 45 51
[10] LUBKEMAN D.L., ZHANG J., GHOSH A.K., JONES R.H.: Field results
for a distribution circuit state estimator implementation,
IEEE Trans. Power Deliv., 2000, 15, (1), pp. 399 406
IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485

[17] United Kingdom Generic Distribution Network (UKGDS) [Online], available at: http://monaco.eee.strath.ac.
uk/ukgds/, accessed June 2008
[18] SINGH R., PAL B.C., VINTER R.B.: Measurement placement in
distribution system state estimation, IEEE Trans. Power
Syst., 2009, 24, (2), pp. 668 675
[19] KOTZ S. , KOZUBOWSKI T.J., PODGORSKI K.: The Laplace
distribution and generalizations (Birkhauser Boston, 2001)
[20] DEMPSTER A.P., LAIRD N.M., RUBIN D.B.: Maximum-likelihood
from incomplete data via the EM algorithm, J. R. Statist.
Soc. Ser. B, 1977, 39, (1), pp. 1 38
[21] REDNER R.A., WALKER H.F.: Mixture densities, maximum
likelihood and the EM algorithm, SIAM Rev., 1984, 26,
(2), pp. 195 239
[22] BILMES J.A.: A gentle tutorial on the EM algorithm and its
application to parameter estimation for Gaussian mixture
and hidden Markov models. Technical Report, International
Computer Science Institute, ICSI-TR-97-021, 1998

Appendix

8.1 Computation of a for various

estimators
The fact that normalised measurement residual r is normally
distributed with zero mean and unit variance can be used to
compute a for the state estimators discussed in Section 3.

8.1.1 Weighted least-squares:

c(r) r, c0 (r) 1

(21)
677

& The Institution of Engineering and Technology 2009

www.ietdl.org
1
E[c2 (r)] p
2p

r 2 e(1=2)r dr Var(r) 1

(22)

1

1
E[c0 (r)] p
2p

e(1=2)r dr 1

(23)

1

E[c2 (r)]
(E[c0 (r)])2

By symmetry of the distribution the above equation can be

expressed as
2
p
2p

av

2 (1=2)r 2

(av sgn(r)) e
1

c(r) sgn(r), c (r) 2d(r)

2
1 1
2
(sgn(r))2 e(1=2)r dr 1
E[c (r)] p
2p 1

(26)

d(t  t0 )f (t)dt f (t0 )

(27)

2
(2a2 v2 F(av) p
2p

(1=2)r 2

2d(r)e
1

E[c2 (r)]
0

(E[c (r)])

r
2
dr
p

p
2

(28)
(29)

c(r)

r
av sgn(r)

if

jrj  av
otherwise

678
& The Institution of Engineering and Technology 2009

r 2 e(1=2)r dr)

(34)

2
p
2p

av

2 (1=2)r 2

r e
0

r
2 2
2
dr  ave(a v =2) (2F(av)  1)
p
(35)

Using the relation F(av) 1  F(av) and substituting

(35) in to (34), we obtain
r
2 2
2
ave(a v =2) 2(a2 v2  1)(1  F(av))
E[c (r)] 1 
p
(36)

(30)

1
if jrj  av
c (r)
2avd(r) 0
otherwise
av
2
1
(av sgn(r))2 e(1=2)r dr
E[c2 (r)] p
2p 1

1 av 2 (1=2)r 2
r e
dr
p
2p av

2
1 1
(av sgn(r))2 e(1=2)r dr
p
2p av
0

av

8.1.3 Schweppe Huber generalised M:



where F is the cumulative probability function. The integral

term in the above equation is given by

in the following expression

1
E[c (r)] p
2p

r 2 e(1=2)r dr
(33)

(25)

We use the fact that

av

(24)

8.1.2 Weighted least absolute value estimator:

1

2
dr p
2p

(31)

1
E[c0 (r)] p
2p

av

e(1=2)r dr 2F(av)  1

(37)

av

In this case a depends on parameters a and v, that is, if

a 1.5 and v 1, the value of a is
(32)

E[c2 (r)]
(E[c0 (r)])2

0:7785
(0:8664)2

1:0371

(38)

IET Gener. Transm. Distrib., 2009, Vol. 3, Iss. 7, pp. 666 678
doi: 10.1049/iet-gtd.2008.0485