PSA Chapter 07

DRAFT and INCOMPLETE
Table of Contents
from
A. P. Sakis Meliopoulos
Power System Modeling, Analysis and Control

Chapter 7 _____________________________________________________________ 3
Operating State Estimation _______________________________________________ 3
7.1 Introduction____________________________________________________________ 3
7.2 SCADA System _________________________________________________________ 4
7.3 System Network Configurator _____________________________________________ 7
7.4 State Estimation ________________________________________________________ 8
7.4.1 Least Squares Solution _______________________________________________________ 13
7.4.2 Least Absolute Deviation Solution (Ll Approach) __________________________________ 18
7.4.3 Chebyshev or Min-Max Solution (L Approach) __________________________________ 21
7.4.4 Summary of the Three State Estimation Approaches ________________________________ 25
7.4.5 Quality of the State Estimate __________________________________________________ 26
7.4.6 Summary and Discussion _____________________________________________________ 34
7.5 Detection and Identification of Bad Data____________________________________ 35

7.5.1 Detection of Bad Data________________________________________________________ 35
7.5.2 Identification of Bad Data_____________________________________________________ 36
7.5.3 Summary and Discussion _____________________________________________________ 47
7.6 Sequential State Estimators ______________________________________________ 48

7.7 State Estimator Observability_____________________________________________ 55
7.7.1 Algebraic Observability ______________________________________________________ 55
7.7.2 Topological Observability ____________________________________________________ 56
7.7.3 Critical Measurements _______________________________________________________ 59
7.8 Limitations and Biases of State Estimation __________________________________ 62

7.8.1 Bias From Unbalanced Operation_______________________________________________ 63
7.8.2 Bias From System Asymmetry _________________________________________________ 64
7.8.3 Bias From Systematic Measurement Errors _______________________________________ 67
7.8.4 Bias from Measurement Time Skews ____________________________________________ 69
7.9 Synchronized Measurements _____________________________________________ 70

7.10 Formulation of the Three-Phase State Estimation___________________________ 72
7.10.1 Three-Phase System State ___________________________________________________ 72
7.10.2 Three-Phase Measurements __________________________________________________ 73
7.10.3 Least Squares Estimation____________________________________________________ 73
Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos
7.10.4 Three-Phase Power System Model ____________________________________________ 74

7.10.5 Observability Analysis Three Phase State Estimation _____________________________ 75
7.10.6 Quality of Three-Phase State Estimator_________________________________________ 77
7.10.7 Discussion of the Three-Phase State Estimator ___________________________________ 77
7.11 Hybrid Three-Phase State Estimator ______________________________________ 77

7.12 Summary and Discussion _______________________________________________ 78
7.13 Problems ____________________________________________________________ 79
Page 2
Copyright A. P. Sakis Meliopoulos 1990-2006
Chapter 7
Operating State Estimation
7.1 Introduction
Effective control and operation of electric power systems requires accurate and reliable
knowledge of the system model and the operating state of the system in real time. For
this purpose, modern power systems are equipped with an extensive data acquisition
system. Local analog and status quantities, such as voltage magnitude, real and reactive
power flows, loads, status of breakers (open/close) etc., are measured and transmitted to a
central location. The measurements are simple, requiring simple instrumentation.
Typical analog measurements are: (a) voltage magnitudes, (b) real and reactive power
flows and (c) current magnitude measurements. Recent technology based on GPS
(Global Positioning System) has made it possible to measure voltage phase angles as
well. Typical status measurements are: (a) breaker status, (b) disconnect switch status,
etc. All measurements are taken every one to several seconds. They are transmitted to a
central location (the Energy Management System (EMS) or the Energy Control Center)
where they are processed to yield the operating state of the system. The process consists
of two analysis problems: (a) determination of network topology, and (b) determination
of operating state. The network topology is constructed from the status of breakers and
disconnect switches. The operating state of the system is constructed from analog
measurements by means of two distinct computational procedures: (a) on-line power flow
and (b) state estimation.
On-line power flows utilize a subset of available measurements which are enough to
define the power flow problem. Solution of the power flow problem yields the operating
state of the system. Because measurements are usually corrupted with error (resulting
from PT or CT inaccuracies, instrument error, transmission error, etc.), the error is
directly transmitted to the computed operating state. It is also possible that one of the
measurements may include a large error (gross error - due to meter malfunctioning or
communication errors) resulting in a non-solvable power flow problem or in a solution
that may be quite different than the actual operating condition of the system. In a real
time environment, it is important to have the ability to identify wrong measurements or
errors in measurements (bad data). For this purpose, it is necessary to take advantage of
redundant measurements. The redundant measurements are utilized to compute the best
estimate of the operating state of the system with statistical methods. The computational
procedure is called state estimation. As we shall discuss, state estimation provides the
mechanism to: (a) determine whether the system state can be computed from existing
data (observability); (b) filter out usual measurement errors and, therefore, compute the
system state with minimum error; (c) identify and reject bad data; and (d) determine the
degree of confidence on the estimated state of the system. The conceptual view of the
Page 3
process is illustrated in Figure 7.1. The two approaches, i.e. on-line power flow and state
estimation, are identified. In addition other applications are possible using the collected
data. These applications are parameter estimation and remote calibration of the meters.
The constituent parts of this process will be discussed next.
Bad Data
Identification
& Rejection
Observability
Analysis
Status
Data
Topology
Processor
Analog
Data
Gross Bad Data

Detection
by Limit Checks
+
Consistency
Checks
On-Line
Power
Flow
Solution
State
Estimation
Best Estimate of
System State
& Model
System State
Parameter
Estimation
Remote
Calibration
Figure 7.1 Conceptual View of Real Time Power System Modeling and
State Estimation
The objective of this chapter is to introduce the models involved in the state estimation
problem and to discuss its various extensions and applications. Because normally there
are redundant measurements, it is possible to use the measurements for the purpose of
determining or improving the system model parameters or to remotely calibrate the
meters.
7.2 SCADA System

The necessary hardware to enable real time modeling of a power system is collectively
referred to as SCADA system. SCADA stands for Supervisory Control And Data
Acquisition system. The supervisory control subsystem consists of hardware and
software which (a) collect status data (i.e. breaker status open/close) and analog data (i.e.
measurements of voltage magnitude, power etc.) and transmit these data to a central
location for processing and display (b) allow remote tripping of breakers, changes of
transformer tap, etc. In most cases, supervisory control is a manual function, i.e., the
dispatcher at the control center will initiate a command to open/close a breaker, etc. The
data-acquisition subsystem consists of remote terminal equipment for interfacing with
power system instrumentation and control devices; interfaces with communication
channels; and equipment for interfacing with the system control center.
Since the SCADA system transmits data from the field to a central location and vice
versa, communication media, protocols and communication speeds are very important. In
Page 4
the past it was customary to have separate communication channels for the Supervisory
Control and Data Acquisition. Today, however it is a unified system sharing a common
two way communication channels that may consist of several physical layers.
Communications are integrated in the remote terminal unit (RTU) which manages data
collection, control functions and communication with a master station. The master
station has multiple communication channels to remote terminal units. Many times a
dedicated channel is assigned to each remote terminal unit. In other cases, there are less
channels than remote terminal units requiring more than one remote to share a channel.
Analog data is scanned periodically, typically every one second to a few seconds. Each
scan is triggered by the system control center at the prescribed interval by using a request
to all remote stations to send in data.
The amount of data collected and transmitted is very large for typical power systems.
These data is transported via communication channels. In order to minimize
communication traffic some form of data compression is utilized. For example, for status
data one can send only changes of status data. This approach minimizes amount of data
transfer and the amount of processing needed at the master station. Analog data can be
also compressed with a number of methods.
Independently of system configuration, SCADA system manufacturer, communication
software and computer configuration, the end result of the SCADA system function will
be the collection of a set of system data every sampling period. The data consists of:
Breaker status
Disconnect switch status
Transformer tap setting
MW flow measurements
MVAR flow measurements
Voltage magnitude (kV) measurements
Current magnitude (kA) measurements
phase angle difference measurements
etc.
A simplified view of a SCADA system is illustrated in Figures 7.2a and 7.2b.
Page 5
G1
G2
MW Flow Measurement
MVAR Flow Measurement
kV Measurement
Disconnect Switch Status
Breaker Status
RTU
Communication
Link with Control
Center
Figure 7.2a Simplified View of a SCADA System - Survey Points

Contact Inputs
Analog Inputs
Contact Outputs
Analog Outputs
Data
RTU
Commands
Master
Station
Figure 7.2b Simplified View of a SCADA System - Configuration

At the central location (energy management system or control center), the data are
managed with the data acquisition software. These software perform the following tasks:
(a) initiate the collection of data and place them in computer memory, (b) gross errorchecking, (c) conversion to engineering units, (d) limit-checking, and (e) generation of a
data base which is interfaced with application programs.
Page 6
The data are utilized to form the system model (Network Configurator) and to estimate
the system operating state (state estimation). The next sections describe these
applications.
7.3 System Network Configurator

Data collected with the SCADA system are utilized in two ways. Status data (circuit
breaker status, interrupt switch status, transformer tap setting, etc.) are utilized to form
the system network configuration and model. The software which take the status data
and computes the system network configuration and model is known as system network
configurator. A typical task performed by the system network configurator is illustrated
in Figures 7.3a and 7.3b. The information received with the SCADA system determines
the status of the breakers. The system network configurator uses prestored information
and the breaker/switch status to determine a bus oriented model, i.e. which circuits are
connected to which bus and what is the model of each circuit. This task is illustrated in
Figure 7.3. Typically, this procedure is executed only when a change in status data
occurs.
The system network configuration and model is next combined with the analog data for
the purpose of determining the operating conditions of the system. Typically, there are
redundant measurements which are used to obtain the best (in some sense to be discussed
later) estimate of the operating state of the system. The computational procedure which
performs this task is known as the state estimator and will be discussed in the next
paragraph.
AutoBank
500kV/230kV
G1
AutoBank
500kV/230kV
G2
Figure 7.3a Breaker Oriented Model Prestored Network Data
Page 7
SG1
SG2
Figure 7.3b Bus Oriented Network Model of the System of Figure 7.3a
7.4 State Estimation

State estimation is a computational procedure which uses a redundant set of
measurements and a bus oriented network model to compute a statistical estimate of the
system operating state. In a power system, the operating conditions are uniquely defined
by the set of variables consisting of all bus voltage magnitudes and angles except the
phase angle at an arbitrarily selected bus which is set equal to zero. We refer to this set
of variables as the state of the system. As an example, the operating condition of the
system of Figure 7.4 is defined with three state variables: V1, V2, and 2. Knowledge of
these three variables is sufficient to determine other quantities of interest, for example,
P12, Q12. For this example, assume that a set of redundant measurements is taken as it is
shown in Figure 7.4. (five measurements: V1, V2, P12, P21, and Q21) A subset of these
measurements is enough to provide the state variables. For example assume that the
following three measurements are selected: V1, P12, and V2, where P12 is the real power
flow on circuit 1-2. From these three measurements, the state variables V1, V2, and 2
can be computed.
-j15.0
Voltage Magnitude Measurement

MW Flow Measurement
MVAr Flow Measurement
Figure 7.4 A Simplified Example System for Power System State

Estimation
In general, a measurable quantity such as P21, Q21, P12, Q12, V1, etc., can be expressed as
a function of the system state. Let zi denote a measured quantity. Then:
Page 8
zi = hi (x )
(7.1)
where x is the system state and hi is a function specific to the measured quantity zi.
Assume that m measurements are taken. Then:
z = h( x )
(7.2)
where
x is the system state - an n x 1 vector
z is a vector of measured quantities - an m x 1 vector
h is a vector function - an m x 1 vector function.
Typically more measurements are taken than the number of state variables to be
determined, i.e. m>n. In this case, the set of Equations (7.2) represents an
overdetermined set of nonlinear equations in real variables.
In general an overdetermined set of equations, such as (7.2) does not have a solution for
x. Only if the system model expressed with the functions h(x) is exact and the
measurements z are extremely accurate, then equations (7.2) have a unique solution for
x. This of course is unlikely in a real system. Yet it is possible to obtain a solution for x
via a procedure known as the state estimation. This procedure will be first introduced by
an example and then it will be discussed in more detail.
Example E7.1: Consider the simplified power system of Figure E7.1. The following
measurements are taken (all in p.u.):
z1 = V1 = 1.0,
z3 = P21 = -1.56,
z2 = V2 = 0.99
z4 = Q21 = 0.01
Formulate the state estimation problem as an overdetermined set of nonlinear equations.
~
V1 =V1e j0
-j15.0
~
V2 = V2e j2

MW Flow Measurement
Figure E7.1
Solution: The state vector x is:
Page 9
V1
x = V2

2
The vector function h(x) is given as:
h1 ( x ) = V1
h2 ( x) = V2
h3 ( x ) = 15.0V1V2 sin 2
h4 ( x ) = 15.0V22 15.0V1V2 cos 2
The state estimation problem is formulated as the following overdetermined set of

equations:
1.0 = V1
0.99 = V 2
1.56 = 15 .0V1V 2 sin 2
0.01 = 15.0V22 15.0V1V2 cos 2

In summary, the state estimation problem is formulated as an over determined set of
nonlinear equations. Specifically, the problem is posed as follows:
Given
b = h(x)
(7.3)
where:
b
h(x)
x
m>n
is a known m x 1 vector
is an m vector of known function
is the unknown n x 1 vector
Compute the vector x.

This problem is known as the nonlinear estimation problem. In case that the functions h
are linear, then the problem collapses to the linear estimation problem defined as follows:
Given
b=Hx
(7.4)
where:
b
H
x
m > n.
Page 10
is a known m x 1 vector
is a known m x n matrix
is the unknown n x 1 vector
Compute the vector x.

In general, the overdetermined set of equations (7.3) or (7.4) does not have a solution, i.e.
a vector x does not exist which satisfies all equation (7.3) or (7.4). In this case, it is
expedient to define the residual vector:
or
r = h ( x ) b (nonlinear case)
(7.5)
r = Hx b (linear case)
(7.6)
and compute a solution x which will minimize the residual vector in some sense. For this
problem, there are three approaches: (a) the least squares solution, (b) the least absolute
deviation solution(L1 approach), and (c) the Chebyshev or min-max solution(L
approach). These methods will be briefly discussed next.
An Alternative Introduction of the State Estimation Problem: An alternative way to
introduce the state estimation problem is by considering the hardware used for obtaining
the measurements. Any metering device comprises an instrumentation channel, a
transducer and an A/D converter (in recent systems the transducers are omitted since
they are not necessary). The intrumentation channel comprises instrument transformers
(Potential Transformers (PT), Current Transformers (CT), Optical Transformers, etc.),
interconnecting cables and possibly attenuators. The transducer may be an analog
device that converts the input signal into a DC output signal proportional to the quantity
measured. The A/D converter samples the signal and convertes it into a ditital form. The
signal can be the output of the transducer or it can be the output of the instrumentation
channel. In the latter case, an intelligent device is required to extract the appropriate
information, i.e. rms value of the voltage waveform, real power flow, etc. A pictorial
view of this arrangement is shown in Figures 7.5a and 7.5b.
PT
Control Cable
Transducer
Output
0-10 V DC
Figure 7.5a Physical System for Analog Measurements - Use of

Transducers
Page 11
PT
Control Cable
A/D
Conversion
Digital
Interface
Figure 7.5b Physical System for Analog Measurements - Transducer-less

Technology
Measurements obtained with the system of Figure 7.5 will contain a measurement error.
The statistics of this error can be quantified from the parameters of the instrumentation
channel, the transducer and the A/D converter. This process is complex and for typical
power system instrumentation channels [???] the statistics may be biased. As a
convenient simplification, we assume that the statistics of the measurement error are
unbiased, they obey a gaussian distribution with a known standard deviation, the
expected mean value is zero and the error of a measurement is uncorrelated to the error of
any other measurement. Specifically, if the error of measurement i is represented with ,
then
E{ i } = 0
E{ i2 }= i2
E{ i j } = 0, i j
(7.7)
(7.8)
(7.9)
Consider a specific measurement bi. This measurement is related to the state of the
system via a known function hi(x). We have named this function the model of the
measurment. We postulate that the difference between the model and the measurement
is the rror of the measurement, i.e.:
i = hi ( x ) bi
(7.10)
It should be recognized that the measurement error is the same as the measurement
residual we have introduced already. The difference is that now we can associate
statistics this residual. In subsequent paragraphs we shall use equations (7.5) and (7.10)
interchangeably and will assume that the residuals r or the errors h have the statistics
stated with equations (7.7), (7.8) and (7.9).
Page 12
7.4.1 Least Squares Solution

The least squares solution of the overdetermined system (7.5) or (7.10) is the vector x
which minimizes the sum of the squares of the components of the residual vector r or the
vector of the measurement errors . Mathematically, this is expressed as follows:
m
Minimize
J = ri 2 = r T r = T
(7.11)
i =1
A variation of this method is the weighted least squares method which minimizes the sum
of the weighted squares of the components of the residual vector r or the vector of the
measurement errors . Mathematically, this is expressed as follows:
m
Minimize
J = wi ri 2 = r T Wr = T W
(7.12)
i =1
where:
wi : the weight for the residual ri
W : a diagonal matrix, the diagonal elements being the weights wi.
A most usual case of weighted least squares is defined as follows. We postulate that we
want to compute the state of the system that minimizes the sum of the squares of the
normalized measurement errors, si. The normalized measurment error is defined with:
si =
i
i
(7.13)
In this case, the problem is stated as follows:

h ( x) zi
J = i
i
i =1
m
Minimize
= s i2 = T W
i =1
(7.14)
where:
1 1
1
W = diag 2 , 2 , ..., 2
m
1 2
Note that this is a weighted least squares formulation with the weights defined as the
inverse of the squared standard deviations. It should be also noted that the formulation in
terms of equation (7.11) is equivalent to assuming that all standard deviations of the error
of all measurements are equal.
In subsequent paragraphs we will consider the weighted least squares approach.
Page 13
Using the nonlinear and linear model equations, the nonlinear and linear state estimation
problem is expressed as follows:
Minimize J = (h( x ) b) T W (h( x ) b) , for the nonlinear case
(7.15)
Minimize J = ( Hx b) T W ( Hx b) , for the linear case
(7.16)
The unknown vector x is obtained from the solution of the necessary conditions, which in
matrix notation are expressed as follows:
dJ
=0
dx
(7.17)
The above problem is fisrt solved for the linear case and then for the nonlinear case.
Linear Case: Direct differentiation of equation (7.16) with respect to x, we obtain:
dJ
d
= [( Hx b) T W ( Hx b)] = 2 H T W ( Hx b) = 0
dx dx
(7.18)
Upon solution of last equation for the state vector x:
x = H T WH
H T Wb
(7.19)
Equation (7.19) provides the solution to the linear estimation problem (7.16).
Nonlinear Case: To obtain the solution to the nonlinear estimation problem (7.15),
assume that an initial guess of the vector x0 is known. The nonlinear model equations
(7.5) are linearized around the point x0 yielding:
r = = h( x 0 ) +
h( x )
|
( x x 0 ) + h.o.t. b
x x = x
0
Where h.o.t. denotes higher order terms. Assuming that the vector x0 is very close to the
solution, then the higher order terms (h.o.t.) are negligibly small and are omitted from
above equation, yielding:
r = =
Let
h( x )
| x = x ( x x 0 ) + h( x 0 ) b
x
0
(7.20)
h( x )
| x = x = H , and b ' = h( x 0 ) + Hx 0 + b
x
0
Page 14
Observe that the vector b ' = h( x 0 ) + Hx 0 + b is known (or computable). Now equation
(7.20) becomes:
r = = Hx b '
Now the problem is identical to the linear estimation problem. Thus, the solution is:
x = H T WH
H T Wb '
Upon substitution of the b vector:
x = H T WH
H T W ( Hx 0 h( x 0 ) + b) = x 0 H T WH
H T W ( h( x 0 ) b)
The last equation is generalized into the following iterative equation:
x +1 = x ( H T WH ) 1 H T W (h( x ) b)
(7.21)
where H is the matrix h ( x )/ x computed at x = x . This is the Jacobian of the vector

function h(x).
In summary, the least squares solution of the linear estimation problem is given by
Equation (7.19) and the least squares solution of the nonlinear estimation problem can be
obtained with the iterative algorithm (7.21). An example will illustrate the method.
Example E7.2: Consider the overdetermined set of equations derived in Example E7.1.
x T = [V1 V2 2 ] , using as an initial guess
Compute the solution
x 0 = [1.0 0.99 0.0] and least squares estimation. Assume that all weights are equal to
1.0.
T
Solution: The problem will be solved with the iterative algorithm (7.21). Since the
weights are all 1.0 the weight matrix W is the identity matrix. The objective function is:
J = (V1 1.0) + (V2 0.99) + (15V1V2 sin 2 + 1.56) + (15V22 15V1V2 cos 2 0.01)
2
Min J = wi (hi ( x ) z i )
i =1
Page 15
The Jacobian matrix H is:

1
0
H =
15V 2 sin 2
15V 2 cos 2
0
1
15V1 sin 2
30V 2 15V1 cos 2
15V1V 2 cos 2
15V1V 2 sin
0
0
The computations follow:

1st Iteration
0.0
0.0
0
0
r = h( x ) b =
1.56
0.1585
0.0
0.0
1.0
0.0
1.0
0.0
H (x0 ) =
0.0 14.85
0.0
14.85 14.7 0.0
0.0
221.5225 218.295
H H = 218.295
217.09
0.0
0.0
220.5225
0.0
T
2.3537
H r = 2.33
23.166
T
0.00538
( H H ) H r = 0.00532
0.10505
T
1.0 0.00538 0.9946

x = 0.99 0.00532 = 0.9953

0.0 0.10505 0.10505
1
2nd Iteration
Page 16
0.0054
0.0053
r 1 = h( x 1 ) b =
0.0030
0.0823
0.0
0.0
1.0
0.0
1.0
0.0
1
H (x ) =
1.5655 1.5644 14.767
14.8472 15.0222 1.557
220.589 0.000647
223.89
H H = 220.589
229.114
46.4911
220.489
0.000647 46.4911
T
1.2320
H r = 1.2369
0.0838
T
(H H )
0.4960 0.4988 0.1052

= 0.4988 0.5063 0.1068
0.1052 0.1068 0.0270
0.00281
( H T H ) 1 H T r 1 = 0.00273
0.000196
0.9974
x = 0.9926
0.1052
2
This completes the solution. It is expedient to compute the residuals (or measurement
errors) using the above computed state:
r1 = V1 1.0 = 0.0026
r2 = V 2 0.99 = 0.0026
r3 = 15.0V1V2 sin 2 + 1.56 = 0.000629
r4 = 15.0V22 15.0V1V2 cos 2 0.01 = 0.000631
Page 17
Note that in this case the residuals or measurement errors are very small.
7.4.2 Least Absolute Deviation Solution (Ll Approach)

The least absolute deviation solution of the overdetermined system (7.5) and (7.6) is the
vector x which minimizes the sum of the absolute deviations of the components of the
residual vector r. Mathematically, this is expressed as follows:
m
Minimize J = | ri |
i =1
A variation of this method is the weighted least absolute deviation method which
minimizes the sum of the weighted sum of the absolute deviations of the components of
the residual vector r. Mathematically, this is expressed as follows:
m
Minimize J = | wi ri |
i =1
The weights can be selected as discussed in the previous method. Again we will consider
the solution of the weighted least absolute deviation approach.
This requirement translates into the following optimization problem:
m
i =1
Subject to: ri = hi ( x ) bi ,
i = 1, 2,...., m
for the nonlinear problem, or:

m
i =1
m
Subject to: ri = hij x j bi ,
i = 1, 2,...., m
j =1
for the linear estimation problem.

Both of the above problems are easily translated into a linear programming problem. The
procedure will be demonstrated for the nonlinear case. For this purpose, the nonlinear
equations are linearized around a given point to yield:
r = h( x ) b H ( x x 0 ) + h( x 0 ) b = Hx b '
Page 18
where:
x = x x 0
b ' = b h( x 0 )
Now the problem becomes

m
i =1
m
Subject to: ri = hij x j bi' ,
i = 1, 2,......, m
j =1
The above problem is transformed into an optimization problem of the linear

programming variety, by replacing the variables r and x with a pair of nonnegative
variables:
r = r+ r,
x = x + x ,
r+,r 0
x+, x 0
The linear programming problem is:

m
Minimize
Subject to:
J = w i (ri+ + ri )
i =1
r r H ( x + x ) = b '
r+ ,r, x+ , x 0
An example will illustrate the method.

Example E7.3: Consider the problem of Example E7.2. Again assume the weights to be
equal to 1.0. Solve this problem using the least absolute deviation method.
Solution: Assuming as a starting solution the point x 0 = [1.0 0.99 0.0] , and
linearizing around this point, the L1 problem is stated as follows:
T
Minimize
Subject to
J = (ri+ + ri )
+
1
+
2
+
3
i =1
r r V1+ + V1 = 0.0
r r V2+ + V2 = 0.0
r r 14.85 2+ + 14.85 2 = 1.56
r4+ r4 + 14.85V1+ 14.85V1 14.7V2+ + 14.7V2 = 0.1585

All variables are non-negative. Note that since the voltage magnitudes are non-negative
Page 19
by definition, there was no need to introduce + and variables for the voltage
magnitudes.
Upon solution of the above linear program we obtain the followings:
V1 = 0.01067
2 = 0.10505
r1 = 0.01067
and all other rs equal to zero
Thus, the new state variables and residuals are:
V1 = 0.98933
V 2 = 0.99
2 = 0.10505
r1 = r1+ r1 = 0.01067
r2 = r2+ r2 = 0.0
r3 = r3+ r3 = 0.0
r4 = r4+ r4 = 0.0
Note that only one residual is nonzero. In general, the LAV solution will result in m-n
maximum number of non-zero residuals. In this case m=4 and n=3, thus 4-3=1 residuals
will be nonzero.
On the other hand, using the nonlinear model equations and the above computed state,
the residuals are computed to be:
r1 = V1 1.0 = 0.01067
r2 = V2 0.99 = 0.0
r3 = 15.0V1V2 sin 2 + 1.56 = 0.01954

r4 = 15.0V22 15.0V1V2 cos 2 0.01 = 0.08138
2nd Iteration
Linearization of the model around the new solution (from the first iteration) yields the
following linear program:
4
Minimize
J = (ri+ + ri )
i =1
Subject to
+
1
+
2
r r1 V1+ + V1 = 0.01067
r r2 V2+ + V2 = 0.0
Page 20
r3+ r3 + 1.5571V1+ 1.5571V1 + 1.556V2+ 1.556V2 14.6101 2+ + 14.6101 2 = 0.01954

r4+ r4 + 14.7681V1+ 14.7681V1 14.9423V2+ + 14.9423V2 + 1.5405 2+ 1.5405 2 = 0.08138
All the variables in above linear program are non-negative.
Upon solution of above problem, we obtain the following state and residuals:
V1 =0.00559
V2 = 0.0
2 = -0.00074
Vl = 0.9949
V2 = 0.99
2 = -0.1058
r1 = r1+ r1 = 0.005112
r2 = r2+ r2 = 0.0
r3 = r3+ r3 = 0.0
r4 = r4+ r4 = 0.0
Using the above computed state and the nonlinear model equations, the actual residuals
are computed to be:
r1 = V1 1.0 = 0.0051
r2 = V 2 0.99 = 0.0
r3 = 15.0V1V2 sin 2 + 1.56 = 0.0002

r4 = 15.0V22 15.0V1V2 cos 2 0.01 = 0.00015
This completes the solution. Note again the residuals (or measurement errors) are low in
this case.
7.4.3 Chebyshev or Min-Max Solution (L Approach)

The min-max solution of the overdetermined system (7.5) or (7.6) is the vector x that
minimizes the absolutely largest (maximum) component of the residual vector r.
Mathematically, this is expressed as follows:
Minimize
r * = max( r1 , r2 ,...., rn
where ri is the i-th component of the residual vector r. This requirement translates into
the following optimization problem:
Minimize
Subject to
J = r*
wi hi ( x ) bi r * , i = 1,2,..., m
Page 21
for the nonlinear estimation problem, or:

Minimize
r*
Subject to
wi
h
j =1
ij
x j bi r * , i = 1,2,..., m
for the linear estimation problem.

One way to solve this problem is by first transforming it into a linear program and
subsequently using the simplex algorithm to solve this problem. The procedure will be
demonstrated for the nonlinear program. Specifically, the stated problem is equivalent
to:
Minimize
Subject to
r*
wi (hi ( x ) bi ) r * , i = 1,2,..., m
wi (hi ( x ) bi ) r * , i = 1,2,..., m
Upon linearization of the residuals:
h( x ) b H ( x x 0 ) + h( x 0 ) b = Hx b '
where:
x = x x 0
b ' = b h( x 0 )
Upon substitution:
Minimize
Subject to
r*
W ( Hx R * ) b '
W ( Hx + R * ) b '
where,
r *
*
r
*
R =
M
*
r
an m x 1 vector.
Next, replace the variable x with nonnegative variables:
x = x + x ,
Page 22
x+, x 0
and introduce the slack (y) and surplus variables (z). Then,
Minimize
r*
Subject to
W ( Hx + Hx R * ) + y = b '
W ( Hx + Hx + R * ) z = b '
x+, x-, r*, y, z 0
Above problem is a standard linear problem. Upon solution, the state estimation problem
is solved. An example will demonstrate the method.
Example E7.4: Consider the problem stated in Example E7.2. Again assume the
weights to be equal to 1.0. Compute the solution x T = [V1 V2 2 ] , using the
Chebyshev method and an initial guess of x 0 = [1.0 0.99 0.0] .

T
Solution: The formulation of the problem is:
Minimize
r*
Subject to
V1 1.0 r *
V2 0.99 r *
15.0V1V2 sin 2 + 1.56 r *
15.0V22 15.0V1V2 cos 2 0.01 r *
Upon removal of absolute values, linearization of the resulting equations around the
operating point x T = [V1 V2 2 ] = [1.0 0.99 0.0] and introduction of slack and
surplus variables, the above problem is translated into:
r1 = V1 1.0 = 0.0
r2 = V 2 0.99 = 0.0
r3 = 15.0V1V2 sin 2 + 1.56 = 1.56

r4 = 15.0V22 15.0V1V2 cos 2 0.01 = 0.1585
Minimize
r*
Subject to:
V1+ V1 r * + y1 = 0.0
V1+ V1 + r * z1 = 0.0
Page 23
V2+ V2 r * + y 2 = 0.0
V2+ V2 + r * z 2 = 0.0
0.0V1+ 0.0V1 + 0.0V2+ 0.0V2 + 14.85 2+ 14.85 2 r * + y 3 = 1.56
0.0V1+ 0.0V1 + 0.0V2+ 0.0V2 + 14.85 2+ 14.85 2 + r * z 3 = 1.56
14.85V1+ + 14.85V1 + 14.7V 2+ 14.7V 2 + 0.0 2+ 0.0 2 r * + y 4 = 0.1585

14.85V1+ + 14.85V1 + 14.7 V 2+ 14.7 V 2 + 0.0 2+ 0.0 2 + r * z 4 = 0.1585
r* 0, y, z 0
Upon solution of above linear program, the following system state is computed:
x T = [V1 V2 2 ] = [0.9948 0.9952 0.1047]

and the following residuals:
rl = | -r* + y1 | = 0.0052
r2= | -r* + y2 | = 0.0052
r3 = | -r* + y3 | = 0.0052
r4 = | -r* + y4 | = 0.00484
Using the above system state and the nonlinear model, the following residuals are
computed using the full nonlinear model:
r1 = V1 1.0 = 0.0052
r2 = V 2 0.99 = 0.0052
r3 = 15.0V1V2 sin 2 + 1.56 = 0.008

r4 = 15.0V22 15.0V1V2 cos 2 0.01 = 0.0773
Second Iteration:
Linearization of the model around the new solution (from the first iteration) yields the
following linear program:
Minimize r*
Subject to:
V1+ V1 r * + y1 = 0.0052
V1+ V1 + r * z1 = 0.0052
V2+ V2 r * + y 2 = 0.0052
V2+ V2 + r * z 2 = 0.0052
Page 24
1.5601V1+ + 1.5601V1 1.5595V2+ + 1.5595V2 + 14.7691 2+ 14.7691 2 r * + y 3 = 0.008

1.5601V1+ + 1.5601V1 1.5595V2+ + 1.5595V2 + 14.7691 2+ 14.7691 2 + r * z 3 = 0.008
14.8463V1+ + 14.8463V1 + 15.0157V 2+ 15.0157V 2 1.552 2+ + 1.552 2 r * + y 4 = 0.0773
14.8463V1+ + 14.8463V1 + 15.0157 V 2+ 15.0157 V 2 1.552 2+ + 1.552 2 + r * z 4 = 0.0773
r* > 0, V1+ , V1 ,... 0, 2+ , 2 0,
y, z 0
Upon solution of above problem, the following system state and residuals are obtained:
V1 =0.0027
V2 = - 0.0027
2 = - 0.0007
Vl = 0.9975
V2 = 0.9925
2 = - 0.1054
rl = | -r* + y1 | = 0.0025
r2= | -r* + y2 | = 0.0025
r3 = | -r* + y3 | = 0.0025
r4 = | -r* + y4 | = 0.0025
Using the above system state and the nonlinear model, the following residuals are
obtained:
r1 = V1 1.0 = 0.0025
r2 = V 2 0.99 = 0.0025
r3 = 15.0V1V2 sin 2 + 1.56 = 0.0023

r4 = 15.0V22 15.0V1V2 cos 2 0.01 = 0.002027
Above completes the solutions.
7.4.4 Summary of the Three State Estimation Approaches

Here we summarize the results of the three approaches for state estimation, namely least
squares (LS), least sum of absolute values (LAV) and minimum maximum residual (minmax). Table 7.1 summaries the solution and the residuals.
Table 7.1 Summary of State Estimation Solutions
Variable
V1
V2
2
r1
LS
0.9974
0.9926
-0.1052
-0.0026
LAV
0.9949
0.99
-0.1058
0.0051
min-max
0.9975
0.9925
-0.1054
0.0025
Page 25
r2
r3
r4
0.0026
0.000629
0.000631
0.0
0.0002
0.00015
0.0025
-0.0023
-0.002027
7.4.5 Quality of the State Estimate

Correct control and operation of the system requires accurate knowledge of the operating
state of the system. State estimation provides the operating state of the system in real
time and, in addition, it provides information about the quality of the state estimate in a
quantitative way. Such analysis provides confidence intervals for the computed estimate
of the state.
With respect to the quality of state estimation, there are two related problems. The first
one relates to the validity of the data (measurements). If the measurements are polluted
with reasonable measurement error (within the specifications of the measuring
instruments), and assuming there is enough redundancy, the state estimate will be
reasonably accurate. However, if one or more data have large errors (due to a number of
reasons), the state estimate will not be accurate. Thus, it is necessary that the state
estimator be smart enough to detect and reject bad data. The second problem relates to
the error transmitted to the state estimate from the measurement error. This error is
measured with the standard deviation of the state estimate. It should be expected that in
the presence of statistically reasonable measurement errors, the standard deviation of the
state estimate should decrease as the redundancy increases.
Consider a measurement of a physical quantity of an electric power system. We have
discussed the fact that this measurement is obtained via an instrumentation channel that
can be complex. The measurement process will exhibit some error. For simplicity, we
introduced a number of assumptions regarding the measurement error. Consider the
normalized error for measurement i, si:
si =
hi ( x ) bi
We have assumed that the normalized errors are Gaussian distributed with standard
deviation 1.0 and zero cross correlation.
Given the meter accuracy defined above, two problems can be defined as follows: (a)
what is the probability that all data are located within expected bounds (Goodness of Fit)
and (b) what is the accuracy of the computed solution? These problems will be
addressed next.
Page 26
7.4.5.1 Goodness of Fit
The goodness of fit is defined as the probability that the distribution of the
measuremenmt errors are within the expected bounds. This probability is computed as
follows. Assume that the state estimate x$ has been computed with the least square
approach. Consider the normalized residuals computed at the solution x$ . We have
postulated that the normalized residuals si are Gaussian random variables with zero mean
and standard deviation 1. Now consider the following variable:
m
= s i2
2
i =1
Since the variables {si, i=1, 2, , m}, are Gaussian random variables, the variable 2 is
also a random variable and it is chi-square distributed [???]. Also since the variables {si,
i=1, 2, , m} are dependent upon only n variables (the state variables x) through a set of
model functions (functions h(x)), the chi-square distributed variable 2 has m-n degrees
of freedom. The chi-square distribution is well known. For example, Table 7.2 tabulates
the probability distribution function of a general chi-square distributed random variable,
Pr( , ) , with degrees of freedom.
Pr( , ) = Pr 2 , where is the degrees of freedom.

Table 7.2 Chi-Square Probability Distribution Function*
p = Pr( 1 , ) = Pr 2 1 , v degrees of freedom]
Page 27
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
.005
.010
.025
.0000393.000157 .000982
.0100 .0201 .0506
.216
.0717 .115
.484
.207
.297
.831
.412
.554
1.24
.676
.872
1.69
.989
1.24
2.18
1.34
1.65
2.70
1.73
2.09
3.25
2.16
2.56
3.82
2.60
3.05
4.40
3.07
3.57
5.01
3.57
4.11
5.63
4.07
4.66
6.26
4.60
5.23
6.91
5.14
5.81
7.56
5.70
6.41
8.23
6.26
7.01
8.91
6.84
7.63
9.59
7.43
8.26
10.3
8.03
8.90
11.0
8.64
9.54
11.7
9.26
10.2
12.4
9.89
10.9
13.1
10.5
11.5
13.8
11.2
12.2
14.6
11.8
12.9
15.3
12.5
13.6
16.0
13.1
14.3
16.8
13.8
15.0
.050
.100
.250
.500
.750
.900
.950
.975
.990
.995
.00393
.103
.352
.711
1.15
1.64
2.17
2.73
3.33
3.94
4.57
5.23
5.89
6.57
7.26
7.96
8.67
9.39
10.1
10.9
11.6
12.3
13.1
13.8
14.6
15.4
16.2
16.9
17.7
18.5
.0158
.211
.584
1.06
1.61
2.20
2.83
3.49
4.17
4.87
5.58
6.30
7.04
7.79
8.55
9.31
10.1
10.9
11.7
12.4
13.2
14.0
14.8
15.7
16.5
17.3
18.1
18.9
19.8
20.6
.102
.575
1.21
1.92
2.67
3.45
4.25
5.07
5.90
6.74
7.58
8.44
9.30
10.2
11.0
11.9
12.8
13.7
14.6
15.5
16.3
17.2
18.1
19.0
19.9
20.8
21.7
22.7
23.6
24.5
.455
1.39
2.37
3.36
4.35
5.35
6.35
7.34
8.34
9.34
10.3
11.3
12.3
13.3
14.3
15.3
16.3
17.3
18.3
19.3
20.3
21.3
22.3
23.3
24.3
25.3
26.3
27.3
28.3
29.3
1.32
2.77
4.11
5.39
6.63
7.84
9.04
10.2
11.4
12.5
13.7
14.8
16.0
17.1
18.2
19.4
20.5
21.6
22.7
23.8
24.9
26.0
27.1
28.2
29.3
30.4
31.5
32.6
33.7
34.8
2.71
4.61
6.25
7.78
9.24
10.6
12.0
13.4
14.7
16.0
17.3
18.5
19.8
21.1
22.3
23.5
24.8
26.0
27.2
28.4
29.6
30.8
32.0
33.2
34.4
35.6
36.7
37.9
39.1
40.3
3.84
5.99
7.81
9.49
11.1
12.6
14.1
15.5
16.9
18.3
19.7
21.0
22.4
23.7
25.0
26.3
27.6
28.9
30.1
31.4
32.7
33.9
35.2
36.4
37.7
38.9
40.1
41.3
42.6
43.8
5.02
7.38
9.35
11.1
12.8
14.4
16.0
17.5
19.0
20.5
21.9
23.3
24.7
26.1
27.5
28.8
30.2
31.5
32.9
34.2
35.5
36.8
38.1
39.4
40.6
41.9
43.2
44.5
45.7
47.0
6.63
9.21
11.3
13.3
15.1
16.8
18.5
20.1
21.7
23.2
24.7
26.2
27.7
29.1
30.6
32.0
33.4
34.8
36.2
37.6
38.9
40.3
41.6
43.0
44.3
45.6
47.0
48.3
49.6
50.9
7.88
10.6
12.8
14.9
16.7
18.5
20.3
22.0
23.6
25.2
26.8
28.3
29.8
31.3
32.8
34.3
35.7
37.2
38.6
40.0
41.4
42.8
44.2
45.6
46.9
48.3
49.6
51.0
52.3
53.7
*Table adopted from CRC book

From basic probability theory we know that the expected value of 2 is:
[ ]
E 2 = = m n , where is the degrees of freedom.

Above statistical properties can be used to compute the probability that the data b is
statistically correct when the state x is computed in the least square sense. We will call
this propability the confidence level of the state estimate. The confidence level is
computed as follows. Consider the least squares solution x$ . This solution minimizes the
sum of the squares of si, i.e. any other state vector x will result in a larger value of 2, i.e.,
m
s
i =1
2
i
( x) = 2 1
where:
m
1 = s i2 ( x )
i =1
Page 28
The probability of above event, 2 1 , is given by the chi-square distribution:
Pr 2 1 = 1.0 Pr 2 1 = 1.0 Pr( 1 , )

This propability expresses how well the normalized residuals si are distributed within the
expected bounds. A high propability value indicates that these residuals are well within
the statistical bounds, i.e. the normalized residuals are gaussian distributed within the
range (-1.0 to 1.0). This means that the actual residuals are comparable to the expected
errors of the measurements. A low propability value indicates that the residuals are
higher than what is statistically expected. This is the reason for calling this propability
the confidence level for the estimation results.
A simple procedure to compute the confidence level is given below.
Step 1. Compute the state estimate, x$ , in the least squares sense.
2
h ( x ) bi
.
Step 2. Compute the value 1 = s ( x ) = i
i
i =1
i =1
2
Step 3. Read the probability Pr 1 = Pr( 1 , ) from the tabulated chi-square
propability distribution function. Note that interpolation can be used or more detailed
data of the distribution function.
Step 3. Compute the probability Pr 2 1 = 1.0 Pr( 1 , ) .
m
2
i
Note that the confidence level can be computed only for the least squares solution.
However, as an approximation, the above procedure can be applied to the other two
solutions for x$ (L1 and L).
Pr( 1 , ) = Pr 2 1
7.4.5.2 Accuracy of Solution
The accuracy of the solution is expressed with the covariance matrix of the state estimate,
x . Specifically, let x be the true but unknown solution, and x be the solution to the
problem (7.5). This solution may be the least squares, L1 or L solution. The definition
of the covariance matrix is:
C x = E[( x x )( x x ) T ]
Note that a linearized expression for x x is as follows (which is obtained by applying
the state estimation algorithm at point x )
x x = ( H T WH ) 1 H T Wr
for least square solution
Page 29
where r = h( x ) b and H is the jacobian matrix. The jacobian is supposed to be

computed at the true state x . However, since the true state is not known, we approximate
the jacobian matrix at the known best estimate of the state, x . As we discussed, the
residuals r represent the measurement errors, i.e. r = . The statistics of the measurment
[ ]
error have been introduced and repeated here E [ ] = 0 and E T = W 1 .
Upon substitution of above into the definition of the covariance matrix, the following is
obtained
C x = E[( H T WH ) 1 H T W T W T H ( H T WH ) T ]
Since the only random variables in above equation is the measurement errors , then
above equation is rewritten as follows:
C x = ( H T WH ) 1 H T W {E T }W T H ( H T WH ) T
Now the above equation is simplified to yield:
C x = ( H T WH ) 1
Once the covariance matrix of the solution has been computed, the standard deviation of
a component of the solution vector x is given with
x = C x (i , i )
i
where, Cx (i, i ) is the ith diagonal entry of the covariance matrix.

The covariance matrix is also known as the information matrix since it provides useful
information on the expected error of the computed state variables. We symbolize the
information matrix with I:
I = ( H T WH ) 1
Other measures of accuracy can be also derived such as E( b$ ), Cov( b$ ). These variables
are derived next. Consider the estimate of the measurements defined with:
b = h( x )
The statistics of the estimate b$ are computed as follows:
[]
E b = h( x )
Page 30
C b = Cov(b) = E[(b b )(b b ) T ]

where b is the true value of the measurements.
Note that:
b b = h( x ) h( x ) = H ( x )( x x ) H ( x )( x x )
Upon substitution:
Cov(b) = E H ( x x )( x x ) T H T = H ( H T WH ) 1 H T
Now lets compute the covariance
Cov(b b) = E (b b)(b b) T
Note that:
b b = (b b ) (b b ) = [h( x ) h( x )] r H ( x )( x x ) r = H ( H T WH ) 1 H T Wr r
Upon substitution and some straightforward manipulations:

Cov ( b b) = W H ( H T WH ) 1 H T
A summary of the state estimation statistics is given in Table 7.3.

Table 7.3 Summary of Statistical Properties of Static Estimators
E [x ] = x
[]
[ ]
E b = h( x )
E b b = 0
E [J ] = m n
Cov( x ) = C x = ( H T WH ) 1
Cov (b) = H ( H T WH ) 1 H T
Cov (b b) = W H ( H T WH ) 1 H T
Pr 2 1 = 1.0 P( 1 , m n)
In summary, the quality of the state estimate is quantified as follows: A measure of data
validity is expressed with the confidence level obtained from the chi-square distribution.
The accuracy of the estimated state variables is given with the diagonal entries of the
information matrix which express the square of the standard deviation.
The evaluation of the quality of a state estimate is illustrated with an example.
Example E7.5. Consider the problem solved in Example E7.2. Compute the confidence
Page 31
level and the standard deviation of the solution components (V1, V2, 2), as well as the
standard deviation of the measurement estimates given that the error of the measurements
is 0.02, 0.03, 0.03, and 0.04 for the measurements of V1, V2, P2l and Q21, respectively.
Use the least square approach.
Solution: First, the least square technique is used to solve the problem starting with an
initial guess:
x T = [V1 V2
2 ] = [1.0 0.99 0.0]
1st Iteration
0.0
0.0
0
r = h( x ) b =
1.56
0.1585
0.0
0.0
1.0
1.0
0.0
0.0
0
H (x ) =
0.0
0.0 14.85
14
.
85
14
.
7
0
.
0
1471.08
H W [h( x ) b] = 1456.22
25740.0
0.0
140327.0 136434.0
H WH = 136434.0 136167.0
0.0
0.0
0.0
245025.0
0.003312
( H WH ) H W [h( x ) b] = 0.007376
0.10505
1.0 0.003312 0.9967
x = 0.99 0.007376 = 0.9974

0.0 0.10505 0.10505
Page 32
2nd Iteration
0.9967 1.0 0.0033
0.9974 0.99 0.0074

1
r = h( x ) b =
=
1.563587 1.56 0.00359
0
.
08268
0
.
01
0
.
0926764
0.0
0.0
1.0
1.0
0.0
0.0
'
H (x ) =
1.5688 1.5677 14.829
14.879 15.0539 1.5636
770.865
H W [h( x ) b] = 792.386
139.950
'
143600.0 137259.0 11308.1
H WH = 137259.0 145479.0 40541.9

11308.1 40541.9 245860.0
0.0017
( H WH ) H W [h( x ) b] = 0.0038
0.000015
'
0.9967 0.0017 0.9984
x = 0.9974 0.0038 = 0.99364

0.10505 0.000015 0.10504
0.9984 1.0 0.0016
0.9936 0.99 0.0036

2
2
r = h( x ) b =
=
1.56014 1.56 0.00014
0.010475 0.01 0.000475

The information matrix is:
Page 33
I = ( H WH )
T
2.7583 2.7649 0.58279
= 10 2.7649 2.8435 0.5961

0.58279 0.5961 0.16577
The standard deviations of the system states are:
V = 2.7583 x 10 -4 = 0.01661
1
V = 2.8435 x 10 -4 = 0.01686
2
= 0.16577 x 10 -4 = 0.00407
2
Note that the standard deviation of the state variables is lower than those of the
measurements.
The confidence level is calculated as follows:
r
1 = i
i =1 i
4
0.0016
0.000475
0.00014
0.0036
=
= 0.020963
+
+
+
0.04
0.03
0.03
0.02
Since = m n = 1 , we obtain from Table 7.2:
Pr 2 0.020963 = Pr(0.020963, ) = 0.115

Thus, the confidence level is:
Pr 2 0.020963 = 1.0 Pr(0.020963, ) = 0.885

This completes the solution.
7.4.6 Summary and Discussion

In summary, a state estimator is a computational procedure by which a statistical estimate
of the state of the system is obtained. Inputs to the state estimator are a set of
measurements, which are characterized with their statistical properties, and a
mathematical model describing the system. The state estimator, in addition to the state
estimate, x$ , computes the covariance of x$ , the estimates of the measurements, b$ , the
covariance matrix of b$ , the residuals b - b$ , and the value of the objective function. The
inputs and outputs of the state estimator are illustrated in Figure 7.6. Of importance is
Page 34
the value of the objective function, which can provide information about the quality of
the computed state estimate. For this purpose the chi-square test is utilized.
Measurements
& Telemetry
Noise
Noiseless
Measurements
System
Model
State
Estimator
Measurement Estimate b
p
Covariance of b,
Residual b b
State Estimate x
I
Covariance of x,
Objective Function Value
Probability of Goodness
of Model Fit
(Confidence Level)
Figure 7.6 Inputs and Outputs of the State Estimator
7.5 Detection and Identification of Bad Data

The presence of bad data deteriorates the performance of the state estimate. It is
imperative, therefore, that bad data be detected, identified, and rejected. This objective is
achieved with information provided from redundant measurements, a characteristic of the
state estimator.
There are two interrelated problems: the first problem is the one of detecting the
existence of bad measurements; the second problem is the one of identifying which data
is bad. These two problems are addressed next.
7.5.1 Detection of Bad Data

Detection of the existence of bad data can be achieved with the chi-square test, i.e. by
computing the confidence level. If the system of Equations (7.5) is free of bad data, the
confidence level will be high. In the presence of one or more bad data, the confidence
level will decrease. This means that whenever the confidence level is low, bad data exist
Page 35
in the measurement set. Note that the chi-square test does not indicate which datum or
data is bad. The identification of the bad data is achieved with other methods to be
described below.
7.5.2 Identification of Bad Data

Identification of bad data normally consists of two steps. In the first step, bad data may
be identified by inspection or simple consistency rules. This step identifies the obviously
bad data and it is very much system dependent. As an example, in power system state
estimation, measurements of voltages, power flow, etc., are known to have specific
ranges. If a measurement is out of this range, it will be classified as a bad measurement
or at least as a measurement suspected of being bad (suspect measurement). In the
second step, bad data are identified with statistical analysis of the residuals and/or its
effects on confidence level. This analysis depends on the selected method for the
solution of Equation (7.5). In the case of least square solution, the possible bad data are
identified with their large residuals. However, it is known that it is possible that: (a) a
measurement with a large residual may not be always a bad measurement and (b) a bad
measurement may have a very small residual (outliers). A rather secure but
computationally demanding way to identify a bad datum is by means of hypothesis
testing. Specifically, assume that a measurement (or a group of measurements) has been
identified as suspect (this characterization may be due to a large normalized residual or
because of failure to pass a consistency check, etc.). For this purpose, the suspect datum
is removed, i.e. the corresponding equation bi = hi(x) is removed from Equations (7.5)
and the least square solution is computed again. Subsequently, the confidence level is
computed. A drastic improvement in the confidence level indicates that the data under
consideration is bad. This procedure tends to be computationally demanding. On the
other hand, the Ll and L solutions tend to be more versatile in identifying the bad data.
For example, the L solution directly provides the measurement with the largest residual.
Detection and identification of bad data will be illustrated with examples. First we
consider a general example to illustrate the effects of outliers. Then we discuss a typical
application of hypothesis testing in the framework of least squares approach. The final
example illustrates bad data detection and identification in the context of the min-max
approach.
Example E7.6. Consider the set of overdetermined equations below resulting from five
measurements. It is desired to estimate the state vector of these equations, determine the
presence or absence of bad measurements and identify the bad measurements if any.
Assume that all measurements have a statistical error of standard deviation 0.1.
2 = x1 + x2
3.1 = 2 x1 + x2
4 = 3 x1 + x2
4.1 = 7 x1 + x2
Page 36
Solution: The equations relating the measurements to the state of the system are linear.
Therefore a linear state estimator provides the solution. The model equations in matrix
form are:
1
2
b = Hx, where : H =
3
1
2
3.1
1
, b=
4
1

1
4.1
The least squares solution of above problem is:

x
x = 1 = H T WH
x 2
0.2891
H T Wb =
2.3602
The residuals are:

0.6494
0.1614
r = b Hx =
0.7723
0.2843
The sum of the squares of the residuals is:
2
r
J = i = 112.5
1 0 .1
4
The probability of goodness of fit is obtained from Table 7.1 with m-n=2 degrees of
freedom:
Pr 0.0
This probability indicates the presence of bad data in this measurement set. Therefore bad
data have been detected. The next step is to identify the bad data.
The bad data identification will be done with hypothesis testing. Specifically, the
following three hypothesis will be examined:
Hypothesis 1: Measurement 3 (largest residual) is bad.
Hypothesis 2: Measurement 1 (next largest residual) is bad.
Page 37
The computations and conclusions follow.

Hypothesis 1: Measurement 3 is removed and the resulting estimation problem is solved.
The solution is:
x
x = 1 = H TWH
x2
0.289
H T Wb =
2.360
The residuals, sum of weighted squared residuals and probability of goodness of fit with
m-n=1 degrees of freedom (Table 7.1) are:
0.36291
r = b Hx = 0.4355 ,
0.0725
r
J = i = 32.66 ,
1 0 .1
4
Pr 0.0
Note that the probability is still low indicating that measurement 3 may not be a bad
measurement, or there are additional bad measurements in the set.
The solution is:
x
x = 1 = H TWH
x2
0.15
H T Wb =
3.133
0.333
r = b Hx = 0.416 ,
0.083
r
J = i = 29.167 ,
1 0. 1
4
Pr 0.0
Note that the probability is still low indicating that measurement 1 may not be a bad
measurement, or there are additional bad measurements in the set.
The solution is:
x
x = 1 = H T WH
x2
Page 38
1 .0
H T Wb =
1.0333
0.0333
r = b Hx = 0.0667 ,
0.0333
r
J = i = 0.667 ,
1 0 .1
4
Pr 0.56
Note that the above probability indicates that measurement 4 is a bad measurement and
should be removed from the set of measurements permanently. For this example, there is
no need to proceed further.
Note that the bad measurement was identified in the third hypothesis test. In the original
state estimation solution the residual of the 4th measurement was not the largest
(absolutely). In other words the solution of the initial state estimation problem failed to
yield a high enough residual for this measurement. Only the hypothesis testing was able
to identify the bad measurement. In this case the probability by which the bad datum was
identified is relatively low (0.56) due to the fact that the number of redundant
measurements is very low. One can provide a geometrical interpretation of this data,
observing that the model equations are of the form:
b = ax1 + x2
Graphing the data on a coordinate system where the horizontal axis represents a and the
vertical axis represents b, Figure E7.6 is obtained. In this figure one can clearly see that
three of the data lie on an almost straight line, while the forth datum lies away from this
line (outlier). In this case, the graphical representation of the data can immediately yield
the bad datum.
6
5
4
b 3
2
1
0
0
Page 39
Figure E7.6 Graphical Representation of the Data of Example E7.6

Example E7.7. Consider the system of Figure E7.7. Five measurements are taken as
follows: V2 = 0.99, Pl2 = 1.4, Ql2 = 0.15, P2l = -1.46, Q21 = 0.01.
-j15.0

MW Flow Measurement
Figure E7.7 A Simple Two Bus Power System

All measurement instruments are known to have an accuracy of 1%. The voltage at bus 1
is 1.0 p.u. with absolute certainty. Perform bad data identification.
Solution: First, the state estimation problem is solved.
computed state estimate is:
Details are omitted.
The
0.09940
x =
0.99497
The covariance matrix of the state estimates is:
0.91 0.09
Cov( x ) = 10 6
0.09 0.90
The measurement estimates are:
0.99497
1.48111
b = 0.14914
1.48111
0.00141
The residuals are:
Page 40
.00497
.08111
b b = .01085
.07880
.01141
The value of the objective function at the estimate is:
J ( x ) = 32.688
which yields a probability of statistically correct estimate:

p 0 .0
The low (zero) probability means that bad data exist in the measurement set (detection
step).
Inspection of the residuals reveals that the second or the fourth measurement may be
bad. To determine which one is bad (or maybe both are bad), hypothesis testing is
employed. Three hypotheses will be examined: (1) the fourth measurement is bad, (2)
the second measurement is bad, and (3) both (second and fourth) measurements are bad.
Hypothesis 1. In this case, the fourth measurement is removed. The state estimate
becomes:
.09419
x =
.99496
and the objective function at the estimate is:

J ( x ) = 1.8358
the probability that the estimate is statistically correct is computed to be (m - n = 2):

p = 0.42
Above probability is too low. With high probability, this measurement is not a bad
datum.
Hypothesis 2. In this case, the second measurement is removed. The state estimate
becomes:
.10475
x =
.99497
Page 41
and the objective function at the estimate is:

J ( x ) = 0.10777
The probability that the estimate is statistically correct is computed to be (m - n = 2):

p = 0.95
Obviously, the second measurement is bad. The degradation of the probability of

statistically correct estimate is due to some loss in data redundancy. Note that the
m
redundancy ( x 100) dropped from 250% to 200%.
n
Hypothesis 3 need not be examined.
Bad data identification in the min-max method can be performed with sensitivity
information provided by the LP algorithm. The procedure is outlined with an example.
Example E7.8. Consider the problem of Example E7.7. Perform bad data identification
using the min-max method.
Solution: First, the state estimate must be computed in the min-max sense. For this
purpose, we start from the known state estimate in the least square sense
0.09940
x =
0.99497
Formulation of the state estimation problem in the min-max sense in terms of the
variables:
V2 = V2 0.99497
2 = 2 + 0.0994
yields:
Minimize r*
Subject to:
V2+ V2 r * + y1 = 0.005
V2+ V2 + r * z1 = 0.005
1.4885V2+ 1.4885V2 14.8509 2+ + 14.8509 2 r * + y2 = 0.0811
1.4885V2+ 1.4885V2 14.8509 2+ + 14.8509 2 + r * z2 = 0.0811
14.926V2+ + 14.926V2 1.4811 2+ + 1.4811 2 r * + y 3 = 0.0009
Page 42
14.926V2+ + 14.926V2 1.4811 2+ + 1.4811 2 + r * z 3 = 0.0009
1.4885V2+ + 1.4885V2 + 14.8509 2+ 14.8509 2 r * + y 4 = 0.0789

1.4885V2+ + 1.4885V2 + 14.8509 2+ 14.8509 2 + r * z 4 = 0.0789
14.9231V2+ 14.9231V2 1.4811 2+ + 1.4811 2 r * + y 5 = 0.0114
14.9231V2+ 14.9231V2 1.4811 2+ + 1.4811 2 + r * z 5 = 0.0114
r * 0,
y 0, z 0
Upon solution of the linear problem, the system state is:
V2 = -0.0046
2 = -0.0004
and
V2 = 0.9904
2 = -0.0998
and the residuals are:

rl = -r* + y1
r2= -r* + y2
r3 = -r* + y3
r4 = -r* + y4
r5 = -r* + y5
= - 0.000364
= - 0.08
= - 0.0689
= - 0.0798
= 0.08
Using the above computed state and the nonlinear model, the residuals are computed to
be:
rl = - 0.0004
r2 = - 0.0802
r3 = - 0.0679
r4 = - 0.0798
r5 = 0.0787
The second iteration is:

Minimize r*
Subject to:
V2+ V2 r * + y1 = 0.0004
V2+ V2 + r * z1 = 0.0004
1.4945V2+ 1.4945V2 14.7821 2+ + 14.7821 2 r * + y 2 = 0.0802
1.4945V2+ 1.4945V2 14.7821 2+ + 14.7821 2 + r * z 2 = 0.0802
14.9254V2+ + 14.9254V2 1.4802 2+ + 1.4802 2 r * + y 3 = 0.0679
14.9254V2+ + 14.9254V2 1.4802 2+ + 1.4802 2 + r * z 3 = 0.0679
Page 43
1.4945V2+ + 1.4945V2 + 14.7821 2+ 14.7821 2 r * + y 4 = 0.0798

1.4945V2+ + 1.4945V2 + 14.7821 2+ 14.7821 2 + r * z 4 = 0.0798
14.7866V2+ 14.7866V2 1.4802 2+ + 1.4802 2 r * + y 5 = 0.0787
14.7866V2+ 14.7866V2 1.4802 2+ + 1.4802 2 + r * z 5 = 0.0787
r * 0,
y 0, z 0
Upon solution of the linear program, the system state is:
V2 = - 0.00009
2 = - 0.0
and
V2 = 0.99031
2 = -0.0998
and the residuals are:

rl =
r2 =
r3 =
r4 =
r5 =
-r* + y1
-r* + y2
-r* + y3
-r* + y4
-r* + y5
= 0.0003
= 0.08
= 0.0692
= 0.08
= 0.08
Using the above system state and the nonlinear model, the residuals are computed to be:
rl = - 0.00031
r2 = 0.08
r3 = - 0.0693
r4 = - 0.08
r5 = 0.08
The sensitivities of the largest residual with respect to the measurements are obtained
from the linear programming problem (see Appendix B):
dr *
= 0.0 ,
db1
dr *
= 0.5 ,
db2
dr *
= 0.0 ,
db3
dr *
= 0.5 ,
db4
dr *
= 0.0
db5
To determine the quality of the state estimate, the chi-square test is performed. Recall
that the meter accuracy is 2%. Then:
Page 44
0.00031
0.08
0.0693 0.08 0.08
J =
+
+
+
+
= 60.0075
0.02
0.02
0.02 0.02 0.02
Since m - n = 3, the confidence level will be:

Pr 0.0
This confidence level indicates the presence of bad data. Examining the above data, it is
observed that the second measurement will have the largest effect on the largest residual,
equal to:
dr *
r2 = 0.0401
db2
Thus, the second measurement is suspected to be wrong. The second measurement is

removed from the measurement set and the L state estimation problem is solved starting
from the previous solution:
V2= 0.99031
2 = - 0.0998
Next, the L problem is solved as follows:
1st Iteration
Minimize r*
Subject to:
V2+ V2 r * + y1 = 0.00031
V2+ V2 + r * z1 = 0.00031
14.9254V2+ + 14.9254V2 1.4800 2+ + 1.4800 2 r * + y 3 = 0.0693
14.9254V2+ + 14.9254V2 1.4800 2+ + 1.4800 2 + r * z 3 = 0.0693
1.4945V2+ + 1.4945V2 + 14.7807 2+ 14.7807 2 r * + y 4 = 0.08

1.4945V2+ + 1.4945V2 + 14.7807 2+ 14.7807 2 + r * z 4 = 0.08
14.7839V2+ 14.7839V2 1.4800 2+ + 1.4800 2 r * + y 5 = 0.08
14.7839V2+ 14.7839V2 1.4800 2+ + 1.4800 2 + r * z 5 = 0.08
r * 0,
y 0, z 0
Upon solution of the linear program, the system state is computed to be:
Page 45
V2 = - 0.00476
2 = - 0.00459
and
V2 = 0.99507
2 = -0.10439
and the residuals are computed to be:

rl = | -r* + y1 | = - 0.005068
r3 = | -r* + y3 | = - 0.005068
r4 = | -r* + y4 | = - 0.005068
r5 = | -r* + y5 | = 0.0286
Using the computed system state and the nonlinear model, the residuals are computed to
be:
rl = - 0.00507
r3 = - 0.0052
r4 = - 0.0047
r5 = - 0.0023
2nd Iteration
Minimize r*
Subject to:
V2+ V2 r * + y1 = 0.00507
V2+ V2 + r * z1 = 0.00507
14.9183V2+ + 14.9183V2 1.5553 2+ + 1.5553 2 r * + y3 = 0.0052
14.9183V2+ + 14.9183V2 1.5553 2+ + 1.5553 2 + r * z3 = 0.0052
1.5630V2+ + 1.5630V2 + 14.8448 2+ 14.8448 2 r * + y4 = 0.0047

1.5630V2+ + 1.5630V2 + 14.8448 2+ 14.8448 2 + r * z 4 = 0.0047
14.9338V2+ 14.9338V2 1.5553 2+ + 1.5553 2 r * + y5 = 0.0023
14.9338V2+ 14.9338V2 1.5553 2+ + 1.5553 2 + r * z5 = 0.0023
r * 0,
y 0, z 0
Upon solution of the linear program, the system state is computed to be:
V2 = - 0.000006
2 = - 0.000026
and
V2 = 0.995076
2 = -0.104364
and the residuals as computed with the linear program, the residuals as computed with
the nonlinear model, the sensitivities and the effect of measurements on the largest
Page 46
residual are:
Residual
LP Model
r1
r2
r3
r4
r5
0.005076
--0.005076
0.005076
0.002256
Nonlinear
Model
- 0.005076
--- 0.0051
- 0.0051
0.0023
Sensitivities
dr*/dbi
0.931749
--0.061779
0.006473
0.0
Note that r* is positive for each measurement.

measurement will deteriorate the state estimate.
r*
0.004724
--0.000321
0.000030
0.0
This means removal of any
To determine the quality of the state estimate, a chi-square test is again performed using
exactly the same performance index. Specifically, rn = 0.10 again.
2
0.005076 0.0051 0.0051 0.0023

J =
+
+
+
= 0.20775
0.02 0.02 0.02 0.02
Now m - n = 2.
Reading on the table, the confidence level is:
Pr = 0.952
The jump in the confidence level from zero to 0.952 indicates that the second
measurement is a bad measurement. Note that the min-max approach immediately
identified the bad measurement. Compare this performance with that of the least square
approach, where the bad measurement was identified at the second hypothesis testing.
7.5.3 Summary and Discussion

In this section we discussed methods for detecting and identifying bad measurements in a
set of data. This process can be accomplished only when there are redundant
measurements. Thus, redundant measurements are needed to enhance: (a) the ability to
detect and reject bad measurements and (b) to obtain the state variables with the smallest
possible error. Typically, the number of measurements is two-to-three times greater than
the number of states variables allowing for a considerable amount of measurement
redundancy (200 to 300%). In this process we tacitly assumed that the system state is
observable from the measurement set.
Page 47
Measurement Set z
Perform State Estimation
Compute Quality of Estimate
Is Estimate
Acceptable?
YES
STOP
NO
Determine Maximum
Normalized Residual sj
Remove Measurement j
Perform State Estimation
Compute Quality of Estimate
Is Performance Improvement
Substantial?
YES
NO
Restore Measurement j
Discard Measurement j
Figure 7.x. Flow Chart for Bad Data Detection, Identification and Rejection
7.6 Sequential State Estimators

Sequential state estimators are mathematical procedures by which the effect of a single
measurement on the state estimate is evaluated. Using this procedure, the state estimate
can be computed by processing each measurement individually until all measurements
have been accounted. Sequential state estimators are useful in two application areas: (a)
in real time applications, where a state estimate has been computed and a set of few new
measurements is received. One way to incorporate the new measurements into the state
estimation process is to add the new measurements to the old ones and recompute the
state estimate using the entire set of measurements. This approach is in general inefficient
from the computational point of view. Another way is to evaluate the effect of the new
Page 48
measurements on the state estimate with the use of the sequential state estimator. (b) in
bad data identification and especially in hypothesis testing, it is required to evaluate the
effect of removal of a measurement on the state estimate. This problem is ideally suited
to sequential state estimators since with minimal computations one can evalauet the
effects of removing one measurement and therefore can determine if the measurement is
bad or not. The sequential state estimator is presented in this section.
Sequential state estimation algorithms can be developed for general nonlinear systems.
For simplicity we will limit the presentation to linear systems. One can expand the linear
sequential state estimation algorithm to nonlinear systems by linearization of the
problem.
Consider the linear system
r = Hx b
Cov( r ) = W 1
where
b :
x :
r :
vector of measurements; dimension m

state vector; dimension n
error of measurements; dimension m.
The estimate of x in the least square sense, i.e., the one that minimizes the objective
J = r T Wr
is given by
x = (H T WH ) H T Wb
1
Now assume that a new scalar measurement is added to the set of measurements
r ' = hT x b'
E ( r ' ) = 0.0
E(r ' r ' ) = 2

The new estimate x$ ' in the least square sense is given by
x ' = x a
where:
= A1h
Page 49
r'
+ hT
A = H T WH
a=
Proof: The addition of the new measurement will augment the model as follows:
r H
b
r ' = h T x b'

r W 1
Cov =
r ' 0
The new state estimate is:
H T W
x ' = T
h 0
T
0 H H W
2 h T h T 0
0 b
2 b'
(7.11)
Note that
T
H W
A' = T
h 0
0 H
= H T WH + hh T 2 = A + hh T 2
2 h T
where
A 1
A' 1
is the old information matrix

is the new information matrix.
Application of the matrix inversion lemma yields:
A' 1 = A + hh T 2
= A1
1
A1hh T A1
T 1
+h A h
2
Also
T
H W
h T 0

0 b
= H T Wb + 2 b' h
2
b'
Substitution of above expressions into (7.11) yields:
1
x ' = A1 2
A1hh T A1 H T Wb + 2 b' h
T 1
+h A h
Page 50
Recall that x = A 1 H T Wb and define = A-1 h

Upon substitution and straightforward manipulations:
Then
r'
x ' = x 2
T
+h
r'
is a scalar which shall be called a. In this way, equation (7.24) is
+ hT
obtained. The above results suggest the following sequential linear state estimator given
a new measurement of b = hTx:
The quantity
Step 1 : Compute = A-1 h

r'
Step 2 : Compute a = 2
+ hT
Step 3 : Compute new estimate x$ = x$ a
Step 4 : Update the inverse of the information matrix A A + 2 hh T
Step 5 : If there are more measurements go to step 1.
Note that above algorithm is not self-starting. Step 1 requires the current information
matrix I (= A-1). When the algorithm starts, the matrix I is undefined. Also, step 2
requires the current estimate x$ . These drawbacks can be overcome by applying the
following initialization procedure: Select n independent measurements. These
measurements, define a deterministic problem: n equations in n-unknowns. Solution of
this problem will provide an initial state estimate x$ . At this solution, compute the
inverse of the information matrix. Use the computed state and information matrix to start
above algorithm.
The sequential linear estimator algorithm is directly applicable to the nonlinear state
estimation of power systems. The basic equations are derived as follows. Let b be a set
of measurements yielding the state estimate x$ . The model of the system is described
with the equations
r = h( x ) b
E ( r ) = 0.0
Cov( r ) = W 1
Assume a new measurement ba is added with the following model
ra = ha ( x ) ba
E ( ra ) = 0.0
Page 51
Cov( ra ) = 2
The new state estimate, assuming that one iteration is sufficient, will be:
x ' = x + ( A' ) 1 H T
W
h
0
0 b h( x )
2 ba ha ( x )
where
A :
A =
A :
A=
H :
h :
the new inverse of the information matrix

A + -2 hhT
the old inverse of the information matrix
(HTWH)
the old Jacobian matrix
the contribution to the Jacobian matrix from the new measurement.
Application of the matrix inversion lemma as in the case of the linear state estimator, and
noting that
A 1 H T W (b h( x ) ) 0,
yields:
x ' = x a
where
= A 1 h
a=
h' ( x ) b'
2 + hT
The above result is directly utilized in a sequential nonlinear state estimator.

procedure will be illustrated with an example.
The
Example E7.9: Consider the simple electric power system of Figure E7.9 with the
illustrated measuring system. The voltage at bus 1 is known to be 1.0 p.u. without error.
Other measurements taken are
b1 = V2 = 0.97 + r1 , Var(r1 ) = (0.02) 2

b2 = P12 = 1.94 + r2 , Var(r2 ) = (0.02) 2
b3 = P21 = 1.93 + r3 , Var (r3 ) = (0.02) 2
b4 = Q21 = 0.48 + r4 , Var(r4 ) = (0.02) 2

ri : measurement error of ith measurement.
Page 52
~
V1 =V1e j0
-j20.0

MW Flow Measurement
Figure E7.9
The state estimate based on above set of measurements is

0.09988
x = 2 =
V 2 0.97028
and the information matrix at the estimate is

1.8735 .2839
6
I 1 ( x ) =
x10
.
2839
.
9168
a) Compute the probability that the estimate is statistically correct.

b) A new measurement is taken: The reactive power flow Q12 is measured to be 0.70 p.u.
The variance of this new measurement error r5 is var(r5) = (0.02)2. Compute the new
state estimate x$ ' . (One iteration is enough.)
c) Compute the probability that the new estimate x$ ' is statistically correct.
Solution: a) At the computed state estimate, the measurement estimate is:
P12 = 20.0V1V 2 sin 2 = 1.93501
P21 = 20.0V1V 2 sin 2 = 1.93501
Q21 = 20.0V22 20.0V1V2 cos 2 = 0.48002

The value of the objective function, computed at the estimate x$ , is
h ( x) bi
J = i
i
i =1
4
= 0.125197
Page 53
Using Table 7.2, the probability that the data are statistically correct is:
Pr = 0.94
b) The model of the new measurement Q12 is:
h5 ( x) = Q12 = 20.0V12 20.0V1V2 cos 2

At the previously computed state estimate
h5 ( x) = Q12 = 0.69111
Thus
b5 h5 ( x) = 0.0089
The linearized model is
hT = [20 V1V2 sin2
- 20 V1 cos2] = [-1.935
-17.9003]
The vector is computed

4.53487
x10 6
23.110592
= I ( x ) 1 h =
Also
hT
= .00086868
= 0.02
= -.0089/.00086868 = -10.245
Application of equation (7.27a) yields:

.09993
x ' =
.97004
c) The estimates of the measurements are computed at the new state estimate x.
V2 = 0.97004
P12 = 1.9355
P21 = -1.9355
Page 54
Q21 = -0.48446
Q12 = 0.69599
The value of the objective function is computed:
J( x$ ' ) = 0.16645
Using Table 7.2, the probability that the data are statistically correct is :
Pr = 0.9825
The increased confidence is due to the addition of one more good measurement.
7.7 State Estimator Observability

One important question in state estimation relates to the observability of the system state.
Simply stated, given a set of measurements can the system state be estimated (observed)
from these measurements? We will consider this question and related issues.
7.7.1 Algebraic Observability

For the state of the system to be observable it is necessary that the number of
measurements be greater or equal to the number of states. This is a necessary condition
but not sufficient. To obtain the sufficiency condition, consider a system on which mmeasurements were taken (vector z) while the system state is an n-vector(vector x). The
linearized model of the measurement is in general
z = Hx +
where:
z
x
H
is an m-vector of measurements
is an n-vector of states
is an m x n matrix
is an m-vector of measurement noise
A sufficient condition to obtain a unique solution for x in a specified sense (i.e. least
squares sense, LAV sense or min-max sense) is that the rank of H be n. This means that
the sufficient conditions for observability are:
m>n
rank(H) = n
Above conditions are sufficient conditions for algebraic observability. Typically, the
term observability always means algebraic observability.
Page 55
In general, the matrix H is not a square matrix. A computationally simple way to

determine the rank of H is based on the observation that
rank ( H ) = rank ( H T H )
The matrix HTH is a square matrix. Now the algebraic observability test becomes
equivalent to determining whether the matrix HTH has full rank. This can be achieved
with an LU factorization procedure as it has been discussed in Appendix A. Specifically,
the matrix H T H is factored into the product of two triangular matrices, L and U.
H T H = LU
Where:
L is a lower triangular matrix,

U is an upper triangular matrix
Note that
det{H T H }= det{L}det{U } = l ii u ii ,
i
where:
the symbol means product,
l represents entries of the lower diagonal (L) matrix and
u represent entries of the upper diagonal (U) matrix.
Thus the matrix HTH is full rank if and only if all the diagonal elements of the L and U
matrices are non zero. For further information see Appendix A.
7.7.2 Topological Observability

For the power system state estimation problem work by Clements [???] and others has
proven that the topology of the matrix H provides a good measure of the rank of the
matrix H and therefore algebraic observability. The process by which we deduct the rank
of matrix H from its topology is known as topological observability. Computationally
topological observability is much more efficient than algebraic observability and it is
preferred. However, it is important to keep in mind that topological observability does
not imply algebraic observability for a general system.
Topological observability analysis is performed as follows. Consider for example a
measurement zi. The model of the system may relate the measurement zi to a number of
system states. Denote these states as the subset X zi of the state vector entries x. Thus the
measurement zi spans the subset X zi . Now a given set of measurements z, span a subset
Page 56
Xz
of system states equal to the union of all subsets X zi corresponding to each
measurement zi, i.e.

X z = U X zi
i
Krumpholz, Clements and Davis [???] have demonstrated with simplified models, that
topological observability is equivalent to algebraic observability for practical power
system estimation problems. Thus, one can use either approach to determine system state
observability. A consequence of this fact is that one can use the topology of the Jacobian
matrix to determine system state observability. As a matter of fact we shall discuss a
method which uses the topology of the information matrix to interpret it as a connectivity
network of the system states. Then using a topology check, the observability can be
determined.
The corresponding topological observability test can be defined as follows. First observe
that the matrix HTH can be interpreted as a network admittance matrix since it is
symmetric. The network has a branch between nodes i and j if and only if the entry ij of
the matrix HTH is non zero. Thus the position of the off-diagonal non zero entries of the
matrix HTH determine the network connectivity of the system states. The system is
observable if and only if this network is fully connected(all states are connected). Highly
efficient algorithms to check connectivity of a network exist, making this approach
practical.
Algebraic and topological observability will be demonstrated with an example.
Example E7.10. Consider the four bus system of Figure E7.10. A number of
measurements are taken as indicated.
3
MW Flow Measurement
Figure E7.10 A Simplified Four Bus System
Page 57
a)
b)
Write the Jacobian matrix H. Only the topology is of interest. Thus set all non zero
entries equal to 1.0.
Write the equation HT(z-b) = (HTH)x and draw a circuit that corresponds to above
equation.
Solution: a)
z = [P12
The total number of measurements is 7. They are:
Q12
P32
Q 32
P24 V1 V 2 V 3 V 4 ]
The system state x is defined as: xT = [V1 2 V2 3 V3 4 V4 ] , the voltage phase

angle at bus 1 is selected to be the reference. The Jacobian matrix with each entry
replaced with 1.0 is obtained as follows:
1
1
0
H = 0
1
0
0
0
b)
1 1 0 0 0 0
1 1 0 0 0 0
1 1 1 1 0 0
1 1 1 1 0 0
1 1 0 0 1 1
0 0 0 0 0 0
0 1 0 0 0 0
0 0 0 1 0 0
0 0 0 0 0 1
The matrix HTH is computed to be:
3
2
T
H H = 0
0
0
0
2 2 0 0 0 0
5 5 2 2 1 1
5 6 2 2 1 1
2 2 2 2 0 0
2 2 2 3 0 0
1 1 0 0 1 1
1 1 0 0 1 2
The topology of an equivalent circuit is illustrated in Figure E7.10a.
Page 58
V1
V2
V3
V4
Figure E7.10a Equivalent Network for Topological Observability
7.7.3 Critical Measurements

State estimation is a real time function. Many times due to malfunction of equipment or
communication channels, certain measurements may be lost resulting to a possibly
unobservable system. In this case, it is important to identify the unobservable states and
to develop procedures for the estimation of the observable states. A related issue is to
identify the measurements which if lost will cause a state of the system to become
unobservable. These measurements will be called critical measurements. The treatment
of this subject will also lead to methods of selecting measurements for maximizing the
reliability of the state estimation. In the discussion of these topics it is expedient to study
the properties of the matrix HTH. Specifically, observe that the matrix HTH is
symmetric and positive semidefinite, i.e.
Given any nonzero x > 0
====>
xTHTHx > 0
Consider also the least square solution of equation (1)

HT(z-b) = (HTH)x
This equation can be interpreted as the nodal equations of a circuit as follows
:
(HTH)
x
:
T
H (z-b) :
admittance matrix
nodal voltages
nodal current injections.
Page 59
Since HTH is symmetric, such a network does exist. The construction of this network has
been demonstrated with an example (Example E7.10). When system observability is of
interest the actual values of the entries of the matrix HTH are not important but rather the
topology of the matrix. In this case it suffices to replace the non-zero entries of the
matrix H with arbitrarily selected values such as 1.0.
The identification of the critical measurements can be performed as follows: Consider a
measurement zi. Assume that the matrix HTH has been computed by assigning the value
of 1.0 to the non zero entries of the Jacobian matrix HTH. Observe that the measurement
zi can be expressed as
zi = hiTx
where hiT is the ith row of the Jacobian matrix H. A direct approach to system state
observability will be to remove the measurement zi, recompute the matrix HTH and
perform the observability test. Observe that when the measurement zi is removed, the
Jacobian matrix H is modified to H
0

L
H ' = H hiT

L
0

Upon computation:
HTH = HTH hihiT
The observability test will consist of checking the topology of the matrix HTH. There
may be three cases:
Case 1: The matrices HTH and HTH have the same number of nonzero entries. In this
case, the connectivity of the network of states does not change. Thus, the
measurement zi is not a critical measurement.
Case 2: The matrix HTH has at least two less nonzero entries than matrix HTH. In this
case, the connectivity of the network of states must be examined to determine
observability.
Case 3: The matrix HTH has odd number less nonzero entries than matrix HTH. In this
case, the system is not observable and the connectivity of the network of states
must be examined to determine the unobservable states.
The procedure will be demonstrated with an example.
Page 60
Example E7.11: Consider the system of example E7.10. Determine the type of the
following three measurements: (a) the bus voltage magnitude measurement at bus 4. (b)
The real power flow measurement at circuit 2 - 4. (c) The reactive power flow
measurement at circuit 3 - 2.
Solution: a)
For measurement V4, hT = [0 0 0 0 0 0 1]:
0
0
hh T = 0
0
0
0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 1
Observe that the matrices HTH and HTH have the same number of nonzero entries.
Therefore, the measurement V4 is not a critical measurement.
b) For measurement P24, hT = [0 1 1 0 0 1 1]:
0
0
T
hh = 0
0
0
0
0 0 0 0 0 0
1 1 0 0 1 1
1 1 0 0 1 1
0 0 0 0 0 0
0 0 0 0 0 0
1 1 0 0 1 1
1 1 0 0 1 1
Note that one of the diagonal entries of HTH becomes zero, so P24 is a critical
measurement.
c) For measurement Q32, hT = [0 1 1 1 1 0 0]:
0
0
T
hh = 0
0
0
0
0 0 0 0 0 0
1 1 1 1 0 0
1 1 1 1 0 0
1 1 1 1 0 0
1 1 1 1 0 0
0 0 0 0 0 0
0 0 0 0 0 0
Page 61
Observe that the matrices HTH and HTH have the same number of nonzero entries.
Therefore, the measurement Q32 is not a critical measurement.
7.8 Limitations and Biases of State Estimation

In preceding paragraphs we presented the traditional state estimation process as it is
applied today in modern energy management systems. This application is a special case
of the general state estimation methodology introduced by Gauss and Legendre (around
1800). The basic idea was to fine-tune state variables by minimizing the sum of the
residual squares. This is the well-known least squares (LS) method, which has become
the cornerstone of classical statistics. The reasons for its popularity are easy to
understand: At the time of its invention there were no computers, and the fact that the LS
estimator could be computed explicitly from the data (by means of some matrix algebra)
made it the only feasible approach. Even now, most statistical packages still use the
same technique because of tradition and computational speed. Also, for one-dimensional
problems, the LS criterion yields the arithmetic mean of the observations, which at that
time seemed to be the most reasonable estimator. Afterwards, Gauss introduced the
normal (or Gaussian) distribution as the error distribution for which LS is optimal. Since
then, the combination of Gaussian assumptions and LS has become a standard
mechanism for the generation of statistical techniques.
In a real time environment, state estimation was applied to power systems by Schweppe
and Wildes in the late 1960s [???]. The initial implementation was based on a single
frequency, balanced and symmetric power system under steady state conditions. Over
the past thirty plus years, the basic structure of power system state estimation has
remained practically the same:
Single phase model

P, Q, V measurement set
Non-simultaneousness of measurements
Single frequency model
The above basic structure of the power system state estimation implies the following
assumptions (which in turn result in a biased state estimator):
all current and voltage waveforms are pure sinusoids with constant frequency and
magnitude
the system operates under balanced three phase conditions
the power system is a symmetric three phase system which is fully described by
its positive sequence network
These assumptions introduce deviations between the physical system and the
mathematical model (bias). Mathematically, it is known that the least squares state
estimation procedure is an unbiased estimator if and only if the model is accurate
(exact) and the measuremnent error is statistically distributed. Both of these
Page 62
conditions do not exist in a practical system and have resulted in practical difficulties
manifested by poor numerical reliability of the iterative state estimation algorithm.
Substantial efforts to fine tune the mathematical models in actual field implementations
are required.
To alleviate the sources of error, new measurement systems, power system model and
estimation methods are needed. For example the first assumption can be met by utilizing
synchronized measurements [???]. Synchronization is achieved via a GPS (Global
Positioning System) which provides the synchronizing signal with accuracy of 1 sec.
Assumption 2 can be met by utilizing three phase measurements. Finally assumption 3
can be met by employing full three phase models.
In this section we first discuss the bias resulting from model inaccuracies and then we
discuss the effect of measurment errors. In particular model inaccuracies result from: (a)
unbalanced operating conditions and (b) asymmetries of power system models. Then, a
state estimator is introduced that is based on the following infrastructure:
Synchronized measurements of voltage and current waveforms
Three phase measurements

Use of full three phase models
The state estimation based on this system is not subject to the usual biases of the
traditional state estimation. This state estimation is formulated in its general form that
allows estimation of all three phase voltages resulting in the three-phase state estimation.
7.8.1 Bias From Unbalanced Operation

An actual power transmission system operates near balanced conditions. The imbalance
may be small or large depending on the design of the system. As an example, Figure 7.7
illustrates the three phase voltages and currents on an actual system. Note for example a
10% difference in the currents of Phases A and B of transmission line to GILBOA. The
voltage in this case has only a 0.2% difference between two phases.
Because of imbalance, the measurements may have an error. We represent this as
follows:
z = z t + z
where z t is the true value of the measured quantity (assuming a balanced system), z is
the measurement error due to imbalance, and z is the measurement.
Application of the LS state estimation procedure, assuming no other error sources, yields:
Page 63
x = x t + ( H T WH ) 1 H T Wz
(7.x)
where x t is the true state of the system or the unbiased state estimate, and the second
term is the bias resulting from the imbalance measurement error. Note that the bias from
unbalanced operation depends on the level of imbalance as well as the system parameters
(matrix H).
Figure 7.7. Actual Three Phase Voltages and Currents in FRASER

Substation
7.8.2 Bias From System Asymmetry

An actual power transmission system is never symmetric. While some power system
elements are designed to be near symmetric, transmission lines are never symmetric. The
impedance of any phase is different than the impedance of any other phase. In many
cases, this imbalance can be corrected with transposition. Because of cost many lines are
not transposed.
The asymmetry may be small or large depending on the design of the system. One power
system component that contributes to the asymmetry is the three phase untransposed line.
As an example, Figure 7.8 illustrates an actual three phase line. For the purpose of
quantifying the asymmetry of this line, two asymmetry metrics are defined:
Page 64
S1 =
1 z max z min
2
z1
S2 =
1 y max y min
2
y1
where z1 is the positive sequence series impedance of the line, zmax and zmin are the max
and min series impedances of the individual phases, y1 is the positive sequence shunt
admittance of the line, ymax and ymin are the max and min shunt admittances of the
individual phases.
The above indices provide in a quantitative manner the level of asymmetry among phases
of a transmission line. As a numerical example, these metrics have been computed for
the line of Figure 7.8 and are presented in Figure 7.9. Note that the asymmetry is in the
order of 5 to 6%.
Page 65
1'-1"
4'
4'
11'-6"
7'-7"
9'-6"
9'-6"
17'-0"
7'-7"
9'-6"
58'-0"
Figure 7.8. Typical Transmission Line Design with Asymmetry
Page 66
Asymmetry Factor
0.06
0.04
Series Admittance
0.02
Shunt Admittance
0.0
180
660
1140
1620
2100
Frequency (Hz)
Figure 7.9. Asymmetry Indices of the Transmission Line of Figure 7.8
Because of the presence of non-symmetric components, the state estimate using single
phase measurment set is biased. An estimate of the bias can be computed as follows.
First observe that because of power system component asymmetry, the relationship of a
measurement to the system model will have an error. Specifically:
z = h ( x ) + h ( x )
where h(x) is the function relating the measurement to the state vector assuming
symmetric power system components, h(x) is the difference between the symmetric
model and the asymmetric model.
Now the jacobian matrix of the measurements becomes:
H = H s + H
where H s is the jacobian matrix assuming symmetric power system elements.
Application of the LS state estimation procedure, assuming no other error sources, yields:
x = ( xt + ( H T WH ) 1 H T Wz )( H T WH ) 1 ( I + 2( H T WH )( H T WH ) 1 ) 1 ( H T WH ) (7.x)
where x t is the state of the system assuming a symmetric model, and the other terms
represent the bias resulting from the system asymmetry.
TO BE COMPLETED
7.8.3 Bias From Systematic Measurement Errors

State estimators are based on the assumption that measurement errors are statistically
distributed with zero mean. The traditional implementation of state estimation uses
sensors of V, P and Q. When the sensors are properly calibrated, the measurement error
is very close to meeting the requirements of state estimation. However, with so many
measurements in a practical power system, there are many opportunities to have some
measurements out of calibration contributing always data with systematic error. Recent
Page 67
trends resulted in the use of sensorless technology for power system measurements.
Sensorless technology refers to the use of A/D converter technology to sample the
voltage and current waveforms. Once the sampled waveforms are available, the required
measurements can be retrieved with numerical computations. These systems need
calibration and again the same comments apply for this technology as before.
Independently of the technology used for measurements, it is important to examine
whether there is bias in the measurements. This can be best achieved by examining the
entire measurement channel of a typical power system instrumentation [???]. The major
sources of error (see Figure 7.5a and 7.5b) are (a) the instrument transformers, (b) the
cables connecting the instrument transformers to the sensors or A/D converters and (c)
the sensors or A/D converters. Figure 7.10 illustrates the transfer functions of a typical
instrument transformer. It can be observed that the characteristics of instrument
transformers near the power frequency are flat. One can conclude that for power
frequency measurements, there is no appreciable measurement bias from instrument
transformers. However, cables and A/D converters can introduced appreciable error at 60
Hz. This error will be a systematic error. Figure 7.11 illustrates the transfer function of a
specific A/D converter. Note the magnitude and phase bias even at power frequency. It is
important to note that the measurement bias is dependent upon the design of the A/D
converter. The measurement bias resulting from control cables is variable depending on
the total length of the cables.
Page 68
Figure 7.10. Magnitude and Phase of Frequency Response of a 200

kV/115/65 Potential Transformer
Figure 7.11. Magnitude and Phase of Frequency Response of the PMU1620 Unit
The measurement bias can be corrected with software. Such methods have been
developed [???], but their use in state estimation is very limited. It is important to note
that the above sources of error cannot be corrected with better (more accurate)
instrumentation. To avoid these sources of error, three phase measurements and a three
phase system model is required.
7.8.4 Bias from Measurement Time Skews

The traditional SCADA system is based on sequential polling the Remote Terminal Units
and therefore the measurements received will not be at exactly the same time. If we
assume that the system operates under steady state conditions, the measurements are all
consistent and the time difference in obtaing one measurement versus another will not be
factor. If, however, system conditions change from the time the first measurement was
obtained to the time of the last measurement, then the data will be inconsistent. This
inconsistency will reflect on the performance of the state estimator, it may not converge,
Page 69
it may provide wrong values, etc. It is very difficult to estimate bounds of the effects of
data time skewness on the state estimator.
The problem can be alleviated by utilization of measurement synchronization. The
technology for this does exist today.
7.9 Synchronized Measurements

Synchronization of the measurements can be achieved with the use of a Global
Positioning System (GPS) receiver and appropriate hardware. A conceptual view of such
a system is illustrated in Figure 7.12. The Global Positioning System (GPS) receiver has
the capability to provide a synchronization signal with precision better than 1
microsecong. This time reference allows the measurement of the phase angle of the
fundamental with accuracy 0.02 degrees on a system-wide common reference. The local
system (RTU) uses input signals from existing instrument transformers. The captured
voltage and current waveforms are time-tagged and transmitted to the energy
management system or the master station. Normally, only the first sample needs to be
time tagged. Knowing the sampling rate, all other information can be easily extracted.
Note that at the energy management system, one can collect all the data with the same
time tag. The local systems can be programmed to obtain a set of measueremnts every 5
seconds, staring at exactly the GPS signal that indicates the arrival of a second.
GPS
Antenna
PMU
V,I
GPS Receiver
Filter
A/D
V,I
Filter
A/D
V,I
Filter
A/D
PMU
Global Time Reference

(GPS)
Serial
Communications
To PC
uP
Computer
PMU
Computer
PMU
Computer
Master Station
Monitoring & Analysis
Power System Dynamics
Geomagnetically Induced Currents
Harmonic Generation and Propagation
Page 70
Control
Center
Computer
Figure 7.12. The Hardware Platform Used by the GPS-Synchrnized

Measurements
Use of synchronized measurements simplifies the state estimation problem. Consider for
example the case in which all measurements are GPS-synchronized. We assume here that
the GPS-synchronized measurements have a time precision of better than one
microsecond, the capabil;ity of the GPS technology. Note that in practice this precision
may not be achievable due to other sources of error (see for example reference [***]).
Consider a set of GPS synchronized measurements, ~
z . These measurements are typically
in the form of voltage phasor measurement and electric current phasor measurement. Let
~
x be the vector of the bus voltage phasors. The measured voltage and electric current
phasors will be functions of the voltage phasors, ~
x . In general, these functions will be
linear. We write:
~
z = Y~
x + ~
where Y is a constant matrix and ~ is the measurement error.
The state estimation problem is now formulated as follows:
Min
~i
J =
i i
= ~ H W~
where: ~ = ~
z Y~
x
The above state estimation problem is linear. The solution is obtained by first separating
the equations into real and imagineray parts yielding:
Min J = rT W r + iT W i
Subject to:
r = z r Y r x r + Yi x i
i = z i Y r x i Yi x r
where the subscript r indicates real and the subscript i indicates imaginary.
Upon substitution and observing that at the solution the derivatives of the function J with
respect to the state variables will vanish:
dJ
= 2Y rT W (z r Y r x r + Yi x i ) 2Yi T W (z i Y r x i + Yi x r ) = 0
dx r
dJ
= 2Yi T W (z r Y r x r + Yi x i ) 2Y rT W (z i Y r x i + Yi x r ) = 0
dx i
Solution of above equations for the state vector provides:
Page 71
xr A
x = B
i
B Y rT Wz r + Y i T Wz i
A Y rT Wz i Y i T Wz r
where:
A = Y rT WYr + Yi T WYi
B = Y rT WYi + YiT WYr
The state estimation problem with GPS synchronized measurements will be demonstrated
with an example.
Example E7.x: Consider the two-bus, single transmission line system of Figure E7.x.
The voltage and electric current at the two ends of the line are measured with GPS
synchronized equipment.
0
0
~
~
V1 = 1.010e + j 61.46
pu
V 2 = 0.985e + j 55.09
pu
0
0
~
~
I 1 = 1.456e + j 73.95
pu
I 2 = 1.485e j143.96
pu
The measurement accuracy is 1% for the magnitude and 0.02 degrees for the phase.
Compute the state estimate for this system.
~
V1 =V1e j1
-j15.0
~
V2 = V2e j2
Solution: The model equations are:
7.10 Formulation of the Three-Phase State Estimation

The formulation is presented with the following postulated model:
z = h(x) +
(7.x)
where z is a vector of three phase measurements; x is a vector of the state (three-phase

state); is a vector of error; h is a vector function depending on the system modeling.
The three-phase state estimator is formulated by selecting the three-phase state, the threephase measurements and the three-phase system model. These are described next.
7.10.1 Three-Phase System State

Similar to the conventional state estimation, the voltages are defined as system state. The
difference is that we use node (phase) voltages versus bus voltages in the conventional
state estimation. Each bus contributes three complex voltages, one for each phase:
Page 72
~
~
~
V a = V ar + jVai , Vb = Vbr + jVbi , Vc = Vcr + jVci
(7.x)
The set of variables above are state variables, one set for each bus. Here the rectangular
coordinate system is used for convenience.
The number of state variables for a bus are 6n, where n is the number of buses.
7.10.2 Three-Phase Measurements

The measurement set consists of synchronized sampled waveforms. The synchronization
ensures the exact time of the sample with accuracy of 1 microsecond. From the sampled
waveforms, the quantities zreal,i and zimag,i are computed which constitute the
measurements in accordance to the following postulated model:
z current , a = I ar + jI ai ,
z current ,b = I br + jI bi ,
z voltage, a = V ar + jV ai ,
z voltage,b = Vbr + jVbi ,
z current ,c = I cr + jI ci
z voltage,c = Vcr + jVci
Real and reactive power measurements should not be used in the presented three-phase
state estimator for the following reason: since voltage and current is measured, and since
the real power and reactive power is derived from these measurements, all the
information needed is included in the V and I waveform measurements. It certainly does
not mean that the real and reactive power measurements can not be processed.
The measurements, z, are related to the state variables with the equations below.
z current = Yx +
z voltage = Tx +
(7.x)
(7.x)
where all variables are complex variables. Matrix Y is an admittance matrix of proper
dimensions. T is a matrix whose entries are either 1 or 0. If the measured state variables
are ordered first in x in the same order as in z, then matrix T has the form I | 0 with
identity matrix I and zero matrix 0 having proper dimensions. Above equations can be
lumped into the one equation below:
~ ~
~
z = H~
x+r
(7.x)
Note that equation (7.x) is linear. Thus, the least square estimation requires only one
iteration(direct solution). This advantage comes from the use of the rectangular
coordinate system.
7.10.3 Least Squares Estimation

The least square estimation is formed as an optimization problem:
Page 73
Minimize: J = ~
r H W~
r
~
~
~
Subject to:
r = z H~
x
(7.x)
where superscript H means Hermitian transpose. By separating the complex variables

into real and imaginary parts, problem (7.x) is transformed to :
Min J = rT W r + iT W i
Subject to:
rr = z r H r x r + H i xi
ri = z i H r xi H i x r
(7.x)
where the subscript r indicates real and the subscript i indicates imaginary.
Upon substitution and observing that at the solution the derivatives of the function J with
respect to the state variables will vanish:
dJ
= 2 H rT W (z r H r x r + H i x i ) 2 H iT W (z i H r x i + H i x r ) = 0
dx r
dJ
= 2 H iT W (z r H r x r + H i x i ) 2 H rT W (z i H r x i + H i x r ) = 0
dx i
Solution of above equations for the state vector provides:
xr A
x = B
i
B H rT Wz r + H iT Wz i
A H rT Wz i H iT Wz r
(7.x)
where:
A = H rT WH r + H iT WH i
B = H rT WH i + H iT WH r
7.10.4 Three-Phase Power System Model

For any estimation problem, a model must be known which relates the measurements to
the state of the system. This task can be achieved by considering individual system
components. The major system components are transmission lines, transformers,
generators, etc. All components are classified into linear and nonlinear. If current
waveform meters are placed at all interfaces with nonlinear devices, the model relating
measurements to the state of the system is linear [???]. For this reason, only linear
devices need to be modeled in the presented three-phase state estimator. The postulated
model for each linear device is:
Page 74
~
~ ~
I d = Yd Vd I sd
where
(7.x)
~
I d are the currents at the terminals of the device
~
Vd are the voltages at the terminals of the device
~
I sd are equivalent current sources
Yd is the admittance matrix of the device

~
Note that for passive devices (i.e. a line), I sd is zero.
The admittance matrix Yd contains the modeling detail for a device. Consider for
example a transmission line. Rigorous modeling of a line yields a matrix Yd which
corresponds to a nonsymmetric system, i.e. each phase exhibits different admittance.
7.10.5 Observability Analysis Three Phase State Estimation

In any state estimation problem, the question of observability is very important. For the
three-phase state estimator it is extremely important because the implementation of the
three-phase state estimator may be initially with a limited set of measurements. In this
case, it is important to know the extent of the system which can be observed. The
number of states to be estimated in the three-phase SE is much higher than the number of
states in the conventional state estimation. Proportionally, the number of measurements
needed is much higher. The computational effort for observability analysis increases
quickly with the problem size. The size of the state in the three-phase SE is 3 times the
size of the state in SE. Thus, the computational effort for three-phase observability
analysis is more than 3 times the required computational effort for the equivalent single
phase analysis. However, we introduce a condition for measurement selection (described
next). When this condition is met, observability algorithms developed for the
conventional state estimation problem can be readily used. The condition is:
Condition: The three-phase measurements are unbiased for three phases (Phase
Unbiased Measurement).
Phase Unbiased Measurement means that if there is a measurement for phase A of bus i,
then there are measurements for phases B and C of bus i respectively. The assumption of
Phase Unbiased Measurements simplifies the observability analysis from three phase
case to single phase case, e.g. phase A. This condition is not always true, e.g. in the
presence of neutral current measurements. However, the whole measurement set can be
divided into two parts, (a) the phase unbiased set, and (b) the phase biased set.
Accordingly, the observability analysis can be separated into two steps. Step 1 only
processes phase unbiased measurements.
Page 75
Step 1: Under the above condition, Matrix
H=
analysis, where subscript A means phase A.

decomposed into:
~
~
~
I1, A
I 1, A
I1, A
~
T
Z 1, A = ~ ,
~x1, A , and ~ = MYA
H
=
x1, A
V1, A
v~1, A
Z% 1,A
x% 1,A
is used to perform observability
The vector
Z% 1, A ,
and matrix H are
~
x1, A
The definition for matrices M, Y, and A can be found in Reference [???]. By proper
v%
ordering of states and measurements: 1,A = In 1 x n 1 0n1 x (n -n1) . In this case,
x% 1,A
H =
M ' Y' A ' T
where
M
M '=
0
0
,Y ' =
I n1 x n1
Y
0
0
AT
T
,and
A
=
'
I n1 x n1
I n1 x n1 | 0 n1 x (n - n1)
This is equivalent to changing all buses with voltage measurements into a topology with
current flow measurement from the node under consideration to ground. Note that this
procedure permits the use conventional topological methods to perform the step 1 of
observability analysis for harmonic state estimation.
After Step 1, the system is decomposed into three parts as shown in Figure 7.13.
Semi-Observable
Island
Observable
Island
Remaining
Part
Semi-Observable
Island
Figure 7.13. System Partition after Step 1 Processing
Page 76
The observable island is an island such that the states at all nodes in this island can be
estimated. The semi-observable island is an island such that if one phase unbiased
voltage measurement is added to any bus in this island, this island can become an
observable island or be merged to the existing observable island.
The remaining part of the system will be unobservable if there is no phase biased
measurement. However, with phase biased measurements, it may be possible to expand
the observable island. This is beyond the scope of this text.
7.10.6 Quality of Three-Phase State Estimator

The overall accuracy and performance of the three-phase state estimator can be evaluated
using the concept of the confidence level as in the case of the traditional state estimator.
Again one has to identify the number of states, the number of measurements for the
purpose of computing the degrees of freedom. Then at the state estimate, the value of the
objective function should be computed. The confidence level can be obtained from Table
7.2. For additional information, refer to section 7.x.
7.10.7 Discussion of the Three-Phase State Estimator

The conventinal state estimation has inherent biases resulting from biases in the
measurements and biases in the power system model (imbalance and asymmetry of
component models). We presented equations for quantifying the biases in conventional
state estimation. In addition, we presented a new state estimation paradigm that is based
on synchronized measuremements, three phase instrumentation and asymmetric three
phase power system model. This state estimator does not exhibit the biases discussed
earlier and it is direct, i.e. it does not require an iterative algorithm to obtain the solution.
Today, the technology exists for implementing this approach. As technology advances,
and the required hardware cost is reduced, this approach becomes attractive. Yet, we
have not seen wide spread efforts to implement three-phase state estimators. Meanwhile,
hybrid state estimators, i.e. estimators that are partially based on the conventional
approach and partially on synchronized three phase measurements are emerging. Hybrid
approaches are not discussed in this text.
7.11 Hybrid Three-Phase State Estimator

Uniform availability of synchronized measurements allows for a linear estimator which
provides a direct (non-iterative) computation of the system state estimate. While the
technology of synchronized measurements is readily available, large scale deployment of
this technology and utilization in the state estimation is not in the immediate future.
However, as this technology is deployed at specific locations, the situation is generated
where some measurements are not synchronized and some are. In this case, the question
is: can we take advantage of the existence of the synchronized measurements. In this
Page 77
case, a hybrid state estimator is emerging. We discuss in this section the hybrid state
estimator.
TO BE CONTINUED
7.12 Summary and Discussion

The operating state of a power system is best computed with the aid of a state estimator
which uses sets of redundant measurements. The state estimator provides:
The best estimate of the operating state

It quantifies the quality of the best estimate with a number of indices, such as (a)
confidence level and (b) standard deviation of the computed operating state.
In a real time environment, it is possible that some of the data (measurements) are
inaccurate or bad. In the presence of data redundancy, it is possible to detect (with the
use of the chi-square test), identify (with a combination of consistency checks and
hypothesis testing), and reject bad measurements. This capability is possible whenever
(a) there is enough redundancy of measurements and (b) the redundant measurements are
well distributed throughout the system. Hypothesis testing is computationally demanding
task. The sequential state estimator provides the algorithm for fast hypothesis testing.
In general, the measurements are selected in such a way that the operating state of the
system is observable. However, there are instances where rejection of bad data or local
malfunction of the SCADA system may result in a set of measurements from which the
state is not observable. Techniques for detecting unobservable states have been
presented. In case the system is unobservable, the introduction of pseudo-measurements
provides the means to perform the state estimation and avoid the ill-conditioning of the
problem in this case.
Finally, the assumptions of the traditional state estimation have been reviewed. While
these assumptions are reasonable, nevertheless they introduce biases into the state
estimation problem. These biases have been discussed and their effect has been
quantified. The biases are avoided if three phase estimation methods are used, i.e. the
model of the system is a full three phase model and the measurement set consists of three
phase measurements.
The introduction of synchronized measurements provides the capability to design a noniterative state estimation algorithm.
The state estimator has many other capabilities and chanllenges. The methodologies
canm be used for remote calibration of meters, parameter and topology estimators and
dynamic estimatyion of system oscillations.
Page 78
7.13 Problems
Problem P7.1: Solve the following set of overdetermined equations using least square
estimation:
x 1 + 2 x 2 = 3.05
x1 + x 2 = 2.08
x1 + x 2 = 0.20
2 x 1 + 0.5 x 2 = 2.39
x1 2 x 2 = 0.95
Solution:
x = (H T H ) H T b
1
where:
2
1
3.05
1
2.08
2
H = 1 1 , and b = 0.2
2 0.5
2.39
1 2
0.95
Upon substitution,
0.9669
x =
1.0252
Problem P7.2 Consider the simplified power system of Figure P7.2 with the indicated
measurements. At a specific instance of time, the following readings are taken:
z 1 = V1 = 1.0 pu , 1 = 0.045 pu
z 2 = V 2 = 0.98 pu , 2 = 0.0 pu
z 3 = P12 = 1.65 pu, 3 = 0.045 pu

z 4 = P21 = 1.62 pu , 4 = 0.045 pu
z 5 = Q21 = 0.23 pu, 5 = 0.045 pu
Note that the standard deviation of the error of the second measurement (z2) is zero,
while the standard deviation of the remaining four measurements is 0.045 p.u.
Page 79
a) Compute the state estimate (two iterations will suffice).

b) Compute the covariance of the state estimate.
c) Compute the probability that the state estimate is statistically correct (confidence
level).
~
V1 =V1e j0
-j15.0
~
V2 = V2e j2

MW Flow Measurement
Figure P7.2
Solution:
(a) The state is:

V
x = 1
2
The measurements and the measurement model are:

1 .0
1.65
,
b=
1.62
0.23
V1
14.7V1 sin 2
,
h(x ) =
14.7V1 sin 2
14.406 14.7V1 cos 2
W = 2 I = 493.82 I
The algorithm is:
x +1 = x ( H T WH ) 1 H T W (h( x ) b)
Staring from flat start:
Iteration 1:
0.99566
x 1 =
0.11122
Iteration 2:
Page 80
1.00183
x 2 =
0.11125
(b) The information matrix is:
9.2707 0.5136
( H T WH ) 1 = 10 6
0.5136 4.7258
(c)
h ( x ) bi
J ( x ) = 1 = i
i
i =1
4
= 0.2238
Degrees of freedom=4-2=2
From table: Prob.=1.0-0.1053=0.8947
Problem P7.3 Consider the simplified two bus electric power system of Figure P7.3.
The following are measured:
z 1 = V1 = 1.01 pu , 1 = 0.0 pu
z 2 = V 2 = 1.0 pu , 2 = 0.0 pu
z 3 = P12 = 1.40 pu, 3 = 0.03 pu

z 4 = Q12 = 0.10 pu , 4 = 0.03 pu
z 5 = P21 = 1.41 pu, 5 = 0.03 pu
z 6 = Q21 = 0.10 pu, 6 = 0.03 pu
Compute the weighted least squares estimate of the system state V1 , V 2 , 2 assuming the
indicated standard deviations of the measurement error. Note that the first and second
measurements have a zero standard deviation which means that the error of these
measurements is negligible.
Compute the probability of goodness of fit (chi-square test).
In case of bad data, identify the bad datum or data by hypothesis testing.
Page 81
-j16.0

MW Flow Measurement
Figure P7.3
Solution: Since the standard deviation for the voltage magnitudes is zero, these variables
are known with certainty. Thus the state of the system consists of one variable only, the
phase angle. The state estimation problem is defined with:
1.40
0.10
,
b=
1.41
0.10
16.16 sin 2
16.3216 16.16 cos

2
,
h( x ) =
16.16 sin 2
16.0 16.16 cos 2
1
W = diag 2 = diag (1,111.1)

Starting from zero, the first iteration yields:

1.40
0.0616
= 0.08694
2 = 0 (522.29 ) 1 [ 16.16 0.0 16.16 0.0]
1.41
0.06
The second iteration yields:
0.00323
0.12264
= 0.08672
2 = 0.08694 (522.28) 1 [ 16.0989 1.4032 16.0989 1.4032]
0.006766
0.001039
Chi square test:
0.000414 0.12232 0.0104 0.00072
J =
= 16.745
+
+
+
0.03
0.03 0.03 0.03
Pr = 0.0
Page 82
This indicates presence of bad data.

Hypothesis testing: Remove data and retest. First candidate is the second measurement.
1.40
b = 1.41 ,
0.10
16.16 sin 2
,
h( x ) = 16.16 sin 2
16.0 16.16 cos 2
1
W = diag 2
= diag (1,111.1)
Starting from 2 = 0.08672 , the first iteration yields:
2 = 0.08672 (520.32)
0.000414
[ 16.0992 16.0992 1.3995] 0.0104 = 0.08705
0.00072
Chi square test:

0.00499 0.00501 0.00119
J =
= 0.05735 , degrees of freedom=2
+
+
0.03 0.03 0.03
2
Pr = 1.0 0.025 = 0.975
Conclusion: Measurement 2 is a bad datum.

Problem P7.3 Consider the simplified two bus electric power system of Figure P7.3.
z 1 = V1 = 1.01 pu , 1 = 0.01 pu
z 2 = V 2 = 1.0 pu , 2 = 0.01 pu
z 3 = P12 = 1.40 pu, 3 = 0.03 pu

z 4 = Q12 = 0.10 pu , 4 = 0.03 pu
z 5 = P21 = 1.41 pu, 5 = 0.03 pu
z 6 = Q21 = 0.10 pu, 6 = 0.03 pu

Compute the weighted least squares estimate of the system state V1 , V 2 , 2 assuming the
indicated standard deviations of the measurement error.
Compute the probability of goodness of fit (chi-square test).
In case of bad data, identify the bad datum or data by hypothesis testing.
Page 83
-j16.0

MW Flow Measurement
Figure P7.3
Solution: The model equations are:
To be continued.
Problem P7.4: Consider an off-nominal tap transformer that belongs to a power system.
The system is illustrated in Figure P7.4. To estimate the value of the tap ratio t, five
measurements are taken as follows:
z 1 = V1 = 0.933 pu , 1 = 0.0 pu
z 2 = V 2 = 0.917 pu , 2 = 0.0 pu
z 3 = P12 = 0.245 pu, 3 = 0.025 pu

z 4 = P21 = 0.240 pu , 4 = 0.025 pu
z 5 = Q21 = 0.170 pu, 5 = 0.025 pu
a) Develop an iterative algorithm for the estimation of the transformer tap setting t.
b) Perform one iteration of above algorithm.
Page 84
Power System
1:t
-j8.0 pu

MW Flow Measurement
Figure P7.4
Solution: The problem is formulated with:
z 3 = 6.8445t sin 2
z 4 = 6.8445t sin 2
z 5 = 6.7271 6.8445t cos 2
t
x=
2
+1
x = x ( H T WH ) 1 H T W (h( x ) b)
where:
6.8445 sin 2 6.8445t cos 2
H = 6.8445 sin 2
6.8445t cos 2 ,
6.8445t sin 2
6.8445 cos 2
0
0
1600
W= 0
1600
0
0
1600
0
1.0
x0 =
0.0
1.0 0.0420 0.9580
x1 =
=
0.0 0.0354 0.0354
Problem P7.5: Consider the simplified two bus electric power system of Figure P7.5.
V1 = 1.01 pu , V 2 = 1.00 pu , P12 = 1.40 pu , P21 = 1.41 pu , Q 21 = 0.10 pu
Compute the weighted least squares estimate of the system state assuming the following
weights: w1 = , w2 = 2500, w3 = 625, w4 = 625, and w5 = 625. Two iterations will
suffice. Note that the first measurement has an infinite weight which means that the error
Page 85
of this measurement is negligible. Hint: determine the minimum number of state

variables and then develop the least squares algorithm.
Compute the probability of goodness of fit (confidence level).
-j16.0

MW Flow Measurement
Figure P7.5
Solution: The state of the system is: x T = [V2
2 ]
Equations:
h1 ( x ) = V 2
h 2 ( x ) = (16.16 )V 2 sin ( 2 )
h3 ( x ) = (16.16)V2 sin( 2 )
h4 ( x ) = 16.0V22 (16.16)V2 cos 2

Algorithm
x +1 = x ( H T WH ) 1 H T W (h( x ) b)
Two iterations starting from (1,0):
1
V 2
1.00372
=

0.08694
2
2
V 2
0.99940
= 0.08706
Sum of normalized residuals
Page 86
J = (2500)( 0.00000601) + (625)(0.00499) + (625)(0.005) + (625)(0.000244) =0.031

3
2
Degrees of freedom = 4-2=2

Chi-square test: Pr=0.93
Problem P7.6: Consider the simplified electric system of the Figure P7.6. At a certain
instant of time a set of measurements are taken as it is indicated in the Figure.
a) Determine whether the system is obervable using topological observability.
b) Determine all the critical measurements it there are any.
c) Suppose that the measurement P21 is removed (i.e. it has ben identified as a bad
measurement). Is the system observable? Are there any critical measurements?
G1
G2
MW Flow Measurement
kV Measurement
Figure P7.6
Problem P7.7: Consider the following set of overdetermined equations:
x1 + 2 x2 = 3.05
x1 + x2 = 2.08
x1 + x2 = 0.05
2 x1 + 0.5 x2 = 2.39
x1 2 x2 = 0.95
3 x1 x2 = 1.38
Note that the values on the right hand side are measurements. It is known that the
standard deviation of the measurement error is 0.1.
Page 87
(a) Compute the best estimate of the variables x using weighted least square
approach.
(b) Compute the confidence level (chi-square test).
(c) It is suspected that the last measurement is a bad measurement. Perform a
hypothesis testing (that the last measurement is a bad measurement). Use the
sequential state estimator to perform the required calculations. (Hint: the new
information matrix will be: A' = H TWH hhT 2 = A hhT 2 , where A is the
information matrix for the initial problem, A is the information matrix for the
state estimation problem with the last measurement removed, and h is the
coefficients of the last model equation).
Solution:
x = (H T WH ) H T Wb
1
where:
2
1
3.05
1
2.08
1
1 1
0.05
H =
, and b =
2
0
.
5
2.39
1 2
0.95
3 1
1.38
Upon substitution,
0.8903
x =
1.0423
(b) Confidence level:
2
r
J = i = 16.7043, = 6 2 = 4, Pr 0.0
i =1 0.1
6
(c) Hypothesis testing: remove sixth measurement and repeat state estimation.
Method 1: Direct state estimation:
0.9877
1
x = (H 'T W ' H ' ) H 'T W ' b ' =
1.0085
2
5
ri
J =
= 2.4253, = 5 2 = 3, Pr 0.49
i =1 0.1
Page 88
Method 2: Use sequential state estimator.

Let: A = H T WH from part (a). Then
A' = H 'T W ' H ' = A hh T 2 ,
x ' = x +
3
where: h = , 2 = 0.1 2 = 100
1
b6 h T x
hT 2
Computations:
1.6956
0.5875
= A 1 h = 10 3
h T x = 1.6286 b6 h T x = 0.2486
h T = 0.0056743 h T 2 = 0.0043
b6 h T x
0.09744
=
T
2
h
0.03376
0.9877
x ' =
, note same as in part (a).
1.0085
Problem P7.8: Present day technology makes it possible to obtain synchronized
measurements at remote parts of a power system with accuracy of less than one
microsecond using the GPS (Global Positioning System). Synchronized measurements
provide the phasor (magnitude and phase) of the measured quantity. Assume that in a
simplified one line power system synchronized measurements provided the following
measurements
0
0
~
~
V1 = 437.621e + j 61.46 kV
V 2 = 432.508e + j 55.09 kV
0
0
~
~
I 1 = 0.6424e + j 73.95 kA
I 2 = 0.6601e j143.96 kA
~
I1
3.18 Ohms
~
V1
j76.72 Ohms
-j2040 Ohms
~
V2
~
I2
Page 89
The voltage meters have an accuracy of 2% on a 440 kV scale and the electric current
meters have an accuracy of 1% on a 1500 A scale. Formulate the state estimation
problem using Cartesian coordinates, i.e. assume that the state of the system is defined in
T
terms of the real and imaginary part of the bus voltages, i.e. x = [x1 x 2 x 3 x 4 ] ,
~
~
V1 = x1 + jx 2 , and V 2 = x 3 + jx 4 .
a) Compute the best estimate of the state in the least squares sense.
b) Compute the confidence level.
c) Compute the standard deviations of the state variables.
Solution: The admittances are:
1
= 0.00054 j 0.01301 S
g + jb =
3.18 + j76.72
1
jb ' =
= j0.00049 S
j 2040.0
(a) The electric currents as functions of voltages are:
~
~
~
I 1 = (0.00054 j0.01252)V1 (0.00054 j0.01301)V2
~
~
~
I 2 = (0.00054 j0.01301)V1 + (0.00054 j0.01252)V2
The measurements as a function of the state variables are:
~
Re V1 = 209.08 = x1
~
Im V1 = 384.44 = x 2
~
Re V2 = 228.62 = x 3
~
Im V2 = 367.15 = x 4
~
Re I 1 = 0.2444 = 0.00054x1 0.01252x 2 0.00054x 3 + 0.01301x 4
~
Im I 1 = 0.5396 = 0.01252x1 + 0.00054x 2 + 0.01301x 3 0.00054x 4
~
Re I 2 = 0.4738 = 0.00054x1 + 0.01301x 2 + 0.00054x 3 0.01252x 4
~
Im I 2 = 0.3642 = 0.01301x1 0.00054x 2 0.01252x 3 + 0.00054x 4
z = Hx
209.083
384.443
1
T
T
kV
x = (H WH ) H Wb =
247.519
354.679
{
{
{
{
{
{
{
{
}
}
}
}
}
}
}
}
(b) The objective function computed at the estimate is:

2
z hiT x
= 0.22014
J = i
i =1 0.02 z b
The confidence level computed from the table with 4 degrees of freedom (8
measurements 4 states=4), is approximately 0.994.
8
(c) The standard deviation of the state variables is computed as follows:
Page 90
0
0
1.0 0
0 1.0 0
0
=
0 1.0 0
0
0
0 1.0
0
I = (H T H )
x = 1.0, x = 1.0, x = 1.0, x = 1.0

1
and
V = x2 + x2 = 1.0 ,
1
V = x2 + x2 = 1.0
Problem P7.9: The purpose of this problem is to investigate the impact of phasor
measurements on the state estimation. Consider problem 7.2. Assume that additional
measurements of the voltage phase angle using synchronized measurements were
obtained. The measurement, in effect, adds the following datum 1 2 = 6.4 0 . The
standard deviation of the measurement error is less than 0.080. Note that in this case, the
measurement set consists of synchronized and non-synchronized measurements (hybrid
system).
a) Compute the state estimate;
b) Compute the confidence level;
c) Compute the standard deviation of the state estimate.
Solution: The added measurement is:
b5 = 0.1117 rads, 5 = 0.0014

(a) The new state estimate is:
1.0018
x =
0.1115
(b) The confidence level is computed as follows.
h ( x ) bi
J ( x ) = 1 = i
i
i =1
5
= 0.5287
Degrees of freedom=5-2=3
From table: Prob.=1.0-0.09=0.91
(c) The covariance matrix is:
Page 91
0.4107 0.0110
Cov ( x ) = 10 5
0.0110 0.1014
Thus:
V = 0.002, = 0.001
1
Problem P7.10: The purpose of this problem is to investigate the impact of phasor
measurements on state estimation. Specifically, we consider a set of synchronized and
non-synchronized measurements taken on a system and we seek the state estimate by
least squares approach. Consider the simplified two bus system of Figure P7.10. The
following measurements are taken.
Measurement
number
1
2
Measured
Quantity
Measured
Value
Measurement
Error
0.02/0.0004
~
V1 = V1e j1
~
V2 = V2 e j 2
0.98e j1.39 pu
0.02/0.0004
P21
Q21
-1.73 pu
-0.19 pu
0.04
0.04
3
4
1.0e j1.51
pu
Note that the first two measurements are phasor measurements (complex), i.e. they
are synchronized measurements.
a)
b)
c)
d)
Formulate the state estimation problem for above system. Hint: Each synchronized
measurement can be considered as two measurements: measurement of the rms value
and of the phase angle. Also define the state of the system as x1 = V1 , x 2 = V 2 and
x3 = 1 2 .
Compute the state estimate.
Compute the confidence level.
Compute the standard deviation for all states.
~
V1 =V1e j0
-j15.0
~
V2 = V2e j2
Figure P7.10
Page 92
Solution: (a) The problem is formulated as follows:

z1 = 1.0 = V1 + 1
z 2 = 0.98 = V 2 + 2
z 3 = 0.12 = + 3
z 4 = 1.73 = 15 .0V1V 2 sin + 4
z 5 = 0.19 = 15V22 15.0V1V2 cos + 5

5
hi ( x ) z i
h ( x) zi
0.12
J = i
+
+
0.02
0.04
0.0004
i =1
i =4
2
Min
(b) The solution is obtained with:
x +1 = x ( H T WH ) 1 H T W (h( x ) b)
1
H =
0
15V 2 sin
15V 2 cos
W = diag (2500
0
1
0
15V1 sin
30V 2 15V1 cos
2500 6250000
15V1V 2 cos
15V1V 2 sin
0
625 625 )
Start from:
1.0
0
x = 0.98
0.12
2
h ( x) zi
= 0.0099 . From tables Prob=0.995 (degrees of freedom: 5-3=2)
(c) J = i
i
i =1
(d) The information matrix is:

5
I = (H T WH )
0.1925 0.0445
0.1898
0.2030 0.0470
= 10 3 0.1925
0.0445 0.0470 0.0181
Thus:
V1 = 0.0138, V2 = 0.0142, = 0.00043
Page 93
Problem P7.11: GPS-synchronized measurements are characterized with high timing

precision (or equivalently phase measurement precision) and usual magnitude precision.
Consider the simplified electric power systm of Figure P7.11. The parameters of the
circuit are: g = 0, b = 0.02 S , bs = 0 . Assume that the following GPS-synchronized
measurements were taken.
0
~
V1 = 134.5e + j10.02 kV
0
~
I 1 = 0.551e + j 0.11 kA
~
V1
0
~
V2 = 129.6e j1.98 kV
0
~
I 2 = 0.558e j179.86 kA
~
I1
jbS
jb
~
I2
~
V2
jbS
The voltage meters have an accuracy of 2% on a 440 kV scale, the electric current meters
have an accuracy of 1% on a 1500 A scale and the time precision is better than 1
microsecond which translates as a precision of phase angle of 0.02 degrees. Formulate
the state estimation problem using Cartesian coordinates, i.e. assume that the state of the
system is defined in terms of the real and imaginary part of the bus voltages, i.e.
~
~
T
x = [x1 x 2 x 3 x 4 ] , V1 = x1 + jx 2 , and V 2 = x 3 + jx 4 .
a) Compute the best estimate of the state in the least squares sense.
b) Compute the confidence level.
c) Compute the standard deviations of the state variables.
d) The above approach does not recognize the preciosion of the phase measurements. To
account for this precision, reformulate the state estimation problem in polar coordinates,
T
i.e. the state variables will be: x = [V1 V2 1 2 ] (only three state variables). Repeat
(a), (b) and (c) for this formulation.
e) State your observations.
Example E7.7. Consider the system of Figure E7.7. Five measurements are taken as
follows: V2 = 0.99, Pl2 = 1.4, Ql2 = 0.15, P2l = -1.46, Q21 = 0.01.
Page 94
-j15.0

MW Flow Measurement
Figure E7.7 A Simple Two Bus Power System
All measurement instruments are known to have an accuracy of 1%. The voltage at bus 1
is 1.0 p.u. with absolute certainty. Perform bad data identification.
Page 95

PSA Chapter 07

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

PSA Chapter 07

Încărcat de

Drepturi de autor:

Formate disponibile

DRAFT and INCOMPLETE

Power System Modeling, Analysis and Control

7.5 Detection and Identification of Bad Data____________________________________ 35

7.6 Sequential State Estimators ______________________________________________ 48

7.8 Limitations and Biases of State Estimation __________________________________ 62

7.9 Synchronized Measurements _____________________________________________ 70

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

7.10.4 Three-Phase Power System Model ____________________________________________ 74

7.11 Hybrid Three-Phase State Estimator ______________________________________ 77

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

Gross Bad Data

7.2 SCADA System

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

A simplified view of a SCADA system is illustrated in Figures 7.2a and 7.2b.

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

Figure 7.2a Simplified View of a SCADA System - Survey Points

Figure 7.2b Simplified View of a SCADA System - Configuration

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

7.3 System Network Configurator

Figure 7.3a Breaker Oriented Model Prestored Network Data

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

7.4 State Estimation

Voltage Magnitude Measurement

Figure 7.4 A Simplified Example System for Power System State

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

Formulate the state estimation problem as an overdetermined set of nonlinear equations.

Voltage Magnitude Measurement

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

The state estimation problem is formulated as the following overdetermined set of

0.01 = 15.0V22 15.0V1V2 cos 2

Compute the vector x.

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

Compute the vector x.

Figure 7.5a Physical System for Analog Measurements - Use of

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

Figure 7.5b Physical System for Analog Measurements - Transducer-less

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

7.4.1 Least Squares Solution

In this case, the problem is stated as follows:

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

Minimize J = ( Hx b) T W ( Hx b) , for the linear case

Upon solution of last equation for the state vector x:

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

Upon substitution of the b vector:

The last equation is generalized into the following iterative equation:

where H is the matrix h ( x )/ x computed at x = x . This is the Jacobian of the vector

Copyright A. P. Sakis Meliopoulos 1990-2006

Power System Modeling, Analysis and Control: Chapter 7, Meliopoulos

The Jacobian matrix H is:

The computations follow: