Observability analysis using linear programming
The set of available measurements are used by the state estimator to estimate the system state. As the measurement and the location of the system is given, the network observability analysis can be determined if the system is observable or not. If the system is not observable then additional meters may have to be placed on the particular location to make the system observable. Prior to the state estimation the observability analysis should be done to ensure that there are enough set of measurements.
Different types of errors and change in topology sometimes may lead to the case where the entire system cannot be estimated. Then the system may contain some isolated observable island. The observability analysis detects all the observable island before the execution of the state estimator.
The methods for the network observability analysis can be divided into two : numerical and topological.
http://txspace.tamu.edu/bitstream/handle/1969.1/1475/etd-tamu-2004C-ELEN-Ding.pdf?sequence=1
1.2 State Estimation, Observability Analysis and Meter Placement
Electric power system state estimation [3-8] was introduced by Fred Schweppe of MIT in 1969. The operating state of a power system is determined by the state
3
estimation (SE) function using a redundant set of real-time measurement data. The SE function is the basis for all advanced applications of an Energy Management System (EMS). The results of state estimation are used to compute various estimates for the line flows, losses and net bus injections. The deregulation of the electric power industry has transformed state estimation from an important application into a critical one [5]. Many critical commercial issues in the power market, such as congestion management, need to be formulated and addressed based on a precise model of the power system, which is derived from the state estimation process. Failure to obtain these quantities in real-time or miscalculating them, should be avoided in order to ensure proper accounting of power transactions as well as for system security. This implies that the state estimator should be made very robust against the topology changes (including branch outages and bus splitting) and temporary loss of measurements or remote terminal units (RTU). Observability analysis is another important procedure closely related to state estimation. State estimation will not be possible if there are not enough measurements. A system is determined to be observable if all the state variables (bus voltage magnitudes and relative phase angles) can be estimated using the available measurements. Various methods proposed for network observability analysis have been well documented in the literature [9-14]. If the power system is unobservable there will be a need to install new meters to make the system observable. Meter placement requires making decisions as to where and what types of meters should be placed. Many approaches are based on Integer Programming (IP) and heuristic solution techniques [15-19].
When installing a new state estimator or upgrading an existing one, measurement configuration will have to be considered in order to ensure that the system will be observable. The paper [20] investigates the meter placement problem with an objective of ensuring network observability against single branch outages. It presented a topological method to install new meters around the system for single branch outages. The papers [21][22] present a systematic procedure by which measurement systems can be optimally upgraded. The proposed method yields a measurement configuration that can withstand any single branch outage or loss of single measurement, without losing
4
network observability. It is a numerical method based on the measurement Jacobian and sparse triangular factorization, making its implementation easy in existing state estimators. However, that method is valid only for loss of single measurements and single branch outages, and also assumes that the original measurement system is observable. In a given practical power system there will be some specified contingencies which will include simultaneous loss of several measurements and/or outages of several branches. A new method [23] will be proposed in Chapter II, which greatly improves the unified approach presented in [21][22] to account for cases involving such contingencies. This is accomplished by extending the IP problem to consider more than one candidate for a given contingency.
1.3 Loss Allocation
The electric power industry is experiencing important changes brought about by the deregulation. Electric power generators and users engage in power transactions which take place over the transmission system and create losses. Transmission losses represent up to 5.10% of the total generation, and are worth millions of dollars per year. Consequently, the problem of “who should pay for losses” arises and the satisfactory sharing of the transmission system utilization costs among all market participants has become a key issue. Unfortunately, losses are expressed as a nonlinear function of line flows, and it becomes almost impossible to calculate exactly the losses that are incurred by each generator, load or transaction in the system.
A number of loss allocation schemes have been proposed in the literature to allocate the system losses to generators / loads in a pool market or to individual transactions in a bi-lateral contracts market. Based on different assumptions and approximations there are mainly four families of schemes: Pro rata methods [24], incremental transmission loss (ITL) methods [25-30], proportional sharing procedures [31-41], and loss formula methods [42-51]. Different loss allocation methods have been compared in [52-55].
http://idea.library.drexel.edu/bitstream/1860/519/9/Dafis_Chris.pdf
1.2.2 Observability in Power Systems
Historically, the observability problem in power systems considers the following
question: Given the output data on a time interval [t0,t1], what is the procedure to
compute the state of the system? In other words, power system observability requires
that enough measurements exist and they are distributed throughout the network in such a
way that the solution to the state estimation problem is possible. As a result, the derived
observability criterion is the solvability condition of the state estimation problem.
Traditionally the state estimation problem is based on a power system model that neglects
the dynamics of the system and assumes the system is at an equilibrium point. Therefore,
the derived observability criteria are also based on the assumption that the system is at
equilibrium and neglects any system dynamics. The general structure for the
observability determination in power systems is given in Figure 7. Algorithms that
compute observability accept inputs from the measurement system and the network
processors to determine known system variables and the topology of the system
respectively. Once the algorithm qualifies whether the system is observable, the next
stage is to quantify the result. The two main approaches in qualifying observability are
numerical and topological approaches.
The topological approach is based on graph theoretic principles and focuses on
determining the maximum spanning tree (a tree that contains every node in the network),
of full rank, where each branch of the tree is assigned to a different measurement [1,2].
With this configuration, all branch flows can be determined from the measurements,
making the network observable (assuming the entire network is incorporated in the
14
spanning tree). For branches where a meter does not exist, the branch is assigned to an
injection measurement at an incident node. A further introduction to these methods is
provided in Chapter 2.
Observability Formulation
Quantifying Result
Network Processor Measurement
System
Figure 7: General Structure of Observability Determination
Contrary to the topological analysis that centers upon the network topology and the
inherent measurement number and placement, the numerical approaches center upon the
Jacobian matrix defined as [1,2]:
i i
k k
i i
k k
P P
V
J
Q Q
V
ϑ
ϑ
∂ ∂
= ∂ ∂ ∂ ∂
∂ ∂
(1.1)
in the case of real (Pi) and reactive (Qi) power measurements, where the states of the
system are the bus voltages Vi and angles è
i. The Jacobian matrix is introduced in
solving the state estimation problem, and in general, when the Jacobian matrix is of full15
rank the network is said to be observable. A Jacobian of full rank renders the state
estimation problem solvable, providing the link between a solvable state estimation
problem and the observability determination of the system. This Jacobian also includes
information necessary for the solution of the power flow problem in power systems. The
power flow Jacobian, in the case of real and reactive measurements, is typically a subset
of the state estimation Jacobian. A further introduction to numerical approaches
addressing observability based on the state estimation Jacobian is provided in Chapter 2.
Both topological and numerical methods have been applied for large systems with
successful results. However, both methods are deficient in the following areas:
• The information is based only on the algebraic system variables and the dynamic
states of the system are ignored. A way to calculate these dynamic states even at
system equilibrium points is not provided. A change in the power system model
is obviously needed to accommodate this capability.
• The inability to track the observability determination of the system during
transitions from one equilibrium to the next.
1.2.3 Observability in System Theory
The study of observability was motivated by the desire to know the entire state of the
system without measuring all the system states, but rather having measurements that are
functions of the system states, or subsets. The idea of a system description as system
states, inputs and outputs (state-space description), as well as the system properties of
16
controllability and observability were first formulated by Kalman in the early 1960s
[10,11].
The early investigations on the observability property of systems focused on linear timeinvariant
systems (LTI), providing a relationship between the system coefficient matrix
(A in Table 1) and the measurement matrix (C in Table 1), where x is the system states
vector, u are the system inputs and p is the system measurements vector. Later studies
evolved to linear time-varying systems (LTV) [12,13] and nonlinear systems [14,15]. A
brief description of the observability properties of these systems is provided in Table 1
(further details can be found in [13]).
The mathematical models used to describe the dynamics vary greatly, and Table 1
provides a mere subset of these models to outline the progression of observability
research. Therefore various other observability tools exist based on the selected
mathematical model to represent the physical system. In general, the observability
property, and other system properties such as stability and controllability, is global for
linear systems, but local in nature for nonlinear systems. Therefore, evaluating a single
operating point in terms of observability can be sufficient for linear systems, and the
entire family of system operating points needs to be examined for nonlinear systems.
For power systems, various mathematical models exist, however, since the objective is to
track the system as it moves along equilibrium points, only nonlinear system models are
considered.
17
Table 1: Sample Observability Methods in System Theory
1.2.4 Differential Algebraic Models of Systems
The dynamics of a power system can be modeled with a combination of nonlinear
differential equations, and nonlinear algebraic equations. The nonlinear differential
equations correspond to the nonlinear dynamics of the system, and the nonlinear
algebraic equations correspond to the algebraic constraints of the system. In a sense, the
algebraic equations provide the region upon which the trajectories defined by the
solutions of the differential equations lie (solution manifold).
Type Model Observability Test Comments
Linear
Time
Invariant
x Ax Bu
p Cx
= +
=
&
[ -1]
( )
n T
O
O
J C CA CA
rank J n
=
=
L
x∈¡n p∈¡q
, & - Constant Coefficient
Matrices
A B C
Linear
Time
Varying
( ) ( )
( )
x A t x B t u
p Ct x
= +
=
&
1
1 ( , ) ( , ) ( )
( ) ( , )
is nonsingular
o
t
T T
O o o
t
o
O
J t t t C
C td
J
ô ô
ô ô ô
= Ö
Ö
∫
x∈¡n p∈¡q
:System Transition
Matrix
Ö
Nonlinear
ODE
Based
( ) ( )
( )
x f x v x u
p h x
= +
=
&
1
1 1
1
[ ( ) ( )
( ) ( )]
( )
o o
O f f q
n n T
f f q
O
J dL h dL h
dL h dL h
rank J n
− −
=
=
L L
L
(.), (.), (.) :
Nonlinear Functions in
x np q
f v h
x
∈¡ ∈¡
( ) ( )
( )
f
L h x h f x
x
Lie Derivative
= ∂
∂
The electricity markets of England, Spain and Brazil are currently using Pro rata schemes to allocate the losses to
5
generators and/or consumers. This type of methods is simple to understand and implement. However, the network topology is never taken into account. Obviously it is not fair for two identical loads, which locate near generators and far away from generators respectively, to be allocated with the same amount of losses.
ITL methodologies use the sensitivities of losses to bus injections to allocate the losses to generators and loads. The paper [28] provides analyses and test results from a practical implementation of an incremental allocation procedure in the Norwegian electric system. The paper [29] solves a system of differential equations by using numerical integration where a distributed slack bus concept is used. The ITL methods depend on the selection of the slack bus and also the slack bus is allocated with no losses.
Proportional sharing methods, sometimes called flow-tracing schemes, assume that the power injections are proportionally shared among the outflows of each bus and trace the electricity down from the generation sources or up from the load sinks. The assumption here “the power flow reaching a bus from any power line splits among the lines evacuating power from the bus proportionally to their corresponding power flows” is neither provable nor disprovable. Also, it is not possible to allocate losses to generators and loads at the same time.
Recently, some loss-formula based methods have been presented. A quadratic loss formula is proposed in [42] to allocate transmission losses among trades. A "physical-power-flow-based" approach expresses the quadratic loss approximation with individual transactions in a multiple-transaction framework in [43]. Another loss allocation method is based on the bus impedance Z-bus matrix [44] and allocates transmission losses among loads and generators assuming a pool dispatch. A
http://www.pserc.org/cgi-pserc/getbig/publicatio/reports/2006report/abur_state_estimators_s22_reports.pdf
1.2 State Estimation Literature Review
Schweppe was one of the first to formulate static state estimation for a power network
based on the power flow model [1]. The idea is to estimate the electrical states of the power network,
mainly voltage magnitudes and phase angles. These states might not be directly observable
based on physical relationships between the measurements and the desired unknown states.
Another advancement in the field of state estimation was the introduction of a weight matrix
to increase the accuracy of the results. Weighting is done to enhance the “input” of accurate
measurements, and de-emphasize the less accurate measurements. It can be shown that the
maximum likelihood estimate utilizes weights that are based on the covariance of the measurement
devices [2]. The more accurate a measurement, the greater is the selected weight in the
state estimator. Weighting is the practice of accounting for the confidence in a measurement.
Over time, the confidence in a measurement may change. A solution to this problem is to auto
tune of the weights of measurements. The suggested method of auto tuning the weights is to look
at recent error variances of the measurements and use these to recalculate the weights of measurements
from a short history [3]. References [2, 4-10] further relate to ideal weighting of measurements
for power system state estimation.
Measurement errors are typically assumed to be statistically distributed with a zero mean
[11]. Due to increase use in “sensorless” technologies such as A/D converters the zero mean as2
sumption is not always true [11]. A suggested method of overcoming this problem is to combine
measurement calibration [12-15] with state estimation. Calibration of the measurements can be
done in parallel with state estimation by noting the error of measurement over several scans of
the measurement. The calibration error will be a constant compared to measurement error which
is typically normally distributed [15].
The process of overcoming measurement noise is inherent in taking physical measurements,
but there are situations in which the data is grossly erroneous. The data that are erroneous
must be identified and eliminated. One method for the detection of bad data is the examination of
the measurements and if the measurements deviate from expected values by some preset threshold
the measurement can be assumed to be bad [16]. Another problem that causes state estimators
inaccuracies is the power system model itself. Generally the simple linear model Hx=z is
used where the H is the measurement model (processing matrix), x is the state vector, and z is the
measurements. If the process matrix is incorrect, the model does not represent what is physically
happening in the system. The detection of both erroneous data or improper formation of the
process matrix may be done by examining the residual of the equation Hx=z [2]. A further modeling
‘error' is a result of linearization. Since the process matrix is truly a function of operating
state, H=h(x). The linearization of the problem results in constant H.
References [4, 6, 10, 17] are textbooks relating to state estimation in power engineering;
references [5, 8, 9, 18] are representative of solutions methods; and [16, 19] are case studies.
1.3 The Pseudoinverse and Least-Squares Estimation
The commonly used model for a linear static system is
Hx=z (1.1)
with H as the process matrix (m by s matrix), x is the state vector (dimension s), and z is the
measurement vector (dimension m) is overdetermined when m is larger than s. References [4, 6,
10, 17, 20] describe Equation (1.1). Equation (1.1) can be “solved” in the least-square sense by
minimizing ||r||2,
r=Hx-z (1.2)
where || ● ||2 refers to the 2-norm [17]. Properties of norms appear in [17] and Appendix A. It
can be shown that || r ||2 is minimized when
x = xˆ = H+z . (1.3)
The notation xˆ is the “estimate” of vector x, H+ pseudoinverse of H. References [4, 6, 10, 17,
20] describe the properties of the pseudoinverse. Equation (1.3) is known as an unbiased least
squares estimator.
Other methods of determining the state variables are under study. One such method is
weighted least absolute value. Unlike weighted least squares there is no explicit formula for the
solution to linear weighted least absolute value. The weighted least absolute value is found by
linear programming [21]. Another method suggested is to find the maximum agreement with
3
measurements. The state estimate agrees with the majority of the measurements taken in the system
[7].
The least squares method of state estimation requires the system to be observable. Observability
can be defined as: given a set of measurements and their locations (i.e., given z and
H), then a unique estimate of the system state vector x, i.e. xˆ , can be found. A basis of observability
analysis is graph theory. To determine which states are unobservable, set the measurement
vector, z, to zero,
Hxˆ = 0 .
