Vous êtes sur la page 1sur 11

SECTION I.

Introduction
In a power system control center, state estimation is crucial for the reliable operation
of the power system. Once the state of the system is determined, real-time contingency
analysis is carried out to make sure the system can survive any single outage. If any
thermal or voltage violations are detected by the contingency analysis, operators take
corrective actions to bring the system back to a secure state. Failure to recognize
violations, either due to invalid state estimation solutions or inaccurate contingency
analysis, can result in blackouts [1].

All the security functions in an energy management system (EMS) rely on the
availability of measurements coming from substations. The loss of a subset of
measurements can render the state estimation unobservable. For example, in large
scale power systems, it is not uncommon to lose one of the Inter-control Center
Communication Protocol (ICCP) links, which is used to share measurements between
control centers, to a neighboring control area creating an unobservable (external) part
of the system [2].

The current practice of most of the EMS vendors in dealing with the loss of
measurements is to continue using the last set of good measurements until new
healthy measurements can be received. The danger with this practice is that it is
completely blind to the contingencies in the external system, i.e., it assumes the
external system topology does not change. As an alternative to using the last set of
good measurements, the unobservable portion of the system can be discarded to
maintain a smaller observable system with the remaining healthy measurements. Even
though discarding the external system gives us state estimation capability on the
remaining internal system, it still does not provide us any insight about the external
system. This paper investigates the use of the internal system state estimation results
and any available phasor measurement units (PMUs) to detect outages in an
unobservable external system due to the loss of measurements. We determine external
outages by using only the internal bus phasors (coming from the state estimation
and/or PMUs). Our formulation allows any external buses with PMU installations to
be considered as internal buses.

The problem of detecting external topology errors is not new and was studied in [3].
Recently, motivated by the increase in the PMU installations in power systems, several
new line outage detection methods have been proposed, which can be grouped into two
broad categories:

1. combinatorially complex exhaustive search methods [6]–[7][8],

2. computatinally efficient sparse estimation methods [9]–[10].

This paper builds on the earlier work [4]–[5][6] in two aspects. First, instead of using
the combinatorially complex integer programming algorithm, the lasso (least absolute
shrinkage and selection operator) algorithm [11]–[12][13] is used to solve the detection
problem. The second innovation is that an iterative detection check is introduced,
where the outage detected by the lasso solution is compared against the injection
values calculated via the line outage injection formula [14]. If the comparison does not
result in a match, the detected outage is eliminated as a false positive and a new
iteration is performed for the remaining potential outages.

The external outage detection formulations in this paper and also in the other papers
given above are not sensitive to the sudden operating condition changes. The
assumption is that the operating condition of the external system remains relatively
constant, i.e., no sudden changes in generation or load. Depending on the time of the
day and changes in generation patterns, the external outage detection algorithms can
become invalid within minutes. For example, areas with high wind farm penetration
are susceptible to sudden generation changes as the weather conditions fluctuate. In
order to test how the proposed algorithm performs under changing operating
conditions, a sensitivity to load variations (2%) is given in Section IV.

The organization of this paper is such that in Section III, the problem formulation is
described. Simulation results for an IEEE 14-bus system [15] are given in Section IV.
In Section V, in addition to the conclusions, an overview of our future work is
provided.

SECTION II.
Problem Formulation

For a power system with buses and transmission lines, the DC powerflow

equations can be used to relate the changes in bus phase angles to

the changes in real power injections as given in (1). The input data
to our external outage detection problem are power flow measurements. Due to the use

of the DC powerflow in formulation, an error term is introduced to


account for the differences between the actual AC measurements and the DC
powerflow model:

View Source where is the admittance matrix and is the error vector.
Two kinds of errors are captured in . The first one is the small changes in power
injections due to variations in load and generation between two measurement scans—
30 samples per second if PMUs are used or 1-3 minutes if regular state estimation
results are used. The second type of error is the model approximation error due to the
use of quantities coming from full AC models, i.e., power injections and bus angles
coming from the AC state estimation solutions. In the simulation results section, test
cases for both types of errors are provided.

The matrix is defined as:

View Source where is the inverse of the reactance between

buses and and is the set of neighbor buses of bus .


Since the goal of this paper is to detect outages in the external system, where
measurements are not available, the system can be partitioned into internal and
external buses, by modifying equation (1)as:

View Source where subscripts and represent the internal and the external
systems respectively.

When one of the external lines goes out-of-service, the matrix and bus angles
change accordingly to reflect the outage in (3). However, in our problem formulation,

the changes in the matrix and external bus angles are the unknowns. Instead of

modifying the matrix, another way to model an outage, as described in Appendix

11A in reference [14], is to add two power injections and

at each end of the line to be taken out-of-service as shown in Fig. 1.

Figure 1.
Line outage modeling using injections
View All

The first observation from Fig. 1 is that the two inserted power injections

and will have the same magnitude but opposite signs.

The other observation from the bottom part of Fig. 1 is that the real power flow

on the line is equal to the magnitude of the inserted power

injections ,
which creates a condition identical to the outage of the line. Using the pre-contingency

powerflow quantities, can be calculated as described in [14]:

View Source

matrix entries

corresponding to the dropped line (between buses to )


reactance of the dropped line

pre-contingency flow of the dropped line

Assuming there are no changes in the power injections in the system, i.e, zero error

case, the post-contingency version of vector in (3) will have only two
non-zeros:

View Source

We can eliminate one of the two unknowns ( and

), in (3) by first isolating in the bottom part of (3):


View Source
Now we can substitute (7) into the top part of (3):

View Source

Moving terms around in (9) to come up with a linear equation with as


the variable:

View Source where,

View Source

The external outage detection problem boils down to sparse estimation of the

vector as given in the linear equation (10).

SECTION III.
Sparse Estimation
In the literature, different optimization techniques are used to solve a linear
formulation of the line outage detection problem (10), e.g., integer programming
in [6], lasso and orthogonal matching pursuit in [9], and exhaustive search in [7].
Similar to [9], this paper takes advantage of the computational efficiency of sparse
estimation techniques. As described in [9], the sparse estimation of linear regression
models can be grouped into two categories. The first one is based on the matching
pursuit algorithm [16] and its improved versions [17] (commonly used in signal

processing). The second group is based on adding an -norm term to the

typical linear regression formulation given in (10). The addition of the -norm
provides a convex relaxation to the problem of finding the minimum number of non-

zeros ( -norm) [18]–[19]. With the addition of the -norm term, the
linear regression formulation in (10) becomes:

View Source where is the tuning parameter.

Instead of directly applying the lasso method to (14), the number of zeros in

can be further reduced to one by taking advantage of the pre-contingency

topology of the external network through the addition of a bus-branch matrix ,


which is defined as:

View Source

With the addition of the matrix, the formulation in (10) becomes:

View Source where, and lasso


equation (14) turns into:

View Source where, .

The choice of in (17) directly impacts the number of non-zeros in

, Since we are looking for only one non-zero in , we initially


start with a large enough parameter to give all zeros in and
reduce it exponentially in each subsequent iteration until we get a single non-zero.
Using Matlab's lasso algorithm [20], the solution of (16) returns a single non-zero

in , but does not guarantee a unique solution, i.e., there can be a


number of false positives. In order to identify and eliminate the false positives, an
iterative scheme is proposed in this paper, where the value of the non-zero coming

from the lasso solution is compared with the value from (4). If these two

values do not match, the entries in the column of the matrix corresponding to the

detected outage are set to zero. In the next iteration using the modified matrix, a

new outage is detected, which is also compared with the new value.
Iterations continue until the value of the non-zero coming from the lasso solution

matches the calculated value.


In [7], similar to this paper, (4) is used in detection of outages by iterating through all
possible outages, which becomes computationally challenging for large scale systems.
In contrast, the method proposed in [9] is computationally efficient since it also uses
sparse estimation techniques, such as lasso. However, there is no post verification of
the detected outages in [9]. Combining computationally efficient lasso method with the
use of (4) to verify the detected outage is the main advantage of this paper.

As shown in [9], the greedy schemes such as optimal matching pursuit


algorithm [17] can provide a faster alternative to lasso. In this paper, for testing the
validity of our proposed method in a 14-bus system, we solve (14) with the lasso
algorithm via coordinate descent (CD) [12]. Solving the same problem with greedy
methods for comparison would be interesting for future work especially for large scale
systems.

A. Line Outage Detection Algorithm


The line outage detection algorithm can be summarized as follows:

1. form

and as given in (2) and (3).

2. form and according


to (11), (12), (13) and (15) respectively.
3. apply the lasso method to (16) to detect the outage.
4. calculate according to (4) and compare it with the lasso solution
found in step 3. if the two values match, the outage is successfully detected and
algorithm terminates. otherwise, go to step 5.
5. set the entries—only two non-zeros: 1 and −1 as given in (15)—of the column of

the matrix, corresponding to the lasso solution, to zero and repeat Step 3

with the updated matrix.


Note that for the error free case, the comparison in Step 3 provides an exact match. As
the value of the error in (3) increases, the tolerance for comparison in Step 3 needs to
be relaxed, which in turn reduces the detection accuracy. Test cases showing the
performance of the algorithm with and without the error term is demonstrated in the
next section.

SECTION IV.
Simulation Results
The proposed algorithm is tested on a modified IEEE 14-bus system as shown in Fig. 2,
where the top and bottom circled parts represent the external (buses 6 through 14) and
internal (buses 1 through 5) systems respectively. The blue square symbols (seven of
them) on the lines are used to label the external outages to be detected—one at a time.
Three of the ten lines (branches 7–8, 9–14 and 13–14) are left out of the detection
algorithm since they are either radials or lines that are connected to the reference bus
(bus 14) whose corresponding rows and columns are removed from all the matrices
used in the problem formulation to guarantee a unique DC powerflow solution through

a full rank matrix. The choice of the reference bus is arbitrary. In an actual large
power system, the reference bus can be chosen far from the boundary buses between
the internal and external systems. For the 14-bus system studied in this paper, the
furthest bus away from the boundary (bus 14) is chosen as the reference bus.
The line parameters in the original IEEE 14-bus case [15] has high impedance lines

with low ratios (in the order of 2 to 5), which are more typical for sub-
transmission or distribution systems instead of a transmission system. To make the
system more representative of an actual transmission network, a number of changes
are made to the original IEEE 14-bus case. The first change is to reduce the resistance

and reactance parameters to get the ratio to around 10. The second change is
to increase the real power load to move more power through the lines—outage of a
lightly loaded line does not have a major impact on the internal power flows. The final
change is to reduce the generation in the internal system so that it becomes an
importing area relying on external lines for import capability.

Figure 2.
IEEE 14-bus test case
View All

The pre-contingency and the post-contingency outage modeled flows, based on (4), are
given in Table I. The values given in the Post-Outage column in Table I are calculations
coming from the line outage modeling using injections as displayed in Fig. 1.

Table I. 14-Bus DC simulation pre- and post-outage flows

A. DC Simulation—Error-Free Case
Using the modified IEEE 14-bus system, DC powerflow simulations are carried out for
the all-lines-in condition and seven external line outages. In this first test case, no
error is modeled to see if all the outages can be detected in an error-free setting. Table
II shows how many iterations it takes to detect the outage and also the value of the

power injection, from Fig. 1, to mimic a line outage without actually

opening the line as given in (4). In the first column, the actual outage and the

values are provided. In the following columns, the detected outage and its

corresponding values are given for each iteration until the algorithm
terminates.
As shown in Table II, the algorithm successfully detects all the outages in the external
system. For four of the outages (branches 10–11, 6–11, 6–13 and 12–13), multiple
iterations are needed to eliminate the false positives, which is due to the

underdetermined nature of the problem. The structure of the matrix from (16) is
such that it has 5 rows and 7 columns corresponding to internal bus angles and
external outages to be detected, but only 2 of the 5 rows have nonzero entries. The

matrix product of and matrices


in (11) and (12) explain the zero rows. In other words, since buses 4 and 5 are the only
connections between the internal and the external systems, only the rows
corresponding to these two buses produce non zero rows in (16). Note that the post-
contingency power flow on the line ends up identical to the actual outage values for the
error-free case. Introduction of the error will make it more difficult to match the actual
and detected outage powerflows as shown in the next case.
Table II. 14-Bus DC simulation (error-free case)

B. DC Simulation—2% Error
As the time step between two measurement scans (or state estimation solutions)
increase, the power injections will change according to the daily load cycle instead of a
random variation. In the literature, such as in [9], it is common to use random (usually
Gaussian) error in modeling load changes. However, a more challenging, but realistic,
way to model changes in loads is to increase all post-contingency power injections by a
certain percentage. To test our algorithm under changing load conditions, the same 14-
bus case is used, but all post-contingency power injections are increased by 2%. The
major difference between the error-free and the 2%-error cases is that in the 2%-error
case as shown in Table III, the injection values found by the algorithm does not exactly
match the actual calculated values in column 1. Therefore, a detection threshold needs
to be set in comparison of actual and detected values.

Table III. 14-Bus DC simulation (2% error case)

For the first outage (branch 7–9) given in Table III, the absolute value of the calculated
value, 103.2 MW, is almost 30% smaller than the absolute value of the detected value,
133.6 MW. Note that if this 30% threshold is used for all the outages, the third outage
(branch 10–11) cannot be detected due to the calculated value (24.1 MW) being almost
50% smaller than the detected value (37.1 MW). However, increasing the threshold
value more will make it difficult to catch false positives and, therefore, reduce the
accuracy of the algorithm.

Another observation from Table III is that the greater the pre-contingency flow on a
line, the easier it is to detect the outage of that line. For instance, for the sixth outage
(branch 6–13), the pre-contingency flow is −72.4 MW as given in Table I and the
calculated (−211.3 MW) and the detected (−219.3 MW) outage values are within 3% of
each other. On the other hand, for the third outage (branch 10–11), the pre-
contingency value is −6.3 MW and the algorithm fails to detect this outage. Not being
able to detect the outage of the lightly loaded lines is not critical for the internal system
since the loss of the lightly loaded lines will not result in major impacts on the flows
and voltages of the internal system. In an actual transmission system, for the reliability
and security of the internal system, it is important to be able to detect outages on the
external lines close to the internal system that carry significant flow, such as greater
than 100 MW.

C. AC Simulation
The DC powerflow based simulations are useful to test the algorithm, but in a real-life
application, the bus phase angles and power injections will be based on full AC models
or PMU measurements. Therefore, in the third simulation, the performance of the
algorithm is shown when AC powerflow quantities are used. Using the same IEEE 14-
bus test case, the post-contingency values come from AC powerflow solutions. The
detection results are given in Table IV.

Table IV. 14-bus AC simulation


All lines were successfully detected using the AC quantities and all the detected outage
values came within 10% of the actual values.

SECTION V.
Conclusions
In this paper, we proposed a new method for identifying external line outages. The new
idea was the post-processing of the detected outages to eliminate false positives. The
algorithm was tested using a modified IEEE 14-bus system for both DC and AC
simulations. The post-processing feature improved the detection accuracy significantly
and allowed us to detect all outages except for one (branch 10–11 in the DC simulation
with 2% error).

Our future work includes testing this algorithm for a large scale utility system in North
America and for different error models. Once verified for larger systems, we plan on
implementing this algorithm to detect external outages for voltage stability limited
load pockets.

Vous aimerez peut-être aussi