Você está na página 1de 6

Application of Distributed Compressive Sensing to

Power System State Estimation


R. Jalilzadeh Hamidi, Student Member, IEEE, H. Khodabandehlou,
H. Livani, Member, IEEE, and M. Sami Fadali, Senior Member, IEEE

Abstract This paper presents an application of distributed which quantitatively assesses the error in a measurement. In
Compressive Sensing (CS) for data recovery/reconstruction in [6], a method is presented for identifying bad data in nearly-
Power System State Estimation (PSSE). Transmitted critical sets (i. e., critical tuples, for 2). In some works,
measurements to power system control centers may disappear bad data identification and state estimation are jointly
due to congestion or disconnection in communication links, accomplished [7], [8]. As mentioned in the existing literature,
sensor failures, and cyber-attacks. Consequently, the state power system bad data are often detectable. However, only a
estimator may encounter problems. In the proposed method, the limited number of methods have addressed the removal of the
identified (Phasor Measurement Unit) PMU bad/missing identified bad measurements and/or the reconstruction of
measurement(s) are reconstructed using CS. Data reconstruction
missing measurements. In [1], bad data is removed when the
exploits the correlation in both time and space among the PMU
measurements using a random projection matrix and a wavelet
estimated error is subtracted from the bad measurement,
dictionary. The linear state estimation is then carried out using provided that there exists only one bad measurement. In [9], an
the available and reconstructed PMU measurements. The approach is presented for recovering bad measurement using
proposed method is evaluated on the IEEE 57-Bus transmission the estimated measurement residuals and the un-detectability
system. The capabilities and limitations of the proposed method index described in [5]. Both [1] and [7] fail to remove large
are also discussed. errors in the bad measurements and also fail to reconstruct a
missing measurement.
KeywordsCompressive sensing, data recovery, PMU data,
state estimation.
Guo et al. applied online sparsity-based missing data
recovery methods in PSSE [10]. While there are several data
I. INTRODUCTION compression techniques such as distributed source coding,
matching pursuit, and other greedy basis pursuit algorithms,
Accurate power system state estimation is a critical task for Compressive Sensing (CS) has received a great deal of
secure and economic operations in power transmission attention recently. It is shown in [11] if a signal is sparse on
networks. The state estimation is performed at monitoring and some known basis, it is possible to recover the signal from a
control centers, which are usually part of a Balancing small number of samples using CS. In [12], a distributed CS
Authority (BA) or an Independent System Operator (ISO). approach is presented to process signals with inter and intra
State estimation requires measurements from meters at correlation. Applications of CS in power systems are studied in
different locations in a transmission network. These [13], [14]. Regarding PSSE, in [15], the power system
measurements include small random errors and bad data due to measurements are compressed and then decompressed using
meter bias and noise. Measurements may also be lost as the CS, and it is shown that the PSSE solutions based on the
results of communication links congestion or disconnection, decompressed data are reliable enough for PSSE applications.
sensor failures, and cyber-attacks. Thus state estimation results In [16], an iterative algorithm is introduced for sparse recovery
may become inaccurate and the system may become from power system measurements based on the system
unobservable in case of missing critical measurements.
admittance matrix ( ). In [17], a regression-based technique
Therefore, removal of bad data and reconstruction of missing
is proposed for power system state forecasting utilizing (Phasor
data are essential for having reliable state estimation results.
Measurement Unit) PMU along with SCADA measurements.
There are several bad data identification and detection
Given the rapid increase in the number of PMUs installed
methods for Power System State Estimation (PSSE). The
in substations in recent years, it is not unrealistic to assume that
simplest but, also, the least accurate is the Chi-squares method
most transmission systems will be fully observable by only
[1]. The Largest Normalized Residuals test is more accurate in
PMU measurements in the future. We assume that the system
bad data detection, but it fails in case of multiple conforming
is fully observable with a minimal number of installed PMUs.
interacting bad data [2]. In [3], the Hypothesis Testing
A large number of data measured by PMUs increases the
Identification technique is introduced for bad data
probability of data loss in communication systems. If any PMU
identification in power systems. This method is based on the
data is lost, the system becomes partially unobservable.
calculated estimates of measurement errors [1], [3]. In [4], an
Referring to Fig. 1, once the bad/missing measurements are
efficient algorithm is introduced to detect bad data resulting
identified, Distributed CS is utilized to reconstruct the
from unobservable/irreducible attacks. However, this algorithm
bad/missing data. Thus, the power system becomes fully
is only applicable if exactly two injection meters are available.
observable and PSSE is performed using the available and the
In [5], a topological-based un-detectability index is introduced
reconstructed measurement sets. The main advantages of this
This paper is based upon work partially supported by the National Science method are as follows: 1) The procedure to find correlation
Foundation under Grant No. IIA-1301726. among the system states is not affected by a change in system
R. J. Hamidi, H. Khodabandehlou, H. Livani, and M. S. Fadali are with
the Electrical and Biomedical Engineering Department; University of Nevada,
Reno; Reno, NV. 89557, USA.
topology, and it is able to find the correlation again after a
short while, 2) the proposed method is computationally cost-
effective and fast, 3) the CS basis functions are updated
according to the last system state to find the best correlation
among the measurements, 4) Wavelets are used for feature
detecting since wavelets have sharp edges that enable them to
handle data with sharp changes.
This paper is organized as follows: Section II describes the
proposed method, as well as the CS theory and its application
to PSSE. Section III presents and discusses the simulation
results for the IEEE 57-bus. The conclusion is given in Section
IV.
II. METHODOLOGY
Referring to Fig. 2, in the proposed method, all the
measurements are collected at a control center. If the control
center identifies bad/missing measurements, they are
reconstructed based on the correlation among the missing and
available measurements using Distributed CS. Consequently, a
state estimator obtains the PSSE solution. Then, CS basis
function, , is updated by the last estimate of the system
measurements which leads to accurate reconstruction of the Fig. 2. The proposed method flowchart.
bad or missing measurements. In the following sections, the CS
theory and its application to PSSE are described. Most traditional methods, such as transform coding and
greedy basis pursuit algorithms, require all the coefficients in
A. Compressive Sensing Theory (1) to determine the largest coefficients. CS uses a finite
Although there are several signal compression techniques number of linear projections to obtain those coefficients. This
like transform coding, distributed source coding, and greedy can be done by calculating the projections of data on a
basis pursuit algorithms, most of these methods are second basis set as follows [12]
computationally complex . CS is one of the well-known , , 1,2, ,
compression techniques that allows us to compress a signal (3)
using a small number of linear projections. Consider a real- If and are incoherent, meaning that none of the
valued measured signal of length . Assume that measured elements of first basis have sparse representation on the terms
signal is sparse on basis , then of the second basis, and when the is sufficiently large, it is
possible to recover those largest coefficients using
projections [12],
(1)
. . (4)
In the matrix form, the equations can be written as while it is enough to have at least 1 measurements to
solve this minimization problem, solving this problem is
(2) time consuming. In fact, it is much easier to solve the
equivalent minimization problem [12],
. . (5)
This is equivalent to a much simpler linear programming
problem. However, changing the norm minimization to
norm minimization results to an oversampling factor. This
means that we need more than 1 measurements to be able
to recover the most significant coefficients from ,
1,2, . . . , [12].
Theorem: Set with 0 1. Then there
exists an oversampling factor 1/ , 1
, such that, for a -parse signal in basis, , the probability of
recovering via Basis Pursuit from random
projections, 0, converges to 1 as .In contrast, the
probability of recovering via Basis Pursuit from
random projections converges to 0 as [12].
Fig. 1. The overall concept of the proposed method.
1) Distributed Compressed Sensing B. State Estimation Method
While the basic compressed sensing theory deals with just Different methods have been introduced for PMU-only
one signal, many real world applications involve more than one linear PSSE. In [20], a non-iterative method uses voltage,
signal. In such cases, one may apply compressed sensing to current, and line admittances in phasor form. In [21], a fast
each signal separately. When there is no correlation between decoupled state estimator utilizes real and imaginary parts of
signals, this solution is simple and efficient, but if the ensemble voltage, current, and line admittances. In this paper, we use the
of signals are mutually correlated, this solution will be same approach as [20]. When a PMU is installed on a bus, the
inefficient. Distributed Compressive Sensing is an expansion bus voltage phasors and all the branch current phasors
of traditional compressed sensing theory to deal with connected to that bus are measured. To utilize the phasor
ensembles of correlated signals. Distributed CS is particularly measurements for PSSE, the SE is formulated as follows [20]
useful and will provide a better compression rate if the
ensemble of signals show strong mutual correlation. (14)
Given signals , 1,2, , , . Assume that
where is the measurement vector consisting of
there exists a basis for on which all the signals can be voltage and current phasors, is the state vector
sparsely represented. Also assume that all signals have a consisting of bus voltage phasors, is the identity matrix with
common sparse component plus an innovation component . the rows corresponding to the buses equipped with a PMU, and
The signals can thus be written as is defined as follows [20]
, 1,2, , (6) (15)
and their components can be written as where , is the series admittance of the
, branch on which the mth current measurement installed, is
(7) the current measurement-bus incidence matrix, and
, , 1,2, , (8) is a current-measurement-by-bus matrix where is
the shunt admittance of the branch that measured by the mth
The main advantage of this approach to distributed
current measurement and connected to bus .
compressed sensing is that it requires a single linear program.
To be able to jointly recover the ensemble of signals using a Let , then the state estimation problem becomes
single linear program we define the following matrices,
(16)
where is the weighting matrix, and . denotes matrix
transposition.
, , (9)
III. SIMULATION RESULTS
In this paper, the simulation results are presented for the
, ,, (10) IEEE 57-Bus which includes 57 buses, 7 generators, and 80
transmission lines [22]. It is assumed that PMUs are utilized as
0 0 the measurement devices and the buses equipped with PMUs
0 are shown in Fig. 3 (in green) and also stated in Table I. The
(11)
0 considered transmission system requires 21 PMUs to become
0 0 fully observable. It is assumed that the PMUs send voltage and
current phasors, and (if applicable) power injections to the
Using the measurement matrix, , compressed
control center every second. As shown in Fig. 2, if the control
measurements are calculated and using both and the center identifies bad and/or missing measurements, the bad
sparse coefficient vector, is calculated, using measurements are reconstructed based on the correlation
. . among the missing and available measurements using
(12)
distributed CS. Also, CS basis function, , is updated by the
where Contains nonzero elements and the last available system state. The simulations are performed in
complete signal can be recovered using MATLAB-2015 environment on a computer with Intel Xeon
E5420 CPU and 16 [GB] of RAM.
(13)
Drastic changes in power system state reduce correlation
If the measured signals have a strong correlation, this among PMU measurements and make it harder for CS to
method allows us to do more compression. In fact, in this reconstruct the bad/missing measurements. In order to evaluate
method, distributed compressed sensing is performed by the performance of our approach, total load is increased by 7%
stacking measurements to form a single vector and running a in 5 [min] as a drastic change in power system operation. We
single linear program. This makes the method easy to also assume that the control center receives measurements and
implement but increases the computational cost. The runs state estimator every second. In order to evaluate the
computational complexity of a linear program is cubic and the performance of distributed CS technique on reconstructing the
computational cost increases dramatically with an increase in power system measurements, no noise is added into the
the size of the vector [12].
measurements (i.e., identity weighting matrix). Therefore, the TABLE I. PMU Placement.
errors in PSSE originate from measurement reconstruction. 1 3 4 9 14 15 18
Three scenarios are simulated. In all three scenarios, an PMU place
20 24 25 27 29 32 34
[bus index]
increase in the total load initiates at 0 [s] and some of the 38 40 41 48 51 53 57
PMUs become unavailable at 30 [s]. Table II shows Totally, 21 PMUs are installed.
unavailable PMUs, missing measurements, and the resulting
unobservable islands. We consider the effect of losing PMUs
TABLE II. Lost Measurements and the Unobservable Islands.
on buses 15 and 38 that are connected to five branches. Unobservable islands
Therefore, their loss has a significant impact on PSSE results. Lost PMU Lost phasor
[bus index] measurements bus
First, to provide an insight into the reconstructed branch
index
measurements by the proposed method, the actual and the V15, I15-1
15-45
reconstructed measurements are presented. Fig. 4 and Fig. 5 Scenario 1 15 I15-3, I15-13 45
44-45
show the reconstructed voltage and current phasors and their I15-14, I15-45
corresponding errors. Fig. 4 shows voltage phasors at bus 15 in V38, I38-22
21-22, 22-23
the first scenario. The PMU data at bus 15 becomes 22 22-38, 36-37
Scenario 2 38 I38-37, I38-44
37 37-38, 37-39
unavailable at 30 [s]. Thus, the missing measurements are I38-48, I38-49
38-44, 44-45
reconstructed using distributed CS. As it can be observed in
V15, V38
Fig. 4, the reconstructed voltage magnitudes closely pursue the I15-1, I15-3
15-45, 44-45
real values. However, the reconstructed voltage angle drifts 22 21-22, 22-23
I15-13, I15-14
Scenario 3 15 & 38 37 22-38, 36-37
after 10 [s]. The absolute errors are used to compare the actual I15-45, I38-22
45 37-38, 37-39
and reconstructed values. The absolute errors are calculated by I38-37, I38-44
38-44, 44-45
I38-48, I38-49
| | (17)
where |. | denotes absolute value, and are actual and
reconstructed values of a measurement.
According to the absolute errors shown in Fig. 4, although
the reconstructed voltage angle drifts from the actual value, it
does not diverge and remains within an acceptable range for
PSSE (max absolute error is less than 0.05 [deg]).
Fig. 5 shows the current magnitude [pu] and angle [deg] for
branch 15-13 in the first scenario. Similar to the voltage angle,
the reconstructed values closely follow the actual values for the
first 10 [s] then start drifting. Although they lose their
consistent pursuit of the actual value, the reconstructed current
magnitude and angle are still reliable for PSSE application. Fig. 4. Voltage phasors and the absolute errors at Bus 15 in the first scenario.
The correlation between the missing measurements and
available ones deteriorates as time passes. Therefore, the
drifting trend in the reconstructed values is normal. Regardless
of the drifts in the measurement reconstructions, the PSSE
results need to be considered for assessing the proposed
method.

Fig. 5. Current phasors and the absolute errors at branch 15-13 in the first
scenario.

Fig. 6 shows the estimated voltage magnitudes and their


corresponding errors according to the actual and reconstructed
data, scenario 1, at 50 [s]. As it is observed in Fig. 6(b),
Fig. 3. The 57-bus IEEE test case, the buses equipped with a PMU are shown
in green, listed as {1, 3, 6, 9, 14, 15, 18, 20, 24, 25, 27, 29, 32, 34, 38, 40, 41,
the largest absolute error of the voltage magnitudes is around
48, 51, 53, 57}. 1.2 10 [pu] which is at buses 19, 50, and 54.
Fig. 8. Voltage magnitude and angle RMSEs in different scenarios over the
time.

shows that the bus angles are generally estimated more


accurately than the bus voltages. It is noteworthy that bus
angles are often more important to system operators. As a
common practice, the Root-Mean-Squared Error (RMSE) is
Fig. 6. a) Voltage magnitudes estimation using the actual and reconstructed calculated in order to evaluate the performance of the proposed
measurements, scenario 1, t=50 [s]. b) Absolute error between the PSSE method on PSSE results [8], [17], [21].
solutions based on the actual and reconstructed measurements.
1
(18)

where is the number of buses, is the difference


between state estimation results using the actual and
reconstructed measurement for the ith bus at time [s].

TABLE III. RMSEs in Different Scenarios.


Scenario 1 Scenario 2 Scenario 3
RMSE RMSE RMSE
T[s] Vol. Ang. T[s] Vol. Ang. T[s] Vol. Ang.
10-6 10-6 10-6 10-6 10-6 10-5
30 3.281 3.310 30 2.957 1.551 30 9.152 4.429
31 1.807 2.944 31 8.064 2.911 31 9.293 2.059
32 3.277 3.284 32 3.295 6.654 32 8.469 1.659
33 2.357 3.087 33 5.451 5.040 33 16.15 4.324
34 2.189 3.002 34 1.081 1.692 34 8.088 1.939
35 2.728 3.129 35 6.609 3.490 35 4.253 4.874
36 3.383 3.279 36 6.173 3.890 36 2.439 2.279
37 1.983 2.962 37 7.527 5.586 37 4.215 1.833
38 2.846 3.177 38 4.807 2.391 38 3.061 1.482
39 2.955 3.210 39 5.361 4.501 39 3.872 4.017
40 3.284 3.299 40 3.745 3.659 40 5.021 3.857
41 4.542 3.698 41 2.791 1.343 41 5.526 6.146
42 5.072 3.845 42 3.878 8.931 42 6.272 3.060
Fig. 7. a) Voltage angles estimation using actual and reconstructed
43 5.767 4.114 43 5.580 1.266 43 4.940 8.654
measurements, scenario 1, t=50 [s]. b) Absolute error between the PSSE
solutions based on the actual and reconstructed measurements. 44 3.058 3.281 44 3.733 3.580 44 8.612 8.060
45 2.816 3.240 45 5.167 2.455 45 7.160 4.706
Fig. 7 shows voltage angles and their corresponding errors 46 3.923 3.443 46 3.779 4.510 46 8.010 1.400
based on the actual and reconstructed measurements, scenario 47 3.173 3.337 47 4.756 3.963 47 5.953 8.399
1, at 50 [s]. The absolute errors of the voltage angles have 48 4.270 3.601 48 7.489 7.761 48 6.808 5.768
a peak around 1 10 [deg] at buses 14, 39, and 50, which is 49 8.143 4.809 49 5.971 5.908 49 4.633 6.509
acceptable for PSSE. Comparing Fig. 6(b) and Fig. 7(b) 50 7.709 4.648 50 5.530 2.093 50 8.721 2.137
Fig. 8 and Table III show RMSEs for voltage magnitudes [6] F. Fusco, General bad data identification and estimation in the presence
[pu] and angles [deg] in all three scenarios from 30 [s] to of critical measurement sets, in proc. IEEE PES General Meeting, pp.
1-5, Jul. 2014.
50 [s] in which the PMU(s) are unavailable. As expected,
[7] V. Kekatos, G. B. Giannakis, Distributed robust power system state
the error is generally larger when two PMU measurements are estimation, IEEE Trans. on Power Systems, vol. 28, no. 2, pp. 1617-
missing but the RMSEs are in an acceptable range. 1626, May 2013.
As PMU-only State Estimation (PSE) is assumed, the [8] M. Gl, A. Abur, LAV based robust state estimation for systems
measured by PMUs, IEEE Trans. on Smart Grid, vol. 5, no. 4, pp.
solution method is linear non-iterative [20]. This method is far 1808-1814, Jul. 2014.
faster than iterative methods and most of the CPU time is due [9] N. G. Bretas, J. B. A. London, Recovering of masked errors in power
to the matrix inversion by Cholesky decomposition method. systems state estimation and measurement gross error detection and
The CPU time comprised the times for data reconstruction and identification proposition, in proc. IEEE PES General Meeting,
state estimation solution. The average CPU times for case Minneapolis, pp. 1-6, Jul. 2010.
studies are, negligibly different, approximately 630 [ms], 450 [10] D. Guo, Z. Liu, X. Qu, L. Huang, Y. Yao, M. Sun, Sparsity-based
for data reconstruction and 180 for state estimator. online missing data recovery using overcomplete dictionary, IEEE
Sensors Journal, vol. 12, no. 7, pp. 2485-2495, Jul. 2012.
It is observed that, the accuracy of the power system state [11] D. L. Donoho, Compressed Sensing, IEEE Trans. on Information
forecasting utilizing our proposed method is improved Theory, vol. 52, no. 4, pp. 1289-1306, Apr. 2006.
compared with the existing forecasting approaches in the [12] D. Baron, M. F. Duarte, S. Sarvotham, M. B. Wakin, R. G. Baraniuk,
literature. Furthermore, the proposed method provides An information-theoretic approach to distributed compressed sensing,
in Proc. 43d Allerton Conf. Communication, Control, and Computing,
reconstructed measurements for several seconds which are Monticello, IL, Sep. 2005.
reliable for PSSE applications. However the proposed method
[13] M. Majidi, M. Etezadi-Amoli, M. S. Fadali, A novel method for single
requires available data from all PMUs before losing some of and simultaneous faults location in distribution networks, IEEE Trans.
the measurements. on Power Systems, pp.1-9, 2014. Early Access.
[14] M. Majidi, M. S. Fadali, M. Etezadi-Amoli, M. Oskuoee, Partial
IV. CONCLUSION discharge pattern recognition via sparse representation and ANN, IEEE
In this paper, we propose a method for reconstructing bad/ Trans. on Dielectrics & Electrical Insulation, Vol. 22, No. 2, pp. 1061-
1070, Apr. 2015.
missing PMU measurements by distributed CS for PSSE
[15] S. S. Alam, B. Natarajan, A. Pahwa, Distribution grid state estimation
(bad/missed data should be detected using any given method from compressed measurements, IEEE Trans. on Smart Grid, vol. 5,
before data reconstruction). The available measurements at a no. 4, pp. 1631-1642, Jul. 2014.
control center are first tested for identifying the probable [16] W. Xu, M. Wang, J. Cai, A. Tang, Sparse error correction from
bad/missing measurement(s). If bad/missing measurement(s) nonlinear measurements with applications in bad data detection for
are identified, then they are reconstructed by distributed CS. power networks, IEEE Trans. on Signal Processing, vol. 61, no. 24, pp.
Then, state estimator obtains the PSSE solution using all 6175-6187, Dec. 2013.
available measurements, including received and reconstructed [17] M. Hassanzadeh, C. Y. Evrenosoglu, Use of PMUs in regression-based
power system dynamic state estimation, in proc. North American
measurements. The CS basis set is updated every second using Power Symposium (NAPS), Champaign, IL, pp. 1-5, 2012.
the last state estimation results to preserve its capability of
[18] S. Foucart and H. Rauhut, A Mathematical Introduction to
accurately reconstructing bad/missing measurements. The Compressive Sensing, Birkhuser Basel, 2013.
proposed method is tested on the IEEE 57-bus test system and [19] Y. C. Eldar and G. Kutyniok, Compressive Sensing, Theory and
the results are discussed. The proposed algorithm will be tested Application, Cambrige University Press, 2012.
on the IEEE 118-Bus and 390-Bus Nevada systems. We will [20] K. D. Jones, Three-phase linear state estimation with phasor
also investigate the application of iteratively reweighted measurements, Master Thesis, Virginia Polytechnic Institute & State
algorithms-based CS to PMU data reconstruction. University, Blacksburg, VA, May 2011.
[21] M. Gol and A. Abur, A fast decoupled state estimator for systems
measured by PMUs, IEEE Trans. on Power Systems, vol. PP, no. 99,
pp. 1-6, 2014.
REFERENCES [22] IEEE 57-bus, Available on-line:
[1] A. Abur and A. G. Exposito, Power system state estimation: theory and https://www.ee.washington.edu/research/pstca/
implementation, Marcel Dekker Inc., 2004.
[2] A. Abur and A. G. Expoosito, Detecting multiple solutions in state
estimation in presence of current magnitude measurements, IEEE
Trans. on Power Systems, vol. 12, no. 1, pp. 370-375, Feb. 1977.
[3] L. Mili, T. custem, M. Ribbens-Pavella, Hypothesis testing
identification:a new method for bad data analysis in power system state
estimation, IEEE Trans. on Power Apparatus And Systems, vol. 103,
no. 11, pp. 3239-3252, Nov. 1984.
[4] A. Giani, E. Bitar, M. Garcia, M. McQueen, P. Khargonekar, and K.
Poolla, Smart grid data integrity attacks: characterizations and
countermeasures, in proc. IEEE International Conference on Smart
Grid Communications (SmartGridComm), pp. 232-237, 2011.
[5] N. G. Bretas, J. B. A. London, L. F. C. Alberto, and R. A. S. Benedito,
Geometrical approaches for gross errors analysis in power systems
state estimation, in proc. IEEE Power Tech Conf., Bucharest, Romania,
pp. 1-7, Jun. 2009.

Você também pode gostar