Fault Identification and Detection Using Compressive Sensing for Lowering the Amount of
Transmitted Data
Vinan-Velasco Ruben University
of Lincoln-Nebraska Department of
Electrical Engineering
Lincoln, Nebraska, USA
[email protected] Abstract—This document exposes the application of the most of the currently research focus on the wide area protection
compressive sensing in the power systems field, specifically in the system, since wide area models use more data.
protection application where a big challenge nowadays is the
One of the biggest challenges is the risk of losing data during
reduction of the amount of data transmitted for monitoring and
the transmission, so several algorithms had been proposed, with
protection system, without compromising the efficiency of the the objective of identify the fault area, even if they are not able
systems. For several instances in the power systems the respective to predict exactly the fault location. The simple fact that the
values of voltage and current are evenly distributed and fault zone can be identified represents already a considerable
therefore the data might well be redundant for the steady state, reduction of the amount of data that needs to be transmitted,
yet during fault the values tend to change significantly and might since a system can be composed of tens to hundreds of regions.
be as low as zero. With this project the reduction of the amount
of data transmitted will be performed using a compressive This project uses the compressive system approach to reduce
sensing approach to transmit a particular marker for the fault the data that is required by the system in case of a fault, by
expressed as a sparse vector and a minimum of data from the sending a sparse vector representing markers for the fault, and
sensing matrix, thus reducing the traffic data required for real based in a dictionary composed by the information of several
time monitoring/protection. other faults in the system. Fundaments for this project can be
found in the reference [], although the development of the
Keywords—faults, power systems, protection systems, algorithm and the final algorithm have a different outcome.
compressive sensing, data traffic, fault zone.
II. PRESENTATION OF PROBLEM
I. INTRODUCTION
For a basic electric system the main goal is to develop a
Power systems usually work with very high magnitudes of program that will allow to obtain a sparse representation of the
voltage and/or current, and the structure of the power network fault that can be used to estimate the allocation and principal
is constantly expanding. The importance of the stability and voltage magnitudes of the fault. The amount of data reduced
reliability of the power system relies in the effect and will be at least half the amount of data points collected from
dependency of power supply of modern society, thus an active the system and try to get a number 2-sparse vector as
and accurate real time tracking of the behavior of all the system representation of the fault vector. The matrix used for the
is crucial, since this data will help operators and computers to problem is composed of the voltage values of the bus-bars in
take actions in order to prevent any failure in the system. The the system, each phase voltage is considered independent and
amount of data to be transmitted is considerably high, since a following the models for protection and control systems for
system usually is composed by tens or hundreds or elements, actual facilities, data points for the cycle previous to the fault
which each provides information for the three phases in the and after the fault.
power system, and the power system needs to received several
data points from the red per second, this is conflictive with the Each fault vector, thus, needs to be reduced from hundreds or
requirements of velocity that a reliable system have. maybe thousands of data points to a number below four. Since
the system will be downgraded from non-sparse to sparse, a
Another problem that comes with the amount of data dictionary and a sampling matrix need to be built according to
transmitted is the capacity of the communication system and the needs of the system, and considering that the final
the devices that compose it. Systems with more capacity tends algorithm problem should not be dependent to specific values
to be more expensive and carry problem of efficiency due to or lengths of transmission lines, fault impedances or voltage
noise or interference, thus it is necessary to find alternatives for levels. Faults will be considered only between 3% and 97%
data transmission that does not sacrifice the operability of the since the protection systems tend to overshoot protection zones
electric network. Research related to this problem has be done, by 5%
III. PRESENTATION OF THE MODEL B. Test Faults
For this project a model of four substations and four Four faults are going to be simulated in order to evaluate
transmission lines is going to be applied. Transmission voltage the performance of the algorithm. Test faults were calculated
is expressed in per unit notation, thus the system is independent at 1% in Line one, 40% in Line two, 60% in Line 3 and 85%
from the level of voltage selected, in the simulation software in Line four.
138 kV was chosen for calculation purposes.
Each substation supply three voltage signals, each which is C. Down-Sampling
composed by 80 data points from the respective voltage cycle.
A full set of data per cycle for the moment before the fault and
after the fault are used for this simulation, thus there In order to obtain a reduced version of the amount of data
4x3x80x2= 960 total data points transmitted per cycle from the that is going to be transmitted a down-sampling matrix is
system. Normally the grid will be much bigger and the system going to be used. A down-sampling matrix is essentially an
more complex, delivering vector with many more data points. identity matrix that have some of the ‘1’values in the diagonal
Impedance values for the lines will be considered from replaced with ‘0’. This matrix is going to be multiplied with
actual values for impedances and power basis is 100 MVA. the fault vector and the dictionary, by doing so the amount of
Direction of power flows are not relevant for the problem since data that will be transmitted is reduced by a ratio of k/N %,
the reference is going to be taken from the order of the faults in where k is the number of values replaced in the diagonal of the
de dictionary. From the diagram in Figure 1, line one runs from down-sampling matrix and N is the number of data points that
substation one to substation four, line two between substations make up the fault vector. According to this matrix we can
four and two, line three from substation one to three and line download the data traffic.
four from substation two to three.
The simulations of the faults were carried out using the D. Algorithm
software tool PowerWorld Student Version.
The algorithm will proceed to calculate a feature fault
training set and a feature predicted sample, this is the product
of the Dictionary and the Fault vector with the down-sampling
matrix, then the following equations are applied:
M’=JM (1)
Y’=Y (2)
Y’=M’X0 (3)
Where J is the down-sampling matrix, Y is the fault vector
and M is the training set.
Figure 1. Power System Model The sparse vector X0 is calculated and that is the marker
representation of the fault. The sparse vector has a distribution
IV. PROGRAM such that the first nine elements represent the first nine feature
faults in Line one, the following nine the faults in line two, and
A. Fault Training Set so. Thus the vector have 36 total values, most of them which are
A set of 36 faults was simulated for constructing the sparse, and can with relative accuracy estimate the fault vector.
Dictionary. Each fault will be simulated a different distances in Two main outcomes can be obtained. First the sensing
the line, considering a three phase balanced fault. Distances matrix and the final vector have a reduction in their size and
chosen were 3%, 10%, 20%, 30%, 50%, 70%, 80%, 90%, and the expression of the fault will have only two non-zero data
97% in each respective line. Thus, the dictionary originally will points, although only one if the fault occurs in one of the line
have dimensions of 960x36. distances used for the dictionary. Additionally it is known for
sure that the two markers i.e. positions in the sparse vector
The same approach can be taken using different faults (Line- must correspond to physical positions, thus the options of
to-Line, Single-Line) or unbalanced loads, an alternative model combinations in the sparse vector will be reduced notably since
including different fault types can be developed, but some the combinations that are close to each other are the ones to be
additional constraints will need to be added in order to avoid considered.
overlapping of faults of different characteristics but with similar
values.
V. SIMULATION RESULTS answer the closest and most accurate point possible, in this case,
the result is constantly converging to the section previous
Simulation was developed for the for different test faults to the fault, in this case it is Line 3 on the last section.
considering 100 runs for several values of data retained from
the original sample. This is in big part due to the proximity of the fault with a
substation, and the differences between the behavior of Line 1
The randomness of the down-sampling matrix creates a and Line 3 due to distribution and physical characteristics.
different set of data points each time, thus the new predicted
sample might have a different outcome each iteration, but the
most accurate solution will occur for positions closer to the The error parameters shown in Table 2 are considerably high,
fault marker in the sparse vector. The average result should and this cannot be accepted as an outcome since the system is
indicate this. working in per unit notation, thus the error in the order of units
might represent a variation of almost 100% in the actual voltage
Three measurements were considered, first the exactitude
values. This might have been somehow expected since the
of the algorithm in the prediction of the location of the fault,
the mean accumulated and per-data point error of the down- location of the fault is not accurate either and this result is the
sampled fault vector, and the mean accumulated and per-data algorithms closest solution within the boundaries of the
point error of the non down-sampled fault vector, i.e., the dictionary.
original fault vector versus the results obtained with the sparse
vector and the whole dictionary. L2
% Exact Next Prev. Other Total
Results are presented in the following tables:
90% 75% 1% 24% 0% 100%
L1
80% 65% 1% 24% 10% 100%
% Exact Next Prev. Other Total
75% 59% 3% 38% 0% 100%
90% 0% 0% 85% 15% 100%
60% 59% 0% 41% 0% 100%
80% 0% 0% 75% 25% 100%
50% 53% 1% 46% 0% 100%
75% 0% 0% 73% 27% 100%
40% 41% 3% 55% 1% 100%
60% 0% 0% 64% 36% 100%
25% 35% 1% 64% 0% 100%
50% 0% 0% 67% 33% 100%
10% 15% 0% 85% 0% 100%
40% 0% 0% 54% 46% 100%
25% 0% 8% 35% 57% 100% Table 3. Results for Test Fault 2 for Different Percentages of
10% 0% 0% 26% 74% 100%
data conserved
Table 1. Results for Test Fault 1 for Different Percentages of
data conserved Error 1 – Down-Sampled
Error 2 – Original Vector
Vector
Mean Mean Error Mean Mean Error
Error 1 – Down-Sampled
Error 2 – Original Vector % Accumulated per Data Accumulated per Data
Vector Error Point Error Point
Mean Mean Error Mean Mean Error
% Accumulated per Data Accumulated per Data 90% 0.07610 7.92708E-05 0.0703 7.32292E-05
Error Point Error Point
80% 0.12590 0.000131146 0.1297 0.000135104
90% 3.7674 0.003924375 0.4704 0.00049
75% 0.2009 0.000209271 0.1738 0.000181042
80% 5.03420 0.005243958 0.7568 0.000788333
60% 0.3009 0.000313438 0.2719 0.000283229
75% 4.9075 0.005111979 0.8246 0.000858958
50% 0.6538 0.000681042 0.4483 0.000466979
60% 6.89946 0.007186938 1.3547 0.001411146
40% 0.7578 0.000789375 0.564 0.0005875
50% 7.493 0.007805208 1.6094 0.001676458
25% 1.0687 0.001113229 0.782 0.000814583
40% 10.1166 0.010538125 2.1637 0.002253854
10% 1.4841 0.001545938 1.1003 0.001146146
25% 14.0385 0.014623438 3.038 0.003164583
Table 4. Error for Test Fault 2
10% 17.33740 0.018059792 3.7952 0.003953333
Fault 2 delivers and accuracy of 53% when using half of
Table 2. Error for Test Fault 1 the data point in the original Matrix and Fault Vector, and most
often tends to point to the previous section of line when the
Results for Test fault 1 seem to contradict the expected outcome amount of data conserved falls below 50%, with 75% the
of the algorithm, but after reviewing results for other faults, and program seems to have a value of about 60% of success.
considering that the fault 1 occurs at 1% of Line 1, which Position in fault 2 is of 40% in Line 2.
means it is outside of any of the regions contained in the Notable, the error for values of data conserved over 50%
dictionary for any Line, the algorithm tries to deliver as an are significantly small per Data Point, which is something that
we require because the magnitudes in the system are expressed L4
in per unit notation. This result contrasts with the error results % Exact Next Prev. Other Total
in fault 1, that even at the highest conservation percentage are
not small enough to guarantee a good approximation of the 90% 60% 0% 40% 0% 100%
fault. 80% 45% 0% 53% 2% 100%
L3 75% 34% 2% 40% 24% 100%
Exact Next Prev. Other Total 60% 36% 7% 51% 6% 100%
%
85% 3% 0% 12% 100% 50% 47% 7% 39% 7% 100%
90%
87% 3% 0% 10% 100% 40% 53% 9% 27% 11% 100%
80%
89% 3% 0% 8% 100% 25% 52% 8% 25% 15% 100%
75%
86% 1% 0% 13% 100% 10% 79% 4% 12% 5% 100%
60%
50% 89% 0% 0% 11% 100% Table 7. Results for Test Fault 4 for Different Percentages of
data conserved
40% 98% 0% 0% 2% 100%
25% 95% 0% 0% 5% 100% Error 1 – Down-Sampled
Error 2 – Original Vector
Vector
10% 93% 0% 0% 7% 100%
Mean Mean Error Mean Mean Error
Table 5. Results for Test Fault 3 for Different Percentages of % Accumulated per Data Accumulated per Data
data conserved Error Point Error Point
Error 1 – Down-Sampled 90% 0.03210 3.34375E-05 0.0333 3.46875E-05
Error 2 – Original Vector
Vector
Mean Mean Error Mean Mean Error 80% 0.10320 0.0001075 0.092 9.58333E-05
% Accumulated per Data Accumulated per Data 75% 0.16750 0.000174479 0.1306 0.000136042
Error Point Error Point
0.73610 0.00077 0.26160 0.00027 60% 0.04067 4.23646E-05 0.2821 0.000293854
90%
0.74160 0.00077 0.37230 0.00039 50% 0.96320 0.001003333 0.5303 0.000552396
80%
0.80550 0.00084 0.43170 0.00045 40% 1.04070 0.001084063 0.6334 0.000659792
75%
0.92170 0.00096 0.59360 0.00062 25% 1.96550 0.002047396 1.0473 0.001090938
60%
50% 0.87210 0.00091 0.64320 0.00067 10% 3.70710 0.003861563 1.7324 0.001804583
40% 0.86400 0.00090 0.71010 0.00074
Table 8. Error for Test Fault 4
Fault 4 brings the location of the fault to 85% of Line 4,
25% 0.95460 0.00099 0.92310 0.00096 and the results are incorrect all the way until 90% of data. Yet
10% 0.99840 0.00104 1.15720 0.00121 the result when wrong are consistently pointing out to the
previous location of the fault. This result is somehow similar to
Table 6. Error for Test Fault 3 the results found in previous cases.
Error is reduced big time from 50% of data conserved and
Fault 3 occurs on the 60% of Line 3 and again the accuracy up, which indicate that although the exact fault location might
of the program for a retention of data higher than 50% is of be wrong, the proximity is good enough to guarantee an
80% and up, whereas below that percentage of data the estimated vector that is close to the actual value.
algorithm tends to point out still to the right position, and the
previous spot as a solution on a reduced number of events. The
trend of pointing out either the exact location of the fault or the VI. CONCLUSIONS
previous spot is something that is repetitive in the three faults
analyzed so far, since in the first case the spot of the actual
fault does not exist in the training set, thus the previous spot is The challenge of finding an accurate way of
the closest the actual solution can be (or the next spot). representation for faults that does not require the use of the
full recorded data was not fulfilled successfully for all cases.
Error magnitudes are a tad bigger than in the previous case, The behavior of the faults in the system although similar in
but are still below the values in case 1. Still with a 50% of
samples retained the error is small enough to guarantee a close effects are not similar for data analysis, thus laying results that
estimate of the non-reduced fault vector. A remarkable from have notable differences between then.
this case is that in any case the solution is the right fault location The accuracy notwithstanding, is good enough for all
but taking the risk of a using smaller conserved data will imply cases of faults that occur within our training set and that’s a
a higher error in the estimation, with an order of good indicator that compressive sensing can be used as a way
10%, this error might be big enough to cause malfunction in to reduce the amount of data transmitted per fault, although
protection and monitoring systems. not always is good for prediction of the location of the fault.
However in cases 2 and 3 the accuracy was high, in 1 and REFERENCES
4 it was not, this might indicate that the methodology applied
in this project is good for faults that occur within the center of [1] LI Bei, HE Jinghan, YIP Tony. Wide Area Power System Fault
Detection Using Compressed Sensing to Reduc the WAN Data Traffics.
the lines but not that good for the faults occurring in the
[2] DING Wei, HE Ben-teng, WANG Hui-fang. Overview of wide-area
extremes or close to the extremes.
relaying protection system. Power System Protection and Control, 2012,
40(1).
Although there is always going to be a cut-off for the [3] XUE Yusheng, LEI Xing, XUE Feng. Review on Wide Area Protection of
efficiency of the algorithm, this is between the minimum and Electric Power Systems, High Voltage Engineering, 2012, 38(3): 513-
maximum points chosen for the dictionary, in actual systems, 520.
this is not an inconvenient since protection system are always [4] OUYANG Fan,LIU Haifeng,ZHAO Yongsheng. Analysis
designed to overshoot between 5% and 20% of their respective onCommunication Network Congestion Occurred in Smart Substation
andPreventive Measures. Power System Technology,2011,35(11):7-11.
region, thus no sacrifice in reliability is present when we discard
[5] DONG Xin-zhou, DING Lei, LIU Kun. System Protection Based
the application in 3% of the data in the extremes. onLocal Information.Proceedings of the CSEE,2010,30(22).
[6] TIAN Congcong,WEN Minghao. A Wide-Area Backup
Also, power simulations designs tend to work with a different ProtectionSystem With High Information Redundancy. Power
algorithm depending on the place where the fault occurs, and SystemTechnology,2011,35(10).
this algorithm focuses in faults that take place in the lines and
no in the bus/generators/loads. A 3% or less of proximity to the
next element will probably require the application of a new
algorithm, since the effect of the impedance in the
transformer/bar have a greater effect in the final behavior of the
fault.
Remarkably, even in the less accurate cases the system
pointed out at worst to the location next to the data point and
this result is a good start point for future work, since the
algorithm can be modified to correct the expected error (i.e.
point out to the previous spot) and improve the accuracy in
this way.
As final conclusion, the algorithm indeed recognize the
correct line where the fault is occurring in every case, and for
any percentage of conserved data over 10%, yet, the exact
location is more difficult to calculate. The location of the fault
can be developed as a region in the line, composed by the
exact location, the previous spot and the next spot, and in this
case the program tend to identify the right location, but not the
section of the line where the fault occurs.
