Simplified Artifwid Neural Network Structure with the Current Transformer Saturation Detector

Provides a Good Estimate of Primary Currents

J. C. Cummins D. C.Yu Lj. A. Kojovic

University of Wisconsin-Milwaukee Cooper Power Systems
Milwaukee, Wisconsin Franksville, Wisconsin
USA USA

Abstract: This paper presents use of artificial neural Several approaches have been proposed for providing estimates
networks (ANN) to correct current transformer (CT) of original waveforms from data obtained from distorted
secondary currents distortions caused by the CT saturation. waveforms caused by the CT saturation. Kalman filters and
The ANN is trained to achieve the inverse transfer function iterative methods have been active areas of research [ 1,2].
of uoncore toroidal CTs, which are widely used in
protective systems. The ANN has been designed as a Kalman filters and iterative methods require development of
simplified structure to minimize use of memory when approximate models of the CT core, and an estimate of flux
implemented in protective devices. To properly estimate levels within the core. Flux levels are used to estimate
primary currents for a saturated transformer, a current magnetizing currents. Magnetizing currents are related to
transformer saturation detector has been added to the errors in CT output due to saturation. Models are non-linear
ANN. The ANN is developed using the MATLAB" and difficult to implement.
program and trained using data generated from actual CTs.
The ANN calculating speed i n d accuracy are satisfactory An alternative approach is to use an artificial neural network
in real-time applications, and provides good estimates of (ANN) to approximate a function that corrects the distorted CT
primary currents. secondary currents caused by CT saturation.

Keywords: Artificial Neural Networks, Current The ANN presented in paper [3] was developed to handle large
Transformers, Saturation, Relays, Protective Equipment dynamic ranges of fault currents using a technique of
employing two sets of network coefficients. Different sets of
coefficients deal with different fault current ranges. The
L Introduction algorithm for running the network was implemented on an
Analog Device ADSP-2101 digital signal processor. The
Iron-Core Current Transformers calculating speed and accuracy were satisfactory in real-time
applications. The ANN was developed using the MATLABm
Iron-core toroidal current transformers are widely used for program and trained using data from EMTP simulations and
measurement and protection in power systems. Their data from actual CTs.
advantages include low cost, galvanic isolation, reliability,
and ease of application. Disadvantages are saturation and The ANN presented in this paper was designed as a simplified
remanent flux in the core. structure to minimize use of memory when implemented in
protective devices. To properly estimate .true (primary) currents
To avoid C T s a t t '" ,it is possible to increase the size of for a saturated transformer, a current transformer saturation
the CT core or use a core material that supports large flux detector was added to the ANN. The ANN was developed using
densities. Both options can affect current transformer cost the MATLAB program and trained using data generated from
and application. actual CTs. The ANN calculating speed and accuracy are
satisfactory in real-time application and provides good
In designs where C T s are incorporated as an integral part estimates of primary currents.
of protection package, dimensions of C T s are determined
by the available space for installation and clearance Section 11 of this paper discusses performance of current
distances to provide adequate insulation. transformers during transient conditions. Section III describes
the ANN design, training, structure optimization, and testing.
Conclusions are presented in Section IV.

0-7803-6420-1/00/$10.00 (c) 2000 IEEE 1373

 .................~ ~ -..............
"- "_.. ......................
II. Performance of Conventional Current I ,

Transformers
When iron-core C T s are used in protective equipment, the
saturation problem may become critical. Most protective
devices, both electromechanical and electronic, make
operating decisions based on fault current RMS values. If
the signal supplied by the CT to the protective equipment
is distorted by its saturation (Figure l), RMS values will be
lower than RMS values of the actual fault currents. Figures
2a and 2b compare relay operation times for non-saturated 4

and saturated CTs. When C T s do not saturate, secondary

currents accurately replicate primary currents (Figure 2a). ._.___ -f ~ "_

1
s i

In this case, relay operation time was 22 ms. Figure 2b , sewdarYcA1: . TalY1064
shows delayed relay tripping if CT saturation is present .... ............ / . ~ . , . _ . , _ _ _ .
under identical test conditions. The relay operation time
was 80 ms. In an actual application, this can cause mis-
coordination with other relays. When CT saturation is
higher, it can even prevent relay tripping. This type of
problem has been widely discussed in the literature as it
applies to various protection schemes [4-71.

Figures 3a and 3b show secondary current harmonic

components for waveforms from Figures 2a and 2b. Since T i ~rl
most digital relays reject DC components, Figures 3a and
(b)
3b also include harmonic components without the DC Figure 2. Comlparison of a Relay Operation (a) without and
component. (b) with CT Saturation
Most phasor-based protective devices estimate current
magnitudes correctly within one cycle. When C T s do not
saturate, they accurately transform the RMS value of ihe
primary current to the secondary. Figure 3a shows the
RMS value of the secondary current for a non-saturated
CT. The protective device operated, in this case, in 22 ms.
When C T s saturate, the RMS value of the secondary
current is smaller than the primary current (Figure 2b) and
the protective device does not "see" the actual current
magnitude. It delays operation until the current increases.
In this case, device operation extends to 80 ms.

------- (a) -l_._.._-r-...__..__ .....

.................
.. ... ... . .

4- ....
RdayTrip
1
0 3 ..... McomponmbmrpIDC . '.. ......... 4t
.. ~ .,.- . - / . . . . - , .... ' 1i
. - , . . . . ( .. . ... i
~,
' TatY1064 1t
i-~.C
-l--.1
-i-
0 om ow 0.06 0.08 0.1 0.12
Time [SI

Figure 1. Primary and Secondary Currents for a (b)

Figure 3. Harmonic Components of (a) Primary Current and
Saturated CT
(b) Secondary Current for a Saturated CT

(c) 2000 IEEE

0-7803-6420-1/00/$10.00 1374
 Another example of relay delayed tripping times caused by vector at the top of the network y. This output is the estimate of
CT saturation is shown in Figure 4. The CT primary the CT primary current. In between the input and output layers
current was 3 A (referred to the secondary). The RMS is a hidden layer connected to the input and output by two
value of the secondary current decreases with an increase matrices wl and w2. These two matrices are analogous to the
in CT saturation. The solid curve represents normally synapses found in biological systems.
inverse TCC curve. When the CT does not saturate, the
relay operates in 0.7 s for 3 A sinusoidal secondary current. The processing of the network consists of the operations of
When the CT saturates, although the primary current has multiplying the input vector by the first weight matrix wl,
the same magnitude, the relay delays tripping depending on adding a constant bias term B, and applying a non-linear
the level of 0 saturation. Figure 4 shows delayed tripping sigmoid function f to the resulting hidden layer vector. The
for three different levels of CT saturation. hidden layer vector is multiplied by the second weight matrix
w2 and the resulting single element vector is operated on by the
same non-linear function. The non-linear function limits the
dynamic range of the elements.

To apply an artificial neural network for correction of CT

saturation, the optimal number of elements in the network and
the values of the matrices must be determined. Like a
biological system, this is accomplished through a learning
process.

I I 1
2 3 4 5
Secondary Omrent [A]
Figure 4. Relay Response Times to Different Levels of
the CT Secondary Current Distortion due to
CT Saturation

III. Artificial Neural Networks

Design

Artificial neural networks am a technology that mimics the

biological neurons and synapses of the brain. Like the Xb) x(n.1) Sa) 0 0 0 0 x(n4l)
brains, they have the ability to learn and to form complex
responses to input stimuli. Figure 5. The Artificial Neural Network Structure

The unique feature of ANN'S exploited for CT saturation

correction is input-output .mapping, or function Training and the ANN Structure Optimization
approximation. It has been demonstrated that a relatively
simple ANN structure can be used to approximate any For this application the method of supervised learning is used
continuous function [a]. The way this feature is exploited is to determine values of the weight matrices. Supervised learning
a number of data samples from the CT secondary current means that the network is presented with sets of input data, and
are used to generate an estimate of the primary current. the values of the weight matrices are adjusted to minimize the
error between the network output and the desired output. The
The A" structure presented in this paper is a single method used to adjust the weights is known as the back-
hidden layer feed-forward. The structure consists of three propagation algorithm.
vectors and two matrices as shown in Figure 5 . The signal
inputs from the CT secondary are presented at the bottom If the error cannot be reduced to an acceptable level for the set
vector x. This vector is analogous to a set of biological of training data, the size of the network must be increased. The
nerve cells that receive a stimulus from the environment. number of elements in the input or hidden layers and the
The signal output is extracted from the single element corresponding dimensions of the weight matrices needs to be

(c) 2000 IEEE

0-7803-6420-1/00/$10.00 1375
 increased until the network can be trained. As in the 6 shows the input waveform to the ANN block and the
biological world, a complex task requires a larger brain. corrected signal, which represents an estimation of the primary
In order for an ANN to be implemented in a protective current. The diita is not smooth due to a low sampling rate.
device, the size of the network must be small enough to
perform the estimation in real time. The matrix
multiplication and application of the non-linear function
can impose a high computational burden on the types of
embedded processors commonly used in protective
devices. In addition to the computational burden, weight
matrices must be stored in memory that is often limited in
protective devices.

To keep the size of the network small, the set of training

data can be restricted to the range of currents producing
significant saturation. This range depends on the CT being
used and the associate burden.

To keep the benefit of a reduced network size produced by Samples

training the ANN only for the CT saturated waveforms and
Figure 6. Input and Output Signals from the ANN Block
yet preserve the correct operation of the protective device
obtaiined during Virtual Testing
for non4aturated waveforms, an algorithm for detecting
saturation was developed.
Real-Time Power System Simulator (RTPSS)Testing. The
In the protective device, the CT secondary current is
RTPSS was U:& to obtain training data from actual C T s and to
processed by saturation detection algorithm prior to being
test the ANIN performance. The RTPSS is a computer-
processed by the ANN. If the waveform is determined to
controlled, analog-based device employing high fidelity voltage
be saturated, the saturation is corrected by the ANN.
and current amplifiers. It has the ability to simulate a variety of
Otherwise the ANN is bypassed and the waveform is input
system configurations and faults. A block diagram shown in
directly to the protection algorithms.
Figure 7 represents setup used for the CT saturation testing. A
large number of system disturbances can be easily generated.
To train the network to correct for CT saturation, actual
Each disturbance type and location can be quickly varied for all
data from C T s used in protective systems was obtained.
inception angles. The RTPSS has models for most all power
Data covered the whole expected range of the CT
system elements including voltage and current transformers [9].
operation, including primary and secondary current
waveforms. This data is known as the training set.

The training set was presented to the ANN one case at ai

time, and the weights were adjusted to minimize the error..
This was repeated until the error reached an acceptable
level for the training set.

Artificial neural networks have the ability to generalize.,

that is to respond correctly to inputs not part of the training
set. However, to assure proper operation, the set of training
cases included data to cover extremes of expected
operation. In other words, intemolation is more easily
accomplished than extrauolation. R o k d i v e De*
Digital R ~ O I ~ R unler Ted

ANN Testing Figure 7. Block Diagram of the Real-Time Power System

Simulator (RTPSS) and the Protective Device
Virtual Testing. To evaluate initial performance of the Testing
ANN, virtual testing was performed prior to analog testing.
Test data was fed to the ANN implemented in the: Two identical protective devices were tested in series to “see”
protective device CPU directly from a PC containing EL the same current during test. One device implemented ANN,
digitized recorded waveform of CT shturation. Output from while the other did not (i.e. without the ANN).
the ANN block was then recorded during playback. Figure

0-7803-6420-1/00/$10.00 (c) 2000 IEEE 1376

 ANN operation was tested in the full range of expected CT influence protective device operation. Both devices operated as
operation. The protective device was connected to the expected.
RTPSS through a 1000/1 A CT. Testing was performed
with different fault current magnitudes, fault closing
angles, and power system X m ratios.

Testing also included verification of the ANN operation for

saturated and non-saturated CTs as well as its operation
with and without the CT saturation detector.

Testing the A N N Operation without the CT Saturation

Detector

Figure 8 shows a test result for a fault current of 3 kAws.

During this test, the CT did not saturate. The expected
operating time of the protective device for this fault was
182 ms. The ANN, which was not trained for these types of
[ - : I : :
0 0.02
-_
0.04
.-......... .."- .
0.06
L.
0.08
.,.
0.1
..
0.12
i

0.14
current waveforms and fault levels, initiated 80 ms T h e [SI
operation. The standard device operated as expected. This
demonstrated the requirement that saturation must be Figure 9. Response of the Protective Device with the ANN and
detected for the reduced sized network Since protective with the CT Saturation Detector to a Fault that did
device operation is not impaired by saturation, the ANN not Saturate the CT
should be bypassed in this situation. This was
accomplished by the CT saturation detection algorithm.
Saturated CT. Figure 10 shows the response of the protective
devices tested with 7 k A ~ current.
s The CT heavily saturated
during this test. The expected operating time was 23 ms. With
the ANN, even under these extreme conditions, the device
operated in 39 ms. The standard device, without the ANN,
operated after 124 ms. The saturation detection algorithm
detected CT saturation leading to the secondary current being
processed by the ANN, improving the response time.

0 0.05 0.15 0.2 035

n m 151
Figure 8. Response of Protective Device with ANN to a
Fault that did not Saturate CT without the CT -I0 i I
Saturation Detector

Testing the ANN Operation with the CT Saturation

Detector
n m bl
Non-Saturated CT. Figure 9 shows protective device Figure 10. Response of the Protective Device with the ANN
response times for a test representing 7 k A ~ fault
s current. and with the CT Saturation Detector to a Fault that
During this test, the CT saturated very slightly. The Saturated CT
expected operating time was 23 ms. The CT saturation
detector properly determined the CT was not saturated and In addition to improving response times, the ANN also
by-passed the ANN. The CT sattiration detector was improves recording fault current magnitudes used by protective
designed to activate at CT saturation levels high enough to device event recorders. In the previous test case, the value of

0-7803-6420-1/00/$10.00 (c) 2000 IEEE 1377

 the fault current, with the ANN was 5959 A. Without the system conditions based on simplified models of the magnetic
ANN, only 2738 A was recorded. circuit.

Figure 11 shows waveforms for primary currents and CT References

secondary currents distorted due to CT saturation. The
same figure also shows corrected CT secondary currents Kang, Y.C.; Park, J.K; Kang, S.H; Johns, A.T.; Aggarwal,
corrected by the ANN. The corrected waveform is very R.K. “Algorithm for compensating secondary currents of
close to the original signal (primary current). current transformers.” IEEE Transactions on Power
Delivery Vol. 12, No. 1, Jan. 1997.
Chen, EL. W.; Glad, S. T. “ Estimation of the primary
current in a saturated transformer,” in Proceedings of the
IEEE Conference on Decision and Control, part 3 (of 3)
Dec. 11 -13 1991, pp. 116-124.
10. C. Yu, J. C. Cummins, Z. Wang, H. J. Yoon, Lj. A.
Kojovic, “Neural Network for Current Transformer
Saturation Correction”, IEEEJPES T&D, April 1999.
Barnett, C.W.; Ghadwick, J.W. et al “Relay Performance
consideirations with low ratio CTS and high fault currents.’’
IEEE Transactions on Power Delivery Vol. 8, No. 3, July
1993, PI?. 884-887.
M. Kezunovic, Lj. A. Kojovic, C. W. Fromen, D. R.
Sevcik, F. Phillips, “Experimental Evaluation of EMTP-
0.02 0.04 0.06 0.08, 0.1 0.12 0.14 Based Current Transformer Models for Protective Relay
[SI
Figure 1 1 . The ANN Correction of the CT Secondary Transient Study”, 93 W M 041-4 PWRD, IEEWES
Current Distortion caused by the CT Winter Meeting, New York, 1993.
Saturation Bozoki, B.; Calhoun, C.M.; et al “Gapped Core Current
Transformer Characteristics And Performance.” IEEE
Transactions on Power Delivery vol 5 num 4 November
Iv. Conclusions 1990 ~€11732-1740.
Wright, A. Current Transformers, Their Transient and
This paper presents use of artificial neural networks (ANN) Steady !State Performance. London, Chapman & Hall 1968,
to correct current transformer (CT)secondary waveform pp. 77.
distortions. Haykin, Simon. “ Neural Networks, A Comprehensive
Foundalion”, IEEE Press New York, MacMillian 1994,
The ANN has been designed as a simplified structure to pp. 181.
minimize use of memory when implemented in protective Lj. A. Kojovic, T. R. Day, “Application of Real-Time
devices. To properly e s h t e primary currents for a Power System Simulators for Testing Protective Relay
saturated transformer, a current transformer saturation System Operational Characteristics”, IEEERES T&D,
detector has been added to the ANN. April 1999.

The ANN was developed using MATLABTMand trained Biograpbiai

using data from actual C T s .
James C. Clummins (M’92) is a research staff engineer in the
ANN calculating speed and accuracy are satisfactory in Department of Electrical Engineering and Computer Science at
real-time applications and provides good estimates of the University of Wisconsin - Milwaukee.
primary currents.
David C. U (M.84) is a professor in the Department of
!
[
In addition to improving operating times, the recording of Electrical Engineering and Computer Science at the University
fault magnitudes used by protective device event recorder of Wisconsi:n - Milwaukee.
is improved.
Ljubomir A. Kojovic (SM’94) currently holds the position
One advantage of this technique is that it produces a Chief Power System Engineer for Cooper Power Systems at the
sample-by-sample estimate of primary current. Another Thomas A. 1 3 s o n Tcchnical Center.
advantage is that the ANN can be trained with data from an
actual system. This avoids making approximations of

0-7803-6420-1/00/$10.00(c) 2000 IEEE 1378