Download Artificial Intelligent Application to Power System Protection

Artificial Intelligent Application to Power System Protection
M.M. Saha
Substation Automation Division
ABB Automation Products AB
SE-721 59 Västerås, SWEDEN
Abstract: The application of Artificial Intelligence (AI)
methods in power system protection has been addressed in
this paper. Particular emphasis has been put on Artificial
Neural Networks (ANN) and Fuzzy Logic (FL). Several
novel concepts have been introduced including ANN application to CT and CVT transients correction, fuzzy criteria signals, fuzzy settings and multi-criteria decision making for digital relays. Attached examples illustrate application of ANN and FL techniques to resolve the selected relaying problems such as the fault classification or CT and
CVT dynamic error correction. Differential protection for
power transformers is selected as an important example to
show efficiency of the proposed concepts of FL and ANN
application.
I. INTRODUCTION
The microprocessor technology brings unquestionable
improvements of the protection relays- criteria signals are
estimated in a shorter time; input signals are filtered-out
more precisely; it is easy to apply sophisticated corrections;
the hardware is standardized and may communicate with
other protection and control systems; relays are capable of
self-monitoring. All this, however, did not make a major
breakthrough in power system protection as far as security,
dependability and speed of operation are considered. The
key reason behind this is that the principles used by digital
relays blindly reproduce the criteria known for decades.
The relaying task, however, may be approached as a pattern recognition problem - by monitoring its inputs, the relay
classifies on-going transients between internal faults and all
the other conditions. Or, the protective relaying may be considered as a decision making problem - the relay should decide whether to trip or retrain itself from tripping. This observation directly leads to AI application in power system
protection [1-4]. Practically, it includes the artificial neural
network approach (pattern recognition), as well as the expert
system and fuzzy logic methods (decision making) [5-8].
This paper briefly reviews the general problems and constraints in power system protection and presents the basics of
AI methods as applied to protective relaying. After general
presentation of the problem (Section II), a brief description
of the AI methods is given (Section III). Some examples of
the AI approach to power system protection are presented in
Section IV and V. The test results are also included in the
paper.
E. Rosolowski
J. Izykowski
Department of Electrical Engineering
Wroclaw University of Technology
50-370 Wroclaw, POLAND
II. PROBLEMS IN POWER SYSTEM PROTECTION
The problems result mainly from the trade-off between
the security demand (no false trippings), and the speed of
operation and the dependability (no missing operations) requirements. The more secure is the relay (both the algorithm
and its particular settings), the more it tends to misoperate or
operate slowly. And vice versa, the faster is the relay, the
more it tends to operate falsely. The problems listed below
reflect the current practice in power system protection.
There are basically two ways to mitigate the problem of
limited recognition power of the classical relaying principles. One of them is to improve and extend the measurements available to a given relay (for example, optical CTs
for improvement and substation integration for extension).
The second way is to improve the recognition process itself
based on what is already available and either:
• search for the new relaying principles, or
• apply several of known principles in one relay to improve the recognition, or
• apply correction of the CT and CVT transient error, or
• improve a type of fault determination by using of the
ANNs classifier, or
• use self-organizing algorithms such as ANNs to find out
automatically a protection principle.
It always takes certain time to estimate the criteria signals accurately enough to base the tripping decision on them.
Either they are measured fast or accurately. There is no perfect digital measuring algorithm that solves this well known
conflict between the speed and the accuracy. Either certain
pre-filtering is applied, or the basic algorithm uses longer
data window; or certain post-filtering is employed (or even a
combination of these three means). There is always a level
of uncertainty in the estimate of the criteria signal at the beginning of a disturbance when the relay operation is mostly
wanted. In some situations, although unprecise, the value of
the criteria signal enables solid decision, but is other cases,
such as a fault at the end of the protection zone, the relay
must wait for more precise estimate of the criteria signals.
III. ARTIFICIAL INTELLIGENCE METHODS
AI is a subfield of computer science that investigates
how the though and action of human beings can be mimicked by machines [5]. Both the numeric, non-numeric and
symbolic computations are included in the area of AI. The
mimicking of intelligence includes not only the ability to
make rational decisions, but also to deal with missing data,
adapt to existing situations and improve itself in the long
time horizon based on the accumulated experience.
Three major families of AI techniques are considered to
be applied in modern power system protection [1,5]:
• Expert System Techniques (XPSs),
• Artificial Neural Networks (ANNs),
• Fuzzy Logic systems (FL).
A. Expert Systems
The first expert systems included a few heuristic rules
based on the expert's experience. In such systems, the
knowledge takes the form of so called production rules
written using the If... then... syntax (knowledge base). The
system includes also the facts which generally describe the
domain and the state of the problem to be solved (data base).
A generic inference engine uses the facts and the rules to deduce new facts which allow the firing of other rules. The
knowledge base is a collection of domain-specific knowledge and the inference system is the logic component for
processing the knowledge base to solve the problem. This
process continues until the base of facts is saturated and a
conclusion has been reached (Fig.1). To guide the reasoning
and to be more efficient, these systems may incorporate
some strategies known as metaknowledge. Rule based systems represent still the majority of the existing expert systems.
There are few applications of XPS to power system protection reported, but all of them solve the off-line tasks such
as settings coordination, post-fault analysis and fault diagnosis [1]. As yet there is no application reported of the XPS
technique employed as a decision making tool in an on-line
operating protective relay. The basic reason for this is that
there is no extensive rule base that describes the reasoning
process applicable to protective relaying. Instead, only a few
rules or criteria are collected [9].
B. Artificial Neural Networks
Knowledge
Base
Inference
Engine
QHXURQ
V\QDSVH
OD\HU
LQSXWV
The ANNs are very different from expert systems since
they do not need a knowledge base to work. Instead, they
have to be trained with numerous actual cases. An ANN is a
set of elementary neurons which are connected together in
different architectures organized in layers what is biologically inspired (Fig.2) [5]. An elementary neuron can be seen
like a processor which makes a simple non linear operation
of its inputs producing its single output. A weight (synapse)
is attached to each neuron and the training enables adjusting
of different weights according to the training set. The ANN
techniques are attractive because they do not require tedious
knowledge acquisition, representation and writing stages
and, therefore, can be successfully applied for tasks not fully
described in advance. The ANN are not programmed or
supported by a knowledge base as are Expert Systems. Instead they learn a response based on given inputs and a required output by adjusting the node weights and biases accordingly. The speed of processing, allowing real time applications, is also an advantage.
Since ANNs can provide excellent pattern recognition,
they are proposed by many researchers to perform different
tasks in power system relaying for signal processing and decision making [2-5,7-8,10,12-13]. The common application
of the ANN technique assumes:
• The ANN is fed either with non-processed samples of the
input signals, or by features of those signals extracted
using certain measuring algorithms (or by a combination).
• The sliding data widow consisting of the recent and a
few historical samples of the signals, is fed to the ANN.
• The output from the ANN encodes the output decision
such as tripping command, type of fault, direction of
fault, etc.
• The training patterns exposed to the ANN cover the most
important operating conditions both internal faults and
other disturbances. Typically, only the selected window
positions are used for training.
• Additional pre- and post-processing may be applied.
The most widespread application of ANNs is in pattern
classification and associative memory where they can learn
to distinguish between classes of inputs and, therefore, they
can be successively used for decision making and phenomena classification. A major problem with ANNs is that no
exact guide exists for the choice to the number of hidden
layers and neurons per hidden layer. On the other hand, the
ability to generalize is one of the main advantage of using
ANNs.
RXWSXW
Solution
Data
Base
Fig.1. Simplified block diagram of an XPS.
Fig.2. Typical three-layer architecture of a feed-forward ANN.
C. Fuzzy Logic
TABLE I. COMPARISON OF AI METHODS IN POWER SYSTEM PROTECTION
With reference to Fig.3 the fuzzy logic approach to protective relaying assumes [6]:
• The criteria signals are fuzzified in order to account for
dynamic errors of the measuring algorithms. Thus, instead of real numbers, the signals are represented by
fuzzy numbers. Since the fuzzification process provides
a special kind of flexible filtering, faster measuring algorithms that speed up the relays may be used.
• The thresholds for the criteria signals are also represented by fuzzy numbers to account for the lack of precision in dividing the space of the criteria signals between the tripping and blocking regions.
• The fuzzy signals are compared with the fuzzy settings.
The comparison result is a fuzzy logic variable between
the Boolean absolute levels of truth and false.
• Several relaying criteria are used in parallel. The criteria
are aggregated by means of formal multi-criteria decision-making algorithms that allow the criteria to be
weighted according to their reasoning ability.
• The tripping decision depends on multi-criteria evaluation of the status of a protected element. Additional decision factors may include the amount of available information, or the expected costs of relay maloperation.
The XPS, ANN and FL approaches have their own advantages and limitations. Table I compares the basic features
of these AI methods.
IV. APPLICATION TO POWER TRANSFORMER
PROTECTION
As an examples authors have considered fuzzy logic and
ANN application to differential transformer protection. The
differential relaying principle in the case of a power transformer shows certain limitations - detection of a differential
current does not provide a clear distinction between internal
faults and other conditions. Inrush magnetizing currents,
stationary overexcitation of a core, external faults combined
with saturation of the CTs and/or CTs and protected transformer ratio mismatch are the most relevant phenomena
which may upset the current balance causing the relay to
maloperate.
A. Fuzzy Logic application
To enhance the performance of transformer protection a
multi-criteria fuzzy logic based relaying frame may be used
Fuzzyfication
Decision
Making
Defuzzyfication
Fig.3. Simplified block diagram of the fuzzy logic approach.
Feature
Knowledge
used
Troubleshooting and
improving a
relay
Self-learning
Handling unclear cases
Robustness
Setting a
relay
XPS
Expert knowledge in the form
of rules, objects,
frames, etc.
Change of rules
required.
Possible.
Possible.
Not-critical and
easy to ensure.
Convenient.
Computations Extensive.
Approach
ANNs
Information
extracted from
the training set
of cases.
Difficult - the
internal signals
are almost impossible to interpret.
Natural.
Natural.
Difficult to
ensure.
Large number
of simulation
required.
Dedicated
hardware.
FL
Expert knowledge
in the form of protection criteria.
Convenient - the
internal signals are
understandable and
analyzable
Possible.
Natural.
Not-critical and
easy to ensure.
Convenient. Both
knowledge and
simulation are used.
Moderate.
[1,6]. The relay presented in [11] uses 12 protection criteria
which indicate the following transformer states (Fig.4):
• Magnetizing inrush
• Stationary overexcitation of a core
• External faults combined with saturation of the CTs
• External faults without saturation of the CTs.
The relay uses fuzzy settings and multi-criteria decision
making methods and is self-set using the EMTP simulations.
The selected example concerning the 140/10.52kV, 5.86
MVA three-phase two-winding transformer demonstrates
both the relay stability and sensitivity. Fig.5 displays the differential and through currents as well as the tripping signal
for the turn-to-turn internal fault occurring after 50ms of
transformer energizing and involving 16% turns of the Yside winding on the column S. The relay activates at the beginning of energizing but remains blocked during inrush
conditions. The tripping command is sent 16ms after the
fault inception.
B. ANN application
The paper [10] presents a wide optimization process involving the type of an ANN as well as both pre- and postprocessing algorithms for the power transformer protection.
M
E
A
S
U
R
I
N
G
U
N
I
T
w1
w2
w3
Ruling-out the hypothesis of
stationary overexcitation
Ruling-out the hypothesis of an
external fault combined with
saturation of CTs
Ruling-out the hypothesis of an
external fault combined with ratio
mismatch
MIN
δ>∆
∆
Fig.4. Simplified block diagram of the considered fuzzy logic relay
for power transformers.
Fig.5. Fuzzy Logic based relay operation under sample turn-to-turn fault occurring during energizing of the transformer.
With reference to Fig.6, a neural network relay (NNR)
consists of three basic units: a pre-processor, an ANN itself
and a post-processor. In the case of a power transformer, a
differential relay measures the currents on all the sides
through the Current Transformers (CTs) - and sometimes also the voltages through the Voltage Transformers (VTs).
These signals are pre-filtered using the analog anti-aliasing
filters and next sampled (in this paper 2nd order RC circuit
with cut-off at 350Hz; sampling at 1kHz). The differential
and restraining currents are next formed according to the art
of differential relaying for power transformers.
Three different levels of pre-processing have been developed and tested: - NNR without essential pre-processing;
- with separation of 12 criteria signals; and - with using of
settings obtained from comparison of the criteria signals
with their appropriate thresholds.
In the first case an ANN is fed just by samples of differential and restraining currents and operates as a protective
relay for power transformer. The length of the sliding data
window and the set of employed signals are the only parameters to be optimized in such an approach.
In the second approach a NNR may combine the advantages of the ANN technology with the expert knowledge
having the form of relaying principles. The relaying criteria
(such as 2nd harmonic restraint) enable to convert the natural relaying signals into the space of features to be fed into
an ANN itself. In the paper, the feature universe consists of
8 criteria signals (per phase), achieved by applying 12 pro&7V
&7V
&%
&%
97V
1(85$/ 1(7:25.
75,33,1* 6,*1$/
35(352&(66,1*
3267352&(66,1*
Fig.6. The basic configuration of a Neural Network based Relay
(NNR) for power transformers
tection criteria (as in fuzzy logic approach).
In the criteria-based approach, the pre-processor converts
the natural relaying signals into 8 features. These signals, are
always real positive and either their low or high values indicate on an internal fault. These criteria signals are next fed
into an ANN. Only the most recent samples are forwarded
(no sliding data window).
To analyze logic signals produced by comparison of an
appropriate settings, instead of using an output logic circuit,
an ANN based classifier is employed. Each criterion is aptimized individually prior to usage with the objective to
minimize the percentage of missing and false indications as
well as the provided average identification time [11]. No
sliding data window is applied - only the most recent samples are fed.
Two methods was analyzed for determination of the final
decision (post-processing) of the relay: natural postprocessing and filter application. Since only two classes of
disturbances are relevant in power system protection (internal faults and other conditions), ANNs operating as protective relays are usually trained to respond with two distinctive
values, respectively. Thus, to analyze the output from an
ANN only the thresholding is needed. Such kind of postprocessing was denoted as natural post-processing. Instead
of this also filters was used: mean-value filter (averaging of
the result) and median filter.
For all considered cases the feed-forward three-layer
fully interconnected sigmoidal ANNs was used. Two basic
configurations of a NNR of a three-phase power transformer
was tested: with single ANN for each transformer phase and
with ANN observing the three phases. The actual number of
training patterns presented to ANNs count in tens of thousands - due to the three-phase structure of a protected transformer and the sliding data window of a NNR.
Wide comparative analysis with different NNR structure
has been provided. Majority of the developed NNRs handle
well the special testing cases [10].
V. APPLICATION TO CT AND CVT CORRECTION
Certain construction limitations of the instrument transformers may in some cases cause maloperation or substantial delay in tripping of the protective relays.
One of the method for correction of CTs saturation con-
L QVQ
a)
Z
3ULPDU\ VHFRQGDU\ DQG FRPSHQVDWHG YROWDJH
100
SULPDU\ YROWDJH
L Q1 VQ
Z
L Q1 VQ
F
L Q
G
VHFRQG YROWDJH
v, V
Z
FRPS YROWDJH
0
Z
-100
L Q1 VQ
F
20
G
F
v, V
G
1
1
1
100
120
140
0
50
60
70
b)
sists in application of CT’s inverse transfer function in the
form of ANN [8]. The correction function and the transfer
function of CT set up in series should assure identity of CT
primary and compensated secondary currents. Since the
CT’s transfer function is nonlinear, usage of the nonlinear
artificial multilayer neural network structure with some form
of feed-back (recurrent network), as presented in Fig.7, is
required. The sigmoidal tangent activation function has been
assigned to neurons in hidden layers and the linear one to the
output neuron of the selected ANN architecture.
The sliding data widow consisting of the recent and a
few historical samples of the signals, is fed to the ANN. Rescaled samples of the CT current are putted on the input
register: iw ( n − N + 1) , ... iw (n ) . ANN output signal
ic ( n − N d ) represents corrected CT current.
Result of compensation the secondary CT current is presented in Fig. 8. Secondary current waveform was obtained
from the EMTP simulation of 3-phase fault at the substation
of 400 kV system [8]. It can be seen that the proposed ANN
corrector almost perfectly reproduces the primary current.
Current magnitude (Fig. 8b) is estimated according to the
full-period Fourier method.
3ULPDU\ VHFRQGDU\ DQG FRPSHQVDWHG FXUUHQW
90
100
110
120
130
140
2
1
0
20
40
60
time, ms
80
100
120
140
Fig. 9. Results of CVT transients error compensation during
phault-to-ground fault at the substation
Very generally, the dynamics of a CVT is determined by
two factors:
• nonlinear oscillations under saturation of magnetic core
of the CVT step-down transformer,
• discharging of the CVT internal energy during short circuits on an associated transmission line.
The greater influence has the second source of transient error. Especially, faults at zero crossing of the primary voltages result in substantial transient errors that, in turn, affect
the operation of supplied relays.
The idea of CVT transients error compensation presented in [12] is based on finding of an inverse transfer
function of the CVT model and reproducing it with an ANN
- what is similar to the approach used for CT correction.
Therefore, the proposed ANN corrector has general structure
as in Fig. 7.
Different ANN sizes and input/feedback connections
have been tested and analyzed. As an example, Fig. 9 presents result of CVT correction during phase-to-ground fault at
the 400 kV substation. The fault is applied when the primary
voltage of the faulted phase crossing zero.
L
80
0DJQLWXGH HVWLPDWLRQ
V, V
Fig.7. ANN structure for CT correction
$
80
-20
40
D
60
20
L Q1 1 VQ
F
40
SULPDU\ FXUUHQW
VHFRQGDU\ FXUUHQW
FRP FXUUHQW
VI. APPLICATION TO FAULT TYPE
CLASSIFICATION
60
$PSOLWXGH HVWLPDWLRQ
E
,
$
WLPH PV
Fig. 8. Plot of phase S currents generated from simulation of 3phase fault at the substation (RF=0Ω). Compensation carried out by
ANN of 5-5-1 structure (a). Amplitude estimation of primary, secondary and compensated current (b).
Fast fault detection and classification of the fault type is
one of the most important task of the protection relays and
fault location function. The basic idea of the fault type estimation consists in analysing of the phase and zero-sequence
voltages and currents. The ANNs have good pattern recognition and classification feature and that is what is here expected.
The proposed neural fault type estimator (NFTE) consists of 4 neural networks: three recurrent nets for particular
phase fault detection and the fourth feedforward one for fault
to ground recognition [13]. The NFTE uses feature vectors
formed by V (voltage) and I (current) trials. The architecture
_
_
XQ5
[2]
_
XQ5− _
< 7Q
5
Q − 1 3K + _
_X
5
_L
Q _
_L
_L
5
Q− _
Y Sn
[3]
Y Rn
$113K
5
Q − 1 3K + _
[4]
67
,Q
[5]
]
67
,Q
=
6
PD[_ LQ _ 6
7
_ LQ − _ _ LQ _ 7
_ LQ − _ [6]
Fig. 10. Recurrent Neural Fault Estimator for a faulty phase
selection
of the NFTE for a faulty phase selection is sketched in
Fig.10. The ANNs are free layer networks with activation
functions of both hidden layers of hyperbolic tangent type
and linear functions in output neurons.
The nets work in parallel indicating faulted phases (the
nets ANN-Ph - Fig.10) and eventually fault to ground
events (the net ANN-G - not presented in Fig.10). Changes
of outputs of particular ANN-Ph classifiers from -1 to 1 indicate fault detection in scanned phase and changes of output of fault to ground detector inform about faults as follows: R-G, S-G, T-G, R-S-G, R-T-G or S-T-G. The decision
threshold in both detectors equal to 0 has been chosen.
Introducing only absolute values of voltage and current
samples at the input (Fig.10) reduces number of different
patterns to be analysed. Usage of ANNs with the feedback
connection makes the output signal from ANN-Ph more stable and the decision taken more reliable [13].
[7]
[8]
[9]
[10]
VII. CONCLUSIONS
The paper reviews the AI approaches to power system
protection and focuses on the application of ANN and fuzzy
logic techniques.
A number of novel application and concepts have been
presented including fuzzy logic approach to differential
transformer protection and ANN application to the transformer protection, CT and CVT transients correction, and.
fault-type classification. Included examples demonstrate application of the AI methods and their features.
REFERENCES
[1]
International Journal of Engineering Intelligent Systems, The special issue on AI applications to power
system protection, edited by M.M. Saha and B.
Kasztenny, Vol.5, No.4, December 1997, pp.185-93.
[11]
[12]
[13]
Dalstein T., Kulicke B., Neural network approach to
fault classification for high speed protective relaying.
IEEE Transactions on Power Delivery, Vol.10, No.2,
April 1995, pp.1002-1009.
Sidhu T.S., Singh H., Sachdev M.S., Design, implementation and testing of an artificial neural network
based fault direction discriminator for protecting
transmission lines. IEEE Transactions on Power Delivery, Vol.11, No.2, April 1995, pp.697-703.
Sanaye-Pasand M., Malik O.P., Performance of a recurrent neural network-based power transmission line
fault directional module. Eng Int Syst (1997) 4, pp.
221-228.
Warwick K., Ekwue A. And Aggarwal R. (ed). Artificial intelligence techniques in power systems. The Institution of Electrical Engineers, London, 1997.
Saha M.M, Kasztenny B., Application of fuzzy logic in
power system protection, International Conference
'Modern Trends in the Protection Schemes of Electric
Power Apparatus and Systems', 28-30 October 1998,
New Delhi, paper IX-1.
Bachmann B., Novosel D., Hart D., Hu Y., Saha M.
M., Application of artificial neural networks for series
compensated line protection, Proc. of the Int. Conf. on
Intelligent System Application to Power Systems, Orlando, January 28 - February 2, 1996, pp.68-73.
Lukowicz M., Rosolowski E., Artificial neural network
based dynamic compensation of current transformer
errors. Proceedings of the 8th International Symposium on Short-Circuit Currents in Power Systems,
Brussels, 8-10 October 1998, pp. 19-24.
Wiszniewski A. and Kasztenny B., Primary protective
relays with elements of expert systems, Proceedings of
the 1992 CIGRE Session, Paris, France, August 1-5,
1992, Paper 34,2,CN.
Kasztenny B., Lukowicz M., Rosolowski E., Selecting
type of a neural network, pre- and post-processing algorithms for power transformer relaying. Proceedings
of the 32nd Universities Power Engineering Conference, Manchester, UK, Sept. 10-12, 1997, vol. 2, pp.
708-712.
Kasztenny B., Rosolowski E., Saha M. and Hillstrom
B., A self-organizing fuzzy logic based protective relay
- an application to power transformer protection, IEEE
Transactions on Power Delivery, Vol.12, No.3, July
1997, pp.1119-27.
Lukowicz M., Rosolowski E., Izykowski J., ANN application to CVT transient errors compensation.
3U]HJOG Elektrotechniczny, R.LXXV 9/1999, s. 224229 (in polish).
Lukowicz M., Rosolowski E., Fault type classification
in high voltage power systems using artificial neural
networks. Proceedings of the 11th International Conference on Power System Protection PSP'98, Bled,
Slovenia, 30. Sept. - 2. Oct. 1998, pp. 141-146.