IEEE TRANSACTIONS ON POWER DELIVERY 1

Classification of Electrical Disturbances

in Real Time Using Neural Networks
Iñigo Monedero, Carlos León, Member, IEEE, Jorge Ropero, Antonio García, José Manuel Elena, Member, IEEE,
and Juan C. Montaño, Senior Member, IEEE

Abstract—Power-quality (PQ) monitoring is an essential service TABLE I

that many utilities perform for their industrial and larger com- TYPES OF DISTURBANCES
mercial customers. Detecting and classifying the different electrical
disturbances which can cause PQ problems is a difficult task that
requires a high level of engineering knowledge. This paper presents
a novel system based on neural networks for the classification of
electrical disturbances in real time. In addition, an electrical pat-
tern generator has been developed in order to generate common
disturbances which can be found in the electrical grid. The classi-
fier obtained excellent results (for both test patterns and field tests)
thanks in part to the use of this generator as a training tool for the
neural networks. The neural system is integrated on a software tool
for a PC with hardware connected for signal acquisition. The tool
makes it possible to monitor the acquired signal and the distur-
bances detected by the system.

transform.
Pr E
Index Terms—Neural networks, power quality (PQ), wavelet

I. INTRODUCTION
E
f
P OWER QUALITY (PQ) is usually defined as the study of
oo
the quality of electric power signals. In recent years, grid
users have detected an increasing number of drawbacks caused
by electric PQ variations [1] and PQ problems have sharpened
IE
because of the increased number of loads sensitive to PQ and
have become more difficult to solve as the loads themselves have
become important causes of the degradation of quality [2]. Thus,
these days, customers demand higher levels of PQ to ensure the
proper and continued operation of such sensitive equipment.
The PQ of electrical power is usually attributed to power-line
disturbances, such as waveshape faults, overvoltages, capac-
itor switching transients, harmonic distortion, and impulse tran- on the side of their customers. The classification and identifica-
sients. Thus, electromagnetic transients, which are momentary tion of each disturbance are usually carried out from standards
voltage surges powerful enough to shatter a generator shaft, can and recommendations depending on where the utilities operate
cause sudden catastrophic damage. Harmonics, sometimes re- (IEEE in the U.S., EN in Europe, etc.). Our own classification,
ferred to as electrical pollution, are distortions of the normal which is given in Table I, is based on the UNE standard in Spain
voltage waveforms found in ac transmission, which can arise at which defines the ideal signal as a single-phase or three-phase
virtually any point in a power system. While harmonics can be sinusoidal voltage signal of 230 and 50 Hz. In addition,
as destructive as transients, often the greatest damage from these we complete from our point of view some aspects which were
distortions lies in the loss of credibility of the power utilities not completely defined in this standard.

II. ARTIFICIAL INTELLIGENCE APPLIED TO PQ

Manuscript received February 16, 2006; revised September 8, 2006. This
work was supported by the Spanish Ministry of Science and Technology
Artificial intelligence (AI) may be broadly defined as the au-
(MCYT: Ministerio de Ciencia y Tecnología) under Project Reference Number tomation of activities that are associated with human thinking,
DPI2002-04420-C03-03. Paper no. TPWRD-00075-2006. such as decision-making, problem-solving, learning, percep-
Color versions of one or more of the figures in this paper are available online tion, and reasoning [3] for the resolution of complex problems.
at http://ieeexplore.ieee.org.
Thus, these days, because of the important development of
Digital Object Identifier 10.1109/TPWRD.2007.899522 computers, AI tools [2] (e.g., artificial neural networks (ANNs)
0885-8977/$25.00 © 2007 IEEE
2 IEEE TRANSACTIONS ON POWER DELIVERY

have emerged as an interesting alternative for the resolution of In multiresolution analysis, the signal is decomposed in a set
problems which require some kind of human reasoning. of approximation wavelet coefficients and another set of detail
In the classification of electrical disturbances, all of the fac- wavelet coefficients. The obtained approximation coefficients
tors that make AI (and, in particular, ANNs) a powerful tool are are, in turn, decomposed in order to increase the level of res-
present. We get information which is massive—electrical sig- olution.
nals are constantly being received—and distorted—there is an The detail coefficients of the lowest levels store the infor-
important noise component—so that a classification of the dis- mation from the fastest changes of the signal while the highest
turbances (which can sometimes be highly complex) must be ones store the low-frequency information. Thus, with the help
carried out. of new mathematic tools, the detection of the electrical distur-
Thus, new and powerful tools for the analysis and operation of bances tends to be easy but the classification is still a difficult
power systems, as well as for PQ diagnosis are currently avail- task in which ANNs play an important role [5]–[12].
able. An excellent overview of these tools is presented in this From that extraction, the problem is a question of pattern
paper [3]. Thus, the main intelligent tools of interest include ex- recognition by means of ANNs. Thus, one of the most impor-
pert systems, fuzzy logic, and ANNs. Expert systems are expen- tant tasks is to generate an adequate number of training patterns
sive in their development and normally slow in their execution in order to train the ANNs correctly so that they can classify fu-
(therefore, they are not good for real-time applications). On the ture inputs appropriately. In particular, in PQ, a great number
other hand, fuzzy-logic schemes are useless for the classifica- of these electrical patterns are necessary due to multiple com-
tions of PQ disturbances due to the fact that they are not well binations of different disturbances which can coincide in one
suited for fuzzy techniques (the distinction among the various or various samples. Another additional problem with ANNs ap-
types of disturbances is discretely defined by standards or rec- plied to PQ is the impossibility of obtaining real useful training
ommendations). Therefore, the schemes of most extended use patterns directly from the power grid due to the irregular appari-

Pr E
are based on ANNs (besides, they have the advantage of being
very recommendable for real-time use due to their low time con-
sumption).
The detection and classification of electric disturbances
requires the preprocessing of data, feature extraction, and final
classification. Thus, in order to extract the signal features,
E tion of these disturbances and the difficulty in capturing them.

III. NN REAL-TIME CLASSIFIER

We have developed a prototype of a real-time system for
the detection and classification of electrical disturbances. The
system is a detector of power-line disturbances whose classifi-

f
ANNs are usually combined with mathematical analysis, such cation kernel is based on AI techniques (particularly on ANNs).
as Fourier and wavelet transforms for the generation of signal The system consists of a PC application which includes an ac-
features which serve as inputs of the network [4]. Signal quisition card, an environment to monitor the acquired signal,
oo
evaluation consists of both spectrum and transient analysis.
The Fourier transform is most commonly used for spectrum
analysis. However, transients are commonly analyzed by means
and an AI kernel to classify possible disturbances. On the other
hand, an electrical pattern generator has been developed as an
auxiliary tool to generate electric patterns for the ANNs.
IE
of digital wavelet transform. It is hardly feasible to localize and
estimate transients by means of the Fourier technique. A. Environment
Although both mathematical analyses (Fourier and wavelet) The environment of the application (Fig. 1) shows the infor-
can be used to generate good signal features, the wavelet al- mation which is acquired and registered by the system. It con-
gorithm has an important advantage when it is a necessary to sists of several windows in which the acquired signal is repre-
quickly analyze the signal (and, therefore, for real-time sys- sented by means of the of the three phase signals, and
tems) due less computational complexity. Thus, whereas the the neutral signal. Other windows show the last detected distur-
complexity of the wavelet algorithm is lineal, the Fourier one is bance, a bar diagram which reports the number and the type of
O(NlogN). The advantages of this type of scheme are opposed to detected disturbances and a historic window which registers the
deterministic methods (for instance, threshold detectors) which date and time of the different events.
are the possibility of classifying successful signals with more We also have other options such as a bar diagram reporting
of a disturbance (by means of accurately training the ANNs) as a temporal graphic view of the disturbances, a more detailed
well as better performance in the classification of those types representation of the last detected disturbance—those are shown
of disturbances in which the measurement of their thresholds is in Fig. 2—or a three-phase diagram and representation of the
highly complex (for instance, in frequency disturbances). Be- signal.
sides, Wavelet analysis is capable of detecting events of data The acquisition card obtains 640 samples every 100 ms. This
that other analysis tools would miss, including trends, break- window width and number of samples proved to be sufficient for
down points, and discontinuities. our aim in the range of detection of disturbances. AUTHOR:
The aim of the preprocessing of the signal in wavelet-ANNs TABLE 4 AND FIGS. 3, 4, AND 8 NEED TO BE CITED IN
schemes is to obtain a feature extraction which provides a TEXT
unique characteristic which can represent every single PQ dis- The acquired samples are shown on the chart and processed
turbance. Thus, for instance, it can be carried out by means of by the AI kernel. When one or more disturbances are detected
a wavelet analysis in different resolutions using the technique within these 100 ms, the corresponding registers are updated,
called multiresolution signal decomposition or multiresolution changing the corresponding windows for the last disturbance,
analysis. the bar diagrams, and the historic.
MONEDERO et al.: CLASSIFICATION OF ELECTRICAL DISTURBANCES 3

For the design and programming of the tool environment, the

tool selected was the Borland C++ Builder 5, which is a pow-
erful tool for the development of visual applications as well as
a robust C++ compiler.

B. Electrical Pattern Generator

An aforementioned important problem with ANNs applied
to PQ is the impossibility of getting real patterns directly from
the power grid due to the irregularity in the apparition of distur-
bances and the problems in their identification. In turn, it is very
difficult to obtain access to databases of electrical disturbances
and, once achieved, to adjust the features of the obtained elec-
trical patterns to the requirements of the system to be designed
(in parameters such as sample time and voltage range defined
for each type of disturbance).
Thus, for the task of training ANNs for the detection and clas-
Fig 1. Classifier environment (I). sification of electrical disturbances, we developed an electrical
pattern generator (EPG) which is patent pending. An unlimited
number of patterns to be used by a classification system can be
generated. This EPG is based on a PC environment which has
been programmed on LabView. The EPG makes it possible to
Pr EE configure parameters, such as the duration of the sample, the
frequency of the signal, and the number of samples in an ideal
cycle (50 Hz or 60 Hz) and to add one or more disturbances on
the display window. From the selected parameters, the gener-
ator creates a text file with the voltage values of the sample. The
structure of the file is simple and consists of a header with the

f
file information (name, number of sample cycles, and sampling
oo period) and three data columns corresponding to the samples of
the three-phase voltages.
On the other hand, the possibility of generating these patterns
as real signals has been added in order to be able to simulate
IE
other detection equipments. Thus, the signal is generated by a
digital-to-analog converter (DAC) card connected to the PC and
then amplified from the output voltage level of the card to a
Fig 2. Classifier environment (II). high-voltage level ( 1500 volts) able for testing actual
loads.
The types of disturbances include impulse, oscillation, sag,
swell, interruption, undervoltage, overvoltage, harmonics, and
frequency variations. Parameters of the amplitude disturbances
(impulses, sags, swells, interruptions, undervoltages, and over-
voltages), such as amplitude, start time, final time, and rising
and falling slope can be configured by the EPG. The edition of
harmonics allows for the configuration of amplitude and phases
as far as 40 harmonic order including the possibility of adding
an offset.
We have generated over 27 000 signal files including one-dis-
turbance signals and two-disturbance signals, and tried to sweep
all types of disturbances.

C. Detection and Classification Kernel

The classification process carried out by the kernel involves
two steps: a feature extraction of the acquired signal and the
ANNs analysis and detection of signal disturbances.
For the feature extraction, the kernel uses a wavelet mul-
Fig 3. Electrical pattern generator. tiresolution analysis of the acquired signal for the generation
of signal features [5]–[11]. The aim of feature extraction by
4 IEEE TRANSACTIONS ON POWER DELIVERY

Pr E
Fig 4. Some examples of disturbances generated from the EPG.
E
f
wavelet transforms is to provide a unique characteristic which
can represent every single PQ disturbance.
For the design of several feature vectors, we have carried out
oo
a study of the variation of the detail coefficients of a wavelet
transform (which stores each level of decomposition the detail
information of the analyzed signal) for the different kinds of
IE
disturbances as well as the variation of these coefficients to pos-
sible noises found in the signal. The aim of this study was to
characterize the signal reaching one or various feature vectors
which, in turn, required a low calculus level in order to fulfil the
real-time specification of the tool.
The study was divided into four parts relating to the three
main types of disturbances to be characterized (voltage distur- Fig 5. Comparison of wavelet detail levels 5, 6, and 7 in a 50-Hz signal (upper
bances, frequency deviation, and harmonic components)—see row) and a 53-Hz signal frequency deviation (lower row).
Table I—and to noisy signals. The procedure of the studies
involved the comparison of the wavelet detail coefficients of
the disturbances with the wavelet detail coefficients of an ideal ysis levels because of the wide variety in the duration of these
voltage signal from the standard (sinusoidal signal of 230 V and kinds of disturbances (trying to capture slow variations with
50 Hz). Thus, we start with the Haar mother wavelet because it high levels and fast ones with low levels).
is the fastest of the Daubechie’s family (speed was very impor- Fig. 5 shows the difference between detail levels 5, 6, and 7 in
tant for real-time applications). Level seven was chosen to be a 50-Hz signal and the same levels in a 53-Hz signal (high-fre-
convenient for the wavelet analysis of the voltage signal. quency deviation). As may be observed, in these last levels, we
Through a study of the variations of the different detail coef- can distinguish the frequency disturbance as much in its wave-
ficients in each level, we proved that the most adequate levels forms as in the range of their amplitude. In level 6, the detection
for characterizing the harmonics and frequency deviations were in the frequency change of the waveform of the signal is clearly
levels 4, 5, 6, and 7. It made sense because these last wavelet observed.
levels are more adequate for detecting low-frequency signal The next step consisted of the characterization of the above
variations. mentioned waveforms by means of an adequate feature vector
For the voltage disturbances, we chose levels 1, 2, 6, and 7 (with the least possible number of inputs) detecting the ampli-
in order to capture both the amplitude components and the fre- tude variations of the selected detail coefficients relative to the
quency variations. In this case, we selected high and low anal- waveform of that same detail coefficient from a signal without
MONEDERO et al.: CLASSIFICATION OF ELECTRICAL DISTURBANCES 5

disturbances. It is important that the detection was independent

of the location of the disturbance within the window (i.e,, the
feature vector corresponding to a selected sag must be the same),
independent of its starting time within the window. Thus, the
complexity in the classification process was simplified to a large
extent.
In order to achieve the above mentioned objectives, the fea-
ture vectors (which would be used as inputs for the ANNs) were
generated carrying out a number of operations on the different
wavelet levels. In particular, we used the following values for
each level of the detail wavelet coefficients.
• The integral of the waveform of the coefficients.
• The maximum of the absolute values of the coefficients.
• The of the waveform of the coefficients.
These values were used as an abstract of the different wave- Fig 6. Block diagram.
forms of the detail levels; in other words, a limited and conve-
nient set of values which were able to identify a wavelet wave-
form in a unique way. Thus, the integral of the waveform gives turbance—or 1—disturbance—. If there is a disturbance, the
the energy amount of the signal. The maximum absolute value ANN inputs are given to three other ANNs, each one special-
of the coefficients identifies possible peaks of the waveform izing in the detection of a class of disturbances. In the same way,
and the carries out a calculus of both amplitude and fre- the outputs of these ANNs are 0 or 1, depending on whether
there is or not that kind of disturbance. These classes are, ac-
Pr E
quency. On the other hand, these values were selected due to
the fact that they are time-independent. In other words, they
do not vary in function of as to which localization of the an-
alyzed waveform extracted was found to be the disturbance per
second. In this way, many patterns were saved for training and
the process was optimized.
E cording to Fig. 1, voltage disturbances—sags, swells, undervolt-
ages and overvoltages, frequency disturbances—slight and se-
vere deviations, and harmonics disturbances.

D. Structure of the NNs

f
Therefore, three values were used by each level. In addition to Important features in the design of ANNs are the study of the
the previously mentioned parameters, for each level of wavelet necessary input pattern, the ANN structures, transfer functions,
detail coefficients, we used the of the sampled signal as and learning algorithms.
an input for every ANN.
oo
In particular, for the previously mentioned second step in the
classification process, we first designed three ANNs: one for
As mentioned in the previous section, we have used the fol-
lowing values as input vector of the disturbance and voltage
ANNs: the of the signal, the integral, the maximum of
IE
amplitude disturbances, another for frequency deviation, and the absolute values, and the of the detail wavelet coeffi-
a third one for harmonics. The main reason for using several cients of level 1, 2, 6, and 7. On the other hand, as inputs for
ANNs is the different architecture of the three main types of dis- frequency and harmonics networks, we used the of the
turbances: voltage, frequency, and harmonics (resulting in dif- voltage signal, the integral, the maximum of the absolute values,
ferent wavelet levels used for the classification of each distur- and the of the detail wavelet coefficients of level 4, 5, 6,
bance type). Besides, a unique ANN with seven outputs—one and 7. In order to obtain faster convergence and better results,
for each type of disturbance—requires too many neurons to these data were scaled so that the minimum was 1 and the
work properly and, consequently, more memory resources. maximum was 1.
Later, we noticed the advantage of designing one new ANN Moreover, the kind of ANN chosen is a multilayer perceptron
(besides those designed for classification) which was able to de- with three hidden layers with a different number of neurons,
tect, with a very high success rate, the appearance or absence depending on the ANN and its number of outputs. A typical
of any type of disturbance in a signal section. Thus, as shown three-layer ANN used for the classification process is shown in
in Fig. 6, the classification system would use this new ANN Fig. 7. The output functions of the layers have been chosen with
(called disturbance ANN) as a filter to detect signals with distur- a logarithmic sigmoid transfer function for all of the layers.
bances and three parallel ANNs to classify the different distur- All of the inputs, structures, functions, and training algo-
bances existing in the signal. The advantages of this filter net- rithms have been obtained after testing the different ones. The
work (which should have a high degree of reliability in the fil- best results until now have been obtained for the ANNs shown
tering process as a basic condition) were that it avoids unneces- in Table II.
sary analyses in the rest of the classification networks, saving We designed two sets of training and test patterns: the first
analysis time and possible errors in the other three networks set had 11501 training patterns and 2914 test patterns whereas
(whose success rate would be doubtlessly lower because of their the second one had 20 254 training patterns and 5088 test sig-
higher number of outputs). nals. The first set included a higher number of signals without
Therefore, the classification process, which is represented in disturbance than the second set (which included a high number
Fig. 6, carries out the following steps: first of all, the inputs are of signals with one or more disturbances). Thus, the disturbance
given to the disturbance ANN, whose output is either 0—no dis- ANN (designed to distinguish between a disturbed signal and a
6 IEEE TRANSACTIONS ON POWER DELIVERY

Once we carried out the test and found a good code for the
preprocessing and the AI kernel, we reprogrammed their algo-
rithms in C++ to optimize the execution time (taking care with
the programmed wavelet transform algorithms which were the
most complex algorithms in the kernel).
As mentioned in Section III, the AI kernel receives 640 sam-
ples every 100 ms. This window width and number of samples
proved to be sufficient for our aim in the range of detection of
disturbances. In addition, two different analyses are carried out
every 100 ms in this way.
This scheme implies that the 640 samples in every 100 ms
must be analyzed twice. It was applied to achieve the detection,
for instance, of a 10-ms sag centered at the end of an acquisition
window. If analyzed as two independent parts (two mini-sags of
5 ms) it would be not consider a sag in the UNE standard (which
defines a sag as lasting at least 10 ms) and, therefore, it would
Fig 7. Three-hidden layers perceptron.
be ignored.
Thus, the time requirements of our system had to achieve the
following terms:

Pr EE In other words, the kernel time and the time for monitoring
the acquired signal and the detected disturbances have to be less
with a margin (thus ensuring the time requirements against pos-
sible delays of the operating system) than the time relative to the
acquisition process of the 640 new samples to analyze. More-

f
Fig 8. Time analysis process of the samples. over
oo
TABLE II
NEURAL-NETWORK STRUCTURE
IE

That is to say, the classification time in the worst conditions

(three-phase disturbed signals) is equal to three times (for each
phase) the preprocessing and the 4 ANNs delay and that result
nondisturbed one and, therefore, requiring a greater balance be- is multiplied by 2 (for each analysis window).
tween both types) was trained using the first set and the rest of We carried out the measurement of the previous times in C++
the ANNs using the second set (designed to classify the kinds code with the following results:
of disturbance in an a priori disturbed signal extract). milliseconds and
E. Time Analysis milliseconds
A last step in the design of the detection and classification The total time in the worst conditions was thus below 46 ms
neural system involved its integration into a real-time environ- which involved a wide margin of 54 ms (taking into account that
ment. First, we had to optimize the different calculus carried the window time was 100 ms).
out in the detection and classification process and then we had
to make sure that the time consumed met the time requirements IV. RESULTS
of our system.
Thus for programming tasks, we have used the MATLAB A. System Simulation
tool to test the different possibilities in the preprocessing of the Before embedding the kernel in the classifier tool, we selected
signal and in the structure of the kernel. We used this tool due to the best training method for the configuration of ANNs. Thus,
the powerful toolboxes it contains with specialized functions, for the training of the networks, we used 80% of the generated
utilizing the signal and the wavelet toolboxes for the prepro- signals as training patterns and 20% as test patterns. On the other
cessing task and the ANNs toolbox for the design of the kernel hand, thresholds were defined in the ANN outputs in order to
[12]. determine whether a particular output value may be considered
MONEDERO et al.: CLASSIFICATION OF ELECTRICAL DISTURBANCES 7

TABLE III TABLE V

DETECTION AND CLASSIFICATION RESULTS TWO-DISTURBANCE SIGNAL RESULTS

TABLE IV TABLE VI
ERRORS IN ANALYSIS FIELD TEST RESULTS

Pr E
as a disturbance. The defined thresholds were 0.3 and 0.7 and,
thus, output values above 0.7 were considered as disturbances
E
f
and below 0.3 as ideal signals. Values found between 0.3 and
0.7 were taken as errors in the detection of the input pattern. near the limit of a disturbance, so it is not a mistake to consider
The distance between the output network and the desired value
oo
was defined as a safety coefficient in the detection.
Two different kinds of results were considered: 1) the gen-
eral results, that is to say, the percentage of success in every
them as the ANN tells us.
Table V shows the results obtained for two-disturbance sig-
nals, with approximately 27 700 signals. About 5400 (the total
IE
for both sets of patterns) were for testing and the rest were for
ANN—see Table III and 2) the particular results, which are training.
more intuitive and consider some particular cases of failure in Conclusions were similar to those obtained for the case of
one of the ANNs. In addition, we have the results for only two- one-disturbance signals although slightly worse. It was due to
disturbance signals (signals with two kinds of disturbances in the greater complexity of the signals and to the fact that the
the same signal window). ANNs had to correctly detect two types of disturbances simul-
The first conclusion we obtain is that the higher number of taneously. Even so, the results remained above 89% in all of the
ANN outputs we use, the higher number of errors we get. This networks, which is considered an excellent result.
is due to the higher complexity introduced by the need to fix all
of the outputs at the same time. In addition, the number of output
B. Connection to the Electrical Pattern Generator
combinations in the voltage ANN (which makes it possible to
detect several voltage disturbances at the same time) is higher For these kinds of field tests, the designed detector equip-
than, for instance, the frequency network (in which both outputs ment was connected to the signal generator (digital-to-analog
cannot be active at the same time). The second conclusion is converter card and amplifier equipment) integrated on the elec-
related to the influence of these errors. As mentioned before, the trical pattern generator.
most important ANN is the one which detects disturbances—the Once each test was configured and generated, the detector
existence of a disturbance is more important than its type—so equipment was programmed to perform the detection and classi-
we have focused our efforts on its correct working. On the one fication of the signals and to show the results obtained for every
hand, we must say that these percentages all refer to the testing ANN. These results are shown in Table VI.
signals but we also have to bear in mind that some of the signals Bearing in mind that the degree of accuracy required in the
that fail are filtered by the disturbance ANN. On the other hand, classification process due to the wide range of signal types as
we have to analyze what kind of signals tend to fail. To illustrate well as the degree of complexity of many of them, 75% of
this point in the following table, we show the results of a study correct classifications—and 89% adding those signals classi-
of 10 samples referring to signals which fail in the voltage ANN. fied as partially correct—appeared as a highly satisfactory re-
Analog results were obtained for the rest of the signals and for sult. Besides, it is necessary to consider that it is always pos-
the other ANNs. As may be observed, the signals which fail are sible to improve these results by means of new trainings of the
8 IEEE TRANSACTIONS ON POWER DELIVERY

ANNs (adding as training patterns the previously mentioned er- been accomplished with very satisfactory results: pattern identi-
rors and other new electrical patterns generated with the EPG). fication was validated and tested with excellent performance (fit
The theory of ANNs proves that the greater the number of pat- of simulated and actual data above 89%). Moreover, results are
terns, the better the success rate of an ANN will be. For this always easily improvable through the generation of new signal
reason, a higher success rate would be reached (which could patterns.
even be 100%). The use of signal preprocessing is quite simple; the program-
The obtained results are inside the reliability degrees reached ming and optimization in C++ of the different algorithms made
by the professional equipment. On the other hand, there are not it possible to achieve the objective of making our system valid
contractual obligations in the UNE standard for these degrees in real time. This makes it possible to detect signal disturbances
and, therefore, it guarantees the possibility of integrating the in time to avoid further problems.
system in a first prototype of electrical disturbance detector and
classifier equipment.
REFERENCES
[1] P. Daponte, M. di Penta, and G. Mercurio, “Transientmeter: A dis-
C. Connection to a Low-Voltage Electric Line tributed measurement system for power quality monitoring,” in Proc.
9th Int. Conf. Harmonics and Quality of Power, 2000, vol. 3, pp.
These types of field tests were carried out by connecting 1017–1022.
the detection equipment to a low-voltage electric line also [2] W. E. Kazibwe and H. M. Sendaula, Expert Systems Targets. Power
connected to professional equipment for the detection of elec- Quality Issues. New York: IEEE, Apr. 1992, 0895-0156/92.
[3] W. R. Anis Ibrahim and M. M. Morcos, “Artificial intelligence and
trical disturbances. The tests were made in two net points in advanced mathematical tools for power quality applications: A survey,”
weekly periods and were of use to refine the set of ANNs. IEEE Trans. Power Del., vol. 17, no. 2, pp. 668–673, Apr. 2002.
This refinement was motivated by the existence of several high [4] G. Zheng, M. X. Shi, D. Liu, J. Yao, and Z. M. Mao, “Power quality dis-
turbance classification based on rule-based and wavelet-multi-resolu-
THD components, but within the standards. The process of
Pr E
refinement consisted, therefore, in the utilization of these signs,
classified incorrectly as “out of standard” by the voltage ANN,
as new patterns for our training universe.
However, none of the systems detected any disturbance in
these tests, so a battery of tests was made to connect the system
E tion decomposition,” presented at the 1st Int. Conf. Machine Learning
and Cybernetics, Beijing, China, Nov. 4–5, 2002.
[5] A. Elmitwally, S. Farghal, M. Kandil, S. Abdelkader, and M. Elkateb,
“Proposed wavelet-neurofuzzy combined system for power quality vi-
olations detection and diagnosis,” Proc. Inst. Elect. Eng., Gen. Transm.
Distrib., vol. 148, no. 1, pp. 15–20, Jan. 2001.
[6] P. K. Dash, S. K. Panda, A. C. Liew, B. Mishra, and R. K. Jena, “New
approach to monitoring electric power quality,” Elect. Power Syst. Res.,

f
to a disturbance generator system. vol. 46, no. 1, pp. 11–20, 1998.
This battery of tests was programmed for a long duration in a [7] A. K. Ghosh and D. L. Lubkeman, “The classification of power system
wind power plant which will predictably generate a high number disturbance waveforms using a neural network approach,” IEEE Trans.
oo
of disturbances. It will make it possible to obtain and publish
new results and, in turn, to study peculiarities in the disturbances
generated by this kind of plant.
Power Del., vol. 10, no. 3, pp. 671–683, Jul. 1990.
[8] J. V. Wijayakulasooriya, G. A. Putrus, and P. D. Minns, “Electric
power quality disturbance classification using self-adapting artificial
neural network,” Proc. Inst. Elect. Eng., Gen. Transm. Distrib., vol.
IE
149, no. 1, pp. 98–101, Jan. 2002.
[9] R. Daniels, “Power quality monitoring using neural networks,” in Proc.
1st Int. Forum Applications Neural Networks Power Syst., 1991, pp.
V. CONCLUSION 195–197.
[10] S. Santoso, J. P. Edward, W. M. Grady, and A. C. Parsons, “Power
These days, it is known that ANNs are a good choice for de- quality disturbance waveform recognition using wavelet-based neural
tecting and classifying electrical power disturbances. In the ex- classifier—Part 1: Theoretical foundation,” IEEE Trans. Power Del.,
vol. 15, no. 2, pp. 222–228, Feb. 2000.
tensive literature [3]–[12] available on the detection of electrical [11] S. Santoso, J. P. Edward, W. M. Grady, and A. C. Parsons, “Power
disturbances, the problem often lies in generating a sufficient quality disturbance waveform recognition using wavelet-based neural
number of training patterns for the ANN to obtain good results classifiers-Part 2: Application,” IEEE Trans. Power Del., vol. 15, no. 1,
pp. 229–235, Feb. 2000.
in future inputs. With the help of the generator, it is possible to [12] M. Mallini and B. Perunicic, “Neural network based power quality
carry out a training of the ANN which is sufficiently complete analysis using MATLAB,” in Proc. Large Eng. Syst. Conf. Power Eng.,
to provide reliable results. Halifax, NS, Canada, 1998, pp. 177–183.
Thus, we have developed a first prototype of a real-time clas-
sifier system based on ANNs. In addition, we have developed
an electrical pattern generator which is capable of generating Iñigo Monedero studied computer science and then was with the Automatics
common disturbances which can be found in the grid with the and Robotics Department for two years.
aim to make the training of ANNs easier. In turn, the generator Currently, he is Assistant Professor in the Electronic Technology Department
of the University of Seville. and in the Electronic Technology Department, he
makes it possible by means of an analog-to-digital converter and is conducting research in the field of artificial intelligence.
an amplifier to generate the configured signal physically.
The advantages of using the generator were the possibility of
easily generating a great number of training patterns with the
Carlos León (M’95) received the physical electronics and computer science
electrical pattern generator in order to obtain perfect training as doctoral degrees from the University of Seville, Seville, Spain, in 1991 and
well as the possibility to carry out the field tests of the detector 1995, respectively.
and classifier. Currently, he is Professor of Electronic Engineering at the University of
Seville, where he has been since 1991. His areas of research are expert systems,
The aim of this project, that is to say, the development of a neural networks, data mining, and fuzzy logic, focusing on utility systems
real-time detector and classifier of electrical disturbances has management.
MONEDERO et al.: CLASSIFICATION OF ELECTRICAL DISTURBANCES 9

Jorge Ropero is an Assistant Professor for the Department of Electronic Tech- José Manuel Elena (M’81) received the physical electronics and a computer
nology at the University of Seville, Seville, Spain, where his special investi- science doctoral degrees from the University of Seville, Seville, Spain, in 1977
gation issues include mainly those related to artificial intelligence, especially and 1979, respectivey.
neural networks and fuzzy logic. He has been Professor of Electronic Engineering at the University of Seville
since 1991. His areas of research are expert systems, neural networks, and fuzzy
logic, focusing on digital communications system management.

Antonio García was born in Seville, Spain, in 1960. He received the physical
electronics degree from the University of Seville, Seville, Spain, in 1982.
He has been Professor of Electronic Engineering in the Electronic Technology Juan Carlos Montaño (SM’00) was born in Sanlúcar (Cádiz), Spain. He
Department since 1984. His areas of research are instrumentation, hardware de- received the Ph.D. degree in physics from the University of Seville, Seville,
sign, and digital signal processing. Spain, in 1972. From 1973 to 1978 he was a Researcher at the Instituto de
Automática Industrial (CSIC—Spanish Research Council), Madrid, Spain,
working on analog signal processing, electrical measurements, and control of
industrial processes.
Since 1978, has been responsible for various projects in connection with re-
search in power theory of nonsinusoidal systems, reactive power control, and
power quality at the IRNAS (CSIC).

Pr EE
f
oo
IE