You are on page 1of 6

A scheme based on PMU data for power quality

disturbances monitoring

Arturo Mejia-Barron, Juan P. Amezquita-Sanchez, Jose R. Razo-Hernandez and David Granados-


Aurelio Dominguez-Gonzalez, and Martin Valtierra- Lieberman
Rodriguez* ENAP-RG, CA-Fuentes Alternas y Calidad de la Energía
ENAP-RG-Facultad de Ingeniería, Universidad Autónoma Eléctrica
de Querétaro, Campus San Juan del Río, ITESI- Departamento de Ingeniería Electromecánica,
Río Moctezuma 249, Col. San Cayetano, 76807, San Juan Instituto Tecnológico Superior de Irapuato (ITESI), Carr.
del Río, Querétaro, México. Irapuato-Silao km 12.5, Colonia El Copal, C. P. 36821,
* martin.valtierra@enap-rg.org Irapuato, Guanajuato, México

Abstract—Power quality (PQ) monitoring has attracted the (DFT) [8], wavelet transform (WT) [9], S transform [10],
interest of many researchers around the world. PQ disturbances Kalman filter (KF) [11], artificial neural networks (ANNs) [6],
(PQDs) such as sags, swells, interruptions, harmonics, notching, and empirical mode decomposition (EMD)-based methods
spikes, and oscillatory transients, among others, have to be [12], among many others, have been presented in literature,
detected and classified in order to apply a proper solution. For where important results have been obtained. Yet, these
their detection and classification, many methodologies in techniques may be affected by several aspects, compromising
literature have been proposed, where a signal processing their performance. For instance, DFT presents excellent results
technique and a pattern recognition algorithm are typically used. for stationary signals; however, when the signal has non-
In this work, a new methodology for detection and classification
stationary features such as the presence of some PQDs, the
of PQDs using a phasor measurement unit (PMU)-based scheme
results of DFT are degraded. WT offers a time-frequency
is presented. In general, the processing of voltage or current
signals is carried out using a phasor estimation model (algorithm representation, but their results depend on a proper selection of
within a PMU), whereas the classification task is performed by both the level of decomposition and the mother wavelet, which
threshold-based rules and an artificial neural network. The is not a straightforward process by considering the diversity of
proposal is validated using synthetic signals. Then, it is tested PQDs. KF and ANN usually use harmonic-based models to
using real measured signals. Results demonstrate its effectiveness perform a tracking of a voltage or current signal. In this regard,
and usefulness. the tracking accuracy depends on the selected harmonic
components and the used updating algorithm/parameter. It is
Keywords— electric disturbances; phasor measurement unit; evident that an a priori selection is sometimes difficult, as a
power quality; neural network. priori information is not available. EMD method has been
proposed as an adaptive technique, i.e., the method could
I. INTRODUCTION directly be used regardless the type of signal. Despite providing
promising results in different fields of science, EMD method
Over the past few years, the topic of power quality (PQ) has suffers of mode-mixing, which can lead to false or wrong
attracted the interest of many researchers around the world [1]- decompositions; besides that, issues such as the computational
[2], as a poor PQ negatively affects utilities and users with time and the inherent requirements of being a batch processing
damages in equipment. In general, voltage or current signals technique may compromise the online PQ monitoring.
have a poor PQ when their nominal values, e.g., amplitude, Therefore, having in mind all the aforementioned ideas, the
phase angle, or frequency, change. These abnormal conditions, development and application of efficient and reliable
typically named PQ disturbances (PQDs), can be categorized approaches in terms of complexity and performance are still
into sags, swells, interruptions, harmonics, notching, and tasks of great importance for the PQ monitoring field.
oscillatory transients, among others [4]-[7]. In order to apply
proper solutions, the main steps are to detect and classify the In this work, a new methodology for detection and
occurrence and type of PQDs. Although many techniques and classification of PQDs is proposed. In general, it consists of: a
methodologies have been proposed in literature [2], the phasor measurement unit (PMU)-based scheme for extraction
growing amount of data in power systems, proliferation of non- of the fundamental component, i.e., the component of 60 Hz, a
linear loads, and integration of renewable energies, among set of threshold-based rules for classification of sags, swells,
other factors, still make necessary the development and and interruptions, and an ANN for classification of harmonics,
application of new methodologies. notching, spikes, and oscillatory transients. The separation of
the fundamental and non-fundamental components allows
In general, methodologies for PQ monitoring perform the somehow a simpler classification of PQDs. In a broad sense, a
automatic detection and classification of PQDs. For detection, PMU is an equipment that performs in a synchronized way the
signal processing techniques such as discrete Fourier transform estimation of amplitude, phase, and frequency of a
This work is supported in part by PRODEP under the register UAQ-PTC
331, PFCE-UAQ, and by the Mexican Council of Science and Technology
(CONACyT) under the scholarship 296868.
l-))) 
voltage/current signal using a phasor estimation algorithm. In From these components, the magnitude, M, and ș are obtained
this work, the algorithm used is the one provided by IEEE as follows:
C37.118 [13]. This algorithm is based on filters, implying a
low computational burden. Other aspects about this algorithm M = X r 2 + X i2 (3)
are explained in the next section. The proposed methodology is
validated using synthetic signals, as the reference values are
known; then, it is tested using real measured signals. The θ = tan −1 ( X i / X r ) (4)
obtained results show that the proposal can be a useful
approach for PQ monitoring.
Also, from ij(t)=Ȧt+ș, the signal frequency can be obtained
II. THEORETICAL BACKGROUND by [13]:
In this section, a brief description about the main topics 1 d ϕ (t )
used in this work is presented. f (t ) = (5)
2π dt
A. Power Quality Disturbances
In voltage or current signals, PQDs are deviations in In order to obtain Xr and Xi, the IEEE standard C37.118.1
magnitude, phase, and/or frequency from nominal values provides a reference signal processing model [13]. A general
during a time interval [14]. The IEEE standard 1159-2009 structure of this model is shown in Fig. 1. Firstly, the input
presents some definitions and typical values of duration and signal, voltage or current, is multiplied by a quadrature
magnitude for different electromagnetic phenomena, e.g., sags, oscillator, i.e., two functions: sine and cosine. Secondly, two
swells, interruptions, oscillatory transients, harmonics, and low-pass (LP) filters are applied. Thus, Xr and Xi are obtained.
notching, among others. Sags, swells, and interruptions are Finally, the magnitude, phase angle, and frequency results are
characterized by root-mean-square (RMS) variations. Sags computed using (3)-(5), respectively.
have magnitudes from 0.1 to 0.9 pu, i.e. per unit, swells have IEEE standard C37.118.1 stablishes two classes of
magnitudes greater than 1.1 pu, and interruptions present performance for a PMU: P class and M class. The former is
magnitudes lower than 0.1 pu. On the other hand, oscillatory intended for applications that require a fast response, whereas
transients, harmonics, and notching are characterized as non- the latter is intended for applications that require a high
fundamental frequency components, i.e., frequency rejection of aliased signals [13]. In Fig. 1, the LP filter design
components different from 60 Hz or 50 Hz. In literature, many depends on the used class. In this work, the P-class filter is
mathematical expressions have been presented for modeling chosen, as it provides a faster response; yet, the proposed
the waveforms of the aforementioned PQDs [15]-[17]. These methodology can be applied/adapted to M-class filters. The P-
synthetic waveforms are typically used to test new monitoring class filter coefficients are determined by:
methodologies in the PQ field. In PQ monitoring, an opportune
detection and a correct classification of PQDs are very § 2 ·
important issues. An opportune detection allows taking fast W (k ) = ¨1 − k¸ (6)
control actions, avoiding or reducing damages for utilities and © N +2 ¹
users. On the other hand, a correct classification allows
applying the proper solution, as different disturbances require where k = -N/2 : N/2 for integer values only and N is the filter
different solutions. order.
B. Phasor Measurement Unit C. Artificial Neural Networks
The sinusoidal waveform of voltage and current signals can ANNs are among the most used algorithms for pattern
be represented by [13]: recognition [18]. There are many types of ANNs. Among them,
feed forward neural networks (FFNNs) are classical tools for

x (t ) = X m cos(ω t + θ ) (1)

where Xm is the amplitude of the signal, Ȧ=2ʌf is the angular


frequency for a frequency f, and ș is the phase angle. The
phasor representation of (1) is given by [13]:

( )
X = X m / 2 e jθ = X r + jX i (2)

where the magnitude, X m / 2 , is the RMS value of the


sinusoidal waveform. Xr and Xi are the real and imaginary parts,
respectively, of a complex value in rectangular components. Fig. 1. Phasor estimation model.


solving classification tasks. In this work, the FFNN structure is original signal. Finally, the remaining samples are presented to
used to classify different PQDs according to their waveforms. the FFNN in order to classify notching, harmonics, and
In the FFNN structure, the input information moves in one oscillatory transients. It is important to mention that the
direction only, from the input neurons, through the hidden classification of PQDs is carried out each period of the
neurons, and to the output neurons [6]. For training, pairs of fundamental signal, i.e., 60 times per second. In fact, PMUs
input-output data are presented to the FFNN; then, an updating have an option called reporting rate, i.e., the number of frames
algorithm adjusts the weights or gain factors in order to per second (FPS), which indicates the number of results that
minimize the error between the calculated output and the are given in a second. Required rates for 60 Hz systems are 10,
desired output. The pairs of data are repeatedly presented to the 12, 15, 20, 30, and 60 FPS [13]. From this point of view, in this
FFNN until the error is acceptable. The output of a neuron is work, the reporting rate of the PMU, i.e., 60 FPS, controls both
given by: the results of the PMU and the time intervals for classification
of PQDs.
§ I ·
z = A f ¨ ¦ wi yi + b ¸ (7)
© i =1 ¹ IV. EXPERIMENTATION AND RESULTS
where Af (·) is the activation function, wi and yi are the weights A. Validation
and inputs, respectively, b is the bias, and I is the total number
of inputs. The mathematical model in (7) applies for each In order to validate the proposal, synthetic signals are used.
neuron of the FFNN structure. A more detailed description For synthetizing them, the mathematical models presented in
about FFNNs can be found in [18]. [16] are used. In his work, the considered PQDs are sags,
swells, interruptions, harmonics, nothing, and oscillatory
transients, as well as the nominal condition. For each PQD,
III. PROPOSED METHODOLOGY
100 signals are randomly generated, giving a set of 700
The proposed methodology for automatic classification of
PQDs is shown in Fig. 2. Firstly, the input signal, i.e., voltage
or current, is processed by the phasor estimation model. This
model corresponds to the one shown in Fig. 1. At this stage, the
magnitude, phase, and frequency parameters using (3)-(5) are
computed. These parameters are inherent information of a
PMU; but, the magnitude can be used for classification of sags,
swells, and interruptions, as these PQDs are characterized by
RMS variations. In this regard, as a second step, the block of
Threshold-based rules applies simple if-then-else rules for
classification of the aforementioned PQDs; for instance, if
magnitude is greater than 1.1 pu then the PQD is a swell else
other rule is explored. Fig. 3 shows a simple flowchart for the
used rules. On the other hand, the magnitude, phase, and
frequency parameters are also used in the block Sine function
in order to construct a sinusoidal signal using (1). This signal is
subtracted from the input signal. If a PQD exists, the result of
the aforementioned operation is a signal with non-fundamental
frequency components, i.e., oscillatory transients, harmonics,
and notching, among others. Then, the resulting signal is Fig. 3. Flowchart for the threshold-based rules.
downsampled in order to reduce the number of samples in the

Fig. 2. Flowchart for the proposed methodology.


signals. The sampling frequency is set to 6000 samples/s or demonstrate that sags, swells, and interruptions can be
100 samples/cycle. Fig. 4 shows an example of each detected using the magnitude tracking, whereas the other
condition. The first column of Fig. 4 shows the input signals. PQDs can be separated from the input signal. Once the PQDs
The second one shows the magnitude tracking of the are separated, the classification is carried out by the threshold-
fundamental component provided by the PMU algorithm, based rules (Fig. 3) and the FFNN. The time window for
where variations associated to sags, swells, and interruptions classification is one cycle or 100 samples, remembering that a
can be observed (Figs. 4b-d, second column). The third one reporting rate of 60 FPS in the PMU-based model is used.
shows the constructed signal using the parameters of Before the application of the FFNN, a downsampling of 5 is
amplitude, phase, and frequency estimated by the PMU-based used; therefore, 20 samples per cycle are obtained. The
model. When the original signal presents either a sag, swell, or structure of the FFNN is 20 inputs (20 samples per cycle), 10
an interruption, the constructed signal presents a change in the neurons in the hidden layer, and 4 outputs (one per each
amplitude. Finally, the fourth one shows the PQDs associated condition: nominal condition, oscillatory transient, harmonics,
to non-fundamental frequency components, i.e., oscillatory and notching). Thus, the activated output neuron indicates the
transients, harmonics, and notching. Results depicted in Fig. 4 condition of the waveform. The number of neurons in the

Fig. 4. Validation signals: a) nominal condition, b) sag, c) swell, d) interruption, e) oscillatory transient, f) harmonics, and g) notching.


hidden layer is obtained by trial and error, where the goal is
the obtaining of a minimum overall error. Table I presents the
classification results (confusion matrix). All correct
classifications are located at the diagonal of Table I
(highlighted in bold). It is observed that a minimum
effectiveness of 88/100 (notching) is obtained. Also, a perfect
match, 100/100, for signals in nominal conditions is obtained.
The low ratios of classification for oscillatory transients,
harmonics, and notching, 89/100, 91/100, and 88/100,
respectively, may be associated to the downsampling effect as
it affects the waveform; however, in the proposal, the lower
the downsampling factor, the higher the FFNN computational
burden. Therefore, a tradeoff is stablished. Although a perfect
classification, i.e., 100% in all cases, is not reached at this
stage of the research, some advantages of the proposal are: i) it
requires neither an a priori model nor a converge time as the
case of adaptive techniques, 2) unlike EMD method or other
batch processing techniques, the proposed PMU-based scheme
allows analysis with a higher time resolution, and 3) the
separation of PQDs at the fundamental and non-fundamental
components simplify somehow the classification task, i.e., a
basic ANN can achieve suitable results. Yet, it should be
pointed out that other strategies based on indicators such as
entropies, fractals, and high-order statistical parameters will be
studied in order to improve the effectiveness.
Fig. 5. Experimentation: a) experimental setup and b) connection circuit.

B. Experimental Setup and Real Measured Signals transient, is shown in Fig. 6b. It is evident that the proposal can
In order to test the proposal under more real conditions, the deal with the simultaneous occurrence of two PQDs. This only
experimental setup shown in Fig. 5a is used. In general, it applies for a sag, swell, or interruption in combination with an
consists of a DC-AC power inverter in a full-bridge oscillatory transient, harmonic, or notching. The combination
configuration that is controlled by a dSPACE DS1103 of oscillatory transients, harmonics, and notching results in an
controller board. The DS1103 controller provides PQD signals unknown waveform for the FFNN, implying a wrong
to the power inverter, previously generated on a personal classification; at this stage of research, the developed
computer (PC) using Matlab. The power inverter is based on methodology does not consider their simultaneous occurrence.
IGBTs that can operate with a DC voltage of 420V and a Sags, swells, and interruptions do not present this problem, as
switching frequency of 20 kHz. In this case, the power inverter they are mutually exclusive. For instance, Fig. 6c shows a
is operated using unipolar voltage switching at 3 kHz, and it is signal that presents a sag, a swell, and harmonics. Sag and
supplied by a DC voltage source of 200V. The LC filter is used swell occur in different time intervals. Harmonics are present
to cut off the switching frequency in the output. It uses an during all the signal. The plots of magnitude show clearly the
inductor of L = 2mH and a capacitor of C=4uF. In order to occurrence of sags and swells. In the fourth column of Fig. 6,
have a constant load, a 600 Ohms resistor is connected to the the separation of oscillatory transients, notching, and
LC filter output. Fig. 5b shows the connection circuit. Fig. 6 harmonics is observed.
shows some acquired signals. A signal with notching is shown
in Fig. 6a. A signal with two PQDs, a sag and an oscillatory V. CONCLUSIONS
For utilities and users, PQ monitoring is an important task.
TABLE I. VALIDATION: CONFUSION MATRIX
PQ is affected by the presence of abnormal conditions in the
NC SAG SW I OT H N waveforms of voltage or current. These conditions are
typically known as PQDs. In order to control them and apply a
NC 100 0 0 0 0 0 0
proper solution, PQDs have to be detected and classified. In
SAG 2 96 0 2 0 0 0 this work, a new methodology based on a PMU-scheme for
SW 2 0 98 0 0 0 0 classification of PQDs is proposed. PQDs such as sags, swells,
I 0 1 0 99 0 0 0 interruptions, oscillatory transients, harmonics, and notching
are considered. The goal of the proposal is to separate them in
OT 0 0 0 0 89 7 4
PQDs that occur at the fundamental frequency (sags, swells, or
H 0 0 0 0 8 91 1 interruptions) and PQDs that occur at non-fundamental
N 2 0 0 0 8 2 88 frequencies (oscillatory transients, harmonics, and notching).
NC: Nominal condition, SW: swell, I: interruption, OT: oscillatory transient, H: The fundamental component is extracted using the phasor
harmonics, and N: notching. estimation model provided by IEEE C37.118.1. In this work,


Fig. 6. Real measured signals: a) notching, b) sag + oscillatory transient, and c) harmonics + sag + swell.
P-class filters are used, exploiting their high performance [7] Z. Oubrahim, V. Choqueuse, Y. Amirat, and M. Benbouzid,
“Classification of three-phase power disturbances based on model order
under steady-state and dynamic conditions. The classification selection in smart grid applications,” in Proc. 42th Annual Conference of
of sags, swells, and interruptions is carried out using if then the IEEE Industrial Electronics Society (IECON), Oct. 2016, pp. 5143-
else rules, implying a low computational burden. On the other 5148.
hand, the classification of oscillatory transients, harmonics, [8] G. T. Heydt, P. S. Fjeld, C. C. Liu, D. Pierce, L. Tu, and G. Hensley,
and notching is carried out by the FFNN. Although the “Applications of the windowed FFT to electric power quality
assessment,” IEEE Trans. Power Deliv., vol. 14 no. 4, pp. 1411-1416,
proposed FFNN is somehow easy to implement, its size Oct. 1999.
implies a high computational burden. In this regard, some [9] J. Barros, M. Apraiz, and R. I. Diego, “A virtual measurement
quantities, such as entropy, energy, and fractals, among others, instrument for electrical power quality analysis using wavelets,”
could be computed over the analyzed time window in order to Measurement, vol. 42, no. 2, pp. 298-307, Feb. 2009.
reduce the number of inputs with the cost of an additional [10] P. K. Dash, B. K. Panigrahi, and G. Panda, “Power quality analysis
processing stage. using S-transform,” IEEE Trans. Power Deliv., vol. 18, no. 2, pp. 406-
411, April 2003.
In future works, other PQDs such as flicker and spikes will
[11] A. A. Abdelsalam, A. A. Eldesouky, and A. A. Sallam,
be considered. Besides that, the classification of simultaneous “Characterization of power quality disturbances using hybrid technique
PQDs will be addressed. On the other hand, different reporting of linear Kalman ¿lter and fuzzy-expert system,” Elect. Power Syst.
rates and M-class filters will also be explored. Res., vol. 83, no. 1, pp. 41–50, Feb. 2012.
[12] M. Lopez-Ramirez, L. Ledesma-Carrillo, E. Cabal-Yepez, C.
REFERENCES Rodriguez-Donate, H. Miranda-Vidales, H., and A. Garcia-Perez,
[1] D. Granados-Lieberman, R. J. Romero-Troncoso, R. A. Osornio-Rios, “EMD-based feature extraction for power quality disturbance
A. Garcia-Perez, and E. Cabal-Yepez, “Techniques and methodologies classification using moments,” Energies, vol. 9, no. 7, pp. 565, July
2016.
for power quality analysis and disturbances classi¿cation in power
systems: A review,” IET Generat. Transmiss. Distrib., vol. 5, no. 4, pp. [13] IEEE standard for synchrophasors measurements for power systems,
519–529, Apr. 2011. IEEE Standard C37.118.1 2011, Dec. 2011.
[2] F. G. Montoya, A. García-Cruz, M. G. Montoya, and F. Manzano- [14] IEEE Recommended Practices for Monitoring Electric Power Quality,
Agugliaro, “Power quality techniques research worldwide: a review,” IEEE Std. 1159-2009, 2009.
Renew. Sustain. Energy Rev., vol. 54, pp. 846-856, Feb. 2016. [15] J. G. M. S. Decanini, M. S. Tonelli-Neto, F. C. V. Malange, and C. R.
[3] M. Bollen and I. Gu, Signal processing of power quality disturbances. Minussi, “Detection and classi¿cation of voltage disturbances using a
Hoboken, NJ, USA: Wiley, 2006. fuzzy-ARTMAP-wavelet network,” Elect. Power Syst. Res., vol. 81, no.
12, pp. 2057–2065, Dec. 2011.
[4] M. Schael and C. Sourkounis, “Influences of power supply quality on
electric equipment in production processes,” in Proc. 39th Annual [16] D. Granados-Lieberman, M. Valtierra-Rodriguez, L. A. Morales-
Conference of the IEEE Industrial Electronics Society (IECON), Nov. Hernandez, R. J. Romero-Troncoso, and R. A. Osornio-Rios, “A Hilbert
2013, pp. 2081-2086. transform-based smart sensor for detection, classi¿cation, and
[5] M. Valtierra-Rodriguez, R. J. Romero-Troncoso, A. Garcia-Perez, and quanti¿cation of power quality disturbances,” Sensors, vol. 13, no. 5, pp.
5507– 5527, Apr. 2013.
R. A. Osornio-Rios, “FPGA-based instantaneous estimation of
unbalance/symmetrical components through the Hilbert transform,” in [17] M. Valtierra-Rodriguez, D. Granados-Lieberman, J. E. Torres-
Proc. 39th Annual Conference of the IEEE Industrial Electronics Fernandez, J. R. Rodriguez-Rodriguez, and J. F. Gomez-Aguilar, “A
Society (IECON), Nov. 2013, pp. 2279-2284. new methodology for tracking and instantaneous characterization of
[6] M. Valtierra-Rodriguez, R. de Jesus Romero-Troncoso, R. A. Osornio- voltage variations,” IEEE Trans. Instrum. Meas., vol. 65, no. 7, pp.
1596-1604, April 2016.
Rios, and A. Garcia-Perez, “Detection and classi¿cation of single and
combined power quality disturbances using neural networks,” IEEE [18] S. Haykin, Neural networks: a comprehensive foundation, 2nd ed.
Trans. Ind. Electron., vol. 61, no. 5, pp. 2473–2482, May 2014. Englewood Cliffs, NJ: Prentice-Hall, 1999.



You might also like