1560

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 19, NO. 4, OCTOBER 2004

Wavelet-Based Neural Network for Power Disturbance Recognition and Classication

Zwe-Lee Gaing, Member, IEEE
AbstractIn this paper, a prototype wavelet-based neuralnetwork classier for recognizing power-quality disturbances is implemented and tested under various transient events. The discrete wavelet transform (DWT) technique is integrated with the probabilistic neural-network (PNN) model to construct the classier. First, the multiresolution-analysis technique of DWT and the Parsevals theorem are employed to extract the energy distribution features of the distorted signal at different resolution levels. Then, the PNN classies these extracted features to identify the disturbance type according to the transient duration and the energy features. Since the proposed methodology can reduce a great quantity of the distorted signal features without losing its original property, less memory space and computing time are required. Various transient events tested, such as momentary interruption, capacitor switching, voltage sag/swell, harmonic distortion, and icker show that the classier can detect and classify different power disturbance types efciently. Index TermsParsevals theorem, power quality, probabilistic neural network, wavelet transform.

I. INTRODUCTION OOR power quality (PQ) may cause many problems for affected loads, such as malfunctions, instabilities, short lifetime, and so on. Poor quality electric power is normally caused by power-line disturbances, such as impulses, notches, glitches, momentary interruptions, wavefaults, voltage swell/sag, harmonic distortion, and icker, resulting in misoperation or failure of end-use equipment. In order to improve power quality, the sources and causes of such disturbances must be known before appropriate mitigating actions can be taken. A feasible approach to achieve this goal is to incorporate detection capabilities into monitoring equipment so that events of interest will be recognized, captured, and classied automatically. Hence, good performance monitoring equipment must have functions which involve the detection, localization, and classication of transient events. In particular, when the disturbance type has been classied accurately, the power-quality engineers can dene the major effects of the disturbance at the load and analyze the source of the disturbances so that an appropriate solution can be formulated [1], [2]. Mo et al. [11] demonstrated how to extract the features from the wavelet transform coefcients at different scales as inputs

Manuscript received February 25, 2003; revised June 17, 2003. Paper no. TPWRD-00073-2003. This work was supported by the National Science Council of Taiwan, R.O.C. The author is with the Electrical Engineering Department, Kao-Yuan Institute of Technology, Kaohsiung 821, Taiwan, R.O.C. (e-mail: zlgaing@ ms39.hinet.net). Digital Object Identier 10.1109/TPWRD.2004.835281

to neural networks for classifying the nonstationary signal type. Santoso et al. [3], [18], [20] proposed to extract the squared wavelet transform coefcients (SWTC) at each scale as inputs to the neural networks for classifying the disturbance type. Perunicic et al. [16] employed the DWT coefcients as inputs to a single-layer self-organizing map neural network to train and classify the transient disturbance type. Elmitwally et al. [17] used the preprocessed DWT coefcients as inputs to a rened neuro-fuzzy network to train and classify the power system disturbance type. Angrisani et al. [12] proposed to employ the continuous wavelet transform (CWT) to estimate the disturbance time duration and the DWT to estimate the disturbance amplitude. The two features thus obtained are then used to classify the transient disturbance type. Santoso et al. [13], [14] presented a wavelet-based neural classier integrating the DWT, learning vector quantization (LVQ) neural network, and decision-making scheme to become an actual power disturbance classier. The classier employed the DWT coefcients as inputs to multiple LVQ neural networks to train and perform waveform recognition, and use the decision-making schemes to classify the transient disturbance type. Chung et al. [19] presented a novel classier using a rule-based method and a wavelet packet-based hidden Markov model (HMM). The rule-based method is employed to classify the time-characterized feature disturbances, while the wavelet packet-based HMM is utilized to categorize the frequency-characterized feature power disturbances. As seen in the above studies, the DWT technology has often been employed to capture the time of transient occurrence and extract frequency features of power disturbance. Integrating the DWT technology with the articial intelligence method or expert system to become a practical power disturbance classier for recognizing accurately the disturbance has attracted much research interest. However, two practical problems must be overcome in the above methods. 1) Adopting directly the DWT coefcients as inputs to the neural networks requires large memory space and much learning time. 2) The decomposition level with the number of extraction features must be reduced to enhance computing efciency and accuracy of recognizing the disturbance type. This paper presents a novel classier consisting of two models. First, the wavelet MRA technique and the Parsevals theorem are employed to extract the energy distribution features of the distorted signal at different resolution levels. Second, the PNN is employed to classify disturbance types according to the transient duration and the detailed energy distribution. By

0885-8977/04$20.00 2004 IEEE

GAING: WAVELET-BASED NEURAL NETWORK FOR POWER DISTURBANCE RECOGNITION AND CLASSIFICATION

1561

using the Parsevals theorem, the number of distorted signal features can be reduced without losing its property. Because the learning efciency of the PNN is very fast, it is suitable for real-time operation for fault diagnosis and signal classication problems. Ninety distorted signals generated by the Power System Blockset in Matlab were employed to test the proposed classier. The accuracy rate and computation time of the proposed method were measured. The experimental results showed that the proposed method could analyze the signals efciently, thus enhancing the performance of the power transient recorder. II. WAVELET TRANSFORM The wavelet analysis block transforms the distorted signal into different time-frequency scales. The wavelet transform (WT) uses the wavelet function and scaling function to perform simultaneously the multiresolution analysis (MRA) decomposition and reconstruction of the measured signal. The wavelet function will generate the detailed version (high-frequency components) of the decomposed signal and the scaling function will generate the approximated version (low-frequency components) of the decomposed signal. The wavelet transform is a well-suited tool for analyzing high-frequency transients in the presence of low-frequency components such as nonstationary and nonperiodic wideband signals [3]. A. Multiresolution Analysis (MRA) and Decomposition The rst main characteristic in WT is the MRA technique that can decompose the original signal into several other signals with different levels (scales) of resolution. From these decomposed signals, the original time-domain signal can be recovered without losing any information. The recursive mathematical representation of the MRA is as follows: (1) where approximated version of the given signal at scale ; detailed version that displays all transient phenomena of the given signal at scale ; denotes a summation of two decomposed signals; is the decomposition level. B. Mathematical Model of DWT Before the WT is performed, the wavelet function and must be dened. The wavelet function scaling function serving as a highpass lter can generate the detailed version of the distorted signal, while the scaling function can generate the approximated version of the distorted signal. In general, the discrete and can be dened as follows: (2) (3)

where is the scaling coefcient at scale , and is the wavelet coefcient at scale . Simultaneously, the two functions must be orthonormal and satisfy the properties as follows:

(4)

Assuming the original signal at scale is sampled at con, the samstant time intervals, thus pling number is . is an integer number. For , its ) DWT mathematical recursive equation (as is presented as follows:

(5) where (6) (7) (8) where is the approximated version at scale , is the detailed version at scale , and is the translation coefcient. According to the orthonormal wavelet functions and (5), the signal can be reconstructed from both and coefcients using the inverse discrete wavelet transform (IDWT) ]. [as Fig. 1 illustrates the three decomposed/reconstructed levels of the DWT algorithm. At each decomposition level, the length of the decomposed signals (e.g., and ) is half that of the in the previous stage. signals III. PARSEVALS THEOREM IN DWT APPLICATION is the In Parsevals theorem, assuming a discrete signal current that ows through the 1- resistance, then the consumptive energy of the resistance is equal to the square sum of the spectrum coefcients of the Fourier transform in the frequency domain (9) where is the sampling period, and cients of the Fourier transform. is the spectrum coef-

1562

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 19, NO. 4, OCTOBER 2004

Fig. 1.

Three decomposed/reconstructed levels of DWT.

To apply the theorem to the DWT, we use (5) and (9) to obtain (10) that is the Parsevals theorem in the DWT application

Fig. 2. Voltage sag with different times of disturbance occurrence. (a) Different times of disturbance occurrence. (b) Energy distribution diagram.

(10) Hence, through the DWT decomposition, the energy of the distorted signal is shown by (10). The rst term on the right of (10) denotes the average power of the approximated version of the decomposed signal, while the second term denotes that of the detailed version of the decomposed signal. The second term giving the energy distribution features of the detailed version of distorted signal will be employed to extract the features of power disturbance.

distorted signal for classifying different power-quality problems. The process can be represented mathematically by (12) (12) is the norm of the expansion coefcient . where Four special properties in (12) need further explanation. The Daubanchie db4 wavelet function was adopted to perform the DWT, thus resulting in the larger energy distributions of the decomposition levels 6, 7, and 8. However, using different wavelet functions will generate different results. The energy distribution remains unaffected by the time of disturbance occurrence. The outline of energy distribution remains the same despite variations in the vibration amplitude of the same disturbance type. The low-level energy distribution will show obvious variations when the distorted signal contains high-frequency elements. On the contrary, the high-level energy distribution will show obvious variations when the distorted signal contains low-frequency elements. To display clearly the characteristics of the above properties, we normalize (12) by (13) (13)

IV. FEATURE EXTRACTION, RECOGNITION, AND CLASSIFICATION A. Feature Extraction 1) Duration of Electromagnetic Transients: In general, when a transient disturbance occurs, the stable power signal will generate a discontinuous state at the start and end points of the disturbance duration. Employing the DWT technique to analyze the distorted signal through one-level decomposition of the MRA will cause the wavelet coefcients at the start and end points of the disturbance to generate severe variation. Therefore, we can easily obtain the start time and end time of the disturbance duration from the variations in absolute and calculate the disturbance duration wavelet coefcients

(11) coefcients need to be To lter the noise and correct , the modied by subtracting the standard deviation of the absolute wavelet coefcients from the absolute wavelet coefcients. 2) Detailed Energy Distribution: As seen in (10), the energy of the distorted signal can be partitioned at different resolution levels in different ways depending on the power-quality problem. of the detailed Therefore, we will examine the coefcient version at each resolution level to extract the features of the For example, Fig. 2(a) shows a voltage sag disturbance in a simulation power system (frequency = 60 Hz, amplitude = 1 p.u.) with three different times of occurrence. Employing (12) and (13) to analyze the three distorted signals (sampling rate is 256 points/per cycle) shows that when the disturbance intervals are the same (ve cycles), the detailed energy distriof the given signals are also the same as butions shown in Fig. 2(b). Similarly, when the amplitude of vibration (0.8 p.u., 0.5 p.u., and 0.2 p.u.) of the sag disturbances are

GAING: WAVELET-BASED NEURAL NETWORK FOR POWER DISTURBANCE RECOGNITION AND CLASSIFICATION

1563

Fig. 6.

Architecture of a PNN.

Fig. 3. Voltage sag with different amplitude magnitudes. (a) Different amplitude magnitude. (b) Energy distribution diagram.

Fig. 4.

Voltage uctuation with low-freqency elements.

It is implemented using the probabilistic model, such as Bayesian classiers. A PNN is guaranteed to converge to a Bayesian classier provided that it is given enough training data. No learning processes are required. No need to set the initial weights of the network. No relationship between learning processes and recalling processes. The differences between the inference vector and the target vector are not used to modify the weights of the network. The learning speed of the PNN model is very fast, making it suitable for fault diagnosis and signal classication problems in real time. Fig. 6 shows the architecture of a PNN model that is composed of the radial basis layer and the competitive layer. In the signal classication application, the training examples are classied according to the their distribution values of probabilistic density function (PDF), which is the basic principle of the PNN. A simple PDF is dened as follows:

(14)
Fig. 5. Distorted voltage with high-frequency elements.

changed as seen in Fig. 3(a), the detailed energy distributions are the same as shown in Fig. 3(b). The energy distribution of the distorted signal with low-frequency elements is shown in Fig. 4, while that with high-frequency elements is shown in Fig. 5. These experimental results depict clearly the properties of energy distribution of Parsevals theorem in DWT applications. B. Probabilistic Neural Network (PNN) The probabilistic neural-network ) (PNN) model is one of the supervised learning networks, and has the following features distinct from those of other networks in the learning processes [21].

Modifying and applying (15) to the output vector hidden layer in the PNN is as below

of the

(15)

The algorithm of the inference output vector follows: and then else

in the PNN is as

(16)

1564

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 19, NO. 4, OCTOBER 2004

Fig. 7.

Diagram of the proposed classier.

where number of input layers; number of hidden layers; number of output layers; number of training examples; number of classications (clusters); smoothing parameter (standard deviation), . In general, is set to be 0.5; input vector; Euclidean distance between the vectors and , i.e. ; connection weight between the input layer and the hidden layer ; is the connection weight between the hidden layer and the output layer . The learning and recalling processes of the PNN for classication problems can be found in [21]. C. Classication of Transient Signals Using PNN Model Though the PNN has some disadvantages, such as a large memory requirement and the recalling time being proportional to the quantity of training samples, we can overcome these drawbacks by employing Parsevals theorem to reduce the training inputs. In this study, we will perform a 13-level decomposition of each discrete distorted signal to obtain the detailed version co. Using (11), we can obtain the disturbance efcients duration by the squared wavelet coefcients of one-level decomposition. Simultaneously, with (12) and (13), we can obtain . These features each detailed energy distribution would be applied to the PNN for recognizing and classifying
Fig. 8. MRA decomposition and the detailed energy distribution of a pure sine wave.

the distorted signals. The calculation procedures of the proposed classier are shown in Fig. 7. V. APPLICATIONS AND RESULTS A. Recognition of Extracted Features To verify the feasibility of the proposed method, we used the Power System Blockset Toolbox in Matlab to generate one pure sine-wave signal (frequency = 60 Hz, amplitude = 1 p.u.) and six sample transient distorted signals. These distorted signals included momentary interruption, capacitor switching, voltage sag/swell, harmonic distortion, and icker. The sampling rate of the digital recorder was 256 points/per cycle. The Daubanchie db4 wavelet was adopted to perform the DWT. The PNN model was provided by the Neural Network Toolbox in Matlab. The proposed method was written in Matlab language and executed on a Pentium III 550 personal computer with 256-MB RAM. Fig. 8 shows the detailed version of a three-level decomand the detailed energy distribution position of a pure sine wave. The X-axis is the sampled signal points and the Y-axis is the magnitude in Fig. 8(a). The X-axis is the decomposition level and the Y-axis is the energy in Fig. 8(b). The disturbance duration could not be detected because the signal is smooth. Figs. 914 show the detailed version of a three-level deand the detailed energy distribution composition

GAING: WAVELET-BASED NEURAL NETWORK FOR POWER DISTURBANCE RECOGNITION AND CLASSIFICATION

1565

Fig. 9.

Detailed energy distribution of a momentary interruption. Fig. 10. Detailed energy distribution of a voltage sag.

of each given distorted signal. In the harmonic distortion and icker cases, the disturbance duration will be set to be very short (0.0001 s) because the voltage waveshape is disturbed periodically. The simulation results are summarized in Table I. From Figs. 814 and Table I, we can categorize three properties of energy distribution of the given distorted signals. These properties become the foundations for classifying the disturbance type. , , and When a sag or swell or interrupt occurs, will show great variations. When the voltage suffers a transient disturbance of the high-frequency elements such as capacitor switching and , , and will show obvious harmonic distortion, variations. When the voltage suffers a transient disturbance of the , , and low-frequency elements such as icker, will show obvious variations. Fig. 15 shows orderly the energy distributions of seven signals on the same three-dimensional (3-D) coordinate axis (using piecewise plot style). Thus, we can clearly observe the differences in energy distribution between different signals.

B. Training Examples and Signal Classication Table I also shows the format of training examples of the PNN model. Each training datum involves 13 energy values

Fig. 11.

Detailed energy distribution of a voltage swell.

1566

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 19, NO. 4, OCTOBER 2004

Fig. 13. Fig. 12. Detailed energy distribution of a capacitor switching.

Detailed energy distribution of a harmonic distortion.

, one disturbance duration time , and one expected classication type. Because of the time of occurrence, duration, and amplitude magnitude of electromagnetic (EM) disturbance in a power system are random, we created randomly ten voltage waves of each given distorted signal that had a different time of occurrence, duration, and amplitude magnitude from the distorted signal property as the training examples. There were 70 distorted signals generated in this paper (involving the pure sine wave). Using (11)(13), we can obtain the features of the 70 signals. They are rearranged as the format shown in Table I to be the training examples for training the PNN model. The numbers of input and output layers in the PNN model were set to be 14 and 7, respectively. First, we employed 35 (half of all training examples) and 70 training examples to train the PNN model, respectively. We also randomly created 20 distorted voltage waves to test the proposed approach. The experimental results are shown in Table II. As can be seen, with 35 training examples were 35, the classied accuracy rate of the distorted signals of the proposed approach was 80% (four distorted signal failures). When the training examples were 70, the classied accuracy rate was 90% (two distorted signal failures). The results show the more the training examples, the better the accuracy rate. Because the PNN model requires little learning time, the proposed approach is suitable for real-time processing in a modern digital recorder. The learning time and recalling time of the PNN model are also shown in Table II.

Fig. 14.

Detailed energy distribution of a icker.

GAING: WAVELET-BASED NEURAL NETWORK FOR POWER DISTURBANCE RECOGNITION AND CLASSIFICATION

1567

TABLE I ENERGY FEATURES AND DISTURBANCE DURATIONS OF DISTORTED SIGNALS

by simulation software, employing real distorted signals measured by the digital recorder to improve the proposed method is one of our future works. ACKNOWLEDGMENT The author gratefully acknowledges the technical support from Taiwan Power Company. REFERENCES
[1] M. B. Hughes, J. S. Chan, and D. O. Koval, Distribution customer power quality experience, IEEE Trans. Ind. Applicat., vol. 29, pp. 12041211, Nov./Dec. 1993. [2] E. W. Gunther and H. Mehta, A survey of distribution system power qualitypreliminary results, IEEE Trans. Power Delivery, vol. 10, pp. 322329, Jan. 1995. [3] S. Santoso, E. J. Powers, W. M. Grady, and P. Hofmann, Power quality assessment via wavelet transform analysis, IEEE Trans. Power Delivery, vol. 11, pp. 924930, Apr. 1996. [4] T. B. Littler and D. J. Morrow, Wavelets for the analysis and compression of power system disturbances, IEEE Trans. Power Delivery, vol. 14, pp. 358364, Apr. 1999. [5] O. Poisson, P. Rioual, and M. Meunier, New signal processing tools applied to power quality analysis, IEEE Trans. Power Delivery, vol. 14, pp. 561566, Apr. 1999. [6] G. T. Heydt, P. S. Fjeld, C. C. Liu, D. Pierce, L. Tu, and G. Hensley, Applications of the windowed FFT to electric power quality assessment, IEEE Trans. Power Delivery, vol. 14, pp. 14111416, Oct. 1999. [7] G. T. Heydt and A. W. Galli, Transient power quality problems analyzed using wavelets, IEEE Trans. Power Delivery, vol. 12, pp. 908915, Apr. 1997. [8] P. Pillay and A. Bhattacharjee, Application of wavelet to model short term power system disturbance, IEEE Trans. Power Syst., vol. 11, pp. 20312037, Nov. 1996. [9] F. H. Magnago and A. Abur, Fault location using wavelets, IEEE Trans. Power Delivery, vol. 13, pp. 14751480, Oct. 1998. [10] D. C. Robertson, O. I. Camps, J. S. Mayer, and W. B. Gish, Wavelets and electromagnetic power system transients, IEEE Trans. Power Delivery, vol. 11, pp. 10501058, Apr. 1996. [11] F. Mo and W. Kinsner, Wavelet modeling of transients in power systems, in Proc. Conf. Communications, Power and Computing Proceedings, Winnipeg, MB, Canada, May 2223, 1997, pp. 132137. [12] L. Angrisani, P. Daponte, M. D Apuzzo, and A. Testa, A measurement method based on the wavelet transform for power quality analysis, IEEE Trans. Power Delivery, vol. 13, pp. 990998, Oct. 1998. [13] S. Santoso, E. J. Powers, W. M. Grady, and A. C. Parsons, Power quality disturbance waveform recognition using wavelet-based neural classierpart 1: theoretical foundation, IEEE Trans. Power Delivery, vol. 15, pp. 222228, Jan. 2000. [14] , Power quality disturbance waveform recognition using waveletbased neural classierpart 2: application, IEEE Trans. Power Delivery, vol. 15, pp. 229235, Jan. 2000. [15] O. Poisson, P. Rioual, and M. Meunier, Detection and measurement of power quality disturbance using wavelet transform, IEEE Trans. Power Delivery, vol. 15, pp. 10391044, July 2000. [16] B. Perunicic, M. Mallini, Z. Wang, and Y. Liu, Power quality disturbance detection and classication using wavelets and artical neural networks, in Proc. 8th Int. Conf. Harmonics and Quality of Power, vol. 1, Oct. 1416, 1998, pp. 7782.

Fig. 15.

Differences in energy distribution of all signals.

TABLE II TEST RESULTS OF THE PROPOSED METHOD

1) 2) 3) 4) 5) 6) 7)

pure sine wave; momentary interruption voltage sag s); s); voltage swell capacitor switching harmonic distortion icker s).

s);

s); s);

VI. DISCUSSION AND CONCLUSION This paper proposed a prototype of wavelet-based neural-network classiers for power disturbance recognition and classication. The proposed method can reduce the quantity of extracted features of distorted signal without losing its property, thus requiring less memory space and computing time for proper classication of disturbance types. The experimental results showed that the proposed method has the ability of recognizing and classifying different power disturbance types efciently, and it has the potential to enhance the performance of the power transient recorder with real-time processing capability. Because the distorted signals in this study were generated

1568

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 19, NO. 4, OCTOBER 2004

[17] A. Elmitwally, S. Farghal, M. Kandil, S. Abdelkader, and M. Elkateb, Proposed wavelet-neurofuzzy combined system for power quality violations detection and diagnosis, Proc. Inst. Elect. Eng., Gen. Transm. Dist., vol. 148, no. 1, pp. 1520, Jan. 2001. [18] S. Santoso, W. M. Grady, E. J. Powers, J. Lamoree, and S. C. Bhatt, Characterization of disturbance power quality events with Fourier and wavelet transforms, IEEE Trans. Power Delivery, vol. 15, pp. 247254, Jan. 2000. [19] J. Chung, E. J. Powers, J. Lamoree, and S. C. Bhatt, Power disturbance classier using a rule-based method and wavelet packet-based hidden Markov model, IEEE Trans. Power Delivery, vol. 17, pp. 233241, Jan. 2002. [20] A. M. Gaouda, M. M. A. Salama, M. R. Sultan, and A. Y. Chikhani, Power quality detection and classication using wavelet-multiresolution signal decomposition, IEEE Trans. Power Delivery, vol. 14, pp. 14691476, Oct. 1999.

[21] D. F. Specht, Probabilistic neural networks, Neural Networks, vol. 3, no. 1, pp. 109118, 1990.

Zwe-Lee Gaing (M02) received the M.S. and Ph.D. degrees from National Sun Yat-Sen University, Kaohsiung, Taiwan, R.O.C., in 1992 and 1996, respectively. Currently, he is an Associate Professor in the Electrical Engineering Department at Kao-Yuan Institute of Technology, Kaohsiung. His research interests include articial intelligence with application to power system operation and control.