A Classifier for Distribution Feeder Overcurrent Analysis

456 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 21, NO. 1, JANUARY 2006 A Classifier for Distribution Feeder Overcurrent Analysis Mesut E. Baran, Member, IEEE, and Jinsang Kim, Student Member, IEEE Abstract—This paper focuses on the problem of differentiating inrush currents from fault currents that are observed for a feeder at a distribution substation. This problem is important for feeder protection and power-quality monitoring purposes. The paper shows, using the actual field data, that it is not always easy to distinguish inrush currents from fault currents–as they do not have always well-defined waveforms. This paper shows that the two approaches–the Fourier transform, and the Wavelet transform can be adopted to extract features that make it possible to distinguish them from each other by using an artificial-neural-network-based classifier. The paper also illustrates how to address the issues for successful implementation of these schemes; such as prescreening of data, how to apply fast Fourier transform (FFT) and wavelet transform on the data, and the training of the artificial neural network in order to maximize the performance of the classifier. These issues are illustrated using the actual field data. The test results indicate that both FFT-based and wavelet-transform (WT)-based classifiers yield good results, but the WT-based classifier has better performance. Index Terms—Pattern classification, power distribution faults, power distribution protection, power quality (PQ). I. INTRODUCTION S UBSTATIONS are becoming the main locations for monitoring power quality (PQ) at distribution levels [1], [2]. PQ monitoring devices placed at these locations provide valuable data about the PQ problems on distribution feeders. However, these raw data are not easy to analyze and interpret as the PQ monitoring devices today are designed mainly to capture a set of predefined disturbances. Recently, methods have been developed to screen these individual events and classify them. Since voltage variations are the main concern for the customer, these methods are geared toward the analysis of voltage variation waveforms. However, utilities have discovered that PQ monitoring should include also the overcurrents as they provide valuable information about the cause of the disturbance, such as the location and type of the disturbance [1]. Overcurrents (OCs) are usually caused by faults. A fault and the protection actions to clear it, usually create more than one event record. Some of these records will correspond to the inrush currents that will follow the circuit re-energizing after a temporary fault. Therefore, a disturbance usually creates a set of fault-current and inrush-current events (records). Since, it is the fault-current events that are of interest to PQ engineers, a screening tool is needed for distinguishing fault currents from inrush currents. The overcurrent protection relays also need such a classifier so that they can decide whether a fault has been cleared or not upon re-energizing (since if the fault is cleared, the relay will see the inrush current rather than the fault current). Inrush currents have dominant second-order harmonics; therefore, methods adopted in protection for their identification utilize Fourier transform (FT) methods which extract the second harmonic component of inrush currents [3]. This method although quite reliable, has limitations, since some of the faults (such as the transformer internal faults) could also contain second harmonics [4]. Recently, new wavelet-transform (WT)-based approaches have been proposed to better identify the inrush currents [5], [6]. Our investigation of actual OC records showed that the identification of feeder inrush currents as seen at the substation will be challenging, as these inrush currents do not always have as a well-defined signature as that of a big substation transformer. Similarly, the fault currents do not always have a well-defined profile either. This paper investigates the use of both FT and WT approaches and develops a new artificial-neural-network (ANN)-based classifier to distinguish feeder inrush currents from fault currents. The proposed method is given in Section II, and the test results on actual field data are presented in Section III. II. FEEDER OVERCURRENT CLASSIFICATION The feeder currents are monitored at distribution substations for protection and, recently, for PQ monitoring. One of the basic tasks involves detecting the overcurrents and then analyzing them. For protection, this must be done automatically by the relay, and for PQ monitoring, the analysis is usually done manually as PQ monitors are designed only to detect/capture the events. As indicated above, our goal is to develop a classifier which can automatically determine if an overcurrent is of fault type or inrush type. Most of the fault and inrush currents have the typical profiles shown in Fig. 1. By visual inspection, we can tell that the two current waveforms are quite different from each other. In signal analysis terms, we say that these two waveforms have different signatures and, therefore, the goal of signal analysis is to classify such waveforms automatically. For this, we need a good feature extraction mechanism and a classifier. The feature extraction helps to reduce the dimension of data, and helps the classifier perform better. Pattern recognition techniques can Manuscript received October 25, 2004; revised December 17, 2004. Paper no. TPWRD-00501-2004. The authors are with the Department of Electrical and Computer Engineering, North Carolina State University, Raleigh, NC 27695-7911 USA (e-mail: baran@ncsu.edu). Digital Object Identifier 10.1109/TPWRD.2005.852310 0885-8977/$20.00 © 2006 IEEE BARAN AND KIM: CLASSIFIER FOR DISTRIBUTION FEEDER OC ANALYSIS 457 Fig. 1. Typical overcurrent events (top: overcurrent, bottom: inrush current). then be employed to classify the events based on the extracted features. The challenge, as pointed out before, in discriminating the inrush currents from fault currents is in the fact that both the fault and inrush current profiles do not always have a well-defined profile, as Fig. 2 illustrates. Also, both events can start at any point on the cycle, making it further challenging. A. Feature Extraction The first step in feature extraction is the prescreening of the signal, as three current signals are captured for each OC eventone for each phase. Considering all three signals together for analysis will potentially help characterize the events better, but it is not desirable from the computational point of view. Our investigation indicated that selecting only one phase can provide effective characterization as well. The phase selected corresponds to the phase with the highest current amplitude. This helps especially identifying the inrush events since the phase with highest current has more of the typical characteristic waveform than the others, as Fig. 3 illustrates. For feature extraction, we considered both the conventional FT and the new WT approaches. These two approaches have the following features. Fig. 2. Two nontypical overcurrent events (top: fault, bottom: inrush current). 1) FT: The FT is widely used to determine the harmonics of a given stationary/periodic signal. These harmonics are the features that identify the signal. The FT is used in protection applications to distinguish inrush currents from normal fault currents, as the second harmonic component of an inrush current is much higher than that of the fault [6]. Since OC waveforms are not stationary, in practice, the windowed FT (WFT) is used by applying a sliding sampling window (usually one cycle of the fundamental frequency). The harmonic coefficients change as the sampling window is moved over the signal; typically the coefficients increase with the beginning of the disturbance and then gradually decrease. The challenge of then adopting the WFT for feature extraction is deciding on which window to use. In PQ event analysis, the PQ events are usually captured such that the records usually contain one pre-event cycle, as Fig. 1 illustrates. However, the initiation of the disturbance on each phase varies. This makes it difficult to precisely locate the beginning of the disturbance and, hence, to choose the window for feature extraction. A simple approach adopted here is to apply 458 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 21, NO. 1, JANUARY 2006 Fig. 4. DWT. Fig. 3. Signature of an inrush current on a phase basis. (a) Phase A and B current. (b) Phase C current. TABLE I HARMONICS OF OVERCURRENT EVENTS OF FIG. 1 fast Fourier transform (FFT) to the second cycle, and use the harmonics of this first disturbance cycle as features. Applying this procedure to the two events in Fig. 1 yields the results shown in Table I, which shows the first seven harmonics (normalized to the first harmonic). These results indicate that harmonics of the fault and inrush signals differ from each other mainly on the second harmonic, so we can use these harmonics as features. 2) WT: A visual inspection of the fault and inrush currents in Fig. 1 indicates that these two signals differ mainly during the first three cycles of the disturbance. Therefore, we need a method that can detect the beginning of the disturbance (i.e., localize the signal in time domain) and also extract the signature of the first few cycles. The WT offers such a capability, as the method is a more general feature extraction method, especially for nonstationary signals [13]. The discrete wavelet transform (DWT) has recently been used to detect and localize the common PQ events, mainly the voltage variations [8]–[11]. Here, the DWT will be used to extract features about the OC events not only in terms of frequency content but also localize variations in time. One key parameter in WT is the selection of the mother wavelet to use. Once the mother wavelet is chosen, then, the DWT is implemented with corresponding high- and low-pass filters [13]. In this scheme, as Fig. 4 illustrates, the signal is first passed through high- and low-pass filters and then downsampled by two by discarding every other sample. The high-pass filter gives the detail coefficients, and the low-pass filter gives the approximation coefficients. Most of the time, this process is applied to a given signal at several levels and then the resulting coefficients are used as features. Current literature on the use of DWT to identify various voltage variations indicate that Daubechies series mother wavelets are a good choice because the associated transforms are fast, stable, and accurate [15], and they are also ideally suited for detecting fast decaying and oscillating type of signals, typical of OC events [16]. It has also been shown that to localize a disturbance accurately, a short mother wavelet such as Daubechies wavelet with four filter coefficients (DB2) is preferred. This is because the DWT is implemented using band-pass filters, which are finite-impulse-response (FIR) filters, which always have inherent time delays associated with them. The longer the filter, the more the delay. On the other hand, slower varying variations, such as a voltage sag, can be localized easier by a longer mother wavelet, such as DB8 [9]. For OC classification, we need to choose a mother wavelet that can localize the disturbance (i.e, determine the beginning of the OC), and also capture the variations in the OC events we want to distinguish. The challenge here is that “short’ wavelets can localize the beginning of the disturbance better, while the long wavelets can characterize the slow changing variations better. Therefore, we considered three candidates: DB2, DB4, and DB8. To illustrate how these wavelets perform, the DWT with these wavelets has been applied to the two typical OC signals of Fig. 1, and Fig. 5 shows the results (detail coefficients at scale 1). Comparing these coefficients indicates that DB2 discriminates the two OC events better than the others, since it yields BARAN AND KIM: CLASSIFIER FOR DISTRIBUTION FEEDER OC ANALYSIS 459 B. Classification ANNs have recently emerged as the preferred choice for classification as they have some appealing features that make them easier to use than the parametric classifiers which requires a good knowledge of the statistical data on the events to be classified. An ANN can learn any arbitrary discriminant function, so it can provide optimal classification in a sense that it can minimize the classification error [14]. Therefore, the ANN is chosen in this application for classification. To use an ANN as a classifier, it first needs to be trained using a sample set of “training” data. For the training to be effective, an appropriate ANN architecture and the training algorithm need to be selected. For this application, the widely used back-propagation scheme was adopted. In [12], it was shown that such an ANN can be trained to discriminate between the inrush and fault currents of a power transformer by using the signal itself (raw data). In this study, we developed two ANNs—both using the feedforward architecture. The first one uses the OC features obtained by WFT—the harmonics corresponding to the first cycle of the disturbance. This ANN consists of eight input neurons, two middle (hidden) layer neurons, and one output neuron. The second ANN uses features extracted via DWT, which provides eight detailed coefficients of DWT. The architecture of this second ANN consists of eight input neurons, two hidden layer neurons, and one output neuron. Both ANNs are fully connected, and the neurons have log-sigmoid transfer functions with a range between 0 and 1. Since we have only two classes to distinguish, one output neuron is used. To avoid the saturation effect during training, the ANNs were trained to have a target value of 0.9 for fault and 0.1 for inrush currents. The training of these ANNs and their performance are given in Section III. III. TEST RESULTS To test and compare the methods, 204 OC events were selected from the actual field database. The data are from three substations serving nine feeders. The feeders do not contain any generation on them. The PQ monitors have been placed to monitor the individual feeder currents at the substation, and they are set up to capture overcurrents that exceed the two times the normal current of the feeder [1]. The current measurements have been made with wye-connected current transformers and, hence, phase currents are directly monitored. (If currents are monitored with delta-connected current transformers (CTs), then the currents are indirectly related to phase currents, but the method can still be used for classification as the method looks at the signatures of the two waveforms, not the phase relationship.) The sampling rate for the data is 16 samples per cycle. The selected data set consisted of 118 faults and 96 inrush events. These events contain various fault types—mostly single-phase, and some two-phase, and a few three-phase faults. We separated the data into two sets of data. The training data set (45 fault, 45 inrush) was used to train classifiers and then the remaining data set was used to test classifier performance. The two ANN-based classifiers have been developed by using this actual data. Fig. 5. DWT detail coefficients for the two typical OC events. (a) Actual waveforms (the first three cycles of waveforms in Fig. 1): left—fault; right—inrush current. (b) Detail coefficients at scale one using DB2. (c) Detail coefficients at scale one using DB4. (d) Detail coefficients at scale one using DB8. coefficients that are quite different for the two events (both in magnitude and profile). Note also that it localizes the beginning of the disturbance better than DB8 as expected. We approximate the beginning of the disturbance corresponding to the first coefficient that “jumps.” For example, in Fig. 5, the first coefficient for the inrush current using DB8 does not jump as much as that of the other ones and, hence, the second coefficient can be chosen as the beginning of the disturbance. The effect of this localization error on classification is that we are not picking the right time window to characterize the disturbance. These results indicate DB2 is an appropriate choice for this application. Since the coefficients at this scale are sufficiently different, further filtering, is not necessary (i.e., coefficients at higher scales are not needed). 460 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 21, NO. 1, JANUARY 2006 Fig. 6. Scatter plot of the second and fifth harmonics of OC training data (45 fault and 45 inrush). A. ANN Classifier 1: Harmonics as Features This ANN classifier uses the features—the harmonics—obtained through FFT. Therefore, the first step of the procedure involved applying FFT to the 204 OC signals to determine the first seven harmonics contained in the first disturbance cycle of the events. Fig. 6 shows the scatter plot of the second and fifth harmonics associated with the training data. The figure confirms that the second harmonic of inrush currents are much higher than that of the faults for most of the cases, and variations on the fifth harmonics are similar for the two currents. The figure also illustrates that the points for the two events are not clearly separated and, hence, it will not be easy to classify all of the events correctly based on these harmonics. The ANN has been trained using the harmonics obtained from the training data (which consisted of 45 fault and 45 inrush events). Part of the test data, called the validation data (which consisted of ten fault and ten re-energizing events), is used during training to observe the performance of the ANN being trained. Note that these validation data are not used for training the ANN. Fig. 7 shows the mean-squared error plot and the percent-wrong plot for training and validation data sets. The percent-wrong plot shows the percentage of incorrectly classified event from an entire dataset. As we can see from Fig. 7(a), the output error on training data gradually decreases through the whole training phase. The error associated with the validation dataset, however, reaches the lowest mean-squared error (MSE) at 4701 epochs and after this point, it starts to increase. This suggests that if we train this ANN beyond this point, it will perform better for the training data set, but will work poorly for the validation data set which is unseen during the training phase. Therefore, the training of ANN is stopped at this point. The performance of the ANN designed above is tested by using the remaining test data (which contains 63 fault and 31 inrush events). Fig. 8 shows the results. In the figure, the first 63 outputs correspond to that of the fault events and the remaining 31 (64–94) correspond to inrush events. Since the target values for fault and re-energizing were 0.9 and 0.1, respectively, the ANN is considered to classify an event as a fault if the corresponding output is greater than 0.5; otherwise, it is considered Fig. 7. Error plots during the ANN 1 training. (a) MSE plot. (b) Percent-wrong plot. Fig. 8. Performance of ANN classifier 1 on test data. to be an inrush event. Note that the threshold of 0.5 corresponds to the midpoint between the centers of the two clusters and, thus, it is usually chosen as the boundary that separates the two clusters (events). Using this threshold, the figure shows that only one (which corresponds to an inrush event) out of 94 events is misclassified. Therefore, the performance of this classifier is 98.94%. BARAN AND KIM: CLASSIFIER FOR DISTRIBUTION FEEDER OC ANALYSIS 461 Fig. 9. Harmonics of two unusual events misclassified by ANN 1. (a) Unusual fault event 18. (b) Unusual inrush event 87. The figure indicates that the ANN clearly misclassifies the inrush event 87, and has difficulty in identifying the fault event 18, as the corresponding output is quite close to the threshold value of 0.5. To gain insight as to why, these two events were examined, and they turned out to be the two “unusual” events given earlier in Fig. 2. The harmonic contents of the fault event are shown in Fig. 9(a); it shows that although this is a fault event, its second harmonic is large compared to a typical fault event. Therefore, the classifier yields 0.6 which is close to the threshold value and, thus, this can be interpreted as the ANN not being sure about this case. Fig. 9(b) shows the harmonic contents of the unusual inrush event. Note that the second harmonics of the unusual inrush event are very small, and this is the reason why this inrush event is misclassified as a fault. B. ANN Classifier 2: Wavelet Coefficients as Features This ANN classifier uses the wavelet coefficients as features. The ANN uses the eight detail coefficients obtained with DWT to distinguish fault events from inrush events. As in the previous case, the ANN is first trained using the training data (45 fault, 45 inrush events), and part of the test data, which are the validating data (ten fault, ten inrush events), is used to monitor the training process. Fig. 10 shows the mean-squared error plot and the percent-wrong plot of training and validation data sets. Note that the percent wrong errors for both training and validation data reach zero at 200 epochs. However, the MSE is still high at 200 epochs. Therefore, to prevent an undertraining problem, the training was continued up to 2000 epochs where the MSE is reasonably low. Fig. 10. Error plots during the ANN 2 training. (a) MSE plot. (b) Percent-wrong plot. Fig. 11. Performance of ANN classifier 2. To test the performance of the ANN, test data (63 fault, and 31 inrush events) are used, and the results are given in Fig. 11. In the figure, the first 63 events correspond to fault and the remaining 31 (64–94) correspond to inrush events. Again, using 0.5 as the threshold to separate the faults from inrush events, the figure shows that all of the events are correctly classified. Therefore, the performance of this ANN is 100%. This high performance indicates that the wavelet coefficients characterize the two events better than the harmonics (i.e., WT is a better feature extraction than FFT). The 100% success of the ANN indicates also that the ANN is able to separate the two events easily, whereas with harmonics, it is not easy. 462 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 21, NO. 1, JANUARY 2006 IV. CONCLUSION This paper shows that the two types of feeder overcurrents—fault and inrush currents—can be challenging to distinguish as they do not always have well-defined waveforms. Nevertheless, the two approaches: the FT and the WT can be adopted to extract features that make it possible to distinguish them from each other by using an ANN-based classifier. This paper also illustrates how to address the issues for successful implementation of these schemes; such as prescreening of data (which phase to choose), how to apply FFT and WT on the data to make sure that the extracted features are distinctive enough for characterization and, finally, the training of the ANN to maximize the performance of the classifier. These issues are illustrated using the actual field data. The test results indicate that both FFT- and WT-based classifiers yield good results, but the WT-based classifier has better performance. ACKNOWLEDGMENT The authors would like to thank Glenn Lampley of Progress Energy, Raleigh, NC, for his valuable help. REFERENCES [1] G. Lampley, “Power quality monitoring to meet customer needs,” in Proc. EPRI Power Quality Analysis Conf., Charlotte, NC, May 1999. [2] IEEE Recommended Practice for Monitoring Electric Power Quality, IEEE Std. 1159-1995, 1995. [3] L. F. Kennedy and C. D. Hayward, “Harmonic-Current-Restrained relays for differential protection,” AIEE Trans., vol. 57, pp. 262–271, May 1938. [4] L. G. Perez, A. J. Flechsig, J. L. Meador, and Z. Obradovic, “Training an artificial neural network to discriminate between magnetizing inrush and internal faults,” IEEE Trans. Power Del., vol. 9, no. 1, pp. 434–441, Jan. 1994. [5] P. Liu, O. P. Malik, D. Chen, G. S. Hope, and Y. Guo, “Improved operation of differential protection of power transformers for internal faults,” IEEE Trans. Power Del., vol. 7, no. 4, pp. 1912–1919, Oct. 1992. [6] M. G. Morante and D. W. Nicoletti, “A wavelet-based differential transformer protection,” IEEE Trans. Power Del., vol. 14, no. 4, pp. 1351–1358, Oct. 1999. [7] S. Santoso, W. M. Grady, E. J. Powers, J. Lamoree, and S. C. Bhatt, “Characterization of distribution power quality events with fourier and wavelet transforms,” IEEE Trans. Power Del., vol. 15, no. 1, pp. 247–254, Jan. 2000. [8] S. Santoso, E. J. Powers, W. M. Grady, and P. Hofmann, “Power quality assessment via wavelet transform analysis,” IEEE Trans. Power Del., vol. 11, no. 2, pp. 924–930, Apr. 1996. [9] S. Santoso, E. J. Powers, and W. M. Grady, “Electric power quality disturbance detection using wavelet transform analysis,” in Proc. IEEE-SP Int. Symp. Time-Frequency Time-Scale Analysis, 1994, pp. 166–169. [10] S. Santoso, J. Lamoree, W. M. Grady, E. J. Powers, and S. C. Bhatt, “A scalable PQ event identification system,” IEEE Trans. Power Del., vol. 15, no. 2, pp. 738–743, Apr. 2000. [11] A. M. Gaouda, M. M. A. Salama, M. R. Sultan, and A. Y. Chikhani, “Power quality detection and classification using wavelet-multiresolution signal decomposition,” IEEE Trans. Power Del., vol. 14, no. 4, pp. 1469–1476, Oct. 1999. [12] A. K. Ghosh and D. L. Lubkeman, “The classification of power system disturbance waveform using a neural network approach,” IEEE Trans. Power Del., vol. 10, no. 1, pp. 109–115, Jan. 1995. [13] O. Rioul and M. Vetterli, “Wavelets and signal processing,” IEEE Signal Process. Mag., vol. 8, no. 4, pp. 14–38, Oct. 1991. [14] J. C. Principe, N. R. Euliano, and W. C. Lefebvre, Neural and Adaptive Systems: Fundamentals Through Simulation. New York: Wiley, 1999, pp. 101–102. [15] A. W. Galli and O. M. Nielsen, “CAP tutorial: Wavelet analysis for power system transients,” IEEE Comput. Appl. Power, vol. 12, no. 1, pp. 16–16, Jan. 1999. [16] P. L. Mao and R. K. Aggarwal, “A novel approach to the classification of the transient phenomena in power transformers using combined wavelet transform and neural network,” IEEE Trans. Power Del., vol. 16, no. 4, pp. 654–660, Oct. 2001. Mesut E. Baran (S’87–M’88) received the Ph.D. degree from the University of California at Berkeley in 1988. Currently, he is an Associate Professor in the Department of Electrical and Computer Engineering, North Carolina State University, Raleigh. His research interests include distribution and transmission system analysis and design. Jinsang Kim (S’00) received the B.S. and M.S. degrees in electrical engineering from SoongSil University, Seoul, Korea, in 1992 and 1996, respectively, and is currently pursuing the Ph.D. degree at North Carolina State University, Raleigh. His research interests include computer applications in power systems, power quality, and power electronics.