Power Quality Analysis Using An Adaptive Decomposition Structure

Transcription

1 Power Quality Analysis Using An Adaptive Decomposition Structure Doğan Gökan Ece 1 and Ömer Nezi Gerek 1 (1) Dept. of Electrical and Elctronics Engineering, Anadolu University, Scool of Engineering and Arcitecture, Eskiseir 6470, Turkey ( dgece@anadolu.edu.tr, ongerek@anadolu.edu.tr) Abstract A new event detection sceme for power quality analysis based on te statistical analysis of adaptive decomposition signals is proposed. Proposed sceme is implemented using MATLAB Simulink DSP Blockset. Te combination of an adaptive prediction filter based subband decomposition structure wit a rule based istogram analysis block produced successful detection and classification results on our real life power system transient data. Keywords Power quality analysis, Statistical metods, Adaptive decomposition. I. INTRODUCTION Event detection and classification are essential processes for explaining and ten correcting te cause of te power quality (PQ) problems in low-voltage, commercial, and industrial applications. Once te voltage and/or current waveforms are captured and stored, an automated post event analysis is needed. Considering te range of equipment operated on te customer's side of te meter, te analysis tool sould be immune to te inerent noisy and armonically ric normal operating conditions wile adapting to te dynamic canges in customer's load to avoid nuisance alarm signals. Recent contributions in te area of PQ analysis use various wavelets suc as Daubecies wavelets, Morlet wavelets, etc., to analyze te disturbances wile pre-event voltage or current waveforms are assumed to be sinusoid [5]-[9]. A specific wavelet may be designed to detect, for example, arcing faults in a sinusoidal pre--fault waveform [10]. However, te selected wavelet may not correspond to te optimal discriminating system for anoter type of a transient event. In tis work, a new event detection sceme for power quality analysis based on te statistical analysis of adaptive decomposition signals is proposed. Te adaptive metod is developed to detect and classify power quality disturbances regardless of te type of te pre-event voltage or current waveforms. Te significance of te proposed metod is tat it provides a way of detecting variety of events witout canging te structure. If we do not ave any prior information on weter te waveform is pure sinusoid, or not, te steady state properties of a waveform can be well approximated using adaptive systems. Te only assumption is tat te pre-event steady state waveform as variations of relatively lower frequency as compared to te noise imposed waveform due to a transient event. Tis idea is utilized to construct a decomposition filter bank structure [1] wic operates on te current or voltage waveforms, and at te same time, adapts its filter bank according to te waveform beavior. Least Mean Squared (LMS) type adaptive filters are used in our filter bank structure [3]. Tese filters are timevarying FIR filters wose coefficients are continuously updated according to te minimization of an error sequence, wic corresponds to one of te subbands in our case. Wen te adaptation converges to a steady state, te disturbance contribution of any transient event on te waveform will take some time for te adaptive filter bank to adapt. Meanwile, te decomposition structure will exibit large adaptation error signals in te ig-pass subband. Time lengt of tis large adaptation error signal is expected to be sort for transient-type events suc as arcing line-to-ground faults, sags, and swells and te adaptation time is expected to be longer for dynamic canges in load. Te teoretical background of te statistical properties of tis structure is explained in detail in [1]. Tis metod is also able to detect canges in te nature of any kind of steady state waveforms. Teoretically, if te steady state waveform as ig frequencies, and if te disturbance causes low frequency components, te adaptation will again be perturbed producing large adaptation errors. However, success for tis case is limited by te ability of convergence speed of te adaptation algoritm (LMS). A general case of relatively lower frequency steady state waveforms generated under normal operating conditions, as compared to iger frequency components generated by PQ events can be safely assumed for te majority of applications. For our tests, real life voltage and current data are obtained from a low-voltage experimental system during various type of staged transient events. Proposed metod is implemented using MATLAB Simulink DSP Blockset [11]. During te detection process, te event data is applied to te system wic is a combination of an adaptive prediction filter based subband decomposition structure and a rule based istogram analysis block. In our experiments, we ave observed tat staged arcing faults, sag, and swell events can be detected, localized in time, and discriminated using tis metod. Te ability to classify and distinguis transients from canges in load makes te proposed metod more flexible as compared to te commonly used transform domain tresolding tecniques for te analysis of power quality events. 1

2 II FILTERBANK STRUCTURE Te signal decomposition consists of an adaptive prediction filter in a polypase structure [1]. In tis aspect, te overall sceme resembles te lifting-style wavelet decomposition due to its filterbank implementation []. However, te basic idea is to produce decomposition signals wic converge to a minimal residual signal tat can be considered as te non-predictable content of te steady state signal. Tis idea is also very new in te signal processing field, and quite recently it as been applied to signal compression [1]. Normally, te wavelet filterbanks decompose te signal according to te frequency content of te filters wit fixed coefficients. Here, te frequency content or spectral decompositions are irrelevant due to te fact tat te adaptive prediction filter constantly canges te filter coefficients. Instead of a spectral decomposition, we split te signal into a) lower resolution, and b) non-predictable portions. Te analysis structure is illustrated in Fig. 1. Bot te lower resolution and non-predictable parts are produced using te two polypase components of te original signal: x1 [] n = x[ n] (1) x [] n = x[ n + 1] () Tese components can be tougt of as even and odd indexed terms of te discrete-time signal. For a signal wit slow variations, te two polypase components ave strong correlation. Terefore one of te polypase compo- x n, can be successfully approximated nents, let s say [ ] using te oter component samples x1 [ n ] and a prediction filer, say, P1 (). In tat case, one can expect te difference n to be relatively small: ε = x n P x n m, L, x n m between te prediction output and x [ ] [ ] ( [ ] [ ]) Comparing te above difference wit Figure 1, it can be seen tat te difference sequence corresponds to te lower x n. branc output: [ ] Fig. 1 Analysis stage of te -cannel adaptive filter bank LMS adaptation attempts to obtain te optimum (in te mean-squared sense) prediction filter by updating te filter coefficients according to te minimization of te error sequence, x [ ] n. Te LMS update is a straigtforward operation for adaptive predictive filtering, and can be found in any kind of literature wic deals wit adaptive processes [3], [4]. However, te use of LMS adaptation inside a polypase decomposition structure is a new approac [1]. Notice tat conventional adaptive predictive filtering uses a causal prediction stage. By using polypase components, te adaptive prediction becomes non-causal, improving te inter-correlation of te samples. In [1], teoretical details are provided, and it as been sown tat tis structure statistically minimizes te error signal energy of te polypase components for a steady state signal. We ave adopted tis new adaptive decomposition approac for te analysis of power system waveforms due to its ability to produce low-magnitude signals wile te input is steady, and to produce ig-magnitude portions wen te steady caracteristics of te input waveform is disturbed. A. Simulink implementation of te adaptive filterbank A commonly used tool for automated post processing of signals is MATLAB Simulink. Te built in functions and blocksets, togeter wit te visual programming environment make Simulink a preferable tool. We developed an integrated detection tool using DSP blocksets. Te adaptive filterbank portion of te integrated layout is sown in te dasboxed portion at te left of Figure. Tis portion can be easily matced to te structure sown in Figure 1. Notice tat te signal is firsed decomposed into polypase components by downsampler and integer delay modules. x n, is directly fed Te above polypase component, [ ] 1 into te LMS block as te input signal. Te Oter component, x [ n ], is delayed by a factor of 10, wic is alf of te filter tap size of te LMS block, and compared to te LMS output using a subtraction module. Te result of tis x n and it is fed back to te difference corresponds to [ ] error input part of te LMS block, by wic te adaptation occurs. Te rest of te Simulink layout deals wit te x n. analysis of te produced [ ] III STATISTICAL ANALYSIS Te residual output, x [ ] n, generated by te adaptive decomposition block carries clearly visible information about te detection of various types of events. Terefore, it may be sufficient to present te above decomposition wic produces necessary features for detection, and leave te detection part to te practicing engineer. Neverteless, we give a sample detection metod to post-process te adaptive decomposition output wit satisfactory results. In tis work, we developed an experimental istogram-based analysis stage wic provides automated detection. Te analysis stage consists of a windowed-istogram generation block and te statistical analysis of te istogram. Statistically, te windowed-istogram provides a sorttime approximation of te density function, pdf. Te pdf

3 naturally carries all te statistical information of a process, terefore its approximation, te istogram, is also observed to be useful for generating te detection rule. In te adaptive decomposition structure explanations, we ave seen tat te residual error x [ n ] becomes large in magnitude wen an event appens. Tis is clearly te point tat must be detected. If we monitor x [ n ] signal in a time-windowed manner, we can see tat te istogram is well centered wen te magnitudes of x [ ] n samples are small. Tis is te case wen te waveform exibits no event. As soon as an event appens, due to large-inmagnitude samples of x [ n ], its istogram becomes no longer centered. Instead, te tails of te istogram becomes eavy. An example for tis case is given in Figure 3. Fig. : Simulink layout of te system. Using te above observation, we developed a simple comparison rule in wic we compare te weigt of te center and tail portions of te istogram. If te tails are weak as compared to te center portion, it means tere is no event. If te tails are eavy, ten we trigger an event alarm. Fig. : Example istogram during event. A. Simulink implementation of te statistical analysis Similar to te previous case, we implemented te statistical analysis using Simulink. Inside te integrated layout, te statistical analysis portion corresponds to te bottom-rigt portion of Figure. Te x [ ] n signal is first buffered to produce a timewindowed portion. Ten tis vector is fed into a component wic calculates te istogram. Next, two vectors, corresponding to two portions of te istogram are extracted from te istogram using te vector selector blocks. Tese portions are te central and te tail portions. Te weigt of tese two vectors are ten calculated and compared. Te output of te comparison is te immediate point of detection of an event. Tere are oter blocks in te integrated system wic are activated by te event. Tese blocks are designed to discriminate arcing faults from sags and swells. RMS canges in te waveform before and after te event triggered calculating blocks sows tat arcing fault type events triggers te system frequently during te event. Terefore te RMS comparisons are performed frequently, and te comparator output produces bursts of noisy sapes. On te oter and, sags and swells trigger te detection block more uniformly. Te difference between a sag and a swell is determined by te direction of te comparator output. Detailed explanations of tis portion togeter wit real life demonstrations 3

4 are presented in te experimental results. IV EXPERIMENTAL RESULTS Experimental system is composed of a tree-pase, 380 V, 50 Hz, 5-wire supply loaded wit RL load banks and tree-pase induction motors coupled wit varying mecanical loads. System also enables te use of adjustable speed drives to control te induction motors wen required. Transient events used for testing te proposed metod are pase-to-ground arcing fault, voltage sag due to an induction motor starting, and momentary cange in load. Wile line-to-line voltages of te experimental system were sampled for te first two types of events, pase currents were sampled for te last type. A line-to-ground arcing fault event data is applied to te Simulink implementation of te proposed system and te are observed. In contrast, as te fault recovers, a rise in RMS value again wit noisy steps are observed. Note tat tese noisy steps are due to te nature of te arcing fault. Finally, te system voltage recovers completely and te cange in RMS value of te system voltage becomes zero following te adaptation to te new state. A voltage sag event due to an induction motor starting is illustrated in Fig. 5. Significant reduction in voltage waveform is observed. Detection of any type of event using an adaptive decomposition sceme, wavelet transformation, and oter frequency domain tecniques would become easier if tere is some ig frequency noise at te start of an event. However, as sown in Figure 5, voltage variation during te sag event is very smoot and free of noise. Even in tis case, tere is a large adaptation Fig. 4 Arcing Fault (Top: faulted line-to-line voltage waveform, Middle: Cange in te rms value of te voltage, Bottom: Adaptation output is illustrated in Figure 4. Altoug we staged a lineto-ground fault, te faulted line-to-line voltage waveform, tat is effected from te fault more, is considered due to te fact tat te ground pat may not always be reacable from te location of a digital event recorder in a real life system. Fault first starts around 0.5 second and strikes again after extinguising for several cycles. Tis is also perfectly visible in te adaptation error signal wic is fed in to te statistical analysis block to trigger te digital RMS voltage measurement block. As expected, some amount of voltage collapse is observed during te fault besides te additive noise due to te nature of arcing fault. Altoug te primary event ere is an arcing fault, tis reduction in te voltage magnitude corresponds to a sag event, too. After te fault extinguises and te system voltage returns to normal, ten te event corresponds to a swell event. Middle waveform in Figure 4 sows te cange in te RMS value of te voltage in successive periods. Wit te fault present, as te adaptation error yields large random bursts, a drop in RMS value wit noisy steps Fig. 5 Voltage sag event (Top: line-to-line voltage waveform, Middle: Cange in te rms value of te voltage, Bottom: Adaptation error wic triggers te RMS voltage measurement block and a sarp drop of RMS voltage magnitude is seen as given in te middle waveform of Figure 5. Tis sarp drop of RMS magnitude of te voltage sould be compared wit te reduction wit noisy steps as observed in arcing fault case given in Figure 4. For te voltage sag, reduction in te RMS voltage is larger tan tat of arcing fault. In addition, te duration of voltage sag due to a motor starting is longer tan tat of arcing fault. Wen te motor reaces its nominal speed, te system voltage recovers and a few cycles later adaptation to te state prior to sag event is restored again. In tis case te adaptation time is longer tan tat of te arcing fault due to te fact tat te variation in te RMS voltage is larger and te filterbank requires more time to adapt. Towards te end of te sag event, some load injected noise is overimposed on te voltage waveform causing a step like increase to its original RMS value. Tis step like increase is also very different tan te noisy steps sown in arcing fault case. Monitoring te pase currents is also possible using te proposed sceme. A pase current of te experimental 4

5 system during a momentary cange in load configuration is sown in Fig. 6. Increasing cange in te RMS value of te current during successive cycles indicates tat an additional load is switced on. For an effective detection and discrimination of power quality events, bot voltage and current waveforms sould be monitored simultaneously. For example, non-linear loads suc as adjustable speed drives distort system voltage by causing notcing. Witout monitoring te system current as well, it is difficult to conclude te source of tis distortion. Altoug it causes a voltage sag event, induction motor starting is a normal event. System current information is also required to distinguis te cause of a voltage sag due to a motor starting or a sort circuit. Note tat if te waveform sown in Fig. 6 were a voltage waveform, te variation would be a swell event and it would be detected as suc. In tis case, increase in te ACKNOWLEDGMENTS Tis work is supported by Anadolu University Researc Foundation under Contract: REFERENCES [1] O. N. Gerek and A. E. Cetin, Adaptive Polypase Subband Decomposition Structure for Image Compression, IEEE Trans., Image Processing, Vol. 9 (10), pp , 000. [] I.Daubecies and W. Sweldens, Factoring wavelet transforms into lifting steps, J. Fourier Anal. Applicat., Vol. 4(3), pp , [3] Simon Haykin, Adaptive Filter Teory, Englewood Cliffs, NJ, Prentice Hall [4] O. Arikan, A. E. Cetin, and Engin Erzin, Adaptive Filtering for non-gaussian Stable Process, IEEE Signal Processing Letters, Vol. 1(11), pp [5] S. J. Huang, C. T. Hsie, and C. L. Huang, Application of Morlet Wavelets to Supervise Power System Disturbances, IEEE Trans. Power Delivery, Vol. 14(1), pp , [6] S. J. Huang and C. T. Hsie, Hig-Impedance Fault Detection Utilizing Morlet Wavelet Transform Approac, IEEE Trans. Power Delivery, Vol. 14(4), pp , [7] A. M. Gaouda, M. A. Salama, M. R. Sultan, and A. Y. Cikani, Power Quality Detection and Classification Using Wavelet- Multiresolution Signal Decomposition, IEEE Trans. Power Delivery, Vol. 14(4), pp , [8] Surya Santoso, W. Mack Grady, Edward J. Powers, J. Lamoree, and S. C. Batt, Caracterization of Distribution Power Quality Events wit Fourier and Wavelet Transforms, IEEE Trans. Power Delivery, Vol. 15(1), pp , 000. [9] L. Angrisani, Pasquale Daponte, and M. D Apuzzo, Wavelet Network-Based Detection and Classification of Transients, IEEE Trans. Inst. Measurement, Vol. 50(5), pp , 001. [10] A. H. Tewfik, D. Sina, and C. Jorgensen, On te Optimal Coice of a Wavelet for Signal Representation, IEEE Trans. Information Teory, Vol. 38(), pp , 199. [11] MATLAB Simulink, Natick, MA, Te Matworks, Inc. Fig. 6 Momentary cange in load (Top: line current waveform, Middle: Cange in te RMS value of te current, Bottom: Adaptation RMS value of te voltage is larger tan tat of arcing fault case. In addition, te duration of te swell event is longer tan te arcing fault event. VI. CONCLUSIONS Our experimental results indicate tat istogram based analysis of te adaptive decomposition outputs can clearly distinguis events suc as faults and abrupt canges from te steady state waveforms. Te adaptive decomposition portions feeds te adaptation error sequence to a timewindowed istogram generating block. Te central and tail istogram portions are ten fed into comparators for an event detection. Once tis statistical analysis portion detects an event, RMS measurement blocks are triggered and local RMS values before and after te trigger time are compared. By applying proper tresolds for te final comparator output, power quality events can be classified and dynamic canges in load can be distinguised. 5