Field Test of a System forAutomated Classification of Power Quality Disturbances

H. Englert, J. StenzelInstitute of Power Systems

Darmstadt University of Technology, Germanyhenglert@eev.tu-darmstadt.dejstenzel@eev.tu-darmstadt.de

Abstract— Power quality monitoring has ad-vanced to an important tool for system evalua-tion and solving power-related problems. As aconsequence the increased amount of recordeddata requires more sophisticated analysis meth-ods. The paper proposes a system for au-tomated recording and classification of powerquality disturbances. The system features sta-tistical classification techniques applied to aframe-based event model. Results and expe-riences made by a test in a low-voltage powersystem are presented.

Keywords— Classification system, field test, powerquality, power quality monitoring.

I. INTRODUCTION

By the increased interest in Power Quality (PQ),monitoring PQ parameters has become an essentialtask for performance evaluation and troubleshootingin power systems. Two major fields of application forPQ monitoring can be quoted.

Permanent monitoring in order to characterizepower system performance is a preventive task. Byunderstanding the normal PQ performance, problemscan quickly be identified. In addition it allows utili-ties to prove the system’s compliance to standards likeEN50160 [1] and EN61000 [2][3] in order to match re-quirements of customers. For competitive reasons thelevel of PQ performance may be a valuable informationfor utilities in the near future.

PQ troubleshooting is a reactive application of PQmonitoring. When facing power-related problems infacilities, the principle strategy is first to inspect thesite and to gather system information. Then the systemis monitored over a certain period of time, usually oneor more weeks. After successful analysis of the recordeddata corrective solutions can be applied to overcomethe problem.

These two applications of PQ monitoring have acommon aspect: the analysis deals with a huge amountof recorded data. Although steady-state measurementsare analyzed automatically, the analysis of recordedevents is still done manually by visual inspection, hencevery time consuming and requiring a high expertise.Here a tool which is able to identify certain types ofevents automatically may reduce this analysis effort.

This paper proposes a method that classifies thetype of PQ event automatically by its typical charac-teristics.

Some approaches have been studied applying arti-ficial intelligence techniques before, one using artificialneural networks (ANN) [4], another applying a rule-based expert system [5]. These two approaches haveachieved useful results, but offer low information aboutclassification reliability.

The present approach uses statistical classificationtechniques on a frame-based event model. The basicmethods have their origin in the field of speech recog-nition, where they have successfully been applied [6][7].The key idea is, that in speech a single word consist of asequence of distinct phones, that characterize the word.Hence identification of an unknown spoken word worksby classification of phones and analyzing the phone se-quence. Translating this method to the problem of PQ,it is assumed, that a PQ event can be split into a se-quence of characteristic waveform frames. Similar tolanguage, the set of possible PQ events defines an eventvocabulary, a certain syntax of waveform frames deter-mines a certain event.

A main objective of this classification system (CS)is robust and reliable type identification of PQ events.Robustness means, that for the benefit of abstractioncapabilities and classification performance very specialevent types are not considered as targets. Reliablemeans, that a criterion has to be introduced, whichassesses the quality of the classification result. As anopen system, the CS should allow to improve its ref-erence data in case of poor classification quality. Alsoa simple possibility for modifications and extensions ofthe event vocabulary should be realized.

First an overview of structure and main featuresof the PQ monitoring system prototype is given. Insection III the individual modules of the monitoringsystem and their operation are discussed. Section IVdescribes a field test of the system and presents resultsand experiences, which have been achieved. Conclu-sions and future improvements are summarized in sec-tion V.

II. SYSTEM OVERVIEW

The PQ monitoring system prototype, which incor-porates the CS for PQ events, consists of three major

modules. The structure is given in figure 1.

DAQ / TRIGGERMODULE

CLASSI-FICATIONSYSTEM

MONITORINGRESULTS

PowerSystem

Threshold &SetupParameters

ReferenceData

Event Log: Type, Magnitude, Duration

Steady-State Measurements

Statistics: Trends, Histograms

online offline

Fig. 1. Modules of the PQ monitoring and classification system

The data acquisition (DAQ) and trigger module es-tablishes the interface of the CS to the power system.The CS is voltage oriented, hence the three phase-to-neutral voltages are connected to the monitor by iso-lating probes. In addition three current signals can bemeasured for more detailed PQ analysis. The DAQ andtrigger module permanently calculates steady-state pa-rameters, such as RMS values of fundamental and har-monic voltages and currents. Moving means of theseparameters are stored for offline analysis. Monitor-ing voltage and current signals for exceeding predefinedthresholds is the major task of the module. When a cer-tain threshold is violated, all input signals are recordedfor analysis of the CS module.

The CS is the core of the monitoring system. Outof the recorded voltage signals characteristic featuresare extracted. The type of the recorded event is de-termined by comparison of its extracted features withprevious trained reference data. Afterwards the qual-ity of the decision is tested. In case of sufficient qualitythe decision is accepted, otherwise the event is notedfor manual classification.

The monitoring result module combines trends, his-tograms of processed steady-state measurements andan event log for a detailed PQ diagnosis. At that timethose events with low classification quality can be clas-sified manually by visual inspection. The result maybe adopted by the reference data to enable a higherquality level in case of similar events.

III. SYSTEM OPERATION

The presented prototype of the PQ monitoring sys-tem is realized on a conventional PC platform with aDAQ board. The monitoring software runs under Win-dows NT in combination with Matlab. The mode ofoperations of the modules in figure 1 is described indetail in the following points.

A. DAQ and Trigger Module

As mentioned in section II the primary tasks of thismodule are:

• Interface between power system and CS.• Detection and recording of PQ variations.• Measuring steady-state values.• Data processing for offline modules.

In low-voltage (LV) power systems voltage and cur-rent signals are directly measured by isolation probesand clamp-on current probes. In medium-voltage (MV)systems the voltage signals of local installed voltagetransformers are connected to the isolation probes. Theprobe signals are the input of the DAQ board. Theboard features 8 differential inputs, 12 bit ADC resolu-tion and 200 kS/s sampling rate.

Starting a PQ monitoring campaign first basic mon-itor parameters have to be set. The number and typeof input signals are to be chosen. Next the samplingrate is set. For PQ event classification usually threeline-to-neutral voltages are selected as input. A maxi-mal sampling rate of 66 kS/s is possible for three chan-nel acquisition, but 20 kS/s per channel proved to besufficient. Other settings include averaging interval ofsteady-state measurements and pre- and post-triggerrecording time. In order to detect PQ variations, thetrigger module monitors four different signal parame-ters for threshold violations:

• Peak value of voltage and current,• RMS value of voltage and current,• relative waveform of voltage,• harmonic voltages.

Peak and RMS values are verified every sample.The RMS value is calculated on the basis of a mov-ing window of a half nominal cycle. The detection ofrelative waveform variations works by taking the sumof an actual voltage sample and its corresponding sam-ple half cycle before. Under normal conditions the sumis about 0 V. When the sum exceeds a previous selectedrange, the number of violations is counted over a pe-riod of a half cycle. If this number is higher than athreshold, the system triggers. In addition it is possi-ble to monitor harmonics of all channels by calculatinga FFT over an interval of 200 ms. The system triggers,if an harmonic violates a predefined threshold.

In case of triggering, all signals are recorded. Whenthe trigger is still active after recording time, newrecordings follow, until the trigger gets inactive. Avoid-ing memory overflow, the system turns in a mode ofsample recordings at regular intervals after an adjustedtime. The system gets in normal monitoring mode ifthe trigger turns inactive.

The recorded events are stored on hard disk and canbe analyzed by the CS.

B. Classification System

The appearance of PQ events of the same typecan be very different. Variations in duration, magni-tude and phase are self-evident. Even superpositionsof different types PQ events may occur. To classify

such events, a reduction to characteristic features isinevitable.

As mentioned in section I a methodology is appliedsimilar to speech recognition [6][7]. The key idea ofthe present CS is that the complexity of PQ events canbe sampled into a sequence of basic phenomena. Forthis purpose an event model has been developed. Thismodel is shown in figure 2 by an example event [8].

Event

EventSignal

Frame

UL1

UL2

UL3

U0

Fig. 2. Modules of a PQ monitoring and classification system

The event model consists of four ”event signals”,three recorded line voltage signals (UL1−3) and the cal-culated zero-sequence voltage (U0). Then each eventsignal is decomposed into small signal intervals, so-called ”frames”.

The classification procedure is divided into fivesteps, which are depicted in figure 3.

EVENT SEGMENTATION

FRAME CLASSIFICATION

CONFIDENCE TEST

EVENT RECONSTRUCTION

EVENT IDENTIFICATION

Fig. 3. Modules of a PQ monitoring and classification system

B.1 Event Segmentation

The events signals are decomposed into frames offixed length of a nominal cycle. The trigger time de-termines the starting point of segmentation includingone pre-trigger frame. Then the frames are serializedand passed to the frame classifier.

B.2 Frame Classification

The frame classifier decides the class of each frame.As classes basic PQ phenomena are chosen. A frame

class consists of a ”main characteristic” and may havea few ”attributes”, if existing. The RMS voltage of theframe determines its main characteristic. Attributesdescribe certain waveform phenomena. For instancethe typical incepting voltage waveform of the faultedline can be expressed as ”sag with oscillating transientand arcing character” for transient earth faults in MVpower systems. This phenomenon is categorized by theframe class Uota. Table I shows the selected frameclasses of the frame classifier.

TABLE I

Frame Classes

Main CharacteristicN RMS voltage in acceptable rangeO RMS voltage > 1.1 puU RMS voltage < 0.9 pu

Attributesa arcing characteristicd voltage distortionit impulsive transientot oscillatory transientr repetitive disturbance

Classification starts with extracting relevant fea-tures from a single frame by signal processing. Thesefeatures are peak and RMS value, THD and distortionindex including inter-harmonics. In addition three fea-tures expressing the frame’s time-frequency behaviorare calculated by a modified discrete wavelet transform.These features are combined to a feature vector, whichis the input of the classifier.

The present classification task requires a classifierwith generalization and learning capabilities [9]. Thatis why rule-based classification is not suited for thepresent application. Statistical classifiers fulfil theserequirements, thus selecting a parallel working Bayesand ”k-nearest neighbor” (kNN) classifier. This struc-ture proved efficiency in comparison to fuzzy and ANNclassification techniques [10].

Each classifier of the parallel structure comparesthe input feature vector with its reference data, whichoriginates from simulations and field recordings. Thatreference class, which fits best to the input vector de-termines the class membership of the present frame.When the two class votes of Bayes and kNN classifiercoincide the decision is accepted, otherwise the classi-fication failed.

B.3 Confidence Test

In order to allow only a reliable decision, its qualityhas to be tested. The strategy is to compare classifi-cation result and feature vector with previous success-fully classified frames. Because the classifier techniquesoperate in a way of distance calculation, a confidencefunction can be calculated on the basis of histogramsof previous classifier results [10]. If the actual result

falls near a point of high accumulation, the confidencelevel is high. A decision has to reach a confidence levelof at least 80%, otherwise the whole event is noted formanual classification

B.4 Event Reconstruction

After successful frame classification, the resultingframe types are arranged in form of the original signal.This sequence is denoted ”waveform vector” and formsa row of the ”waveform matrix”. Figure 4 shows framesegmentation and classification of a sample event, anearth fault in a MV power system [8]. The results of

-2

-1

0

1

2

volta

ge (

pu)

-2

-1

0

1

2

volta

ge (

pu)

-2

-1

0

1

2

volta

ge (

pu)

-0.02 0 0.02 0.04 0.06 0.08

-1

-0.5

0

0.5

1

time (s)

volta

ge (

pu)

N Oot Od Od N N

N Uota Ua Uota N N

N Oot Od Od N N

N Oot Od Od N N

Uota Ua

Oot Od

Oot Od

Oot Od

Uota

UL1

UL2

UL3

U0

Fig. 4. Classification of frames and resulting waveform vectorsof a sample PQ event

frame classification are drawn under the plot for eachsignal voltage. Afterwards the waveform vectors arecondensed. This means that same consecutive frametypes are reduced to a single instance (in case of eventtype ”Od” for signals UL2, UL3, U0). Frames of type”N” are removed. The condensed waveform vectors areshown on the right side of the plot.

B.5 Event Identification

The procedure for type identification of the sam-ple event is illustrated in figure 5. First the signaltypes are identified. The waveform vectors a comparedto signal patterns in a look-up table. For example thewaveform vector of line 1 matches with the pattern”Fault, arcing” by the use of wildcards (”*”). Thewaveform vectors of line 2 and 3 can be identified as”Swell, transient”. The waveform vector of the zerosequence voltage is only tested for voltage unbalance.The results of signal identification form the ”signal vec-tor” and are compared to entries in the event patterntable. The present sample event would be identified asshown in the figure. When no matches can be found

L1L2L3

Uot* U*a U*

Fault, arcing Fa

O*t* O*

Swell, transient Ot

. . .

. . .

. . .

. . .

. . .

. . .

Fa O* O*

Phase to ground fault,arcing, single phase

Ot N* N*

Capacitor energizing,single phase

Example ofWaveform

Vectors

Table ofSignal Patterns

Table ofEvent Patterns

‚0‘

ASY

ASY

SYN

ASY!N

Uota Ua

Oot Od

Oot Od

Oot Od

Uota

Fig. 5. Modules of a PQ monitoring and classification system

in signal or event pattern table, the event is noted formanual classification.

The advantage of these look-up tables is their sim-ple expandability. Like a construction kit, existent pat-terns can be modified, new signal and event pattern canbe added.

C. Results Module

As mentioned in section II the results of steady-state measurements and of the event classification sys-tem are brought together to enable an assessment ofthe monitoring campaign. Trend and histogram plotsof RMS voltages, harmonics, THD and voltage unbal-ance can be generated. But most important is the eventlog, where trigger time, magnitude, duration and ofcourse the type together with its confidence level islisted. In case of unknown event types, it is possible toplot the recorded event and classify it manually. De-pending where the classification failed, in frame clas-sification or event identification, unknown waveformscan improve the reference data or new signal and eventpatterns can be added to the look-up tables, respec-tively. A summary of occurred event types concludesthe PQ analysis module.

IV. FIELD TEST

A. Monitoring Location

A field test was carried out at two different low-voltage power systems of the university campus. Fig-ure 6 shows the diagram of the campus power system.An internal 6 kV-power system distributes the electricenergy to several low-voltage systems.

At two different service entrances of low-voltage sys-tems – locations A and B – the PQ monitoring sys-tem was installed. The two investigated power sys-tems supply different loads. The system at location Asupplies two office buildings and a laboratory. Hencelocation A is mainly characterized by lighting and com-puter loads and in addition by a 150 kW induction mo-tor with power factor correction, which is installed atthe laboratory. The power system at location B sup-plies experimental set-ups for adjustable speed drives(ASD).

2 x 5 MVAuk = 6%

20 kV

6 kV630 kVAuk = 4%

150 kVAuk = 4%

0.4 kV

M MA B

Fig. 6. One-line diagram of monitored power system

The three phase-to-neutral voltages were monitoredat the two service entrances over a week period, respec-tively. Monitoring could not performed simultaneously,because only one PQ monitoring system was available.

B. Results

B.1 Steady-State Measurements

Figure 7 shows the trend of the steady-state-measurements recorded at location A. The extrema

0.4

0.6

0.8

1

0

0.02

0.04

0.06

0.08

Sa Su Mo Tu We Th Fr Sa0

0.01

0.02

0.03

0.04Voltage Unbalance

time (d)

abso

lute

abso

lute

Vol

tage

(pu

)

THD

RMS Voltage

Fig. 7. Trends at location A, 09/09/2000 - 09/16/2000

of RMS voltage and THD can be seen. The lower plotshows the trend of voltage unbalance. During a weekwith thunderstorms, events in form of voltage sags areobvious in the RMS voltage trend. Effects of thesesevents are visible in THD and voltage unbalance plot.According to standard EN50160 [1], 95% of the steady-state measurement are to be in acceptable range ofRMS voltage (0.9 − 1.1 pu), THD (< 8%) and volt-age unbalance (< 2%). Even with those voltage sagscompliance to EN50160 is kept.

In comparison to location A the trends measuredat location B, depicted in figure 8, show no signif-icant events. The RMS voltage shows a lower level,

0.9

0.95

1

1.05

1.1

0

0.01

0.02

0.03

0.04Voltage Unbalance

abso

lute

abso

lute

Vol

tage

(pu

)

THD

RMS Voltage

Sa Su Mo Tu We Th Fr Satime (d)

0

0.02

0.04

0.06

0.08

Fig. 8. Trends at location B, 11/04/2000 - 11/11/2000

because of the transformers tap changer remained onUn = 380 V. Also THD has a lower level, which canbe mainly traced back to missing computer loads andfluorescent lights. In addition with a low voltage un-balance trend the compliance of the trend to EN50160is proved.

B.2 Event Analysis

Table II shows the summary of the classification ofthe events recorded at location A.

TABLE II

Events at location A

Event type #Voltage sag, balanced 3Voltage sag, unbalanced 1Capacitor energizing 47unknown 9

total: 60

Totally 60 events were recorded. The majority wasclassified as capacitor energizing, that can be mostprobably traced back to the power factor correctionof the induction motor. Four voltage sags are identi-fied. Knowing the transformer type, interpretations ofthe events can be drawn. Three were balanced sags,in spite of unequal phase-to-neutral voltages. Thusthey can be assumed as result of remote faults in trans-mission systems. The unbalanced sag was an internalevent. Nine events kept unknown, because of unsuffi-cient confidence level after frame classification. Visualinspection of those unknown events identified them asslight waveform distortions.

At location B totally 23 events were recorded. Afterautomated classification all events remained unknownafter frame classification. By visual inspection the un-known events were identified as typical events of ASD

-0.005 0 0.005 0.01 0.015 0.02 0.025-1.5

-1

-0.5

0

0.5

1

1.5

time (s)

volta

ge (

pu)

Fig. 9. Added frame class ”Nasd”

starting. A sample frame is depicted in figure 9. Thesolution was to introduce a new frame class ”Nasd” byadopting an event frame to the reference data. TableIII shows the results after reclassification the events.

TABLE III

Events at location B

Event type #ASD starting 22unknown 1

total: 23

The single unknown event showed a slight waveformdistortion.

V. CONCLUSIONS

This paper presented a power quality monitoringsystem, which features an automated event classifica-tion. Main goal of the system is robust and reliableclassification of power quality events.

The system allows modifications and extensions of itsknowledge data by a type of construction kit for powerquality events. After description of structure and op-eration of the system prototype, results of a field testwere presented.

The field test proved efficency of the system andmade the advantage of the open classification systemobvious. By introducing a new frame type nearly allevents could be classified successfully. So it is possi-ble to build a reference data of a specific power sys-tem. This allows a simple identification of new un-known event types in a set of recorded events.

Future work will include the extension of the eventreference data and the signal and event pattern tables.It is also intended to gain experience by system testsin medium-voltage power systems.

References

[1] EN 50160, ”Voltage characteristics of electricity suppliedby public distribution systems”, European Standard CEN-ELEC, November, 1994.

[2] EN 61000-2-2, ”Electromagnetic Compatibility, Section 2:Compatibility levels for low frequency conducted distur-bances and signaling in low-voltage power supply systems”,(IEC 1000-2-2: 1990), German Version EN 61000-2-2, 1993.

[3] EN 61000-2-4, ”Electromagnetic Compatibility, Section 4:Compatibility levels in industrial plants for low frequencyconducted disturbances”, (IEC 1000-2-2: 1994), GermanVersion EN 61000-2-4, 1994.

[4] A. Gosh, D. Lubkeman, ”The classification of power systemdisturbance waveforms using a neural network approach”,IEEE Transactions on Power Delivery, Vol. 10, No. 1, Jan-uary, 1995, pp. 109-115.

[5] S. Santoso, J. Lamoree, M. Grady, E. Powers, S. Bhatt, ”Ascalable PQ Event Identification System”, IEEE Transac-tions on Power Delivery, PE-109PRD (09-99), 1999.

[6] R. Hoffmann, Signalanalyse und -erkennung, Springer,Berlin-Heidelberg, 1998.

[7] L. R. Rabiner, B.-H. Juang, Fundamentals of speech recog-nition, Prentice Hall, Englewood Cliffs, 1993.

[8] IEEE Working Group, power quality event char-acterization (1159.2), test waveform ”wave7.xls”,http://grouper.ieee.org/groups/1159/2/testwave.html.

[9] H. Englert, J. Stenzel, ”Design of a classification system forpower quality disturbances”, 12th International Conferenceon Power System Protection, Bled, Slovenia, 2000.

[10] H. Englert, J. Stenzel, ”A parallel classifier structure foridentification of power quality disturbances”, accepted paperfor 2000 National Power System Conference, December 20-22, Bangalore, India.