APPLICATION OF ABRUPT CHANGE DETECTION IN AUTOMATIC DISTURBANCE RECOGNITION IN ELECTRICAL POWER SYSTEMS

by

Abhisek Ukil

Notes for Tutorial 3 Second World Congress on Lateral Computing Bangalore, India 16th – 18th December, 2005

Contact Details: Abhisek Ukil Tshwane University of Technology, Dept. of Mathematical Technology, 175 Nelson Mandela Drive, Private Bag X680, Pretoria, 0001, South Africa. E-mail: [email protected] Tel: +27 72 736 9557 Fax: +27 86 655 1449

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

2

CONTENTS 1

INTRODUCTION....................................................................................................................................... 3

2

ANALYSIS OF DISTURBANCE DATA.................................................................................................. 3 2.1 2.2

3

AUTOMATED ANALYSIS SYSTEM...................................................................................................... 5 3.1 3.2 3.3 3.4 3.5 3.6

4

DISTURBANCE ANALYSIS SCHEME ....................................................................................................... 4 EXISTING DISTURBANCE ANALYSIS SYSTEM ....................................................................................... 5 FUNCTIONAL SPECIFICATION ............................................................................................................... 5 SYSTEM OVERVIEW .............................................................................................................................. 6 RECORDINGS FROM DFRS .................................................................................................................... 7 ABRUPT CHANGE DETECTION-BASED SIGNAL SEGMENTATION ........................................................... 8 FEATURE VECTOR CONSTRUCTION ...................................................................................................... 8 PATTERN MATCHING AND DISTURBANCE RECOGNITION ..................................................................... 9

ABRUPT CHANGE DETECTION......................................................................................................... 11 4.1 VARIOUS TECHNIQUES ....................................................................................................................... 11 4.2 RECURSIVE IDENTIFICATION METHOD ............................................................................................... 12 4.3 WAVELET TRANSFORM METHOD ....................................................................................................... 13 4.3.1 Wavelet Transform Analysis ......................................................................................................... 13 4.3.2 Signal Decomposition ................................................................................................................... 14 4.3.3 Application of Threshold Method ................................................................................................. 14 4.3.4 Application Result......................................................................................................................... 15

5

ADAPTIVE ABRUPT CHANGE DETECTION ................................................................................... 16 5.1 ADAPTIVE WHITENING FILTER METHOD ............................................................................................ 16 5.2 ADJUSTED HAAR WAVELET METHOD ................................................................................................ 19 5.2.1 Introduction .................................................................................................................................. 19 5.2.2 Overview of the Haar Wavelet...................................................................................................... 19 5.2.3 Adjustment to the Haar Wavelet ................................................................................................... 20 5.2.4 Application of Adjusted Haar Wavelet.......................................................................................... 21

6

COMPLETE ALGORITHM ................................................................................................................... 22

7

VARIOUS APPLICATIONS ................................................................................................................... 24 7.1 7.2 7.3 7.3.1 7.3.2 7.3.3 7.4 7.4.1 7.4.2 7.4.3 7.4.4 7.4.5 7.4.6 7.5

8

EVENT-SPECIFIC SIGNAL SEGMENTATION .......................................................................................... 24 SYNCHRONIZATION ............................................................................................................................ 24 RELAY PERFORMANCE MONITORING ................................................................................................. 25 Fastest Relay Operating Time ...................................................................................................... 25 Auto-Reclosing of the Circuit-Breakers........................................................................................ 26 Main-1 and Main-2 Relay Operation............................................................................................ 27 ADDITIONAL DISTURBANCE ANALYSIS .............................................................................................. 27 Cleared Single Phase Fault .......................................................................................................... 27 Uncleared Single Phase Fault ...................................................................................................... 28 Circuit-breaker Restrike ............................................................................................................... 28 Reactor Ring Down....................................................................................................................... 29 Capacitive Voltage Transformer Transient Behavior................................................................... 29 Energizing of a Transformer......................................................................................................... 30 ANALYSIS OF POWER SIGNALS FROM THE MEXICAN NETWORK ........................................................ 30

CONCLUSION.......................................................................................................................................... 31

REFERENCES ................................................................................................................................................... 31

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

1

3

Introduction

The analysis of faults and disturbances in power systems is a fundamental foundation for a secure and reliable electrical power supply. The Automatic Disturbance Recognition Project focuses on automatic disturbance recognition and analysis based on the disturbance signals obtained from the digital fault recorders in the power transmission network in South Africa. In an automated recognition-oriented task, typically, the first step is to apply the abrupt change detection algorithms to segment the fault recordings into different segments, such as, the pre-fault segment, fault, after the circuit-breaker opening, after auto-reclosure of the circuit-breakers, and so forth. The next step involves constructing the appropriate feature vectors for the different segments. Finally, pattern-matching algorithms are applied using these feature vectors to accomplish the disturbance recognition and analysis tasks. The main focus of this tutorial is to take the first step towards an automatic disturbance analysis, namely abrupt change detection-based segmentation. The aim is to accurately estimate the time-instants of the changes in the signal model parameters during different events of the disturbance and accordingly segment the signal. This is critical for improving the fault recognition rate and automatic analysis quality. It also provides scope for analyzing certain kinds of disturbances directly from the segmented recordings. This work proposes and establishes various new and customized techniques of detecting the abrupt changes in the electrical power systems disturbance signals. Using those techniques, event-specific automatic signal segmentation and various applications like synchronization, relay performance monitoring and such like are performed. Commercial implementation of the project as Application Service Provider (ASP) solution is also proposed.

2

Analysis of Disturbance Data

In 1998, a 24 hour support service for ESKOM (South African power utility) National Control was introduced to supply operating personnel with information from disturbance records analysis with an aim to enhance operational decisions, immediately identify potential risk and improve response time to latent weaknesses or failures [1]. Since 1993 ESKOM is installing centralized per substation disturbance recorders, especially digital fault recorders (DFRs) on the transmission lines (currently installed at 136 substations). The DFRs are installed on the feeder bays, with an additional few installed on the Static Var Compensators (SVCs) [2]. The recorders monitor all voltages and currents and 30 selected protection operation outputs per line. The recorders are triggered by protection operation pulses and, some of them, by dv/dt (change of phase voltage). The recording duration varies from 2 to 12 seconds with scanning frequency of 2.5 kHz. The implementation of the X.25 communication facility was initiated in 1995 and presently all recorders are remotely accessible [3]. In line with the commissioning of new recorders the amount of mathematical analysis of recorded signals was continuously growing. Storage and retrieval of disturbance data became a real challenge (presently approximately 2000 records a year).

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

4

Presently semi-automatic software is available for data acquisition and storage in the performance database. Results from the analysis of the digital fault records are also captured in a database called ‘TIPPS’ (Transmission Integrated Plant Performance System) [4]. In the fully developed automatic system, mathematical analysis of all (presently known) patterns of incorrect behavior (disturbances) should be performed automatically and only a short message should be issued for the controllers (operating personnel) summarizing the required information and knowledge.

2.1

Disturbance Analysis Scheme

Fig. 1 shows the disturbance analysis scheme employed at the National Control, ESKOM, South Africa. In Fig. 1, the blocks in solid-lines indicate the existing systems, whereas the blocks in dotted-lines indicate the proposed automatic disturbance recognition and analysis systems.

Fig. 1: Disturbance Analysis Scheme at National Control, ESKOM, South Africa.

As per Fig. 1, when an event (Tx-event) on the power transmission network occurs National Control (NC) will be informed via a SCADA (Supervisory Control And Data Acquisition) network and also the DFRs will be triggered. The protection engineers provide a 24-hour support through the analysis of these records. It is possible to download the fault record via the X.25 communication network. The protection engineer has to do a manual analysis to create a report for National Control. The manual analysis is cumbersome and takes a lot of time, typically from 1 to 10 hour or more, depending on the complexity and severity of the event. Additionally an automatic scanning PC downloads the records. They need to be saved

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

5

manually at a file server. Results from the analysis of the digital fault records are also captured in the TIPPS database [4]. Due to large amount of historical data available trend analysis on the performance of primary and secondary equipment can be done. An automated analysis system (shown in dotted lines in Fig. 1) would be much faster, and it could be placed in two different ways. The more efficient way is to do the automated analysis direct at the substation and only to transmit a report to NC. This is the fastest way to do it but additional PCs and the communication infrastructure (for example a LAN) are needed. The other way is to run the automated analysis on a central PC. The disadvantage is that it takes much more time to download the whole record than to send a report. The big advantage is the cost factor because only one PC is needed and the communication network is much simpler. The automated analysis system should also be able to write the analysis results to the TIPPS database [5].

2.2

Existing Disturbance Analysis System

Majority of digital fault recorders in ESKOM are from Siemens: ‘SIMEAS-R’ and ‘OSCILLOSTORE P531’ recorders. ‘SIMEAS R’ comes with 16-bit resolution and a 12.8 kHz maximum scanning frequency per channel. There are two types of central unit: one with 8 analog and 16 binary channels, and one with 32 analog and 64 binary channels [3]. ‘OSCILLOSTORE P531’ comes with 8- and 12-bit resolution and a 5 kHz maximum scanning frequency. It has 31 Data acquisition units (DAUs) per central unit, i.e., 124 analog or 992 binary data acquisition channels [3]. Every feeder uses 3 Data acquisition unit (DAU): 2 x ADAU (analog data acquisition units used for 4 x voltage and 4 x current signals), and 1 x BDAU (binary data acquisition unit used for up to 32 digital signals) [3]. The 32 Binary values are either stored as a ‘0’ or a ‘1’ and indicates the status of a contact e.g., breaker auxiliary contact. Analog values indicate the magnitude of an analog signal (voltage or current) measured at a specific point in time. Following the IEEE COMTRADE standard [6], the DFR recordings are provided as input to the existing semi-automatic fault analysis software which uses Discrete Fourier Analysis and Superimposed Current Quantities [3]. All the disturbance analysis are done manually which is quite complex and time-consuming process [3]. Also some information from the existing analysis system, especially the circuit breaker binaries are very often not complete, that is why they are not so trustworthy and need to be handled with care. A fault analysis system based only on the binaries should be avoided [5].

3

Automated Analysis System

3.1

Functional Specification

The first requirement for the Automated Analysis System (AAS) is that it must be able to read a COMTRADE file [6]. The AAS must be able to automatically import the COMTRADE file from a user defined directory, do the analysis and produce the results in a format which can be viewed by any text editor. The distribution of the results should be done

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

6

through any of the following media: fax, e-mail, SMS, print or Web [4]. The AAS must extract from the COMTRADE file the following information [4]: ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ

Faulted phase(s) Fault type Total fault duration Main 1 protection operating time Main 2 protection operating time Fault location Fault resistance DC offset Breaker operating time Auto re-close time.

Apart from the above the AAS must also determine and report on the following as described in [4]: ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ ƒ

3.2

Was the fault on the specific feeder? Did the main 1 relay operate? Did the main 1 relay operate correctly? Did the main 2 relay operate? Did the main 2 relay operate correctly? Was the main 1 permissive carrier signal sent, and was this correct? Was the main 2 permissive carrier signal sent, and was this correct? Was the main 1 permissive carrier received, and was this correct? Was the main 2 permissive carrier received, and was this correct? How did the breaker operate – 1 pole or 3 poles? Did the breaker auto re-close (ARC)? What was the magnitude of the fault current? What was the magnitude of the neutral current? What was the magnitude of the healthy phase currents during the fault? What was the depression in voltage on the faulted phase/s? What was the depression in voltage on the healthy phase/s? What were the dominant frequencies before and during the fault? Did any of the breaker poles re-strike?

System Overview

In the direction towards an automatic recognition-oriented task, we would first apply the abrupt changes detection algorithms on the fault recordings. This will segment the fault recordings into different event-specific segments, namely, pre-fault segment, after circuitbreaker opening, after auto-reclosure of the circuit-breakers. Then we would construct the appropriate feature vectors for the different segments; finally the pattern-matching algorithm would be applied using those feature vectors to accomplish the fault recognition and associated tasks [5].

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

7

The complete architecture of the proposed automated analysis system is shown in Fig. 2 [5]. In Fig. 2, the architecture of the automated analysis system is a sequential top-down block diagram. Each of the blocks is described in the following sections.

Fig. 2: Architecture of the Automated Analysis System.

3.3

Recordings from DFRs

DFRs are highly accurate recording instruments providing sampled waveform and contact data using relatively high sampling rate, typically above 5 kHz (a sample every 0.2 milliseconds). As per IEEE COMTRADE standard [6], each event shall have three types of files associated with it [5]. Each of the three types carries a different class of information: header (*.HDR), configuration (*.CFG) and data (*.DAT). Each record (row) in the data file (*.DAT) obtained from the ESKOM DFRs has 42 data items (columns): first column is the sample number, second column is the time in microseconds from the beginning of the record, next 8 columns are the 4 analog voltage and 4 analog current recordings, next 32 columns are the 32 binary points representing the contact changes [5]. Fig. 3 shows a sample record from a data file.

Fig. 3: Digital Fault Recorder data file format.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

3.4

8

Abrupt Change Detection-based Signal Segmentation

Detection of abrupt changes in signal characteristics is a much-studied subject with many different approaches. It has a significant role to play in failure detection and isolation (FDI) systems and segmentation of signals in recognition-oriented signal processing [7]. Many of these signals are quasi-stationary in nature, that is, they are composed of segments of stationary behavior with abrupt changes in their characteristics in the transitions between different segments. It is important to find the time instants when the changes occur and to develop models for the different segments during which the system does not change [8]. Segmentation of the fault recordings by detecting the abrupt changes in the characteristics of the fault recordings, obtained from the DFRs of the power network in South Africa, is the first step towards automatic disturbance recognition and analysis. Abrupt change detection algorithms collectively segment the fault recordings into different event-specific segments, namely, pre-fault segment, after initiation of fault, after circuit-breaker opening and after auto-reclosure of the circuit-breakers. Ukil and Zivanovic [8] conducted a comparative study of the different technologies for abrupt change detection and categorized the techniques are broadly as simple methods, linear model-based approaches, model-free approaches and nonparametric approaches. Utilizing the segmented recording and working on the specific segments such as, pre-fault, fault, after circuit-breaker opening, etc, is critical for improving the fault recognition rate and automatic analysis quality. Besides facilitating further complex feature vector analysis and pattern recognition leading to automated disturbance recognition and analysis, the abrupt change detection-based segmentation procedure followed by the synchronization step can be directly used to monitor the relay performance as studied by Ukil and Zivanovic [9]. Abrupt change detection based segmentation also provides huge scope for analyzing certain kinds of disturbances directly from the segmented recordings before conforming with any further significant and complex feature vector analysis [10].

3.5

Feature Vector Construction

The main objectives of this step are as follows: • • •

Development of the parameter identification techniques for various models. Development of parameter identification technique in presence of non-linearity, e.g., case with the slow time-varying amplitude and frequency. Feature extraction from the other sources apart from analog signals.

Practical recorded signal is corrupted with noise and nuisance components (higher frequencies due to wave transients and measurement system bias). Link between the system model and the signal model could be established through the State Space Theory. To filter noise and bias, extended version of the Structured Total Least Squares method [11] will be used.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

9

To include network nonlinearities, the system can be modeled using the system of linear differential equations perturbed with nonlinearity (quasi-linear system). Here in the modeling we can use Averaging approximation technique (only the first order approximation) called in engineering the Describing Function method [12]. The signal model has the same structure as described above but the parameters are slow timevarying. Time varying version of the Structured Total Least Squares method should be adopted here. Higher harmonics associated with a non-linearity will be treated as nuisance components and filtered out in the estimation process. Link between the time-varying signal model and the system model will be established using the Averaging Theory [12]. Other sources for feature extraction process are binary disturbance records (opening & closing of various contacts associated with recorded event) and expert knowledge about disturbances provided by ESKOM specialists. Zivanovic [13,14,15] also proposed a frequency estimation algorithm based on local polynomial approximation technique for use in the feature vector construction. The semi-parametric estimation approach for the feature vector construction is depicted in Fig. 4.

Fig. 4: Feature vector construction using semi-parametric approach.

3.6

Pattern Matching and Disturbance Recognition

The main objectives of this step are as follows: •

The best set of features (obtained from parameter identification process and other sources) should be selected to represent disturbances.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

• •

10

Adopting the classification algorithm and training using recorded data and human expertise. The simple demonstration prototypes of the following analysis algorithms will be developed: Fault location algorithm, Analysis for resonant conditions in the system with trapped energy, Protection and control system performance analysis, Electromechanical oscillation and system stability analysis.

Development in this step will rely on the available techniques from the Statistical Learning field [16]. The classification task should deal with data of mixed type (continuous values, breaker open & close, expert knowledge etc). Keller, Henze and Zivanovic proposed a Support Vector Machine (SVM)-based fault classifier in [4]. SVM is a training algorithm for learning classification and regression rules from data, first introduced by Vapnik in the 1960s [16]. The SVM-based fault classifier shall separate ground faults from non-ground faults. Detail mathematical description of the fault classification system can be found in the practical training report by Henze [17]. The input feature space is two-dimensional and consists of the magnitude of the 50 Hz component of the neutral current (X1) and the zero sequence Voltage (X2). Typically 10 ground faults and 10 phase-to-phase faults are used as training data [4]. For the output, Yi = 1 means ground fault and Yi = –1 means none ground fault. After training the SVM classifier, another data set is used to test it. Typical test data set consists also of 10 ground and 10 non-ground faults [4]. Fig. 5 shows the algorithm of the SVM fault classifier as a flowchart.

Fig. 5: SVM-based fault classifier algorithm flowchart.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

4

11

Abrupt Change Detection

Detection of abrupt changes in signal characteristics is a much studied subject with many different approaches. A possible approach to recognition-oriented signal processing consists of using an automatic segmentation of the signal based on abrupt changes detection as the first processing step. A segmentation algorithm splits the signal into homogeneous segments, the lengths of which are adapted to the local characteristics of the analyzed signal. This can be achieved either on-line or off-line [7]. Mathematically, we can consider a sequence of observations depending only upon one scalar parameter θ . Before the unknown change time t 0 , the parameter θ is equal to θ 0 , and after the change it is equal to θ 1 ≠ θ 0 . The problems are then to detect and estimate the changes in the parameter and the change time-instant t 0 [8].

4.1

Various Techniques

In the scope of this tutorial, we focus on power system fault and disturbance signals. Many of these signals are quasi-stationary in nature i.e., these signals are composed of segments of stationary behavior with abrupt changes in their characteristics in the transitions between different segments. It is important to find the time-instants when the changes occur and to develop models for the different segments during which the system does not change. To accomplish the abrupt changes detection, hence segmentation of the fault and disturbance signals, the following approaches are considered. •

Simple methods o Superimposed Current Quantities [2] o Linear Prediction Error Filter [18] o Adaptive Whitening Filter [19]

•

Linear Model-based approach o Additive Spectral Changes [7] o Autoregressive (AR) Modeling and Joint Segmentation [20] o State-Space Modeling and Recursive Parameter Identification [21, 22]

•

Model-free approach o Support Vector Machines [23]

•

Nonparametric approach o Discrete Fourier Transform [24] o Wavelet Transform [25, 26].

Among the different techniques studied, the following are the most promising ones in respect to our application domain. •

Linear Model-based approach o State-Space Modeling and Recursive Parameter Identification

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

•

4.2

12

Nonparametric approach o Wavelet Transform.

Recursive Identification Method

As power system has a deterministic system model, we can apply the model based approach for abrupt change detection. In the proposed Recursive Identification [21] technique, several Kalman filters that estimate the parameters are run in parallel, each of them corresponding to a particular assumption about when the system actually changed. The relative reliability of these assumed system behaviors is constantly judged, and unlikely hypotheses are replaced by new ones [22]. This is followed by smoothing filtering operation. A typical recursive identification algorithm is:

θˆ(t ) = θˆ(t − 1) + K (t )[ y (t ) − yˆ (t )] θˆ(0) = θ 0 ,

(1)

where θˆ(t ) is the estimate of the parameter vector θ at time t , and y (t ) is the observed output at time t . yˆ (t ) is a prediction of the value y (t ) based on observations up to time t − 1 and also based on the current model (and possibly also earlier ones) at time t − 1 . The gain K (t ) determines in what way the current prediction error y (t ) − yˆ (t ) affects the update of the parameter estimate. An optimal choice of the gain K (t ) can be computed using the Kalman filter [22]. In Fig. 6, we show the application results for a fault signal obtained from ESKOM DFRs implemented using MATLAB® based on recursive identification algorithm.

Fig. 6: Segmentation of RED-Phase Voltage signal using the Recursive Identification method.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

13

In Fig. 6, the upper plot shows the original DFR recording for the voltage during a phase-toground fault in the RED-Phase, sampled at a sampling frequency of 2.5 kHz. The middle plot shows the modeling of the system using the recursive identification technique. The lower plot shows the time-instants of the changes in the signal characteristics, marked by the impulse indicators, indicating the different signal segments owing to different events during the fault. For example, segment A indicates the pre-fault section and the fault inception, segment B the fault, segment C the opening of the circuit-breaker, segment D auto-reclosing of the circuitbreaker and system restore [22]. 200 disturbance records were tested with the proposed algorithm and 95% accuracy in correct segmentation was achieved. The proposed algorithm for the abrupt change detection and signal segmentation took an average computation time of 5.585 seconds. An Intel® Celeron® 1.9 GHz, 256 MB RAM computer was used for all the application tests [22].

4.3

Wavelet Transform Method

Wavelet transform is particularly suitable for the power system disturbance and fault signals which may not be periodic and may contain both sinusoidal and impulse components. In particular, the ability of wavelets to focus on short intervals for high-frequency components and long intervals for low-frequency components improves the decomposition of the fault signals into finer and detailed scales, facilitating further effective signal processing and analysis [25]. In this application, wavelet transform is used to transform the original fault signal into finer wavelet scales, followed by a progressive search for the largest wavelet coefficients on that scale [25]. Large wavelet coefficients that are co-located in time across different scales provide estimates of the changes in the signal parameter. The change time-instants can be estimated by the time-instants when the wavelet coefficients exceed a given threshold (which is equal to the ‘universal threshold’ of Donoho and Johnstone [27] to a first order of approximation).

4.3.1 Wavelet Transform Analysis Wavelet analysis is the breaking up of a signal into shifted and scaled versions of the original (mother) wavelet which is an oscillatory waveform of effectively limited duration that has an average value of zero. The continuous wavelet transform (CWT) is defined as the sum over all time of the signal multiplied by scaled, shifted versions of the wavelet function ψ . The CWT of a signal x(t ) is defined as ∞

CWT (a, b) =

∫ x(t )ψ

* a ,b

(t )dt ,

(2)

ψ a ,b (t ) =| a | −1 / 2 ψ ((t − b) / a ) .

(3)

−∞

where,

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

14

ψ (t ) is the mother wavelet, the asterisk in (2) denotes a complex conjugate, and a, b ∈ R, a ≠ 0, ( R is a real continuous number system) are the scaling and shifting parameters respectively. By choosing a = a 0m , b = na 0m b0 , t = kT in (2), where T = 1.0 and k , m, n ∈ Z , and Z is the set of positive integers, we get the discrete wavelet transform (DWT), DWT (m, n) = a 0− m / 2

(∑ x[k ]ψ

*

)

[(k − na 0m b0 ) / a 0m ] .

(4)

4.3.2 Signal Decomposition The Multiresolution Signal Decomposition (MSD) technique [25] decomposes a given signal into its detailed and smoothed versions. MSD technique can be realized with cascaded Quadrature Mirror Filter (QMF) banks [25]. A QMF pair consists of two finite impulse response filters, one being a lowpass filter and the other a highpass filter. The output of the lowpass filter is the smoothed version of the input signal and used as the next QMF pair’s input. The output of the highpass filter is the detailed version of the original signal. For our application, Daubechies 1 & 4 [25] mother wavelets are used. Using these mother wavelets and MSD technique the original fault signal is transformed into the smoothed and detailed versions. We use the detailed version for threshold checking to estimate the change time-instants.

4.3.3 Application of Threshold Method After transforming the original fault signal using the wavelet transform, we will search progressively across the finer wavelet scales for the largest wavelet coefficients on that scale [9]. As wavelet coefficients are changes of averages, so a coefficient of large magnitude implies a large change in the original signal. Large wavelet coefficients that are co-located in time across different scales provide estimates of the cusp points [9] hence time-instants of abrupt changes. The change time-instants can be estimated by the instants when the wavelet coefficients exceed a given threshold which is equal to the ‘universal threshold’ of Donoho and Johnstone [27] to a first order of approximation. The universal threshold T is given by T = σ 2 log e n ,

(5)

where σ can be the median absolute deviation of the wavelet coefficients, or standard deviation, and n is the number of samples of the wavelet coefficients. Median absolute deviation is a good choice because median is hardly influenced by a small fraction of extreme values [25]. However, experimenting with the real disturbance signals, standard deviation is finally chosen to cover a wider range of disturbance signals than that covered by the median absolute deviation.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

15

4.3.4 Application Result In Fig. 7, we show the application results for a fault signal obtained from ESKOM DFRs implemented using MATLAB® based on the wavelet transform.

Fig. 7: Segmentation of the RED-Phase Current Signal using the Wavelet Transform method.

In Fig. 7, the original DFR recording for the current during the phase-to-ground fault in the RED-Phase is shown in the top section, wavelet coefficients for this fault signal (in blue) and the universal threshold (in black, dashed) are shown in the middle section. The change timeinstants computed using the threshold checking (middle section) followed by smoothing filtering are shown in the bottom section. It should be noted that only the highpass filter output of the QMF pair is shown, so the wavelet coefficients in the middle section indicate half of the total samples of the original signal. The time-instants of the changes in the signal characteristics, in the lower plot in Fig. 7, indicate the different signal segments owing to different events during the fault, for example, segment A indicates the pre-fault section and the fault inception, segment B the fault, segment C the opening of the circuit-breaker, segment D autoreclosing of the circuit-breaker and system restore [25]. 200 disturbance records were tested with the proposed algorithm and 99% accuracy in correct segmentation was achieved. The proposed algorithm for the abrupt change detection and signal segmentation took an average computation time of 0.431 seconds. An Intel® Celeron® 1.9 GHz, 256 MB RAM computer was used for all the application tests [25, 26].

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

5

16

Adaptive Abrupt Change Detection

The proposed recursive identification and wavelet method estimates exactly the time-instants of the changes in the signal model parameters during the pre-fault condition, after the initiation of fault, after the circuit-breaker opening and auto-reclosure etc. These are effective for about 60% of the disturbance signals, having distinct abrupt changes in the signal model parameters. However, about 40% of the disturbance signals do not show distinct abrupt changes in the signal parameters. In those cases, we have to apply adaptive abrupt change detection. For the adaptive abrupt change detection, we propose two methods, 1. Adaptive Whitening Filter [26] 2. Adjusted Haar Wavelet [28].

5.1

Adaptive Whitening Filter Method

The Adaptive whitening filter based on the adjusted Fourier filter [19] is used to pre-filter the original fault signal. The wavelet transform method is then used, as discussed in the chapter 5, to transform the pre-filtered fault signal into the finer wavelet scales, followed by a progressive search for the largest wavelet coefficients on that scale [26] to perform the abrupt change detection, as depicted in Fig. 8.

Fig. 8: Architecture of Abrupt Change Detection based on the Adaptive Whitening Filter.

The optimal Linear Prediction Error (LPE) filter for transient monitoring would perfectly decorrelate the signal, leaving only white noise (whitening filter) [19]. The finite impulse response (FIR) whitening filter can be defined by A( z −1 ) = 1 − z − C ,

(6)

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

17

where A indicates the finite impulse response of the whitening filter, z indicates the Z-transform and C is the rounded number of samples per cycle at the nominal network frequency.  f C = round  s f  fund

 ,  

(7)

where f s is the sampling frequency and f fund is the fundamental frequency; (C = 50 at 50 Hz for a sampling frequency of 2.5 kHz). This filter compares the current signal value xk to the value approximately one cycle before, x k −C . The zeroes of this filter are the harmonics of the fundamental signal, up to the Nyquist frequency, they are evenly spaced on the unit cycle [19]. This whitening filter is called the Fourier filter [18]. When the true fundamental frequency is close to f s / C , but not exactly equal to this quantity, the zeroes of the whitening filter move away from the true harmonics, however, it is still possible to force a filter zero at the estimated fundamental frequency with some more computation. This leads to the Adjusted Fourier filter [18] in case of a non-integer frequency ratio. Using the same definition of C as before, let A( z −1 ) = 1 − α z − C +1 − β z − C ,

(8)

where α and β , functions of the network frequency, are given by zC − α z − β = 0

for

z = e ± jω0TS ,

(9)

where TS is the sampling period and ω 0 is the network pulsation. We obtain the following formulae [19] for computing the coefficients α and β :

α=

sin(ω 0TS C ) , sin(ω 0TS )

β = cos(ω 0TS C ) − α cos(ω 0TS ) .

(10) (11)

Adaptive whitening filter is based on the Adjusted Fourier filter [18] with the fact that the output of the filter must be minimum when its coefficients are well adapted. A Least Mean Square (LMS) estimate of α and β is carried out to minimize the output, with the constraint of exactly filtering the dc component of the filter [19]. More detailed discussion on the adaptive whitening filter along with the associated optimal Linear Prediction Error (LPE) filter, and the adjusted Fourier filter can be found in [18] and [19].

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

18

The fault signals from the DFRs are first filtered with the adaptive whitening filter primary focus on the 50 Hz component. Hence we set the network pulsation frequency at 51 Hz, at close proximity to 50 Hz. This choice of the network pulsation frequency is optimal for this application, following several tests with different fault signals [26]. The pre-filtering operation extracts the main frequency component to be used for the signal decomposition, strongly attenuating its harmonics and the dc component. This operation increases the precision and sensitivity of the next operations, namely signal decomposition and abrupt change detection [26]. This pre-filtering operation also helps to distinguish a fault from a transient recovery, a short-term swing and the like, which are otherwise extremely difficult to estimate using only the wavelet transform [26]. This is followed by signal decomposition using the wavelet transform, threshold checking on the decomposed signal and smoothing filtering as in [25] to perform the abrupt change detection. Fig. 9 shows a result for the fault signal obtained from the DFRs in the electrical power network of ESKOM, South Africa, during a phase-to-phase fault. The fault signal is sampled at a sampling frequency of 2.5 kHz.

Fig. 9: Segmentation of RED-Phase Voltage Signal using the Adaptive Whitening Filter and the Wavelet Transform Method.

In Fig. 9, the original DFR recording for the voltage in the RED-Phase during the fault involving the RED- and BLUE-Phases is shown in plot (i), pre-filtering of the fault signal using the adaptive whitening filter in plot (ii), wavelet coefficients for the filtered fault signal (in continuous line) and the universal threshold (in dashed line) in plot (iii), and the abrupt

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

19

change time-instants computed using the threshold checking followed by smoothing filtering in plot (iv). Note that because only the highpass filter output of the QMF pair is shown for the wavelet coefficients, so the wavelet coefficients in plot (iii) indicate half of the total number of samples of the original signal [26]. Fig. 9 shows an example in which the fault signal follows a resistive decay, which does not produce the sharp transitions among the different segments caused by the different events during the fault. Since we cannot determine the different sharp segments in these cases, we rather focus on correctly estimating the time-instant of the fault-inception for further signal processing and analysis based on it [26]. In Fig. 9, this is shown in plot (iv), where the impulse indicator shows the fault-inception time-instant, thus segmenting the fault signal into two segments, pre- and post-fault, indicated by segments A and B respectively. The application of the adaptive whitening filter for pre-filtering the fault signal is of particular importance for these kinds of signals. As shown in plot (ii) of Fig. 9, the filtered fault signal clearly suppresses the resistive decay part and highlights the fault-inception time-instant, which improves the accuracy of detection of the fault-inception time-instant using the subsequent wavelet decomposition and threshold operations [26].

5.2

Adjusted Haar Wavelet Method

5.2.1 Introduction The adaptive whitening filter technique does not work well for all the special disturbance signals without distinct abrupt changes in the signal parameters. It sometimes also degrades the accuracy of the wavelet transform method for normal disturbance signals that have distinct abrupt changes. Also, for these special disturbance signals not showing distinct abrupt changes in the signal parameters, standard mother wavelets such as, Haar (Daubechies 1), Daubechies 4 [29] fail to achieve correct event-specific signal segmentations. Hence Ukil and Zivanovic propose a new adjustment technique to the standard Haar wavelet by introducing 2n adjusting zeroes in the Haar wavelet scaling filter in [28]. This technique is fairly effective in segmenting these fault signals into pre- and post-fault segments, and is an improvement on the standard mother wavelets for this application.

5.2.2 Overview of the Haar Wavelet Haar wavelet was first introduced by Alfred Haar in 1910 [30]. It has the following mathematical description. The scaling function φ ( x) is defined as

φ ( x) = 1, if x ∈ [0,1] , φ ( x) = 0, if x ∉ [0,1]. The wavelet function ψ (x) for this scaling function is defined as

(12)

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

20

ψ ( x) = 1, if x ∈ [0,0.5] , ψ ( x) = −1, if x ∈ [0.5,1] , ψ ( x) = 0, if x ∉ [0,1].

(13)

Fig. 10 shows the scaling (a) and wavelet (b) function for the Haar wavelet.

Fig. 10: Haar wavelet: (a) Scaling function and (b) Wavelet function.

5.2.3 Adjustment to the Haar Wavelet In this section, we discuss the proposed adjusted Haar wavelet in terms of the key properties of the wavelets. In general, the FIR scaling filter for the Haar wavelet is h = 0.5[1 1] , where 0.5 is the normalization factor. As an adjustment to and hence an improvement of the characteristics of the Haar wavelet, we propose to introduce 2n zeroes (n is a positive integer) in the Haar wavelet scaling filter, keeping the first and last coefficients 1 as proposed by Ukil and Zivanovic in [28]. Following the orthogonality property for the scaling filter, the filter length has to be even [31]. It is thus necessary to introduce 2n adjusting zeroes, n being the adjustment parameter [28]. The Haar wavelet corresponds to n = 0 . The introduced additional zeroes in the filter kernel have zero coefficients. The scaling filter kernel for the adjustment parameter is shown below. h = 0.5[1 1]

for n = 0

h = 0.5[1 0 0 1] for n = 1 h = 0.5[1 0 0 0 0 1] for n = 2

(14)

. . . It should to be noted that the original Haar wavelet scaling filter corresponds to n = 0 , and additional complex conjugate pairs of zeroes for each n > 0 are introduced [28]. It can be shown that the adjusting zeroes improve the Haar wavelet characteristics, especially for application in this research. Also, it has been shown mathematically in [28] that the introduction of the adjusting zeroes does not violate the key wavelet properties like compact support, orthogonality and perfect reconstruction.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

21

A theorem has been proven in [28] which states: “The introduction of the 2n adjusting zeroes to the Haar wavelet scaling filter improves the frequency characteristics of the adjusted wavelet function by an order of 2n+1.” Following the proof of the above-mentioned theorem [28], it can be shown that the adjusted wavelet function Ψn (ω ) of the adjusted Haar wavelet becomes 2 { sin ((2n + 1) ω 4 )} Ψn (ω ) =

(2n + 1) ω 4

<

4 . (2n + 1)ω

(15)

The factor 2n+1 in the denominator of (15) improves the frequency characteristics of the adjusted Haar wavelet function, by decreasing the ripples (as n > 0 ) [28].

5.2.4 Application of Adjusted Haar Wavelet After normalizing the original fault signal using its mean value, it is transformed into the smoothed and detailed version using the wavelet transform, with the adjusted Haar wavelet as the mother wavelet. Adjustment parameter n=2 is applied, that is, four adjusting zeroes are included in the adjusting Haar wavelet scaling filter [28]. The threshold method [25] is then applied to the detailed version to determine the abrupt change time-instants. This is followed by smoothing filter operations [25, 26] to indicate the change time-instants as unit impulses. Fig. 11 and 12 show the comparative results of the application of the original Haar wavelet (plot a) and the adjusted Haar wavelet (plot b) on the fault signals, sampled at a sampling frequency of 2.5 kHz [2], obtained from the DFRS of ESKOM, South Africa, during various disturbances. For these special disturbance signals, the original Haar wavelet fails to achieve correct segmentation whereas the adjusted Haar wavelet correctly segments the fault signals into pre- and post-fault segments shown as A and B respectively in the plot (b) of Fig. 11 and 12 [28]. Fig. 11 and 12 show segmentation of the voltage waveform during phase-to-ground and phase-to-phase faults respectively. The voltage waveform is shown in continuous lines and segmentations as vertical dashed unit impulses. Plot (a) shows the incorrect segmentation using the Haar wavelet as the mother wavelet, and plot (b) shows the correct segmentation using the adjusted Haar wavelet. In plot (b), the fault signal is segmented into pre- and postfault segments, indicated as A and B respectively, based on the fault inception timing. One-hundred and fifty critical disturbance records, which were unsuccessfully segmented using the Haar wavelet, were tested with the proposed adjusted Haar wavelet. The accuracy of correct segmentation, in terms of pre- and post-fault segments, was 98%.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

22

Fig. 11: Segmentation of the Voltage Waveform during a Phase-to-Ground Fault. Plot (a) shows the Result using the Haar Wavelet, while Plot (b) shows the Result using the Adjusted Haar Wavelet.

Fig. 12: Segmentation of the Voltage Waveform during a Phase-to-Phase Fault. Plot (a) shows the Result using the Haar Wavelet, while Plot (b) shows the Result using the Adjusted Haar Wavelet.

6

Complete Algorithm

The complete segmentation algorithm utilizes all the individual algorithms developed above in an optimized manner. The complete algorithm is tested and optimized using real disturbance signals, to perform segmentation for every kinds of disturbance signals originating from power systems faults. Fig. 13 shows the architectural block diagram of the complete algorithm [32]. The sequential blocks are: 1. 2. 3. 4. 5.

Disturbance Signal Read Module Signal Representation Algorithms Threshold Checking Algorithm Heuristic Smoothing Filtering Decision-Making Algorithm.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

23

Fig. 13: Top-down architecture of the complete segmentation algorithm.

Disturbance signal read module reads the COMTRADE [6] data file. The read signal is effectively represented using the different algorithms next. This is followed by threshold checking [25, 26] to perform the segmentation. Heuristic smoothing filtering [22, 25, 26] is applied to validate the segmentation. The decision making algorithm optimizes the selection of signal representation algorithms for the best possible result considering the specific disturbance signal to be processed. If segmentation is incorrect and needs to be refined, decision making algorithm chooses the proper signal representation algorithm and performs the segmentation again and until it is ok. All the algorithms are tested and validated using real disturbance signals. Table-I shows the performance of different algorithms [32]. TABLE I PERFORMANCE

OF DIFFERENT ALGORITHMS

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

7

Various Applications

7.1

Event-specific Signal Segmentation

24

The primary application of the automatic segmentation is for automatic disturbance recognition. Due to the event-specific segmentation of the disturbance signals, we can concentrate on the specific segments in the subsequent operating, e.g., feature vector construction and pattern matching, instead of processing the whole signal. This improves the accuracy, recognition rate, performance time [32]. Examples of this are shown above in Fig. 6, 7.

7.2

Synchronization

Usually many DFRs, employed for different distance protection zones, trigger for any abnormal condition in the power network. All these simultaneous recordings must be synchronized before any further global analysis to prevent any erroneous analysis. The authors discussed the synchronization technique and its application in [9]. Fig. 14 shows example of three different DFRs triggering for the same disturbance.

Fig. 14: Three Segmented but Unsynchronized DFR recordings for the same event.

The different DFR recordings for the same event, segmented by detecting the abrupt changes, are synchronized based on the fault inception timing for further analysis. The recording with the minimum fault inception time is chosen as the reference one and the rest of the recordings are synchronized with the reference recording by equating their fault inception timings with the reference fault inception timing; i.e., the rest of the recordings are left-shifted, their fault

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

25

inception time equated to the reference one. The synchronized recordings are again segmented [9]. Fig. 15 shows the synchronized, segmented BLUE-Phase voltage recordings during fault in the transmission network of South Africa, compared with the unsynchronized and segmented recordings shown in Fig. 14. In Fig. 14 and 15, for all the plots, segment A indicates the prefault section and the fault inception, segment B the fault, segment C the opening of the circuit-breaker, segment D the auto-reclosing of the circuit-breaker and system restore.

Fig. 15: Three Segmented and Synchronized DFR recordings for the same event.

7.3

Relay Performance Monitoring

Using the synchronized, segmented disturbance recordings, performance of the protective relays can be effectively monitored as discussed in [9]. Parameters like the fastest relay operating time, auto-reclosing length of the circuit-breakers can be estimated and main-1 and main-2 relay operation can be monitored.

7.3.1 Fastest Relay Operating Time To determine the fastest relay operating time, first we need to determine the fault duration. This can be done by estimating the duration of the segment B (fault) in the synchronized analog plots as shown in Fig. 15. Table II lists the fault duration times of the three DFR recordings shown in Fig. 15 in terms of number of samples and milliseconds.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

26

TABLE II FAULT DURATION OF DIFFERENT DFR RECORDINGS

For determining the fastest relay operation we need to select the minimum fault duration from the different synchronized DFR recordings. In this case, this is the DFR-3 fault duration (70.4 milliseconds) from the Table II. The formula for calculating the fastest relay operating time is the following: Fastest Relay Operating Time = Fault Duration – Trip Time.

(16)

Most of the circuit-breakers in the ESKOM transmission system are two cycle breakers [2], i.e., the expected tripping time is in the region of 40 milliseconds (50 Hz system). Using this information, we can compute the fastest relay operating time during the disturbance, which for our example case is 70.4 – 40 = 30.4 milliseconds. The fastest relay operating time gives a fair amount of idea whether the relays are operating correctly or they require maintenance.

7.3.2 Auto-Reclosing of the Circuit-Breakers From the synchronized, segmented analog signals and their matched binaries, it is possible to analyze auto-reclosing of the circuit-breakers. By comparing the signal parameter values of the segment A and D in the synchronized analog plots as shown in Fig. 15, it can be determined whether or not the auto-reclosing is successful, following the relay operation. In this case, matching the segment A and D signal parameters values of the synchronized analog plots in Fig. 15 yields that auto-reclosing of the circuit-breakers have been successful. Length of the auto-reclosing can be determined by estimating the duration of the segment C in the synchronized analog plots as shown in Fig. 15. Using the synchronized analog plots of Fig. 15, we can compute the length of auto-reclosing (segment C) for all the three DFR recordings, as listed in Table III. TABLE III AUTO-RECLOSING LENGTH OF CIRCUIT-BREAKERS

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

27

7.3.3 Main-1 and Main-2 Relay Operation Every feeder in ESKOM transmission system is equipped with two identical relays, main-1 and main-2 with the same settings for distance protection [4]. The idea behind it is to avoid the possibility of fault on the protection side when the incident occurs. Information regarding main-1 and main-2 relay operations can be obtained from the synchronized, matched binaries depending on the distance protection scheme. Analysis of main-1 and main-2 relay operation data gives the idea whether the relays employed for the distance protection are working properly or not [9]. The relay operating times can be used to obtain relationship between the fault location and the tripping speed. Using that relationship and system impedance ratio, impedance plots for the relays can be drawn, which can be used to analyze the distance protection. Distance protection at ESKOM have three zones of protection, each having a certain impedance reach with respect to the impedance of the line; Zone 1 - 80%, Zone 2 - 120% and Zone 3 - 150% of the line impedance [9].

7.4

Additional Disturbance Analysis

Analysis of certain kinds of disturbances can be performed directly from the segmented recordings before conforming to any further significant and complex feature vector analysis. These include cleared and uncleared single-phase faults, circuit-breaker restrike, reactor ring down, capacitive voltage transformer (CVT) transient behavior, energizing of a transformer [10].

7.4.1 Cleared Single Phase Fault Fig. 16 shows an example of the cleared single phase fault. In Fig. 16, segment A indicates the pre-fault section and the fault inception and segment D the auto-reclosing of the circuitbreaker and system restore. Comparison of the signal parameter values of segments A and D yields that the fault was cleared successfully [10].

Fig. 16: Segmented Analog Voltage recording during a Phase-to-Ground fault successfully cleared.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

28

7.4.2 Uncleared Single Phase Fault Fig. 17 shows an example of the uncleared single phase fault. In Fig. 17, the current during the phase-to-ground fault in the BLUE-phase is shown along with the segmentations. Re-appearance of the fault not cleared is evident from the comparison of the segments B and D, representing the fault and segment C, representing the breaker opening and reclosing [10].

Fig. 17: Segmented Analog Voltage recording during a Phase-to-Ground fault not cleared successfully.

7.4.3 Circuit-breaker Restrike In Fig. 18, we show the voltage and current recordings during a circuit-breaker restrike in the upper and lower plot respectively. Using the abrupt change detection, both the voltage and current recordings are segmented into two segments, namely, A and B, i.e., before and after the restrike of the circuit-breaker. It is to be noted that the segments of the voltage and the current are matched. A closer look at segment B of the segmented current plot also shows the spike which indicates the restriking end [10].

Fig. 18: Voltage (upper plot) and Current (lower plot) recordings during a Circuit-breaker Restrike.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

29

7.4.4 Reactor Ring Down When a line reactor is connected to a line and the circuit-breakers are opened at the both ends, the voltage does not disappear, instead, an oscillating voltage waveform can be found which slowly reduces in magnitude. This phenomenon is called reactor ring down [10]. It is a result of the interaction between the reactor and the capacitance of the line. Fig. 19 shows the segmented voltage recording during a rector ring down incident, the segmentations shown as vertical dashed lines [10].

Fig. 19: Voltage recording during a Reactor Ring Down, oscillating and slowly reducing in magnitude.

7.4.5 Capacitive Voltage Transformer Transient Behavior Capacitive Voltage Transformer (CVT) transient behavior is due to the discharge of energy stored in the capacitive and inductive elements of the CVT when there is a sudden change in the primary voltage. Fig. 20 shows a segmented voltage recording for the transient behavior of the CVT. From Fig. 20, it is to be noted that the transient behavior is reflected in the segment B. For zero or small burdens the transients are very prominent as can be seen at the start of the segment B and start of the segment C [10].

Fig. 20: Segmented Voltage recording for the transient behavior of the CVT.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

30

7.4.6 Energizing of a Transformer Energizing of a transformer often goes together with high magnetizing inrush currents. Transformer protection must be set so that the transformer does not trip for this inrush current [10]. The segmented current recording for the energizing of a transformer is shown in Fig. 21. In Fig. 11, the current signal is segmented into two segments, A and B, i.e., before and after the energizing of the transformer. A closer look at the segment B yields the fact of high magnetizing inrush currents.

Fig. 21: Segmented Current recording reflecting the Energizing of a Transformer.

7.5

Analysis of Power Signals from the Mexican Network

Ruiz-Vega, Messina and Enriquez-Harper discussed about the use of nonlinear, non-stationary analysis techniques to characterize forced inter-area oscillations problem in power systems, recorded in the Mexican interconnected system [33]. In collaboration with the Mexican team, we also tested our segmentation algorithms on the power oscillation signals obtained from the Mexican interconnected system (MZD-DGD) [32]. Fig. 22 shows the result of the test using the adjusted Haar wavelet method [28] which achieves satisfactory result.

Fig. 22: Segmentation of the power oscillation signal from Mexican power network.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

8

31

Conclusion

The analysis of faults and disturbances in power systems is a fundamental foundation for a secure and reliable electrical power supply. In the project, Automatic Disturbance Recognition in Electrical Power Systems, we focus on automated disturbance recognition and analysis for the power transmission network of South Africa, in collaboration with South African power utility, ESKOM. The purpose of this study is to augment the existing fault analysis and recognition system at National Control, ESKOM, with more robust and accurate algorithms and automated techniques in an effort to make it fully automated. This could streamline the complex task of converting massive amounts of disturbance data into the knowledge which is extremely complex, time-consuming and slow when done manually. In this tutorial, we have highlighted the urgent need for the automatic disturbance recognition in the power industry. We have also identified the problem of lack of fast, fully automated processing tools for the massive amount of disturbance records available from advanced instruments such as the digital fault recorders. Following the industrial motivation for the whole project and hence proper specification, we described the idea of an efficient, robust and advanced automated analysis system. Then we looked into the detail architecture of the proposed automated analysis system and sub-divided the overall project of Automatic Disturbance Recognition in Electrical Power Systems into four major submodules. Typically the first processing step in a recognition-oriented signal processing can be automatic segmentation of the signal. We proposed a similar approach in this work by performing the event-specific signal segmentation based on detection of abrupt changes in the disturbance signal model parameters. This is the first step towards the proposed automatic disturbance analysis and primary focus of this tutorial, followed by feature vector construction for the specific segments and applying pattern-matching algorithms to accomplish the automatic disturbance recognition and analysis tasks. We have described different abrupt change detection algorithms. The recursive identification method and the wavelet transform-based methods are mainly used. Adaptive abrupt change detection techniques are utilized for the disturbance signals not showing distinct abrupt changes in their signal parameters. The various applications explained in this tutorial show the prospective uses of abrupt change detection based automatic signal segmentation as the critical first processing step in the recognition oriented signal processing for the purpose of automatic disturbance recognition and analysis in electrical power systems.

References [1] A. Bartylak, “Application of Disturbance Recorders as Near Real Time Information Support for National Control in ESKOM,” DPSP Conference, Amsterdam, Netherlands, 2002. [2] E. Stokes-Waller, “Automated Digital Fault Recording Analysis on the Eskom Transmission System,” Southern African Conference on Power System Protection, South Africa, 1998. [3] E. Stokes-Waller, and P. Keller, “Power Network and Transmission System based on Digital Fault Records,” Southern African Conference on Power System Protection, South Africa, 1998.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

32

[4] P. Keller, C. Henze, and R. Zivanovic, “Automated Analysis of Digital Fault Records: A Utility Perspective,” SAUPEC Conference, Johannesburg, South Africa, Jan 2005. [5] R. Zivanovic, A. Ukil, and J. Jordaan, “A Novel Approach to the Automatic Analysis of Disturbance Data: Technology and Implementation,” WSEAS Conference, Tenerife Island, Spain, Dec 2005. [6] IEEE Standard Common Format for Transient Data Exchange (COMTRADE) for Power Systems, IEEE Standard C37.111-1991, Version 1.8, Feb 1991. [7] M. Basseville, and I.V. Nikoforov, Detection of Abrupt Changes–Theory and Applications, Englewood Cliffs, NJ : Prentice-Hall, 1993. [8] A. Ukil and R. Zivanovic, "Detection of Abrupt Changes in Power System Fault Analysis: A Comparative Study," SAUPEC Conference, Johannesburg, South Africa, Jan. 2005. [9] A. Ukil and R. Zivanovic, “Application of Abrupt Change Detection in Relay Performance Monitoring,” in Int. Univ. Power Engg (UPEC) Conference, Cork, Ireland, Sept 2005. [10] A. Ukil and R. Zivanovic, “Application of Abrupt Change Detection in Power Systems Disturbance Analysis and Relay Performance Monitoring,” IEEE Transactions on Power Delivery (Under review). [11] R. Zivanovic, “Analysis of Recorded Transients on 765kV Lines with Shunt Reactors,” IEEE Power Tech Conference, St. Petersburg, Russia, 2005. [12] A. Geleb, and W.E.V. Velde, Multiple-input Describing Functions and Nonlinear System Design, McGrawHill, 1968. [13] R. Zivanovic, “Nonparametric Frequency Estimation for Power System Applications,” IEEE Power Tech. Conference, Porto, Portugal, 2001. [14] J.A. Jordaan and R. Zivanovic, “Time-varying Phasor Estimation in Power Systems by Using a Non-quadratic Criterium”, Transactions of the South African Institute of Electrical Engineers (SAIEE), vol. 95, no. 1, pp. 35-41, Mar 2004, ERRATA: vol. 94, no. 3, p.171-172, Sept. 2004. [15] R. Zivanovic, “Frequency Estimation Algorithm based on Local Polynomial Approximation,” UPEC Conference, Leicester, UK, 1999. [16] V. Vapnik, The Nature of Statistical Learning Theory, New York: Springer, 1995. [17] C. Henze, “Automatic Fault Classification using Support Vector Machines,” Practical Training Report: Eskom Transmission, Germiston, South Africa, 2005 (Unpublished). [18] L. Philippot, “Parameter Estimation and Error Estimation for Line Fault Location and Distance Protection in Power Transmission Systems”, PhD dissertation, Université Libre de Bruxelles, February 1996. [19] D. Wiot, “A New Adaptive Transient Monitoring Scheme for Detection of Power System Events”, IEEE Transactions on Power Delivery, vol. 19, no. 1, 2004. [20] I.S. Caballero, C.P. Prieto, A.R.F. Vidal, “Joint Segmentation and AR Modeling of Quasistationary Signals using the EM Algorithm,” In Proc. of IEEE Workshop on Nonlinear Signal and Image Processing (NSIP’97), Mackinac Island, Michigan, 1997. [21] L. Ljung and T. Söderström, Theory and Practice of Recursive Identification, Cambridge, MA: MIT Press, 1986. [22] A. Ukil and R. Zivanovic, “The Detection of Abrupt Changes using Recursive Identification for Power System Fault Analysis,” Electric Power Systems Research, Elsevier (Under review). [23] F. Desobry and M. Davy, “Support Vector-Based Online Detection of Abrupt Changes”, IEEE ICASSP Conference, Hong Kong, China, 2003.

Notes for Tutorial 3, WCLC, 2005 Abhisek Ukil

33

[24] A.V. Oppenheim and R.W. Schafer, Discrete-Time Signal Processing, Englewood Cliffs, NJ: Prentice Hall, 1989. [25] A. Ukil and R. Zivanovic, “Abrupt Change Detection in Power System Fault Analysis using Wavelet Transform,” Int. Power Systems Transient Conference, Montréal, Canada, Jun 2005. [26] A. Ukil and R. Zivanovic, “Abrupt Change Detection in Power System Fault Analysis using Adaptive Whitening Filter and Wavelet Transform,” Electric Power Systems Research, Elsevier (To be published). [27] D. L. Donoho and I. M. Johnstone, “Ideal Spatial Adaptation by Wavelet Shrinkage,” Biometrika, vol. 81, no. 3, pp. 425-455, 1994. [28] A. Ukil and R. Zivanovic, “Adjusted Haar Wavelet for application in the Power Systems Disturbance Analysis,” EURASIP Journal of Applied Signal Processing (Under review). [29] I. Daubechies, Ten Lectures on Wavelets, Philadelphia: Society for Industrial and App. Mathematics, 1992. [30] A. Haar, “Zur Theorie der orthogonalen Funktionen-Systeme,” Math. Ann., 69: 331 – 371, 1910. [31] G. Strang and T. Nguyen, Wavelets and filter banks, Wellesley-Cambridge Press, Wellesley, MA, 1996. [32] A. Ukil and R. Zivanovic, “Automatic Signal Segmentation based on Abrupt Change Detection for Power Systems Applications,” IEEE ISCCSP Conference, Marrakech, Morocco, 2006 (Under review). [33] D.R. Vega, A.R. Messina, and G.E. Harper, “Analysis of Inter-area Oscillations via Non-linear Time Series Analysis Techniques,” in Power Systems Computation Conf., Liege, Belgium, Aug 2005.

Biography of the presenter Abhisek Ukil received the B.E. degree in electrical engineering from the Jadavpur University, Kolkata, India, in 2000 and the M.Sc. degree in electronic systems and engineering management from the University of Applied Sciences, South Westphalia, Soest, Germany, and Bolton Institute, UK, in 2004. He is currently pursuing his PhD at Tshwane University of Technology, Pretoria, South Africa, working on abrupt change detection-based signal segmentation for automatic disturbance recognition and analysis, jointly being done with Eskom.