Distribution feeder fault diagnosis classifier using SVM with PSO optimization

Distribution Feeder Fault Diagnosis Classifier Using SVM with PSO Optimization PhD student: Thi Thom Hoang Advisor: Prof.. Ming-Yuan Cho Department of Electrical Engineering, National Ka

Distribution Feeder Fault Diagnosis Classifier

Using SVM with PSO Optimization

研 究 生： 黃 氏 香 (Thi Thom Hoang)

指導教授： 卓明遠 博士 (Dr Ming-Yuan Cho)

Using SVM with PSO Optimization

研 究 生： 黃 氏 香

指導教授： 卓明遠 博士

A Dissertation Submitted to Department of Electrical Engineering National Kaohsiung University of Applied Sciences

in Partial Fulfilment of the Requirements

for the Degree of Doctor of Engineering

in Electrical Engineering

January 2018 Kaohsiung, Taiwan, Republic of China

中華民國 一零七 年 一 月

以 PSO 最佳化為基礎之配電饋線故障診斷 SVM 分類器

博士研究生: 黃氏香 指導教授: 卓明遠 博士

國立高雄應用科技大學 電機工程系博士班

博士論文

摘要

本文針對電力配電系統中的十種短路故障進行分類，我們的目標是開發一種增強型支援向量機，眾所周知支援向量機在分析非線性系統問題上是一套強而有力的工具，其特性是將原始資料轉換到更高維度以及僅需少量訓練的樣本。在基於結構風險最小化原則和統計機器學習理論相結合的理論基礎上。解決有關機器學習過程中如何選取最佳特徵和核心參數的這個目標正是許多相關研究人員致力解決的重要課題。

在此工作中，透過時域反射法 (TDR) 分析得到包括 10 種類型配電系統短路故障之人工神經網絡/支援向量機 (ANN/SVM) 分類器的數據，提取 12 個特徵作為輸入特徵進行分類。接下來，使用從 TDR 響應獲得的訓練和驗證數據組對ANN/SVM 進行訓練和驗證。然後，粒子群優化演算法首次被用來提高人工神經網絡和多層支援向量機的性能，藉由特徵和徑向基函數核參數選擇，從而診斷配電網故障。粒子群優化 (PSO) 算法用於提高分類精度，它的能力在於同時去除可能混淆 ANN /SVM 分類器及選擇最佳參數的無關聯的輸入特徵。最後，發展了PSO 的一些新的變種，包含突變 PSO，差分 PSO 和擾動 PSO，幫助粒子逃逸局部最小值以獲得更高品質之分類問題的解決方案。

透過使用 PSO 演算法，ANN 和 SVM 分類器的效能得到明顯提升兩者均超

過 93％以上。SVM 的分類結果比神經網絡的分類結果更佳。特別是使用 PSO 變

體（如 DPSO 和 PPSO）的 SVM 的效率上升到 97％以上。模擬結果顯示了我們所提出的 PSO-SVM 方法在提高分類成功率和計算速度方面的優越性。

具有高效選取最佳特徵及核心參數之 PSO 演算法的神經網絡/支援向量機分類器可被應用於配電系統的故障診斷。一些 PSO 的新型變體被開發來推動粒子使其具備較好的局部搜索能力。以 PSO 為基礎的支援向量機可以用來有效地解決饋線間負載平衡和電力系統復電等電力工程問題。

關鍵詞: 故障診斷, 粒子群優化, 配電系統, 時域反射法, 支援向量機

Distribution Feeder Fault Diagnosis Classifier Using SVM with PSO Optimization

PhD student: Thi Thom Hoang Advisor: Prof Ming-Yuan Cho

Department of Electrical Engineering, National Kaohsiung University of

Applied Sciences, Kaohsiung 80778, Taiwan

ABSTRACT

To classify ten types of short-circuit faults for an electric power distribution system, we aim to develop an enhanced support vector machine (SVM), which has been well known as a powerful tool for nonlinearly problems that have high dimensionalities with a small number of training samples It has a solid theoretical foundation based on a combination between the structural risk minimization (SRM) principle and statistical machine learning theory (SLT) Solving the problems related to select the optimal feature and kernel parameters in machine learning has been considered by many researchers

In this work, the dataset of artificial neural network (ANN)/SVM classifier including ten types of short-circuit faults in a distribution system is obtained by time domain reflectometry (TDR) analysis, and 12 features are extracted as input features for classification Next, the ANN/SVM is trained and validated using the training set and validation set that are obtained from TDR responses Then, particle swarm optimization (PSO) algorithms has been investigated to improve the performance of an ANN and a multi-layer SVM by feature and radial basis function (RBF) kernel parameter selection

in order for fault diagnosis of a distribution network for the first time The PSO algorithm is applied to increase the classification accuracy; which are capacity of

removing unrelated input features that may be confusing the ANN/SVM classifiers and selecting the optimal parameters at the same time Finally, some novel variants of PSO, including Mutant PSO, differential PSO (DPSO) and perturbed PSO (PPSO) are developed to help particle escaping the local minima in order to obtain a higher quality solution in classification problems

By using the PSO algorithm, the performance of ANN and SVM classifiers are improved significantly The success rates of both reach over 93%, respectively, in which classification results for SVM was better than those of ANN Especially, this rate increases to over 97% for SVM using variants of PSO, such as DPSO and PPSO Simulation results show the superiority of the proposed PSO-SVM approach in increasing the success rate of classification as well as computational speed

An effective PSO algorithm in optimal feature and parameter selection of ANN/SVM classifier was established for fault diagnosis in a power distribution system Some novel variants of PSO have been developed to push particles to escape from the local minima Suggestions for PSO-based SVM can be used for effectively solving electrical engineering problems, such as load balance among feeders and power system restoration in the future research

Keywords: fault diagnosis, particle swarm optimization, power distribution system, time

domain reflectometry, support vector machine

Contents

摘要 i

ABSTRACT iii

Contents v

List of Figures x

List of Abbreviations xii

Chapter 1 Introduction 1

1.1 Motivation 1

1.2 Literature review 3

1.3 Contributions of this study 6

1.4 Dissertation organization 7

Chapter 2 Distribution system operation 8

2.1 Introduction 8

2.2 Feeder automation system (FAS) 9

2.2.1 Structure and control functions of FDIR 9

2.2.2 The operation of FDIR 11

2.3 Types of fault 14

2.3.1 Single line-to-ground fault 15

2.3.2 Double line-to-ground fault 16

2.3.3 Line-line fault 17

2.3.4 Three phase-to-ground fault 18

2.4 Time Domain Reflectometry (TDR) and Pseudo Random Binary Sequence (PRBS) 20

2.5 Summary 23

Chapter 3 Support Vector Machine 24

3.1 Introduction 24

3.2 The Optimal Hyperplane 24

3.3 The optimal hyperplane for inseparable case 31

3.4 Non-linear SVM 35

3.5 Examples of SVM using the different kernel functions 43

3.6 Summary 45

Chapter 4 Particle Swarm Optimization 48

4.1 Introduction 48

4.2 Brief history 49

4.3 Concepts & Formulation 51

4.3.1 Basic Concepts 51

4.3.2 Particle Swarm Optimization in Real Number Space 52

4.3.3 Discrete Particle Swarm Optimization 58

4.4 The popular variants of PSO 60

4.5 The proposed variants of PSO 61

4.5.1 Mutant-PSO 61

4.5.2 DPSO 64

4.5.3 PPSO 66

4.6 Summary 69

Chapter 5 The PSO-based SVM 71

5.1 Introduction 71

5.2 The proposed fault diagnosis methods 73

5.3 Summary 89

Chapter 6 Simulation Results 91

6.1 Introduction 91

6.2 Testing Run for the PSO Algorithms on Benchmark Problems 94

6.3 Results of SVM classifiers using PSO algorithms 97

6.3.1 Results of PSO-based ANN/SVM classifier 98

6.3.2 Results of Mutant PSO-based SVM classifier 100

6.3.3 Results of DPSO-based SVM classifier 102

6.3.4 Results of PPSO-based SVM classifier 109

6.4 Summary 112

Chapter 7 Conclusion and Future Research 114

7.1 Conclusion 114

7.2 Future Research 116

REFERENCES 118

Publication List Since 2016 129

List of Tables

the substation 93

problems 10 independent runs 95

benchmark problems 10 independent runs 96

problems 10 independent runs 96

problems 10 independent runs 97

Optimization Techniques 105

Dataset Division Patterns 106

Dataset Division Patterns 106

technique 109

variants 110

divisions 112

List of Figures

Figure 1.1 Diagram of a two-branched distribution system using TDR 2

Figure 2.1 Comparison of feeder systems before and after automation 9

Figure 2.2 FDIR system configuration 10

Figure 2.3 A simple feeder automation system 13

Figure 2.4 A single line-to-ground fault 15

Figure 2.5 Interconnected sequence network 16

Figure 2.6 A double line-to-ground fault 17

Figure 2.7 Interconnected sequence network 17

Figure 2.8 A line-to-line fault 18

Figure 2.9 Interconnected sequence network 18

Figure 2.10 A three phase-to-ground fault 19

Figure 2.11 Interconnected sequence network 19

Figure 2.12 Approximate modelling of a typical distribution line 21

Figure 3.1 The optimal separating hyperplane 25

Figure 3.2 The mapping structure of SVM 36

Figure 3.3 A two-layer support vector machine 38

Figure 3.4 An example which not linearly separable 40

Figure 3.5 An example of training error 41

Figure 3.6 The mechanism of SVM classifier 46

Figure 4.1 Flowchart of PSO algorithm 55

Figure 4.2 Swarm topologies 56

Figure 4.6 Flowchart of the proposed DPSO 65

Figure 4.5 Proposed DPSO search mechanism of pth particle at kth iteration in a multi-dimensional search space 66

Figure 4.7 Dynamics of particle in the multi-dimensional search space in PSO algorithm 67

Figure 4.8 Flowchart of the proposed PPSO 68

Figure 5.1 Block diagram of the proposed PSO-based ANN/SVM classifier 71

Figure 5.2 The overall structure of ANN/SVM classifiers 74

Figure 5.3 Single-line diagram of a radial distribution system 74

Figure 5.4 The overall structure of the proposed PSO-based SVM classifier 76

Figure 5.5 Flowchart of the proposed PSO-SVM approach 78

Figure 5.6 The overall structure of the proposed Mutant PSO-based SVM classifier for fault diagnosis 79

Figure 5.7 The overall structure of the proposed single-stage DPSO-SVM 81

Figure 5.8 The overall structure of the proposed multiple-stage DPSO-SVM classifier for fault diagnosis 82

Figure 5.9 The overall structure of the proposed PPSO-based SVM classifier 83

Figure 5.10 Flowchart of the proposed PPSO-based SVM approach 85

Figure 6.1 A typical two-branched distribution line diagram of the sample system 91

Figure 6.2 Convergence characteristic of the proposed PSO 100

Figure 6.3 Convergence characteristic of the proposed Mutant PSO 102

Figure 6.4 Convergence characteristic of the proposed DPSO in the single-stage SVM 104

Figure 6.5 Convergence characteristic of the proposed DPSO in the multiple-stage SVM 104

Figure 6.6 Convergence characteristic of the proposed single-stage DPSO-SVM classifier for different dataset division patterns 107

Figure 6.7 Convergence characteristic of the proposed multiple-stage DPSO-SVM classifier for different dataset division patterns 108

Figure 6.8 Comparative convergence characteristics of various PSO algorithm 111

List of Abbreviations

Mutant PSO: Mutant Particle Swarm Optimization

Chapter 1 Introduction

1.1 Motivation

A distribution system is one of the most important parts of an electrical power system As a large number of connections are present in any distribution system, systems are highly susceptible to various types of electrical short-circuit faults These electrical faults not only interrupt the electrical power supply, but also have the potential

to severely damage the system Therefore, it is necessary to locate the fault and classify its type in order to guarantee the reliability of the system However, fault location methods proposed for transmission systems is not an easy task to applied for distribution networks due to multi-branch topology, unbalance operation and wide variation load [1,2] Because of multi-lateral topology, the distance obtained match to several fault locations in the distribution network and causes problems to the maintenance team It is really difficult to define the real fault location in a high spread system, delaying the restoration of the power service

Faults in distribution networks can be divided into two groups: incipient faults and permanent faults The incipient faults are as result of the aging process and incipient faulty systems are usually protected by overcurrent relays By contrast, in permanent fault detecting process, these relays guide the circuit breaker are locked, and hence the offline method is used The offline methods handle the special instruments to monitor the out of service line in the field One of the most popular instruments for permanent fault diagnosis is known as Time Domain Reflectometry (TDR) [3-5] It relies on the working principle of radar, in which a single pulse is applied to a cable and then some

of the pulse energies are reflected by any fault as given in Fig 1 Since the velocity of propagation along the line is considered constant, the time taken for the pulse to return from the impedance mismatches is a measure of distance to the source of the fault reflection interface

Unfortunately, fault diagnosis methods based on TDR is inherently imprecise due

to quickly attenuation of single pulse along the line [6] In addition, the pulse width is

Furthermore, the TDR technique is very sensitive to the corruptive interference effects

of inductively coupled link noise, which can cause the loss of reflections from long distance faults [8]

distribution transformer or load

Figure 1.1 Diagram of a two-branched distribution system using TDR

To overcome the drawback of traditional reflectometry method, a novel TDR using pseudo random binary sequence (PRBS) stimulus is proposed to identify fault types and its location However, it is quite difficult to apply any TDR approach to find faults in distribution systems because various reflected responses may occur in the reflectometry trace, resulting from various junctions and ends of branched distribution network [9] Therefore, an intelligent algorithm is required to extract fault location information on a multiple-branched network from the reflectometry trace provided With the capacity of strong robust and nonlinear mapping, ANN has been widely applied to solve fault classification problems [10,11] However, the shortcomings of over fitting and sinking into the local optimal are the major drawbacks of ANNs [12] Moreover, there is not a universal technique to specify an ANN structure with the number of hidden layers as well as the total number of neurons in those layers Compare

to the ANN, the SVM algorithm firstly developed by Vapnik and Cortes has emerged as

a powerful tool for fault classification [13] SVM roots on the basic of combining between the structural risk minimization (SRM) principle and statistical machine

learning theory (SLT) It has successfully addressed many problems of classification and regression in a wide range of applications

Some parameters in the SVM need to be optimized for every case during processing to improve generalization property of systems Some approaches have been proposed to select of feature and parameter of SVM in order to improve the efficiency and accuracy of the system These optimization algorithms have resulted in certain accuracy, but they often get trapped in local optima

In this dissertation, a novel method based upon PSO technique and its variants is developed to simultaneously optimize input features and parameters of a SVM classifier for purpose of classifying the fault types found in a distribution network These fault-types can be divided into ten classes; including single phase-to-ground faults (AG, BG, CG), line-to-line faults (AB, AC, BC), double line-to-ground faults (ABG, ACG, BCG) and three-phase short-circuit faults (ABC) The PSO-based technique automatically selects the features providing the most significant information for the classifiers, and removes the redundant features confusing to the classifiers As a result, the performance and robustness is significantly improved both classification accuracy as well as training speed Although PSO algorithm results in better accuracy than existing methods, but they often get trapped in local optima To overcome this disadvantage, some variants of PSO, including Mutant PSO, DPSO and PPSO are developed in this work, which not only significantly improve the performance of classifiers, but also avoid the trapped status in local optima

1.2 Literature review

In power systems, distribution networks deliver electrical energy from generating stations through transmission networks to consumers Electrical faults are one of the most common undesirable phenomena which may interrupt the energy supply Once an electrical fault occurs in any distribution system, immediate fault classification plays an important role in post-fault analysis and power supply restoration The accuracy of the fault type information not only assists the fault

power-as well power-as the reliability of the system [14]

A variety of approaches have been developed to build an effective fault classifier

in electric power distribution networks These studies can be divided into three separate categories, as follows: impedance based method [15-17], travelling wave based method [18-20], and artificial intelligence based method [21-23] Among them, TDR is one of the most popular methods for finding faults in distribution networks [4-6,24,25] However, fault diagnosis methods based on TDR is inherently imprecise, and hence requires the support of intelligent algorithms With the capacity of strong robust and nonlinear mapping, ANN has been widely applied to solve fault classification problems [11,12] However, the shortcomings of over fitting and sinking into the local optimal are the major drawbacks of ANN Further, there is not a universal technique to specify an ANN structure with the number of hidden layers as well as the total number of neurons

in those layers In a recent paper, an adaptive neural fuzzy inference system (ANFIS) based technique has been proposed for fault classification [26] In this method, features extracted from recorded signal are used for training of neural fuzzy inference system (NFIS); however, results show that it takes a large amount of time to optimize parameters Compare to ANN and ANFIS, SVM algorithm has been known as a powerful tool for fault classification [13] SVM roots on the basic of combining between the structural risk minimization principles (SRM) and statistical machine learning theory (SLT) It has successfully addressed many problems of classification and regression in a wide range of applications, such as face recognition [27], time series forecasting [28], fault detection [29], and the modelling of nonlinear dynamic systems [30] Especially, it is successful used for nonlinear systems with huge dimensional but a small number of training samples It gives a unique solution and is strongly regularized, thus SVM is a superior technique to solve classification problems

To build a SVM classifier, the aspect of feature subset selection plays an important role in detecting relevant variables in classification spaces An appropriate feature set can greatly increase the efficiency of classifiers Principal component analysis (PCA) [31] and multidimensional scaling (MDS) [32] are two traditional methods applied to remove redundant variables in the original feature vectors Fisher’s

linear discriminant is the most common goodness-score function applied in feature extraction The main advantage of this method is very easy to implement; moreover, there is no strict assumption in the feature distribution To increase the computational speed, some of search schemes have been proposed, including knock-out [33] and back-track tree [34] Authors in [35] proposed a Hadoop approach to extract feature in parallel; in which hundreds of mappers are composed An ANN-based technique has been developed; in which the features are removed one by one and then the training procedure of ANN is repeated [36] The paper [37] proposed a system that is able to automatically select features by using genetic algorithm (GA) In a recent paper [38], Tianan Ma and Dong Xiao Niu used the firework algorithm to select input features by removing redundant influence in order to improve the icing forecasting of high voltage transmission line

In addition to feature subset selection, the optimal set of SVM parameters also plays an important role in the distribution of samples in a given search space Chapelle and Vapnik showed that SVM is very sensitive to the trade-off regularization parameter

C and kernel function parameter such as gamma γ for the radial basis function (RBF), so

they must be selected carefully when building a fault classifier [39] It is noted that the

error penalty parameter C controls the trade-off cost between the complexity of model

and the training error Hence, set small or excessive values of C will reduce the generalization ability of SVM A SVM classifier can achieve the best generalization

capability with the best C value Also, the RBF kernel parameter γ represents the

distribution of training samples, so it determines both the generalization capability and the accuracy of classification In other words, the selection of SVM parameters plays an important role in improving the classification accuracy as well as the training speed In [40], the grid search method (GSM) has been proposed to find the optimal parameters

by attempting different values and selecting the values possessing the least testing error However, this method is both time-consuming and unable to find the best parameters

To overcome these issues, various optimization approaches have been proposed in selecting the optimum parameters, including genetic algorithm (GA) [41,42], a

an adaptive charged system search (ACSS) algorithm [44], a stochastic variable neighbourhood algorithm [45], artificial bee colony (ABC) algorithm [46] Although these optimization algorithms have resulted in better accuracy than non-optimizing methods, they often get trapped in local optima

1.3 Contributions of this study

The selection of features and kernel parameters that has been optimized by using PSO-based technique is investigated in this work Architecture is proposed to diagnose fault situation in a distribution network using two classifiers; namely, ANN and SVM with PSO-based feature and kernel parameter selection from TDR dataset After PSO algorithm is applied, it can be clearly seen that the classification accuracy on testing dataset has significantly been improved, which reaches over 93% for both classifiers Especially, this rate increases to over 97% for SVM using variants of PSO, such as DPSO and PPSO The test results show the main advantages of the proposed PSO approaches that can be used for a widely range of application in optimization problem areas

The main contributions of this dissertation are:

1 Applying TDR with PRBS stimulus to generate a reliable dataset for classifiers, in which 12 features are extracted;

2 Using the PSO for classification problem to construct an innovative hybrid system with ANN/SVM classifiers;

3 Developing variants of PSO, namely mutant particle swarm optimization (Mutant PSO), by replace the worst particle in the classical PSO to ensure particles escaping the local minima;

4 Developing a modified version of PSO, namely differential particle swarm optimization (DPSO), by adding a feature in the classical PSO to increase the classification accuracy as well as the training speed

5 Developing perturb particle swarm optimization (PPSO) variant to obtain a higher quality solution in classification problems; moreover, it can help particle

to escape from local minima

1.4 Dissertation organization

Chapter 2 presents overview of distribution system operation, in which the major

components and functionality of distribution automation system (DAS) is discussed in detail

Chapter 3 discusses about overview of support vector machine, whose theoretical

foundation is based on a combination between the structural risk minimization principle (SRM) and statistical machine learning theory (SLT) Then, those concepts are integrated with novel optimization manner

to create superior hybrid architecture for classification problems

Chapter4 reviews of an evolutionary computation technique developed by Eberhart

and Kennedy in 1995; namely particle swarm algorithm (PSO), was inspired

by the social behaviour of bird flocking and fish schooling Further, three variants of PSO are developed in the work

Chapter 5 provides the proposed methods This article proposes an innovative

approach based on particle swarm algorithms, which can improve the ability

of SVM and ANN to diagnose faults of distribution system Details about the PSO-based classifiers are discussed in this chapter

Chapter 6 then presents the simulation results based on the proposed methods and

confirms that the PSO-based SVM is an innovative technique that is very effective for the TDR data classification

Chapter 7 gives a summary of the work represented in this dissertation Main

contributions of this dissertation are also shown Further, this chapter also points the directions for future researches

Chapter 2 Distribution system operation

2.1 Introduction

Distribution networks deliver electrical energy from transmission systems to consumers, are important and integral part of all power systems Automation in the distribution field is employed to enhance efficiency, reliability, and quality of electric service

A variety of distribution automation researches have been developed to enhance the reliability and operation efficiency of distribution system The authors in [47] divide

a distribution system into three different zones to implement distribution automation The paper [48] shows the drawback of present automation techniques applied in distribution systems, and then presents challenges of implementing new technologies The commercially devices used for the purpose of distribution automation are listed in [49-51] As the technology advances, there are possible solutions to develop advanced distribution automation system The requirements and implementation of Advanced Distribution Automation (ADA) is explained in [52,53] A novel technique of using IEC

61850 at distribution automation level is proposed in [54]

The distribution automation system (DAS) can be separated into two fields: distribution substation & feeder automation and consumer location automation In this work, we only focus distribution automation system implementation at feeder to quickly determine the location of faults, isolate them, and automatically restore the healthy feeders promptly Among all functions achieved by feeder automation, Fault Detection,

FDIR is to find a fault, localize it on the feeder, open the switches around the fault, and restore un-faulted sources via the substation and alternative sources as available

This chapter describes an actual utility’s FDIR automation scheme and their component Further, the types of short-circuit fault in distribution systems and the most popular method of fault diagnosis, namely TDR are also discussed in detail in this chapter

2.2 Feeder automation system (FAS)

2.2.1 Structure and control functions of FDIR

Compare to traditional distribution system, a distribution automation system in FAS is implemented by integrating the computer master station, the communication system and the remote controlling equipments to increase the service reliability of distribution systems, as given in Fig 2.1 A fully integrated FDIR in Fig 2.2 is designed

to include a Master Station (MS) with application software, Remote Terminal Units (RTUs) in the substations, Feeder Terminal Units (FTUs), and automatic line switches along primary feeders [55] It can be noted that the master controlling station of FAS system to perform the function of FDIR can accelerate the process of service restoration for fault contingency

Switch operated manually

Switch operation controlled by software

Repair and restore on-site

Auto switch station A

Downstream feeder section restoration

Fault occurred Diagnosis of fault area Fault area isolated

Reparation and restoration

Power supply restored

Fault occurred Diagnosis of fault area Fault area isolated

Reparation and restoration

Power supply restored

Figure 2.1 Comparison of feeder systems before and after automation

SWITCH

ROUTER

SWITCH ROUTER

SWITCH ROUTER

SWITCH ROUTER

SWITCH ROUTER

TPC DDCS

Server group of

system data

Server group of application

Workstation

of dispatcher

TPS s PC and OMS server

firewall

SWITCH ROUTER

FTU: Feeder Terminal Unit (located at feeder line)

Main sub-control center

Secondary sub-control center

Simulation server

Fault detection, isolation and restoration (FDIR) application is designed for automated fault handling on distribution systems with radial or open loop configurations The distribution system can consist of feeder sections with three phases, two phases or single phase It provides the following fundamental control functions:

- Detect and isolate feeder faults

- Automatically restore service to the feeder sections upstream of the fault (Primary Restoration)

- Provide a switching sequence that can be activated by the dispatcher to automatically restore service to the feeder sections downstream of the fault (Secondary Restoration)

- Provide a switching sequence that can be activated by the dispatcher to automatically connect the de-energized feeder sections to available alternative sources if faults appear on the substation side of the feeder breaker

- Provide a switching sequence that can be activated by the dispatcher to return the distribution system to the normal configuration when previously faulted areas become available for service

2.2.2 The operation of FDIR

The feeder fault detection and isolation function of the FDIR software can be automatically initiated when a feeder breaker trip signal is received If the breaker has reclosing relays, FDIR will not start execution until it detects subsequent failures of fast reclosing of the feeder breaker to clear the fault

When the presence of a fault in the feeder is detected, FDIR will identify the fault location by logically analysing the real time data from the RTUs on the faulted feeder; automatically isolate the faulted feeder section if the related line switches are in remote control mode After ensuring the fault isolation, FDIR starts to make the primary restoration by re-energizing the feeder sections upstream of the fault It can be noted that the total time for FDIR to complete all the necessary actions from fault detection to the display of feeder re-energized data is normally less than 30 seconds

While single faults occurring simultaneously on several different feeders, they can be handled in a parallel manner and each one can be completed in a few seconds The FDIR can process with the number of single faults on different feeders occurring simultaneously at the same time

FDIR will begin to perform secondary restoration after completion of the fault isolations and primary restorations for all the faulted feeders with no new faults occurring within a dispatcher specified time interval It will accomplish this by determining a switching sequence through reconfiguring of the associated and available tie switches to re-energize the feeder sections downstream of the faults The switching sequence is developed with the objective of restoring as many of the de-energized feeder sections as possible with minimum number of switching actions within the allowed overload and voltage drop limits (either normal limits or emergency limits can

be applied) of the impacted feeders and energy sources Alarm messages will be provided if there exist non restorable feeder sections due to topology constraints or overload and voltage drop limits The recommended switching sequence will not cause service interruption on other feeder sections whenever possible The recommended switching sequence can be displayed in a tabular list along with their impacts on feeder loads The resultant configuration of the impacted feeders, if the switching sequence would be implemented, can be graphically displayed with blinking symbols for the dispatcher to review

Upon review, the dispatcher is able to reject or accept the sequence or modify one

or more steps of the recommended switching procedures before it is initiated for FDIR

to implement the switching procedures in either sequential or automatic or manual step

by step mode The dispatcher would also be able to abort and reverse the switching sequence at any time during or after the execution of the switching sequence The implementation of switching sequence can also be automatically aborted, if desired, when a new feeder fault occurs during the execution

The recommended switching sequence can be developed in a few seconds, depending upon the numbers of involved tie switches and feeder sections related to the faults during the restoration A recommended switching sequence will be automatically

cancelled if it is not implemented in a dispatcher specified time interval In this case, a pending flag will be set for each non restored feeder section such that a new switching plan can be developed for these sections by FDIR at any time under dispatcher's request FDIR also provides the capability to return the network to its pre-fault configuration under the dispatcher's request when the previously faulted feeders are available for service

Based on the dispatcher's requirement, FDIR will develop a recommended switching sequence to reconfigure the related feeder sections back to the default configuration This switching sequence can be carried out automatically or in semi-automatic mode requiring operator approval at each step

Consider a simple feeder automation system, given as Fig 2.3:

Figure 2.3 A simple feeder automation system

MTR: main transformer

LBS：Load Break Switch for Main Circuit Breaker of underground 4-Way

CB：Circuit Breaker for Branch Circuit of underground 4-Way Switch

NO：Normal Open Switch

Assume a fault occurs in Zone 2, the recloser completes its three operation and

locks out All four zones are de‐ energized The computer system at the control center starts to execute FDIR function immediately：

from the circuit

Feeder closes if capacity is available Zone 4 is re-energized

such, Zones 1, 3 and 4 are restored It can be noted that the time to energize Zones 1, 3, and 4 is less than 60 seconds and this process is implemented without human support

corrective action if necessary Checks for subsequent faults on restored zones

Logic automatically returns the circuit to normal configuration

As such, circuit is back to normal configuration

2.3 Types of fault

A distribution system is one of the most important parts of an electrical power system As a large number of connections are present in any distribution system, systems are highly susceptible to various types of electrical short-circuit faults These electrical faults not only interrupt the electrical power supply, but also have the potential

to severely damage the system Once an electrical fault occurs in any distribution systems, immediate fault classification plays an important role in post-fault analysis and power supply restoration The accuracy of the fault type information not only assists the fault diagnosis system to locate the electrical faults promptly but also to ensure power quality as well as reliability of the system

For a three phase line, short circuit faults are classified into four categories [56]:

1 Single line-to-ground fault

2 Line-to-line fault

3 Double line-to-ground fault

4 Balance 3 phase-to-ground fault

2.3.1 Single line-to-ground fault

For single line-to-ground fault, only one phase has non-zero fault current The following three types of single-phase-to-ground faults are experienced:

1 Phase A-to-ground fault

2 Phase B-to-ground fault

3 Phase C-to-ground fault

Consider a single line-to-ground fault from phase A to ground at the general three-phase bus shown in Fig 2.4 Fig 2.5 shows the sequences networks interconnection diagram

f a

2.3.2 Double line-to-ground fault

Double line-to-ground fault occurs when two line conductors come in contact both with each other and ground Two-phase-to-ground faults are of the following three types

1 Phase B and phase C-to-ground faults

2 Phase C and phase A-to-ground faults

3 Phase A and phase B-to-ground faults

Consider a line-to-line fault from phase B to C, shown in Fig 2.6 Fig 2.7 shows the sequences networks interconnection diagram

For double line-to-ground fault, the positive-sequence fault current is:

1 Phase B-to-phase C fault

2 Phase C-to-phase A fault

3 Phase A-to-phase B fault

impedance Zf to ground is shown in Fig 2.8 Fig 2.9 shows the sequences networks interconnection diagram

ABCF

Ia=0

ZfFigure 2.8 A line-to-line fault

F 2

Z 2

N 2

V a2 +

Three phase-to-ground faults are the least probable faults but yet the most severe

A general representation of a balanced three-phase fault is shown in Fig 2.10 where F is

interconnection diagram

A B C F

to ground fault (ABG, ACG, BCG) and three phase-to-ground faults (ABC)

2.4 Time Domain Reflectometry (TDR) and Pseudo Random Binary Sequence (PRBS)

As mentioned above, it is essential to monitor distribution system carefully and constantly, and fault diagnosis must be carried out on a timely way so that alarms can be given as early as possible Faults in distribution networks can be divided into two groups: incipient faults and permanent faults The incipient faults are as result of the aging process and incipient faulty systems are usually protected by overcurrent relays

By contrast, as detecting short circuit fault that the most popular type of permanent fault, these relays guide the circuit breaker are locked, and hence the offline method is used The offline methods handle the special instruments to monitor the out of service line in the field

TDR is one of the most common offline methods used for fault classification and location It relies on the working principle of radar; in which a single pulse is applied to

a cable and then some of the pulse energies are reflected by any impedance mismatches These impedance mismatches can be the fault, tee joint or the line terminal Therefore, these obtained echo responses are used for identifying the fault nature

Let us assume an enclosed coaxial distribution line can be modelled by an equivalent circuit, as shown in Fig 2.12

The voltage and current traveling along the line can be expressed as:

In eqns (2.13) and (2.14), v(x,t) and i(x,t) are the forward travelling voltage and

current waves respectively

Figure 2.12 Approximate modelling of a typical distribution line

Using the Laplace transform and differential equation, we can obtain:

x+∆x C.∆x

x

to the impact of noise In addition, the pulse width is one of the factors that effects on the accuracy of the reflectometry method Using narrow pulse is the ideal measurement because of easily increasing to sharp traces; however, it is quickly attenuated on signal pathway resulting from the relationship between pulse width and bandwidth Alternatively, a wide pulse stimulus can propagate further due to not so quickly attenuated and hence it is useful for locating longer fault distance, otherwise, this wide pulse stimulus produces more rounded trace features that are not easy to discern and lead to incorrect fault location Furthermore, the TDR technique is very sensitive to the corruptive interference effects of inductively coupled link noise that causes the loss of reflections from faults occurring at long distance

Thus, a PRBS perturbation is used to identify fault and its location to overcome the drawback of traditional reflectometry method For this purpose, instead of using a single pulse as in the traditional TDR, a binary sequence whose logic state changes

between value of 0 and 1, corresponding to voltage levels –V and +V is used, namely

PRBS The PRBS is applied in a large range of fields, such as encoder and transducer testing [57], frequency response testing[58], communication area [59,60], ADC

converter testing [61], etc

PRBS is constructed by a linear shift feedback register (LSFR) Each register includes n stages, in which each stage can hold a bit The bits are shifted to the right

after every clock period Therefore, the maximum length of PN sequence is L=2n-1,

equal to 2n-1 bit

To reduce the impact of coupled noise and the accentuation of fault signature, the cross-correlation of the reflected response with the incident excitation can be carried out over an integral number of PRBS sequences as follows:

1 ( )

is the reflected waves

The cross-correlation of the reflected fault response with the incident PRBS is This TDR with PRBS stimulus; however, is only applied for fault diagnosis on transmission line It is not easy to use TDR for detecting fault in distribution network because of muti-branch topology In this work, CCR along with the reflected signals obtained from TDR instrument wave are used for SVM training phase for fault

diagnosis in distribution network

2.5 Summary

In power systems, distribution networks deliver electrical energy from generating stations through transmission networks to consumers Condition monitoring

power-of distribution system is gaining important owing to the need to increase reliability and

to decrease possible loss of production because of power breakdown by accident This chapter introduces the DAS at feeder level The DAS implementation reduces outage duration in order to increase the reliability as well as quality of power supply Moreover, DAS reduces human involvement and better manages of system and component loading

In addition, this chapter describes ten types, which need to be classified Furthermore, a brief concept of TDR using PRBS is also included in this chapter

Chapter 3 Support Vector Machine

3.1 Introduction

SVM [62,63] was first mentioned by Vapnik in 1995, and it has become one of the most optimal techniques for data classification It has a solid theoretical foundation based on a combination between SRM and SLT SVM has been known as a powerful tool for classification and regression problems in a wide range of applications, such as face recognition, fault detection, nonlinear equalization and text categorization [64-69] Especially, it is successful used for nonlinear systems with huge dimensional but a small number of training samples, so called support vectors SVM maximizes the boundary margin by using a separating hyperplane in order to produce the global optimization and high generalization ability [70] For fault classification problem, various approaches have been investigated With good learning ability, SVM overcomes over-fitting problems and provides sparse solutions in comparison to existing methods such as artificial neuron network (ANN) and decision trees (DT) in fault classification [71]

SVM maps input vectors into a very high-dimensional feature space and builds an optimal hyperplane to separate samples from two classes For non-separation data problems, SVM uses two methods Firstly, SVM is employed with a soft margin hyperplane and a penalty function of training errors Secondly, a linear separating hyperplane is constructed in the high dimensional kernel space, which is transformed from the original feature space The concepts of SVM are discussed in the following sections (Part of this chapter is modified from [62,63,66])

3.2 The Optimal Hyperplane

In standard linear classification problem, for example, one should separate the set

label vectors Support vector learning finds a separating hyperplane (x∙ϕ) = c that separates the positive subset I (y = 1) from the negative subset II (y= -1) with the largest

margin

Assume all the data satisfy the constraints:

where (x∙y) is the inner product between vectors x and y.

For each vector ϕ, two values of constant c are defined:

positive subset I from the negative subset II with the largest margin This hyperplane is called as name “the maximal margin hyperplane” or “the optimal hyperplane” (Fig.3.1)

x

x x

x x

H1 H H2

Figure 3.1 The optimal separating hyperplane