- Email: [email protected]
Non-intrusive fault identification of power distribution systems in intelligent buildings based on power-spectrum-based wavelet transform
Accepted Manuscript Title: Non-intrusive fault identification of power distribution systems in intelligent buildings based on power-spectrum-based wavelet transform Author: Hsueh-Hsien Chang PII: DOI: Reference:
S0378-7788(16)30538-2 http://dx.doi.org/doi:10.1016/j.enbuild.2016.06.050 ENB 6785
To appear in:
ENB
Received date: Revised date: Accepted date:
11-11-2015 27-3-2016 16-6-2016
Please cite this article as: Hsueh-Hsien Chang, Non-intrusive fault identification of power distribution systems in intelligent buildings based on power-spectrum-based wavelet transform, Energy and Buildings http://dx.doi.org/10.1016/j.enbuild.2016.06.050 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Non-intrusive fault identification of power distribution systems in intelligent buildings based on power-spectrum-based wavelet transform Hsueh-Hsien Chang Dept. of Electronic Engineering, Jin-Wen University of Science and Technology, New Taipei, Taiwan [email protected]
Highlights
• Non-intrusive monitoring techniques are proposed in real-time load-bus faults and transmission-line faults detection. • A power-spectrum-based wavelet transform and ANNs are proposed. • Parseval’s Theorem is adopted to reduce the number of WTCs representing fault transients. • The proposed method improves significantly the performances of the distribution system fault detection in intelligent buildings. • The proposed method is scarcely influenced to the fault inception angles, fault resistances, and system voltage variations.
Abstract A new approach for protection of power distribution systems in intelligent buildings has been presented in this paper. Directly adopting the wavelet transform coefficients (WTCs) requires longer computation time and larger memory requirements for the non-intrusive fault monitoring (NIFM) identification process. However, the WTCs contain plenty of information needed for the symmetric and asymmetric transient signals of fault events. To effectively reduce the number of WTCs representing fault transient signals without degrading performance, a power spectrum of the WTCs in different scales calculated by Parseval’s Theorem is proposed in this paper. In this paper, artificial neural networks (ANNs), in combination with power-spectrum-based wavelet transform, are used to identify fault types and locations in power distribution systems of industrial buildings by using NIFM. The high success rates of fault event recognition for load-bus 1
faults and transmission-line faults from simulations have proved that the proposed algorithm is applicable to fault identifications of non-intrusive monitoring applications.
Keywords: Artificial neural networks (ANNs); Parseval’s Theorem; wavelet transform; non-intrusive fault monitoring (NIFM); intelligent buildings.
Acronyms ANFIS:
Adaptive Neuro-Fuzzy Inference System
ANN:
Artificial Neural Network
BP-ANN:
Back-Propagation Artificial Neural Network
DLGF:
Double Line-to-Ground Fault
DWT:
Discrete Wavelet Transform
EMTP:
Electromagnetic Transient Program
ESE:
Electrical Service Entry
FTED:
Fault Transient Event Detection
LLF:
Line-to-Line Fault
LLLGF:
Three-Line-to-Ground Fault
MDMS:
Meter Database Management System
MFNN:
Multi-layer Feed-forward Neural Network
NIFM:
Non-Intrusive Fault Monitoring
NILM:
Non-Intrusive Load Monitoring
PSO:
Particle Swarm Optimization
PSs:
Power Signatures
SCADA:
Supervisory Control and Data Acquisition
SE:
Spectral Envelope
2
SLGF:
Single Line-to-Ground Fault
STFT:
Short-Time Fourier Transform
UT:
Turn-on Transient Energy Algorithm
WFNN:
Wavelet Fuzzy Neural Network
WMRA:
Wavelet Multi-Resolution Analysis
WPT:
Wavelet Packet Transform
WTCs:
Wavelet Transform Coefficients
1. Introduction In protection analyses of power system, power system fault is one of the most important research topics. These faults which are mainly due to short circuit phenomena can drastically fail the operations of power systems and cause excessively high currents to flow which causes damages to devices and fire accidents to buildings. Types of faults, like a short circuit condition in a power system distribution network, will result in severe economic losses and reduce the reliability of the electrical system. Different types of transient phenomena occur on the transmission line. From these transient phenomena, faults on transmission lines and load buses need to be detected, located, classified accurately, and cleared as fast as possible. In earth fault and short-circuit protections, faulty phase identification and location of fault are the two most important items which need to be addressed in a reliable and accurate manner. In multiple phases system, the faults are basically divided into symmetric (balanced) and asymmetric (un-balanced) types. The type of symmetric fault is only a balanced three-line-to-ground fault (LLLGF). Different types of asymmetric faults are single line-to-ground fault (SLGF), line-to-line fault (LLF), and double line-to-ground fault (DLGF). 1.1. Research background Due to the serious consequences of fault damages in power systems, many researchers
3
have devoted their efforts to this area and reported their findings in articles. Chunju et al. [1] have proposed a SLGF location method for employing wavelet fuzzy neural network (WFNN) to extract fault characteristics from the fault signals in an industrial distribution power system. However, this method cannot significantly increase the fault location accuracy because it does not use the high frequency information of pre-fault current and voltage. In fact, when fault location changes, the equivalent capacitance will change, the charging and discharging currents will be also changed, namely the high frequency currents are different. The high frequency currents and voltages are related with the distance from the relaying point to the fault location. Borghetti et al. [2] have built specific mother wavelets inferred from the recorded fault-originated voltage transient waveforms to improve the wavelet analysis. In this paper, the authors assumed that the network topology and the traveling wave speeds of the various propagation modes are known. However, as concluded in [2], this method is expected to improve the algorithm accuracy by means of proper integration of time-domain fault location approaches. Reference [3] has used wavelet energy and entropy criterion of the wavelet packet transform (WPT) coefficients for every faulty current and voltage signal to extract features and reduce the size of data sets for training and testing of artificial neural networks (ANNs). However, this method only implements in SLGF of power systems. Reddy et al. [4] have performed well to use a wavelet multi-resolution analysis (WMRA) technique to extract the features of the transient current signals and employed computational intelligence techniques as an adaptive neuro-fuzzy inference system (ANFIS) and ANN in conjunction with GPS to identify fault location. This paper focuses on amplitudes of the second- and third-order harmonics generated during fault current occurrence to track the fault location. Therefore, among different coefficients pertaining to different decomposition levels, only the summation of the fifth-level detailed
4
coefficients (d5) is considered for the sampling rate of 6 kHz. Bezerra Costa [5] has presented a wavelet-based methodology for real-time detection of fault-induced transients in transmission lines, where the wavelet coefficient energy takes into account the border effects of the sliding windows. As a result, the performance of the proposed energy analysis is not affected by the choice of the mother wavelet and presenting no time delay in real-time fault detection. The concept of non-intrusive load monitoring (NILM) was originally introduced at MIT by Hart [6] in late 80’s. The NILM, in contrast with the traditional supervisory control and data acquisition (SCADA)-based system, can drastically reduce hardware and maintenance costs because only one set of voltage and current sensors is needed at the electrical service entry (ESE) [6]. Although several NILM algorithms were developed during the past two decades, recognition accuracy and computational efficiency have been remaining as challenges. Hart proposed a method for disaggregating electrical loads by examining only the appliance specific power consumption signatures within the aggregated load data [7]. However, the voltage variations of the utility may result in overlapping of loads in the P-Q plane [8], different loads may consume the same real and reactive powers or they are with the simultaneous start and power signatures (PSs) that have non-discrete changes in power consumption may not be adequately measured only from steady-state parameters [6], [9]. In addition to the steady-state power draw, Leeb et al. have used the “spectral envelope” (SE) concept [10] for a NILM system. The SE is a vector of the first several coefficients of the short-time Fourier transform (STFT) of the transient current signal. Although this method can detect numerous appliances including the variable loads, it requires extensive training for each appliance before classification and monitoring can be performed. Yang et al. [11] employed a turn-on transient energy algorithm (UT) in a non-intrusive monitoring of industrial electrical loads. However, this
5
algorithm requires a high sampling rate to accurately capture the detailed variations in the transient energy. Recently, there has been a growing interest in improving the performance of recognition for using artificial intelligence methods such as ANNs. Srinivasan et al. [12] proposed a NN-based approach to identify harmonic sources. However, the method does not incorporate the various operation modes in each load. In [13], the authors have applied particle swarm optimization (PSO) to optimize the parameters of training algorithms in NN for improving the recognition accuracy. Nevertheless, the resulting training time could be quite long. 1.2. Problems and contributions However, as far as the author is well aware that fault identification for the distribution system of an industrial building by using non-intrusive monitoring techniques is still in the primary stage of development. Based on the concept of the NILM, this paper proposes new fault identification approaches for load-bus faults and transmission-line faults on low-voltage power distribution systems in intelligent buildings by using non-intrusive monitoring techniques. The non-intrusive fault monitoring (NIFM) refers to a method of detecting the current waveforms of a distribution system using a single set of sensors at ESE. Fig. 1 shows a typical NIFM system on power distribution systems in an industrial building. Three-phase three-wire and three-phase four-wire systems are generally used to run a typical power system distribution network. As shown in the figure, the secondary side of the distribution transformer, the voltage and current at ESE are measured and sent to the meter database management system (MDMS). This data is processed by the NIFM algorithms to identify the operation statuses and fault events for the individual feeder. As in the existing methods mentioned above, the wavelet transform (WT) technology has been often used to capture the time of transient occurrence [14]. Integrating the WT technology with the artificial intelligence method or expert system becomes a promising
6
approach to improve the recognition accuracy of fault events. However, the following issues must be overcome before their practicality can be realized for a non-intrusive classifier in the NIFM system [15]. 1) Directly adopting the WT coefficients as inputs of the NNs requires large memory space and long training time; 2) The decomposition level with the number of extraction features must be reduced to enhance computation efficiency of the NIFM system; 3) Not only can symmetric and asymmetric fault events be identified but load-bus faults and transmission-line faults can also be detected. 4) The reliability of the electrical system cannot be reduced even only one set of voltage and current sensors is needed at ESE. In an effort to overcome some of the above-mentioned problems, a new NIFM system is proposed in this paper. This paper presents a novel classifier consisting of two models as shown in Fig. 2. First, the WMRA technique and Parseval’s Theorem are utilized to extract the energy distribution features of the fault transient signal from Electromagnetic Transient Program (EMTP) simulation and ESE at different resolution levels. Second, the back-propagation artificial neural network (BP-ANN) is employed to classify fault locations and types according to the detailed energy distribution. By using Parseval’s Theorem, the number of transient power features can be reduced without losing their fidelity. The accuracy rate and the computation requirement of the proposed method were tested in a simulated model system by using EMTP software. These results showed that the proposed methods can efficiently and effectively analyze the signals, thus enhance the performance of the fault identifications in the NIFM system to compare with other extraction methods.
7
2.
The proposed NIFM system in intelligent buildings The block diagram in Fig. 2 includes the algorithms of fault transient event detection
(FTED), the processes of the WT technique, the feature extractions from symmetric and asymmetric fault transient signals, and the BP-ANN classification network. The procedures of recognizing fault transient signals belonging to certain kinds of the fault events are described as follows. First, a fault occurrence of transient event is detected by the algorithm of FTED.
Then, a complete fault event signal can be acquired by
monitoring at ESE, i.e. the steady-state signal ahead and the transient signal back from the fault inception. The current waveforms of the complete fault transient signals are transferred using WT to obtain the different scales of the wavelet transform coefficients (WTCs). Although all scales of the WTCs of a complete fault transient signal may contain some important features of the faults, the identification system would require many memory spaces if all the WTCs are adopted directly as the features. As shown in the Fig. 2, Parseval’s Theorem is therefore employed to extract the power spectrum of the WTCs as the features of non-intrusive fault identification system. To equip the identification system with the ability to learn any arbitrary mapping of network inputs to desired outputs, the BP-ANN which has been trained and tested from the power spectrum inputs of EMTP simulation and previous fault waveforms measured at ESE is utilized in this paper. 3.
Proposed methods
3.1. WT Morlet and Grossman [16] used the WT to manage the extended training conditions for applying STFT. By applying the band-pass filters’ special properties, it is proposed that the WT can be used to detect and locate the transience phenomena with high-frequency components of power signals with different scales of the WTCs in NIFM systems.
8
ψa,b (t )
t b ψ , a R , b R a a
1
(1)
where ψ a, b t is the daughter wavelet; ψ t is the mother wavelet; a is the scaling factor; and b is the shift factor. Principally, decompositions of signals as a series of daughter wavelets to be analyzed are achieved by using scaling and shift factors of discrete WT (DWT). Thus, it is proper for transient analyses to be carried out [17]. As statements for the DWT, owning to different frequency components, the WTCs scales have frequently been used in determining the fault transient occurrences by focusing on the high-frequency components on power systems. These high-frequency components are often the results of the signals coming from the symmetric and asymmetric transients. Daubechies is used as the mother wavelet for the WT in this research. 3.2. WTCs’ power spectrum In order to depict the unitary characteristics of any Fourier transform in physics and engineering, the Parseval’s Theorem is applied. By utilizing a Fourier series, a periodic signal f (t) is decomposed into the sum of a set of harmonic oscillators, and can be derived as follows: f (t )
1 a0 2
an cos(
n 1
nπt nπt bn sin( ) ) T T n 1
1 T a0 f ( t ) dt T T T nπt a 1 ) dt f ( t ) cos( n T T T T 1 nπt ) dt f ( t ) sin( bn T T T
(2)
(3)
where T is the time of a cycle. From the energy perspective, (2) can be formulated to become (4). 9
1 T
T
T
1 f ( t ) dt a 02 2 2
(a
2 n
bn2 )
(4)
n 1
Suppose that the scaling function and the mother wavelet are orthogonal basis sets, the relationship between the power spectrum of the discrete signal f [n] and each of the WTCs can be set up by Parseval’s Theorem. The WT can isolate the signal energy in time and frequency domains. According to the MRA characteristics of the WT, a complete fault transient signal can be depicted by each scale of the WTCs.
WTCs_ f [n] [ CJ ,k d1,k d2,k ... dJ ,k ]
(5)
where J is a number of the WT scale; CJ ,k and d j,k are the coefficients of approximate and detailed versions of a complete fault transient signal, respectively. Hence, owning to the DWT decomposition, the energy of the complete fault transient signal is formulated in (6). The first term on the right side of (6) describes the average power of the approximated version of the decomposed signal, and the second term describes the sum of the average power of the detailed version of the decomposed signal. 1 N
f [n ]
2
nN
1 NJ
k
C J ,k
2
1 N j j 1 J
k
d
2 j ,k
(6)
To filter out the noise, by subtracting the standard deviation of the absolute WTCs from the WTCs, the d j , k coefficients are modified into a new one ( d j , k ). As mentioned above, the high-frequency components of the second term in (6), will be used to extract the features of the fault event in NIFM systems, and formulated as Psignal in (7). The power spectrum of a fault event signal described by Parseval’s Theorem and the WTCs, the power spectrum of Psignal of each WTC, can thus be established as shown in (8). 10
Psignal
1 N j
d j , k
k
Psp ectru m Psig n a l
1 2
2
d j Nj
d j Nj
2
2
(7) 1
2
(8)
In the end, as (8) indicates, the energy of the event signal can be separated at different resolution levels in different processes dependent on the sampling frequency. For instance, the sampling frequency of 15 kHz, the decomposition levels can be 1 to 8 in order to examine the coefficient of the detailed version at each resolution level for extracting the fault signal features inherent in the NIFM system. Besides, the following confirmed important properties from EMTP simulations and experiments resulted from (8) for recognition features are described below: 1) The energy distribution remains the same at different timing or inception angle of fault event occurrence; 2) The outline of the energy distribution remains similar to variations in the different system voltage amplitude and fault resistances of the same fault type; 3) The low-level energy distribution will denote obvious changes when there is high-frequency components appearing in the event signal. For instance, Fig. 3(a) indicates a complete transient current waveform with three different occurrences (0.05s, 0.1s, and 0.15s) of a SLGF on phase A at BUS 1 as in Fig. 1. Investigations on the three signals show that the detailed energy distributions (scales 1~ scale 8) of the given signals are identical to those shown in Fig. 3(b). When the different current amplitudes (-20%, 0%, and +20%) of the fault event appear as those indicated in Fig. 4(a), the detailed version of energy distributions (scale 1~scale 8) is similar to that shown in Fig. 4(b). Moreover, when the different fault resistances (0pu, 0.002pu, and
11
0.02pu) appear as those indicated in Fig. 5(a), the detailed version of energy distributions (scale 1~scale 8) is also similar to that shown in Fig. 5(b). As a consequence, the fault detection is scarcely influenced to the fault inception angles, fault resistances, and system voltage variations for using the proposed methods. Accordingly, by using differences of the detailed version of energy distribution, the features of fault signals can demarcate different fault locations. As shown in Fig. 1 and Fig. 6, five different loads on different buses labeled as bus 1 to bus 5 are one 45-hp induction motor, one 55-hp induction motor, one 30-hp induction motor, one 20-hp induction motor driven by variable-voltage drives, and one load bank supplied by a six-pulse thyristor rectifier from ac power, respectively. The power spectrums in Fig. 6(a) and 6(b) depict that the five different buses have different energy distributions from scale 1 to scale 8 for SLGF transient voltage and current waveforms, respectively. Consequently, in terms of the small-scale power spectrum (i.e. in the high-frequency domain); different fault buses can be classified using particular energy spectra as the features in the NIFM system. Evidently, it is better to select the fault transient current waveform than the fault transient voltage waveform for fault identification of power distribution systems. 3.3. ANNs There is a wide range of power system applications of neural networks in operation and control processes, including stability assessment, security monitoring, load forecasting, state estimation, load flow analysis, contingency analysis, emergency control actions, etc. [18]. 1) Training algorithms and fitness function Over-fitting is a typical difficulty occurring during ANN trainings. Bayesian regularization typically provides better generalization performance than early stopping
12
due to its not requiring a validation dataset to be segregated from the training dataset. In other words, all training datasets are required for the input training data. This benefit is particularly critical when the available dataset is small. The representative fitness function applied in training a feedforward neural network is the mean sum of squares of network errors, alternatively called MSE. It is the average squared error between network outputs and desired outputs.
MSE(l )
1 (t j (l ) a j (l ))2 2 jC
(9)
where the variable t j (l) is the desired output for the lth iteration at node j , and variable
aj (l) is the network output for the lth iteration at node j. 2) Multi-layer feed-forward neural network Most BP-ANN applications apply single- or multilayer perceptron networks by using gradient-descent training methods combined with learning through backward propagation. These multi-layer perceptrons can be trained under supervision using analytical functions to activate network nodes (“neurons”); besides, by applying a backward error-propagation algorithm to update interconnecting weights and thresholds, a sufficient recognition capability can later be achieved. A trainable classifier applies the BP classifier in a multi-layer feed-forward neural network (MFNN) in this study. “Classification” in this context denotes a mapping from a feature space to a set of class labels, which are the names of fault points. Generally speaking, a supervised MFNN is partitioned into three layers: input, hidden, and output layers. These neurons are connected by linked weights selected to meet the desired relationship between input and output neurons. The MFNN in this study is employed to identify faults in the NIFM system. The input, output, and hidden layers of the BP-ANN are described as follows: 13
a) Input layer: PSs information, the energy spectra of a complete fault event signal, for an ESE or EMTP simulation serves as inputs. b) Output layer: the number of output neurons is the same as that of individual fault feeder identified. Each binary bit serves as a load-bus or transmission-line indicator for fault status. c) Hidden layer: two hidden layers are used in this study. Some heuristics that have been developed can determine the number of neurons in a hidden layer [19]. The common number of neurons in a hidden layer is the sum of the number of neurons in an input layer and that in an output layer. 4.
Experimental results
4.1. Study environments Experimental data sets were generated by preprocessing the data on the current waveforms of the total feeders. Each final sample consists of (T60256) samples obtained over a period of T. Each example of the power feature includes a voltage variation from − 3% to +3% at 0.5% intervals, yielding thirteen examples of power features for each bus or feeder and (N13) raw data for each power signature ( I a , I b , I c ) , given that N load buses can only be a short circuit for one bus or line at any time in three-phase electric power distribution systems. The full input dataset (N13)(T 60256)
matrix includes the training dataset and the test dataset. The training data and test data are randomly selected from all data. To confirm the inferential power of the neural networks, the full raw data set creates, respectively, a (N13)/ 2 matrix as the training data set and test dataset. A simulated NIFM system constructed by the EMTP software program monitors the current waveforms in a three-phase ESE powering a collection of loads representative of the five load classes in an industrial building. As shown in Fig. 1 for load-bus faults and 14
Fig. 7 for transmission-line faults, five different loads on different buses include a 45-hp induction motor, a 55-hp induction motor, a 30-hp induction motor, a 20-hp induction motor driven by variable-voltage drives, and a bank of loads supplied by a six-pulse thyristor rectifier for ac power. A neural network simulation program was designed using MATLAB. The program was run to identify faults on a PC equipped with a 3.10 GHz Intel Core i5- 4440 CPU. Each entry in the BP-ANN represents 10 different trials, and each trial uses random initial weights. In each case, the network is trained until the MSE is less than 0.00001 or the number of iterations reaches 5000. The proposed method is able to identify the fault locations and types instantly. The results of four different fault types for load-bus faults and transmission-line faults show the feasibility and superiority of the proposed methods compared with other extraction methods such as standardized statistical moments based WT (Z-score based WT) [20] and turn-on transient energy algorithm (UT) [21]. 4.2. Results 1)Case Study 1, Load-Bus Faults: In case study 1, the neural network algorithm in the NIFM system identifies SLGF on phase A, DLGF and LLF on phase A and B, and symmetric fault (LLLGF) at five different 220-V load buses based upon the proposed power spectrum method (Pspectra), as shown in Fig. 1. There are basically five 220V distributed lines in the model system of Fig. 1 because the distance is short between the 480V/220V distribution transformer and the load bus. The lines are modeled as lumped parameter sections in EMTP program. Table 1 shows that values for the training accuracy of fault identification are 100% for all current signatures with the proposed Pspectra. Furthermore, the test accuracy of fault identification is higher than 97.82% in current signature Ib of LLF. The values for the training accuracy of fault identification are also 100% for all current signatures with the
15
Z-score based WT. Moreover, the test accuracy of fault identification is higher than 18.44% in current signature Ib of LLF. Finally, the values for the training accuracy of fault identification are higher than 94.55% in phase B of DLGF with the UT. Besides, the test accuracy of fault identification is higher than 79.38% in phase A of LLF. Obviously, when the proposed method is used, the test accuracy of fault identification is higher than that of using other methods from the results. In general, the average test accuracy of fault identification for using the proposed method is also higher than that of using other methods from the results for different faults except for the LLLGF. The average test accuracy of using the proposed method is higher than 99.27% in LLF. Regarding the computation time, when the UT is used, the average time of training is longer than that of other methods for different faults except for the LLLGF. Furthermore, the time of test for all methods and different faults is approximately 0.004 seconds. 2)Case Study 2, Transmission-Line Faults: In case study 2, the neural network algorithm in the NIFM system identifies SLGF on phase A, DLGF and LLF on phase A and B, and symmetric fault (LLLGF) at five different 220-V transmission lines based upon the proposed method, as shown in Fig. 7. There are essentially five 220V transmission lines in the model system of Fig. 7 because the distance is long between the 480V/220V distribution transformer and the load bus. Each line is 3 kilometers long and there are two sections per line, each section being 1.5 kilometers in length. This allows the user to apply faults at the section junctions. The lines are modeled using a constant parameter line model in EMTP program. The line conductor is a 1AWG with a 8.4-millimeter diameter and a dc resistance of 0.5426Ω/kilometer at 25°C. The line parameters are calculated at 60Hz with an earth resistivity of 100Ω-m. Table 2 shows that values for the training accuracy of fault identification are 100% for all current signatures with the proposed Pspectra. Besides, the test accuracy of fault
16
identification is higher than 71.25% in current signature Ic of LLF. The values for the training accuracy of fault identification are higher than 93.94% in current signature Ib and Ic of SLGF with the Z-score based WT. Moreover, the test accuracy of fault identification is higher than 4.375% in current signature Ia of SLGF. Finally, the values for the training accuracy of fault identification are higher than 45.15% in phase A of DLGF with the UT. Furthermore, the test accuracy of fault identification is higher than 22.81% in phase A of DLGF. Apparently, when the proposed method is used, the test accuracy of fault identification is higher than that of using other methods from the results. In general, the test accuracy of fault identification for using the proposed method is also higher than that of using other methods from the results of average for different faults. In addition, the average test accuracy of using the proposed method is higher than 79.9% in SLGF, even 98.13% in DLGF. In terms of the computation time, when the UT is used, the average time of training is longer than that of other methods for different faults. Moreover, the time of test for all methods and different faults is also approximately 0.004 seconds. 4.3. Discussions To prove the feasibility of the proposed Pspectra, not only can it identify the fault locations and types, but it can also distinguish faulty phases for power system transmission-line faults instantly. The results of two different feature extraction techniques show the superiority of the proposed method, i.e., Z-score based WT and UT, as shown in Table 2. To simplify these full input waveforms of power signatures, a number of statistical features were generated using moments of the fault data. The normalization of statistical moments constitutes a unit-free measure which can be used to compare observations
17
measured with different units [20]. The turn-on real-power transients differ from one another because the transient properties of a typical power system fault are mainly determined by the physical task that the load performs and the fault location changes. Therefore, the transient is the dominant state directly after fault inception. In Table 2 for the proposed method Pspectra in asymmetric faults, the test accuracies of the faulted phases are higher than those of the un-faulted phases. It is quite obvious that the proposed method can not only identify fault locations and types but it can recognize the faulty phases to compare with the Z-score based WT and UT. Fig. 8 and Fig. 9 show the current distributions of each order by using the Z-score based WT and the turn-on real-power transient waveforms for the SLGF on phase A at five different load buses, respectively. Furthermore, it is apparently observed that the proposed method can improve the test recognition accuracy of fault identification by comparing Fig. 6(b) with Fig. 8 and Fig. 9. Based on the reaction time of protection equipment, the device is usually to be used as short-circuit protection, the fuse or circuit breaker must interrupt the fault quickly (generally less than 4 milliseconds) in order to give the maximum protection to equipment and personnel [22]. Underwriters Laboratories Inc. (UL) defines breaker current limitation as a breaker that interrupts and isolates a fault in less than 1/2 of an ac cycle. 1/2 a cycle is completed in 8.3 milliseconds [23]. In these case studies of the paper, the time of test for the proposed method at different faults is approximately 0.004 seconds for 32 test examples. The time of test for each example is about 0.125 milliseconds. The time of test can react quickly the interruption and isolation of short-circuit fault for the fuse or circuit breaker. 5.
Conclusions
Based on EMTP simulation of NIFM, this paper presents a novel NIFM algorithm to
18
improve the associated recognition accuracy and enhance the capability of fault identification for power systems. The proposed NIFM algorithm applies Parseval's Theorem to the WTCs of the transient waveforms of fault events to find out their energy spectra. The advantage of using the derived energy spectra from the PSs is that it can select the important features of the transient signals without requiring excessive memory space and training time. The energy spectra are then processed by the proposed BP-ANN for training and testing. To verify the validity of the proposed method, four different fault types are investigated and two other feature extraction methods are compared for the load-bus faults and transmission-line faults in the paper. Even the cases include some of the most challenging scenarios for a NIFM system to identify such as system voltage variations, the time or inception angle changes of fault occurrence and different fault resistances, the proposed method yield better recognition accuracy and capability for fault locations, types, and classifications, as compared with other existing methods such as Z-score based WT and UT methods. Acknowledgements
The author would like to thank the Ministry of Science and Technology of Taiwan, Republic of China, for financially supporting this research under Contract No. MOST 104-2221-E-228-004. References
[1] F. Chunju, K. K. Li, W. L. Chan, Y. Weiyong, and Z. Zhaoning, “Application of wavelet fuzzy neural network in locating single line to ground fault (SLG) in distribution lines,” Electr. Pow. Energ. Syst., vol.29, pp.497-503, 2007. [2] A. Borghetti, M. Bosetti, M. D. Silvestro, C. A. Nucci, and M. Paolone, “Continuous-wavelet transform for fault location in distribution power networks:
19
definition of mother wavelets inferred from fault originated transients,” IEEE Trans. Power Syst., vol.23, no.2, pp.380-388, May 2008. [3] S. Ekici, S. Yildirim, and M. Poyraz, “Energy and entropy-based feature extraction for locating fault on transmission lines by using neural network and wavelet packet decomposition,” Expert Syst. Appl., vol.34, pp.2937-2944, 2008. [4] M. J. B. Reddy, D. V. Rajesh, P. Gopakumar, and D. K. Mohanta, “Smart fault location for smart grid operation using RTUs and computational intelligence techniques,” IEEE Syst. J., vol.8, no.4, pp.1260-1271, Dec. 2014. [5] F. Bezerra Costa, “Fault-induced transient detection based on real-time analysis of the wavelet coefficient energy,” IEEE Trans. Power Del., vol. 29, no. 1, pp. 140–153, Feb. 2014 [6] G. W. Hart, “Nonintrusive appliance load monitoring,” Proc. IEEE, vol. 80, no. 12, pp. 1870–1891, Dec. 1992. [7] A. Zoha, A. Gluhak, M. A. Imran, and S. Rajasegarar, “Non-intrusive load monitoring approaches for disaggregated energy sensing: A survey,” Sensors, vol. 12, no. 12, pp. 16838–16866, December 2012. [8] M. Akbar and Z. Khan, “Modified nonintrusive appliance load monitoring for nonlinear devices,” in Proc. INMIC, 2007, pp. 1–5. [9] H. H. Chang, K. L. Chen, Y. P. Tsai, and W. J. Lee, “A new measurement method for power signatures of non-intrusive demand monitoring and load identification,” IEEE Trans. Ind. Appl., vol. 48, no. 2, pp. 764-771, Mar./Apr. 2012. [10] S. B. Leeb, S. R. Shaw, and J. L. Kirtley, “Transient event detection in spectral envelope estimates for nonintrusive load monitoring,” IEEE Trans. Power Del., vol. 10, no. 3, pp. 1200–1210, Jul. 1995.
20
[11] H. T. Yang, H. H. Chang, and C. L. Lin, “Load identification in neural networks for a non-intrusive monitoring of industrial electrical loads,” Comput. Supported Coop. Work in Design IV, Lecture Notes Comput. Sci., vol. 5236, pp. 664-674, 2008. [12] D. Srinivasan, W. S Ng, and A C. Liew, "Neural-network-based signature recognition for harmonic source identification," IEEE Trans. Power Del., vol. 21, no. 1, pp. 398 – 405, Jan. 2006. [13] H. H. Chang, L. S. Lin, N. Chen, and W. J. Lee, “Particle-swarm-optimization-based nonintrusive demand monitoring and load identification in smart meters,” IEEE Trans. Ind. Appl., vol. 49, no. 5, pp. 2229-2236, Sep./Oct. 2013. [14] H. H. Chang, “Non-intrusive demand monitoring and load identification for energy management systems based on transient feature analyses,” Energies, vol. 5, pp. 4569-4589, 2012. [15] Z. L. Gaing, “Wavelet-based neural network for power disturbance recognition and classification,” IEEE Trans. Power Del., vol. 19, no. 4, pp. 1560-1568, Oct. 2004. [16] Y. Meyer, Wavelets: algorithms and applications, SIAM, Philadelphia, 1993. [17] D. C. Robertson, O. I. Camps, J. S. Mayer, andW. B. Gish, Sr., “Wavelets and electromagnetic power system transients,” IEEE Trans. Power Del., vol. 11, no. 2, pp. 1050–1056, Apr. 1996. [18] A.J. Mazon, I. Zamora, J. Gracia, K.J. Sagastabeitia, and J.R. Saenz, “Selecting ANN structures to find transmission faults,” IEEE Computer Applications in Power, pp. 44–48, Jul. 2001. [19] Y. Y. Hong and B. Y. Chen, “Locating switched capacitor using wavelet transform and hybrid principal component analysis network,” IEEE Trans. Power Del., vol. 22, no. 2, pp. 1145-1152, 2007.
21
[20] H. Guo, L. B. Jack, and A. K. Nandi, “Feature generation using genetic programming with application to fault classification,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 35, no. 1, pp. 89-99, Feb. 2005. [21] H. H. Chang, “Genetic algorithms and non-intrusive energy management system based economic dispatch for cogeneration units,” Energy, vol. 36, pp. 181-190, 2011. [22] Fuse selection guide, OptiFuse - Rev A 01/2010. [23] Miniature circuit breaker-Application guide, Low Voltage Products & Systems, ABB Inc., Apr. 2009.
22
Figure Captions
Fig. 1. Fault identification system on load-bus faults for a NIFM system. Fig. 2. Block diagram of a fault event identification system. Fig. 3. SLGF event with different occurrence times for BUS 1. (a) Different times of fault current occurrence (0.05s, 0.1s, and 0.15s), (b) Power distribution diagram. Fig. 4. SLGF event with different current amplitude magnitudes for BUS1. (a) Different amplitude magnitudes of fault occurrence (-20%, 0%, and +20%), (b) Power distribution diagram. Fig. 5. SLGF event with different fault resistances for BUS1. (a) Different fault resistances of fault occurrence (0pu, 0.002pu, and 0.02pu), (b) Power distribution diagram. Fig. 6. Distributions of each scale for five different load buses. (a) Voltage. (b) Current. Fig. 7. Fault identification system on transmission-line faults for a NIFM system. Fig. 8. Current distributions of each order for the SLGF on five different load buses. Fig.9.
Turn-on real-power transient waveforms for the SLGF on five different load buses.
23
Fig. 1. Fault identification system on load-bus faults for a NIFM system.
24
Fig. 2. Block diagram of a fault event identification system.
25
(a)
(b) Fig. 3. SLGF event with different occurrence times for BUS 1. (a) Different times of fault current occurrence (0.05s, 0.1s, and 0.15s), (b) Power distribution diagram.
26
(a)
(b) Fig. 4. SLGF event with different current amplitude magnitudes for BUS1. (a) Different amplitude magnitudes of fault occurrence (-20%, 0%, and +20%), (b) Power distribution diagram.
27
(a)
(b) Fig. 5. SLGF event with different fault resistances for BUS1. (a) Different fault resistances of fault occurrence (0pu, 0.002pu, and 0.02pu), (b) Power distribution diagram.
28
(a)
(b) Fig. 6. Distributions of each scale for five different load buses. (a) Voltage. (b) Current.
29
Fig. 7. Fault identification system on transmission-line faults for a NIFM system.
30
Fig. 8. Current distributions of each order for the SLGF on five different load buses.
Fig. 9. Turn-on real-power transient waveforms for the SLGF on five different load buses.
31
Table 1 The results of fault identification in case study 1 Fault Type
SLGF
DLGF
LLF
LLLGF
Methods Monitoring on Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s)
Ia
Pspectrum Ib Ic
Average
Ia
Z-score based WT Ib Ic
UT Average
Phase A
Phase B
Phase C
Average
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
95
95.63
97.19
95.94
100
96.88
100
98.96
0.6129 0.0039
0.5688 0.0039
0.5725 0.0040
0.5847 0.0039
0.4996 0.0040
0.5023 0.0040
0.5403 0.0041
0.5141 0.0040
0.5693 0.0050
0.7449 0.0041
0.7638 0.0044
0.6927 0.0045
100
100
100
100
100
100
100
100
100
94.55
100
98.18
100
100
100
100
74.06
35.94
86.88
65.63
100
82.50
100
94.17
0.6662 0.0039
0.5759 0.0039
0.6719 0.0040
0.6380 0.0039
0.5429 0.0042
0.7056 0.0043
0.6474 0.0042
0.6320 0.0042
0.5615 0.0042
25.249 0.0040
0.6014 0.0041
8.8040 0.0041
100
100
100
100
100
100
100
100
95.45
100
98.79
98.08
100
97.82
100
99.27
55
18.44
98.13
57.19
79.38
100
90
89.79
0.5756 0.0062
0.6015 0.0040
0.5803 0.0040
0.5858 0.0047
0.5975 0.0042
0.6050 0.0042
0.6909 0.0042
0.6311 0.0042
29.005 0.0042
0.5334 0.0041
9.0808 0.0041
12.873 0.0041
100
100
100
100
100
100
100
100
100
100
100
100
100
100
99.38
99.79
93.44
30.63
93.13
72.40
100
100
100
100
0.5990 0.0040
0.5876 0.0041
0.9930 0.0040
0.7265 0.0040
0.5688 0.0041
0.5565 0.0042
0.5812 0.0041
0.5688 0.0041
0.5316 0.0043
0.5226 0.0042
0.5242 0.0042
0.5261 0.0042
32
Table 2 The results of fault identification in case study 2 Fault Type
SLGF
DLGF
LLF
LLLGF
Methods Monitoring on Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in
Ia
Pspectrum Ib Ic
Average
Ia
Z-score based WT Ib Ic
UT Average
Phase A
Phase B
Phase C
Average
100
100
100
100
100
93.94
93.94
95.96
76.36
89.09
100
88.49
88.13
75.94
75.63
79.90
4.375
18.44
32.81
18.54
59.06
64.38
93.13
72.19
0.6250 0.0040
0.6291 0.0040
0.6659 0.0041
0.6400 0.0040
1.4026 0.0043
60.330 0.0043
41.610 0.0043
34.448 0.0043
46.788 0.0042
41.926 0.0043
2.1864 0.0041
30.300 0.0042
100
100
100
100
100
100
100
100
45.15
51.52
83.03
59.90
99.38
97.81
97.19
98.13
36.25
62.19
65.31
54.58
22.81
33.13
59.69
38.54
0.6291 0.0041
0.6466 0.0042
0.6997 0.0043
0.6585 0.0042
0.5630 0.0042
0.6179 0.0043
0.6048 0.0042
0.5952 0.0042
40.870 0.0042
40.028 0,0041
43.826 0.0048
41.575 0.0045
100
100
100
100
100
100
100
100
63.94
46.67
68.79
59.80
90.63
93.44
71.25
85.11
27.50
45.31
35.00
35.94
41.25
28.13
48.75
39.38
0.6977 0.0042
0.6713 0.0041
0.7978 0.0042
0.7223 0.0042
0.7654 0.0041
0.5642 0.0041
0.7871 0.0042
0.7056 0.0041
42.470 0.0040
40.610 0.0041
43.330 0.0043
42.137 0.0041
100
100
100
100
100
100
100
100
60.00
73.08
75.00
69.36
72.31
88.85
93.85
85.00
17.69
12.69
52.69
27.69
44.23
48.08
53.08
48.46
33
Test (%) Training Time (s) Test Time (s)
0.5525 0.0041
0.5450 0.0042
0.5124 0.0041
0.5366 0.0041
0.7326 0.0041
0.78681 0.0041
34
0.4879 0.0040
0.6691 0.0041
36.410 0.0041
39.946 0.0042
40.566 0.0042
38.974 0.0042
S0378-7788(16)30538-2 http://dx.doi.org/doi:10.1016/j.enbuild.2016.06.050 ENB 6785
To appear in:
ENB
Received date: Revised date: Accepted date:
11-11-2015 27-3-2016 16-6-2016
Please cite this article as: Hsueh-Hsien Chang, Non-intrusive fault identification of power distribution systems in intelligent buildings based on power-spectrum-based wavelet transform, Energy and Buildings http://dx.doi.org/10.1016/j.enbuild.2016.06.050 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Non-intrusive fault identification of power distribution systems in intelligent buildings based on power-spectrum-based wavelet transform Hsueh-Hsien Chang Dept. of Electronic Engineering, Jin-Wen University of Science and Technology, New Taipei, Taiwan [email protected]
Highlights
• Non-intrusive monitoring techniques are proposed in real-time load-bus faults and transmission-line faults detection. • A power-spectrum-based wavelet transform and ANNs are proposed. • Parseval’s Theorem is adopted to reduce the number of WTCs representing fault transients. • The proposed method improves significantly the performances of the distribution system fault detection in intelligent buildings. • The proposed method is scarcely influenced to the fault inception angles, fault resistances, and system voltage variations.
Abstract A new approach for protection of power distribution systems in intelligent buildings has been presented in this paper. Directly adopting the wavelet transform coefficients (WTCs) requires longer computation time and larger memory requirements for the non-intrusive fault monitoring (NIFM) identification process. However, the WTCs contain plenty of information needed for the symmetric and asymmetric transient signals of fault events. To effectively reduce the number of WTCs representing fault transient signals without degrading performance, a power spectrum of the WTCs in different scales calculated by Parseval’s Theorem is proposed in this paper. In this paper, artificial neural networks (ANNs), in combination with power-spectrum-based wavelet transform, are used to identify fault types and locations in power distribution systems of industrial buildings by using NIFM. The high success rates of fault event recognition for load-bus 1
faults and transmission-line faults from simulations have proved that the proposed algorithm is applicable to fault identifications of non-intrusive monitoring applications.
Keywords: Artificial neural networks (ANNs); Parseval’s Theorem; wavelet transform; non-intrusive fault monitoring (NIFM); intelligent buildings.
Acronyms ANFIS:
Adaptive Neuro-Fuzzy Inference System
ANN:
Artificial Neural Network
BP-ANN:
Back-Propagation Artificial Neural Network
DLGF:
Double Line-to-Ground Fault
DWT:
Discrete Wavelet Transform
EMTP:
Electromagnetic Transient Program
ESE:
Electrical Service Entry
FTED:
Fault Transient Event Detection
LLF:
Line-to-Line Fault
LLLGF:
Three-Line-to-Ground Fault
MDMS:
Meter Database Management System
MFNN:
Multi-layer Feed-forward Neural Network
NIFM:
Non-Intrusive Fault Monitoring
NILM:
Non-Intrusive Load Monitoring
PSO:
Particle Swarm Optimization
PSs:
Power Signatures
SCADA:
Supervisory Control and Data Acquisition
SE:
Spectral Envelope
2
SLGF:
Single Line-to-Ground Fault
STFT:
Short-Time Fourier Transform
UT:
Turn-on Transient Energy Algorithm
WFNN:
Wavelet Fuzzy Neural Network
WMRA:
Wavelet Multi-Resolution Analysis
WPT:
Wavelet Packet Transform
WTCs:
Wavelet Transform Coefficients
1. Introduction In protection analyses of power system, power system fault is one of the most important research topics. These faults which are mainly due to short circuit phenomena can drastically fail the operations of power systems and cause excessively high currents to flow which causes damages to devices and fire accidents to buildings. Types of faults, like a short circuit condition in a power system distribution network, will result in severe economic losses and reduce the reliability of the electrical system. Different types of transient phenomena occur on the transmission line. From these transient phenomena, faults on transmission lines and load buses need to be detected, located, classified accurately, and cleared as fast as possible. In earth fault and short-circuit protections, faulty phase identification and location of fault are the two most important items which need to be addressed in a reliable and accurate manner. In multiple phases system, the faults are basically divided into symmetric (balanced) and asymmetric (un-balanced) types. The type of symmetric fault is only a balanced three-line-to-ground fault (LLLGF). Different types of asymmetric faults are single line-to-ground fault (SLGF), line-to-line fault (LLF), and double line-to-ground fault (DLGF). 1.1. Research background Due to the serious consequences of fault damages in power systems, many researchers
3
have devoted their efforts to this area and reported their findings in articles. Chunju et al. [1] have proposed a SLGF location method for employing wavelet fuzzy neural network (WFNN) to extract fault characteristics from the fault signals in an industrial distribution power system. However, this method cannot significantly increase the fault location accuracy because it does not use the high frequency information of pre-fault current and voltage. In fact, when fault location changes, the equivalent capacitance will change, the charging and discharging currents will be also changed, namely the high frequency currents are different. The high frequency currents and voltages are related with the distance from the relaying point to the fault location. Borghetti et al. [2] have built specific mother wavelets inferred from the recorded fault-originated voltage transient waveforms to improve the wavelet analysis. In this paper, the authors assumed that the network topology and the traveling wave speeds of the various propagation modes are known. However, as concluded in [2], this method is expected to improve the algorithm accuracy by means of proper integration of time-domain fault location approaches. Reference [3] has used wavelet energy and entropy criterion of the wavelet packet transform (WPT) coefficients for every faulty current and voltage signal to extract features and reduce the size of data sets for training and testing of artificial neural networks (ANNs). However, this method only implements in SLGF of power systems. Reddy et al. [4] have performed well to use a wavelet multi-resolution analysis (WMRA) technique to extract the features of the transient current signals and employed computational intelligence techniques as an adaptive neuro-fuzzy inference system (ANFIS) and ANN in conjunction with GPS to identify fault location. This paper focuses on amplitudes of the second- and third-order harmonics generated during fault current occurrence to track the fault location. Therefore, among different coefficients pertaining to different decomposition levels, only the summation of the fifth-level detailed
4
coefficients (d5) is considered for the sampling rate of 6 kHz. Bezerra Costa [5] has presented a wavelet-based methodology for real-time detection of fault-induced transients in transmission lines, where the wavelet coefficient energy takes into account the border effects of the sliding windows. As a result, the performance of the proposed energy analysis is not affected by the choice of the mother wavelet and presenting no time delay in real-time fault detection. The concept of non-intrusive load monitoring (NILM) was originally introduced at MIT by Hart [6] in late 80’s. The NILM, in contrast with the traditional supervisory control and data acquisition (SCADA)-based system, can drastically reduce hardware and maintenance costs because only one set of voltage and current sensors is needed at the electrical service entry (ESE) [6]. Although several NILM algorithms were developed during the past two decades, recognition accuracy and computational efficiency have been remaining as challenges. Hart proposed a method for disaggregating electrical loads by examining only the appliance specific power consumption signatures within the aggregated load data [7]. However, the voltage variations of the utility may result in overlapping of loads in the P-Q plane [8], different loads may consume the same real and reactive powers or they are with the simultaneous start and power signatures (PSs) that have non-discrete changes in power consumption may not be adequately measured only from steady-state parameters [6], [9]. In addition to the steady-state power draw, Leeb et al. have used the “spectral envelope” (SE) concept [10] for a NILM system. The SE is a vector of the first several coefficients of the short-time Fourier transform (STFT) of the transient current signal. Although this method can detect numerous appliances including the variable loads, it requires extensive training for each appliance before classification and monitoring can be performed. Yang et al. [11] employed a turn-on transient energy algorithm (UT) in a non-intrusive monitoring of industrial electrical loads. However, this
5
algorithm requires a high sampling rate to accurately capture the detailed variations in the transient energy. Recently, there has been a growing interest in improving the performance of recognition for using artificial intelligence methods such as ANNs. Srinivasan et al. [12] proposed a NN-based approach to identify harmonic sources. However, the method does not incorporate the various operation modes in each load. In [13], the authors have applied particle swarm optimization (PSO) to optimize the parameters of training algorithms in NN for improving the recognition accuracy. Nevertheless, the resulting training time could be quite long. 1.2. Problems and contributions However, as far as the author is well aware that fault identification for the distribution system of an industrial building by using non-intrusive monitoring techniques is still in the primary stage of development. Based on the concept of the NILM, this paper proposes new fault identification approaches for load-bus faults and transmission-line faults on low-voltage power distribution systems in intelligent buildings by using non-intrusive monitoring techniques. The non-intrusive fault monitoring (NIFM) refers to a method of detecting the current waveforms of a distribution system using a single set of sensors at ESE. Fig. 1 shows a typical NIFM system on power distribution systems in an industrial building. Three-phase three-wire and three-phase four-wire systems are generally used to run a typical power system distribution network. As shown in the figure, the secondary side of the distribution transformer, the voltage and current at ESE are measured and sent to the meter database management system (MDMS). This data is processed by the NIFM algorithms to identify the operation statuses and fault events for the individual feeder. As in the existing methods mentioned above, the wavelet transform (WT) technology has been often used to capture the time of transient occurrence [14]. Integrating the WT technology with the artificial intelligence method or expert system becomes a promising
6
approach to improve the recognition accuracy of fault events. However, the following issues must be overcome before their practicality can be realized for a non-intrusive classifier in the NIFM system [15]. 1) Directly adopting the WT coefficients as inputs of the NNs requires large memory space and long training time; 2) The decomposition level with the number of extraction features must be reduced to enhance computation efficiency of the NIFM system; 3) Not only can symmetric and asymmetric fault events be identified but load-bus faults and transmission-line faults can also be detected. 4) The reliability of the electrical system cannot be reduced even only one set of voltage and current sensors is needed at ESE. In an effort to overcome some of the above-mentioned problems, a new NIFM system is proposed in this paper. This paper presents a novel classifier consisting of two models as shown in Fig. 2. First, the WMRA technique and Parseval’s Theorem are utilized to extract the energy distribution features of the fault transient signal from Electromagnetic Transient Program (EMTP) simulation and ESE at different resolution levels. Second, the back-propagation artificial neural network (BP-ANN) is employed to classify fault locations and types according to the detailed energy distribution. By using Parseval’s Theorem, the number of transient power features can be reduced without losing their fidelity. The accuracy rate and the computation requirement of the proposed method were tested in a simulated model system by using EMTP software. These results showed that the proposed methods can efficiently and effectively analyze the signals, thus enhance the performance of the fault identifications in the NIFM system to compare with other extraction methods.
7
2.
The proposed NIFM system in intelligent buildings The block diagram in Fig. 2 includes the algorithms of fault transient event detection
(FTED), the processes of the WT technique, the feature extractions from symmetric and asymmetric fault transient signals, and the BP-ANN classification network. The procedures of recognizing fault transient signals belonging to certain kinds of the fault events are described as follows. First, a fault occurrence of transient event is detected by the algorithm of FTED.
Then, a complete fault event signal can be acquired by
monitoring at ESE, i.e. the steady-state signal ahead and the transient signal back from the fault inception. The current waveforms of the complete fault transient signals are transferred using WT to obtain the different scales of the wavelet transform coefficients (WTCs). Although all scales of the WTCs of a complete fault transient signal may contain some important features of the faults, the identification system would require many memory spaces if all the WTCs are adopted directly as the features. As shown in the Fig. 2, Parseval’s Theorem is therefore employed to extract the power spectrum of the WTCs as the features of non-intrusive fault identification system. To equip the identification system with the ability to learn any arbitrary mapping of network inputs to desired outputs, the BP-ANN which has been trained and tested from the power spectrum inputs of EMTP simulation and previous fault waveforms measured at ESE is utilized in this paper. 3.
Proposed methods
3.1. WT Morlet and Grossman [16] used the WT to manage the extended training conditions for applying STFT. By applying the band-pass filters’ special properties, it is proposed that the WT can be used to detect and locate the transience phenomena with high-frequency components of power signals with different scales of the WTCs in NIFM systems.
8
ψa,b (t )
t b ψ , a R , b R a a
1
(1)
where ψ a, b t is the daughter wavelet; ψ t is the mother wavelet; a is the scaling factor; and b is the shift factor. Principally, decompositions of signals as a series of daughter wavelets to be analyzed are achieved by using scaling and shift factors of discrete WT (DWT). Thus, it is proper for transient analyses to be carried out [17]. As statements for the DWT, owning to different frequency components, the WTCs scales have frequently been used in determining the fault transient occurrences by focusing on the high-frequency components on power systems. These high-frequency components are often the results of the signals coming from the symmetric and asymmetric transients. Daubechies is used as the mother wavelet for the WT in this research. 3.2. WTCs’ power spectrum In order to depict the unitary characteristics of any Fourier transform in physics and engineering, the Parseval’s Theorem is applied. By utilizing a Fourier series, a periodic signal f (t) is decomposed into the sum of a set of harmonic oscillators, and can be derived as follows: f (t )
1 a0 2
an cos(
n 1
nπt nπt bn sin( ) ) T T n 1
1 T a0 f ( t ) dt T T T nπt a 1 ) dt f ( t ) cos( n T T T T 1 nπt ) dt f ( t ) sin( bn T T T
(2)
(3)
where T is the time of a cycle. From the energy perspective, (2) can be formulated to become (4). 9
1 T
T
T
1 f ( t ) dt a 02 2 2
(a
2 n
bn2 )
(4)
n 1
Suppose that the scaling function and the mother wavelet are orthogonal basis sets, the relationship between the power spectrum of the discrete signal f [n] and each of the WTCs can be set up by Parseval’s Theorem. The WT can isolate the signal energy in time and frequency domains. According to the MRA characteristics of the WT, a complete fault transient signal can be depicted by each scale of the WTCs.
WTCs_ f [n] [ CJ ,k d1,k d2,k ... dJ ,k ]
(5)
where J is a number of the WT scale; CJ ,k and d j,k are the coefficients of approximate and detailed versions of a complete fault transient signal, respectively. Hence, owning to the DWT decomposition, the energy of the complete fault transient signal is formulated in (6). The first term on the right side of (6) describes the average power of the approximated version of the decomposed signal, and the second term describes the sum of the average power of the detailed version of the decomposed signal. 1 N
f [n ]
2
nN
1 NJ
k
C J ,k
2
1 N j j 1 J
k
d
2 j ,k
(6)
To filter out the noise, by subtracting the standard deviation of the absolute WTCs from the WTCs, the d j , k coefficients are modified into a new one ( d j , k ). As mentioned above, the high-frequency components of the second term in (6), will be used to extract the features of the fault event in NIFM systems, and formulated as Psignal in (7). The power spectrum of a fault event signal described by Parseval’s Theorem and the WTCs, the power spectrum of Psignal of each WTC, can thus be established as shown in (8). 10
Psignal
1 N j
d j , k
k
Psp ectru m Psig n a l
1 2
2
d j Nj
d j Nj
2
2
(7) 1
2
(8)
In the end, as (8) indicates, the energy of the event signal can be separated at different resolution levels in different processes dependent on the sampling frequency. For instance, the sampling frequency of 15 kHz, the decomposition levels can be 1 to 8 in order to examine the coefficient of the detailed version at each resolution level for extracting the fault signal features inherent in the NIFM system. Besides, the following confirmed important properties from EMTP simulations and experiments resulted from (8) for recognition features are described below: 1) The energy distribution remains the same at different timing or inception angle of fault event occurrence; 2) The outline of the energy distribution remains similar to variations in the different system voltage amplitude and fault resistances of the same fault type; 3) The low-level energy distribution will denote obvious changes when there is high-frequency components appearing in the event signal. For instance, Fig. 3(a) indicates a complete transient current waveform with three different occurrences (0.05s, 0.1s, and 0.15s) of a SLGF on phase A at BUS 1 as in Fig. 1. Investigations on the three signals show that the detailed energy distributions (scales 1~ scale 8) of the given signals are identical to those shown in Fig. 3(b). When the different current amplitudes (-20%, 0%, and +20%) of the fault event appear as those indicated in Fig. 4(a), the detailed version of energy distributions (scale 1~scale 8) is similar to that shown in Fig. 4(b). Moreover, when the different fault resistances (0pu, 0.002pu, and
11
0.02pu) appear as those indicated in Fig. 5(a), the detailed version of energy distributions (scale 1~scale 8) is also similar to that shown in Fig. 5(b). As a consequence, the fault detection is scarcely influenced to the fault inception angles, fault resistances, and system voltage variations for using the proposed methods. Accordingly, by using differences of the detailed version of energy distribution, the features of fault signals can demarcate different fault locations. As shown in Fig. 1 and Fig. 6, five different loads on different buses labeled as bus 1 to bus 5 are one 45-hp induction motor, one 55-hp induction motor, one 30-hp induction motor, one 20-hp induction motor driven by variable-voltage drives, and one load bank supplied by a six-pulse thyristor rectifier from ac power, respectively. The power spectrums in Fig. 6(a) and 6(b) depict that the five different buses have different energy distributions from scale 1 to scale 8 for SLGF transient voltage and current waveforms, respectively. Consequently, in terms of the small-scale power spectrum (i.e. in the high-frequency domain); different fault buses can be classified using particular energy spectra as the features in the NIFM system. Evidently, it is better to select the fault transient current waveform than the fault transient voltage waveform for fault identification of power distribution systems. 3.3. ANNs There is a wide range of power system applications of neural networks in operation and control processes, including stability assessment, security monitoring, load forecasting, state estimation, load flow analysis, contingency analysis, emergency control actions, etc. [18]. 1) Training algorithms and fitness function Over-fitting is a typical difficulty occurring during ANN trainings. Bayesian regularization typically provides better generalization performance than early stopping
12
due to its not requiring a validation dataset to be segregated from the training dataset. In other words, all training datasets are required for the input training data. This benefit is particularly critical when the available dataset is small. The representative fitness function applied in training a feedforward neural network is the mean sum of squares of network errors, alternatively called MSE. It is the average squared error between network outputs and desired outputs.
MSE(l )
1 (t j (l ) a j (l ))2 2 jC
(9)
where the variable t j (l) is the desired output for the lth iteration at node j , and variable
aj (l) is the network output for the lth iteration at node j. 2) Multi-layer feed-forward neural network Most BP-ANN applications apply single- or multilayer perceptron networks by using gradient-descent training methods combined with learning through backward propagation. These multi-layer perceptrons can be trained under supervision using analytical functions to activate network nodes (“neurons”); besides, by applying a backward error-propagation algorithm to update interconnecting weights and thresholds, a sufficient recognition capability can later be achieved. A trainable classifier applies the BP classifier in a multi-layer feed-forward neural network (MFNN) in this study. “Classification” in this context denotes a mapping from a feature space to a set of class labels, which are the names of fault points. Generally speaking, a supervised MFNN is partitioned into three layers: input, hidden, and output layers. These neurons are connected by linked weights selected to meet the desired relationship between input and output neurons. The MFNN in this study is employed to identify faults in the NIFM system. The input, output, and hidden layers of the BP-ANN are described as follows: 13
a) Input layer: PSs information, the energy spectra of a complete fault event signal, for an ESE or EMTP simulation serves as inputs. b) Output layer: the number of output neurons is the same as that of individual fault feeder identified. Each binary bit serves as a load-bus or transmission-line indicator for fault status. c) Hidden layer: two hidden layers are used in this study. Some heuristics that have been developed can determine the number of neurons in a hidden layer [19]. The common number of neurons in a hidden layer is the sum of the number of neurons in an input layer and that in an output layer. 4.
Experimental results
4.1. Study environments Experimental data sets were generated by preprocessing the data on the current waveforms of the total feeders. Each final sample consists of (T60256) samples obtained over a period of T. Each example of the power feature includes a voltage variation from − 3% to +3% at 0.5% intervals, yielding thirteen examples of power features for each bus or feeder and (N13) raw data for each power signature ( I a , I b , I c ) , given that N load buses can only be a short circuit for one bus or line at any time in three-phase electric power distribution systems. The full input dataset (N13)(T 60256)
matrix includes the training dataset and the test dataset. The training data and test data are randomly selected from all data. To confirm the inferential power of the neural networks, the full raw data set creates, respectively, a (N13)/ 2 matrix as the training data set and test dataset. A simulated NIFM system constructed by the EMTP software program monitors the current waveforms in a three-phase ESE powering a collection of loads representative of the five load classes in an industrial building. As shown in Fig. 1 for load-bus faults and 14
Fig. 7 for transmission-line faults, five different loads on different buses include a 45-hp induction motor, a 55-hp induction motor, a 30-hp induction motor, a 20-hp induction motor driven by variable-voltage drives, and a bank of loads supplied by a six-pulse thyristor rectifier for ac power. A neural network simulation program was designed using MATLAB. The program was run to identify faults on a PC equipped with a 3.10 GHz Intel Core i5- 4440 CPU. Each entry in the BP-ANN represents 10 different trials, and each trial uses random initial weights. In each case, the network is trained until the MSE is less than 0.00001 or the number of iterations reaches 5000. The proposed method is able to identify the fault locations and types instantly. The results of four different fault types for load-bus faults and transmission-line faults show the feasibility and superiority of the proposed methods compared with other extraction methods such as standardized statistical moments based WT (Z-score based WT) [20] and turn-on transient energy algorithm (UT) [21]. 4.2. Results 1)Case Study 1, Load-Bus Faults: In case study 1, the neural network algorithm in the NIFM system identifies SLGF on phase A, DLGF and LLF on phase A and B, and symmetric fault (LLLGF) at five different 220-V load buses based upon the proposed power spectrum method (Pspectra), as shown in Fig. 1. There are basically five 220V distributed lines in the model system of Fig. 1 because the distance is short between the 480V/220V distribution transformer and the load bus. The lines are modeled as lumped parameter sections in EMTP program. Table 1 shows that values for the training accuracy of fault identification are 100% for all current signatures with the proposed Pspectra. Furthermore, the test accuracy of fault identification is higher than 97.82% in current signature Ib of LLF. The values for the training accuracy of fault identification are also 100% for all current signatures with the
15
Z-score based WT. Moreover, the test accuracy of fault identification is higher than 18.44% in current signature Ib of LLF. Finally, the values for the training accuracy of fault identification are higher than 94.55% in phase B of DLGF with the UT. Besides, the test accuracy of fault identification is higher than 79.38% in phase A of LLF. Obviously, when the proposed method is used, the test accuracy of fault identification is higher than that of using other methods from the results. In general, the average test accuracy of fault identification for using the proposed method is also higher than that of using other methods from the results for different faults except for the LLLGF. The average test accuracy of using the proposed method is higher than 99.27% in LLF. Regarding the computation time, when the UT is used, the average time of training is longer than that of other methods for different faults except for the LLLGF. Furthermore, the time of test for all methods and different faults is approximately 0.004 seconds. 2)Case Study 2, Transmission-Line Faults: In case study 2, the neural network algorithm in the NIFM system identifies SLGF on phase A, DLGF and LLF on phase A and B, and symmetric fault (LLLGF) at five different 220-V transmission lines based upon the proposed method, as shown in Fig. 7. There are essentially five 220V transmission lines in the model system of Fig. 7 because the distance is long between the 480V/220V distribution transformer and the load bus. Each line is 3 kilometers long and there are two sections per line, each section being 1.5 kilometers in length. This allows the user to apply faults at the section junctions. The lines are modeled using a constant parameter line model in EMTP program. The line conductor is a 1AWG with a 8.4-millimeter diameter and a dc resistance of 0.5426Ω/kilometer at 25°C. The line parameters are calculated at 60Hz with an earth resistivity of 100Ω-m. Table 2 shows that values for the training accuracy of fault identification are 100% for all current signatures with the proposed Pspectra. Besides, the test accuracy of fault
16
identification is higher than 71.25% in current signature Ic of LLF. The values for the training accuracy of fault identification are higher than 93.94% in current signature Ib and Ic of SLGF with the Z-score based WT. Moreover, the test accuracy of fault identification is higher than 4.375% in current signature Ia of SLGF. Finally, the values for the training accuracy of fault identification are higher than 45.15% in phase A of DLGF with the UT. Furthermore, the test accuracy of fault identification is higher than 22.81% in phase A of DLGF. Apparently, when the proposed method is used, the test accuracy of fault identification is higher than that of using other methods from the results. In general, the test accuracy of fault identification for using the proposed method is also higher than that of using other methods from the results of average for different faults. In addition, the average test accuracy of using the proposed method is higher than 79.9% in SLGF, even 98.13% in DLGF. In terms of the computation time, when the UT is used, the average time of training is longer than that of other methods for different faults. Moreover, the time of test for all methods and different faults is also approximately 0.004 seconds. 4.3. Discussions To prove the feasibility of the proposed Pspectra, not only can it identify the fault locations and types, but it can also distinguish faulty phases for power system transmission-line faults instantly. The results of two different feature extraction techniques show the superiority of the proposed method, i.e., Z-score based WT and UT, as shown in Table 2. To simplify these full input waveforms of power signatures, a number of statistical features were generated using moments of the fault data. The normalization of statistical moments constitutes a unit-free measure which can be used to compare observations
17
measured with different units [20]. The turn-on real-power transients differ from one another because the transient properties of a typical power system fault are mainly determined by the physical task that the load performs and the fault location changes. Therefore, the transient is the dominant state directly after fault inception. In Table 2 for the proposed method Pspectra in asymmetric faults, the test accuracies of the faulted phases are higher than those of the un-faulted phases. It is quite obvious that the proposed method can not only identify fault locations and types but it can recognize the faulty phases to compare with the Z-score based WT and UT. Fig. 8 and Fig. 9 show the current distributions of each order by using the Z-score based WT and the turn-on real-power transient waveforms for the SLGF on phase A at five different load buses, respectively. Furthermore, it is apparently observed that the proposed method can improve the test recognition accuracy of fault identification by comparing Fig. 6(b) with Fig. 8 and Fig. 9. Based on the reaction time of protection equipment, the device is usually to be used as short-circuit protection, the fuse or circuit breaker must interrupt the fault quickly (generally less than 4 milliseconds) in order to give the maximum protection to equipment and personnel [22]. Underwriters Laboratories Inc. (UL) defines breaker current limitation as a breaker that interrupts and isolates a fault in less than 1/2 of an ac cycle. 1/2 a cycle is completed in 8.3 milliseconds [23]. In these case studies of the paper, the time of test for the proposed method at different faults is approximately 0.004 seconds for 32 test examples. The time of test for each example is about 0.125 milliseconds. The time of test can react quickly the interruption and isolation of short-circuit fault for the fuse or circuit breaker. 5.
Conclusions
Based on EMTP simulation of NIFM, this paper presents a novel NIFM algorithm to
18
improve the associated recognition accuracy and enhance the capability of fault identification for power systems. The proposed NIFM algorithm applies Parseval's Theorem to the WTCs of the transient waveforms of fault events to find out their energy spectra. The advantage of using the derived energy spectra from the PSs is that it can select the important features of the transient signals without requiring excessive memory space and training time. The energy spectra are then processed by the proposed BP-ANN for training and testing. To verify the validity of the proposed method, four different fault types are investigated and two other feature extraction methods are compared for the load-bus faults and transmission-line faults in the paper. Even the cases include some of the most challenging scenarios for a NIFM system to identify such as system voltage variations, the time or inception angle changes of fault occurrence and different fault resistances, the proposed method yield better recognition accuracy and capability for fault locations, types, and classifications, as compared with other existing methods such as Z-score based WT and UT methods. Acknowledgements
The author would like to thank the Ministry of Science and Technology of Taiwan, Republic of China, for financially supporting this research under Contract No. MOST 104-2221-E-228-004. References
[1] F. Chunju, K. K. Li, W. L. Chan, Y. Weiyong, and Z. Zhaoning, “Application of wavelet fuzzy neural network in locating single line to ground fault (SLG) in distribution lines,” Electr. Pow. Energ. Syst., vol.29, pp.497-503, 2007. [2] A. Borghetti, M. Bosetti, M. D. Silvestro, C. A. Nucci, and M. Paolone, “Continuous-wavelet transform for fault location in distribution power networks:
19
definition of mother wavelets inferred from fault originated transients,” IEEE Trans. Power Syst., vol.23, no.2, pp.380-388, May 2008. [3] S. Ekici, S. Yildirim, and M. Poyraz, “Energy and entropy-based feature extraction for locating fault on transmission lines by using neural network and wavelet packet decomposition,” Expert Syst. Appl., vol.34, pp.2937-2944, 2008. [4] M. J. B. Reddy, D. V. Rajesh, P. Gopakumar, and D. K. Mohanta, “Smart fault location for smart grid operation using RTUs and computational intelligence techniques,” IEEE Syst. J., vol.8, no.4, pp.1260-1271, Dec. 2014. [5] F. Bezerra Costa, “Fault-induced transient detection based on real-time analysis of the wavelet coefficient energy,” IEEE Trans. Power Del., vol. 29, no. 1, pp. 140–153, Feb. 2014 [6] G. W. Hart, “Nonintrusive appliance load monitoring,” Proc. IEEE, vol. 80, no. 12, pp. 1870–1891, Dec. 1992. [7] A. Zoha, A. Gluhak, M. A. Imran, and S. Rajasegarar, “Non-intrusive load monitoring approaches for disaggregated energy sensing: A survey,” Sensors, vol. 12, no. 12, pp. 16838–16866, December 2012. [8] M. Akbar and Z. Khan, “Modified nonintrusive appliance load monitoring for nonlinear devices,” in Proc. INMIC, 2007, pp. 1–5. [9] H. H. Chang, K. L. Chen, Y. P. Tsai, and W. J. Lee, “A new measurement method for power signatures of non-intrusive demand monitoring and load identification,” IEEE Trans. Ind. Appl., vol. 48, no. 2, pp. 764-771, Mar./Apr. 2012. [10] S. B. Leeb, S. R. Shaw, and J. L. Kirtley, “Transient event detection in spectral envelope estimates for nonintrusive load monitoring,” IEEE Trans. Power Del., vol. 10, no. 3, pp. 1200–1210, Jul. 1995.
20
[11] H. T. Yang, H. H. Chang, and C. L. Lin, “Load identification in neural networks for a non-intrusive monitoring of industrial electrical loads,” Comput. Supported Coop. Work in Design IV, Lecture Notes Comput. Sci., vol. 5236, pp. 664-674, 2008. [12] D. Srinivasan, W. S Ng, and A C. Liew, "Neural-network-based signature recognition for harmonic source identification," IEEE Trans. Power Del., vol. 21, no. 1, pp. 398 – 405, Jan. 2006. [13] H. H. Chang, L. S. Lin, N. Chen, and W. J. Lee, “Particle-swarm-optimization-based nonintrusive demand monitoring and load identification in smart meters,” IEEE Trans. Ind. Appl., vol. 49, no. 5, pp. 2229-2236, Sep./Oct. 2013. [14] H. H. Chang, “Non-intrusive demand monitoring and load identification for energy management systems based on transient feature analyses,” Energies, vol. 5, pp. 4569-4589, 2012. [15] Z. L. Gaing, “Wavelet-based neural network for power disturbance recognition and classification,” IEEE Trans. Power Del., vol. 19, no. 4, pp. 1560-1568, Oct. 2004. [16] Y. Meyer, Wavelets: algorithms and applications, SIAM, Philadelphia, 1993. [17] D. C. Robertson, O. I. Camps, J. S. Mayer, andW. B. Gish, Sr., “Wavelets and electromagnetic power system transients,” IEEE Trans. Power Del., vol. 11, no. 2, pp. 1050–1056, Apr. 1996. [18] A.J. Mazon, I. Zamora, J. Gracia, K.J. Sagastabeitia, and J.R. Saenz, “Selecting ANN structures to find transmission faults,” IEEE Computer Applications in Power, pp. 44–48, Jul. 2001. [19] Y. Y. Hong and B. Y. Chen, “Locating switched capacitor using wavelet transform and hybrid principal component analysis network,” IEEE Trans. Power Del., vol. 22, no. 2, pp. 1145-1152, 2007.
21
[20] H. Guo, L. B. Jack, and A. K. Nandi, “Feature generation using genetic programming with application to fault classification,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 35, no. 1, pp. 89-99, Feb. 2005. [21] H. H. Chang, “Genetic algorithms and non-intrusive energy management system based economic dispatch for cogeneration units,” Energy, vol. 36, pp. 181-190, 2011. [22] Fuse selection guide, OptiFuse - Rev A 01/2010. [23] Miniature circuit breaker-Application guide, Low Voltage Products & Systems, ABB Inc., Apr. 2009.
22
Figure Captions
Fig. 1. Fault identification system on load-bus faults for a NIFM system. Fig. 2. Block diagram of a fault event identification system. Fig. 3. SLGF event with different occurrence times for BUS 1. (a) Different times of fault current occurrence (0.05s, 0.1s, and 0.15s), (b) Power distribution diagram. Fig. 4. SLGF event with different current amplitude magnitudes for BUS1. (a) Different amplitude magnitudes of fault occurrence (-20%, 0%, and +20%), (b) Power distribution diagram. Fig. 5. SLGF event with different fault resistances for BUS1. (a) Different fault resistances of fault occurrence (0pu, 0.002pu, and 0.02pu), (b) Power distribution diagram. Fig. 6. Distributions of each scale for five different load buses. (a) Voltage. (b) Current. Fig. 7. Fault identification system on transmission-line faults for a NIFM system. Fig. 8. Current distributions of each order for the SLGF on five different load buses. Fig.9.
Turn-on real-power transient waveforms for the SLGF on five different load buses.
23
Fig. 1. Fault identification system on load-bus faults for a NIFM system.
24
Fig. 2. Block diagram of a fault event identification system.
25
(a)
(b) Fig. 3. SLGF event with different occurrence times for BUS 1. (a) Different times of fault current occurrence (0.05s, 0.1s, and 0.15s), (b) Power distribution diagram.
26
(a)
(b) Fig. 4. SLGF event with different current amplitude magnitudes for BUS1. (a) Different amplitude magnitudes of fault occurrence (-20%, 0%, and +20%), (b) Power distribution diagram.
27
(a)
(b) Fig. 5. SLGF event with different fault resistances for BUS1. (a) Different fault resistances of fault occurrence (0pu, 0.002pu, and 0.02pu), (b) Power distribution diagram.
28
(a)
(b) Fig. 6. Distributions of each scale for five different load buses. (a) Voltage. (b) Current.
29
Fig. 7. Fault identification system on transmission-line faults for a NIFM system.
30
Fig. 8. Current distributions of each order for the SLGF on five different load buses.
Fig. 9. Turn-on real-power transient waveforms for the SLGF on five different load buses.
31
Table 1 The results of fault identification in case study 1 Fault Type
SLGF
DLGF
LLF
LLLGF
Methods Monitoring on Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s)
Ia
Pspectrum Ib Ic
Average
Ia
Z-score based WT Ib Ic
UT Average
Phase A
Phase B
Phase C
Average
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
95
95.63
97.19
95.94
100
96.88
100
98.96
0.6129 0.0039
0.5688 0.0039
0.5725 0.0040
0.5847 0.0039
0.4996 0.0040
0.5023 0.0040
0.5403 0.0041
0.5141 0.0040
0.5693 0.0050
0.7449 0.0041
0.7638 0.0044
0.6927 0.0045
100
100
100
100
100
100
100
100
100
94.55
100
98.18
100
100
100
100
74.06
35.94
86.88
65.63
100
82.50
100
94.17
0.6662 0.0039
0.5759 0.0039
0.6719 0.0040
0.6380 0.0039
0.5429 0.0042
0.7056 0.0043
0.6474 0.0042
0.6320 0.0042
0.5615 0.0042
25.249 0.0040
0.6014 0.0041
8.8040 0.0041
100
100
100
100
100
100
100
100
95.45
100
98.79
98.08
100
97.82
100
99.27
55
18.44
98.13
57.19
79.38
100
90
89.79
0.5756 0.0062
0.6015 0.0040
0.5803 0.0040
0.5858 0.0047
0.5975 0.0042
0.6050 0.0042
0.6909 0.0042
0.6311 0.0042
29.005 0.0042
0.5334 0.0041
9.0808 0.0041
12.873 0.0041
100
100
100
100
100
100
100
100
100
100
100
100
100
100
99.38
99.79
93.44
30.63
93.13
72.40
100
100
100
100
0.5990 0.0040
0.5876 0.0041
0.9930 0.0040
0.7265 0.0040
0.5688 0.0041
0.5565 0.0042
0.5812 0.0041
0.5688 0.0041
0.5316 0.0043
0.5226 0.0042
0.5242 0.0042
0.5261 0.0042
32
Table 2 The results of fault identification in case study 2 Fault Type
SLGF
DLGF
LLF
LLLGF
Methods Monitoring on Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in Test (%) Training Time (s) Test Time (s) Recognition Accuracy in Training (%) Recognition Accuracy in
Ia
Pspectrum Ib Ic
Average
Ia
Z-score based WT Ib Ic
UT Average
Phase A
Phase B
Phase C
Average
100
100
100
100
100
93.94
93.94
95.96
76.36
89.09
100
88.49
88.13
75.94
75.63
79.90
4.375
18.44
32.81
18.54
59.06
64.38
93.13
72.19
0.6250 0.0040
0.6291 0.0040
0.6659 0.0041
0.6400 0.0040
1.4026 0.0043
60.330 0.0043
41.610 0.0043
34.448 0.0043
46.788 0.0042
41.926 0.0043
2.1864 0.0041
30.300 0.0042
100
100
100
100
100
100
100
100
45.15
51.52
83.03
59.90
99.38
97.81
97.19
98.13
36.25
62.19
65.31
54.58
22.81
33.13
59.69
38.54
0.6291 0.0041
0.6466 0.0042
0.6997 0.0043
0.6585 0.0042
0.5630 0.0042
0.6179 0.0043
0.6048 0.0042
0.5952 0.0042
40.870 0.0042
40.028 0,0041
43.826 0.0048
41.575 0.0045
100
100
100
100
100
100
100
100
63.94
46.67
68.79
59.80
90.63
93.44
71.25
85.11
27.50
45.31
35.00
35.94
41.25
28.13
48.75
39.38
0.6977 0.0042
0.6713 0.0041
0.7978 0.0042
0.7223 0.0042
0.7654 0.0041
0.5642 0.0041
0.7871 0.0042
0.7056 0.0041
42.470 0.0040
40.610 0.0041
43.330 0.0043
42.137 0.0041
100
100
100
100
100
100
100
100
60.00
73.08
75.00
69.36
72.31
88.85
93.85
85.00
17.69
12.69
52.69
27.69
44.23
48.08
53.08
48.46
33
Test (%) Training Time (s) Test Time (s)
0.5525 0.0041
0.5450 0.0042
0.5124 0.0041
0.5366 0.0041
0.7326 0.0041
0.78681 0.0041
34
0.4879 0.0040
0.6691 0.0041
36.410 0.0041
39.946 0.0042
40.566 0.0042
38.974 0.0042