Fault diagnosis system for tapped power transmission lines

Electric Power Systems Research 80 (2010) 599–613

Contents lists available at ScienceDirect

Electric Power Systems Research journal homepage: www.elsevier.com/locate/epsr

Fault diagnosis system for tapped power transmission lines E.A. Mohamed a,∗,1 , H.A. Talaat a,2 , E.A. Khamis b a b

Elect. Power & Machines Dept., Ain Shams Univ., Cairo, Egypt EEA, Nasr City, Cairo, Egypt

a r t i c l e

i n f o

Article history: Received 28 July 2008 Received in revised form 10 April 2009 Accepted 25 October 2009 Available online 5 December 2009 Keywords: Tapped transmission lines Fault diagnosis ANN Protective relaying

a b s t r a c t This paper presents a design for a fault diagnosis system (FDS) for tapped HV/EHV power transmission lines. These lines have two different protection zones. The proposed approach reduces the cost and the complexity of the FDS for these types of lines. The FDS consists basically of fifteen artificial neural networks (ANNs). The FDS basic objectives are mainly: (1) the detection of the system fault; (2) the localization of the faulted zone; (3) the classification of the fault type; and finally (4) the identification of the faulted phase. This FDS is structured in a three hierarchical levels. In the first level, a preprocessing unit to the input data is performed. An ANN, in the second level, is designed in order to detect and zone localize the line faults. In the third level, two zone diagnosis systems (ZDS) are designed. Each ZDS is dedicated to one zone and consists of seven parallel-cascaded ANN’s. Four-parallel ANN’s are designed in order to achieve the fault type classification. While, the other three cascaded ANN’s are designed mainly for the selection of the faulted phase. A smoothing unit is also configured to smooth out the output response of the proposed FDS. The proposed FDS is designed and evaluated using the local measurements of the three-phase voltage and current samples acquired at only one side. The sampling rate was taken 16 samples per cycle of the power frequency. Data window of 4 samples was utilized. These samples were generated using the EMTP simulation program, applied to the High-Dam/Cairo 500 kV tapped transmission line. All possible shunt fault types were considered. The effect of fault location and fault incipience time were also included. Moreover, the effect of load and capacitor switchings on the FDS performance was investigated. Testing results have proved the capability as well as the effectiveness of the proposed FDS. © 2009 Elsevier B.V. All rights reserved.

1. Introduction Protective relaying is one of the basic and necessary elements of an electric power system. The protective relaying role is to cause the prompt removal from service of any element, when it suffers a short circuit or when it starts to operate in an abnormal manner that may cause damage to the power system. In fact, in power systems, most of the faults occur on the transmission lines. Faults MVA levels are usually high, and if they are not cleared rapidly they may cause system instability as well as damage and hazards to equipment and persons. Hence, the proper diagnosis or classification of line faults is essential to the appropriate operation of power systems. Therefore, the fault type classification is a crucial protective relaying feature due to its significant effect on the operation enhancement of relaying scheme. The correct operation of

∗ Corresponding author. Tel.: +20 2 26399371; fax: +20 2 26826636. E-mail address: [email protected] (E.A. Mohamed). 1 Currently with Qassim University, Saudi Arabia. 2 Currently with King Saud University, Saudi Arabia. 0378-7796/$ – see front matter © 2009 Elsevier B.V. All rights reserved. doi:10.1016/j.epsr.2009.10.030

major protective relays depends mainly on the fault classification feature [1,2]. On the other hand, the faulted phase selection is as important as the fault detection. It would lead to an increase in the system stability and availability by allowing the healthy phases to operate using the single pole (or phase) tripping [3]. Single pole tripping has many benefits such as improving the transient stability and reliability of the power system, reducing the switching over-voltages and mitigating of the shaft torsional oscillations of large thermal units [3]. Fig. 1 shows the essential structural modules of a modern protective relay, where the fault diagnosis modules are considered very important ones. Tapped transmission lines are those lines that are tapped usually through transformer bank primarily to supply loads. They are usually economical in their breakers requirements, but they need a complex relaying scheme for adequate protection and operation. These lines have two different protection zones, and faults must be detected and isolated in each zone individually. Simply, two different FDS’s can be designed in order to achieve the standard requirements. There are few research studies that have been carried out on these types of lines, due to the need of a very difficult

600

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

Fig. 1. Essential modules of a protective relay.

protection scheme [1]. This study is trying to explore the design details of a proposed fault diagnosis module [4], based on the application of ANN technology, as a part of the protection scheme required for these types of lines. The conventional analytical based classification approaches are expected to be affected by the system operating conditions. Also, a complete faulted phase selection cannot be achieved through

these approaches. Moreover, these approaches are also time consuming [1–4]. On the other hand, artificial intelligence techniques (expert systems, pattern recognition, ANN, and fuzzy logic) in general and ANN in particular provides a very interesting and valuable alternative [4]. ANN [5–7] can efficiently deal with most situations, which are not defined sufficiently for deterministic algorithms to exe-

Fig. 2. Architecture of the suggested FDS.

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

601

Table 1 ANN’s role, output levels and its triggering network. Designed ANN

Fault type

Faulted zone

Triggering ANN

Output levels High

Medium

Low

ANN#1

All fault types

Zone#1 or zone#2

–

Fault on zone#1

Fault on zone#2

Normal

Fault type classification ANN11 (or ANN21) ANN12 (or ANN22) ANN13 (or ANN23) ANN14 (or ANN24)

L-G L-L 2L-G 3L-G

Zone#1 (or zone#2) Zone#1 (or zone#2) Zone#1 (or zone#2) Zone#1 (or zone#2)

ANN#1 ANN#1 ANN#1 ANN#1

L-G L-L 2L-G 3L-G

– – – –

Normal or other fault types Normal or other fault types Normal or other fault types Normal or other fault types

Faulted phase selection ANN111 (or ANN211) ANN121 (or ANN221) ANN131 (or ANN231)

L-G L-L 2L-G

Zone#1 (or zone#2) Zone#1 (or zone#2) Zone#1 (or zone#2)

ANN11 (or ANN21) ANN12 (or ANN22) ANN13 (or ANN23)

R-G R-S RS-G

S-G S-T ST-G

T-G R-T RT-G

Fig. 3. System under study.

cute. ANN can also accurately handle highly nonlinear tasks. Furthermore, ANN’s are paralleled data-processing tools, capable of learning functional dependencies of data. ANN’s are robust with respect to incorrect or missing data. Protective relaying based on ANN is not affected by a change in the system operating conditions. ANN also has fast computation rates, large input error tolerance and adaptive capabilities. Many ANN based approaches have been developed in different applications in the power industry [8–13]. On the other hand, many TL fault diagnosis or classification applications based on ANN technique have been introduced [14–25]. For example, an approach [14] of two modules was developed: (a) the first for the fault type classification and phase selection and (b) the second for the classification of arcing and non-arcing faults. Samples of the threephase voltages and currents with 1 kHz sampling rate were used. In the first, one ANN (30 × 20 × 15 × 11) of 30 input nodes, two hidden layers of 20 and 15 neurons, and 11 neurons in the output layer, was designed and a classification time in the range of 5–7 ms was achieved. In the second, three ANN’s (20 × 15 × 10 × 1)

Fig. 4. Steps of ANN design.

Fig. 5. Adopted ANN structure.

602

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

Fig. 6. Three phase voltages and currents at FDS and corresponding ANN outputs due to S-G fault at HD500. Table 2 Faulted phase selection on zone#1. ANN no.

ANN111

ANN121

ANN131

Fault location Phase Target

HDNH SG 0.5

HDNH RG 0.1

HDNH TG 0.9

HDNH RS 0.1

HD500 ST 0.5

NH500 RT 0.9

HDNH RSG 0.7

HD500 RTG 0.9

NH500 STG 0.5

ANN output

0.38060 0.45535 0.49774 0.53031 0.49129 0.45884 0.48467 0.48370 0.49652 0.46369 0.47267 0.48868

0.09675 0.06084 0.08480 0.05153 0.07699 0.05311 0.08512 0.05962 0.07766 0.05372 0.08478 0.06106

0.90307 0.90559 0.90907 0.92546 0.90534 0.92398 0.90611 0.91464 0.90381 0.92335 0.90542 0.91042

0.1164 0.0954 0.0772 0.0768 0.0990 0.0916 0.0922 0.0845 0.0931 0.0942 0.0976 0.0887

0.5703 0.4130 0.4422 0.5476 0.5230 0.4976 0.4878 0.5038 0.5063 0.4985 0.4965 0.4967

0.8728 0.7703 0.9128 0.8107 0.9289 0.0515 0.9188 0.7934 0.9238 0.8448 0.9222 0.7909

0.1352 0.1038 0.0900 0.0615 0.0849 0.0881 0.0940 0.1009 0.0836 0.0896 0.1053 0.1072

0.9341 0.6434 0.8993 0.8949 0.9258 0.6822 0.8977 0.8290 0.9276 0.6964 0.8987 0.8078

0.2825 0.5402 0.5793 0.5446 0.4905 0.5044 0.5069 0.5217 0.5063 0.5036 0.4865 0.509

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

603

Fig. 7. Three phase voltages and currents at FDS and corresponding ANN outputs due to S-G fault at AS500.

were designed and a classification time of 25–70 ms was reached. Also, different hybrid TL fault diagnosis techniques have been implemented. In [26,27], a fuzzy/ANN hybrid classification approach is provided, a radial basis function (RBF) based ANN is employed for the TL classification system [28], and a fuzzy/wavelet hybrid classification system for power system TL relaying is introduced [29,30]. In [31] a wavelet transform technique is utilized for the protection of parallel lines. The modular ANN is used for the design of the TL directional protection [32]. In [33,34] a hybrid of fuzzy/ANN technique is implemented for the design of the classification scheme of power transmission system. A wavelet/ANN technique is also introduced for the design of the TL boundary protection system [35]. The purpose of this study is to design a fault diagnosis system (FDS) using ANN capabilities, for the HV/EHV power tapped lines. The conventional protection scheme for these types of lines demands less requirement of breakers and higher complexity of

relaying devices [1–3]. Simply, these types of lines have two protection zones. Therefore, two different FDS’s are required, in order to detect and isolate the fault in each zone individually. The proposed FDS is a one device scheme, designed to detect the fault, localize the faulted zone, classify the fault type and finally identify the faulted phase. Thus, the proposed FDS based relaying and protection scheme is less in cost and requirements. It consists basically of three hierarchical levels. It employs the modular ANN concept [32]. It has an output smoothing unit. Moreover, it is designed based on the simulation of the tapped power lines using the EMTP program. Samples of the three phase voltages and currents at only the sending end are used as the input measurements. The HighDam/Cairo 500 kV tapped transmission line is the tested power line in this study. All possible fault types are considered, the fault location and the fault incipience time are also implemented. Also, the effect of load and capacitor switchings on the FDS performance is evaluated. Testing results have proved the capability as well as the effectiveness of the proposed FDS.

604

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

2. FDS design procedure An important module of the modern protective relay is the fault diagnosis module. The different functions required to be performed by the proposed FDS are: the fault detection, the faulted zone localization, the fault type classification and finally the faulted phase selection. Since the design procedure is a multi-task problem, it is preferable to decompose it into individual sub-problems, where each sub-problem is dedicated to one task. Using one ANN per task (modular ANN [32]) make it more powerful and will increase the learnability power of the ANN. Therefore, the proposed FDS consists of modular ANN, multi-ANN that are arranged to be working in parallel, which makes the proposed FDS more fast, efficient, robust, accurate and reliable. The proposed FDS consists of 15 modular ANN’s constructed in three hierarchical levels as illustrated in Fig. 2. The first level is a preprocessing stage involving data filtering, data conversion, and data normalization. In the second level, one ANN (ANN#1) designed to detect the faulted zone. It has a three output levels: low level

(<0.3) for normal condition, medium level (0.3–0.7 inclusive) for zone#2 faults, and high level (>0.7) for zone#1 faults (nearest to the location of FDS). The third level contains two parallel zone diagnosis systems (ZDS1 and ZDS2). Each ZDS consists of 7 modular ANN’s, designed to classify the fault type as well as to select the faulted phase in its zone. ZDS1 is the fault diagnosis system for zone#1. It consists of four parallel ANN’s, each ANN of the first three, is cascaded with another ANN. The inputs to all ANN’s are derived from the voltage and current samples acquired locally at the bus where the FDS is connected. The function of each ANN can be described as follows:

(a) ANN11: The task of this ANN is to identify the phase to ground fault on zone#1. It has two output levels: high level (>0.5) for phase to ground fault and low level (≤0.5) for normal condition and other fault types. (b) ANN12: This ANN is intended to classify the phase-to-phase fault on zone#1. It has two output levels: high level (>0.5) for

Fig. 8. Three phase voltages and currents at FDS and corresponding ANN outputs due to S-T fault at NH500.

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

phase to phase fault and low level (≤0.5) for normal case and other fault types. (c) ANN13: This ANN is designed to recognize the double phase to ground fault at zone#1. Its output is high level (>0.5) for double phase to ground fault and low level (≤0.5) for normal case and other fault types. (d) ANN14: This ANN is used to identify three phase to ground faults at zone#1. Its output is high level (>0.5) for three phases to ground fault and low level (≤0.5) for else. (e) ANN111, ANN121, ANN131: These modular networks are designed to identify the faulted phase selection in case of single phase to ground, phase to phase, and double phase to ground faults respectively. The output of each ANN has three levels according to the faulted phases: low level (<0.3), medium level (0.3–0.7 inclusive) and high level (>0.7).

605

three output levels: high for faults on zone#1, medium for faults on zone#2, and low for normal conditions. Also, ANN11 (or ANN21) is designed, for the classification of fault type, with two output levels: high for L-G on zone#1 (or zone#2) and low for normal or other fault type conditions. These two networks are also triggered by ANN#1 in case of fault condition. Moreover, ANN111 (or ANN211) is designed, for the selection of faulted phase, with three output levels: high for R-G fault on zone#1 (or zone#2), medium for S-G fault on zone#1 (or zone#2), and low for T-G fault on zone#1 (or zone#2). This network is triggered by ANN11 (or ANN21) in case of a L-G fault. Similarly, the rest of similar networks are following a similar procedure. 3. Numerical application results 3.1. System under study

Similarly, ZDS2 is the zone fault diagnosis system for zone#2. Moreover, Table 1 summarizes the role of design for each modular ANN, its output levels and its triggering mechanism. For example, ANN#1 is designed, for the detection of faulted zone, with

The present study is concerned with the protection of the HighDam/Cairo 500 kV tapped transmission system. The section from HD500 to NH500 is taken as zone#1 and the length of this section

Fig. 9. Three phase voltages and currents at FDS and corresponding ANN outputs due to S-T fault at NHAS.

606

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

is 235 km. Zone#2 is taken as the section from NH500 to AS500 and the length of this section is 185 km. This system is selected as an application example to design and evaluate the performance of the proposed FDS. Fig. 3 shows the system arrangement and the location of the FDS as well. The power system data and operating conditions are given in [11] and Appendix A. 3.2. Data samples The simulation of the selected power system is carried out using the electromagnetic transient program (EMTP) [36]. The simulation is used to generate the fault patterns needed for the ANN design and evaluation processes. Different fault types (R-G, S-G, T-G, RS, S-T, R-T, RS-G, ST-G, RT-G, and RST-G) are considered. For each fault type, the fault locations are selected at each bus and on the midpoints, as well as the fault incipience time is taken at the zero

crossing and the peak time Number of fault conditions = 10 fault types × 5 fault locations × 2 fault incipience times

Therefore, there are 100 case studies. Each case study contains three-cycles during the fault and three-cycles after the fault clearance. In addition, one-cycle before the fault is taken into consideration for only two case studies, one for zone#1 and the other for zone#2. Accordingly, the number of samples is (2 cases × 7 cycles × 16 samples/cycle + 98 cases × 6 cycles × 16 samples/cycle = 9632 samples). The generated samples are then divided into two sets; the design (training) set which is composed of 6752 samples (70 case studies), and the evaluation (testing) set is composed of 2880 samples (30 different case studies).

Fig. 10. Three phase voltages and currents at FDS and corresponding ANN outputs due to RS-G fault at HDNH.

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

3.3. ANN design procedure Case studies are generated using the EMTP, and then are loaded into the MATLAB software. The voltages and currents are normalized and reshaped in the form of group of patterns, where each pattern is composed of a predefined number of consecutive samples. Different data windows are examined (one, two, or four samples). The Neural Network-Toolbox of the MATLAB software [37] is used to design each modular ANN, that is assigned in the proposed design scheme of the FDS. Fig. 4 shows the different steps required for the design procedure for each modular ANN (see Appendix B). 3.4. ANN adopted structure There is no definite way for pre-determining the optimum network structure without testing different configurations. The three layer feed-forward ANN was found satisfactory for pattern classification problems [14–19]. Therefore, it is chosen for this application. The tan-sigmoid and the log-sigmoid activation functions are differentiable as well as monotonic functions; therefore, they are employed as the network activation functions. Thus, the network structure will be one hidden layer with the tan-sigmoid neurons followed by an output layer with the log-sigmoid neurons.

607

The number of input variables and number of neurons in the hidden layer are decided using the experimentation process that involves different network configurations. The process will be terminated when a suitable network structure is achieving a satisfactory performance. Each ANN is designed and evaluated using the three-phase voltage and current samples with the sampling rate of 16 samples per cycle. Different data windows (number of samples per pattern) one sample, two samples, one-RMS sample, two-RMS samples and four consecutive samples are examined. The best FDS performance was obtained using a data window of four consecutive samples of the three-phase voltages and currents. Based on the above, the appropriate structure for each modular ANN used in this application is: 24 input nodes, one hidden layer with 24 tan-sigmoid neurons, an output layer with one log-sigmoid neuron. Fig. 5 shows the adopted structure for each ANN. 4. FDS performance evaluation 4.1. Localization of faulted zone and classification of fault type After completing the design phase, the FDS is evaluated at different fault conditions, i.e. at different fault locations, different fault types and different fault incipience time (i.e. 30 case studies). These cases are different from those used for the design.

Fig. 11. Three phase voltages and currents at FDS and corresponding ANN outputs due to ST-G fault at NHAS.

608

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

(1) Phase to ground faults: Fig. 6 shows the three-phase voltages and currents, due to S-G fault at HD500, ANN#1-output and the outputs of other ANN’s (ANN11, ANN12, ANN13, and ANN14), which are responsible for classifying the fault type at zone#1. As shown in the figure, when the fault is triggered the voltage of phase S collapses and its corresponding phase current increases. The fault is then cleared after 3 cycles. ANN#1-output flags at a high level ≈0.9 (as expected for a fault on zone#1) during the fault duration and falls to a low level ≈0.1 when the fault is cleared after a duration of other 3 cycles. This proves that the network has localized the faulted zone accurately. The outputs of other ANN’s are at low level ≈0.1 except ANN11, which is designed for detecting L-G faults. The output of ANN11 is almost at high level ≈0.9 during the fault duration and falls to a low level ≈0.1 when the fault is cleared. Therefore, the performance of the FDS in this case is superior. It must be noted here that all ANN outputs are

smoothed out using the output unit which performs an average over each 4 samples. Therefore, each ANN output is 24 response points (pattern index). Fig. 7 shows a distorted voltage waveform and the current of phase S increasing due to S-G fault at AS500. ANN#1-output is at a medium level ≈0.5 (as expected for fault on zone#2) during the fault duration then falls to a low level ≈0.1 when the fault is cleared. This proves that the network detection capability of the faulted zone is accurate. The output of ANN21, designed for classifying L-G faults on zone#2, is almost at a high level ≈0.9 during the fault and at a low level ≈0.1 after clearing the fault. The outputs of other three ANN’s (ANN22, ANN23 and ANN24) are at a low level ≈0.1 all over the fault and fault cleared periods, as it is expected. (2) Phase to phase faults Fig. 8 shows the output of ANN#1 at almost a high level ≈0.9 (as expected for fault on zone#1) during the S-T fault at

Fig. 12. Three phase voltages and currents at FDS and corresponding ANN output due to 3L-G fault at HDNH.

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

NH500, then falls to a low level ≈0.1 after clearing the fault. The output of ANN12 is at a high ≈0.9 and low ≈0.1 levels during and after the fault, respectively. The outputs of the other three ANN’s are all at a low level ≈0.1 all over the period, as it should be. Fig. 9 shows the output response, due to S-T fault at NHAS, of ANN#1 as well as the outputs of the other ANN’s (ANN21, ANN22, ANN23 and ANN24), which are responsible for classifying the fault type on zone#2. ANN#1 output is at a medium level ≈0.5 during fault then falls to a low level ≈0.1 after the fault is cleared. The output of other ANN’s are all at a low level ≈0.1 except ANN22 that is designed to detect L-L faults, whose output is at a high level ≈0.9 only during fault. (3) Double phase to ground faults Fig. 10 shows the FDS response due to RS-G fault at HDNH, where ANN#1-output responds a high level ≈0.9 during fault and then falls to a low level ≈0.1 after clearing fault. The ANN13 response is at high level ≈0.9 during fault and low level ≈0.1 after clearing the fault. Other three ANN’s outputs are at low levels ≈0.1 all over the period. Similarly, Fig. 11 shows the FDS response due to ST-G fault at NHAS. The output of ANN#1 is almost at a medium level ≈0.5 during fault and falls to a low level ≈0.1 after the fault is

609

cleared. The outputs of other ANN’s are also at low level ≈0.1 except ANN23, which is designed to detect 2L-G faults, is at high level ≈0.9 during fault. (4) Three phase to ground faults Fig. 12 presents, the three-phase voltages and currents due to 3L-G fault at HDNH, the output of ANN#1 and the outputs of other ANNs. ANN#1 output is at a high level ≈0.9 during fault and falls to a low level ≈0.1 after clearing the fault. The outputs of other ANN’s are all at low level ≈0.1 except ANN14, which is designed to detect the 3L-G faults, is at high level ≈0.9 during fault. Similarly, Fig. 13 shows, due to 3L-G fault at AS500, the output of ANN#1 and the outputs of other ANN’s (which are responsible for classifying the fault type at zone#2). ANN#1 output is at medium level ≈0.5 during fault and falls to a low level ≈0.1 after clearing the fault. The outputs of other ANN’s are all at low level ≈0.1 except ANN24, designed to detect the 3L-G faults, is almost at high level ≈0.9 during fault. 4.2. Phase selection The testing results of the faulted phase selection on zone#1 are summarized in Table 2, while those for zone#2 are summarized

Fig. 13. Three phase voltages and currents at FDS and corresponding ANN outputs due to 3L-G fault at AS500.

610

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

Fig. 15. Three phase voltages and currents at FDS and corresponding ANN outputs due to load switching at NH500. Fig. 14. Three phase voltages and currents at FDS and corresponding ANN outputs due to capacitor switching at NH500.

4.3. Effect of load and capacitor switching in Table 3. From these results, it can be seen that for each ANN of the three specified ANN’s per zone (e.g. ANN111, ANN121, and ANN131 for zone#1), the response during fault (12 data points) matches well with the corresponding target value. This means that the ANN response is fairly satisfactory regarding the phase selection for each fault. Therefore, it can be stated that the FDS phase selection capability is very promising.

Load or capacitance switching may generate transients in the power system, which can be misclassified as fault case. Hence, the performance of the proposed FDS is evaluated using these two cases of switchings. Fig. 14 shows the three-phase voltages and currents and the corresponding output response of ANN#1, due to the switching of a capacitor at NH500 bus. On the other hand, Fig. 15 shows the three-phase voltages and currents and the associated output response for ANN#1, due to the switching

Table 3 Faulted phase selection on zone#2. ANN no.

ANN211

Fault location Phase Target

NHAS RG 0.1

NHAS SG 0.5

AS500 TG 0.9

ANN221 NHAS RS 0.1

NHAS ST 0.5

AS500 RT 0.9

ANN231 NHAS RSG 0.1

AS500 RTG 0.9

NHAS STG 0.5

ANN output

0.1692 0.0952 0.1041 0.0820 0.1077 0.0816 0.1003 0.0853 0.1052 0.0803 0.0998 0.0868

0.2552 0.3409 0.6302 0.5561 0.4498 0.4950 0.5080 0.5004 0.4852 0.5072 0.4942 0.4909

0.9121 0.8233 0.8851 0.8975 0.8816 0.8948 0.8585 0.8893 0.8782 0.8980 0.8523 0.8844

0.3374 0.2192 0.1241 0.1677 0.1713 0.1740 0.1466 0.2036 0.1716 0.1770 0.1624 0.2180

0.4920 0.4018 0.5543 0.5674 0.4996 0.4497 0.5218 0.5227 0.4913 0.4629 0.4995 0.5032

0.8771 0.8054 0.9285 0.9125 0.9186 0.8270 0.9086 0.8794 0.9170 0.8234 0.9029 0.8594

0.0698 0.0932 0.1638 0.0788 0.1122 0.0977 0.1022 0.0716 0.0837 0.0963 0.0977 0.0727

0.8195 0.8922 0.9305 0.8467 0.9224 0.8859 0.9053 0.8660 0.9204 0.8852 0.8894 0.8688

0.5159 0.3516 0.2424 0.6064 0.5118 0.4409 0.3792 0.5295 0.5141 0.4679 0.4762 0.5067

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

of an inductive load at NH500 bus. As shown from these figures, the voltage and current waveforms are affected but the output of ANN#1 is still low all over the period indicating a normal condition. Therefore, the proposed FDS is insensitive to the load and capacitor switchings. 5. Conclusions A modular ANN based approach was proposed to design a fault diagnosis system (FDS), as a main function of a high speed protective relaying scheme for the HV/EHV tapped transmission lines. The FDS main features are: identify the faulted zone, classify the type of fault and select the faulted phase. The proposed FDS consists of three hierarchical levels. The first level is a preprocessing stage, where the input measurements are processed via filtration, conversion, and normalization. The second level consists of one ANN, which is responsible for detecting the faulted zone. In the third level there are two zone diagnosis systems. Each system consists of seven ANN’s and is responsible for the classification of fault type as well as the selection of faulted phase in its zone. The investigation of the performance of the proposed FDS under various fault conditions leads to the following conclusions: 1. The proposed FDS reduces the cost and complexity of the relaying scheme for the tapped lines. 2. The proposed architecture of the proposed FDS is constructed from 15 modular ANN’s, with the advantage of assigning one task to each ANN, to enhance its learning ability and leading to a high quality performance. 3. The adequate length of the used data window is 1/4 cycle (4 samples of three phase voltages and currents at a sampling rate of 16 samples per cycle). 4. The time response of the proposed FDS is very fast due to its parallel structure. The expected time delay of the decision is about 1/3 cycle. 5. The selected design for each modular ANN is suitable (24 input nodes, 24 hidden-neurons and one output-neuron). 6. The accuracy in the training phase was perfect (100%), irrespective to fault location, fault type, and fault incipience time, while that of the testing phase is in the range 92–100%. 7. The proposed FDS has proved high capability in classifying the transients produced from the capacitor and load switchings as normal cases. 8. The FDS can be used as a part of new generation of ultra-highspeed protective relaying schemes. Appendix A. Power system simulation using EMTP A.1. General The electro-magnetic transient program (EMTP) is used for simulating the transients in power system elements. The power system considered here for this study is the Upper Egypt Power System (UEPS). It consists of generating stations, substation transformers, power lines, and loads. The power line, starting from the High-Dam 500 kV “HD500” generating station and ending at Cairo 500 kV “CA500” substation, is a double circuit line each of length 788 km. The line passes through Nagh-Hammady substation (NH500) at distance 235 km from the High-Dam and reaches Samalut substation (SA500) at distance of 343 km from NH500 and 209 km from CA500. The main transformer substation is the HighDam generating-station substation, it consists of 12 units, each of 15.75/500 kV, 206 MVA capacity. On the other hand, the High-Dam generating-station contains twelve generators each of 175 MW capacities.

611

A.2. EMTP digital simulation The EMTP is a well known world-wide program, used to solve the equations describing the system under transient conditions. It consists of a number of functional modules, which separately contains mathematical models of various components. The major functional modules are transformers, transmission lines, switches, surge arresters, control systems, and electric machinery. The synchronous machine is simulated using the EMTP developing transient module. All overhead transmission lines are represented by their ␲ equivalent network with their series R-L and shunt G-C lumped parameters as a suitable representation for the phenomenon under study. All loads are simulated by their equivalent R-L-C parameters as constant impedances. Each three-phase transformer is simulated by three single-phase transformers, each of them is represented by its transient model. The EMTP can be used to represent an infinite bus with a constant voltage, and constant frequency. A.3. System data The system data and the operating conditions are given as follows: Generator data: 

Xd 0.4



Xd 0.37



Xd 1.2



Xq 0.36



Xq 0.16

Xq 0.79



d 2.8

H 6

D 0.18



where Xd , Xd and Xd are the d-axis subtransient, transient and   synchronous reactance’s respectively. Xq , Xq and Xq are the q-axis subtransient, transient and synchronous reactance’s respectively.   d is the d-axis transient time constant (s); H is the inertia time constant (s); D is the damping factor. Transformer data: 206 MVA, 15.75 /500Y kV, X(LV) = 0.006 , X(HV) = 1.67 . Transmission lines data: #

From bus

To bus

R ()

X ()

B (␮s)

1 2 3 4

HD500 NH500 AS500 SA500

NH500 AS500 SA500 CA500

2.56 2.00 1.75 2.17

34.80 27.28 23.75 29.50

460.2 360.8 308.1 407.6

Operating conditions: Bus

V (kV)

Load (MVA)

HD500 NH500 AS500 SA500 CA500

523.38 ∠22.9 507.70 ∠11.8 505.34 ∠6.5 496.70 ∠3.2 475.00 ∠0

10.32 + j 103.32 198.30 + j 132.17 85.12 + j 42.56 45.44 + j 40.89 157.43 + j 107.6

Appendix B. ANN design algorithm The back-propagation learning rules are used to adjust the weights and biases of networks so as to minimize the sum squared error of network. This is done by continually changing the values of the network weights and biases in the direction of steepest descent with respect to error. This is called a gradient descent procedure. Changes in each weight and bias are proportional to that element’s effect on the sum-squared error of the network. Typically, a new input will lead to an output similar to the correct output for input vectors used in training that are similar to the new input being presented. This generalization property makes it possible to train a network on a representative set of input/target pairs and get good results for new inputs without training the network on all possible

612

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

Fig. B1. The commonly used activation functions and corresponding MATLAB functions.

input/output pairs. A summary of steps for the back-propagation technique may be as follows: 1. Initialize the network weights and biases by small random elements. 2. Present a data pattern from the training set with input and desired output/target pairs. 3. The actual outputs will be calculated by the log/tan sigmoid neurons. The log-sigmoid is a squashing activation function f(x)

that maps the input to the (0, 1) interval, while the tan-sigmoid is also a squashing activation function f(x) that maps the input to the (−1, 1) interval, where (see Fig. B1). (log-sig f (x) = 1/(1 + exp(−x)) and tan -sig f (x) = (1 − exp(−2x))/(1 + exp(−2x))) 4. Calculate the error, the difference between target and actual output, to be minimized.

Fig. B2. The ANN design program flow chart.

E.A. Mohamed et al. / Electric Power Systems Research 80 (2010) 599–613

5. Derivatives of error (called delta vectors) are calculated for the output layer and then back propagated through the network until delta vectors are available to the hidden layers. For the sigmoid transfer function ıj = (tj − oj )oj (1 − oj )



ıj = oj (1 − oj )

wij ıj

for node j on the output layer for node j on the hidden layer

(1) (2)

i

6. The weights and biases changes are calculated recursively backwards from the output layer towards the input layer. Once all changes are calculated the weights and biases are updated as follows: wji (k + 1) = wji (k) + wji (k)

(3)

bj (k + 1) = bj (k) + bj (k)

(4)

wji (k) = ıj &oi bj (k) = ıj

(5)

where wji is the weight change from a node in ith layer to a node in jth layer, wji is the weight from layer i to layer j, bj is the baises for the layer j, ıj is the delta vector of jth layer, oj is the output of network jth layer, oi is the input to network jth layer,  is the learning rate, k is the iteration number and t is the target vector. 7. A technique called momentum is used to make it less likely for a back-propagation network to get caught in a shallow minimum. It allows a network to respond not only to the local gradient but also to recent trends in the error surface. When the momentum constant (˛) is 0 a weight change is based slowly on the gradient, but when it is 1 the new weight change is set equal to the last weight change. Back propagation with momentum is expressed mathematically as: wji (k + 1) = wji (k) + ıj oi + ˛wji (k)

(6)

8. Using an adaptive learning rate, that attempts to keep the learning step as large as possible, leads to reduce the training time. 9. Initial conditions for two-layer network can be chosen more favorably than by using purely random numbers. If you can pick more favorable initial conditions the learning rule does not need to work as hard and the training time will be reduced. 10. Present the next pattern in the training set and repeat steps from 1 to 9. 11. Check the error goal or a predefined maximum number of iterations. Steps from 1 to 10 will be repeated until one of them is achieved. (See program flow chart in Fig. B2). References [1] Applied Protective Relaying, Westinghouse Electric Corporation, Relay Investment-Division, Coral Springs, FL 33065, 1982. [2] S.H. Horowitz, A.G. Phadke, Power System Relaying, RSP Ltd., England, 1992. [3] IEEE working group—power system relaying committee, Single phase tripping and autoreclosing of transmission lines, IEEE Trans. Power Deliv. 7 (1) (1992). [4] Y. Sekine, et al., Fault diagnosis of power systems, Proc. IEEE 80 (May(5)) (1992) 673–683. [5] CIGRE TF 38.06.06, Artificial neural networks for power systems, Electra 159 (April) (1995) 77–101. [6] M.A. El-Sharkawi, D. Niebur, A tutorial course on artificial neural networks with applications to power systems, IEEE Copyright (1996). [7] M.A. El-Sharkawy, Overview of neural network application to power systems, in: Proceedings of the 5th, MEPCON’97, Alexandria, Egypt, January 4–6, 1997. [8] R.K. Aggarwal, A.T. Johns, Y.H. Song, R.W. Dunn, D.S. Fitton, Neural network based adaptive single-pole auto-reclosure technique for ehv transmission systems, IEE Proc. -C 141 (March(2)) (1994) 155–160.

613

[9] M.R. Zaman, M.A. Rahman, Experimental testing of an artificial neural network based protection of power transformer, IEEE Trans. Power Deliv. 13 (2) (1998) 510–517. [10] H.E.A. Talaat, Artificial Intelligence Applications in Digital Protection: A Review Article, Ain-Shams University, Faculty of Engineering, 2000, February. [11] E.A. Khamis, Power transmission system fault classification using AI technique, M.Sc. Thesis, Ain-Shams University, September 2000. [12] E.A. Mohamed, A.Y. Abdelaziz, A.S. Mustafa, A neural network based scheme for fault diagnosis of power transformers, Electric Power Syst. Res. J. 75 (July) (2005) 29–39. [13] A.J. Al-Shareef, E.A. Mohamed, E. Al-Judaibi, One hour ahead load forecasting using artificial neural network for the western area of Saudi Arabia, Int. J. Electric. Syst. Sci. Eng. 1 (1) (2008) 35–40. [14] T. Dalstein, B. Kulicke, Neural network approach to fault classification for high speed protective relaying, IEEE Trans. Power Deliv. 10 (April(2)) (1995) 1002–1011. [15] T. Dalsten, T. Friedrick, B. Kulicke, D. Sobajic, Multi-neural network based fault area estimation for high speed protective relaying, IEEE Trans. Power Deliv. 11 (April(2)) (1996). [16] M. Kezunovic, I. Rikalo, Detect and classify faults using neural nets, IEEE Trans. Comput. Appl. Power 9 (October(4)) (1996). [17] Z. Chen, J.C. Maun, Artificial neural network approach to single-ended fault locator for transmission lines, in: Proceedings, Power Industry Computer Applications, Conference, PICA’97, Columbus, OH, May, 1997, pp. 125–131. [18] K.L. Butler, J.A. Momoh, D.J. Sobajic, Field studies using a neural net-based approach for fault diagnosis in distribution networks, IEE Proc. Gen. Trans. Disbn. 144 (5) (1997) 429–436. [19] Y.H. Song, Q.Y. Xuan, A.T. Johns, Comparison studies of five neural network based fault classifiers for complex transmission lines, Electric Power Syst. Res. 43 (1997) 125–132. [20] M. Sanaye-Passand, O.P. Malik, High speed transmission system directional protection using an Elman network, IEEE Trans. Power Deliv. 13 (October(4)) (1998) 1040–1045. [21] M. Sanaye-Passand, O.P. Malik, Implementation and laboratory test results of an Elman network-based transmission line directional relay, IEEE Trans. Power Deliv. 14 (3) (1999) 782–788. [22] A. Poeltl, K. Frohlich, Two new methods for very fast fault type detection by means of parameter fitting and artificial neural networks, IEEE Trans. Power Deliv. 14 (4) (1999) 1269–1275. [23] W.M. Lin, C.D. Yang, J.H. Lin, M.T. Tsay, A fault classification method by RBF neural network with OLS learning procedure, IEEE Trans. Power Deliv. 16 (4) (2001) 473–477. [24] M. Sanaye-Passand, H. Khorashadi-Zadeh, O.P. Malik, High speed accurate transmission line distance protection using ANNs, in: Conference Proceedings CD, IEEE PES, 2001 Summer Meeting, 15–19, Vancouver, 2001. [25] W. Qi, G. Swift, P.G. McLaren, A. Castro, Distance protection using an artificial neural network, Proceedings, IEE Developments in Power System Protection Conference, IEE Publication No. 434, pp. 286–290. [26] A.L.O. Fernandez, N.K.I. Ghonaim, A novel approach using a FIRANN for fault detection and direction estimation for high-voltage transmission lines, IEEE Trans. Power Deliv. 17 (October) (2002) 894–900. [27] S.M. Yeo, C.H. Kim, K.S. Hong, Y.B. Lim, R.K. Aggarwal, A.T. Johns, M.S. Choi, A novel algorithm for fault classification in transmission lines using a combined adaptive network and fuzzy inference system, Electric Power Energy Syst. 25 (2003) pp 747–758. [28] R.N. Mohanty, P.B. Dutta Gupta, Application of RBF neural network to fault classification and location in transmission lines, IEE Proc. Gen. Trans. Disbn. 151 (2) (2004) 201–212. [29] Y.A.S. Omar, Combined fuzzy-logic wavelet-based fault classification technique for power system relaying, IEEE Trans. Power Deliv. 19 (2) (2004) 582–589. [30] A.K. Pradhan, A. Routray, S. Pati, D.K. Pradhan, Wavelet fuzzy combined approach for fault classification of a series-compensated transmission line, IEEE Trans. Power Deliv. 19 (4) (2004) 1612–1618. [31] A.H. Osman, O.P. Malik, Protection of parallel transmission lines using wavelet transform, IEEE Trans. Power Deliv. 19 (1) (2004) 49–55. [32] U. Lahiri, A.K. Pradhan, S. Mukhopadhyaya, Modular neural network-based directional relay for transmission line protection, IEEE Trans. Power Syst. 20 (4) (2005) 2154–2155. [33] S. Vasilic, M. Kezunovic, Fuzzy ART neural network algorithm for classifying the power system faults, IEEE Trans. Power Deliv. 20 (2 (pt. 2)) (2005) 1306–1314. [34] M.J. Reddy, D.K. Mohanta, Adaptive-neuro-fuzzy inference system approach for transmission line fault classification and location incorporating effects of power swings, IET Gener. Transm. Distrib. 2 (2) (2008) 235–244. [35] N. Zhang, M. Kezunovic, Transmission line boundary protection using wavelet transform and neural network, IEEE Trans. Power Deliv. 22 (April(2)) (2007) 859–869. [36] EMTP Developed Coordination Group; Electric Power Research Institute EMTP Rule Book, Version 2.1, Sections 6–10, 1993. [37] H. Demuth, M. Beal, Neural network—toolbox, using MATLAB The MathWorks Inc., Natick, MA, 2002 [Online]. Available: http://www.mathworks.com/.