Australian Journal of Basic and Applied Sciences

Transcription

1 AENSI Journas Austraian Journa of Basic and Appied Sciences ISSN: Journa home page: Improvement of Faut Identification and Locaization Using Gustafson-Kesse Agorithm In Adaptive Neuro-Fuzzy Inference System Amaina Abduah, Channarong Banmongko and Naebboon Hoonchareon Chuaongkorn University, 254 Phayathai Road, Pathumwan, Bangkok A R T I C L E I N F O Artice history: Received 25 January 2014 Received in revised form 12 March 2014 Accepted 14 Apri 2014 Avaiabe onine 25 Apri 2014 Keywords: Transmission ine, faut identification, cassification and ocaization, Waveet Transform, adaptive neurofuzzy inference system, Gustafson- Kesse A B S T R A C T Most of the techniques on identifying faut ocation depend on parameters of power transmission ine. Thus, a compex mathematica soution wi be considered which at such conditions, the dependence on ine parameters wi imit performance of agorithms. An independent or parameters free agorithm is an option to overcome this probem by using an artificia inteigent technique. This paper presents a recent scheme incuding of faut identification, and faut ocaization to estimate the distance of faut. A 115 kv parae transmission ine system has been used to deveop and impement the proposed scheme. We used the approach of hybrid inteigent method caed adaptive neuro-fuzzy inference system (ANFIS) combine with discrete waveet transform (DWT) to obtain a great performance. A nove approach of impementing Gustafson- Kesse (GK) custering agorithm has improved the stabiity and accuracy of performance. Resuts show that the proposed scheme can contributes as an efficient too to overcome the obstaces in anaysis of faut data AENSI Pubisher A rights reserved. To Cite This Artice: Amaina Abduah, Channarong Banmongko and Naebboon Hoonchareon., Improvement of Faut Identification and Locaization Using Gustafson-Kesse Agorithm In Adaptive Neuro-Fuzzy Inference System. Aust. J. Basic & App. Sci., 8(5): , 2014 INTRODUCTION One of the probems in power system protection is to accuratey cassify the faut data and estimate the exact ocation where the faut occurs.the increasing compexities of the modern power transmission systems have greaty raised the significance of the faut ocation research studies. It is very important issue in order to cear fauts quicky and restore power suppy as soon as possibe with minimum interruption. Soution to this probem wi hep the service operator, increase the reiabiity of power utiity performance and reduces economic damage to the network by reducing restoration time. Most of the techniques on identifying faut ocation depend on parameters of transmission ine. Thus, a compex mathematica soution wi be considered which at such conditions, dependabe on ine parameters wi imit the performance of agorithms. Faut identification and diagnosis based on conventiona deterministic approach face a few probems in achieving the desired speed, seectivity and accuracy. This is due to inabiity to adapt the dynamicay to the variation of operating conditions. Exampes of the probems are, system oading eve, faut inception instance and faut resistance, etc. An independent or parameters free agorithm is an option to overcome these probems by using an artificia inteigent technique such as Fuzzy ogic based by (Biswarup et a and Mahanty et a. 2007), Artificia neura network (ANN) based by (Samantray et a. 2007), Fuzzy neura network based by (Huisheng Wang et a. 1998), Waveet based by (Chunju Fan et a. 2006, Osman H. et a. 2004, E Safty S. et a. 2009), and combined Waveet-ANN techniques based by (Sami Ekici et a. 2008, Chiradeja P. et a. 2009, Nan Zhang et a. 2007, Abdoahi A. et a. 2010, and Meisam et a. 2011). The advantage of this soution wi increase the robustness and fexibiity of such techniques. This paper presents a recent scheme incuding of data reduction, faut identification, and faut ocaization. A 115 kv parae transmission ine system has been used to deveop and impement the proposed scheme. We used the approach of hybrid inteigent method caed adaptive neuro-fuzzy inference system (ANFIS) combine with discrete waveet transform (DWT) to obtain a great performance. ANFIS is inteigent techniques based on combination of artificia neura network (ANN) and fuzzy inference system (FIS) has been used on detection and cassification of fauts since the ast 20 years. Recenty, the researchers have started the investigation on its Corresponding Author: Amaina Abduah, Chuaongkorn University, 254 Phayathai Road, Pathumwan, Bangkok E-mai: amaina1979@gmai.com

2 456 Amaina Abduah et a, 2014 capabiity on estimating faut ocaization. Waveet transform is one of the efficient toos for anayzing non stationary signas such as transients, and has been widey appied to sove numerous probems in power systems. DWT with the mother waveet of Daubechies-4 is used to extract the features of data. The data is pertinent to singe ine to ground faut which is the most faut in transmission and distribution networks. More than 700 data is used in this study incuding the pre-faut data. The deveopment of simuation data invoved variety vaue of faut resistances and faut ocations to increase the accuracy performance. A nove approach of impementing Gustafson-Kesse (GK) custering agorithm (R. Babuska et a 2005) has improved the stabiity and accuracy of performance. Resuts show that the proposed scheme can contributes as an efficient too to overcome the obstaces in anaysis of faut data. MATERIAL AND METHODS A. System Structures: Transmission ine is a component of eectrica power systems that are most frequenty experienced fauts compared to other power system eements. Parae transmission ines are used in the transmission system due to their economic and environmenta benefits. The robustness and reiabiity of doube circuit/ parae ine is important since the fauts are commony happen at this type of system votage. However, the protection of parae ines are aways chaenging due to the mutua couping effect of the two ines. B. Waveet transforms impementation: Waveet transforms are fast and efficient means of anayzing transient votage and current signas (S. Sajedi et a 2011). This technique has extensivey used in AI approach but aso for improving two-termina faut ocation agorithm in conventiona approach (Wutthikorn et a 2010). The waveet transform not ony decomposes a signa into frequency bands, but aso, provides a non uniform division of the frequency domain. The appication of waveet transform in engineering areas usuay requires discrete waveet transform (DWT), which represented as: DWT m, n = 1 m x k ψ k nb m 0a 0 k a 0 am (1) 0 Waveet Coefficients are used for pre-processing of data input to the proposed scheme. These coefficients are obtained by appying DWT on the data under study. The DWT consideraby simpifies the input signa which reduces the voume of input data without oss of information. In order to reduce the input data, in connection with accuracy and speed but retaining important feature of the waveet signas, features have been extracted. This dramaticay reduces the training stage in the artificia neura network ANN and increases the overa performance of the digita reay. Technique of identifying maximum waveet coefficient of ¼ cyces of the signas immediatey after faut occurs has been used. It can improve agorithm s performance if they are used as inputs instead of actua vaues of votages and currents of three phase transmission ine system. One of the most popuar mother waveets found for power system transient anaysis in the iterature is Daubechies s waveet famiy. Waveet anaysis deas with expansion of functions in terms of a set of basic functions (waveets) which are generated from a mother waveet by operations of diatations and transations (C. Sydney et a 1998). The mother waveet daubechies (db4) is empoyed to decompose high frequency components from the signa. After appying DWT to votages and currents signas, coefficients are obtained. Waveet anaysis is a reativey new signa processing too and is appied recenty by many researchers in power systems due to its strong capabiity of time and frequency domain anaysis (C.H. Kim and R. Aggarwa, 2000 and 2001). C. Modeing of Adaptive Neuro-Fuzzy Inference System (ANFIS): ANFIS is a hybrid inteigent system which combines the fuzzy ogic quaitative approach and adaptive neura network capabiities towards better performance (Jang 1993). ANFIS has been used in faut detection and faut cassification in power system during ast 2 decades (Ramadoni 2013). Recenty, the researchers have spend their interest on its capabiity for faut ocation in transmission ine and distribution network (Ramadoni 2013, Javad et a 2009, E Sayed 2010). A few advantages in appication of ANFIS are better baance between credibiity and understandabe, aso it is proficient to adjust the membership function parameters and inguistic rues directy from the data. Other advantage is that the neura networks become more transparent, whie fuzzy systems become competent for earning. The basic structure of Fuzzy Inference System (FIS) as shown in Figure 1 is a mode that maps input characteristics to input membership functions, input membership function to rues, rues to a set of output characteristics, output characteristics to output membership functions, and the output membership function to a singe vaued output or a decision associated with the output. Fuzzy inference can aso be empoyed to mode systems whose rue structure is essentiay predetermined by the user s interpretation of the characteristics of variabes in the mode. However, it cannot be distinguish what the FIS membership functions shoud ook ike simpy from the data for some situations. Parameters of FIS coud be

3 457 Amaina Abduah et a, 2014 chosen so as to taior the membership functions to the input and output data rather than choosing the parameters associated with a given membership function arbitrariy in order to account for these types of variations in the vaues of data. D. Impementation of Gustafson-Kesse agorithm: The Gustafson-Kesse (GK) agorithm (D.E. Gustafson et a 1979) is a powerfu custering technique with a arge number of appications in various domains incuding image processing, cassification and system identification (R.N. Dave et a 1992, R. Babuska 1992). The important characteristic of this agorithm is the oca adaptation of the distance metric to the shape of the custer by estimating the custer covariance matrix. Another main feature is abiity on adapting the distance-inducing matrix correspondingy, when the GK agorithm is impemented in the extraction of Takagi Sugeno fuzzy mode from data. Overfitting is abe to be reduced when the number of training sampes is ow in comparison to the number of custers. This is achieved by adding a scaed unity matrix to the cacuated covariance matrix. Mosty, the Gustafson-Kesse custering agorithm is appied to identify Takagi Sugeno modes (D.E. Gustafson et a 1979). A drawback of this agorithm is that ony custers with equa voumes can be found and the resuted custers cannot be directy used to form membership functions. The Gustafson-Kesse agorithm aows cacuating an adaptive distance as a way to detect custers with different geometries within a data set. A specia GustafonKesse agorithm, caed modified GK agorithm (Robert Babuska, 2005) has been used. The advantage of this agorithm is it can improve the performance of training stage and abe to find a partitioning of the data. The expanation beow describes the steps on impementing modified Gustafson Kesse agorithm. RULES crisp inputs FUZZIFICATION DEFUZZIFICATION crisp outputs INFERENCE Fig. 1: Basic structure of the Fuzzy Inference System. step 1: Compute custer prototypes means : V i () = N k=1( ( 1) ) m Z k N (μ ( 1) k=1 ik ) m, 1 i K step 2: Compute the custer coverience matrices: F i = N k=1 1 m (z k V i )(z k V i ) T N k=1 1 m, 1 i K. Add a scaed identity matrix: F i = 1 γ F i + γ det F 0 1 n I, Extract eigenvauesλ ij and eigenvectorsφ ij fromf i. Findλ i max = max j λ ij and set: λ ij = λ i max / β j for whichλ i max / λ i j > β Reconstruct F i by F i = φ i1... φ in diag(λ i1,..., λ in ) φ i1... φ in 1 Step 3: Compute the distances

4 458 Amaina Abduah et a, D ik Ai = (z k V i ) T [ρ i det F 1 i nf 1 i ] (z k V i ), 1 i K, 1 k N. Step 4: Update the partition matrix: for 1 k N if D ik Ai > 0 for 1 i K, = 1 K j =1 (D ik Ai /D jk Ai ) 2/(m 1), Otherwise = 0 ifd ik Ai > 0, and [0, 1] with K i=1 = 1 otherwise. Unti U U ( 1) < Faut Identification: The inputs, which are the six waveet coefficients of the singe ine to ground faut data. Then it custers the data into two custers using modified GK agorithm which is better than FCM or hard partitioning ike K-means agorithm. Membership of each data point is cacuated for each defined cass. Since it is the fuzzy custering, the data point may beong to more than one cass with certain degree of membership. The membership to one cass is assigned based on the highest vaue of the membership vaue. For exampe, if the data point 1 has the foowing membership in two custers; Custer 1 : 0.8 Custer 2 : 0.2 (the sum shoud be equa to 1) Then, it is assigned to custer 1 due to higher membership vaue. Since there are two custers, the task was to create a custer for a faut case, and a custer for no faut case. Faut Locaization: In this part, the same strategy has been used. We knew that our aim is to identify fauts data pertinent to phase to ground (AG), phase B to ground (BG) and phase C to ground (CG). These are the cass of faut data. The input is six waveet coefficients. The same strategy on membership vaue and custering are used in the program. The distance of faut ocation wi be predicted based on the deveoped agorithm. RESULTS AND DISCUSSION Fig. 2: One-ine diagram of the system. Modeing using ANFIS V Waveet transform Faut identification Faut cassification Faut ocation I Contro unit Trip signa

5 459 Amaina Abduah et a, 2014 Fig. 3: Proposed mode of protection scheme. Figure 2 shows a one-ine diagram of 115 kv doube-circuit transmission ines used in this study. The transmission ines with 100 km in ength are suppied from both sides. They are modeed and simuated by using Eectromagnetic Transient Program (EMTP) to obtain a arge amount of data of faut currents and votages. A faut is created at every 2 km aong the fu ength. The purpose on creating faut at a ot of ocations is to increase the reiabiity on accuracy of faut distance estimation. For each distance taking into account, faut resistance is varying with the vaues of 0.5 Ω, 5 Ω, 25 Ω, 100 Ω and 200 Ω. The simuation faut data is then transferred to Matabenvironment for further process and anaysis. To accompish the task, features of votages and currents signas are extracted based on discrete waveet transform with daubechies4 (db4) as mother waveet. These features are then used as an input to train and test the agorithm. Technique of identifying maximum waveet coefficient of ¼ cyce of the signas immediatey after faut occurs has been used. Figure 3 shows the proposed mode of protection scheme discussed in this paper. It begin with raw signa from the faut pertinent to votages and currents. The signas then being through data reduction process using waveet transform method. Now, the data is ready for the input for the system of ANFIS. The occurrence of faut wi be identified to confirm the existence of faut. Then, the faut wi be cassify either singe ine to ground, doube ine, doube ine to ground or three phase. In this paper, faut cassification wi concern on specific of singe ine to ground; AG, BG or CG. Then, distance of faut wi be estimated. Basicay, the estimated resut wi be compared with the actua distance to identify the accuracy of performance. The steps beow summarize the proposed mode in genera description. Step 1: Create the architecture of fuzzy inference system (FIS) in Matab environment. The architecture consist of two inputs (i.e. votages and currents faut measured in the ocator end of transmission ine) and one output. Step 2: Determine the membership functions for ANFIS input and output. In this work, gbe membership function has been chosen for each input and output of FIS architecture. Step 3: Coecting or producing suitabe information (data) to train ANFIS. The data for training process shoud have same form and the various conditions of a rea power system is incuded. A power system simuation using EMTP has been carried for achieving the suitabe data. Figure 4 in the foowing part shows the G-Be membership function of faut ocaization. The performance of a be membership functions for specifying fuzzy sets. It is important to attain smoothness and symmetrica shape of potting. The parameter ocates the center of the curve width of the curve. It can be seen in figure 4 that impementation of GK agorithm has improved the shape of membership function. Thus, the further resuts wi be much better and reiabe. Another resuts in figure 5 shows the prediction from Takagi-Sugeno for training and testing pertinent to faut ocaization. Figure 5a shows that the potting is more scatter and ampitude is high. Thus, the resut is ess reiabe. A better potting can be seen in figure 5b. The ampitude is aso reduced significanty. Numerica outputs in Matab aso proved the resuts. With this performance, further resuts wi be more stabe and obtain better accuracy. a) b)

6 460 Amaina Abduah et a, 2014 Fig. 4: G-Be membership function a) Without GK agorithm b) With GK agorithm impementation. a) b) Fig. 5: Prediction from Takagi-Sugeno a) Without GK agorithm b) With GK agorithm impementation. The percentage of error is then cacuated based on the formua given beow (E Sayed et a 2006) Actuafautocation Predictedfautocation % Error = 100 totaineengt Tabe 1 shows the recorded faut distance estimation for singe ine to ground studied in this paper. The distances are at 2km, 20km, 40km, 60km, 80km and 100km. Five different vaues of faut resistance are invoved to observe the effect on accuracy. The resuts show a reasonaby good performance for distance estimation of faut at power transmission ine. Tabe 1: Faut distance estimation for singe ine to ground faut (random seection of distances). Actua distance Estimation error (%) (km) Rf = 0.5 Ω Rf =5 Ω Rf = 25 Ω Rf = 100 Ω Rf =200 Ω Concusion: This paper has proposed a waveet adaptive neuro fuzzy inference system approach that abe to identify the faut, cassify the types and estimate the distance of faut ocation. The input is based on votages and currents which then extracted by using waveet coefficient. The proposed agorithm is different from conventiona agorithms that are based on deterministic computations on a we-defined mode to be protected. The simuation resuts show that the proposed scheme provides a fast, accurate and reiabe technique for faut identification and diagnosis. The successfu of impementing neuro-fuzzy is heaviy depends on prior knowedge of the system and the training data. The estimation error for singe ine to ground faut with the variation of faut resistances is ess than 0.1%. The proposed scheme has been deveoped with aim for generaization output to avoid over fitting resuts. REFERENCES Abdoahi, A. and S. Seyedtabaii, Transmission ine faut ocation estimation by fourier and waveet transforms using ANN. Internationa Power Engineering and Optimization Conf., pp:

7 461 Amaina Abduah et a, 2014 Biswarup, Das and Vitta, J. Reddy, Fuzzy ogic based faut cassification scheme for digita distance protection. IEEE Transactions on Power deivery, 20(2): Sydney Burrus, C., R.A. Gopinath and H. Guo, Introduction to Waveets and Waveet Transforms, Engewood Ciffs, NJ: Prentice-Ha. Kim, C.H. and R. Aggarwa, Waveet transforms in power systems Part 1: Genera introduction to the waveet transforms, Power Engineering Journa, 14(2): Kim, C.H. and R. Aggarwa, Waveet transforms in power systems Part 2: Exampes of appication to actua power system transients, Power Engineering Journa, 15(4): Chunju Fan, Li, K.K., W.L. Chan and Yu, Weiyong, Study of protection scheme for transmission ine based on waveet transient energy. Eectrica Power and Energy Systems, 28: Chiradeja, P. and A. Ngaopitakku, Identification of fauttypes for singe circuit transmission ine using discrete waveet transform and artificia neura networks. Proceedings of the Internationa Muti Conference of Engineers and Computer Scientists, 2. Gustafson, D.E. and W.C. Kesse, Fuzzy custering with a fuzzy covariance matrix, In Proc. IEEE CDC, pages , San Diego, CA, USA. E-Safty, S. and A. E-Zonkoy, Appying waveet entropy principe in faut cassification. Eectrica Power and Energy Systems, 31: E-Sayed Mohamed Tag, Faut Location for Series Compensated Transmission Line Based on waveet Transform and an ANFIS. E-Sayed Tag E-Din, M.M. Abde Aziz, D.K. Ibrahim, M. Giany, Faut ocation scheme for combined overhead ine with underground power cabe, Eectr. Power Syst. Res., 76: Huisheng Wang and W.W.L. Keerrthipaa, Fuzzy neuro approach to faut cassification for transmission ine protection. IEEE Transactions on Power Deivery, 13(4): Jang, J.S., ANFIS: adaptive network based fuzzy inference system. Javad Sadeh and Hamid Afradi, A New and accurate Faut Location Agorithm for combined Transmission Line Using ANFIS. Mahanty, R.N. and P.B. Dutta, Gupta, A fuzzy ogic based faut cassification approach using current sampes ony. Eectric Power System Research, 77: Meisam Pourahamdi and Safavi, Ai Akbar, Path characteristic frequency based faut ocating in radia distribution systems using waveets and neura network, IEEE Trans. Power Deivery, 26: Nan Zhang and Kezunovic, Maden, Transmission ine boundary protection using waveet transform and neura network. IEEE Transaction on Power deivery, 22(2): Osman, H. and O.P. Maik, Protection of parae transmission ines using waveet transform. IEEE Transactions on Power Deivery, 19(1): Ramadoni Syahputra, A Neuro-fuzzy Approach For the Faut Location Estimation on Unsynchronized Two-Termina Transmission Line. Babu ska, R., P.J. van der Veen and U. Kaymak, "Improved covariance estimation for Gustafson- Kesse custering," Internationa Conference on Fuzzy Systems, Honouu, Hawaii, pp: Dave, R.N., Boundary detection through fuzzy custering, In IEEE InternationaConference on Fuzzy Systems, pages , San Diego, USA. Babuˇska, R., Fuzzy Modeing for Contro, Kuwer Academic Pubishers, Boston, USA. Sajedi, S., F. Khaifeh, Z. Khaifeh, T. Karimi, Appication Of Waveet Transform For Identification Of Faut Location On Transmission Lines, Austraian Journa of Basic and Appied Sciences, 5(12): , ISSN Samantray, S.R., P.K. Dash and G. Panda, Distance reaying for transmission ine using support vector machine and radia basis function neura network. Eectrica Power and Energy System, 29: Sami Ekici, Yidirim, Secuk and Poyraz, Mustafa, Energy and entropy based feature extraction for ocating faut on transmission ines by using neura network and waveet packet decomposition. Expert Systems with Appications, 34: Wutthikorn Threevithayanon and Naebboon Hoonchareon, Faut Data Synchronization Using Waveet for Improving Two-termina Faut Location Agorithm, Power and Energy Engineering Conference (APPEEC), Chengdu, March.