Partial Discharge Pattern Recognition of Cast Resin Current Transformers Using Radial Basis Function Neural Network

Transcription

1 J Electr Eng Technol Vol. 9, No. 1: , ISSN(Prnt) ISSN(Onlne) Partal Dscharge Pattern Recognton of Cast Resn Current Transformers Usng Radal Bass Functon Neural Network Wen-Yeau Chang Abstract Ths paper proposes a novel pattern recognton approach based on the radal bass functon (RBF) neural network for dentfyng nsulaton defects of hgh-voltage electrcal apparatus arsng from partal dscharge (PD). Pattern recognton of PD s used for dentfyng defects causng the PD, such as nternal dscharge, external dscharge, corona, etc. Ths nformaton s vtal for estmatng the harmfulness of the dscharge n the nsulaton. Snce an nsulaton defect, such as one resultng from PD, would have a correspondng partcular pattern, pattern recognton of PD s sgnfcant means to dscrmnate nsulaton condtons of hgh-voltage electrcal apparatus. To verfy the proposed approach, experments were conducted to demonstrate the feld-test PD pattern recognton of cast resn current transformer (CRCT) models. These tests used artfcal defects created n order to produce the common PD actvtes of CRCTs by usng feature vectors of feld-test PD patterns. The sgnfcant features are extracted by usng nonlnear prncpal component analyss (NLPCA) method. The expermental data are found to be n close agreement wth the recognzed data. The test results show that the proposed approach s effcent and relable. Keywords: Partal dscharge, Pattern recognton, Radal bass functon neural network, Cast resn current transformer, Defect of nsulaton, Nonlnear prncpal component analyss 1. Introducton PD measurement and pattern recognton are mportant tools for mprovng the relablty of hgh-voltage nsulaton systems [1]. The pattern recognton of PD ams to dentfy potental nsulaton defects from the measured data, and the potental defects can then be used for estmatng the rsk of nsulaton falure n hgh-voltage electrcal apparatus. In the presence of a suffcently strong electrc feld, a sudden local dsplacement of electrons and ons wll lead to a PD f a defect n an nsulator exsts. A PD event that occurs n the epoxy resn nsulator of hgh-voltage electrcal apparatus would have harmful effects on nsulaton that may fnally cause power system blackout. A defect n hgh-voltage electrcal apparatus, resultng n PD, wll have a correspondng partcular pattern, and so pattern recognton of PD s a sgnfcant technque for evaluatng the condton of the nsulaton n hgh-voltage electrcal apparatus [2]. Because there has been substantal progress n the physcal understandng of PD durng the last decade, ths knowledge can now be exploted to support the nterpretaton of nsulaton defects. Recently, several methods have been employed for the pattern recognton of PD, ncludng neural network, expert systems, self organzng maps, wavelet analyss, and the grey clusterng Correspondng Author: Dept. of Electrcal Engneerng, St. John s Unversty, Tawan. (changwy@mal.sju.edu.tw) Receved: September 20, 2012; Accepted: September 24, 2012 method. The applcaton of neural network to pattern recognton and system dentfcaton has become a major trend n fault dagnoss [3]. Neural network have been appled for PD classfcaton of epoxy resn power transformer, PD pattern recognton of current transformers, and PD montorng technque of gas nsulated substaton. Although the speed of neural network allows real-tme operaton wth comparable accuracy, the tranng process of multlayer neural network s often very slow, and the tranng data must be suffcent and compatble. The recognton of PD pattern and the evaluaton of nsulaton performance are relatvely complcated, a task whch often must be completed by decson tree method. Decson tree method for the dagnostcs of SF 6 decomposton products have been developed [4]. In contrast to other clusterng mappng methods for unsupervsed data, the mappng relatonshp of a self organzng map can be hghly nonlnear, drectly showng the smlar nput vectors n the source space by ponts close to the two-dmensonal target space. Along wth the smlarty of the nput data, a self organzng map may lead to classfcaton results, and ths technque has been appled for PD pattern recognton of CRCT [5]. The wavelet analyss method s a useful tool n fault detecton and de-nosng, and ths method has also been appled to analyss of power transformer partal dscharge sgnals [6]. Grey system theory s a useful methodology for systems wth ncomplete nformaton. Grey relatonal analyss can be used to analyss the relatonshps between 293

2 Partal Dscharge Pattern Recognton of Cast Resn Current Transformers Usng Radal Bass Functon Neural Network one major sequence and the other comparatve ones n a gven set. The applcaton of grey clusterng approach has been proposed for recognzng partal dscharge patterns of the hgh-voltage equpment [7]. In ths paper the PD patterns are measured usng a commercal PD detector. A set of features, used as operators, for each PD pattern s extracted through statstcal schemes. The sgnfcant features of statstcal operators are extended extracted by usng the NLPCA scheme. After feature extracton, ths paper proposes the applcaton of RBF neural network to recognze partal dscharge patterns of CRCT. Ths paper s organzed as follows. Creaton of the PD pattern dataset and the extracton of phase-related dstrbutons are descrbed n Secton 2. The algorthm of statstcal feature extracton s descrbed n Secton 3. The NLPCA features extracton algorthm s descrbed n Secton 4. The prncples of RBF neural network and the operaton flowchart of the proposed pattern recognton scheme are gven n the next secton. The expermental results and the analyss usng 250 sets of feld-test PD patterns from hgh-voltage CRCTs are presented n Secton 6. From the test results, the effectveness of the proposed scheme to mprove the recognton accuracy has been demonstrated. The paper s concluded n Secton 7. cavty dscharge caused by an ar cavty nsde the epoxy resn nsulator on the hgh-voltage sde (VH), as shown n Fg. 2; (3) nternal cavty dscharge caused by two cavtes nsde the epoxy resn nsulator on the low-voltage sde (VL), as shown n Fg. 3; (4) nternal fssure dscharge caused by an ar fssure nsde the epoxy resn nsulator on the hgh-voltage sde (FH), as shown n Fgs. 4; and (5) nternal dscharge caused by a metal-lne mpurty nsde the epoxy resn nsulator on the hgh-voltage sde (MH), as shown n Fg. 5. The PD events were detected by a PD detectng system set up n our laboratory. The structure of the PD detectng system s shown n Fg. 6. It ncludes a step-up transformer, capactor couplng crcut, PD detector, and the CRCT under test. Through the testng processes, all the data measured were dgtally converted n order to store them n 2. PD Patterns Database Creaton In order to nvestgate the PD features and to verfy the classfcaton capabltes of the proposed RBF neural network based pattern recognton approach for dfferent PD types commonly occurrng n hgh-voltage electrcal apparatus, a PD dataset s needed. The PD dataset for ths study was collected from laboratory PD tests on a seres of model CRCTs. The materals and process used to manufacture these hgh-voltage CRCTs were exactly the same as that of makng the feld equpment. The appearance of a 12kV, 40VA model CRCT s shown n Fg. 1. Fve types of expermental models wth artfcal defects embedded were made to produce fve common PD events n the CRCT. The fve PD actvtes nclude (1) normal PD actvty n standard CRCT (NM); (2) nternal Fg. 2. VH on the hgh-voltage sde of CRCT Fg. 3. VL on the low-voltage sde of CRCT Fg. 1. The appearance of model CRCT Fg. 4. FH on the hgh-voltage sde of CRCT 294

3 Wen-Yeau Chang Fg. 8. Typcal phase-related dstrbutons of PD for VL Fg. 5. MH on the hgh-voltage sde of CRCT Fg. 9. Typcal phase-related dstrbutons of PD for FH Fg. 6. System confguraton of the PD detectng system the computer memory. Then, the phase-related dstrbutons of PD derved from the orgnal PD data are obtaned n relaton to the waveform of the feld test hgh voltage. The hgh voltage n the feld tests s assumed to be held constant and the voltage phase angle s dvded nto a sutable number of wndows (blocks). The PD detectng system, shown n Fg. 6, s used for acquston of all the ndvdual quasntegrated pulses and quantfyng each of these PD pulses by ther dscharge magntude (q), the correspondng phase angle (φ), at whch PD pulses occur and the number of dscharge (n) over the chosen block. The analyss software plots these data as functons of the phase postons. The three phase-related dstrbutons refer to the peak pulse magntude dstrbuton H qmax (φ), the average pulse magntude dstrbuton H qn (φ), and the number of pulse dstrbuton H n (φ). The typcal phase-related dstrbutons of PD patterns for the four types of defects (VH, VL, FH, and MH) of the nsulaton models are shown n Fgs. 7 to 10, respectvely. As shown n Fgs. 7 to 10, the PD patterns of deferent defects dsplay dscrmnatve features. Fg. 10. Typcal phase-related dstrbutons of PD for MH 3. Statstcal Feature Extracton In PD pattern recognton, feature extracton s a technque essental to reduce the dmenson of the orgnal data [8]. The features are ntended to denote the characterstcs of dfferent PD statuses. Several statstcal methods of feature extracton are descrbed n ths secton, and fve statstcal operators are extracted from the phaserelated dstrbutons. Defntons of the operators are descrbed below. The profle of all these dscrete dstrbuton functons can be put n a general framework,.e., y = f (x ) [9]. The statstcal operators of mean (μ) and varance (σ 2 ) can be computed as follows: x = f ( x ) μ (1) f ( x ) 2 ( x μ) f ( x ) f ( x ) 2 σ = (2) Fg. 7. Typcal phase-related dstrbutons of PD for VH Skewness (S k ) s extracted from each phase-related dstrbuton of PD to denote the asymmetry of dstrbuton. It can be expressed as: 295

4 Partal Dscharge Pattern Recognton of Cast Resn Current Transformers Usng Radal Bass Functon Neural Network Sk 3 ) p ( x μ = (3) 3 σ Kurtoss (K u ) s extracted to descrbe the sharpness of dstrbuton as: 4 ( ) = x μ p K u 3 (4) 4 σ In (3) and (4), x s the statstcal value n the phase wndow, p s the related probablty of appearance. Peaks (P e ) count the number of peaks n the postve or negatve half of a cycle of the dstrbuton. Asymmetry (D a ) represents the asymmetrcal characterstc of partal pulses n postve and negatve cycles. It can be expressed as: + N q D = a (5) + N q where N - s the number of PD pulses n the negatve cycle, N + s the number of PD pulses n the postve cycle, q - s the ampltude of the PD pulse at the phase wndow n the negatve cycle, and q + s the ampltude of the PD pulse at the phase wndow n the postve cycle. Cross correlaton factor (C c ) ndcated the dfference n sharp of the dstrbutons n the postve and negatve half cycles. C c = 1 means that the sharps are totally symmetrc and C c = 0 means that sharps are totally asymmetrc. The cross correlaton factor can be expressed as: x y x y / n C = c (6) ( x ( x ) / n) ( y ( y ) / n) where x s the statstcal value n the phase wndow of the postve half cycle, y s the statstcal value n the correspondng wndow of the negatve half cycle, and n s the number of phase wndows per half cycle. Upon applyng S k, K u and P e to both postve and negatve cycles of H qmax (φ), H qn (φ), and H n (φ), a total of 18 features can be extracted from a PD pattern. However, upon applyng D a and C c to ndcate the dfference or asymmetry n postve and negatve cycles of H qmax (φ), H qn (φ), and H n (φ), a total of 6 features can be extracted from a PD pattern. Therefore, after the procedure of feature extracton, a feature set of 24 statstcal features s bult for each PD pattern. The use of statstcal features rather than recordng the dstrbuton profles can sgnfcantly reduce the dmenson of the database. To a certan extent, they can be used for characterzng PD patterns wth reasonable dscrmnaton [10]. 4. NLPCA Based Feature Extracton Scheme The statstcal feature extracton methods were used to extract 24 statstcal features for patterns. But snce some of the statstcal features are futle for pattern recognton, feature extracton s necessary n the PD pattern recognton to reduce dmenson of orgnal data and make effectve dscrmnaton of the statstcal feature patterns for dfferent PD status. In ths paper, the sgnfcant features are extracted from statstcal features by usng the NLPCA scheme [11]. The NLPCA s based on the structure of dual multplayer neural networks model (DMNN), whch contans fve layers of neurons, as shown n Fg. 11. In Fg. 11, the DMNN for NLPCA contans two subnetworks of mappng network and demappng network. The mappng from data space to feature space s referred to as the mappng network and the reverse mappng as the demappng network. The neurons at layers 1 and 3 of the network have sgmod actvaton functons. In tranng, the output vector x = [x 1,x 2,..,x n ], where n s the number of the neurons at the output and nput layers, s antcpated to approach to the nput data vector x = [x 1,x 2,..,x n ] at the nput layer. As noted, the nput layer of the mappng network has neurons equal to the dmensonalty of the nput data. In ths paper n s set to be Demappng Network Mappng Network x1' x2' xn' f1 x1 xn-1 xn Output Layer Layer 2 (Feature Layer) Input Layer Fg. 11. Archtecture of the DMNN n the NLPCA fm Layer 3 Layer 1 296

5 Wen-Yeau Chang 24 whch s the number of statstcal features. After the network s traned, the m neurons at layer 2 (feature layer) represent lower-dmensonal nonlnear features f = [f 1,f 2,..,f m ] extracted from the nput data set. The NLPCA attempts to fnd the mappngs from multdmensonal data space to lower-dmensonal feature space. In the process, the reconstructon error between nput x and output x of the dual networks s mnmzed [12]. X0 X1 h0 h1 h2 O0 2 J = x x' (7) X2 O1 The whole network, consstng of the dual networks n the NLPCA, s an auto-assocatve network where the output vector corresponds to the nput vector. The man advantage of NLPCA over prncpal component analyss s that NLPCA has the ablty to stand for nonlnear relatonshps among the data set of varables. X3 Xd-1 hg-1 On-1 Input Layer Hdden Layer Output Layer 5. RBF Neural Network Based PD Pattern Recognton Approach The RBF neural network s a useful methodology for systems wth ncomplete nformaton. It can be used to analyze the relatonshps between one major sequence and the other comparatve ones n a gven set. In ths secton, the algorthms of RBF neural network and the RBF neural network-based PD pattern recognton scheme are descrbed. The PD recognton through a RBF neural network n multdmensonal feature space s also valdated on the bass of the features extracted by the NLPCA scheme, as mentoned above. 5.1 Prncpals of RBF neural network The RBF neural network s a forward network models wth unversal approxmaton capabltes, and whch s employed to approxmate the functon [13]. It s a multnput, mult-output system consstng of an nput layer, a hdden layer, and an output layer. Durng data processng, the hdden layer performs nonlnear transforms for the feature extracton and the output layer gves a lnear combnaton of output weghts [14]. The structure s shown n Fg. 12. The network actually performs a nonlnear mappng from the nput space R d to the output space R n. The mappng relatonshp between nput vector and output vector of RBF neural network s based on the followng functon: RBF Neural d n R R Network : (8) x o Fg. 12. Archtecture of the RBF neural network system where nput vector x ={x, for = 0,1,2,,d-1}, output vector o = {o, for = 0,1,2,,n-1}. Each hdden neuron computes a Gaussan functon n the followng equaton T 2 ( x μ j ) h j ( x) = exp[ ], for j = 0,1,2,..., g 1 2 2σ j where μ j and σ j are, respectvely, the center and the wdth of the Gaussan potental functon of the jth neuron n the hdden layer. Each output neuron of the RBF neural network computes a lnear functon n the followng form: (9) g 1 ok = wkjh j ( x) θ k, for k = 0,1,2,..., n 1 (10) j= 0 where o k s output of the kth node n the output layer, w kj s weght between jth node n the hdden layer and kth node n the output layer, h j (x) s output from the jth node n the hdden layer, θ k s bas of the kth node n the output layer. 5.2 Tranng procedure of RBF neural network The tranng procedure of RBF neural network s composed of a two-step decomposton: estmatng μ j and σ j and estmatng the weghts between the hdden layer and output layer. The procedure of μ j and σ j estmaton s descrbed brefly n the followng steps: Step 1 Take any pont μ j and ts assocated wdth σ j (ntally σ j = 0); 297

6 Partal Dscharge Pattern Recognton of Cast Resn Current Transformers Usng Radal Bass Functon Neural Network Step 2 Use the Eucldean dstance to fnd the nearest pont μ l of the same class; Step 3 Compute the mean of these two ponts to obtan a new pont wth ts assocated wdth by usng followng equaton σ = ( μ j, μl ) / 2 + σ j ; Step 4 Compute the dstance D from the new mean to the nearest pont of all other classes; Step 5 If D < 2σ, then accept the merge of μ j and μ l and start agan from Step 2; f D 2σ,, reject the merge and recover the two orgnal ponts and ther wdths, restart from Step 1; Step 6 Repeat Steps 1 5 untl all clusters of each class are used. The procedure of estmatng the weghts between the hdden and output layer s descrbed brefly as follows. After the Gaussan functon centres and wdths are computed from tranng vectors, the connecton weghts between the hdden and output layers can be calculated usng the pseudo-nverse matrx method. No Start PD Patterns Data Base Creaton Statstcal Features Extracton Features Extracton Usng NLPCA Scheme Prepare the Tranng Set Intal Settng the RBF Neural Network Tranng the RBF Neural Network Tranng Procedure Fnshed? Stop Yes Save the Traned RBF Neural Network PD Pattern Recognton for Equpments Fg. 13. Flowchart of the RBF neural network based recognton scheme 5.3 RBF neural network based PD pattern recognzng procedure The proposed PD pattern recognton scheme based on RBF neural network has been successfully mplemented usng PC-based software for the PD recognton. The overall flowchart s shown n Fg. 13, and the proposed recognton scheme s descrbed brefly n the followng steps: Step 1 Create data base of the phase-related dstrbutons of PD patterns; Step 2 Extract the statstcal features from phase-related dstrbutons; Step 3 Extract the sgnfcant features from statstcal features by usng the NLPCA scheme; Step 4 Prepare the tranng set for the RBF neural network; Step 5 Use the tranng set to tran the RBF neural network for PD pattern recognton; Step 6 Save the Gaussan functons centres, wdths and connecton weghts between the hdden and output layers of traned RBF neural network when the tranng procedure s fnshed; Step 7 Use the traned RBF neural network to dentfy the defect types of PD patterns. Even though a concluson may revew the man results or contrbutons of the paper, do not duplcate the abstract or the ntroducton. For a concluson, you mght elaborate on the mportance of the work or suggest the potental applcatons and extensons. 6. Experment Results To verfy the proposed approach, a practcal experment s conducted to demonstrate the effectveness of the PD pattern recognton scheme. The expermental tests were carred out on model CRCTs. The test results show that the proposed approach s able to accurately recognze the testng defects. Fve types of expermental models wth artfcal defects are purposely embedded to produce fve common PD events n CRCT. The proposed approach has been mplemented accordng to the feld-test PD patterns collected from our laboratory. The nput data for ths PD recognton system are the peak pulse magntude dstrbuton H qmax (φ), the average pulse magntude dstrbuton H qn (φ), and the number of pulse dstrbutons H n (φ). Assocated wth ther real defect types, there are a total of 250 sample data for dfferent PD events. Each PD event contans 50 patterns of sample data, of whch 30 patterns are tranng data and 20 patterns are testng data. Statstcal feature extracton methods are used to extract 24 statstcal features for each pattern. Upon applyng S k, K u, P e, D a and C c to both postve and negatve cycles of 298

7 Wen-Yeau Chang Table 1. Recognton performance of tranng data n CRCT tests Pattern Defect Types Accuracy Rate NM 100% Tranng Data VH 100% (150 patterns) VL 100% FH 100% MH 100% Table 2. Recognton performance of testng data n CRCT tests Pattern Defect Types Accuracy Rate NM 100% Testng Data VH 95% (100 patterns) VL 100% FH 95% MH 95% H qmax (φ), H qn (φ), and H n (φ), a total of 24 statstcal features have been extracted from a PD pattern. The sgnfcant features are extracted from statstcal features by usng the NLPCA scheme. There are 10 extracted features for feature vector extracted from 24 statstcal features usng the NLPCA scheme n ths experment. After the feature extracton process, all the features n the feature vectors were normalzed to set up the tranng sets. After settng up the tranng sets, the tranng procedure of RBF neural network was started. The tranng data consst of 150 feature vectors, whch are randomly chosen from the 250 feature vectors of sample data. The remanng 100 feature vectors were used as the testng data. To verfy the tranng results of RBF neural network, the tranng data were appled to the traned RBF neural network agan. Table 1 shows the test results of the tranng data. The data n Table 1 show that the proposed approach has 100% accuracy for the 150 tranng feature vectors, because the tranng process of RBF neural network was stop under the error s lower than Table 2 demonstrates the promsng performance when 100 testng patterns were tested. Table 2 shows that among the 100 testng patterns, there were only 3 errors of recognton, one for VH, one for FH, and the other for MH defects. The total accuracy rate of 100 testng patterns s 97%. 7. Concluson Ths paper has proposed an RBF neural network based pattern recognton technque for PD of hgh-voltage equpment. The effectveness of the proposed technque has been verfed usng expermental results. It has been shown that through the NLPCA feature extracton procedure, the extracted feature vectors can sgnfcantly reduce the sze of the PD pattern database. In addton, the PD pattern recognton scheme, based on RBF neural network s very effectve for clusterng the defects of hghvoltage equpment. The content of PD dataset nfluences the accuracy of pattern recognton. To further mprove the recognton accuracy of the proposed approach, more extensve PD dataset creaton methods wll be examned n future studes. References [1] J. Y. Jeong, D. S. Kang, J. H. Sun, J. C. Heo and C. H. Park, Assessment of 23 kv Capactve Coupler for On-lne Partal Dscharge Measurements, Journal of Electrcal Engneerng & Technology, vol. 4, no.1, pp , March [2] A. Rodrgo, P. Llovera, V. Fuster and A. Qujano, Study of Partal Dscharge Charge Evaluaton and the Assocated Uncertanty by Means of Hgh Frequency Current Transformers, IEEE Trans. Delectrcs and Electrcal Insulaton, vol. 19, no. 2, pp , Aprl [3] M. Oskuoee, A.R. Yazdzadeh and H.R. Mahdan, A New Feature Extracton and Pattern Recognton of Partal Dscharge n Sold Materal by Neural Network, n Proceedngs of the Eghth Internatonal Conference on Natural Computaton, Chongqng, Chna, May [4] J. Tang, F. Lu, Q.H. Meng, X.X. Zhang and J.G. Tao, Partal Dscharge Recognton Through an Analyss of SF6 Decomposton Products Part 2: Feature Extracton and Decson Tree-Based Pattern Recognton, IEEE Trans. Delectrcs and Electrcal Insulaton, vol. 19, no. 1, pp , February [5] W.Y. Chang and H.T. Yang, Applcaton of Self Organzng Map Approach to Partal Dscharge Pattern Recognton of Cast-Resn Current Transformers, WSEAS Trans. Computer Research, vol. 3, no. 3, pp , March [6] S.W. Gao, M. Zhang, C.H. Yuan, X.T. Sun and C. Zhang, Wavelet Packet Analyzng of Power Transformer Partal Dscharge Sgnals, n Proceedngs of the 2011 Internatonal Conference on Control, Automaton and Systems Engneerng, Sngapore, July [7] W.Y. Chang, Applcaton of Grey Clusterng Approach and Genetc Algorthm to Partal Dscharge Pattern Recognton, WSEAS Trans. Systems, vol. 8, no. 12, pp , December [8] R.J. Lao, K. Wang, L.J. Yang, T.C. Zhou and S.X. Zheng, Study on Thermal Agng Condton Assessment of Ol-paper Insulaton Based on Statstcal Features of Partal Dscharge, n Proceedngs of the IEEE 9th Internatonal Conference on Propertes and Applcatons of Delectrc Materals, Harbn, Chna, July [9] N.C. Sahoo and M.M.A. Salama, Trends n Partal Dscharge Pattern Classfcaton: A Survey, IEEE Trans. Delectrcs and Electrcal Insulaton, vol. 12, no. 2, pp , Aprl

8 Partal Dscharge Pattern Recognton of Cast Resn Current Transformers Usng Radal Bass Functon Neural Network [10] R.E. James and B.T. Phung, Development of Computer-based Measurements and Ther Applcaton to PD Pattern Analyss, IEEE Trans. Delectrcs and Electrcal Insulaton, vol. 2, no. 5, pp , October [11] T.A. Reddy, K.R. Dev and S.V. Gangashetty, Nonlnear Prncpal Component Analyss for Sesmc Data Compresson, n Proceedngs of the 1st Internatonal Conference on Recent Advances n Informaton Technology, Dhanbad, Inda, March [12] T. Stamkopoulos, K. Damantaras, N. Maglaveras, and M. Strntzs, ECG Analyss Usng Nonlnear PCA Neural Networks for Ischema Detecton, IEEE Trans. Sgnal Processng, vol. 46, no. 11, pp , November [13] Al Karam, Radal Bass Functon Neural Network for Power System Transent Energy Margn Estmaton, Journal of Electrcal Engneerng & Technology, vol. 3, no.4, pp , December [14] Y. Zhang, Q. Zhou, C. Sun, S. Le, Y. Lu and Y, Song, RBF Neural Network and ANFIS-based S hort-term Load Forecastng Approach n Real-tme P rce Envronment, IEEE Trans. Power Systems, vol. 23 no. 3, pp , August Wen-Yeau Chang He receved the Ph.D. degree n electrcal engneerng from Natonal Cheng-Kung Unversty, Tanan, Tawan n Hs current research nterests are hgh voltage engneerng, renewable generaton systems, power electroncs, and dstrbuted generators. 300