Electrical Power and Energy Systems 26 (2004) 199205 www.elsevier.

com/locate/ijepes

A fuzzy-expert system for classifying power quality disturbances

Yuan Liaoa,*, Jong-Beom Leeb
b a ABB USCRC, 1021 Main Campus, Raleigh, NC 27606, USA Department of Electrical Electronic and Information Engineering, Wonkwang University, 344-2 Sinyong-dong, Iksan, Jeonbuk 570-749, South Korea

Abstract This paper presents a fuzzy-expert system for automated detection and classication of power quality disturbances. The types of concerned disturbances include voltage sags, swells, interruptions, switching transients, impulses, ickers, harmonics, and notches. It is assumed that the analyzed waveforms are available in sampled form. Fourier transform and wavelet analysis are utilized to obtain unique features for the waveforms. A fuzzy-expert system is designed for making a decision regarding the type of the disturbance. Simulation studies are presented to verify the accuracy of the proposed approach. Comparison studies between an Articial Neural Network based classication technique and the proposed approach are also reported to show the advantages of the proposed approach. q 2004 Elsevier Ltd. All rights reserved.
Keywords: Fuzzy-expert system; Power quality disturbance classication; Wavelet analysis; Articial neural network

1. Introduction In the new open-access and competitive power market, electricity consumers are in a unique position to demand a higher quality of service. The utilities or other power providers have to ensure a high quality of their service to remain competitive and retain/attract the customers. Efcient power quality (PQ) monitoring and analysis system can help to achieve this goal. PQ monitoring can usually be a complex task involving hardware instrumentation and software packages. The instrumentation entails communication software to accomplish the expected functionality. Various types of intelligent electronic devices can be used for collecting the desired PQ data. Examples include dedicated digital fault recorders and protective relays [1 5]. After PQ data of interest are obtained, a comprehensive PQ assessment can be carried out depending on the purpose of the study. This may include power system and equipment modeling verication, PQ problem mitigation and optimization, and data analysis [4,5]. In most cases, automated PQ assessment is desirable because manual analysis may be difcult to carry out due to lack of time and special
* Corresponding author. Address: ABB USCRC, 1021 Main Campus, Raleigh, NC 27606, USA. Tel: 919-807-5787; Fax: 82-63-850-6735. E-mail addresses: ipower@wonkwang.ac.kr, yuan.liao@us.abb.com (Y. Liao). 0142-0615/$ - see front matter q 2004 Elsevier Ltd. All rights reserved. doi:10.1016/j.ijepes.2003.10.012

expertise. Specialized software tools can make use of intelligent techniques to automate the PQ assessment for improved accuracy and efciency [5]. This paper is focusing on automated detection and classication of PQ disturbances that may facilitate the overall PQ assessment. Intelligent techniques like fuzzy logic and expert system as well as signal processing techniques like Fourier transform and wavelet analysis have been utilized for developing the tools. When carrying out these studies, various assumptions made will be illustrated where appropriate. PQ disturbance detection and classication has been addressed for many years [5,6]. The denition of PQ disturbance types is illustrated in Ref. [2]. Efforts for automating the detection and classication process can be found in Refs. [3 5]. A performance comparison study on various available approaches is presented in a recent dissertation [5]. For many applications, the accuracy and efciency resulting from the existing approaches may not be adequate. Improved accuracy is desirable. Aiming at this goal, this paper proposes a fuzzy-expert system for making a decision based on the features extracted from both Fourier and wavelet transform. The paper is organized as follows. The distinctive features extracted by the proposed approach are illustrated rst. Then the design of the fuzzy-expert system is proposed. Simulation study results are presented next. Concluding remarks are given at the end.

200

Y. Liao, J.-B. Lee / Electrical Power and Energy Systems 26 (2004) 199205

2. Proposed new features This section describes the proposed new features based on the Fourier and wavelet analysis for PQ disturbance analysis. A number of PQ disturbances of various types have been simulated and corresponding waveforms obtained. The following eight distinct features inherent to different types of PQ disturbances have been identied: the fundamental component Vn ; phase angle shift an ; total harmonic distortion THDn ; number of peaks of the wavelet coefcients Nn ; energy of the wavelet coefcients EWn ; oscillation number of the missing voltage Osn ; lower harmonic distortion TSn ; and oscillation number of the rms variations (RN). The formulae for computing these features are given as follows [5,7] p Vn 2absV n 1=N 1 2

as an array composed of WCn k for k 1; 2; ; le; with le being the length of WCn : vmiss i vi 2 2=N absV 1 1 cos{angleV 1 1 2pi 2 1=N} 10

vs miss is dened as an array composed of vmiss i; i 0; 1; ; L 2 1 v u N21 u1 X 2 t v i n 2 1N N i0

n Vrms

11

an angleV n 1 2 angleV 1 1
v uintN=2 u X t {absV n k}2 THDn
k2

s n Vrms is dened as an array composed of Vrms ; n 1; 2; ; 10 abs gives the absolute value of the argument intN=2 equals N=2 if N is even, and N 2 1=2 if N odd angle returns the phase angle of the input argument root returns the number of roots (or zero-crossings) of the argument peak returns the number of peaks of the argument mean gives the mean of the argument

V 1 1

Nn peakabsWCs
le X k1

EWn

absWCn k

OSn rootvs miss v uX u 10 t {absV n k}2 TSn

k2

V 1 1

7 8

s s RN rootVrms 2 meanVrms

In the above equations, V n k is the Discrete Fourier Transform for the samples contained in the nth data window dened as V n k
N21 X i0

vi n 2 1Ne2j2pki=N

vi represents the sampled input signal i 0; 1; ; L 2 1; with L the length of the signal N is the number of samples in one data window (one cycle) j is the imaginary unit, and n 1; 2; ; 10 WCn are the wavelet coefcients for the samples contained in the nth data window, and WCs is dened

In the work presented here, 10 cycles of samples of the three-phase voltage signals (in per unit) are used. The Daubechies-4 wavelet family and the second scale wavelet detail coefcients are utilized. Detailed analysis of the wavelet transform and multiresolution decomposition techniques is presented in Ref. [7]. As examples, typical analysis results of the selected features for several types of disturbances are presented as follows. For convenience of labeling, in the following gures, cD1 is used for designating the rst scale detail wavelet coefcient, V for fundamental component, PAS for phase angle shift; THD for total harmonic distortion, N for number of peaks of the wavelet coefcients, EW for energy of the wavelet coefcients, OS for oscillation number of the missing voltage, and TS for lower harmonic distortion. For the studies, the oscillation number of the rms variation RN is not drawn because it is calculated only once for the whole sampling set, but its value is given. A switching transient signal and its features are shown in Fig. 1. It is clearly seen that cD1 has large values during the transient. The cD1 is oscillatory due to the oscillatory nature of a switching transient signal. The main characteristics include small phase angle shift, almost constant fundamental component, large oscillation number of the missing voltage, etc. Also RN 0: Fig. 2 shows a sag signal and its features. It can be seen that cD1 has a noticeable jump at the occurrence of the sag while it does not have signicant outputs during the sag period. Its dominant characteristics include reduced fundamental component, large phase angle shift, small energy of wavelet coefcients, etc. Also RN 0: Other types of PQ disturbances can be analyzed similarly. Extensive studies have evinced that the extracted

Y. Liao, J.-B. Lee / Electrical Power and Energy Systems 26 (2004) 199205

201

3. The proposed fuzzy-expert system for PQ disturbance detection and classication This section rst briey presents the concepts of the fuzzy logic and fuzzy-expert system, and then illustrates the proposed rule sets for detection and classication of PQ disturbances. 3.1. Introduction to fuzzy-expert system Fuzzy logic refers to a logic system that generalizes the classical two-valued logic for reasoning under uncertainty. It is motivated by observing that human reasoning can utilize concepts and knowledge that do not have welldened or sharp boundaries [8]. Unlike the classical Boolean set allowing only 0 or 1 value, the fuzzy set is a set with a smooth boundary allowing partial membership. The degree of membership in a set is expressed by a number between 0 and 1 with 0 indicating entirely not in the set, 1 indicating completely in the set and a number in between meaning partially in the set. In this way, a smooth and gradual transition from the regions outside the set to those in the set can be described. A fuzzy set can thus be dened by a function that maps objects in the domain of concern (i.e. the universe of discourse) to their membership values in the set. Such a function is called the membership function. The two most widely used membership functions are the triangular and trapezoidal functions. A fuzzy-expert system is an expert system that uses a collection of fuzzy sets and rules, instead of Boolean sets for reasoning about data. The rule in the fuzzy-expert system usually takes the form If x is low and y is high; then z medium where x and y are input variables, z is the output variable, and low, high and medium are membership functions dened for x; y and z; respectively. The antecedent (the rules premise) describes the degree that the rule applies, while the conclusion (the rules consequent) assigns a membership function to the output variable. The set of rules in a fuzzy-expert system is known as the rule base or knowledge base. The computation of the output variable usually takes the following steps [8]. (1) Fuzzication. This step is also called Fuzzy Matching, which calculates the degree that the input data match the conditions of the fuzzy rules. (2) Inference. Calculate the fuzzy set of the rules conclusion based on its matching degree. There are two common approaches for the inference, namely the clipping method and the scaling method. Both methods generate the conclusion by suppressing the membership function of the consequent. The extent to which they suppress the membership function depends on the degree to which the rule is matched. The lower the matching degree, the more severe the suppression

Fig. 1. A switching transient signal and its features.

parameters display distinctive patterns under different types of disturbances. Next, the statistical properties of the parameters for various PQ disturbances can be obtained. Based on these distinctive patterns, appropriate fuzzy rules can be established for distinguishing between them as shown below.

Fig. 2. A voltage sag signal and its features.

202

Y. Liao, J.-B. Lee / Electrical Power and Energy Systems 26 (2004) 199205

of the membership functions. The clipping method cuts off the top of the membership function, whose value is higher than the matching degree. The scaling method scales down the membership function in proportion to the matching degree. The scaling method is used here. (3) Composition. Because a fuzzy rule-based system consists of a set of fuzzy rules with partially overlapping conditions, a particular input to the system often triggers multiple fuzzy rules (i.e. more than one rule will match the input to a nonzero degree). Therefore, the composition is needed to combine the inference results of all the triggered rules to form a single fuzzy subset for the output variable. The fuzzy disjunction operator Max is commonly used for constructing the output fuzzy set by taking the point-wise maximum over all the fuzzy subsets generated from the inference step, and is used here. (4) Defuzzication. This step is to convert the fuzzy set of the output variable to a crisp number. Among the various types of defuzzication methods, the Center of Area (COA, or Centroid) and Maximum are the two most widely used techniques. The COA derives the crisp number by calculating the weighted average of the output fuzzy set while the Maximum method chooses the value with maximum membership degree as the crisp number. The maximum method is used here. The preceding steps can be mathematically stated as follows [8]. Suppose that the following rules describe a mapping from U1 U2 Ur to W with Ui ; W being the universe of discourse of the according variables Ri : if xi is Ai1 and and xr is Air then y is Ci : where xj j 1; 2; ; r are the input variables, y is the output variable, and Aij and Ci are fuzzy sets for xj and y; respectively. Given inputs of the form x1 is A01 ; x2 is A02 ;, and xr is A0r where A01 ; A02 ; ; A0r are fuzzy subsets of U1 ; U2 ; ; Ur (e.g. fuzzy numbers). The contribution of rule Ri to the output is a fuzzy set, whose membership function is computed by

the maximum defuzzication method as follows:

mC y max{mC01 y; mC02 y; ; mC0L y}

Finally, the crisp value yc can be obtained as:

14

mC yc sup mC y
y

15

3.2. The proposed fuzzy-expert system Based on the distinct features extracted previously, the core of the rule set of the fuzzy-expert system is illustrated as follows [5,8]. (1) Detection. One rule follows Rule 1: if THDn is A2 or PSn is B2 or Vn is C3 or Vn is C1 then DETECT 1 (2) Classication. Fifteen rules follow Rule 1: Vn1 is A4 and Nn is F1 and OSn is G1 then IMPULSE 1 Rule 2: Vn is A1 or Vn1 is A1 then INTERRUPTION 1 Rule 3: Vn is A6 or Vn1 is A6 then SWELL 1 Rule 4: Vn is A5 and PSn is C1 and PSn1 is C1 and EWn1 is D1 and {TSn1 is H2 or [TSn1 is H4 & TSn2 is H1 ]} then SWELL 1 Rule 5: Vn1 is A5 and {PSn is C2 or PSn1 is C2 } then SWELL 1 Rule 6: Vn1 is A2 then SAG 1 Rule 7: Vn1 is A3 and {PSn is C2 or PSn1 is C2 } then SAG 1 Rule 8: Vn1 is A3 and {PSn is C1 and PSn1 is C1 } and {THDn1 is B1 or [THDn1 is B2 and OSn1 is G4 ]} then SAG 1 Rule 9: Vn1 is A3 and PSn is C1 and PSn1 is C1 and OSn is G2 and THDn1 is B2 and THDn2 is B2 and THDn3 is B2 then NOTCH 1 Rule 10: Vn1 is A3 and Nn is F2 and OSn is G2 then NOTCH 1 Rule 11: Vn1 is A4 and PSn is C1 and PSn1 is C1 and THDn is B3 and THDn3 is B1 and {OSn is G4 or OSn1 is G4 } then TRANSIENT 1 Rule 12: Vn1 is A4 and TSn1 is H3 and TSn2 is H3 and TSn3 is H3 and OSn1 is G4 then HARMONIC 1 Rule 13: THDn1 is B4 and THDn2 is B4 and THDn3 is B4 and OSn2 is G4 then HARMONIC 1 Rule 14: TSn1 is H4 and TSn2 is H4 and TSn3 is H4 and OSn2 is G4 then HARMONIC 1 Rule 15: If RN is K1 then FLICKER 1 In the above rules, Ai; Bi ; Ci ; Di ; Fi; Gi ; Hi ; and Ki are the membership functions for the input patterns, and the commonly used triangular and trapezoidal functions are utilized.

mC0i y ai1 ^ ai2 ^ ^ ain ^ mCi y

12

where aij is the matching degree between xj and Ri s condition about xj is calculated as follows

aij sup mA0j xj ^ mAij xj

xj

13

where ^ denotes the fuzzy conjunction operator and sup denotes the continuous maximum operator. Then the fuzzy set of the output variable can be obtained by aggregating the outputs from all the rules using

Y. Liao, J.-B. Lee / Electrical Power and Energy Systems 26 (2004) 199205

203

It can be seen that the inputs of the fuzzy rule based system are: RN, THDn ; THDn1 ; THDn2 ; THDn3 Vn ; Vn1 ; PSn ; PSn1 ; EWn1 ; Nn ; OSn ; OSn1 ; OSn2 ; TSn ; TSn1 ; TSn2 ; and TSn3 : And the outputs of the systems are: Detect: indicates whether the disturbance exists or not. Flicker, Impulse, Interruption, Swell, Sag, Notch, Transient, and Harmonic: indicates the degree to which the disturbance belongs to each of these categories. The type of the disturbance will be the one with the largest membership. In cases where two or more types of disturbances have the same largest membership value, all of them will be outputted for further analysis. The values for the outputs can be obtained using Eqs. (12 15). The fuzzy partitions and the according membership functions of the inputs for detection and classication are obtained based on the statistical results of the features as well as on the experts knowledge.

Fig. 3. A feed-forward neural network based classier.

4. Case studies In this section, evaluation studies are performed to verify the accuracy and efciency of the proposed system for PQ disturbance detection and classication. A neural network-based method has also been implemented, and its performance is compared with that of the proposed fuzzy logic-based system [5]. Eight hundred disturbance waveforms of various types have been generated using both the algebraic equations and Electromagnetic Transients Program (EMTP) [9]. The advantage of using algebraic equations over using EMTP for evaluation is the exibility of adjusting signal noise contents as well as various waveform parameters such as the disturbance occurrence time, harmonic contents, sag depth, etc. The classication results of the generated disturbances using the proposed system are shown in Table 1. It can be seen that the new system results in a correct identication
Table 1 Classication results yielded by the developed fuzzy-expert system Type of PQ Number of Number of Number of Correct disturbances disturbances cases correctly cases mistakenly identiidentied identied cation rate (%) Sag Swell Interruption Harmonic Switching transient Impulse Flicker Notch Sum 100 100 100 100 100 100 100 100 800 98 100 100 100 98 98 100 97 791 2 0 0 0 2 2 0 3 9

99

rate of 99%. These studies show that the proposed methods for feature extraction and decision making are efcient for PQ disturbance detection and classication. As shown in Fig. 3, a feed-forward one-hidden layer neural network with 17 input neurons, 10 hidden neurons, and eight output neurons is implemented [10]. The same features as those used in the fuzzy-expert system are utilized as inputs to the neural network. The output variables are sag, swell, interruption, notch, switching transient, impulse, harmonic and icker. Hence, the input will be a 17dimension vector. The output will be an eight-dimension vector. The transfer functions for the rst layer and second layer are the hyperbolic tangent sigmoid function and linear function, respectively [10]. The training patterns are obtained as follows. Another eight hundred new disturbance waveforms of various types are generated using simulation approaches. Then 800 vectors of features with dimension of 17 are generated according to Eqs. (1 11). These features are treated as inputs to the neural network. The target values are determined as follows. The output with the same name as the input disturbance waveform will have a value of one, and all the other output will be set to zero. For example, for the sag disturbance input, the target vector will be 1; 0; 0; 0; 0; 0; 0; 0: For the interruption disturbance input, the target vector will be 0; 0; 1; 0; 0; 0; 0; 0: Each input feature vector and the corresponding target vector will form one pair of training pattern. In this way, 800 pairs of training patterns will be obtained. Then the neural network is trained using the back-propagation approach. The mean-squared error performance function is used. A value of 0.01 of the performance function is reached after 40,000 iterations. Then, the same disturbances as shown in Table 1 are used to test the accuracy of the trained neural network. The classication results are shown in Table 2. It is seen that

204

Y. Liao, J.-B. Lee / Electrical Power and Energy Systems 26 (2004) 199205

Table 2 Classication results yielded by the developed neural network Type of PQ Number of Number of Number of Correct disturbances disturbances cases correctly cases mistakenly identiidentied identied cation rate (%) Sag Swell Interruption Harmonic Switching transient Impulse Flicker Notch Sum 100 100 100 100 100 100 100 100 800 97 96 98 96 95 97 99 97 775 3 4 2 4 5 3 1 3 25

incorporated into the rule sets. In contrast, neural network functions as a black box and is more difcult to understand. It is worth noting that another new decision making technique namely the neuro-fuzzy system has been developed recently by combining the neural network and fuzzy logic. In this system, the parameters of the membership functions of the fuzzy variables can be automatically determined through the training of the neural network. However, its suitability for PQ detection and classication has to be investigated before a conclusive remark can be made, which is out of the scope of the paper. Authoritative references on the topic of neuro-fuzzy system can be found in Ref. [8].

97

the correct identication rate is 97%. This demonstrates that the combination of Fourier and wavelet analysis as well as the fuzzy logic techniques seems a better choice than the combination of Fourier and wavelet analysis and the neural network techniques. The feasibility of the proposed decision making system is further justied as follows. Essentially speaking, both the fuzzy logic and neural network realizes a non-linear functional mapping from one space (the space of features) to another space (the space of the types of the disturbances) [8,10]. The advantages of using fuzzy logic over using neural network are illustrated as follows. First, the performance of the neural network heavily depends on the training process, during which the weights of the neural network are adaptively changed according to specic learning rules until certain criterion is met. In order for the neural network to generalize well after being trained, the training pairs used in the training process have to be chosen to be sufciently representative. On the other hand, it may be difcult to consider all the possible situations that may happen in the real world. Therefore, choosing training pairs is one difculty when using neural network. Furthermore, the training process may take a long time to converge. In comparison, fuzzy logic eliminates the need for training. Although typical cases may be used for generating the rule sets (or knowledge base), they are not mandatory because the rule sets can be improved or entirely generated by the experts with the required knowledge. Second, there is no theoretical guidance for choosing the size of the network, which has to be decided by error-andtrial. Once the neural network is trained, its structure cannot be changed. Otherwise, the neural network needs to be retrained. On the other hand, in the fuzzy-expert system, it is quite exible to determine the number of the fuzzy rules in the rule set, and the rules can be conveniently modied if needed in the future. Moreover, fuzzy rules are more comprehensible than the neural network. Experts knowledge can be readily

5. Conclusions A fuzzy-expert system has been proposed in this paper for detecting and classifying PQ disturbances. Eight types of common PQ disturbance are considered. A neural network based approach is also implemented for comparison purposes. Comparison studies using simulated data have demonstrated the effectiveness of the proposed approach for detection and classication of PQ disturbances.

References
[1] IEEE Project 1346 Working Group, Electric power system compatibility with industrial process equipment. Part 1. Voltage sags. IEEE Industrial and Commercial Power Systems Technical Conference, Irvine, CA, USA, May 15; 1994. p. 2616. [2] IEEE Std 1159-1995, IEEE recommended practice for monitoring electric power quality. New York: IEEE; 1995. [3] Kezunovic M, Heydt G, Schilleci J, Lambert-Torres G, Key T, Kreis D. Intelligent system applications to power quality and substation automation. Panel Session at the IEEE Power Engineering Society Summer Meeting, Edmonton, Alberta, Canada, July 1822, vol. 1.; 1999. p. 468 75. [4] Kezunovic M, Liao Y. Use of intelligent techniques in the power quality assessment applications. International Conference of Intelligent Systems Applications to Power Systems, 1SAP, 2001. Budapest, Hungary, June 18 21 2001. [5] Liao Y. Automated analysis of power quality data and transmission line fault location. PhD Dissertation. Texas A&M University; 2000. [6] Lee J, Won Y, Jeong J, Nam S. Classication of power disturbance using feature extraction in time-frequency plane. Electron Lett 2002; 38(15):8335. [7] Daubechies I. Ten lectures on wavelets. Society for Industrial & Applied Mathematics; May 1992. [8] Yen J, Langari R. Fuzzy logic: intelligence, control, and information. Englewood Cliffs, NJ: Prentice-Hall; 1999. [9] Leuven KU. Alternative transient program, user manual and rulebook. Heverlee, Belgium: EMTP Center; 1987. [10] Hassoun MH. Fundamental of articial neural networks. Cambridge, MA: MIT Press; 1995.

Y. Liao, J.-B. Lee / Electrical Power and Energy Systems 26 (2004) 199205

205

Yuan Liao received his PhD degree from Texas A&M University, College Station, in electrical engineering in 2000. He joined ABB as a consulting R&D engineer in 2000. His research interests include power system monitoring, protection, control, and system planning. He can be reached at: ABB US Corporate Research Center, Raleigh, NC 27606, USA.

Jong-Beom Lee received BS, MS and PhD degree in Electrical Engineering from Hanyang University, Korea, in 1981, 1983 and 1986, respectively. He has been worked in Korea Electrotechnology Research Institute from 1987 to 1990. He has been a visiting scholar in Technical University of Berlin, Germany, City University, UK and Texas A&M University in 1993, 1995 and 1997, respectively. He is currently a professor of Department of Electrical, Electronic and Information Engineering, Wonkwang University, Korea. His current research interests are power system operation, transient analysis of power cable system, protective relaying, application of neural networks and fuzzy application to power system.