Identication of multiple harmonic sources in power systems using independent component analysis and mutual information

Masoud Farhoodnea , Azah Mohamed, Hussain Shareef

This paper presents a novel technique for accurate determination of the probable locations of multiple harmonic sources in power systems. In the proposed methodology, independent component analysis (ICA), which is one of the blind source separation techniques, is applied for estimating the proles of injected harmonic currents produced by nonlinear loads. By using the ICA algorithm, the reconstruction of the injected harmonic currents can be implemented easily without prior knowledge about the system components. From the reconstructed harmonic current proles, the mutual information theory is then applied for determining the exact location of harmonic sources by computing the pair-wise mutual information between the extracted current trends and bus voltages. Numerical simulations were made to prove the accuracy of the proposed method in locating multiple harmonic sources in the IEEE 30 and 34 bus test distribution systems. Results proved that the proposed method can accurately determine the probable location of harmonic sources in radial and non radial power systems with mean square error less than 0.00035. Keywords: Harmonic distortion; harmonic source location; power quality; multiple harmonic sources.

1.

INTRODUCTION

Due to the penetration of various distributed nonlinear and harmonic producing loads, the diffusion of harmonics in power systems is escalating and has become an important power quality problem. In addition harmonics are the main reason for causing resonance problems, overheating in capacitor banks, transformers and conductors, and telecommunication conictions. These problems may decrease the reliability and increase the maintenance costs of power systems. From the above-mentioned issues, it is important for power providers to identify the source of harmonic distortions so as to solve
E-mail: [email protected]

or lessen the problems related to harmonic distortion. Methods for identifying the location of harmonic sources in power systems are generally categorized as single point and multiple point strategies. The real power direction method [1] is one of the earliest single point method proposed in locating harmonic sources. However, the real power direction method has been found to be only 50 percent reliable and therefore the accuracy of the method is questionable [2]. Other single point methods for harmonic source localization include incentive-based method [34], the critical impedance method [5] and voltage magnitude comparison method [6]. These methods require complete knowledge of system parameters in harmonic frequency or implementation of switching tests for obtaining the harmonic impedances of the system. However, such data are

IDENTIFICATION OF MULTIPLE HARMONIC SOURCES IN POWER SYSTEMS

not available in most of the time, and implementing switching tests is not possible in practical power systems. More recent single point based methods such as total harmonic distortion (THD) method [7], harmonic vector method (HVM) [8], and IEEE14592000 standard based method [9] tried to solve the above mentioned problems in the previous methods. Nonetheless, these current methods still suffer from some problems such as the use of unsuitable index harmonic modeling and inability to determine the location of harmonic sources for individual harmonic orders. In general, implementing single point based methods are very economical in comparison with other methods, but such methods are unable to work in practical systems because harmonic sources exist at different points in a power network. To tackle this problem, several methods have been developed for locating multiple harmonic sources in power systems. Harmonic state estimation (HSE) [1011] is one of the most popular multiple point based methods. But, the HSE based methods require complete knowledge about the system parameters for each individual harmonic order, which are usually unknown or difcult to calculate in practice. In addition, the measurement placement technique [1213] which is widely applied in HSE for optimizing the number of measurements in the system needs to have prior knowledge about harmonic sources and their locations. Another concern with HSE is that this approach needs various types of harmonic measurements such as voltage, real and reactive powers which are costly for large systems. In [14], the weighted least square estimation technique and Euclidean norm have been applied to recognize the location of multiple harmonic sources. In this method, the weighted least square estimation technique and HSE are used to locate the possible buses with harmonic sources while the Euclidean norm determines the exact buses with harmonic sources. However, the disadvantage of this method is the use of all types of harmonic measurements at all buses which is not economical. Another multiple point based method is the Independent component analysis (ICA) method [15]. In this method, measured voltages at the selected buses, determined by the measurement placement technique are used for estimating the harmonic impedance matrix of the system, and harmonic current proles of the harmonic sources at each harmonic frequency. The minimum electrical distances between the estimated impedance matrix and the actual impedance matrix are then obtained by using the exhaustive search technique. The buses with minimal distances are considered as the location of harmonic sources. However, this method is not quite practical because it needs to determine the actual impedance matrix of a system at each harmonic frequency. In addition, this method requires prior information about system parameters and historical records of the loads for solving indeterminacies related to ICA algorithm. This paper presents a new method based on ICA and mutual information theory (MI) to determine the accurate location of multiple harmonic sources in a power system. In this method, ICA is used to rebuild or estimate the harmonic currents produced by the harmonic sources while the MI theory is used for determining the locations of harmonic sources, respectively. The proposed method for identifying the location of harmonic sources is different from the previous ICA method [15], in the fact that it uses the mutual information theory for nal decision making stage in locating the harmonic sources.

In addition, unlike the above mentioned techniques, the proposed method only requires voltage measurements of all buses and eliminates other kinds of measurements. In addition, it is assumed that there are no prior information about the system parameters, harmonic impedance matrix and characteristics of harmonic sources. Hence estimation will be done completely in blind situation. This assumption is signicant in the deregulated power systems where complete power system parameters, load behavior and utility response is not known because of economic and security restrictions in electricity market.

2.

INDEPENDENT COMPONENT ANALYSIS

Independent component analysis (ICA) is one of the blind source separation techniques (BSS) based on the statistical independence between the source signals. Due to the generality of the ICA model, it has found many applications in different areas such as audio separation, econometrics, brain imaging, and telecommunication data separation. Basically, ICA is a technique that transforms the observed signals into a linear transform of source signals that are statistically and mutually independent from each other. To reconstruct the source signals from the observed signals, certain assumptions are considered such that the source signals are statistically independent, all but one source signals must have non gaussian distributions and the number of observations should be greater or equal to the number of sources [16]. Assuming that there are N sources and M measurements, the linear mixing model of ICA can be written as xj = aj 1 s1 + aj 2 s2 + ... + aj n sn for all j = 1, 2, ..., M The matrix representation of Eq. (1) is X = AS where, S(t) = [s1 (t), s2 (t), ..., sn (t)]T : tor of unknown source signals X(t) = [x1 (t), x 2 (t), ..., xm (t)]T vector of observed signals N -dimensional vec: M dimensional (2)

(1)

A : M N full column coefcient matrix aij called as the mixing matrix In ICA, the objective is to nd estimates of the S and A from the available observation vector X. The unknown source matrix S can be estimated as Sest = W X where, Sest : estimate of the sources S with N T dimension W : N M separating matrix which is the pseudo inverse of the mixing matrix A (3)

M. FARHOODNEA ET AL

For estimating the model of ICA given by Eq. (3), several optimization techniques can be used such as maximization of nongaussianity, minimization of mutual information, and maximum likelihood estimation [17]. Here, the ICA algorithm with maximization of nongaussianity of the source signal has been applied. Based on this approach, the entropy of a random variable which is related to the information that the observation of the variable gives is considered. The more random the variable is, the larger is its entropy [17]. The entropy, H of a random vector, y with density function, py () is dened as: H (y) = py () log py ()d (4)

where D and E are the diagonal matrix of eigenvalues and the orthogonal matrix of eigenvectors, respectively. They are obtained from the covariance matrix of the observed vector X. For maximizing the contrast functions in Eq. (7), the FastICA algorithm [18] is applied and it is described in terms of a owchart as shown in Figure 1. Note that for convergence of the algorithm, the old and new values of w should be in the same direction.

3.

MUTUAL INFORMATION THEORY

where y is a random variable with mean and standard deviation . It should be noted that a gaussian random variable has the largest entropy among all random variables with equal variance. Therefore, entropy can be used as a measure of nongaussianity of random variables, which is always nonnegative. An important measure of nongaussianity is given by the normalized version of differential entropy which is called as negentropy J and is dened as: J (y) = H (ygauss ) H (y) (5)

Basically, mutual information (MI) is a measure of the information that members of a set of random variables have on the other random variables in the set. Using the concept of entropy dened for continuous random variables in Eq. (4), the discrete form of entropy is derived as,
N

H (y) =
y =1

p(y) log p(y)

(9)

where ygauss is a gaussian random vector with the same covariance matrix as y . Negentropy is always nonnegative, and it is zero if y has a gaussian distribution. Estimating negentropy of Eq. (5) is very difcult because it needs to estimate the probability functions of the source signals. Hence, by using only one quadratic function, the approximation of J becomes [18]: J (y) [E {G(y)} E {G(v)}]2 (6)

The entropy in Eq. (9) can be interpreted as a measure of uncertainty of the events with probability function p . In other words, entropy is small for events with probability close to 0 or 1, and large for probabilities between 0 and 1. In fact, the entropy of a random variable can be dened as the degree of information that the observation of the variable gives. To determine uncertainty about variable A in subsequence of trials in which B occurs, the conditional entropy is used [19]. Assuming random variables A and B consisting of the N elements of ai and bi for i = 1, 2, ..., N , the conditional entropy can be dened as:
N

where G is practically any nonquadratic function with slow variation such as tanh(u), v is a gaussian variable with zero mean and unit variance and y is a random variable with zero mean and unit variance. To estimate the independent components, maximization of the estimated negentropy (6) is necessary. By maximizing the sum of N one unit contrast functions and taking into account the constraint of decorrelation, an optimization problem is obtained as follows:
n

H (A |bi ) =
i =1

p(ai |bi ) log p(ai |bi )

(10)

Assuming that B is the average of H (A |bi ), the conditional entropy of variables A and B can be written as:
N

H (A |B ) =
i =1

p(bi )H (A |bi )

(11)

Maximize
i =1

T J (wi x)

T T under constraint E (wk x)(wj x) = j k

(7)

By using Eq. (10) and (11), the mutual information I of variable A and B can be obtained as: I (A,B) = H (A) H (A |B ), I (A,B) 0 (12)

where wi for i = 1, 2, ..., N are the rows of the separating matrix, W which is inverse of coefcient matrix A. Before implementing ICA, the observed vector X is preprocessed by centering and whitening. Preprocessing in ICA is known to be a suitable technique for simplifying the ICA estimation .[16]. By centering, the mean of vector X is subtracted and transformed to zero-mean variable, while by whitening; the observed vector X is transformed linearly to a new vector where its components are uncorrelated and their variances X are unity. In whitening, Eq. (8) is considered, = ED 1/2 E T X X (8)

Equation (12) is symmetric and can be interpreted as the information about A contained in B and vice versa. If A and B are independent, their mutual information should be zero. In addition, mutual information is a better function for measuring dependency between two random variables rather than the correlation function [20]. The reason is that mutual information is able to measure general dependency between two variables, while the correlation function can only measure linear dependency between them.

IDENTIFICATION OF MULTIPLE HARMONIC SOURCES IN POWER SYSTEMS

Start

tion under non-sinusoidal condition which can be written in matrix form as: (13) V h = ZhI h where V , Z and I are the bus voltage, impedance and bus injected current vectors, respectively and h is the harmonic order of the frequency. If voltage V and impedance Z are identied, then the harmonic current I can be computed using Eq. (13). However, it is difcult to form the impedance matrix because of complexity of obtaining the values of harmonic impedance of the system for individual harmonic orders. Assuming that in the absence of any information of the system parameters and characteristics of non linear loads, the only available information is the bus voltages which are obtained by distributed harmonic voltage measurement devices for each harmonic order. The linear system of equations for the harmonic load ow (13) under non sinusoidal condition is in close similarity with the model in Eq. (2) which makes ICA a useful instrument for multiple harmonic source estimation. In addition, fast uctuations of the loads consumption during specic time that are related to second to second or minute to minute variations, have non gaussian characteristics and act independently [21 23], which satisfy the preliminary conditions for applying ICA algorithm. Applying ICA to Eq. (13), X represents the harmonic voltage measurement vectors V, the mixing matrix A represents the harmonic impedance matrix Z in the harmonic domain, and the estimated signal S represents the bus injected current vector. In Eq. (13), the current matrix I is zero for linear loads and non zero for harmonic or non-linear loads. From (13), the injected harmonic currents generated by the non-linear loads at the respective buses have a relationship with the bus voltages. By using the mutual information theory, the produced harmonic current at each bus has maximum mutual information with its own bus voltage and the mutual information between this harmonic current and other bus voltages is reduced due to the current division between branches. The procedure in implementing the proposed method using ICA and mutual information theory for estimating the location of multiple harmonic sources is summarized as follows: Measure harmonic voltages at all buses. I. Estimate current traces of the harmonic sources using the fast ICA algorithm (Figure 1). II. Calculate MI between the reconstructed harmonic current traces and harmonic bus voltages.

Choose the number of independent components

Centering of the observed vector X

Whitening of the observed vector X using Eq. (8)

Initializing weight vector w randomly with unit norm

Update separating matrix w using gradient optimization algorithm to maximize negentropy by taking the gradient of Eq. (7)

Normalizing separating matrix w to a new matrix with unite norm

No

If converged

Yes
Repeat the algorithm for each independent component

End

III. Estimate the location of harmonic sources by using computed mutual information.

Figure 1 Implementation procedure of the FastICA.

5. 4. HARMONIC SOURCE LOCALIZATION USING ICA AND MI

SIMULATION RESULTS

For estimating the probable location of multiple harmonic sources in power systems rstly, consider the system equa-

To validate the performance and accuracy of the proposed method in radial and non radial distribution systems, two test systems include the non radial IEEE 30 bus and radial IEEE 34 bus test systems are used. In both test systems, three harmonic sources are considered at the selected buses. In addition, for both cases it is assumed that all loads are with constant power

M. FARHOODNEA ET AL

Figure 2 IEEE 30 bus non radial test system.

Figure 3 Actual and reconstructed harmonic currents at bus 7, 16, and 30 for the 5th harmonic order for the 30 bus test system.

factor and the harmonic loads are modeled as harmonic current injection sources with harmonic spectrums given in [24]. The procedure of estimating the location of harmonic sources using ICA begins with harmonic voltage measurements. Harmonic power ow simulations are performed to compute bus harmonic voltages and these voltages are considered as measurement vector, X in the ICA algorithm. The injected harmonic currents are then reconstructed as illustrated in Figure 1. As we use random number for initialization of the separating matrix w, an average of ten runs of the algorithm is considered. The locations of harmonic sources are then determined by using the mutual information theory. The test results for the two test systems are presented in the following subsections.

5.1

Results of the Non Radial 30 Bus Test System

In this subsection, to verify the accuracy of the proposed method, a non radial IEEE 30 bus test system shown in Figure 2 is used [25]. In this test system, three harmonic sources containing the 5th , 7th , 11th , 13th , and 17th harmonic orders are placed at bus 7, 16, and 30. The PCFLO software [26] is employed to perform harmonic power ow simulations so as to determine the harmonic voltages at all the buses. To create variations in the harmonic measurement vectors, harmonic power ows are simulated for different loading conditions. The loads are then multiplied by the Laplace distributed random variables to create different operating conditions for each

IDENTIFICATION OF MULTIPLE HARMONIC SOURCES IN POWER SYSTEMS

Figure 4 Actual and reconstructed harmonic currents at bus 7, 16, and 30 for the 7th harmonic order for the 30 bus test system.

Figure 5 Actual and reconstructed harmonic currents at bus 7, 16, and 30 for the 11th harmonic order for the 30 bus test system.

load at each sampling time. Harmonic bus voltages are computed by solving the linear system equation (13) for each harmonic frequency of interest. In this case, 350 harmonic voltage samples have been generated to represent the harmonic measurement vector, X. Since the number of harmonic measurements is greater than the number of harmonic sources, preprocessing procedures are taken into account to reduce the dimension of the harmonic measurement vectors, X [17]. For reconstructing variations in the harmonic currents produced by the harmonic sources, the Fast ICA algorithm programmed in Matlab is used. Figures 3 to 7 show the actual and the extracted current traces at bus 7, 16, and 30 for the 5th , 7th , 11th , 13th and 17th harmonic orders, respectively. In the gures, the red lines represent the actual or original harmonic currents while the blue lines represent the reconstructed harmonic currents over the 350 samples. It should be noted that for ease of comparison, both signals in each gure are normalized with their largest values in per-unit (p.u).

To evaluate the accuracy of the ICA algorithm, correlation coefcients of the actual and the reconstructed signals are computed as shown in Table 1. From the table, it is clear that all the correlation coefcients are within the acceptable range, that is, close to 1. This implies high accuracy between the estimated and the actual harmonic currents. Table 2 shows the mean square error of the signals so as to quantify the difference between the estimated and the actual harmonic currents. The results of Table 1 and 2 show that the Fast ICA algorithm can rebuild signals with high accuracy in which the mean square error of the estimated and the actual harmonic currents is less than 0.0015. To identify the location of harmonic sources, the mutual information theory has been applied using Eq. (15) to calculate the pair-wise mutual information between the extracted current trends and bus voltage for each harmonic frequency. The results of the mutual information are shown in Table 3, in which each column and row represents the estimated cur-

M. FARHOODNEA ET AL

Figure 6 Actual and reconstructed harmonic currents at bus 7, 16, and 30 for the 13th harmonic order for the 30 bus test system.

Figure 7 Actual and reconstructed harmonic currents at bus 7, 16, and 30 for the 17th harmonic order for the 30 bus test system.

Table 1 Correlation coefcients between the actual and the reconstructed signals for the 30 bus test system.

Table 2 Mean square error between the actual and reconstructed signals for the 30 bus test system.

Harmonic order 5 7 11 13 17

Bus 7 0.9811 0.9829 0.9847 0.9846 0.9696

Bus 16 0.7789 0.9316 0.8851 0.8422 0.9389

Bus 30 0.9878 0.9930 0.9887 0.9885 0.9842

Harmonic order 5 7 11 13 17

Bus 7 4.61e-4 9.76e-5 8.76e-5 8.81e-5 1.73e-4

Bus 16 1.52e-3 3.90e-4 6.56e-4 9.01e-4 3.48e-4

Bus 30 5.69e-4 4.00e-5 6.45e-5 6.54e-5 9.05e-5

rent and the bus harmonic voltage for individual harmonic order, respectively. The intersection between each column and row represents the computed mutual information between the three estimated currents at each harmonic frequency, and the specied bus harmonic voltage. As shown in Table 3, the underlined-bold mutual information values imply that the related bus has the greatest mutual information between the

estimated harmonic currents and voltages and therefore this bus is identied as the location of harmonic source. For example, in Table 3 the mutual information of 2.387 between the rst estimated current of the 5th harmonic order (rst column) and the harmonic voltage at bus 7 (7th row) has the greatest value in the related column. Therefore, the location of one of the possible harmonic sources (for 5th harmonic order) is at

IDENTIFICATION OF MULTIPLE HARMONIC SOURCES IN POWER SYSTEMS

Figure 8 IEEE 34 bus radial test system.

Table 3 Mutual information between estimated currents and bus voltages for the IEEE30 bus test system.

Bus 1 Bus 2 Bus 3 Bus 4 Bus 5 Bus 6 Bus 7 Bus 8 Bus 9 Bus 10 Bus 11 Bus 12 Bus 13 Bus 14 Bus 15 Bus 16 Bus 17 Bus 18 Bus 19 Bus 20 Bus 21 Bus 22 Bus 23 Bus 24 Bus 25 Bus 26 Bus 27 Bus 28 Bus 29 Bus 30

1 1.644 1.990 1.628 1.635 2.335 2.338 2.387 2.385 0.966 0.754 0.966 0.612 0.612 0.782 0.907 0.850 0.875 0.828 0.776 0.746 0.760 0.735 0.860 0.915 0.818 0.792 0.895 2.059 0.782 0.782

5 2 0.747 0.700 0.755 0.754 0.753 0.776 0.753 0.757 0.728 0.625 0.728 0.564 0.564 0.704 0.836 0.688 0.791 0.742 0.683 0.717 0.651 0.615 0.720 0.674 1.350 1.302 1.791 0.817 2.335 2.406

3 1.026 0.946 1.036 1.031 0.856 0.864 0.856 0.872 1.353 1.581 1.353 1.603 1.603 2.413 1.499 2.541 1.360 2.392 2.205 2.163 1.799 1.820 1.867 1.856 0.929 0.860 0.876 0.984 0.811 0.841

1 0.739 0.754 0.739 0.733 0.782 0.798 0.782 0.786 0.759 0.711 0.759 0.682 0.682 0.698 0.684 0.679 0.718 0.655 0.710 0.714 0.715 0.709 0.586 0.749 1.463 1.410 2.330 0.779 2.767 2.945

7 2 0.820 0.793 0.820 0.835 0.726 0.775 0.726 0.756 1.007 1.218 1.007 1.493 1.493 1.476 1.449 1.726 1.464 1.357 1.338 1.305 1.216 1.176 1.172 1.068 0.728 0.762 0.729 0.705 0.662 0.670

Harmonic order 11 3 1 2 1.575 0.836 1.909 1.732 0.830 2.169 1.575 0.836 1.909 1.581 0.833 1.918 1.730 0.768 2.269 1.918 0.785 2.327 2.330 0.768 2.469 1.788 0.838 2.417 1.066 0.746 0.779 0.865 0.741 0.933 1.066 0.746 0.779 0.787 0.705 0.737 0.787 0.705 0.909 0.794 0.705 0.909 0.800 0.705 0.943 0.738 0.667 0.930 0.821 0.696 0.860 0.835 0.692 0.943 0.846 0.704 0.951 0.822 0.741 0.976 0.886 0.766 0.918 0.828 0.747 0.944 0.638 0.686 0.838 0.875 0.705 0.968 0.728 1.162 1.033 0.713 1.141 1.002 0.700 1.474 0.951 1.441 0.807 2.313 0.773 2.069 0.867 0.708 2.311 0.829

3 0.903 0.911 0.903 0.905 0.766 0.861 0.766 0.841 1.438 1.692 1.438 1.671 1.767 1.767 1.708 1.784 1.721 1.720 1.727 1.751 1.684 1.655 1.298 1.436 1.028 1.003 0.911 0.890 0.873 0.820

1 1.649 1.763 1.649 1.643 1.183 1.952 2.182 1.846 1.115 0.962 1.115 0.922 0.922 0.897 0.726 0.899 0.900 0.920 0.911 0.906 0.946 0.895 0.900 0.943 0.832 0.801 0.748 1.604 0.775 0.777

13 2 0.769 0.773 0.769 0.766 0.766 0.729 0.767 0.770 0.755 0.734 0.755 0.709 0.709 0.699 0.638 0.713 0.689 0.775 0.740 0.759 0.748 0.763 0.740 0.815 1.196 1.105 1.635 0.762 2.538 2.553

3 0.920 0.850 0.920 0.918 0.779 0.812 0.778 0.809 1.213 1.543 1.213 1.779 1.779 1.835 1.684 1.856 1.738 1.791 1.786 1.710 1.610 1.600 1.344 1.502 0.966 0.944 0.778 0.819 0.702 0.732

1 0.869 0.861 0.868 0.853 0.762 0.850 0.762 0.856 1.344 2.015 1.346 2.065 2.065 2.084 2.050 2.421 1.529 2.002 1.999 2.022 1.917 1.905 2.208 2.088 1.826 1.787 1.557 0.829 0.774 0.749

17 2 1.793 1.844 1.790 1.790 1.746 1.767 1.846 1.770 0.992 0.922 0.994 0.890 0.890 0.892 0.898 0.909 0.959 0.892 0.896 0.907 0.921 0.901 0.896 0.889 0.914 0.906 0.913 1.738 0.927 0.796

3 0.831 0.854 0.824 0.822 0.785 0.847 0.785 0.860 0.810 0.792 0.807 0.738 0.738 0.742 0.734 0.850 0.882 0.752 0.770 0.774 0.757 0.758 0.680 0.734 0.854 0.858 0.892 0.850 1.967 2.739

M. FARHOODNEA ET AL

Figure 9 Actual and reconstructed harmonic currents at bus 15, 25, and 33 for the 5th harmonic order for the 34 bus test system.

Figure 10 Actual and reconstructed harmonic currents at bus 15, 25, and 33 for the 7th harmonic order for the 34 bus test system.

bus 7. From the results shown in Table 3, it is evident that harmonic source locations are located at bus 7, 16, and 30 and this proves the accuracy of the proposed harmonic source location method.

5.2

Results of the Radial 34 bus Test System

To prove the accuracy and ability of proposed method for locating multiple harmonic sources in radial distribution networks, the radial IEEE 34 bus test system shown in Figure 8 is used [27]. In the test system, three harmonic sources containing the 5th , 7th , 11th , 13th , and 17th harmonic orders are located at bus 15, 25, and 33. In order to use the ICA algorithm for reconstructing the injected harmonic currents, harmonic voltage measurements are generated using the harmonic power ow algorithm for radial systems [28-29] which has been implemented using the Matlab codes.

Similar to the 30 bus test system, harmonic power ows are simulated for different loading conditions by means of multiplying the Laplace distributed random variables at all loads to create different operating conditions for each load and at each sampling time. Again, harmonic bus voltages are computed by solving the linear system for each harmonic order. In order to reconstruct injected harmonic currents caused by the harmonic sources, 500 harmonic voltage samples have been generated at each bus so as to represent harmonic measurement vector, X. by using preprocessing procedures, the dimension of the harmonic measurement vectors X can be decreased so as to omit the redundant information of the data. Figures 9 to 13 show the extracted current traces by using the Fast ICA algorithm in comparison with the actual signals at bus 15, 25, and 33 for the 5th , 7th , 11th , 13th and 17th harmonic orders, respectively, in which the red lines represent the actual harmonic currents and the blue lines represent the reconstructed

IDENTIFICATION OF MULTIPLE HARMONIC SOURCES IN POWER SYSTEMS

Figure 11 Actual and reconstructed harmonic currents at bus 15, 25, and 33 for the 11th harmonic order for the 34 bus test system.

Figure 12 Actual and reconstructed harmonic currents at bus 15, 25, and 33 for the 13th harmonic order for the 34 bus test system.

harmonic currents over 500 samples and all signals are normalized with their largest values in per-unit (p.u). To prove high accuracy between the estimated and the actual harmonic currents, the obtained correlation coefcients and mean square error of the signals are shown in Tables 4 and 5. The results of Tables 4 and 5 imply that compatibility degree between the reconstructed and original signals is greater than 98% and mean square error of estimation is less than 4.33e5. It is shown that the 34 bus test system has correlation coefcient values closer to 1, and lesser mean square error value in comparison with the results of the 30 bus test system. Hence, the accuracy of the results is due to the difference in the sample size used in the two systems. To identify the location of harmonic sources, the pair-wise mutual information between the extracted current trends and bus voltage have been obtained for each harmonic frequency as shown in Table 6. In the Table, the underlined-bold mutual information values imply that the related bus has the greatest mutual information between the estimated harmonic currents

Table 4 Correlation coefcients between the actual and the reconstructed signals for the 34 bus test system.

Harmonic order 5 7 11 13 17

Bus 15 0.9892 0.9931 0.9892 0.9936 0.9930

Bus 25 0.9906 0.9909 0.9907 0.9972 0.9909

Bus 33 0.9945 0.9978 0.9944 0.9951 0.9944

Table 5 Mean square error between the actual and reconstructed signals for the 34 bus test system.

Harmonic order 5 7 11 13 17

Bus 15 4.33e-5 2.24e-5 4.32e-5 2.57e-5 2.79e-5

Bus 25 3.74e-5 3.81e-5 3.73e-5 1.11e-5 3.62e-5

Bus 33 2.21e-5 8.90e-6 2.22e-5 1.94e-5 2.23e-5

M. FARHOODNEA ET AL

Table 6 Mutual information between estimated currents and bus voltages for the 34 bus test system.

5 1 Bus 1 Bus 2 Bus 3 Bus 4 Bus 5 Bus 6 Bus 7 Bus 8 Bus 9 Bus 10 Bus 11 Bus 12 Bus 13 Bus 14 Bus 15 Bus 16 Bus 17 Bus 18 Bus 19 Bus 20 Bus 21 Bus 22 Bus 23 Bus 24 Bus 25 Bus 26 Bus 27 Bus 28 Bus 29 Bus 30 Bus 31 Bus 32 Bus 33 Bus 34 0.000 0.814 0.814 0.954 1.066 1.108 0.958 0.956 0.866 0.829 0.829 0.829 0.815 0.709 0.701 0.701 1.201 1.284 1.338 1.414 1.476 1.564 1.607 1.705 1.741 1.641 1.541 0.958 0.958 0.958 0.818 0.827 0.795 0.795 2 0.000 0.911 0.911 0.770 0.752 0.753 0.759 0.679 0.687 0.649 0.649 0.649 1.075 1.328 1.483 1.443 0.739 0.744 0.718 0.733 0.714 0.714 0.669 0.664 0.680 0.680 0.680 0.759 0.759 0.759 0.617 0.663 0.695 0.695 3 0.000 0.758 0.758 0.866 0.905 0.931 1.004 1.042 1.135 1.184 1.184 1.184 0.723 0.667 0.625 0.625 0.875 0.855 0.815 0.765 0.743 0.721 0.730 0.717 0.729 0.729 0.729 1.004 1.004 1.004 1.257 1.321 1.335 1.235 1 0.000 0.747 0.747 0.904 1.013 1.057 1.160 1.249 1.377 1.460 1.460 1.460 0.685 0.667 0.652 0.652 0.977 0.991 0.943 0.917 0.870 0.852 0.850 0.829 0.793 0.793 0.793 1.160 1.160 1.160 1.524 1.596 1.632 1.432

7 2 0.000 1.046 1.046 0.900 0.814 0.747 0.740 0.717 0.711 0.674 0.674 0.674 1.321 1.695 1.846 1.546 0.744 0.788 0.767 0.745 0.734 0.712 0.753 0.742 0.706 0.706 0.706 0.740 0.740 0.740 0.689 0.672 0.665 0.665

Harmonic order 11 3 1 2 0.000 0.000 0.000 0.703 0.712 1.173 0.703 0.712 1.173 0.718 0.822 0.916 0.820 0.931 0.842 0.880 0.967 0.789 0.807 1.068 0.752 0.744 1.130 0.775 0.780 1.257 0.737 0.696 1.306 0.703 0.696 1.306 0.703 0.696 1.306 0.703 0.653 0.623 1.462 0.647 0.630 1.808 0.659 0.635 1.988 0.659 0.635 1.888 0.928 0.919 0.834 0.998 0.942 0.803 1.043 0.896 0.767 1.094 0.867 0.750 1.152 0.856 0.760 1.191 0.810 0.741 1.233 0.805 0.702 1.297 0.805 0.727 1.322 0.780 0.721 1.122 0.780 0.721 1.022 0.780 0.721 0.807 1.068 0.752 0.807 1.068 0.752 0.807 1.068 0.752 0.711 1.357 0.733 0.696 1.385 0.722 0.677 1.443 0.723 0.677 1.343 0.723

3 0.000 0.711 0.711 0.722 0.824 0.887 0.843 0.835 0.745 0.740 0.740 0.740 0.671 0.664 0.671 0.671 0.974 0.991 1.084 1.095 1.129 1.195 1.232 1.313 1.324 1.242 1.127 0.843 0.843 0.843 0.755 0.727 0.704 0.704

1 0.000 0.670 0.670 0.700 0.680 0.782 0.770 0.820 0.769 0.707 0.707 0.707 0.632 0.631 0.629 0.629 0.840 0.878 0.917 0.964 1.025 1.086 1.091 1.125 1.196 1.155 1.124 0.770 0.770 0.770 0.671 0.698 0.724 0.724

13 2 0.000 1.407 1.407 1.037 0.881 0.860 0.819 0.831 0.753 0.773 0.773 0.773 1.681 1.998 2.144 2.005 0.813 0.828 0.817 0.776 0.784 0.786 0.753 0.734 0.767 0.767 0.767 0.819 0.819 0.819 0.739 0.718 0.723 0.723

3 0.000 0.699 0.699 0.781 0.876 0.973 1.109 1.160 1.273 1.333 1.333 1.333 0.657 0.629 0.627 0.627 0.952 0.916 0.890 0.901 0.884 0.868 0.845 0.874 0.839 0.839 0.839 1.109 1.109 1.109 1.364 1.421 1.500 1.325

1 0.000 0.608 0.608 0.719 0.732 0.788 0.904 0.977 1.055 1.084 1.084 1.084 0.603 0.585 0.585 0.585 0.831 0.796 0.815 0.802 0.785 0.781 0.750 0.697 0.710 0.710 0.710 0.904 0.904 0.904 1.128 1.161 1.190 0.968

17 2 0.000 0.707 0.707 0.779 0.814 0.895 0.860 0.869 0.828 0.782 0.782 0.782 0.672 0.678 0.664 0.664 0.972 1.037 1.124 1.194 1.272 1.311 1.344 1.432 1.480 1.365 1.312 0.860 0.860 0.860 0.782 0.779 0.747 0.747

3 0.000 1.321 1.321 0.990 0.902 0.838 0.792 0.804 0.774 0.754 0.754 0.754 1.585 1.938 2.094 1.973 0.794 0.764 0.802 0.792 0.757 0.727 0.717 0.725 0.731 0.731 0.731 0.792 0.792 0.792 0.768 0.792 0.741 0.741

Table 7 Mean square errors of the mutual information values for IEEE 30 and 34 bus test system.

Harmonic orders IEEE 30 bus test system IEEE 34 bus test system

5th 1.29e-4 6.96e-5

7th 2.24e-4 8.58e-5

11th 2.69e-4 9.94e-5

13th 3.45e-4 1.21e-4

17th 1.26e-4 9.81e-5

IDENTIFICATION OF MULTIPLE HARMONIC SOURCES IN POWER SYSTEMS

Figure 13 Actual and reconstructed harmonic currents at bus 15, 25, and 33 for the 17th harmonic order for the 34 bus test system.

and voltages and this bus should identify as one of the possible location of harmonic source. For example, the mutual information of 1.741 between the rst estimated current of the 5th harmonic order (rst column) and the harmonic voltage at bus 25 (25th row) has the greatest value in the related column, which implies that bus 25 is the location of one of the possible harmonic sources in this test system. From the results shown in Table 6, it is evident that harmonic source locations are located at bus 15, 25, and 33 for the 34 bus test system, because these buses have the highest mutual information values. Thus, the harmonic source locations have been accurately estimated by the proposed method using ICA and mutual information theory. Table 7 shows the mean square errors of the mutual information values for both 30 and 34 bus test systems. The results shows that the mutual information algorithm can estimate exact locations of the harmonic sources with the mean square errors less than 0.00035 and 0.00012 for the 30 and 34 bus test systems, respectively.

REFERENCES
1. K. Islam and A. Samra, "Identication of harmonic sources in power distribution systems," IEEE Southeastcon97.Engineering new New Century., Proceedings, pp. 301 303, 1997. 2. W. Xu, X. Liu, and Y. Liu, "An investigation on the validity of power direction method for harmonic source determination," IEEE Power Engineering Review, vol. 22, pp. 6262, 2002. 3. X. Wilsun and L. Yilu, "A method to determine customer harmonic contributions for incentive-based harmonic control applications," in Power Engineering Society Summer Meeting, 1999. IEEE, 1999, pp. 361366 vol.1. 4. X. Wilsun and L.Yilu, "A method for determining customer and utility harmonic contributions at the point of common coupling," IEEE Transactions on Power Delivery, vol. 15, pp. 804811, 2000. 5. C. Chaoying, L. Xiuling, D. Koval, X. Wilsun, and T. Tayjasanant, "Critical impedance method - a new detecting harmonic sources method in distribution systems," IEEE Transactions on Power Delivery, vol. 19, pp. 288297, 2004. 6. N. Hamzah, A. Mohamed, and A. Hussain, "Harmonic source location at the point of common coupling based on voltage magnitude," in TENCON 2004. 2004 IEEE Region 10 Conference, 2004, pp. 220223 Vol. 3. 7. S. F. de Paula Silva and J. C. de Oliveira, "The sharing of responsibility between the supplier and the consumer for harmonic voltage distortion: A case study," Electric Power Systems Research, vol. 78, pp. 19591964, 2008. 8. T. Pfajfar, B. Blazic, and I. Papic, "Harmonic Contributions Evaluation With the Harmonic Current Vector Method," IEEE Transactions on Power Delivery, vol. 23, pp. 425433, 2008. 9. A. Cataliotti and V. Cosentino, "A New Measurement Method for the Detection of Harmonic Sources in Power Systems Based on the Approach of the IEEE Std. 1459–2000," IEEE Transactions on Power Delivery, vol. 25, pp. 332340, 2010. 10. Z. P. Du, J. Arrillaga, N. R. Watson, and S. Chen, "Identication of harmonic sources of power systems using state estimation," Generation, Transmission and Distribution, IEE Proceedings, vol. 146, pp. 712, 1999. 11. G. DAntona, C. Muscas, and S. Sulis, "State Estimation for the Localization of Harmonic Sources in Electric Distribution

6.

CONCLUSION

In this paper, a novel technique has been presented for locating multiple harmonic sources in radial and non radial distribution systems, without prior information about the network components and branch parameters. In the proposed method, ICA which is one of the blind source separation techniques is applied for estimating the injected harmonic currents produced by harmonic sources and the mutual information theory is applied to identify the location of possible harmonic sources. ICA is also used to estimate the load proles of the harmonic sources with different sample sizes. The results from the case studies conrm that the proposed method can accurately identify the location of multiple harmonic sources in both radial and non radial distribution systems. The proposed method is promising because it uses only bus voltage measurements and there is no need for real and reactive power measurements and other restricted power system information.

M. FARHOODNEA ET AL

12.

13.

14.

15.

16.

17. 18.

19. 20.

Systems," IEEE Transactions on Instrumentation and Measurement, vol. 58, pp. 14621470, 2009. C. Madtharad, S. Premrudeepreechacharn, N. Watson, and R. Saeng-Udom, "An optimal measurement placement method for power system harmonic state estimation," IEEE Transactions on Power Delivery, vol. 20, pp. 15141521, 2005. C. Rakpenthai, S. Premrudeepreechacharn, S. Uatrongjit, and N. Watson, "Measurement placement for power system state estimation using decomposition technique," Electric Power Systems Research, vol. 75, pp. 4149, 2005. A. Kumar, B. Das, and J. Sharma, "Determination of location of multiple harmonic sources in a power system," International Journal of Electrical Power & Energy Systems, vol. 26, pp. 73 78, 2004. E. Gursoy, "Independent component analysis for harmonic source identication in electric power systems," Drexel University, 2007. E. Oja and A. Hyvarinen, "Independent component analysis: algorithms and applications," Neural Networks, vol. 13, pp. 411 430, 2000. A. Hyvar nen, J. Karhunen, and E. Oja, "Independent Component Analysis," 4th ed: John Wiley, 2001. A. Hyvarinen, "Fast and robust xed-point algorithms for independent component analysis," IEEE Transactions on Neural Networks, vol. 10, pp. 626634, 1999. A. Papoulis and S. Pillai, Probability, Random Variables, and Stochastic Processes: McGraw-Hill, 2002. W. Li, "Mutual information functions versus correlation functions," Journal of Statistical Physics, vol. 60, pp. 823837, 1990.

21. E. Duggan and R. E. Morrison, "Probabilistic methods applied to harmonic distortion," in Probabilistic Methods Applied to Electric Power Systems, 1991., Third International Conference on, 1991, pp. 4346. 22. L. Huaiwei and D. Niebur, "Exploring independent component analysis for electric load proling," in Neural Networks, 2002. IJCNN 02. Proceedings of the 2002 International Joint Conference on, 2002, pp. 21442149. 23. L. Huaiwei and D. Niebur, "Load prole estimation in electric transmission networks using independent component analysis," IEEE Transactions on Power Systems, vol. 18, pp. 707715, 2003. 24. R. Abu-Hashim, et al., "Test systems for harmonics modeling and simulation," IEEE Transactions on Power Delivery, vol. 14, pp. 579587, 1999. 25. H. Saadat, "Power system analysis, WCB," ed: McGraw-Hill, Boston, 1999. 26. M. Grady, "PCFLO User Manual," ECE Dept, The University of Texas at Austin, . http://www.ece.utexas.edu/grady.,2005. 27. S. F. Mekhamer, S. A. Soliman, M. A. Moustafa, and M. E. ElHawary, "Application of fuzzy logic for reactive-power compensation of radial distribution feeders," IEEE Transactions on Power Systems, vol. 18, pp. 206213, 2003. 28. T. Jen-Hao and C. Chuo-Yean, "Backward/Forward SweepBased Harmonic Analysis Method for Distribution Systems," IEEE Transactions on Power Delivery, vol. 22, pp. 16651672, 2007. 29. J. Teng, "A network-topology-based three-phase load ow for distribution systems," Proc Natl. Sci. Counc. ROC (A), vol. 24, pp. 259264, 2000.