System and method for electric load identification and classification employing support vector machine

Abstract

A method identifies electric load types of a plurality of different electric loads. The method includes providing a support vector machine load feature database of a plurality of different electric load types; sensing a voltage signal and a current signal for each of the different electric loads; determining a load feature vector including at least six steady-state features with a processor from the sensed voltage signal and the sensed current signal; and identifying one of the different electric load types by relating the load feature vector including the at least six steady-state features to the support vector machine load feature database.

Claims

1. A method of identifying electric load types of a plurality of different electric loads, said method comprising: providing a support vector machine load feature database of a plurality of different electric load types; sensing a voltage signal and a current signal for each of said different electric loads; determining a load feature vector including at least six steady-state features with a processor from said sensed voltage signal and said sensed current signal; and identifying one of said different electric load types by relating the load feature vector including the at least six steady-state features to the support vector machine load feature database.

2. The method of claim 1 further comprising: training said support vector machine load feature database as a multi-class one-against-all support vector machine for each of a plurality of different load classes or for each of said different electric load types.

3. The method of claim 2 further comprising: employing as said different load classes seven different load classes having distinct steady-state features, said seven different load classes including resistive loads, reactive predominant loads, electronic loads with a power factor correction circuit, electronic loads without a power factor correction circuit, linear power supplies using a transformer to boost voltage, phase angle controllable loads, and complex structures.

4. The method of claim 2 further comprising: employing said different electric load types having distinct steady-state features.

5. The method of claim 1 further comprising: employing a Gaussian radial basis functions kernel with said support vector machine load feature database.

6. The method of claim 1 further comprising: representing voltage and current waveforms corresponding to said sensed voltage signal and said sensed current signal, respectively, by Fourier series.

7. The method of claim 1 further comprising: employing said at least six steady-state features selected from the group consisting of RMS current value, displacement power factor, total harmonic distortion of current, power factor, current crest factor, current K-factor, admittance, and normalized current third and fifth harmonics.

8. A method of identifying electric load types of a plurality of different electric loads, said method comprising: providing a database including a first layer formed by a supervised self-organizing map database and a second layer formed by a support vector machine database; clustering a plurality of different load classes having a plurality of different load features in the first layer; providing a plurality of different electric load types under each of the different load classes in the second layer; placing different ones of said different electric load types having similar load feature vectors into a same one of the different load classes; sensing a voltage signal and a current signal for each of said different electric loads; determining a load feature vector including a plurality of steady-state features with a processor from said sensed voltage signal and said sensed current signal; and identifying by a support vector machine one of said different electric load types by relating the determined load feature vector including the steady-state features in the second layer of said database.

9. The method of claim 8 further comprising: employing as said different load classes seven different load classes having distinct steady-state features, said seven different load classes including resistive loads, reactive predominant loads, electronic loads with a power factor correction circuit, electronic loads without a power factor correction circuit, linear power supplies using a transformer to boost voltage, phase angle controllable loads, and complex structures.

10. The method of claim 8 further comprising: training the supervised self-organizing map database employing data corresponding to the different electric load types; training the support vector machine database as a multi-class one-against-all support vector machine for each of said plurality of different load classes; identifying said one of said different electric load types as being in one of said plurality of different load classes with said supervised self-organizing map database; and identifying said one of said different electric load types with the trained support vector machine database for said one of said plurality of different load classes.

11. The method of claim 10 further comprising: extracting information from the trained support vector machine database and storing simplified information in a trained neuron grid; employing as said different load classes a plurality of different load categories; determining one of the different load categories employing the determined load feature vector; and employing a support vector machine discriminator function for each the different load categories to identify said one of said different electric load types.

12. The method of claim 10 further comprising: employing as said different load classes seven different load classes having distinct steady-state features, said seven different load classes including resistive loads, reactive predominant loads, electronic loads with a power factor correction circuit, electronic loads without a power factor correction circuit, linear power supplies using a transformer to boost voltage, phase angle controllable loads, and complex structures.

13. The method of claim 8 further comprising: representing voltage and current waveforms corresponding to said sensed voltage signal and said sensed current signal, respectively, by Fourier series.

14. The method of claim 8 further comprising: selecting the steady-state features of said determined load feature vector from the group consisting of RMS current value, displacement power factor, total harmonic distortion of current, power factor, current crest factor, current K-factor, admittance, and normalized current third and fifth harmonics.

15. The method of claim 8 further comprising: including with the steady-state features of said determined load feature vector voltage-current trajectory features selected from the group consisting of area, eccentricity, and Hausdorff distance.

16. A system for identifying electric load types of a plurality of different electric loads, said system comprising: a database including a first layer formed by a supervised self-organizing map database and a second layer formed by a support vector machine database, a plurality of different load classes having a plurality of different load features being clustered in the first layer, a plurality of different electric load types being under each of the different load classes in the second layer, different ones of said different electric load types having similar load feature vectors being placed into a same one of the different load classes; a plurality of sensors structured to sense a voltage signal and a current signal for each of said different electric loads; and a processor structured to determine a load feature vector including a plurality of steady-state features from said sensed voltage signal and said sensed current signal, and identify by a support vector machine one of said different electric load types by relating the determined load feature vector including the steady-state features in the second layer of said database.

17. The system of claim 16 wherein said processor is further structured to train the supervised self-organizing map database employing data corresponding to the different electric load types, train the support vector machine database as a multi-class one-against-all support vector machine for each of said plurality of different load classes, identify said one of said different electric load types as being in one of said plurality of different load classes with said supervised self-organizing map database, and identify said one of said different electric load types with the trained support vector machine database for said one of said plurality of different load classes.

18. The system of claim 17 wherein said processor is further structured to extract information from the trained support vector machine database and store simplified information in a trained neuron grid, employ as said different load classes a plurality of different load categories; determine one of the different load categories employing the determined load feature vector, and employ a support vector machine discriminator function for each the different load categories to identify said one of said different electric load types.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) A full understanding of the disclosed concept can be gained from the following description of the preferred embodiments when read in conjunction with the accompanying drawings in which:

(2) FIG. 1 is a diagram of a support vector machine (SVM) framework employing kernel functions.

(3) FIGS. 2A and 2B are plots of voltage and current profiles, respectively, of a DVD player.

(4) FIGS. 2C and 2D are plots of voltage and current profiles, respectively, of an LCD TV.

(5) FIGS. 2E and 2F are plots of voltage and current profiles, respectively, of an oscillating fan.

(6) FIGS. 3A and 3B are plots of an output grid of a supervised Self-Organizing Map (SSOM) using a set of six example steady-state features and harmonics as features, respectively.

(7) FIG. 4 is a block diagram of a cross validation mechanism for testing performance of a one-against-all SVM identifier in accordance with an embodiment of the disclosed concept.

(8) FIG. 5 is a block diagram of a support vector machine (SVM) based load classification/identification system in accordance with embodiments of the disclosed concept.

(9) FIG. 6 is a flowchart of a hybrid SSOM/SVM classifier process in accordance with an embodiment of the disclosed concept.

(10) FIG. 7 is a block diagram of a self-organizing map (SSOM)/support vector machine (SVM) based load classification/identification system in accordance with embodiments of the disclosed concept.

(11) FIGS. 8A and 8B are plots of voltage and current profiles, respectively, of an LCD TV used by a trained SSOM in accordance with an embodiment of the disclosed concept.

(12) FIGS. 8C and 8D are plots of voltage and current profiles, respectively, of an LED TV used by a trained SSOM in accordance with an embodiment of the disclosed concept.

(13) FIGS. 8E and 8F are plots of voltage and current profiles, respectively, of a plasma TV used by a trained SSOM in accordance with an embodiment of the disclosed concept.

DESCRIPTION OF THE PREFERRED EMBODIMENTS

(14) As employed herein, the term number shall mean one or an integer greater than one (i.e., a plurality).

(15) As employed herein, the term processor shall mean a programmable analog and/or digital device that can store, retrieve, and process data; a computer; a workstation; a personal computer; a microprocessor; a microcontroller; a microcomputer; a central processing unit; a mainframe computer; a mini-computer; a server; a networked processor; or any suitable processing device or apparatus.

(16) The disclosed concept provides a support vector machine (SVM) based advanced identification system and method for miscellaneous electric loads (MELs). Embodiments including applying only SVM as well as a combination of SVM and a supervised self-organizing map (SSOM) are disclosed. This applies the supervised SSOM to classify and identify MELs. A relatively large number of MELs are classified into several clusters. Within each cluster, a SVM can be applied to separate MELs with similar but not identical features.

(17) In the second embodiment, SSOM first clusters a relatively large number of MELs into several classes. MELs with similar feature values fall into the same class. SVM is then employed to identify similar MELs. The combination of SVM and SSOM shows satisfactory accuracy in tests using real-world data. The hybrid SSOM/SVM identifier can achieve better performance in the sense of accuracy, robustness and applicability. The SSOM identifier first extracts information from a relatively large amount of training data and stores that simplified information in a trained neuron grid. When an input feature vector is presented to the hybrid SSOM/SVM identifier, it first determines which load category it falls into and then utilizes an SVM discriminator function for each category to make an identification decision.

(18) Multi-Class Support Vector Machine (SVM) Classifier

(19) A method for multi-class classification is known from basic SVMs. In the disclosed concept, a MEL identification method employs a SVM based load feature database formed from sensed current and voltage signals. Basic SVMs are inherently designed for classification of two classes, .sub.1 and .sub.2, of patterns which are described by feature vectors extracted from data in a predefined manner. If x is such a feature vector, then the SVM utilizes a (typically nonlinear) mapping from an input feature vector space to a high-dimensional (possibly infinite-dimensional) Euclidean space H, as shown by Equation 1.
:xR.sup.l.fwdarw.(x)H (Eq. 1)
wherein:

(20) x is a feature vector of dimension l;

(21) R is the standard notation of the vector space of real numbers;

(22) R.sup.l is the vector space of real numbers (of dimension l); and

(23) is a mapping function, mapping from the l-dimension space of real numbers to the space H.

(24) In turn, the two classes can be satisfactorily separated by a hyperplane as defined by Equation 2.
g(x)=.sup.Tx+.sub.0 (Eq. 2)
wherein:

(25) and .sub.o are co-efficients defining the hyperplane g(x);

(26) superscript T denotes the transpose of a vector; and

(27) g(x) is a function describing a hyperplane in high dimensional space.

(28) Both x and , for example, are 2-by-1 vectors, then .sup.T is of dimension 1-by-2, and g(x) is a straight line in two-dimensional space.

(29) Once an optimal hyperplane (, .sub.0) has been determined, classification of which class an unknown feature vector x* (i.e., a different vector other than x) belongs to is performed based on the sign of g(x*). The SVM training algorithm only depends on the training data through inner products in H (i.e., on functions of the form shown in Equation 3).
K(x.sub.i,x.sub.j)=(x.sub.i),(x.sub.j) (Eq. 3)
wherein:

(30) K is usually called a kernel function; it is common that an SVM needs only to specify K before its training instead of knowing the explicit form of ;

(31) x.sub.i, i=1, 2, . . . , are feature vectors in the training data; for each x.sub.i denote the corresponding class indicator by y.sub.i (+1 for .sub.1 and 1 for .sub.2); and

(32) the function K needs two inputs, thus there is another index besides i, that index is j.

(33) Once an appropriate kernel has been adopted, the optimal hyperplane (, .sub.0) can be determined from Equation 4:

(34) $\begin{array}{cc}\underset{}{\max }\left(\underset{i}{\overset{}{.Math.}}{}_i-\frac{1}{2}\underset{i,j}{\overset{}{.Math.}}{}_i{}_j{y}_i{y}_{j}K\left(x_i,x_j\right)\right)& \left(\mathrm{Eq}.4\right)\end{array}$
subject to Equation 5:

(35) $\begin{array}{cc}0{}_{i}C,\underset{i}{\overset{}{.Math.}}{}_i{y}_i=0i=1,2,\mathrm{.Math.}N& \left(\mathrm{Eq}.5\right)\end{array}$
wherein:

(36) is the vector of nonnegative Lagrange multipliers .sub.i;

(37) C is a parameter to be chosen by the user with a relatively larger C corresponding to assigning a relatively higher penalty to errors; and

(38) N is the total number of Lagrange multipliers, N=l, i.e., the dimension of feature vector x.

(39) The resulting classifier assigns x to .sub.1 (.sub.2) if Equation 6 is met.

(40) $\begin{array}{cc}g\left(x\right)=\underset{i=1}{\overset{N}{.Math.}}{}_i{y}_{i}K\left(x_i,x\right)+{}_0>(<)0& \left(\mathrm{Eq}.6\right)\end{array}$

(41) FIG. 1 depicts a SVM framework 2 that employs kernel functions, K(.,.) 4, 6, 8, where each x.sub.i 10 is an input to one of the kernel functions, each .sub.iy.sub.i 12 is a weight to the output of one of the functions, and .sub.0 14 is the bias.

(42) For multi-class SVM kernel selection, which is an M-class problem, common extensions are to either consider it as a set of M two-class problems (one-against-all) or train M(M1)/2 basic SVM classifiers (one-against-one). In the disclosed concept, the one-against-all technique is employed. For each .sub.i of the M classes, the one-against-all SVM aims at determining an optimal discriminator function, g.sub.i(x), i=1, 2, . . . , M, so that g.sub.i(x)>g.sub.j(x) for all ji and x.sub.i.

(43) The classification rule is then defined by Equation 7.

(44) $\begin{array}{cc}x{}_i\mathrm{if}i=\arg \underset{k}{\max }\left\{g_k\left(x\right)\right\}& \left(\mathrm{Eq}.7\right)\end{array}$
wherein:

(45) k is an index other than i, since here i is the output number, and k denotes the index during the searching of such an output i.

(46) Many kernels are available for SVM, such as polynomials, RBF, and hyperbolic tangent. In the disclosed concept, the Gaussian RBF kernel is employed, which is also the most commonly adopted kernel in pattern recognition problems.

(47) Electric Load Identification Framework Using Only SVM

(48) Known proposals apply SVM to identify harmonic sources. The features used in these proposals are high frequency harmonic components.

(49) For the purpose of MELs identification, a different set of steady-state features with practical meaning are adopted with the voltage (V(t)) and current (I(t)) waveforms represented by Fourier series of the form shown by Equations 8 and 9, respectively.

(50) $\begin{array}{cc}V\left(t\right)=\underset{k=1}{\overset{}{.Math.}}{V}_k\sin \left(k{}_{0}t+{}_k\right)& \left(\mathrm{Eq}.8\right)\\ I\left(t\right)=\underset{k=1}{\overset{}{.Math.}}{I}_k\sin \left(k{}_{0}t+{}_k\right)& \left(\mathrm{Eq}.9\right)\end{array}$
wherein:

(51) .sub.0 is frequency;

(52) k is an index of the order of harmonics; and

(53) .sub.k, .sub.k are the phase angles of the k-th order harmonic.

(54) The following six steady-state features are considered.

(55) First, the RMS current value, I.sub.RMS, gives equivalent information about the average power.

(56) Second, the average displacement power factor is shown by Equation 10.
pf.sub.disp=cos(.sub.1.sub.1) (Eq. 10)
wherein:

(57) pf.sub.disp is the displacement power factor; and

(58) (.sub.1.sub.1) is the fundamental power factor angle.

(59) Third, the average total harmonic distortion (THD) of current (THD.sub.i) is shown by Equation 11.

(60) $\begin{array}{cc}{\mathrm{THD}}_I=\frac{\sqrt{\underset{k=2}{\overset{}{.Math.}}{I}_k^2}}{I_1}100\%& \left(\mathrm{Eq}.11\right)\end{array}$
wherein:

(61) k is an index of the order of harmonics;

(62) I.sub.k is the k-th order harmonic in current; and

(63) I.sub.1 is the fundamental, i.e., first order harmonic.

(64) Fourth, the average power factor (pf) is determined by calculating displacement power factor and the current THD using the fast Fourier transform (FFT) of the current waveform as shown by Equation 12.

(65) $\begin{array}{cc}\mathrm{pf}=\frac{{\mathrm{pf}}_{\mathrm{disp}}}{\sqrt{1+{\mathrm{THD}}_1^2}}& \left(\mathrm{Eq}.12\right)\end{array}$

(66) Fifth, the crest factor (cf) or peak-to-average ratio (PAR) is determined by Equation 13.

(67) $\begin{array}{cc}\mathrm{cf}=\frac{.Math.I_{\mathrm{peak}}.Math.}{I_{\mathrm{RMS}}}& \left(\mathrm{Eq}.13\right)\end{array}$
wherein:

(68) I.sub.peak is the current's peak amplitude; and

(69) I.sub.rms is the current's RMS value.

(70) The example current crest factor or PAR or peak-to-RMS power ratio (PAPR) is a measurement of a waveform, calculated from the peak amplitude of the waveform divided by the RMS value of the waveform. It is therefore a dimensionless quantity. Crest factor can be used to detect the existence of a current pulse. A sharp peak corresponds to a relatively higher value of crest factor.

(71) Sixth, there are the normalized 3.sup.rd (and 5.sup.th) harmonics of current.

(72) The fact that the above example set of steady-state features is better for the purpose of MELs identification can be validated by both SVM and SSOM. For three example MELs, a DVD player (D), an LCD TV (T) and a oscillating fan (F), their voltage profiles 22,26,30 and current profiles 24,28,32 are shown in FIGS. 2A-2F.

(73) FIGS. 3A-3B show output grids 42,44 of SSOM using two sets of different features: (1) the disclosed set of six example steady-state features, and (2) harmonics as features. The SSOM output grid 42 gives a better clustered grid using the disclosed set of steady-state features. For SVM, unlike SSOM, the training output is parameters describing a high-dimensional hyper-plane, which cannot be visualized in two dimensions.

Example 1

(74) Therefore, the advantage of the disclosed set of steady-state features over using harmonics is shown by tests as summarized in Table 1. In this example, the total number of available feature vectors for training and testing is 3600, and three different cases are tested and compared. The results are generated by solving multi-class one-against-all SVMs. This compares testing success rate of different features sets using a multi-class one-against-all SVMs.

(75) TABLE-US-00001 TABLE 1 Success rate 270 points 540 points 1080 points for training, for training, for training, and 3330 and 3060 and 2520 points points points for testing for testing for testing Disclosed 100% 100% 100% set of six example steady- state features Harmonics 99.56% 99.53% 99.43% as features
It is clear to see from Table 1 that simply using harmonics cannot guarantee a 100% success rate even with only three relatively distinct MELs.

(76) Data was collected, processed and employed from commercially available MELs 48. Steady-state features of an example set of 42 types of MELs, with 5 to 7 brands per type, were evaluated. For the purpose of accuracy and convenience for FFT, the sampling frequency of the data acquisition (DAQ) system was set to 30.72 kS/sec. Lower sampling frequencies, such as 7.68 kS/sec and 3.84 kS/sec, were also tested with the cross validation results being relatively the same. Similar to a known SSOM classifier, a cross validation mechanism tested the performance of the one-against-all SVM identifier 50, as shown in FIG. 4. This shows the process of how to cross-validate the performance of a one-against-all SVM classifier 58. The collected data 52 is divided into two sets: training 54 and testing 56. The SVM classifier 58 is first trained by the training data and then performs identification 60 on the testing data. The results 62,64 are compared with the actual labeling of the testing data to evaluate the success rate 66 and the performance of the SVM identifier 50.

(77) Referring to FIG. 5, a system 70 provides V/I waveform measurement 72, high-dimensional feature extraction 74, a multi-class one-against-all SVM 76 employing load feature clustering by load types 78, an SVM classifier construction 80, and load identification results 82. The system 70 also includes a processor 84 structured to determine at least six different load features from a sensed voltage signal and a sensed current signal for a corresponding one of a plurality of different electric loads, and identify, at 82, a load type of the different electric load types by relating the different load features to the SVM database 76.

(78) Hybrid SSOM/SVM Classifier

(79) The self-organizing map (SOM) is an unsupervised artificial neural network that is trained using competitive learning. That is, all neurons compete for the right to respond to the input data but only one neuron will win at a time. The training result of a basic SOM is a low-dimensional (typically two-dimensional), discretized grid of neurons with similar characteristics as the training samples. MELs that are similar or share common features in the input space are mapped to neurons that are positioned close to one another to form a cluster 90, as shown in FIG. 3A. With a relatively large number of MELs types, the output grid of the SSOM would get very crowded and thus unreadable, as shown in the output grid 44 of FIG. 3B.

(80) For the purpose of MELs identification, different types of loads with similar power supply units or features are partitioned into the same cluster. For example, DVD players and set-top boxes are very similar in both front-end power supply units and steady-state operating characteristics. Therefore, in a trained SSOM, DVD players and set-top boxes are classified into one cluster. However, the disadvantage is that it is difficult for a SSOM classifier to distinguish DVD players from set-top boxes. The SVM classifier 58 (FIG. 4), on the other hand, performs well when handling similar, but not identical, sets of feature vectors. Therefore, a hybrid SSOM/SVM achieves better performance than only the SSOM classifier or only a multi-class one-against-all SVM classifier.

(81) Based on front-end electronic circuit topologies of MELs, electrical operation principles or the functional nature of MELs, and user usage behaviors, MELs are divided into seven example load categories with distinct steady-state features: resistive loads (R); reactive predominant loads (X); electronic loads with power factor correction circuit (P); electronic loads without power factor correction circuit (NP); linear power supply using transformer to boost voltage (T); phase angle controllable loads (PAC); and complex structures (M). For example and without limitation, different electric load types having distinct steady-state features, such as resistive appliances, motor driven appliances, electronically fed appliances, non-linear loads with direct AC connections, and other unknown load types can be employed.

(82) The architecture of the hybrid SSOM/SVM classifier 100 is shown in FIG. 6. First, MELs data are collected to form a database at 102. Then, the SSOM is trained with all available MELs data to form a clustered grid, and a multi-class one-against-all SVM is trained for each of the seven example load categories at 104. Next, an unknown MEL is identified as being one of the seven example load categories at 106. Finally, using the trained SVM for the identified load category, the MEL type 109 is identified at 108.

(83) Referring to FIG. 7, a system 110 includes a processor 111, V/I waveform measurement 112, high-dimensional feature extraction 114, a hybrid hierarchical SSOM/SVM 116 including a top-layer SSOM 118 providing load feature clustering by load categories 119 and a second-layer SVM 120 providing load feature clustering by load type under each load category 121, SSOM/SVM classifier construction 122, and load identification results 124. For example and without limitation, the steady-state features of the determined load feature vector, as part of the high-dimensional feature extraction 114, can include voltage-current trajectory features selected from the group consisting of area, eccentricity, and Hausdorff distance.

(84) Area, A, refers to the area enclosed by a V-I trajectory. Area is proportional to the magnitude of the phase shift between the voltage and the current. If current leads voltage, then Area has a positive sign. If current lags voltage, then Area becomes negative. Area is directly calculated from the coordinates of the voltage and current points, (x.sub.i, y.sub.i), on the V-I trajectory.

(85) Eccentricity, E, is the measure of the aspect ratio of a shape, and is the ratio of the length of the major axis to the length of the minor axis. This feature helps to identify the shape of a voltage or current waveform.

(86) The Hausdorff distance, or Hausdorff metric, also called Pompeiu-Hausdorff distance, measures how far two subsets of a metric space are from each other. It turns the set of non-empty compact subsets of a metric space into a metric space in its own right. The Hausdorff distance is the longest distance one can be forced to travel by an adversary who chooses a point in one of the two sets, from where you then must travel to the other set. In other words, it is the farthest point of a set that you can be to the closest point of a different set.

Example 2

(87) Table 2 summarizes several tests to show the performance of the disclosed method and system 110 for MELs identification, including the testing success rate. These test the performance of the multi-class one-against-all SVM classifier 58 of FIG. 4 on a relatively large number of MELs including a selection of typical MELs.

(88) TABLE-US-00002 TABLE 2 Identification MELs Success Rate (%) Compact 98.67 Fluorescent Lights Fluorescent 100.00 Lights Incandescent 100.00 Lights Fan 100.00 Printer 99.66 Cellphone 100.00 Charger DVD player 98.66 Heater 100.00 LCD TV 99.72 LED TV 93.33 Microwave 100.00 Plasma TV 89.66 Set Top Box 100.00

Example 3

(89) The multi-class one-against-all SVM classifier 58 of FIG. 4 is compared against SSOM classifier performance with different amounts of data. One advantage of the disclosed multi-class one-against-all SVM classifier 58 is that it employs a relatively small amount of training data compared with other classifiers. Some test results containing three loads and different choices of cross-validation are shown in Table 3, in which the ratios between training data and testing data are indicated in the first column. The following tests are carried out when both classifiers are trained by feature vectors from 32 different models of 12 types of MELs, with 3600 feature points for each model. For all of the tests disclosed herein, 512-point FFTs are done to calculate the features.

(90) TABLE-US-00003 TABLE 3 Success Rate DVD TV Fan Total SVM 100% 93.77% 100% 97.92% (5% for training, 95% for test) SVM 100% 93.43% 100% 97.81% (10% for training, 90% for test) SVM 100% 100% 100% 100% (20% for training, 80% for test) SVM 100% 100% 100% 100% (30% for training, 70% for test) SSOM 49.36% 81.99% 97.34% 76.23% (5% for training, 95% for test) SSOM 91.30% 87.81% 97.69% 92.26% (10% for training, 90% for test) SSOM 96.74% 94.03% 93.92% 94.90% (20% for training, 80% for test) SSOM 99.75% 99.83% 98.00% 99.19% (67% for training, 33% for test)

(91) From Table 3, it is clear to see that the multi-class one-against-all SVM classifier 58 can get a 100% testing success rate with only 20% of the total data, which is much better than a SSOM classifier.

Example 4

(92) The following tests are for the hybrid SSOM/SVM classifier 100 of FIG. 6. These tests are carried out with the SSOM 118 (FIG. 7) being trained by feature vectors from 32 different models of 12 types of MELs, with 3600 feature points for each model. In this case, the output grid of the SSOM 118 gets relatively very crowded and unreadable. Therefore, it is not shown. In the trained SSOM 118, MELs with similar features are mapped into the same cluster. The voltage profiles 132,136,140 and current profiles 134,138,142 of three types of TVs (e.g., LCD, LED and plasma) that would be clustered together are shown in FIGS. 8A-8F. These three MELs share similar front-end power supply units as well as feature vectors, and thus it is very difficult to distinguish between them.

(93) The different types of TVs are grouped into one cluster by the SSOM classifier. However, the SSOM classifier gets an average success rate around only about 85% to identify each type of TV. In contrast, within the hybrid SSOM/SVM classifier 100, the average testing success rate is greater than 95%. These success rates, for similar MELs, are shown in Table 4.

(94) TABLE-US-00004 TABLE 4 Success Rate LCD TV LED TV Plasma TV Average SSOM 80.17% 97.85% 85.25% 85.28% identifier Hybrid 98.30% 78.89% 98.96% 92.05% SSOM/SVM identifier (20% data for training) Hybrid 95.99% 90.95% 98.85% 95.26% SSOM/SVM identifier (30% data for training)

(95) From Table 4, it is clear to see that the more training data for the SVM 120 in the hybrid SSOM/SVM identifier 116, the better performance it has. But the SVM training in the hybrid SSOM/SVM identifier 116 requires far less data than a pure SSOM classifier.

(96) Compared with methods, such as only SSOM or only SVM classifiers, the disclosed hybrid SSOM/SVM identifier 116 provides better performance in the sense of accuracy, robustness and applicability. The SSOM identifier 118 first extracts information from the relatively large amount of training data and stores that simplified information in the trained neuron grid. When an input feature vector is presented to the hybrid SSOM/SVM identifier 116, it first determines which load category it falls into, and then employs an SVM discriminator function 120 for each category to get a robust and correct identification decision.

(97) The disclosed concept provides a hybrid SSOM/SVM classifier 116 for the purpose of intelligent and nonintrusive MELs classification and identification with relatively high identification accuracy, robustness and applicability with respect to the diversity of different models of each type of MEL. This hybrid classifier 116 employs the power of an SSOM classifier 118 for MELs to first classify the relatively large amount of MELs models into several clusters. Within each cluster, a more accurate identification decision is made by a multi-class one-against-all SVM classifier 120. This hybrid SSOM/SVM classifier 116 employs steady-state conditions. For MELs with similar power supply units, the disclosed hybrid SSOM/SVM identifier 116 provides hard (or absolute) decisions. Preferably, soft (or probabilistic) decisions should be provided for the electric load identification problem.

(98) While specific embodiments of the disclosed concept have been described in detail, it will be appreciated by those skilled in the art that various modifications and alternatives to those details could be developed in light of the overall teachings of the disclosure. Accordingly, the particular arrangements disclosed are meant to be illustrative only and not limiting as to the scope of the disclosed concept which is to be given the full breadth of the claims appended and any and all equivalents thereof.