SYSTEM AND METHOD EMPLOYING A SELF-ORGANIZING MAP LOAD FEATURE DATABASE TO IDENTIFY ELECTRIC LOAD TYPES OF DIFFERENT ELECTRIC LOADS

• US 20130138651A1
• Filed: 11/28/2011
• Published: 05/30/2013
• Est. Priority Date: 11/28/2011
• Status: Active Grant

• Current Assignee: Eaton Intelligent Power Limited (Eaton Corporation PLC.), Georgia Tech Research Corporation (University System of Georgia)
• Original Assignee: Bin Lu, Liang Du
• Inventors: LU, BIN, Harley, Ronald G., Yang, Yi, Sharma, Santosh K., Zambare, Prachi, Madane, Mayura A., Du, Liang
• Granted Patent: US 8,756,181 B2
• Time in Patent Office: Days
• US Class Current: 707/737
• CPC Class Codes: G01D 2204/24 Identification of individual loads, e.g. by analysing current/voltage waveforms
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning
• CPC Class Codes: Y04S 20/30 Smart metering, e.g. specially adapted for remote reading

Abstract

A method identifies electric load types of a plurality of different electric loads. The method includes providing a self-organizing map load feature database of a plurality of different electric load types and a plurality of neurons, each of the load types corresponding to a number of the neurons; employing a weight vector for each of the neurons; sensing a voltage signal and a current signal for each of the loads; determining a load feature vector including at least four different load features from the sensed voltage signal and the sensed current signal for a corresponding one of the loads; and identifying by a processor one of the load types by relating the load feature vector to the neurons of the database by identifying the weight vector of one of the neurons corresponding to the one of the load types that is a minimal distance to the load feature vector.

Claims

1. A method of identifying electric load types of a plurality of different electric loads, said method comprising:

providing a self-organizing map load feature database of a plurality of different electric load types and a plurality of neurons, each of said different electric load types corresponding to a number of said neurons;

employing a weight vector for each of said neurons;

sensing a voltage signal and a current signal for each of said different electric loads;

determining a load feature vector comprising at least four different load features from said sensed voltage signal and said sensed current signal for a corresponding one of said different electric loads; and

identifying by a processor one of said different electric load types by relating the load feature vector to the neurons of said self-organizing map load feature database by identifying the weight vector of one of said neurons corresponding to said one of said different electric load types that is a minimal distance to the load feature vector.

2. The method of claim 1 further comprising:

employing the self-organizing map load feature database as a hierarchical and scalable self-organizing map load feature database.

3. The method of claim 1 further comprising:

employing with said load feature vector load characteristics selected from the group consisting of steady state current waveform and power quality related quantities, voltage-current trajectory graphical representations under start-up and steady-state conditions, short-term harmonic content transient state characteristics of load current, transient power content transient state characteristics of load current, load event detection patterns, operating status profile patterns, and operating modes patterns.

4. The method of claim 1 further comprising:

employing with said self-organizing map load feature database a first layer including load feature clustering by a plurality of load categories, and a second layer including load feature clustering by load type under each of the load categories.

5. The method of claim 4 further comprising:

employing said load categories selected from the group consisting of resistive appliances, motor driven appliances, electronically fed appliances, non-linear loads with direct alternating current connections, and unknown loads.

6. The method of claim 4 further comprising:

employing as said second layer a plurality of sub-self-organizing maps, each of which corresponds to one of the load categories having a plurality of different load types.

7. The method of claim 1 further comprising:

mapping said different load features of said load feature vector to a two-dimensional or three-dimensional space to cluster and organize load features of said different electric load types.

8. The method of claim 1 further comprising:

including with said load feature vector steady state features selected from the group consisting of displacement power factor, true power factor, current total harmonic distortion, current K-factor, current crest factor, admittance, and voltage-current trajectory graphical representation.

9. The method of claim 8 further comprising:

including with said load feature vector voltage-current trajectory features selected from the group consisting of area, eccentricity, and Hausdorff distance.

10. The method of claim 1 further comprising:

providing said identifying by the processor in real-time.

11. The method of claim 1 further comprising:

employing a U-matrix via a distance metric with said identifying the weight vector.

12. The method of claim 1 further comprising:

providing the self-organizing map load feature database with a self-organizing map defining a plurality of load clusters and a plurality of boundaries between the load clusters.

13. The method of claim 12 further comprising:

automatically defining un-identified loads among said load clusters based upon said boundaries.

14. The method of claim 1 further comprising:

employing with said identifying the weight vector an average squared Euclidean distance to a plurality of neurons in a class corresponding to said one of said plurality of different electric load types.

15. The method of claim 14 further comprising:

employing i as an index;

employing ω

_i, as said class;

employing x as said load feature vector;

for each of said plurality of different electric load types, employing a group of values of said self-organizing map load feature database having a mean y_i and a square covariance matrix Σ

_i;

employing with said identifying the weight vector a point-to-cluster function of average squared Euclidean distance from said load feature vector to every point in said class, ω

_i, corresponding to said one of said plurality of different electric load types;

employing Tr( ) as a trace of the square covariance matrix Σ

_i; and

determining the average squared Euclidean distance from;

d_E(x,ω

_i)²=(x−

y_i)^T(x−

y_i)+Tr(Σ

_i).

16. The method of claim 1 further comprising:

training said self-organizing map load feature database with a statistical distance measure.

17. The method of claim 16 further comprising:

employing i as an index;

training said self-organizing map load feature database with x as an input load feature vector;

for each class of said plurality of different electric load types, employing a group of values of said self-organizing map load feature database having a mean y_i and a square covariance matrix Σ

_i;

employing as said statistical distance measure a Mahalanobis distance from said input load feature vector to the weight vector, y, of each of said neurons of said self-organizing map load feature database;

employing an average of a within-class square covariance matrix Σ

_i for the class that the weight vector of a corresponding one said neurons belongs to as the square covariance matrix of both said input load feature vector and the last said weight vector; and

determining the Mahalanobis distance from;

d_M(x,y)²=(x−

y)^TΣ

<sup>−

1</sup>(x−

y).

18. The method of claim 17 further comprising:

providing a plurality of samples of extracted load features from said sensed voltage signal and said sensed current signal for one of said different electric loads;

employing k as an index;

employing ω

_i as said class;

quantizing said determined load feature vector as x′

[k] based on a predetermined resolution;

identifying said quantified determined load feature vector x′

[k] appearing in said samples a plurality, T_k, of times;

identifying said one of said different electric load types as a best match to said ω

_i a plurality, T_ki, of times; and

determining likelihood of said identifying by the processor of said one of said different electric load types from;

19. The method of claim 1 further comprising:

training said self-organizing map load feature database to comprise;

a plurality, M, of classes of said different electric load types,

a plurality, K, of said neurons, and

a plurality, l, of dimensions of said weight vector for each of said neurons;

classifying a subset of a plurality, K_s, of said neurons, with said K_s being less than said K; and

including for each of said classes a mean and a covariance matrix.

20. The method of claim 19 further comprising:

comparing said load feature vector to all of said plurality, K_s, of said neurons for each of said classes; and

said identifying by the processor comprises identifying said weight vector of one of said neurons of one of said classes that has a minimal distance to said load feature vector.

21. The method of claim 1 further comprising:

employing said different load features selected from the list consisting of displacement power factor, current total harmonic distortion, current RMS value, current crest factor, current third order harmonic amplitude, current third order harmonic phase angle with respect to voltage, current fifth order harmonic amplitude, and current fifth order harmonic phase angle with respect to voltage.

22. A system comprising:

a self-organizing map load feature database of a plurality of different electric load types and a plurality of neurons, each of said different electric load types corresponding to a number of said neurons, each of said neurons having a corresponding weight vector;

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 comprising at least four different load features from said sensed voltage signal and said sensed current signal for a corresponding one of said different electric loads, and

identify one of said different electric load types by relating the load feature vector to the neurons of said self-organizing map load feature database by identifying the weight vector of one of said neurons corresponding to said one of said different electric load types that is a minimal distance to the load feature vector.

23. The system of claim 22 wherein said self-organizing map load feature database comprises a first layer including load feature clustering by a plurality of load categories, and a second layer including load feature clustering by load type under each of the load categories.

24. The system of claim 23 wherein said second layer comprises a plurality of sub-self-organizing maps, each of which corresponds to one of the load categories having a plurality of different load types.