US 11,232,361 B1Granted

PMU date correction using a recovery method

AssigneeNorth China Electric Power University

InventorsTianshu Bi, Hao Liu, Zhiwei Yang, Ke Jia

Life of this patent

Estimating the term…

Abstract

A data-driven PMU bad data detection algorithm based on spectral clustering using single PMU data is described. The described algorithm does not require the system topology and parameters. First, a data identification method based on a decision tree is described to distinguish event data and bad data by using the slope feature of each set of data. Then, a bad data detection method based on spectral clustering is described. By analyzing the weighted relationships among all the data, this method can detect the bad data that has a small deviation.

Claims1

Dependent claims

A method comprising:

receiving, at a processor of a phasor data concentrator (PDC), input data from a phasor measurement unit (“PMU”);
calculating, using the processor, a slope for each data point in the input data, wherein the slope of each data point ki is calculated as follows:
k
i
=


X
i
+
1

-

X
i

t
i
+
1
-
t
i

(
1
)
wherein during a first mode of operation, the slope for each data point is smaller than a first threshold and during a second mode of operation, the slope for each data point is larger than the first threshold;
generating, using the processor, an output of a decision tree as a binary classification model, wherein the processor is configured to:
set a depth p and a threshold of information gain ratio ε;
calculate (2) a gain ratios of features (a, b, c, d) for each data point, wherein the gain ratio of a feature b is calculated as follows:
o
⁡
(
D
,
b
,
s
i
)
=
O
⁡
(
D
,
b
,
s
i
)
I
⁡
(
b
)
(
5
)
I
⁡
(
b
)
=
-
∑
β
∈
{
-
,
+
}
⁢

D
s
i
β


D

⁢
log
2
⁢

D
s
i
β


D

(
6
)
where I(b) is the intrinsic value;
select (2) a largest gain ration to compare with ε;
if the gain ration is greater than ε,
use (3) a corresponding feature as a feature of the first division;
place (4) data Xi whose feature bi≤sb in a first collection;
and
place (5) the data Xi whose feature bi>sb in a second collection;
and
if the gain ration is less than ε, repeat the above steps (1)-(5) recursively up to p times until labels of data in one of the first collection and the second collection are the same;
wherein a total information entropy of the training data D is calculated by:
Z
⁡
(
D
)
=
-
∑
i
=
1
2
⁢
z
i
⁢
log
2
⁢
z
i
(
2
)
where z1 indicates a proportion of step point in D, z2 indicates a proportion of non-step point in D, and Z(D) is an uncertainty of the data label;
determine, using the processor and the output, whether a test data D′ contains a step point;
filter, using the processor, an amplitudes of data points before the step point occurs and an amplitudes of data points after the step point occurs using a 3σ rule, as follows:
P(∥Xi|−μ|≤3σ)≤99.73% (8)
where μ represents the mean value of the amplitudes, and σ is the standard deviation of the amplitudes;
construct a Laplacian matrix L:
L=Dd−W (15)
wherein a similarity matrix W is established according to a similarity between any two data points, and the similarity of any two data points is defined as follows:
ω
ij
=
{
exp
⁡
(
-

r
i
-
r
j

2
δ
2
)
⁢
,
i
≠
j
0
,
i
=
j
(
13
)
where δ is a scale parameter;
wherein a degree matrix Dd is a diagonal matrix as follows:
D
d
=
[
d
1
d
2
⋱
d
n
-
1
d
n
]
,
d
i
=
∑
j
=
1
n
⁢
ω
ij
(
14
)
decompose, using the processor, eigenvalues of the Laplacian matrix L to find smallest m eigenvectors, wherein a minimum value of the following is equal to a sum of the m smallest eigenvalues of the Laplacian matrix L:
{
min
H
∈
•
n
×
k
Tr
⁡
(
H
T
⁢
L
⁢
H
)
s
.
t
.
H
T
⁢
H
=
I
(
22
)
construct, using the processor, a matrix F composed of eigenvectors corresponding to a minimum m eigenvalues of the Laplacian matrix L;
calculate, using the processor, C1 which contains normal data, and C2 which contains bad data using use K-means to cluster the matrix F;
output, using the processor, C1 and C2;
and
transmit, from the PDC using the processor, C1 and C2 to a server at a power station to:
correct the input data using a recovery method;
monitor an operating status of a power system;
and
perform closed-loop control on the power system.

247 more elements

Dependencies are inferred from the claim text.

Facts

Application: 17/376,353Filed 15 Jul 2021 · continuation
Priority: 25 Feb 2021
Granted: 25 Jan 2022
Term ends: Not estimatedThe source records don’t support an estimate.

Classification15 codes

G06N 20/00
Machine learning
G01R 19/2513
Using digital measurement techniques: arrangements for monitoring electric power systems, e.g. power lines or loads; Logging
G06F 18/21322
Based on discrimination criteria, e.g. discriminant analysis: rendering the within-class scatter matrix non-singular
G06F 18/22
Analysing: matching criteria, e.g. proximity measures

Family

PCT/CN2021/077901Parent of this continuation · filed 25 Feb 2021
Application17/215,724Parent of this continuation · filed 29 Mar 2021
US 11,170,304 B1Parent of this continuation
US 11,232,361 B1This patent

From the US filing history. Foreign family members aren’t listed.

Timeline

Citations

Open elsewhere

patentgrep

Loading patent…

Abstract

A method comprising:

receiving, at a processor of a phasor data concentrator (PDC), input data from a phasor measurement unit (“PMU”);
calculating, using the processor, a slope for each data point in the input data, wherein the slope of each data point ki is calculated as follows:
k
i
=


X
i
+
1

-

X
i

t
i
+
1
-
t
i

(
1
)
wherein during a first mode of operation, the slope for each data point is smaller than a first threshold and during a second mode of operation, the slope for each data point is larger than the first threshold;
generating, using the processor, an output of a decision tree as a binary classification model, wherein the processor is configured to:
set a depth p and a threshold of information gain ratio ε;
calculate (2) a gain ratios of features (a, b, c, d) for each data point, wherein the gain ratio of a feature b is calculated as follows:
o
⁡
(
D
,
b
,
s
i
)
=
O
⁡
(
D
,
b
,
s
i
)
I
⁡
(
b
)
(
5
)
I
⁡
(
b
)
=
-
∑
β
∈
{
-
,
+
}
⁢

D
s
i
β


D

⁢
log
2
⁢

D
s
i
β


D

(
6
)
where I(b) is the intrinsic value;
select (2) a largest gain ration to compare with ε;
if the gain ration is greater than ε,
use (3) a corresponding feature as a feature of the first division;
place (4) data Xi whose feature bi≤sb in a first collection;
and
place (5) the data Xi whose feature bi>sb in a second collection;
and
if the gain ration is less than ε, repeat the above steps (1)-(5) recursively up to p times until labels of data in one of the first collection and the second collection are the same;
wherein a total information entropy of the training data D is calculated by:
Z
⁡
(
D
)
=
-
∑
i
=
1
2
⁢
z
i
⁢
log
2
⁢
z
i
(
2
)
where z1 indicates a proportion of step point in D, z2 indicates a proportion of non-step point in D, and Z(D) is an uncertainty of the data label;
determine, using the processor and the output, whether a test data D′ contains a step point;
filter, using the processor, an amplitudes of data points before the step point occurs and an amplitudes of data points after the step point occurs using a 3σ rule, as follows:
P(∥Xi|−μ|≤3σ)≤99.73% (8)
where μ represents the mean value of the amplitudes, and σ is the standard deviation of the amplitudes;
construct a Laplacian matrix L:
L=Dd−W (15)
wherein a similarity matrix W is established according to a similarity between any two data points, and the similarity of any two data points is defined as follows:
ω
ij
=
{
exp
⁡
(
-

r
i
-
r
j

2
δ
2
)
⁢
,
i
≠
j
0
,
i
=
j
(
13
)
where δ is a scale parameter;
wherein a degree matrix Dd is a diagonal matrix as follows:
D
d
=
[
d
1
d
2
⋱
d
n
-
1
d
n
]
,
d
i
=
∑
j
=
1
n
⁢
ω
ij
(
14
)
decompose, using the processor, eigenvalues of the Laplacian matrix L to find smallest m eigenvectors, wherein a minimum value of the following is equal to a sum of the m smallest eigenvalues of the Laplacian matrix L:
{
min
H
∈
•
n
×
k
Tr
⁡
(
H
T
⁢
L
⁢
H
)
s
.
t
.
H
T
⁢
H
=
I
(
22
)
construct, using the processor, a matrix F composed of eigenvectors corresponding to a minimum m eigenvalues of the Laplacian matrix L;
calculate, using the processor, C1 which contains normal data, and C2 which contains bad data using use K-means to cluster the matrix F;
output, using the processor, C1 and C2;
and
transmit, from the PDC using the processor, C1 and C2 to a server at a power station to:
correct the input data using a recovery method;
monitor an operating status of a power system;
and
perform closed-loop control on the power system.

247 more elements

Facts

Application

17/376,353Filed 15 Jul 2021 · continuation

Priority

25 Feb 2021

Granted

25 Jan 2022

Term ends

Not estimatedThe source records don’t support an estimate.

Classification15 codes

G06N 20/00
Machine learning
G01R 19/2513
Using digital measurement techniques: arrangements for monitoring electric power systems, e.g. power lines or loads; Logging
G06F 18/21322
Based on discrimination criteria, e.g. discriminant analysis: rendering the within-class scatter matrix non-singular
G06F 18/22
Analysing: matching criteria, e.g. proximity measures