User electricity stealing detection method based on data driving

Through the Gram angle field method and principal component analysis method combined with the K-means clustering method, Pearson correlation coefficient is used to judge abnormal users, and the problems of low accuracy and high miss detection rate in the existing technology are solved, achieving high efficiency and low Cost of user power theft detection.

CN118916740APending Publication Date: 2024-11-08国网福建省电力有限公司营销服务中心 +1

Patent Information

Application Number
CN202410950124.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing data-driven user power theft detection method has problems such as low accuracy and high missed detection rate, and is relatively high in training costs, strong engineering application limitations and high cost.

Method used

The Gram angle field method is used to convert the electricity consumption data into a second-order matrix, and the dimensionality is reduced by principal component analysis method, and the K-means clustering method is used to combine Pearson's correlation coefficient to judge abnormal users, reducing calculation costs.

Benefits of technology

It improves the accuracy of electricity abnormality recognition, reduces the leakage detection rate, and is simple and efficient, with low application cost.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916740A_ABST
    Figure CN118916740A_ABST
Patent Text Reader

Abstract

The invention provides a user electricity stealing detection method based on data driving, and the method comprises the following steps: 1, obtaining the electricity utilization data of n users in a certain transformer area through an electricity utilization information collection system, obtaining the load data of a 96-point power curve every day, and carrying out the data preprocessing; 2, converting the preprocessed discrete point data into a second-order matrix by adopting a GASF (Grubrum Angle Field), and extracting data features; 3, performing dimension reduction on the GASF matrix by adopting a principal component analysis method to obtain a data feature matrix; step 4, classifying the data feature matrix after dimension reduction by using a K-means clustering method to obtain a clustering center sequence; and 5, calculating the similarity of the clustering center sequence by adopting a Pearson correlation coefficient, and setting a threshold value to judge an abnormal user. According to the technical scheme, the accuracy of electricity consumption abnormity identification can be improved, and the omission ratio is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of electric energy metering and collection, and in particular to a data-driven method for detecting electricity theft by users. Background Art

[0002] With the development and improvement of electricity consumption information collection systems, the increasing amount of user data has made data-driven user electricity theft detection a research hotspot. If users' electricity theft behaviors are discovered in time, relevant measures can be taken in advance to reduce losses. At present, the main data-driven user electricity theft detection methods are:

[0003] (1) Clustering-based electricity theft detection. Based on the user's electricity consumption data (voltage, current, power factor, power consumption, line loss rate, etc.), characteristic indicators that characterize the user's electricity consumption behavior are obtained. The user's electricity consumption characteristic indicators are clustered in an unsupervised manner to find abnormal users who deviate from the normal user clustering cluster. However, traditional clustering methods are easily affected by noise and have low accuracy.

[0004] (2) Classification-based electricity theft detection. The supervised methods such as decision trees, support vector machines and deep learning are used to train samples with known category labels, build an abnormal electricity consumption behavior recognition model, and detect abnormal electricity users. However, the classification-based electricity theft detection method has high training costs and requires a large amount of data. In addition, the number of samples with electricity theft labels is usually far less than that of normal electricity usage samples, resulting in a high missed detection rate.

[0005] (3) Electricity theft detection based on state estimation. The state estimation method is used to estimate the state quantity of the target user based on the measured values ​​of other nodes (such as the node voltage and injected power of each node), so as to perform electricity theft detection based on the state estimation of the distribution network. However, in addition to mastering the detailed network topology and parameters, electricity theft detection based on state estimation also requires that the relationship between the line transformer and the user meter is correct and strictly consistent with the actual system. It has strong limitations in engineering applications and is costly. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide a data-driven method for detecting electricity theft by users, so as to improve the accuracy of identifying abnormal electricity consumption and reduce the missed detection rate, while having a lower computational cost.

[0007] To achieve the above object, the present invention adopts the following technical solution: a data-driven user electricity theft detection method, comprising the following steps:

[0008] Step 1: Obtain the electricity consumption data of all users in a certain area through the electricity consumption information collection system, obtain the load data of the 96-point power curve every day, and perform data preprocessing;

[0009] Step 2: Use Gram angular field GASF to convert the preprocessed discrete point data into a second-order matrix and extract data features;

[0010] Step 3: Use principal component analysis to reduce the dimension of the Gram angle field GASF matrix to obtain the data feature matrix;

[0011] Step 4: Use the K-means clustering method to classify the reduced-dimensional data feature matrix to obtain a cluster center sequence;

[0012] Step 5: Use the Pearson correlation coefficient to calculate the similarity of the cluster center sequence and set a threshold to determine abnormal users.

[0013] In a preferred embodiment, the data preprocessing method in step 1 is specifically as follows: for user load data with a data missing amount not exceeding 20%, the load data of 96 points per day are supplemented by interpolation method:

[0014]

[0015] In the formula, is the i-th load data on the t-th day, is the i+1th load data on the tth day, is the i-1th load data on the tth day, represents supplementary data, and φ represents an empty set; if there is a point set that does not meet the above conditions, all load data for that day will be eliminated.

[0016] In a preferred embodiment, the step 2 specifically includes:

[0017] Step 21: For each user, based on the completed load data, convert the daily load sequence of length 96 on the tth day into Normalized to the interval [-1,1] to form a new load series Among them are:

[0018]

[0019] In the formula, max(X t ) is the maximum load value in the Xt load sequence; min(X t ) is the minimum load value in the Xt load sequence;

[0020] Step 22: Convert the normalized load sequence to the polar coordinate system, correspond the load value to the cosine value of the angle, and the timestamp to the radius, that is:

[0021]

[0022] In the formula, t i For point timestamp, N = 96 is the number of all time points contained in a load sequence; where the polar axis r i Preserve the temporal relationship of the original sequence; polar angle Preserve the numerical relationship of the original sequence;

[0023] Step 23: Use the cosine function of the sum of two angles to represent the correlation of load data in different time intervals, and convert the discrete load data of the tth day into a second-order matrix G t :

[0024]

[0025] In a preferred embodiment, the step 3 specifically includes:

[0026] Step 31: Transform the second-order matrix G t It is represented as an N×N matrix X,

[0027]

[0028] According to the correlation coefficient formula r ij = Cov(x i ,x j ), calculate the correlation coefficient, that is:

[0029] The correlation coefficient is expressed as the covariance matrix R,

[0030]

[0031] Step 32: Perform eigendecomposition on the covariance matrix R to obtain the eigenvalue λ and the corresponding eigenvector υ of the covariance matrix, that is, Rυ=λυ;

[0032] Step 33: Calculate the contribution rate and cumulative contribution rate of each principal component, and select the first z principal components so that the cumulative contribution rate is greater than 90%; arrange the eigenvectors from large to small according to the corresponding eigenvalues, and take the first z columns to form a matrix W, that is, an N×z-order matrix; among them, the contribution rate P of each principal component i for:

[0033]

[0034] Cumulative contribution rate S(P Z )for

[0035]

[0036] Step 34: Finally, calculate the data feature matrix after reducing to the z dimension, that is, an N×z-order matrix.

[0037] In a preferred embodiment, the step 4 specifically includes:

[0038] Step 41: The principal component data obtained in step 3 is used to form a data set, and k objects are selected from the data set as initial clustering centers;

[0039] Step 42: Calculate the Euclidean distance between the remaining data points in the data set and each initial cluster center;

[0040] Step 43: assigning the data point to the cluster represented by the nearest cluster center according to the Euclidean distance;

[0041] Step 44: Calculate the average value of the objects in each cluster and update the cluster center;

[0042] Steps 42 to 44 are repeated until each cluster no longer changes, and finally the cluster center sequence is obtained:

[0043] C={C1,C2,...,C K} (12)

[0044] In the formula, C i is the sequence of each cluster center.

[0045] In a preferred embodiment, step 5 specifically includes: using the Pearson correlation coefficient to calculate the similarity ρ of each user's daily cluster center sequence C, when ρ is lower than the similarity experience value η, that is, ρ<η, it is considered that the user's electricity consumption is abnormal.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. A data-driven method for detecting electricity theft is proposed, which can extract features with large contrast from discrete point data of user profiles, improve the accuracy of identifying abnormal electricity consumption, and reduce the missed detection rate.

[0048] 2. The Gram's angle field method, principal component analysis method and clustering method used in the data-driven user electricity theft detection method proposed in the present invention are simple, efficient and low in application cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 The figure is a schematic diagram of a method flow of a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0051] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0052] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0053] The above method can accurately extract the power load characteristics of each user. When the user's power consumption characteristics are fixed, the daily similarity is high. When a user is suspected of stealing electricity, the cluster center will shift significantly from the date of the theft. From the date of the theft, the similarity is significantly reduced compared with before the theft. Therefore, the date of the theft can be accurately located.

[0054] The purpose of the present invention is to provide a data-driven method for detecting electricity theft by users. Figure 1 , including the following steps:

[0055] Step 1: Obtain the electricity consumption data of all users in a certain area through the electricity consumption information collection system, obtain the load data of the 96-point power curve every day, and perform data preprocessing.

[0056] Step 2: Use Gram Angular Field (GASF) to convert the preprocessed discrete point data into a second-order matrix and extract data features.

[0057] Step 3: Use principal component analysis to reduce the dimension of the GASF matrix and obtain the data feature matrix.

[0058] Step 4: Use the K-means clustering method to classify the reduced-dimensional data feature matrix to obtain a cluster center sequence.

[0059] Step 5: Use the Pearson correlation coefficient to calculate the similarity of the cluster center sequence and set a threshold to determine abnormal users.

[0060] In the above data-driven user electricity theft detection method, the data preprocessing method in step 1 is specifically as follows:

[0061] In order to prevent missing data from affecting the calculation, for user load data with a missing amount of no more than 20%, the interpolation method is used to complete the load data of 96 points per day:

[0062]

[0063] In the formula, is the i-th load data on the t-th day, is the i+1th load data on the tth day, is the i-1th load data on the tth day, represents supplementary data, and φ represents an empty set. If there is a point set that does not meet the above conditions, all load data for that day will be removed.

[0064] In the above-mentioned electricity theft detection method based on data dimension reduction and cluster analysis, the implementation of the GASF method described in step 2 includes the following steps:

[0065] (1) For each user, based on the complete load data, the daily load sequence of length 96 on day t is Normalized to the interval [-1,1] to form a new load sequence Among them are:

[0066]

[0067] In the formula, max(X t ) is the maximum load value in the Xt load sequence; min(X t ) is the minimum load value in the Xt load sequence.

[0068] (2) Convert the normalized load sequence to the polar coordinate system, correspond the load value to the cosine value of the angle, and the timestamp to the radius, that is:

[0069]

[0070] Where, t i For point timestamp, N = 96 is the number of all time points contained in a load sequence. i The temporal relationship of the original sequence is retained; the polar angle The numerical relationship of the original sequence is preserved.

[0071] (3) In order to extract data features, the cosine function of the sum of two angles is used to represent the correlation of load data in different time intervals. The discrete load data of the tth day is converted into a second-order matrix G t :

[0072]

[0073] In the above data-driven method for detecting electricity theft by users, the dimension reduction step of the principal component analysis method in step 3 is specifically as follows:

[0074] (1) For the convenience of demonstration, the second-order matrix G t It is represented as an n×m matrix X,

[0075]

[0076] According to the correlation coefficient formula r ij = Cov(x i ,xj ), calculate the correlation coefficient, that is:

[0077]

[0078] The correlation coefficient is expressed as the covariance matrix R,

[0079]

[0080] (2) Establish the characteristic equation |λI-R=0|, perform eigendecomposition on the covariance matrix R, and find the eigenvalue λ of the covariance matrix k and the corresponding eigenvector υ k , that is, Rυ k =λ k υ k .

[0081] (3) Calculate the contribution rate and cumulative contribution rate of each principal component, and select the first z principal components so that the cumulative contribution rate is greater than 90%. Arrange the eigenvectors from large to small according to the corresponding eigenvalues, and take the first z columns to form the matrix W, that is, an m×k-order matrix. Among them, the contribution rate P of each principal component is i for:

[0082]

[0083] Cumulative contribution rate S(P k )for

[0084]

[0085] (4) Finally, the data feature matrix after being reduced to the z dimension is calculated, that is, an N×z-order matrix.

[0086] In the above data-driven user electricity theft detection method, the K-means clustering method in step 4 is specifically:

[0087] (1) The principal component data obtained in step 3 constitute a data set, and k objects are selected from the data set as initial clustering centers.

[0088] (2) Calculate the Euclidean distances dist(x) between the remaining data points in the data set and the initial cluster centers. i ,x j ),

[0089]

[0090] In the formula, x i is the cluster center, x j are other data points, and D is the dimension of the data point.

[0091] (3) According to the Euclidean distance, the data points are assigned to the cluster represented by the closest cluster center.

[0092] (4) Calculate the average value of objects in each cluster and update the cluster center.

[0093] Repeat (2) to (4) until each cluster no longer changes, and finally obtain the cluster center sequence:

[0094] C={C1,C2,...,C K} (12)

[0095] In the formula, C i is the sequence of each cluster center.

[0096] In the above data-driven method for detecting electricity theft by users, the similarity calculation method and abnormal user identification described in step 5 are specifically as follows:

[0097] The Pearson correlation coefficient is used to calculate the similarity of each user's daily cluster center sequence C, where C k =(x k ,y k ), k=1,2,...,n. For two sets of data X={x1,x2,...,x n} and Y={y1,y2,...,y n}, its Pearson correlation coefficient ρ XY for:

[0098]

[0099] Where E(X) and E(Y) are means, Cov(X,Y) is covariance, σ X and σ Y is the standard deviation.

[0100] A judgment threshold η is set through experience, when ρ XY When it is lower than the similarity experience value η, that is, ρ XY When <η, it is considered that the user's electricity consumption is abnormal.

Claims

1. A data-driven method for detecting electricity theft by users. It is characterized in that The following steps are involved: Step 1: Obtain the electricity consumption data of all users in a certain area through the electricity consumption information collection system, obtain the load data of the 96-point power curve every day, and perform data preprocessing; Step 2: Use Gram angular field GASF to convert the preprocessed discrete point data into a second-order matrix and extract data features; Step 3: Use principal component analysis to reduce the dimension of the Gram angle field GASF matrix to obtain the data feature matrix; Step 4: Use the K-means clustering method to classify the reduced-dimensional data feature matrix to obtain a cluster center sequence; Step 5: Use the Pearson correlation coefficient to calculate the similarity of the cluster center sequence and set a threshold to determine abnormal users.

2. According to the data-driven method for detecting electricity theft by users according to claim 1, It is characterized in that The data preprocessing method in step 1 is as follows: for user load data with a data missing amount not exceeding 20%, the load data of 96 points per day are supplemented by interpolation method: In the formula, is the i-th load data on the t-th day, is the i+1th load data on the tth day, is the i-1th load data on the tth day, represents supplementary data, and φ represents an empty set; if there is a point set that does not meet the above conditions, all load data for that day will be eliminated.

3. According to the data-driven method for detecting electricity theft by users according to claim 1, It is characterized in that The step 2 specifically includes: Step 21: For each user, based on the completed load data, convert the daily load sequence of length 96 on the tth day into Normalized to the interval [-1,1] to form a new load series Among them are: In the formula, max(X t ) is X t The maximum load value in the load sequence; min(X t ) is X t The minimum load value in the load sequence; Step 22: Convert the normalized load sequence to the polar coordinate system, correspond the load value to the cosine value of the angle, and the timestamp to the radius, that is: Where, t i For point timestamp, N = 96 is the number of all time points contained in a load sequence; where the polar axis r i Preserve the temporal relationship of the original sequence; polar angle Preserve the numerical relationship of the original sequence; Step 23: Use the cosine function of the sum of two angles to represent the correlation of load data in different time intervals, and convert the discrete load data of the tth day into a second-order matrix G t :

4. According to the data-driven method for detecting electricity theft by users according to claim 1, It is characterized in that The step 3 specifically includes: Step 31: Transform the second-order matrix G t It is represented as an N×N matrix X, According to the correlation coefficient formula r ij = Cov(x i ,x j ), calculate the correlation coefficient, that is: The correlation coefficient is expressed as the covariance matrix R, Step 32: Perform eigendecomposition on the covariance matrix R to obtain the eigenvalue λ and the corresponding eigenvector υ of the covariance matrix, that is, Rυ=λυ; Step 33: Calculate the contribution rate and cumulative contribution rate of each principal component, and select the first z principal components so that the cumulative contribution rate is greater than 90%; arrange the eigenvectors from large to small according to the corresponding eigenvalues, and take the first z columns to form a matrix W, that is, an N×z-order matrix; among them, the contribution rate P of each principal component i for: Cumulative contribution rate S(P Z )for Step 34: Finally, calculate the data feature matrix after reducing to the z dimension, that is, an N×z-order matrix.

5. According to the data-driven method for detecting electricity theft by users according to claim 1, It is characterized in that The step 4 specifically includes: Step 41: The principal component data obtained in step 3 is used to form a data set, and k objects are selected from the data set as initial clustering centers; Step 42: Calculate the Euclidean distance between the remaining data points in the data set and each initial cluster center; Step 43: assigning the data point to the cluster represented by the nearest cluster center according to the Euclidean distance; Step 44: Calculate the average value of the objects in each cluster and update the cluster center; Steps 42 to 44 are repeated until each cluster no longer changes, and finally the cluster center sequence is obtained: C={C 1 ,C 2 ,...,C K } (12) In the formula, C i is the sequence of each cluster center.

6. According to the data-driven method for detecting electricity theft by users as described in claim 1, It is characterized in that The step 5 specifically includes: using the Pearson correlation coefficient to calculate the similarity ρ of the daily cluster center sequence C of each user, when ρ is lower than the similarity experience value η, that is, ρ<η, it is considered that the user's electricity consumption is abnormal.

Citation Information

Patent Citations

  • Power grid peak load clustering extraction method based on Pearson coefficient and MapReduce parallel computing

    CN110069467A

  • Transformer area user-transformer relation abnormity diagnosis method and device

    CN112699913A

  • Electricity stealing monitoring method based on user electricity consumption behavior

    CN116777124A

  • User electricity stealing detection method and system

    CN117076997A

  • Method and system for predicting a day-ahead wind power of wind farms

    US20230136352A1

Cited By

  • Zone area load component identification method based on image coding technology

    CN119693708A