CN114926303A - Electric larceny detection method based on transfer learning - Google Patents

Abstract

The invention discloses a transfer learning-based electricity stealing detection method, which comprises the following steps: acquiring a target domain test set, a source domain data set and a target domain training set; preprocessing a target domain test set, a source domain data set and a target domain training set; dividing a source domain data set into a training set and a test set; building a source domain and target domain neural network model; pre-training and evaluating a source domain neural network model; migrating parameters of a source domain neural network model to a target domain neural network model for initialization; training a target domain neural network model; and carrying out electricity stealing detection classification on the target domain test set. The method can accurately identify the electricity stealing users in a short time according to the electricity consumption data through the transfer learning technology under the condition that a small amount of electricity consumption data of the electricity stealing state label is contained, and the problem that a deep learning method depends on a large amount of label data with the electricity stealing state is solved; in addition, the date, week and month cycle data are used as the input of the neural network model, so that the performance of the electricity stealing detection model is improved.

Description

A method for detecting electricity theft based on transfer learning

technical field

The invention relates to the technical field of electricity stealing detection, in particular to a method for detecting electricity stealing based on migration learning.

Background technique

Electricity theft brings very serious consequences and harm to power companies, society and the country. Electricity theft not only causes huge economic losses in various countries every year, but also affects the stability of the power grid. With the development of the smart grid, the power system is gradually digitized, and a large amount of user power consumption data will be obtained, which provides a basis for the detection of electricity theft based on deep learning. However, a single smart meter does not have the function of electricity theft detection, and can only provide a large amount of electricity consumption data without electricity tampering status labels. Electrical status tags require huge labor costs. Therefore, the amount of electricity consumption data with electricity stealing status labels is very small, which makes it difficult for the deep learning electricity stealing detection method relying on data labels to be effectively trained in actual situations, unable to exert its detection performance, and unable to correctly detect electricity stealing users.

SUMMARY OF THE INVENTION

In order to overcome the problem of insufficient amount of training data caused by the above-mentioned prior art requiring that the electricity consumption data contain a large number of electricity stealing state labels, the present invention proposes a method for detecting electricity stealing based on migration learning.

The primary purpose of the present invention is to solve the above-mentioned technical problems, and the technical scheme of the present invention is as follows:

A method for detecting electricity theft based on migration learning, comprising the following steps:

S1: Obtain the sub-meter readings of all users in the detection target area, and construct the target domain test set D _target ; divide the target area into multiple sub-regions, detect the sub-region master meter readings, and determine All user sub-meter readings, power technical losses and error thresholds determine whether there is electricity stealing in the sub-region, and set the electricity stealing status label to construct the source domain data set D _source ; obtain a small number of historical users' electricity consumption data and electricity stealing status labels, Construct the target domain training set D _train ;

S2: Perform data preprocessing on the target domain test set, the source domain data set, and the target domain training set described in step S1, respectively, and convert all the electricity consumption data sequences into electricity consumption data with daily, weekly, and monthly time periods respectively. matrix, and then perform missing value recovery, data cleaning, and data normalization;

S3: Divide the source domain data set described in step S2 into a source domain training set and a source domain test set;

S4: Build a source domain and a target domain CNN neural network model, where the CNN neural network model includes three input layers, multiple convolution layers, and multiple fully connected layers;

S5: Use the source domain training set to pre-train the source domain CNN neural network model, and use the source domain test set to evaluate the source domain CNN neural network model;

S6: Save the source domain CNN neural network model parameters that are qualified in the evaluation of step S5, and migrate them to the target domain CNN neural network model to be trained, and initialize the target domain CNN neural network model parameters;

S7: Use the electricity consumption data of all users in the target domain training set and their corresponding electricity stealing state labels to train the initialized target domain CNN neural network model described in step S6;

S8: Input the electricity consumption data of all users in the target domain test set into the target domain CNN neural network model trained in step S7, classify user types, and find electricity stealing users in the target area.

In this solution, the target domain test set D _target , the source domain data set D _source and the target domain training set D _train described in step S1 are composed of the following steps:

S101: Record the readings of all user sub-meters in the target area as the electricity consumption data of the corresponding users, and construct the target domain test set D _target as follows:

Among them, d _m represents the power consumption data sequence of user m; d _m,n represents the data recorded by the nth sample of user m;

S102: Divide the target area into a plurality of sub-areas, detect the sub-area total meter reading, and use the sub-area electricity consumption data as follows:

Among them, d _{sub, g} represents the power consumption data sequence of sub-region g; d _{sub, g, n} represents the data recorded by the nth sampling of sub-region g;

S103: Add up the readings of the sub-meters of users in each sub-area according to the area, and obtain the total electricity consumption data of users in each sub-area as follows:

Among them, d _reg,g represents the total electricity consumption data sequence of users in sub-region g; d _{reg, g, n} represents the data recorded by the nth sample in sub-region g;

S104: Calculate the electric energy technical loss d _TL,g of the transmission line between the master meter of each sub-region and the user sub-meter; calculate the electricity stealing status label of each sub-region according to the following formula:

Wherein, y _reg,g represents the electricity stealing state label of sub-region g, y _{reg, g} =1 represents the electricity stealing state, y _{reg, g} =0 represents the normal state; α is the error threshold;

S105: Construct a source domain data set D _source by combining the user's total electricity consumption data sequence of each sub-area and its electricity stealing state label:

S106: According to the electricity consumption data of a small number of historical users and their electricity stealing state labels, construct the target domain training set D _train as follows:

Among them, d _his,k represents the electricity consumption data sequence of the historical user k; d _his,k,n represents the data of the nth sampling record of the historical user k; y _his,k represents the electricity stealing status label of the historical user k.

In this scheme, the data preprocessing described in step S2, its specific composition steps are as follows:

S201: For the target domain test set, the source domain data set, and the target domain training set described in step S1, convert all the power consumption data sequences into a power consumption data matrix with day, week, and month as time periods, respectively. Each row of the electricity consumption data matrix represents electricity consumption data in a single time period. The daily period data matrix D _day , the weekly period data matrix D _week , and the monthly period data matrix D _month are constructed as follows:

Among them, o represents the number of days for meter sampling; p represents the number of weeks for meter sampling; q represents the number of months for meter sampling;

S202: Perform missing value recovery, data cleaning, and data normalization processing on the daily period data matrix D _day , the weekly period data matrix D _week , and the monthly period data matrix D _month .

In this solution, the construction of the source domain and target domain CNN neural network model described in step S4, its specific composition steps are as follows:

S401: The CNN neural network model has three input layers, including daily cycle data input X _day , weekly cycle data input X _week and monthly cycle data input X _month , whose sizes are o×(24×60/n), p×(24×60×7/n) and q×(24×60×30/n);

S402: Use multiple convolution layers to perform convolution operations on three sets of input data:

S ₁ =f(WX _day +b)

S ₂ =f(WX _week +b)

S ₃ =f(WX _month +b)

Among them, S ₁ , S ₂ and S ₃ respectively input three sets of feature outputs after multiple convolutional layers; W and b represent the weight and bias of the convolutional layer, respectively; f(.) represents the activation function of the convolutional layer ;

S403: Use a fusion layer to fuse S ₁ , S ₂ and S ₃ to obtain S ₄ ;

S404: Use multiple convolutional layers and fully connected layers to perform feature extraction _on S4, and output the type of detection target:

S ₅ =f(WS ₄ +b)

y=g(VS ₅ +c)

Among them, S5 is the feature output of multiple convolutional layers after the fusion layer; V and _c represent the weight and bias of the fully connected layer, respectively; g(.) represents the activation function of the fully connected layer; y represents the detection target category, y=1 means stealing electricity, y=0 means normal.

In this solution, in step S5, the source domain training set is used to pre-train the source domain CNN neural network model, and the source domain test set is used to evaluate the source domain CNN neural network model. The specific composition steps are as follows:

S501: Use the daily period data matrix D _day , the weekly period data matrix D _week , and the monthly period data matrix D _month of each subregion in the source domain training set and the source domain test set as the daily data matrix of the source domain CNN neural network model described in step S4, respectively. Cycle data input X _day , weekly cycle data input X _week and monthly cycle data input X _month ;

S502: Pre-train and evaluate the CNN neural network model described in step S4 by using the electricity consumption data of the source domain training set and the source domain test set and the electricity stealing state label.

In this solution, initializing the parameters of the target domain CNN neural network model described in step S6 is specifically: migrating the weights and biases of the source domain CNN neural network model described in step S5 to the target domain CNN neural network model as its Initialization values for weights and biases.

In this solution, step S7 trains the initialized target domain CNN neural network model, specifically: the daily cycle data matrix D _day , the weekly cycle data matrix D _week , and the monthly cycle data matrix D of a small number of historical users in the target domain training set _month is respectively used as the daily period data input X _day , the weekly period data input X _week and the monthly period data input X _month of the target domain CNN neural network model in step S6 .

Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

The transfer learning technology and CNN neural network model are used to detect users in the target area, which overcomes the problem that the existing deep learning electricity stealing detection method relies on a large amount of data containing electricity stealing status labels, and only needs a small amount of electricity consumption data containing electricity stealing status labels. In addition, only a small amount of source domain data and target domain training data are used to train the CNN neural network model, which greatly reduces the training time of the deep neural network for electricity stealing detection, and can accurately identify Electricity theft users. The daily, weekly and monthly cycle data are used as the input of the CNN neural network model, and the daily, weekly and monthly cycle features are extracted respectively, so that the extracted features can more accurately describe the electricity consumption behavior, thereby improving the detection performance of the CNN neural network model. .

Description of drawings

1 is a flowchart of a method for detecting electricity theft based on migration learning proposed by the present invention;

FIG. 2 is a CNN neural network model according to an embodiment of the present invention.

Detailed ways

In order to understand the above objects, features and advantages of the present invention more clearly, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments may be combined with each other in the case of no conflict.

Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific details disclosed below. Example limitations.

In a specific embodiment, as shown in FIG. 1 , a method for detecting electricity theft based on transfer learning includes the following steps:

S1: Obtain the sub-meter readings of all users in the detection target area, and construct the target domain test set D _target ; divide the target area into multiple sub-regions, detect the sub-region master meter readings, and determine All user sub-meter readings, power technical losses and error thresholds determine whether there is electricity stealing in the sub-region, and set the electricity stealing status label to construct the source domain data set D _source ; obtain a small number of historical users' electricity consumption data and electricity stealing status labels, Construct the target domain training set D _train ;

S2: Perform data preprocessing on the target domain test set, the source domain data set, and the target domain training set described in step S1, respectively, and convert all the electricity consumption data sequences into electricity consumption data with daily, weekly, and monthly time periods respectively. matrix, and then perform missing value recovery, data cleaning, and data normalization;

S3: Divide the source domain data set described in step S2 into a source domain training set and a source domain test set;

S4: Build source domain and target domain CNN neural network models, the CNN neural network model includes three input layers, multiple convolution layers and multiple fully connected layers;

S5: Use the source domain training set to pre-train the source domain CNN neural network model, and use the source domain test set to evaluate the source domain CNN neural network model;

S6: Save the source domain CNN neural network model parameters that are qualified in the evaluation of step S5, and migrate them to the target domain CNN neural network model to be trained, and initialize the target domain CNN neural network model parameters;

S7: Use the electricity consumption data of all users in the target domain training set and their corresponding electricity stealing state labels to train the initialized target domain CNN neural network model described in step S6;

S8: Input the electricity consumption data of all users in the target domain test set into the target domain CNN neural network model trained in step S7, classify user types, and find electricity stealing users in the target area.

In this solution, the target domain test set D _target , the source domain data set D _source and the target domain training set D _train described in step S1 are composed of the following steps:

S101: Record the readings of all user sub-meters in the target area as the electricity consumption data of the corresponding users, and construct the target domain test set D _target as follows:

Among them, d _m represents the power consumption data sequence of user m; d _m,n represents the data recorded by the nth sample of user m;

S102: Divide the target area into a plurality of sub-areas, detect the sub-area total meter reading, and use the sub-area electricity consumption data as follows:

Among them, d _{sub, g} represents the power consumption data sequence of sub-region g; d _{sub, g, n} represents the data recorded by the nth sampling of sub-region g;

S103: Add up the readings of user sub-meters in each sub-area according to the area, and obtain the total electricity consumption data of users in each sub-area as follows:

Among them, d _reg,g represents the total electricity consumption data sequence of users in sub-region g; d _{reg, g, n} represents the data recorded by the nth sample in sub-region g;

S104: Calculate the electric energy technical loss d _TL,g of the transmission line between the master meter of each sub-region and the user sub-meter; calculate the electricity stealing status label of each sub-region according to the following formula:

Wherein, y _reg,g represents the electricity stealing state label of sub-region g, y _{reg, g} =1 represents the electricity stealing state, y _{reg, g} =0 represents the normal state; α is the error threshold;

S105: Construct a source domain data set D _source by combining the user's total electricity consumption data sequence of each sub-area and its electricity stealing state label:

S106: According to the electricity consumption data of a small number of historical users and their electricity stealing state labels, construct the target domain training set D _train as follows:

Among them, d _his,k represents the electricity consumption data sequence of the historical user k; d _his,k,n represents the data of the nth sampling record of the historical user k; y _his,k represents the electricity stealing status label of the historical user k.

In this scheme, the data preprocessing described in step S2, its specific composition steps are as follows:

S201: For the target domain test set, the source domain data set, and the target domain training set described in step S1, convert all the power consumption data sequences into a power consumption data matrix with day, week, and month as time periods, respectively. Each row of the electricity consumption data matrix represents electricity consumption data in a single time period. The daily period data matrix D _day , the weekly period data matrix D _week , and the monthly period data matrix D _month are constructed as follows:

Among them, o represents the number of days for meter sampling; p represents the number of weeks for meter sampling; q represents the number of months for meter sampling;

S202: Perform missing value recovery, data cleaning, and data normalization processing on the daily period data matrix D _day , the weekly period data matrix D _week , and the monthly period data matrix D _month .

In this solution, the construction of the source domain and target domain CNN neural network models described in step S4 is shown in Figure 2, and its specific composition steps are as follows:

S401: The CNN neural network model has three input layers, including daily cycle data input X _day , weekly cycle data input X _week and monthly cycle data input X _month , whose sizes are o×(24×60/n), p×(24×60×7/n) and q×(24×60×30/n);

S402: Use multiple convolution layers to perform convolution operations on three sets of input data:

S ₁ =f(WX _day +b)

S ₂ =f(WX _week +b)

S ₃ =f(WX _month +b)

Among them, S ₁ , S ₂ and S ₃ respectively input three sets of feature outputs after multiple convolutional layers; W and b represent the weight and bias of the convolutional layer, respectively; f(.) represents the activation function of the convolutional layer ;

S403: Use a fusion layer to fuse S ₁ , S ₂ and S ₃ to obtain S ₄ ;

S404: Use multiple convolutional layers and fully connected layers to perform feature extraction _on S4, and output the type of detection target:

S ₅ =f(WS ₄ +b)

y=g(VS ₅ +c)

Among them, S5 is the feature output of multiple convolutional layers after the fusion layer; V and _c represent the weight and bias of the fully connected layer, respectively; g(.) represents the activation function of the fully connected layer; y represents the detection target category, y=1 means stealing electricity, y=0 means normal.

In this solution, in step S5, the source domain training set is used to pre-train the source domain CNN neural network model, and the source domain test set is used to evaluate the source domain CNN neural network model. The specific composition steps are as follows:

S501: Use the daily period data matrix D _day , the weekly period data matrix D _week , and the monthly period data matrix D _month of each subregion in the source domain training set and the source domain test set as the daily data matrix of the source domain CNN neural network model described in step S4, respectively. Cycle data input X _day , weekly cycle data input X _week and monthly cycle data input X _month ;

S502: Pre-train and evaluate the CNN neural network model described in step S4 by using the electricity consumption data of the source domain training set and the source domain test set and the electricity stealing state label.

In this solution, initializing the parameters of the target domain CNN neural network model described in step S6 is specifically: migrating the weights and biases of the source domain CNN neural network model described in step S5 to the target domain CNN neural network model as its Initialization values for weights and biases.

In this solution, step S7 trains the initialized target domain CNN neural network model, specifically: the daily cycle data matrix D _day , the weekly cycle data matrix D _week , and the monthly cycle data matrix D of a small number of historical users in the target domain training set _month is respectively used as the daily period data input X _day , the weekly period data input X _week and the monthly period data input X _month of the target domain CNN neural network model in step S6 .

Claims (7)

1. A power stealing detection method based on transfer learning is characterized by comprising the following steps:

s1: obtaining sub-table readings of all users in a detection target area, and constructing a target area test set D _target (ii) a Dividing a target area into a plurality of subareas, detecting reading of a subarea master meter, judging whether the subareas have electricity stealing according to the reading of the subarea master meter, the reading of all user sub-meters in the subareas, electric energy technical loss and an error threshold value, setting an electricity stealing state label, and constructing a source area data set D _source (ii) a Acquiring power consumption data of a small number of historical users and power stealing state labels thereof, and constructing a target domain training set D _train ；

S2: respectively carrying out data preprocessing on the target domain test set, the source domain data set and the target domain training set in the step S1, converting all power utilization data sequences into power utilization data matrixes respectively with time periods of day, week and month, and then carrying out missing value recovery, data cleaning and data normalization;

s3: dividing the source domain data set in the step S2 into a source domain training set and a source domain test set;

s4: building a source domain and a target domain CNN neural network model, wherein the CNN neural network model comprises three input layers, a fusion layer, a plurality of convolution layers and a plurality of full connection layers;

s5: pre-training the source domain CNN neural network model by using a source domain training set, and evaluating the source domain CNN neural network model by using a source domain testing set;

s6: saving the qualified source domain CNN neural network model parameters evaluated in the step S5, transferring the qualified source domain CNN neural network model parameters to a target domain CNN neural network model to be trained, and initializing the target domain CNN neural network model parameters;

s7: training the initialized target domain CNN neural network model in the step S6 by adopting the power consumption data of all users in the target domain training set and corresponding power stealing state labels;

s8: and inputting the electricity utilization data of all users in the target domain test set into the target domain CNN neural network model trained in the step S7, classifying the user types, and searching for electricity stealing users in the target region.

2. The method for detecting electricity stealing according to claim 1, wherein the target domain test set D of step S1 is _target Source domain data set D _source And a target domain training set D _train The method comprises the following specific steps:

s101: recording the readings of all the user sub-tables in the target area as the electricity utilization data of the corresponding users, and constructing a target area test set D _target The following were used:

wherein d is _m A power usage data sequence representing a user m; d is a radical of _m,n Data representing the nth sample record of user m;

s102: dividing a target area into a plurality of subareas, and detecting reading of a subarea summary table as subarea electricity data as follows:

wherein d is _sub,g Representing a power consumption data sequence of the subarea g; d _sub,g,n Data representing the nth sampling record of the partition g;

s103: adding the reading of the sub-user tables in each subarea according to the area to obtain the total power consumption data of the users in each subarea as follows:

wherein d is _reg,g Representing a user total electricity data sequence of the subarea g; d is a radical of _reg,g,n Data representing the nth sampling record of the subarea g;

s104: calculating the electric energy technical loss d of the transmission line from each sub-regional general meter to the sub-meters of the user _TL,g (ii) a Calculating the electricity stealing state labels of the sub-areas according to the following formula:

wherein, y _reg,g Label indicating electricity stealing status of zone g, y _reg,g 1 indicates an electricity stealing status, y _reg,g 0 indicates a normal state; alpha is an error threshold value;

s105: the user total electricity data sequences of all the subareas and the electricity stealing state labels thereof are combined to construct a source area data set D _source ：

S106: according to the electricity consumption data of a small number of historical users and the electricity stealing state labels thereof, a target domain training set D is constructed _train The following were used:

wherein d is _his,k A power consumption data sequence representing a historical user k; d is a radical of _his,k,n Data representing the nth sampling record of the historical user k; y is _his,k A power stealing status tag that represents historical user k.

3. The method for detecting electricity stealing based on transfer learning of claim 1, wherein the data preprocessing of step S2 comprises the following steps:

s201: aiming at the target domain test set, the source domain data set and the target domain training set in the step S1, converting all the electricity utilization data sequences into electricity utilization data matrixes by taking days, weeks and months as time periods respectively, wherein each row of the electricity utilization data matrixes represents electricity utilization data in a single time period, and constructing a day period data matrix D _day A periodic data matrix D _week Monthly cycle data matrix D _month The following were used:

wherein o represents the number of days of electric meter sampling; p represents the number of weeks of sampling of the electric meter; q represents the number of months of the electric meter sampling;

s202: will day period data matrix D _day Periodic data matrix D _week Monthly cycle data matrix D _month Carrying out missing value recovery, data cleaning,And (6) data normalization processing.

4. The method for detecting electricity stealing based on transfer learning according to claim 1, wherein the step S4 of building a source domain and target domain CNN neural network model specifically comprises the following steps:

s401: the CNN neural network model has three input layers including daily period data input X _day Cycle data input X _week Date entry X of the moon cycle _month The sizes thereof are o × (24 × 60/n), p × (24 × 60 × 7/n) and q × (24 × 60 × 30/n), respectively;

s402: performing convolution operations on three sets of input data using a plurality of convolution layers, respectively:

S ₁ ＝f(WX _day +b)

S ₂ ＝f(WX _week +b)

S ₃ ＝f(WX _month +b)

wherein S is ₁ 、S ₂ And S ₃ Respectively inputting the characteristic output after passing through a plurality of convolution layers by three groups; w and b represent the weight and bias of the convolutional layer, respectively; f (.) represents the activation function of the convolutional layer;

s403: using a fused layer pair S ₁ 、S ₂ And S ₃ Carrying out fusion to obtain S ₄ ；

S404: using multiple convolutional and fully-connected layer pairs S ₄ And (3) performing feature extraction, and outputting the type of the detection target:

S ₅ ＝f(WS ₄ +b)

y＝g(VS ₅ +c)

wherein S is ₅ Outputting the characteristics of the plurality of convolution layers after the layers are fused; v and c represent the weight and bias of the fully connected layer, respectively; g (.) denotes the activation function of the fully connected layer; y represents the type of the detection target, y 1 represents power stealing, and y 0 represents normal.

5. The method according to claim 1, wherein the step S5 of pre-training the CNN neural network model using the source domain training set and evaluating the CNN neural network model using the source domain test set comprises the following steps:

s501: collecting the source area training set and the source area test set and collecting the daily period data matrix D of each subarea _day A periodic data matrix D _week Monthly cycle data matrix D _month Respectively as daily cycle data input X of the source domain CNN neural network model in step S4 _day Cycle data input X _week Date entry X of the moon cycle _month ；

S502: and pre-training and evaluating the CNN neural network model in the step S4 by using the power utilization data of the source domain training set and the source domain testing set and the power stealing state labels thereof.

6. The method according to claim 1, wherein the initializing of the target domain CNN neural network model parameters in step S6 is specifically: and migrating the weights and the offsets of the source domain CNN neural network model obtained in the step S5 to the target domain CNN neural network model as initialization values of the weights and the offsets.

7. The method for detecting electricity stealing according to claim 1, wherein step S7 trains the initialized target domain CNN neural network model, specifically: a daily period data matrix D of a small number of historical users in a target domain training set _day A periodic data matrix D _week Monthly cycle data matrix D _month Respectively as daily cycle data input X of the target domain CNN neural network model in step S6 _day Cycle data input X _week Date of Suyue period input X _month 。

Images