A CNN-DBSCAN method for detecting intermittent electricity theft in high-loss areas based on correlation analysis

Abstract

The present invention belongs to the technical field of power system automation, and discloses a CNN-DBSCAN high-loss area intermittent electricity theft detection method based on correlation analysis. The multi-scale CNN model is used to preliminarily select the period in which the maximum correlation analysis result may exist between the power consumption data of all users to be tested and the line loss data of the area. After obtaining the time period in which all users to be tested are most likely to steal electricity, the time period is compared with the time period in which the line loss rate data actually has anomalies, and the Pearson correlation analysis result under time distance weighting is obtained; the correlation analysis results of all users are clustered by the DBSCAN clustering algorithm, and the outliers therein are identified as electricity theft users. The method described in the present invention can effectively screen intermittent electricity thieves whose electricity theft behavior is more concealed and whose electricity theft characteristics are weaker, and can stably and quickly detect electricity theft, thereby ensuring the safe and stable operation of the power system.

Description

Technical Field

[0001] The invention belongs to the technical field of power system automation, and in particular relates to a CNN-DBSCAN high-loss station area intermittent electricity theft detection method based on correlation analysis. Background Art

[0002] Line loss is an important indicator for evaluating the efficiency of power systems. Excessive line loss may lead to reduced stability of power systems. Long-term abnormal line loss will shorten the service life of equipment and reduce the quality of electric energy, further reducing the satisfaction of electricity users. In addition, excessive line loss usually represents a certain degree of economic loss for power companies. Therefore, the ability to regulate line loss indicators is one of the capabilities that smart grid systems must have. The power system realizes the entire process from power generation to power consumption through a hierarchical architecture of generation, transmission, transformation, distribution, and use. Each of the above stages is accompanied by the generation of line loss, among which the line loss generated in the power distribution stage accounts for 65% of the total line loss of the power system. In actual systems, line loss is usually expressed as the sum of technical loss and non-technical loss, but in low-voltage power distribution systems, non-technical losses caused by power theft are the main cause of abnormal line loss. Therefore, detecting illegal power theft users in low-voltage distribution stations with abnormal line loss indicators is one of the most effective ways to try to reduce power system line loss and safeguard the interests of power companies under the condition of limited audit resources.

[0003] Compared with other links in the power system, the power distribution system is directly facing all industrial, commercial and residential electricity users, so it is at the greatest risk of being harmed by electricity theft. Existing methods of electricity theft include the undervoltage method, undercurrent method and phase shift method of intrusive destruction of electronic energy meters, as well as non-intrusive bypassing, interference and tampering of smart meter measurement, wire wrapping, strong magnetic interference and network attacks. With the continuous improvement of information and communication technology, the power distribution system has built an electricity information collection system architecture based on smart meters. The area monitoring data and user electricity data recorded and collected by this architecture provide strong support for the analysis and identification of electricity theft users. Most of the existing electricity theft detection methods are based only on the long-term periodic characteristics of the electricity consumption data of the electricity theft users for electricity theft detection, which is easily affected by the randomness of the electricity consumption behavior of the users, resulting in a certain false alarm rate of the electricity theft detection method; and when the electricity theft detection is only based on the periodicity of the electricity consumption data of the electricity theft users, the physical meaning of the electricity theft criterion is not clear enough, the method has low interpretability, and the detection effect of intermittent electricity theft is poor. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a CNN-DBSCAN high-loss substation intermittent electricity theft detection method based on correlation analysis. By comparing the long-term and short-term characteristics of the joint data, the intermittent electricity theft users in the distribution substation area can be effectively identified, thereby further improving the accuracy of the electricity theft detection method and effectively reducing the false alarm rate of the electricity theft detection method, providing a solution to the problem of electricity theft detection with relatively hidden electricity theft characteristics.

[0005] The present invention discloses a CNN-DBSCAN high-loss area intermittent electricity theft detection method based on correlation analysis, comprising the following steps:

[0006] Step 1: pre-process the power consumption data of each power user in the high-loss distribution area and the line loss data of the distribution area within the detection period, and divide them into long-term and short-term joint data of different resolutions according to short-term and daily time intervals, build and train a multi-scale CNN model, and obtain the maximum period of correlation between the power consumption data of each power user and the line loss data of the distribution area based on the long-term and short-term joint data, that is, the period with the greatest suspicion of power theft;

[0007] Step 2: According to the actual abnormal time of the line loss data, the maximum period of suspected electricity theft for each user output by the multi-scale CNN model is compared to calculate the time distance offset weight; then, the Pearson correlation analysis algorithm is used to obtain the suspected electricity theft evaluation index size of each user during the abnormal period of the line loss data;

[0008] Step 3: Use the DBSCAN algorithm to perform cluster analysis on the suspected electricity theft evaluation index values ​​of all electricity users in the high-loss distribution station area, identify the clustered outliers as electricity thieves in the station area, and assign staff to conduct on-site inspections of the electricity theft behaviors of the electricity thieves.

[0009] Furthermore, in step 1, the multi-scale CNN model includes a short-term feature extraction module, a long-term feature extraction module and a long-term and short-term feature comparison module; the short-term feature extraction module and the long-term feature extraction module are connected with the long-term and short-term feature comparison module through a connection layer;

[0010] The short-term feature extraction module consists of 5 convolutional layers, which are used to extract the short-term correlation features between the user power consumption data and the area line loss data;

[0011] The long-term feature extraction module consists of three convolutional layers, which are used to extract the daily correlation features between user power consumption data and station area line loss data;

[0012] The long-term and short-term feature comparison module is composed of 4 fully connected layers, which are used to compare the differences between long-term and short-term features, and further determine the time period when there is a strong correlation between user power consumption data and station line loss data.

[0013] Furthermore, the input data of the short-term feature extraction module is the user power consumption data at the short-term measurement frequency corresponding to each week in a month. And short-term line loss in the area , ; The input data of the long-term feature extraction module is the user power consumption data D and the daily line loss S of the substation under the daily measurement frequency in months; Among them, short-term refers to the highest collection frequency of the power consumption information collection system, which may be 1 hour, 30 minutes, 15 minutes, 5 minutes, etc., which can be adjusted according to the specific application; n refers to dividing the short-term data within a month into 4 weeks and sending them to the short-term feature extraction module in units of weeks;

[0014] Short-term line loss within the distribution station area The general calculation method of daily line loss S is as follows:

[0015] ,

[0016] in, The total power supply of the distribution area in the time period corresponding to different measurement resolutions; It is the sum of the power consumption data recorded by the smart meters of each user in the distribution area during the time periods corresponding to different measurement resolutions, and it is the total power sales in the distribution area; is the line loss data of the substation in the time period corresponding to different measurement resolutions, including the technical loss caused by the loss of the transmission line itself and the non-technical loss caused by the theft of electricity; t represents the time interval for calculation under different measurement frequencies; when When the short-term line loss is calculated ,when When , the daily line loss S is calculated;

[0017] The multi-scale CNN model uses the long-term feature extraction module and the short-term feature extraction module to identify the long-term and short-term correlations between each user's power consumption data and the line loss data in the substation area. Finally, the long-term and short-term feature comparison module maps the difference between local features and global features. The difference is used to output the probability of the period of maximum correlation between each user's high-resolution power consumption data and the substation line loss data every week within each month. , and according to The maximum value of the correlation is obtained from the time period with the maximum correlation. .

[0018] Furthermore, in step 1, the loss function of the multi-scale CNN model during training is as follows:

[0019] ,

[0020] in, Indicates the probability that the weekly power consumption data and the area line loss data have the greatest correlation, represents the true label.

[0021] Furthermore, in step 2, the abnormal time period T of the line loss data is determined, specifically:

[0022] Taking line loss rate as the monitoring indicator, the calculation is as follows:

[0023] ,

[0024] in, is the specific acquisition frequency corresponding to the short-time measurement resolution; The total power supply of the distribution area within the time period corresponding to the short-term measurement resolution; It is the line loss data of the substation in the time period corresponding to the short-term measurement resolution, including the technical loss part TL caused by the loss of the transmission line itself and the non-technical loss NTL caused by the electricity theft behavior; It is the line loss rate of the distribution station area in the time period corresponding to the short-term measurement resolution, in units of %;

[0025] If there are no electricity theft users in the distribution station area, the NTL value is 0 under ideal conditions; if there are cut-point users in the distribution station area, the NTL increases, causing the line loss rate of the station area to exceed the normal range; suspected electricity theft users in the high-loss station area are located, and the abnormal time period T when the line loss rate exceeds the normal range is monitored.

[0026] Furthermore, in step 2, the time distance weight is calculated as follows:

[0027] ,

[0028] in, is the time distance weight; and They are the start and end time nodes of the abnormal line loss rate period in the substation area; and are the start and end time nodes of the time period with the maximum correlation for each user to be tested output by the multi-scale CNN model. It can be seen from formula (4) that the time distance weight is 1 only when the time period with the maximum correlation of the user completely overlaps with the abnormal time period of the line loss rate of the station area, and the weight values ​​at other times are all less than 1.

[0029] Furthermore, in step 2, the offset between the maximum correlation time period output by the multi-scale CNN model and the abnormal line loss rate time period of the substation is first calculated by the size of the time distance weighted value, so as to obtain the maximum result that can be obtained when the correlation analysis is performed between the user power consumption data and the substation line loss data under the premise of abnormal line loss rate; the Pearson correlation analysis algorithm based on time distance weighting is shown as follows:

[0030] ,

[0031] in, It is the Pearson correlation analysis result of the user power consumption data and the area line loss data in the time period with the maximum correlation output by the model under the time distance weighting, that is, the maximum correlation analysis result that can be obtained for each user to be tested in the time period with abnormal area line loss rate; Indicates the time period with the maximum correlation for each user under test output by the multi-scale CNN model Line loss data of the substation area within Indicates the maximum correlation time period for each user under test output by the multi-scale CNN model User power consumption data within.

[0032] Furthermore, in step 3, the DBSCAN algorithm is used to perform cluster analysis on the suspected electricity theft evaluation index values ​​of all electricity users in the high-loss distribution station area, specifically:

[0033] Step 3-1: The time-distance weighted correlation analysis results of all users to be tested are used as cluster samples. Based on the characteristics of the samples and the expected density of clusters, the minimum number of samples MinPts required to form a cluster is selected;

[0034] Step 3-2: The time-distance weighted correlation analysis results of all users to be tested are used as cluster samples to obtain the neighborhood radius Eps, so as to accurately identify outlier electricity theft users. Specifically:

[0035] a) Calculate the Euclidean distance between each sample point and other sample points to obtain the distance matrix between all samples;

[0036] b) For each sample point, select the distance to its kth nearest neighbor as the parameter of all sample points, and plot it according to the order of the sample points to obtain the k-distance curve;

[0037] c) In the k-distance graph, determine the inflection point of the curve and set the distance corresponding to the inflection point as the neighborhood radius Eps.

[0038] The beneficial effects described in the present invention are as follows: based on the power consumption data of each user in the area recorded by the distribution station control center and the line loss data of the substation area, the present invention effectively evaluates the time period with the maximum correlation between the power consumption data of each power user in the distribution station area and the line loss data of the substation area through the constructed multi-scale CNN model, and combines the proposed time distance weight, and obtains the maximum correlation analysis result that can be obtained by each user to be tested in the abnormal time period of the substation line loss rate according to the time distance weighted Pearson correlation analysis. By using the DBSCAN algorithm to perform cluster analysis on the Pearson correlation analysis results of all users to be tested after time distance weighting, the key feature differences between positive and abnormal users are highlighted, thereby achieving effective detection of electricity thieves. When the intermittent electricity thieves have no obvious characteristics of electricity theft, the method of the present invention effectively investigates the electricity theft behavior through the correlation between the power consumption data of the electricity thieves and the line loss data of the substation area, and can provide an accurate and fast electricity theft detection method when the system is monitored, further ensuring the safe and stable operation of the power system. Compared with other electricity theft detection methods, the electricity theft criterion used in the method of the present invention is more reliable and can more effectively identify intermittent electricity theft users. Therefore, the proposed method has higher accuracy and lower false alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flow chart of the method of the present invention;

[0040] Figure 2 is an application flow chart of the method of the present invention;

[0041] Figure 3 Flowchart for multi-scale CNN model training;

[0042] Figure 4 Apply a flowchart for the multi-scale CNN model;

[0043] Figure 5 It is a flow chart for cluster analysis application based on DBSCAN algorithm;

[0044] Figure 6 This is a schematic diagram of the changes in line loss rate in each substation within one month;

[0045] Figure 7 This is a clustering effect diagram based on the correlation analysis results under time distance weighting. DETAILED DESCRIPTION

[0046] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0047] like Figure 1 and Figure 2As shown, the CNN-DBSCAN high-loss area intermittent electricity theft detection method based on correlation analysis described in the present invention includes the following steps:

[0048] Step 1: pre-process the power consumption data of each power user in the high-loss distribution area and the line loss data of the distribution area within the detection period, and divide them into long-term and short-term joint data with different resolutions according to time intervals of 15 minutes and 1 day, respectively. Use the multi-scale CNN model to obtain the maximum period of correlation between the power consumption data of each power user and the line loss data of the distribution area based on the above data, that is, the period with the greatest suspicion of power theft;

[0049] Step 2: According to the actual abnormal time of the line loss data, the maximum period of suspected electricity theft for each user output by the multi-scale CNN model is compared to calculate the time distance offset weight; then, the Pearson correlation analysis algorithm is used to obtain the suspected electricity theft evaluation index size of each user during the abnormal period of the line loss data;

[0050] Step 3: Use the DBSCAN algorithm to perform cluster analysis on the suspected electricity theft evaluation index values ​​of all electricity users in the high-loss distribution station area, and determine the outliers after clustering as electricity thieves in the station area; and assign staff to conduct on-site inspections of the electricity theft behaviors of electricity thieves.

[0051] The historical line loss data of the distribution station area and the power consumption data of the power users required for the power theft detection in the present invention are provided by the control center of the distribution station area; the multi-scale CNN model is trained accordingly according to the simulation data or the real data.

[0052] In step 1, the multi-scale CNN model is constructed, and its training and application include Figure 3 and Figure 4 After the above processing, all the steps shown in the flowchart can be used to determine the period when all electricity users in the distribution station area are most likely to steal electricity based on the user's power consumption data and the line loss data of the station area. This can effectively identify the period when all electricity users are most suspected of stealing electricity during the monitoring period, providing an effective solution to ensure the effective execution of electricity theft detection when intermittent electricity theft occurs.

[0053] In step 3 of the present invention, the DBSCAN clustering algorithm is used to perform cluster analysis on the suspected electricity theft evaluation index values ​​of each user in the abnormal time period of the line loss index obtained by the Pearson correlation analysis algorithm under the time distance offset weight, such as Figure 5 As shown, after the above processing, outlier electricity thieves can be found among all electricity users, so that even in the case of intermittent electricity thieves with relatively hidden electricity theft identification characteristics, electricity theft behavior can still be effectively detected.

[0054] In order to verify the effect of the detection method proposed in the present invention, the present invention simulates high-loss distribution substations under different levels of electricity theft based on the real power consumption data in the UCI power load data set. The UCI power load data set contains the real power consumption data of 370 normal users from 2011 to 2014. The time interval of each user's power consumption data recording is 15 minutes and the unit is kwh. The present invention uses the above real data to construct the following two scenarios, which are used to generate artificial electricity theft samples containing line loss information under different situations, so as to train the multi-scale correlation maximum period selection model based on CNN and verify the detection effect of the proposed method in the intermittent electricity theft scenario.

[0055] Scenario (1): The present invention divides the electricity consumption data of 370 normal users in the UCI power load data set into 28 days a month, and simulates the power supply process of the distribution station area under different conditions with a monthly cycle. The technical loss of the substation in each simulation cycle is set to 4% of the total power supply of the substation, and the non-technical loss part is caused by randomly selected power thieves in the substation. In order to further reduce the adverse effects of sample imbalance on the model detection effect, the number of power thieves in each cycle is set to 50% of the total number of users in the current substation. Each power thief tampers the original power consumption data according to the typical power theft method. The specific power theft parameters of each cycle are uniformly selected according to the final line loss rate of the substation of 8% to 20%. In addition, the power theft duration of the power thief in each cycle is set to any 3 to 14 days in 28 days to simulate the randomness of intermittent power theft behavior.

[0056] Scenario (2): Since the proportion of electricity thieves in the actual system is extremely small, and the proportion of electricity thieves in the real data set is only 8.63% of the total number of users, the present invention sets the proportion of electricity thieves to 10% of the total number of users in the simulated substation in scenario (2), thereby simulating the actual electricity consumption of the distribution substation in the actual system. 600 segments of 28 days of data that do not contain abnormal values ​​and are not selected as electricity thieves in scenario (1) are randomly selected as electricity users that simulate the actual electricity consumption of the distribution substation, and three simulated distribution substations with 200 users are constructed. The method for constructing electricity theft samples in each simulated substation is the same as that in scenario (1), but the extreme values ​​of the line loss rate caused by electricity thieves are three specified values ​​with a larger span, thereby simulating high-loss substations under different loss degrees in actual situations, so as to facilitate the next step of applying the proposed method to perform intermittent electricity theft detection. The detailed settings are shown in Table 1, and the line loss rate fluctuations in each simulated substation are shown in Table 1. Figure 6 shown.

[0057] Table 1 Distribution of positive and abnormal users and line loss rate in each simulated area

[0058]

[0059] In order to verify the superiority and detection effect of the multi-scale CNN model, the present invention divides the artificial positive anomaly samples containing the line loss information of the substation area generated in scenario (1) into a training set and a test set at a ratio of 8:2, and simultaneously selects the single short-term CNN model in the proposed method of the logistic regression model (LR), random forest (RF), support vector machine (SVM), and the single long-term CNN model in the proposed method as comparison items for comparative experiments. The input of the single model is the single-scale data within the detection period, while the input of all multi-scale models is a combination of long-term and short-term data. After the training obtains the optimal parameters, the selection effect of the maximum correlation period of each model is shown in Table 2.

[0060] Table 2 Effect of selecting the maximum correlation period for each model

[0061]

[0062] According to the evaluation index results of each model in Table 2, the model built based on single-cycle data cannot accurately capture the difference between the local and global features of users when they steal electricity intermittently, so the detection effect is poor. Compared with the long-term CNN model, the short-term CNN recognizes a finer cycle, and the learned local features can play a more critical role in selecting the period with the maximum correlation. Therefore, the detection effect of the short-term CNN is better between the two. However, due to the lack of global features in the input data, the short-term CNN model still cannot be compared and analyzed through the differences between data of different scales. Therefore, compared with the multi-scale models, the detection accuracy is still low. Among the four multi-scale correlation maximum period selection models based on different algorithms (LR, RF, SVM, CNN), the multi-scale CNN achieves effective extraction of high-dimensional data features with the help of a 2D convolution module. Under the comparison of global and local features, it can most effectively select the maximum period of correlation between the power consumption data of each measured user and the line loss in the station area.

[0063] The method proposed in the present invention is used to detect electricity theft users in the high-loss simulated substation in the above scenario (2). The DBSCAN clustering algorithm is used to perform cluster analysis on the correlation analysis results between the power consumption data of all simulated users in each simulated substation and the line loss of the corresponding substation. The outliers after clustering can be judged as electricity theft users. The classification results of positive and abnormal users after clustering all data in the three substations are as follows: Figure 7 shown.

[0064] From the clustering results, it can be seen that after the multi-scale CNN model blindly selects the maximum period of correlation with the line loss data of each area, the power consumption data of all users in the three areas are compared with the period of abnormal line loss rate in each area, and then the Pearson correlation coefficient after time distance weighting is calculated, and finally the maximum correlation analysis result that can be obtained in the abnormal line loss rate period of each area is obtained. Since the non-technical losses in each area are directly determined by the amount of electricity stolen by all electricity thieves, there is a strong correlation between the power consumption data of the electricity thieves and the line loss data of the area. The correlation analysis results of all outlier electricity thieves in each area range from [0.68, 0.93]. On the other hand, although in theory there is no strong correlation between the power consumption data of normal users in each area and the line loss of the area, the randomness of user power consumption behavior may cause the power consumption data of users to have the same change trend as the line loss of the area in a certain period of time. Therefore, the correlation analysis results between the power consumption data of all normal users in each area and the line loss data of the area range from [0.09, 0.17]. Proven, Figure 7 Most of the outliers after clustering are electricity thieves set up in the corresponding substation area. Therefore, the method of the present invention can effectively detect intermittent electricity thieves based on the key features of electricity theft under the interference of randomness of user electricity consumption behavior.

[0065] For the electricity thieves in the above three simulated substations, the methods used in the prior art: transfer entropy density clustering (TE-DBSCAN), correlation analysis for precise positioning of electricity theft (CAPET), fast clustering algorithm based on density peak (MIC-CFSFDP), wide deep convolutional neural network (WDCNN) and the CNN-DBSCAN method of the present invention were used as comparison items for detection. The comparison results of each method after detection are shown in Table 3.

[0066] Table 3 Comparative experiments with other similar methods

[0067]

[0068] The comparison results in Table 3 show that the periodic characteristics of intermittent electricity theft users are more complex than those of long-term electricity theft users, resulting in a large deviation when WDCNN detects electricity theft based only on the user's power consumption data, so the detection effect is poor. Since some intermittent electricity thieves in the simulated substation have a short electricity theft cycle and the proportion of electricity theft is lower than the original data, TE-DBSCAN and MIC-CFSFDP directly perform electricity theft detection based on global data through the correlation analysis between user power consumption data and substation line loss data, which is easy to cause a certain degree of false detection, and the MIC in MIC-CFSFDP is not suitable for correlation analysis between short-term data, so the detection effect in intermittent electricity theft samples is also poor. When applied, the above two methods failed to accurately capture the local strong correlation characteristics between user power consumption data and substation line loss, so some intermittent electricity theft users were not effectively identified. However, the method CNN-DBSCAN described in the present invention can accurately identify the maximum correlation period that may exist for each user based on the multi-scale CNN model, so it can identify intermittent electricity thieves with a low proportion of the above-mentioned electricity theft cycle. Among all the comparison methods, CAPET can effectively detect intermittent electricity theft users, but it only judges users with strong correlation as electricity theft users based on the global correlation analysis results, and cannot exclude the time periods when the power consumption data and the line loss data of the substation have the same trend due to the randomness of the normal user's electricity consumption behavior, so there is still a high false alarm rate. In fact, the substation line loss rate anomaly screening should be carried out for the time periods when the user's power consumption data and the substation line loss data have the same trend, so as to determine the existence of electricity theft users based on the correlation analysis results in the time period of abnormal line loss rate.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to be a further limitation of the present invention. All equivalent changes made using the contents of the present specification and drawings are within the protection scope of the present invention.

Claims

1. A CNN-DBSCAN high-loss area intermittent electricity theft detection method based on correlation analysis, characterized in that: The following steps are involved: Step 1: pre-process the power consumption data of each power user in the high-loss distribution area and the line loss data of the distribution area within the detection period, and divide them into long-term and short-term joint data of different resolutions according to short-term and daily time intervals, build and train a multi-scale CNN model, and obtain the maximum period of correlation between the power consumption data of each power user and the line loss data of the distribution area based on the long-term and short-term joint data, that is, the period with the greatest suspicion of power theft; Step 2: According to the actual abnormal time of the line loss data, the maximum period of suspected electricity theft for each user output by the multi-scale CNN model is compared to calculate the time distance offset weight; then, the Pearson correlation analysis algorithm is used to obtain the suspected electricity theft evaluation index size of each user during the abnormal period of the line loss data; Among them, firstly, the offset between the maximum correlation time period output by the multi-scale CNN model and the abnormal line loss rate time period of the substation is calculated by the size of the time distance weighted value, so as to obtain the maximum result that can be obtained when the correlation analysis is performed on the user power consumption data and the substation line loss data under the premise of abnormal line loss rate; the Pearson correlation analysis algorithm based on time distance weighting is shown as follows: ， in, It is the Pearson correlation analysis result of the user power consumption data and the area line loss data in the time period with the maximum correlation output by the model under the time distance weighting, that is, the maximum correlation analysis result that can be obtained for each user to be tested in the time period with abnormal area line loss rate; Indicates the time period with the maximum correlation for each user under test output by the multi-scale CNN model Line loss data of the substation area within Indicates the time period with the maximum correlation for each user under test output by the multi-scale CNN model Power consumption data of users in is the time distance offset weight; Step 3: Use the DBSCAN algorithm to perform cluster analysis on the suspected electricity theft evaluation index values ​​of all electricity users in the high-loss distribution station area, identify the clustered outliers as electricity thieves in the station area, and assign staff to conduct on-site inspections of the electricity theft behaviors of the electricity thieves.

2. According to claim 1, a CNN-DBSCAN high-loss area intermittent electricity theft detection method based on correlation analysis is characterized in that: In step 1, the multi-scale CNN model includes a short-term feature extraction module, a long-term feature extraction module and a long-term and short-term feature comparison module; the output results of the short-term feature extraction module and the long-term feature extraction module are jointly sent to the long-term and short-term feature comparison module; The short-term feature extraction module consists of 5 convolutional layers, which are used to extract the short-term correlation features between the user power consumption data and the area line loss data; The long-term feature extraction module consists of three convolutional layers, which are used to extract the daily correlation features between user power consumption data and station area line loss data; The long-term and short-term feature comparison module is composed of 4 fully connected layers, which are used to compare the differences between long-term and short-term features, and further determine the time period when there is a strong correlation between user power consumption data and station line loss data.

3. According to claim 2, a CNN-DBSCAN high-loss area intermittent electricity theft detection method based on correlation analysis is characterized in that: The input data of the short-term feature extraction module is the user power consumption data under the short-term measurement frequency corresponding to each week in a month. And short-term line loss in the area , ; The input data of the long-term feature extraction module is the user power consumption data D and the daily line loss S of the area at the daily measurement frequency in months; Short-term line loss within the distribution station area The general calculation method of daily line loss S is as follows: ， in, The total power supply of the distribution area in the time period corresponding to different measurement resolutions; It is the sum of the power consumption data recorded by the smart meters of each user in the distribution area during the time periods corresponding to different measurement resolutions, and it is the total power sales in the distribution area; is the line loss data of the substation in the time period corresponding to different measurement resolutions, including the technical loss caused by the loss of the transmission line itself and the non-technical loss caused by the theft of electricity; t represents the time interval for calculation under different measurement frequencies; when When the short-term line loss is calculated ,when When , the daily line loss S is calculated; The multi-scale CNN model uses the long-term feature extraction module and the short-term feature extraction module to identify the long-term and short-term correlations between each user's power consumption data and the line loss data in the substation area. Finally, the long-term and short-term feature comparison module maps the difference between local features and global features. The difference is used to output the probability of the period of maximum correlation between each user's high-resolution power consumption data and the substation line loss data every week within each month. , and according to The maximum value of the correlation is obtained from the time period with the maximum correlation. .

4. According to claim 1, a CNN-DBSCAN high-loss area intermittent electricity theft detection method based on correlation analysis is characterized in that: In step 1, the loss function of the multi-scale CNN model during training is as follows: ， in, Indicates the probability that the weekly power consumption data and the area line loss data have the greatest correlation, represents the true label.

5. The CNN-DBSCAN high-loss area intermittent electricity theft detection method based on correlation analysis according to claim 1 is characterized in that: In step 2, the abnormal time period T of the line loss data is determined, specifically: Taking line loss rate as the monitoring indicator, the calculation is as follows: ， in, is the specific acquisition frequency corresponding to the short-time measurement resolution; The total power supply of the distribution station area within the time period corresponding to the short-term measurement resolution; It is the line loss data of the substation in the time period corresponding to the short-term measurement resolution, including the technical loss part TL caused by the loss of the transmission line itself and the non-technical loss NTL caused by the theft of electricity; It is the line loss rate of the distribution station area in the time period corresponding to the short-term measurement resolution, in units of %; If there are no electricity theft users in the distribution station area, the NTL value is 0 under ideal conditions; if there are electricity theft users in the distribution station area, the NTL increases, causing the line loss rate of the station area to exceed the normal range; suspected electricity theft users in high-loss stations are located, and the abnormal time period T when the line loss rate exceeds the normal range is monitored.

6. The CNN-DBSCAN high-loss area intermittent electricity theft detection method based on correlation analysis according to claim 5 is characterized in that: In step 2, the time distance offset weight is calculated as: ， in, is the time distance offset weight; and They are the start and end time nodes of the abnormal line loss rate period in the substation area; and They are respectively the start and end time nodes of the time period with the maximum correlation for each user to be tested output by the multi-scale CNN model.

7. The CNN-DBSCAN high-loss area intermittent electricity theft detection method based on correlation analysis according to claim 1 is characterized in that: In step 3, the DBSCAN algorithm is used to perform cluster analysis on the suspected electricity theft evaluation index values ​​of all electricity users in the high-loss distribution station area, specifically: Step 3-1: The time-distance weighted correlation analysis results of all users to be tested are used as cluster samples. Based on the characteristics of the samples and the expected density of clusters, the minimum number of samples MinPts required to form a cluster is selected; Step 3-2: The time-distance weighted correlation analysis results of all users to be tested are used as cluster samples to obtain the neighborhood radius Eps, so as to accurately identify outlier electricity theft users. Specifically: a) Calculate the Euclidean distance between each sample point and other sample points to obtain the distance matrix between all samples; b) For each sample point, select the distance to its kth nearest neighbor as the parameter of all sample points, and plot it according to the order of the sample points to obtain the k-distance curve; c) In the k-distance graph, determine the inflection point of the curve and set the distance corresponding to the inflection point as the neighborhood radius Eps.

Images