A granular early warning control method for power marketing audit

Abstract

The present application provides a granular early warning control method for power marketing audit, including: obtaining multi-granular electricity consumption data of typical user groups in different regions, grouping users according to their geographical location coordinates, obtaining typical electricity consumption patterns of groups in different regions, and extracting electricity consumption behavior characteristics of peak hours, valley-to-peak ratios, and load curve shapes of each group; analyzing the influencing factors of differences in electricity consumption patterns between groups in different regions, including residents' living electricity consumption habits, industrial and commercial electricity consumption layout, and climatic conditions, constructing electricity consumption portraits of regional groups, and determining the corresponding existing electricity theft methods and electricity theft time characteristics for differences in electricity consumption patterns of groups in different regions; comparing the real-time multi-granular electricity consumption data of each user with the typical electricity consumption patterns of the group to which it belongs at different time scales, calculating the similarity of multi-dimensional load curves, and if the similarity is lower than a preset dynamic threshold, it is determined that the user has abnormal electricity consumption behavior, and an abnormal electricity consumption warning is triggered.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a granular early warning control method for power marketing audit. Background Art

[0002] The data granularity early warning control in power marketing audit faces a complex technical challenge. When monitoring massive smart meter data in real time, the system needs to process multi-dimensional information at the same time, including meter readings, power consumption changes, and load curves. However, these data often have discontinuities in time series and uneven spatial distribution, which makes it difficult for anomaly detection algorithms to accurately identify real power theft. Specifically, during the collection process, smart meter data may be missing or delayed due to communication interruptions, equipment failures, etc., forming breakpoints in the time series. At the same time, the power consumption patterns of different regions and different user types are significantly different, showing uneven spatial distribution. This data characteristic makes traditional statistical methods and machine learning algorithms face huge challenges in processing. For example, when identifying economic power theft, it is necessary not only to analyze the long-term changes in users' power consumption patterns, but also to analyze and judge the group power theft behavior of users. However, due to the discontinuity of the data, it is difficult to build a stable power consumption baseline model. When detecting direct power theft, it is necessary to compare the power consumption data of adjacent areas in real time, but the uneven spatial distribution leads to a high false alarm rate. In addition, electricity theft itself is concealed and diverse, and criminals may steal electricity in an intermittent or gradual manner, further increasing the difficulty of detection. Therefore, how to design an algorithm model that can adapt to data characteristics and accurately identify various types of electricity theft in a complex data environment has become a technical problem that needs to be solved urgently. Summary of the invention

[0003] The present invention provides a method for early warning control of power marketing audit granularity, which mainly includes:

[0004] Obtain multi-granular electricity consumption data of typical user groups in different regions, group users according to their geographic location coordinates, obtain typical electricity consumption patterns of groups in different regions, and extract electricity consumption behavior characteristics of each group in terms of peak hours, valley-peak ratio, and load curve shape;

[0005] Analyze the factors that influence the differences in electricity consumption patterns among different regional groups, including residents' electricity consumption habits, industrial and commercial electricity consumption layout, and climate conditions, build electricity consumption portraits for regional groups, and determine the corresponding electricity theft methods and time characteristics based on the differences in electricity consumption patterns among different regional groups;

[0006] According to the typical electricity consumption pattern of each regional group, a multivariate load curve feature is constructed. Combined with the collective electricity theft methods, electricity theft time characteristics and electricity consumption portraits of similar user groups in adjacent regions, a library of electricity theft disguise patterns for different groups is established, including short-term electricity theft disguise, long-term low-volume electricity theft disguise, electricity theft disguise at different time periods and electricity theft disguise at peak time;

[0007] For each user's real-time multi-granularity electricity consumption data, compare it with the typical electricity consumption pattern of the group to which it belongs at different time scales, and calculate the similarity of multi-dimensional load curves. If the similarity is lower than the preset dynamic threshold, it is judged that the user has abnormal electricity consumption behavior, triggering abnormal electricity consumption warning;

[0008] Obtain the historical multi-granular electricity consumption data of the warning users, and determine whether they have long-term low-power consumption behavior through time series decomposition and pattern matching. Combined with the mutation characteristics of their load curves, identify suspected economic electricity theft users, and match the corresponding electricity theft disguise pattern according to their electricity consumption behavior characteristics;

[0009] Obtain multi-granular electricity consumption data of other users in the area where the suspected economic electricity theft users are located and in the adjacent areas, calculate the group electricity theft risk index of the area, identify the spatial distribution characteristics of electricity theft users, and compare them with the typical electricity theft disguise patterns of the group in the area to determine the regional electricity theft risk;

[0010] Based on the abnormal electricity usage behavior characteristics of suspected economic electricity theft users, electricity theft disguise patterns and the group electricity theft risk index in the area where they are located, rule-based classification is used to classify electricity theft users, and combined with their electricity load curves and electricity usage data, the electricity loss of each type of electricity theft user is estimated and pushed to the marketing audit system.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a granular early warning control method for power marketing audit. Firstly, the method analyzes and processes the power consumption data of typical user groups in different regions by using multi-granularity data, realizes user grouping and extraction of typical power consumption patterns, and improves the accuracy and comprehensiveness of the analysis; secondly, by multi-dimensionally comparing the real-time power consumption data of users with the typical patterns of the groups, identifying abnormal power consumption behaviors, considering the factors of multiple dimensions such as individual users, regional groups, time and space, making the power theft detection more comprehensive and accurate; in addition, combining historical data analysis and regional group characteristics and analyzing the seasonal power consumption characteristics and long-term power consumption trends of regional groups, not only individual power theft can be detected, but also collective, regional and progressive power theft behaviors can be discovered; finally, according to the abnormal characteristics of users, camouflage patterns and regional risks, the power theft users are classified and the power loss can be estimated, and the power loss due to power theft can be estimated, which provides a strong basis for subsequent processing. In summary, the present invention realizes the accurate identification of power theft behaviors of different types and scales by processing massive data. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flow chart of a method for granular early warning control of power marketing audit.

[0014] Figure 2 It is a schematic diagram of a power marketing audit granularity early warning control method of the present invention.

[0015] Figure 3 It is another schematic diagram of a power marketing audit granularity early warning control method of the present invention. DETAILED DESCRIPTION

[0016] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] like Figure 1-3 In this embodiment, a method for early warning control of power marketing audit granularity may specifically include:

[0018] Step S101, obtain multi-granular electricity consumption data of typical user groups in different regions, group users according to their geographical location coordinates, obtain typical electricity consumption patterns of groups in different regions, and extract electricity consumption behavior characteristics of peak hours, valley-peak ratios, and load curve shapes of each group.

[0019] The multi-granularity electricity consumption data of users in different regions are obtained, and the multi-granularity electricity consumption data includes minute-level, hour-level and daily-level electricity consumption data, meter readings and load curves. At the same time, the geographical location coordinate information of the users is collected to establish the correlation between the user electricity consumption data and the geographical location. According to the geographical location coordinates of the users, spatial analysis is performed using the geographic information system, and the users are grouped using the K-means clustering algorithm. For the electricity consumption data of each group, the average daily electricity consumption, peak electricity consumption and standard deviation of electricity consumption indicators are calculated, and the peak electricity consumption period and valley-peak ratio electricity consumption behavior characteristics of each group are extracted. Based on the clustering results, the load curves of each user group are morphologically analyzed, and the similarity of load curves between different user groups is calculated using dynamic time warping or Euclidean distance. The typical electricity consumption patterns of groups in different regions are summarized, and the daily load curves, weekly load curves and monthly load curves of each group are drawn. For each user group, seasonal trend decomposition is used to decompose the trend component, seasonal component and random component of its electricity consumption data, and the periodic characteristics of electricity consumption of each group are extracted, including the difference in electricity consumption patterns between weekdays and weekends and seasonal electricity consumption changes.

[0020] Specifically, analyze the long-term trend, identify abnormal fluctuations, and construct a power consumption feature vector containing these features. When obtaining multi-granularity power consumption data of users, the real-time collection system of smart meters is used to record power data once a minute, summarize it once an hour, and generate daily data at 23:59 every day. At the same time, the user's geographic coordinates are obtained through the GPS positioning module, accurate to six decimal places. Use the geographic information system for spatial analysis, and divide the users into 10 km × 10 km grids according to longitude and latitude. In each grid, the K-means clustering algorithm is applied, the number of clusters is set to 5, and iterates 100 times to divide the users into 5 groups. For each group, the average daily power consumption, peak power consumption and standard deviation are calculated. For example, the average daily power consumption of a group is 15 kWh, the peak power consumption is 3 kW, and the standard deviation is 0.5 kWh. Extract the peak power consumption period, such as 18:00-22:00, and calculate the valley-peak ratio as 0.4. Perform morphological analysis on the load curve, use the dynamic time warping algorithm to calculate the similarity, and set the window size to 24 hours. Draw the daily, weekly, and monthly load curves for each group, such as the average load on weekdays is 1.2 kilowatts and on weekends is 0.8 kilowatts. Use the seasonal trend decomposition method, set the cycle to 168 hours, or one week, and decompose the electricity consumption data into three components: trend, season, and random. Extract periodic features, such as the electricity consumption on weekdays in summer is 20% higher than in winter. Analyze long-term trends, such as an average annual growth rate of 3%. Identify abnormal fluctuations, such as a sudden increase in electricity consumption of more than 50%. Finally, construct an electricity consumption feature vector, which includes 10 eigenvalues ​​such as average daily electricity consumption, peak-to-valley ratio, and seasonal index.

[0021] Step S102, analyzing the factors affecting the differences in electricity consumption patterns among different regional groups, including residents' electricity consumption habits, industrial and commercial electricity consumption layout and climatic conditions, constructing electricity consumption portraits for regional groups, and determining the corresponding electricity theft methods and electricity theft time characteristics based on the differences in electricity consumption patterns among different regional groups.

[0022] The electricity consumption data of different regional groups are obtained, and the data are associated according to the geographical location information, and a multi-dimensional regional electricity consumption feature data set is constructed based on the data. The principal component analysis method is used to reduce the dimension of the regional electricity consumption feature data set, and the key features that affect the difference in electricity consumption patterns are obtained by calculating the covariance matrix, solving the eigenvalues ​​and eigenvectors. The C4.5 decision tree algorithm is used to construct a regional group electricity consumption portrait, and the electricity consumption portrait contains nodes for information on electricity consumption, electricity consumption time period, and type of electricity equipment. For the electricity consumption portrait, the isolation forest algorithm is used for anomaly detection, and the anomaly score of each sample is calculated by constructing an isolated tree. If the anomaly score exceeds the preset threshold of the characteristic of electricity theft behavior, it is determined that there is a risk of electricity theft; according to the anomaly score and threshold, the electricity theft means and electricity theft time characteristics of different regional groups are judged, and an electricity theft risk assessment report containing high-risk time periods and suspicious electricity consumption behaviors is generated.

[0023] Specifically, the principal component analysis method is used to reduce the dimension of the regional electricity consumption feature data set, and then the data is standardized, the covariance matrix is ​​calculated, and the eigenvalues ​​and eigenvectors are solved. The principal components with a cumulative contribution rate of 85% are selected to extract the main factors affecting the differences in electricity consumption patterns, calculate the contribution rate of each factor, and obtain the key features that affect the differences in electricity consumption patterns. According to the extracted key features, the C4.5 decision tree algorithm is used to construct the electricity consumption portrait of the regional group, set the information gain ratio as the feature selection standard, and use pruning technology to prevent overfitting. The generated decision tree contains information nodes such as electricity consumption, electricity consumption period, and type of electricity equipment, forming a characteristic description of electricity consumption behavior of different regional groups. Based on the constructed electricity consumption portrait, the isolation forest algorithm is used for anomaly detection, and the abnormal sample ratio threshold is set to 0.1%, and 100 isolated trees are constructed. The abnormal score of each sample is calculated, and the characteristic threshold of electricity theft behavior is set. When obtaining regional electricity consumption data, smart meters are used to collect electricity consumption every 15 minutes, and residents' living habits information such as family population and work and rest time are collected through questionnaires. The industrial and commercial layout data were obtained from the urban planning department, including the type, size and geographical location of enterprises. The climate data were obtained from the meteorological station, including the average daily temperature, humidity and precipitation. These data were associated according to the geographical coordinates to construct a data set containing 50 features. The principal component analysis was applied to the data set, and the covariance matrix was calculated after standardization, and the eigenvalues ​​and eigenvectors were solved. The first 10 principal components were selected, and the cumulative contribution rate reached 87%, and the main factors affecting the difference in electricity consumption patterns were extracted. The C4.5 decision tree algorithm was used to construct the electricity consumption portrait, and the minimum information gain ratio was set to 0.01, the maximum tree depth was 10, and the minimum number of leaf node samples was 100. The generated decision tree contains nodes. If the average daily electricity consumption is >15kWh and the proportion of electricity consumption from 18:00 to 22:00 is >40%, it is a high-energy-consuming household user. Based on the electricity consumption portrait, the isolation forest algorithm was used for anomaly detection, and 100 trees were set, and the subsampling size of each tree was 256. The anomaly score was calculated and the threshold was set to 0.65. If the user's abnormal score is > 0.65 and the power consumption between 23:00 and 5:00 at night increases by more than 200%, it is considered suspected power theft. Finally, a power theft risk report is generated, including high-risk periods such as 2:00-4:00 in the morning and suspicious power consumption behaviors such as power consumption not matching power consumption equipment.

[0024] Step S103, constructing multivariable load curve characteristics for the typical electricity consumption pattern of each regional group, combining the electricity theft means of the adjacent regional groups, electricity theft time characteristics and electricity consumption portraits of similar user groups, and establishing a library of electricity theft disguise patterns for different groups, including short-term electricity theft disguise, long-term low-volume electricity theft disguise, electricity theft disguise at different time periods and electricity theft disguise at peak time.

[0025] Obtain the electricity consumption data of each regional group, and construct a multivariate load curve including electricity consumption, power factor, and harmonic content based on the electricity consumption data. Use fast Fourier transform to extract frequency domain features from the multivariate load curve, and select the first ten main frequency components as features. Calculate the mean, variance, skewness, and kurtosis of the features to obtain an eighty-dimensional feature vector as the typical electricity consumption pattern feature vector of each regional group. Use the K-means clustering algorithm to group users. Calculate the electricity consumption similarity of users in the group, using the cosine similarity metric. Construct an electricity consumption portrait of similar user groups, including ten features: average daily electricity consumption, peak-to-valley electricity consumption ratio, and electricity consumption time distribution. Based on the typical electricity consumption pattern feature vector and the electricity consumption portrait of the similar user group, use the support vector machine algorithm to establish an electricity theft behavior classifier. Use the radial basis function kernel to optimize the parameters C and gamma through grid search.

[0026] Specifically, the electricity theft features are extracted from adjacent regional groups, including sudden changes in electricity consumption, abnormal power factor, and abnormal harmonic content. The K-means clustering algorithm is used to group users, with the number of clusters set to 5 and the number of iterations set to 100. The training data set contains normal electricity consumption samples and known types of electricity theft samples, with a ratio of 9:1. The model performance is evaluated using cross-validation, with an accuracy rate of more than 95%. A library of electricity theft camouflage patterns is constructed for the various types of electricity theft behaviors identified. The autoregressive integral moving average model is used to generate simulated electricity theft load curves, including camouflage patterns such as short-term sudden increases, long-term small fluctuations, periodic anomalies, and peak anomalies. 100 simulation curves are generated for each pattern to form a complete library of electricity theft camouflage patterns. When obtaining regional group electricity consumption data, smart meters are used to record electricity consumption, power factor, and harmonic content every 5 minutes. Fast Fourier transform is applied to 24-hour data, and the 10 frequency components with the largest amplitude are selected. Their mean, variance, skewness, and kurtosis are calculated to obtain an 80-dimensional feature vector. For example, the first four dimensions of the feature vector of a residential area are [10.5, 0.92, 0.15, 2.3], which represent the mean of the main frequency components, the mean of the power factor, the mean of the harmonic content, and the kurtosis of the power consumption. The K-means algorithm is used to group users, and the number of clusters is set to 5, the maximum number of iterations is set to 100, and the convergence threshold is set to 0.001. The cosine similarity of users in the group is calculated, and the threshold is set to 0.85 to construct a power consumption profile. The profile of a user group in a commercial area includes features such as an average daily power consumption of 250 kWh, a peak-to-valley ratio of 1.8, and 75% of the power consumption from 9:00 to 18:00. The support vector machine algorithm is used to establish a classifier for power theft behavior, using the radial basis function kernel, and the optimal parameters C = 10 and gamma = 0.01 are determined through grid search. The training set contains 9000 normal samples and 1000 power theft samples, and the cross-validation accuracy rate reaches 96.5%. The autoregressive integrated moving average model is used to generate the electricity theft load curve, with the parameters set to p = 2, d = 1, q = 2. The short-term surge mode increases electricity consumption by 75% within a random 3-hour period; the long-term small fluctuation mode increases electricity consumption by 7.5% every day within 30 days; the periodic abnormal mode increases electricity consumption by 25% from 22:00 to 6:00; the peak abnormal mode increases electricity consumption by 50% from 14:00 to 16:00. Each mode generates 100 curves, and a total of 400 simulated electricity theft load curves constitute the electricity theft camouflage mode library.

[0027] Step S104, for each user's real-time multi-granularity electricity consumption data, compare it with the typical electricity consumption pattern of the group to which it belongs at different time scales, and calculate the similarity of the multi-dimensional load curve. If the similarity is lower than the preset dynamic threshold, it is judged that the user has abnormal electricity consumption behavior, and an abnormal electricity consumption warning is triggered.

[0028] The method obtains the user's real-time multi-granularity electricity consumption data, wherein the multi-granularity electricity consumption data includes minute-level data, hour-level data and daily-level data; extracts the characteristics of electricity consumption, power factor and harmonic content according to the multi-granularity electricity consumption data to obtain a multi-dimensional load curve; adopts a sliding window method to segment the multi-dimensional load curve; uses a dynamic time warping algorithm to calculate the similarity between the multi-dimensional load curve and a preset load curve template at each time scale, and calculates the Euclidean distance as a cost function; adopts an adaptive threshold algorithm to dynamically adjust the similarity judgment threshold, and the adaptive threshold algorithm is based on the statistical distribution of historical data and time factors; if the similarity is lower than the similarity judgment threshold, it is judged that the user has abnormal electricity consumption behavior; triggers an abnormal electricity consumption warning, and records the abnormal period, abnormal degree and abnormal type; pushes the abnormal electricity consumption warning to the monitoring module in JSON format, and the JSON format contains user ID, abnormal time, abnormal type and abnormal degree fields.

[0029] Specifically, the real-time multi-granularity electricity consumption data of users is obtained, including one data point every 5 minutes at the minute level, hourly and daily data, and the characteristics such as electricity consumption, power factor, harmonic content, etc. are extracted to construct a multi-dimensional load curve. The sliding window method is used to segment the data of different time scales, and the window size is set to 24 hours and the step length is 1 hour. For minute-level data, each window contains 288 data points. According to the typical electricity consumption pattern of the group to which the user belongs, the corresponding load curve template is retrieved from the electricity consumption pattern library. The dynamic time warping algorithm is used to calculate the similarity between the actual load curve of the user and the template curve at each time scale. The window size is set to 5, the Euclidean distance is calculated as the cost function, and the optimal path is solved using dynamic programming. The adaptive threshold algorithm is used to dynamically adjust the similarity judgment threshold based on the statistical distribution of historical data and time factors. The similarity mean μ and standard deviation σ of the same period in the past 30 days are calculated, and the initial threshold is set to μ-2σ. The threshold is adjusted according to the time factor, and the threshold on weekdays is reduced by 5% and the threshold on weekends is increased by 5%. Different user groups are set differently according to their electricity consumption stability, and the threshold for groups with large fluctuations is relaxed by 10%. If the calculated similarity is lower than the dynamic threshold, the user is judged to have abnormal electricity consumption behavior. Trigger abnormal electricity consumption warning, record the abnormal period accurate to the hour, the abnormal degree (i.e. the difference between the similarity and the threshold), and the abnormal type (including short-term sudden increase, long-term small fluctuation, periodic abnormality and peak abnormality). The warning information is pushed to the relevant monitoring module in JSON format, including fields such as user ID, abnormal time, abnormal type and abnormal degree. The smart meter collects electricity consumption data every 5 minutes, recording the power consumption, power factor and harmonic content. For example, the power consumption of a user from 8:00 to 8:05 is 2.5kWh, the power factor is 0.95, and the harmonic content is 3%. Use a 24-hour window with a step size of 1 hour to process the data, and each window contains 288 data points. For the residential area group to which the user belongs, retrieve the typical load curve template. Using the dynamic time warping algorithm, the window size is set to 5, and the similarity between the user curve and the template is calculated. For example, the similarity between 8:00 and 9:00 is 0.92. The adaptive threshold algorithm calculates the similarity mean μ=0.95 and the standard deviation σ=0.03 ​​for the same period in the past 30 days, and the initial threshold is set to 0.89. The threshold is adjusted to 0.8455 on weekdays and 0.9345 on weekends. If the power consumption of the user group fluctuates greatly, the threshold is further relaxed to 0.76095. Assuming that the calculated similarity is 0.75, which is lower than the threshold of 0.76095, it is judged as abnormal power consumption. The abnormal period is recorded as 8:00-9:00, the abnormality degree is 0.01095, and the abnormality type is short-term sudden increase. Generate JSON format warning information: {"User ID":"U001","Abnormal time":"2024-07-2508:00","Abnormal type":"Short-term sudden increase","Abnormal degree":0.01095}, and push it to the monitoring module.

[0030] Step S105, obtain the historical multi-granular electricity consumption data of the warning user, and determine whether it has long-term low-power consumption behavior through time series decomposition and pattern matching. Combined with the mutation characteristics of its load curve, identify suspected economic electricity theft users, and match the corresponding electricity theft disguise pattern according to their electricity consumption behavior characteristics.

[0031] The historical multi-granularity electricity consumption data of the warning user is obtained, and the multi-granularity electricity consumption data includes minute-level data, hour-level data and daily-level data. The multi-granularity electricity consumption data is decomposed by seasonal and trend decomposition methods to obtain trend items, seasonal items and residual items. The trend items are pattern matched by a dynamic time warping algorithm and compared with the preset low-power consumption mode, and the preset low-power consumption mode includes a continuous decline type, a periodic trough type and a sudden drop stable type. If the similarity of the pattern matching exceeds the preset threshold, it is judged that the user has a long-term and continuous low-power consumption behavior. The load curve of the warning user is subjected to mutation feature analysis, and the singular point of the load curve is detected by the wavelet transform method. If the mutation amplitude of the singular point exceeds the preset proportion of the user's average load, and the frequency is greater than the preset number of times, combined with the judgment result of the long-term and continuous low-power consumption behavior, it is identified as a suspected economic electricity theft user. According to the electricity consumption behavior characteristics of the suspected economic electricity theft user, the most similar electricity theft disguise pattern is matched from the pre-established electricity theft disguise pattern library, which contains short-term sudden increase, long-term small fluctuation, periodic anomaly and peak anomaly patterns. The cosine similarity algorithm is used to calculate the matching degree, and the pattern with the highest matching degree is selected as the electricity theft disguise pattern of the suspected economic electricity theft user.

[0032] Specifically, the seasonal and trend decomposition methods are used to decompose the time series, and the seasonal cycle is set to 24 hours to obtain the trend term, seasonal term and residual term, and extract the long-term electricity consumption trend characteristics of the user, including the trend slope, periodic fluctuation amplitude, etc. The similarity score is calculated, and the threshold is set to 0.85. If the similarity exceeds the threshold, it is judged that the user has a long-term and continuous low-power consumption behavior. The number of decomposition layers is set to 5, and the wavelet coefficients of each layer are calculated. The mutation point is identified by the threshold method, and the mutation amplitude and frequency are calculated. If the mutation amplitude exceeds 50% of the user's average load and the frequency is greater than 3 times a week, combined with the long-term low-power consumption behavior judgment results, it is identified as a suspected economic electricity theft user. According to the electricity consumption behavior characteristics of the suspected economic electricity theft users identified, including the duration of low power consumption and load mutation characteristics, the most similar electricity theft camouflage pattern is matched from the pre-established electricity theft camouflage pattern library. For the warning user A, its electricity consumption data for the past 90 days is obtained, including electricity consumption every 5 minutes, average power per hour, and total electricity per day. The seasonal and trend decomposition method is used, and 24 hours is set as a cycle. The trend item shows that the electricity consumption decreases by 5% month by month, and the seasonal item shows the peak electricity consumption from 14:00 to 18:00 every day. Using the dynamic time warping algorithm, the window size is set to 48 hours and the step size is 1 hour. The trend item is compared with the preset low-power mode, and the similarity with the continuous decline mode is calculated to be 0.89, which exceeds the threshold of 0.85, and it is determined that there is a long-term low-power consumption behavior. The load curve is decomposed by 5 layers of wavelet, and the threshold is set to 20% of the user's average load. 12 mutation points are detected, of which 8 mutation amplitudes exceed 50%, and an average of 3.5 times per week. Combined with the low-power behavior, user A is identified as a suspected economic electricity theft user. Feature vectors are extracted from the electricity theft camouflage pattern library, including low power for 30 days, mutations 3-4 times a week, and mutation amplitudes of 50%-80%. The cosine similarity algorithm is used to calculate the matching degree between user A and each mode, and the matching degree with the long-term small fluctuation mode is the highest, which is 0.92, and it is determined to be the electricity theft disguise mode of this user.

[0033] Step S106, obtaining multi-granular electricity consumption data of other users in the area where the suspected economic electricity theft user is located and in the adjacent area, calculating the group electricity theft risk index of the area, identifying the spatial distribution characteristics of the electricity theft users, and comparing them with the typical electricity theft disguise patterns of the group in the area to determine the regional electricity theft risk.

[0034] The multi-granularity electricity consumption data of other users in the area where the suspected economic electricity theft user is located and in the adjacent area are obtained, and the multi-granularity electricity consumption data includes minute-level, hour-level and daily-level data; a multi-dimensional data set including electricity consumption, power factor, harmonic content and geographic coordinates is constructed based on the multi-granularity electricity consumption data, and each user record in the multi-dimensional data set includes historical data of a preset number of days; a local anomaly factor algorithm is used to calculate the anomaly score of each user in the multi-dimensional data set, and the anomaly score is based on a preset K nearest neighbor parameter and anomaly threshold; the proportion of abnormal users is counted according to the anomaly score, and the group electricity theft risk index of the area is calculated in combination with the user density and geographical distribution; the spatial distribution characteristics of the abnormal users are calculated using the spatial autocorrelation analysis method. The spatial distribution characteristics include the global Moran index and the local Gittes index; according to the spatial distribution characteristics, the clustering areas and hotspot distribution of electricity theft users are identified, and the clustering areas and hotspot distribution are based on a preset spatial weight matrix and significance level; the typical electricity theft camouflage pattern characteristics of the regional group are extracted from the historical data, and the typical electricity theft camouflage pattern characteristics include the rate of change of electricity consumption, power factor fluctuation and abnormal harmonic content; the dynamic time warping algorithm is used to calculate the similarity between the electricity consumption behavior of each user and the typical electricity theft camouflage pattern characteristics, and the similarity is based on a preset window size, step size and similarity threshold; combined with the group electricity theft risk index and the spatial distribution characteristics, the weighted average method is used to comprehensively judge the regional electricity theft risk.

[0035] Specifically, we obtain multi-granular electricity consumption data of other users in the area where the suspected economic electricity theft users are located and in the adjacent areas, including minute-level (one data point every 5 minutes), hour-level and daily-level data. Combined with the user's geographic location information, we construct a multidimensional data set containing electricity consumption, power factor, harmonic content and geographic coordinates. Each user record in the data set contains 30 days of historical data. Based on the constructed multidimensional data set, the local anomaly factor algorithm is used to calculate the anomaly score of each user, and the K nearest neighbor parameter is set to 20 and the anomaly threshold is set to 1.5. The proportion of abnormal users is counted, and the group electricity theft risk index of the area is calculated by combining the number of users per square kilometer and the geographical distribution of user density. The risk index calculation formula is risk index = abnormal user proportion * user density * geographical distribution coefficient. Using the spatial autocorrelation analysis method, we calculate the spatial distribution characteristics of abnormal users. First, we calculate the global Moran index to determine the overall spatial correlation, and then calculate the local Gittes index to identify the clustering areas and hotspot distribution of electricity theft users. The spatial weight matrix is ​​set to inverse distance weighting, and the significance level is 0.05. The dynamic time warping algorithm is used to calculate the similarity between the electricity consumption behavior of each user and the typical pattern. The window size is set to 24 hours, the step size is 1 hour, the similarity threshold is set to 0.8, and the proportion of users with high similarity is counted. Combined with the group electricity theft risk index and spatial distribution characteristics, the weighted average method is used to comprehensively judge the regional electricity theft risk. For the suspected economic electricity theft users in area A, the system automatically obtains 30 days of electricity consumption data of 1,000 users in the area and within the adjacent 5 kilometers, including electricity consumption, power factor and harmonic content every 5 minutes, and records the user's geographic coordinates. The local anomaly factor algorithm is used, K=20 is set, the anomaly score of each user is calculated, the anomaly threshold is set to 1.5, and 50 abnormal users are identified. The user density is calculated to be 200 households / square kilometer, the proportion of abnormal users is 5%, the geographical distribution coefficient is 0.8, and the group electricity theft risk index is 8. Through spatial autocorrelation analysis, the global Moran index I=0.6 is calculated, and the p value is <0.05, indicating that there is significant spatial correlation. The local Gittens index was calculated to identify three high-risk clusters, with the largest hotspot containing 15 abnormal users. Typical electricity theft camouflage patterns were extracted from historical data, such as a 10% decrease in daily electricity consumption, a 20% increase in power factor fluctuations, and a 30% excess of harmonic content. The dynamic time warping algorithm was used with a window size of 24 hours and a step length of 1 hour to calculate the similarity between the user's electricity consumption behavior and the typical pattern. The threshold was set to 0.8, and 75 highly similar users were found. Finally, the system combined various indicators and determined that the risk level of electricity theft in the area was high and required priority treatment.

[0036] Analyze the seasonal electricity consumption characteristics of different types of regional groups, identify abnormal seasonal fluctuations that are inconsistent with the group characteristics, including peak electricity consumption in industrial parks during the non-production season and abnormally high electricity consumption in residential areas in spring and autumn, compare the seasonal electricity consumption patterns of each group with historical data and similar regional groups, and determine collective electricity theft behavior.

[0037] The electricity consumption data of different types of regional groups in the past five years are obtained. The seasonal components of the electricity consumption data are extracted by seasonal and trend decomposition methods to obtain the seasonal electricity consumption characteristic model of each group. According to the seasonal electricity consumption characteristic model, the K-means clustering algorithm is used to classify each regional group, and similar regional groups are identified based on geographical location, electricity consumption scale and industrial type characteristics. If the characteristics of similar regional groups meet the preset conditions, a group characteristic label library is established. The difference between the current electricity consumption data and the seasonal electricity consumption characteristic model is compared by the local anomaly factor algorithm combined with the group characteristic label library. If the anomaly score exceeds the anomaly score threshold, it is judged as abnormal electricity consumption behavior. If the industrial park has a peak electricity consumption in the non-production peak season, and the electricity consumption exceeds the preset proportion of the annual average, it is marked as abnormal seasonal fluctuation. If the electricity consumption in the residential area in spring and autumn exceeds the preset proportion of the summer and winter average, it is marked as abnormal seasonal fluctuation. The abnormal seasonal fluctuation is compared with the historical data of the same period and the electricity consumption pattern of similar regional groups, and the dynamic time warping algorithm is used to calculate the similarity score. If the similarity score is lower than the preset threshold and the abnormal pattern is consistent among users above the preset proportion in the area, it is determined to be collective electricity theft.

[0038] Specifically, the electricity consumption data of different types of regional groups in the past five years, including industrial parks and residential areas, are obtained. The seasonal components are extracted using seasonal and trend decomposition methods. The seasonal cycle is set to 12 months, and the seasonal electricity consumption characteristic model of each group is constructed. The K-means clustering algorithm is used to classify each regional group, and the number of clusters is set to 5. Based on the characteristics of geographical location such as longitude and latitude, electricity consumption scale such as annual total electricity consumption, and industry type such as industry code, similar regional groups are identified and a group characteristic label library is established. Each group label contains information such as regional type, electricity consumption scale level, and main industry type. Through the local anomaly factor algorithm, combined with the group characteristic label, the difference between the current electricity consumption data and the seasonal electricity consumption characteristic model is compared, and the anomaly score threshold is set to 1.5. If the industrial park has a peak electricity consumption in the non-production peak season according to the industry characteristics, such as exceeding 150% of the annual average, or the residential area in the spring and autumn seasons such as March-May and September-November exceeds 120% of the summer and winter average, it is marked as abnormal seasonal fluctuations. The identified abnormal power consumption pattern is compared with the historical data of the same period and the power consumption pattern of similar regional groups. The dynamic time warping algorithm is used to calculate the similarity score, and the judgment threshold is set to 0.7. If the similarity is lower than the threshold, and the abnormal pattern is consistent among more than 50% of users in the area, such as the power consumption pattern similarity is higher than 0.9, it is judged as collective power theft. Taking City A as an example, the system automatically obtains power consumption data from 2019 to 2023, including 10 industrial parks and 50 residential areas. The seasonal and trend decomposition method is used, and the 12-month cycle is set to extract the seasonal components to obtain the power consumption characteristic model of each region. The K-means clustering algorithm is applied to these 60 regions, with k=5, and the latitude and longitude, annual power consumption and industry code are used as features to divide them into 5 categories. For example, the first category is a large manufacturing park with an annual power consumption of >100 million kWh and the main industry code is C26. The local anomaly factor algorithm is used, and the threshold is set to 1.5 to compare the current power consumption with the characteristic model. It was found that Industrial Park B belonged to Category 1 and its electricity consumption in April reached 210 million kWh, which was 150% higher than the annual average of 140 million kWh. The industry usually has a peak production season from August to October, and the system automatically marked it as abnormal. At the same time, the electricity consumption of residential area C reached 7.2 million kWh in March, which was 120% higher than the average of 6 million kWh in summer and winter, and was also marked as abnormal. The system then used a dynamic time warping algorithm to calculate the similarity of these abnormal electricity consumption patterns with the same period in history and similar areas. Park B scored 0.65, which was lower than the threshold of 0.7, and 60% of the enterprises in the park showed similar patterns, such as similarity>0.9, and the system automatically determined it as collective electricity theft.

[0039] Track and analyze the long-term electricity consumption trends of user groups in specific areas, identify unreasonable slow growth patterns, identify electricity consumption growth patterns that do not match the group's development stage, including unreasonable continuous growth in mature communities, compare the group's electricity consumption growth rate with typical development indicators of this type of area, and judge group progressive electricity theft behaviors that deviate from normal trajectories.

[0040] The monthly electricity consumption data of the user group in the specific area for the past 10 years is obtained, and the trend items of the electricity consumption data are extracted by seasonal and trend decomposition methods. The annual electricity consumption growth rate and cumulative growth rate are calculated according to the trend items, and the growth trend in the next two years is predicted by the exponential smoothing method to obtain the electricity consumption trend model. According to the user group characteristics and development stage indicators of the specific area, the K-means clustering algorithm is used to group different types of areas. The matching degree of the electricity consumption trend model and the group development stage label is compared by the isolation forest algorithm. If the abnormal score of the matching degree exceeds the preset threshold, it is determined to be an abnormal growth pattern. The abnormal growth pattern is compared with the typical development indicators of this type of area, and the Z-score is calculated as the deviation score. If the absolute value of the Z-score exceeds the preset threshold, and the abnormal pattern is consistent among users in the area, it is determined to be a group progressive electricity theft behavior.

[0041] Specifically, the monthly electricity consumption data of the user group in a specific area for the past 10 years is obtained, and the seasonal cycle is set to 12 months. Based on the characteristics of the user group such as population density, number of commercial facilities and building age, and development stage indicators such as GDP growth rate and population growth rate, the K-means clustering algorithm is used to group different types of areas, and the number of clusters is set to 5. A group development stage label library is established, including labels such as 0-5 years for new areas, 5-15 years for rapid development areas, 15-30 years for mature areas, and more than 30 years for old areas. The matching degree between the electricity consumption trend model and the group development stage label is compared through the isolation forest algorithm. The abnormal score threshold is set to 0.7. If the average annual growth rate of the mature area from 15 to 30 years continues to exceed 5%, or the average annual growth rate of the new area from 0 to 5 years is less than 10%, it is marked as an abnormal growth pattern. The identified abnormal growth pattern is compared with the typical development indicators of this type of area, such as per capita electricity consumption and energy consumption per unit GDP, and the Z-score is calculated as the deviation score. The judgment threshold is set to 2. If the absolute value of the Z-score exceeds 2, and the abnormal mode shows a consistent growth rate difference of less than 1 percentage point among more than 50% of users in the area, it is judged as a collective progressive electricity theft. Taking District B of City A as an example, the system automatically obtains the monthly electricity consumption data from 2013 to 2022, and uses the seasonal and trend decomposition method, setting 12 months as a cycle to extract trend items. The average annual growth rate of District B is calculated to be 3.2%, and the cumulative growth rate is 36.8%. Using the exponential smoothing method, α = 0.3, the growth rates in 2023-2024 are predicted to be 3.5% and 3.7% respectively. The system collects characteristics such as the population density of District B of 8,000 people / square kilometer, the number of commercial facilities of 500, and the average building age of 20 years, as well as development indicators such as GDP growth rate of 4% and population growth rate of 1.5%. Applying the K-means algorithm, k = 5, District B is classified as a mature area of ​​15-30 years. Using the isolation forest algorithm, setting the pollution rate to 0.1 and the sample size to 256, the anomaly score for Area B was calculated to be 0.75, exceeding the 0.7 threshold. As Area B is a mature area, the average annual growth rate of 3.2% does not exceed 5% and is not marked as abnormal. The system further calculated the Z-score of Area B's per capita electricity consumption of 2500kWh / year and unit GDP energy consumption of 0.5kWh / yuan, which were 1.8 and -0.5 respectively, both of which did not exceed the threshold of 2. At the same time, the growth rate difference of 75% of users in the area was within 0.8 percentage points, which was lower than the consistency standard of 1 percentage point. Comprehensively judging, Area B was not identified as a collective progressive electricity theft behavior.

[0042] Step S107, based on the abnormal electricity usage behavior characteristics of suspected economic electricity theft users, electricity theft disguise patterns and the group electricity theft risk index in the area, the electricity theft users are classified using rule-based classification, and combined with their electricity load curves and electricity usage data, the electricity theft loss of each type of electricity theft user is estimated and pushed to the marketing audit system.

[0043] The abnormal electricity consumption behavior characteristics, electricity theft disguise mode and group electricity theft risk index of the suspected economic electricity theft user are obtained. A feature vector is constructed based on the abnormal electricity consumption behavior characteristics, electricity theft disguise mode and regional group electricity theft risk index, and the feature vector includes the abnormal degree of electricity consumption, the type of electricity theft behavior and the regional risk score. The feature vector is used to establish a rule classifier based on a decision tree. If the abnormal degree of electricity consumption is greater than a preset threshold and the regional risk score is greater than a preset risk threshold, it is determined to be a high risk; if the abnormal degree of electricity consumption is between a first preset interval and the regional risk score is between a second preset interval, it is determined to be a medium risk; the rest of the cases are determined to be low risks. The electricity load curve and electricity usage data of each type of electricity theft user are extracted, and the deviation of the electricity load curve and electricity usage data from the normal electricity consumption mode is calculated. For the electricity load curve and electricity usage data, an autoregressive integrated moving average model is used to perform time series analysis, set a confidence interval, and identify abnormal electricity consumption time periods and abnormal electricity consumption that exceed the confidence interval. According to the abnormal power consumption period and abnormal power consumption, combined with the user type and power consumption characteristics, the weighted average method is used to estimate the power loss of various types of power theft users. An audit report containing user identification, risk level, power theft type and power loss is generated and pushed to the marketing audit system in a preset format.

[0044] Specifically, the abnormal electricity consumption behavior characteristics of suspected economic electricity theft users are obtained, such as sudden changes in electricity consumption and abnormal power factor, electricity theft disguise patterns such as short-term sudden increases and long-term small fluctuations, and the risk index of group electricity theft in the region. Construct a feature vector, including the degree of abnormal electricity consumption 0-1 points, the type of electricity theft behavior code 1-5, and the regional risk score 0-100 points. Establish a rule classifier based on a decision tree and set a classification rule set. If the degree of abnormal electricity consumption > 0.8 and the regional risk score > 80, it is judged as high risk; if 0.5 < abnormal electricity consumption ≤ 0.8 and 50 < regional risk score ≤ 80, it is judged as medium risk; the rest are judged as low risk. Extract the electricity load curves and electricity usage data of various types of electricity theft users, and calculate the deviation from the normal electricity consumption pattern. Use the autoregressive integrated moving average model for time series analysis, set the confidence interval to 95%, and identify abnormal electricity consumption periods and abnormal electricity consumption that exceed the prediction interval. For high-risk users, 80% of abnormal power consumption is taken as power loss; for medium-risk users, 60%; and for low-risk users, 40%. Generate an audit report containing user ID, risk level, power theft type, and power loss, and push it to the marketing audit system in JSON format. Taking user A as an example, the system obtains its abnormal power consumption behavior characteristics: power consumption mutation amplitude 50%, power factor fluctuation 0.2. The power theft disguise mode is a short-term surge, and the regional group power theft risk index is 85. Construct a feature vector with an abnormal power consumption degree of 0.85, a power theft behavior type code of 2, and a regional risk score of 85 points. Apply the decision tree rule classifier to determine that user A is high risk. The system extracts A's power load curve and finds that the average daily power consumption increased from 200kWh to 300kWh from July 15th to 20th. Using the autoregressive integrated moving average model, with parameters p=1, d=1, q=1 and a confidence interval of 95%, we identified July 15-20 as an abnormal power consumption period, with a total abnormal power consumption of 500kWh. According to the high-risk user standard, 80% of the abnormal power consumption is taken as the power loss, that is, 400kWh. The system automatically generates an audit report in JSON format: {"User ID":"A001","Risk Level":"High","Power Theft Type":"Short-term Sudden Increase","Power Loss":400}, and pushes it to the marketing audit system.

[0045] The above only lists some preferred embodiments of the present invention, but the present invention is not limited thereto, and many improvements and changes can be made. As long as the improvements and changes are made on the basis of the basic principles of the present invention, they should be regarded as falling within the protection scope of the present invention.

Claims

1. A method for early warning control of power marketing audit granularity, characterized in that: The method comprises: Obtain multi-granular electricity consumption data of typical user groups in different regions, group users according to their geographic location coordinates, obtain typical electricity consumption patterns of groups in different regions, and extract the peak hours, valley-peak ratios, and load curve shapes of each group as electricity consumption behavior characteristics; Analyze the factors that influence the differences in electricity consumption patterns among different regional groups, including residents' electricity consumption habits, industrial and commercial electricity consumption layout, and climate conditions, build electricity consumption portraits for regional groups, and determine the corresponding electricity theft methods and time characteristics based on the differences in electricity consumption patterns among different regional groups; According to the typical electricity consumption pattern of each regional group, a multivariate load curve feature is constructed. Combined with the collective electricity theft methods, electricity theft time characteristics and electricity consumption portraits of similar user groups in adjacent regions, a library of electricity theft disguise patterns for different groups is established, including short-term electricity theft disguise, long-term low-volume electricity theft disguise, electricity theft disguise at different time periods and electricity theft disguise at peak time; For each user's real-time multi-granularity electricity consumption data, compare it with the typical electricity consumption pattern of the group to which it belongs at different time scales, and calculate the similarity of multi-dimensional load curves. If the similarity is lower than the preset similarity judgment threshold, it is judged that the user has abnormal electricity consumption behavior, triggering abnormal electricity consumption warning; Obtain the historical multi-granular electricity consumption data of the warning users, and determine whether they have long-term low-power consumption behavior through time series decomposition and pattern matching. Combined with the mutation characteristics of their load curves, identify suspected economic electricity theft users, and match the corresponding electricity theft disguise pattern according to their electricity consumption behavior characteristics; Obtain multi-granular electricity consumption data of other users in the area where the suspected economic electricity theft user is located and in the adjacent area, calculate the group electricity theft risk index in the area where the suspected economic electricity theft user is located, identify the spatial distribution characteristics of the electricity theft users, and compare them with the typical electricity theft disguise patterns of the group in the area where the suspected economic electricity theft user is located to determine the regional electricity theft risk. The group electricity theft risk index calculation formula is: group electricity theft risk index = abnormal user ratio × user density × geographical distribution coefficient; Based on the abnormal electricity usage behavior characteristics of suspected economic electricity theft users, electricity theft disguise patterns and the group electricity theft risk index in the area where they are located, rule-based classification is used to classify electricity theft users, and combined with their electricity load curves and electricity usage data, the electricity loss of each type of electricity theft user is estimated and pushed to the marketing audit system.

2. The method according to claim 1, wherein: The method of obtaining multi-granular electricity consumption data of typical user groups in different regions, grouping users according to their geographical location coordinates, obtaining typical electricity consumption patterns of groups in different regions, and extracting peak hours, valley-peak ratios, and load curve forms of each group as electricity consumption behavior features includes: Acquire multi-granularity electricity consumption data of users in different regions, the multi-granularity electricity consumption data includes minute-level, hour-level and daily-level electricity consumption data, meter readings and load curves, and collect the geographical location coordinate information of the users to establish a correlation between the user electricity consumption data and the geographical location; According to the geographical location coordinates of the users, spatial analysis is performed using a geographic information system, and the users are grouped using a K-means clustering algorithm; For each group's electricity consumption data, calculate the average daily electricity consumption, peak electricity consumption and standard deviation of electricity consumption, and extract the peak electricity consumption period and valley-peak ratio of each group; Based on the clustering results, the load curve of each user group is analyzed in morphology, and the similarity of load curves between different user groups is calculated using dynamic time warping or Euclidean distance; Summarize the typical electricity consumption patterns of groups in different regions, and draw daily, weekly and monthly load curves for each group; For each user group, seasonal trend decomposition is used to decompose the trend component, seasonal component and random component of their electricity consumption data, and the periodic characteristics of electricity consumption of each group are extracted, including the difference in electricity consumption patterns between weekdays and weekends and seasonal changes in electricity consumption.

3. The method according to claim 1, wherein: The analysis of the factors affecting the differences in electricity consumption patterns among different regional groups includes residents' electricity consumption habits, industrial and commercial electricity consumption layout, and climatic conditions, and constructs a regional group electricity consumption portrait. According to the differences in electricity consumption patterns among different regional groups, the corresponding electricity theft methods and electricity theft time characteristics are determined, including: Obtain electricity consumption data of different regional groups, associate the data according to geographic location information, and construct a multi-dimensional regional electricity consumption feature data set based on the data; The principal component analysis method is used to reduce the dimension of the regional electricity consumption characteristic data set, and the key features that affect the differences in electricity consumption patterns are obtained by calculating the covariance matrix, solving the eigenvalues ​​and eigenvectors; The C4.5 decision tree algorithm is used to construct a regional group electricity consumption profile, which includes information nodes on electricity consumption, electricity consumption period, and type of electricity consumption equipment; For the electricity consumption portrait, an isolation forest algorithm is used to perform anomaly detection, and an anomaly score of each sample is calculated by constructing an isolation tree; If the abnormal score exceeds the preset characteristic threshold of electricity theft behavior, it is determined that there is a risk of electricity theft; According to the abnormal score and the characteristic threshold of electricity theft behavior, the electricity theft means and time characteristics of electricity theft existing in different regional groups are judged, and an electricity theft risk assessment report including high-risk time periods and suspicious electricity usage behaviors is generated.

4. The method according to claim 1, wherein: The typical electricity consumption pattern of each regional group is used to construct multivariate load curve characteristics, and combined with the group electricity theft means, electricity theft time characteristics and electricity consumption portraits of similar user groups in adjacent regions, a library of electricity theft disguise patterns for different groups is established, including short-term electricity theft disguise, long-term low-volume electricity theft disguise, electricity theft disguise at different time periods and electricity theft disguise at peak time, including: Obtaining electricity consumption data for each regional group, and constructing a multivariate load curve including electricity consumption, power factor, and harmonic content based on the electricity consumption data; Extracting frequency domain features from the multivariable load curve using fast Fourier transform, and selecting the first ten main frequency components as features; Calculate the mean, variance, skewness and kurtosis of the feature to obtain an eighty-dimensional feature vector as a typical electricity consumption pattern feature vector of each regional group; Use K-means clustering algorithm to group users; Calculate the electricity consumption similarity of users in the group, using cosine similarity measurement; Build electricity consumption profiles for similar user groups, including average daily electricity consumption, peak-to-valley ratio, and electricity consumption time distribution characteristics; Based on the typical electricity consumption pattern feature vector and the electricity consumption portrait of the similar user group, a power theft behavior classifier is established using a support vector machine algorithm; The radial basis function kernel is used and the parameters C and gamma are optimized by grid search.

5. The method according to claim 1, wherein: The real-time multi-granularity electricity consumption data of each user is compared with the typical electricity consumption pattern of the group to which it belongs at different time scales, and the similarity of the multi-dimensional load curve is calculated. If the similarity is lower than the preset similarity judgment threshold, it is judged that the user has abnormal electricity consumption behavior, and abnormal electricity consumption warning is triggered, including: Acquire real-time multi-granularity electricity consumption data of users, wherein the multi-granularity electricity consumption data includes minute-level data, hour-level data and day-level data; Extracting the characteristics of power consumption, power factor and harmonic content according to the multi-granularity power consumption data to obtain a multi-dimensional load curve; Using a sliding window method to process the multi-dimensional load curve in sections; Using a dynamic time warping algorithm to calculate the similarity between the multi-dimensional load curve and a preset load curve template at each time scale, and calculating the Euclidean distance as a cost function; Adopting an adaptive threshold algorithm to dynamically adjust the similarity judgment threshold, the adaptive threshold algorithm is based on the statistical distribution of historical data and time factors; If the similarity is lower than the similarity judgment threshold, it is judged that the user has abnormal electricity usage behavior; Trigger abnormal power consumption warning, record abnormal time period, abnormal degree and abnormal type; The abnormal power consumption warning is pushed to the monitoring module in JSON format, and the JSON format includes user ID, abnormal time, abnormal type and abnormal degree fields.

6. The method according to claim 1, wherein: The method of obtaining the historical multi-granular electricity consumption data of the warning user, judging whether the user has long-term low-power consumption behavior through time series decomposition and pattern matching, identifying the suspected economic electricity theft user in combination with the mutation characteristics of the load curve, and matching the corresponding electricity theft disguise pattern according to the electricity consumption behavior characteristics, includes: Acquire historical multi-granularity electricity consumption data of the warning user, wherein the multi-granularity electricity consumption data includes minute-level data, hour-level data, and day-level data; Decomposing the multi-granularity electricity consumption data using seasonality and trend decomposition methods to obtain trend terms, seasonal terms and residual terms; Performing pattern matching on the trend item through a dynamic time warping algorithm and comparing it with a preset low-power consumption mode, wherein the preset low-power consumption mode includes a continuous decline type, a periodic valley type, and a sudden drop stable type; If the similarity of the pattern matching exceeds a preset similarity threshold, it is determined that the user has a long-term and continuous low-power consumption behavior; Perform mutation characteristic analysis on the load curve of the warning user, and use wavelet transform method to detect singular points of the load curve; If the mutation amplitude of the singular point exceeds the preset proportion of the user's average load, and the frequency is greater than the preset number of times, combined with the long-term low-power consumption behavior judgment result, it is identified as a suspected economic electricity theft user; According to the electricity usage behavior characteristics of the suspected economic electricity theft user, the most similar electricity theft disguise pattern is matched from a pre-established electricity theft disguise pattern library, wherein the electricity theft disguise pattern library also includes abnormal patterns of short-term sudden increase, long-term small fluctuation, periodic anomaly and peak anomaly; The cosine similarity algorithm is used to calculate the matching degree, and the pattern with the highest matching degree is selected as the electricity theft disguise pattern of the suspected economic electricity theft user.

7. The method according to claim 6, wherein: The method of obtaining multi-granular electricity consumption data of other users in the area where the suspected economic electricity theft user is located and in the adjacent area, calculating the group electricity theft risk index in the area where the suspected economic electricity theft user is located, identifying the spatial distribution characteristics of the electricity theft users, and comparing them with the typical electricity theft disguise patterns of the group in the area where the suspected economic electricity theft user is located, and judging the regional electricity theft risk includes: Obtain multi-granularity electricity consumption data of other users in the area where the suspected economic electricity theft user is located and in adjacent areas, wherein the multi-granularity electricity consumption data includes minute-level data, hour-level data, and day-level data; constructing a multidimensional dataset including power consumption, power factor, harmonic content and geographic coordinates according to the multi-granularity power consumption data, wherein each user record in the multidimensional dataset includes historical data of a preset number of days; Using a local anomaly factor algorithm to calculate an anomaly score for each user in the multidimensional data set, the anomaly score being obtained based on a preset K nearest neighbor parameter and a preset anomaly score threshold; According to the abnormal scores, the proportion of abnormal users is counted, and the group electricity theft risk index of the area is calculated in combination with the user density and geographical distribution; Calculating the spatial distribution characteristics of the abnormal users by using a spatial autocorrelation analysis method, wherein the spatial distribution characteristics include a global Moran index and a local Gittes index; Identify the clustering areas and hotspot distribution of electricity theft users according to the spatial distribution characteristics, wherein the clustering areas and hotspot distribution are based on a preset spatial weight matrix and significance level; Extracting typical electricity theft camouflage pattern characteristics of the regional group from the historical data, wherein the typical electricity theft camouflage pattern characteristics include power consumption change rate, power factor fluctuation and abnormal harmonic content; Based on the preset window size, step size and similarity threshold, the dynamic time warping algorithm is used to calculate the similarity between each user's electricity usage behavior and the typical electricity theft disguise pattern characteristics; Combining the group electricity theft risk index and the spatial distribution characteristics, a weighted average method is used to comprehensively judge the regional electricity theft risk.

Images