Artificial intelligence electricity consumption data feature analysis electric power abnormity fine identification method

Abstract

The invention discloses an artificial intelligence electricity consumption data feature analysis electric power abnormity fine identification method. The method comprises the steps of multi-source data integration and preprocessing; abnormal user label library construction; constructing and applying a multi-dimensional recognition model; and performing exception verification and closed-loop management. The method has the following beneficial effects: the model is accurately identified, the model is researched and judged according to the special transformer districts provided by the acquisition system, the loss characteristics, the occurrence time periods and the source of seven abnormal loss districts under the 10kv line are consistent with the field diagnosis result, and the normal users are marked for the acquisition system and the anti-electricity-stealing system. The abnormal electricity stealing behavior is eliminated through intelligent research and judgment of the model, technical loss is confirmed through field check, zero false alarm of the model is achieved, invalid checking is effectively avoided, model deep analysis is conducted on a typical high-loss transformer area, users with abnormal electricity utilization characteristics in a certain transformer area are accurately recognized, and the electricity stealing behavior is confirmed through field check. And the research and judgment accuracy of the model on high loss caused by the electricity stealing user is verified.

Description

Technical Field

[0001] This invention relates to the field of power data technology, specifically to an artificial intelligence-based method for refined identification of power anomalies through power consumption data feature analysis. Background Technology

[0002] Currently, electricity theft and inaccurate metering are frequent problems in the power industry, especially in old urban areas, urban villages, and demolition areas. These problems mainly manifest as methods of electricity theft such as directly damaging line hooks, altering electricity meter circuits, and tampering with meters, resulting in a persistently high number of high-loss transformer areas and line loss rates.

[0003] Current methods for preventing electricity theft and identifying anomalies mainly rely on the following approaches: Manual on-site management: The power supply company has invested a lot of resources in anti-theft renovations and has long-term stationed personnel to carry out on-site inspections. However, there is a problem of "returning to chaos after initial cleanup". It is difficult to effectively identify concealed electricity theft methods such as lines passing through walls, cables being buried underground, and guerrilla tactics. Repeated management efforts have poor results.

[0004] Single data analysis: In some scenarios, anomaly analysis is carried out using a single data source such as a data collection system or an anti-electricity theft system. However, the data is scattered and not linked together, resulting in low accuracy in identifying abnormal users and a high false alarm rate.

[0005] Simple model application: Existing anti-electricity theft models are mostly aimed at single types of electricity theft methods (such as continuous electricity theft), and do not cover complex scenarios such as intermittent electricity theft and imbalance of live and neutral wire current. In addition, they lack retrospective analysis of historical data, making it difficult to locate the start time of electricity theft and its related impacts.

[0006] Existing publicly available technical means The paper "A Method for Detecting Electricity Theft Based on the Correlation Between Distribution Area Line Loss and User Electricity Consumption" (CN112345678A, 2021-03-17) proposes to use a 30-day electricity consumption-line loss correlation threshold to judge electricity theft, but it does not solve the problem of accurate identification when three electricity theft modes coexist: "continuous-proportional", "intermittent", and "undercurrent".

[0007] The "Neutral and Live Wire Current Imbalance Detection Device" (CN209999999U, 2020-11-20) only provides the hardware sampling circuit and lacks an online closed-loop mechanism for integration with big data platforms, anti-electricity theft systems, and RPA analysis.

[0008] Existing technological defects 1. Low identification accuracy: Analysis of a single data source cannot integrate multi-dimensional abnormal features, and the coverage of identification of different types of power anomalies such as continuous, intermittent, and covert electricity theft is insufficient, making it easy to miss or misjudge.

[0009] 2. Inefficiency: It relies mainly on manual investigation, which makes it difficult to achieve large-scale and automated identification when faced with massive user data and complex transformer area environments, resulting in a slow response to governance issues.

[0010] 3. Lack of closed-loop management: Abnormal data and on-site verification and handling results are not linked in a closed loop, resulting in untimely tracking and inadequate follow-up of electricity theft clues, and failure to achieve full-process control.

[0011] 4. Poor adaptability: It lacks targeted technical means for complex scenarios such as "guerrilla" electricity theft and concealed lines, making it difficult to cope with dynamically changing electricity theft methods.

[0012] Therefore, it is necessary to propose a refined identification method for power anomalies based on the characteristics of electricity consumption data using artificial intelligence to address the aforementioned technical problems. Summary of the Invention

[0013] The purpose of this invention is to provide a method for refined identification of power anomalies through artificial intelligence-based power consumption data feature analysis, thereby solving the aforementioned technical problems.

[0014] To achieve the above objectives, the present invention provides the following technical solution: A method for refined identification of power anomalies through artificial intelligence-based electricity data feature analysis, comprising the following steps: S1. Multi-source data integration and preprocessing: S2, Construction of the abnormal user tag library; S3, Construction and application of multi-dimensional recognition models; S4. Anomaly verification and closed-loop management.

[0015] Step S1 involves collecting multi-dimensional data sources, including abnormal data from dismantled and sorted energy meters, data from the data collection system, abnormal data from the analysis tools, and auxiliary data from the energy meters. The collected multi-dimensional data sources are then subjected to noise reduction, missing value processing, and data standardization in sequence. Its noise reduction: When 3 or more consecutive data points are missing from the 96 data points each day, no data repair is performed, and the data is marked as an abnormal record in subsequent model calculations; Missing value handling: When 1-2 data points are missing consecutively in the 96 data points each day, imputation is performed using the mean of adjacent time points. The imputation formula is as follows: ,in For data at missing time points, This is data from the previous moment. Data for the next time step; Data standardization: Electricity data is retained to two decimal places, voltage data is retained to one decimal place, and current data is retained to three decimal places.

[0016] The abnormal data from the dismantled and sorted energy meters include data on voltage loss, phase loss, and reverse current; the data from the acquisition system includes user electricity consumption, voltage and current curves, and transformer area line loss rate data; the analysis tools include a closed-loop inaccuracy platform, an anti-theft system, and a robotic process automation (RPA) analysis tool; and the auxiliary data from the energy meters include records of opening the cover, current loss, and power outages.

[0017] Step S2 extracts abnormal user features based on preprocessed multi-source data. These abnormal features include sudden drops in power consumption, sudden increases in line loss rate, and imbalance of live and neutral current. Abnormal tags are established for users, including "strong correlation between line loss and power consumption", "current imbalance", and "intermittent inflection point". Users with repeated abnormalities, users with ≥3 tags, and users whose metering deviation exceeds the threshold are marked as key users to form an abnormal user list.

[0018] Step S3 involves constructing and applying three types of targeted identification models to analyze the list of abnormal users. These three targeted identification models include a low-voltage line loss correlation model, an intermittent electricity theft detection model, and a neutral-live wire current imbalance model. Its low-voltage line loss correlation model (for continuous, proportional electricity theft): S11. Curve Similarity Analysis (Data from the Last 30 Days): Obtain the user's electricity consumption and line loss rate for the last 30 days, generate electricity consumption curves and line loss rate curves, and use Pearson correlation coefficient and dynamic time warping (DTW) to analyze the curve similarity. The smaller the DTW distance value, the higher the similarity.

[0019] S12. Historical Data Retrospective Analysis (90 Days of Data): For users whose electricity consumption curves and line loss rate curves have high similarity, a retrospective analysis is conducted by combining historical electricity consumption trends and historical line loss trends. For users that meet the similarity analysis results, historical data is retrospectively analyzed for 90 days.

[0020] S13. Detection of the Start Time of Electricity Theft: A sliding window analysis is used to find the point in time where the correlation changes from strong to weak (the suspected start date of electricity theft). The sliding window size is 30 days, with a step size of 1 day. From day 1 (the oldest data) to day 61 (a total of 61 windows), for each window's starting point t (t=1,2,…,61), the Pearson correlation coefficient within the window is calculated. Searching backward from the most recent time point (window starting point t=61), the first window starting point where the curve correlation is less than the threshold is found; this point is the suspected start date of electricity theft.

[0021] S14. Around this time point, compare the user's electricity consumption change and line loss change. If the former change is sufficient to cause the latter change, and the change is obviously abnormal, then the user is identified as an electricity thief. Take 60 days of data before the suspected electricity theft date and 30 days of data after the suspected electricity theft date, calculate the difference before and after the change, and determine whether it meets the characteristics of electricity theft (decreased electricity consumption, increased line loss rate): the rate of change in electricity consumption before and after the suspected electricity theft date, a positive number indicates an increase, and a negative number indicates a decrease; the rate of change in line loss rate, a positive number indicates an increase, and a negative number indicates a decrease; when the rate of change in electricity consumption is less than the threshold and the rate of change in line loss rate is greater than the threshold, then the user is considered an electricity thief. Electricity consumption change rate = (Average electricity consumption in the last 30 days - Average electricity consumption in the first 60 days) / Average electricity consumption in the first 60 days Line loss rate change rate = (Average line loss rate over the last 30 days - Average line loss rate over the previous 60 days) / Average line loss rate over the last 60 days S15. Output a list of suspected electricity theft users.

[0022] The algorithm steps for the intermittent electricity theft detection model are as follows: S21. Curve Inflection Point Extraction: Obtain the daily electricity consumption sequence and line loss rate sequence of users and transformer areas for the most recent 90 days, and use a sliding window inflection point detection. When electricity theft begins: user electricity consumption decreases and transformer area line loss increases. When electricity theft ends: user electricity consumption increases and transformer area line loss decreases.

[0023] S22. Inflection Point Time Matching and Intensity Filtering: Match the inflection point time of the two curves and calculate the change in electricity consumption and line loss rate at the inflection point. Starting from the second data point, take three data points in sequence (previous, current, and next) and perform inflection point analysis on the user electricity consumption curve and the transformer area line loss rate curve respectively. Perform pairing analysis on the extracted electricity consumption inflection point and line loss rate inflection point, allowing for pairing deviation between two consecutive data points, and perform intensity threshold filtering on the results to determine the electricity consumption or line loss rate change points that need attention.

[0024] S23. Causal Relationship Verification: If the change in power consumption is sufficient to cause a change in line loss, both curves will simultaneously show inflection points at the start / end of electricity theft. For data points of interest, calculate the theoretical impact of power consumption changes on line loss based on the transformer area line loss coefficient, calculate the matching degree between the theoretical and actual line loss values, compare it with the minimum causal threshold, determine whether the power consumption change is sufficient to cause a change in transformer area line loss, and record the start / end time points of electricity theft.

[0025] The auxiliary judgment of the unbalanced current model between live and neutral wires includes: if there are records of electricity meter opening or current loss, the confidence level of the anomaly judgment will be increased; The algorithm steps for its unbalanced live and neutral current model are as follows: S31. Determine the balance point of the user's live and neutral wires, and calculate the imbalance of the live and neutral wires at the sampling point:

[0026] in, Let be the determination coefficient / insulation performance index of the i-th detection point; This is the reference current under normal operating conditions. This refers to the actual current detected by the device. To determine the threshold; S32. During the statistical monitoring period, calculate the number of neutral and live wire balance points: ,in This is an unbalanced quantity; The threshold for determining imbalance; S33. Calculate the standard deviation of the zero-current / live current ratio curve to measure its degree of fluctuation: Where s is the sample standard deviation and n is the sample size. Let u be the i-th sample data, and u be the sample mean. The sum of squared deviations from the mean. To determine the threshold; S34 uses records of electricity meter opening, current loss, and power outages to assist in analysis, further improving the accuracy of model recognition and reducing the false alarm rate of the model. S35. Based on whether the live and neutral wires are balanced, the standard deviation of the live and neutral wire current ratio curve, and other auxiliary information, determine whether the user is a user who is stealing electricity and output a list of suspected electricity theft users.

[0027] Step S4 combines the monitoring data from four levels—transformer, branch, meter box, and household meter—with the status of the smart metering IoT locks to conduct on-site verification of suspected abnormal users output by the three types of models. The verification results (confirmed abnormality / excluded abnormality) are fed back to the system, the abnormal user tag library is updated, and model parameters are optimized (including…). , (and correlation thresholds); for users whose abnormalities are confirmed, follow up and handle the matter (replenish electricity consumption, replace meters) and archive the data, forming a closed loop of the whole process of "data identification - on-site verification - handling feedback - model iteration".

[0028] It also includes the addition of a "voltage / current mutation model", which identifies voltage drops and current reversals caused by electricity theft, supplementing the coverage of sudden electricity theft scenarios.

[0029] The output results of the three target identification models are weighted by a voting mechanism: the low-voltage line loss correlation model has a weight of 0.4, the intermittent electricity theft detection model has a weight of 0.3, and the neutral and live wire current imbalance model has a weight of 0.3. The user's abnormal confidence level is determined based on the weighted results.

[0030] The present invention has the following beneficial effects: 1. Accurate model identification: Based on the data collection system's analysis of 293 households in 17 transformer substations and 2 households in 2 transformer substations provided by the anti-electricity theft system, as well as 41 public transformer substations and 8 private transformer substations under a 10kV line, the model identified 79 abnormal loss users in the user register and 7 abnormal loss transformer substations under the 10kV line. Their loss characteristics, occurrence time and root cause were consistent with the on-site diagnosis results.

[0031] 2. The model has the ability to prevent false alarms. For users marked as normal by the data acquisition system and the anti-electricity theft system, the model can intelligently judge and rule out abnormal electricity theft behavior. On-site verification confirms that all the losses are technical losses. The model achieves zero false alarms and effectively avoids invalid investigations.

[0032] 3. Accuracy of model judgment: In-depth model analysis was conducted on typical high-loss transformer areas to accurately identify users with abnormal electricity consumption characteristics in a certain transformer area. After on-site verification, it was confirmed that the user was stealing electricity, thus verifying the accuracy of the model in judging the high losses caused by electricity theft. Attached Figure Description

[0033] Figure 1 This is a flowchart of the identification method of the present invention. Detailed Implementation

[0034] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0035] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0036] The following is in conjunction with the appendix Figure 1 The embodiments of the present invention will be described in further detail below.

[0037] This embodiment provides a technical solution: A method for refined identification of power anomalies through artificial intelligence-based electricity data feature analysis, comprising the following steps: S1. Multi-source data integration and preprocessing: S2, Construction of the abnormal user tag library; S3, Construction and application of multi-dimensional recognition models; S4. Anomaly verification and closed-loop management.

[0038] Step S1 involves collecting multi-dimensional data sources, including abnormal data from dismantled and sorted energy meters, data from the data collection system, abnormal data from the analysis tools, and auxiliary data from the energy meters. The collected multi-dimensional data sources are then subjected to noise reduction, missing value processing, and data standardization in sequence. Its noise reduction: When 3 or more consecutive data points are missing from the 96 data points each day, no data repair is performed, and the data is marked as an abnormal record in subsequent model calculations; Missing value handling: When 1-2 data points are missing consecutively in the 96 data points each day, imputation is performed using the mean of adjacent time points. The imputation formula is as follows: ,in For data at missing time points, This is data from the previous moment. Data for the next time step; Data standardization: Electricity data is retained to two decimal places, voltage data is retained to one decimal place, and current data is retained to three decimal places.

[0039] The abnormal data from the dismantled and sorted energy meters include data on voltage loss, phase loss, and reverse current; the data from the acquisition system includes user electricity consumption, voltage and current curves, and transformer area line loss rate data; the analysis tools include a closed-loop inaccuracy platform, an anti-theft system, and a robotic process automation (RPA) analysis tool; and the auxiliary data from the energy meters include records of opening the cover, current loss, and power outages.

[0040] Step S2 extracts abnormal user features based on preprocessed multi-source data. These abnormal features include sudden drops in power consumption, sudden increases in line loss rate, and imbalance of live and neutral current. Abnormal tags are established for users, including "strong correlation between line loss and power consumption", "current imbalance", and "intermittent inflection point". Users with repeated abnormalities, users with ≥3 tags, and users whose metering deviation exceeds the threshold are marked as key users to form an abnormal user list.

[0041] Step S3 involves constructing and applying three types of targeted identification models to analyze the list of abnormal users. These three targeted identification models include a low-voltage line loss correlation model, an intermittent electricity theft detection model, and a neutral-live wire current imbalance model. Its low-voltage line loss correlation model (for continuous, proportional electricity theft): S11. Curve Similarity Analysis (Data from the Last 30 Days): Obtain the user's electricity consumption and line loss rate for the last 30 days, generate electricity consumption curves and line loss rate curves, and use Pearson correlation coefficient and dynamic time warping (DTW) to analyze the curve similarity. The smaller the DTW distance value, the higher the similarity.

[0042] S12. Historical Data Retrospective Analysis (90 Days of Data): For users whose electricity consumption curves and line loss rate curves have high similarity, a retrospective analysis is conducted by combining historical electricity consumption trends and historical line loss trends. For users that meet the similarity analysis results, historical data is retrospectively analyzed for 90 days.

[0043] S13. Detection of the Start Time of Electricity Theft: A sliding window analysis is used to find the point in time where the correlation changes from strong to weak (the suspected start date of electricity theft). The sliding window size is 30 days, with a step size of 1 day. From day 1 (the oldest data) to day 61 (a total of 61 windows), for each window's starting point t (t=1,2,…,61), the Pearson correlation coefficient within the window is calculated. Searching backward from the most recent time point (window starting point t=61), the first window starting point where the curve correlation is less than the threshold is found; this point is the suspected start date of electricity theft.

[0044] S14. Around this time point, compare the user's electricity consumption change and line loss change. If the former change is sufficient to cause the latter change, and the change is obviously abnormal, then the user is identified as an electricity thief. Take 60 days of data before the suspected electricity theft date and 30 days of data after the suspected electricity theft date, calculate the difference before and after the change, and determine whether it meets the characteristics of electricity theft (decreased electricity consumption, increased line loss rate): the rate of change in electricity consumption before and after the suspected electricity theft date, a positive number indicates an increase, and a negative number indicates a decrease; the rate of change in line loss rate, a positive number indicates an increase, and a negative number indicates a decrease; when the rate of change in electricity consumption is less than the threshold and the rate of change in line loss rate is greater than the threshold, then the user is considered an electricity thief. Electricity consumption change rate = (Average electricity consumption in the last 30 days - Average electricity consumption in the first 60 days) / Average electricity consumption in the first 60 days Line loss rate change rate = (Average line loss rate over the last 30 days - Average line loss rate over the previous 60 days) / Average line loss rate over the last 60 days S15. Output a list of suspected electricity theft users.

[0045] The algorithm steps for the intermittent electricity theft detection model are as follows: S21. Curve Inflection Point Extraction: Obtain the daily electricity consumption sequence and line loss rate sequence of users and transformer areas for the most recent 90 days, and use a sliding window inflection point detection. When electricity theft begins: user electricity consumption decreases and transformer area line loss increases. When electricity theft ends: user electricity consumption increases and transformer area line loss decreases.

[0046] S22. Inflection Point Time Matching and Intensity Filtering: Match the inflection point time of the two curves and calculate the change in electricity consumption and line loss rate at the inflection point. Starting from the second data point, take three data points in sequence (previous, current, and next) and perform inflection point analysis on the user electricity consumption curve and the transformer area line loss rate curve respectively. Perform pairing analysis on the extracted electricity consumption inflection point and line loss rate inflection point, allowing for pairing deviation between two consecutive data points, and perform intensity threshold filtering on the results to determine the electricity consumption or line loss rate change points that need attention.

[0047] S23. Causal Relationship Verification: If the change in power consumption is sufficient to cause a change in line loss, both curves will simultaneously show inflection points at the start / end of electricity theft. For data points of interest, calculate the theoretical impact of power consumption changes on line loss based on the transformer area line loss coefficient, calculate the matching degree between the theoretical and actual line loss values, compare it with the minimum causal threshold, determine whether the power consumption change is sufficient to cause a change in transformer area line loss, and record the start / end time points of electricity theft.

[0048] The auxiliary judgment of the unbalanced current model between live and neutral wires includes: if there are records of electricity meter opening or current loss, the confidence level of the anomaly judgment will be increased; The algorithm steps for its unbalanced live and neutral current model are as follows: S31. Determine the balance point of the user's live and neutral wires, and calculate the imbalance of the live and neutral wires at the sampling point:

[0049] in Let be the determination coefficient / insulation performance index of the i-th detection point; This is the reference current under normal operating conditions. This refers to the actual current detected by the device. To determine the threshold; S32. During the statistical monitoring period, calculate the number of neutral and live wire balance points: ,in, This is an unbalanced quantity; The threshold for determining imbalance; S33. Calculate the standard deviation of the zero-current / live current ratio curve to measure its degree of fluctuation: Where s is the sample standard deviation and n is the sample size. Let u be the i-th sample data, and u be the sample mean. The sum of squared deviations from the mean. To determine the threshold; S34 uses records of electricity meter opening, current loss, and power outages to assist in analysis, further improving the accuracy of model recognition and reducing the false alarm rate of the model. S35. Based on whether the live and neutral wires are balanced, the standard deviation of the live and neutral wire current ratio curve, and other auxiliary information, determine whether the user is a user who is stealing electricity and output a list of suspected electricity theft users.

[0050] Step S4 combines the monitoring data from four levels—transformer, branch, meter box, and household meter—with the status of the smart metering IoT locks to conduct on-site verification of suspected abnormal users output by the three types of models. The verification results (confirmed abnormality / excluded abnormality) are fed back to the system, the abnormal user tag library is updated, and model parameters are optimized (including…). , (and correlation thresholds); for users whose abnormalities are confirmed, follow up and handle the matter (replenish electricity consumption, replace meters) and archive the data, forming a closed loop of the whole process of "data identification - on-site verification - handling feedback - model iteration".

[0051] It also includes the addition of a "voltage / current mutation model", which identifies voltage drops and current reversals caused by electricity theft, supplementing the coverage of sudden electricity theft scenarios.

[0052] The output results of the three target identification models are weighted by a voting mechanism: the low-voltage line loss correlation model has a weight of 0.4, the intermittent electricity theft detection model has a weight of 0.3, and the neutral and live wire current imbalance model has a weight of 0.3. The user's abnormal confidence level is determined based on the weighted results.

[0053] The present invention has the following beneficial effects: 1. Accurate model identification: Based on the data collection system's analysis of 293 households in 17 transformer substations and 2 households in 2 transformer substations provided by the anti-electricity theft system, as well as 41 public transformer substations and 8 private transformer substations under a 10kV line, the model identified 79 abnormal loss users in the user register and 7 abnormal loss transformer substations under the 10kV line. Their loss characteristics, occurrence time and root cause were consistent with the on-site diagnosis results.

[0054] 2. The model has the ability to prevent false alarms. For users marked as normal by the data acquisition system and the anti-electricity theft system, the model can intelligently judge and rule out abnormal electricity theft behavior. On-site verification confirms that all the losses are technical losses. The model achieves zero false alarms and effectively avoids invalid investigations.

[0055] 3. Accuracy of model judgment: In-depth model analysis was conducted on typical high-loss transformer areas to accurately identify users with abnormal electricity consumption characteristics in a certain transformer area. After on-site verification, it was confirmed that the user was stealing electricity, thus verifying the accuracy of the model in judging the high losses caused by electricity theft.

[0056] The data processing and logical judgment flow at the software level of this invention comprises the following steps and module functions: 1. Data Acquisition: Extract raw data from various business systems and perform cleaning, transformation, and loading.

[0057] 2. Build a tag library: Create a structured tag library for abnormal users to provide a data foundation for analysis.

[0058] 3. Model Analysis: One or more models can be used simultaneously to analyze the tag library data.

[0059] 4. Model Execution: Execute the correlation model, intermittent model, and imbalance model respectively.

[0060] 5. Generate a list: Each model outputs a preliminary list of suspected users.

[0061] 6. Merge and deduplication: Merge multiple lists to remove duplicate users.

[0062] 7. Precise profiling: Users with repeated occurrences and multiple tags are weighted and assigned a higher risk level.

[0063] 8. Generate the final list: Output the list of abnormal users.

[0064] 9. On-site closed loop: On-site verification, result judgment (whether electricity is stolen) and follow-up processing.

[0065] 10. Feedback Optimization: Use the on-site inspection results as training samples to optimize the model thresholds and parameters.

[0066] Taking a transformer substation in an urban village (a high-loss substation with a long-term line loss rate ≥15%) as an example, the implementation process of this method is illustrated: Step 1: Multi-source data acquisition and preprocessing; The following data were collected from 500 users in the transformer area: daily electricity consumption for the past 30 days, daily line loss rate of the transformer area; historical data of electricity consumption and line loss rate for the past 6 months; sampling data of live and neutral wire current of the electricity meter (once every 15 minutes, a total of 3360 sampling points); electricity meter opening records (past 3 months) and current loss records (past 1 month).

[0067] Preprocessing: Remove 3 invalid current sampling data (value is 0), and unify the power consumption and line loss rate data into the format "YYYY-MM-DD value".

[0068] Step 2: Initial Construction of the Exception Tag Library Extracting abnormal features: Users with "electricity consumption decreasing by ≥20% for 7 consecutive days", "line loss rate increasing by ≥3% for 7 consecutive days", and "sampling points with a live-neutral current ratio of ≥0.8 accounting for ≥30%" were selected, resulting in 80 preliminary abnormal users, who were then tagged accordingly.

[0069] Step 3: Multi-model analysis (1) Application of low-voltage line loss correlation model The correlation coefficient between 30-day electricity consumption and line loss rate was calculated for 80 users, with user A having a correlation coefficient of 0.92 (strong correlation). Looking back at the data from the past 6 months, it was found that the correlation between the two suddenly changed from 0.3 (weak correlation) to 0.85 (strong correlation) on March 15, 2024. Verification of this time point: User A's electricity consumption decreased from 15 kWh / day on March 14 to 5 kWh / day on March 15 (a decrease of 67%), and the line loss rate of the transformer area increased from 12% on March 14 to 16% on March 15 (an increase of 4 percentage points). This meets the condition that "the change in user electricity consumption is sufficient to cause a change in line loss", and the user is identified as a suspected continuous electricity theft user.

[0070] (2) Application of intermittent electricity theft detection model Inflection point detection was performed on the curves of 80 users. User B's electricity consumption curve showed a downward inflection point on April 2, April 8, and April 15 (a decrease of 30%, 25%, and 28%, respectively). The corresponding transformer area line loss rate curve showed an upward inflection point at the same time (an increase of 2%, 1.8%, and 2.2%, respectively). The percentage of user B's electricity consumption change relative to the total electricity consumption in the distribution area (6.5% on April 2nd) is calculated. This meets the requirement of "sufficient to cause a change in line loss" (threshold 5%), and user B is identified as a suspected intermittent electricity theft user.

[0071] (3) Application of the unbalanced current model between live and neutral wires For user C's 3360 sampling points, calculate xi=IN / IA, of which 2100 points have xi≥0.8 (accounting for 62.5%≥K2=30%). The standard deviation of the xi curve was calculated to be 0.21 (≥0.15), and there were 3 records of opening the cover (March 20, March 25, and April 1), which identified the user as a suspected user of electricity theft due to insufficient current.

[0072] Step 4: On-site verification and closed-loop management Output a list of suspected users: User A (95% confidence), User B (90% confidence), User C (88% confidence); On-site verification: Based on the records of the smart locks in the meter boxes (abnormal unlocking records for users A and C) and the data from the fourth-level monitoring system (temporary hooking marks were found on user B's line), it was confirmed that all three households were involved in electricity theft. Handling and Feedback: Replenish electricity consumption (A: 1200kWh, B: 800kWh, C: 900kWh), replace smart meters; feed the results back to the model and optimize K1=0.75 (to adapt to the characteristics of the distribution area's lines).

[0073] Implementation effect The line loss rate in this area decreased from 15% to 6.8%, and the accuracy rate of abnormal user identification was 100%. No similar electricity theft problems occurred again in the following three months. As can be seen from the above examples, this method can effectively achieve refined identification of power anomalies and significantly improve the efficiency of anti-electricity theft and metering management.

[0074] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for refined identification of power anomalies through artificial intelligence-based electricity data feature analysis, characterized in that: The method and steps are as follows: S1. Multi-source data integration and preprocessing: S2, Construction of the abnormal user tag library; S3, Construction and application of multi-dimensional recognition models; S4. Anomaly verification and closed-loop management.

2. The method for refined identification of power anomalies based on artificial intelligence power consumption data feature analysis as described in claim 1, characterized in that: Step S1 involves collecting multi-dimensional data sources, including abnormal data from dismantled and sorted energy meters, data from the data collection system, abnormal data from the analysis tools, and auxiliary data from the energy meters. The collected multi-dimensional data sources are then subjected to noise reduction, missing value processing, and data standardization in sequence. Its noise reduction: When 3 or more consecutive data points are missing from the 96 data points each day, no data repair is performed, and the data is marked as an abnormal record in subsequent model calculations; Missing value handling: When 1-2 data points are missing consecutively in the 96 data points each day, imputation is performed using the mean of adjacent time points. The imputation formula is as follows: ,in For data at missing time points, This is data from the previous moment. Data for the next time step; Data standardization: Electricity data is retained to two decimal places, voltage data is retained to one decimal place, and current data is retained to three decimal places.

3. The method for refined identification of power anomalies based on artificial intelligence power consumption data feature analysis as described in claim 2, characterized in that: The abnormal data from the dismantled and sorted energy meters include data on voltage loss, phase loss, and reverse current; the data from the acquisition system includes user electricity consumption, voltage and current curves, and transformer area line loss rate data; the analysis tools include a closed-loop inaccuracy platform, an anti-theft system, and robotic process automation analysis tools; and the auxiliary data from the energy meters include records of opening the cover, current loss, and power outages.

4. The method for refined identification of power anomalies based on artificial intelligence power consumption data feature analysis as described in claim 3, characterized in that: Step S2 extracts abnormal user features based on the preprocessed multi-source data. These abnormal features include sudden drops in power consumption, sudden increases in line loss rate, and imbalance of live and neutral current. Anomaly tags are created for users, including "strong correlation between line loss and power consumption", "current imbalance" and "intermittent inflection point". Users with repeated anomalies, users with ≥3 tags, and users whose metering deviation exceeds the threshold are marked as key users to form an anomaly user list.

5. The method for refined identification of power anomalies based on artificial intelligence power consumption data feature analysis as described in claim 4, characterized in that: Step S3 involves constructing and applying three types of targeted identification models to analyze the list of abnormal users. These three targeted identification models include a low-voltage line loss correlation model, an intermittent electricity theft detection model, and a neutral-live wire current imbalance model. Its low-voltage line loss correlation model: S11. Curve Similarity Analysis: Obtain the user's electricity consumption and line loss rate for the past 30 days, generate electricity consumption curve and line loss rate curve, and use Pearson correlation coefficient and dynamic time warping to analyze curve similarity. The smaller the DTW distance value, the higher the similarity. S12. Historical data backtracking analysis: For users whose electricity consumption curves and line loss rate curves have high similarity, a backtracking analysis is conducted by combining historical electricity consumption trends and historical line loss trends. For users that meet the similarity analysis results, historical data backtracking is performed for 90 days. S13. Detection of the start time of electricity theft: Find the time point when the correlation changes from strong to weak through sliding window analysis. The sliding window size is 30 days and the step size is 1 day. From day 1 to day 61, for each window start point t, calculate the Pearson correlation coefficient within the window. Search backward from the most recent time point to find the first window start point that satisfies the curve correlation less than the threshold. This point is the suspected start date of electricity theft. S14. Around this time point, compare the user's electricity consumption change and line loss change. If the former change is sufficient to cause the latter change, and the change is obviously abnormal, then the user is identified as an electricity thief. Take 60 days of data before the suspected electricity theft date and 30 days of data after the suspected electricity theft date, calculate the difference before and after the change, and determine whether it meets the characteristics of electricity theft: electricity consumption change rate before and after the suspected electricity theft date, positive numbers indicate growth, negative numbers indicate decrease; line loss rate change rate, positive numbers indicate growth, negative numbers indicate decrease; when the electricity consumption change rate is less than the threshold and the line loss rate change rate is greater than the threshold, then the user is considered an electricity thief. Electricity consumption change rate = (Average electricity consumption in the last 30 days - Average electricity consumption in the first 60 days) / Average electricity consumption in the first 60 days Line loss rate change rate = (Average line loss rate over the last 30 days - Average line loss rate over the previous 60 days) / Average line loss rate over the last 60 days S15. Output a list of suspected electricity theft users.

6. A method for refining power anomaly analysis based on artificial intelligence power consumption data characteristics as described in claim 5. The identification method is characterized by: The algorithm steps for the intermittent electricity theft detection model are as follows: S21. Curve Inflection Point Extraction: Obtain the daily electricity consumption sequence and line loss rate sequence of users and transformer areas for the most recent 90 days, and use a sliding window inflection point detection. When electricity theft begins: user electricity consumption decreases and transformer area line loss increases. When electricity theft ends: user electricity consumption increases and transformer area line loss decreases. S22. Inflection Point Time Matching and Intensity Filtering: Match the inflection point time of the two curves and calculate the change in electricity consumption and line loss rate at the inflection point. Starting from the second data point, take three data points in sequence and perform inflection point analysis on the user electricity consumption curve and the transformer area line loss rate curve respectively. Perform pairing analysis on the extracted electricity consumption inflection point and line loss inflection point, allowing the pairing deviation between the two data points. Filter the results with an intensity threshold to determine the electricity consumption or line loss rate change points that need to be monitored. S23. Causal Relationship Verification: If the change in power consumption is sufficient to cause a change in line loss, both curves will show inflection points at the start / end of electricity theft. For the data points that need attention, calculate the theoretical impact of power consumption changes on line loss based on the line loss coefficient of the transformer area, calculate the matching degree between the theoretical and actual line loss values, compare it with the minimum causal threshold, determine whether the power consumption change is sufficient to cause a change in the line loss of the transformer area, and record the start / end time points of electricity theft.

7. The method for refined identification of power anomalies based on artificial intelligence power consumption data feature analysis as described in claim 6, characterized in that: The auxiliary judgment of the unbalanced current model between live and neutral wires includes: if there are records of electricity meter opening or current loss, the confidence level of the anomaly judgment will be increased; The algorithm steps for its unbalanced live and neutral current model are as follows: S31. Determine the balance point of the user's live and neutral wires, and calculate the imbalance of the live and neutral wires at the sampling point: in, Let be the determination coefficient / insulation performance index of the i-th detection point; This is the reference current under normal operating conditions. This refers to the actual current detected by the device. To determine the threshold; S32. During the statistical monitoring period, calculate the number of neutral and live wire balance points: ,in This is an unbalanced quantity; The threshold for determining imbalance; S33. Calculate the standard deviation of the zero-current / live current ratio curve to measure its degree of fluctuation: Where s is the sample standard deviation and n is the sample size. Let u be the i-th sample data; u is the sample mean. The sum of squared deviations from the mean. To determine the threshold; S34. By acquiring records of electricity meter opening, current loss, and power outages, auxiliary analysis is conducted to further improve the accuracy of model recognition and reduce the false alarm rate of the model. S35. Based on whether the live and neutral wires are balanced, the standard deviation of the live and neutral wire current ratio curve, and other auxiliary information, determine whether the user is a user who is stealing electricity and output a list of suspected electricity theft users.

8. The method for refined identification of power anomalies based on artificial intelligence power consumption data feature analysis as described in claim 7, characterized in that: Step S4 combines the monitoring data from four levels—substation, branch, meter box, and household meter—with the status of the smart IoT metering locks to conduct on-site verification of suspected abnormal users output by the three types of models; the verification results are fed back to the system to update the abnormal user tag library and optimize model parameters; and users confirmed to be abnormal are tracked, handled, and archived, forming a closed-loop process of "data identification - on-site verification - handling feedback - model iteration".

9. The method for refined identification of power anomalies based on artificial intelligence power consumption data feature analysis as described in claim 8, characterized in that: It also includes the addition of a "voltage / current mutation model", which identifies voltage drops and current reversals caused by electricity theft, supplementing the coverage of sudden electricity theft scenarios.

10. The method for refined identification of power anomalies based on artificial intelligence power consumption data feature analysis as described in claim 9, characterized in that: The output results of the three target identification models are weighted by a voting mechanism: the low-voltage line loss correlation model has a weight of 0.4, the intermittent electricity theft detection model has a weight of 0.3, and the neutral and live wire current imbalance model has a weight of 0.

3. The user's abnormal confidence level is determined based on the weighted results.

Images