Box-level line loss decomposition and abnormal metering coordination identification method for low-voltage electric energy metering box

Abstract

The present application belongs to the technical field of abnormal perception of metering equipment, and relates to a box-level line loss decomposition and abnormal metering collaborative identification method for a low-voltage electric energy metering box, which comprises the following steps: collecting box-level total quantity data and each meter position sub-metering data, performing time alignment, effective meter position set screening and number shortage correction, and constructing a box-level conservation residual; then, the box-level conservation residual is decomposed into four residual components: a persistent residual, an event coupling residual, an inter-phase imbalance residual and a number shortage compensation residual; further, residual attribution coefficients, event coupling coefficients, phase deviation coefficients, number shortage influence coefficients and zero sequence correlation coefficients of each meter position are calculated to form a meter position abnormal evidence vector; finally, a lightweight abnormal collaborative identification model is used, a fixed dimension feature vector including the residual components and the meter position abnormal evidence vector is input, and an abnormal category probability, a doubtful meter position ranking and a processing suggestion are output, so that fine identification and collaborative attribution of abnormal line loss in a low-voltage electric energy metering box scenario are realized.

Description

Technical Field

[0001] This invention belongs to the field of abnormal sensing and intelligent operation and maintenance technology of low-voltage power metering equipment, and relates to a method for box-level line loss decomposition and abnormal metering collaborative identification of low-voltage power metering boxes. Background Technology

[0002] During long-term operation, low-voltage energy metering boxes often experience discrepancies between the total box reading and the individual meter readings. Causes of this discrepancy may include communication errors, meter malfunctions, wiring abnormalities, bypass power supply, abnormal power loss, and short-term fluctuations in operating conditions. Simply relying on the difference between the total box reading and the individual meter readings for alarm purposes is insufficient to further differentiate the source of the anomaly, easily leading to false alarms, missed alarms, and low troubleshooting efficiency.

[0003] Existing methods for analyzing abnormal line losses typically employ single difference thresholds, manual judgment based on experience, or isolated anomaly detection methods for single meter positions. These methods fail to fully utilize conservation constraints at the metering box level, the correlation between multiple meter positions, three-phase imbalance, and the impact of communication data loss and re-reading on box-level residuals. Consequently, they have limited ability to distinguish between different types of line loss anomalies caused by factors such as bypass power supply, wiring abnormalities, communication data loss, and meter malfunctions.

[0004] Furthermore, abnormal metering problems in low-voltage power metering boxes often exhibit multi-timescale characteristics. Communication errors are more likely to result in abrupt residual changes on a short timescale, meter inaccuracies are more likely to lead to persistent deviations, and bypass power draws or abnormal losses may manifest as a persistent conservation mismatch that varies with load conditions. If the box-level residuals cannot be decomposed in a calculable and structured manner, and the decomposition results cannot be collaboratively identified with the meter position-level evidence vectors, it will be difficult to generate stable, interpretable, and easily executable analysis results for operation and maintenance. Summary of the Invention

[0005] To address the shortcomings of existing methods for analyzing abnormal line losses in low-voltage energy metering boxes, which rely solely on simple differential alarms, struggle to differentiate between multiple anomaly sources, and lack interpretability in complex multi-meter-position operating conditions, this invention proposes a box-level line loss decomposition and abnormal metering collaborative identification method for low-voltage energy metering boxes. This method is applicable to single-phase multi-meter-position metering boxes, three-phase multi-meter-position metering boxes, and mixed-meter-position low-voltage metering equipment in scenarios involving inconsistencies between total and individual meters, abnormal line loss identification, suspicious meter position sorting, and anomaly type differentiation, providing online analysis and maintenance assistance. The specific technical solution is as follows:

[0006] A method for co-identifying box-level line loss decomposition and abnormal metering in low-voltage power metering boxes includes:

[0007] First, collect the total data at the box level and the sub-measurement data at each station, then perform time alignment, screening of the effective station set and correction of missing data, and construct the box-level conservation residuals;

[0008] Then, the box-level conserved residuals are decomposed into four residual components: persistence residuals, event coupling residuals, phase imbalance residuals, and missing number compensation residuals;

[0009] Further calculations are made of the residual attribution coefficient, event coupling coefficient, phase deviation coefficient, missing number influence coefficient, and zero-order correlation coefficient for each epitope to construct an epitope anomaly evidence vector;

[0010] Finally, a lightweight anomaly collaborative identification model is used. The input consists of a fixed-dimensional feature vector including residual components and anomaly evidence vectors of table positions. The output includes anomaly category probability, suspicious table position ranking, and processing suggestions, thereby achieving fine identification and collaborative attribution of abnormal line losses in low-voltage power metering box scenarios.

[0011] Furthermore, the total data at the box level includes: total active power. Total electrical energy Phase voltage and current and zero-sequence current The meter readings for each meter position include: voltage. Current Active power Power factor Electricity Timestamp, communication integrity rate Supplementing and copying marks and the phase to which the epitope belongs Among them, the supplementary copying marks The data used to indicate that the corresponding sampling point was obtained by a copying strategy or by correcting missing data is different from real-time acquired data.

[0012] Furthermore, the aforementioned time alignment, effective tabletop set filtering, and missing number correction, and the construction of bin-level conserved residuals, specifically involve: according to a unified sampling time... Time alignment is performed on bin-level and table-level data; for sampling time... If table position Time tolerance If a real-time sampled value exists in memory, then that real-time sampled value is used as the reference power, i.e., the aligned position power. and table Add to the set of valid epitopes If the sampling time If the sample is located between two adjacent valid sampling points, linear interpolation is used to obtain the result. and table join in If table position At sampling time There are no valid real-time sampled values ​​and Then the reference power before the missing number correction will be... Set as , and according to After performing missing number correction, the table position will be... join in ,in , and These are the two most recent valid sampling time points, This represents the upper limit of the trend coefficient.

[0013] like And exist ,but ;

[0014] like And it does not exist. Then the table will be Mark as invalid table and from Excluded from the middle;

[0015] According to the expression Construct box-level conserved residuals.

[0016] Furthermore, the persistent residual is calculated from the box-level conserved residual using the sliding window mean, and the expression is as follows: ,in The duration is the length of the persistent window.

[0017] The event coupling residual is obtained through event window determination, and its expression is: ; among them This is an event window indicator, at the sampling time. Event window length before and after Within the scope, there are power outage events, maintenance operation events, and communication restoration events, or events that meet the following conditions. For load mutation events, the event window indicator value ,otherwise , To prevent tiny positive numbers with a denominator of zero, Threshold for determining load mutations;

[0018] The interphase unbalanced residuals are obtained through unbalance constraints, and the expression is:

[0019]

[0020] in The unbalanced decomposition coefficient, Voltage imbalance;

[0021] Based on the aforementioned persistent residuals, event-coupled residuals, and interphase imbalance residuals, the missing number compensation residual is: .

[0022] Furthermore, when the metering box is a three-phase metering box, the voltage imbalance is calculated based on the average deviation of the three-phase voltage; when the metering box is a single-phase metering box, the voltage imbalance is taken as 0 or based on the single-phase voltage relative to the rated voltage. Deviation calculation.

[0023] Furthermore, the residual attribution coefficient is used to measure epitopes. The degree of attribution in persistent residuals and missing data compensation residuals; the event coupling coefficient, used to measure the degree of synchronization between the position power mutation and the event window; the phase deviation coefficient, used to measure the degree of deviation of the position voltage and power factor relative to their respective phases; the missing data impact coefficient, used to measure the impact of communication missing data and copy correction on anomaly attribution; the zero-sequence correlation coefficient, used to measure the degree of correlation between position power changes and box-level zero-sequence current.

[0024] Furthermore, the lightweight anomaly collaborative identification model employs a lightweight gradient boosting classification model composed of 60 to 100 CART classification trees. The input fixed-dimensional feature vector specifically includes: normalized values ​​of persistence residual, event coupling residual, phase imbalance residual, and missing number compensation residual relative to the total active power; as well as the table position anomaly evidence vector, communication integrity rate, re-copying mark, the phase to which the table position belongs, and the most recent... Risk statistics within each sampling period; the output anomaly category probabilities specifically include: communication data shortage anomaly probability, meter anomaly probability, wiring anomaly probability, bypass suspicion probability, and abnormal loss probability.

[0025] Furthermore, the epitope risk index is calculated by weighting and summing the probabilities of each anomaly category and the anomaly persistence intensity index separately. Furthermore, the epitope risk index is categorized according to... The suspicious point tabletop ranking is generated; wherein, the anomaly persistence intensity index is:

[0026] ;

[0027] For the first The epitope in the th epitope Each sampling period corresponds to an anomaly category The probability, This is the anomaly probability threshold. This is an indicator function.

[0028] Furthermore, when the meter location risk index exceeds the preset sorting threshold, the corresponding meter location is listed as a priority review location, and corresponding handling suggestions are output based on the anomaly category with the highest probability: when the probability of communication data loss is the highest, suggestions for communication data replenishment or data acquisition link verification are output; when the probability of meter anomaly is the highest, suggestions for meter verification are output; when the probability of wiring anomaly is the highest, suggestions for phase and wiring relationship verification are output; when the probability of bypass suspicion is the highest, suggestions for bypass power supply investigation are output; when the probability of abnormal loss is the highest, suggestions for continuous tracking and on-site verification of abnormal loss are output.

[0029] The beneficial effects of this invention include:

[0030] First, reduce the misleading effect of communication omissions on box-level residuals by using a clear omission correction formula;

[0031] Second, the residuals are decomposed into a structured form using the moving average, event window, and imbalance constraints to improve the interpretability of anomaly sources.

[0032] Third, by fusing five types of evidence—residual attribution, event coupling, phase deviation, missing data impact, and zero-sequence correlation—the ability to distinguish between meter anomalies, wiring anomalies, bypass suspicions, communication missing data, and abnormal losses is improved.

[0033] Fourth, outputting the ranking and handling suggestions of suspicious sites without increasing the amount of dedicated detection hardware helps to narrow down the scope of on-site investigation and improve the efficiency of line loss management. Attached Figure Description

[0034] Figure 1 This is a flowchart of a method for decomposing low-voltage power metering box-level line loss and collaboratively identifying abnormal metering.

[0035] Figure 2 This diagram illustrates the collaborative identification results of suspicious meter locations, showing the probability output for different meter locations under the categories of missing communication data, meter malfunction, wiring malfunction, suspected bypass, and abnormal loss.

[0036] Figure 3 This is a diagram illustrating the risk ranking of suspicious epitopes, showing the order based on... and The resulting priority review tabletop list.

[0037] Figure 4 This diagram illustrates the performance comparison of different methods in online anomaly identification, showing the performance comparison between the present invention and the single difference threshold method and the single epitope isolation detection method. Detailed Implementation

[0038] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0039] like Figure 1 As shown, the present invention provides a method for decomposing box-level line losses and collaboratively identifying abnormal metering in a low-voltage power metering box, comprising the following steps:

[0040] S1. Collect multi-source data of box level and meter position: Collect the total box level data and the sub-item measurement data of each meter position within the most recent preset time window.

[0041] In this embodiment, a single-phase multi-position low-voltage energy metering box and a three-phase multi-position low-voltage energy metering box are selected as the implementation objects. The rated operating voltage of the metering box is 220 / 380V, the frequency is 50Hz, the sampling period is preferably 1 minute, and the most recent preset time window is preferably the most recent 60 minutes to 120 minutes, so as to take into account both short-term event analysis and continuous deviation extraction.

[0042] Specifically, the total active power at the acquisition box level Total electrical energy Phase A voltage Phase B voltage C-phase voltage With corresponding phase current and zero sequence current Simultaneously, collect sub-metering data from each meter position, including voltage. Current Active power Power factor Electricity Timestamp, communication integrity rate Supplementing and copying marks and the phase to which the epitope belongs Among them, the supplementary copying marks The data used to indicate that the corresponding sampling point was obtained by a data replenishment strategy or corrected from missing data is distinct from real-time acquired data. For example, when a data replenishment marker is used... This indicates that the data for this sampling point comes from a data recovery strategy or is obtained by correcting for missing data. This indicates that the sampling point is collecting data in real time. The sampling time.

[0043] S2. Perform time alignment and construct bin-level conserved residuals: Perform time alignment of bin-level total data and the sub-item measurement data of each measurement station, filter the effective station set, correct missing data, and construct bin-level conserved residuals.

[0044] Specifically, first follow the unified sampling time Perform time alignment on box-level and table-level data.

[0045] Then, the missing or supplemented power data of the meter positions are corrected to obtain the corrected active power of the meter positions. To form the sampling time Effective tabletop set Specifically: for the sampling time If table position Time tolerance If a real-time sampled value exists in memory, then that real-time sampled value is used as the reference power, i.e., the aligned position power. and table Add to the set of valid epitopes If the sampling time If the sample is located between two adjacent valid sampling points, linear interpolation is used to obtain the result. and table join in If table position At sampling time There are no valid real-time sampled values ​​and Then the reference power before the missing number correction will be... Set as , and according to After performing missing number correction, the table position will be... join in ,in , and These are the two most recent valid sampling time points, This represents the upper limit of the trend coefficient. This represents the timestamp corresponding to sampling time k. This represents the timestamp corresponding to the most recent valid sampling time;

[0046] like And exist ,but ;

[0047] like And it does not exist. Then the table will be Mark as invalid table and from Excluded from the middle;

[0048] Thus, the sampling time is obtained. Effective tabletop set .

[0049] The expression for the constructed box-level conserved residual is as follows: .

[0050] S3. Perform multi-time-scale line loss decomposition: Perform multi-time-scale decomposition on the box-level conserved residuals to obtain four residual components with clear physical meaning and calculation methods, including: persistence residuals, event coupling residuals, phase imbalance residuals, and missing number compensation residuals.

[0051] The persistent residual is calculated from the box-level conserved residual using the mean of a sliding window, and the expression is: ,in This is the length of the persistent window.

[0052] The event coupling residual is obtained through event window determination, and its expression is: Among them, if the sampling time Event window length before and after Within the scope, there are power outage events, maintenance operation events, and communication restoration events, or events that meet the following conditions. For load mutation events, the event window indicator value ,otherwise .

[0053] The interphase unbalanced residuals are obtained through unbalance constraints, and the expression is:

[0054]

[0055] in The unbalanced decomposition coefficient, This refers to the voltage imbalance.

[0056] Voltage imbalance of three-phase metering box The expression for calculating based on the average deviation of the three-phase voltage is:

[0057] ,

[0058] in, This represents the average voltage of phases A, B, and C.

[0059] Single-phase metering box Take 0 or use single-phase voltage Relative to rated voltage deviation calculate.

[0060] The expression for the basic compensation amount for missing numbers is:

[0061]

[0062] The expression for missing number compensation residuals is:

[0063]

[0064] Then It represents the degree to which compensation for missing numbers dominates.

[0065] In this embodiment, the event window length The duration ranges from 1 minute to 30 minutes, with a continuous window length of [missing information]. For 30 to 120 sampling points, the unbalanced decomposition coefficients The threshold for determining load mutation ranges from 0.05 to 0.50. The value ranges from 0.05 to 0.30. To prevent tiny positive numbers with a denominator of zero.

[0066] S4. Construction of the anomaly evidence vector for each metering station: Based on the corrected power, event status, phase deviation, communication integrity rate, and zero-sequence current relationship of each metering station, that is, by calculating the residual attribution coefficient, event coupling coefficient, phase deviation coefficient, missing number influence coefficient, and zero-sequence correlation coefficient of each metering station, the anomaly evidence vector for each metering station is constructed.

[0067] In this embodiment, the residual attribution coefficient is calculated for each epitope. Event coupling coefficient Phase deviation coefficient Missing number influence coefficient Correlation coefficient with zero order And form an epitope anomaly evidence vector.

[0068] in Used to measure epitopes Attribution degree in persistent residuals and missing-count compensated residuals Used to measure the degree of synchronization between epitope power mutations and event windows. Used to measure the degree of deviation of the voltage and power factor relative to their respective phases. Used to measure the impact of missing communication data and data correction on anomaly attribution. Used to measure the correlation between changes in meter power and zero-sequence current at the tank level. Specifically:

[0069]

[0070]

[0071]

[0072]

[0073] in, Indicates will Limited to between 0 and 1, For epitope Average power over similar historical periods , For tabletop The power factor, and Tablets The phase voltage and average power factor of the respective phases, For window The correlation coefficient between the change in internal surface power and the zero-sequence current at the tank level. The zero-sequence current reference value determined for a normally operating sample.

[0074] S5. Perform collaborative identification of abnormal measurement: Input the residual components and the anomaly evidence vector of the meter position into the lightweight collaborative identification model to obtain the anomaly category probability corresponding to each measurement meter position.

[0075] Specifically, will Relative to The normalized value, and Communication integrity rate Supplementing and copying marks Phase to which the table belongs And recently Risk statistics within each sampling period constitute a fixed-dimensional feature vector, which is input into a lightweight anomaly collaborative identification model. The model preferably employs a lightweight gradient boosting classification model, consisting of 60 to 100 CART classification trees, each with a maximum depth of 2 to 4, a learning rate of 0.02 to 0.10, and a minimum leaf node sample count of 3 to 10. In this embodiment, the model can consist of 80 CART classification trees, with a maximum depth of 3, a learning rate of 0.05, a minimum leaf node sample count of 5, and a total model storage footprint not exceeding 256kB. Figure 2 The diagram illustrates the collaborative identification results of suspicious epitopes under different anomaly categories.

[0076] The model outputs the probability of communication missing numbers. Meter malfunction probability , probability of wiring abnormalities Probability of bypassing and abnormal loss probability The probabilities of each anomaly category satisfy:

[0077] .

[0078] The training samples include standard wiring and normal communication samples, manually set communication missing data samples, manually set meter drift samples, manually set wiring abnormal samples, manually set bypass load samples, abnormal loss samples, and historical abnormal samples confirmed by on-site manual verification.

[0079] S6. Output Risk Ranking and Rectification Suggestions: Calculate and output risk level, suspicious point ranking and handling suggestions based on the probability of abnormality category and the intensity of abnormality persistence index.

[0080] In this embodiment, the abnormal persistence intensity index is:

[0081] ;

[0082] in For the first The epitope in the th epitope Each sampling period corresponds to an anomaly category The probability, This is the threshold for the probability of anomalies. This is an indicator function.

[0083] The epitope risk index is:

[0084] ,

[0085] to The weights are non-negative and satisfy the following conditions: .

[0086] And further according to Generate a list of suspicious positions. When When the number of cases exceeds the preset sorting threshold, the corresponding table position will be included in the priority review table position, and corresponding processing suggestions will be output based on the anomaly category with the highest probability. Figure 3 The diagram illustrates the risk ranking results for the suspicious tabletops.

[0087] when When the probability is highest, output a suggestion for communication re-reading or data acquisition link verification; when... When the probability is at its maximum, output meter verification recommendations; when When the probability is maximized, output phase and wiring relationship verification suggestions are provided; when When the probability is at its maximum, output a bypass power supply troubleshooting suggestion; when When the probability is at its maximum, output suggestions for continuous tracking of abnormal losses and on-site verification.

[0088] In this embodiment, normal samples are derived from standard wiring, normal communication, and stable loss operation conditions; abnormal samples are derived from manually set communication failures, meter drift, wiring abnormalities, and suspected bypass conditions, as well as abnormal samples confirmed by manual verification on-site. During the verification phase, the top 3 hit rates of suspected meter positions, the accuracy rate of abnormal category identification, the misattribution rate of communication failures, the recall rate of suspected bypass, the false alarm rate, the average identification time, and the model storage usage are statistically analyzed. Figure 4 The results of the comparison of different methods are presented.

[0089] In summary, this invention not only separates the trend and mutation terms in the box-level conservation residuals, but also incorporates the unique data transfer markings, communication integrity rate, effective meter location set, phase assignment, and zero-sequence current information of low-voltage power metering boxes into the same evidence construction process. Specifically, this is achieved through the box-level conservation residuals in S2 and S3. and The executable rules of residual decomposition can transform persistent anomalies, event synchronization anomalies, imbalance correlation anomalies, and missing number compensation anomalies into features that can be used for epitope ranking. This is achieved through the fixed-dimensional anomaly evidence vector in S4. It can maintain stable model input even when the number of meter positions changes or some meter positions are invalid, making it easy to deploy in low-computing-power electricity metering box monitoring terminals.

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for decomposing box-level line losses and collaboratively identifying abnormal metering in a low-voltage power metering box, characterized in that, include: First, collect the total data at the box level and the sub-measurement data at each station, then perform time alignment, screening of the effective station set and correction of missing data, and construct the box-level conservation residuals; The aforementioned time alignment, effective tabletop set filtering, and missing number correction, and the construction of bin-level conserved residuals, specifically involve: according to a unified sampling time... Time alignment is performed on bin-level and table-level data; for sampling time... If table position Time tolerance If a real-time sampled value exists in memory, then that real-time sampled value is used as the reference power, i.e., the aligned position power. and table Add to the set of valid epitopes If the sampling time If the sample is located between two adjacent valid sampling points, linear interpolation is used to obtain the result. and table join in If table position At sampling time There are no valid real-time sampled values ​​and the supplementary recording markers in the header-level total data. Then the reference power before the missing number correction will be... Set as , and according to After performing missing number correction, the table position will be... join in ,in , and These are the two most recent valid sampling time points, This represents the upper limit of the trend coefficient. This represents the timestamp corresponding to sampling time k. This represents the timestamp corresponding to the most recent valid sampling time; like And exist ,but ; like And it does not exist. Then the table will be Mark as invalid table and from Excluded from the middle; According to the expression Construct box-level conserved residuals; Then, the box-level conserved residuals are decomposed into residual components: persistence residuals, event-coupled residuals, phase-to-phase imbalance residuals, and missing number compensation residuals; Further calculations are made of the residual attribution coefficient, event coupling coefficient, phase deviation coefficient, missing number influence coefficient, and zero-order correlation coefficient for each epitope to construct an epitope anomaly evidence vector; Finally, a lightweight anomaly collaborative identification model is used, with inputs including a fixed-dimensional feature vector containing residual components and epitope anomaly evidence vectors, and outputs anomaly category probabilities, suspicious epitope ranking, and processing suggestions.

2. The method as described in claim 1, characterized in that, The total data at the box level includes: total active power. Total electrical energy Phase voltage and current and zero-sequence current The meter readings for each meter position include: voltage. Current Active power Power factor Electricity Timestamp, communication integrity rate Supplementing and copying marks and the phase to which the epitope belongs Among them, the supplementary copying marks The data used to indicate that the corresponding sampling point was obtained by a copying strategy or by correcting missing data is different from real-time acquired data.

3. The method as described in claim 1, characterized in that, The persistent residual is calculated from the box-level conserved residual using the mean of a sliding window, and the expression is: ,in The duration is the length of the persistent window. The event coupling residual is obtained through event window determination, and its expression is: ; among them This is an event window indicator, at the sampling time. Event window length before and after Within the scope, there are power outage events, maintenance operation events, and communication restoration events, or events that meet the following conditions. For load mutation events, the event window indicator value ,otherwise , To prevent tiny positive numbers with a denominator of zero, Threshold for determining load mutations; The interphase unbalanced residuals are obtained through unbalance constraints, and the expression is: in The unbalanced decomposition coefficient, Voltage imbalance; Based on the aforementioned persistent residuals, event-coupled residuals, and interphase imbalance residuals, the missing number compensation residual is: .

4. The method as described in claim 3, characterized in that, When the metering box is a three-phase metering box, the voltage imbalance is calculated based on the average deviation of the three-phase voltage; when the metering box is a single-phase metering box, the voltage imbalance is taken as 0 or based on the single-phase voltage relative to the rated voltage. Deviation calculation.

5. The method as described in claim 3, characterized in that, The residual attribution coefficient is used to measure epitopes. The degree of attribution in persistent residuals and missing data compensation residuals; the event coupling coefficient, used to measure the degree of synchronization between the position power mutation and the event window; the phase deviation coefficient, used to measure the degree of deviation of the position voltage and power factor relative to their respective phases; the missing data impact coefficient, used to measure the impact of communication missing data and copy correction on anomaly attribution; the zero-sequence correlation coefficient, used to measure the degree of correlation between position power changes and box-level zero-sequence current.

6. The method as described in claim 3, characterized in that, The lightweight anomaly collaborative identification model employs a lightweight gradient boosting classification model composed of 60 to 100 CART classification trees. The input fixed-dimensional feature vector specifically includes: normalized values ​​of persistence residual, event coupling residual, phase imbalance residual, and missing number compensation residual relative to the total active power; as well as the table position anomaly evidence vector, communication integrity rate, re-copying mark, the phase to which the table position belongs, and the most recent... Risk statistics within each sampling period; the output anomaly category probabilities specifically include: communication data shortage anomaly probability, meter anomaly probability, wiring anomaly probability, bypass suspicion probability, and abnormal loss probability.

7. The method as described in claim 6, characterized in that, The epitope risk index is calculated by summing the weighted probabilities of each anomaly category and the anomaly persistence intensity index. Furthermore, the epitope risk index is categorized according to... The suspicious point tabletop ranking is generated; wherein, the anomaly persistence intensity index is: ； For the first The epitope in the _ ... Each sampling period corresponds to an anomaly category The probability, This is the anomaly probability threshold. This is an indicator function.

8. The method as described in claim 7, characterized in that, When the risk index of a meter position exceeds the preset sorting threshold, the corresponding meter position will be listed as a priority review meter position, and corresponding handling suggestions will be output according to the anomaly category with the highest probability: when the probability of communication data loss is the highest probability, suggestions for communication data replenishment or data acquisition link review will be output; when the probability of meter anomaly is the highest probability, suggestions for meter verification will be output. When the probability of wiring abnormality is at its maximum, suggestions are made to review the output phase and wiring relationship; when the probability of suspected bypass is at its maximum, suggestions are made to investigate the output bypass power supply; when the probability of abnormal loss is at its maximum, suggestions are made to continuously track the output abnormal loss and conduct on-site verification.

Images