A method for intelligent fusion and anomaly identification of multi-source electricity meter data

Abstract

The present invention provides a method for intelligent fusion and anomaly identification of multi-source meter data; the method collects original electrical parameter data of multiple smart meters under a unified time base, constructs a cross-device measurement data set indexed by meter number, and completes electrical parameter classification and associated information generation in combination with electrical wiring information and functional parameters; extracts electrical parameter change curves through continuous time windows, calculates numerical change consistency indicators, and further introduces response offset feature vectors for joint judgment to achieve accurate identification of abnormal meters and abnormal time periods; for smart meters identified as abnormal, constructs estimated values ​​through other meter data, and calculates residuals between the estimated values ​​and the original values ​​and adjacent estimated values ​​based on the main conductive parameters to achieve double residual verification; judges the anomaly type based on the consistency indicator and the residual result, and outputs the anomaly identification result; the method improves the accuracy of anomaly detection in multi-meter collaborative scenarios.

Description

Technical Field

[0001] The present invention relates to the technical field of electric meters, and in particular to a method for intelligent fusion and anomaly identification of multi-source electric meter data. Background Art

[0002] In existing technologies, anomaly detection methods based on energy balancing, meter comparison, or single-meter event analysis are widely used for the large number of smart meters deployed in distribution systems. Some technical solutions calculate deviations based on the energy difference between the central meter and the customer's sub-meters, or identify anomalies by analyzing the voltage and current trends of a single meter. Other solutions use fixed rules or simple threshold settings to classify specific anomaly types. These methods have achieved certain results in practical applications and have been integrated into distribution automation systems.

[0003] However, existing technologies generally suffer from insufficient data fusion capabilities when processing heterogeneous meter data. They lack systematic modeling of the logical relationships between electrical parameters across multi-source measurement data, making it difficult to effectively extract consistent evolutionary features across meters and time windows. Furthermore, anomaly detection often relies on single features or crude rules, making it difficult to adapt to metering anomaly detection requirements in areas with dynamic load changes or complex topologies. This can easily lead to misjudgments or missed detections, impacting system diagnostic accuracy and response efficiency.

[0004] In view of this, there is an urgent need to provide a data fusion and recognition method that can be used for multi-source heterogeneous smart meters and has spatiotemporal collaborative analysis capabilities and an abnormal information reconstruction mechanism. Summary of the Invention

[0005] This application provides a method for intelligent fusion and anomaly identification of multi-source electricity meter data to improve the diagnostic accuracy and traceability of hidden faults in the power distribution system.

[0006] This application provides a method for intelligent fusion and anomaly identification of multi-source electricity meter data, including:

[0007] Collecting raw electrical parameter data from multiple heterogeneous smart meters under a unified time base, and organizing the raw electrical parameter data into a cross-device measurement data set indexed by smart meter number;

[0008] Classify and label each raw electrical parameter data in the cross-device measurement data set according to the electrical wiring information and functional parameters of each smart meter, and generate correlation information between the electrical parameters;

[0009] Segmenting the cross-device measurement data set into continuous time windows, extracting the electrical parameter change curve of each smart meter, and calculating the value change consistency index within each time window based on the correlation information to obtain a consistency index sequence;

[0010] Calculating the response offset characteristic vector between the electrical parameter change curves, performing a joint judgment in combination with the consistency index, and identifying the abnormal smart meter number and its corresponding continuous time window;

[0011] For the smart meter number identified as abnormal, the original electrical parameter data of other smart meters are called to generate estimated electrical parameter values, and a first residual between the estimated value and the original value, and a second residual between the estimated value and the adjacent estimated values ​​are calculated based on the main conductive parameters;

[0012] According to the consistency index, the first residual and the second residual, determine whether there is a wiring error, current reverse connection, metering abnormality or other abnormality type, and output an abnormality identification result record including the abnormality type, smart meter number and continuous time window.

[0013] The beneficial effects of the technical solution provided by this application include:

[0014] (1) By constructing a cross-device measurement data set and electrical parameter correlation information, the data of single-phase, three-phase and bidirectional metering smart meters are effectively integrated, improving the scope and accuracy of data consistency analysis. (2) By introducing the electrical parameter change consistency index and response offset feature vector, anomalies can be identified from the two dimensions of timing change and topological relationship, significantly improving the recognition accuracy of hidden faults such as wiring errors and current reverse connection. (3) By simultaneously calculating the first residual between the estimated value and the original value, as well as the second residual between the estimated values, the risk of misjudgment caused by a single error indicator is reduced, achieving more reliable anomaly confirmation. (4) It can accurately identify the time period when the anomaly occurred and the associated meter number, providing efficient and traceable data support for fault tracing, status assessment and operation and maintenance scheduling of the distribution system. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flowchart of a method for intelligent fusion and anomaly identification of multi-source electricity meter data provided by the first embodiment of the present application. DETAILED DESCRIPTION

[0016] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of the present application. Therefore, the present application is not limited to the specific implementations disclosed below.

[0017] The first embodiment of the present application provides a method for intelligent fusion and abnormality identification of multi-source electricity meter data. Figure 1 , which is a schematic diagram of the first embodiment of this application. Figure 1A method for intelligent fusion and anomaly identification of multi-source electricity meter data is described in detail in the first embodiment of the present application.

[0018] Step S101: Under a unified time base, original electrical parameter data is collected from multiple heterogeneous smart meters, where the heterogeneous smart meters include single-phase smart meters, three-phase smart meters, and composite smart meters supporting bidirectional metering. The original electrical parameter data includes instantaneous voltage, current, active power, reactive power, electric energy, and power factor, and the original electrical parameter data is organized into a cross-device measurement data set indexed by device number.

[0019] When implementing step S101, it is necessary to synchronously collect raw electrical parameter data from multiple heterogeneous smart meters under a unified time base. This unified time base means that the data collection time points of all meters must be aligned throughout the entire data collection cycle, meaning that data is acquired at the same timestamp to ensure comparability between different meters. This step should rely on a high-precision clock synchronization mechanism, such as the IEEE 1588 Precision Time Protocol (PTP) or Network Time Protocol (NTP), to align the clocks of all meters in the system and ensure consistent sampling start times and intervals across all devices.

[0020] The multiple heterogeneous smart meters include but are not limited to three types: the first type is a single-phase smart meter, which is suitable for ordinary residences or light-load terminals, and collects voltage, current and power information in single-phase circuits; the second type is a three-phase smart meter, which is suitable for commercial, industrial loads or multi-phase power supply scenarios, and can simultaneously collect all electrical parameters of phases A, B and C; the third type is a composite smart meter that supports bidirectional metering. This type of device can simultaneously measure power consumption and reverse power transmission, and is suitable for distributed power grid-connected environments to capture bidirectional power flow information.

[0021] The collected raw electrical parameter data should include at least the following: instantaneous voltage, instantaneous current, active power, reactive power, electric energy, and power factor. Specifically, the instantaneous voltage and current can be sampled at high frequency to obtain the effective value of each phase at the current moment; active power and reactive power are calculated by the instantaneous phase relationship between voltage and current; electric energy is the integral of power per unit time; and power factor represents the cosine value between the current phase and the voltage phase, reflecting the nature of the load. Each of the above electrical parameters must be marked with structured fields such as the corresponding meter number, phase (such as A / B / C phase), measurement channel (such as incoming / outgoing line), and sampling timestamp to ensure accurate data traceability and subsequent processing.

[0022] All collected electrical parameter data must be organized according to the device's unique identification code and organized into a cross-device measurement data set indexed by device number. Each device number should uniquely correspond to a smart meter, for example, distinguished by its device ID, installation location code, or meter communication address. The structure of this cross-device measurement data set can take the form of a two-dimensional or three-dimensional matrix, where each row of the matrix represents a time point and each column represents a different electrical parameter field or meter number combination. This allows for rapid retrieval of the electrical parameters corresponding to all meters at any point in time, enabling efficient comparison of cross-device data.

[0023] In specific project deployments, edge computing devices or concentrators can be configured as data aggregation nodes to periodically poll the measurement channels of all meters and cache raw data. After preliminary verification, this cached data is uploaded to the master system for subsequent fusion analysis and processing. To ensure data integrity and timeliness, an automatic re-collection policy for missing data should be implemented, and abnormal communication, bit errors, or breakpoint retransmission should be flagged. This ensures a clearly structured, time-consistent, and complete set of raw measurement data.

[0024] By completing step S101 in the above manner, a unified, accurate, and traceable raw measurement basis for multi-source heterogeneous electricity meters can be provided for subsequent data fusion, classification and labeling, consistency analysis, and anomaly identification. The accuracy and synchronization of this step directly determine the reliability of the entire system's fusion and identification capabilities.

[0025] Step S102: Classify and label each original electrical parameter data in the cross-device measurement data set according to the electrical wiring information and functional parameters of each smart meter, and generate correlation information between the electrical parameters.

[0026] When implementing step S102, it is necessary to further combine the cross-device measurement data set that has been uniformly collected and archived with device numbers in step S101 with the electrical wiring information and functional parameters corresponding to each smart meter, and systematically classify and label the collected raw electrical parameter data to achieve clear distinction and semantic understanding of the data meaning, source, and contextual relationship in the subsequent analysis process.

[0027] The so-called electrical wiring information refers to the physical wiring method, access type, connected load type, and phase relationship of each smart meter in the power system. For example, a three-phase smart meter may use a three-phase four-wire wiring method, and its corresponding electrical parameters should include the voltage, current, active power, and reactive power of phases A, B, and C, while a single-phase smart meter may only measure the voltage and current between LN. For composite smart meters that support bidirectional metering, it is necessary to clarify their forward and reverse power metering channels, and whether they have harmonic analysis, power factor correction, or independent three-phase metering capabilities. The above wiring information is usually entered by the system when the meter is installed, and can also be modeled and confirmed by the master station system through an automatic identification program.

[0028] Functional parameters include the meter's technical specifications, measurement accuracy, sampling frequency, data communication method, supported electrical parameters, and measurement range. These parameters are used to determine the reliability and applicability of a particular electrical parameter. For example, when analyzing reactive power trends over a specific time period, if a meter only supports active power measurement or has insufficient reactive power accuracy, the system should automatically exclude its reactive power data from subsequent analysis to avoid introducing bias.

[0029] Specifically, for the original electrical parameter data set corresponding to each smart meter number, the system needs to establish a structured labeling system to mark the original data fields according to the electrical parameter type (such as voltage, current, power, etc.), phase (such as phase A, phase B, phase C or single phase), measurement direction (forward, reverse), data nature (instantaneous value, average value, integral value, etc.), measurement point location (incoming side, outgoing side), sampling frequency, metering accuracy level and other information, and record the associated wiring method, equipment role (such as main table, branch table, relay table) and its regional topology number and other meta-information.

[0030] Ultimately, based on the aforementioned classification and tag management mechanisms, the system generates an electrical parameter association information table that maps the functional relationships between electrical parameters across meters, channels, phases, and directions. For example, it indicates that the power factor of a three-phase load should be derived from the active and reactive power of phases A, B, and C, or that missing phase C current data from a meter can be compensated for using phase matching data from an adjacency table.

[0031] The successful implementation of this step provides a solid data structure and semantic foundation for subsequent operations such as time series consistency analysis, extraction of response differences between multiple tables, and residual verification, ensuring that the system can achieve accurate and controllable fusion analysis and anomaly identification in complex data environments.

[0032] Here's a specific example: Assume the power distribution system includes three smart meters, numbered M1001, M1002, and M1003. M1001 is a three-phase, four-wire meter installed on the main incoming line of a commercial building. M1002 is a single-phase smart meter installed on a branch line of the same building. M1003 is a bidirectional meter connected to the distributed photovoltaic grid-connected interface. After executing step S101, the system has collected raw electrical parameter data from each meter at a specific point in time, including voltage, current, active power, reactive power, energy, and power factor.

[0033] In step S102, the system first retrieves the preconfigured wiring information and functional parameters from the meter management platform. For example, it identifies M1001 as a three-phase meter supporting A / B / C voltage and current sampling with 0.5S-level metering accuracy; M1002 as a single-phase meter capable of only LN voltage and current sampling with a 15-minute sampling frequency; and M1003 as a composite meter supporting forward and reverse metering, distinguishing directionality for both active and reactive power, and supporting 1-second high-frequency data transmission. The system then annotates each of these data fields based on a pre-set tagging template and categorizes them into structured measurement semantic tags.

[0034] For example, for the field "I_B" of M1001, the system marks it as: "M1001, three-phase four-wire, phase B current, instantaneous value, forward channel, trunk node, sampling frequency 1Hz, accuracy 0.5S", and for the field "P_reverse" of M1003, it is marked as: "M1003, single-phase bidirectional, reverse active power, average value, distributed interface, sampling frequency 1Hz, accuracy level 1.0".

[0035] In the electrical parameter association information table finally generated by the system, the A / B / C phase power data of M1001 is marked as "three-phase synthetic source", the data of M1002 is marked as "branch reference point", and the forward and reverse power of M1003 is marked as "boundary flow point". It is also clear that there is a "time-space comparison relationship" between it and the upstream node M1001, thereby supporting subsequent data mapping and abnormal linkage analysis between multi-source meters.

[0036] Step S103: segmenting the cross-device measurement data set into continuous time windows, extracting the electric parameter change curve of each smart meter, and calculating the value change consistency index within each time window based on the correlation information to obtain a consistency index sequence.

[0037] When implementing step S103, it is necessary to perform sequence segmentation processing on the cross-device measurement data set constructed in step S101 in the time dimension based on the structured electrical parameter labels and their associated information generated in step S102, so as to extract the change curve reflecting the evolution behavior of the electrical parameters of each smart meter, and on this basis quantify the coordinated change characteristics of the electrical parameters between different meters in the same time window.

[0038] First, the system should set a fixed-length time window parameter, such as 15 minutes, 1 minute, or 10 seconds, and divide the measurement data set into multiple non-overlapping or sliding overlapping time periods based on the sampling frequency of the smart meter. Each time window contains several consecutive sampling points to ensure that each meter has a complete record of the original electrical parameters within that time period. If a meter has missing data within a certain window, it can be supplemented through temporal neighbor interpolation, mean filling, or structured missing markup to ensure that the data set involved in the analysis has a consistent structure and uniform dimensions.

[0039] Next, within each time window, the system extracts time series data for target electrical parameters based on the categorized and labeled fields. For example, it extracts the Phase A active power curve, total reactive power curve, and power factor curve for all meters. These curve data are not primarily based on single-point values, but rather on sequential trajectories formed within the window. This constructs a trajectory for each electrical parameter of each device within the current time period. The system normalizes these trajectories so that values ​​from different meter types (e.g., single-phase and three-phase) can be compared using a unified dimension.

[0040] To evaluate the operational consistency of multiple smart meters within the same time window, it is necessary to calculate the consistency index of the numerical changes between devices based on the correlation information between the electrical parameters. This index is not a simple difference or average error, but reflects the degree of coordination between multiple electrical parameter trajectories. The calculation method can be implemented using various methods such as sliding correlation coefficient, dynamic time warping (DTW) distance, average slope deviation, cross standard deviation, etc. For example, if three meters are connected to the same branch and have highly correlated load change characteristics, their phase A current curves should have a synchronous upward or downward trend; if one of the meters suddenly experiences abnormal fluctuations or becomes unresponsive, its consistency index will drop significantly.

[0041] The system constructs a consistency score vector for each set of electrical parameters, including the correlation metrics between each meter within the current window. It then weights and combines the consistency scores of all electrical parameters to generate a numerical indicator representing the overall consistency level within that time window, known as the consistency index. This index can be a real number or a graded level, used to indicate the possibility of abnormal behavior.

[0042] This consistency indicator sequence will eventually form a continuous time trajectory over time, becoming an important basis for identifying abnormal trends and diagnosing anomaly onsets. Subsequent step S104 will further utilize this sequence in conjunction with response offset characteristics to determine anomalies. Therefore, step S103 not only achieves temporal structuring and sequence abstraction of the measurement data but also lays the core foundation for behavioral modeling and quantitative evaluation throughout the intelligent recognition process.

[0043] To facilitate understanding of the specific implementation of step S103, the following example illustrates this. Assume that three smart meters, numbered M2001, M2002, and M2003, are deployed in a certain substation. M2001 is the substation's master meter, while M2002 and M2003 are sub-meters for two branch users, respectively. All three meters support one-minute data reporting, collecting electrical parameters such as three-phase current, active power, and power factor. The system sets a time window length of 15 minutes and uses non-overlapping segmentation to divide all data into continuous time windows.

[0044] During the time window of 10:00 to 10:15 on May 1, 2025, the system extracts the Phase A active power curves of the three meters and confirms, based on the tag information provided in step S102, that all three meters are three-phase, four-wire systems and belong to the trunk and branch nodes of the same power supply branch. Next, the system normalizes each power curve so that its value range within the window is normalized to the [0,1] interval and uses the sliding correlation coefficient method to calculate the power change trend correlation between M2001 and M2002, M2001 and M2003, and M2002 and M2003. Assuming the correlation coefficients obtained are 0.93, 0.89, and 0.15, respectively, the system determines that there is a significant deviation between M2002 and M2003.

[0045] The system then combines the three sets of correlation values ​​to construct a consistency score vector of [0.93, 0.89, 0.15]. The weighted average yields a consistency index value of 0.656 for the current time window, which is lower than the set threshold (e.g., 0.80). The system records this value and writes it to the consistency index sequence, which is used in subsequent steps to identify anomaly occurrence points and infer anomaly evolution trends.

[0046] Step S104: Calculate the response offset characteristic vector between the electrical parameter change curves, perform joint judgment in combination with the consistency index, and identify the abnormal smart meter number and its corresponding continuous time window.

[0047] During the implementation of step S104, the system needs to further analyze the response differences between the electrical parameter change trajectories based on the obtained consistency indicator sequence, and capture the micro-timing mismatch, hysteresis change or sudden offset in the meter behavior by introducing the response offset feature vector, thereby enhancing the sensitivity and accuracy of anomaly identification.

[0048] Specifically, response offset refers to the time difference between synchronized changes in the measured values ​​of different meters when faced with events such as system load changes, power supply disturbances, or operating state changes. In theory, if multiple smart meters are located at consecutive locations on the topological path of the same substation, their active power or current should respond to external disturbances in a nearly synchronized manner. However, in real-world scenarios, due to communication delays, device response time differences, or wiring differences, the change trajectories of some meters may lag or advance significantly. If this time offset does not match its physical topological relationship, it may indicate potential problems such as abnormal wiring, clock drift, or data errors.

[0049] When calculating the response offset eigenvector, the system first selects a key electrical parameter for analysis, such as active power or phase A current, and extracts the corresponding change curves for all meters within each time window. Based on this, the system accurately calculates the response time differences between meters by matching the curve's changing trends and inflection points, using a sliding window delay comparison method or cross-correlation peak location method. This time difference, expressed in milliseconds or as a number of sampling points, forms a response offset eigenvector indexed by the meter number.

[0050] For example, within a 15-minute window, if a load increase occurs on the main trunk meter M3001 and two branch meters M3002 and M3003, but M3002's response lags behind M3001 by two sampling points, while M3003's response is one sampling point ahead, the response offset characteristic vector can be represented as {M3001: 0, M3002: +2, M3003: -1}. If M3003 is physically located downstream of M3001, its "early response" behavior will be considered an abnormal offset, triggering further system review.

[0051] The system then performs a joint judgment on the obtained response offset feature vector and the consistency index sequence calculated in step S103. On the one hand, if a meter exhibits a large response offset and a low consistency index within a certain time window, it indicates that the behavior of the meter is obviously inconsistent with other meters, and its abnormality possibility is significantly increased. On the other hand, even if the consistency index is high, if the response offset value is seriously inconsistent with the expected topological response logic, it may also indicate the presence of a structural anomaly. The system can set a dual-condition judgment rule, for example, "the response offset exceeds ±2 sampling points and the consistency index is lower than 0.6" to trigger a preliminary abnormality mark.

[0052] This judgment mechanism, combining response behavior with consistency of change, enables the system to identify issues difficult to detect using traditional energy difference methods, such as incorrect wiring sequences, meter address mapping anomalies, sampling time base drift, and local power supply disturbances. The final anomaly determination output includes the smart meter ID, the consecutive time window number where the anomaly occurred, and the corresponding response offset value range, providing a data foundation for subsequent residual verification and anomaly classification.

[0053] The design of this step ensures that the anomaly identification process not only focuses on the differences between numerical values, but also more comprehensively considers the dynamic behavior of electrical parameters in the time dimension, making the anomaly judgment have higher timing sensitivity and topological explanatory power.

[0054] To facilitate understanding of this embodiment, the following example is provided. Assume that a distribution area contains three smart meters, numbered M3001, M3002, and M3003. M3001 is the master meter for the trunk line, while M3002 and M3003 are the user sub-meters for the two terminal branches. Physically, these three meters form a typical master-branch-branch topology. Within a certain continuous time window, the system detects a significant upward trend in the active power of phase A on all three meters.

[0055] The system first extracts the rising inflection point for each meter based on the electrical parameter variation curve within the time window. Using the cross-correlation peak location method, the system finds that the power rise for M3001 occurs at time T+4, for M3002 at T+6, and for M3003 at T+2. Taking M3001 as the reference, the response of M3002 lags by two sampling points, while the response of M3003 leads by two sampling points. This results in a response offset feature vector of {M3001: 0, M3002: +2, M3003: –2}.

[0056] The system then reviewed the substation's wiring topology and confirmed that M3003 was physically located after M3001, making it a typical downstream node. Normally, it should have responded later or more synchronously with M3001, but it responded earlier, violating expected logic. The system also determined that the consistency index between M3003 and other meters during that time window was only 0.41, far below the average for a typical substation.

[0057] Based on the pre-set dual-condition judgment logic (response deviation exceeding ±2 sampling points and consistency index below 0.6), the system determined that the M3003 had abnormal response behavior within the time window. Further tracking of its historical data confirmed that the meter had been replaced several days ago. A possible construction wiring error caused the topology configuration to be inconsistent with the sampled response. Based on this, the system automatically marked the meter number and time window, generated an abnormality record, and pushed it to the master station operation and maintenance system to trigger an on-site verification task.

[0058] Furthermore, the calculating of the response offset characteristic vector between the electrical parameter change curves includes:

[0059] In a cross-device measurement data set segmented by continuous time windows, for each continuous time window, the electrical parameter change curves of all smart meters under the marked main conductive parameter dimension are extracted, and the main response inflection point sequence in each electrical parameter change curve is identified based on the correlation information between the electrical parameters;

[0060] Calculate the response delay of each smart meter within the current continuous time window based on the relative position difference between the main response inflection point sequence corresponding to each smart meter number and the main response inflection point sequence of a reference smart meter, where the reference smart meter is the smart meter with the highest value change consistency index within the current continuous time window;

[0061] Combined with the electrical wiring information corresponding to each smart meter in the association information, a topological path directed distance matrix is ​​constructed, and based on the constructed topological path directed distance matrix, the response delay of each smart meter in the current continuous time window is converted into a topological offset distance;

[0062] The response delay of each smart meter in the current continuous time window and its topological offset distance are jointly encoded to generate a response offset feature vector indexed by the smart meter number.

[0063] First, we need to clarify the essential meaning of the "response offset feature vector": This feature vector is an ordered vector that characterizes the response time differences of multiple smart meters to electrical parameter changes within the same continuous time window. Combined with their physical wiring structure, it is used to subsequently identify abnormal response behavior. The construction of this feature vector requires a complete process based on a collection of cross-device measurement data and electrical parameter semantic labeling, combined with topological structure, inflection point identification, and timing analysis.

[0064] First, in the step of extracting the electrical parameter change curve within the continuous time window, the field marked as "main conductive parameter" should be used as the basis. For example, when detecting metering anomalies, active power is usually selected. If reverse connection is determined, current direction or power factor is used first. For each smart meter, the sampling value sequence of the main conductive parameter within the current time window is extracted to form the electrical parameter change curve of the meter under this window. After the curve is constructed, the main response inflection point sequence of the meter is extracted through the first-order derivative extreme value detection or trend turning point detection algorithm in combination with the correlation information between the electrical parameters (such as the voltage-current correspondence, the power calculation path, the three-phase symmetry constraint, etc.). The time index position of each main response inflection point is used to characterize the dynamic response position of the meter to load changes.

[0065] Next, the system needs to determine the reference smart meter within this window as the benchmark for response alignment. This reference meter should not be pre-specified, but should be dynamically selected by calculating the consistency index of the numerical changes of all meters within the current continuous time window. The consistency index can be evaluated using methods such as the Pearson correlation coefficient, dynamic time warping (DTW) similarity, or trend angle deviation, and the one with the highest index is used as the reference smart meter. On this basis, the system compares the difference between the main response inflection point sequence of each meter and the inflection point sequence of the reference meter on the time axis, and calculates the response delay of each meter, usually in units of sampling points or milliseconds, with positive values ​​indicating lag and negative values ​​indicating advance.

[0066] The system then calls the association information table generated in step S102 to extract the upstream and downstream relationships and line path numbers for each smart meter in the physical wiring structure. Combined with electrical wiring information (such as incoming / outgoing line direction, phase difference, and wiring method), it constructs a topological path directed distance matrix. Each element in this matrix represents the path length or path level difference between the reference meter and the target meter, quantifying the delivery order in the physical topology. For example, if M7011 is the reference meter, and its path to M7012 is a level 1 arrival, and its path to M7013 is a level 2 arrival, the path values ​​in the matrix are 1 and 2, respectively. These path values ​​are used to determine whether the response delay conforms to the physical delivery logic.

[0067] Finally, the response delay of each smart meter within the current continuous time window and its corresponding topological path distance are jointly encoded. This encoding can use a vector concatenation structure, sequentially combining the response delay and topological offset distance of each meter into a binary tuple, such as {M7012: [+2, 1], M7013: [–1, 2]}. Alternatively, it can be further converted into a single offset score, such as the offset residual of the response delay minus the topological path value. This vector structure, indexed by the meter number, forms a complete response offset feature vector, which serves as the input for subsequent identification of abnormal smart meter numbers and their corresponding continuous time windows.

[0068] To facilitate understanding, a specific example is provided below to demonstrate the complete construction process of the response offset feature vector in an actual substation anomaly monitoring scenario, ensuring that those skilled in the art can accurately understand and implement it.

[0069] Assume that three smart meters, numbered M8011, M8012, and M8013, are deployed in a city's distribution substation. M8011 is the main incoming meter for the substation, while M8012 and M8013 are the branch user meters. Between 2:00 PM and 2:15 PM on September 12, 2025, the system detects a significant load change, entering this continuous time window for processing.

[0070] First, the system extracts the electrical parameter change curves for the three smart meters, under the dimension of active power, the primary conductive parameter, from a cross-device measurement data set indexed by device number. The M8011 curve shows a sharp increase in power at 2:03 PM. The M8012 curve shows the same trend around 2:03:5 PM, and a change occurs at 2:02:5 PM for the M8013 curve. Using an inflection point detection algorithm, the system identifies the primary response inflection point in each electrical parameter change curve: t = 180 seconds for M8011, t = 183 seconds for M8012, and t = 177 seconds for M8013. With a sampling period of 1 second, M8012 lags behind M8011 by 3 sampling points, while M8013 leads by 3 sampling points, marking the response delays as +3 and –3, respectively.

[0071] The system then evaluated the consistency of the values ​​of the three meters within the current continuous time window. The results showed that the consistency index between M8011 and M8012 was 0.91, between M8011 and M8013 was 0.76, and between M8012 and M8013 was 0.68. Because M8011's change trend was most consistent with the other meters, it was selected as the reference smart meter.

[0072] Next, the system retrieves the electrical wiring information for M8011, M8012, and M8013. From the associated information table, it learns that M8011 is the incoming main meter, while M8012 and M8013 are the first- and second-level downstream sub-meters, respectively. Based on this information, the system constructs a directed distance matrix for topological paths, where the path from M8011 to M8012 is level 1, and the path to M8013 is level 2. The topological path values ​​are recorded as 1 and 2, respectively.

[0073] The response delay of each meter is jointly encoded with its topological path value to generate the following response offset feature vector:

[0074] M8012: [+3, 1] → offset residual = +2;

[0075] M8013: [–3, 2] → offset residual = –5;

[0076] If the offset residual is used to construct a single score value (e.g., residual = response delay – topological path value), the complete response offset feature vector can be written as:

[0077] {

[0078] M8012: +2,

[0079] M8013: –5

[0080] }

[0081] The vector shows that M8012 exhibits a slight response lag, which is generally consistent with expectations for its primary path. However, despite being located on a more distant secondary branch, M8013 responds early, with an offset residual of -5 significantly deviating from the topological logic, suggesting that this device may have a wiring error or an abnormal address configuration.

[0082] This vector is then used as input for subsequent steps to identify anomalies. In this example, the offset feature not only quantifies the timing characteristics of the response but also integrates the response time difference with the physical structure, avoiding misjudgments caused by simple timing comparisons. This has clear engineering significance and application value.

[0083] Step S105: Number the smart meter identified as abnormal, call the original electrical parameter data of other smart meters to generate estimated electrical parameter values, and calculate the first residual between the estimated value and the original value, and the second residual between the estimated value and the adjacent estimated value based on the main conductive parameter.

[0084] During step S105, the system uses the abnormal smart meter number identified in step S104 and its corresponding continuous time window to retrieve the raw electrical parameter data of the associated smart meter, generate estimated electrical parameter values ​​for the abnormal meter, and construct a residual vector for error correction. This process not only serves as a secondary verification of the abnormality determination but also forms a crucial foundation for classifying and quantifying the causes of the abnormality.

[0085] First of all, the so-called "estimated electrical parameter value" refers to the prediction of the normal measurement results of the abnormal meter during the same period by fusing the original electrical parameter data of other non-abnormal smart meters in the same area. This estimated value is not a simple average, but is obtained by weighted combination, interpolation, extrapolation or model fitting, taking into account factors such as topology, adjacent physical relationships, load similarity, and electrical parameter correlation. For example, if a branch meter is identified as abnormal, its upstream trunk meter and other user meters on the same branch can provide estimated reference values. The system inputs its active power, phase A current or power factor and other indicators into the reconstruction function and outputs the estimated electrical parameters of the target meter.

[0086] To ensure that those skilled in the art can clearly implement the process of estimating electrical parameter values, the following provides a feasible specific algorithm example for the process, which is applicable to the meter deployment structure with a trunk and branch relationship in most distribution substations.

[0087] After identifying a smart meter numbered M5012 as exhibiting abnormal behavior during a specific time window, the system needs to estimate the meter's primary conduction parameters (e.g., active power) during that time period. To do this, the system first selects the upstream trunk meter M5001 and the lateral user meter M5013 on the same branch line from among other normal smart meters in the same substation with a clear physical connection to M5012 as reference meters.

[0088] Next, the system extracts the historical active power measurements of M5001, M5012, and M5013 over the same time period over several consecutive days (for example, the past 30 days) to construct three time series. Based on these time series, the system calculates the Pearson correlation coefficients between M5001 and M5012, and between M5013 and M5012, respectively. Assume that the calculation results are: Corr(M5001, M5012) = 0.82, and Corr(M5013, M5012) = 0.64. Based on this, the system determines the estimated contribution weight of each reference meter to M5012, specifically:

[0089] The weight of M5001 is 0.82 / (0.82 + 0.64) ≈ 0.5625;

[0090] The weight of M5013 is 0.64 / (0.82 + 0.64) ≈ 0.4375.

[0091] Then, the system extracts the real-time active power data of M5001 and M5013 within the current estimated time window, which are 1.80 kW and 1.20 kW respectively. Based on the weighted average method, the system calculates the estimated active power of M5012 as:

[0092] Estimate(M5012) = 1.80 × 0.5625 + 1.20 × 0.4375 = 1.53 kW.

[0093] This estimated value is the combined prediction of the M5012 within that time window, based on data from physically adjacent meters and historical correlations. If the M5012's actual measurement during that time period is 2.10 kW, the first residual is +0.57 kW, and the residual percentage is +37.25%, significantly exceeding the system's preset ±10% warning threshold. This indicates a preliminary determination of an abnormal deviation in the meter.

[0094] Next, we need to clarify the meaning of "primary conductive parameters." When distinguishing different types of anomalies, the system should select the electrical parameter that is most sensitive and representative of that type of anomaly as the dominant indicator. For wiring errors and reverse current, the direction of phase A current or power is most critical; for metering anomalies, active power or electrical energy is more appropriate. Therefore, based on the previously labeled semantic labels of electrical parameters, the system automatically identifies the primary conductive parameters for the current scenario and uses them as the primary input for residual analysis.

[0095] The system then calculates the first residual between the estimated and original measured values ​​based on the primary conductive parameter. This residual, known as the "prediction error," is the deviation between the target meter's measured behavior within the current time window and the predicted behavior based on neighboring meters. If this deviation exceeds a normal statistical range or threshold (e.g., exceeding ±10% or ±2σ), it indicates that the meter's behavior may be affected by abnormal factors.

[0096] To avoid misjudgments due to estimation errors in the first residual, the system also introduces a "second residual" as an auxiliary judgment basis. The second residual refers to the deviation between the estimated value and its neighboring estimated values. It is used to verify whether the estimated value is consistent with other estimated data within the entire estimation scenario. For example, if the estimated values ​​of multiple adjacent meters form a consistent sequence, but the estimated value of an abnormal meter deviates significantly from these sequences, it may indicate that the anomaly of the meter is an independent disturbance rather than a group fluctuation.

[0097] The system ultimately combines the first and second residuals to construct a residual feature vector, which is then cross-validated in multiple dimensions using response offsets and consistency metrics. Residual values ​​can be expressed in the same units as the original data, such as kWh, kW, or A, or normalized to a percentage for unified judgment. The positive or negative value of the residual can also reflect the directionality of anomalies, such as energy backflow or current reverse flow.

[0098] In summary, step S105 not only provides a refined abnormal error quantification mechanism, but also realizes a verifiable, explainable, and classifiable meter abnormality confirmation process through the selection of main conductive parameters and the construction of double residuals.

[0099] To facilitate understanding of the specific implementation of step S105, the following provides a detailed example based on a real-world scenario. Assume that multiple smart meters are deployed within a residential complex. M4011 is the master meter on the main line, while M4012 and M4013 are the user sub-meters on two branch lines. In step S104, the system identifies that M4013 exhibits abnormal behavior during the time window from 10:00 AM to 10:15 AM on July 15, 2025. This behavior manifests as an early response and a power trend significantly inconsistent with that of other meters.

[0100] During step S105, the system first retrieves the raw electrical parameter data corresponding to M4013 within the time window from M4011 and M4012, primarily including fields such as Phase A active power, active energy, and power factor. Combining the previously determined electrical parameter semantic tags, the system determines that the primary electrical parameter of this abnormal event is "active power," as this parameter is highly responsive to wiring errors, reverse current, and metering inaccuracies.

[0101] The system then compared topological paths to confirm that M4011 and M4013 were connected in an upstream-downstream power supply relationship. M4012 physically belonged to the same branch as M4013, but were not directly connected in series. Based on this structure, the system constructed an estimation model and reconstructed M4013's estimated active power using a weighted summation method. Weights were set based on topological distance and historical correlation, for example, M4011 had a weight of 0.7 and M4012 had a weight of 0.3. Calculations revealed that M4013's estimated active power during the time window was 1.35 kW, while the actual measured value was 1.89 kW. The difference was 0.54 kW, representing 40% of the estimated value and significantly exceeding the system's set 10% error threshold, resulting in the first residual error.

[0102] Furthermore, the system calculated the estimated values ​​of M4013, reconstructed using the same method for M4011 and M4012 during the same time period, to be 1.32 kW and 1.37 kW, respectively. The average deviation between these two values ​​and the current estimated value of 1.35 kW was less than 2%, indicating that the estimated values ​​themselves were stable and reliable among neighboring devices. However, this estimated value deviated by 40% from the original data for M4013, indicating that the anomaly was more likely due to fluctuations in M4013 itself rather than in the estimated source, thus forming a secondary residual, further validating the independence anomaly for M4013.

[0103] Ultimately, the system records the first residual (+0.54 kW) and the second residual (highly consistent but conflicting with the original value) as a residual feature vector, and inputs them into the anomaly classifier together with the consistency index (0.42) and response offset (–3 sampling points) formed in the previous step, confirming that M4013 has high-confidence abnormal behavior in the wiring direction of the electricity meter. The system automatically generates an anomaly identification record containing the device number, anomaly type, time window, and a comparison table of estimated values ​​and original values ​​for subsequent on-site operation and maintenance.

[0104] Furthermore, the calculating, based on the main conductive parameter, a first residual between the estimated value and the original value, and a second residual between the estimated value and an adjacent estimated value, includes:

[0105] Extracting, within the current continuous time window, original electrical parameter data of the main conductive parameter corresponding to the smart meter number identified as abnormal based on the generated estimated electrical parameter value, and calculating a difference between the original electrical parameter data and the estimated electrical parameter value; and obtaining a standardized residual value of the smart meter number within the continuous time window by dividing the difference by the estimated electrical parameter value, which is used as a basis for constructing a first residual;

[0106] Calling the topological path directed distance matrix, screening multiple smart meter numbers that are topologically adjacent to the smart meter number identified as abnormal from the cross-device measurement data set, obtaining estimated electrical parameter values ​​generated based on the main conductive parameters within the continuous time window, and performing weighted combination according to the topological path directed distance to construct a set of adjacent estimated values ​​within the continuous time window;

[0107] A weighted difference calculation is performed on the estimated electrical parameter value of the smart meter number identified as abnormal within the continuous time window and the set of neighboring estimated values. The weighting coefficient is determined by the inverse proportion of the directed distance of the topological path, and combined with the sampling stability index of the main conductive parameter, a second residual of the smart meter number within the continuous time window is generated to characterize the degree of deviation of the estimated electrical parameter value relative to the set of neighboring estimated values.

[0108] In the method described in this invention, the "first residual" and "second residual" are not simply recalculations of a single numerical error. Instead, they serve two logical purposes: first, they measure the relative deviation between a single meter's current behavior and its expected value; second, they assess the credibility of the current estimate itself, avoiding misidentification of anomalies due to unstable estimation sources. This establishes a robust, bidirectionally cross-validated anomaly identification mechanism. The following describes the residual construction process.

[0109] First, the system must enter the continuous time window defined by the previous steps, assuming it's from 4:00 PM to 4:15 PM on October 12, 2025. Within this time window, the system has identified a group of smart meter numbers marked as "possibly abnormal" based on its anomaly recognition mechanism, such as M9012. This number is supported by the following data structures: tag information for its primary conductive parameter (such as "active power" or "phase B current"), the raw electrical parameter sampling sequence within the continuous time window, the wiring paths and directed distances derived from the topological graph, and estimated electrical parameter values ​​constructed by the system by fusing data from multiple neighboring meters.

[0110] Before calculating the first residual, it is necessary to ensure that the estimated electrical parameter values ​​are engineering-reasonable. These values ​​are derived from the topologically adjacent, non-anomalous smart meters in step S105. They are generated by combining the associated information table with the directed distance matrix of the topological path, using a weighting algorithm (such as Pearson correlation weighting, inverse path weighting, or historical load similarity weighting). Based on the estimated electrical parameter values, the system then extracts the raw electrical parameter data sequence for the M9012 in the primary conductive parameter dimension within the current continuous time window, such as its minute-by-minute active power measurement values: {3.2 kW, 3.4 kW, 3.6 kW, 3.5 kW}.

[0111] The system uses the estimated electrical parameter value sequence as a reference and calculates the difference between the original measured value and the estimated value point by point. For example, if the estimated values ​​are: {2.9 kW, 3.1 kW, 3.2 kW, 3.3 kW}, then the difference values ​​are: {+0.3, +0.3, +0.4, +0.2}. To avoid distortion of absolute value errors at different magnitudes (for example, the judgment criteria for a 10 kW error of 0.5 is the same as that for a 1 kW error of 0.5), the system divides each difference by the estimated value itself to form a standardized proportional error sequence, that is:

[0112] The first residual sequence = {(3.2–2.9) / 2.9 = 10.3%, (3.4–3.1) / 3.1 ​​= 9.7%, (3.6–3.2) / 3.2 = 12.5%, (3.5–3.3) / 3.3 = 6.1%}.

[0113] The system can further take the mean or maximum value of the residual sequence as the first residual of M9012 within the time window. For example, the maximum standardized residual value of 12.5% ​​is selected as the first residual. This value is used to measure whether the meter behavior deviates significantly from the normal range inferred by the system and is one of the basic dimensions of anomaly identification.

[0114] After completing the first residual, the system enters the second residual construction process. This process aims to determine whether the estimated electrical parameter value has reasonable internal consistency, that is, whether the estimated value itself is reliable. If the estimated value fluctuates due to large fluctuations or low correlation with neighboring meter data, even if the first residual is large, it cannot be directly judged as a device anomaly.

[0115] The system first selects neighboring smart meters from the cross-device measurement data set whose path level is less than or equal to 2 in the topological path directed distance matrix. These are assumed to be M9009 and M9010, with path distances of 1 and 2, respectively. Within the current continuous time window, the system obtains the estimated electrical parameter values ​​of M9009 and M9010 in the primary conductive parameter dimension: M9009: {3.0, 3.2, 3.4, 3.3}; M9010: {2.8, 3.0, 3.2, 3.2}.

[0116] The system performs a weighted fusion of these proximity estimation sequences, with the weight determined by the inverse of the path distance. For example, if the path is 1, the weight is 1.0, and if the path is 2, the weight is 0.5. The proximity estimation value at each moment can be expressed as a weighted average. For example:

[0117] Neighborhood estimate (point 1) = (3.0×1 + 2.8×0.5) / (1+0.5) = 2.93Neighborhood estimate (point 2) = (3.2×1 + 3.0×0.5) / (1+0.5) = 3.13……

[0118] Calculate the difference between the estimated electrical parameter value of M9012 and the adjacent estimated value sequence point by point to obtain the deviation sequence. For example, if the estimated value of M9012 is {2.9, 3.1, 3.2, 3.3}, the difference is:

[0119] {2.9–2.93 = –0.03,3.1–3.13 = –0.03,3.2–3.27 = –0.07,3.3–3.18 = +0.12}.

[0120] The system calculates the mean square error or mean absolute error of the difference sequence, for example, taking the mean absolute error as the second residual.

[0121] To improve the accuracy of the credibility judgment, the system also considers the sampling stability of the main conductive parameter within the current continuous time window, namely the standard deviation or the rate of change within the sliding window. If the estimated value fluctuates significantly, it is not considered stable even if the deviation is small. Assuming that the system sets the sliding rate of change to no more than 5%, the stability of the estimated value can be further verified. If it meets the requirements, the current second residual is valid.

[0122] Finally, the system outputs the first residual (12.5%) and the second residual (for example, 0.04) as the residual output results of the smart meter number M9012 identified as abnormal in the current continuous time window, and records them in a structured manner for subsequent abnormality classification and judgment.

[0123] This implementation path not only completely closes the loop on data dependencies, ensuring that the primary residual depends on the estimated value and the original electrical parameter data, and the secondary residual depends on the neighboring estimated values ​​and the topological structure, but also avoids the risk of misjudgment that "high error means anomaly" in existing technologies through standardized error and stability control mechanisms. Its residual model is not a static numerical comparison, but rather an intelligent power grid judgment logic that integrates physical structure, data correlation, and dynamic sampling behavior.

[0124] Furthermore, the method of generating estimated electric parameter values ​​by calling original electric parameter data of other smart meters for the smart meter number identified as abnormal includes:

[0125] In a current continuous time window, selecting, from the cross-device measurement data set, a plurality of smart meter numbers that have a finite path distance from the smart meter number identified as abnormal in the topological path directed distance matrix, and extracting raw electrical parameter data of these smart meter numbers in a primary conductive parameter dimension to form a topologically adjacent electrical parameter data set;

[0126] Calculating a correlation weight vector based on the Pearson correlation coefficient between the main conductive parameters corresponding to the numbers of each topologically adjacent smart meter and the number of the smart meter identified as abnormal in multiple historical continuous time windows, so as to determine the linear combination contribution coefficient of each adjacent meter to the estimated electrical parameter value;

[0127] Combining the electrical wiring information with the topological path directed distance matrix, determining the path directionality and level difference of each adjacent smart meter number relative to the smart meter number identified as abnormal, and constructing a topological proportional correction vector for adjusting the linear combination contribution coefficient according to the path direction, wherein the adjustment coefficient of the upstream node is less than 1 and the adjustment coefficient of the downstream node is greater than 1;

[0128] The correlation weight vector and the topology proportion correction vector are element-wise multiplied and fused to obtain a weighted correction coefficient vector, and the weighted correction coefficient vector is used as a weighting factor to perform a weighted sum operation on the topological adjacent electrical parameter data set to generate the estimated electrical parameter value of the smart meter number identified as abnormal within the current continuous time window.

[0129] In smart distribution networks, raw electrical parameter data from some smart meters may experience sudden changes, abnormal offsets, or temporary failures due to factors such as wiring errors, reverse current connection, local disturbances, meter inaccuracies, or communication anomalies. To improve the system's ability to identify and tolerate these abnormal electrical parameters, a mechanism for generating estimated electrical parameters that incorporates physical constraints, historical correlations, and topological interpretation is essential, especially in operating environments where manual verification is lacking or where external measurement equipment cannot be relied upon for redundant measurement.

[0130] This method integrates the temporal data associations, physical wiring relationships, and topological path characteristics of multiple smart meters to achieve a robust, dynamic, and non-line-loss electrical parameter estimation method. When a smart meter is identified as abnormal, this method uses the raw electrical parameter data of its neighboring smart meters, combined with historical behavioral similarities and path structure differences, to reconstruct the estimated electrical parameter values ​​for the meter in the current period. This provides input support for subsequent residual analysis, anomaly classification, and diagnosis.

[0131] In actual deployment, the system first identifies the anomalous smart meter ID and locates the continuous time window within it, following the previous processing flow. Assume that the current processing target is smart meter M4056, and the time window is from 17:30 to 17:45 on November 3, 2025. At this point, it is necessary to construct estimated electrical parameter values ​​for M4056's main conductive parameters during this time window. To do this, the system selects several smart meter IDs from the cross-device measurement data set that have valid physical proximity to M4056 in the directed distance matrix of the topological path.

[0132] The so-called "topological path directed distance matrix" is a predefined structural matrix that reflects the wiring paths and level differences between each smart meter in the substation based on electrical wiring information, load connection diagram and communication configuration information. Each element in the matrix It represents the number of node hops or electrical path length required to reach smart meter j along the distribution network structure from smart meter i. For example, if M4056 is a backbone node, its neighboring nodes may include M4052, M4054, M4060, and so on, with path distances of 1, 2, and 1, respectively.

[0133] From this matrix, the system retrieves all smart meter numbers whose topological distance from M4056 within the time window is less than or equal to a threshold (e.g., ≤2). Assume that four numbers are screened: M4052, M4053, M4054, and M4060. The system then extracts the raw electrical parameter data for these four meters in the primary conductive parameter dimension from the cross-device measurement data set. For example, if the primary conductive parameter is "active power," its raw sampled values ​​are extracted every minute between 5:30 PM and 5:45 PM, forming a dataset consisting of multiple sample sequences corresponding to the numbers, i.e., the topologically adjacent electrical parameter data set.

[0134] However, simply averaging the raw data from these neighboring meters ignores historical behavioral characteristics and structural positional differences, making it difficult to reflect the system's true response. Therefore, the present invention further introduces historical data correlation as a control factor for estimation weights. Beyond the current window, the system retrieves the raw electrical parameter sequences of these neighboring meters and the target meter M4056 in the primary conductive parameter dimension from multiple consecutive time windows within the past day, week, or defined period, and calculates the Pearson correlation coefficient through point-by-point matching.

[0135] The Pearson correlation coefficient r reflects the degree of linear correlation between two time series and has a range of [–1, +1]. To avoid directional interference, the absolute value is taken and normalized to [0, 1] to form a correlation weight vector. Assuming that the correlation coefficients of M4052, M4053, M4054, and M4060 with M4056 are 0.83, 0.51, 0.93, and 0.70, respectively, after normalization, they can be set to {0.30, 0.18, 0.34, 0.26}, which serve as the contribution weights of each adjacent meter in the linear combination.

[0136] Next, to further improve the accuracy of the estimated electrical parameter values' response to the actual physical path, the system needs to combine the wiring pattern with the directed distance matrix of the topological path to construct a proportional correction mechanism. This mechanism aims to introduce directional differentiation: the measured value of the upstream node should be appropriately weakened when transmitted to the target meter, and the response value of the downstream node should be proportionally compensated for load distribution or convergence effects.

[0137] Based on the actual topology information, the system can establish a set of path direction determination rules. For example, if the path from node i to node j is downstream of node i, it is recorded as a "forward path", otherwise it is a "reverse path". The longer the path distance, the more obvious the signal attenuation. The proportional adjustment coefficient is defined based on the path distance, for example:

[0138] Path distance = 1, upstream correction factor = 0.95, downstream correction factor = 1.05;

[0139] Path distance = 2, upstream correction factor = 0.90, downstream correction factor = 1.10;

[0140] The system then performs an element-by-element multiplication of the topology scale correction vector with the previously obtained correlation weight vector to obtain the final weighted correction coefficient vector. For example, M4052 is an upstream node on path 1; M4053 is upstream on path 2; M4054 is a parallel node; and M4060 is downstream on path 1. Their correction vectors are {0.95, 0.90, 1.00, 1.05}, respectively. Consequently, the weighted correction coefficient vector is {0.285, 0.162, 0.340, 0.273}.

[0141] In generating the final estimated electrical parameter values, the system uses the weighted correction coefficient vector as the weighting factor to perform a weighted summation of the main electrical parameter values ​​of the four adjacent meters at each time point. For example, at 17:32, the sampling values ​​of each meter are:

[0142] M4052: 3.5kW, M4053: 3.1kW, M4054: 3.6kW, M4060: 3.4kW.

[0143] The estimated electrical parameter values ​​of M4056 at 17:32 are:

[0144] 3.5×0.285 + 3.1×0.162 + 3.6×0.340 + 3.4×0.273 =0.9975 + 0.5022 +1.224 + 0.9282 = 3.6519 kW;

[0145] In this way, the system can generate a complete estimated electrical parameter sequence of M4056 in the entire continuous time window and write it into the numbered index structure for subsequent residual construction steps to call.

[0146] It is worth noting that this estimation method is fundamentally different from the "global averaging" method widely used in existing technologies. In existing technologies, the value of each meter is often estimated by directly subtracting the sum of the branch meters from the total meter data, or by uniformly averaging the data of all smart meters. This ignores the actual electrical location, load response characteristics, and historical synergy between meters. This method not only considers proximity but also uses historical synergy as a weight and topological path directionality as a proportional correction, making the estimated value highly personalized and physically interpretable.

[0147] Furthermore, this method is easy to deploy at the implementation level. Correlation coefficient calculations can be preprocessed using a sliding window on historical data. Topological path matrices and electrical wiring information can be generated by power system topology modeling tools and adjusted in real time as the system is updated. Various weighted operations and proportional corrections can also be efficiently implemented at the database level or in edge computing nodes, supporting real-time processing capabilities across large-scale substations.

[0148] In summary, this estimation method provides an electrical parameter reconstruction mechanism with a clear structure, traceable logic, and strong dynamic response capabilities. It is particularly suitable for the task of identifying anomaly in smart meters with multi-source coordination at the substation level. This method effectively reduces recognition errors caused by sample loss, sensor drift, or data communication interruptions, while providing reliable and physically consistent input data for the construction of the first and second residuals, significantly improving the robustness and accuracy of anomaly identification.

[0149] Step S106: Determine whether there is a wiring error, current reverse connection, metering abnormality or other abnormality type based on the consistency index, the first residual and the second residual, and output an abnormality recognition result record including the abnormality type, the smart meter number and the continuous time window.

[0150] During the implementation of step S106, the system needs to comprehensively utilize the three core analysis results of consistency index, first residual and second residual obtained in the previous stage to perform multi-dimensional cross-validation and logical attribution, so as to clarify whether the anomaly actually exists, and further judge its type and nature, and finally generate a structured anomaly recognition result output.

[0151] In this step, the system must set multidimensional judgment rules. For example, normalizing the three indicators mentioned above to the same dimension and aggregating them using a weighted scoring function can also be done. A three-dimensional threshold model can also be constructed, such as automatically determining a high-confidence anomaly when the consistency index is less than 0.6, the first residual is greater than ±15%, and the second residual is less than ±5%. Each judgment threshold should be dynamically adjusted based on the actual load characteristics of the substation, the measurement accuracy level of the equipment, and operation and maintenance experience.

[0152] After identifying abnormal behavior, the system needs to further combine the response offset feature vector, device type, electrical parameter category and topology information to perform abnormal type identification operations. For example:

[0153] If the current direction is reversed, the power sign is reversed, the consistency index is normal, and the first residual is prominent, it is likely that the current transformer is connected in reverse;

[0154] If the consistency index of one of the three phases A, B, and C suddenly drops, the power change amplitude is abnormally high, and the time lag is significant, it may be a single-phase wiring error or phase sequence disorder;

[0155] If all indicators deviate, the estimated value differs greatly from the neighboring meter, or even the estimation itself is unstable, it may be due to device configuration errors, meter address mapping errors, or communication channel confusion;

[0156] If all indicators fluctuate only slightly within a specific time window, it can be judged as a short-term disturbance or transient load jump, which is an ignorable event to avoid false alarms.

[0157] Finally, the system will format and output the recognition results, including the abnormality type (such as reverse connection, wrong phase, misalignment), meter number, start and end time window, main conductive parameter type, actual measurement value, estimated value, residual value and confidence level, etc., and write it into the abnormality recognition result record table, which will be synchronously pushed to the main station system or abnormal data alarm platform for further processing by the dispatching system, operation and maintenance personnel or data analysis module.

[0158] To facilitate understanding of the application process of step S106, a specific example is provided below.

[0159] For example, an industrial area has multiple smart meters deployed, including M6011 as the main incoming meter and M6012 and M6013 as sub-meters for parallel branch users. During the time window from 2:30 PM to 2:45 PM on August 10, 2025, the system identified abnormal behavior in M6012 in step S104. The phase A current curve of this meter during this window was completely out of sync with the other meters, with a response offset of –3 sampling points and a consistency index of 0.35, far below the system's normal threshold of 0.8.

[0160] After entering step S105, the system retrieves the raw electrical parameter data from M6011 and M6013 to estimate the main conducting parameter (phase A active power) for M6012 during that period. The estimated value is 4.10 kW, while the actual measured value for M6012 is 5.02 kW. The resulting first residual is +0.92 kW, representing 22.4% of the estimated value. This residual is significantly higher than the ±15% deviation threshold set by the system, constituting a significant deviation. Furthermore, the system further examines the average deviation between this estimated value and the estimated sequence for M6011 and M6013, finding only ±0.03 kW. This indicates a high degree of stability in the estimated value itself, leading to a low second residual, eliminating the influence of modeling errors.

[0161] After entering step S106, the system performs a cross-judgment based on the aforementioned results. First, the consistency index is extremely low (0.35), the first residual is significantly out of limit (+22.4%), and the second residual is extremely small, meeting the "high confidence anomaly" condition. Subsequently, the system further analyzes the sign of the meter's current active power and finds that its power direction has changed from positive to negative, and the current direction has also reversed. The system confirms in the topology diagram that the load connected to the meter is a unidirectional motor. In theory, there should be no reverse power phenomenon, eliminating the possibility of load feedback. Therefore, based on preset rules, combined with the characteristic pattern of abnormal power direction and reverse current, the system determines that the event is a reverse connection of the current transformer.

[0162] The system formats and outputs the above identification results, including: anomaly type: "reverse current connection," smart meter ID: M6012, anomaly start and end time window: "2025-05-10 14:30–14:45," the main conducting parameter: "Phase A active power," actual value: 5.02 kW, estimated value: 4.10 kW, first residual: +0.92 kW, residual percentage: +22.4%, consistency index: 0.35, response offset: -3 sampling points, and confidence level: "high." This result is written into the anomaly identification result table and uploaded to the master station system, where it is used by the dispatch system to issue maintenance instructions.

[0163] From this example, it can be seen that the comprehensive determination logic and classification identification mechanism described in step S106 can be directly embedded in the existing distribution automation platform to support local abnormality self-diagnosis or centralized alarm linkage mechanism.

[0164] Furthermore, determining whether there is a wiring error, current reverse connection, measurement abnormality, or other abnormality type based on the consistency index, the first residual, and the second residual includes:

[0165] Extract the numerical change consistency index, first residual, and second residual corresponding to each smart meter number identified as abnormal within the current continuous time window, and construct an abnormal quantitative feature vector for the smart meter number within the continuous time window. The abnormal quantitative feature vector includes three standardized indicator values, which are used to comprehensively measure the degree of behavioral deviation, estimation stability, and group consistency;

[0166] Based on the response offset feature vector, the response delay and topology offset distance corresponding to the smart meter number in the current continuous time window are extracted. In combination with the sign change trend of the main conductive parameter in the continuous time window, a response behavior feature label of the smart meter number in the continuous time window is constructed to reflect its time consistency state relative to the physical topology transmission logic;

[0167] The abnormality quantization feature vector and the response behavior feature label are used as input items and substituted into a preset abnormality type discrimination rule set. The abnormality type discrimination rule set is structured by the system to define the combined mapping relationship between the main conductive parameter change law, electrical wiring information, response offset direction and error distribution characteristics, and is used to achieve attribution matching of wiring errors, current reverse connection, measurement abnormalities or other abnormality types;

[0168] The matching result of the abnormality type discrimination rule set is used as the judgment output to determine whether the smart meter number has a wiring error, current reverse connection, measurement abnormality or other abnormality types within the current continuous time window.

[0169] In actual engineering implementation, it is first necessary to take the number of the smart meter identified as "abnormal" as the processing object, relying on the processing structure established in the previous steps, especially the following three key quantities generated in each continuous time window: First, the numerical change consistency index, which reflects the operational coordination of the target meter relative to other smart meters, that is, whether it presents a similar change trend with other meters; second, the first residual, that is, the standardized deviation between the original measured value and the estimated electrical parameter value of the main conductive parameter of the target meter in the time window, which characterizes the degree of abnormality of its measurement behavior relative to the predicted behavior; third, the second residual, which is used to evaluate whether the estimated electrical parameter value of the meter is reliable, that is, whether there is a deviation from the estimated value of the adjacent meter on the topological path. If the second residual is large, the current estimated value may be from an unstable source or the surrounding load fluctuates violently.

[0170] The system combines these three indicators into a unified format for anomaly quantification. For example, for smart meter number M8032, the current continuous time window is set to 2:00 PM–2:15 PM on October 25, 2025. Within this window, the system extracts a consistency index of 0.41, a primary residual of 18.2%, and a secondary residual of 4.7%. These three values ​​are then combined into a three-dimensional feature vector: {0.41, 0.182, 0.047}. This is then combined with the smart meter number M8032 and the time window ID to form a structured record for subsequent attribution.

[0171] Next, the system retrieves the response offset feature vector, constructed through inflection point extraction, response delay calculation, and topological path mapping. Taking the M8032 as an example, its primary conducting parameter is the phase B current. Within the current time window, the system has calculated that its response delay is –2 sampling points, indicating an early response, a topological offset distance of 1, and that the meter should be physically located in the downstream path. The system then further analyzes the sign change trend of the primary conducting parameters. For example, if the phase B current changes from positive to negative, and the active power changes from positive to negative, this indicates a clear sign of power reversal. The system then constructs a response behavior feature label for this meter: "Early response + Power reversal + Negative topological offset." This label is used to infer whether the meter's time response conforms to the actual electrical structure transmission logic.

[0172] After completing the above two feature constructions, the system inputs the abnormality quantification feature vector and the response behavior feature label into the abnormality type discrimination rule set. This rule set is a decision mapping matrix summarized through a large number of sample training and simulation models. The core basis includes:

[0173] If the consistency index is low, the first residual is greater than the set threshold (such as 15%), the second residual is small (such as less than 5%), the response delay is less than 0 (early response), the power direction is reversed, and the topology offset is negative, then it is determined that the current transformer is reversely connected;

[0174] If the consistency index is extremely low (e.g. <0.2), and only one of the three phases has a significant deviation in the main conducting parameters, and the response is severely delayed, then combined with the wiring method analysis, it can be determined that there is a single-phase wiring error or phase mismatch;

[0175] If the consistency index is moderate (0.4-0.7), the first residual exceeds 20%, but the second residual fluctuates greatly (greater than 10%), and the estimated value of the main conductive parameter itself fluctuates dramatically (large sliding standard deviation), then it is determined that the estimated path is unstable or the anomaly is caused by an adjacent disturbance and cannot be directly classified as an equipment abnormality.

[0176] If all three indicators are near the critical values, but intermittent deviations occur for a long time, and the response behavior is normal, then the preliminary judgment is that the measurement of the main conductive parameters is inaccurate, which needs to be confirmed in combination with long-term behavior.

[0177] The rule set is not a single threshold judgment, but rather a set of logical decisions based on a combination of indicators. The system can use multidimensional Boolean rule chains, fuzzy membership functions, or small interpretable rule network models for matching. Each rule ultimately outputs an anomaly type label, limited to the aforementioned "wiring error," "reverse current," "metering anomaly," or "other anomaly type."

[0178] Based on the judgment results, the system writes the label "reverse current" into the structured identification record of M8032, and packages and binds its main conductive parameter as "phase B current" and the response behavior as "early response + power direction reversal", providing a complete traceable evidence chain for subsequent result output and operation and maintenance processing.

[0179] Through the above approach, this embodiment not only constructs a complete closed-loop structure of "anomaly quantification - response behavior - attribution matching" from the data perspective, but more importantly, it breaks through the limitations of traditional meter anomaly identification methods that only rely on numerical deviations or fixed rule judgments. It introduces the three dimensions of topological structure, response behavior directionality, and estimation confidence to achieve truly explainable and attributable anomaly classification and identification.

[0180] A second embodiment of the application provides an electronic device, comprising:

[0181] processor;

[0182] The memory is used to store a program. When the program is read and executed by the processor, it executes a method for intelligent fusion and abnormality identification of multi-source electricity meter data provided in the first embodiment of the present application.

[0183] The third embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the method for intelligent fusion and anomaly identification of multi-source electricity meter data provided in the first embodiment of the present application is executed.

[0184] Although the present application is disclosed as above with the preferred embodiments, it is not intended to limit the present application. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A method for intelligent fusion and anomaly identification of multi-source electricity meter data, characterized in that: include: Collecting raw electrical parameter data from multiple heterogeneous smart meters under a unified time base, and organizing the raw electrical parameter data into a cross-device measurement data set indexed by smart meter number; Classify and label each raw electrical parameter data in the cross-device measurement data set according to the electrical wiring information and functional parameters of each smart meter, and generate correlation information between the electrical parameters; Segmenting the cross-device measurement data set into continuous time windows, extracting the electrical parameter change curve of each smart meter, and calculating the value change consistency index within each time window based on the correlation information to obtain a consistency index sequence; Calculating the response offset characteristic vector between the electrical parameter change curves, performing a joint judgment in combination with the consistency index, and identifying the abnormal smart meter number and its corresponding continuous time window; For the smart meter number identified as abnormal, the original electrical parameter data of other smart meters are called to generate estimated electrical parameter values, and a first residual between the estimated value and the original value, and a second residual between the estimated value and the adjacent estimated values ​​are calculated based on the main conductive parameters; According to the consistency index, the first residual and the second residual, determine whether there is a wiring error, current reverse connection, metering abnormality or other abnormality type, and output an abnormality identification result record including the abnormality type, smart meter number and continuous time window.

2. The method for intelligent fusion and anomaly identification of multi-source electric meter data according to claim 1 is characterized in that: The calculating of the response offset characteristic vector between the electrical parameter change curves includes: In a cross-device measurement data set segmented by continuous time windows, for each continuous time window, the electrical parameter change curves of all smart meters under the marked main conductive parameter dimension are extracted, and the main response inflection point sequence in each electrical parameter change curve is identified based on the correlation information between the electrical parameters; Calculate the response delay of each smart meter within the current continuous time window based on the relative position difference between the main response inflection point sequence corresponding to each smart meter number and the main response inflection point sequence of a reference smart meter, where the reference smart meter is the smart meter with the highest value change consistency index within the current continuous time window; Combined with the electrical wiring information corresponding to each smart meter in the association information, a topological path directed distance matrix is ​​constructed, and based on the constructed topological path directed distance matrix, the response delay of each smart meter in the current continuous time window is converted into a topological offset distance; The response delay of each smart meter in the current continuous time window and its topological offset distance are jointly encoded to generate a response offset feature vector indexed by the smart meter number.

3. The method for intelligent fusion and anomaly identification of multi-source electric meter data according to claim 2, characterized in that: The step of calculating a first residual between an estimated value and an original value, and a second residual between an estimated value and an adjacent estimated value, based on the main conductive parameter, includes: Extracting, within the current continuous time window, original electrical parameter data of the main conductive parameter corresponding to the smart meter number identified as abnormal based on the generated estimated electrical parameter value, and calculating a difference between the original electrical parameter data and the estimated electrical parameter value; and obtaining a standardized residual value of the smart meter number within the continuous time window by dividing the difference by the estimated electrical parameter value, which is used as a basis for constructing a first residual; Calling the topological path directed distance matrix, screening multiple smart meter numbers that are topologically adjacent to the smart meter number identified as abnormal from the cross-device measurement data set, obtaining estimated electrical parameter values ​​generated based on the main conductive parameters within the continuous time window, and performing weighted combination according to the topological path directed distance to construct a set of adjacent estimated values ​​within the continuous time window; A weighted difference calculation is performed on the estimated electrical parameter value of the smart meter number identified as abnormal within the continuous time window and the set of neighboring estimated values. The weighting coefficient is determined by the inverse proportion of the directed distance of the topological path, and combined with the sampling stability index of the main conductive parameter, a second residual of the smart meter number within the continuous time window is generated to characterize the degree of deviation of the estimated electrical parameter value relative to the set of neighboring estimated values.

4. The method for intelligent fusion and anomaly identification of multi-source electric meter data according to claim 3 is characterized in that: The determining, based on the consistency index, the first residual, and the second residual, whether there is a wiring error, current reverse connection, measurement abnormality, or other abnormality type includes: Extract the numerical change consistency index, first residual, and second residual corresponding to each smart meter number identified as abnormal within the current continuous time window, and construct an abnormal quantitative feature vector for the smart meter number within the continuous time window. The abnormal quantitative feature vector includes three standardized indicator values, which are used to comprehensively measure the degree of behavioral deviation, estimation stability, and group consistency; Based on the response offset feature vector, the response delay and topology offset distance corresponding to the smart meter number in the current continuous time window are extracted. In combination with the sign change trend of the main conductive parameter in the continuous time window, a response behavior feature label of the smart meter number in the continuous time window is constructed to reflect its time consistency state relative to the physical topology transmission logic; The abnormality quantization feature vector and the response behavior feature label are used as input items and substituted into a preset abnormality type discrimination rule set. The abnormality type discrimination rule set is structured by the system to define the combined mapping relationship between the main conductive parameter change law, electrical wiring information, response offset direction and error distribution characteristics, and is used to achieve attribution matching of wiring errors, current reverse connection, measurement abnormalities or other abnormality types; The matching result of the abnormality type discrimination rule set is used as the judgment output to determine whether the smart meter number has a wiring error, current reverse connection, measurement abnormality or other abnormality types within the current continuous time window.

5. The method for intelligent fusion and anomaly identification of multi-source electric meter data according to claim 4 is characterized in that: The method of generating estimated electric parameter values ​​by calling original electric parameter data of other smart meters for the smart meter number identified as abnormal includes: In a current continuous time window, selecting, from the cross-device measurement data set, a plurality of smart meter numbers that have a finite path distance from the smart meter number identified as abnormal in the topological path directed distance matrix, and extracting raw electrical parameter data of these smart meter numbers in a primary conductive parameter dimension to form a topologically adjacent electrical parameter data set; Calculating a correlation weight vector based on the Pearson correlation coefficient between the main conductive parameters corresponding to the numbers of the topologically adjacent smart meters and the number of the smart meter identified as abnormal in multiple historical continuous time windows, so as to determine the linear combination contribution coefficient of each adjacent meter to the estimated electrical parameter value; Combining the electrical wiring information with the topological path directed distance matrix, determining the path directionality and level difference of each adjacent smart meter number relative to the smart meter number identified as abnormal, and constructing a topological proportional correction vector for adjusting the linear combination contribution coefficient according to the path direction, wherein the adjustment coefficient of the upstream node is less than 1 and the adjustment coefficient of the downstream node is greater than 1; The correlation weight vector and the topology proportion correction vector are element-wise multiplied and fused to obtain a weighted correction coefficient vector, and the weighted correction coefficient vector is used as a weighting factor to perform a weighted sum operation on the topological adjacent electrical parameter data set to generate the estimated electrical parameter value of the smart meter number identified as abnormal within the current continuous time window.

Images