Analytical Model and Verification Method for the Hierarchical Relationship of Electricity Meters

Abstract

This invention discloses an analysis model and verification method for the hierarchical relationship of electricity meters, relating to the field of energy management technology. The model includes: a high-frequency electricity meter data acquisition module that uses a distributed architecture to collect electricity meter data in batches and in parallel, encrypts and cleans it, and stores it in a time-series database; a spatiotemporal dynamic graph construction module that dynamically constructs edges using electricity meters as nodes; a spatiotemporal dynamic graph convolutional network analysis module that inputs the graph and outputs the association probability; a hierarchical relationship verification module that verifies and corrects relationships from multiple dimensions; and an active feature blind search confirmation module that generates features to blindly search and confirm associations when no data is available. This invention constructs a spatiotemporal dynamic graph combined with a graph convolutional network, which can automatically capture complex electricity meter relationships, accurately and quickly identify hierarchical relationships, and has strong parallel capabilities and good adaptability when processing large-scale data. Simultaneously, it eliminates false judgments through multi-dimensional verification and uses active feature blind search confirmation when no auxiliary data is available, exhibiting excellent robustness and flexibility, providing strong support for power grid operation and maintenance and safe operation.

Description

Technical Field

[0001] This invention relates to the field of energy management technology, specifically to an analysis model and verification method for the hierarchical relationship of electricity meters. Background Technology

[0002] With the rapid development of smart grids, the scale and complexity of power grids have significantly increased. As the core terminal for power data collection, the accurate identification of the hierarchical relationships of electricity meters has become crucial for ensuring the safe operation of the power grid, optimizing resource allocation, and achieving accurate line loss analysis. Traditional power grid topology management mainly relies on manual inspections and static records. However, in the face of dynamically changing power grid structures and massive amounts of electricity meter data, manual methods are difficult to maintain relationships in real time and accurately. At the same time, the access of new loads such as distributed energy and electric vehicle charging piles has further exacerbated the dynamism and uncertainty of the power grid, resulting in low efficiency and high error rates when traditional methods deal with complex electrical connections. Therefore, there is an urgent need for an automated and intelligent technical means to achieve dynamic perception and accurate verification of electricity meter relationships.

[0003] Traditional meter association identification technologies suffer from several significant drawbacks: First, they rely on manual inspection and static topology files. When the power grid structure changes dynamically, data needs to be updated manually, leading to information lag and susceptibility to errors. Second, single-dimensional verification methods are easily affected by data noise or random factors. For example, two independent meters may be misjudged as associated due to similar electricity usage habits, even when there is no actual electrical connection. Third, they lack proactive detection capabilities. In areas without auxiliary data or with unknown topology, traditional methods cannot effectively confirm associations and require on-site verification, which is costly and inefficient. Furthermore, traditional technologies struggle to handle large-scale power grid data. When dealing with tens of thousands of meters, the computational complexity increases exponentially, failing to meet real-time requirements. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an analysis model and verification method for the hierarchical relationship of electricity meters. By collecting electricity meter data at high frequency, a dynamic graph structure is constructed. Combined with graph attention mechanism and gated loop unit, spatiotemporal features are extracted to achieve intelligent identification of the relationship. The model introduces a multi-dimensional verification mechanism, including load similarity, power conservation and abnormal event transmission analysis, and has the ability to actively search for features blindly. It can efficiently handle complex relationship problems in dynamic power grid environments.

[0005] To solve the above-mentioned technical problems, this invention provides the following technical solution: Firstly, an analysis model for the hierarchical relationship of electricity meters, comprising: a high-frequency electricity meter data acquisition module: adopting a distributed architecture, deploying high-speed communication acquisition terminals, and batch parallelly acquiring real-time and historical electricity consumption data from electricity meters. After encrypted transmission and cleaning, the data is stored in a time-series database in a standard format containing timestamps, meter identifiers, and multi-dimensional parameters; and a spatiotemporal dynamic graph construction module: using electricity meters as nodes, assigning them real-time electricity consumption parameters, historical trend characteristics, and their own attributes, dynamically constructing edges based on potential electrical connections, determining edge weights for known topological regions according to cable parameters, and constructing edges for unknown regions according to electricity load correlation, updating node attributes every 5 minutes. With edge weights; Spatiotemporal dynamic graph convolutional network analysis module: Inputs the spatiotemporal dynamic graph into a network model containing spatial and temporal feature extraction layers and a fusion output layer. The spatial layer uses a graph attention mechanism to aggregate neighbor information, the temporal layer uses gated recurrent units to model time-series features, and the fusion layer outputs the probability of meter hierarchical association. The model is trained using labeled historical topology data; Hierarchical relationship verification module: Collects auxiliary data such as power grid topology to form a candidate association list, verifies from multiple dimensions such as load curve similarity, power conservation, and abnormal event conduction, corrects and outputs a confirmed hierarchical relationship list; Active feature blind search confirmation module: Defines the meter detection range when there is no auxiliary data, and generates harmonic injection and pulse in stages. Based on the characteristics of electricity consumption and phase shift, and combined with the collected data, a blind search and narrowing process is used to confirm the relationship between upper and lower level meters. The specific process of blindly searching and narrowing the range based on the collected data to confirm the relationship between upper and lower level meters is as follows: First stage: Within the defined meter detection range, pulse electricity consumption characteristics are generated for candidate lower level meters, and power sequence data of all upper level candidate meters are collected simultaneously through the meter data high-frequency acquisition module; the collected power sequence data is compared with a preset pulse characteristic template, and upper level candidate meters with a signal matching the pulse characteristic in the power sequence are selected, while meters without a matching signal are eliminated, completing the first round of range narrowing; Second stage: For the upper level candidate meters after the first round of narrowing, and the corresponding pulse characteristics are compared with the pulse characteristics of the lower level meters, the upper level candidate meters with the pulse characteristics are selected, and the meters without a matching signal are eliminated, completing the first round of range narrowing; For the lower-level meters, harmonic injection characteristics are created, and current signal data from the upper-level candidate meters is collected through the high-frequency acquisition module. Spectral analysis is performed on the collected current signal data to screen out upper-level meters exhibiting harmonic injection characteristics in the spectrum, while meters without detected harmonic characteristics are eliminated, further narrowing the range. In the third stage, for the further narrowed meter pairs, phase shift characteristics are created for the lower-level single-phase meters, and power factor data from the upper-level meters is collected. The changes in the power factor of the upper-level meters are monitored, and meter pairs exhibiting abrupt changes in power factor corresponding to the phase shift characteristics are screened out, while meter pairs without corresponding abrupt changes are eliminated, ultimately confirming the upper-lower level correlation between the meters.

[0006] Furthermore, in the high-frequency data acquisition module of the electricity meter, the high-speed communication acquisition terminal is an intelligent acquisition device based on 5G or fiber optic communication. Each terminal is connected to the electricity meter via RS-485 bus or wireless LoRa technology. The acquired real-time electricity consumption data includes voltage, current, active power, reactive power, power factor, and frequency. The historical electricity consumption data includes the daily electricity consumption curve, monthly total electricity consumption, seasonal electricity consumption characteristics data, and electricity consumption records during extreme weather periods for the past 5 years. The electricity meter identification includes the electricity meter number, installation location, and transformer substation affiliation information.

[0007] Furthermore, the specific steps for constructing the spatiotemporal dynamic graph in the spatiotemporal dynamic graph construction module are as follows: First, determine the node set, using a single electricity meter as the basic node of the graph structure, and assign multi-dimensional attribute features to each node, including real-time collected 30-second-level electricity consumption parameters, historical electricity consumption trend features, and the meter's own attributes. Second, employ a spatiotemporal feature attenuation fusion algorithm to fuse the real-time collected features of the node with historical features to obtain the spatiotemporal fusion features of the node. Third, when constructing edges between nodes, for areas with known transformer topology, determine the existence of edges based on cable connection paths, using cable length, diameter, material, and laying method as the basic weight parameters of the edges. Fourth, for areas with unknown topology, calculate the correlation coefficient of the meter to the electricity consumption parameters using a time-series correlation algorithm for electricity consumption parameters; when the absolute value of the correlation coefficient is higher than a set threshold, generate temporary edges. Fifth, employ a two-factor dynamic edge weight algorithm, combining the spatiotemporal fusion feature similarity of nodes with inherent connection coefficients, to calculate the dynamic weights of edges between nodes. Sixth, update the node attribute values ​​every 5 minutes based on the latest collected electricity consumption data, and synchronously adjust the edge weight parameters based on load fluctuations to achieve dynamic updates of the spatiotemporal dynamic graph.

[0008] Furthermore, in the spatiotemporal dynamic graph construction module, the calculation formula for the spatiotemporal feature attenuation fusion algorithm is as follows: ,in, Indicates electricity meter At any moment The spatiotemporal fusion characteristics, It is an electricity meter At any moment Real-time acquisition of feature vectors, It is an electricity meter At any moment Historical feature vectors These are real-time feature weight coefficients. It is the historical characteristic attenuation coefficient. It is the length of the historical time window. It's the meter indicator. It is a timestamp parameter.

[0009] Furthermore, in the spatiotemporal dynamic graph construction module, the calculation formula for the time-series correlation algorithm of power parameters is as follows: ,in For electricity meters With electricity meter The correlation coefficient of electricity consumption parameters For electricity meters At any moment The power consumption parameter values, For electricity meters At any moment The power consumption parameter values, For electricity meters In the time window Average power consumption parameters within the area For electricity meters In the time window Average power consumption parameters within the area The time window length for calculating the correlation coefficient, This is the timestamp parameter, with a set threshold of 0.65. At that time, at the electricity meter With electricity meter Temporary edges are generated between them.

[0010] Furthermore, in the spatiotemporal dynamic graph construction module, the calculation formula for the two-factor dynamic edge weight algorithm is as follows: Sim Con ,in, It is a moment Electricity meter With electricity meter Dynamic weights of edges between them, Sim It is a similarity calculated based on the spatiotemporal fusion features of nodes. , Electricity meters , At any moment The spatiotemporal fusion characteristics, It is the inherent connection coefficient between the two meters. It is the feature similarity weight coefficient. , These are the identifiers of the two meter nodes whose edge weights are to be calculated.

[0011] Furthermore, in the spatiotemporal dynamic graph convolutional network analysis module, the network model specifically includes: a spatial feature extraction layer employing an enhanced graph attention mechanism, setting three attention heads, each with a hidden layer dimension of 64, and calculating the attention weights between nodes and aggregating neighbor node information through a graph attention weight algorithm. This algorithm combines the physical distance of the electricity meter with the similarity of electricity consumption characteristics to determine the attention weights; a temporal feature extraction layer employing an improved gated recurrent unit, setting a two-layer recurrent structure, with each hidden layer having a hidden layer dimension of 128, performing temporal modeling on the historical electricity consumption sequence of each node, with the input being the spatiotemporal fusion feature sequence of the node; a fusion output layer including a spatiotemporal feature fusion sublayer and an association probability calculation sublayer, the spatiotemporal feature fusion sublayer performing weighted fusion of the output features of the spatial feature extraction layer and the temporal feature extraction layer through a multi-layer attention mechanism, setting a spatiotemporal attention weight matrix; and an association probability calculation sublayer containing two fully connected layers, the first layer having 256 neurons and the second layer having 64 neurons, compressing and mapping the fused features through the fully connected layers, and using a feature concatenation logistic regression algorithm to output the upper and lower level association probability values ​​between electricity meters.

[0012] Furthermore, in the spatiotemporal dynamic graph convolutional network analysis module, the calculation formula for the graph attention weight algorithm is as follows: ,in, It is an electricity meter Its neighboring nodes Attention weights This is the attention weight matrix. Represents a node and The spatiotemporal fusion feature vectors are concatenated, LeakyReLU is used as the activation function, and then the neighbor node information is weighted and aggregated based on this attention weight. It is an electricity meter The neighbor node identifier, It is an electricity meter The set of neighboring nodes, Electricity meters , At any moment The spatiotemporal fusion characteristics.

[0013] Furthermore, in the spatiotemporal dynamic graph convolutional network analysis module, the calculation of the feature concatenation logistic regression algorithm is as follows: ,in It is a node and Feature concatenation after spatiotemporal convolution It is a weight vector. It is a bias term. It is an electricity meter At any moment Feature vector after spatiotemporal convolution It is an electricity meter At any moment Feature vector after spatiotemporal convolution and These are the node identifiers of the two meters whose correlation probability is to be calculated, consistent with the node definition in the spatiotemporal dynamic diagram. This is the moment for probability calculation.

[0014] On the other hand, the method for verifying the relationship between upper and lower level electricity meters includes the following steps: S100, building a candidate list: collecting power grid topology diagrams and electricity meter installation archives, combining the probability of upper and lower level electricity meter relationships output by the spatiotemporal dynamic graph convolutional network analysis module, and screening out electricity meter pairs with relationship probabilities higher than a preset probability threshold to form a candidate relationship list; S200, load similarity verification: verifying from the load curve similarity dimension, calculating the load curve similarity of each pair of upper and lower level electricity meters in the candidate relationship list within the same time period, using the dynamic time warping algorithm to calculate the distance between the two load curves, the smaller the distance, the higher the similarity, and if the similarity is lower than the set similarity threshold, marking the electricity meter pair as pending confirmation; S300, energy conservation verification: verifying from the energy conservation dimension, statistically analyzing the upper level electricity meters in the candidate relationship list within 1 The total power supply during the week is compared with the total power consumption of all lower-level meters corresponding to the upper-level meter in the same period. The power loss rate is calculated. If the power loss rate exceeds the reasonable loss range, the association relationship of the meter pair is re-evaluated. S400, Abnormal Conduction Verification: Verify from the dimension of abnormal event conduction. When a voltage drop or current overload occurs in the power grid, record the time difference between the detection of abnormal events by the upper and lower level meters in the candidate association list. If the time difference does not conform to the conduction law of electrical connection, it is judged that there is a problem with the association relationship of the meter pair. S500, Correction List Output: For meter pairs marked as pending confirmation or with problematic association relationships after verification in the above three dimensions, analyze the data obtained by the active feature blind search confirmation module, correct the upper and lower level association relationships of the meters, and finally output the confirmed upper and lower level relationship list.

[0015] Compared with existing technologies, the analysis model and verification method for the hierarchical relationship of electricity meters have the following advantages: First, by constructing a spatiotemporal dynamic graph and combining it with graph convolutional networks for deep analysis, the model can automatically capture complex relationship patterns between electricity meters, including direct and indirect electrical connections. Compared with traditional methods, this invention avoids the tediousness and uncertainty of manual investigation, and achieves accurate and rapid identification of the hierarchical relationship of electricity meters. Especially when processing large-scale power grid data, this model demonstrates powerful parallel processing capabilities and dynamic adaptability, and can update the relationship in real time as the power grid structure changes, providing strong technical support for the operation and maintenance management of the power grid.

[0016] Second, through comprehensive verification of multiple dimensions such as load curve similarity, power conservation, and abnormal event transmission, the model can comprehensively evaluate the correlation between electricity meters, effectively eliminating misjudgments caused by data interference or accidental factors. In addition, in the absence of auxiliary data, the model can further confirm the correlation through active feature blind search technology, demonstrating its strong robustness and flexibility. This multi-dimensional verification method not only improves the accuracy of identification, but also provides a more solid guarantee for the safe operation of the power grid.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a module architecture diagram for the analysis model of the hierarchical relationship between electricity meters. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1: Example of an analysis model for the hierarchical relationship of electricity meters.

[0022] This embodiment uses a large urban residential community as an application scenario. The community includes one main electricity meter, 12 building electricity meters, and 500 household electricity meters. The system of this invention needs to determine the hierarchical relationship between the main meter, building meters, and household meters, solving the problems of low efficiency and inaccurate topology records in traditional manual inspection. The specific system module deployment and operation process are as follows: Figure 1As shown, a distributed architecture is adopted, deploying high-speed communication data acquisition terminals in the community's power distribution room, building distribution boxes, and residents' homes. The acquisition terminals in the community's power distribution room and building distribution boxes are based on 5G communication technology, enabling high-speed, low-latency data transmission and ensuring rapid uploading of real-time electricity consumption data. The acquisition terminals in residents' homes are based on wireless LoRa technology, offering advantages such as low power consumption and wide coverage, suitable for long-term stable operation in home scenarios. All terminals are connected to their corresponding electricity meters via an RS-485 bus, ensuring the stability and reliability of data transmission.

[0023] The data acquisition terminal collects electricity meter data in batches at a frequency of 30 seconds per acquisition. Real-time electricity consumption data includes voltage, current, active power, reactive power, power factor, and frequency. This data can reflect the current operating status of the electricity meters in real time, providing a foundation for subsequent analysis of real-time correlations between electricity meters. Historical electricity consumption data retrieves daily electricity consumption curves, monthly total electricity consumption, seasonal electricity consumption characteristics, and electricity consumption records during extreme weather periods from the past 5 years. This rich historical data can help uncover long-term electricity consumption patterns of the electricity meters and improve the accuracy of correlation analysis.

[0024] The collected data is transmitted to the data processing center using the AES encryption algorithm, effectively preventing the data from being stolen or tampered with during transmission and ensuring the security of electricity consumption data. Outliers are removed through data cleaning to avoid interference from abnormal data with subsequent analysis results and ensure data quality. Finally, the data is stored in the InfluxDB time-series database in the standard format of "timestamp + meter identifier + multi-dimensional electricity consumption parameters". This database is designed specifically for time-series data and can efficiently store and query massive amounts of electricity consumption data, providing support for subsequent modules to quickly retrieve data.

[0025] The node set was determined as follows: A total of 513 electricity meters, including one main meter, 12 building meters, and 500 residential meters, were used as the basic nodes of the graph structure. Each node was assigned multi-dimensional attributes, including real-time electricity consumption parameters at the 30-second level, historical electricity consumption trends, and the meter's own attributes. By clearly defining the nodes and their attributes, a foundation was laid for constructing the relationships between the meters, ensuring that the characteristics of each meter were clearly described and avoiding biases in the association analysis due to missing information.

[0026] Node Feature Fusion: A spatiotemporal feature attenuation fusion algorithm is used to fuse the real-time collected features of each node with historical features. The calculation formula for the spatiotemporal feature attenuation fusion algorithm is as follows: ,in, Indicates electricity meter At any moment The spatiotemporal fusion characteristics, It is an electricity meter At any moment Real-time acquisition of feature vectors, It is an electricity meter At any moment Historical feature vectors These are real-time feature weight coefficients. It is the historical characteristic attenuation coefficient. It is the length of the historical time window. It's the meter indicator. This refers to timestamp parameters. For example, for electricity meters in three buildings, by combining their current real-time active power with historical active power characteristics from the past week, month, and three months, an algorithm balances the weights of real-time features and historical features from different time periods to obtain the spatiotemporal fusion characteristics of the building's electricity meters. This fusion method can reflect both the current electricity consumption status of the meters and long-term electricity consumption patterns, avoiding the interference of short-term fluctuations caused by relying solely on real-time data, or the inability to capture the latest electricity consumption changes by relying solely on historical data. This provides a more comprehensive basis for the construction of subsequent edges, supporting node features.

[0027] Construction of edges between nodes: Known topology region: The cable connection paths from the community's power distribution room to each building are clear, belonging to the known transformer area topology region. The existence of edges is determined based on cable length, diameter, material, and laying method, and these parameters are used as the basic weight parameters of the edges. Constructing edges based on the known topology and cable parameters can make full use of existing power grid information, ensure that the edge construction conforms to the actual electrical connection situation, and improve the initial accuracy of the association relationship.

[0028] Unknown Topology Area: Due to missing early construction records, the connection between residential electricity meters and building electricity meters falls within the unknown topology area. The correlation coefficient of electricity consumption parameters between residential and building electricity meters is calculated using a time-series correlation algorithm. The calculation formula for the time-series correlation algorithm is as follows: ,in For electricity meters With electricity meter The correlation coefficient of electricity consumption parameters For electricity meters At any moment The power consumption parameter values, For electricity meters At any moment The power consumption parameter values, For electricity meters In the time window Average power consumption parameters within the area For electricity meters In the time window Average power consumption parameters within the area The time window length for calculating the correlation coefficient, This is the timestamp parameter, with a set threshold of 0.65. At that time, at the electricity meter With electricity meter Temporary edges are generated between them. If a resident's electricity meter and the electricity meters of two buildings show highly consistent trends in current and power changes during peak electricity consumption periods of 8-9 am and 7-8 pm, and the calculated correlation coefficient is 0.82, which is higher than the set threshold of 0.65, a temporary edge is generated between them. This algorithm can discover potential connections through the correlation of electricity consumption data in the absence of topological archives, solving the problem that traditional methods cannot determine the correlation of electricity meters in unknown topological areas, and expanding the applicability of the system.

[0029] Dynamic edge weight calculation: A two-factor dynamic edge weight algorithm is used, which combines the spatiotemporal fusion feature similarity of nodes with the inherent connectivity coefficient to calculate the edge weight. The calculation formula of the two-factor dynamic edge weight algorithm is as follows: Sim Con ,in, It is a moment Electricity meter With electricity meter Dynamic weights of edges between them, Sim It is a similarity calculated based on the spatiotemporal fusion features of nodes. , Electricity meters , At any moment The spatiotemporal fusion characteristics, It is the inherent connection coefficient between the two meters. It is the feature similarity weight coefficient. , These are the identifiers of the two electricity meter nodes whose edge weights are to be calculated. For example, the spatiotemporal fusion feature similarity between the main electricity meter and the electricity meter of Building 1 is 0.91, and the inherent connection coefficient is 0.88. The dynamic weight of the edge between them is obtained through weighted calculation using an algorithm. For the residential electricity meter and the electricity meter of Building 2 that generate a temporary edge, the dynamic edge weight is also calculated based on their spatiotemporal fusion feature similarity and the inherent connection coefficient of the temporary edge. Dynamic edge weights can reflect the tightness of the connection between electricity meters in real time, avoiding the problem that fixed edge weights cannot adapt to changes in electricity load, and allowing subsequent network analysis to more accurately capture the dynamic changes in the connection relationship between electricity meters.

[0030] Dynamic updates: Every 5 minutes, the attribute values ​​of all nodes are updated based on the latest electricity consumption data collected in the time-series database, and the weight parameters of the corresponding edges are adjusted synchronously based on load fluctuations. This high-frequency dynamic update mechanism ensures that the spatiotemporal dynamic graph remains synchronized with the actual operating status of the power grid, promptly reflects the impact of changes in electricity load on the meter correlation, avoids inaccurate analysis results due to graph structure lag, and provides real-time and accurate graph data support for subsequent network analysis.

[0031] The constructed spatiotemporal dynamic graph is input into the network model, which includes a spatial feature extraction layer, a temporal feature extraction layer, and a fusion output layer. Through multi-dimensional feature extraction and fusion, the probability of hierarchical association between electricity meters is accurately calculated. The specific operation process and function are as follows: Spatial feature extraction: The spatial layer adopts an enhanced graph attention mechanism, setting three attention heads, each with a hidden layer dimension of 64. The attention weight between nodes is calculated using a graph attention weight algorithm, combining the physical distance between electricity meters and the similarity of electricity consumption characteristics. The calculation formula for the graph attention weight algorithm is: ,in, It is an electricity meter Its neighboring nodes Attention weights This is the attention weight matrix. Represents a node and The spatiotemporal fusion feature vectors are concatenated, LeakyReLU is used as the activation function, and then the neighbor node information is weighted and aggregated based on this attention weight. It is an electricity meter The neighbor node identifier, It is an electricity meter The set of neighboring nodes, Electricity meters , At any moment The spatiotemporal fusion feature aggregates neighbor node information. For example, when analyzing the spatial characteristics of electricity meters in a building, the focus is on aggregating the information of residents' electricity meters that are physically close and have similar electricity consumption characteristics, while weakening the influence of meters that are far away and have large feature differences, thus obtaining the spatial feature vector of the building's electricity meters. This spatial feature extraction based on the attention mechanism can automatically focus on key neighbor nodes, avoid interference from irrelevant node information, improve the targeting and accuracy of spatial features, and provide a reliable spatial dimension basis for subsequent correlation analysis.

[0032] Temporal Feature Extraction: The time layer employs an improved gated recurrent unit with a two-layer recurrent structure, each hidden layer having a dimension of 128. Using the spatiotemporal fusion feature sequence of nodes as input, time-series modeling is performed on the historical electricity consumption sequence of each node to capture the temporal dependencies of electricity consumption data, such as the peak electricity consumption patterns of residential meters at 8 AM and 7 PM on weekdays. The output is the node's temporal feature vector. The improved gated recurrent unit effectively processes long-sequence data, accurately capturing the temporal trends and periodic patterns of electricity consumption data. This overcomes the shortcomings of traditional time-series analysis methods in effectively uncovering long-term temporal dependencies, providing comprehensive temporal dimension support for correlation analysis.

[0033] Feature fusion and association probability output: The spatiotemporal feature fusion sublayer of the fusion output layer uses a multi-layer attention mechanism to weight and fuse spatial and temporal feature vectors, setting a spatiotemporal attention weight matrix. This fusion method can dynamically adjust the spatiotemporal feature weights according to the meter type and actual scenario, ensuring that the fused features fully reflect the core influencing factors of the association between meters. The association probability calculation sublayer contains two fully connected layers, with 256 neurons in the first layer and 64 neurons in the second layer. The fully connected layers compress and map the fused features, reducing data redundancy and improving feature expressiveness. Finally, a feature concatenation logistic regression algorithm is used to concatenate the fused features of the meter pairs to be analyzed and output the upper and lower level association probabilities of the two. The calculation of the feature concatenation logistic regression algorithm is as follows: ,in It is a node and Feature concatenation after spatiotemporal convolution It is a weight vector. It is a bias term. It is an electricity meter At any moment Feature vector after spatiotemporal convolution It is an electricity meter At any moment Feature vector after spatiotemporal convolution and These are the node identifiers of the two meters whose correlation probability is to be calculated, consistent with the node definition in the spatiotemporal dynamic diagram. This refers to the moment of probability calculation. For example, the association probability between the main electricity meter and the electricity meter of Building 1 is 0.98, and the association probability between the electricity meter of Building 1 and a resident's electricity meter is 0.95, both falling within the high probability range. This algorithm can transform high-dimensional fused features into intuitive association probabilities, providing a clear basis for judgment in subsequent verification modules. Furthermore, the logistic regression algorithm has good interpretability, making it easy for staff to understand the calculation logic of association probabilities.

[0034] Through multi-dimensional verification, the association probability results are further screened and corrected to ensure that the final output of the hierarchical relationship is accurate and reliable, and to solve the misjudgment problem that may exist if only the network model is relied upon. The specific operation process and function are as follows: Constructing a candidate association list: Collect auxiliary data such as the community power grid topology map and meter installation files. Combined with the association probability output by the above network module, meter pairs with association probabilities higher than the preset threshold of 0.8 are screened to form a candidate association list. With the help of auxiliary data and high association probability, the verification scope can be narrowed by using existing power grid information, and the meter pairs in the candidate list can be ensured to have high initial association credibility, thereby reducing the workload of subsequent verification and improving verification efficiency.

[0035] Multi-dimensional verification: Load curve similarity verification: For each pair of meters in the candidate list, the distance between the load curves within the same time period is calculated using a dynamic time warping algorithm. If the distance between the load curves of a building's meter and a resident's meter is 0.12, which is lower than the set similarity threshold of 0.3, it indicates that their load changes are highly synchronized, and the verification passes. If the distance between the load curves of a resident's meter and the meters of two buildings is 0.58, which is higher than the threshold, it is marked as pending confirmation. Load curves can intuitively reflect the electricity consumption patterns of meters. Similarity verification can judge the rationality of the correlation from the perspective of consistency in electricity consumption behavior. The dynamic time warping algorithm can effectively handle the time offset problem of load curves, improve the accuracy of similarity calculation, and avoid misjudgments caused by small differences in peak electricity consumption times.

[0036] Energy conservation verification: The total power supply of the upstream meter in the candidate list within one month, and the total electricity consumption of all its corresponding downstream candidate meters within one month, are statistically analyzed to calculate the energy loss rate. If the total power supply of a building's meters is 50,000 kWh and the total electricity consumption of residents is 48,000 kWh, the loss rate is 4%, which is within a reasonable loss range, and the verification passes. If the loss rate of a building's meters is 8%, exceeding the reasonable range, the correlation between the building's meters and the residents' meters needs to be reassessed. Verification based on the principle of energy conservation in the power grid can determine the correctness of the correlation from the perspective of energy transfer. The setting of a reasonable loss range fully considers the objective existence of power grid line losses, avoiding misjudgments caused by normal losses, and ensuring that the verification results conform to the actual operation of the power grid.

[0037] Anomaly propagation verification: When a voltage drop occurs in the community power grid due to an external line fault, the time difference between the detection of the anomaly by candidate meters is recorded. If the main meter detects the anomaly at 14:00:00 and the meter in Building 1 detects the anomaly at 14:00:01, the time difference conforms to the cable propagation law, and the verification passes. However, if a resident's meter detects the anomaly only at 14:00:05, the time difference does not conform to the law, indicating a problem with the correlation. Anomaly propagation has clear electrical laws. Time difference verification can determine whether there is an actual electrical connection between meters from a physical connection perspective, compensating for the shortcomings of the previous two verifications in judging physical connections and further improving the accuracy of correlation.

[0038] For meter pairs marked as pending confirmation or with questionable associations during verification, in the absence of supporting data, potential associations are uncovered by actively generating characteristic signals. This solves the problem that traditional methods cannot further confirm associations without supporting data. The specific operation process is as follows: Figure 1 As shown:

[0039] For the meter pairs marked as pending confirmation or with problematic associations in the aforementioned verification, due to the lack of detailed installation records for some buildings, the active feature blind search confirmation module was activated. First, the detection range was defined, including the 10 residential meters in the unit containing the resident's meter, as well as the meters in buildings 2 and 3. A reasonable scope avoids inefficiency caused by too many detection targets, while ensuring no potential associated objects are overlooked. Feature signals were generated in stages: the first stage injected harmonic signals, which have unique frequency characteristics facilitating subsequent identification and matching; the second stage generated pulsed power consumption characteristics, whose obvious changes quickly reflect the response correlation between meters; the third stage adjusted the phase offset, as phase characteristics are a core indicator of meter electrical parameters, and their correlation changes accurately reflect electrical connection relationships. Combining the collected feature response data, the feature correlation between the resident's meter pending confirmation and the meters in buildings 2 and 3 was analyzed. By comparing the consistency of different meters' responses to the feature signals, the actual associated objects could be clearly identified. If the response trend of the resident's electricity meter is highly consistent with that of the electricity meters in the three buildings when subjected to harmonic injection, pulsed power consumption, and phase shift, the detection range is gradually narrowed down, ultimately confirming that the resident's electricity meter is connected to the electricity meters in the three buildings, thus correcting the correlation in the original candidate list. This module actively intervenes to create identifiable features, breaking through the limitations of traditional passive data collection. It can accurately confirm correlations even in scenarios without auxiliary data, significantly improving the applicability and reliability of the system.

[0040] Based on the operation and verification of all the above modules, and after correcting the correlation relationships, a list of hierarchical relationships of the community's electricity meters is output. It clarifies that the lower level of the main electricity meter is the electricity meters of 12 buildings, and the lower level of each building's electricity meter is the residential electricity meter of the corresponding building. This clear list of hierarchical relationships provides accurate basis for electricity consumption monitoring of the community's power grid, allowing staff to quickly locate the source of abnormal electricity consumption in a certain area. At the same time, it provides a clear direction for fault diagnosis. When a residential electricity meter malfunctions, it can be quickly traced back to the corresponding building's electricity meter and the main electricity meter according to the list, shortening the fault diagnosis time, reducing power grid operation and maintenance costs, and ensuring the safe and stable operation of the community's power grid.

[0041] Example 2: Verification of the relationship between upper and lower levels of electricity meters in industrial parks.

[0042] This embodiment uses an industrial park as an application scenario. The park includes one main electricity meter, five workshop main electricity meters, 20 production line meters, and 80 key equipment meters. The meter hierarchy relationship verification method of this invention is used to verify and correct the initially determined relationship of "main electricity meter - workshop electricity meter - production line electricity meter - equipment electricity meter". This solves the problems of traditional manual topology verification in industrial parks being time-consuming and easily affected by complex power environment interference, leading to misjudgment of the relationship. It ensures that the final output relationship conforms to the actual power grid electrical connection logic. The specific steps and functions are as follows: First, collect the power grid topology map and meter installation files of the industrial park (recording the meter number, installation location, corresponding production line and equipment affiliation, such as "workshop 1 main electricity meter (number G100) corresponds to production lines 1-5, and each production line is matched with 4 equipment meters"). This auxiliary data can provide the initial physical connection basis for the relationship and avoid subsequent analysis from deviating from the actual power grid layout. Next, the association probabilities of all meter pairs were obtained from the previously deployed meter hierarchy relationship analysis model (including data acquisition, spatiotemporal dynamic graph construction, and network analysis modules). For example, the association probability between the main meter of the park (G000) and the main meter of workshop 1 (G100) was 0.97, the association probability between the main meter of workshop 1 and the meter of production line 1 (G110) was 0.94, and the association probability between the meter of production line 1 and the meter of equipment 1 (G111) was 0.92. The association probability quantifies the tightness of the relationship between meters, providing an intuitive numerical reference for screening. Subsequently, an association probability threshold of 0.85 was set, and meter pairs with association probabilities higher than this threshold were selected to form a candidate association list, such as (G000-G100, 0.97), (G100-G110, 0.94), (G110-G111, 0.92), etc., totaling 105 meter pairs. By using both supplementary data and high correlation probability for screening, we can not only use existing power grid information to narrow down the scope of subsequent verification and reduce invalid verification work, but also ensure that the meter pairs in the candidate list have high initial credibility, laying a reliable foundation for subsequent multi-dimensional verification.

[0043] First, the verification period was determined to be one week of normal production in the industrial park (Monday to Sunday, 8:00-20:00 daily). This period covers weekday peak production (e.g., 9:00-11:00, 14:00-16:00), lunch break off-peak (12:00-13:00), and weekend shutdown, which can comprehensively reflect the electricity consumption patterns under different operating conditions and avoid the bias caused by data from a single operating condition. Then, the load curve data of each pair of meters in the candidate list within the above time period was extracted from the time series database. The load curve records the active power value at 15-minute intervals. For example, the active power of the main meter (G100) in workshop 1 is 500kW at 8:00 on Monday, 520kW at 8:15, and 100kW at 12:00 (lunch break). The power of the meter (G110) on production line 1 during the same period is 100kW, 105kW, and 20kW. These data can intuitively reflect the electricity consumption trend of the meters. Finally, a dynamic time warping algorithm is used to calculate the distance between the load curves of each pair of upstream and downstream meters. This algorithm can effectively handle the time offset problem caused by the fine-tuning of the load curve due to the production process (such as a production line starting 10 minutes earlier, resulting in a slightly earlier peak load) and accurately measure the similarity of the curves. If the distance between the load curves of G100 and G110 is 0.15, which is lower than the set similarity threshold of 0.4, it indicates that their electricity consumption trends are highly synchronized (synchronous increase during peak production and synchronous decrease during lunch break), and the verification is successful. If the distance between the load curves of a production line meter (G130) and the main meter of workshop 1 (G100) is 0.62, which is higher than the threshold, it indicates that their electricity consumption patterns are significantly different (such as G130 maintaining a high load during lunch break, which is inconsistent with the low trend of G100), and the meter pair (G100-G130) is marked as pending confirmation. Through load similarity verification, the rationality of the association can be judged from the perspective of consistency of electricity consumption behavior, eliminating the problem of mismatched electricity consumption patterns caused by misjudgment of meter ownership, and further narrowing down the scope of suspected associations.

[0044] The first step is to determine the statistical period as one calendar month (30 days), which includes 22 working days and 8 rest days, covering the complete production plan (such as fixed equipment maintenance days every week and monthly production task adjustments). This ensures that the electricity data reflects the long-term stable energy transfer pattern and avoids interference from short-term special operating conditions (such as single-day equipment failures) causing abnormal electricity consumption to affect the verification results. The second step is to retrieve the total power supply of the upper-level electricity meters and the total electricity consumption of all corresponding lower-level candidate electricity meters from the electricity meter data acquisition system. For example, the total power supply of the main electricity meter (G200) in workshop 2 is 300,000 kWh in one month, and the total electricity consumption of its six lower-level production line meters (G210-G260) is 45,000 kWh, 48,000 kWh, 46,000 kWh, 47,000 kWh, 49,000 kWh, and 45,000 kWh, respectively, totaling 280,000 kWh. These data directly reflect the transfer and consumption of energy between the upper and lower level electricity meters. The third step calculates the loss rate based on the formula: "Electricity Loss Rate = (Total Power Supply from Upper Level - Total Power Consumption from Lower Level) / Total Power Supply from Upper Level × 100%". The loss rate of the total electricity meter in Workshop 2 is approximately (300,000 - 280,000) / 300,000 × 100% ≈ 6.67%, while the reasonable loss range for the workshop-level power grid in the industrial park is 5%-8%. This loss rate falls within the reasonable range, and the verification is successful. If the loss rate of the total electricity meter in a certain workshop is 12%, exceeding the reasonable range, it indicates that there may be missing lower-level meters (such as newly added meters on a production line not included in the candidate list) or metering malfunctions (such as metering deviations in lower-level meters). The correlation between the total electricity meter in that workshop and the corresponding production line meters needs to be reassessed. Verification based on the principle of power conservation in the power grid can judge the correctness of the correlation from the perspective of the essence of energy transfer, compensating for the shortcomings of load similarity verification which only focuses on trends and does not involve actual energy quantification, ensuring that the correlation conforms to the laws of physical energy.

[0045] First, the power grid operation status of the park is monitored in real time through the power grid monitoring system, and abnormal events are recorded, such as current overload caused by a short circuit in the production line equipment at 10:00 on a certain workday (the current in workshop 3 suddenly increased from 200A to 350A), and voltage drop caused by an external power supply line fault at 15:00 (the park voltage dropped from 380V to 320V). These abnormal events have clear occurrence times and characteristics, and can be used as "signal sources" to verify the physical connection between the meters. Next, the time records of the meters in the candidate association list for detecting the above-mentioned abnormal events are retrieved. For example, if the main meter (G300) of workshop 3 detects a current overload at 10:00:00, the meter of its subordinate production line 3 (G330) detects an overload at 10:00:00.5, and the meter of production line 4 (G340) detects an overload at 10:00:00.6; in the case of a voltage drop event, the main meter of the park (G000) detects an anomaly at 15:00:00, and the main meters of each workshop successively detect anomalies between 15:00:00.2 and 15:00:00.5. The time difference data can directly reflect the transmission speed of the abnormal signal between the meters. Finally, based on the electrical connection principle, the transmission law is determined: the transmission time of current and voltage anomalies in the cable is positively correlated with the cable length. In the same workshop, the production line meters and the workshop main meter are relatively close, and the time difference of anomaly detection should be within 1 second. The time difference between the main power meter in Workshop 3 and the production line power meters is within 0.6 seconds, which conforms to the conduction law and passes the verification. However, if a production line power meter (G350) detects an anomaly 10 seconds after the main power meter in Workshop 3 detects an overload, far exceeding the reasonable time difference, it indicates that there is no direct electrical connection between the two, suggesting a problem with the correlation between the two meters (G300-G350). The anomaly conduction verification verifies the correlation from the perspective of physical signal transmission, eliminating misjudgments based solely on similar electricity consumption patterns without actual electrical connection, further improving the accuracy of the correlation.

[0046] First, for the meter pairs to be confirmed in S200 (such as G100-G130) and the meter pairs with questionable association in S400 (such as G300-G350), the active feature blind search confirmation module is activated to obtain supplementary data: For G100-G130, the meter pairs of some production lines in Workshop 1 and Workshop 2 are designated as the detection range (to avoid inefficiency due to an excessively large range). Harmonic signals are injected in stages, the characteristics of pulse power consumption are manufactured, and the phase offset is adjusted. By collecting the characteristic response data of each meter, it is found that G130 responds in all characteristic tests in the same way as the main meter (G200) in Workshop 2 (such as synchronous changes in harmonic amplitude and the same time of occurrence of pulse power peak), rather than the main meter (G100) in Workshop 1. It is confirmed that the superior of G130 is G200. For G300-G350, the feature blind search verification shows that their actual circuit connection belongs to Workshop 4. It is confirmed that the superior is the main meter (G400) in Workshop 4. The proactive feature blind search confirmation module actively generates identifiable electrical features, overcoming the limitation of being unable to further verify without auxiliary data, and providing key evidence for correcting association relationships. Next, based on the above verification results, the candidate association list is corrected, adjusting (G100-G130) to (G200-G130) and (G300-G350) to (G400-G350), ensuring that all meter pairs conform to the multi-dimensional verification results. Finally, the corrected meter pairs are organized to form a list of hierarchical relationships between meters in the industrial park, clearly defining complete association chains such as "Park Main Meter (G000) - Workshop 1 Main Meter (G100) - Production Line 1 Meter (G110) - Equipment 1 Meter (G111)" and "Park Main Meter (G000) - Workshop 2 Main Meter (G200) - Production Line 3 Meter (G130) - Equipment 13 Meter (G131)". This list provides accurate data support for electricity cost accounting in industrial parks (such as allocating electricity costs by workshop and production line), and also lays the foundation for fault location (such as quickly tracing back to the corresponding production line and workshop electricity meter when a certain equipment meter is abnormal, shortening the investigation time) and power grid operation and maintenance optimization (such as rationally planning the line maintenance sequence according to the correlation), ensuring the safe and efficient operation of the park's power grid.

[0047] In summary, this embodiment addresses the scenario of verifying the hierarchical relationship between electricity meters in industrial parks, following a process of "constructing a candidate list - multi-dimensional verification - correcting the output." First, a candidate list is formed by combining the power grid topology map, installation records, and association probabilities to screen high-confidence meter pairs. Then, load similarity verification (dynamic time warping algorithm), energy conservation verification, and anomaly propagation verification are used to examine the rationality of associations from the dimensions of electricity consumption patterns, energy transfer, and physical connections. Finally, an active feature blind search confirmation module is used to correct questionable associations, outputting an accurate list of hierarchical relationships. This entire process solves the problems of low efficiency and susceptibility to misjudgment associated with traditional manual verification, providing reliable support for electricity cost accounting, fault location, and power grid operation and maintenance in industrial parks.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An analysis model of the superior-inferior relationship of electric meters, characterized by, The model comprises: Electricity meter data high-frequency acquisition module: adopt distributed architecture, deploy high-speed communication acquisition terminal, batch parallel acquisition of electricity meter real-time and historical power consumption data, data is transmitted after encryption, cleaning, and stored in time series database according to the standard format containing timestamp, electricity meter identification, and multi-dimensional parameters; The spatio-temporal dynamic graph construction module: the electricity meter is taken as the node, and real-time power consumption parameters, historical trend characteristics and its own attributes are given, the edges are dynamically constructed according to the potential electrical connection, the known topological region is determined by the cable parameter, and the unknown region is constructed by the power load correlation, and the node attributes and edge weight are updated every 5 minutes; The spatio-temporal dynamic graph convolution network analysis module: the spatio-temporal dynamic graph is input into the network model containing spatial and temporal feature extraction layer and fusion output layer, the spatial layer uses the graph attention mechanism to aggregate neighbor information, the time layer uses the gated recurrent unit to model the time sequence characteristics, and the fusion layer outputs the upper and lower correlation probability of the electricity meter, and the model is trained through the labeled historical topological data; The upper and lower relationship verification module: collect the power grid topological data to form a candidate correlation list, verify from the multi-dimensional load curve similarity, power conservation and abnormal event transmission, correct and output the confirmed upper and lower relationship list; The active feature blind search confirmation module: without auxiliary data, the electricity meter detection range is determined, the characteristics of harmonic injection, pulse power consumption and phase shift are manufactured in stages, the range is blindly searched and contracted combined with the collected data, and the upper and lower correlation of the electricity meter is confirmed, In the spatio-temporal dynamic graph construction module, the specific steps of constructing the spatio-temporal dynamic graph are: Determine the node set, use a single electricity meter as the basic node of the graph structure, give each node multi-dimensional attribute characteristics, including 30-second-level power consumption parameters collected in real time, historical power consumption trend characteristics and electricity meter attributes; Adopt the spatio-temporal feature attenuation fusion algorithm to fuse the real-time collection characteristics and historical characteristics of the node to obtain the spatio-temporal fusion characteristics of the node; Build the edge between nodes, for the region with known substation area topology, determine the existence of the edge according to the cable connection path, and take the cable length, wire diameter, material and laying method as the basic weight parameters of the edge; for the unknown topology region, calculate the correlation coefficient of the electricity meter to the power consumption parameter through the power consumption parameter time sequence correlation algorithm, and generate a temporary edge when the absolute value of the correlation coefficient is higher than the set threshold; Adopt the double-factor dynamic edge weight algorithm to calculate the dynamic weight of the edge between nodes combined with the node spatio-temporal fusion characteristic similarity and the inherent connection coefficient; Update the node attribute value according to the latest collected power consumption data every 5 minutes, and adjust the weight parameters of the edge based on the load fluctuation to realize the dynamic update of the spatio-temporal dynamic graph, In the spatiotemporal dynamic graph construction module, the calculation formula for the spatiotemporal feature attenuation fusion algorithm is as follows: ,in, Indicates electricity meter At any moment The spatiotemporal fusion characteristics, It is an electricity meter At any moment Real-time acquisition of feature vectors, It is an electricity meter At any moment Historical feature vectors These are real-time feature weight coefficients. It is the historical characteristic attenuation coefficient. It is the length of the historical time window. It's the meter indicator. It is a timestamp parameter.

2. The model for analyzing the parent-child relationship of the electric meter according to claim 1, wherein, In the electricity meter data high-frequency acquisition module, the high-speed communication acquisition terminal is an intelligent acquisition device based on 5G or optical fiber communication, each terminal is connected with the electricity meter through RS-485 bus or wireless LoRa technology; the real-time power consumption data collected includes voltage, current, active power, reactive power, power factor and frequency, the historical power consumption data includes daily power consumption curve, monthly power consumption total, seasonal power consumption characteristic data and extreme weather period power consumption record, and the electricity meter identification includes electricity meter number, installation position and substation attribution information.

3. The model for analyzing the parent-child relationship of the electric meter according to claim 1, wherein, The calculation formula of the time sequence correlation algorithm of the power consumption parameter in the spatiotemporal dynamic graph construction module is as follows: Wherein is the power consumption parameter correlation coefficient of the electric meter and the electric meter is the power consumption parameter value of the electric meter at the time point is the power consumption parameter value of the electric meter at the time point is the average value of the power consumption parameter of the electric meter in the time window is the average value of the power consumption parameter of the electric meter in the time window is the time window length for calculating the correlation coefficient is the time stamp parameter, and the set threshold value is 0.

65. When , a temporary edge is generated between the electric meter and the electric meter .​​​​​ 4. The model for analyzing the parent-child relationship of the electric meter according to claim 1, wherein, The calculation formula of the double-factor dynamic edge weight algorithm in the space-time dynamic graph construction module is: Sim Con wherein, is the space-time fusion feature of the electric meter at the moment is the dynamic weight of the edge between the electric meter and the electric meter , Sim is the similarity calculated based on the space-time fusion feature of the node, , respectively, is the space-time fusion feature of the electric meter , at the moment , is the inherent connection coefficient between the two electric meters is the feature similarity weight coefficient, , is the node identifier of the two electric meters of the edge weight to be calculated.

5. The model for analyzing the parent-child relationship of the electric meter according to claim 1, wherein, The network model in the space-time dynamic graph convolution network analysis module comprises: The spatial feature extraction layer adopts an enhanced graph attention mechanism, sets three attention heads, the hidden layer dimension of each attention head is 64, and neighbor node information is aggregated by calculating attention weight between nodes through a graph attention weight algorithm; The time feature extraction layer adopts an improved gated recurrent unit, sets two layers of recurrent structures, the hidden layer dimension of each layer is 128, the historical power consumption sequence of each node is modeled in time sequence, and the input is a space-time fusion feature sequence of the node; The fusion output layer comprises a space-time feature fusion sublayer and a correlation probability calculation sublayer, the space-time feature fusion sublayer performs weighted fusion on the output features of the spatial feature extraction layer and the time feature extraction layer through a multi-layer attention mechanism, and a space-time attention weight matrix is set; the correlation probability calculation sublayer comprises two fully connected layers, the first layer has 256 neurons, the second layer has 64 neurons, the fusion features are compressed and mapped in dimension through the fully connected layers, and a feature splicing logistic regression algorithm is used to output the upper and lower correlation probability values between the electric meters.

6. The model for analyzing the parent-child relationship of the electric meter according to claim 1, wherein, The calculation formula of the graph attention weight algorithm in the space-time dynamic graph convolution network analysis module is: wherein, is an electric meter The attention weight of its neighbor node , is an attention weight matrix, represents the spatio-temporal fusion feature vector splicing of node and , LeakyReLU is an activation function, and neighbor node information is weighted and aggregated based on the attention weight, is the neighbor node identifier of the electric meter , is the neighbor node set of the electric meter , , respectively, , the spatio-temporal fusion feature at time .

7. The model for analyzing the parent-child relationship of the electric meter according to claim 1, wherein, In the spatio-temporal dynamic graph convolution network analysis module, the calculation of the feature splicing logistic regression algorithm is: wherein is a node and the feature splicing after spatio-temporal convolution, is a weight vector, is a bias term, is an electric meter at time the feature vector after spatio-temporal convolution, is an electric meter at time the feature vector after spatio-temporal convolution, and are two electric meter node identifiers to be calculated for the correlation probability, consistent with the node definition in the spatio-temporal dynamic graph, is the time for probability calculation.

8. The method for verifying the correlation between the electric meter and its superior and subordinate, which is applicable to the analysis model of the correlation between the electric meter and its superior and subordinate according to any one of claims 1-7, characterized in that, The specific steps of the method are as follows: S100, candidate list building: collecting power grid topology map, electric meter installation file data, combining the upper and lower correlation probability of the electric meter output by the space-time dynamic graph convolution network analysis module, screening out the electric meter pairs with a correlation probability higher than a preset probability threshold, and forming a candidate correlation list; S200, load similarity verification: verifying from the load curve similarity dimension, calculating the load curve similarity of each pair of upper and lower electric meters in the candidate correlation list in the same time period, calculating the distance between two load curves by using a dynamic time warping algorithm, the smaller the distance, the higher the similarity, and if the similarity is lower than a set similarity threshold, marking the electric meter pair as a to-be-confirmed state; S300, power conservation verification: verifying from the power conservation dimension, calculating the power loss rate by counting the total power supply of the upper electric meter in the candidate correlation list in one week and the total power consumption of all lower electric meters corresponding to the upper electric meter in the same period, and reevaluating the correlation relationship of the electric meter pair if the power loss rate exceeds a reasonable loss range; S400, abnormal conduction verification: verifying from the abnormal event conduction dimension, recording the time difference between the upper and lower electric meters in the candidate correlation list when voltage sag and current overload occur in the power grid, and if the time difference does not conform to the conduction law of electrical connection, it is judged that the correlation relationship of the electric meter pair has a problem; S500, correction output list: analyzing the electric meter upper and lower correlation relationship by combining the data obtained by the active feature blind search confirmation module for the electric meter pairs marked as a to-be-confirmed state or having a problem in the correlation relationship after the three-dimensional verification, and finally outputting a confirmed upper and lower relationship list.

Images