A power energy consumption monitoring method and system of an electric energy metering box

Abstract

The application discloses a kind of electric energy metering box electric power energy consumption monitoring method and system, belong to electric power energy monitoring technical field.The present monitoring mode is extensive, abnormal positioning is inefficient, network pressure is big, and the problems are solved, including: according to the division strategy, divide monitoring area, for each branch configuration monitor and cross collection adjacent branch electric power parameter;Grouping subarea monitoring network and selecting main node, monitor carries out edge pretreatment and uploads data to main node, and main node converges and analyzes and uploads cloud server;Cloud server is analyzed by global AI model fusion, locates exception and root branch and outputs alarm.The system corresponds including subarea, monitor, monitoring network construction etc.Module.The application realizes electric energy metering box branch refinement monitoring, improves abnormal positioning efficiency, reduces network load, can also realize predictive maintenance, and is suitable for electric energy metering box energy consumption monitoring scene.

Description

Technical Field

[0001] This application relates to the field of power energy consumption monitoring technology, and in particular to a method and system for monitoring power energy consumption of an electricity metering box. Background Technology

[0002] As a critical node in the power supply and distribution chain of a power system, the monitoring of power consumption in each branch of the electricity metering box is essential for achieving refined management, rapid fault location, energy efficiency optimization, and electricity safety. With the development of IoT and AI technologies, traditional methods based on manual inspection or independent meter reading are gradually evolving towards automated and intelligent monitoring. Current technologies typically involve installing independent intelligent monitoring modules for important branches within the metering box, or centrally collecting data via a bus and uploading it to a backend system for analysis. These solutions improve the automation level of data acquisition to a certain extent.

[0003] However, in practical applications, it has been found that the existing technical solutions still have several significant drawbacks: First, the monitoring method is relatively crude, and each monitoring unit is usually only responsible for measuring its own circuit. When an anomaly occurs in a branch, it is difficult to quickly distinguish whether it is a fault of its own or affected by adjacent or related circuits, resulting in low positioning efficiency. Second, the full upload of massive real-time data puts enormous pressure on the communication network, and the centralized cloud processing mode is difficult to meet the real-time requirements for rapid response to local anomalies.

[0004] Therefore, there is an urgent need for an innovative method and system for monitoring the power consumption of electricity metering boxes, which can achieve refined, highly reliable, and intelligent monitoring of the status of each branch circuit in the box, and has the ability to quickly locate anomalies and diagnose root causes, so as to overcome the above-mentioned shortcomings of existing technologies. Summary of the Invention

[0005] The purpose of this invention is to achieve refined, highly reliable, and intelligent monitoring of the status of each branch circuit within the box, and to possess the ability to quickly locate anomalies and diagnose their root causes, thereby overcoming the above-mentioned shortcomings of the prior art. This application provides a method and system for monitoring the power consumption of an electricity metering box.

[0006] To achieve the above objectives, the power consumption monitoring method and system for an electricity metering box provided in this application adopts the following technical solution:

[0007] Firstly, a method for monitoring the power consumption of an electricity metering box is disclosed, including: obtaining a preset zoning strategy for the electricity metering box, and dividing at least two branches in the electricity metering box into a monitoring area according to the zoning strategy;

[0008] Each branch in each monitoring area is equipped with a monitor. The power supply channel of the monitor is connected to one of the branches, and the monitoring channel of the monitor is connected to another branch adjacent to the current branch to obtain the power parameter data of the other branch.

[0009] Each monitor forms a zoned monitoring network based on the Internet of Things protocol, and a master node is selected in the monitoring network of each monitoring zone;

[0010] Each monitor continuously collects power parameter data of the branch circuits connected to the monitoring channel and performs edge preprocessing;

[0011] Within each monitoring zone, all monitors except the master node send the pre-processed power parameter data to the master node. The master node then aggregates and performs preliminary edge analysis on the data within its monitoring zone to obtain the initial analysis results and characteristic data for that monitoring zone.

[0012] Each master node uploads the preliminary analysis results and feature data to the cloud server;

[0013] The cloud server utilizes a global energy consumption analysis and fault tracing model to perform a fusion analysis based on event sequences and topology on the preliminary analysis results and feature data uploaded by each master node, and identifies abnormal energy consumption data, fault propagation paths and root cause branches across monitoring areas.

[0014] Based on abnormal energy consumption data, fault propagation paths, and root cause branches, generate and output location alarm information that includes root cause branch identifiers, abnormality types, and fault impact ranges.

[0015] Preferably, the zoning strategy for obtaining electricity metering boxes includes:

[0016] Collect historical power parameter data for each branch within a set period;

[0017] Cluster analysis is performed on historical power parameter data to identify candidate branch groups with high correlation based on the change patterns of power parameter data, load curves, and power characteristics of branches.

[0018] Based on the correlation analysis results of candidate route groups, and combined with the physical adjacency constraints between routes, routes with high correlation and physical adjacency are assigned to the same monitoring area, generating an optimized zoning strategy.

[0019] Preferably, equipping each branch within the monitoring area with a monitor includes:

[0020] Each branch circuit within the power metering box is configured with an independent node unit, and all node units are physically arranged to form a point array. The node units corresponding to branches belonging to the same monitoring area are electrically connected through a continuous series detection link. Each monitor is integrated into a node unit, and its monitoring channel is used to detect the electrical continuity or impedance characteristics of the series detection link at this node unit. When an electrical abnormality occurs in the branch circuit corresponding to any node unit in the series detection link, the series detection link will show a state interruption or a sudden change in characteristic parameters at that node. The link status of each node unit is queried through the communication mechanism within the monitoring network, and the abnormal branch is located based on the location information of the link state interruption or the sudden change in characteristic parameters.

[0021] Preferably, edge preprocessing includes:

[0022] Each monitor acquires the original analog signal of the monitoring channel at a fixed sampling rate, and obtains the original power parameter data stream after analog-to-digital conversion;

[0023] Data quality assessment and primary filtering are performed on the raw power parameter data stream to remove outlier points, and a moving average filtering algorithm is used to smooth the data.

[0024] Based on a preset feature extraction template, multi-dimensional features are extracted in parallel from the filtered data stream;

[0025] Based on the current IoT network load and the historical anomaly probability of the routing, a compression algorithm is dynamically selected to compress and encode the multi-dimensional feature dataset, generating a compressed feature data packet.

[0026] The compressed feature data packet, the type of compression algorithm used, the data acquisition time information, and the corresponding branch identification information are encapsulated to form a standardized edge preprocessing data packet.

[0027] Preferably, a master node is selected in the monitoring network of each monitoring area, including:

[0028] Each monitor broadcasts its own status information packet during network initialization or according to a preset period. The status information packet includes at least the device performance score, signal quality strength, and remaining energy level.

[0029] Each monitor within the monitoring area receives and evaluates the status information packets of adjacent branches;

[0030] Based on a pre-defined comprehensive weighting algorithm, the election score of each monitor is calculated;

[0031] The monitor with the highest election score is selected as the master node of this monitoring area.

[0032] Preferably, preliminary edge analysis includes:

[0033] The master node constructs a district-level multivariate time series feature matrix at the current detection time based on the preprocessed data packets received from all branches within the monitoring area.

[0034] The pre-set district-level edge anomaly localization model is invoked, and the district-level multivariate time series feature matrix is ​​compared and analyzed with the baseline pattern of inter-path correlation learned by the monitoring area under the historical normal operation state.

[0035] The district-level edge anomaly localization model identifies one or more candidate anomaly branches that are significantly different from the behavioral characteristics of other branches in the district by calculating the deviation of each branch data vector from the baseline pattern of the correlation relationship.

[0036] The identifiers of candidate abnormal paths and their corresponding deviation characteristics are used as key components of the preliminary analysis results.

[0037] Preferably, after the master node obtains the preliminary analysis results and feature data of this monitoring area, adaptive data upload is performed, including:

[0038] Based on the preliminary analysis results obtained from the initial edge analysis, the operating status of this monitoring area is determined. If it is determined to be in a normal state, the master node uploads the filtered summary feature data and low-frequency summary data to the cloud server. If it is determined to be in a potentially abnormal state or the certainty of the state is lower than the preset threshold, the master node increases the data upload frequency to the cloud server and uploads more detailed raw data fragments or complete standardized edge preprocessing data packets. The cloud server dynamically adjusts the computing resources allocated to the corresponding monitoring area according to the characteristics of the data uploaded by each monitoring area.

[0039] Preferably, it also includes predictive maintenance steps: The cloud server's global energy consumption analysis and fault tracing model predicts the performance aging trend and remaining service life of key electrical components in each branch based on historical data from long-term monitoring; When the predicted remaining service life is lower than a preset threshold, predictive maintenance suggestions are generated and output.

[0040] Preferably, the global energy consumption analysis and fault tracing model performs a time-series and topology-based fusion analysis on the primary analysis results and feature data uploaded by each master node, including:

[0041] Using the data streams uploaded by each monitoring area and their physical topology connections within the energy metering box as input, each monitoring area or branch is abstracted as a graph node, and the electrical connections and physical adjacency relationships are abstracted as graph edges to construct a graph neural network model. The graph neural network model is used to perform graph embedding learning on the global energy consumption correlation pattern under normal conditions, and to detect abnormal subgraph structures in the actual data that deviate from this normal pattern, thus locating the root cause branch causing the anomaly. Based on the verification information of the root cause branch causing the anomaly and the external input, the internal parameters of the global energy consumption analysis and fault tracing model are continuously adjusted, and the deviation judgment threshold used by the district-level edge anomaly location model is optimized and updated.

[0042] Secondly, this application discloses a power consumption monitoring system for an electricity metering box, applied to the power consumption monitoring method for the electricity metering box as described in the first aspect, comprising:

[0043] The partitioning module is used to obtain the preset partitioning strategy of the power metering box and divide at least two branches in the power metering box into a monitoring area according to the partitioning strategy.

[0044] The monitor module is used to acquire power parameter data from another branch.

[0045] The monitoring network construction module is used by each monitor to form a partitioned monitoring network based on the Internet of Things protocol, and a master node is selected in the monitoring network of each monitoring area;

[0046] The edge preprocessing module is used for each monitor to continuously collect current parameter data of the branch circuits connected to the detection channel;

[0047] The data aggregation and preliminary analysis module is used by monitors other than the master node to send the pre-processed power parameter data from the edge to the master node. The master node aggregates and performs preliminary edge analysis on the data within its monitoring area to obtain the preliminary analysis results and feature data of the monitoring area.

[0048] The cloud server module is used to receive the preliminary analysis results and feature data uploaded by each master node;

[0049] The alarm and output module is used to generate and output location alarm information containing root source branch identifier, anomaly type and fault impact range based on abnormal energy consumption data, fault propagation path and root source branch.

[0050] The predictive maintenance module generates and outputs predictive maintenance recommendations when the predicted remaining useful life is lower than a preset threshold.

[0051] Compared with the prior art, the present invention provides a method and system for monitoring the power consumption of an electricity metering box, which has the following beneficial effects:

[0052] 1. By implementing the zoning strategy of the power metering box through the zoning module, the monitoring area is divided by combining historical power parameter clustering analysis and branch physical adjacency constraints. The monitoring module cross-collects the power parameters of adjacent branches, and the node units in the same monitoring area are electrically connected through series detection links. The monitoring device can detect the link continuity status and impedance characteristics, effectively solving the defects of the existing monitoring method and greatly improving the accuracy and efficiency of branch anomaly location.

[0053] 2. By controlling each monitor to build a zoned monitoring network based on the Internet of Things (IoT) protocol through the monitoring network construction module, and selecting the master node through a comprehensive weight algorithm, the power parameter data collected by the monitor is preprocessed by the edge preprocessing module, including filtering, outlier removal, and dynamic compression encoding. Then, the data aggregation and preliminary analysis module completes the data aggregation and preliminary edge analysis of the monitoring area, which can significantly reduce the transmission load of the IoT network and improve the real-time performance of power parameter data processing.

[0054] 3. The cloud server module receives data uploaded by each master node, and uses a global energy consumption analysis and fault tracing model combined with a graph neural network model to perform fusion analysis based on event sequences and topology. This accurately identifies abnormal energy consumption data, fault propagation paths, and root cause branches across monitoring areas. The adaptive data upload mechanism dynamically adjusts the upload frequency and data type, and rationally allocates computing resources. Furthermore, the predictive maintenance module predicts the aging trend of key electrical components based on historical monitoring data and outputs maintenance suggestions to achieve proactive protection. At the same time, the model optimization unit can continuously optimize the analysis accuracy. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating a method for monitoring the power consumption of an electricity metering box according to an embodiment of this application.

[0056] Figure 2 This is a flowchart illustrating a method for monitoring the power consumption of an electricity metering box according to an embodiment of this application, in which a monitor is installed on each branch circuit within the monitoring area.

[0057] Figure 3 This is a flowchart of an edge preprocessing method for a power energy consumption monitoring method for an electricity metering box according to an embodiment of this application. Detailed Implementation

[0058] The following is in conjunction with the appendix Figure 1-3 This application will be described in further detail.

[0059] This application discloses a method and system for monitoring the power consumption of an electricity metering box.

[0060] Firstly, referring to Figure 1This application discloses a method for monitoring the power consumption of an electricity metering box, comprising:

[0061] S1. Obtain the preset energy metering box zoning strategy, and divide at least two branches in the energy metering box into a monitoring area according to the zoning strategy.

[0062] S2. Equip each branch in each monitoring area with a monitor. The power supply channel of the monitor is connected to one of the branches, and the monitoring channel of the monitor is connected to another branch adjacent to the current branch to obtain the power parameter data of the other branch.

[0063] S3. Each monitor forms a zoned monitoring network based on the Internet of Things protocol, and a master node is selected in the monitoring network within each monitoring zone;

[0064] S4. Each monitor continuously collects power parameter data of the branch circuits connected to the detection channel and performs edge preprocessing;

[0065] S5. Within each monitoring area, all monitors except the master node send the pre-processed power parameter data to the master node. The master node aggregates and performs preliminary edge analysis on the data within its monitoring area to obtain the primary analysis results and characteristic data of its monitoring area.

[0066] The preliminary edge analysis is performed by the main node of the monitoring area. The main node first constructs a district-level multivariate time-series feature matrix based on the standardized power parameter data of all branches in the monitoring area after edge preprocessing. Then, it calls the pre-trained and solidified district-level edge anomaly localization model and compares the matrix with the baseline pattern of the correlation between branches under normal operation of the monitoring area learned by the model. By calculating the deviation of each branch data vector from the baseline pattern, candidate abnormal branches are identified. Finally, the preliminary analysis results containing candidate abnormal branch identifiers and deviation features are output.

[0067] The primary analysis results are the output data after the main nodes of each monitoring area of ​​the power metering box complete the preliminary edge analysis of the preprocessed power parameter data of all branches in their respective areas. The results include the operation status judgment results of the monitoring area, the unique identifier of the candidate abnormal branch, the abnormal deviation characteristics of the corresponding branch, and the accompanying simplified time series feature summary.

[0068] S6. Each master node uploads the preliminary analysis results and feature data to the cloud server;

[0069] S7. Pre-build a global energy consumption analysis and fault tracing model. The cloud server uses the global energy consumption analysis and fault tracing model to perform a fusion analysis based on event sequence and topology on the primary analysis results and feature data uploaded by each master node, and identify abnormal energy consumption data, fault propagation paths and root cause branches across monitoring areas.

[0070] S8. Based on abnormal energy consumption data, fault propagation paths, and root cause branches, generate and output location alarm information including root cause branch identifiers, abnormal types, and fault impact ranges.

[0071] The strategy for obtaining the zoning of the electricity metering box includes:

[0072] Collect historical power parameter data for each branch within a set period;

[0073] Historical power parameter data refers to time-series data collected from each branch of the power metering box during the initial deployment of the system or the historical operation phase, which can reflect the characteristics of its power consumption behavior.

[0074] Cluster analysis is performed on historical power parameter data to identify candidate branch groups with high correlation based on the change patterns of power parameter data, load curves, and power characteristics of branches.

[0075] Clustering analysis algorithms, such as K-means or hierarchical clustering, are used to learn from the historical power parameter data mentioned above.

[0076] The analysis object is a behavioral feature vector abstracted from the variation pattern, load curve, and power characteristics of each branch.

[0077] The analysis objective is to use clustering algorithms to automatically discover and group circuits with highly similar behavior patterns into the same monitoring area; for example, all lighting circuits form one monitoring area.

[0078] Based on the correlation analysis results of candidate route groups, and combined with the physical adjacency constraints between routes, highly correlated and physically adjacent routes are assigned to the same monitoring area to generate an optimized zoning strategy.

[0079] Physical adjacency constraints between branches refer to the spatial layout and wiring positions of branches within the actual electricity metering box. If the air switches or terminals of each branch are physically adjacent within the box, then physical adjacency is satisfied.

[0080] The principle for generating an optimized partitioning strategy is to prioritize classifying candidate routes that are judged to have high correlation and routes that are physically adjacent to each other into the same monitoring area. The final optimized partitioning strategy is a clear mapping table.

[0081] The power parameter data includes:

[0082] Basic electrical parameters, such as voltage, current, active power, reactive power, apparent power, quantity, and frequency;

[0083] Power and power factor parameters, such as power factor, harmonic content, and total harmonic distortion.

[0084] Advanced features and status parameters, such as voltage / current harmonic frequencies, voltage deviation, fluctuations and flicker, three-phase imbalance, load curves / load characteristics, peak values, valley values, average values ​​and volatility, waveform characteristics and other related electrical indicators.

[0085] Among them, reference Figure 2 Equipping each branch within the monitoring area with a monitor includes:

[0086] S21. Each branch circuit in the power metering box is equipped with an independent node unit, and all node units are physically arranged into a point array.

[0087] The node units corresponding to the branches belonging to the same monitoring area are electrically connected through a continuous series detection link;

[0088] S22. Each monitor is integrated into a node unit, and its monitoring channel is used to detect the electrical on / off state or impedance characteristics of the series detection link at this node unit.

[0089] S23. When an electrical abnormality occurs in the branch corresponding to any node unit in the series detection link, the series detection link will show a state interruption or a sudden change in characteristic parameters at that node.

[0090] For example, when a specific type of severe fault occurs in a branch, it will directly and inevitably affect the processing of the series detection link by the corresponding node unit, that is:

[0091] If a branch circuit short-circuits its own air switch due to overload or short circuit, the node unit where it is located can be designed to synchronously disconnect the series detection link, causing the series detection link to be in an open circuit state at that node unit.

[0092] If a branch circuit experiences loose contact, oxidation, or overheating, the node unit where the branch circuit is located can be designed to increase impedance to reflect this change, causing a sudden change in the impedance characteristic parameter of the link at that point. This allows the electrical abnormality of the branch circuit to be directly and one-to-one converted into a clearly located physical signal change measured on the shell of the series detection link.

[0093] S24. By monitoring the communication mechanism within the network, query the link status of each node unit, and locate abnormal branch based on the location information of link status interruption or sudden change in characteristic parameters.

[0094] In this process, the master node in the monitoring area periodically or randomly sends status query commands to all node units through the established IoT detection network. Each node unit reports the status of the series link it perceives. After the master node collects the reports from all node units, it only needs to identify the node whose series detection link is interrupted or whose impedance value has changed abruptly. The location of the node directly corresponds to the specific branch where the anomaly occurred.

[0095] The IoT monitoring node is designed as a self-diagnostic node unit that can be connected in series. An independent series detection link is used to map branch faults into changes in link status. The link status can be queried in an extremely simple and direct way to achieve second-level positioning.

[0096] Furthermore, refer to Figure 3 Edge preprocessing includes:

[0097] S41. Each monitor acquires the original analog signal of the monitoring channel at a fixed sampling rate, and obtains the original power parameter data stream after analog-to-digital conversion;

[0098] Each monitor acquires the analog signals of voltage and current of the branch circuits connected to the current monitor channel in real time at a fixed frequency that conforms to the Nyquist sampling theorem. The continuous analog waveform is converted into a discrete digital sequence through analog-to-digital conversion to form the original time-series data stream.

[0099] S42. Perform data quality assessment and primary filtering on the raw power parameter data stream to remove outlier points, and use a moving average filtering algorithm to smooth the data.

[0100] S43. Based on the preset feature extraction template, extract multi-dimensional features in parallel from the filtered data stream;

[0101] The types and algorithms for feature extraction are pre-configured based on the monitoring target and include at least:

[0102] Key electrical parameters and statistical characteristics;

[0103] Frequency domain component characteristics, which can be calculated by fast Fourier transform, include the amplitude and content of the fundamental frequency and major harmonics.

[0104] S44. Based on the current IoT network load and the historical anomaly probability of the branch, dynamically select a compression algorithm to compress and encode the multi-dimensional feature dataset to generate compressed feature data packets.

[0105] S45. Encapsulate the compressed feature data packet, the compression algorithm type information used, the data acquisition time information, and the corresponding branch identification information to form a standardized edge preprocessing data packet.

[0106] In this application, the feature extraction template is set as follows:

[0107] N1. Basic analysis based on partitioning strategy and routing;

[0108] First, it should be noted that the basic analysis based on partitioning strategy and routing is the data foundation and strategic starting point of the feature extraction template. It is dominated by the cloud server and directly relies on and expands upon the results of the partitioning strategy.

[0109] N11. Deep Branch Profile Generation: When performing cluster analysis, in addition to dividing the candidate branch groups with high correlation, it is also necessary to perform deep feature analysis on the historical power parameter data of each branch.

[0110] The analysis objectives include: identifying the load type, power consumption behavior patterns, and typical fault precursor characteristics of the branch circuits, forming a multi-dimensional profile of the branch circuit operation.

[0111] N12, Optimization of associated partitioning strategy;

[0112] By inputting the route operation profile and physical adjacency constraints together, the final partitioning strategy is optimized and generated. At the same time, the route operation profile will become the direct basis for subsequent feature template assignment.

[0113] N13. Construct a basic feature template library;

[0114] The cloud server has a built-in expandable feature algorithm library. Based on different monitoring scenarios and the identified path types, it pre-configures multiple benchmark feature extraction templates, such as the basic template. Harmonic template Transient template .

[0115] N2, intelligent template adaptation and assignment, binds a general template for feature extraction templates to a specific monitoring object;

[0116] N21. Rule-based matching: Automatically match the branch operation profile with the baseline feature template. For example, when a branch is identified as a switching power supply type load, it is automatically matched with the harmonic template. .

[0117] N22, Risk-weighted adjustment: Based on the historical anomaly probability of this branch, which comes from the historical database, the matching template is fine-tuned. For high-risk road closures, the sampling frequency of the feature extraction template will be increased.

[0118] N23. ​​Unified Monitoring Area: To ensure data consistency within the same monitoring area, the cloud server performs collaborative calibration on the assignment templates of all branches within the same monitoring area, aligning the output feature vectors in terms of dimension and timestamp, thus forming a unified area-level feature extraction standard for the monitoring area.

[0119] N3, cloud-based delivery and edge activation;

[0120] N31, Command Encapsulation and Secure Transmission: The cloud server encapsulates the assigned template and target branch / monitoring area identifier to generate encrypted policy configuration commands.

[0121] Furthermore, the policy configuration command is sent to the master node of the target monitoring area, and then distributed to other branches within the same monitoring area through the master node.

[0122] N32. Edge device loading and operation: After receiving the instruction, the monitor verifies and stores the template.

[0123] N4. Closed-loop optimization based on operational feedback;

[0124] Furthermore, a master node is selected in the monitoring network of each monitoring area, including:

[0125] Each monitor broadcasts its own status information packet during network initialization or according to a preset period. The status information packet includes at least the device performance score, signal quality strength, and remaining energy level.

[0126] The status information packet contains at least the following:

[0127] Equipment performance rating;

[0128] Signal quality strength;

[0129] Remaining energy level; characterizes the stability of the monitor's energy storage capacitor or power supply.

[0130] Each monitor within the monitoring area receives and evaluates the status information packets of adjacent branches;

[0131] The monitoring channel of the monitors in the same monitoring area receives status information packets broadcast by all other monitors in the monitoring area, and then parses and verifies the status information packets.

[0132] Based on a pre-defined comprehensive weighting algorithm, the election score of each monitor is calculated;

[0133] Each monitor calculates its election score based on the same pre-defined comprehensive weighting algorithm;

[0134] The calculation formula is:

[0135]

[0136] Where Score is the campaign score, Q is the performance score, A is the signal quality strength, and Z is the remaining energy level;

[0137] in, The preset weighting coefficients, and .

[0138] The monitor with the highest election score is selected as the master node of this monitoring area.

[0139] Further preliminary edge analysis includes:

[0140] The master node constructs a district-level multivariate time series feature matrix at the current detection time based on the preprocessed data packets received from all branches within the monitoring area.

[0141] The master node collects time-synchronized, standardized, multi-dimensional power parameter feature data of all branches within the monitoring area after edge preprocessing. Using the time-series sampling points within the preset detection window as matrix rows and each standardized power parameter feature of each branch as matrix columns, it completes the alignment, regularization, and splicing of the time-series data of all branches, ultimately forming a standardized district-level multivariate time-series feature matrix adapted to the input requirements of the district-level edge anomaly localization model.

[0142] For example, structured multi-dimensional features are extracted from each standardized edge processing data packet, such as the effective voltage value of a certain branch, the third harmonic content, the effective current value of another branch besides a certain branch, the waveform distortion rate, etc., and the feature values ​​of all branches at the same time are arranged according to a fixed branch order and feature dimensions to form a two-dimensional matrix.

[0143] Each row of the two-dimensional matrix represents a branch, and each column represents a feature dimension. The values ​​in the two-dimensional matrix are the variable values ​​of each feature of each branch at that moment.

[0144] The pre-set district-level edge anomaly localization model is invoked, and the district-level multivariate time series feature matrix is ​​compared and analyzed with the baseline pattern of inter-path correlation learned by the monitoring area under the historical normal operation state.

[0145] The correlation baseline model is generated by the global energy consumption analysis and fault tracing model of the cloud server based on the long-term historical normal power parameter data of the detection area. It is then distributed to the master node of the current detection area for storage and retrieval. It can be a parameter of a multivariate Gaussian distribution, or the encoder part of a trained lightweight autoencoder.

[0146] The district-level edge anomaly localization model identifies one or more candidate anomaly branches that are significantly different from the behavioral characteristics of other branches in the district by calculating the deviation of each branch data vector from the baseline pattern of the correlation relationship.

[0147] If the baseline model of the association relationship is a multivariate normal distribution model, the multivariate normal distribution model will calculate the Mahalanobis distance of the feature vector of each branch relative to the multivariate normal distribution, which is the deviation.

[0148] Anomaly identification includes:

[0149] Overall deviation: First, calculate the overall deviation of the state matrix of the entire monitoring area;

[0150] Individual localization: The district-level edge anomaly localization model is used to inversely deduce the contribution of each branch feature vector to the overall deviation.

[0151] Candidate identification: Paths with significantly higher deviations than other paths are marked as candidate abnormal paths.

[0152] The identifiers of candidate abnormal paths and their corresponding deviation characteristics are used as key components of the preliminary analysis results.

[0153] Furthermore, after the master node obtains the preliminary analysis results and feature data for this monitoring area, it performs adaptive data upload, including:

[0154] Based on the preliminary analysis results obtained from the initial edge analysis, the operating status of this monitoring area is determined. If it is determined to be in a normal state, the master node uploads the filtered summary feature data and low-frequency summary data to the cloud server. If it is determined to be in a potentially abnormal state or the certainty of the state is lower than the preset threshold, the master node increases the data upload frequency to the cloud server and uploads more detailed raw data fragments or complete standardized edge preprocessing data packets. The cloud server dynamically adjusts the computing resources allocated to the corresponding monitoring area according to the characteristics of the data uploaded by each monitoring area.

[0155] Triggering and Basis for Adaptive Decision-Making: Once the master node within the monitoring area completes preliminary edge analysis and generates primary analysis results and feature data for the monitoring area, the adaptive data upload process is initiated. The core basis for decision-making is the primary analysis results. The master node analyzes the details of the results and extracts key judgment information, mainly including:

[0156] Overall abnormal status indicator: Based on the comprehensive deviation of all branches, determine whether the monitoring area is in a normal or abnormal state.

[0157] Candidate Anomaly Path List: The specific path identifier for the identified behavioral anomalies;

[0158] State certainty index: A quantitative value that reflects the confidence level of the analysis results. For example, this index value is low when the deviation of a certain branch fluctuates just around the threshold, or when the data quality is temporarily reduced due to interference.

[0159] Specific implementation of the dual-mode upload strategy:

[0160] The first level is the determination of operational status and decision-making.

[0161] If the status is determined to be normal: this indicates that the preliminary edge analysis did not find any significant abnormalities and the monitoring area is operating stably. In this case, the system adopts the low-frequency summary upload mode.

[0162] If a potential abnormal state is identified or the certainty of the state is below a preset threshold, this means that suspicious signs have been detected, or the current data is insufficient to make a clear judgment. In this case, the system immediately switches to high-frequency detailed data upload mode.

[0163] The second level involves dynamic adjustment of data content and frequency;

[0164] Implementation of low-frequency summary upload mode under normal conditions:

[0165] Uploaded content: The master node performs secondary filtering and aggregation on the aggregated feature data, and generates and uploads the filtered summary feature data and low-frequency summary data.

[0166] Selected summary feature data: Usually only the core features that can reflect the operating trend are uploaded, such as the average hourly power of each branch, daily power consumption, and main features, such as the daily average value of total harmonic distortion, which greatly reduces the amount of data.

[0167] Low-frequency aggregated data: The upload cycle has been significantly extended, for example, from real-time uploads to status reports uploaded every 15 minutes or hour.

[0168] Objective: To maximize the conservation of network bandwidth and cloud storage resources while ensuring that the basic operational status of the monitoring area is monitored from the cloud.

[0169] Implementation of high-frequency detailed data upload mode under abnormal or uncertain conditions:

[0170] Triggered actions: The master node automatically performs two actions: increasing the data upload frequency and switching the data upload content.

[0171] Upload content and frequency:

[0172] Feature data: Continuously upload complete, high-frequency, multi-dimensional feature datasets for in-depth trend analysis in the cloud;

[0173] Raw data fragments: During the critical time period following the trigger, more detailed raw data fragments are uploaded synchronously. These more detailed raw data fragments are uncompressed or low-compressed raw waveform data, which are crucial for accurate fault type diagnosis and waveform analysis in global energy consumption analysis and fault tracing models.

[0174] Complete preprocessed data packet: For candidate outlier routes, upload the complete standardized edge preprocessed data packet to provide the cloud with the most comprehensive feature view of the route.

[0175] Frequency increase: The upload interval may be shortened from minutes to seconds or even sub-seconds, enabling intensive sampling of this abnormal process.

[0176] Objective: To provide sufficient, high-quality, and timely data for cloud servers to perform global energy consumption analysis and fault tracing model fusion analysis, ensuring that faults can be accurately diagnosed and traced.

[0177] Cloud-based resource collaborative scheduling:

[0178] Dynamic resource adjustment: Based on this judgment, the cloud server's resource scheduler dynamically adjusts the computing resources allocated to the corresponding monitoring area.

[0179] For monitoring areas in low-frequency summary mode, basic computing resources are allocated for data archiving and long-term trend statistics.

[0180] For monitoring areas that trigger high-frequency detailed data mode, more CPU, memory, and AI inference resources are immediately and elastically allocated to prioritize the processing of the detailed data uploaded therein, accelerate the execution of global fusion analysis and alarm generation, and achieve rapid cloud response to abnormal events.

[0181] Synergy with the overall technical solution: Through intelligent data reporting strategies, the system's monitoring depth and response speed for abnormal events are significantly improved at the same communication cost; or, at the same performance requirements, the long-term communication and cloud resource overhead of the system is greatly reduced.

[0182] Furthermore, it also includes predictive maintenance steps: the cloud server's global energy consumption analysis and fault tracing model predicts the performance aging trend and remaining service life of key electrical components in each branch based on long-term monitoring historical data; when the predicted remaining service life is lower than a preset threshold, predictive maintenance suggestions are generated and output.

[0183] Its logical flow and system collaboration relationship are as follows:

[0184] Data Foundation: The cloud server's global energy consumption analysis and fault tracing model continuously aggregates and stores long-term historical data uploaded from master nodes in various monitoring areas, after preprocessing and preliminary analysis. This data includes not only alarm records but also massive amounts of normal operating parameters and their subtle changes.

[0185] Model Learning and Prediction: The global energy consumption analysis and fault tracing model is not only used for immediate anomaly diagnosis, but also operates as a deep time-series prediction model. It analyzes key electrical components of each branch, such as circuit breaker contacts, cable joints, and metering modules, and related characteristic parameters, such as contact resistance trends, local temperature rise characteristics, specific harmonic growth patterns, and the long-term variation sequence of slowly changing insulation leakage current.

[0186] Predictive output: By learning the complete evolution pattern of components from normal to failure in historical data, the model can predict the performance aging trend of specific components in each branch and calculate an estimate of their remaining service life.

[0187] Decision and Output: When the model predicts that the remaining service life of a component is lower than a preset safety threshold, the system will not wait for a fault to occur before issuing an alarm, but will proactively generate and output predictive maintenance suggestions. These suggestions accurately target specific branches and suspected components, and provide an estimated risk time window.

[0188] Furthermore, the global energy consumption analysis and fault tracing model performs a time-series and topology-based fusion analysis on the preliminary analysis results and feature data uploaded by each master node, including:

[0189] Using the data streams uploaded by each monitoring area and their physical topology connections within the electricity metering box as input, each monitoring area or branch is abstracted as a graph node, and the electrical connection relationships and physical adjacency relationships are abstracted as graph edges to construct a graph neural network model. The graph neural network model is used to perform graph embedding learning on the global energy consumption correlation pattern under normal conditions, and to detect abnormal subgraph structures in the actual data that deviate from the normal pattern, thereby locating the root cause branch causing the anomaly. Based on the verification information of the root cause branch causing the anomaly and the external input, the internal parameters of the global energy consumption analysis and fault tracing model are continuously adjusted, and the deviation judgment threshold used by the district-level edge anomaly location model is optimized and updated.

[0190] The district-level edge anomaly localization model is a lightweight time-series anomaly detection model deployed on the main nodes of each monitoring area and adapted to edge computing power limitations. It is used for the preliminary identification of candidate anomaly paths within the monitoring area. The construction method is as follows:

[0191] First, collect time-series data of standardized power parameters of normal operating branches and district-level multivariate time-series feature matrix for no less than 30 days in the corresponding monitoring area. Then, supplement abnormal operating condition samples labeled with abnormal types and abnormal branch identifiers within the area to complete the construction and unified preprocessing of the training dataset.

[0192] Furthermore, a lightweight temporal convolutional network backbone architecture adapted to edge computing power is built to extract the temporal correlation features of power parameters of each branch in this monitoring area;

[0193] Correspondingly, the completed dataset is used to supervise the training of the graph neural network, so that the district-level edge anomaly localization model can learn and solidify the baseline pattern of the correlation between the power parameter changes of each branch under normal operating conditions in the monitoring area, and complete the model weight convergence and core function solidification.

[0194] A deviation calculation module based on Mahalanobis distance and an initial deviation judgment threshold are configured for the district-level edge anomaly localization model. Candidate anomaly branch identification is achieved by calculating the deviation between the real-time branch data vector and the baseline mode. Finally, a parameter adaptation unit is configured for the district-level edge anomaly localization model so that it can receive optimized parameters issued by the global energy consumption analysis and fault tracing model, and dynamically update the deviation judgment threshold and the lightweight weight of the district-level edge anomaly localization model. These will not be described in detail here.

[0195] The global energy consumption analysis and fault tracing model is as follows:

[0196] Specifically, the deep analysis model, deployed on a cloud server and based on the fusion of graph neural network (GNN) and time-series event sequence analysis, has sufficient cloud computing power to support it. It is adapted to the needs of global anomaly tracing, fault propagation path identification and system-wide optimization across monitoring areas. The input of the global energy consumption analysis and fault tracing model is the primary analysis results and feature data uploaded by the master nodes of each monitoring area, as well as the physical topology connection relationship, electrical connection relationship and physical adjacency relationship data of all branches in the power metering box and the monitoring area.

[0197] Furthermore, the global energy consumption analysis and fault tracing model has a built-in graph construction unit that can abstract each monitoring area and each branch into graph nodes at different levels. It can also abstract the electrical connection relationship, physical adjacency relationship, and correlation of historical operation data between branches / monitoring areas into weighted graph edges, construct a heterogeneous graph structure that is completely mapped to the physical topology of the power metering box, and complete graph embedding learning. It can then fuse topological relationships and time-series features into a unified graph feature vector.

[0198] Correspondingly, the pre-training stage of the global energy consumption analysis and fault tracing model uses the target energy metering box and the full normal operation history dataset of the same type of metering box, as well as the labeled dataset of various abnormal working conditions to complete supervised training, learn and solidify the energy consumption correlation benchmark pattern and fault propagation characteristics across monitoring areas and across branches under the global normal operation state of the system.

[0199] Meanwhile, during the inference phase, the global energy consumption analysis and fault tracing model compares the real-time input data from each monitoring area with the pre-trained global correlation benchmark pattern. It identifies abnormal subgraph structures that deviate from the normal pattern through a graph attention mechanism. Combined with the event sequence analysis algorithm, it accurately identifies abnormal energy consumption data and fault propagation paths across monitoring areas based on the chronological order of anomalies and the correlation of feature transmission. At the same time, it locates the root cause branch that leads to global anomalies through a causal inference algorithm. Finally, it outputs the analysis results of the root cause branch identifier, anomaly type, and fault impact range to the alarm and output module to generate corresponding alarm information.

[0200] In addition, the global energy consumption analysis and fault tracing model has a built-in model optimization unit that can continuously iterate and adjust the weight parameters and graph edge weight coefficients within the global energy consumption analysis and fault tracing model based on the on-site feedback results of location alarm information and the manual verification information input from the outside. At the same time, it can send the optimized deviation judgment threshold to the district-level edge anomaly location model in the main node of each monitoring area to complete the adaptation optimization of the district-level model, forming a closed-loop optimization mechanism of cloud global optimization and edge district-level adaptation. Furthermore, based on the full historical data of long-term monitoring, it can predict the performance aging trend and remaining service life of key electrical components in each branch, supporting the realization of predictive maintenance functions.

[0201] The input is integrated: The input of the global energy consumption analysis and fault tracing model comes from the primary analysis results and feature data streams uploaded by each monitoring area, and the physical topology structure defined by the zoning strategy and electrical drawings, which reflects the actual connection relationship between branches, is also input simultaneously, forming the data foundation for spatiotemporal fusion analysis.

[0202] Modeling and analysis of graph neural networks:

[0203] Modeling: Construct a graph neural network model.

[0204] Specifically, each monitoring area or key branch is abstracted as a graph node, and the electrical connections, physical adjacencies, and other relationships between them are abstracted as graph edges.

[0205] Learning: Local energy consumption analysis and fault tracing model. By learning from long-term normal historical data, we can grasp the global energy consumption correlation pattern between each branch of the entire metering box under healthy conditions.

[0206] Detection: When an anomaly occurs, the graph neural network model inputs real-time data into the learned normal graph pattern for comparison to detect abnormal subgraphs in the actual data that deviate from the normal pattern, that is, to identify abnormal collaborative changes in the connection relationships between branches.

[0207] Root cause localization: By analyzing the abnormal subgraph, isolated abnormal points are identified, and possible propagation paths of the abnormality on the topology network are inferred. This distinguishes between end-load failures and problems with upstream lines or common equipment, ultimately locating the single root cause branch that triggered the chain reaction.

[0208] Specifically, the following describes the method for constructing a global energy consumption analysis and fault tracing model, including:

[0209] S71. Collect all historical power parameter data of each branch and monitoring area in the target power metering box within the set historical period, as well as the corresponding primary analysis results and feature data. Simultaneously collect data on the physical topology connection relationship, electrical connection relationship, and physical adjacency relationship between the branch and the monitoring area in the power metering box. At the same time, supplement the collection of abnormal operating condition sample data marked with abnormal type, root cause branch, and fault propagation path to complete the construction and standardization preprocessing of the dataset.

[0210] It should be noted here that the preliminary analysis results and feature data refer to the standardized time-series dataset with unique identifiers generated by the main nodes of each monitoring area of ​​the power metering box based on the standardized edge preprocessing data packets uploaded by all branch monitors in this monitoring area, after data aggregation and multi-dimensional feature statistical calculations.

[0211] The full historical power parameter data includes the effective value of branch current, active power, reactive power, load curve, power change rate, and impedance characteristic value.

[0212] The preliminary analysis results and feature data include the identification of candidate abnormal branches, the deviation characteristics of the branches, and the monitoring area-level multivariate time series feature matrix;

[0213] Among them, the physical topology connection relationship includes the subordinate relationship between the branch and the monitoring area, the electrical connection relationship between the branches, and the adjacent relationship of the physical installation position of the branches in the metering box;

[0214] The abnormal operating condition sample data includes the timestamp of the abnormality, the type of abnormality, the unique identifier of the root cause branch, the time sequence path of the fault propagation and the scope of impact. The types of abnormalities include abnormal energy consumption, line faults and poor contact.

[0215] Based on the above, the structured definition expression for the dataset is as follows:

[0216]

[0217] In the formula:

[0218] D represents the complete training dataset for the global energy consumption analysis and fault tracing model;

[0219] A complete historical power parameter dataset, and N is the total number of branches in the power metering box, T is the total number of sampling points in the set historical period, and 6 is the full historical power parameter data;

[0220] This is the preliminary analysis results and feature dataset, and M represents the total number of monitoring areas, T represents the total number of sampling points within the set historical period, and F represents the feature dimension, which refers to the total number of independent feature variables that characterize the primary analysis results and feature data of the monitoring area without redundancy. Specifically, it corresponds to the total number of independent feature terms contained in the candidate anomaly branch identifier, branch deviation feature, and district-level multivariate time series feature matrix.

[0221] It is a topological relation dataset, and These correspond to the connection relationship matrices between branch paths and monitoring intervals, respectively. A matrix element of 1 indicates the existence of a corresponding connection / adjacency relationship, while a matrix element of 0 indicates the absence of a corresponding connection / adjacency relationship.

[0222] This is a dataset of abnormal operating conditions.

[0223] For the preprocessing of the full historical power parameter dataset, Z-score normalization is used to eliminate differences in the magnitude of different parameters. The expression is as follows:

[0224]

[0225] In the formula:

[0226] x represents the original continuous characteristic value, such as current, power, and deviation characteristics;

[0227] This is the mean of the feature across all normal training samples;

[0228] The standard deviation of this feature in the full set of normal training samples, to avoid division by zero, when At that time, take = ;

[0229] These are the standardized feature values.

[0230] In the preprocessing of the topological relation dataset, adjacency matrix normalization is used to adapt to the input requirements of graph neural networks. The expression is as follows:

[0231]

[0232] In the formula,

[0233] A is the original topological adjacency matrix, corresponding to ;

[0234] I is the identity matrix;

[0235] D is the degree matrix;

[0236] The normalized adjacency matrix serves as the topological input for the global energy consumption analysis and fault tracing model.

[0237] For the preprocessing of the abnormal operating condition sample dataset, ordinal encoding is used to convert the classification labels into a numerical format recognizable by the global energy consumption analysis and fault tracing model. The expression is as follows:

[0238]

[0239] In the formula, These are raw text labels, such as exception type and root cause identifier;

[0240] A mapping function from labels to unique serial numbers;

[0241] The encoded label value is used for supervised training of the global energy consumption analysis and fault tracing model.

[0242] S72. Each monitoring area and each branch is abstracted into graph nodes of different levels. The electrical connection relationship, physical adjacency relationship and correlation of historical operation data between branches and monitoring areas are abstracted into weighted graph edges. A heterogeneous graph structure is constructed that maps to the physical topology of the target power metering box. The heterogeneous graph structure maps to the physical topology, electrical connection and operation characteristics of the target power metering box.

[0243] The different levels of graph nodes include:

[0244] Upper-level nodes: Monitoring area level nodes, each node uniquely corresponds to one monitoring area, and the total number of nodes is equal to the total number of monitoring areas M;

[0245] Lower-level nodes: Branch-level nodes, each node uniquely corresponds to one branch, the total number of nodes is equal to the total number of branches N, and each branch node uniquely belongs to the upper-level node of the corresponding monitoring area;

[0246] The heterogeneous graph structure is defined as follows:

[0247]

[0248] In the formula,

[0249] A heterogeneous graph structure that is completely mapped to the physical topology of the target energy metering box;

[0250] For different levels of graph node sets, and , For monitoring the set of nodes at the hierarchical level, A set of nodes at the branching level;

[0251] For the graph edge set, ;

[0252] in, For electrical connection relationships, connect branch nodes / monitoring area nodes that have electrical connections;

[0253] For physically adjacent edges, connect branch nodes / monitoring area nodes that are physically adjacent in installation location;

[0254] For historical operational data correlation edges, connect branch nodes / monitoring area nodes where operational data have a strong correlation;

[0255] Let be the edge weight matrix, and .

[0256] Furthermore, the weights for the aforementioned electrical connection edges, physical adjacency edges, and historical operational data correlation edges are calculated as follows:

[0257] A. Edge weights for electrical connections: The weight is 1 if an electrical connection exists, and 0 if no electrical connection exists.

[0258]

[0259] B. Weight of physical adjacency: If the physical locations are directly adjacent, the weight is 1; if they are separated by one branch within the same monitoring area, the weight is 0.5; otherwise, the weight is 0.

[0260]

[0261] C. Historical operational data correlation edge weights: The Pearson correlation coefficient is used to quantify the correlation between operational data of different routes, normalized to the range [0,1].

[0262]

[0263] In the formula:

[0264] ) represents the Pearson correlation coefficient of the historical power parameter time series data of node i and node j, with a value range of [-1, 1].

[0265] The normalized operational data correlation edge weights have a value range of [0,1], corresponding to the correlation of historical operational data.

[0266] S73. Construct a graph neural network that integrates graph attention mechanism and temporal convolution module. Through graph neural network, perform graph embedding learning on heterogeneous graph structures, and integrate the topological relationship features of branch and monitoring areas and the temporal power parameter features into a unified graph feature vector.

[0267] Specifically, the graph neural network adopts a two-branch fusion architecture, which will be referred to as branch 1 and branch 2 below:

[0268] Branch 1: Temporal Convolutional Module (TCN), which is responsible for extracting the temporal power parameter features of the branch / monitoring area;

[0269] Branch 2, Graph Attention Mechanism (GAT) module, is responsible for extracting topological relationship features of heterogeneous graphs;

[0270] End point: Feature fusion layer, which merges temporal features and topological features into a unified graph feature vector.

[0271] Furthermore, the expression for temporal feature extraction of the temporal convolution module is as follows:

[0272]

[0273] In the formula,

[0274] The time-series power parameter characteristics output by the time-series convolutional module at time t;

[0275] K is the kernel size, preferably 3, which can be adjusted in the range of 2 to 7 depending on the sampling interval;

[0276] d is the expansion coefficient, and the preferred value is [value missing]. , The number of convolutional layers is set to ensure coverage of long time periods;

[0277] = This is the set of convolution kernel weights for the temporal convolution module;

[0278] for Time-series power parameter data after time standardization.

[0279] For topological feature extraction in the graph attention mechanism module, a multi-head attention mechanism is adopted to capture the relationships between nodes, and its expression is:

[0280]

[0281] In the formula,

[0282] The topological relationship features output by node i;

[0283] H represents the number of attention heads, which is preferably 4 in this embodiment;

[0284] Let i be the set of neighboring nodes of node i;

[0285] Under the h-th attention head, node i's node association attention with its neighbor node j.

[0286] Force weight set;

[0287] Let h be the set of feature mapping weight matrices for the h-th attention head;

[0288] Let be the initial feature vector of node j;

[0289] ·) is a non-linear activation function, preferably the ReLU function.

[0290] For feature fusion and graph embedding output, temporal features and topological features are concatenated and fused to obtain a unified graph feature vector, thus completing graph embedding learning. The expression is:

[0291]

[0292] In the formula,

[0293] The unified graph feature vector output by graph embedding learning;

[0294] This is a feature concatenation operation to ensure the complete fusion of temporal and topological features; the output... These features serve as input for subsequent supervised training.

[0295] S74. The constructed dataset is used to supervise the training of the graph neural network, so that the trained global energy consumption analysis and fault tracing model learns and solidifies the global energy consumption correlation benchmark pattern and fault propagation characteristic law under the normal operation state of the target power meter box. The trainable parameters of the global energy consumption analysis and fault tracing model are converged and the core functions are solidified. The trainable parameters include the node association attention weights corresponding to the graph attention mechanism, the convolution kernel weights of the temporal convolution module, the feature mapping weights of graph embedding learning, and the deviation calculation weights of abnormal mode judgment.

[0296] It should first be noted that in this embodiment, the training environment uses the Adam optimizer, and the initial learning rate is preferably... The optimal batch size is 32, the optimal number of training rounds is 100, and the early stop strategy is to stop training if the loss does not decrease after 20 consecutive rounds of verification.

[0297] Furthermore, the global energy consumption correlation benchmark model is a benchmark Gaussian distribution of the correlation relationship and fluctuation range of power parameter changes in each branch / monitoring interval, which is learned by the global energy consumption analysis and fault tracing model on the full set of normal samples.

[0298] Meanwhile, the fault propagation characteristics are learned from the global energy consumption analysis and fault tracing model on abnormal samples, and the time sequence, correlation weight, and characteristic change magnitude of the abnormal characteristics propagating from the root cause branch to adjacent branches / across monitoring areas are quantified.

[0299] The expression for the set of trainable parameters is:

[0300]

[0301] In the formula,

[0302] It is a complete set of trainable parameters for global energy consumption analysis and fault tracing models;

[0303] This refers to the set of attention weights associated with the nodes in the graph attention mechanism.

[0304] This is the set of convolution kernel weights for the temporal convolution module;

[0305] The set of feature mapping weights for graph embedding learning;

[0306] Calculate the set of weights for the deviation of the abnormal mode determination.

[0307] Correspondingly, by calculating the weights based on the deviation, the expression for determining the abnormal mode is derived:

[0308]

[0309] In the formula,

[0310] The score is the abnormal deviation score for node i. The higher the score, the greater the probability of abnormality.

[0311] The deviation calculation weight vector for abnormal mode determination;

[0312] Let i be the real-time graph feature vector of node i;

[0313] This represents the baseline mean of the graph feature vectors under normal operating conditions, corresponding to the global energy consumption correlation baseline mode.

[0314] Furthermore, it is necessary to construct a supervised training joint loss function, employing a two-branch joint loss function, to simultaneously fit the normal baseline pattern and the fault propagation law, with the expression as follows:

[0315]

[0316] In the formula,

[0317] The normal mode fitting loss uses mean squared error loss to enable the model to learn and solidify the global energy consumption correlation benchmark pattern. The formula is as follows:

[0318]

[0319] In the formula,

[0320] For fault classification loss, cross-entropy loss is used to enable the model to learn and solidify the fault propagation characteristics, accurately identify anomaly types and root cause paths. The expression is:

[0321]

[0322] in, This refers to the anomaly prediction results output by the global energy consumption analysis and fault tracing model. The actual label is used for marking.

[0323] The loss weighting balancing coefficient is preferably set to 0.3, and can be adjusted within the range of 0.1 to 0.5 according to the ratio of normal samples to abnormal samples.

[0324] Furthermore, when the global energy consumption analysis and fault tracing model simultaneously satisfy the following two conditions, they will be considered converged, thus completing the solidification of trainable parameters and core functions:

[0325] C1. Loss value stability condition: The total loss value change over 10 consecutive training rounds is less than a preset convergence threshold, expressed as:

[0326]

[0327] In the formula, Let be the total loss value for the nth round of training;

[0328] In this embodiment, the preferred convergence threshold is... .

[0329] S75. Configure a parameter optimization unit for the trained global energy consumption analysis and fault tracing model, so that the global energy consumption analysis and fault tracing model can continuously adjust the trainable parameters according to the feedback results of the location alarm information and the external input verification information.

[0330] The parameter optimization unit is an online learning module built into the global energy consumption analysis and fault tracing model. The triggering conditions are: when 100 alarm messages with feedback are received, or every 30 days at a fixed period, or when batch labeled data with manual verification is received, the parameters are updated.

[0331] The feedback loss function is expressed as follows:

[0332]

[0333] In the formula,

[0334] This is a feedback loss function constructed based on feedback and verification information;

[0335] B represents the batch size of the validation samples with feedback.

[0336] These are the actual labels obtained from on-site feedback / manual verification, such as anomaly type and root cause branch.

[0337] This is the original prediction result of the global energy consumption analysis and fault tracing model.

[0338] Based on the above feedback loss function expression, the parameters can be iteratively trained, where the iterative update expression for the trainable parameters is:

[0339]

[0340] In the formula,

[0341] The set of trainable parameters before the update;

[0342] The optimized and updated set of trainable parameters;

[0343] For incremental learning, the preferred learning rate is [value]. , which is the initial training learning rate .

[0344] To provide feedback on the gradient of the loss function with respect to the trainable parameters, and to achieve continuous adaptive adjustment of the trainable parameters.

[0345] Secondly, this application discloses a power energy consumption monitoring system for an electricity metering box, which is applied to the power energy consumption monitoring method for the electricity metering box as described in the first aspect.

[0346] include:

[0347] The partitioning module is used to obtain the preset partitioning strategy of the power metering box and divide at least two branches in the power metering box into a monitoring area according to the partitioning strategy.

[0348] The monitor module is used to acquire power parameter data from another branch.

[0349] The monitoring network construction module is used by each monitor to form a partitioned monitoring network based on the Internet of Things protocol, and a master node is selected in the monitoring network of each monitoring area;

[0350] The edge preprocessing module is used for each monitor to continuously collect current parameter data of the branch circuits connected to the detection channel;

[0351] The data aggregation and preliminary analysis module is used by monitors other than the master node to send the pre-processed power parameter data from the edge to the master node. The master node aggregates and performs preliminary edge analysis on the data within its monitoring area to obtain the preliminary analysis results and feature data of the monitoring area.

[0352] The cloud server module is used to receive the preliminary analysis results and feature data uploaded by each master node;

[0353] The alarm and output module is used to generate and output location alarm information containing root source branch identifier, anomaly type and fault impact range based on abnormal energy consumption data, fault propagation path and root source branch.

[0354] The predictive maintenance module generates and outputs predictive maintenance recommendations when the predicted remaining useful life is lower than a preset threshold.

[0355] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for monitoring the power consumption of an electricity metering box, characterized in that, include: Obtain the preset zoning strategy for the electricity metering box, and divide at least two branches in the electricity metering box into a monitoring area according to the zoning strategy; Each branch in each monitoring area is equipped with a monitor. The power supply channel of the monitor is connected to one of the branches, and the monitoring channel of the monitor is connected to another branch adjacent to the current branch to obtain the power parameter data of the other branch. Each monitor forms a zoned monitoring network based on the Internet of Things protocol, and a master node is selected in the monitoring network of each monitoring zone; Each monitor continuously collects power parameter data of the branch circuits connected to the monitoring channel and performs edge preprocessing; Within each monitoring zone, all monitors except the master node send the pre-processed power parameter data to the master node. The master node then aggregates and performs preliminary edge analysis on the data within its monitoring zone to obtain the initial analysis results and characteristic data for that monitoring zone. Each master node uploads the preliminary analysis results and feature data to the cloud server; A global energy consumption analysis and fault tracing model is pre-built. The cloud server uses the global energy consumption analysis and fault tracing model to perform a fusion analysis based on event sequence and topology on the primary analysis results and feature data uploaded by each master node, and identifies abnormal energy consumption data, fault propagation paths and root cause branches across monitoring areas. The global energy consumption analysis and fault tracing model performs a fusion analysis based on event sequences and topology structure on the preliminary analysis results and feature data uploaded by each master node, including: Using the data streams uploaded by each monitoring area and their physical topology connections within the power metering box as input, each monitoring area or branch is abstracted as a graph node, and the electrical connection relationship and physical adjacency relationship are abstracted as graph edges to construct a graph neural network model. The graph neural network model is used to learn the global energy consumption correlation pattern under normal conditions and detect abnormal subgraph structures that deviate from the normal state in the actual data to locate the root cause of the abnormality. Based on the root cause of the anomaly and the verification information of the external input, the internal parameters of the global energy consumption analysis and fault tracing model are continuously adjusted, and the deviation judgment threshold used by the district-level edge anomaly localization model is optimized and updated. Based on abnormal energy consumption data, fault propagation paths, and root cause branches, generate and output location alarm information that includes root cause branch identifiers, abnormality types, and fault impact ranges.

2. The method for monitoring the power consumption of an electricity metering box according to claim 1, characterized in that, The zoning strategy for obtaining electricity metering boxes includes: Collect historical power parameter data for each branch within a set period; Cluster analysis is performed on historical power parameter data to identify candidate branch groups with high correlation based on the change patterns of power parameter data, load curves, and power characteristics of branches. Based on the correlation analysis results of candidate route groups, and combined with the physical adjacency constraints between routes, routes with high correlation and physical adjacency are assigned to the same monitoring area, generating an optimized zoning strategy.

3. The method for monitoring the power consumption of an electricity metering box according to claim 2, characterized in that, Equipping each branch within the monitoring area with a monitor includes: Each branch circuit in the electricity metering box is configured with an independent node unit, and all node units are physically arranged to form a point array; The node units corresponding to the branches belonging to the same monitoring area are electrically connected through a continuous series detection link; Each monitor is integrated into a node unit, and its monitoring channel is used to detect the electrical on / off state or impedance characteristics of the series detection link at this node unit; When an electrical abnormality occurs in the branch corresponding to any node unit in the series detection link, the series detection link will show a state interruption or a sudden change in characteristic parameters at that node. By monitoring the communication mechanisms within the network, the link status of each node unit is queried, and abnormal branching is located based on the location information of link status interruption or sudden change in characteristic parameters.

4. The method for monitoring the power consumption of an electricity metering box according to claim 1, characterized in that, Edge preprocessing is performed, including: The raw analog signals from the monitoring channel are acquired at a fixed sampling rate and then converted from analog to digital to obtain the raw power parameter data stream. Data quality assessment and primary filtering are performed on the raw power parameter data stream to remove outlier points, and a moving average filtering algorithm is used to smooth the data. Based on a preset feature extraction template, multi-dimensional features are extracted in parallel from the filtered data stream; Based on the current IoT network load and the historical anomaly probability of the routing, a compression algorithm is dynamically selected to compress and encode the multi-dimensional feature dataset, generating a compressed feature data packet. The compressed feature data packet, the type of compression algorithm used, the data acquisition time information, and the corresponding branch identification information are encapsulated to form a standardized edge preprocessing data packet.

5. The method for monitoring the power consumption of an electricity metering box according to claim 1, characterized in that, In the monitoring network of each monitoring area, a master node is selected, including: Each monitor broadcasts its own status information packet during network initialization or according to a preset period. The status information packet includes at least the device performance score, signal quality strength, and remaining energy level. Each monitor within the monitoring area receives and evaluates the status information packets of adjacent branches; Based on a pre-defined comprehensive weighting algorithm, the election score of each monitor is calculated; The monitor with the highest election score is selected as the master node of this monitoring area.

6. The method for monitoring the power consumption of an electricity metering box according to claim 4, characterized in that, Preliminary marginal analysis includes: The master node constructs a district-level multivariate time series feature matrix at the current detection time based on the preprocessed data packets received from all branches within the monitoring area. The pre-set district-level edge anomaly localization model is invoked, and the district-level multivariate time series feature matrix is ​​compared and analyzed with the baseline pattern of inter-path correlation learned by the monitoring area under the historical normal operation state. The district-level edge anomaly localization model identifies one or more candidate anomaly branches that are significantly different from the behavioral characteristics of other branches in the district by calculating the deviation of each branch data vector from the baseline pattern of the correlation relationship. The identifiers of candidate abnormal paths and their corresponding deviation characteristics are taken as key components of the preliminary analysis results; Among them, the district-level edge anomaly localization model is a lightweight time-series anomaly detection model deployed in the main nodes of each monitoring area.

7. A method for monitoring the power consumption of an electricity metering box according to claim 6, characterized in that, After the master node obtains the preliminary analysis results and feature data for this monitoring area, it performs adaptive data upload, including: Based on the preliminary analysis results obtained from the initial edge analysis, the operational status of this monitoring area is determined. If the status is determined to be normal, the master node uploads the filtered summary feature data and low-frequency summary data to the cloud server. If a potential abnormal state is identified or the certainty of the state is lower than a preset threshold, the master node increases the frequency of data uploads to the cloud server and uploads more detailed raw data fragments or complete standardized edge preprocessing data packets. The cloud server dynamically adjusts the computing resources allocated to the corresponding monitoring area based on the characteristics of the data uploaded from each monitoring area.

8. A method for monitoring the power consumption of an electricity metering box according to claim 3, characterized in that, It also includes predictive maintenance steps: The global energy consumption analysis and fault tracing model of cloud servers is based on historical data monitored over a long period of time to predict the performance aging trend and remaining service life of key electrical components in each branch. When the predicted remaining useful life is lower than a preset threshold, predictive maintenance recommendations are generated and output.

9. A method for monitoring the power consumption of an electricity metering box according to claim 1, characterized in that, Construct a global energy consumption analysis and fault tracing model, including: Collect all historical power parameter data of each branch and monitoring area in the target power metering box within a set historical period, and generate corresponding primary analysis results and feature data. Simultaneously collect data on the physical topology connection relationship, electrical connection relationship, and physical adjacency relationship between the branch and the monitoring area in the power metering box. At the same time, supplement the collection of abnormal operating condition sample data marked with anomaly type, root cause branch, and fault propagation path to complete the construction and standardization preprocessing of the dataset. Each monitoring area and branch circuit is abstracted into graph nodes of different levels. The electrical connection relationship, physical adjacency relationship and correlation of historical operation data between branches and monitoring areas are abstracted into weighted graph edges. A heterogeneous graph structure that maps to the physical topology of the target power metering box is constructed. A graph neural network integrating graph attention mechanism and temporal convolution module is constructed. The graph neural network performs graph embedding learning on heterogeneous graph structures, and integrates the topological relationship features of the branch and monitoring area and the temporal power parameter features into a unified graph feature vector. The constructed dataset is used to supervise the training of the graph neural network, so that the trained global energy consumption analysis and fault tracing model learns and solidifies the global energy consumption correlation benchmark pattern and fault propagation feature law under the normal operation state of the target power meter box. The trainable parameters of the global energy consumption analysis and fault tracing model are converged and the core functions are solidified. The trainable parameters include the node association attention weights corresponding to the graph attention mechanism, the convolution kernel weights of the temporal convolution module, the feature mapping weights of graph embedding learning, and the deviation calculation weights of abnormal mode judgment. A parameter optimization unit is configured for the trained global energy consumption analysis and fault tracing model, so that the global energy consumption analysis and fault tracing model can continuously adjust the trainable parameters based on the feedback results of the location alarm information and the external input verification information.

10. A power consumption monitoring system for an electricity metering box, applied to the power consumption monitoring method for the electricity metering box as described in any one of claims 1-9, characterized in that, include: The partitioning module is used to obtain the preset partitioning strategy of the power metering box and divide at least two branches in the power metering box into a monitoring area according to the partitioning strategy. The monitor module is used to acquire power parameter data from another branch. The monitoring network construction module is used by each monitor to form a partitioned monitoring network based on the Internet of Things protocol, and a master node is selected in the monitoring network of each monitoring area; The edge preprocessing module is used for each monitor to continuously collect current parameter data of the branch circuits connected to the detection channel; The data aggregation and preliminary analysis module is used by monitors other than the master node to send the pre-processed power parameter data from the edge to the master node. The master node aggregates and performs preliminary edge analysis on the data within its monitoring area to obtain the preliminary analysis results and feature data of the monitoring area. The cloud server module is used to receive the preliminary analysis results and feature data uploaded by each master node; The alarm and output module is used to generate and output location alarm information containing root source branch identifier, anomaly type and fault impact range based on abnormal energy consumption data, fault propagation path and root source branch. The predictive maintenance module generates and outputs predictive maintenance recommendations when the predicted remaining useful life is lower than a preset threshold.

Images