A fault classification and tracing method based on electric energy meter event log

Abstract

The application relates to the field of smart grids, and discloses a fault classification and tracing method based on an electric energy meter event log, which comprises the following steps: performing millisecond-level dynamic collection on power-off, recovery and abnormal voltage data of user electric meters and adjacent electric meters, and performing second-level time alignment based on a transformer area reference clock signal; performing multi-meter cross-time sequence correlation analysis on an event priority sequence, and constructing a power grid event multi-dimensional feature tensor; adopting a hierarchical causal reasoning model to perform abnormal aggregation on local modes in the tensor according to the power grid event multi-dimensional feature tensor; introducing a time sequence dependent attention mechanism on the basis of a candidate fault propagation path, and dynamically evaluating the contribution degree of different electric meter events to a potential fault; dividing electric meter nodes and affected branches corresponding to a candidate power-off reason set into a fault sensitive area, and outputting a tracing result; and the application has the advantages of real-time judgment of power-off reasons and automatic generation of a solution.

Description

Technical Field

[0001] This invention relates to the field of smart grids, specifically to a fault classification and tracing method based on electricity meter event logs. Background Technology

[0002] With the continuous development of power systems, electricity meters are widely used in distribution networks to record user electricity consumption and power outage events. However, in some remote areas, due to long power grid lines, complex geographical environments, and limited maintenance personnel, frequent power outages within a short period are common. These frequent outages not only severely impact users' electricity experience but also increase the pressure on power companies to handle customer complaints. In existing technologies, smart meters typically record outage and recovery events and upload logs to a backend system for analysis by maintenance personnel. However, this method has significant shortcomings: existing methods often rely on manual experience or simple rules for analysis, lacking automation, real-time performance, and interpretability, making it difficult to meet the needs of rapid tracing and handling of high-frequency power outages in remote areas. Therefore, it is essential to design a fault classification and tracing method based on electricity meter event logs that can determine the cause of power outages in real time and automatically generate solutions. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a fault classification and tracing method based on electricity meter event logs, which has the advantages of real-time determination of power outage causes and automatic generation of solutions, thus solving the problems mentioned in the background technology.

[0004] To achieve the aforementioned goal of real-time determination of power outage causes and automatic generation of solutions, this invention provides the following technical solution: a fault classification and tracing method based on electricity meter event logs, comprising the following steps:

[0005] The system dynamically collects power outage, restoration, and abnormal voltage data from user meters and nearby meters at the millisecond level, initiates a local clock deviation adaptive correction algorithm, performs second-level time alignment based on the transformer area reference clock signal, and generates an event priority sequence according to both fault urgency and data reliability.

[0006] Perform multi-table cross-time series correlation analysis on the event priority sequence, and construct a multi-dimensional feature tensor of power grid events by combining historical fault modes, event consistency of neighboring meters in the distribution area and communication quality indicators;

[0007] For the multidimensional feature tensor of power grid events, a hierarchical causal reasoning model is used to aggregate local patterns in the tensor to form anomalies, and then combined with the topological constraints of the distribution area to form candidate fault propagation paths.

[0008] Based on the candidate fault propagation path, a time-dependent attention mechanism is introduced to dynamically evaluate the contribution of different meter events to potential faults, and a candidate set of power outage causes is generated based on the topology propagation direction.

[0009] The meter nodes and affected branches corresponding to the candidate power outage cause set are divided into fault-sensitive areas. By utilizing the event similarity, self-consistency rules and the differential characteristics of the boundary neighboring tables within the area, a hierarchical local tracing graph is constructed, and the tracing results are output.

[0010] Preferably, the process of initiating the local clock skew adaptive correction algorithm is as follows:

[0011] Continuously collect data on voltage, current, power status, and event changes of user meters and nearby meters;

[0012] Based on the collected data, local clock drift characteristics are extracted, and adaptive time deviation correction is performed in combination with the local area reference clock signal;

[0013] Continuous monitoring of corrected event data to identify sudden power outages, power restorations, and abnormal voltage events;

[0014] Perform basic preprocessing on the event data, including outlier removal, missing value imputation, and filtering;

[0015] The preprocessed event data is stored according to meter identification and time series.

[0016] Preferably, the process of generating an event priority sequence based on both fault urgency and data reliability is as follows:

[0017] The meter events after adaptive time correction are classified into power outage, recovery, and voltage anomaly.

[0018] The categorized events are batch aggregated according to the set time window, and the urgency level is graded and assigned according to the event type and frequency of occurrence.

[0019] The reliability of the data is evaluated by weighting the communication quality indicators collected from the event and the consistency among nearby meters.

[0020] Output an event priority sequence with time consistency labels and risk stratification weights.

[0021] Preferably, the multi-table cross-time-series join sub-process for the event priority sequence is as follows:

[0022] The event priority sequence of each meter is divided according to time segments to establish a time overlap matrix between meters;

[0023] A cross-table matching method based on dynamic time warping eliminates misalignment caused by communication delays;

[0024] By combining the physical distance between adjacent meters and the power distribution relationship of the transformer substation, the synchronicity and lag between event priority sequences are compared and analyzed.

[0025] Generate a cross-temporal correlation vector set with temporal consistency measure, spatial proximity weight and topological constraint factor.

[0026] Preferably, the process of constructing the multidimensional feature tensor of power grid events is as follows:

[0027] Based on the cross-temporal correlation vector set, the historical fault case library is called to compare and extract typical pattern templates that match the characteristics of the current event, and output the pattern matching score.

[0028] Using the event consistency test method, the event consistency rate of neighboring meters within the same window is calculated and embedded as a consistency factor;

[0029] Based on the real-time packet loss rate and latency indicators of the communication link, a communication quality correction coefficient is generated;

[0030] By embedding pattern matching scores, consistency factors, and communication quality correction coefficients together with cross-time series correlation results into the tensor space, a multidimensional feature tensor of power grid events is constructed.

[0031] Preferably, the process of using a hierarchical causal reasoning model to aggregate local patterns in the tensor and then combining this with the topological constraints of the transformer area to form candidate fault propagation paths is as follows:

[0032] The multidimensional feature tensor of power grid events is sliced ​​into layers to extract local temporal anomalies, spatial adjacency anomalies, and cross-layer coupling anomalies respectively.

[0033] Attribution aggregation is performed on different anomaly types based on a causal reasoning rule base, and redundant factors unrelated to historical patterns are eliminated.

[0034] By combining the physical topology constraints of the distribution lines in the transformer area, event links with causal relationships are screened to form candidate fault propagation paths.

[0035] Preferably, the process of dynamically evaluating the contribution of different meter events to potential faults and generating a candidate set of outage causes based on the topology propagation direction is as follows:

[0036] A time-dependent attention mechanism is introduced on the candidate fault propagation path to calculate the weight distribution of each meter event in different time segments.

[0037] The contribution of events in path propagation is dynamically adjusted based on the weight distribution, and potential fault propagation starting points are located by combining the upstream and downstream directional constraints of the transformer area topology.

[0038] Event nodes whose contribution exceeds a set threshold are identified as candidate causes of power outages, and a set of candidate causes of power outages is output.

[0039] Preferably, the process of dividing the meter nodes and affected branches corresponding to the candidate power outage cause set into fault-sensitive areas is as follows:

[0040] Based on the geographical location and power topology of each meter node in the candidate power outage cause set, determine its direct impact range;

[0041] Similarity calculations and consistency checks are performed on meter events within the affected area, and areas with highly consistent events are designated as sensitive core areas;

[0042] Differential characteristic analysis is performed on meter events near the boundary, and they are marked as sensitive edge areas, thus forming a hierarchical fault-sensitive area.

[0043] Preferably, the process of constructing a hierarchical local source map is as follows:

[0044] Using sensitive core areas and sensitive edge areas as subgraph units, establish event causal links and topological connectivity relationships;

[0045] Consistency rules and anomaly indicators are introduced within the subgraph to correct causal relationships, generating local causal subgraphs with confidence weights;

[0046] Multiple local causal subgraphs are merged according to their topological connections to construct a hierarchical local source graph covering the candidate regions.

[0047] Preferably, the process for outputting the source tracing results is as follows:

[0048] Feature fusion is performed on the hierarchical local source map to extract time-aligned multimodal features, including power quality fluctuation features, communication link stability indicators and meter operating status features;

[0049] Based on the fusion feature call fault classification model, the event type in the candidate area is identified and the power outage cause classification result is output;

[0050] The classification results and the source tracing path are combined to generate a visual source tracing result.

[0051] Compared with existing technologies, the present invention provides a fault classification and tracing method based on electricity meter event logs, which has the following beneficial effects:

[0052] This invention, through high-precision time alignment and cross-meter correlation analysis of electricity meter event logs, achieves second-level synchronization based on millisecond-level data acquisition, significantly improving the temporal consistency and data reliability of event sequences. Combining hierarchical causal reasoning and transformer area topology constraints, it accurately eliminates redundant anomalies in complex distribution networks, quickly converging to possible fault propagation paths. Introducing a time-dependent attention mechanism enables dynamic contribution assessment of different meter events during the source tracing process, thereby improving the precision of power outage cause identification. Through fault-sensitive area division and hierarchical local source tracing map construction, it achieves step-by-step source tracing from point to surface, balancing local accuracy with global integrity. Supported by multimodal feature fusion, it not only achieves high-precision classification of fault types but also outputs intuitive and interpretable source tracing results, improving the automation and real-time response capabilities of fault diagnosis, as well as enhancing the transparency and visualization of results, providing efficient, intelligent, and reliable technical support for power grid operation and maintenance. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1: Please refer to Figure 1 As shown in the embodiment of the present invention, a fault classification and tracing method based on an electricity meter event log includes the following steps:

[0056] S1: Perform millisecond-level dynamic acquisition of power outage, restoration, and abnormal voltage data of user electricity meters and nearby electricity meters, start the local clock deviation adaptive correction algorithm, perform second-level time alignment based on the transformer area reference clock signal, and generate an event priority sequence according to the dual dimensions of fault urgency and data reliability.

[0057] The process of activating the local clock skew adaptive correction algorithm in S1 is as follows:

[0058] The system continuously collects data on the voltage, current, power status, and event changes of user meters and nearby meters. Under the control of the data acquisition terminal or concentrator in the distribution area, the system uses the built-in sampling module of the electricity meter to perform millisecond-level periodic sampling of the effective voltage value, effective current value, active power, reactive power, and event triggering indicators such as power outage, recovery, overvoltage, and undervoltage. The sampling frequency can be dynamically adjusted according to the system configuration to ensure that the collected data covers both normal operation and abnormal fluctuation stages.

[0059] Based on the collected data, local clock drift characteristics are extracted, and adaptive time deviation correction is performed in combination with the substation reference clock signal. By comparing the offset of sampling timestamps of different meters, the drift characteristics of the local crystal oscillator of the meter caused by factors such as ambient temperature and electromagnetic interference are extracted. This offset is compared with the substation reference clock (such as GPS time synchronization, Beidou time synchronization or unified power grid time synchronization signal). The timestamp is dynamically adjusted using an adaptive correction algorithm. The algorithm can use a combination of sliding window weighted average and drift trend prediction to achieve time alignment at the second level or even sub-second level.

[0060] The corrected event data is continuously monitored to identify sudden power outages, power restorations, and abnormal voltage events. In the data stream with unified timestamps, an event-triggered detection model is used to identify features such as voltage drops, power outages, power restorations, and voltage over-limits in real time. This is achieved by combining threshold judgment with rate of change monitoring. For example, when the voltage drop exceeds a set percentage and the duration exceeds a certain number of sampling periods, it is determined to be a power outage event, and the precise time point and duration of the event are recorded.

[0061] Basic preprocessing is performed on the event data, including outlier removal, missing value imputation, and filtering. In response to possible communication packet loss, sampling noise, or hardware errors during the acquisition process, abnormally deviated sampling points are removed, and a small number of missing sampling points are filled by interpolation or repaired by prediction methods based on adjacent time series. At the same time, low-pass filtering, moving average filtering, and other methods are used to suppress high-frequency noise to ensure that the preprocessed data has continuity, smoothness, and high reliability.

[0062] The preprocessed event data is stored according to the meter identifier and time series. The event data that has been clock aligned and cleaned is indexed and stored according to the unique meter identifier (such as meter number or physical address) and the aligned unified timestamp to form a structured time series table. The storage format can adopt a relational database or a distributed time series database to support subsequent cross-table comparison, time series analysis and multi-source feature fusion processing.

[0063] The process of generating an event priority sequence based on both fault urgency and data reliability in S1 is as follows:

[0064] After adaptive time correction, the meter events are classified into power outage, recovery, and voltage anomaly. After time deviation correction is completed, the event data streams generated by different meters are classified into categories. Events reflecting power outages are classified into power outages, events reflecting power recovery are classified into recovery events, and events with voltage exceeding a set threshold are classified into voltage anomalies. The classification can be combined with event identifiers, threshold judgments, and timestamp labels to make a comprehensive judgment, thereby ensuring that the same type of event has a unified definition and judgment standard and avoiding cross-category interference.

[0065] Events are batched according to a set time window, and their urgency is graded and assigned based on event type and frequency. Within a given time window (e.g., 1 second, 5 seconds, or 1 minute), meter events of the same category are batched in chronological order, and the frequency, scope of impact, and event category of events within the batch are weighted. For example, if a large number of meters experience power outages in a short period of time, the batch is assigned the highest level of urgency; if only a few meters experience slight voltage fluctuations, a lower level of urgency is assigned. Through this grading mechanism, a multi-level urgency system from high to low risk can be formed.

[0066] By combining communication quality indicators from event collection with consistency among neighboring meters, data credibility is evaluated using a weighted approach, outputting an event priority sequence with time consistency labels and risk stratification weights. Considering quality factors of the event collection link, such as data packet loss rate, latency jitter rate, and communication signal strength, and comparing the event consistency of neighboring meters within the same time period, if multiple neighboring meters exhibit similar events within the same time window, the event credibility is considered high. If a meter experiences an anomaly independently despite poor communication quality, the credibility needs to be reduced. The weighted evaluation can employ the analytic hierarchy process (AHP) or weighted average method to comprehensively calculate different indicators, thereby obtaining a credibility score for each event batch. The urgency level and credibility score are then integrated to generate a two-dimensional weighted result. Each event sequence is further labeled with a time consistency label (to indicate whether it has undergone unified clock alignment) and a risk stratification weight (to characterize the risk level under different combinations of urgency and credibility). The output event priority sequence can not only serve as input for subsequent cross-time series correlation analysis but also provide an intuitive risk stratification basis for power grid dispatching and emergency response.

[0067] S2: Perform multi-table cross-time series correlation analysis on the event priority sequence, and construct a multi-dimensional feature tensor of power grid events by combining historical fault modes, consistency of events of adjacent meters in the distribution area and communication quality indicators;

[0068] The multi-table cross-time sequence join sub-process for the event priority sequence in S2 is as follows:

[0069] The event priority sequences of each electricity meter are divided according to time segments to establish a time overlap matrix between electricity meters. The event priority sequences from different electricity meters are divided according to preset time segments, such as one minute or five minutes. Each segment contains all events and their corresponding priority information. For each electricity meter, its event segments are compared with the segments of other electricity meters to count the time overlap of each segment and form a time overlap matrix between electricity meters. This matrix uses rows and columns to represent the electricity meter number, and the element values ​​represent the degree of overlap of events of two electricity meters within the same time window, providing basic structural data for subsequent cross-time series matching.

[0070] The cross-meter matching method based on dynamic time warping eliminates misalignment caused by communication delays. To address the issue of event misalignment between meters due to communication delays or clock drift, the dynamic time warping method is used to flexibly adjust the event sequence of one meter along the time axis to match the time points of another meter sequence. DTW dynamically aligns the event priority sequences by calculating the minimum matching cost, ensuring that high-priority events correspond correctly in cross-meter analysis, thereby guaranteeing the synchronization between time series. At the same time, the time offset is recorded for subsequent analysis.

[0071] By combining the physical distance between adjacent meters and the power distribution relationship within the transformer area, the synchronicity and lag between event sequences are compared and analyzed. The physical spatial location of each pair of meters and the topological connection information of the power distribution network are incorporated into the analysis. The synchronicity of aligned event sequences is evaluated. By comparing the occurrence time and time offset of events, the lag time and synchronization consistency index between event sequences are calculated. For adjacent meters or meters within the same transformer area, events with excessively long lag times are marked, and a spatial proximity weight matrix is ​​generated to incorporate physical proximity constraints in cross-meter association, thereby improving the accuracy of the analysis.

[0072] A cross-temporal correlation vector set with time consistency metric, spatial proximity weight, and topology constraint factor is generated. By combining the time overlap matrix, DTW matching results, and spatial topology weight, a cross-temporal correlation vector is generated for each pair of meters. Each vector contains information such as time consistency metric, event priority weight, spatial proximity weight, and transformer area topology constraint factor, forming a unified multi-dimensional vector set. This vector set can be directly used for subsequent construction of multi-dimensional feature tensors of power grid events and analysis of potential fault source evolution trajectories, realizing the preparation of basic data for cross-temporal correlation analysis of multiple meters.

[0073] The process of constructing the multidimensional feature tensor of power grid events in S2 is as follows:

[0074] Based on the cross-time series correlation vector set, the historical fault case library is invoked to compare and extract typical pattern templates that match the features of the current event, and the pattern matching score is output. Each cross-meter event sequence in the cross-time series correlation vector set is traversed and compared with the typical fault pattern templates stored in the historical fault case library. The comparison methods include time series morphological matching, event type consistency, and correspondence analysis of event occurrence order. By calculating the similarity between each current event sequence and the historical template, the pattern matching score is obtained, and the data is indexed and recorded according to meter nodes and time windows to provide reference features for subsequent multidimensional tensor construction.

[0075] Using an event consistency test method, the event consistency rate of neighboring meters within the same window is calculated and embedded as a consistency factor. Within each time window, the event occurrence of the target meter and its spatially neighboring meters is statistically analyzed. By comparing the consistency of the event types and occurrence times of neighboring meters, the event consistency rate is calculated and quantified as a consistency factor. This factor is used to reflect the synchronicity of neighborhood events and is embedded together with other features during tensor construction to increase spatial coupling information.

[0076] Based on the real-time packet loss rate and latency indicators of the communication link, a communication quality correction coefficient is generated. The data link of each meter is monitored in real time to obtain the communication packet loss rate, data latency and fluctuation. According to the collected link status, the communication quality is quantified into a correction coefficient by setting a calculation formula to reflect the impact of communication instability on event records. This correction coefficient is used to weight the credibility of cross-time series data in the feature tensor.

[0077] Event matching scores, consistency factors, and communication correction coefficients are embedded together with cross-time series correlation results into a tensor space to construct a multi-dimensional feature tensor for power grid events. The generated feature vectors are uniformly mapped to the cross-time series correlation vector set, and a tensor structure is formed through multi-dimensional indexing. Each dimension of the tensor contains event matching scores, consistency factors, communication correction coefficients, and cross-time series coupling information to achieve multi-dimensional feature fusion. This tensor can be used as input for subsequent graph neural network or causal analysis models for potential fault source evolution trajectory analysis and power outage cause determination.

[0078] S3: For the multidimensional feature tensor of power grid events, a hierarchical causal reasoning model is used to aggregate local patterns in the tensor to form anomalies, and then combined with the topological constraints of the distribution area to form candidate fault propagation paths.

[0079] The process in S3, which employs a hierarchical causal reasoning model to aggregate anomalies in local patterns within a tensor and then combines this with transformer area topology constraints to form candidate fault propagation paths, is as follows:

[0080] The multidimensional feature tensor of power grid events is sliced ​​hierarchically to extract local temporal anomalies, spatial adjacency anomalies, and cross-layer coupling anomalies. The multidimensional feature tensor of power grid events is partitioned according to the time dimension, spatial adjacency dimension, and cross-layer coupling dimension. In the time dimension, potential abnormal fluctuations in the meter event sequence are extracted, such as short-term large voltage drops and frequent power outages and recovery switching. In the spatial adjacency dimension, inconsistencies in event responses between adjacent meters are extracted, such as most meters on a line being out of power while a few meters remain powered. In the cross-layer coupling dimension, abnormal interactions between different features are extracted, such as voltage fluctuations and communication delays occurring simultaneously and exhibiting a coupling relationship. Through this hierarchical slicing process, different categories of abnormal features hidden in the original high-dimensional data can be clearly separated.

[0081] Based on a causal reasoning rule base, different anomaly types are aggregated for attribution, and redundant factors unrelated to historical patterns are eliminated. Causal mapping relationships between various anomalies are established. The causal reasoning rule base is pre-constructed in conjunction with historical fault patterns. For example, line short circuits are often related to sudden drops in synchronous voltage in the area, and communication anomalies are often related to link blockages at the concentrator level. During the attribution process, the degree of matching between the current anomaly and historical patterns is compared, and irrelevant redundant factors, such as single isolated meter anomalies or occasional low-impact events, are eliminated. This preserves the causal chains that are highly correlated with typical fault patterns, thereby narrowing the scope of analysis and improving the accuracy and efficiency of source tracing.

[0082] By combining the physical topology constraints of the distribution lines in the transformer substation, event links with causal relationships are screened to form candidate fault propagation paths. The meter events are mapped to the line structure of the distribution network in the transformer substation to verify whether the causal propagation paths conform to the direction of power flow and the topological connection relationship. For example, when an upstream meter on a branch line malfunctions first and a downstream meter is subsequently affected, the causal link is considered to conform to the topological constraints. If a downstream meter malfunctions first and an upstream meter malfunctions later, it is determined to be inconsistent with the topological logic and is eliminated. Through this constraint condition, false causal paths can be effectively avoided, ensuring that the final candidate fault propagation paths are reasonable and engineering interpretable.

[0083] S4: Based on the candidate fault propagation path, a time-dependent attention mechanism is introduced to dynamically evaluate the contribution of different meter events to potential faults and generate a set of candidate outage causes based on the topology propagation direction.

[0084] The process in S4 of dynamically evaluating the contribution of different meter events to potential faults and generating a set of candidate outage causes based on the topology propagation direction is as follows:

[0085] A time-dependent attention mechanism is introduced on the candidate fault propagation path to calculate the weight distribution of each meter event in different time segments. Each meter event is input into the attention calculation module in chronological order to evaluate its abnormal amplitude, duration and correlation with context events in different time segments, and generate the corresponding weight distribution. The attention mechanism can adaptively capture long-term dependency and short-term burst features, and ensure the comparability of weight distribution through normalization.

[0086] The contribution of events in path propagation is dynamically adjusted based on the weight distribution, and potential fault propagation starting points are located by combining the upstream and downstream directional constraints of the transformer area topology. The adjustment process not only considers the importance of individual events, but also combines the upstream and downstream directional constraints of the transformer area topology to ensure that the causal influence of upstream events on downstream events is given priority, while the contribution of events that do not conform to the laws of physical propagation is suppressed. A time decay factor is introduced when calculating the contribution, so that the weight of events that are far away from the current time is reduced in the contribution allocation, while events that are consistent with the propagation direction of the topology and occur continuously are enhanced, thus better reflecting the actual characteristics of power grid operation.

[0087] Event nodes whose contribution exceeds a set threshold are identified as candidate causes of power outages, and a set of candidate causes of power outages is output. The threshold can be a preset fixed value or can be adaptively determined based on the event noise level and the operating characteristics of the transformer area. During the identification process, not only are candidate cause nodes output, but also their corresponding affected branches and supporting evidence information are recorded so that they can be displayed in the visualization interface and manually verified later. Through the above processing, the accuracy and practicality of fault tracing can be effectively improved while ensuring the interpretability of the results.

[0088] S5: Divide the meter nodes and affected branches corresponding to the candidate power outage cause set into fault-sensitive areas. Utilize the event similarity, self-consistency rules and differential features of neighboring tables within the area to construct a hierarchical local source map and output the source tracing results.

[0089] The process in S5 of dividing the meter nodes and affected branches corresponding to the candidate power outage cause set into fault-sensitive areas is as follows:

[0090] Based on the geographical location and power topology of each meter node in the candidate power outage cause set, the direct impact range is determined; the route of the line where the candidate meter node is located, the transformer access point and the connection relationship of the adjacent branch are extracted from the topology database of the distribution substation, and combined with the coordinate location in the geographic information system, the mapping relationship from node to branch is constructed. Through this mapping, the power supply coverage of each candidate node and the specific meter group affected downstream can be clearly defined, thus forming the direct impact area of ​​the candidate cause node;

[0091] Similarity calculation and consistency test are performed on meter events within the affected area. Areas with highly similar events are marked as sensitive core areas. After synchronizing the time of meter events within the same affected area, the time interval difference, the duration difference of power outage or restoration, and the amplitude difference of voltage fluctuation are calculated one by one to obtain a multidimensional similarity index between events. Consistency test is performed on event clusters with high similarity to determine whether these events maintain a high degree of consistency in statistical distribution, evolution trend, and spatial distribution. If the consistency result meets the set threshold, the area is marked as a sensitive core area to reflect the range where the fault has the most direct and obvious impact.

[0092] Differential characteristic analysis is performed on boundary-adjacent meter events, identifying them as sensitive edge areas, thus forming a hierarchical fault-sensitive region. Boundary meter nodes refer to meters located at the boundary of the direct influence range of candidate cause nodes and adjacent to other unaffected areas. Differential analysis is performed on the event characteristics of these boundary meters, comparing their abnormal fluctuation amplitude, response delay, and recovery speed with those of core area events to determine whether they are in a partially affected state. Based on the analysis results, areas with significant differences are identified as sensitive edge areas, which, together with the core area, form a hierarchical fault-sensitive region.

[0093] The process of constructing the hierarchical local source graph in S5 is as follows:

[0094] Using sensitive core areas and sensitive edge areas as subgraph units, causal links and topological connectivity relationships of events are established; the meter nodes and affected branches in the obtained fault-sensitive areas are regarded as vertices of the subgraph, and the edges between vertices are determined by using candidate fault propagation paths and distribution topology of the transformer area, forming a preliminary causal link structure. In this way, each subgraph not only reflects the spatial distribution of the core area and edge area, but also reflects the direction of power flow and the logical order of potential fault propagation.

[0095] Consistency rules and anomaly indicators are introduced within the subgraph to correct causal relationships, generating a local causal subgraph with confidence weights. Indicators such as event synchronicity between nodes, differences in anomaly amplitude, duration, and historical failure mode matching degree are used to score and adjust the initially established causal edges. At the same time, consistency rules, such as the order of upstream and downstream events, the consistency of responses of neighboring nodes, and the coupling of cross-layer events, are applied to strengthen or correct the causal chain. Each causal edge is given a confidence weight, forming a local causal subgraph that can quantify the influence relationship between nodes.

[0096] Multiple local causal subgraphs are merged according to topological connections to construct a hierarchical local source graph covering the candidate region. Based on shared nodes, branch connections, and topological continuity between subgraphs, the subgraphs are stitched together into a complete graph structure, while preserving confidence information and hierarchical relationships within the subgraphs, clearly presenting the core area, edge area, and their causal relationships. This hierarchical local source graph not only covers the impact range of candidate power outage causes but also provides multi-level fault causal information.

[0097] The process of outputting the source tracing results in S5 is as follows:

[0098] Feature fusion is performed on the hierarchical local source map to extract time-aligned multimodal features, including power quality fluctuation features, communication link stability indicators, and meter operating status features. The power quality data, communication link status indicators, and meter operating status features corresponding to each meter node and its associated events in the local source map are uniformly aligned and fused to form a comprehensive multimodal feature vector. During the fusion process, time alignment is used to ensure the consistency of data from different sources in the time series, ensuring that subsequent analysis can fully reflect the dynamic evolution of events and the mutual influence between nodes.

[0099] The fault classification model is invoked based on fusion features to identify the type of events in the candidate area and output the classification results of the power outage cause. The fault classification model adopts a deep learning model to determine the type of each node or event chain, such as line short circuit, equipment failure, and instantaneous power outage. The model comprehensively considers time series features, spatial adjacency, communication link quality and historical fault modes during analysis, so as to ensure that the classification results are accurate and reliable and provide directly usable information for power grid operation and maintenance.

[0100] The classification results and the tracing path are combined to generate a visual tracing result. The candidate power outage cause nodes, fault-sensitive areas, causal links of hierarchical local tracing diagrams and classification labels are presented through a visual interface, so that operation and maintenance personnel can intuitively view the fault occurrence nodes, propagation paths and type information. The visualization results support interactive operations, such as zooming, node filtering and path highlighting.

[0101] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fault classification and tracing method based on electric energy meter event log, characterized in that, The method comprises the following steps: Millisecond-level dynamic collection of power-off, recovery, and abnormal voltage data of user meters and adjacent meters, starting of a local clock deviation adaptive correction algorithm, second-level time alignment based on a substation reference clock signal, generation of an event priority sequence in two dimensions of fault emergency and data reliability; Multi-meter cross-time sequence correlation analysis is performed on the event priority sequence, historical fault patterns, consistency of events of adjacent meters in a substation, and communication quality indicators are combined to construct a power grid event multi-dimensional feature tensor; For the power grid event multi-dimensional feature tensor, a hierarchical causal reasoning model is used to perform abnormal aggregation on local patterns in the tensor, and a candidate fault propagation path is formed in combination with the topological constraints of the substation; On the basis of the candidate fault propagation path, a time sequence dependent attention mechanism is introduced to dynamically evaluate the contribution of different meter events to potential faults, and a candidate power-off cause set is generated based on the topological propagation direction; The meter nodes and affected branches corresponding to the candidate power-off cause set are divided into a fault sensitive area, and a hierarchical local traceability graph is constructed using the event similarity within the area, self-consistency rules, and the differentiated features of the boundary adjacent table, and a traceability result is output; The multi-meter cross-time sequence correlation analysis process for the event priority sequence is as follows: The event priority sequences of the meters are divided according to time sequence segments, and a time overlap matrix between the meters is established; Based on a cross-meter matching method of dynamic time warping, misalignment caused by communication delay is eliminated; The synchronicity and hysteresis between the event priority sequences are compared and analyzed in combination with the physical distance between adjacent meters and the distribution relationship of the substation; A cross-time sequence correlation vector set is generated with time consistency measurement, spatial proximity weight, and topological constraint factor; The process of constructing the power grid event multi-dimensional feature tensor is as follows: Based on the cross-time sequence correlation vector set, the historical fault case library is called to compare and extract typical pattern templates matching the current event features, and a pattern matching score is output; The event consistency test method is used to calculate the event consistency rate of adjacent meters in the same window, which is embedded as a consistency factor; A communication quality correction coefficient is generated according to the real-time packet loss rate and delay indicators of the communication link; The pattern matching score, consistency factor, and communication quality correction coefficient are embedded in the tensor space together with the cross-time sequence correlation result to construct the power grid event multi-dimensional feature tensor; The process of using a hierarchical causal reasoning model to perform abnormal aggregation on local patterns in the tensor and forming a candidate fault propagation path in combination with the topological constraints of the substation is as follows: The power grid event multi-dimensional feature tensor is sliced hierarchically to extract local time sequence anomalies, spatial adjacency anomalies, and cross-layer coupling anomalies; Different anomaly types are attributed and aggregated based on a causal reasoning rule library, and redundant factors unrelated to historical patterns are removed; In combination with the physical topological constraints of the substation distribution line, event links with causal relationships are screened to form a candidate fault propagation path.

2. The fault classification and tracing method based on the event log of the electric energy meter according to claim 1, characterized in that, The process of starting the local clock deviation adaptive correction algorithm is as follows: Continuous collection of voltage, current, power state, and event changes of user meters and adjacent meters; Based on the collected data, local clock drift features are extracted, and adaptive time deviation correction is performed in combination with the substation reference clock signal; Continuously monitor the corrected event data to identify power outage, recovery and abnormal voltage events; Perform basic preprocessing on the event data, including outlier removal, missing value filling and filtering; Store the preprocessed event data by meter identifier and time series.

3. The fault classification and tracing method based on the event log of the electric energy meter according to claim 2, characterized in that, The process of generating an event priority sequence in two dimensions of fault urgency and data reliability is as follows: Classify the meter events after adaptive time correction into power outage, recovery and voltage anomaly; Batch aggregate the classified events according to the set time window, and assign the urgency level according to the event type and occurrence frequency; Combine the communication quality indicators and the consistency between adjacent meters to evaluate the data reliability; Output the event priority sequence with time consistency labels and risk stratification weights.

4. The fault classification and tracing method based on the event log of the electric energy meter according to claim 3, characterized in that, The process of dynamically evaluating the contribution of different meter events to potential faults and generating a candidate power outage cause set based on the propagation direction of the topology is as follows: Introduce a time-dependent attention mechanism on the candidate fault propagation path to calculate the weight distribution of each meter event in different time segments; Dynamically adjust the contribution of events in path propagation according to the weight distribution, and locate the potential fault propagation starting point by combining the upstream and downstream direction constraints of the transformer area topology; Mark the event nodes with contribution exceeding the set threshold as candidate power outage causes, and output the candidate power outage cause set.

5. The fault classification and tracing method based on the event log of the electric energy meter according to claim 4, characterized in that, The process of dividing the meter nodes and affected branches corresponding to the candidate power outage cause set into fault-sensitive areas is as follows: Determine the direct impact range of each meter node in the candidate power outage cause set according to its geographical location and power topology location; Perform similarity calculation and consistency test on the meter events within the impact range, and mark the areas with highly consistent events as sensitive core areas; Differential feature analysis is performed on the boundary adjacent meter events, which are marked as sensitive edge areas, thus forming a hierarchical fault-sensitive area.

6. The fault classification and tracing method based on the event log of the electric energy meter according to claim 5, characterized in that, The process of constructing a hierarchical local traceability graph is as follows: Establish the event causal link and topological connectivity relationship with the sensitive core area and sensitive edge area as subgraph units; Introduce consistency rules and abnormal indicators to correct the causal relationship within the subgraph, generating a local causal subgraph with confidence weight; Merge multiple local causal subgraphs according to the topological connection relationship to construct a hierarchical local traceability graph covering the candidate area.

7. The fault classification and tracing method based on the event log of an electric energy meter according to claim 6, characterized in that, The process of outputting the traceability result is as follows: Perform feature fusion on the hierarchical local traceability graph, extract time-aligned multi-modal features, including power quality fluctuation features, communication link stability indicators and meter operating state features; Based on the fusion features, call the fault classification model to identify the type of events in the candidate area and output the power outage cause classification result; Generate a visual traceability result combining the classification result and the traceability path.

Images