A power metering device fault prediction method, device, equipment and medium

Abstract

This application discloses a method, apparatus, device, and medium for predicting faults in power metering equipment, relating to the field of power metering automation technology. The method includes: spatiotemporal alignment and feature fusion of multi-source heterogeneous metering data from power metering equipment to generate a multimodal spatiotemporal feature tensor; constructing anomaly evidence items based on the multimodal spatiotemporal feature tensor; selecting anomaly detection models corresponding to the anomaly evidence items and dynamically orchestrating the anomaly detection models to generate anomaly detection model execution chains; inputting metering data into the anomaly detection model execution chains to extract anomaly features; organizing multiple anomaly feature extraction results into an anomaly evidence chain, and inputting the anomaly evidence chain and the anomaly feature extraction results into a fault prediction model to predict faults in the metering equipment based on a knowledge graph of the metering fault domain. This application can improve the accuracy of anomaly identification in metering equipment and the penetration of root cause analysis, reducing missed detections and false detections.

Description

Technical Field

[0001] This application relates to the field of power metering automation technology, and in particular to a method, device, equipment and medium for predicting faults in power metering equipment. Background Technology

[0002] Smart grids are the core infrastructure for ensuring power supply and energy distribution in modern society. As data sensing endpoints in all aspects of power generation, transmission, distribution and consumption, the safe and reliable operation of electricity metering devices / equipment is a key link in maintaining fair electricity trading and stable grid operation.

[0003] However, with the deepening of smart grid construction, the deployment scale of electricity metering devices / equipment has experienced explosive growth. Operating under complex long-term field conditions, they are highly susceptible to various factors such as natural aging of components, harsh environments (e.g., thunderstorms, high temperature and humidity, strong magnetic interference), and human interference, leading to operational anomalies such as voltage loss, current loss, and electricity theft. These anomalies not only cause serious grid energy loss and economic losses but are also important precursors to complete damage to metering equipment or localized transformer area failures. If not detected and addressed in a timely manner, they may trigger wider-ranging power supply security problems. Currently, the monitoring and maintenance of abnormal operation of electricity metering equipment mainly involves the following technical solutions: The first type is the traditional method of manually inspecting and maintaining electricity metering equipment on-site. This method is not only labor-intensive, has a delayed discovery process, and a high rate of invalid work orders, but it is also prone to missing or misjudging abnormal features when faced with massive amounts of time-series electrical data containing high levels of noise.

[0004] The second type is automated alarm systems based on fixed thresholds. These systems preset upper and lower thresholds for electrical parameters such as voltage, current, and power factor, triggering an alarm when real-time collected data exceeds these thresholds. For example, by comparing the acquired parameters with preset thresholds and historical databases, a comparative method is used to determine the fault type of the smart meter. This method improves the reliability of fault diagnosis to some extent and avoids biased judgment. However, this method mainly relies on a simple comparison of fixed threshold exceedances and preset historical data. Its output only reaches the superficial alarm level (such as identifying electrical states like voltage or current loss), completely lacking penetrating analysis of the deeper causes of the fault. Furthermore, after receiving such alarms, on-site maintenance personnel still cannot directly distinguish whether the anomaly stems from aging components in the equipment itself, external environmental factors, or human theft of electricity, resulting in a high rate of invalid dispatches for subsequent on-site investigations and failing to achieve preventative measures and accurate root cause tracing.

[0005] The third category is anomaly detection methods based on machine learning or deep learning. This method comprehensively considers the records, operating status, and grid topology location of the electricity metering device / equipment, and uses graph neural networks and attention mechanisms to model the assessment as a sequence prediction problem. While this type of method improves the accuracy of fault risk assessment, these purely data-driven deep learning models are black-box algorithms, lacking the rigid constraints of electrical engineering physics rules (such as energy conservation and Kirchhoff's laws). When faced with complex and fluctuating field conditions, the model is prone to unexplained misjudgments, even drawing conclusions that violate the objective physical principles of power systems. Furthermore, the model itself lacks error correction and self-healing mechanisms, resulting in insufficient reliability and credibility of anomaly detection, failing to meet the high reliability and safety requirements of power systems. Summary of the Invention

[0006] In view of this, the purpose of this application is to provide a method, device, equipment, and medium for predicting faults in electricity metering equipment. This method enables automated equipment fault prediction and a shift from post-event handling to pre-event prevention. It also improves the accuracy of anomaly identification and the penetration of root cause analysis in electricity metering equipment, reducing missed and false positives caused by single data sources or fixed threshold judgments. Furthermore, by dynamically selecting anomaly detection models adapted to the current operating conditions, it avoids wasting computing resources. The specific solution is as follows: In a first aspect, this application discloses a method for predicting faults in electricity metering equipment, including: Collect current multi-source heterogeneous metering data of power metering equipment in the target distribution substation area, and perform spatiotemporal alignment and feature fusion on the current multi-source heterogeneous metering data to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates. Anomaly evidence terms are constructed based on the multimodal spatiotemporal feature tensor; the anomaly evidence terms are used to describe the support and / or disproving relationships between the equipment anomaly and the candidate root causes; The abnormal evidence items are input into a pre-created model orchestrator to select multiple target anomaly detection models of corresponding anomaly types from a preset anomaly detection model pool, and to dynamically orchestrate the multiple target anomaly detection models to generate an anomaly detection model execution chain; the anomaly detection models in the anomaly detection model pool are models obtained by training a preset neural network using historical econometric data corresponding to different anomaly types. The current multi-source heterogeneous measurement data is input into the execution chain of the anomaly detection model to extract anomaly features from the current multi-source heterogeneous measurement data, and anomaly feature extraction results are obtained; the anomaly feature extraction results include anomaly feature labels, candidate root causes and corresponding root cause confidence scores; The multiple abnormal feature extraction results output by different target anomaly detection models are organized into an anomaly evidence chain, and the anomaly evidence chain and the abnormal feature extraction results are input into the trained fault prediction model to perform fault prediction on the power metering equipment based on the metering fault domain knowledge graph, and obtain the fault prediction result; the fault prediction model is a model obtained by training a pre-trained language model using historical multi-source heterogeneous metering data.

[0007] Optionally, constructing anomaly evidence terms based on the multimodal spatiotemporal feature tensor includes: Extract the abnormal evidence information corresponding to each abnormal phenomenon in the multimodal spatiotemporal feature tensor; the abnormal evidence information includes abnormal evidence type, abnormal evidence source, associated metering equipment, abnormal occurrence time, abnormal topological location, abnormal feature value, abnormal evidence credibility, and the set of supporting candidate root causes and the set of counter-evidence candidate root causes corresponding to the abnormal phenomenon; Different pieces of evidence information in the abnormal evidence information are combined to obtain an abnormal evidence item that includes the abnormal evidence information.

[0008] Optionally, selecting multiple target anomaly detection models of corresponding anomaly types from a preset anomaly detection model pool includes: Obtain the metering business context information related to the power metering equipment, and calculate the confidence level of each candidate root cause in the abnormal evidence item based on the credibility of the abnormal evidence item to obtain the prior confidence level. Based on the abnormal evidence items, the metering business context information, and the prior confidence level, an anomaly detection model of the corresponding anomaly type is selected from the preset anomaly detection model pool to obtain multiple target anomaly detection models.

[0009] Optionally, organizing the multiple anomaly feature extraction results output by different target anomaly detection models into an anomaly evidence chain includes: Using the abnormal event as the central node, the abnormal evidence type corresponding to the abnormal feature extraction result is taken as the evidence node, and directed association edges are established between different evidence nodes according to the preset evidence association relationship. A mapping relationship is also established between different evidence nodes and the abnormal feature extraction result of the corresponding abnormal type to obtain the abnormal evidence chain.

[0010] Optionally, the fault prediction method for the power metering equipment further includes: Named entity recognition technology is used to analyze knowledge including power industry standard equipment management rules, metering equipment expert troubleshooting manuals and historical metering equipment failure cases, in order to construct a triplet structured graph with fault phenomena, fault physical causes and fault troubleshooting strategies as the core, and obtain a knowledge graph in the field of metering faults.

[0011] Optionally, the step of performing fault prediction on the power metering equipment based on a knowledge graph of metering fault domains to obtain fault prediction results includes: The abnormal evidence chain and the abnormal feature extraction results are converted into vector representations to obtain the abnormal query vector; Based on the anomaly query vector, semantic similarity retrieval is performed on the knowledge graph of metering fault domain to obtain anomaly retrieval results. The anomaly retrieval results are then used as prompt words to analyze the anomaly feature extraction results, thereby obtaining the fault prediction results of the power metering equipment. The fault prediction results include the predicted root cause of the fault and the path of the corresponding anomaly evidence chain, the root cause confidence, the fault risk level, and early warning and handling suggestions.

[0012] Optionally, the current multi-source heterogeneous metering data includes electrical operation data, abnormal event alarm data, transformer area topology data, equipment ledger data, environmental meteorological data, and on-site operation data; Accordingly, the step of performing spatiotemporal alignment and feature fusion on the current multi-source heterogeneous metrological data to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates includes: Using the highest frequency data sampling period as a benchmark, the current multi-source heterogeneous metering data is resampled in the frequency domain to obtain the sampled multi-source metering data; The sampled multi-source measurement data is filtered and smoothed to generate a processed multi-source measurement data stream. The processed multi-source measurement data stream is fused on both the temporal and spatial scales to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates.

[0013] Secondly, this application discloses a fault prediction device for electricity metering equipment, comprising: The data acquisition module is used to collect current multi-source heterogeneous metering data of the power metering equipment in the target distribution area; The data processing module is used to perform spatiotemporal alignment and feature fusion on the current multi-source heterogeneous measurement data to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates. The construction module is used to construct anomaly evidence items based on the multimodal spatiotemporal feature tensor; the anomaly evidence items are used to describe the support relationship and / or disproving relationship between the device anomaly and the candidate root cause; The first input module is used to input the abnormal evidence items into a pre-created model orchestrator to select multiple target abnormality detection models of corresponding abnormality types from a preset abnormality detection model pool, and to dynamically orchestrate the multiple target abnormality detection models to generate an abnormality detection model execution chain; the abnormality detection models in the abnormality detection model pool are models obtained by training a preset neural network using historical econometric data corresponding to different abnormality types. The second input module is used to input the current multi-source heterogeneous measurement data into the anomaly detection model execution chain to extract anomaly features from the current multi-source heterogeneous measurement data and obtain anomaly feature extraction results; the anomaly feature extraction results include anomaly feature labels, candidate root causes and corresponding root cause confidence levels; The organization module is used to organize multiple abnormal feature extraction results output by different target anomaly detection models into an anomaly evidence chain; The third input module is used to input the abnormal evidence chain and the abnormal feature extraction results into the trained fault prediction model, so as to perform fault prediction on the power metering equipment according to the knowledge graph of the metering fault domain and obtain the fault prediction result; the fault prediction model is a model obtained by training a pre-trained language model using historical multi-source heterogeneous metering data.

[0014] Thirdly, this application discloses an electronic device, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the aforementioned method for predicting faults in an energy metering device.

[0015] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned method for predicting faults in power metering equipment.

[0016] As can be seen, this application first collects current multi-source heterogeneous metering data of electricity metering equipment in the target distribution substation, and performs spatiotemporal alignment and feature fusion on the current multi-source heterogeneous metering data to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates. Then, based on the multimodal spatiotemporal feature tensor, anomaly evidence items (used to describe the support relationship and / or disproving relationship between equipment anomalies and candidate root causes) are constructed. Next, the anomaly evidence items are input into a pre-created model orchestrator to select multiple target anomaly detection models of corresponding anomaly types from a preset anomaly detection model pool, and dynamically orchestrate the multiple target anomaly detection models to generate an anomaly detection model execution chain. The anomaly detection models in the anomaly detection model pool utilize historical metering data corresponding to different anomaly types. The model is obtained by training a preset neural network with the current multi-source heterogeneous metering data; then, the current multi-source heterogeneous metering data is input into the execution chain of the anomaly detection model to extract anomaly features from the current multi-source heterogeneous metering data, and obtain anomaly feature extraction results (including anomaly feature labels, candidate root causes and corresponding root cause confidence scores); finally, the multiple anomaly feature extraction results output by different target anomaly detection models are organized into anomaly evidence chains, and the anomaly evidence chains and the anomaly feature extraction results are input into the trained fault prediction model to perform fault prediction on the power metering equipment based on the metering fault domain knowledge graph to obtain fault prediction results; the fault prediction model is a model obtained by training a pre-trained language model using historical multi-source heterogeneous metering data.

[0017] This application first performs spatiotemporal alignment and feature fusion on multi-source heterogeneous metering data from electricity metering equipment, and constructs anomaly evidence items based on the multimodal spatiotemporal feature tensors obtained after fusion. Then, a model orchestrator selects multiple anomaly detection models corresponding to the anomaly evidence items from the anomaly detection model pool, and dynamically orchestrates these multiple anomaly detection models into a model execution chain. The model execution chain then extracts anomaly features from the multi-source heterogeneous metering data, and organizes the anomaly feature extraction results output by different anomaly detection models into an anomaly evidence chain. Finally, a pre-trained language model combined with a knowledge graph of the metering fault domain is used to analyze the anomaly evidence chain and anomaly feature extraction results, thereby obtaining fault prediction results. Therefore, this application is not based on a fixed... Instead of relying on fixed thresholds or single models for fault prediction, this approach dynamically generates corresponding model execution chains based on the characteristics of multi-source heterogeneous metering data collected in real time. This allows for feature extraction from the multi-source heterogeneous metering data, and the construction of anomaly evidence chains based on the feature extraction results. Finally, a pre-trained language model (combined with a knowledge graph) is used for comprehensive fault prediction. Through this method, automated equipment fault prediction is achieved, shifting from post-event handling to pre-event prevention. This also improves the accuracy of anomaly identification in power metering equipment and the penetration of root cause analysis, reducing missed and false judgments caused by single data sources or fixed threshold judgments. Furthermore, by dynamically selecting anomaly detection models that are adapted to the current working conditions, the waste of computing resources can be avoided. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 This is a flowchart of a fault prediction method for power metering equipment disclosed in this application; Figure 2 This is a schematic diagram of the structure of a fault prediction device for electricity metering equipment disclosed in this application; Figure 3 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] This application discloses a method for predicting faults in electricity metering equipment. (See also...) Figure 1 As shown, the method includes: Step S11: Collect current multi-source heterogeneous metering data of the power metering equipment in the target distribution area, and perform spatiotemporal alignment and feature fusion on the current multi-source heterogeneous metering data to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates.

[0022] It should be noted that the fault prediction method for electricity metering equipment proposed in this application can be specifically applied to a fault prediction system for metering equipment, which can be deployed on computer equipment, servers, and other equipment.

[0023] In this embodiment, firstly, full metering data of the power metering equipment in the target distribution transformer area is collected from multiple data sources to obtain the current multi-source heterogeneous metering data. The current multi-source heterogeneous metering data includes, but is not limited to, electrical operation data, abnormal event alarm data, transformer area topology data, equipment ledger data, environmental meteorological data (such as temperature, humidity, thunderstorms, etc.) and on-site operation data (such as construction work order information).

[0024] Electrical operation data refers to real-time dynamic operation data of electricity metering equipment, including but not limited to high-frequency time-series sequences such as three-phase voltage, current, active power / reactive power, power factor, and harmonic distortion rate. Distribution area topology data may include equipment ledger attributes extracted based on the Common Information Model (CIM), the power grid physical topology hierarchy based on the transformer-line-transformer-customer structure, and the geographic spatial coordinates of the equipment, such as GIS (Geographic Information System) spatial coordinates. Furthermore, the current multi-source heterogeneous metering data undergoes spatiotemporal alignment and feature fusion to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates.

[0025] Specifically, the step of performing spatiotemporal alignment and feature fusion on the current multi-source heterogeneous metrology data to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates may include: resampling the current multi-source heterogeneous metrology data in the frequency domain based on the highest frequency data sampling period to obtain sampled multi-source metrology data; filtering and smoothing the sampled multi-source metrology data to generate a processed multi-source metrology data stream; and performing feature fusion on the processed multi-source metrology data stream at both the temporal and spatial scales to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates. In this embodiment, considering that the sampling periods of the multi-source heterogeneous metrology data may be inconsistent and there may be noise, a batch-streaming hybrid mechanism can be used for data cleaning and alignment. For example, on the one hand, frequency domain resampling is performed, using the sampling period of the highest frequency electrical operation data as the reference fundamental frequency, and forward fill or linear interpolation algorithm is used to resample and align the low-frequency spatiotemporal environmental data (such as geospatial coordinates); on the other hand, filtering and smoothing are performed, and Hampel filters (a type of filter based on the median and absolute deviation of the median) can be applied to identify distribution characteristics and smooth replacement for missing values ​​(NaN) and abnormal abrupt changes generated during data transmission, thereby generating a standard continuous data stream without discontinuities (i.e., the processed multi-source metering data stream).

[0026] Furthermore, feature fusion is performed on the smoothed multi-source metering data stream at both the temporal and spatial scales to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates. For example, spatiotemporal dual mapping eliminates the isolation of data in each dimension. For temporal fusion, a dynamic sliding window technique can be used to extract time window slices along the historical direction based on the timestamp of the alarm event, and to extract the temporal precursor features before the anomaly occurs. For spatial fusion, the aforementioned temporal precursor features are jointly mapped with the distribution network physical topology data (belonging to transformer area topology data) and GIS spatial coordinates, thereby attaching spatial context labels to each group of electrical operation data. After the above spatiotemporal joint processing, the original multi-source heterogeneous metering data is fused and transformed into a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates. This feature tensor can serve as a standardized underlying data structure, directly acting as the input source for subsequent differentiated anomaly detection models.

[0027] Step S12: Construct anomaly evidence terms based on the multimodal spatiotemporal feature tensor; the anomaly evidence terms are used to describe the support relationship and / or disproven relationship between the device anomaly and the candidate root cause.

[0028] In this embodiment, after generating a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates, a corresponding anomalous evidence item can be constructed for each anomalous phenomenon based on the multimodal spatiotemporal feature tensor. That is, the underlying multimodal spatiotemporal feature tensor is transformed into a standardized anomalous evidence item, thereby solving the problems of difficulty in uniformly expressing multi-source data, difficulty in tracing the source of evidence, and lack of structured constraints in model input during the traditional anomaly judgment process.

[0029] Among them, the abnormal evidence item is used to describe the support relationship and / or disproven relationship between the abnormal phenomenon of the equipment and the candidate root cause.

[0030] In this embodiment, constructing anomaly evidence items based on the multimodal spatiotemporal feature tensor specifically includes: extracting anomaly evidence information corresponding to each anomaly phenomenon in the multimodal spatiotemporal feature tensor; and combining different evidence information in the anomaly evidence information to obtain anomaly evidence items containing the anomaly evidence information. In this embodiment, firstly, anomaly evidence information corresponding to each anomaly phenomenon in the multimodal spatiotemporal feature tensor is extracted, including but not limited to anomaly evidence type, anomaly evidence source, associated metering equipment, anomaly occurrence time, anomaly topological location, anomaly feature value, anomaly evidence credibility, and the supporting candidate root cause set and the counter-evidence candidate root cause set corresponding to the anomaly phenomenon. Then, different evidence information in the anomaly evidence information is combined to obtain anomaly evidence items containing the anomaly evidence information. Specifically, the anomaly evidence item corresponding to each anomaly problem can be represented as: ; In the formula, Evidence number; Indicates the type of abnormal evidence; Indicates the source of unusual evidence; Indicates the associated metering equipment; Indicates the time when the anomaly occurred; Indicates the abnormal topology location (i.e., the current location of the metering device in the low-voltage area or power supply topology); Indicates abnormal characteristic values; Indicates the credibility of unusual evidence; This represents the set of supporting candidate root causes corresponding to the abnormal phenomenon. This represents the set of candidate root causes for counter-evidence corresponding to the abnormal phenomenon.

[0031] The abnormal evidence types include at least electrical timing evidence, event alarm evidence, topological clustering evidence, energy residual evidence, environmental correlation evidence, equipment batch evidence, and work record evidence.

[0032] For example, when multiple energy metering devices in the same distribution area experience communication interruptions or voltage anomalies simultaneously within a similar time window, and there are thunderstorm records in the environmental meteorological data, a spatiotemporal cluster evidence item supporting the external environmental root cause can be generated; when a user's current curve shows an abnormal sudden drop while the voltage remains stable, and the residual line loss in the distribution area increases synchronously, an energy anomaly evidence item supporting the human factor root cause can be generated; when phase anomalies, rate anomalies, or power factor abrupt changes occur shortly after the completion of meter replacement, parameter issuance, or transformer replacement work, an operation-related evidence item supporting the operation quality root cause can be generated.

[0033] By constructing the aforementioned standardized anomaly evidence items, the original multi-source heterogeneous measurement data can be transformed into model input data with a unified data structure, clear evidence sources, traceable supporting relationships, and rebuttal relationships, thereby providing a data foundation for the dynamic orchestration of subsequent anomaly detection models and the controlled inference of pre-trained language models.

[0034] Step S13: Input the abnormal evidence items into the pre-created model orchestrator to select multiple target abnormality detection models of the corresponding abnormality type from the preset abnormality detection model pool, and dynamically orchestrate the multiple target abnormality detection models to generate an abnormality detection model execution chain; the abnormality detection models in the abnormality detection model pool are models obtained by training a preset neural network using historical econometric data corresponding to different abnormality types.

[0035] In this embodiment, the above-mentioned abnormal evidence items can be input into a pre-created model orchestrator so that the model orchestrator can select an abnormal detection model corresponding to the current abnormal type from the preset abnormal detection model pool to obtain multiple target abnormal detection models. Then, the multiple target abnormal detection models are dynamically orchestrated to generate an abnormal detection model execution chain.

[0036] Among them, the anomaly detection models in the anomaly detection model pool are models obtained by training a preset neural network (NN) using historical econometric data corresponding to different anomaly types.

[0037] In this embodiment, selecting multiple target anomaly detection models of corresponding anomaly types from a preset anomaly detection model pool can specifically include: acquiring metering business context information related to the electricity metering equipment, and calculating the confidence level of each candidate root cause in the anomaly evidence item based on the credibility of the anomaly evidence item to obtain a prior confidence level; selecting anomaly detection models of corresponding anomaly types from the preset anomaly detection model pool based on the anomaly evidence item, the metering business context information, and the prior confidence level to obtain multiple target anomaly detection models. In this embodiment, to address the problems of high computational consumption, poor real-time performance, and feature overload when traditional general-purpose large models directly process massive amounts of raw electrical operation data, a set of lightweight, decoupled, and differentiated anomaly detection models (stored in a model pool) is pre-constructed as a pre-feature extraction layer. The anomaly detection model is a model trained and deployed for a specific metering task. To avoid wasting computing resources by constantly calling all models and to avoid the problem that a single model cannot adapt to complex field conditions, this application further sets up a dynamic model orchestrator for different anomaly detection models. This orchestrator can automatically select and combine corresponding differentiated anomaly detection models based on a standardized set of anomaly evidence items, the prior confidence of candidate root causes, and the measurement business context information, thereby forming a model execution chain for the current anomaly event.

[0038] Specifically, the anomaly detection model pool can be represented as: ; In the formula, to These represent different types of anomaly detection models, based on the set of anomaly evidence items E and the prior confidence of candidate root causes. In addition to the metering business context information C, the execution chain of the anomaly detection model corresponding to the current anomaly event can be generated. : ; Router stands for Dynamic Model Orchestrator; This represents the execution chain formed by the selected anomaly detection model.

[0039] The formula for calculating the prior confidence level is as follows: ; in, Indicates the k-th type of candidate root cause. Indicating abnormal evidence items For candidate root causes Support weights, Indicating abnormal evidence items Credibility.

[0040] It should be noted that, based on different abnormal evidence items, an anomaly detection model corresponding to the anomaly type can be selected from the model pool. For example, when the abnormal evidence item manifests as a sudden change in voltage, current, phase angle, power factor, or harmonic distortion rate, the electrical characteristic anomaly detection model can be prioritized; when the abnormal evidence item manifests as multiple alarms occurring simultaneously in the same transformer area, the same line, or adjacent geographical areas, the event spatiotemporal clustering model based on topology association can be prioritized; when the abnormal evidence item manifests as an imbalance in the total and sub-meter power consumption in the transformer area, abnormal line loss or residual power, or abnormal energy flow, the energy analysis model based on physical topology can be prioritized; when the abnormal evidence item manifests as a concentrated occurrence of the same type of anomaly in equipment from the same manufacturer, model, batch, or software version, the equipment batch quality association model can be prioritized; and when the abnormal evidence item is temporally related to meter installation / replacement, transformer replacement, parameter issuance, seal changes, or on-site commissioning operations, the operation quality association model can be prioritized.

[0041] The electrical anomaly detection model can employ an Isolation Forest anomaly detection algorithm or a Long Short-Term Memory (LSTM) autoencoder to perform real-time streaming analysis on high-frequency sampled time-series data such as voltage, current, and phase angle uploaded by metering devices. This model does not rely on fixed thresholds but learns the distribution patterns of normal electrical operation data to automatically capture minute waveform distortions, numerical abrupt changes, or trend anomalies, and outputs anomaly confidence levels. For example, when using an LSTM autoencoder, if the input vector corresponding to the multidimensional electrical operation data of the i-th metering device at time t is... (Including voltage, current, phase, etc.), then the output vector after encoder-decoder reconstruction is: Specifically, reconstruction error can be used as an anomaly confidence level. The calculation formula can be expressed as: .

[0042] Specifically, the topology-based event spatiotemporal clustering model integrates power grid GIS geographic information (such as GIS spatial coordinates) and low-voltage transformer area topology data (such as distribution network physical topology data). It employs an improved DBSCAN algorithm (Density-Based Spatial Clustering of Applications with Noise) to perform multidimensional spatiotemporal clustering of massive discrete alarm events. Furthermore, this model can automatically identify batch events with temporal and spatial correlations (e.g., multiple meters in the same transformer area simultaneously triggering alarms at similar times), thus classifying them as regional common faults (such as lightning strikes, construction damage, or transformer area power outages), effectively eliminating false alarms caused by occasional single-meter alarms. To integrate spatiotemporal multidimensional features, the model constructs a joint spatiotemporal distance metric. It is used to measure two abnormal alarm events, i.e., two different abnormal evidence items. and The degree of correlation between them: ; In the formula, , These are the GIS coordinate vectors of the two abnormal alarm event occurrence points; , These are the timestamps of the abnormal alarm events; It is a geospatial distance function; and These are the normalized weights for spatial distance and time interval, respectively. If If the clustering neighborhood radius is used, it is determined that the two events have direct density and belong to the same spatiotemporal fault cluster.

[0043] Based on the aforementioned joint spatiotemporal distance metric, the model executes an improved spatiotemporal clustering algorithm. The specific clustering steps are as follows: First, for any received alarm event... If its spatiotemporal neighborhood radius Internal (i.e., satisfying) Adjacent alarm events If the total number of events is greater than or equal to the preset density threshold MinPts, then the event is... The events are marked as spatiotemporal core events. Next, starting from any unvisited spatiotemporal core event, a connected graph traversal algorithm is used to find all sets of events that are density-reachable from it, and these interconnected events are merged into the same spatiotemporal fault cluster. Isolated events that do not belong to any spatiotemporal fault cluster are marked as occasional noise. Finally, clustering feature vectorization is performed, transforming the massive discrete alarm events into a set of clustered fault events. For each event, a structured clustering label (e.g., localized lightning strike power outage cluster, isolated occasional events, etc.) and cluster size feature are output. This clustering label will be directly input into the subsequent pre-trained language model as a strong feature of the spatial dimension, serving as the core basis for the model to determine whether there are regional common factors interfering.

[0044] To address the issue of traditional black-box models neglecting physical constraints, this application employs a physical topology-based energy analysis model. This model is constructed based on the physical topology of the low-voltage distribution area and Kirchhoff's laws. It collects active power and current data from the main meter and individual meters of the low-voltage distribution area in real time, and dynamically calculates the deviation between the theoretical and actual line losses by combining line impedance parameters. Unlike simple threshold judgments, this model calculates the line loss residual using a sliding window algorithm. When the residual exceeds the dynamic confidence interval, it outputs a feature label indicating an energy anomaly and the corresponding moment of line loss abrupt change. This output information serves as a strong physical feature, directly inputting into a subsequent pre-trained language model to assist in determining whether energy loss behaviors such as current theft or line leakage exist.

[0045] To address the difficulty of identifying batch defects in traditional metering equipment anomaly prediction, this application constructs an equipment batch quality correlation model. This model identifies the clustering trend of similar anomalies occurring in different distribution areas, users, and operating environments for metering equipment of the same manufacturer, model, batch, and software version. Specifically, the model takes equipment manufacturer, model, batch number, commissioning time, calibration records, software version, historical fault records, operating environment, and event alarm type as inputs to calculate the anomaly concentration, anomaly type consistency, version consistency, and environmental exclusion degree of the target batch of equipment. The model can output an equipment batch quality risk score for the current electricity metering equipment, calculated using the following formula: ; In the formula, This indicates the batch quality risk score of the equipment. This indicates the frequency of anomalies occurring in the target batch of equipment. Indicates consistency of exception types. This indicates software version or hardware batch consistency. Indicates external environmental interference factors. Indicates the interference factor in on-site operations. , , , , These are the corresponding weights.

[0046] It should be noted that when the frequency of anomalies in the same batch of equipment is significantly higher than the baseline anomaly frequency of similar equipment, and the anomaly types are consistent, while excluding the influence of regional environmental factors and operational factors, the model will output feature labels of equipment quality anomalies, batch risk scores, and a list of associated equipment. This output will serve as important evidence for subsequent pre-trained language models to determine the root causes of equipment quality failures.

[0047] In addition, to address the potential metering anomalies caused by on-site meter installation / replacement, transformer replacement, parameter issuance, seal changes, and data acquisition debugging, this application pre-constructs an operation quality correlation model. This model is used to identify the temporal, typological, and causal correlations between abnormal events and on-site operations. Specifically, the model can be based on on-site operation data (such as on-site work order data), meter installation / replacement records, transformer replacement records, parameter issuance records, seal change records, data acquisition debugging records, and operation completion time. It constructs a pre-operation time window and a post-operation time window centered on the operation completion time. Then, based on these two time windows, it compares and analyzes the voltage, current, phase angle, power factor, power curve, event alarms, and data acquisition success rate before and after the operation to obtain an operation quality correlation score. The specific calculation formula is as follows: ; In the formula, W represents the job quality correlation score. This indicates the proximity of the time of the abnormal event to the time of task completion. This indicates the degree of matching between the exception type and the job type. The value of H represents the magnitude of change in electrical characteristics before and after the operation, and H represents the historical frequency of anomalies of the metering equipment. , , , These are the corresponding weights.

[0048] It should be noted that when an abnormal event occurs within a preset time window after the completion of on-site work, and there is a preset mapping relationship between the abnormality type and the work content, the model will increase the confidence of the root cause of the work quality abnormality. For example, if a sudden change in phase angle, abnormal power factor, abnormal rate, or a cliff-like change in the power curve occurs within a short period of time after meter replacement (within a preset time), the model will output the feature label of the work quality abnormality and provide candidate abnormality causes (i.e., candidate root causes) such as suspected wiring errors, reverse polarity of the current transformer, incorrect rate parameter configuration, or incorrect distribution of acquisition parameters.

[0049] Step S14: Input the current multi-source heterogeneous measurement data into the anomaly detection model execution chain to extract anomaly features from the current multi-source heterogeneous measurement data and obtain anomaly feature extraction results; the anomaly feature extraction results include anomaly feature labels, candidate root causes and corresponding root cause confidence levels.

[0050] In this embodiment, through the aforementioned dynamic orchestration mechanism, different model combinations can be adaptively selected according to different anomaly scenarios, thereby realizing the transformation from fixed model judgment to scenario-based, evidence-based, and dynamic judgment. Next, using the prior confidence of abnormal evidence items, candidate root causes, and metrological business context information as input, the dynamic orchestrator selects the corresponding anomaly detection model and performs dynamic orchestration. Furthermore, the current multi-source heterogeneous metrological data is input into the anomaly detection model execution chain to perform anomaly feature extraction and preliminary anomaly detection using the orchestrated model chain, thereby outputting anomaly feature extraction results including anomaly feature labels, candidate root causes and corresponding anomaly confidence, and counter-evidence relationships.

[0051] Step S15: Organize the multiple abnormal feature extraction results output by different target anomaly detection models into an abnormal evidence chain, and input the abnormal evidence chain and the abnormal feature extraction results into the trained fault prediction model to perform fault prediction on the power metering equipment based on the metering fault domain knowledge graph to obtain the fault prediction result; the fault prediction model is a model obtained by training a pre-trained language model using historical multi-source heterogeneous metering data.

[0052] In this embodiment, for the various anomaly feature extraction results output by different anomaly detection models in the anomaly detection model execution chain, the multimodal features contained within them can be vectorized and output first. For example, the numerical anomaly indicators, clustering labels, and preliminary judgment results output by the models can be transformed into a unified high-dimensional structured feature vector through normalization processing and embedding technology. This feature vector serves as the standard input for subsequent pre-trained language models (i.e., Large Language Models), providing accurate feature support for the model's deep reasoning. The high-dimensional structured feature vector can include not only the anomaly feature labels and anomaly confidence scores output by each anomaly detection model, but also information such as candidate root causes, evidence sources, refutation relationships (e.g., a set of candidate root causes for refutation), topological location, and time windows, thus providing a unified input for the subsequent construction of the anomaly evidence chain.

[0053] Next, the abnormal feature extraction results, including abnormal feature labels, abnormal confidence, candidate root causes, abnormal evidence sources, abnormal timestamps (i.e., the time of abnormal occurrence), abnormal topological locations, and counter-evidence relationships, are organized into an abnormal evidence chain.

[0054] In one specific implementation, organizing the multiple anomaly feature extraction results output by different target anomaly detection models into an anomaly evidence chain may specifically include: taking an anomaly event as the central node, using the anomaly evidence type corresponding to the anomaly feature extraction result as evidence nodes, establishing directed association edges between different evidence nodes according to a preset evidence association relationship, and establishing a mapping relationship between different evidence nodes and the anomaly feature extraction results of the corresponding anomaly type to obtain the anomaly evidence chain. For example, the anomaly evidence chain can take an anomaly event as the central node, using electrical timing evidence, topological clustering evidence, energy residual evidence, environmental association evidence, equipment batch evidence, and work record evidence as evidence nodes, and establishing directed association edges according to time sequence, topological subordination, energy flow direction, environmental impact, equipment batch, and business operation relationships. The anomaly evidence can specifically be represented as: ; In the formula, V represents the set of evidence nodes; A represents the set of evidence-related edges, which includes at least time sequence edges, topological subordination edges, energy flow edges, equipment batch edges, operation-related edges, and counter-evidence constraint edges.

[0055] It should be noted that for each candidate root cause, the evidence support degree of the corresponding candidate root cause can also be calculated based on the positive evidence nodes, negative evidence nodes and their associated edge weights in the abnormal evidence chain.

[0056] By constructing the above-mentioned chain of abnormal evidence, scattered abnormal phenomena, monitoring indicators, equipment attributes, topological relationships, environmental factors, and work records can be transformed into traceable, interpretable, and verifiable structured reasoning evidence, thereby providing a clear evidence path for the controlled reasoning of pre-trained language models.

[0057] Furthermore, the abnormal evidence chain and abnormal feature extraction results are input into the fault prediction model obtained by training the pre-trained language model using historical multi-source heterogeneous metering data. In this way, the fault prediction of the current power metering equipment is performed based on the pre-created metering fault domain knowledge graph, and the fault prediction result is obtained.

[0058] Additionally, this can include: using named entity recognition (NER) technology to analyze knowledge including power industry standard equipment management rules, expert troubleshooting manuals for metering equipment, and historical metering equipment failure cases to construct a triplet structured graph centered on fault phenomena, physical causes of faults, and troubleshooting strategies, thus obtaining a knowledge graph for the metering fault domain. In this embodiment, to improve the prediction accuracy of the fault prediction model, a metering fault domain knowledge graph is pre-created to ensure that the prediction results meet power business rules. Specifically, named entity recognition (NER) technology can be used to analyze unstructured power industry standards (such as technical management regulations for electricity metering devices), expert troubleshooting manuals, and historical typical fault cases to construct a triplet structured graph centered on fault phenomena-physical causes-troubleshooting strategies, serving as an external hard constraint knowledge base for fault prediction model reasoning based on a pre-trained language model.

[0059] In this embodiment, the step of predicting the fault of the power metering equipment based on the knowledge graph of the metering fault domain to obtain the fault prediction result may specifically include: converting the abnormal evidence chain and the abnormal feature extraction result into a vector representation to obtain an abnormal query vector; performing semantic similarity retrieval on the knowledge graph of the metering fault domain based on the abnormal query vector to obtain an abnormal retrieval result; and using the abnormal retrieval result as a prompt word to analyze the abnormal feature extraction result to obtain the fault prediction result of the power metering equipment; the fault prediction result includes the predicted root cause of the fault and the path of the corresponding abnormal evidence chain, the root cause confidence, the fault risk level, and early warning and handling suggestions. In this embodiment, the abnormal evidence chain and abnormal feature extraction results are first mapped to a high-dimensional query vector through a vectorized model, such as the embedding layer, to obtain the abnormal query vector. Then, based on the abnormal query vector, semantic similarity retrieval is performed in the knowledge graph of the fault measurement domain to match the top-N most relevant fault mechanism rules (i.e., abnormal retrieval results). These rules are used as context words and are input together with the abnormal query vector into the fault prediction model for retrieval-augmented generation (RAG) inference. At this time, the model performs inference within the scope defined by expert experience, rather than free divergence.

[0060] As can be seen, the fault prediction model in this application does not perform free-form judgments on raw electrical operation data, but rather relies on controlled inference based on structured chains of anomalous evidence, prior confidence levels of candidate root causes, expert rule retrieval results, and historical similar cases. The fault prediction result R output by the model can be represented in a structured form, for example: ; In the formula, Root represents the predicted root cause of the failure, Conf represents the confidence level of the root cause, EvidencePath represents the path of the abnormal evidence chain supporting the root cause, ConflictSet represents the set of counter-evidence that conflict with the root cause, RiskLevel represents the failure risk level, and Suggestion represents the early warning and handling suggestions.

[0061] By using the above structured output method, we can avoid the fault prediction model from only giving natural language conclusions without traceable evidence. This ensures that the judgment of each root cause can be linked to the corresponding data source, abnormal feature extraction results, expert rules, and physical verification basis.

[0062] Furthermore, to prevent fault prediction models from generating conclusions that violate electrical physics laws or metering rules under complex operating conditions, a counter-verification set based on physical consistency and expert rules can be set to verify the fault prediction results. Specifically, the counter-verification set can be represented as: ; In the formula, to These represent different types of anomaly verification items, including at least energy conservation verification, phase logic verification, voltage-current correlation verification, transformer area topology consistency verification, metering business rule verification, and historical case similarity verification.

[0063] For the root causes of faults predicted by the fault prediction model, proof by contradiction can be performed one by one. When the predicted root cause is inconsistent with energy conservation, Kirchhoff's laws, phase relationship, voltage-current correspondence, or transformer topology, the inconsistent root cause can be marked as evidence of contradiction, and the corresponding root cause confidence level can be corrected. The corrected root cause confidence level is: ; In the formula, This represents the corrected root cause confidence level. This represents the original root cause confidence level of the model output. This represents the root cause of the predicted failure for the j-th proof by contradiction. The intensity of the conflict This indicates the weight of the corresponding proof-of-contrast verification item.

[0064] If the corrected root cause confidence level is lower than a preset confidence threshold, or if the confidence difference between the highest-confidence root cause and the second-candidate root cause is lower than a preset distinction threshold, a backtracking process is triggered. The backtracking process specifically includes: re-searching the knowledge graph of the metering fault domain, adjusting the weights of abnormal evidence items, rearranging differentiated anomaly detection models, supplementing and calling equipment batch quality association models or operation quality association models, and regenerating the abnormal evidence chain. For example, if the model initially determines that a metering device has a voltage loss anomaly, but physical consistency verification finds that the phase voltage, current, power factor, and quantity changes remain normal, and there are no event alarm records corresponding to the voltage loss, then this conclusion is marked as a physically conflicting conclusion, and a backtracking process is triggered to re-search for other candidate fault causes such as acquisition anomalies, communication anomalies, or data loss. Through the above-mentioned backtracking verification mechanism, physical constraints and business rule constraints can be applied to the model's reasoning results, avoiding diagnostic conclusions that are unexplainable or violate the objective laws of the power system.

[0065] Furthermore, based on the above consistency verification results, fault prediction results with higher confidence can be selected. For example, the abnormal event features that have passed physical consistency verification (including temporal mutation features, spatiotemporal clustering labels, line loss energy residuals, etc.) are used as anchor points. Multi-hop reasoning is performed in the knowledge graph of the metering fault domain to extract historical confirmed cases and expert reasoning paths that highly match the current abnormal composite features. The extraction results are then input into the fault prediction model. The model relies on its semantic understanding and generalization reasoning capabilities to perform deep logical deconstruction on ambiguous alarm appearances and calculate the probability distribution (such as confidence scores) of the current abnormality on each potential root cause dimension. Then, based on the probability distribution output by the model, the dimension with the highest confidence score is selected as the dominant root cause (while retaining secondary associated fault root causes), thereby accurately mapping and tracing the alarm information to the five preset root cause dimensions, namely equipment quality dimension, operating condition dimension, external environment dimension, human factor dimension, and operation quality dimension. Among these, the equipment quality dimension is used to comprehensively analyze the correlation between equipment batches and temporal waveform mutations to identify inherent problems such as component aging and batch-specific hardware and software defects; the operating condition dimension is used to combine the power grid physical topology and electrical load curves to identify deteriorating power grid-side operating conditions such as transformer overload, three-phase imbalance, and severe harmonic pollution; the external environment dimension is used to correlate spatiotemporal fault cluster labels with meteorological GIS data to identify environmental damage factors such as lightning overvoltage, strong magnetic interference, and continuous high temperature and humidity; the human factors dimension is used to integrate physical energy analysis residual characteristics with users' historical electricity consumption profiles to identify illegal human interference behaviors such as diversion electricity theft, meter reversal theft, and malicious damage; and the work quality dimension is used to trace recent construction order records and phase logic error characteristics to identify non-standard construction and maintenance issues such as loose installation wiring, reverse polarity connection, and incorrect parameter issuance. It should be noted that dedicated prompt templates can be set for root cause classification in different dimensions.

[0066] To avoid the five-dimensional root cause tracing being limited to static classification, positive evidence sets, negative evidence sets, triggering conditions, and verification rules can be set for each type of root cause. Specifically, the comprehensive confidence level of each candidate root cause can be calculated based on the abnormal evidence chain, and the dominant root cause and secondary related root causes can be retained. Specifically, for equipment quality root causes, the consistency of equipment batches, the consistency of anomaly types, the consistency of software versions, and the degree of environmental exclusion can be calculated; for depressive operating conditions root causes, the load fluctuation, three-phase imbalance, harmonic distortion, overload degree, and the influence range of transformer topology can be calculated; for external environment root causes, the spatiotemporal overlap between abnormal events and external factors such as thunderstorms, high temperatures, humidity, and construction can be calculated; for human factor root causes, the line loss residual, the deviation of user load profile, the abnormal energy flow direction, and historical audit records can be calculated; for operation quality root causes, the proximity of the anomaly occurrence time to the on-site operation time, the matching degree between the anomaly type and the operation content, and the magnitude of changes in electrical characteristics before and after the operation can be calculated.

[0067] Through the above methods, the five-dimensional root cause tracing is transformed from a simple fault classification into a computable judgment process based on abnormal evidence chains, abnormal detection model output (i.e., abnormal feature extraction results), and physical counter-evidence rules.

[0068] After identifying the dominant root cause of the current anomaly in the metering equipment, further risk propagation simulations can be conducted based on the topological relationships of the low-voltage distribution area, equipment batch relationships, GIS spatial proximity relationships, similarity of user load profiles, and correlations with on-site operations. This allows for the identification of a set of potential risky equipment. Specifically, based on the chain of evidence and the confidence level of the root cause, the anomaly can be pinpointed to dimensions such as equipment quality, operating conditions, external environment, human factors, and operational quality. This leads to the generation of a list of potential risky equipment and corresponding handling strategies. For example, for equipment quality-related root causes, horizontal expansion can be performed along equipment manufacturers, models, batches, software versions, and commissioning times to identify the potential risk range of equipment in the same batch, model, or version; for external environment-related root causes, regional expansion can be performed along GIS spatial proximity relationships, meteorological influence range, and transformer area topology; for operational condition-related root causes, expansion can be performed along power supply topology, transformer area load structure, and three-phase imbalance characteristics; for work quality-related root causes, expansion can be performed along the same operators, the same construction batch, tasks issued with the same parameters, and similar work time windows; for human factor-related root causes, expansion can be performed along abnormal energy flow, similarity of user load profiles, and historical audit records. In a specific implementation, the risk score of candidate equipment with potential risks can be calculated based on the dominant root cause, and the specific calculation formula is as follows: ; In the formula, This represents the risk score of candidate device d. Indicates the confidence level of the dominant root cause. Indicates the candidate device d and the abnormal device Similarity in device attributes, user profiles, or operational characteristics between them. Indicates the candidate device d and the abnormal device The topological correlation between them This represents the historical anomaly factor of candidate device d. to These are the corresponding weights.

[0069] Furthermore, the system can rank potentially risky equipment based on risk scores and output risk levels, impact ranges, priority inspection lists, and handling strategies. For equipment quality risks, it can generate suggestions for batch inspection or rotation of equipment in the same batch; for operational quality risks, it can generate on-site verification work orders, prompting verification of wiring, multiplier, phase sequence, transformer polarity, and parameter configuration; for human factor risks, it can generate audit and evidence collection suggestions; and for external environmental risks, it can generate regional inspection and reinforcement suggestions. Through the above-mentioned risk propagation and early warning mechanism, the results of single-point anomaly analysis can be expanded into risk early warning results for related equipment, thereby realizing the transformation from post-fault repair to proactive prevention and control.

[0070] Furthermore, to enhance the system's adaptability during long-term operation, a closed-loop feedback mechanism for on-site handling results can be further implemented. The system receives on-site verification results, fault confirmation results, form replacement detection results, audit conclusions, work order closed-loop results, and manual review opinions, and writes these back as feedback tags to the abnormal evidence chain sample library. When the on-site handling result matches the system's prediction result (i.e., the fault root cause prediction result), the weights of the corresponding abnormal evidence item, the output result of the anomaly detection model, expert rules, and evidence chain path can be increased; when the on-site handling result is inconsistent with the system's prediction result, the weight of the conflicting evidence path can be reduced, and new rebuttal rules or supplementary samples can be generated. Moreover, the aforementioned closed-loop feedback mechanism can also be used to continuously optimize the credibility of abnormal evidence items, the dynamic orchestration strategy of the anomaly detection model, the root cause confidence calculation parameters, and risk propagation early warning rules, thereby enabling the system to continuously improve the accuracy of judgment and the reliability of early warning as business data accumulates. Through the aforementioned proactive early warning mechanism, single-point anomaly handling can be expanded to batch hidden danger early warning and proactive operation and maintenance control, thereby shortening the on-site investigation cycle and improving the allocation efficiency of operation and maintenance resources.

[0071] In one specific implementation, when the line loss analysis model detects a slight increase in the line loss rate of a low-voltage distribution area (e.g., from 4.5% to 6.2%, but not reaching the 8% alarm threshold), and simultaneously, the electrical anomaly detection model uses the LSTM algorithm to identify an abnormal pulse-like drop in the load curve of a user in that area at night, which deviates significantly from the user's historical electricity consumption habits (user profile), further, the fault prediction model performs inference. First, it calls the distribution area topology data to confirm that the user has no distributed photovoltaic access (to rule out reverse generation interference), then it calls the spatiotemporal environmental data to confirm that there was no extreme weather at the time (to rule out line leakage). Next, it retrieves the characteristic pattern of electricity theft by current diversion from the metering fault domain knowledge graph, which highly matches the current load waveform, triggering an energy conservation check. It calculates the integral of the difference between the total meter and the sub-meter electricity consumption, and finds that the trend of the difference change is highly positively correlated with the time of the user's load drop, thus determining it to be human interference. Finally, the root cause is determined to be human factors (suspected electricity theft by diverting current), and a confidence level (e.g., 94%) is output. At the same time, handling suggestions are generated, such as an evidence collection work order, which is used to prompt inspectors to focus on checking the sealing status of the user's meter box and the concealed interface of the incoming line, and recommend the best time window for investigation (e.g., 20:00-22:00 at night).

[0072] As can be seen, the embodiments of this application first perform spatiotemporal alignment and feature fusion on the multi-source heterogeneous metering data of the power metering equipment, and construct anomaly evidence items based on the multimodal spatiotemporal feature tensors obtained after fusion. Then, a model orchestrator selects multiple anomaly detection models corresponding to the anomaly evidence items from the anomaly detection model pool, and dynamically orchestrates the multiple anomaly detection models into a model execution chain. The model execution chain then extracts anomaly features from the multi-source heterogeneous metering data, and organizes the anomaly feature extraction results output by different anomaly detection models into an anomaly evidence chain. Finally, a pre-trained language model combined with a knowledge graph of the metering fault domain is used to analyze the anomaly evidence chain and the anomaly feature extraction results, thereby obtaining the fault prediction result. Therefore, the embodiments of this application... Instead of fault prediction based on fixed thresholds or a single model, this method dynamically generates corresponding model execution chains based on the characteristics of multi-source heterogeneous metering data collected in real time. This allows for feature extraction of the multi-source heterogeneous metering data, and the construction of anomaly evidence chains based on the feature extraction results. Then, a pre-trained language model (combined with a knowledge graph) is used for comprehensive fault prediction. Through this approach, automated equipment fault prediction is achieved, shifting from post-event handling to pre-event prevention. This also improves the accuracy of anomaly identification in power metering equipment and the penetration of root cause analysis, reducing missed and false judgments caused by single data sources or fixed threshold judgments. At the same time, by dynamically selecting anomaly detection models that are adapted to the current working conditions, the waste of computing resources can be avoided.

[0073] Accordingly, this application also discloses a fault prediction device for electricity metering equipment, see [link to relevant documentation]. Figure 2 As shown, the device includes: The acquisition module 11 is used to acquire current multi-source heterogeneous metering data of the power metering equipment in the target distribution area; Data processing module 12 is used to perform spatiotemporal alignment and feature fusion on the current multi-source heterogeneous measurement data to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates; Construction module 13 is used to construct anomaly evidence items based on the multimodal spatiotemporal feature tensor; the anomaly evidence items are used to describe the support relationship and / or disproving relationship between the device anomaly and the candidate root cause; The first input module 14 is used to input the abnormal evidence items into a pre-created model orchestrator to select multiple target abnormal detection models of corresponding abnormal types from a preset abnormal detection model pool, and to dynamically orchestrate the multiple target abnormal detection models to generate an abnormal detection model execution chain; the abnormal detection models in the abnormal detection model pool are models obtained by training a preset neural network using historical econometric data corresponding to different abnormal types. The second input module 15 is used to input the current multi-source heterogeneous measurement data into the anomaly detection model execution chain to extract anomaly features from the current multi-source heterogeneous measurement data and obtain anomaly feature extraction results; the anomaly feature extraction results include anomaly feature labels, candidate root causes and corresponding root cause confidence levels; Organization module 16 is used to organize multiple abnormal feature extraction results output by different target anomaly detection models into an anomaly evidence chain; The third input module 17 is used to input the abnormal evidence chain and the abnormal feature extraction result into the trained fault prediction model, so as to perform fault prediction on the power metering equipment according to the knowledge graph of the metering fault domain and obtain the fault prediction result; the fault prediction model is a model obtained by training a pre-trained language model using historical multi-source heterogeneous metering data.

[0074] The specific workflow of each of the above modules can be found in the relevant content disclosed in the foregoing embodiments, and will not be repeated here.

[0075] Furthermore, embodiments of this application also disclose an electronic device, Figure 3 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0076] Figure 3This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the power metering equipment fault prediction method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0077] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0078] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0079] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the power metering device fault prediction method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0080] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned method for predicting faults in electricity metering equipment. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0081] Furthermore, embodiments of this application also disclose a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the power metering equipment fault prediction method disclosed above.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0083] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0084] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0085] Finally, 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0086] The above provides a detailed description of the fault prediction method, apparatus, equipment, and medium for electricity metering equipment provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting faults in electricity metering equipment, characterized in that, include: Collect current multi-source heterogeneous metering data of power metering equipment in the target distribution substation area, and perform spatiotemporal alignment and feature fusion on the current multi-source heterogeneous metering data to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates. Anomaly evidence terms are constructed based on the multimodal spatiotemporal feature tensor; the anomaly evidence terms are used to describe the support and / or disproving relationships between the equipment anomaly and the candidate root causes; The abnormal evidence items are input into a pre-created model orchestrator to select multiple target abnormal detection models of the corresponding abnormal type from a preset abnormal detection model pool, and to dynamically orchestrate the multiple target abnormal detection models to generate an abnormal detection model execution chain. The anomaly detection model in the anomaly detection model pool is a model obtained by training a preset neural network using historical measurement data corresponding to different anomaly types. The current multi-source heterogeneous measurement data is input into the execution chain of the anomaly detection model to extract anomaly features from the current multi-source heterogeneous measurement data, and anomaly feature extraction results are obtained; the anomaly feature extraction results include anomaly feature labels, candidate root causes and corresponding root cause confidence scores; The multiple abnormal feature extraction results output by different target anomaly detection models are organized into an anomaly evidence chain, and the anomaly evidence chain and the abnormal feature extraction results are input into the trained fault prediction model to perform fault prediction on the power metering equipment based on the metering fault domain knowledge graph, and obtain the fault prediction result; the fault prediction model is a model obtained by training a pre-trained language model using historical multi-source heterogeneous metering data.

2. The fault prediction method for electricity metering equipment according to claim 1, characterized in that, The construction of anomaly evidence items based on the multimodal spatiotemporal feature tensor includes: Extract the abnormal evidence information corresponding to each abnormal phenomenon in the multimodal spatiotemporal feature tensor; the abnormal evidence information includes abnormal evidence type, abnormal evidence source, associated metering equipment, abnormal occurrence time, abnormal topological location, abnormal feature value, abnormal evidence credibility, and the set of supporting candidate root causes and the set of counter-evidence candidate root causes corresponding to the abnormal phenomenon; Different pieces of evidence information in the abnormal evidence information are combined to obtain an abnormal evidence item that includes the abnormal evidence information.

3. The fault prediction method for electricity metering equipment according to claim 2, characterized in that, The step of selecting multiple target anomaly detection models of corresponding anomaly types from a preset anomaly detection model pool includes: Obtain the metering business context information related to the power metering equipment, and calculate the confidence level of each candidate root cause in the abnormal evidence item based on the credibility of the abnormal evidence item to obtain the prior confidence level. Based on the abnormal evidence items, the metering business context information, and the prior confidence level, an anomaly detection model of the corresponding anomaly type is selected from the preset anomaly detection model pool to obtain multiple target anomaly detection models.

4. The fault prediction method for electricity metering equipment according to claim 2, characterized in that, The step of organizing multiple anomaly feature extraction results output by different target anomaly detection models into an anomaly evidence chain includes: Using the abnormal event as the central node, the abnormal evidence type corresponding to the abnormal feature extraction result is taken as the evidence node, and directed association edges are established between different evidence nodes according to the preset evidence association relationship. A mapping relationship is also established between different evidence nodes and the abnormal feature extraction result of the corresponding abnormal type to obtain the abnormal evidence chain.

5. The method for predicting faults in electricity metering equipment according to claim 1, characterized in that, Also includes: Named entity recognition technology is used to analyze knowledge including power industry standard equipment management rules, metering equipment expert troubleshooting manuals and historical metering equipment failure cases, in order to construct a triplet structured graph with fault phenomena, fault physical causes and fault troubleshooting strategies as the core, and obtain a knowledge graph in the field of metering faults.

6. The fault prediction method for electricity metering equipment according to claim 5, characterized in that, The step of performing fault prediction on the power metering equipment based on a knowledge graph of metering fault domains to obtain fault prediction results includes: The abnormal evidence chain and the abnormal feature extraction results are converted into vector representations to obtain the abnormal query vector; Based on the anomaly query vector, semantic similarity retrieval is performed on the knowledge graph of metering fault domain to obtain anomaly retrieval results. The anomaly retrieval results are then used as prompt words to analyze the anomaly feature extraction results, thereby obtaining the fault prediction results of the power metering equipment. The fault prediction results include the predicted root cause of the fault and the path of the corresponding anomaly evidence chain, the root cause confidence, the fault risk level, and early warning and handling suggestions.

7. The method for predicting faults in electricity metering equipment according to any one of claims 1 to 6, characterized in that, The current multi-source heterogeneous metering data includes electrical operation data, abnormal event alarm data, transformer area topology data, equipment ledger data, environmental meteorological data, and on-site operation data; Accordingly, the step of performing spatiotemporal alignment and feature fusion on the current multi-source heterogeneous metrological data to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates includes: Using the highest frequency data sampling period as a benchmark, the current multi-source heterogeneous metering data is resampled in the frequency domain to obtain the sampled multi-source metering data; The sampled multi-source measurement data is filtered and smoothed to generate a processed multi-source measurement data stream. The processed multi-source measurement data stream is fused on both the temporal and spatial scales to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates.

8. A fault prediction device for electricity metering equipment, characterized in that, include: The data acquisition module is used to collect current multi-source heterogeneous metering data of the power metering equipment in the target distribution area; The data processing module is used to perform spatiotemporal alignment and feature fusion on the current multi-source heterogeneous measurement data to generate a multimodal spatiotemporal feature tensor with unified spatiotemporal coordinates. The construction module is used to construct anomaly evidence items based on the multimodal spatiotemporal feature tensor; the anomaly evidence items are used to describe the support relationship and / or disproving relationship between the device anomaly and the candidate root cause; The first input module is used to input the abnormal evidence items into a pre-created model orchestrator to select multiple target abnormal detection models of the corresponding abnormal type from a preset abnormal detection model pool, and to dynamically orchestrate the multiple target abnormal detection models to generate an abnormal detection model execution chain. The anomaly detection model in the anomaly detection model pool is a model obtained by training a preset neural network using historical measurement data corresponding to different anomaly types. The second input module is used to input the current multi-source heterogeneous measurement data into the anomaly detection model execution chain to extract anomaly features from the current multi-source heterogeneous measurement data and obtain anomaly feature extraction results; the anomaly feature extraction results include anomaly feature labels, candidate root causes and corresponding root cause confidence levels; The organization module is used to organize multiple abnormal feature extraction results output by different target anomaly detection models into an anomaly evidence chain; The third input module is used to input the abnormal evidence chain and the abnormal feature extraction results into the trained fault prediction model, so as to perform fault prediction on the power metering equipment according to the knowledge graph of the metering fault domain and obtain the fault prediction result; the fault prediction model is a model obtained by training a pre-trained language model using historical multi-source heterogeneous metering data.

9. An electronic device, characterized in that, It includes a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the fault prediction method for power metering equipment as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when the computer program is executed by a processor, it implements the fault prediction method for power metering equipment as described in any one of claims 1 to 7.

Images