Method and system for real-time monitoring of power metering load fluctuation

Abstract

This invention relates to the field of power monitoring, specifically to a method and system for real-time monitoring of load fluctuations in power metering. The method includes the following steps: synchronously accessing three types of data: load time series, scenario attributes, and power grid operating conditions, and outputting a compliance feature sequence; classifying scenarios based on historical standardized data, filtering normal samples through unsupervised clustering, constructing baselines for each scenario through time series decomposition, fitting and calculating initial anomaly judgment boundaries, and forming a scenario-based baseline library; real-time matching of the scenario corresponding to the current operating condition and loading the baseline, updating the baseline incrementally in unsupervised manner based on normal samples within a sliding window, dynamically calibrating the anomaly judgment boundaries, and setting hard constraints for boundary adjustment; establishing a small-sample fault prototype library, and performing real-time sequence anomaly pattern matching based on the small-sample fault prototype library, making a judgment based on a dual-dimensional fusion of boundary exceedance and pattern matching results, and outputting a graded alarm. This application achieves accurate and efficient real-time monitoring of load fluctuations in power metering.

Description

Technical Field

[0001] This invention belongs to the field of power monitoring, specifically relating to a method and system for real-time monitoring of load fluctuations in power metering. Background Technology

[0002] Current electricity metering for load fluctuation monitoring typically uses fixed threshold alarms. Traditional fixed threshold alarms can lead to numerous false alarms during peak summer seasons, Spring Festival, and extreme weather events due to the sudden increase in normal load fluctuations. Conversely, lenient thresholds can result in missed detections of minor anomalies. The "normal / abnormal" boundary of load fluctuations is dynamically evolving, and labeled data for extreme operating conditions and small-sample faults are extremely scarce. Therefore, how to dynamically update the anomaly judgment boundary in real-time streaming data, while ensuring a high recall rate for identifying latent anomalies with an extremely low false alarm rate, has become a significant technical challenge. Summary of the Invention

[0003] The purpose of this invention is to provide a method for real-time monitoring of load fluctuations in electricity metering, and to provide a system for real-time monitoring of load fluctuations in electricity metering, so as to solve the problems mentioned in the background art.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0005] A method for real-time monitoring of load fluctuations in electricity metering, including the following steps:

[0006] Synchronously access three types of data: load time sequence, scenario attributes, and power grid operating conditions, unify the sampling period, complete missing value completion, bad data removal and feature standardization, and output compliant feature sequences;

[0007] Based on historical standardized data, scene labeling was completed, normal samples were screened through unsupervised clustering, baselines for each scene were constructed through time-series decomposition, and initial anomaly judgment boundaries were calculated to form a scene baseline library.

[0008] Real-time matching of the current working condition and corresponding scenario and loading of baseline; based on normal samples within the sliding window, unsupervised incremental update of baseline, dynamic calibration of anomaly judgment boundary, and setting hard constraints for boundary adjustment.

[0009] A small sample fault prototype library is established, and real-time sequence anomaly pattern matching is performed based on the small sample fault prototype library. The judgment is based on the dual-dimensional fusion of boundary exceedance and pattern matching results, and a graded alarm is output.

[0010] Furthermore, the synchronous access to the three types of data—load time series, scenario attributes, and power grid operating conditions—includes: load time series data, which includes electrical-related data collected in real time by the power metering device; scenario attribute data, which includes meteorological data and time-dimensional feature data; and power grid operating condition data, which includes power grid operating condition data.

[0011] Furthermore, the process of classifying scenarios based on historical standardized data specifically includes: extracting historical standardized feature data from the time-series database, using scenario attribute features as the basis for classification, performing dimensional labeling processing on the historical data, classifying basic electricity consumption and operating condition scenarios, and establishing a corresponding association between scenarios and historical load fluctuation data.

[0012] Furthermore, the step of screening normal samples through unsupervised clustering includes: performing clustering processing using unsupervised clustering on the historical load sequences under each single basic scenario that have been divided, screening samples within the core area of ​​the cluster as the normal load sample set for the corresponding scenario, and removing outlier abnormal samples at the edge of the cluster.

[0013] Furthermore, the fitting calculation of the initial anomaly determination boundary to form a scenario-based baseline library includes: performing time-series decomposition on the normal load sample set screened by unsupervised clustering under each scenario, decomposing the load time-series data into multiple time-series components, fitting the probability distribution of normal load samples based on the residual terms of the normal sample set, calculating the upper and lower limits of the corresponding confidence interval as the initial anomaly determination boundary for the corresponding scenario, and then storing the scenario label, baseline component parameters, initial anomaly determination boundary and normal sample statistics into the scenario adaptive baseline library to form a scenario-based baseline library.

[0014] Furthermore, the real-time matching of the current operating condition corresponding to the scenario and loading of the baseline includes: extracting the scenario attributes, power grid operating conditions and time dimension features of the standardized feature stream of the current sampling period in real time, generating real-time scenario labels, performing multi-dimensional matching degree calculation with the full scenario labels of the scenario adaptive benchmark baseline library, the matching dimensions include time attributes, meteorological conditions and power grid operating conditions, the weight of each dimension is executed according to preset rules, and according to different matching results, the corresponding scenario baseline and boundary are judged respectively, and the scenario matching result, effective baseline parameters and abnormal judgment boundary values ​​are output.

[0015] Furthermore, the unsupervised incremental update of the baseline and dynamic calibration of the anomaly judgment boundary based on normal samples within a sliding window includes: setting a sliding update window with a step size consistent with the sampling period of the load data; performing initial screening on real-time load samples within the window based on the currently effective anomaly judgment boundary, and including samples that do not exceed the boundary into the normal incremental update sample set to exclude interference from samples exceeding the limit; disassembling the normal incremental sample set, performing unsupervised incremental updates on the baseline based on the disassembled components to correct the baseline deviation; fitting the probability distribution of the residual terms of normal samples, and dynamically calibrating the upper and lower limits of the anomaly judgment boundary based on the change in the bandwidth of the residual distribution.

[0016] Furthermore, the real-time sequence anomaly pattern matching based on the small sample fault prototype library includes: real-time extraction of load time series sequences and related features that are homologous to the baseline incremental update and have undergone standardization, ensuring that they are consistent with the sample processing rules of the fault prototype library; generating real-time feature vectors using homologous encoding and incorporating scenario and working condition weights, firstly screening the prototype library in a scenario-based manner, and then matching through cosine similarity traversal to output similarity and corresponding fault information.

[0017] Furthermore, the method of determining the boundary exceedance and pattern matching results from two dimensions and outputting graded alarms includes: performing boundary exceedance determination based on real-time standardized load data, simultaneously performing abnormal pattern determination based on abnormal pattern matching results, implementing three-level graded alarms according to abnormal confidence and risk level, pushing alarm information along with traceability data to the monitoring system and storing it, and simultaneously implementing hard constraints on alarm frequency.

[0018] This application also discloses an electronic device, including:

[0019] At least one processor; and

[0020] A memory communicatively connected to the at least one processor; wherein,

[0021] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the above-described method for real-time monitoring of power metering load fluctuations of the present invention.

[0022] This application also discloses a real-time monitoring system for electricity load fluctuations, including:

[0023] The data synchronization module is used to synchronize the access of three types of data and unify the sampling period;

[0024] The data preprocessing module is used to complete missing values, remove bad data, standardize features, and output compliant feature sequences.

[0025] The scene tagging module is used to classify scenes based on historical data;

[0026] The baseline construction module is used to screen normal samples, build a benchmark baseline, fit the initial abnormal boundary, and form a scenario-based baseline library; the baseline update module is used to match the working condition loading baseline, incrementally update the baseline and dynamically calibrate the abnormal boundary, and set boundary adjustment hard constraints.

[0027] The fault prototype library module is used to build a small sample fault prototype library;

[0028] The anomaly detection module is used to match anomaly patterns, fuse boundary exceedance and pattern matching results, and output hierarchical alarms.

[0029] The model iteration module is used to iteratively optimize the model based on verification feedback, forming a closed loop of operation.

[0030] Beneficial Effects: This application addresses the problems of false alarms and missed alarms caused by traditional fixed threshold alarms, difficulties in dynamically adapting to abnormal boundaries, and a lack of small-sample fault annotations, achieving accurate and efficient real-time monitoring of load fluctuations in electricity metering. Through multi-source data synchronization and alignment and standardized preprocessing, it ensures data compliance and usability. By employing scenario-based labeling and baseline construction, it accurately adapts to complex operating conditions such as peak summer seasons, holidays, and extreme weather, fundamentally solving the problem that fixed thresholds cannot adapt to large normal fluctuations. Relying on unsupervised clustering to screen normal samples and time-series decomposition to construct baselines, it eliminates the need for extensive manual data annotation, overcoming the challenge of insufficient annotation data for extreme operating conditions and small-sample faults. Through unsupervised incremental updates of the baseline via sliding windows and dynamic calibration of anomaly detection boundaries, coupled with hard constraints for boundary adjustments, it adapts to the dynamic evolution of load patterns while avoiding excessive boundary offset distortion, improving the detection capability of small anomalies while strictly controlling the false alarm rate. A small-sample fault prototype library is established to achieve latent anomaly pattern matching. Combining boundary exceedance and pattern matching for dual-dimensional fusion judgment, it takes into account both explicit exceedance anomalies and latent feature anomalies, significantly improving the anomaly identification recall rate. Tiered alarms can accurately distinguish risk levels. At the same time, a model iteration closed loop is formed based on review feedback to continuously optimize the baseline, boundary, and fault prototypes, and continuously improve monitoring accuracy. It truly achieves real-time monitoring with a high recall rate and extremely low false alarm rate, and stably adapts to the changing operating conditions of the power grid. Attached Figure Description

[0031] Figure 1 This is a flowchart of the real-time monitoring method for power load fluctuations in this invention;

[0032] Figure 2 This is a schematic diagram of the process of real-time working condition scene matching and corresponding baseline loading in this invention;

[0033] Figure 3 This is a schematic diagram of the unsupervised incremental baseline update and dynamic calibration of anomaly determination boundaries in this invention;

[0034] Figure 4 This is a schematic diagram illustrating the construction process of the small sample fault prototype library in this invention;

[0035] Figure 5 This is a schematic diagram of the real-time sequence anomaly pattern matching process based on a small sample fault prototype library in this invention.

[0036] Figure 6 This is a bar chart comparing the false alarm rates of various methods used in different scenarios in this invention. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0038] This invention provides a method for real-time monitoring of load fluctuations in electricity metering, such as... Figure 1 As shown, the steps include:

[0039] Synchronously access three types of data: load time sequence, scenario attributes, and power grid operating conditions, unify the sampling period, complete missing value completion, bad data removal and feature standardization, and output compliant feature sequences;

[0040] Based on historical standardized data, scene labeling was completed, normal samples were screened through unsupervised clustering, baselines for each scene were constructed through time-series decomposition, and initial anomaly judgment boundaries were calculated to form a scene baseline library.

[0041] Real-time matching of the current working condition and corresponding scenario and loading of baseline; based on normal samples within the sliding window, unsupervised incremental update of baseline, dynamic calibration of anomaly judgment boundary, and setting hard constraints for boundary adjustment.

[0042] A small sample fault prototype library is established, and real-time sequence anomaly pattern matching is performed based on the small sample fault prototype library. The judgment is based on the dual-dimensional fusion of boundary exceedance and pattern matching results, and a graded alarm is output.

[0043] The synchronous access to three types of data—load time series, scenario attributes, and power grid operating conditions—in practice specifically includes: synchronously accessing these three core data sources at millisecond-level speeds. Load time series data is real-time data collected by electricity metering devices, including active power, reactive power, current, and voltage data. Scenario attribute data includes ambient temperature, relative humidity, meteorological data such as extreme weather warning levels, and time characteristic data such as Gregorian and lunar calendar dates, time period types, holiday types, and peak summer and peak winter time period markers. Power grid operating condition data includes the load rate of the relevant lines, the operating status of distribution transformers, and the overall power grid operating condition data. Simultaneously, the sampling period of all accessed data sources is unified to achieve synchronous alignment of multi-source data.

[0044] The unified sampling period, which completes missing value completion, bad data removal, and feature standardization, outputs a compliant feature sequence. Specifically, this includes: adjusting the sampling period of the three core data sources—load time series, scenario attributes, and power grid operating conditions—to 15 minutes to achieve millisecond-level synchronization and alignment of multi-source data; filling missing values ​​in the data with the average normal load within a 7-day sliding window for the same scenario; removing bad data samples caused by acquisition link failures using the 3σ criterion, retaining valid load fluctuation samples; performing min and max standardization on all numerical features, mapping them uniformly to the [0,1] interval to eliminate feature dimension differences; simultaneously performing feature dimension splitting, dividing the standardized features into three categories: basic load features, scenario attribute features, and power grid operating condition features; and finally outputting a standardized compliant feature sequence, which is synchronously written to the time series database for persistent storage.

[0045] The process of classifying scenarios based on historical standardized data specifically includes: extracting at least 12 months of historical standardized feature data from the time-series database, using the scenario attribute features as the core classification basis, performing dimensional labeling processing on the historical data, and identifying six basic scenarios: regular weekdays, regular weekends, statutory holidays (including Spring Festival), peak summer seasons, peak winter seasons, and extreme weather conditions. A correspondence between scenarios and historical load fluctuation data is established, laying the foundation for subsequent unsupervised clustering to screen normal samples and construct scenario-based baselines.

[0046] The method of screening normal samples through unsupervised clustering specifically includes: for the historical load sequences under each single basic scenario that have been divided, DBSCAN unsupervised clustering is used to carry out clustering processing, and samples within the core area of ​​the cluster are selected as the normal load sample set of the corresponding scenario. Outlier abnormal samples at the edge of the cluster are removed. The automatic initial screening of normal and abnormal samples can be achieved without manual annotation, providing clean normal sample data for subsequent time series decomposition and baseline construction.

[0047] The fitting calculation of the initial anomaly determination boundary to form a scenario-based baseline library specifically includes: performing STL time series decomposition on the normal load sample set obtained by unsupervised clustering in each scenario, decomposing the load time series data into three components: trend, period, and residual. Based on the residual of the normal sample set, the kernel density estimation method is used to fit its probability distribution, and the upper and lower limits corresponding to the 99% confidence interval are calculated. This value is used as the initial anomaly determination boundary for the corresponding scenario. Then, the scenario label, the parameters of each component of the baseline, the initial anomaly determination boundary, and the normal sample statistics are uniformly stored in the scenario adaptive baseline library, finally forming a scenario-based baseline library that includes all scenarios.

[0048] The real-time matching of the current working condition and loading of the baseline, such as Figure 2 In practice, this includes: extracting scene attribute features and power grid operating condition features from the standardized feature stream of the current sampling period in real time, synchronously associating them with the corresponding time dimension features, generating complete real-time scene labels, and performing multi-dimensional matching degree calculations between the generated real-time scene labels and the full scene labels stored in the scene adaptive baseline library. The matching degree calculation dimensions include three major categories: time attributes, meteorological conditions, and power grid operating conditions. The matching degree weights of each dimension are executed according to preset fixed values, among which the time attribute dimension is the core matching dimension, with a weight ratio of no less than 50%, to ensure the matching accuracy of the core fluctuation characteristics of the scene.

[0049] If the matching calculation yields a unique high-matching scenario with a matching degree ≥ 95%, then the baseline parameters corresponding to that scenario and the currently effective anomaly detection boundary are directly loaded.

[0050] If the matching degree values ​​of multiple scenarios are all ≥85% after matching calculation, it is determined to be a multi-scenario superposition condition. Scenario weight fusion is adopted, using the matching degree value corresponding to each scenario as the weight coefficient, and performing weighted fusion calculation on the baseline trend item, period item, residual item component parameters of the corresponding scenario, as well as the upper and lower limit values ​​of the anomaly judgment boundary, to generate a composite baseline and fused anomaly judgment boundary that are fully adapted to the current superposition condition.

[0051] If the matching degree calculation results of all scenarios are <85%, it is determined that no valid scenario has been matched. The baseline of the general common scenario and the corresponding anomaly judgment boundary are loaded first to ensure that the real-time monitoring process is not interrupted. At the same time, the real-time condition is marked as a new scenario to be learned, and the full feature data corresponding to the scenario is recorded in the time series database to provide complete data support for the subsequent baseline construction and storage of new scenarios.

[0052] After completing the entire process of scene matching and baseline loading, the system synchronously outputs the scene matching result labels, the full set of parameters of the loaded baseline, and the anomaly detection boundary values, providing accurate matching basis for subsequent baseline incremental updates, boundary dynamic calibration, and anomaly identification.

[0053] The method involves unsupervised incremental baseline updates based on normal samples within a sliding window and dynamic calibration of anomaly detection boundaries. Figure 3 In practice, this includes setting a 24-hour sliding update window with a step size consistent with the load data sampling period. For real-time load samples within the window, initial screening of samples is completed based on the anomaly judgment boundaries that have taken effect under the current matching loading scenario. Samples that do not exceed the upper and lower limits of the boundaries are included in the normal incremental update sample set of the corresponding scenario, eliminating the interference of samples exceeding the limits on baseline and boundary updates.

[0054] Incremental STL time series decomposition is adopted to decompose the normal incremental update sample set into time series components to obtain the corresponding trend, periodic and residual components. Based on the component calculation results of the incremental samples, unsupervised incremental updates are performed on the trend and periodic components of the baseline to correct the long-term trend and periodic fluctuation characteristics of the baseline in real time and eliminate the baseline lag bias caused by the long-term evolution of load electricity consumption patterns.

[0055] Online kernel density estimation is used to fit the probability distribution of the residual terms of normal samples within a sliding window in real time. The upper and lower limits of the anomaly judgment boundary are dynamically calibrated based on the bandwidth change of the residual distribution. When the bandwidth of the residual distribution continues to increase within three consecutive sliding windows, the boundary is adaptively widened to the upper and lower limits corresponding to the 99% confidence interval of the residual distribution of normal samples in that scenario, so as to avoid batch false alarms caused by a sudden increase in the amplitude of normal load fluctuations. When the bandwidth of the residual distribution continues to narrow within three consecutive sliding windows, the boundary is adaptively tightened to the upper and lower limits corresponding to the 99.5% confidence interval of the residual distribution of normal samples in that scenario, so as to avoid small abnormal fluctuations being missed.

[0056] The setting of boundary adjustment hard constraints specifically includes: determining the scope of application of hard constraint rules across all scenarios, including the native anomaly judgment boundary of a single scenario, the fusion boundary under the superposition of multiple scenarios, and the temporary boundary loaded in the learning state of a new scenario, to ensure that boundary adjustment behavior in all scenarios is subject to constraint control.

[0057] For boundary adjustments triggered after a single sliding update window is completed, a hard threshold for the magnitude of a single adjustment is set. For example, the magnitude of a single upward widening or downward tightening of the same boundary in the same scene must not exceed ±15% of the effective value of the boundary before the adjustment. If the adjustment magnitude calculated by online kernel density estimation exceeds this threshold, the adjustment is performed at the maximum allowable magnitude of ±15%. At the same time, the triggering scenario, data characteristics and calculation results of this over-adjustment are recorded and included in the optimization scope of subsequent periodic model verification.

[0058] Using natural days as the statistical period, a hard threshold is set for the cumulative adjustment range per day. For example, for the same scenario and the same anomaly judgment boundary, the total upward widening range per day shall not exceed +30% of the initial effective boundary value of the day, and the total downward tightening range per day shall not exceed -30% of the initial effective boundary value of the day. If the cumulative adjustment range reaches this threshold, the same-direction adjustment operation of the boundary shall be suspended for the rest of the day, and only the reverse callback operation shall be allowed to avoid the failure of anomaly judgment caused by continuous unidirectional offset of the boundary.

[0059] Set hard constraints on sample quality for boundary updates. The input samples for boundary adjustments are limited to the normal incremental update sample set after initial screening by the currently effective boundary. For example, if the proportion of valid normal samples in a single sliding window is less than 90% of the total sample size in the window, the boundary update operation corresponding to that window is immediately paused, and the boundary parameters that have already taken effect in the previous window are directly used. At the same time, the data anomaly verification process is triggered to investigate interference factors such as acquisition link failures and batch anomaly events, and to prevent abnormal samples from being mixed into the update set, which would cause baseline offset and boundary distortion.

[0060] For work conditions marked as new scenarios to be learned, special boundary adjustment constraints are set. For example, the temporary loaded general conventional scenario boundary is only allowed to be fine-tuned by no more than ±10% during the new scenario sample accumulation period, and the cumulative adjustment range per day shall not exceed ±20%. Only after the new scenario has accumulated no less than 7 days of continuous normal sample data, completed the construction of unsupervised baseline and the initial boundary calculation, and passed the offline test set verification to meet the core performance indicators, can it be included in the full boundary dynamic adjustment range.

[0061] Set a hard constraint for a mandatory callback triggered by false alarms. For example, if the number of false alarm samples triggered by the same alarm rule in the same scenario on a single day exceeds 0.5% of the total number of samples in that scenario on that day, a mandatory boundary callback will be triggered immediately. This will roll back the anomaly judgment boundary of that scenario to the parameter state that was initially effective on that day, suspend the dynamic adjustment function of that scenario, and simultaneously start the targeted analysis of false alarm samples and boundary calibration. For example, after the calibrated parameters are verified by the offline test set and meet the core indicators of alarm false alarm rate ≤0.5% and anomaly recognition recall rate ≥99%, the dynamic adjustment permission can be restored.

[0062] The establishment of a small sample fault prototype library, such as Figure 4 In practice, this includes: determining the scope and criteria for inclusion of fault samples; including load time-series samples of all types of electricity metering-related faults that have been manually annotated and confirmed in the past, such as hidden faults of electricity metering devices, abnormal electricity theft, abnormal line loss, abnormal light load / overload of distribution transformers, abnormal wiring of metering devices, and excessive error of instrument transformers; supporting the entry of small sample data with a minimum sample size of no more than 50 samples for a single type of fault; and requiring inclusion samples to meet the basic requirements of complete sampling period, clear labeling of fault time period, and no bad data interference caused by failure of acquisition link.

[0063] All fault samples entering the database undergo standardized preprocessing that is consistent with the real-time streaming data. The sampling period of the samples is uniformly aligned to 15 minutes. Missing values ​​are filled with the average normal load of the same scenario and time period. Invalid bad data is removed by using the 3σ criterion. All numerical features are standardized by min and max and uniformly mapped to the [0,1] interval. The dimensional decomposition of basic load features, scenario attribute features, and power grid operating condition features is completed simultaneously to ensure that the feature dimensions and processing rules of fault samples and real-time streaming data are completely consistent.

[0064] The standardized fault samples are managed using structured classification labels. Multi-dimensional labels are added based on fault root cause type, abnormal impact range, and typical fault occurrence scenarios. A mapping relationship is established between fault samples and six basic scenarios in the scenario adaptive baseline library, providing support for subsequent scenario-based accurate matching.

[0065] Using the same feature encoding logic as the real-time sequence, and based on prototype network few-shot learning, the standardized sample sequence of each type of fault is uniformly feature-encoded, the mean of all sample encoding vectors of that type is calculated, and a unique, quantifiable feature prototype vector corresponding to the fault type is generated. The prototype vector generation time, corresponding sample size, iteration version number and other traceability information are recorded synchronously.

[0066] The fault type structured labels, original labeled sample sets, corresponding feature prototype vectors, typical fault scenario labels, and similarity judgment reference thresholds are uniformly stored in the fault feature prototype library. This establishes a structured storage architecture that can be quickly retrieved and called in real time, ensuring that the corresponding prototype vector can be retrieved in milliseconds when matching real-time abnormal patterns.

[0067] Establish strict verification standards for prototype entry into the database. Each newly added or iteratively updated fault feature prototype must pass cross-validation on an offline test set. For example, the verification pass standard is that the recognition recall rate of the corresponding fault type is not less than 99% and the cross-scenario mismatch rate is not more than 0.5%. Prototypes that do not meet the standards will not be included in the database. Establish a dynamic iterative update mechanism for the prototype database, and reserve interfaces for sample expansion and prototype optimization. Based on the abnormal samples confirmed in the alarm review process, the expansion of the corresponding fault type sample set and the iterative update of the feature prototype vector can be completed. At the same time, it supports the sample entry, feature encoding and prototype generation of new fault types for database entry, adapts to the dynamic evolution characteristics of power metering fault modes, and forms a complete linkage with the whole process model closed-loop optimization mechanism.

[0068] The method of real-time sequence anomaly pattern matching based on a small sample fault prototype library, such as... Figure 5The implementation specifically includes: real-time extraction of complete load time-series sequences that are from the same source as the incremental update of the baseline, within the current 24-hour sliding window, and after standardized preprocessing; synchronous association with the scene attribute features and power grid operating condition features within the corresponding window; ensuring that the sampling period, feature dimension, and standardized processing rules of the input sequence are completely consistent with the preprocessing rules of the samples in the fault feature prototype library, eliminating matching deviations caused by differences in feature dimension and scale; and using a prototype network feature encoding algorithm that is completely from the same source as the fault feature prototype generation to perform dual-dimensional fusion judgment based on boundary exceedance and pattern matching results, and outputting graded alarms.

[0069] The real-time load sequence within the sliding window is subjected to end-to-end feature encoding to generate a real-time feature vector that is completely consistent with the dimension of the fault feature prototype vector. The encoding process simultaneously incorporates the weight adaptation of the current real-time scenario attributes and power grid operating conditions, thereby enhancing the identification of fault features under the same scenario and reducing the probability of mismatch caused by normal load fluctuations across scenarios.

[0070] First, the fault feature prototype library is initially screened based on the current real-time scene labels, prioritizing fault feature prototypes that have an association mapping with the current scene. Then, a full traversal matching is performed on all fault feature prototype vectors in the library. Cosine similarity is used to calculate the spatial similarity between the real-time feature vector and each fault feature prototype vector in the library. The similarity value, the fault type label with the highest matching degree, and the corresponding prototype source information are output for each type of fault. The full matching calculation process is completed within the real-time stream processing framework. The time taken for a single round of matching does not exceed 10% of a single sampling period, strictly meeting the core requirement that the end-to-end processing delay does not exceed one sampling period.

[0071] Based on the similarity calculation results, hierarchical judgment and labeling are performed. When the similarity between the real-time feature vector and a certain fault prototype is ≥90%, the real-time sequence sample is labeled as a high-confidence latent abnormal sample, and the corresponding fault type, abnormal feature interval and core abnormal indicators are locked simultaneously. When the similarity is between 80% and 90%, the sample is labeled as a medium-confidence suspected abnormal sample, and the matched fault type, similarity value and feature deviation item are recorded simultaneously. When the similarity is <80%, the sample is labeled as a normal mode sample and is not included in the output range of abnormal matching results.

[0072] Simultaneously, all feature vectors, similarity calculation results, and hierarchical label data from the matching process are written to a time-series database for persistent storage. This provides complete source data support for subsequent dual-dimensional fusion anomaly detection, alarm review feedback, and fault prototype iterative optimization. At the same time, it ensures that this process is executed in parallel with the baseline incremental update and boundary dynamic calibration processes without interference, thus guaranteeing the continuity and stability of the entire real-time monitoring process.

[0073] The method of using a dual-dimensional fusion of boundary exceedance and pattern matching results to determine and output tiered alarms includes the following specific implementation: Based on the real-time load data corresponding to the anomaly determination boundary after standardization and dynamic calibration within the current sampling period, the first dimension of boundary exceedance determination is performed. The residual values ​​of the real-time load time series data after STL decomposition are compared point by point with the upper and lower limits of the anomaly determination boundary currently in effect for the corresponding scenario. Real-time samples whose residual values ​​exceed the upper and lower limits of the boundary are uniformly marked as boundary exceedance anomaly samples, and the exceedance magnitude, exceedance period, corresponding scenario label, and effective baseline parameters are recorded simultaneously.

[0074] Based on the real-time sequence anomaly pattern matching results completed in parallel, the second dimension of anomaly pattern judgment is performed. Samples marked as high-confidence latent anomalies and medium-confidence suspected anomalies in the matching results are uniformly included in the anomaly sample category of anomaly pattern matching. The corresponding fault type label, cosine similarity value, core anomaly feature interval and fault prototype source information are locked simultaneously.

[0075] Based on the judgment results of the two dimensions, a full-scale dual-dimensional fusion cross-judgment is performed. Three-level hierarchical control is completed according to the anomaly confidence level and risk level. Among them, samples that simultaneously meet the boundary over-limit anomaly mark and the high-confidence latent anomaly pattern matching mark are directly judged as confirmed anomalies, triggering a first-level real-time alarm. The alarm information is accompanied by full-scale anomaly tracing data, which can be directly pushed to the alarm handling module of the power metering and monitoring system without secondary verification.

[0076] Samples that only meet the boundary exceedance anomaly marker and do not trigger the anomaly pattern matching anomaly marker are judged as high-probability anomaly samples to be verified. A secondary verification of the load fluctuation pattern in the same scenario and at the same time is immediately performed. For example, normal load sequence samples of the same time period of the previous 7 natural days in the corresponding scenario are retrieved, and the fluctuation pattern deviation between the sample to be verified and the normal sample set in the same period is calculated. When the deviation is ≥80%, it is judged as a high-probability anomaly and a level 2 alarm is triggered. When the deviation is <80%, it is marked as a sample to be manually reviewed and no alarm is triggered. It is only included in the daily inspection and review list.

[0077] Samples that only meet the medium confidence level suspected anomaly pattern matching mark and do not trigger the boundary over-limit anomaly mark are judged as suspected latent anomalies, triggering a level 3 warning. The warning information is simultaneously pushed to the hidden danger investigation module to identify hidden faults and hidden dangers within the boundary in advance.

[0078] All triggered Level 1 alarms, Level 2 alarms, and Level 3 warnings are simultaneously accompanied by corresponding scenario tags, effective baseline parameters, dynamically calibrated anomaly judgment boundary values, anomaly characteristic data, fault type matching information, full original sampled data, and source tracing information. After being structured and encapsulated, they are synchronously output to the corresponding power metering and monitoring system in milliseconds to achieve real-time push and visualization. At the same time, all alarm and warning samples and corresponding judgment data are completely written into the time-series database for persistent storage, providing complete data source support for subsequent alarm feedback loop and continuous model iteration and optimization.

[0079] Synchronous execution of hard constraint control over alarm trigger frequency. For example, when the number of similar alarm samples triggered in a single sampling period under the same scenario exceeds 0.5% of the total number of samples in that scenario, a batch alarm verification mechanism is immediately triggered, the external push of the batch of alarms is suspended, and scenario matching verification, boundary parameter verification, and pattern matching accuracy verification are automatically performed to investigate the risk of batch false alarms caused by scenario switching, baseline offset, boundary distortion, etc. Alarm push is only executed after the verification is passed. The core indicator requirements of alarm false alarm rate ≤0.5% and full-type anomaly identification recall rate ≥99% are strictly controlled from the judgment stage.

[0080] The process of completing model iteration and optimization based on review feedback to form an operational closed loop includes, in practice: establishing a structured manual review feedback channel that is deeply integrated with the electricity metering and monitoring system; defining the specifications for mandatory fields in the review feedback, including full-dimensional traceability information, such as unique alarm / early warning identifiers, review result classifications, fault types marked by reviewers, descriptions of abnormal root causes, supplementary scenario labels, and review time traceability information. The feedback channel simultaneously supports two operation modes: single-sample review and batch alarm review. The review result data is synchronized to the model iteration and optimization module in real time and synchronously and completely written to the time-series database for persistent storage, ensuring that the data throughout the entire process is traceable and verifiable.

[0081] For samples marked as false alarms, the precise location of the sample's scenario is first completed. The scenario label, effective baseline parameters, and anomaly judgment boundary values ​​corresponding to the sample when the alarm was triggered are matched. After standardized preprocessing, the false alarm sample is included in the normal sample set of the corresponding scenario. Simultaneously, the directional calibration process of the scenario's baseline and boundary is triggered. Based on the newly added normal samples, incremental STL time series decomposition is re-executed to correct the trend and periodic parameters of the baseline. Kernel density estimation is used to refit the residual distribution of the normal samples in the scenario, tightening the boundary thresholds of the corresponding confidence interval. The statistical distribution characteristics of the normal samples in the scenario are updated simultaneously to directionally correct the baseline offset and boundary looseness issues that cause false alarms. At the same time, the false alarm sample is included in the false alarm feature library of the corresponding scenario. For example, when the same type of false alarm in the same scenario is triggered more than 3 times within 7 days, the full baseline offline verification and optimization of the scenario is automatically triggered to avoid repeated triggering of the same type of false alarm.

[0082] For samples marked as confirmed anomalies, a structured match between the anomaly type and the root cause of the fault is first performed. If the fault type corresponding to the sample is a type already included in the fault feature prototype library, the sample is preprocessed with the same source as the real-time streaming data and then included in the small sample library of the corresponding fault type. The feature encoding is completed again using prototype network small sample learning. The feature prototype vector corresponding to the fault type is iteratively updated, and the full traceability information of the prototype is updated synchronously. The iterated prototype must pass the offline test set cross-validation. For example, the hard standard for validation is that the recall rate of the corresponding fault type recognition is ≥99% and the cross-scenario mismatch rate is ≤0.5%. Only after the validation is passed can the original prototype be replaced and added to the library.

[0083] If the fault type corresponding to the sample is a newly added fault type not included in the fault feature prototype library, after standardization preprocessing and structured classification labeling of the sample and related samples of the same type, it is entered into the small sample fault sample library. The feature encoding is completed using the same prototype network algorithm to generate the feature prototype vector of the newly added fault type. After offline test set verification to meet the core performance indicators, it is added to the fault feature prototype library. At the same time, the association mapping between the newly added fault type and the six basic scenarios in the scenario adaptive baseline library is established to improve the ability to include all types of faults.

[0084] For the review results of samples marked as pending review samples and suspected latent abnormal samples of Level 3 warning, if the review confirms that the sample is normal, it is included in the normal sample set of the corresponding scenario after standardized preprocessing to optimize the boundary calibration accuracy of the scenario. If the review confirms that the sample is abnormal, the fault prototype library is iteratively optimized according to the processing procedure for confirmed abnormal samples. At the same time, the sample is included in the abnormal sample verification set of the corresponding scenario for constraint verification of subsequent boundary adjustment.

[0085] A full offline verification and optimization mechanism for the model is set up with a fixed cycle of 7 days. All normal samples, verified abnormal samples, and false alarm samples are extracted within the cycle. The scenario adaptive baseline library is fully optimized. The scenario clustering granularity is optimized based on the newly added normal samples. The baseline and initial anomaly judgment boundary of the newly marked scenarios within the cycle are supplemented. Failed scenario baselines with no matching records or valid sample support for 30 consecutive days are removed. At the same time, the baseline and anomaly judgment boundary of the entire scenario are fully re-verified to ensure that the baseline of the entire scenario accurately adapts to the current load fluctuation pattern.

[0086] Simultaneously perform full verification and optimization on the fault feature prototype library, optimize the scene weight adaptation of the feature encoding algorithm based on the fault samples added within the period, merge duplicate fault prototypes with vector cosine similarity ≥98%, and remove failed prototypes with consistently unsatisfactory matching recall rates for two consecutive verification periods, thereby improving the accuracy and efficiency of prototype matching.

[0087] Establish a model iteration parameter admission verification mechanism. All targeted update parameters triggered by review feedback, baseline parameters generated by periodic optimization, boundary parameters, and fault prototype vectors must pass full verification with a dedicated offline test set. The offline test set must completely include 6 major categories of basic scenarios, all types of fault samples, historical false alarm samples, and extreme working condition samples. The hard standards for passing the verification are: alarm false alarm rate ≤ 0.5%, full-type anomaly identification recall rate ≥ 99%, and end-to-end processing latency ≤ 1 sampling period. Update parameters that do not meet the standards will not be implemented online and will be directly rolled back to the stable parameters of the previous version. At the same time, parameter optimization deviation analysis will be triggered to locate abnormal interference factors in the optimization process.

[0088] Establish a real-time performance monitoring mechanism for the entire model lifecycle, continuously monitoring the model's core metrics, including scene matching accuracy, false alarm rate, anomaly recognition recall rate, boundary adjustment compliance rate, and prototype matching accuracy. When a core metric fails to meet the preset standard for three consecutive verification cycles, the model is automatically retrained. Based on all historical data and newly added verification and annotation data, the entire process of scene labeling, unsupervised clustering to select normal samples, construction of a full-scene baseline, and generation of fault feature prototypes is completed. After passing full verification, the model version iteration is completed. At the same time, all historical version model parameters and performance data are retained, forming a traceable, rollbackable, and continuously optimized closed-loop operation to ensure the long-term stability and recognition accuracy of the model.

[0089] The original incremental STL time series decomposition and online kernel density estimation technology combination in this application has three major shortcomings in the field of power time series monitoring: First, incremental STL is prone to mode aliasing for non-stationary and strongly abrupt load time series (extreme weather, sudden operating conditions), and baseline updates are lagging; second, online KDE has a large fitting deviation for small sample data distributions in a 24-hour sliding window, and the judgment logic for three consecutive windows further exacerbates the boundary calibration lag; third, it cannot adaptively balance the weights of historical baseline patterns and new load characteristics, which easily leads to baseline drift or underfitting.

[0090] Therefore, this application proposes a new boundary calibration method, which is customized for power metering load fluctuation scenarios and significantly improves baseline update accuracy and the real-time performance and robustness of boundary calibration, including:

[0091] Input data:

[0092] Within the current 24-hour sliding window, the normal load time series sample set after initial screening is as follows: , Sampling time, ;

[0093] The baseline modal parameters output by the previous sliding window: ( The modal number is explicitly defined and determined based on the load time-series energy ratio method. , For trend items, It is a periodic term. (for residual terms)

[0094] The conjugate prior hyperparameter of the residual distribution of the previous window: (initial value) Explicitly defined: , (Based on historical normal sample statistics)

[0095] Memory decay coefficient: (Dimensionless, fixed empirical value, dynamic response speed adapted to load fluctuations).

[0096] Then, construct the variational optimization objective function (correcting fuzzy terms and supplementing constraint logic):

[0097] ;

[0098] Constraints: ;

[0099] Parameter definition:

[0100] : No. The intrinsic mode functions (IMFs) obtained from the decomposition are dimensionless. (Trend item) (Periodic term) (residual term) );

[0101] : No. The center frequency of each mode is dimensionless (due to input normalization) and characterizes the oscillation period of the mode.

[0102] The Dirac function is dimensionless and used to generate analytic signals.

[0103] Imaginary unit ( ;

[0104] Convolution operation;

[0105] Regarding time Partial derivative operations;

[0106] The square of the norm is used to measure modal energy and is dimensionless.

[0107] Previous window The components of each mode (from) ), dimensionless;

[0108] ;

[0109] in The relative rate of change (dimensionless) of the standard deviation of the residuals between the current window and the previous window is calculated using the following formula: ;

[0110] The standard deviation of the current window residuals. The standard deviation of the residuals from the previous window. (Dimensionless, empirical value).

[0111] The above variational problem is solved using the alternating direction method of multipliers (ADMM):

[0112] 1. Initialize modal components Center frequency Lagrange multipliers 2. Iterative Updates , , until the convergence condition is met. 3. After iterative convergence, output the modal components of the current window. .

[0113] Output baseline results:

[0114] Baseline (Dimensionless): Obtained by combining the trend term and the period term, formula:

[0115] ;

[0116] residuals (Dimensionless): The difference between the load sample and the baseline, formula:

[0117] ;

[0118] Update modal parameters (For the next window iteration).

[0119] Then, residual distribution modeling is performed:

[0120] Set residual terms Follows a normal distribution: ,in The mean of the residuals (dimensionless) The NIG distribution is selected as the residual variance (dimensionless). and The conjugate prior (ensuring the posterior distribution is of the same type as the prior distribution):

[0121] ;

[0122] Then, sequentially update the hyperparameters:

[0123] Based on the current window residual samples Sequential update hyperparameters The formulas and parameters have the following meanings:

[0124] ;

[0125] ;

[0126] ;

[0127] ;

[0128] Parameter definition:

[0129] The mean of the current window residuals (dimensionless) is calculated using the following formula:

[0130] : Current window sample size (fixed value, dimensionless);

[0131] The mean (dimensionless) of the updated residual distribution represents the overall shift of the residuals.

[0132] mean The confidence parameter (dimensionless) has a larger value, and the more reliable the mean estimate is.

[0133] Shape parameters (dimensionless) of the inverse gamma distribution (IG), and control variance. The degree of concentration of distribution;

[0134] The scaling parameter (dimensionless) and variance of the inverse gamma distribution (IG) The mean is positively correlated; : Hyperparameters updated in the previous window (from ).

[0135] Then, the anomaly detection boundary is dynamically calibrated: the variance of the residual posterior distribution is calculated from the updated hyperparameters. (dimensionless) (To ensure positive variance) Based on residual fluctuations, dynamically calibrate the boundary (clearly define quantile values ​​and connect hard constraints):

[0136] when (Current load fluctuations are increasing): The boundary is widened to a 99% confidence interval, formula:

[0137] ;

[0138] in (99.5th quantile of standard normal distribution, dimensionless);

[0139] when (Current load fluctuations are narrowing): The boundary is tightened to the 99.5% confidence interval, formula:

[0140] ;

[0141] in (99.75th quantile of standard normal distribution, dimensionless).

[0142] Then proceed with boundary hard constraints: retain the original boundary and adjust the hard constraints (adjustment range per transaction). Daily cumulative adjustment range (etc.), the above calibrated , Substitute the constraints; if the value exceeds the constraint range, adjust according to the maximum allowable range to ensure that the boundary does not shift or become distorted.

[0143] Then proceed and output the result:

[0144] Updated scene baseline: ;

[0145] Anomaly detection boundary after dynamic calibration: upper boundary Lower boundary (dimensionless);

[0146] Updated modal parameters: (For use in the next window iteration);

[0147] Updated hyperparameters: (Used for updating the residual distribution in the next window).

[0148] Figure 6 A bar chart comparing the false alarm rates of the three methods under the six core power consumption scenarios defined by the patent is presented. The horizontal axis represents the load operation scenario, and the vertical axis represents the false alarm rate. The red dashed line represents the upper limit of the false alarm rate of this patent. The results show that the false alarm rate of the method of this invention is ≤0.42% in all scenarios, fully meeting the hard indicators. Even in scenarios with strong load fluctuations such as extreme weather and peak summer, the false alarm rate is still stably controlled within 0.5%. In contrast, the false alarm rate of existing technologies soars significantly in complex scenarios. The false alarm rate of the traditional fixed threshold method reaches 3.12% under extreme weather conditions. This chart intuitively verifies the core innovative value of the invention's scenario labeling, scenario-based baseline library, and dynamic boundary calibration. It completely solves the industry pain points of traditional fixed thresholds, such as missed alarms with strict thresholds and false alarms with wide thresholds, demonstrating the excellent adaptability and anti-false alarm performance of the solution across all scenarios.

[0149] This application also provides an embodiment of an electronic device. The electronic device is manifested in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: one or more processors or processing units, memory, and buses connecting different components (including memory and processing units).

[0150] A bus refers to one or more of several bus architectures, including memory buses or memory controllers, peripheral buses, graphics acceleration ports, processors, or local buses using any of the various bus architectures. Examples of these architectures include, but are not limited to, Industry Standard Architecture (ISA) buses, Micro Channel Architecture (MCA) buses, Enhanced ISA buses, Video Electronics Standards Association (VESA) local buses, and Peripheral Component Interconnect (PCI) buses.

[0151] Electronic devices typically include a variety of computer-readable media. These media can be any available media that can be accessed by the electronic device, including volatile and non-volatile media, and removable and non-removable media.

[0152] The memory may include computer-readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. Electronic devices may further include other removable / non-removable, volatile / non-volatile computer device storage media. By way of example only, the storage system may be used to read and write non-removable, non-volatile magnetic media.

[0153] The electronic device can also communicate with one or more external devices (e.g., keyboard, pointing device, camera, etc.), may include a display, and may communicate with one or more devices that enable a user to interact with the electronic device, and / or with any device that enables the electronic device to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed via an input / output (I / O) interface. Furthermore, the electronic device can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN) and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the electronic device via a bus. The processor executes various functional applications and data processing by running programs stored in memory, such as implementing the real-time monitoring method for power metering load fluctuations provided in the above embodiments of the present invention.

[0154] This application also discloses a real-time monitoring system for electricity load fluctuations, including:

[0155] The data synchronization module is used to synchronize the access of three types of data and unify the sampling period;

[0156] The data preprocessing module is used to complete missing values, remove bad data, standardize features, and output compliant feature sequences.

[0157] The scene tagging module is used to classify scenes based on historical data;

[0158] The baseline construction module is used to screen normal samples, build a benchmark baseline, fit the initial abnormal boundary, and form a scenario-based baseline library; the baseline update module is used to match the working condition loading baseline, incrementally update the baseline and dynamically calibrate the abnormal boundary, and set boundary adjustment hard constraints.

[0159] The fault prototype library module is used to build a small sample fault prototype library;

[0160] The anomaly detection module is used to match anomaly patterns, fuse boundary exceedance and pattern matching results, and output hierarchical alarms.

[0161] The model iteration module is used to iteratively optimize the model based on verification feedback, forming a closed loop of operation.

[0162] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time monitoring of load fluctuations in electricity metering, characterized in that, Includes the following steps: Synchronously access three types of data: load time sequence, scenario attributes, and power grid operating conditions, unify the sampling period, complete missing value completion, bad data removal and feature standardization, and output compliant feature sequences; Based on historical standardized data, scene labeling was completed, normal samples were screened through unsupervised clustering, baselines for each scene were constructed through time-series decomposition, and initial anomaly judgment boundaries were calculated to form a scene baseline library. Real-time matching of the current working condition and corresponding scenario and loading of baseline; based on normal samples within the sliding window, unsupervised incremental update of baseline, dynamic calibration of anomaly judgment boundary, and setting hard constraints for boundary adjustment. A small sample fault prototype library is established, and real-time sequence anomaly pattern matching is performed based on the small sample fault prototype library. The judgment is based on the dual-dimensional fusion of boundary exceedance and pattern matching results, and a graded alarm is output.

2. The method for real-time monitoring of load fluctuations in electricity metering according to claim 1, characterized in that, The synchronous access to three types of data—load time series, scenario attributes, and power grid operating conditions—includes: load time series data, which includes electrical-related data collected in real time by the power metering device; scenario attribute data, which includes meteorological data and time-dimensional feature data; and power grid operating condition data, which includes power grid operating condition data.

3. The method for real-time monitoring of load fluctuations in electricity metering according to claim 2, characterized in that, The process of classifying scenarios based on historical standardized data includes: extracting historical standardized feature data from the time series database, using scenario attribute features as the basis for classification, performing dimensional labeling on the historical data, classifying basic electricity consumption and operating conditions scenarios, and establishing a corresponding relationship between scenarios and historical load fluctuation data.

4. The method for real-time monitoring of load fluctuations in electricity metering according to claim 2, characterized in that, Normal samples are screened through unsupervised clustering, including: for the historical load sequences under each single basic scenario that have been divided, unsupervised clustering is used to carry out clustering processing, and samples within the core area of ​​the cluster are selected as the normal load sample set of the corresponding scenario, while outlier abnormal samples at the edge of the cluster are removed.

5. The method for real-time monitoring of load fluctuations in electricity metering according to claim 2, characterized in that, The process of fitting and calculating the initial anomaly detection boundary to form a scenario-based baseline library includes: performing time-series decomposition on the normal load sample set screened by unsupervised clustering under each scenario, decomposing the load time-series data into multiple time-series components, fitting the probability distribution of normal load samples based on the residual terms of the normal sample set, calculating the upper and lower limits of the corresponding confidence intervals as the initial anomaly detection boundary for the corresponding scenario, and then storing the scenario label, baseline component parameters, initial anomaly detection boundary and normal sample statistics into the scenario adaptive baseline library to form a scenario-based baseline library.

6. The method for real-time monitoring of load fluctuations in electricity metering according to claim 1, characterized in that, The real-time matching of the current operating condition and loading of the baseline includes: extracting scene attributes, power grid operating conditions and time dimension features from the standardized feature stream of the current sampling period in real time, generating real-time scene labels, performing multi-dimensional matching degree calculation with the full scene labels in the scene adaptive baseline library, the matching dimensions include time attributes, meteorological conditions and power grid operating conditions, the weight of each dimension is executed according to preset rules, and the corresponding scene baseline and boundary are determined according to different matching results, and the scene matching result, effective baseline parameters and abnormal judgment boundary values ​​are output.

7. The method for real-time monitoring of load fluctuations in electricity metering according to claim 6, characterized in that, Based on normal samples within a sliding window, the baseline is updated incrementally without supervision, and the anomaly detection boundary is dynamically calibrated. This includes: setting a sliding update window with a step size consistent with the load data sampling period; performing initial screening on real-time load samples within the window based on the currently effective anomaly detection boundary, and including samples that do not exceed the boundary into the normal incremental update sample set to exclude interference from samples that exceed the limit; disassembling the normal incremental sample set and performing unsupervised incremental updates on the baseline based on the disassembled components to correct the baseline deviation; fitting the probability distribution of the residual terms of normal samples and dynamically calibrating the upper and lower limits of the anomaly detection boundary based on the change in the bandwidth of the residual distribution.

8. The method for real-time monitoring of load fluctuations in electricity metering according to claim 7, characterized in that, Real-time sequence anomaly pattern matching is performed based on a small sample fault prototype library. This includes: real-time extraction of load time series sequences and related features that are homologous to the baseline incremental update and have undergone standardization, ensuring that they are consistent with the sample processing rules of the fault prototype library; generating real-time feature vectors using homologous encoding and incorporating scenario and working condition weights, first screening the prototype library in a scenario-based manner, and then performing traversal matching through cosine similarity to output similarity and corresponding fault information.

9. The method for real-time monitoring of load fluctuations in electricity metering according to claim 1, characterized in that, The method of determining the boundary exceedance based on the dual-dimensional fusion of boundary exceedance and pattern matching results and outputting hierarchical alarms includes: performing boundary exceedance determination based on real-time standardized load data, performing abnormal pattern determination simultaneously based on abnormal pattern matching results, implementing three-level hierarchical alarms according to abnormal confidence and risk level, pushing alarm information along with traceability data to the monitoring system and storing it, and simultaneously implementing hard constraints on alarm frequency.

10. A system utilizing the real-time monitoring method for power metering load fluctuations as described in claim 1, characterized in that, include: The data synchronization module is used to synchronize the access of three types of data and unify the sampling period; The data preprocessing module is used to complete missing values, remove bad data, standardize features, and output compliant feature sequences. The scene tagging module is used to classify scenes based on historical data; The baseline construction module is used to screen normal samples, build benchmark baselines, fit initial abnormal boundaries, and form a scenario-based baseline library. The baseline update module is used to match the loading baseline under the working condition, incrementally update the baseline and dynamically calibrate abnormal boundaries, and set hard constraints for boundary adjustment. The fault prototype library module is used to build a small sample fault prototype library; The anomaly detection module is used to match anomaly patterns, fuse boundary exceedance and pattern matching results, and output hierarchical alarms. The model iteration module is used to iteratively optimize the model based on verification feedback, forming a closed loop of operation.

Images