Methods and Systems for Analyzing Anomalies in Electricity Meters

Abstract

This invention relates to the field of electricity meter metering technology, and discloses a method and system for electricity meter metering anomaly analysis. The method includes: collecting data such as voltage waveforms to generate a metering feature vector set and constructing an anomaly detection rule base; setting a set of scenario parameter types and establishing an anomaly association judgment model; dynamically correcting thresholds based on the model to generate an optimized error threshold set and an anomaly triggering condition set; updating the metering analysis strategy to generate an anomaly judgment scheme and correcting the data verification timing logic. The system includes modules for data acquisition, anomaly rule base construction, scenario parameter configuration, association model training, dynamic threshold optimization, strategy updating, and timing correction. This invention improves the accuracy and adaptability of electricity meter metering anomaly detection through multi-dimensional data modeling, scenario-based threshold configuration, and dynamic timing calibration, and is suitable for metering anomaly analysis in diverse electricity consumption scenarios in smart grids.

Description

Technical Field

[0001] This invention relates to the field of electricity meter measurement technology, specifically to a method and system for analyzing electricity meter measurement anomalies. Background Technology

[0002] In the construction and operation of smart grids, electricity meters, as the core equipment for electricity metering, directly affect the economic benefits of power companies, fair electricity use for users, and the stable operation of the power grid. With the increasing complexity of power systems and the diversification of electricity consumption scenarios, such as residential electricity consumption, industrial production electricity consumption, and commercial complex electricity consumption, the electricity consumption characteristics under different scenarios vary significantly. This places higher demands on the metering accuracy of electricity meters under various complex operating conditions.

[0003] Traditional methods for analyzing electricity meter anomalies typically employ fixed threshold detection, lacking adaptability to dynamic electricity usage scenarios. For example, in industrial settings with significant load fluctuations, fixed threshold values ​​are insufficient to accurately capture anomalies caused by load changes. Furthermore, in special electricity environments with harmonic interference, traditional methods cannot effectively identify the impact of harmonics on metering characteristics such as voltage waveforms and current phases, easily leading to misjudgments or missed detections. Moreover, existing technologies, when processing multi-dimensional metering data, often lack in-depth analysis of the temporal correlations and feature coupling relationships between data points, making it difficult to establish comprehensive and accurate anomaly detection models, resulting in insufficient timeliness and accuracy in anomaly analysis.

[0004] Meanwhile, with the increasing frequency and surge in data collection from electricity meters in smart grids, the limitations of traditional methods in terms of data processing efficiency and dynamic threshold adjustment are becoming increasingly apparent. Fixed detection rules and thresholds cannot be dynamically optimized based on real-time operational data, making it difficult to adapt to real-time changes in grid operating conditions and resulting in insufficient detection capabilities for emerging anomaly patterns. Therefore, there is an urgent need for an electricity meter anomaly analysis method and system that can combine multi-dimensional metering data, adapt to different electricity consumption scenarios, and achieve dynamic threshold optimization and time-series logic calibration, in order to improve the accuracy, timeliness, and adaptability of metering anomaly detection and ensure the safe and reliable operation of smart grids. Summary of the Invention

[0005] The purpose of this invention is to provide a method for analyzing anomalies in electricity meter readings, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for analyzing metering anomalies in electricity meters, the method comprising: Collect voltage waveforms, current phases, and power pulse data within multiple electricity meter metering cycles to generate a set of metering feature vectors; An anomaly detection rule base is constructed based on the set of measurement feature vectors; Set the threshold types of metering parameters for different electricity consumption scenarios to form a set of scenario parameter types; Based on the set of scenario parameter types, error distribution, fluctuation frequency and abnormal trigger records in historical measurement data are obtained, and an abnormal correlation judgment model is established. Extract the threshold types of metering parameters corresponding to the current electricity meter operation scenario, combine them with the anomaly correlation judgment model to perform dynamic threshold correction, and generate an optimized error threshold set and anomaly triggering condition set. The measurement analysis strategy is updated based on the optimized error threshold set, anomaly triggering condition set, and anomaly detection rule base to generate an anomaly determination scheme; the timing logic of data verification in the anomaly determination scheme is synchronously corrected.

[0007] Preferably, the generation of the measurement feature vector set includes: Define multiple metering characteristic dimensions, including voltage distortion coefficient, current imbalance, and cumulative power deviation; Collect real-time sampling data corresponding to each dimension and perform standardization processing to generate a standardized measurement feature matrix; The standardized quantitative feature matrix is ​​input into the anomaly detection model generation module, and an anomaly detection rule base is constructed through a pre-set clustering algorithm.

[0008] Preferably, the establishment of the anomaly correlation determination model includes: Set parameter types for different power consumption scenarios, including load fluctuation range, harmonic content range, and metering error tolerance; Based on the set of scene parameter types, obtain the error distribution characteristics, data fluctuation cycle and alarm event records when historical anomalies are triggered, and construct a historical anomaly feature matrix, a historical fluctuation dataset and an alarm tag set. An initial anomaly correlation model is trained using the historical anomaly feature matrix, historical fluctuation dataset, and alarm label set to generate an anomaly correlation determination model.

[0009] Preferably, the initial anomaly association model for training includes: An initial anomaly correlation model is established based on the set of scenario parameter types and the dimensions of measurement features; Input the historical anomaly feature matrix and historical fluctuation dataset item by item into the initial anomaly association model to generate a set of historical alarm prediction results. Set the error judgment tolerance range and calculate the matching deviation between the historical alarm prediction result set and the alarm tag set; When the matching deviation exceeds the tolerance range, the density clustering algorithm is used to iteratively optimize the judgment boundary of the initial abnormal association model until the deviation reaches the preset standard, thus forming an abnormal association judgment model.

[0010] Preferably, the density clustering algorithm employs an adaptive neighborhood radius adjustment mechanism.

[0011] Preferably, the dynamic threshold correction includes: Obtain the real-time fluctuation characteristics and error distribution of the current electricity meter's metering parameters, and generate a dynamic parameter feature set; Set anomaly detection and response criteria, input the dynamic parameter feature set into the anomaly association detection model, and calculate the current anomaly probability value; When the current anomaly probability value exceeds the response standard, the error threshold set is updated using a sliding window to generate an optimized error threshold set.

[0012] Preferably, the sliding window update includes: Create a threshold optimization queue and define the window length and sliding step parameters; Extract the statistics of the dynamic parameter feature set on the time series and generate a feature rolling average matrix; Construct a threshold adjustment function based on the anomaly probability value; The historical threshold data in the queue is updated by a window sliding mechanism to generate an optimized error threshold set.

[0013] Preferably, the anomaly detection scheme includes: The optimized error threshold set and anomaly triggering condition set are loaded into the metrology analysis unit to generate an anomaly determination scheme. Identify the time dependency of the data verification process in the anomaly detection scheme and perform logical calibration based on the device clock synchronization mechanism.

[0014] Preferably, the logic calibration employs a timestamp alignment algorithm, which dynamically adjusts the verification trigger timing based on the data acquisition frequency and transmission delay.

[0015] Preferably, the present invention also includes an electricity meter metering anomaly analysis system, the system comprising the following modules: The data acquisition module is used to collect voltage waveforms, current phases, and power pulse data within multiple electricity meter metering cycles, and generate a set of metering feature vectors. The anomaly rule base construction module trains an anomaly detection rule base based on the measured feature vector set and generates initial anomaly judgment rules through feature clustering and pattern matching. The scenario parameter configuration module allows you to set the threshold types of metering parameters for different electricity consumption scenarios, forming a set of scenario parameter types that includes voltage deviation, phase shift, and power pulse frequency. The association model training module extracts error distribution, fluctuation frequency and abnormal trigger records from historical measurement data based on the scenario parameter type set, and establishes an abnormal association judgment model through a time-series association algorithm. The dynamic threshold optimization module matches the corresponding metering parameter threshold type according to the current electricity meter operating scenario, calls the anomaly association judgment model to dynamically correct the error threshold, and generates an optimized error threshold set and an anomaly triggering condition set. The strategy update module integrates the optimized error threshold set, the abnormal triggering condition set, and the abnormal detection rule base to generate an abnormal judgment scheme that includes data verification logic. The timing correction module synchronously calibrates the timing logic of data verification in the anomaly determination scheme to ensure the dynamic consistency between the anomaly detection rules and the correction threshold in each measurement cycle.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Through multi-dimensional data acquisition and feature modeling, key metering features such as voltage waveforms, current phases, and power pulses during the operation of electricity meters can be comprehensively captured. An anomaly detection rule base is constructed through standardized processing and clustering algorithms, enabling effective identification of complex metering anomaly patterns. By introducing a set of scenario parameter types, differentiated metering parameter threshold types are set for different electricity usage scenarios (such as load fluctuation range and harmonic content range). An anomaly correlation judgment model is established by combining error distribution, fluctuation frequency, and anomaly trigger records from historical metering data. This allows the system to deeply analyze the temporal correlation and causal relationship between metering parameters and anomaly events in different scenarios, significantly improving the targeting and accuracy of anomaly detection.

[0017] The dynamic threshold correction mechanism acquires the fluctuation characteristics and error distribution of metering parameters in real time, calculates the anomaly probability value by combining it with an anomaly correlation judgment model, and dynamically adjusts the error threshold using a sliding window update strategy. This achieves adaptive matching between the threshold setting and the real-time operating status of the power grid, effectively solving the problem of insufficient adaptability of traditional fixed threshold detection methods to dynamic scenarios and reducing the false positive and false negative rates. Simultaneously, a timestamp alignment algorithm synchronously corrects the temporal logic of data verification, ensuring the dynamic consistency between the anomaly detection rules and the correction threshold within each metering cycle, thus improving the timeliness and reliability of anomaly analysis.

[0018] Furthermore, the collaborative work of various modules within the system (such as the data acquisition module, the anomaly rule base construction module, and the scenario parameter configuration module) forms a complete closed loop from data acquisition, feature modeling, model training to strategy updates, achieving full-process automation and intelligence in metering anomaly analysis. This method and system not only effectively address the diverse metering anomaly detection needs in existing smart grids across various electricity consumption scenarios, but also provide reliable technical support for the future development of power grids towards higher levels of intelligence, demonstrating significant economic and social benefits. Attached Figure Description

[0019] Figure 1This is a schematic diagram illustrating the working principle of the electricity meter metering anomaly analysis method described in this invention. Figure 2 Design diagrams generated for the quantitative feature vector set; Figure 3 Design diagram for sliding window-style updates; Figure 4 This is a system block diagram for analyzing anomalies in electricity meter readings. Detailed Implementation

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

[0021] Please see Figures 1-4 The method for analyzing metering anomalies involved in this invention comprises the following specific steps: Voltage waveforms, current phases, and power pulse data are collected over multiple electricity metering cycles to generate a set of metering feature vectors. Sensors or data acquisition devices deployed at the electricity meters collect various data types in real time at a preset sampling frequency (e.g., 100 times per second), covering at least 10 consecutive metering cycles (e.g., each cycle is one month), ensuring the data has temporal continuity and statistical significance. After collection, the raw data is preprocessed to remove obviously abnormal noise data, forming an initial dataset. Feature engineering is then used to generate a vector set containing multi-dimensional features.

[0022] An anomaly detection rule base is constructed based on the set of measurement feature vectors. Using the feature vectors in the initial dataset, data mining techniques are employed to distinguish between normal measurement patterns and abnormal patterns, extracting key feature combinations and corresponding rules that can characterize abnormal states. For example, when voltage waveform distortion exceeds a certain benchmark value and current phase offset reaches a specific range, it is determined to be a potential anomaly. Such rules are structured and stored to form a rule base.

[0023] Define the threshold types of metering parameters corresponding to different electricity consumption scenarios to form a set of scenario parameter types. Analyze the load characteristics of typical electricity consumption scenarios (such as residential daily electricity consumption, commercial peak electricity consumption, and industrial inductive load electricity consumption), and define the parameter threshold types that match each scenario, including the allowable range of voltage deviation, the upper limit of current phase fluctuation, and the normal range of power pulse frequency, to form a set of parameter types covering multiple scenarios.

[0024] Based on the set of scenario parameter types, error distribution, fluctuation frequency, and abnormal trigger records from historical measurement data are obtained to establish an abnormal correlation judgment model. Measurement data for each scenario are retrieved from the historical database, and the probability of error occurrence, frequency of fluctuation, and actual triggered abnormal alarm records under different parameter combinations are statistically analyzed. By analyzing the correlation between data, a model that can reflect the causal relationship between parameter changes and abnormal events is constructed to predict the probability of anomalies under the current parameter state.

[0025] The system extracts the threshold types of metering parameters corresponding to the current electricity meter operating scenario, and performs dynamic threshold correction based on the anomaly correlation judgment model to generate an optimized set of error thresholds and anomaly triggering conditions. It identifies the current electricity consumption scenario of the electricity meter in real time (e.g., through load curve pattern matching), calls the initial threshold parameters corresponding to that scenario, and simultaneously inputs the currently collected metering data into the anomaly correlation judgment model to calculate the impact of parameter fluctuations on anomaly occurrence. Based on this, it adjusts the initial thresholds to form optimized thresholds and triggering conditions adapted to real-time operating conditions.

[0026] The metrological analysis strategy is updated based on the optimized error threshold set, anomaly triggering condition set, and anomaly detection rule base to generate an anomaly determination scheme. The timing logic of data verification in the anomaly determination scheme is then synchronously corrected. The optimized thresholds and conditions are integrated with the existing rule base, and the judgment logic and priority in the metrological analysis process are adjusted to form a new anomaly determination scheme. Simultaneously, the timing of verification triggers and the execution order are calibrated based on a clock synchronization mechanism to ensure the temporal consistency of the detection logic, taking into account the time sequence of each stage in the data verification process.

[0027] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0028] Example 1: In generating the metering feature vector set, several metering feature dimensions are first defined, including voltage distortion coefficient, current imbalance, and cumulative power deviation. The voltage distortion coefficient is obtained by calculating the difference between the actual voltage waveform and the standard sine wave, reflecting the degree of voltage quality distortion; the current imbalance is calculated based on the difference rate of the effective values ​​of the three-phase current, used to measure the balance of the three-phase load; and the cumulative power deviation is determined by comparing the difference between the actual cumulative power and the theoretically calculated power, reflecting the accuracy of the metering results.

[0029] After collecting real-time sampling data for each dimension, standardization is performed to eliminate the influence of dimensions. Specific operations include: normalizing the original voltage distortion coefficient value, mapping it to the [0,1] interval; standardizing the current imbalance using Z-score, making its mean 0 and standard deviation 1; and scaling the cumulative power deviation proportionally, converting it to a percentage relative to the rated power. After standardization, the data for each dimension are arranged in a time series, forming a standardized metrological feature matrix containing timestamps and various eigenvalues. Each row of the matrix corresponds to a multi-dimensional feature vector at a sampling time.

[0030] The standardized metrological feature matrix is ​​input into the anomaly detection model generation module, and an anomaly detection rule base is constructed using a pre-defined clustering algorithm. The clustering algorithm employs DBSCAN (density-based spatial clustering of applications with noise), which, by setting a density threshold and neighborhood radius, divides density-connected sample points in the feature space into the same cluster, where anomalies are identified as isolated points in low-density regions. Based on the clustering results, the central feature value and boundary range of each cluster are extracted to generate anomaly detection rules for different feature combinations. For example, when the voltage distortion coefficient is greater than 0.3 and the current imbalance exceeds 15%, it is determined to be a voltage and current anomaly cluster, triggering the corresponding detection rule.

[0031] In constructing the anomaly detection rule base, the parameters of the DBSCAN algorithm are dynamically adjusted. First, based on the distribution characteristics of historical data, the density threshold and neighborhood radius parameters are initialized. Then, during the model training phase, a grid search combined with cross-validation is used to find the optimal parameter combination within a preset parameter range. Specifically, the historical dataset is divided into training and validation sets. For each parameter combination, the model is trained on the training set, and its performance is evaluated on the validation set. The parameter combination that achieves the best balance between anomaly detection accuracy and recall is selected.

[0032] During feature vector generation, the collected raw data undergoes quality assessment and preprocessing. Quality assessment includes data integrity checks, reasonableness verification, and outlier detection. Missing data points are filled using linear interpolation or Kalman filtering; outliers significantly exceeding reasonable ranges are identified and corrected through sliding window statistical analysis. The preprocessing stage also includes data smoothing, employing moving average filtering or wavelet transform to remove high-frequency noise and preserve true signal variation characteristics.

[0033] Based on the standardized econometric feature matrix, feature selection and dimensionality reduction are performed. First, the correlation matrix between features is calculated to identify highly correlated feature pairs. For feature pairs with a correlation exceeding a preset threshold (e.g., 0.8), features with higher information entropy are retained, while redundant features are removed. Then, Principal Component Analysis (PCA) is used for feature dimensionality reduction, mapping the high-dimensional feature space to a low-dimensional subspace, reducing computational complexity while preserving the main information of the original data. The dimensionality-reduced feature matrix is ​​used as input to further improve the training efficiency and generalization ability of the anomaly detection model.

[0034] The anomaly detection rule base update mechanism includes two modes: periodic update and triggered update. In periodic update mode, the system collects new econometric data at preset time intervals (e.g., weekly), retrains the anomaly detection model, and updates the parameters and boundary conditions in the rule base. In triggered update mode, when the system detects multiple consecutive anomalies or frequent occurrences of specific types of anomalies, it automatically initiates the rule base update process, analyzes the matching degree between the current anomaly pattern and existing rules, and adjusts relevant rules or adds new detection rules accordingly.

[0035] When constructing the anomaly detection rule base, expert knowledge is incorporated for rule optimization. Experts in the power system field are invited to evaluate and revise the initially generated rules, supplementing them with judgment rules based on their professional experience. For example, experts may add anomaly detection rules targeting specific equipment faults or operating states based on the operating characteristics of the power system, improving the comprehensiveness and practicality of the rule base. Simultaneously, a rule conflict resolution mechanism is established. When different rules produce contradictory judgments on the same anomaly, a comprehensive judgment is made based on the priority and confidence level of the rules, ensuring the consistency of anomaly detection results.

[0036] In the feature vector generation process, the impact of different electricity consumption scenarios on the features is considered. Feature extraction models are established separately for different scenarios such as residential, commercial, and industrial electricity consumption. For example, in industrial electricity consumption scenarios, the focus is on features such as harmonic content and load abrupt changes; in residential electricity consumption scenarios, the emphasis is on the periodicity and stability of electricity consumption patterns. By differentiating scenarios, the ability of feature vectors to represent abnormal states is improved, enhancing the targeting and accuracy of anomaly detection.

[0037] Regarding the management of the anomaly detection rule base, a rule version control mechanism has been established. Each updated rule base is version-marked, recording the update time, content, and reason. If a new rule causes a performance degradation, a quick rollback to a previous stable version is possible. Simultaneously, an audit log for the rule base is established, recording metrics such as rule usage frequency, trigger count, and false positive rate, providing data support for continuous rule optimization.

[0038] An adaptive standardization method is employed during the standardization process. Standardization parameters are dynamically adjusted based on the statistical characteristics of real-time data, enabling the standardization process to adapt to changes in data distribution. For example, when significant changes in the mean and variance of the data are detected, the standardization parameters are recalculated to ensure that the standardized data possesses good distribution characteristics, thereby improving the effectiveness of subsequent cluster analysis.

[0039] In the feature vector generation process, a time dimension feature extraction is added. In addition to extracting static features at the current moment, dynamic features such as the rate of change and trend are also calculated. For example, the trend of voltage distortion coefficient over the past 10 sampling periods is calculated to determine whether it is gradually increasing or stabilizing. By introducing time dimension features, potential abnormal trends can be detected earlier, improving the timeliness of anomaly detection.

[0040] In constructing the anomaly detection rule base, an ensemble learning approach is employed. The DBSCAN algorithm is combined with other clustering algorithms (such as K-means and hierarchical clustering) to perform multi-faceted analysis of the same dataset. By fusing the clustering results of different algorithms, more comprehensive and accurate anomaly detection rules are generated. For example, when multiple algorithms classify a data point as an anomaly, the confidence level of that anomaly classification is increased; when different algorithms disagree, data features are further analyzed to determine a more reasonable classification result.

[0041] In the feature vector generation process, multi-scale analysis is performed on the data. Wavelet transform is used to decompose the original data into sub-signals of different frequencies, and features are extracted at each scale. Features at different scales can reflect the changes in data at different time granularities; for example, high-frequency features reflect instantaneous fluctuations in data, while low-frequency features reflect long-term trends. By fusing multi-scale features, the ability of the feature vector to represent abnormal states is improved, and different types of anomalous events can be captured.

[0042] In the application of the anomaly detection rule base, an anomaly level classification mechanism is established. Based on the severity and scope of impact of anomalies, they are divided into different levels (e.g., Level 1, Level 2, and Level 3). Different levels of anomalies trigger different response mechanisms. For example, Level 1 anomalies trigger an emergency alarm and activate backup power; Level 2 anomalies trigger an early warning and notify operations and maintenance personnel; and Level 3 anomalies are only logged for subsequent analysis. This anomaly level classification enables tiered handling of anomalies, improving system response efficiency and resource utilization.

[0043] Example 2: When establishing an anomaly correlation judgment model, it is first necessary to set parameter types according to the characteristics of the electricity consumption scenario, including load fluctuation range, harmonic content range, and metering error tolerance. Taking the residential electricity consumption scenario as an example, considering its daily electricity consumption regularity, the load fluctuation range is set to 20%-80% of the rated power. Considering the nonlinear characteristics of household appliances, the harmonic content is limited to a total harmonic distortion rate of less than 5%, and based on national standards and metering accuracy requirements, the metering error tolerance is determined to be ±2%. For industrial motor load scenarios, due to the frequent start-stop of production equipment and large differences in power demand, the load fluctuation range is extended to 10%-100% of the rated power, the upper limit of the allowable harmonic content is increased to 10%, and the metering error tolerance is relaxed to ±5% to adapt to the strong interference and high dynamism of industrial scenarios.

[0044] Based on the scenario parameter type set, multi-dimensional data on historical anomaly triggering in various scenarios is extracted from the historical database. Taking the peak commercial electricity consumption scenario as an example, all records of abnormal alarms triggered by voltage drops within the past 12 months are retrieved, including the voltage deviation value at the time of each alarm (such as the difference between the actual voltage and the nominal voltage), the duration of fluctuation (the time from voltage drop to recovery), and the error jump amplitude (the amount of change in metering error before and after the anomaly). This data is structured according to the event dimension to construct a historical anomaly feature matrix (each row corresponds to one anomaly event, and each column contains feature values ​​such as voltage deviation, harmonic content, and load fluctuation), a historical fluctuation dataset (recording real-time parameter values ​​in a time series), and an alarm tag set (marking the anomaly type such as "voltage drop" and "harmonic exceedance" and the time of occurrence).

[0045] When training the initial anomaly correlation model, the tree-based XGBoost algorithm was selected. This algorithm iteratively generates decision trees through a gradient boosting mechanism, effectively capturing the nonlinear relationships between features. The historical anomaly feature matrix was used as input features, and the alarm labels as output labels. Simultaneously, the time-series features in the historical fluctuation dataset underwent frequency domain transformation: Fourier transform was used to decompose the time-series data into different frequency components, and the amplitude and phase of each frequency component were extracted as frequency domain features. These features were then concatenated with static features such as voltage and current before being input into the model to capture the periodicity and suddenness of parameter changes.

[0046] A cross-validation mechanism is implemented during model training: the historical dataset is divided into multiple subsets in chronological order, and a rolling window is used to sequentially use the preceding subset as the training set and the subsequent subset as the validation set, avoiding the leakage of temporal information caused by random partitioning. By adjusting hyperparameters such as the learning rate (controlling the contribution of each tree), maximum tree depth (limiting the complexity of the tree), and regularization parameters (preventing overfitting), the matching degree between the prediction results on the validation set and the alarm labels is observed until the model's performance in identifying abnormal events tends to stabilize, generating the final anomaly association judgment model.

[0047] In the scenario parameter setting stage, the granularity of electricity usage scenarios is further refined. For example, for residential electricity usage scenarios, it is divided into summer high-load period and winter stable period according to the season: in summer, due to the use of high-power equipment such as air conditioners, the load fluctuation range is adjusted to 15%-90%, and the upper limit of harmonic content is increased to 7%; in winter, the original fluctuation range of 20%-80% and the harmonic limit of 5% are maintained. For industrial scenarios, a distinction is made between continuous production enterprises and intermittent processing enterprises. The former sets stricter upper and lower limits for load fluctuation (such as 15%-95%) to monitor the stability of equipment operation, while the latter relaxes it to 10%-100% to adapt to the periodic start-up and shutdown requirements.

[0048] When constructing the historical anomaly feature matrix, data cleaning and feature derivation operations are performed. First, duplicate records are identified and removed using statistical methods. Missing voltage and current data are filled using linear interpolation or by means based on adjacent periods. Then, derived features are calculated, such as the power factor change rate (the ratio of the power factor difference between adjacent times to the time interval) and voltage fluctuation frequency (the number of times the voltage exceeds the normal range per unit time). All features are standardized, and Z-score transformation is used to make the mean of each feature 0 and the standard deviation 1, eliminating the influence of dimensional differences on model training.

[0049] In terms of time-series feature processing, in addition to Fourier transform, wavelet transform is introduced for multi-scale analysis. For non-stationary signals such as voltage drops, wavelet transform can decompose them into sub-bands of different frequencies (such as low-frequency trend components and high-frequency detail components), extract the energy proportion and singular point positions of each sub-band as features, effectively capture the instantaneous change characteristics and long-term trend changes of abnormal signals, and improve the model's ability to identify complex abnormal patterns.

[0050] To prevent overfitting during model training, an early stopping strategy is employed: the maximum number of iterations is set to 500. Training automatically terminates and saves the current optimal model parameters when the anomaly detection accuracy on the validation set fails to improve and the loss function no longer decreases after 20 consecutive iterations. Simultaneously, the importance of each feature is analyzed using SHAP values. For example, in peak commercial scenarios, voltage deviation has the largest absolute SHAP value, indicating its highest contribution to anomaly detection, which can be used to optimize the priority of monitoring parameters.

[0051] In practical applications, an online model update mechanism is established: new abnormal event data is collected weekly, and the model is fine-tuned using incremental learning. When the false negative rate of a certain type of abnormal event (such as excessive harmonics) exceeds a preset threshold (such as 15%) for three consecutive days, a full data retraining process is triggered. By dynamically adjusting the decision tree structure and feature weights, the model adapts to new abnormal patterns or changes in data distribution.

[0052] When processing historical fluctuation datasets, the changing trends and rate characteristics of parameters are additionally recorded. The changing trend is characterized by the sign of the difference between adjacent time points (positive, negative, zero), and the rate of change is measured by the absolute change per unit time. For example, although the current value is within the normal range at a certain moment, if it shows a downward trend for five consecutive sampling periods and the rate of change exceeds 5A / minute, it may indicate abnormal load attenuation before equipment failure. Such characteristics can trigger early warning mechanisms.

[0053] For newly introduced electricity consumption scenarios (such as electric vehicle charging stations), due to insufficient historical abnormal data, transfer learning technology is adopted: based on the pre-trained model of industrial nonlinear load scenario, its underlying feature extraction layer is retained, and only the terminal classification layer is fine-tuned. Adapted models are quickly generated through a small amount of new scenario labeled data, reducing the dependence on large-scale samples and shortening the model deployment cycle.

[0054] The correlation between parameters also needs to be considered during the model construction process: by calculating the Pearson correlation coefficient matrix, strongly correlated feature pairs (such as the correlation coefficient between voltage deviation and power factor > 0.7) are identified. When inputting the model, avoid directly stacking strongly correlated features, and instead use principal component analysis to extract comprehensive features, reduce the impact of multicollinearity on model stability, and improve computational efficiency.

[0055] Example 3: During the dynamic threshold correction process, the real-time fluctuation characteristics and error distribution of the current electricity meter's metering parameters are first acquired through a real-time data acquisition module, generating a dynamic parameter feature set. Real-time values ​​of voltage, current, and power are collected on a minute-by-minute basis. The maximum, minimum, and standard deviation of voltage within each minute are calculated as fluctuation characteristics. Simultaneously, the mean, variance, and number of times the error exceeds the initial threshold within that time period are statistically analyzed as the error distribution. For example, in industrial load scenarios, if the current harmonic content is detected to continuously increase over multiple minute cycles and the load fluctuation amplitude exceeds the preset range for the scenario, the system integrates this real-time data and derived characteristics into a dynamic parameter feature set.

[0056] Set anomaly detection and response criteria, such as setting the anomaly probability threshold to 0.8 (i.e., triggering a response when the model predicts an anomaly occurrence probability exceeding 80%). Input the dynamic parameter feature set into the trained anomaly association and detection model. The model outputs the anomaly probability value for the current state through its internal feature weights and decision logic. Taking a peak commercial electricity consumption scenario as an example, if the currently collected voltage deviation value is close to the upper limit of the initial threshold for the scenario, and the fluctuation frequency increases significantly, the model may output an anomaly probability value of 0.85, indicating that the current state has a high risk of triggering an anomaly.

[0057] When the current anomaly probability exceeds the response standard, a sliding window threshold update mechanism is initiated. First, a threshold optimization queue is created to store historical threshold data for the most recent N metering cycles (e.g., N=5), with each cycle corresponding to a time window. The window length is defined as M (e.g., M=3), and the sliding step size is 1 (i.e., the window moves one cycle at a time), ensuring that each update is based on the latest M cycles of data. Statistical measures of the dynamic parameter feature set over the time series are extracted, such as the mean voltage deviation, median current imbalance, and power error fluctuation range over the past M cycles, generating a feature rolling average matrix that reflects the recent trend of parameter fluctuations.

[0058] A threshold adjustment function is constructed based on the anomaly probability value. This function is designed based on the mapping relationship between the anomaly probability and the threshold adjustment range. For example, when the anomaly probability is 0.8-0.85, the voltage deviation threshold is adjusted from ±5% to ±4.5%; when the anomaly probability exceeds 0.85, it is further adjusted to ±4%, to gradually improve the detection sensitivity. The specific parameters of the threshold adjustment function are determined through correlation analysis of the anomaly probability and the actual error distribution in historical data, ensuring that the adjusted threshold can both capture anomalies in a timely manner and avoid false alarms due to oversensitivity.

[0059] The historical threshold data in the queue is updated using a sliding window mechanism: the latest calculated threshold adjustment result is added to the head of the queue, and the oldest historical threshold data at the tail of the queue is removed, ensuring that the queue always retains threshold information from the most recent N periods. After each window slide, the feature rolling average is recalculated based on the M window data in the queue, and the threshold adjustment function is called according to the current anomaly probability value to generate an optimized error threshold set containing the latest threshold. For example, in the third period, the window covers the data from periods 1-3, the feature average for these three periods is calculated, and the threshold is adjusted; when the data from the fourth period arrives, the window slides to periods 2-4, and the above process is repeated to achieve dynamic updating of the threshold.

[0060] A data quality filtering mechanism is introduced during the generation of the dynamic parameter feature set. For the collected raw data, validity is first verified, removing abrupt or unreasonable values ​​caused by communication failures. For missing sampling points, forward imputation or linear fitting based on data collected within the same period is used to complete the set, ensuring the integrity and reliability of the dynamic parameter feature set. Simultaneously, feature values ​​are normalized to ensure they fall within the numerical range required by the model input, avoiding the impact of data scale differences on the accuracy of anomaly probability calculations.

[0061] In the anomaly probability calculation stage, the model output is based not only on the dynamic parameter characteristics of the current cycle but also on historical characteristics from previous cycles for time-series correlation analysis. For example, if the voltage deviation value of the current cycle does not exceed the initial threshold, but the deviation values ​​of the previous two cycles show a gradually increasing trend, the model will increase the anomaly probability prediction value based on historical trend characteristics, providing early warning of potential anomalies. This time-series correlation analysis is achieved through the recursive structure or time-series feature encoding within the model, enhancing the foresight of the threshold correction.

[0062] The threshold optimization queue management mechanism includes setting a queue capacity limit and a data expiration policy. When the number of periods in the queue exceeds N, the oldest historical data is automatically deleted, ensuring that the computational and storage resources used by the queue are within a controllable range. Simultaneously, the threshold data in the queue is version-marked, recording the update time and adjustment reason for each threshold version. This facilitates subsequent tracing and analysis of the historical trajectory of threshold changes, providing a reference for model optimization and parameter adjustment.

[0063] During the sliding window update process, manual intervention in threshold adjustment is permitted. Power system operation and maintenance personnel can manually adjust the parameters of the threshold adjustment function or directly specify the threshold adjustment range based on actual operating experience or the power supply requirements for specific periods. For example, during power supply guarantee periods for important events, operation and maintenance personnel can temporarily tighten the voltage deviation threshold to ±3% to increase sensitivity to voltage anomalies. The manually adjusted threshold will take effect before the automatically calculated results from the model and will revert to automatic update mode after the power supply guarantee period ends.

[0064] To address the dynamic characteristics of different power consumption scenarios, scenario-adaptive sliding window parameters are designed. For scenarios with relatively stable load fluctuations, such as residential electricity consumption, a larger window length M (e.g., 5) and a smaller sliding step size (e.g., 1) are set to smooth the threshold adjustment frequency and avoid frequent threshold changes due to small fluctuations. For highly dynamic scenarios such as industrial loads, a smaller window length M (e.g., 2) and a larger sliding step size (e.g., 1) are used to enable the threshold to quickly respond to real-time parameter changes and improve the timeliness of anomaly detection.

[0065] When generating the optimized error threshold set, the anomaly triggering condition set is updated synchronously. Anomaly triggering conditions are not only based on threshold exceeding a single parameter, but also include multi-parameter combination conditions. For example, an anomaly alarm is triggered only when the voltage deviation exceeds the adjusted threshold and the current imbalance simultaneously exceeds the corresponding threshold, thus reducing the false alarm rate of single parameters through combined conditions. The logical relationships (such as "AND" and "OR") of multi-parameter combination conditions are determined based on the correlation analysis of parameters in historical anomaly events, ensuring the rationality and accuracy of the triggering conditions.

[0066] The entire process of dynamic threshold correction is recorded in the system log, including the time of each threshold adjustment, the threshold before / after adjustment, the anomaly probability value that triggered the adjustment, and the feature data involved in the calculation. The log data can be used for subsequent threshold adjustment effect evaluation and model performance analysis. For example, by comparing the number and types of anomaly detections before and after threshold adjustment, the effectiveness of the threshold correction strategy can be analyzed, providing data support for further optimization of sliding window parameters and threshold adjustment functions.

[0067] Example 4: When generating an anomaly detection scheme, the optimized error threshold set and anomaly triggering condition set are first loaded into the metering analysis unit. Based on a preset logical framework, the metering analysis unit integrates the new thresholds and conditions with rules in the anomaly detection rule base. For example, if the original rule base contains a rule that "triggers an early warning when the voltage distortion coefficient exceeds 0.25," and the optimized voltage deviation threshold is adjusted from ±5% to ±4%, the resulting composite judgment condition is: "When the voltage distortion coefficient exceeds 0.25 and the voltage deviation exceeds ±4%, it is judged as a voltage anomaly." The corresponding alarm method (such as audible and visual alarms, SMS notifications, etc.) is then configured according to the anomaly level.

[0068] The system identifies the time dependencies in the data verification process within the anomaly detection scheme. The data verification process includes data acquisition, preprocessing, feature extraction, and rule matching, each with strict temporal logic. For example, preprocessing requires denoising and normalizing the acquired raw data, feature extraction relies on the standardized preprocessed data, and rule matching compares the feature extraction results with thresholds and rules. The system analyzes the state transition diagram of the verification process to identify the input-output relationships and temporal order of each step, forming a process topology structure containing nodes (steps) and directed edges (temporal dependencies).

[0069] Based on the device clock synchronization mechanism, a timestamp alignment algorithm is used to logically calibrate the verification trigger timing. The timestamp alignment algorithm first obtains the clock source of the data acquisition module (such as a precise GPS-based clock or an IEEE 1588 synchronization clock) and assigns a timestamp accurate to the microsecond level to each acquired raw data sample. In the verification process, the system executes each step sequentially according to the timestamp order: when the preprocessing stage receives the raw data with timestamp t1, it processes it and generates preprocessed data with timestamp t2 (t2≥t1), which is then passed to the feature extraction stage; the feature extraction stage completes feature calculation at timestamp t3 (t3≥t2) and outputs it to the rule matching stage. For data samples that arrive late due to network latency or other reasons, the algorithm automatically inserts them into the queue position corresponding to the timestamp order, waiting for the data with the preceding timestamp to be processed before execution, avoiding feature calculation errors or rule matching failures due to timing disorder.

[0070] A version management mechanism is established when loading optimized thresholds and conditions. Each optimized error threshold set and anomaly triggering condition set includes metadata such as version number, effective time, and adjustment reason. When a new version is loaded into the measurement and analysis unit, the system automatically backs up the old version data so that it can quickly roll back if the new version causes abnormal detection results. At the same time, the version management mechanism supports querying and retrieving historical threshold versions by time range, meeting the needs of retrospective analysis of historical data.

[0071] The time dependency identification in the data verification process also includes handling parallel steps. For example, in the feature extraction stage, voltage feature extraction and current feature extraction can be performed in parallel, with no direct time dependency between them. By marking the independent execution attribute of parallel steps, the system allows them to start calculations simultaneously after obtaining preprocessed data copies, improving the execution efficiency of the verification process. The results of parallel steps are summarized before the rule matching stage to ensure the normal execution of the multi-feature fusion judgment logic.

[0072] The parameters of the timestamp alignment algorithm are dynamically adjusted based on the data acquisition frequency and transmission delay. For example, when the data acquisition frequency increases from 100 times per second to 500 times per second, the algorithm improves the timestamp precision from milliseconds to microseconds and shortens the processing time interval between each stage to match higher real-time requirements; when the average transmission delay is detected to increase from 50ms to 100ms, the algorithm automatically extends the data reception waiting time window to avoid data loss due to premature triggering of subsequent stages.

[0073] In the rule fusion process of the anomaly detection scheme, a rule priority mechanism is introduced. For rules at different levels (such as basic threshold rules, composite scenario rules, and expert experience rules), a priority order from high to low is set. When rules of different priorities conflict in their determination of the same data sample, the result of the higher-priority rule prevails. For example, in expert experience rules, the anomaly detection conditions for a specific model of electricity meter have a higher priority than the basic threshold rules, ensuring the accuracy of the detection logic in special scenarios.

[0074] During the timing logic calibration process, clock synchronization verification is performed periodically. The system automatically detects the clock deviation of each device node every week. When the deviation exceeds a preset threshold (such as 10ms), a global clock synchronization process is triggered. The clocks of all devices are calibrated through Network Time Protocol (NTP) or a dedicated synchronization link to ensure the consistency and accuracy of timestamps and avoid timing logic chaos caused by the accumulation of clock drift.

[0075] When generating anomaly detection schemes, custom configuration of the verification process is supported. Power system operation and maintenance personnel can manually adjust the execution order, parallel strategies, and timeout thresholds of each step through a human-machine interface. For example, in industrial scenarios with extremely high real-time requirements, some computational tasks in the feature extraction step can be completed in the preprocessing step to reduce overall processing latency; in scenarios with strict data quality requirements, the number of iterations in the data cleaning step can be increased to ensure the accuracy of data in the input rule matching step.

[0076] The timestamp alignment algorithm is implemented based on a distributed system time synchronization framework. In a multi-node metering and analysis system, each node maintains a local clock and synchronizes with the master clock node via a heartbeat mechanism. When a node detects that the data transmission delay exceeds expectations, it automatically requests time calibration from the master node and adjusts its local verification trigger timing to ensure that the verification logic timing of each node is consistent in a distributed environment.

[0077] The anomaly detection scheme outputs detailed detection logs, recording the processing time chain for each data sample (the time consumed at each stage from data collection to completion of the detection), triggered rule entries, threshold comparison results, and other information. This log data is used for subsequent process performance analysis; for example, by statistically analyzing the average processing time of each stage, bottlenecks in the verification process can be identified, providing a basis for system optimization.

[0078] In the rule matching stage, caching technology is used to store frequently accessed threshold data and rule entries. For example, the optimized error threshold set for the current scenario is loaded into the memory cache to avoid reading data from the disk every time a rule is matched, thus improving judgment efficiency. An expiration time is set for the cached data; when a threshold version update is detected, the cache content is automatically refreshed to ensure that the latest judgment conditions are used.

[0079] Example 5: When generating an anomaly detection scheme, the optimized error threshold set and anomaly triggering condition set are first loaded into the metering analysis unit. Based on a preset logical framework, the metering analysis unit integrates the new thresholds and conditions with rules in the anomaly detection rule base. For example, it combines the adjusted voltage deviation threshold with the existing voltage distortion coefficient rule to form a new composite judgment condition: when the voltage distortion coefficient exceeds 0.25 and the voltage deviation exceeds ±4%, it is judged as a voltage anomaly. During the integration process, the system defines the combination relationship between rules through logical operators (such as "AND", "OR", and "NOT") and assigns a unique identifier to each rule for traceability.

[0080] The system identifies the time dependencies in the data verification process within the anomaly detection scheme. The data verification process includes data acquisition, preprocessing, feature extraction, and rule matching, each of which must be executed sequentially. The system establishes an execution queue for each step, using timestamps as a basis to determine the processing order. For example, the preprocessing step must wait until data with timestamps is received before proceeding. After processing the raw data, the process starts and generates a timestamp. Standardized data ( (and then pass it to the feature extraction stage.)

[0081] Based on the device clock synchronization mechanism, a timestamp alignment algorithm is used to logically calibrate the verification trigger timing. This algorithm is implemented through the following steps: Timestamp allocation: Attach a timestamp of the acquisition time to each data sample. Accuracy down to the microsecond level.

[0082] Sequential queue management: Establish a first-in, first-out (FIFO) queue, and sort it according to... The data samples to be processed are stored sequentially to ensure that the data collected later does not enter the processing stage before the data collected earlier.

[0083] Delay compensation: For samples that arrive late due to transmission delay (assuming the actual arrival time is...) ,like ,in (If a preset delay tolerance threshold is set), it will be automatically marked as an abnormal sample and a resampling mechanism will be triggered, or it will be inserted into the end of the queue to wait for processing.

[0084] When loading optimized thresholds and conditions, the system performs rule conflict detection. For example, if the rule "cumulative power deviation exceeds ±3%" and the rule "abnormal power pulse frequency" logically overlap in a certain scenario, the system compares the rule priorities (which are determined by attributes such as rule type and application scenario) and retains the higher-priority rule, while masking the conflicting parts of the lower-priority rule. The priority is defined as: expert experience rules > composite scenario rules > basic threshold rules.

[0085] The time dependency identification process in the data verification workflow also includes handling parallel steps. For example, if there is no data dependency between voltage feature extraction and current feature extraction, they can be executed in parallel to improve efficiency. The system achieves parallel processing through multi-threading technology and merges the results before the rule matching step. During merging, the feature data with the latest timestamp is used to ensure the synchronization of multiple features.

[0086] The core formula of the timestamp alignment algorithm is: in: This is the start time of the current processing step; The timestamp of the current data sample; This refers to the completion time of the preceding steps.

[0087] This formula ensures that the current step only starts after the preceding step is completed and the data samples arrive, avoiding calculation errors caused by timing discrepancies. For example, in the feature extraction step... The preprocessing stage must be completed simultaneously. ) and data samples arrived ( The later of the two times is taken as the actual start time.

[0088] When generating anomaly detection schemes, it is supported to configure timeout thresholds for each step via a human-machine interface. For example, the maximum allowable time for the feature extraction step can be set to [value missing]. If the processing time for a certain sample in this step exceeds If the sample fails to pass the test, an alarm will be triggered and the sample will be skipped (or marked as pending retry) to prevent the processing delay of a single sample from blocking the entire process.

[0089] During timing logic calibration, the system clock is synchronized periodically (e.g., daily). This is achieved by connecting to a Network Time Protocol (NTP) server to align the local clocks of each device with the global time base, ensuring... To ensure accuracy and consistency, and avoid errors caused by clock drift. Calculate the deviation.

[0090] The anomaly detection scheme employs a forward chain reasoning mechanism in its rule matching stage. Starting with the basic features of the data samples, it progressively matches conditions from the rule base. For example, it first determines whether the voltage deviation exceeds a threshold, then whether the current imbalance is abnormal, and finally generates a judgment conclusion based on the logical combination results. This mechanism ensures that rule matching is executed in a preset order, improving judgment efficiency.

[0091] The data verification process log includes timestamps and processing times for each step, for example: Collection time: Preprocessing completion time: ,time consuming Feature extraction completion time: ,time consuming By analyzing this data, resource allocation between stages can be optimized, such as adding computation threads for time-consuming feature extraction stages.

[0092] In distributed deployment scenarios, timestamp alignment algorithms achieve cross-node time synchronization through distributed coordination services (such as Zookeeper). Each node adjusts its local processing rhythm by listening to global time signals to ensure that similar processes on different nodes (such as the rule matching process between node A and node B) execute consistent judgment logic for data samples with the same timestamp.

[0093] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0094] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing abnormality of an electric energy meter, characterized by, Includes the following steps: Collect voltage waveforms, current phases, and power pulse data within multiple electricity meter metering cycles to generate a set of metering feature vectors; Based on the measurement feature vector set, data mining techniques are used to distinguish between normal measurement patterns and abnormal patterns, extract key feature combinations and corresponding rules that can characterize abnormal states, and store such rules in a structured manner to build an anomaly detection rule library. Set the threshold types of metering parameters for different electricity consumption scenarios to form a set of scenario parameter types; Based on the set of scenario parameter types, the error distribution, fluctuation frequency and abnormal trigger records in historical measurement data are obtained. By analyzing the correlation between data, a model that can reflect the causal relationship between parameter changes and abnormal events is constructed to predict the probability of anomalies under the current parameter state, thereby establishing an abnormal correlation judgment model. Extract the metering parameter threshold types corresponding to the current electricity meter operating scenario, and perform dynamic threshold correction based on the anomaly correlation judgment model to generate an optimized error threshold set and an anomaly triggering condition set. The optimized error threshold set and anomaly triggering condition set refer to optimized thresholds and triggering conditions adapted to real-time operating conditions. The dynamic threshold correction includes: Obtain the real-time fluctuation characteristics and error distribution of the current electricity meter's metering parameters, and generate a dynamic parameter feature set; Set anomaly detection response criteria, input the dynamic parameter feature set into the anomaly association detection model, and calculate the current anomaly probability value; When the current anomaly probability value exceeds the response standard, the error threshold set is updated using a sliding window to generate an optimized error threshold set. The measurement analysis strategy is updated based on the optimized error threshold set, anomaly triggering condition set, and anomaly detection rule base to generate an anomaly determination scheme; the timing logic of data verification in the anomaly determination scheme is synchronously corrected.

2. The electric energy metering anomaly analysis method according to claim 1, characterized in that, The generated measurement feature vector set includes: Define multiple metering characteristic dimensions, including voltage distortion coefficient, current imbalance, and cumulative power deviation; Collect real-time sampling data corresponding to each dimension and perform standardization processing to generate a standardized measurement feature matrix; The standardized quantitative feature matrix is ​​input into the anomaly detection model generation module, and an anomaly detection rule base is constructed through a pre-set clustering algorithm.

3. The electric energy metering anomaly analysis method according to claim 2, characterized in that, The establishment of the anomaly correlation determination model includes: Set parameter types for different power consumption scenarios, including load fluctuation range, harmonic content range, and metering error tolerance; Based on the scene parameter type set, obtain the error distribution characteristics, data fluctuation cycle and alarm event records when historical anomalies are triggered, and construct a historical anomaly feature matrix, a historical fluctuation dataset and an alarm tag set. An initial anomaly correlation model is trained using the historical anomaly feature matrix, historical fluctuation dataset, and alarm label set to generate an anomaly correlation determination model.

4. The electric energy metering anomaly analysis method according to claim 3, characterized in that, The initial anomaly association model for training includes: An initial anomaly correlation model is established based on the set of scenario parameter types and the dimensions of measurement features; Input the historical anomaly feature matrix and historical fluctuation dataset item by item into the initial anomaly association model to generate a set of historical alarm prediction results. Set the error judgment tolerance range and calculate the matching deviation between the historical alarm prediction result set and the alarm tag set; When the matching deviation exceeds the tolerance range, the density clustering algorithm is used to iteratively optimize the judgment boundary of the initial abnormal association model until the deviation reaches the preset standard, thus forming an abnormal association judgment model.

5. The electric energy metering anomaly analysis method according to claim 4, characterized in that: The density clustering algorithm employs an adaptive neighborhood radius adjustment mechanism.

6. The electric energy metering anomaly analysis method of claim 1, wherein, The sliding window update includes: Create a threshold optimization queue and define the window length and sliding step parameters; Extract the statistics of the dynamic parameter feature set on the time series and generate a feature rolling average matrix; Construct a threshold adjustment function based on the anomaly probability value; The historical threshold data in the queue is updated by a window sliding mechanism to generate an optimized error threshold set.

7. The electric energy metering anomaly analysis method according to claim 6, characterized in that, The anomaly detection scheme includes: The optimized error threshold set and anomaly triggering condition set are loaded into the metrology analysis unit to generate an anomaly determination scheme. Identify the time dependency of the data verification process in the anomaly detection scheme and perform logical calibration based on the device clock synchronization mechanism.

8. The electric energy metering anomaly analysis method according to claim 7, characterized in that: The logical calibration employs a timestamp alignment algorithm, which dynamically adjusts the verification trigger timing based on the data acquisition frequency and transmission delay.

9. An electric energy metering anomaly analysis system, characterized by, Includes the following modules: The data acquisition module is used to collect voltage waveforms, current phases, and power pulse data within multiple electricity meter metering cycles, and generate a set of metering feature vectors. The abnormal rule base construction module uses data mining technology to distinguish between normal measurement patterns and abnormal patterns based on the measurement feature vector set, extracts key feature combinations and corresponding rules that can characterize abnormal states, stores such rules in a structured manner to train the abnormal detection rule base, and generates initial abnormal judgment rules through feature clustering and pattern matching. The scenario parameter configuration module allows you to set the threshold types of metering parameters for different electricity consumption scenarios, forming a set of scenario parameter types that includes voltage deviation, phase shift, and power pulse frequency. The correlation model training module extracts error distribution, fluctuation frequency and abnormal trigger records from historical measurement data based on the scenario parameter type set. By analyzing the correlation between data, it constructs a model that can reflect the causal relationship between parameter changes and abnormal events, which is used to predict the probability of anomalies under the current parameter state, thereby establishing an abnormal correlation judgment model. The dynamic threshold optimization module matches the corresponding metering parameter threshold type according to the current electricity meter operating scenario, calls the anomaly association judgment model to dynamically correct the error threshold, and generates an optimized error threshold set and an anomaly triggering condition set. The optimized error threshold set and anomaly triggering condition set refer to optimized thresholds and triggering conditions adapted to real-time operating conditions. The dynamic threshold correction includes: Obtain the real-time fluctuation characteristics and error distribution of the current electricity meter's metering parameters, and generate a dynamic parameter feature set; Set anomaly detection response criteria, input the dynamic parameter feature set into the anomaly association detection model, and calculate the current anomaly probability value; When the current anomaly probability value exceeds the response standard, the error threshold set is updated using a sliding window to generate an optimized error threshold set. The strategy update module integrates the optimized error threshold set, the abnormal triggering condition set, and the abnormal detection rule base to generate an abnormal judgment scheme that includes data verification logic. The timing correction module synchronously calibrates the timing logic of data verification in the anomaly determination scheme to ensure the dynamic consistency between the anomaly detection rules and the correction threshold in each measurement cycle.

Images