Electric energy meter state evaluation and error prediction system based on data driving

Abstract

The invention relates to the technical field of electric energy meter evaluation and prediction, and discloses an electric energy meter state evaluation and error prediction system based on data driving. According to the system, dynamic operation data and static attribute data of the electric energy meter are fused, a comprehensive state evaluation model is constructed, and accurate description of the performance state of the electric energy meter and effective traceability of error causes are achieved. And after error trend prediction is carried out based on a real-time state evaluation result, a dynamic change trend of prediction deviation is further analyzed, and a specific differentiation operation and maintenance decision is automatically generated, so that closed-loop management from state monitoring and prediction to active intervention is completed. The method is mainly applied to the field of operation and maintenance of metering equipment of an intelligent power grid, the comprehensiveness and accuracy of state evaluation can be improved, the operation and maintenance mode is promoted to be converted from regular verification to predictive maintenance, and the intelligent level of asset management is enhanced.

Description

Technical Field

[0001] This invention relates to the field of electricity meter assessment and prediction technology, specifically to a data-driven electricity meter condition assessment and error prediction system. Background Technology

[0002] In the operation of a power system, the electricity meter, as the core device for measuring electricity consumption, directly affects the fairness of electricity trade settlement and the overall operating efficiency of the power system due to the stability of its operating status and the accuracy of its measurement errors. With the advancement of smart grid construction, the number of electricity meters has increased significantly, and their operating environment has become increasingly complex. Factors such as changes in temperature and humidity in different regions, fluctuations in grid voltage, and differences in the aging of equipment can all affect the operating status of electricity meters, leading to fluctuations in measurement errors.

[0003] Methods for assessing the condition and predicting errors of electricity meters largely rely on periodic manual inspections and offline testing. Manual inspections require staff to check each meter on-site, which is not only costly in terms of manpower and time, but also suffers from long inspection cycles, limited coverage, and difficulty in capturing real-time dynamic changes in the meter's operating status. Offline testing requires removing the meter from the field and sending it to a laboratory for testing. This process interrupts electricity metering, affecting users' normal electricity consumption, and the test results only reflect the meter's condition at the time of testing, failing to continuously track its status changes and error trends during operation.

[0004] Existing technologies that assess the status of electricity meters using single parameters or simple models often overlook the diversity and correlation of meter operating data. For example, judging the status of an electricity meter solely based on voltage or current data cannot comprehensively reflect potential problems such as aging internal components or poor wiring connections. Furthermore, traditional statistical models tend to suffer from insufficient generalization ability and low error prediction accuracy when processing large amounts of time-series operating data, failing to meet the real-time and accuracy requirements of smart grids for electricity meter status assessment and error prediction. In addition, the existing assessment process lacks a systematic analysis and strategy generation stage for error deviation parameters, resulting in assessment results that cannot effectively guide electricity meter operation and maintenance decisions, further reducing the efficiency of power system operation and management. Summary of the Invention

[0005] The purpose of this invention is to provide a data-driven electricity meter condition assessment and error prediction system to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a data-driven system for assessing the condition and predicting errors of electricity meters, the system comprising:

[0007] The module includes an electricity meter operation data acquisition module, a state assessment model construction module, a real-time state estimation module, an error prediction module, an assessment strategy analysis module, and an assessment result generation module.

[0008] The electricity meter operation data acquisition module is configured to collect dynamic and static operation datasets of the electricity meter.

[0009] The state assessment model construction module is configured to construct an energy meter state assessment model based on the dynamic and static operation dataset.

[0010] The real-time state estimation module is configured to use the energy meter state assessment model to perform state matching and parameter estimation on the current energy meter operating data to obtain the real-time state parameters of the energy meter.

[0011] The error prediction module is configured to predict the error trend based on the real-time status parameters of the energy meter and output the error deviation parameters.

[0012] The evaluation strategy parsing module is configured to judge the changing trend of the error deviation parameter and analyze the strategy to generate a state evaluation strategy.

[0013] The evaluation result generation module is configured to analyze the real-time status parameters and error deviation parameters of the electricity meter based on the status evaluation strategy, and generate electricity meter status evaluation results and error prediction results.

[0014] Preferably, the electricity meter operation data acquisition module acquires dynamic and static operation datasets through the following operations: acquiring electricity time-series data and metering curves in parallel from multiple electricity meter devices; performing real-time stream processing on the acquired raw data, including data verification and format unification; applying anomaly detection algorithms to identify and remove outliers; using interpolation methods to complete missing data; performing data dimensionality reduction processing to extract key features through principal component analysis; and finally scaling the data to a uniform range using a standardization algorithm to form a standardized dynamic and static operation dataset.

[0015] Preferably, the state assessment model building module constructs the electricity meter state assessment model through the following steps: calculating time series feature indicators from dynamic and static operating datasets, including sliding window statistics and trend components; initializing a long short-term memory network architecture for meter value prediction; adapting the pre-trained model to new electricity meter data using an adversarial transfer learning algorithm; adjusting the network hyperparameters using a Bayesian optimization method; and evaluating the model performance using cross-validation, ultimately solidifying the model parameters.

[0016] Preferably, the real-time state estimation module obtains the real-time state parameters of the electricity meter through the following operations: calculating the feature similarity matrix between the current electricity meter operating data and the electricity meter state assessment model; applying a recursive least squares algorithm with multiple forgetting factors to recursively estimate the state parameters; introducing a regularization term to prevent matrix ill-conditionedness; and iteratively updating until the parameters converge, outputting stable real-time state parameters of the electricity meter.

[0017] Preferably, the error prediction module outputs error deviation parameters through the following steps: using a collaborative filtering algorithm to jointly solve for the line loss of the transformer area and the operating error of the electricity meter; extracting over-limit events from the error sequence based on the peak over-threshold model; fitting a generalized Pareto distribution to calculate the over-limit probability; generating an error deviation parameter sequence and analyzing its dynamic changes through a sliding window.

[0018] Preferably, the joint solution operation in the error prediction module is further optimized through the following steps: constructing a Tikhonov regularization matrix to balance the estimated bias and variance; using leave-one-out cross-validation to select regularization parameters; and iteratively updating the estimated values ​​using gradient descent until the error is minimized.

[0019] Preferably, the evaluation strategy parsing module generates a state evaluation strategy through the following operations: performing time series decomposition on the error deviation parameter to extract trend components, seasonal components, and residual components; formulating a horizontal adjustment strategy based on the trend components, a vertical adjustment strategy based on the seasonal components, and a lateral adjustment strategy based on the residual components; and integrating multiple strategies to form a comprehensive state evaluation strategy.

[0020] Preferably, the strategy formulation operation in the evaluation strategy parsing module is further enhanced by the following steps: establishing a fuzzy judgment matrix to evaluate the importance of each indicator; applying the eigenvector method to calculate the initial weights; dynamically adjusting the weights based on the data freshness index using a variable weight function; and outputting the final weight set for strategy weighted fusion.

[0021] Preferably, the evaluation result generation module generates the electricity meter status evaluation result and error prediction result through the following operations: using the K-means clustering algorithm to classify the real-time status parameters of the electricity meter into health levels; generating a rolling prediction interval based on the error deviation parameter; calculating the risk value of risk to assess the error risk; and outputting a complete evaluation result including health level labels and prediction intervals.

[0022] Preferably, the evaluation result generation module also integrates a visualization platform unit, which displays the results through the following steps: rendering an interactive dashboard to show the status evaluation results; dynamically updating the error prediction curve; providing drill-down functionality for detailed analysis; and supporting multi-dimensional data filtering and real-time updates.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] This data-driven electricity meter condition assessment and error prediction system establishes a complete process by setting up multiple functional modules that work together. Among them, the electricity meter operation data acquisition module can selectively collect dynamic and static operation datasets, covering various key data during the electricity meter's operation. This provides a comprehensive data foundation for subsequent model building and condition assessment, avoiding the biased assessment problems caused by single data sets in traditional methods.

[0025] The condition assessment model building module constructs the model based on the collected dynamic and static operating datasets. Relying on a data-driven approach, it fully explores the inherent correlations between data, enabling the constructed electricity meter condition assessment model to better fit the actual operating characteristics of the electricity meter. Compared with traditional statistical models, it is more adaptable to the needs of electricity meter condition analysis under different operating environments, effectively improving the model's ability to characterize the electricity meter condition.

[0026] The real-time status estimation module uses the existing energy meter status assessment model to perform status matching and parameter estimation on the current energy meter operating data. It can obtain the real-time status parameters of the energy meter in real time, breaking the limitations of traditional manual inspection and offline detection that cannot monitor in real time. This allows staff to keep abreast of changes in the operating status of the energy meter and facilitates a rapid response when abnormalities occur.

[0027] The error prediction module predicts error trends based on real-time status parameters and outputs error deviation parameters. It can detect the direction of change in the metering error in advance, avoid inaccurate metering caused by error accumulation, ensure fairness in the electricity trade settlement process, and reduce disputes caused by metering errors.

[0028] The evaluation strategy analysis module judges the changing trend of error deviation parameters and analyzes the strategy to generate corresponding state evaluation strategies. This provides clear directional guidance for the subsequent evaluation result generation, making the evaluation process more targeted and avoiding the problem of low practicality of evaluation results caused by the lack of strategy guidance in traditional evaluations.

[0029] The assessment result generation module, based on a state assessment strategy and combining real-time state parameters and error deviation parameters, produces state assessment results and error prediction results. This transforms complex data and model analysis into intuitive assessment conclusions, providing clear reference for the operation and maintenance of electricity meters. It helps maintenance personnel develop more reasonable maintenance plans, optimize the allocation of maintenance resources, reduce unnecessary maintenance costs, and ensure the overall stability and reliability of the power system. Through the close cooperation of its modules, the entire system achieves full-process coverage from data acquisition, model building, real-time monitoring to result generation, forming a closed-loop electricity meter state management mechanism. This effectively addresses the characteristics of a large number of electricity meters and a complex operating environment in smart grids, meeting the actual needs of the power system for electricity meter state assessment and error prediction. Attached Figure Description

[0030] Figure 1 This is a schematic diagram illustrating the working principle of the data-driven energy meter status assessment and error prediction system described in this invention.

[0031] Figure 2 A flowchart for collecting dynamic and static operation datasets for the electricity meter operation data acquisition module;

[0032] Figure 3 A flowchart for constructing a state assessment model for an energy meter;

[0033] Figure 4 A multi-scale estimation curve of the state parameters of an electricity meter;

[0034] Figure 5 A graph showing the relationship between the percentage improvement in electricity meter error and adjustment strategies. Detailed Implementation

[0035] 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.

[0036] Please see Figure 1This invention provides a data-driven system for assessing the condition and predicting errors of electricity meters. The system includes: an electricity meter operation data acquisition module responsible for collecting dynamic and static operation datasets from the electricity meter device, including raw information such as power time-series data and metering curves; a condition assessment model construction module using the dynamic and static operation datasets to establish an electricity meter condition assessment model, capturing data features through machine learning methods; a real-time condition estimation module applying the electricity meter condition assessment model to process the current electricity meter operation data and outputting real-time condition parameters; an error prediction module performing trend analysis based on the real-time condition parameters to generate error deviation parameters; an assessment strategy parsing module parsing and judging the error deviation parameters to form a condition assessment strategy; and an assessment result generation module integrating the real-time condition parameters and error deviation parameters to generate the final electricity meter condition assessment result and error prediction result.

[0037] Example 1: See Figure 2The electricity meter operation data acquisition module collects dynamic and static operation datasets through a distributed acquisition node group deployed in the power data center. This distributed acquisition node group establishes a communication connection with the smart meters installed on-site via a dedicated power fiber optic network. Each distributed acquisition node is equipped with an industrial-grade communication interface that supports the DL / T645 protocol and Modbus protocol to adapt to the data exchange needs of different electricity meter models. Data acquisition tasks are triggered periodically according to a preset acquisition cycle. After the acquisition command is issued, the smart meter encapsulates the power time-series data and metering curve into data frames and uploads them through the communication network. The power time-series data includes second-level or minute-level sampled values ​​of parameters such as voltage, current, active power, and reactive power. The metering curve records the periodic change trajectory of the electricity meter's cumulative electricity consumption. When the data stream enters the distributed acquisition nodes, it undergoes preliminary format parsing and validity verification. Invalid data packets are marked with a timestamp and source information and then transferred to an exception handling queue. The raw data enters the real-time stream processing engine through a message queue middleware. The real-time stream processing engine adopts a streaming computing framework based on the Apache Flink architecture. The data verification operation checks whether the numerical range of each data point is within the range of the electricity meter. The numerical range check includes ensuring the voltage value is not lower than 70% and not higher than 120% of the rated voltage, and the current value is not less than 0 and does not exceed the maximum allowable current. The format unification operation maps data fields parsed from different protocols to a unified data model. The data model definition includes standard fields such as meter identification, acquisition time, data value, and quality code. The anomaly detection algorithm uses an improved local outlier factor algorithm to identify outliers. This improved algorithm adapts to the density distribution characteristics of the power data by adjusting the neighborhood parameter k value. Outlier determination is based on the ratio of local reachability density to the reachability density of neighboring points. Data records marked as outliers are not immediately deleted but are transferred to a manual review cache. The review cache retains outlier data from the most recent 72 hours for maintenance personnel to review.

[0038] Missing data completion employs a piecewise interpolation method based on time series similarity. This method first uses a dynamic time warping algorithm to find the most similar data pattern to the current missing segment in historical data, and then uses a cubic spline interpolation algorithm to fit curves based on the feature points of the similar patterns. Data dimensionality reduction uses kernel principal component analysis (KPCA) to extract key features. KPCA maps the original data to a high-dimensional feature space using radial basis functions, and then performs principal component extraction in this high-dimensional space to capture nonlinear features. The data processing pipeline adopts a modular design, with each processing unit running independently in a Docker container and communicating via an event-driven architecture. The data validation unit outputs a quality report recording the pass rate of each batch of data, with a standardized format and a unit-maintained reduction mapping table supporting dynamic loading of new reductions. The anomaly detection unit provides a parameter configuration interface allowing adjustment of detection sensitivity, and the missing data completion unit records the confidence score of each interpolation operation. The dimensionality reduction unit generates a feature importance report listing the variance contribution rate of each principal component, and the standardization unit saves scaling parameters for real-time processing of subsequent new data. The operational status of all processing units is displayed in real time through a monitoring panel, including information such as processing throughput, processing latency, and error logs.

[0039] Standardized dynamic and static operation datasets are stored in a time-series database, which employs a partitioned storage strategy to shard data according to meter identifiers and time ranges. The data access layer provides a unified query interface supporting data retrieval by time range, meter group, data type, and other criteria. The data caching layer uses a Redis cluster to cache hot data, improving query response speed. Metadata information for the dynamic and static operation datasets is registered in a metadata management repository, including descriptive information such as data version, creation time, data size, and quality score. Data security employs a role-based access control policy, granting different users different levels of data access permissions. Encrypted data transmission uses the TLS protocol to ensure communication security. Distributed acquisition nodes have a reconnection mechanism; when the network is interrupted, acquired data is temporarily stored in a local cache, and automatically synchronized to the data center after network recovery. Performance monitoring metrics for the acquisition module include acquisition success rate, data latency, and system load, which are visualized through a dashboard with threshold alarm settings. The update frequency of the dynamic and static operation datasets can be configured according to business needs, supporting acquisition plans set at different granularities such as minutes, hours, and days. Historical data archiving strategies transfer data that exceeds a certain time range to cold storage media, freeing up online storage space while maintaining the queryability of historical data.

[0040] The electricity meter operation data acquisition module and the status assessment model construction module exchange data via an enterprise service bus, which encapsulates data messages in JSON format. The data exchange protocol defines a complete message header and payload structure. The message header includes information such as message identifier, timestamp, and source module identifier, while the payload carries the binary stream or index information of the dynamic and static operation datasets. Digital signature technology is used to verify data integrity during data transmission, preventing data tampering. The acquisition module provides a data subscription service, allowing other modules to register and listen for specific types of data change events, actively pushing new data to subscribers when it is generated. The backup strategy for the dynamic and static operation datasets adopts a multi-replica mechanism, with the primary replica stored on a high-performance solid-state drive array and the replicas distributed across servers in different racks. The data lifecycle management strategy defines the entire process of data creation, storage, archiving, and destruction; expired data is anonymized and cleaned up according to a prescribed process. The acquisition module's operation and maintenance management interface provides complete operation logs, recording details of user logins, parameter modifications, data exports, etc., for easy auditing and tracing. The electricity meter operation data acquisition module adopts a microservice architecture to achieve high availability. The failure of a single node will not affect the overall acquisition function. Service mesh technology realizes load balancing and automatic fault transfer.

[0041] Example 2: See Figure 3The state assessment model construction module extracts time-series feature indicators from the dynamic and static operation datasets output by the electricity meter operation data acquisition module. The calculation of time-series feature indicators is based on a sliding window mechanism to segment the electricity time-series data. The sliding window size is set to 144 data points corresponding to one day's electricity meter sampling data, and the window sliding step is 1 data point to achieve continuous coverage. Within each sliding window, 12 statistical features are calculated, including mean, variance, skewness, kurtosis, root mean square, waveform factor, peak factor, impulse factor, margin factor, standard deviation, mean deviation, and coefficient of variation. The trend component is obtained by fitting the slope values ​​of the data points within the window using the linear least squares method. The feature indicator calculation process uses vectorized operations to improve processing efficiency, and the calculation results are stored in the form of a feature matrix for easy model input. The long short-term memory network architecture adopts a 3-layer hidden layer structure, with each layer containing 128 memory units. The number of input layer nodes corresponds to the number of dimensions of the feature matrix, and the number of output layer nodes corresponds to 1 metering prediction value. The Long Short-Term Memory (LSTM) network architecture is initialized using the Xavier method to set the weight matrix. The Xavier method automatically adjusts the initial weight range based on the number of input and output nodes to prevent gradient vanishing. The forget gate bias term is initially set to 1.0 to retain more historical information during the early stages of model training, while the input and output gate bias terms are initially set to 0.0. The loss function for the LSM network architecture is defined as the mean squared error function, and the Adam algorithm is chosen as the optimizer with an initial learning rate of 0.001. The batch size is set to 32 samples, the maximum number of training epochs is 200, and early stopping is employed to prevent overfitting.

[0042] The adversarial transfer learning algorithm comprises two components: a feature extractor and a domain discriminator. The feature extractor consists of the first few layers of a pre-trained Long Short-Term Memory (LSTM) network architecture. The pre-trained model is trained on three years of historical data from 100,000 electricity meters, capturing the general operating patterns of these meters. The domain discriminator is a fully connected neural network with three hidden layers, each with 64 nodes. The domain discriminator outputs a similarity score between the new electricity meter data and the pre-trained data. During adversarial training, the feature extractor learns to generate feature representations that confuse the domain discriminator, which continuously optimizes its domain classification capabilities. A gradient inversion layer multiplies the gradient of the domain discriminator by -1 during backpropagation to achieve the adversarial objective. The training iteration count is set to 5000. After the adversarial transfer learning algorithm is completed, the feature extractor parameters are fixed, and the last two layers of the LSM network architecture are fine-tuned on fully connected layers to adapt to the characteristics of the new electricity meters.

[0043] Bayesian optimization constructs a Gaussian process surrogate model to describe the mapping relationship between hyperparameters and model performance. The hyperparameter search space includes a learning rate with a logarithmic uniform distribution between 0.0001 and 0.01, a hidden layer number with an integer uniform distribution between 1 and 5, a unit number per layer with an integer uniform distribution between 30 and 200, and a dropout rate with a uniform distribution between 0.1 and 0.5. The acquisition function is selected as the expected improvement function to balance exploration and utilization. The expected improvement function calculates the expected improvement value of each hyperparameter combination compared to the current optimal solution. Bayesian optimization iterates 50 times, evaluating a set of hyperparameter combinations each iteration. The evaluation uses the mean squared error of 5-fold cross-validation as the objective function. The hyperparameter optimization process runs in parallel on multiple GPUs to shorten the search time. Finally, the hyperparameter combination with the smallest cross-validation error is selected to configure the Long Short-Term Memory network architecture. Cross-validation evaluates model performance using temporal cross-validation to maintain the temporal order of the data. Temporal cross-validation divides the dataset into 5 consecutive subsets after sorting by timestamp. Each iteration uses the first 4 subsets as the training set and the last subset as the test set, progressively sliding the validation window backward. Model performance metrics include root mean square error, mean absolute error, mean absolute percentage error, and coefficient of determination. For each metric, the mean and standard deviation are calculated across five test sets. The model parameter solidification process involves retraining once on all training data. After training, the weight matrix and bias terms are saved as binary files. Simultaneously, standardized feature parameters are saved for new data preprocessing. These standardized feature parameters include the mean and standard deviation for each feature dimension.

[0044] The real-time state estimation module loads the electricity meter state assessment model generated by the state assessment model building module. This model is deployed in the embedded inference engine to support real-time computation. The feature similarity matrix calculation uses a dynamic time warping algorithm to align the time series of current electricity meter operating data with the model training data. The dynamic time warping algorithm finds the optimal curved path between the two time series. The similarity score is calculated based on the cumulative distance of the curved path, using the Euclidean distance formula. The feature similarity matrix has an m×n dimension, where m is the number of current data points and n is the number of model reference sequences. Each element of the matrix stores the similarity value of the corresponding sequence pair. A recursive least squares algorithm with multiple forgetting factors sets three forgetting factors to correspond to short-term, medium-term, and long-term trends: a short-term forgetting factor of 0.9 corresponding to the most recent hour's data, a medium-term forgetting factor of 0.95 corresponding to the current day's data, and a long-term forgetting factor of 0.98 corresponding to historical data. The initial value of the covariance matrix in the recursive least squares algorithm is the identity matrix multiplied by 1000, and the gain vector update step size is automatically adjusted according to the information sequence. The recursive estimation of state parameters processes one new data point at a time, and the recursive formula includes prediction and correction steps. The prediction step uses the transition matrix of the state-space model to calculate the prior estimate, and the correction step updates the posterior estimate based on the new observation. The multiple forgetting factor fuses the estimation results from different time scales through a weighted approach, with a weight of 0.6 for short-term estimates, 0.3 for medium-term estimates, and 0.1 for long-term estimates.

[0045] The regularization term employs Tikhonov regularization to prevent matrix ill-conditionedness, and the optimal value of the Tikhonov regularization parameter is determined using the L-curve method. The regularization matrix uses a first-order difference operator to smooth state parameter changes, and the regularization strength parameter is set to 0.01 to balance goodness of fit and parameter smoothness. The matrix condition number is monitored in real-time, and the regularization strength is automatically increased when the condition number exceeds 10,000. Singular value decomposition (SVD) is applied to the matrix inversion process, ignoring singular values ​​less than 0.001 to improve numerical stability. The iterative update process uses an adaptive convergence criterion, with a parameter change rate threshold set to 0.001. After each iteration, the change in the Euclidean norm of the parameter vector is calculated; convergence is considered achieved when the change is below the threshold for five consecutive iterations. The maximum number of iterations is limited to 100 to prevent infinite loops, and a backup estimation algorithm is activated if convergence fails. The backup estimation algorithm uses a sliding window least squares method, with the window size set to the 100 most recent data points. Before outputting stable real-time status parameters of the electricity meter, a rationality check is performed. Parameter value range checks include verifying whether the positive and negative signs and numerical values ​​conform to physical constraints. The real-time status parameters of the electricity meter are updated every 5 minutes, and the uncertainty measure of the parameter estimation is recorded simultaneously. The uncertainty measure is calculated using the trace of the parameter covariance matrix; a quality alarm is triggered when the trace value exceeds a threshold. The status estimation results are stored in a circular buffer to save the historical trajectory for the past 7 days, supporting trend backtracking analysis. The real-time status estimation module and the error prediction module exchange data through shared memory, reducing communication overhead and improving system response speed.

[0046] See Figure 4 In the electricity meter condition assessment system, the state parameter estimation results are visualized through a multi-timescale fusion algorithm. Specifically, the real-time condition estimation module uses a recursive least squares algorithm with multiple forgetting factors to recursively calculate the electricity meter operating data: the short-term estimate (orange dashed line) is based on the most recent hour's data (forgetting factor 0.9), the medium-term estimate (green dotted line) corresponds to the current day's data (forgetting factor 0.95), and the long-term estimate (red dotted line) covers historical data (forgetting factor 0.98). A comprehensive estimate (blue solid line) is generated through weighted fusion (weights of 0.6, 0.3, and 0.1 respectively). The uncertainty interval (light gray shading) is dynamically calculated from the trace of the parameter covariance matrix, reflecting the confidence range of the estimated value.

[0047] Example 3: The error prediction module receives real-time state parameters of the electricity meter from the real-time state estimation module. These parameters include multi-dimensional features such as error coefficients, operational stability indicators, and health scores. A collaborative filtering algorithm constructs a user-item rating matrix to describe the correlation between line loss in the distribution area and the electricity meter's operational error. The user dimension corresponds to different power supply areas, while the item dimension includes variables such as line loss rate, error value, and load characteristics. Matrix imputation technology uses alternating least squares to solve for missing values, fixing one matrix and iteratively optimizing the other. The number of latent factors is set to twenty hidden features captured from the data, and a regularization coefficient of 0.01 is chosen to prevent overfitting. The objective function of the collaborative filtering algorithm is defined as the Frobenius norm of the rating matrix and the reconstructed matrix plus a regularization term. The optimization process uses stochastic gradient descent to update the factor matrix.

[0048] The peak over-threshold model uses a dynamic threshold mechanism to identify out-of-limit events in the error sequence. The dynamic threshold is calculated based on the rolling quantiles of the error values, with a window size of one hundred data points. A quantile level of 95% means that an error value exceeding the historical 95th quantile is considered an out-of-limit event. An out-of-limit event must meet a persistence requirement of exceeding the threshold for three consecutive sampling points to be recorded. A generalized Pareto distribution is used to fit the amplitude distribution of the out-of-limit events. The probability density function of the generalized Pareto distribution is:

[0049]

[0050] in: Indicates the magnitude of the out-of-limit event. The position parameter is taken as a dynamic threshold. The scale parameter describes the degree of dispersion of the distribution. The shape parameter controls the thickness of the distribution tail. The scale parameter is solved using the probabilistic weighted moment method for parameter estimation. and shape parameters Closed solution, position parameters The threshold is fixed as a dynamic threshold. The probability of exceeding the limit is calculated based on a fitted generalized Pareto distribution to obtain the probability value of exceeding a given range. The probability value is mapped to a risk score of 1 to 100 for subsequent processing.

[0051] The error bias parameter sequence generation employs a sliding window analysis to analyze dynamic changes, with the sliding window size set to 24 data points corresponding to a one-day analysis period. Within each window, the mean, variance, autocorrelation coefficient, and trend slope of the error bias parameters are calculated. The window sliding step size is one data point to achieve continuous analysis. The error bias parameter sequence is stored in time series format with timestamps and quality markers. The quality markers indicate the reliability of the data based on goodness-of-fit test results. Outlier detection uses the Grubbs test to identify statistically significant outliers, which are replaced using linear interpolation to maintain sequence continuity. The joint solution operation in the error prediction module introduces a Tikhonov regularization matrix to balance the estimation of bias and variance. The Tikhonov regularization matrix is ​​constructed as a symmetric positive definite matrix with diagonal elements of 1.0 and off-diagonal elements of 0.1. Regularization parameter selection uses leave-one-out cross-validation, where one sample is removed at a time, and the remaining samples are used to train the model to predict the removed sample. The sum of squared prediction errors is used as the objective function to search for the optimal value of the regularization parameter in a grid search, with the search range being 0.001 to 1.0, uniformly selected from fifty points. Gradient descent iteratively updates the estimated values, using a momentum term to accelerate convergence; the momentum coefficient is set to 0.9, and the initial learning rate is 0.01, adaptively adjusted. The loss function is defined as a weighted mean squared error function, with recent errors given a higher weight and an exponential decay coefficient of 0.95 over time. The error bias parameter output interface is designed in a standardized data format, including point estimates, confidence intervals, and timestamps. Data sequence compression storage uses differential encoding to reduce storage space. The real-time prediction function supports on-demand querying of error prediction values ​​at specified time points, and historical data backtracking provides querying of error bias parameter sequences for the past 30 days. The data quality control module monitors the integrity of the input data, triggering a data quality alarm when the missing data ratio exceeds 10%. The prediction model is periodically updated, retraining the model every 7 days, and model version management records the timestamp and performance metrics of each update. Data transmission between the error prediction module and the evaluation strategy parsing module uses an asynchronous message queue, with the message format including a header and payload. The message header records information such as message type, timestamp, and data length. The payload uses Protocol Buffers serialization format to improve transmission efficiency. The error handling mechanism includes retry logic and a dead-letter queue; messages that fail to transmit are transferred to the dead-letter queue for manual processing.

[0052] Module performance monitoring metrics include prediction latency, computational throughput, and memory utilization, with real-time data displayed on the operations dashboard. The alert function has multiple threshold levels; different levels of alerts are sent when error deviation parameters exceed these thresholds. Thresholds are categorized as a 3% attention threshold, a 5% warning threshold, and a 8% severity threshold, corresponding to yellow, orange, and red alert colors. Notification channels support SMS, email, and internal system messages; the recipient list is configurable and can be specified by region and responsibility. A data archiving strategy transfers historical prediction results to a data warehouse, retaining archived data for three years for long-term trend analysis. Access control is role-based, with different roles having different data viewing and operation permissions. System logs record all critical operations, including model updates, parameter adjustments, and data queries, and are retained for six months to meet auditing requirements. The error prediction module is deployed in a container environment using a microservice architecture, supporting horizontal scaling to handle high concurrency requests. A health check interface periodically monitors the module's operational status, automatically restarting the service in case of anomalies to ensure high availability.

[0053] Example 4: The evaluation strategy parsing module receives the error deviation parameter sequence from the error prediction module. This sequence includes fields such as timestamp, error value, confidence interval, and risk score. The time series decomposition operation employs a seasonal trend decomposition method, using the LOESS smoothing algorithm to split the sequence into trend, seasonal, and residual components. The trend component, reflecting the long-term direction of error deviation parameter changes, is extracted using a moving average smoothing window of 30 data points. The seasonal component captures periodic fluctuation patterns with a period length set to 24 data points corresponding to a daily cycle. The residual component represents the random fluctuation portion after removing trend and seasonal factors. Normality tests and autocorrelation analyses are performed on the residual component to assess the sequence's stability. The decomposition results are stored in a structured data format containing the time series data of the three components and quality assessment indicators. The horizontal adjustment strategy is formulated based on the slope and direction of the trend component. The slope of the trend component is calculated using linear regression, with a regression coefficient significance level set at 5%. A positive slope exceeding a threshold of 0.01 triggers an upward adjustment strategy to increase monitoring frequency, while a negative slope below a threshold of -0.01 triggers a downward adjustment strategy to reduce monitoring level. The horizontal adjustment strategy includes monitoring frequency adjustment rules, dynamic threshold setting rules, and early warning triggering rules. The monitoring frequency is divided into five adjustable levels, ranging from per minute to per hour. Dynamic threshold setting is based on the magnitude of trend component changes; the larger the change, the smaller the threshold relaxation and the tighter the threshold. The early warning triggering rules combine the trend duration and rate of change; an early warning is triggered immediately when the duration exceeds three cycles and the rate of change increases.

[0054] The vertical adjustment strategy is designed for the amplitude and phase characteristics of the seasonal components. Amplitude calculation uses the difference between peak and trough values, standardized to the range of zero to one. When the amplitude exceeds 0.5, a reinforced seasonal adjustment model is activated; when the amplitude is below 0.1, a simplified seasonal adjustment model is used. Phase characteristics are identified through peak time distribution analysis to detect mode shifts; if the phase shift exceeds two hours, a phase correction algorithm is activated. The vertical adjustment strategy includes period matching rules, amplitude compensation rules, and phase synchronization rules. The period matching rule compares real-time data with historical models from the same period. The amplitude compensation rule adjusts the compensation coefficient according to the amplitude magnitude, and the phase synchronization rule calibrates timing deviations to maintain model consistency. The lateral adjustment strategy handles the dispersion and anomalies of the residual components. Dispersion is calculated using the ratio of standard deviation to mean. When the dispersion coefficient exceeds 0.3, a robust estimation method is activated to reduce the impact of outliers; when the dispersion coefficient is below 0.1, an accurate estimation method is used to improve sensitivity. Anomalies are described by kurtosis and skewness indices; when the kurtosis is greater than 3, a peak suppression algorithm is activated for distribution spikes. The lateral adjustment strategy includes fluctuation smoothing rules, anomaly handling rules, and distribution correction rules. The fluctuation smoothing rules use a low-pass filter to remove high-frequency noise. The anomaly handling rules perform winsorizing on extreme values, replacing 5% of the first and last values. The distribution correction rules use a Box-Cox transformation to make the distribution approximate a normal distribution.

[0055] The multi-strategy fusion employs a weighted comprehensive evaluation matrix to integrate three adjustment strategies. Rows in the matrix correspond to the adjustment strategy type, and columns correspond to the evaluation indicators. Initial strategy weights are determined based on expert ratings: 0.5 for horizontal adjustment strategies, 0.3 for vertical adjustment strategies, and 0.2 for lateral adjustment strategies. The fusion process calculates the dot product of the score vector and weight vector for each strategy, normalizing the score vector to the range of 0 to 100. The comprehensive state evaluation strategy output is a strategy configuration set containing parameter settings and rule sets, encapsulated in JSON format for easy transmission and parsing. The fuzzy judgment matrix construction involves five core indicators: error stability, trend significance, seasonal intensity, residual normality, and prediction confidence. The relative importance comparison between indicators uses a nine-scale method, which categorizes importance into five levels: equally important, slightly important, significantly important, strongly important, and extremely important. The fuzzy judgment matrix is ​​a 5x5 matrix with diagonal elements set to 1 to maintain reflexivity. The upper triangular matrix is ​​filled in by expert evaluation, while the lower triangular matrix is ​​the reciprocal of the matrix to maintain reciprocal symmetry. The consistency check calculates a consistency ratio value that must be less than 0.1. If this is not met, the matrix elements are readjusted until the consistency requirement is met. The eigenvector method is used to solve for the weight vector using a power iteration algorithm. This algorithm initializes random positive vectors and iteratively calculates the eigenvector corresponding to the largest eigenvalue of the matrix. The number of iterations is set to one hundred, with a precision requirement of 0.0001. The eigenvectors are normalized so that the sum of all components is 1. The weight vector output is retained to four decimal places to meet the precision requirements. Sensitivity analysis is performed on the weight vector to evaluate the impact of changes in the indicators on the weights.

[0056] Referring to Table 1, the variable weighting function introduces a data freshness index to dynamically adjust the weights. The data freshness index is calculated based on the data collection time and the current time interval. A freshness index of 1 is assigned to intervals less than one hour, and 0 is assigned to intervals greater than twenty-four hours, with a linear decay in between. The variable weighting function is designed as a product of the baseline weight and the freshness index. A weight decay mechanism is activated when the freshness index falls below 0.5. The variable weighting function parameters include a decay coefficient and a freshness threshold; the decay coefficient is set to 0.9 for a smooth transition. The final weight set is generated after adjustment using the variable weighting function and is updated every ten minutes to reflect the latest data status.

[0057] Table 1: Importance Assessment Table of Fuzzy Judgment Matrix Indicators

[0058]

[0059] The strategy weighted fusion uses a weighted average method to synthesize the outputs of various strategies, with the weights derived from the final weight set. The fusion result is normalized and mapped to an evaluation score from 0 to 100. The evaluation score is divided into four levels: Excellent (above 85 points), Good (70-85 points), Satisfactory (60-70 points), and Unsatisfactory (below 60 points). After the status evaluation strategy is generated, it is stored in the strategy library and marked with a version number. The strategy library retains historical versions and supports rollback operations. The strategy activation mechanism has a time delay, setting new strategies to take effect one hour later. Simulation verification is performed before activation to check the strategy's effectiveness. The evaluation strategy parsing module and the evaluation result generation module communicate via a message middleware. The message format includes fields such as strategy type, parameter list, and activation time. The exception handling mechanism monitors errors during strategy execution. When the number of errors exceeds a threshold, a backup strategy is switched to ensure continuous system operation.

[0060] See Figure 5 This study illustrates the distribution of percentage improvements in electricity meter error under different adjustment strategies. The horizontal adjustment strategy achieved a 15% improvement, the vertical adjustment strategy achieved a 10% improvement, the lateral adjustment strategy achieved an 8% improvement, and the combined strategy achieved a 28% improvement. This demonstrates that the combined application of horizontal, vertical, and lateral adjustment strategies can achieve more significant results in improving electricity meter error, showcasing the advantages of multi-strategy integration in electricity meter condition assessment and error optimization, and providing data support for strategy selection in electricity meter error management.

[0061] Example 5: The evaluation result generation module receives the state evaluation strategy from the evaluation strategy parsing module and the real-time state parameters of the electricity meter from the real-time state estimation module. These real-time state parameters include twelve dimensions of features such as error coefficient, stability index, runtime, and environmental adaptability. The K-means clustering algorithm initializes four cluster centers corresponding to four categories of health levels: healthy state, sub-healthy state, warning state, and fault state. The clustering features are selected using the first three principal components after dimensionality reduction using principal component analysis (PCA), which retains over 90% of the variance information of the original features. The iterative process of the K-means clustering algorithm calculates the Euclidean distance from each data point to the cluster center and assigns the data points to the nearest cluster. The cluster centers are updated to the mean vector of the data points within each cluster. The iteration stops when the change in cluster centers is less than 0.001 or when the maximum number of iterations (100) is reached. Health level labels are assigned based on the cluster center feature values; the cluster with the best feature value is labeled as healthy, and the cluster with the worst feature value is labeled as fault state. The rolling forecast interval generation employs a bootstrap aggregation algorithm combined with an autoregressive integral moving average (AMA) model. The bootstrap aggregation algorithm randomly selects a subset of 1000 samples with replacement from the error deviation parameter sequence. Each subset trains an AMA model, whose parameters are determined using the Akaike information criterion. The forecast interval is calculated using the 2.5% and 97.5% quantiles of the 1000 predicted values, forming a 95% confidence interval. The rolling window size is set to 30 time points, updating the forecast results by sliding the window forward one time point at a time. The forecast interval width is dynamically adjusted to reflect forecast uncertainty; a forecast reliability warning is triggered when the interval width exceeds a threshold. The Value at Risk (VaR) metric is calculated based on the distribution characteristics of historical error deviation parameters, with a 95% confidence level and a one-day time period. The historical simulation method uses error data from the past 250 trading days to construct an empirical distribution, calculating the 5th percentile of this distribution as the VaR metric value. The expected shortage indicator simultaneously calculates the average loss exceeding the VaR metric value; the expected shortage indicator is the conditional expected value at the tail of the loss distribution. The Value at Risk (VaR) measure is updated daily and risk exposure is adjusted based on current market volatility.

[0062] The visualization platform unit utilizes a WebGL-based graphics rendering engine to implement an interactive dashboard. The dashboard layout is divided into four main areas: a health status overview area, an error prediction area, a risk analysis area, and a detailed information area. The health status overview area displays the current health level in a dashboard format, with the pointer color changing according to the health status: green for healthy and red for faulty. The error prediction area plots a time-series curve showing historical error values ​​and future prediction ranges, with a semi-transparent band used to enhance visualization of the prediction range. The risk analysis area uses a heatmap to display the value-at-risk (VAT) measure at different confidence levels, with a color gradient from green to red indicating increasing risk. The detailed information area lists the specific values ​​of all feature parameters in a table format, supporting sorting by time and feature filtering. A dynamic update mechanism uses the WebSocket protocol to push data in real time, with a configurable update frequency supporting refreshes as low as seconds. A mouse hover tooltip displays the specific value and time point on the error prediction curve; clicking on a data point allows drill-down to view a detailed analysis report for that time point. The chart zoom function supports dynamic adjustment of the time range, with eight scales ranging from minutes to months. Multi-dimensional data filtering provides joint queries based on conditions such as time range, regional distribution, and electricity meter type, and the query results are refreshed instantly without page reload.

[0063] The drilling function features a three-level drilling depth design to meet different levels of analysis needs. The first level displays a summary of health levels, including the number and proportion of energy meters in each state. The second level displays detailed operating parameter trend curves for individual energy meters. The third level provides raw data queries and abnormal event logs. Each drilling level maintains a navigation path for easy user backtracking, and the drilling operation response time is controlled within 0.5 seconds to ensure a good user experience. The data export function supports three formats: PNG images, PDF reports, and Excel data. The exported content maintains a visually consistent appearance with the screen display. User access management is based on a role-based access control model, with different roles having different data viewing and operation permissions. Operation logs record all user interactions, including queries, drilling, and exports, and are retained for three years to meet auditing requirements. The system integration interface provides a RESTful API for other systems to call, and the API interface uses the OAuth 2.0 protocol for authentication. Performance monitoring tracks key indicators such as page load time and data query response time in real time, and sets threshold alarms to automatically notify maintenance personnel when performance indicators are abnormal. Mobile adaptation adopts a responsive layout design to maintain optimal display effects on screens of different sizes.

[0064] The visualization theme offers two color schemes: light and dark, which users can freely switch between depending on the environment. Accessibility features cater to colorblind users by adding pictorial markers alongside color differentiation. Multilingual support includes both Chinese and English interface languages, with instant language switching without page refresh. Adjustable font sizes cater to different reading habits, with four adjustment levels ranging from 12px to 20px. A tiered caching architecture is used for data caching, storing frequently accessed data in memory to improve access speed. Historical data is stored in a distributed file system, with indexes optimizing query performance. A full backup is performed daily at midnight, retaining backup data for 30 days. The disaster recovery solution is designed with a multi-active architecture, automatically switching to a backup center when one data center fails. A user feedback channel is provided in the lower right corner of the interface, allowing users to submit user experience suggestions and problem reports. Version updates use a canary release strategy, gradually expanding the scope of new features to a small number of users first. Usage statistics collect anonymized user behavior data to improve system functionality and user experience. Help documentation is integrated into the system, providing detailed function descriptions and operation guides. The online customer service function provides real-time consultation during working hours, and switches to message mode outside of working hours. The system announcement bar displays the latest updates and maintenance notices, and important announcements use pop-up reminders to ensure users are aware of them.

[0065] 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.

[0066] 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 data-driven system for assessing the condition and predicting errors of electricity meters, characterized in that, The system includes an energy meter operation data acquisition module, a state assessment model construction module, a real-time state estimation module, an error prediction module, an assessment strategy parsing module, and an assessment result generation module. The electricity meter operation data acquisition module is configured to collect dynamic and static operation data sets of the electricity meter; The state assessment model construction module is configured to construct an energy meter state assessment model based on the dynamic and static operation dataset. The real-time state estimation module is configured to use the energy meter state assessment model to perform state matching and parameter estimation on the current energy meter operating data to obtain the real-time state parameters of the energy meter. The error prediction module is configured to predict the error trend based on the real-time status parameters of the energy meter and output the error deviation parameters. The evaluation strategy parsing module is configured to judge the changing trend of the error deviation parameter and analyze the strategy to generate a state evaluation strategy. The evaluation result generation module is configured to analyze the real-time status parameters and error deviation parameters of the electricity meter based on the status evaluation strategy, and generate electricity meter status evaluation results and error prediction results.

2. The data-driven energy meter condition assessment and error prediction system as described in claim 1, characterized in that, The electricity meter operation data acquisition module acquires dynamic and static operation datasets through the following operations: it acquires electricity time-series data and metering curves from multiple electricity meter devices in parallel as raw data; it performs real-time stream processing on the acquired raw data, including data verification and format standardization; it applies anomaly detection algorithms to identify and remove outliers; it uses interpolation methods to complete missing data; it extracts key features through principal component analysis to reduce the dimensionality of the execution data; and finally, it uses standardization algorithms to scale the data to a uniform range, forming a standardized dynamic and static operation dataset.

3. The data-driven energy meter condition assessment and error prediction system as described in claim 1, characterized in that, The state assessment model construction module constructs the energy meter state assessment model through the following steps: combining the temporal characteristics of energy meter data, a long short-term memory network architecture is designed in a targeted manner; Time series feature indicators, including sliding window statistics and trend components, are calculated from dynamic and static running datasets; a long short-term memory network architecture is initialized for meter value prediction; an adversarial transfer learning algorithm is used to adapt the pre-trained model to the new electricity meter data; network hyperparameters are adjusted using a Bayesian optimization method; and model performance is evaluated using cross-validation, ultimately solidifying the model parameters.

4. The data-driven energy meter condition assessment and error prediction system as described in claim 1, characterized in that, The real-time state estimation module obtains the real-time state parameters of the electricity meter through the following operations: calculating the feature similarity matrix between the current electricity meter operating data and the electricity meter state assessment model; A recursive least squares algorithm with multiple forgetting factors is applied to recursively estimate the state parameters; a regularization term is introduced to prevent matrix ill-conditioning; and stable real-time state parameters of the energy meter are output through iterative updates until the parameters converge.

5. The data-driven energy meter condition assessment and error prediction system as described in claim 1, characterized in that, The error prediction module outputs error deviation parameters through the following steps: using a collaborative filtering algorithm to jointly solve for the line loss of the transformer area and the operating error of the electricity meter; and combining the continuous data of the electricity meter operating error arranged in chronological order into an error sequence. The peak-to-threshold model is used to extract out-of-limit events from the error sequence; the generalized Pareto distribution is fitted to calculate the out-of-limit probability; the error deviation parameter sequence is generated and its dynamic changes are analyzed through a sliding window.

6. The data-driven energy meter condition assessment and error prediction system as described in claim 5, characterized in that, The joint solution operation in the error prediction module is further optimized through the following steps: constructing a Tikhonov regularization matrix to balance the estimated bias and variance; using leave-one-out cross-validation to select regularization parameters; and iteratively updating the estimated values ​​using gradient descent until the error is minimized.

7. The data-driven energy meter condition assessment and error prediction system as described in claim 1, characterized in that, The evaluation strategy parsing module generates the state evaluation strategy through the following operations: performing time series decomposition on the error deviation parameter to extract the trend component, seasonal component and residual component; Horizontal adjustment strategies are formulated based on trend components, vertical adjustment strategies are formulated based on seasonal components, and lateral adjustment strategies are formulated based on residual components. Integrate multiple strategies to form a comprehensive state assessment strategy.

8. The data-driven energy meter condition assessment and error prediction system as described in claim 7, characterized in that, The strategy formulation operation in the evaluation strategy analysis module is further enhanced by the following steps: establishing a fuzzy judgment matrix to evaluate the importance of each indicator; The initial weights are calculated using the eigenvector method; the data freshness index reflects the timeliness of the data, and the variable weight function is a mathematical function that dynamically adjusts the weights accordingly; the weights are dynamically adjusted based on the data freshness index combined with the variable weight function. The final set of output weights is used for policy weighted fusion.

9. The data-driven energy meter condition assessment and error prediction system as described in claim 1, characterized in that, The evaluation result generation module generates the energy meter status evaluation result and error prediction result through the following operations: using the K-means clustering algorithm to divide the real-time status parameters of the energy meter into health levels; and generating a rolling prediction interval based on the error deviation parameter. Calculate the risk of error in the Value at Risk (VaR) metric assessment; output a complete assessment result including health level labels and prediction ranges.

10. The data-driven energy meter condition assessment and error prediction system as described in claim 9, characterized in that, The evaluation result generation module also integrates a visualization platform unit, which displays the results through the following steps: rendering an interactive dashboard to show the status evaluation results; dynamically updating the error prediction curve; providing drill-down functionality for detailed analysis; and supporting multi-dimensional data filtering and real-time updates.

Images