Error self-correction method and system for power meter cluster collaborative metering

Abstract

The application discloses an error self-correction method and system for cluster coordination metering of electric energy meters, and relates to the technical field of electric power metering. The method comprises the following steps: collecting historical error event data of the electric energy meter cluster, identifying and classifying a plurality of error types; extracting correction features of the error types, and performing transfer learning on the initialized correction channel based on the correction features to obtain a plurality of self-correction channels; reading a coordination metering data set to identify real-time errors, and clustering the real-time errors according to error types; establishing a connection between the clustering results and the self-correction channels, and performing error self-correction. The application solves the technical problem that the existing error correction method for the electric energy meter cluster cannot respond to dynamically changing error types in real time, resulting in insufficient electric power metering accuracy, and achieves the technical effect of improving electric power metering accuracy through real-time error identification and dynamic correction of the self-correction channel to realize dynamic error self-correction of the electric energy meter cluster.

Description

Technical Field

[0001] This invention relates to the field of power metering technology, specifically to an error self-correction method and system for collaborative metering of electricity meters. Background Technology

[0002] In modern power systems, electricity meter clusters are widely used in smart grids, enabling large-scale electricity metering and data acquisition. However, with the increasing number of meters and the increasing complexity of operating conditions, such as environmental changes, equipment aging, and communication interference, various errors often occur in the metering process. These errors not only affect the accuracy of electricity metering but may also affect the stability of the power grid. Traditional error correction methods typically rely on manual detection and periodic calibration, which are inefficient and cannot respond to error changes in real time, making it difficult to meet the requirements of high-precision and high-efficiency metering. Summary of the Invention

[0003] This application provides a method and system for error self-correction in collaborative metering of electricity meter clusters, which solves the technical problem that existing error correction methods for electricity meter clusters cannot respond in real time to dynamically changing error types, resulting in insufficient accuracy of electricity metering.

[0004] The first aspect of this application provides an error self-correction method for collaborative metering of an energy meter cluster. The method includes: collecting a historical error event set of the energy meter cluster and identifying multiple error types in the historical error event set; extracting multiple sets of error correction features for the multiple error types, and performing transfer learning on an initial correction channel based on the multiple sets of error correction features to obtain multiple self-correction channels, wherein each set of error correction features includes correction control rate, correction control duration, correction control gradient, and correction control stability; reading the collaborative metering dataset of the energy meter cluster, performing real-time error identification based on the collaborative metering dataset, and clustering the real-time error types according to the multiple error types to obtain multiple clustering results; establishing connections between the multiple clustering results and the multiple self-correction channels, and performing error self-correction based on the multiple self-correction channels and the multiple clustering results.

[0005] A second aspect of this application provides an error self-correction system for collaborative metering of a cluster of electricity meters. The system includes: an error type identification module for collecting a historical set of error events from the electricity meter cluster and identifying multiple error types from the historical error event set; a transfer learning module for extracting multiple sets of error correction features from the multiple error types, and performing transfer learning on an initialization correction channel based on the multiple sets of error correction features to obtain multiple self-correction channels, wherein each set of error correction features includes correction control rate, correction control duration, correction control gradient, and correction control stability; a real-time error clustering module for reading the collaborative metering dataset of the electricity meter cluster, performing real-time error identification based on the collaborative metering dataset, and clustering the real-time error types according to the multiple error types to obtain multiple clustering results; and an error self-correction module for establishing connections between the multiple clustering results and the multiple self-correction channels, and performing error self-correction based on the multiple self-correction channels.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0007] The error self-correction method and system for collaborative metering of electricity meter clusters provided in this application relate to the field of power metering technology. By collecting historical error events of the electricity meter cluster, multiple error types are identified and correction features are extracted. Based on these features, transfer learning is performed on the initial correction channel to obtain multiple self-correction channels. Through real-time error identification and clustering, a connection is established between the clustering results and the self-correction channels, realizing dynamic error self-correction of the electricity meter cluster. This solves the technical problem that existing electricity meter cluster error correction methods cannot respond to dynamically changing error types in real time, leading to insufficient power metering accuracy. It achieves dynamic error self-correction of the electricity meter cluster through real-time error identification and dynamic correction of the self-correction channels, thereby improving the technical effect of power metering accuracy. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A schematic flowchart of the error self-correction method for collaborative metering of electricity meter clusters provided in this application embodiment;

[0010] Figure 2 This is a schematic diagram of the error self-correction system for collaborative metering of electricity meters provided in an embodiment of this application.

[0011] Figure labeling: Error type identification module 11, transfer learning module 12, real-time error clustering module 13, error self-correction module 14. Detailed Implementation

[0012] This application provides a method and system for error self-correction in collaborative metering of electricity meter clusters, which solves the technical problem that existing error correction methods for electricity meter clusters cannot respond in real time to dynamically changing error types, resulting in insufficient accuracy of electricity metering.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, this application provides an error self-correction method for collaborative metering of electricity meters clusters, the method comprising:

[0016] P10: Collect the historical error event set of the electricity meter cluster and identify multiple error types in the historical error event set. These multiple error types include metering proportional error, clock synchronization error, temperature drift system error, dynamic error caused by load changes, and offset error due to communication link anomalies.

[0017] Specifically, the first step is to collect a set of historical error events from the electricity meter cluster and conduct a detailed analysis of these events to identify the various error types. Collecting this set of error events is a crucial preliminary step, providing data support for subsequent error type identification and error correction implementation.

[0018] First, the collection of historical error events relies on the monitoring or data acquisition system of the electricity meter cluster. To ensure data integrity and accuracy, the acquisition system periodically retrieves data from each electricity meter in the cluster. These meters typically use built-in sensors and metering chips to record energy consumption and various errors that may occur during measurement, including error event data such as error timestamps, error types, and error magnitudes. This data provides a detailed reflection of the meter's performance in actual operation. During the collection process, a standardized data format can be used for recording, ensuring data integrity and accuracy. Data can be aggregated locally or transmitted to a backend data platform via a remote communication system.

[0019] The collected historical error event set contains various types of errors, which can affect the metering accuracy of electricity meters to varying degrees. This application primarily identifies the following error types: metering proportional error, clock synchronization error, temperature drift system error, dynamic error caused by load changes, and offset error due to communication link anomalies. Among these, metering proportional error refers to a fixed proportional deviation between the metered value and the actual electricity consumption value, typically caused by inaccuracies in the internal metering chip. To identify metering proportional error, it is necessary to compare and analyze the actual measured value of the electricity meter with a known standard value. By calculating the proportional deviation of the error, the existence of this error type can be accurately determined.

[0020] Clock synchronization error refers to the difference between the time recorded by the electricity meter and the actual time. To detect this type of error, it is necessary to compare the meter's measurement time with a standard time source, such as GPS time or network time. By analyzing time-related deviations in historical error data, errors caused by clock inaccuracy can be identified.

[0021] Temperature drift systematic error is a systematic error caused by the influence of temperature changes on the performance of the metering components of an electricity meter. As ambient temperature changes, the performance of the meter's sensors may change, leading to deviations in measurement results. For example, the sensitivity of voltage and current sensors may differ in high and low temperature environments, resulting in lower measurement accuracy. To identify temperature drift error, it is necessary to analyze the error performance of the electricity meter under different ambient temperatures and compare the relationship between temperature changes and error amplitude. If the error shows a strong correlation with temperature changes, it is very likely caused by temperature drift error.

[0022] Dynamic errors typically occur when there are significant load changes. The measurement accuracy of an electricity meter can be affected by load fluctuations, especially when the load changes drastically. The meter's response may not be timely, resulting in dynamic errors. For example, in cases of a sudden increase or decrease in load, the meter's response time may be insufficient, leading to low measurement accuracy. To detect this type of error, the load changes of the electricity meter and the corresponding metering errors can be analyzed to identify dynamic errors.

[0023] Offset errors due to communication link anomalies occur because of abnormalities in the communication link during data transmission, leading to deviations or distortions in the metering data. Electricity meter clusters typically rely on wireless or wired communication networks to transmit measurement data to a central system. During transmission, any communication anomalies can affect data accuracy. Communication link anomalies may cause error shifts, data loss, or delays, all of which impact the metering results. To identify this type of error, one can analyze indicators such as the quality of electricity meter data transmission, latency, and packet loss rate, combined with the timestamps of error events, to determine if a communication link anomaly exists.

[0024] For example, to accurately identify the various error types mentioned above, this application employs a machine learning-based classification algorithm. First, the collected historical error event data is preprocessed to remove noise and outliers, and then normalized so that different types of error data can be compared and analyzed on the same scale. Next, features related to the error type are extracted from the preprocessed data, such as the proportional deviation between the measured value and the standard value for metering ratio errors, and the time deviation between the electricity meter clock and the standard time source for clock synchronization errors. Then, the labeled historical error event data is used as training samples to train a multi-class machine learning model. This model can employ algorithms such as Support Vector Machines, Random Forests, or Convolutional Neural Networks. During training, the model's classification performance is optimized by adjusting its hyperparameters, such as the penalty parameter C and kernel function parameter γ in Support Vector Machines, and the number and depth of trees in Random Forests. Finally, the trained model is used to identify each type of error individually.

[0025] By identifying the above five common error types, the errors of electricity meter clusters can be classified in detail, providing an accurate basis for subsequent error correction.

[0026] P20: Extract multiple sets of error correction features for the multiple error types, and perform transfer learning on the initial correction channel based on the multiple sets of error correction features to obtain multiple self-correction channels. Each set of error correction features includes correction control rate, correction control duration, correction control gradient, and correction control stability.

[0027] Optionally, after collecting the historical error event set of the electricity meter cluster and identifying the error types, the next step is to extract multiple sets of error correction features corresponding to various error types. These features will serve as important parameters in the subsequent correction process. Extracting each set of error correction features involves in-depth analysis of different error types to ensure precise control of various error correction operations during the correction process. These features include correction control rate, correction control duration, correction control gradient, and correction control stability, all of which are important factors affecting the correction effect.

[0028] The correction control rate defines the magnitude of error correction per unit time. For example, for proportional measurement error, if the error ratio is high, a higher correction control rate may be needed to quickly adjust the measurement value; while for temperature drift system error, since its change is relatively slow, the correction control rate can be set lower to avoid overcorrection.

[0029] The calibration control duration determines the length of time each calibration operation lasts. Different error types may require different calibration times. For example, for dynamic errors, due to the unpredictability of load changes, a longer time may be needed to gradually adjust the meter's measurements to restore accuracy. In contrast, proportional metering errors, due to their direct and stable nature, may require only a shorter calibration time.

[0030] The correction control gradient represents the magnitude of adjustment during each correction process; in other words, it's the rate of change of the error correction relative to the current error during each correction operation. For some error types, such as dynamic errors caused by load changes, which exhibit large load fluctuations, a larger correction gradient may be needed to quickly respond to load changes. Conversely, for other error types, such as temperature drift errors, the correction process may be more gradual, requiring a smaller correction gradient for stable adjustment. The correction gradient should be flexibly adjusted according to the characteristics of each error type to ensure that the correction process can correct errors promptly without over-adjusting and causing oscillations.

[0031] Correction control stability measures the stability of the control signal during the correction process. A stable correction control signal reduces fluctuations during correction and avoids introducing new errors due to overcorrection or unstable adjustments. For example, the correction of clock synchronization errors should have high stability because frequent clock adjustments can lead to significant data timestamp offsets. For dynamic errors with large load fluctuations, the correction process can tolerate larger fluctuations, but it must still ensure that it eventually returns to a stable state.

[0032] Based on the extracted multiple sets of error correction features, this application employs transfer learning to optimize the initial correction channels, thereby generating multiple self-correcting channels. Transfer learning is a machine learning technique that allows knowledge learned from one task (source task) to be transferred to another related task (target task). In this application, the source task is to optimize correction channels based on known error types and correction features, while the target task is to generate specialized self-correcting channels for each error type.

[0033] For example, firstly, an initial calibration channel model is established, containing basic calibration control parameters and algorithms. The purpose of initializing the calibration channel is to provide a starting point for transfer learning; its parameters can be preset according to general calibration needs. Then, the extracted error calibration features are mapped to the parameter space of the initial calibration channel. This process can be achieved through a feature transformation function, ensuring that features of different error types can be adapted to the parameter system of the calibration channel. Next, using a transfer learning algorithm, the calibration knowledge learned from the source task, such as calibration strategies and parameter optimization methods for specific error types, is transferred to the target task. Transfer learning algorithms can employ methods such as instance-based transfer learning, feature-based transfer learning, or model-based transfer learning.

[0034] In this application, a model-based transfer learning method is chosen. By adjusting the parameters of the initial calibration channel, it is adapted to the calibration requirements of different error types. Specifically, during the transfer learning process, the parameters of the calibration channel are optimized to ensure that each self-calibrating channel can effectively correct errors according to its corresponding error type and calibration characteristics. The optimization process can be implemented through iterative algorithms, such as gradient descent or genetic algorithms, to minimize calibration errors and improve calibration accuracy. After transfer learning and parameter optimization, a dedicated self-calibrating channel is generated for each error type. Each self-calibrating channel has specific calibration control parameters for its corresponding error type, enabling precise error correction. For example, the self-calibrating channel for metering proportionality errors will have a high calibration control rate and a moderate calibration control duration; while the self-calibrating channel for temperature drift system errors will have a lower calibration control rate and a longer calibration control duration to adapt to the slow characteristics of temperature changes.

[0035] Through the above process, this application can extract corresponding error correction features for various error types in electricity meter clusters and generate multiple self-correction channels using transfer learning technology. This process not only ensures the relevance and effectiveness of the correction channels but also improves the automation and accuracy of electricity meter cluster error self-correction, providing reliable technical support for achieving high-precision collaborative metering of electricity meter clusters.

[0036] Furthermore, based on the multiple sets of error correction features, transfer learning is performed on the initial correction channel to obtain multiple self-correction channels. In this embodiment, step P20 further includes:

[0037] P21: Determine the first set of error correction features corresponding to the first error type based on the frequency of occurrence of the multiple error types corresponding to the multiple sets of error correction features; P22: Perform initialization learning according to the first set of error correction features to obtain the first self-correction channel corresponding to the first error type, and mark the first self-correction channel as the initialization correction channel; P23: Calculate the Q feature difference distribution matrices between the first set of error correction features after standardization and the remaining set of error correction features; P24: Using the initialization correction channel as the source domain and the remaining self-correction channels as the target domain, perform transfer learning according to the Q feature difference distribution matrices to obtain multiple self-correction channels.

[0038] It should be understood that the process of transfer learning on the initial correction channel based on multiple sets of error correction features can be further refined to achieve more accurate self-correction channel construction.

[0039] First, based on the frequency of occurrence of multiple error types corresponding to multiple sets of error correction features, the first set of error correction features corresponding to the first error type is determined. In this stage, by analyzing the frequency of occurrence of different error types in historical error events, it is assessed which error types are most common in the operation of the energy meter cluster. Frequently occurring error types will be given higher priority because the correction of these types of errors is crucial to improving the overall metering accuracy of the energy meter cluster. For example, the error type with the highest frequency in the energy meter cluster is identified as the first error type, and a self-correction channel is built preferentially for this main error type. The frequency statistics can be completed by analyzing data from a set of historical error events. For example, if the metering ratio error occurs most frequently in historical data, it is identified as the first error type, and its corresponding error correction features are extracted as the first set of error correction features.

[0040] Next, initial learning is performed according to the first set of error correction features to obtain the first self-correction channel corresponding to the first error type, and this channel is marked as the initial correction channel. This process can be regarded as the initial training stage of the model, using the first set of features to initialize and train the error correction model. For example, a network model architecture suitable for error correction is first constructed. This architecture is based on a deep learning framework, such as a convolutional neural network or a recurrent neural network. Its input consists of metering data and operating status information related to the first error type, and its output is the corrected metering data. The design of the network model needs to consider the complexity and real-time requirements of error correction to ensure efficient processing and correction of metering errors in the electricity meter cluster.

[0041] Next, a correction feature loss constraint function is introduced, including rate, duration, gradient, and stability loss functions, to optimize the network model's performance and adapt it to the correction requirements of different error types. Before training, the data is preprocessed, including noise removal, missing value imputation, and normalization, to improve data quality. Subsequently, the first set of error correction features is used as training data, and the model is trained through forward and backpropagation until the loss value no longer decreases significantly or the preset number of iterations is reached. After training, the model performance is evaluated using a validation dataset, and its effectiveness is verified through metrics such as accuracy, recall, and F1 score. Finally, the correction channels corresponding to the trained network model are marked as initial correction channels, serving as the starting point for further transfer learning.

[0042] Next, Q feature difference distribution matrices are calculated between the standardized first group of error-corrected features and the remaining groups of error-corrected features. In this step, the first group of error-corrected features and the other remaining features are first standardized to eliminate differences in dimensions and scales, ensuring they can be compared and computed within the same model. Then, the differences between these features are analyzed by calculating the Euclidean distance, cosine similarity, or other statistics between each group of features and the first group of error-corrected features, resulting in Q feature difference distribution matrices, where each matrix corresponds to a feature difference distribution for a set of remaining error types. These feature difference distribution matrices will provide the necessary input data for subsequent transfer learning, helping the model to efficiently transfer knowledge between the source and target domains.

[0043] Finally, using the initial calibration channel as the source domain and the remaining self-calibration channels as the target domain, transfer learning is performed according to the calculated Q feature difference distribution matrices. The transfer learning process transfers knowledge from the source domain (initial calibration channel) to the target domain (self-calibration channels for other error types). Specifically, based on each feature difference distribution matrix, the parameters of the corresponding target domain self-calibration channel are adjusted to adapt to the calibration requirements of the corresponding error type. The transfer learning algorithm can employ a feature-based transfer learning method, minimizing the difference between the target domain model and the source domain model by adjusting the parameters of the target domain model, while considering the feature differences reflected in the feature difference distribution matrix, to achieve accurate calibration for different error types. In this way, multiple self-calibration channels will be optimized under the calibration requirements of different error types, thereby improving the overall calibration capability of the electricity meter cluster.

[0044] Furthermore, in this embodiment of the application, step P22 further includes obtaining the first self-calibration channel corresponding to the first error type:

[0045] P22-1: Initialize the network model architecture. The network model architecture introduces a correction feature loss constraint function to initialize and learn the first set of error correction features, thereby obtaining the first self-correcting channel that has been learned to convergence. P22-2: Wherein, the correction feature loss constraint function includes a rate loss constraint function, a duration loss constraint function, a step size loss constraint function, and a stability loss constraint function. The rate loss constraint function is used to adjust the depth of the network model architecture. The duration loss constraint function is used to adjust the temporal attention interval of the network model architecture. The step size loss constraint function is used to adjust the gradient descent magnitude of the network model architecture. The stability loss constraint function is used to adjust the regularization parameter of the network model architecture.

[0046] Optionally, to further optimize the process of obtaining the first self-correction channel corresponding to the first error type, the initialization of the network model architecture and the correction feature loss constraint function can be introduced to ensure that the model can perform accurate adaptive correction for the selected error type.

[0047] First, the network model architecture is initialized, providing a basic structural framework for the first self-calibrating channel. The choice of network model architecture should be adjusted according to the complexity of the calibration task and the characteristics of the error type. During initialization, the network architecture will be initialized by introducing different calibration feature loss constraint functions. Specifically, the initialized network architecture should contain multiple layers to effectively handle multi-dimensional error features. The model training process will learn from these features, gradually adjusting the network parameters to ultimately achieve error correction. This process optimizes the model through backpropagation until the model's loss function converges, resulting in a first self-calibrating channel that has been learned and reached a stable state.

[0048] The calibration feature loss constraint function is a crucial tool for ensuring that the network model can accurately correct based on error characteristics. The introduction of this loss constraint function helps the model adjust to different error characteristics during training, ensuring that the correction process considers not only the magnitude of the error but also other correction requirements. Specifically, these loss constraint functions include rate loss constraint function, duration loss constraint function, step size loss constraint function, and stability loss constraint function, each designed with a specific purpose.

[0049] The rate loss constraint function adjusts the depth of the network model architecture, adapting to the rate requirements of the correction control by controlling the number of layers and neurons. The duration loss constraint function adjusts the temporal attention range of the network model architecture, adapting to the duration requirements of the correction control by controlling the network's focus on time-series data. The step size loss constraint function adjusts the gradient descent magnitude of the network model architecture, adapting to the gradient requirements of the correction control by controlling the step size of each parameter update. The stability loss constraint function adjusts the regularization parameters of the network model architecture, adapting to the stability requirements of the correction control by controlling the regularization strength. Through the combined effect of these loss constraints, the network model architecture can effectively learn and optimize for the first set of error correction features.

[0050] Next, initial learning is performed until the network model converges, thus obtaining the first self-correcting channel. During the initial learning process, the network model calculates the predicted values ​​of the error correction features through forward propagation and updates the network parameters through backpropagation. In each iteration, the network model calculates the loss value according to the aforementioned correction feature loss constraint function and adjusts the network parameters based on this loss value. When the loss value of the network model no longer decreases significantly, or when the preset number of iterations is reached, the network model is considered to have converged. At this point, the network model is able to generate an effective first self-correcting channel based on the first set of error correction features, which can perform accurate error correction for the first error type.

[0051] Through the loss constraint function described above, the network model can dynamically adjust its parameters and structure during training, ensuring high accuracy and stability in correction tasks of different types of errors.

[0052] Furthermore, calculating the Q feature difference distribution matrix between the first group of error correction features and the remaining group of error correction features after standardization, step P23 of this embodiment also includes:

[0053] P23-1: Introduce a high-dimensional feature space, and perform radial basis function nonlinear mapping on the first group of error correction features and the remaining group of error correction features after standardization according to the high-dimensional feature space to obtain the first group of high-dimensional error correction features and the remaining group of high-dimensional error correction features; P23-2: Recalculate Q feature difference distribution matrices based on the first group of high-dimensional error correction features and the remaining group of high-dimensional error correction features.

[0054] Specifically, the calculation process of the feature difference distribution matrix can be further optimized. First, a high-dimensional feature space is introduced, and the first group of error correction features and the remaining groups of error correction features after standardization are nonlinearly mapped using radial basis functions (RBFs). RBFs are a commonly used nonlinear mapping method that can map low-dimensional data to a high-dimensional space, thereby increasing the separability of the data. Through this nonlinear mapping, the original error correction features can be transformed into high-dimensional error correction features. Specifically, for the first group of error correction features and each remaining group of error correction features, radial basis functions are applied for mapping to obtain the first group of high-dimensional error correction features and the remaining groups of high-dimensional error correction features. Through RBF nonlinear mapping, the error features originally in the low-dimensional space are transformed into high-dimensional error correction features. In the high-dimensional space, the features of different error types are more apparent, helping the system to better understand and distinguish the features of different error types.

[0055] Next, Q feature difference distribution matrices are recalculated based on the first group of high-dimensional error correction features and the remaining groups of high-dimensional error correction features. In the high-dimensional feature space, the differences between features can be quantified by calculating Euclidean distance, cosine similarity, or other statistics. Here, taking Euclidean distance as an example, the distance between two high-dimensional error features is calculated as follows: for each pair of features in the first group of high-dimensional error correction features and the remaining groups of high-dimensional error correction features, their Euclidean distance in the high-dimensional space is calculated. This yields the difference degree between each pair of features. Then, the difference degrees between multiple features are represented in matrix form, generating Q feature difference distribution matrices. The Q feature difference distribution matrices reflect the relationships between all error feature groups.

[0056] In transfer learning, these difference matrices help the model identify and utilize the differences between features in the source and target domains, thereby achieving efficient feature transfer. By using these matrices, the system can perform effective knowledge transfer between the self-correction channels of the source and target domains, further improving the accuracy of error correction.

[0057] Furthermore, transfer learning is performed according to the Q feature difference distribution matrices to obtain multiple self-correcting channels. Step P24 in this embodiment of the application further includes:

[0058] P24-1: Calculate the Q feature difference indices of the Q feature difference distribution matrices; P24-2: Based on the Q feature difference indices, obtain q1 feature transformation matrices less than or equal to a preset difference index and q2 feature transformation matrices greater than the preset difference index, where Q = q1 + q2, and Q, q1, and q2 are all positive integers greater than or equal to 0; P24-3: Perform direct transfer learning on the initial correction channel based on the q1 feature transformation matrices to obtain q1 self-correcting channels; P24-4: Perform indirect transfer learning on the initial correction channel based on the q2 feature transformation matrices to obtain q2 self-correcting channels.

[0059] It should be understood that the transfer learning process based on the feature difference distribution matrix can be further optimized. Through effective transfer learning strategies, different learning methods, such as direct transfer learning and indirect transfer learning, can be employed based on the differences between error features to optimize the performance of the self-correcting channel.

[0060] First, calculate Q feature difference indices for the Q feature difference distribution matrices. These indices quantify the degree of feature difference between different error types. Specifically, each feature difference distribution matrix can be used to calculate a feature difference index by calculating its internal statistics, such as mean, variance, or maximum value. Each feature difference index represents the degree of similarity or difference between different error feature groups. The smaller the feature difference index, the higher the similarity between error features; the larger the feature difference index, the more significant the difference between error features.

[0061] Next, based on the Q feature difference indices, the feature difference distribution matrix is ​​divided into two groups: one group consists of q1 feature transformation matrices whose difference indices are less than or equal to a preset difference index, and the other group consists of q2 feature transformation matrices whose difference indices are greater than the preset difference index. Specifically, the Q feature difference matrices are divided into q1 and q2 matrices according to their degree of difference, where q1 represents feature transformation matrices with smaller error feature differences, and q2 represents feature transformation matrices with larger error feature differences. Here, q1 and q2 must satisfy the relationship Q = q1 + q2, and both q1 and q2 are positive integers greater than or equal to 0. The division is based on the similarity between error features: when the difference index is less than the preset value, the differences between features are small, which is suitable for direct transfer learning; while when the difference index is greater than the preset value, the differences between features are large, which is suitable for indirect transfer learning.

[0062] Then, direct transfer learning is performed on the initialized correction channels based on q1 feature transformation matrices to obtain q1 self-correcting channels. Direct transfer learning is suitable for scenarios with small feature differences, where the feature distributions between the source and target domains are similar, allowing for transfer through a lightweight and fast parameter reuse strategy. This method primarily focuses on efficiency and stability, ensuring rapid adaptation to the error characteristics of the source domain during the transfer process while maintaining high computational performance. In practice, a pre-trained correction model from the source domain can be quickly transferred to the target domain for fine-tuning, resulting in q1 self-correcting channels. The advantages of direct transfer learning lie in its efficiency and stability, making it suitable for error types with small feature differences.

[0063] Finally, indirect transfer learning is performed on the initialized correction channels based on q² feature transformation matrices to obtain q² self-correcting channels. Indirect transfer learning is suitable for scenarios with large feature differences because the feature distributions of these error types differ significantly. Direct transfer learning may not be effective in handling large distribution differences, thus requiring more complex and deeper feature alignment and adaptation strategies. Indirect transfer learning methods ensure effective knowledge transfer under large differences by aligning the feature spaces of the source and target domains more deeply, thereby improving correction performance and generalization ability. Strategies for indirect transfer learning include aligning features of the source and target domains using deep learning models or transforming the feature space using complex adaptation networks, thereby ensuring that the transferred self-correcting channels achieve good performance in the target domain.

[0064] Through the above steps, the system can flexibly select either direct or indirect transfer learning strategies based on the degree of difference between error features. This flexible transfer learning method not only improves the efficiency of transfer learning but also optimizes the error correction process, enabling the system to accurately correct different types of errors based on their feature differences.

[0065] Furthermore, based on the q1 feature transformation matrices, direct transfer learning is performed on the initialization correction channel. Step P24-3 in this embodiment further includes:

[0066] P24-31: Obtain the source domain channel parameters of the first self-calibrating channel; P24-32: Load the source domain channel parameters of the first self-calibrating channel into q1 target domain channel parameters; P24-33: Extract q1 sets of error correction features corresponding to the q1 feature transformation matrices as training input samples to perform feedback optimization on the q1 target domain channel parameters until q1 self-calibrating channels are obtained.

[0067] Optionally, the process of direct transfer learning has been further refined.

[0068] First, the source domain channel parameters of the first self-calibration channel are obtained. These source domain channel parameters are based on source domain data, i.e., obtained from the error correction model already trained on the source domain, and represent the internal state of the model in the source domain calibration task, including network weights, bias terms, and other information. The source domain channel parameters reflect the adjustment strategy of the source domain data features and contain the correction rules learned for specific error types, such as metering errors or clock synchronization errors. By obtaining these channel parameters, the system can reuse the knowledge of the source domain during transfer learning, improving the initial accuracy of the target domain channels.

[0069] Next, the source domain channel parameters of the first self-calibrating channel are loaded into the q1 target domain channel parameters. Since the error characteristics between the source and target domains are relatively small in direct transfer learning, the calibration parameters from the source domain can be directly transferred to the target domain, quickly initializing the channel parameters for the target domain. This process essentially applies the calibration strategy learned in the source domain to the self-calibrating channel of the target domain, enabling the target domain to possess the calibration capabilities of the source domain from the initialization stage. The loading process can be achieved by copying the source domain parameters to the corresponding positions in the target domain model, or by using parameter sharing, ensuring that the self-calibrating channel of the target domain can adapt quickly.

[0070] Finally, q1 sets of error correction features corresponding to the q1 feature transformation matrices are extracted as training input samples. These samples are then used to perform feedback optimization on the parameters of the q1 target domain channels. Specifically, each feature transformation matrix corresponds to a set of error correction features, which serve as input samples for adjusting and optimizing the target domain channel parameters. Specifically, for each target domain channel, its corresponding error correction features are extracted, including correction control rate, duration, gradient, and stability, and these features are used as training input samples. Through feedback optimization algorithms, such as gradient descent, the parameters of the target domain channels are adjusted based on the training input samples to adapt to the correction requirements of the corresponding error type.

[0071] During training, feedback optimization continuously updates the weights of the target domain channels, enabling the self-calibrating channels in the target domain to gradually adapt to the specific error characteristics of the target domain. Training continues until the correction effect reaches a predetermined standard and the model converges. Through this process, the q1 self-calibrating channels will be able to perform accurate error correction tasks in the target domain, utilizing knowledge from the source domain and data from the target domain for efficient error correction. This process, based on the transfer of knowledge from the source domain and the combination of data from the target domain, ensures the efficiency and accuracy of transfer learning.

[0072] Furthermore, based on the q2 feature transformation matrices, indirect transfer learning is performed on the initialization correction channel. Step P24-4 of this embodiment further includes:

[0073] P24-41: Obtain the source domain channel parameters of the first self-calibrating channel; P24-42: Decouple the source domain channel parameters of the first self-calibrating channel to obtain fixed channel parameters and dynamic channel parameters; P24-43: Re-learn the dynamic channel parameters based on the q2 feature transformation matrices to obtain differentiated channel parameters; P24-44: Load the fixed channel parameters and the differentiated channel parameters into the q2 target domain channel parameters; P24-45: Extract the q2 sets of error correction features corresponding to the q2 feature transformation matrices as training input samples to perform feedback optimization on the q2 target domain channel parameters until q2 self-calibrating channels are obtained.

[0074] Specifically, the process of indirect transfer learning can be further refined. Since q2 represents the error type with large feature differences, indirect transfer learning employs complex feature alignment and adaptation strategies to ensure that features between the source and target domains can be effectively matched, and further optimizes the correction capability of the target domain channels through differential learning.

[0075] First, the source domain channel parameters of the first self-calibrating channel are obtained. This first self-calibrating channel is obtained through initialization learning and is a calibration channel for the first error type, i.e., the error type that occurs most frequently. The domain channel parameters of this channel include the network model's weights, biases, and learning rate. These parameters are trained based on the first set of error correction features and can effectively correct the first error type. These parameters will serve as the starting point for transfer learning and will be used to subsequently generate self-calibrating channels for other error types.

[0076] Next, the source domain channel parameters of the first self-calibrating channel are decoupled, dividing the network model parameters into two parts: fixed channel parameters and dynamic channel parameters. The fixed channel parameters remain unchanged during transfer learning and typically include the network's basic structural parameters and some global parameters, which are universal across different error types. The dynamic channel parameters, on the other hand, are highly correlated with specific error types and are adjusted according to different error types during transfer learning. These parameters require differentiated learning based on the specific error type.

[0077] Then, the dynamic channel parameters are re-learned differentially based on the q² feature transformation matrices to obtain differentially learned channel parameters. The purpose of differential learning is to adjust the dynamic channel parameters using deep learning algorithms, such as deep neural networks, to adapt to the feature distribution of different error types, targeting error types with significant feature differences. Specifically, for each feature transformation matrix, the corresponding error correction features are used for training, and the dynamic channel parameters are adjusted to achieve feature alignment and adaptation. Through differential learning, the correction capability of the dynamic channels is enhanced, enabling better handling of error types in the target domain.

[0078] Next, the fixed channel parameters and the differentiated channel parameters are loaded into the q² target domain channel parameters. In this stage, the fixed channel parameters remain unchanged, while the differentiated channel parameters are adjusted and loaded into the target domain calibration model. By combining these two sets of parameters, the target domain calibration model retains both the general calibration capability of the source domain and the dynamic calibration capability specifically adjusted according to the characteristics of the target domain, enabling the model to better handle error types in the target domain.

[0079] Finally, q2 sets of error correction features corresponding to q2 feature transformation matrices are extracted as training input samples to perform feedback optimization on the parameters of q2 target domain channels until q2 self-calibrating channels are obtained. Specifically, for each target domain channel, its corresponding error correction features are extracted, including correction control rate, duration, gradient, and stability, and these features are used as training input samples. Using a feedback optimization algorithm, such as gradient descent, the parameters of the target domain channels are adjusted based on the training input samples to adapt to the correction requirements of the corresponding error type. The optimization process continues until the correction performance of the target domain channels reaches a preset threshold or convergence condition, thus obtaining q2 self-calibrating channels.

[0080] The key to this process is optimizing dynamic channel parameters through differential learning, enabling the target domain to better adapt to error characteristics that differ significantly from the source domain. By combining fixed and differential channel parameters, the system can effectively transfer correction capabilities even when there are significant feature differences between the source and target domains, thereby improving the performance and generalization ability of the self-calibrating channel.

[0081] P30: Read the collaborative metering dataset of the electricity meter cluster, perform real-time error identification based on the collaborative metering dataset, and cluster the real-time error types according to the multiple error types to obtain multiple clustering results.

[0082] Optionally, after completing the collection of historical error events from the electricity meter cluster, error type identification, and generation of self-calibration channels, the real-time error identification and cluster analysis phase begins. First, the collaborative metering dataset of the electricity meter cluster needs to be read. This dataset contains metering data generated by the electricity meters during real-time operation, such as key parameters like energy consumption, current, and voltage, as well as related operating status information, such as the meter's internal temperature, operating time, and hardware status. This data is acquired in real-time through a pre-configured data acquisition system and stored in a local database to ensure data integrity and real-time performance.

[0083] Based on the aforementioned collaborative metering dataset, real-time error identification is performed. The purpose of real-time error identification is to detect potential errors in the current metering data and preliminarily determine their type. These errors may originate from multiple factors, such as hardware failures of the electricity meter, environmental factors, load changes, and communication problems. The system uses error identification algorithms, such as threshold-based rules and machine learning models, to determine whether errors exist in the current data and their specific types, based on the error types and characteristics in historical data and combined with real-time data. Error identification is a classification problem; the model analyzes the real-time data and outputs a label for the error type. For example, if the electricity consumption in the real-time data deviates significantly from the expected value, and this deviation matches the characteristics of a known error type, the model can identify that error type.

[0084] After identifying real-time errors, they are clustered according to multiple error types. The purpose of clustering is to group real-time errors with similar characteristics into one category, so that targeted correction measures can be taken for different error types. Clustering algorithms can include K-means, hierarchical clustering, or DBSCAN. These algorithms can classify errors into multiple categories based on their characteristics, such as error magnitude, frequency of occurrence, and changes in related parameters. For example, if multiple real-time errors exhibit characteristics of measurement proportionality errors, such as a fixed proportional deviation between measured and actual values, these errors will be grouped into one category; if other real-time errors exhibit characteristics of clock synchronization errors, such as a deviation between timestamps and standard time, these errors will be grouped into another category. In this way, multiple clustering results can be obtained, each representing a category of real-time errors with similar characteristics.

[0085] In this way, the system can obtain multiple clustering results, each representing a specific type of error. The clustering results can include multiple error categories, with errors within each category exhibiting similar characteristics and behavioral patterns. Through cluster analysis, not only can real-time errors be identified, but appropriate correction channels can also be assigned to each error category for adaptive correction in subsequent steps.

[0086] P40: Establish connections between the multiple clustering results and the multiple self-correction channels, and perform error self-correction of the multiple clustering results according to the multiple self-correction channels.

[0087] Furthermore, step P40 in this embodiment of the application also includes:

[0088] P41: Obtain the power grid topology of the electricity meter cluster; P42: Identify multiple clustering nodes based on the power grid topology according to the multiple clustering results, establish transient correction connections with the multiple self-correction channels according to the multiple clustering nodes, and reset the connection relationships of the multiple clustering nodes after the correction is completed.

[0089] It should be understood that after completing the reading of the collaborative metering dataset of the electricity meter cluster, real-time error identification, and cluster analysis, the error self-correction stage begins. First, connections need to be established between multiple clustering results and multiple self-correction channels. Each clustering result represents a class of real-time errors with similar characteristics, while each self-correction channel is pre-generated for a specific error type. By connecting the clustering results with the corresponding self-correction channels, targeted correction for different error types can be achieved. After the connection is complete, the self-correction channels will perform error correction tasks based on the connected clustering results, continuously adjusting the measured values ​​of the electricity meters until the error is corrected.

[0090] To further optimize the calibration process, the grid topology of the electricity meter cluster is then acquired. Grid topology refers to the connection and distribution of the electricity meter cluster within the power grid, including the connections between meters, branch structures, and connections to other grid equipment. The grid topology can be obtained by reading geographic information system (GIS) data of the power grid, the grid network topology database, or by real-time monitoring of the communication and data transmission relationships between meters. This information helps the system understand how the individual meters influence each other and clarifies how errors propagate and correlate in different areas of the power grid.

[0091] Next, multiple clustering nodes based on the power grid topology are identified using various clustering results. A clustering node refers to a set of electricity meters with similar error characteristics within the power grid topology. By analyzing the clustering results and the power grid topology, it can be determined which electricity meters belong to the same clustering node. For example, if a group of electricity meters are interconnected in the power grid topology and their real-time errors are clustered into the same type, these meters can be considered a single clustering node. Each clustering node may contain error types from one or more electricity meters, and the system needs to perform appropriate correction connections for these nodes.

[0092] Next, transient correction connections are established between multiple cluster nodes and multiple self-calibration channels. Transient correction connections are temporary relationships established during error correction to connect the electricity meters in the cluster nodes to their corresponding self-calibration channels. These connections are dynamic and only valid during error correction. Through these transient connections, the system can flexibly adjust and apply correction strategies based on the error characteristics of different electricity meters in the cluster nodes. The connection relationships of each cluster node are reset after error correction is completed, ensuring that the power grid topology is not unnecessarily altered and that each electricity meter can independently perform error correction tasks. The reset process includes disconnecting transient correction connections and restoring the electricity meters' default communication and metering parameters.

[0093] By analyzing the power grid topology and clustering results, the system can effectively distribute the correction task to different electricity meters and regions. This process not only improves the accuracy and efficiency of the electricity meter cluster error self-correction, but also enhances the system's adaptability and robustness.

[0094] In summary, the embodiments of this application have at least the following technical effects:

[0095] This application achieves accurate identification and correction of various error types by collecting and analyzing historical error data, combined with real-time error identification and clustering methods, significantly improving the metering accuracy of electricity meter clusters. Utilizing transfer learning and adaptive correction channels, the error correction process is automated, reducing manual intervention and improving correction efficiency. By analyzing the collaborative metering data of the electricity meter cluster in real time, flexible corrections are performed for different error types, enabling the system to cope with dynamic errors caused by changes in the power grid environment and load. Furthermore, based on the power grid topology, the error correction strategy is precisely adjusted to ensure that the correction process conforms to the actual situation of the power grid layout and electricity meter distribution, improving the system's intelligence and adaptability, and enhancing its ability to cope with complex power grid environments.

[0096] This technology achieves the technical effect of improving the accuracy of electricity metering by realizing dynamic error self-correction of the electricity meter cluster through real-time error identification and self-correction channels.

[0097] Example 2, based on the same inventive concept as the error self-correction method for collaborative metering of electricity meter clusters in the foregoing examples, such as... Figure 2 As shown, this application provides an error self-correction system for collaborative metering of electricity meters. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0098] Error type identification module 11 is used to collect a set of historical error events of the electricity meter cluster and identify multiple error types of the set of historical error events.

[0099] The transfer learning module 12 is used to extract multiple sets of error correction features for the multiple error types, and to perform transfer learning on the initial correction channel based on the multiple sets of error correction features to obtain multiple self-correction channels. Each set of error correction features includes correction control rate, correction control duration, correction control gradient, and correction control stability.

[0100] The real-time error clustering module 13 is used to read the collaborative metering dataset of the electricity meter cluster, perform real-time error identification based on the collaborative metering dataset, and cluster the real-time error types according to the multiple error types to obtain multiple clustering results.

[0101] Error self-correction module 14 is used to establish the connection between the multiple clustering results and the multiple self-correction channels, and to perform error self-correction of the multiple clustering results according to the multiple self-correction channels.

[0102] Furthermore, in the error type identification module 11:

[0103] The various error types include metering ratio error, clock synchronization error, temperature drift system error, dynamic error caused by load changes, and offset error due to communication link anomalies.

[0104] Furthermore, the transfer learning module 12 is also used to perform the following steps:

[0105] Based on the frequency of occurrence of the multiple error types corresponding to the multiple sets of error correction features, a first set of error correction features corresponding to the first error type is determined; initial learning is performed according to the first set of error correction features to obtain a first self-correcting channel corresponding to the first error type, and the first self-correcting channel is marked as an initial correction channel; Q feature difference distribution matrices are calculated between the first set of error correction features after standardization and the remaining sets of error correction features; using the initial correction channel as the source domain and the remaining self-correcting channels as the target domain, transfer learning is performed according to the Q feature difference distribution matrices to obtain multiple self-correcting channels.

[0106] Furthermore, the transfer learning module 12 is also used to perform the following steps:

[0107] Calculate the Q feature difference indices of the Q feature difference distribution matrices; based on the Q feature difference indices, obtain q1 feature transformation matrices less than or equal to a preset difference index and q2 feature transformation matrices greater than the preset difference index, where Q = q1 + q2, and Q, q1, and q2 are all positive integers greater than or equal to 0; perform direct transfer learning on the initial correction channel based on the q1 feature transformation matrices to obtain q1 self-correcting channels; perform indirect transfer learning on the initial correction channel based on the q2 feature transformation matrices to obtain q2 self-correcting channels.

[0108] Furthermore, the transfer learning module 12 is also used to perform the following steps:

[0109] Obtain the source domain channel parameters of the first self-calibrating channel; load the source domain channel parameters of the first self-calibrating channel into q1 target domain channel parameters; extract q1 sets of error correction features corresponding to the q1 feature transformation matrices as training input samples to perform feedback optimization on the q1 target domain channel parameters until q1 self-calibrating channels are obtained.

[0110] Furthermore, the transfer learning module 12 is also used to perform the following steps:

[0111] Obtain the source domain channel parameters of the first self-calibrating channel; decouple the source domain channel parameters of the first self-calibrating channel to obtain fixed channel parameters and dynamic channel parameters; re-perform differential learning on the dynamic channel parameters based on the q2 feature transformation matrices to obtain differential channel parameters; load the fixed channel parameters and the differential channel parameters into the q2 target domain channel parameters; extract q2 sets of error correction features corresponding to the q2 feature transformation matrices as training input samples to perform feedback optimization on the q2 target domain channel parameters until q2 self-calibrating channels are obtained.

[0112] Furthermore, the transfer learning module 12 is also used to perform the following steps:

[0113] A high-dimensional feature space is introduced, and the first group of error correction features and the remaining group of error correction features after standardization are nonlinearly mapped by radial basis functions according to the high-dimensional feature space to obtain the first group of high-dimensional error correction features and the remaining group of high-dimensional error correction features; Q feature difference distribution matrices are recalculated based on the first group of high-dimensional error correction features and the remaining group of high-dimensional error correction features.

[0114] Furthermore, the transfer learning module 12 is also used to perform the following steps:

[0115] The network model architecture is initialized, and the network model architecture is introduced with a correction feature loss constraint function to initialize and learn the first set of error correction features, thereby obtaining a first self-correcting channel that has been learned to convergence. The correction feature loss constraint function includes a rate loss constraint function, a duration loss constraint function, a step size loss constraint function, and a stability loss constraint function. The rate loss constraint function is used to adjust the depth of the network model architecture, the duration loss constraint function is used to adjust the temporal attention interval of the network model architecture, the step size loss constraint function is used to adjust the gradient descent magnitude of the network model architecture, and the stability loss constraint function is used to adjust the regularization parameter of the network model architecture.

[0116] Furthermore, the error self-correction module 14 is also used to perform the following steps:

[0117] Obtain the power grid topology of the electricity meter cluster; identify multiple clustering nodes based on the power grid topology according to the multiple clustering results; establish transient correction connections with the multiple self-correction channels according to the multiple clustering nodes; and reset the connection relationships of the multiple clustering nodes after the correction is completed.

[0118] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0119] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0120] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. An error self-correction method for collaborative metering of electricity meters in a cluster, characterized in that, The method includes: Collect a set of historical error events from a cluster of electricity meters and identify multiple error types in the set of historical error events; Multiple sets of error correction features for the multiple error types are extracted. Based on the multiple sets of error correction features, the initial correction channel is transferred to obtain multiple self-correction channels. Each set of error correction features includes correction control rate, correction control duration, correction control gradient, and correction control stability. Based on the aforementioned multiple sets of error correction features, transfer learning is performed on the initial correction channel to obtain multiple self-correction channels. The method includes: Based on the frequency of occurrence of the multiple error types corresponding to the multiple sets of error correction features, the first set of error correction features corresponding to the first error type is determined; Initialization learning is performed according to the first set of error correction features to obtain the first self-correction channel corresponding to the first error type, and the first self-correction channel is marked as the initialization correction channel. Calculate the Q feature difference distribution matrix between the first group of error correction features and the remaining groups of error correction features after standardization; Using the initial calibration channel as the source domain and the remaining self-calibration channels as the target domain, transfer learning is performed according to the Q feature difference distribution matrices to obtain multiple self-calibration channels; Read the collaborative metering dataset of the electricity meter cluster, perform real-time error identification based on the collaborative metering dataset, and cluster the real-time error types according to the multiple error types to obtain multiple clustering results; Establish connections between the multiple clustering results and the multiple self-correction channels, and perform error self-correction of the multiple clustering results based on the multiple self-correction channels; The method for establishing connections between the multiple clustering results and the multiple self-calibrating channels includes: Obtain the power grid topology of the energy meter cluster; Based on the multiple clustering results, identify multiple groups of clustering nodes based on the power grid topology, establish transient correction connections with the multiple self-correction channels based on the multiple groups of clustering nodes, and reset the connection relationships of the multiple groups of clustering nodes after the correction is completed.

2. The method as described in claim 1, characterized in that, Identify multiple error types in the set of historical error events, including metering ratio error, clock synchronization error, temperature drift system error, dynamic error caused by load changes, and offset error due to communication link anomalies.

3. The method as described in claim 1, characterized in that, Transfer learning is performed on the Q feature difference distribution matrices respectively to obtain multiple self-correcting channels. The method includes: Calculate the Q feature difference indices of the Q feature difference distribution matrices; Based on the Q feature difference indices, obtain q1 feature transformation matrices that are less than or equal to the preset difference indices and q2 feature transformation matrices that are greater than the preset difference indices, where Q = q1 + q2, and Q, q1, and q2 are all positive integers greater than or equal to 0; Based on the q1 feature transformation matrices, the initialization correction channel is directly transferred to obtain q1 self-correction channels; Indirect transfer learning is performed on the initial correction channel based on the q2 feature transformation matrices to obtain q2 self-correction channels.

4. The method as described in claim 3, characterized in that, The method involves performing direct transfer learning on the initialization correction channel based on the q1 feature transformation matrices, including: Obtain the source domain channel parameters of the first self-calibration channel; Load the source domain channel parameters of the first self-calibration channel into q1 target domain channel parameters; The q1 error correction features corresponding to the q1 feature transformation matrices are extracted as training input samples to perform feedback optimization on the q1 target domain channel parameters until q1 self-correcting channels are obtained.

5. The method as described in claim 3, characterized in that, Indirect transfer learning is performed on the initial correction channels using the q2 feature transformation matrices to obtain q2 self-correcting channels. The method includes: Obtain the source domain channel parameters of the first self-calibration channel; The source domain channel parameters of the first self-calibration channel are decoupled to obtain fixed channel parameters and dynamic channel parameters; Based on the q2 feature transformation matrices, the dynamic channel parameters are relearned differentially to obtain the differential channel parameters; The fixed channel parameters and the differentiated channel parameters are loaded into q2 target domain channel parameters; The q2 sets of error correction features corresponding to the q2 feature transformation matrices are extracted as training input samples to perform feedback optimization on the q2 target domain channel parameters until q2 self-correcting channels are obtained.

6. The method as described in claim 1, characterized in that, The method for calculating the Q feature difference distribution matrix between the first group of error correction features and the remaining groups of error correction features after standardization also includes: A high-dimensional feature space is introduced, and the first set of error correction features and the remaining set of error correction features after standardization are nonlinearly mapped by radial basis functions according to the high-dimensional feature space to obtain the first set of high-dimensional error correction features and the remaining set of high-dimensional error correction features. Based on the first set of high-dimensional error correction features and the remaining sets of high-dimensional error correction features, recalculate the Q feature difference distribution matrices.

7. The method as described in claim 1, characterized in that, The method includes initial learning based on the first set of error correction features to obtain the first self-correction channel corresponding to the first error type, and the method includes: Initialize the network model architecture, wherein the network model architecture introduces a correction feature loss constraint function to initialize and learn the first set of error correction features, thereby obtaining the first self-correcting channel that has been learned to converge. The correction feature loss constraint function includes a rate loss constraint function, a duration loss constraint function, a step size loss constraint function, and a stability loss constraint function. The rate loss constraint function is used to adjust the depth of the network model architecture, the duration loss constraint function is used to adjust the temporal attention interval of the network model architecture, the step size loss constraint function is used to adjust the gradient descent magnitude of the network model architecture, and the stability loss constraint function is used to adjust the regularization parameter of the network model architecture.

8. An error self-correction system for collaborative metering of electricity meters in a cluster, characterized in that, The system is used to implement the error self-correction method for clustered collaborative metering of electricity meters according to any one of claims 1-7, the system comprising: An error type identification module is used to collect a set of historical error events from a cluster of electricity meters and identify multiple error types in the set of historical error events. The transfer learning module is used to extract multiple sets of error correction features for the multiple error types, and to perform transfer learning on the initial correction channel based on the multiple sets of error correction features to obtain multiple self-correction channels. Each set of error correction features includes correction control rate, correction control duration, correction control gradient, and correction control stability. The real-time error clustering module is used to read the collaborative metering dataset of the energy meter cluster, identify real-time errors based on the collaborative metering dataset, and cluster the real-time error types according to the multiple error types to obtain multiple clustering results. An error self-correction module is used to establish connections between the multiple clustering results and the multiple self-correction channels, and to perform error self-correction on the multiple clustering results according to the multiple self-correction channels.

Images