Electric quantity error online retreating and supplementing method and system for electric energy metering

Abstract

The invention discloses an electric quantity error online returning and supplementing method and system for electric energy metering, and relates to the technical field of electric power metering, and the method comprises the steps: obtaining a multi-modal data set in an electric energy metering process, carrying out the cross-scale long time sequence feature fusion, and constructing a cross-scale fusion feature time chain; performing electric energy metering abnormal evolution analysis to obtain a plurality of abnormal evolution event streams; performing anomaly evaluation to obtain a plurality of anomaly evolution reliability coefficients; performing mapping extraction on the plurality of abnormal evolution event flows to obtain a plurality of real abnormal evolution event flows; and performing electric quantity error return and supplement decision analysis to obtain a target electric quantity error online return and supplement parameter. The technical problems that in the prior art, the electric energy metering abnormity recognition accuracy is poor, and the efficiency and accuracy are low due to the fact that electric quantity error return and supplement mostly depend on manual operation are solved, and the technical effects that accurate recognition of electric energy metering abnormity and intelligent return and supplement of electric quantity errors are achieved, and the metering return and supplement efficiency and accuracy are improved are achieved.

Description

Technical Field

[0001] This invention relates to the field of power metering technology, and more specifically to an online error correction and compensation method and system for power metering. Background Technology

[0002] In the field of electricity metering, with the widespread adoption of smart meters and the increasing complexity of electricity usage scenarios, metering data is easily affected by factors such as equipment aging, environmental interference, and fluctuations in electricity load, leading to metering deviations. Existing technologies mostly rely on single-dimensional data for anomaly detection, making it difficult to capture cross-timescale correlation features, resulting in low accuracy in anomaly identification and a tendency to misjudge or miss real metering problems. Furthermore, the error correction and refund process often requires manual retrieval of historical data, analysis of anomaly causes, and calculation of refund amounts. This is not only time-consuming and labor-intensive, but also prone to inaccuracies due to differences in human experience, failing to meet the power system's demand for intelligent and efficient metering management.

[0003] Existing technologies suffer from a lack of accuracy in identifying abnormalities in electricity metering, and the reliance on manual operation for correcting electricity errors leads to low efficiency and accuracy. Summary of the Invention

[0004] This application provides an online method and system for correcting and compensating for electricity metering errors, which addresses the technical problems of insufficient accuracy in identifying abnormalities in electricity metering and the low efficiency and accuracy caused by the reliance on manual operation for correcting and compensating for electricity errors in the prior art.

[0005] In view of the above problems, this application provides an online error correction and compensation method and system for electricity metering.

[0006] The first aspect of this application provides an online error correction method for electricity metering, the method comprising: A multimodal data set is acquired during the electricity metering process. Cross-scale long-time-series feature fusion is performed on the multimodal data set to construct a cross-scale fused feature time chain. Based on the cross-scale fused feature time chain, electricity metering anomaly evolution analysis is performed to obtain multiple anomaly evolution event streams. Anomaly assessment is performed by traversing the multiple anomaly evolution event streams to obtain multiple anomaly evolution reliability coefficients. Based on the anomaly evolution reliability coefficients that are greater than or equal to a preset coefficient threshold, the multiple anomaly evolution event streams are mapped and extracted to obtain multiple real anomaly evolution event streams. Based on the multiple real anomaly evolution event streams, electricity error correction decision analysis is performed to obtain target electricity error online correction parameters.

[0007] A second aspect of this application provides an online error correction system for electricity metering, the system comprising: The system comprises the following modules: a time chain construction module, used to acquire a multimodal data set during the electricity metering process, perform cross-scale long-time-series feature fusion on the multimodal data set, and construct a cross-scale fused feature time chain; an anomaly evolution analysis module, used to perform electricity metering anomaly evolution analysis based on the cross-scale fused feature time chain, and obtain multiple anomaly evolution event streams; an anomaly evaluation module, used to traverse the multiple anomaly evolution event streams to perform anomaly evaluation, and obtain multiple anomaly evolution reliability coefficients; a mapping extraction module, used to map and extract the multiple anomaly evolution event streams based on anomaly evolution reliability coefficients that are greater than or equal to a preset coefficient threshold, and obtain multiple real anomaly evolution event streams; and a decision analysis module, used to perform electricity error refund and compensation decision analysis based on the multiple real anomaly evolution event streams, and obtain target electricity error online refund and compensation parameters.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This method acquires multimodal data sets during the electricity metering process, performs cross-scale long-time-series feature fusion, and constructs a cross-scale fused feature time chain. It then performs electricity metering anomaly evolution analysis to obtain multiple anomaly evolution event streams. These event streams are traversed for anomaly assessment to obtain multiple anomaly evolution reliability coefficients. The event streams are then mapped and extracted to obtain multiple real anomaly evolution event streams. Finally, based on these real anomaly evolution event streams, an electricity error refund / compensation decision analysis is performed to obtain online refund / compensation parameters for the target electricity error. This method achieves accurate identification of electricity metering anomalies and intelligent refund / compensation for electricity errors, improving the efficiency and accuracy of metering refund / compensation. Attached Figure Description

[0009] 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.

[0010] Figure 1 This is a schematic diagram of the online error correction and compensation method for electricity metering provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an online error correction and compensation system for electricity metering provided in an embodiment of this application.

[0011] Figure labeling: Time chain construction module 10, anomaly evolution analysis module 20, anomaly assessment module 30, mapping extraction module 40, decision analysis module 50. Detailed Implementation

[0012] This application provides an online method and system for correcting and compensating for electricity metering errors, which addresses the technical problems of insufficient accuracy in identifying abnormalities in electricity metering and the low efficiency and accuracy caused by the reliance on manual operation for correcting and compensating for electricity errors in the existing technology.

[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] Example 1, as Figure 1 As shown, this application provides an online error correction method for electricity metering, the method comprising: Step S100: Obtain a multimodal data set during the electricity metering process, perform cross-scale long-time-series feature fusion on the multimodal data set, and construct a cross-scale fused feature time chain.

[0015] Specifically, a multimodal data set is obtained from the electricity metering process by classifying data sources from multiple dimensions: metering measurement data such as voltage, current, electricity consumption, and power factor are collected from the metering side; load curves are collected from the user side, including load-related data such as peak load characteristics, off-peak load characteristics, and the start-stop status of electrical equipment, while user behavior data is extracted according to daily electricity consumption patterns and holiday electricity consumption patterns, covering the characteristics of periodic and holiday patterns; equipment health, aging degree, and error codes are extracted from the metering equipment based on daily operation logs; and environmental disturbance data such as temperature, humidity, and weather disturbance factors are collected from the environmental monitoring end through multi-source sensors. The above-mentioned metering measurement, load curve, electricity consumption behavior, equipment condition, environmental disturbance, periodic pattern, and holiday pattern related data are summarized and integrated to form a complete multimodal data set. Subsequently, cross-scale long-time-series feature fusion was performed on the multimodal dataset to construct a cross-scale fused feature time chain: First, the multimodal dataset was aligned with time windows to ensure consistency in the time dimension of data from different sources, resulting in an aligned multimodal dataset; then, the aligned multimodal dataset was segmented into windows according to micro-windows (e.g., seconds, minutes), meso-windows (e.g., hours), and macro-windows (e.g., days, weeks), and features within each window were extracted to generate micro-window feature time chains, meso-window feature time chains, and macro-window feature time chains; finally, through a hierarchical integration method of micro-mapping-meso-assignment-macro-embedding, the micro-window feature time chains were mapped to the full time axis at second / minute resolution, and the meso-window feature time chains were... The feature time chains are assigned values ​​on the full time axis according to the corresponding time intervals. The macro window feature time chains are embedded into the full time axis that has completed micro and meso processing to form an integrated full time axis. Based on the integrated full time axis, hierarchical fusion is carried out. First, cross-scale fusion is performed with macro window features corresponding to meso window feature sets. Feature similarity is calculated, normalization is performed to obtain the fusion matrix, and the meso features are mapped to enhance the fusion matrix. The integrated full time axis is then updated. Next, the fused meso window features are fused with micro window feature sets, and the integrated full time axis is updated again. After all time intervals are updated, the updated time chains corresponding to the micro windows are extracted. Finally, a cross-scale fused feature time chain that can comprehensively reflect the spatiotemporal correlation features of multi-dimensional data is constructed.

[0016] Step S200: Perform anomaly evolution analysis of power metering based on the cross-scale fusion feature time chain to obtain multiple anomaly evolution event streams.

[0017] Specifically, the core analysis uses a constructed cross-scale fused feature time chain as its core analytical framework. This time chain integrates features from micro, meso, and macro scales, including metering measurements, load curves, electricity consumption behavior, equipment operating conditions, and environmental disturbances, comprehensively reflecting the spatiotemporal correlation and dynamic changes in electricity metering data. First, a feature deviation detection algorithm extracts multiple outliers in the cross-scale fused feature time chain that deviate from the normal fluctuation range. These outliers may correspond to metering errors, equipment malfunctions, or environmental interference causing metering data anomalies. Then, for each outlier, based on its position in the cross-scale fused feature time chain, the anomalous magnitude of the outlier is extracted—that is, the degree and direction of deviation from the normal range, i.e., whether the data is too high or too low. Based on this retrieval, a correlation and retrospective retrieval of multi-scale feature data before and after the anomaly point in the cross-scale fusion feature time chain is performed: at the micro scale, the changes in measurement data at the second / minute level before and after the anomaly occurred are traced; at the meso scale, the load curve and equipment operating condition fluctuations during the anomaly duration are analyzed; at the macro scale, the origin, development process and impact range of the anomaly are determined by combining the electricity consumption behavior during the same period, such as periodic patterns, holiday patterns and environmental disturbance factors, such as sudden temperature changes and weather interference. By connecting the correlated anomalies in chronological order and causal relationships, the complete trajectory of the anomaly from its inception, development to stabilization or regression is sorted out, and finally, multiple anomaly evolution event flows that can completely characterize the development process of different anomaly events are obtained.

[0018] Step S300: Traverse the multiple abnormal evolution event streams to perform anomaly assessment and obtain multiple abnormal evolution reliability coefficients.

[0019] Specifically, a core indicator system for anomaly assessment is defined, which includes anomaly magnitude indicators, suddenness indicators, and contextual consistency indicators. The anomaly magnitude indicator is used to quantify the degree to which anomalous data deviates from the normal range in the anomaly evolution event stream, reflecting the significance of the anomaly. The suddenness indicator is used to assess the rate of data change when an anomaly occurs, distinguishing between gradual fluctuations and sudden anomalies. The contextual consistency indicator is used to verify the matching degree between the anomaly evolution event stream and concurrent multi-dimensional data, such as the periodic patterns / holiday patterns of electricity consumption, equipment operating conditions, and environmental disturbance factors, to determine whether the anomaly is caused by a real metering problem, rather than accidental data fluctuations. Subsequently, for each abnormal evolution event stream, the three indicators mentioned above are quantitatively calculated: based on the micro, meso, and macro characteristic data in the cross-scale fusion feature time chain, the specific values ​​of the abnormal amplitude are calculated, such as the standard deviation multiple of the deviation from the mean; the instantaneous change rate at the time of the abnormality is statistically analyzed to characterize the suddenness; and the differences in electricity consumption behavior, equipment status, and environmental conditions between the abnormal period and the normal period are compared to determine the contextual consistency. Then, through a weighted fusion algorithm, the weights of each indicator are assigned according to the importance of each indicator in different scenarios, and the quantitative results of the three indicators are integrated into a single evaluation value. This evaluation value is the abnormal evolution reliability coefficient of the corresponding abnormal evolution event stream. The higher the coefficient, the stronger the credibility of the event stream in reflecting the real electricity metering anomaly. Finally, all abnormal evolution event streams are traversed to complete the evaluation and obtain multiple abnormal evolution reliability coefficients corresponding to each event stream.

[0020] Step S400: Based on the abnormal evolution reliability coefficients among the multiple abnormal evolution reliability coefficients that are greater than or equal to a preset coefficient threshold, the multiple abnormal evolution event streams are mapped and extracted to obtain multiple real abnormal evolution event streams.

[0021] Specifically, based on the actual operation and maintenance needs of the electricity metering scenario, the statistical patterns of historical anomaly data, and the metering accuracy requirements, a reasonable threshold for the anomaly evolution reliability coefficient is preset. This threshold needs to balance avoiding missing real anomalies with reducing interference from false anomalies, and is usually determined by calibration using the reliability coefficient distribution characteristics of historical real anomaly events. Subsequently, multiple anomaly evolution reliability coefficients are compared one by one with the preset threshold, and anomaly evolution reliability coefficients with coefficients greater than or equal to the threshold are selected. The anomaly evolution event streams corresponding to these coefficients mean that they meet the criteria for judging real anomalies in the three core evaluation indicators of anomaly amplitude, suddenness, and contextual consistency, such as significant anomaly amplitude, suddenness characteristics consistent with metering fault patterns, and high contextual matching degree with electricity consumption behavior, equipment operating conditions, and environmental disturbances. Finally, through mapping and extraction operations, the selected reliability coefficients are correlated one-to-one with the original multiple abnormal evolution event streams. The abnormal evolution event streams with coefficients that meet the standards are accurately extracted, while low-confidence event streams with coefficients below the threshold are eliminated, such as false anomalies caused by accidental data fluctuations or temporary sensor interference. In the end, multiple real abnormal evolution event streams that can truly reflect the abnormal problems of electricity metering are obtained, providing accurate abnormal event basis for subsequent electricity error refund and compensation decisions.

[0022] Step S500: Perform power error correction and compensation decision analysis based on the multiple real abnormal evolution event streams to obtain the target power error online correction and compensation parameters.

[0023] Specifically, deep feature analysis is performed on multiple real-world abnormal evolution event streams. Each event stream is traversed and key information is extracted. Abnormality types are identified through semantic recognition, such as metering deviations caused by aging metering equipment or temporary data anomalies triggered by environmental disturbances. Abnormality amplitude quantification determines the specific numerical deviation of metering data from the normal range during abnormal periods. Abnormality trend analysis reveals the duration, rate of change, and impact cycle of the anomalies. Subsequently, based on these multi-dimensional anomaly identification results, combined with historical data from normal electricity consumption scenarios during the same period, such as load curves and electricity consumption behavior characteristics of the same cycle or holiday patterns, time-series prediction algorithms are employed, such as prediction models based on load patterns and normal equipment operating conditions. The system performs abnormal power consumption prediction, calculates the theoretical power consumption value if no metering anomaly occurs during the abnormal period, i.e., multiple benchmark power consumption prediction results. Finally, using the benchmark power consumption prediction results as a reference, it compares the actual metered power consumption during the abnormal period to calculate the power consumption error value caused by the anomaly. At the same time, it comprehensively considers the credibility of the anomaly, and performs refund and compensation rule matching based on factors such as the previous reliability coefficient, error duration, and user power consumption type, such as determining the refund and compensation ratio and refund and compensation cycle. Finally, it integrates the above information through a decision analysis model to obtain the target power consumption error online refund and compensation parameters, which include key contents such as the error power consumption value, refund and compensation duration, and refund and compensation execution standards, providing data support for the accurate execution of online power consumption refund and compensation.

[0024] In one possible implementation, step S100 further includes: Step S110: Collect the voltage, current, energy consumption and power factor measured by the electricity meter from the metering side to obtain metering side data.

[0025] Step S120: Collect user load curves, peak load characteristics, off-peak load characteristics, and start / stop status of electrical equipment to obtain load-side data.

[0026] Step S130: Collect temperature, humidity and weather disturbance factors through environmental multi-source sensors to obtain environmental data.

[0027] Step S140: Extract health status, aging level and error codes from the daily logs of the metering equipment to obtain equipment operating condition data.

[0028] Step S150: Extract user behavior data according to daily electricity consumption patterns and holiday electricity consumption patterns to obtain user behavior data.

[0029] Step S160: Summarize the metering data, load data, environmental data, equipment operating data, and user behavior data to obtain the multimodal data set.

[0030] Specifically, the core data acquisition on the metering side involves establishing a real-time data interaction link with the electricity meter to directly collect key metering parameters recorded during the meter's operation. These parameters include voltage and current data reflecting the state of electricity transmission, electricity data reflecting the total amount of electricity consumed, and power factor data measuring the efficiency of electricity utilization. These data form the basis for judging the accuracy of metering and together constitute the metering side data.

[0031] Load-side data collection is carried out, with user electricity terminals as the data collection object. The load monitoring device records the changes in user electricity load at different times to generate continuous user load curves. At the same time, feature information is extracted from the load curves, including peak load characteristics during peak electricity consumption periods, low-valley load characteristics during low electricity consumption periods, and start-up and shutdown status of major electrical equipment such as air conditioners and water pumps obtained through the equipment status monitoring module. The data are combined to form load-side data that can reflect user electricity demand and equipment operation patterns.

[0032] Environmental data collection is conducted by relying on multi-source sensors deployed around metering equipment and in the user's power environment, such as temperature and humidity sensors and weather monitoring sensors, to collect environmental parameters that affect the operation of metering equipment and power consumption behavior in real time. These parameters include ambient temperature, relative humidity, and weather disturbance factors such as rainstorms, strong winds, and high temperatures that may interfere with metering accuracy or change power consumption patterns. The collected data is then integrated into environmental data.

[0033] Extract equipment operating condition data by reading the daily operation logs of metering devices, such as electricity meters and data acquisition terminals, and extract key information that characterizes the operating status of the equipment: health level calculated based on equipment runtime and fault records, aging degree assessed based on component aging curves, and error codes generated when the equipment reports an error, such as communication fault codes and metering anomaly codes. This information together constitutes equipment operating condition data that reflects the health status of the metering equipment.

[0034] By acquiring user behavior data and analyzing historical electricity consumption data, the data is categorized and extracted according to electricity consumption scenarios: on the one hand, the daily electricity consumption patterns of users during regular periods such as weekdays and weekends are identified; on the other hand, special electricity consumption patterns during holidays are captured, such as increases or decreases in electricity consumption and shifts in electricity consumption periods. The electricity consumption characteristics corresponding to the two types of patterns are integrated into user behavior data to reflect the periodicity and particularity of users' electricity consumption habits.

[0035] The multimodal dataset is aggregated by unifying the format and aligning the acquired metering, load, environmental, equipment condition, and user behavior data in terms of time. This ensures consistency in time and compatibility of data formats across different sources. Ultimately, the dataset is integrated to form a complete multimodal dataset covering metering, load, environment, equipment, and user behavior, providing comprehensive data support for subsequent cross-scale long-term time-series feature fusion.

[0036] In one possible implementation, step S100 further includes: Step S170: Perform time window alignment on the multimodal data set to obtain an aligned multimodal data set.

[0037] Step S180: Perform window segmentation and feature extraction within the window on the aligned multimodal data set according to micro-window, meso-window and macro-window to obtain micro-window feature time chain, meso-window feature time chain and macro-window feature time chain.

[0038] Step S190: Perform cross-scale fusion of the microscopic window feature time chain, the mesoscopic window feature time chain, and the macroscopic window feature time chain to obtain a cross-scale fused feature time chain.

[0039] Specifically, due to the differences in the collection frequency and time recording standards of data in various dimensions of the multimodal dataset, for example, voltage and current data on the metering side are often collected at high frequencies of seconds or minutes, while temperature and humidity data on the environmental side are mostly collected at minutes or hours, user behavior data, such as daily / holiday electricity consumption patterns, are statistically analyzed at the daily or weekly level, and equipment operating condition data, such as health status and error codes, are dynamically adjusted according to changes in equipment status. If used directly, the feature association may be distorted due to time misalignment. Therefore, a unified timeline needs to be constructed based on the timestamps corresponding to the highest acquisition frequency in the multimodal data, typically the second / minute level data from the metering side. For datasets with acquisition frequencies lower than the benchmark, such as environmental and user behavior data, linear interpolation and nearest-neighbor fill methods are used to complete the data, ensuring that there is corresponding data within each benchmark time window. At the same time, the timestamp format of all data is standardized and calibrated to correct time recording errors caused by equipment clock deviations. Through the above time window alignment operations, the data from the metering side, load side, environmental side, equipment operating condition side, and user behavior side are aligned one-to-one at each time node, ultimately obtaining an aligned multimodal data set with strong time consistency that can be directly used for subsequent multi-scale feature extraction.

[0040] After obtaining a time-consistent aligned multimodal dataset, based on the temporal characteristics and analysis requirements of the electricity metering data, a three-level time window dimension is set and layered processing is carried out: Micro-windows are at the second or minute level, and the aligned multimodal dataset is continuously segmented according to this time granularity. For each micro-window, the instantaneous fluctuation amplitude of the metering-side data is extracted, such as the voltage / current peak difference; the instantaneous change rate of the load-side data, such as the load curve slope; and the real-time values ​​of the environmental-side data, such as the current temperature. These features are then concatenated in chronological order to form a micro-window feature time chain that reflects short-term dynamic changes in the data. Meso-windows are at the hour level, and after similarly segmenting the aligned data, the mean of the metering-side data within each window is extracted, such as average power and the peak-to-valley difference of the load-side data. The time-series features, such as the difference between peak and off-peak loads and the time-period average of environmental data, are integrated into a meso-level window feature time chain that reflects the medium-term change pattern of the data. The macro-level window is divided into daily or weekly units, and the total amount of metering data in each window is extracted, such as daily electricity consumption, the pattern matching degree of load data (i.e., the similarity with the load curve of the same period in history), the periodic characteristics of user behavior data (such as whether it conforms to the daily / holiday electricity consumption pattern), and the trend value of equipment operating condition data (such as the periodic characteristics of the weekly change rate of health status). These are then connected in time sequence to form a macro-level window feature time chain that reflects the long-term change trend of the data. Finally, through three-level window processing, the core features of the aligned multimodal data are mined from the short-term, medium-term, and long-term dimensions to obtain three types of feature time chains at different scales.

[0041] Based on the micro-window feature time chain, a full time axis is constructed with a time resolution of seconds or minutes. Micro-window features are directly mapped to their corresponding time nodes, ensuring that the fused data retains fine-grained information about short-term dynamic changes. Next, for the meso-window feature time chain, based on the correspondence between its hourly time intervals and the full time axis, each meso-window feature is assigned to all micro-time nodes within that interval of the full time axis, achieving coverage of the micro-time axis by meso-window features and allowing short-term data to incorporate medium-term change patterns. Finally, the macro-window feature time chain is embedded into the full time axis, which has integrated micro- and meso-window features, according to daily or weekly time intervals, allowing long-term periodic features, such as user daily / holiday usage, to be reflected in the data. The electrical patterns, equipment health trends, and short-to-medium-term characteristics are correlated. Based on this, hierarchical fusion is carried out: first, the similarity between macroscopic window features and corresponding mesoscopic window feature sets is calculated, and a fusion matrix is ​​obtained after normalization. This matrix is ​​used to map and enhance mesoscopic features, updating the mesoscopic features in the full time axis. Then, based on the enhanced mesoscopic window features, the similarity calculation and fusion enhancement operation are repeated in the corresponding microscopic window feature sets to further update the full time axis. After all time intervals have been fused and updated, the updated feature sequences corresponding to the microscopic windows in the full time axis are extracted, and finally, a cross-scale fused feature time chain with microscopic fine granularity, mesoscopic temporality, and macroscopic periodicity, and with deep correlation between features at each scale, is obtained.

[0042] In one possible implementation, step S190 further includes: Step S191: Map the micro-window feature time chain directly to the full time axis at a resolution of seconds or minutes.

[0043] Step S192: Assign values ​​to the full time axis according to the corresponding time intervals of the mesoscopic window feature time chain.

[0044] Step S193: Embed the macroscopic window feature time chain into the full time axis after mapping and amplitude of the mesoscopic and microscopic windows according to the corresponding time intervals to obtain the integrated full time axis.

[0045] Step S194: Perform cross-scale fusion based on the integrated full time axis to obtain the cross-scale fused feature time chain.

[0046] Specifically, the fusion foundation is built around the micro-window feature time chain. Since the micro-window feature time chain has a high time resolution of seconds or minutes, it can accurately reflect the short-term dynamic changes of electricity metering data. Therefore, it is directly mapped onto the pre-constructed full time axis. The time interval of the full time axis is consistent with that of the micro-window, at the second or minute level. Each time node corresponds to a fine-grained feature in the micro-window feature time chain, ensuring that the fused data retains the accuracy of short-term data.

[0047] The full time axis is assigned values ​​to the time chain of meso-level window features. The meso-level window is in hourly units, and each meso-level window corresponds to multiple consecutive micro-level time nodes on the full time axis. For example, a 1-hour meso-level window corresponds to 60 1-minute micro-level nodes. Based on the correspondence between the time intervals of the meso-level window and the full time axis, the features of each meso-level window, such as hourly average power and load peak-valley difference, are assigned to all micro-level time nodes within that interval. This allows the time-varying features of the meso-level to cover the micro-time axis, and allows short-term data to be integrated into medium-term change patterns.

[0048] The process involves embedding the macroscopic window feature time chain, with macroscopic windows measured in daily or weekly units. Each macroscopic window corresponds to a longer continuous time interval on the full time axis, such as one daily macroscopic window corresponding to a 24-hour mesoscopic window and 1440 one-minute microscopic windows. Features of each macroscopic window, such as daily electricity consumption, electricity consumption pattern matching degree, and equipment health trend, are embedded into the corresponding intervals of the full time axis. This combines the periodic features of the macroscopic scale with the full time axis that has completed microscopic mapping and mesoscopic assignment, forming an integrated full time axis that covers features of short, medium, and long time scales.

[0049] Macro-meso fusion is performed: a single macro window feature is extracted from the integrated full time axis, such as the daily electricity consumption of a certain day, the user's electricity consumption pattern on holidays, and the set of all meso window features covered by the macro window, such as the average power and load peak-valley difference of the day at the 24-hour level. The cosine similarity algorithm is used to traverse and calculate the similarity between the macro window feature and each meso window feature to generate the first macro-meso scale similarity cluster. The similarity cluster is subjected to Min-Max normalization to map the similarity value to the interval [0,1] to obtain the first macro-meso scale fusion matrix set. The meso window feature set is weighted and enhanced by weighted mapping using the fusion matrix set as the weight. That is, each meso feature is multiplied by the corresponding weight and then concatenated with the macro feature to generate the first fused meso window feature set, and the meso feature region in the integrated full time axis is updated accordingly. Next, meso-micro fusion is performed: From the updated integrated full time axis, a single first fused meso-window feature and the set of all micro-window features covered by that meso-window are extracted, such as voltage fluctuation amplitude and instantaneous current values ​​at the minute level over 60 minutes within one hour. The Pearson correlation coefficient algorithm is used to calculate the correlation between the fused meso-feature and each micro-feature, generating a meso-micro cross-scale correlation cluster. Z-score normalization is applied to the correlation cluster to obtain a meso-micro cross-scale fusion matrix set. Based on this matrix set, the micro-window feature set is weighted and fused to generate the first micro-meso-window feature set, which is then used to update the micro-feature regions in the integrated full time axis. This process is repeated for all macro- and meso-windows in the integrated full time axis. After completing the fusion update for all time intervals, the final feature sequence corresponding to the micro-windows in the full time axis is extracted, which is the cross-scale fused feature time chain.

[0050] In one possible implementation, step S194 further includes: Step S1941: Extract the first macroscopic window features and the corresponding first mesoscopic window feature set from the integrated full time axis.

[0051] Step S1942: Perform cross-scale fusion of the first meso-level window feature set based on the first macro-level window features to obtain the first fused meso-level window feature set.

[0052] Step S1943: Update the integrated full time axis based on the first fused mesoscopic window feature set.

[0053] Step S1944: Extract the first fusion mesoscopic window features from the first fusion mesoscopic window feature set, and extract the microscopic window features of the corresponding time interval in the integrated full time axis to obtain the first microscopic window feature set.

[0054] Step S1945: Perform cross-scale fusion of the first micro-window feature set based on the first fused meso-window features to obtain the first micro-meso-window feature set.

[0055] Step S1946: Update the integrated full time axis according to the first micro-meta-window feature set. After each time interval in the integrated full time axis has been updated, extract the updated time chain corresponding to the micro-window to obtain the cross-scale fusion feature time chain.

[0056] Specifically, the first macro-time interval corresponding to the first macro-time interval is selected from the integrated full time axis as the first macro-window feature, such as the total daily electricity consumption on a certain day and the user's electricity consumption mode label on that day. At the same time, all meso-time interval features covered by the macro-window are extracted to form the first meso-window feature set, such as the average power at the 24-hour level, the load peak-valley difference, and the hourly average ambient temperature, thus clarifying the correlation range between macro-features and corresponding meso-features.

[0057] Based on the first macroscopic window features, cross-scale fusion is carried out on the first mesoscopic window feature set. First, the cosine similarity algorithm is used to calculate the similarity between the first macroscopic window features and each mesoscopic feature in the first mesoscopic window feature set, generating a macro-mesoscopic cross-scale similarity cluster. The similarity cluster is then subjected to Min-Max normalization to obtain a macro-mesoscopic cross-scale fusion matrix, where the matrix elements are the association weights between each mesoscopic feature and the macroscopic feature. Then, based on the fusion matrix, each mesoscopic feature and the macroscopic feature are weighted and concatenated. The mesoscopic feature is multiplied by its corresponding weight and then merged with the vector dimension of the macroscopic feature to generate a first fused mesoscopic window feature set that combines macroscopic periodicity and mesoscopic time periodity.

[0058] The integrated full time axis is updated based on the first fused mesoscopic window feature set. The mesoscopic feature regions corresponding to the first macroscopic window in the original full time axis are replaced with the newly generated first fused mesoscopic window features, ensuring that the mesoscopic scale features have been integrated into the related information of the macroscopic features.

[0059] The first fused mesoscopic feature is selected from the updated first fused mesoscopic window feature set as the first fused mesoscopic window feature. At the same time, the microscopic time interval corresponding to this mesoscopic window in the integrated full time axis is located, and all microscopic scale features in this interval are extracted, such as the voltage fluctuation amplitude at the minute level, the instantaneous current value, and the power factor change, to form the first microscopic window feature set.

[0060] Based on the first fused mesoscopic window features, the first microscopic window feature set is fused across scales. The Pearson correlation coefficient algorithm is used to calculate the correlation between the first fused mesoscopic window features and each microscopic feature in the first microscopic window feature set, generating mesoscopic-microscopic cross-scale correlation clusters. The correlation clusters are normalized by Z-Score to obtain the mesoscopic-microscopic cross-scale fusion matrix. Using this matrix as weights, the microscopic features are enhanced and their dimensions are merged. The microscopic features are multiplied by their corresponding weights and then concatenated with the fused mesoscopic features to obtain the first microscopic-mesoscopic window feature set that combines mesoscopic time-varying characteristics with microscopic fine-grained features.

[0061] Replace the original features of the corresponding micro-interval in the full time axis with the feature set of the first micro- and meso-level window to complete the update of the micro-interval; then, follow the above process to traverse all macro- and meso-level windows in the full time axis, repeating the operations from S1941 to S1945 until each time interval in the full time axis has been fused and updated; finally, extract the feature sequences corresponding to all micro-level windows in the updated full time axis. This sequence is the cross-scale fused feature time chain that covers macro-level periodicity, meso-level temporality, and micro-level fine granularity, and has deep correlation between features at each scale.

[0062] In one possible implementation, step S1942 further includes: The feature similarity between the first macroscopic window feature set and the first mesoscopic window feature set is calculated by iterating through the data to obtain the first macroscopic-mesoscopic cross-scale similarity cluster.

[0063] The cross-scale similarity clusters in the first macro are normalized within the set to obtain the cross-scale fusion matrix set in the first macro.

[0064] The first mesoscopic window feature set is mapped, fused, and enhanced based on the first macroscopic cross-scale fusion matrix set to obtain the first fused mesoscopic window feature set.

[0065] Specifically, for the first macroscopic window features, such as the total daily electricity consumption and the feature vector corresponding to the daily electricity consumption pattern label, and the first mesoscopic window feature set, such as the feature vector set consisting of the average power, load peak-valley difference, and hourly average ambient temperature for that day, a cosine similarity algorithm is used to traverse and calculate the similarity between the two. Each mesoscopic feature vector in the first macroscopic window feature set and the first mesoscopic window feature set is input into the similarity calculation model. The similarity value between each mesoscopic feature and the macroscopic feature is obtained by dividing the vector dot product by the vector magnitude product. All similarity values ​​are arranged in the time order of the mesoscopic features to form the first macro-mesoscopic cross-scale similarity cluster. This cluster intuitively reflects the degree of correlation between the macroscopic features and each mesoscopic feature.

[0066] Next, the cross-scale similarity clusters in the first macro are normalized within the set: Considering that the numerical ranges of different similarity values ​​may differ, in order to avoid the impact of different numerical levels on the subsequent fusion weight allocation, the Min-Max normalization method is adopted to map all values ​​in the similarity cluster to the interval [0, 1]. By calculating the formula (single similarity value - minimum value of similarity cluster) / (maximum value of similarity cluster - minimum value of similarity cluster), each similarity value is standardized. The normalized similarity values ​​after transformation constitute the cross-scale fusion matrix set in the first macro. Each matrix element corresponds to a normalized correlation weight between a meso-level feature and a macro-level feature. The matrix dimension is consistent with the number of features in the first meso-level window feature set.

[0067] Finally, a neural network model is used for mapping, fusion, and enhancement: the feature vectors of the first meso-level window feature set and the weight matrix of the first macro-level cross-scale fusion matrix set are taken as input and fed into a pre-defined lightweight fully connected neural network model, such as a neural network with 2-3 hidden layers, using ReLU as the activation function; the model first performs dimensional transformation and nonlinear mapping on the meso-level feature vectors through the hidden layers, and at the same time uses the weight values ​​of the fusion matrix as attention weights to weight and highlight the transformed meso-level features, giving higher weights to meso-level features with high correlation and strengthening their feature expression; then, the weighted meso-level features are concatenated with the feature vectors of the first macro-level window through dimensional concatenation, and feature fusion and dimensional unification are completed through the output layer, finally generating a first fused meso-level window feature set that has both macro-level periodic features and meso-level time-varying features with significantly enhanced feature correlation.

[0068] In one possible implementation, step S200 further includes: Step S210: Extract outliers from the cross-scale fusion feature time chain to obtain multiple outliers.

[0069] Step S220: Based on the abnormal amplitude and abnormal direction of the multiple abnormal points, perform a back-and-forth correlation retrieval of the cross-scale fused feature time chain to obtain multiple abnormal evolution event streams.

[0070] Specifically, relying on the micro, meso, and macro multi-scale features integrated in the cross-scale fusion feature time chain, including related information such as metering measurements, load curves, equipment operating conditions, environmental disturbances, and user behavior, a dual anomaly judgment mechanism based on statistical thresholds and feature correlation is constructed. On the one hand, the 3σ statistical principle is adopted to calculate the mean and standard deviation of the fusion features of each micro window based on historical normal data in the time chain, and the mean ± 3 times the standard deviation is set as the basic normal fluctuation range. On the other hand, correlation verification rules are established in combination with meso and macro scale features. For example, if the feature value of a certain micro window does not exceed the 3σ range, but there is a significant conflict with the load trend of the meso window during the same period, such as the power characteristics that should be present during peak load periods, the user behavior pattern of the macro window, such as the difference in electricity consumption patterns between weekdays and holidays, or the equipment operating condition, such as the metering deviation trend caused by equipment aging, it is also included in the anomaly candidate range. Subsequently, each micro-window of the cross-scale fusion feature time chain is traversed. First, it is determined whether the fusion feature value exceeds the 3σ basic range. Then, the consistency with the meso- and macro-level features is verified through feature association verification rules. For micro-windows that simultaneously meet the conditions of exceeding the basic range or having conflicting associated features, the corresponding data points are marked as outliers, and the key information of the outlier is recorded, including the timestamp, the micro / meso / macro time interval to which it belongs, the feature deviation value, and the associated meso / macro-level anomaly basis. After the traversal is completed, all marked outlier data points are summarized, and finally, multiple outlier points that can accurately reflect the anomalies in the electricity metering data are obtained.

[0071] For each anomaly, its magnitude and direction are extracted. The magnitude quantifies the degree to which the fused feature value deviates from the normal range average, while the direction distinguishes between positive anomalies (feature values ​​above the normal range) and negative anomalies (feature values ​​below the normal range). The timestamp and the corresponding micro, meso, and macro time intervals are recorded for each anomaly. Then, using this anomaly as the core, a bidirectional backtracking correlation is performed on the cross-scale fused feature time chain: When tracing forward, the micro-window fused features before the anomaly are examined sequentially in reverse chronological order. Data points consistent with the current anomaly magnitude trend and matching the corresponding meso-window load characteristics and macro-window user behavior characteristics are selected to identify precursory information in the anomaly's nascent stage. When tracing backward, the micro-window fused features after the anomaly are tracked sequentially in forward chronological order to observe whether the anomaly magnitude changes, whether the anomaly persists, and to correlate with subsequent meso-window equipment operating conditions and macro-window environmental disturbance characteristics to understand the anomaly's development or regression process. By linking and integrating the relevant data points traced back to the previous and current anomalies in chronological order, a trajectory is formed that covers the entire process of anomaly initiation, occurrence, and development, with the magnitude and direction of the anomalies remaining consistent at each stage—that is, an anomaly evolution event flow. After traversing all anomalies according to the above process and removing duplicate trajectories, multiple anomaly evolution event flows corresponding to different electricity metering anomalies are finally obtained.

[0072] In one possible implementation, step S300 further includes: Step S310: Obtain anomaly assessment indicators, wherein the anomaly assessment indicators include anomaly magnitude indicators, suddenness indicators, and context consistency indicators.

[0073] Step S320: Based on the anomaly evaluation index, perform anomaly evaluation on the multiple anomaly evolution event streams to obtain multiple anomaly evolution reliability coefficients.

[0074] Specifically, the core indicator system for anomaly assessment is first defined, which includes three key indicators: anomaly magnitude indicator, used to measure the degree to which anomalous data deviates from the normal range in the anomaly evolution event stream, reflecting the significance of the anomaly; suddenness indicator, used to assess the rate of data change when an anomaly occurs, distinguishing between gradual normal fluctuations and sudden anomalies; and contextual consistency indicator, used to verify the matching degree between the anomaly evolution event stream and the multi-dimensional correlation features of the same period, and to determine whether the anomaly is caused by a real measurement problem rather than accidental interference.

[0075] For each abnormal evolution event stream, three indicators are quantified: For the abnormality magnitude indicator, the difference between the fused feature value of all abnormal points in the event stream and the mean value of the normal feature in the corresponding time interval is extracted, the absolute average of all differences is calculated, and then the average value is mapped to the [0, 1] interval through Min-Max normalization to obtain the abnormality magnitude quantification value; For the suddenness indicator, the first abnormal point in the event stream is located, the change in feature value between the abnormal point and the previous normal data point is calculated, and the instantaneous change rate is obtained by dividing by the time interval between the two points. The ratio of the change rate to the threshold of the maximum change rate in the historical normal period is used as the suddenness quantification value, and the part exceeding the threshold is counted as 1; For the context consistency indicator, a multi-dimensional associated feature library for the abnormal period is constructed, including user behavior patterns, equipment operating data, and environmental disturbance factors. The cosine similarity algorithm is used to calculate the matching degree between the abnormal features of the event stream and the associated feature library. The matching degree result is the context consistency quantification value. Next, based on the requirements of the electricity metering scenario, weights are assigned to the three indicators, such as anomaly amplitude (0.4), suddenness (0.3), and context consistency (0.3). The reliability coefficient of a single event stream is calculated using the weighted summation formula: Reliability Coefficient = Anomaly Amplitude Quantification Value × 0.4 + Suddenness Quantification Value × 0.3 + Context Consistency Quantification Value × 0.3. Following the above implementation method, all anomaly evolution event streams are traversed, completing the indicator calculation and weight integration for each event stream, ultimately obtaining multiple anomaly evolution reliability coefficients corresponding one-to-one with each event stream.

[0076] In one possible implementation, step S500 further includes: Step S510: Traverse the multiple real abnormal evolution event streams to identify abnormal semantics, abnormal amplitude, and abnormal trend features, and obtain multiple multi-dimensional abnormal identification results.

[0077] Step S520: Based on the multiple multidimensional anomaly identification results, perform anomaly-free power prediction to obtain multiple benchmark power prediction results.

[0078] Step S530: Perform power error correction and compensation decision analysis based on the multiple benchmark power prediction results to obtain the target power error online correction and compensation parameters.

[0079] Specifically, each real abnormal evolution event stream is traversed one by one. For a single event stream, abnormal semantic recognition is first performed: cross-scale fusion feature data associated with the event stream is retrieved, including equipment operating conditions, environmental disturbances, and user behavior information. Abnormal features in the event stream are matched with a pre-set abnormal semantic library, covering common semantic tags and corresponding feature templates such as metering equipment aging deviation, environmental temperature interference, and abnormalities caused by electrical equipment failures. The most fitting abnormal semantic is determined through a feature similarity algorithm to clarify the core cause of the abnormality. Next, abnormal amplitude recognition is performed: the fusion feature values ​​of all abnormal points in the event stream are extracted, and the deviation of each abnormal point from the mean feature value of the corresponding normal period is calculated. The deviation is measured by the average and maximum values ​​of the deviation, which quantifies the overall severity and extreme deviations of the anomaly. Finally, anomaly trend identification is carried out: the changes in the deviation of the anomaly points in the event stream are sorted out in chronological order, and the anomaly is determined by linear fitting or sliding window analysis to determine whether the anomaly is gradually aggravated, continuously stable, or gradually receding. The complete duration of the anomaly from start to finish is recorded. The anomaly semantics, anomaly amplitude statistics, anomaly trend type, and duration identified above are integrated into a set of structured data, which is the multidimensional anomaly identification result corresponding to the event stream. After traversing all real anomaly evolution event streams, multiple multidimensional anomaly identification results consistent with the number of event streams are finally obtained.

[0080] For each multidimensional anomaly identification result, key information is extracted. The specific time period of the anomaly includes the start and end times, the corresponding user electricity consumption scenario (e.g., weekday commuting hours, all day on holidays), and normal electricity consumption characteristics before the anomaly (e.g., load curves and average electricity consumption 1-3 days prior to the anomaly). This clarifies the prediction time range and scenario basis. A prediction dataset is then constructed: historical data from the same period in the historical electricity consumption database that matches the current anomaly scenario and has no metering anomalies is selected, such as electricity consumption data from the same type of weekday in the same season and under similar ambient temperatures. Normal electricity consumption characteristics before the anomaly are used as supplementary samples to form the training and validation dataset for the neural network model. Next, a pre-defined LSTM (Long Short-Term Memory) neural network model is used, with time features, user behavior features, and environmental features from the training data as input features, and historical normal electricity consumption as the output label. The model is trained and its parameters optimized to ensure it can accurately learn the electricity consumption variation patterns under normal scenarios. Finally, the time and scene features of the abnormal period to be predicted are input into the trained model. The model calculates the theoretical electricity consumption value if no metering abnormality occurs during the period by learning the normal electricity consumption pattern. This value is the baseline electricity prediction result of the corresponding multi-dimensional anomaly identification result. After traversing all multi-dimensional anomaly identification results, multiple baseline electricity prediction results are finally obtained.

[0081] First, the difference between the predicted baseline electricity consumption and the actual metered electricity consumption during the corresponding abnormal period is calculated to obtain the electricity error value. Then, the electricity calculation model based on the LSTM (Long Short-Term Memory) algorithm is called: historical metering, load, environmental data, and time-series abnormal time and duration are input into the LSTM model, allowing the model to learn the electricity consumption change pattern under normal electricity consumption scenarios, while fitting the dynamic correlation between abnormal scenarios and compensation parameters. By memorizing long-term electricity consumption patterns and short-term abnormal features, the model first outputs the theoretical compensation baseline value for the abnormal period, and then automatically matches the compensation ratio for the corresponding abnormal type and adjusts the compensation cycle parameters based on the learned compensation strategy logic. Finally, the target electricity error online compensation parameters, including compensation electricity consumption, execution time, and rule standards, are integrated and output.

[0082] Example 2, based on the same inventive concept as the online error correction method for electricity metering in the foregoing examples, such as... Figure 2 As shown, this application provides an online error correction system for electricity metering. The system and method embodiments in this application are based on the same inventive concept. The system includes: The time chain construction module 10 is used to acquire a multimodal data set during the electricity metering process, perform cross-scale long-time-series feature fusion on the multimodal data set, and construct a cross-scale fused feature time chain.

[0083] The anomaly evolution analysis module 20 is used to perform anomaly evolution analysis of power metering based on the cross-scale fusion feature time chain to obtain multiple anomaly evolution event streams.

[0084] The anomaly assessment module 30 is used to traverse the multiple anomaly evolution event streams to perform anomaly assessment and obtain multiple anomaly evolution reliability coefficients.

[0085] The mapping extraction module 40 is used to map and extract multiple abnormal evolution event streams based on the abnormal evolution reliability coefficients among the multiple abnormal evolution reliability coefficients that are greater than or equal to a preset coefficient threshold, so as to obtain multiple real abnormal evolution event streams.

[0086] The decision analysis module 50 is used to perform power error refund and compensation decision analysis based on the multiple real abnormal evolution event streams to obtain the target power error online refund and compensation parameters.

[0087] Furthermore, the system is also used to implement the following functions: The system collects voltage, current, energy consumption, and power factor from the electricity meter to obtain metering-side data; it also collects user load curves, peak load characteristics, off-peak load characteristics, and the start / stop status of electrical equipment to obtain load-side data; it collects temperature, humidity, and weather disturbance factors from multi-source environmental sensors to obtain environmental-side data; it extracts health status, aging level, and error codes from the daily logs of metering equipment to obtain equipment operating condition data; it extracts user behavior data according to daily and holiday electricity consumption patterns to obtain user behavior data; and it summarizes the metering-side data, load-side data, environmental-side data, equipment operating condition data, and user behavior data to obtain the multimodal data set.

[0088] Furthermore, the system is also used to implement the following functions: The multimodal dataset is aligned with time windows to obtain an aligned multimodal dataset. The aligned multimodal dataset is then segmented into windows and in-window features are extracted according to micro, meso, and macro windows to obtain micro-window feature time chains, meso-window feature time chains, and macro-window feature time chains. The micro-window feature time chains, meso-window feature time chains, and macro-window feature time chains are then fused across scales to obtain a cross-scale fused feature time chain.

[0089] Furthermore, the system is also used to implement the following functions: The microscopic window feature time chain is directly mapped to the full time axis at a resolution of seconds or minutes; the mesoscopic window feature time chain is assigned values ​​on the full time axis according to the corresponding time intervals; the macroscopic window feature time chain is embedded into the full time axis after the mapping and amplitude of the mesoscopic and microscopic windows according to the corresponding time intervals to obtain the integrated full time axis; cross-scale fusion is performed based on the integrated full time axis to obtain the cross-scale fused feature time chain.

[0090] Furthermore, the system is also used to implement the following functions: Extract the first macroscopic window features and the corresponding first mesoscopic window feature set from the integrated full time axis; perform cross-scale fusion on the first mesoscopic window feature set based on the first macroscopic window features to obtain a first fused mesoscopic window feature set; update the integrated full time axis based on the first fused mesoscopic window feature set; extract the first fused mesoscopic window features from the first fused mesoscopic window feature set, and extract the microscopic window features of the corresponding time intervals in the integrated full time axis to obtain a first microscopic window feature set; perform cross-scale fusion on the first microscopic window feature set based on the first fused mesoscopic window features to obtain a first microscopic mesoscopic window feature set; update the integrated full time axis according to the first microscopic mesoscopic window feature set; after each time interval in the integrated full time axis has been updated, extract the updated time chain corresponding to the microscopic window to obtain a cross-scale fused feature time chain.

[0091] Furthermore, the system is also used to implement the following functions: The feature similarity between the first macroscopic window feature set and the first mesoscopic window feature set is calculated by iterating through the data to obtain the first macroscopic cross-scale similarity cluster; the first macroscopic cross-scale similarity cluster is normalized within the set to obtain the first macroscopic cross-scale fusion matrix set; the first mesoscopic window feature set is mapped and fused to enhance it according to the first macroscopic cross-scale fusion matrix set to obtain the first fused mesoscopic window feature set.

[0092] Furthermore, the system is also used to implement the following functions: Anomalies are extracted from the cross-scale fusion feature time chain to obtain multiple anomalies. Based on the anomaly magnitude and anomaly direction of the multiple anomalies, the cross-scale fusion feature time chain is back-tracked to obtain multiple anomaly evolution event streams.

[0093] Furthermore, the system is also used to implement the following functions: Obtain anomaly assessment metrics, including anomaly magnitude metrics, suddenness metrics, and context consistency metrics; perform anomaly assessment on the multiple anomalous evolution event streams based on the anomaly assessment metrics to obtain multiple anomalous evolution reliability coefficients.

[0094] Furthermore, the system is also used to implement the following functions: The system iterates through multiple real abnormal evolution event streams to identify abnormal semantics, abnormal amplitude, and abnormal trend features, obtaining multiple multi-dimensional abnormal identification results; based on the multiple multi-dimensional abnormal identification results, it performs abnormal power prediction to obtain multiple benchmark power prediction results; and based on the multiple benchmark power prediction results, it performs power error correction decision analysis to obtain target power error online correction parameters.

[0095] 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.

[0096] 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.

[0097] 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. A method for online error correction and compensation in electricity metering, characterized in that, The method includes: A multimodal data set is acquired during the electricity metering process, and cross-scale long-time-series feature fusion is performed on the multimodal data set to construct a cross-scale fused feature time chain; Based on the cross-scale fusion feature time chain, anomaly evolution analysis of power metering is performed to obtain multiple anomaly evolution event streams; The multiple abnormal evolution event streams are traversed to perform anomaly assessment and obtain multiple abnormal evolution reliability coefficients; Based on the abnormal evolution reliability coefficients that are greater than or equal to a preset coefficient threshold among the multiple abnormal evolution reliability coefficients, the multiple abnormal evolution event streams are mapped and extracted to obtain multiple real abnormal evolution event streams. Based on the multiple real abnormal evolution event streams, power error refund and compensation decision analysis is performed to obtain the target power error online refund and compensation parameters.

2. The online error correction method for electricity metering as described in claim 1, characterized in that, Acquire multimodal data sets during the electricity metering process, including: The metering side collects voltage, current, energy consumption, and power factor data measured by the electricity meter to obtain metering side data; Collect user load curves, peak load characteristics, off-peak load characteristics, and the start / stop status of electrical equipment to obtain load-side data; Environmental data is obtained by collecting temperature, humidity, and weather disturbance factors through multi-source environmental sensors; Health status, aging level and error codes are extracted from the daily logs of metering equipment to obtain equipment operating condition data; User behavior data is extracted based on daily electricity consumption patterns and holiday electricity consumption patterns. The multimodal data set is obtained by summarizing the metering data, load data, environmental data, equipment operating condition data, and user behavior data.

3. The online error correction method for electricity metering as described in claim 1, characterized in that, The multimodal dataset is subjected to cross-scale long-time-series feature fusion to construct a cross-scale fused feature time chain, including: The multimodal data set is aligned with a time window to obtain an aligned multimodal data set; The aligned multimodal dataset is segmented into micro-windows, meso-windows, and macro-windows, and features within each window are extracted to obtain micro-window feature time chains, meso-window feature time chains, and macro-window feature time chains. The microscopic window feature time chain, mesoscopic window feature time chain and macroscopic window feature time chain are fused across scales to obtain a cross-scale fused feature time chain.

4. The online error correction method for electricity metering as described in claim 3, characterized in that, Cross-scale fusion is performed on the microscopic window feature time chain, the mesoscopic window feature time chain, and the macroscopic window feature time chain to obtain a cross-scale fused feature time chain, including: The micro-window feature time chain is directly mapped to the full time axis at a resolution of seconds or minutes. Assign values ​​to the full time axis according to the corresponding time intervals of the mesoscopic window feature time chain; The macroscopic window feature time chain is embedded into the full time axis after the mapping and amplitude of the mesoscopic and microscopic windows according to the corresponding time intervals to obtain the integrated full time axis; Cross-scale fusion is performed based on the integrated full time axis to obtain a cross-scale fused feature time chain.

5. The online error correction and compensation method for electricity metering as described in claim 4, characterized in that, Based on the integrated full timeline, cross-scale fusion is performed to obtain a cross-scale fused feature time chain, including: Extract the first macroscopic window features and the corresponding first mesoscopic window feature set from the integrated full time axis; Based on the first macroscopic window features, the first mesoscopic window feature set is fused across scales to obtain the first fused mesoscopic window feature set. The integrated full time axis is updated based on the first fused mesoscopic window feature set; Extract the first fused mesoscopic window features from the first fused mesoscopic window feature set, and extract the microscopic window features of the corresponding time interval in the integrated full time axis to obtain the first microscopic window feature set; Based on the first fused mesoscopic window features, the first microscopic window feature set is fused across scales to obtain the first microscopic mesoscopic window feature set; The integrated full time axis is updated based on the first micro- and meso-level window feature set. After each time interval in the integrated full time axis has been updated, the updated time chain corresponding to the micro-window is extracted to obtain the cross-scale fusion feature time chain.

6. The online error correction method for electricity metering as described in claim 5, characterized in that, Based on the first macroscopic window features, the first mesoscopic window feature set is fused across scales to obtain a first fused mesoscopic window feature set, including: The feature similarity between the first macroscopic window feature set and the first mesoscopic window feature set is calculated to obtain the first macroscopic cross-scale similarity cluster. The cross-scale similarity clusters in the first macro are subjected to set-in-set normalization to obtain the set of cross-scale fusion matrices in the first macro. The first mesoscopic window feature set is mapped, fused, and enhanced based on the first macroscopic cross-scale fusion matrix set to obtain the first fused mesoscopic window feature set.

7. The online error correction method for electricity metering as described in claim 1, characterized in that, Based on the aforementioned cross-scale fused feature time chain, anomaly evolution analysis of power metering was performed to obtain multiple anomaly evolution event streams, including: Extract outliers from the cross-scale fusion feature time chain to obtain multiple outliers; Based on the abnormal amplitude and abnormal direction of the multiple anomalies, the cross-scale fused feature time chain is back-tracked and correlated to obtain multiple abnormal evolution event streams.

8. The online error correction method for electricity metering as described in claim 1, characterized in that, Anomaly assessment is performed by traversing the multiple anomalous evolution event streams to obtain multiple anomalous evolution reliability coefficients, including: Obtain anomaly assessment indicators, wherein the anomaly assessment indicators include anomaly magnitude indicators, suddenness indicators, and context consistency indicators; Anomaly assessment is performed on the multiple anomalous evolution event streams based on the aforementioned anomaly assessment index to obtain multiple anomalous evolution reliability coefficients.

9. The online error correction method for electricity metering as described in claim 1, characterized in that, Based on the multiple real abnormal evolution event streams, power error refund and compensation decision analysis is performed to obtain the target power error online refund and compensation parameters, including: By traversing the multiple real abnormal evolution event streams, abnormal semantics, abnormal amplitude, and abnormal trend features are identified, and multiple multi-dimensional abnormal identification results are obtained. Based on the multiple multidimensional anomaly identification results, anomaly-free power prediction is performed to obtain multiple benchmark power prediction results. Based on the multiple benchmark power prediction results, a power error correction and compensation decision analysis is performed to obtain the target power error online correction and compensation parameters.

10. An online error correction and compensation system for electricity metering, characterized in that, The system is used to implement the online error correction and compensation method for electricity metering as described in any one of claims 1-9, and the system comprises: The time chain construction module is used to acquire a multimodal data set during the electricity metering process, perform cross-scale long-time-series feature fusion on the multimodal data set, and construct a cross-scale fused feature time chain. An anomaly evolution analysis module is used to perform anomaly evolution analysis of power metering based on the cross-scale fused feature time chain, and obtain multiple anomaly evolution event streams; An anomaly assessment module is used to traverse the multiple anomaly evolution event streams to perform anomaly assessments and obtain multiple anomaly evolution reliability coefficients. The mapping extraction module is used to map and extract multiple abnormal evolution event streams based on the abnormal evolution reliability coefficients that are greater than or equal to a preset coefficient threshold among the multiple abnormal evolution reliability coefficients, so as to obtain multiple real abnormal evolution event streams. The decision analysis module is used to perform power error refund and compensation decision analysis based on the multiple real abnormal evolution event streams to obtain the target power error online refund and compensation parameters.

Images