Edge computing electric energy metering data real-time analysis method and system

Abstract

This invention relates to the field of power monitoring, specifically to a method and system for real-time analysis of edge computing power metering data. The method includes the following steps: collecting raw sampling data from all metering nodes in a distribution area, completing hardware-level clock synchronization and lossless spatiotemporal alignment; dividing causal-related node groups according to property topology, extracting and filtering low-dimensional core causal features; embedding rigid constraints in power physics, constructing a lightweight baseline causal graph, and calibrating the baseline effect thresholds for endogenous normal causal paths and exogenous abnormal causal paths; collecting real-time incremental metering data, completing standardized preprocessing and feature extraction; performing lightweight causal inference based on the baseline causal graph, calculating real-time causal effect values; and distinguishing between normal bidirectional power flow and abnormal events through two-level threshold verification, completing the classification and determination of electricity theft or metering faults. This invention effectively solves the pain points of traditional analysis, such as inaccurate data, insufficient computing power, ambiguous judgments, and lack of source tracing.

Description

Technical Field

[0001] This invention belongs to the field of power monitoring, specifically relating to a method and system for real-time analysis of edge computing power metering data. Background Technology

[0002] With the large-scale integration of distributed photovoltaics, residential energy storage, and charging piles, the metering of distribution areas has changed from one-way electricity consumption to two-way power flow. Traditional anomaly identification models based on correlation are prone to misjudging the backflow of new energy sources and bidirectional load fluctuations as electricity theft or metering failures. Accurate identification must be achieved by relying on causal inference models. However, the high computing power requirements of high-precision causal inference models are in unavoidable conflict with the low power consumption and highly limited hardware resources of edge nodes, which is the biggest technical obstacle for edge metering analysis to adapt to the new power system. Summary of the Invention

[0003] The purpose of this invention is to provide a method for real-time analysis of edge computing power metering data, and at the same time, to provide a system for real-time analysis of edge computing power metering data, so as to solve the problems mentioned in the background art.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for real-time analysis of edge computing power metering data, including the following steps: Collect raw sampling data from all metering nodes in the distribution area and complete hardware-level clock synchronization and lossless spatiotemporal alignment; Divide causal relationship node groups according to property rights topology, and extract and filter low-dimensional core causal features; By embedding rigid constraints in electrical physics, a lightweight baseline causal graph is constructed, and the baseline effect thresholds for endogenous normal causal paths and exogenous abnormal causal paths are calibrated. Collect real-time incremental metering data and complete standardized preprocessing and feature extraction; Lightweight causal inference is performed based on a baseline causal graph, and real-time causal effect values ​​are calculated. By using two levels of threshold verification, normal bidirectional power flow can be distinguished from abnormal events, thus completing the classification and determination of electricity theft or metering faults.

[0005] Furthermore, the process of collecting raw sampling data from all metering nodes in the distribution area and achieving hardware-level clock synchronization and lossless spatiotemporal alignment includes: collecting raw sampling data from all legally mandated metering nodes within the distribution area through the distribution area edge gateway, including the distribution area main meter, household electricity meters, distributed photovoltaic grid-connected meters, household energy storage bidirectional meters, and charging pile meters; synchronizing the hardware-level clock of the distribution area edge gateway with that of all metering nodes; removing noise and outliers from the raw sampling data; and achieving spatiotemporal alignment for sampling nodes with different frequencies, unifying the time granularity.

[0006] Furthermore, the step of dividing causal relationship node groups according to property rights topology includes: grouping all metering nodes in the entire transformer area according to the physical topology of the transformer area and the ownership of metering nodes; grouping all metering nodes under the same property owner into the same causal relationship node group; ensuring that nodes within the group have power transfer causal relationships and property rights consistency; setting the transformer area gateway master table as a global reference node and not including it in the user-side node group; decomposing the overall causal calculation of the transformer area into multiple independent units through grouping; reducing the search for invalid causal paths and reducing the amount of computation; and assigning a unique topology identifier to each causal relationship node group and binding it with the corresponding data and feature sequences.

[0007] Furthermore, the extraction and screening of low-dimensional core causal features includes: based on the causal association node groups divided in the initialization phase, extracting core causal features with physical causal relationships using a single node group as a unit; performing redundancy screening on the extracted initial features to remove features without statistical significance; and binding the screened low-dimensional features with time series sequences and node topology attribution identifiers to generate a time series sequence of low-dimensional core causal features.

[0008] Furthermore, the embedded rigid constraints of power physics include: based on Kirchhoff's laws of the power system and the physical topology of the transformer area, three types of rigid hard constraints are embedded: causal directionality constraints, node group power balance constraints, and causal graph acyclic constraints; the causal directionality constraints are embedded to ensure that the causal direction of power transmission is consistent with the power flow direction; the node group power balance constraints are embedded to define the power transmission association within the same owner's node group as an endogenous normal causal path that needs to be retained; and the causal graph acyclic constraints are embedded to ensure that the causal graph conforms to the physical logic of power transmission; the three types of constraints are embedded throughout the entire causal structure learning process, and the search space is limited before the conditional independence test.

[0009] Furthermore, the construction of the lightweight baseline causal graph includes: performing conditional independence tests on the low-dimensional core causal features of each causal association node group based on the search space limited by rigid constraints of power physics; constructing an initial causal graph for the distribution area based on the test results; calculating the causal dependencies and path coefficients between feature variables; dividing the causal paths into endogenous normal causal paths and exogenous abnormal causal paths according to the bidirectional power flow characteristics of the distribution area and completing the calibration; performing lightweight pruning on the initial causal graph to remove redundant causal edges; calibrating the baseline causal effect threshold of each causal path based on the historical operating data of the distribution area; and generating a lightweight baseline causal graph adapted to the low computing power environment of the edge side after verification of computing power occupancy.

[0010] Furthermore, the benchmark effect thresholds for calibrating endogenous normal causal paths and exogenous abnormal causal paths include: based on the lightweight initial causal graph constructed during initialization, benchmark effect thresholds are calibrated for the two types of causal paths according to the historical normal operation data and abnormal evidence data of the transformer area; the thresholds of each path are calibrated separately according to the laws of power physics and the operating characteristics of the transformer area; after calibration, the thresholds are bound to the benchmark causal graph with the path classification label and topology identifier, and the accuracy of the thresholds is verified. If the requirements are not met, the thresholds are recalibrated until they are met.

[0011] Furthermore, the step of completing lightweight causal inference based on the benchmark causal graph and calculating the real-time causal effect value includes: using the initialized lightweight benchmark causal graph of the transformer area as a benchmark, for the preprocessed real-time incremental low-dimensional causal feature sequence, using a lightweight inference algorithm, performing causal inference with a single causal association node group as a unit, following the rigid constraints of power physics, quantifying the causal contribution of each core causal feature to the change of the node group's metering value, completing the binding of the real-time causal effect value with relevant identifiers, and forming a set for threshold verification.

[0012] Furthermore, the two-level threshold verification distinguishes between normal bidirectional power flow and abnormal events, and completes the classification and judgment of electricity theft or metering faults. This includes: conducting two-level progressive threshold verification based on real-time causal effect values, path labels of the baseline causal graph, and thresholds. The first level verifies whether the endogenous normal causal path effect value is within the threshold range to determine whether it is a normal bidirectional power flow. The second level verifies the identification of abnormal events and distinguishes between electricity theft and metering faults based on the causal contribution decomposition results and abnormal electrical characteristics. The abnormal classification, path tracing, time and node information are bound to the analysis results to provide a basis for abnormal handling and evidence preservation.

[0013] This application also discloses an electronic device, including: At least one processor; and

[0014] A memory communicatively connected to the at least one processor; wherein,

[0015] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the above-described edge computing power metering data real-time analysis method of the present invention.

[0016] This application also discloses a real-time analysis system for edge computing power metering data, including: The data acquisition and synchronization module is used to collect raw sampling data from all metering nodes in the distribution area and complete hardware-level clock synchronization and lossless spatiotemporal alignment. The node feature processing module is used to divide causal relationship node groups according to property rights topology and extract and filter low-dimensional core causal features. The causal graph construction module is used to embed rigid constraints in electrical physics, construct a lightweight baseline causal graph, and calibrate the baseline effect thresholds for endogenous normal causal paths and exogenous abnormal causal paths. The real-time data preprocessing module is used to collect real-time incremental metering data and complete standardized preprocessing and feature extraction. The causal reasoning calculation module is used to perform lightweight causal reasoning based on the baseline causal graph and calculate the real-time causal effect value. The anomaly detection module is used to distinguish between normal bidirectional power flow and abnormal events through two-level threshold verification, and to classify and determine electricity theft or metering faults.

[0017] Beneficial effects: This application collects raw data from all metering nodes and achieves hardware-level clock synchronization and lossless spatiotemporal alignment. Combined with property topology grouping and decomposition of computing units, it reduces invalid causal searches and lowers computing power consumption, adapting to low-computing-power edge environments. It incorporates rigid constraints of power physics and filters low-dimensional core causal features, making the causal graph conform to the power operation logic and improving the accuracy of causal inference. Relying on a lightweight benchmark causal graph and inference algorithm, coupled with two-level progressive threshold verification, it can accurately distinguish between normal bidirectional power flow and abnormal events, clearly define electricity theft and metering faults, and realize abnormal path tracing, node and time information binding, providing a reliable basis for anomaly handling and evidence preservation, effectively solving the pain points of inaccurate data, insufficient computing power, ambiguous judgment, and lack of tracing in traditional analysis. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of the edge computing power metering data real-time analysis method of the present invention; Figure 2 is a flowchart of the steps for dividing causal association node groups according to property rights topology in this invention; Figure 3 is a flowchart of the steps for extracting and screening low-dimensional core causal features in this invention; Figure 4 is a flowchart of the steps for embedding rigid electrical physics constraints and constructing a lightweight baseline causal graph in this invention; Figure 5 is a flowchart of the steps of causal path benchmark effect threshold calibration, real-time incremental measurement data acquisition and preprocessing and feature extraction in this invention; Figure 6 is a flowchart of the steps of lightweight causal reasoning calculation, two-level threshold verification and anomaly classification judgment in this invention; Figure 7 This is a comparison chart of the false alarm rates in a two-way power flow scenario between this application and existing technologies; Figure 8 This is a graph comparing the inference time and computing power usage of this application and existing technologies on the dual Y-axis. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] This invention provides a method for real-time analysis of edge computing power metering data, such as... Figure 1 As shown, the steps include: Collect raw sampling data from all metering nodes in the distribution area and complete hardware-level clock synchronization and lossless spatiotemporal alignment; Divide causal relationship node groups according to property rights topology, and extract and filter low-dimensional core causal features; By embedding rigid constraints in electrical physics, a lightweight baseline causal graph is constructed, and the baseline effect thresholds for endogenous normal causal paths and exogenous abnormal causal paths are calibrated. Collect real-time incremental metering data and complete standardized preprocessing and feature extraction; Lightweight causal inference is performed based on a baseline causal graph, and real-time causal effect values ​​are calculated. By using two levels of threshold verification, normal bidirectional power flow can be distinguished from abnormal events, thus completing the classification and determination of electricity theft or metering faults.

[0021] The acquisition of raw sampling data from all metering nodes in the distribution area specifically includes: performing raw sampling data acquisition operations on all metering nodes in the distribution area through the distribution area edge gateway. The acquisition scope includes all legally mandated metering nodes within the distribution area, specifically including the distribution area main meter, household electricity meters, distributed photovoltaic grid-connected meters, household energy storage bidirectional meters, and charging pile meters. The unified sampling frequency for all nodes is 1kHz. The acquired data items include instantaneous voltage, instantaneous current, active power, reactive power, energy readings, power flow direction indicators, and sampling timestamps for each metering node. The data acquisition process is compatible with DL / T645, IEC61850, and Modbus power industry standard protocols to ensure the complete acquisition and compliant access of raw sampling data from all nodes.

[0022] The implementation of hardware-level clock synchronization and lossless spatiotemporal alignment specifically includes: using the IEEE 1588PTP precision clock synchronization protocol to complete hardware-level clock synchronization between the edge gateway of the distribution area and all metering nodes in the entire distribution area, controlling the synchronization error of the sampling timestamps of all nodes within 1ms, and providing a unified and compliant time reference for all metering data.

[0023] The sliding window 3σ criterion is used to remove spike noise and outliers from the acquired raw sampling data. The removal process only deletes invalid sampling points caused by electromagnetic interference, without performing interpolation to complete the data, thus fully preserving the originality and legal validity of the measurement data.

[0024] For different frequency sampling nodes, the power flow inflection point matching method is used to complete the equal-interval spatiotemporal alignment. Sampling point matching and time granularity are only performed at the inflection point where the power flow direction is reversed. The time granularity of all node data is unified to 10ms level, avoiding the additional computing power consumption caused by full resampling. Finally, a time series metering dataset with unique timestamp identifier, power flow direction identifier, and node topology attribution identifier is generated.

[0025] The division of causal relationship node groups according to property rights topology, such as Figure 2 As shown, the specific implementation includes: performing grouping and division of all metering nodes in the entire transformer area based on the physical topology of the transformer area and the legal ownership information of the metering nodes.

[0026] All legally registered metering nodes corresponding to the same property owner are grouped into the same independent causal node group, specifically including household electricity meters, distributed photovoltaic grid-connected meters, household energy storage bidirectional meters, and charging pile meters under the name of the property owner, to ensure that the metering nodes in a single node group have a direct causal relationship for power transfer and consistency in property ownership.

[0027] The overall table of gateways in the entire distribution area is set as a global reference node for the entire distribution area, and is not included in any user-side causal relationship node group. It serves as a unified reference for power balance verification and causal path benchmark comparison for the entire distribution area. After grouping, the causal calculation scope of all nodes in the entire distribution area is decomposed into causal calculation units within multiple independent node groups. This eliminates meaningless causal path searches between node groups of different ownership entities and compresses the computational scale of causal structure learning and causal inference at the topology level.

[0028] A unique node topology affiliation identifier is assigned to each causal association node group. This identifier is then bound and matched one-to-one with the original sampling data and time-series feature sequences of all metering nodes within the corresponding node group, providing a unified topology identifier benchmark for subsequent feature extraction and causal analysis of the node groups.

[0029] The extraction and screening of low-dimensional core causal features, such as Figure 3 As shown, the implementation specifically includes: based on the independent causal relationship node groups that have completed topological partitioning and property ownership matching in the initialization phase, causal feature extraction is performed on a single node group as an independent unit. Feature extraction is performed throughout the entire process to ensure that the extracted features have interpretable physical causal relationships, rather than simple statistical correlation features.

[0030] For each independent causal node group, six types of core causal features with fixed dimensions are extracted, including power flow direction timing features, active power bidirectional fluctuation amplitude and slope features, node group power balance deviation features, voltage-current phase difference timing features, current harmonic distortion rate features, and metering loop on / off status features. Each type of feature has a direct physical causal correspondence with the power transfer path, power flow direction changes, and metering loop operation status within the node group.

[0031] For the full set of initial features extracted from a single node group, the maximum information coefficient (MIC) is used to perform feature redundancy screening and verification. The verification process has a fixed significance level of 0.05. The statistical correlation between each feature and the change of the corresponding node group measurement value is quantitatively calculated. Only features that have a statistically significant correlation with the change of the node group measurement value are retained, and redundant features without statistical significance are removed.

[0032] After completing the redundancy screening, the feature dimension of a single causal node group is fixed to within 10 dimensions. A unique feature saliency identifier is assigned to each retained effective feature. The selected low-dimensional feature time series sequence is bound and matched one-to-one with the node topology attribution identifier of the corresponding node group to generate a low-dimensional core causal feature time series sequence of sub-node groups with node topology association labels and feature saliency identifiers. This ensures that the dimension of the output feature sequence is completely matched with the input dimension requirements of the subsequent causal structure learning and real-time causal inference stages.

[0033] The implanted electrical physical rigid constraint, such as Figure 4 As shown, the specific implementation includes: based on Kirchhoff's fundamental laws of power systems and the established physical topology of the transformer substations, three insurmountable rigid hard constraints are implanted to limit the effective search space for causal structure learning, eliminating the possibility of generating false causal paths from the physical principle level, significantly compressing the computational scale of causal structure learning, and adapting to the computing power constraints of edge nodes, including: By imposing rigid constraints on causal direction, the causal direction of power transfer is strictly consistent with the power flow direction identifier collected by the metering node, prohibiting reverse causal paths that are opposite to the power flow direction, thus eliminating false causal relationships that violate the physical laws of power transfer from the root. By implanting rigid constraints on the power balance of node groups, within the causal relationship node group corresponding to the same property owner, there is a rigid causal relationship between the output of distributed photovoltaic power, the charging and discharging power of household energy storage, the power consumption of charging piles, the power consumption of household loads and the metering value of household electricity meters in that node group. This relationship path is defined as an endogenous normal causal path, which is a benchmark path that must be retained in causal structure learning. It is forbidden to remove or tamper with this core causal relationship. By embedding a loop-free rigid constraint in the causal graph, the final causal graph must fully conform to the physical logic of power transfer in the power system, prohibiting the appearance of false causal loops that violate the laws of power transfer, and ensuring the physical interpretability and compliance of the causal structure.

[0034] Three rigid hard constraints are embedded throughout the entire process of causal structure learning. The search space is limited before the conditional independence test is performed. Subsequent causal structure learning operations are only performed within the effective range defined by the constraints, avoiding the computational power consumption caused by unbounded full causal search. At the same time, it ensures that the constructed benchmark causal graph fully conforms to the physical nature of the power operation of the transformer area and is fully matched with the endogenous and exogenous causal path verification logic of the subsequent real-time causal reasoning stage.

[0035] The construction of a lightweight baseline causal graph, such as Figure 4 As shown, the specific implementation includes: based on the effective search space defined by the rigid hard constraints of the embedded power physics, performing conditional independence tests on the low-dimensional core causal features screened within each independent causal relationship node group. The conditional independence test adopts a parallelized t-test method adapted to the parallel processing capabilities of the multi-core MCU of the edge gateway, with a fixed significance level of 0.05. The test calculation is only completed within the effective search range defined by the rigid constraints, avoiding the high computational cost caused by the unbounded full causal search.

[0036] Based on the output of the conditional independence test, an initial causal graph of the distribution area is constructed. The causal dependencies between each characteristic variable in the initial causal graph and the path coefficients of the corresponding causal paths are quantitatively calculated and determined. At the same time, based on the physical logic of power transfer in the power system and the bidirectional power flow operation characteristics of the distribution area, all causal paths in the initial causal graph are classified and labeled. The causal paths corresponding to bidirectional power flow fluctuations caused by distributed photovoltaic power backfeeding, household energy storage charging and discharging, and charging pile start-up and shutdown are defined as endogenous normal causal paths. The causal paths corresponding to abnormal deviations in metering values ​​caused by electricity theft and metering equipment failure are defined as exogenous abnormal causal paths. This provides a standardized classification benchmark for path verification and anomaly identification in the subsequent normalized real-time analysis stage.

[0037] A lightweight pruning operation is performed on the initial causal graph after path classification and labeling. Based on the feature saliency label and the causal path contribution verification results, redundant causal edges that have no statistical significance or physical causal relationship are removed. After pruning, the number of nodes in the causal graph corresponding to a single causal relationship node group is controlled within 10, and the number of causal edges is controlled within 15, to ensure that the structural complexity of the causal graph is fully adapted to the hardware resource constraints of the edge nodes.

[0038] Based on the full-node compliant metering dataset of the historical normal operation period of the transformer substation, the baseline causal effect threshold is calibrated on the trimmed causal graph. For each endogenous normal causal path, the upper and lower limits of the causal effect threshold in its normal operation range are quantified and calibrated. At the same time, the abnormal trigger judgment threshold of the exogenous abnormal causal path is calibrated to ensure that the threshold calibration results fully conform to the actual source-load-storage access characteristics and operation rules of the corresponding transformer substation.

[0039] After completing the calibration of the baseline causal effect threshold, a special verification of computing power consumption is performed on the final lightweight baseline causal graph of the transformer area. By simulating the real-time causal inference operation of the entire process, it is confirmed that the computing power consumption of the real-time inference process corresponding to the baseline causal graph does not exceed 10% of the rated computing power of the edge gateway MCU. After the verification is passed, the corresponding endogenous or exogenous path classification labels, baseline causal effect thresholds, and node topology affiliation identifiers are bound to all causal paths of the baseline causal graph. Finally, a lightweight baseline causal graph of the transformer area that is fully adapted to the low computing power operation environment of the edge side is generated, providing a unified benchmark model basis for causal inference calculation and accurate identification of abnormal events in the subsequent normalized real-time analysis stage.

[0040] The baseline effect thresholds for calibrating endogenous normal causal pathways and exogenous abnormal causal pathways, such as... Figure 5 As shown, the specific implementation includes: based on the lightweight initial causal graph of the transformer area constructed in the initialization phase, and based on the full-node compliant metering dataset of the transformer area's historical normal operation period, and based on the endogenous normal causal path and exogenous abnormal causal path that have been classified and calibrated, performing a quantitative calibration operation on the benchmark causal effect threshold.

[0041] For each endogenous normal causal path, the causal effect value corresponding to the path during the historical normal operation of the transformer area is extracted. The normal distribution range of the causal effect value of the path is calculated using the sliding window statistical method. The upper and lower limits of the distribution range are used as the benchmark effect threshold of the endogenous normal causal path to ensure that the threshold can accurately include the causal effect change range corresponding to normal bidirectional power flow fluctuations such as new energy backflow, energy storage charging and discharging, and charging pile start-up and shutdown.

[0042] For each exogenous abnormal causal path, based on the evidence data of historical abnormal events (electricity theft, metering failure) in the transformer area, the minimum causal effect value of the exogenous abnormal causal path when such an abnormal event occurs is extracted. This minimum causal effect value is determined as the abnormal triggering baseline effect threshold of the exogenous abnormal causal path, ensuring that the abnormality judgment is triggered only when the real-time causal effect value reaches or exceeds this threshold.

[0043] During the threshold calibration process, the baseline effect threshold of each causal path is calibrated separately in strict accordance with the physical laws of power transmission in the power system and the source-load-storage access characteristics of the distribution area, so as to avoid confusion of thresholds for different paths. After calibration, the baseline effect threshold of each causal path is bound to the classification label and node topology affiliation identifier of the path and embedded into the lightweight baseline causal graph. This ensures that the baseline threshold of the corresponding path can be quickly matched during subsequent real-time causal effect verification. At the same time, all calibrated baseline effect thresholds are verified to confirm that the threshold calibration accuracy meets the requirement that the anomaly identification accuracy rate is not less than 99%. If the verification fails, historical data is extracted again for secondary calibration until the accuracy requirement is met. Finally, the calibration of the baseline effect thresholds of all endogenous normal causal paths and exogenous abnormal causal paths is completed.

[0044] The process involves collecting real-time incremental metering data, performing standardized preprocessing and feature extraction, such as... Figure 5 As shown, the implementation specifically includes: incremental real-time sampling data acquisition of all metering nodes in the distribution area is performed through the distribution area edge gateway at a fixed 1-second cycle. The acquisition scope is consistent with the initialization phase, namely all legally mandated metering nodes within the distribution area, specifically including the distribution area main meter, household electricity meters, distributed photovoltaic grid-connected meters, household energy storage bidirectional meters, and charging pile meters. The acquired data items remain consistent with the initialization phase, including instantaneous voltage, instantaneous current, active power, reactive power, energy readings, power flow direction indicators, and sampling timestamps for each metering node. The sampling frequency is maintained at 1kHz, and the data acquisition protocol is compatible with DL / T645, IEC61850, and Modbus power industry standard protocols, ensuring complete acquisition and compliant access of incremental data.

[0045] For the acquired real-time incremental sampling data, standardized preprocessing operations are performed in complete consistency with the initialization phase. Hardware-level clock synchronization is completed using the IEEE 1588PTP precision clock synchronization protocol to ensure that the synchronization error between the real-time incremental data timestamp and the reference dataset is controlled within 1ms. The sliding window 3σ criterion is used to remove spike noise and outliers caused by electromagnetic interference. No interpolation is performed to complete the data, preserving the originality and legal validity. For inter-frequency sampling nodes, the power flow inflection point matching method is used to complete the equal-interval spatiotemporal alignment, unifying the time granularity to the 10ms level to avoid additional computing power consumption, and outputting the aligned real-time incremental time series dataset.

[0046] Based on the causal node groups and assigned node topology identifiers defined in the initialization phase, core causal feature extraction and redundancy filtering operations, identical to those in the initialization phase, are performed on a single node group as an independent unit. Six types of core causal features are extracted: power flow direction time series features, active power bidirectional fluctuation amplitude and slope features, node group power balance deviation features, voltage-current phase difference time series features, current harmonic distortion rate features, and metering loop on / off status features. Redundancy filtering is performed using the maximum information coefficient (MIC), with a significance level fixed at 0.05. Redundant features without statistical significance are eliminated, ensuring that the feature dimension retained by a single node group is controlled within 10 dimensions. A feature significance identifier is bound to each effective feature, ultimately generating a real-time incremental low-dimensional causal feature sequence with timestamps and node group identifiers that perfectly matches the input dimension of the baseline causal graph. This ensures that the preprocessing and feature extraction processes are lightweight throughout, without additional computing power burden, and are fully compatible with the input requirements of subsequent real-time causal inference stages.

[0047] The lightweight causal inference based on the baseline causal graph is completed, and the real-time causal effect value is calculated, such as... Figure 6 As shown, the implementation specifically includes: using the lightweight baseline causal graph of the transformer area built during the initial deployment phase as the sole baseline model, and for the real-time incremental low-dimensional causal feature sequence collected and preprocessed during the normalized real-time analysis phase, based on the fixed causal path and node topology attribution identifier in the baseline causal graph corresponding to each causal association node group, the linear non-Gaussian acyclic model LiNGAM lightweight inference algorithm is used to perform causal inference operations. The algorithm only performs matrix multiplication operations and does not involve complex iterative operations, ensuring that the time taken for a single round of inference does not exceed 100ms, which is fully adapted to the low computing power constraints of edge nodes.

[0048] The reasoning process uses a single causal association node group as an independent computing unit. For the real-time incremental low-dimensional causal feature sequence within the node group, each causal path in the baseline causal graph is matched one by one. The causal contribution of each core causal feature to the change of the measurement value of the node group is quantified and calculated. This contribution is the real-time causal effect value.

[0049] The calculation process strictly follows the rigid constraints of electrical physics embedded in the baseline causal graph to ensure that the causal reasoning direction is consistent with the power flow direction, conforms to the power balance logic, and has no false causal loops. After the calculation is completed, the real-time causal effect value of each causal association node group is bound one by one with the classification label, node topology attribution identifier, and timestamp of the corresponding causal path to form a set of real-time causal effect values ​​with complete identification information. This ensures that the set can be directly used for subsequent two-level threshold verification and abnormal event identification, while ensuring that the entire calculation process is lightweight and the computing power consumption does not exceed 10% of the rated computing power of the edge gateway MCU.

[0050] The system uses two-level threshold verification to distinguish between normal bidirectional power flow and abnormal events, thus classifying and determining whether electricity theft or metering faults are occurring. Figure 6 As shown, the specific implementation includes: taking the set of real-time causal effect values ​​calculated in the normalized real-time analysis phase as the core basis, and performing a two-level progressive threshold verification operation based on the classification labels of each causal path in the lightweight baseline causal graph and the calibrated baseline effect threshold, accurately distinguishing between normal bidirectional power flow and abnormal events, and completing the classification and judgment of electricity theft or metering failure.

[0051] In the first-level verification operation, for each causal relationship node group, the real-time causal effect value set is compared one by one with the real-time causal effect value of all endogenous normal causal paths in the baseline causal graph of that node group and the corresponding baseline effect threshold. If all causal contributions of the fluctuation of the measurement value of that node group are 100% from the endogenous normal causal path, and the real-time causal effect value of each endogenous normal causal path is within the range of its corresponding baseline effect threshold, it is determined to be a normal bidirectional power flow fluctuation event, and the two-level verification operation is immediately terminated without triggering any abnormal alarms.

[0052] In the second-level verification operation, if the causal contribution of the fluctuation in the metering value of the node group includes an exogenous abnormal causal path, or if the real-time causal effect value of any endogenous normal causal path exceeds its corresponding baseline effect threshold range, it is judged as an abnormal event. For events judged as abnormal, based on the causal contribution decomposition results of the exogenous abnormal causal path and the endogenous normal causal path, a fixed rule for abnormal classification judgment is executed: if the causal contribution of the exogenous abnormal causal path is ≥80%, it is judged as electricity theft. If the real-time causal effect value of the endogenous normal causal path exceeds the baseline threshold, and is accompanied by abnormal voltage-current phase difference characteristics and current harmonic distortion rate characteristics, it is judged as a metering fault.

[0053] Once the determination is completed, the abnormal event classification label, the causal path tracing results corresponding to the abnormality, the time of occurrence of the abnormality and the corresponding measurement node information are bound to the real-time analysis results, providing a basis for subsequent local closed-loop handling and compliant evidence storage of abnormal events.

[0054] Figure 7The graph compares the false alarm rates of this application with those of existing technologies in a two-way power flow scenario. The coordinates cover the entire scenario of source-load-storage access conditions from 0% to 100%, and the vertical axis represents the average false alarm rate for anomaly identification. The three sets of curves show significant differences: the false alarm rate of the method of this invention remains stable in the range of 0.2%-0.8% throughout, and is still below 1% even in a 100% strong two-way power flow scenario, with no significant fluctuations; the false alarm rate of the traditional correlation model increases linearly with the access ratio, reaching a maximum of 18.6%; although the traditional non-lightweight causal model is better than the former, the highest false alarm rate still reaches 4.9%. This result fully verifies that this patent solves the core pain points of high false alarm rate and inability to adapt to new power system distribution areas of traditional solutions through rigid constraints of power physics, lightweight causal reasoning, and two-level threshold verification, while meeting the low computing power operation requirements of the edge side.

[0055] Figure 8 The graph shows a comparison of the inference time and computing power usage of this application and the prior art on the dual Y-axis. The inference time per sample of the method of this invention is only 82ms, which is far lower than the 100ms threshold set by the patent and is 61.9% lower than the optimal traditional model; the computing power usage rate is only 8.7%.

[0056] This application also provides an embodiment of an electronic device. The electronic device is manifested in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: one or more processors or processing units, memory, and buses connecting different components (including memory and processing units).

[0057] A bus refers to one or more of several bus architectures, including memory buses or memory controllers, peripheral buses, graphics acceleration ports, processors, or local buses using any of the various bus architectures. Examples of these architectures include, but are not limited to, Industry Standard Architecture (ISA) buses, Micro Channel Architecture (MCA) buses, Enhanced ISA buses, Video Electronics Standards Association (VESA) local buses, and Peripheral Component Interconnect (PCI) buses.

[0058] Electronic devices typically include a variety of computer-readable media. These media can be any available media that can be accessed by the electronic device, including volatile and non-volatile media, and removable and non-removable media.

[0059] The memory may include computer-readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. Electronic devices may further include other removable / non-removable, volatile / non-volatile computer device storage media. By way of example only, the storage system may be used to read and write non-removable, non-volatile magnetic media.

[0060] The electronic device can also communicate with one or more external devices (e.g., keyboard, pointing device, camera, etc.), may include a display, and may communicate with one or more devices that enable a user to interact with the electronic device, and / or with any device that enables the electronic device to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed via an input / output (I / O) interface. Furthermore, the electronic device can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN) and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the electronic device via a bus. The processor executes various functional applications and data processing by running programs stored in memory, such as implementing the edge computing power metering data real-time analysis method provided in the above embodiments of the present invention.

[0061] This application also discloses a real-time analysis system for edge computing power metering data, including: The data acquisition and synchronization module is used to collect raw sampling data from all metering nodes in the distribution area and complete hardware-level clock synchronization and lossless spatiotemporal alignment. The node feature processing module is used to divide causal relationship node groups according to property rights topology and extract and filter low-dimensional core causal features. The causal graph construction module is used to embed rigid constraints in electrical physics, construct a lightweight baseline causal graph, and calibrate the baseline effect thresholds for endogenous normal causal paths and exogenous abnormal causal paths. The real-time data preprocessing module is used to collect real-time incremental metering data and complete standardized preprocessing and feature extraction. The causal reasoning calculation module is used to perform lightweight causal reasoning based on the baseline causal graph and calculate the real-time causal effect value. The anomaly detection module is used to distinguish between normal bidirectional power flow and abnormal events through two-level threshold verification, and to classify and determine electricity theft or metering faults.

[0062] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time analysis of edge computing power metering data, characterized in that, Includes the following steps: Collect raw sampling data from all metering nodes in the distribution area and complete hardware-level clock synchronization and lossless spatiotemporal alignment; Divide causal relationship node groups according to property rights topology, and extract and filter low-dimensional core causal features; By embedding rigid constraints in electrical physics, a lightweight baseline causal graph is constructed, and the baseline effect thresholds for endogenous normal causal paths and exogenous abnormal causal paths are calibrated. Collect real-time incremental metering data and complete standardized preprocessing and feature extraction; Lightweight causal inference is performed based on a baseline causal graph, and real-time causal effect values ​​are calculated. By using two levels of threshold verification, normal bidirectional power flow can be distinguished from abnormal events, thus completing the classification and determination of electricity theft or metering faults.

2. The real-time analysis method for edge computing power metering data according to claim 1, characterized in that, The process of collecting raw sampling data from all metering nodes in the distribution area and achieving hardware-level clock synchronization and lossless spatiotemporal alignment includes: collecting raw sampling data from all legally mandated metering nodes within the distribution area through the distribution area edge gateway, including the distribution area main meter, household electricity meters, distributed photovoltaic grid-connected meters, household energy storage bidirectional meters, and charging pile meters; synchronizing the hardware-level clock of the distribution area edge gateway with all metering nodes; removing noise and outliers from the raw sampling data; and achieving spatiotemporal alignment for sampling nodes with different frequencies, unifying the time granularity.

3. The real-time analysis method for edge computing power metering data according to claim 1, characterized in that, The method of dividing causal association node groups according to property rights topology includes: grouping all metering nodes in the entire transformer area according to the physical topology of the transformer area and the ownership of metering nodes; grouping all metering nodes under the same property owner into the same causal association node group; ensuring that nodes in the group have power transfer causal association and property rights consistency; setting the transformer area gateway master table as a global reference node and not including it in the user-side node group; decomposing the overall causal calculation of the transformer area into multiple independent units through grouping; reducing the search for invalid causal paths and reducing the amount of computation; and assigning a unique topology identifier to each causal association node group and binding it with the corresponding data and feature sequences.

4. The real-time analysis method for edge computing power metering data according to claim 3, characterized in that, The extraction and screening of low-dimensional core causal features includes: based on the causal association node groups divided in the initialization phase, extracting core causal features with physical causal relationships using a single node group as a unit; performing redundancy screening on the extracted initial features, removing features without statistical significance; and binding the screened low-dimensional features with time series sequences and node topology attribution identifiers to generate a time series sequence of low-dimensional core causal features.

5. The real-time analysis method for edge computing power metering data according to claim 1, characterized in that, The embedded rigid constraints of power physics include: based on Kirchhoff's laws of the power system and the physical topology of the transformer area, three types of rigid hard constraints are embedded: causal directionality constraints, node group power balance constraints, and causal graph acyclic constraints. The causal directionality constraints ensure that the causal direction of power transmission is consistent with the power flow direction. The node group power balance constraints define the power transmission associations within the same owner's node group as endogenous normal causal paths that need to be retained. The causal graph acyclic constraints ensure that the causal graph conforms to the physical logic of power transmission. The three types of constraints are embedded throughout the entire causal structure learning process, and the search space is limited before the conditional independence test.

6. The real-time analysis method for edge computing power metering data according to claim 5, characterized in that, The construction of the lightweight baseline causal graph includes: performing conditional independence tests on the low-dimensional core causal features of each causal association node group based on the search space limited by rigid constraints of power physics; constructing an initial causal graph for the distribution area based on the test results; calculating the causal dependencies and path coefficients between feature variables; dividing the causal paths into endogenous normal causal paths and exogenous abnormal causal paths according to the bidirectional power flow characteristics of the distribution area and completing the calibration; performing lightweight pruning on the initial causal graph to remove redundant causal edges; calibrating the baseline causal effect threshold of each causal path based on the historical operating data of the distribution area; and generating a lightweight baseline causal graph adapted to the low computing power environment of the edge side after verification of computing power occupancy.

7. The real-time analysis method for edge computing power metering data according to claim 6, characterized in that, The benchmark effect thresholds for calibrating endogenous normal causal paths and exogenous abnormal causal paths include: based on the lightweight initial causal graph constructed during initialization, benchmark effect thresholds are calibrated for the two types of causal paths according to the historical normal operation data and abnormal evidence data of the transformer area; the thresholds of each path are calibrated separately according to the laws of power physics and the operating characteristics of the transformer area; after calibration, the thresholds are bound to the benchmark causal graph with the path classification label and topology identifier, and the accuracy of the thresholds is verified. If the requirements are not met, the thresholds are recalibrated until they are met.

8. The real-time analysis method for edge computing power metering data according to claim 7, characterized in that, The lightweight causal inference based on the benchmark causal graph and the calculation of real-time causal effect values ​​include: using the initialized lightweight benchmark causal graph of the transformer area as a benchmark, for the preprocessed real-time incremental low-dimensional causal feature sequence, using a lightweight inference algorithm, performing causal inference with a single causal association node group as a unit, following the rigid constraints of power physics, quantifying the causal contribution of each core causal feature to the change of the node group's metering value, completing the binding of real-time causal effect values ​​with relevant identifiers, and forming a set for threshold verification.

9. The real-time analysis method for edge computing power metering data according to claim 8, characterized in that, The method involves a two-level threshold verification to distinguish between normal bidirectional power flow and abnormal events, thereby classifying and determining whether electricity theft or metering faults occur. This includes: conducting a two-level progressive threshold verification based on real-time causal effect values, path labels of the baseline causal graph, and thresholds. The first level verifies whether the endogenous normal causal path effect value is within the threshold range to determine whether it is a normal bidirectional power flow. The second level verifies and identifies abnormal events, and distinguishes between electricity theft and metering faults based on the causal contribution decomposition results and abnormal electrical characteristics. The method binds the anomaly classification, path tracing, time and node information with the analysis results to provide a basis for anomaly handling and evidence preservation.

10. A system utilizing the real-time analysis method for edge computing power metering data as described in claim 1, characterized in that, include: The data acquisition and synchronization module is used to collect raw sampling data from all metering nodes in the distribution area and complete hardware-level clock synchronization and lossless spatiotemporal alignment. The node feature processing module is used to divide causal relationship node groups according to property rights topology and extract and filter low-dimensional core causal features. The causal graph construction module is used to embed rigid constraints in electrical physics, construct a lightweight baseline causal graph, and calibrate the baseline effect thresholds for endogenous normal causal paths and exogenous abnormal causal paths. The real-time data preprocessing module is used to collect real-time incremental metering data and complete standardized preprocessing and feature extraction. The causal reasoning calculation module is used to perform lightweight causal reasoning based on the baseline causal graph and calculate the real-time causal effect value. The anomaly detection module is used to distinguish between normal bidirectional power flow and abnormal events through two-level threshold verification, and to classify and determine electricity theft or metering faults.

Images