A method for monitoring electricity larceny of an electric energy metering box

Abstract

This application relates to power system security monitoring technology, specifically to a method for monitoring electricity theft in electricity metering boxes. The aim is to address the problem in existing technologies where electricity thieves use advanced algorithms such as generative adversarial networks to generate highly realistic false load curves, accurately simulating fundamental characteristics of normal electricity consumption such as voltage ripple fluctuation frequency, current phase shift, and load curve temporal distribution entropy. This invention achieves cross-box collaborative feature learning through federated learning to capture the spatiotemporal correlation of electricity theft behavior. It combines blockchain data anchoring and smart contract verification to ensure data credibility, and dynamically adjusts parameters and triggers tiered interventions based on credible data. This effectively solves the technical bottlenecks of traditional monitoring methods in dealing with intelligent disguised electricity theft and distributed collaborative electricity theft.

Description

Technical Field

[0001] This application relates to power system security monitoring technology, and more specifically, to a method for monitoring electricity theft prevention in electricity metering boxes. Background Technology

[0002] In the operation of power systems, the monitoring of electricity metering boxes to prevent electricity theft faces increasingly complex challenges. With the deep integration of electricity theft techniques with artificial intelligence and edge computing, traditional monitoring mechanisms have revealed significant shortcomings.

[0003] On the one hand, electricity thieves use advanced algorithms such as generative adversarial networks to generate highly realistic fake load curves, which can accurately simulate the basic characteristics of normal electricity use, such as voltage ripple mutation frequency, current phase shift and load curve time distribution entropy. They can even tamper with local metering data in real time through edge computing devices, making it difficult for single-box monitoring systems to distinguish between real electricity use and intelligent spoofing behavior, causing traditional methods based on historical template matching and local feature extraction to fail.

[0004] On the other hand, distributed collaborative electricity theft is becoming increasingly common. Electricity thieves operate across regions and time periods, linking multiple electricity metering boxes. This is akin to multiple users within the same area coordinating to adjust their load to distribute the total amount of stolen electricity, creating a concealed characteristic of "normal local data but abnormal overall grid." Existing isolated monitoring systems only focus on single-point data, failing to capture the spatiotemporal collaborative characteristics, load transfer characteristics, and parameter consistency characteristics across different boxes, resulting in monitoring blind spots. Furthermore, the dynamic parameter adjustment process is severely disconnected from data reliability verification. The system relies solely on local data from a single box for parameter correction. If this local data is maliciously tampered with, it will directly mislead the adjustment direction. Simultaneously, the lack of macro-level grid-side characteristic data, such as regional total load and line losses, means that parameter tolerance settings lack a global basis, fundamentally limiting the overall anti-theft mechanism's resistance to evasion and accuracy.

[0005] These problems are particularly prominent in the context of the digital transformation of the power industry. The traditional logic of single-point monitoring plus local feature matching can no longer cope with new electricity theft methods. The core contradiction lies in the superimposed effect of isolated monitoring architecture and data unreliability risk, and there is an urgent need to build an innovative solution that integrates multi-dimensional data collaboration and reliable verification. Summary of the Invention

[0006] (a) Technical problems to be solved The purpose of this application is to provide a method for monitoring electricity theft in an electricity metering box, which has the advantages of improving the accuracy and reliability of electricity theft monitoring, effectively identifying intelligent disguised electricity theft behavior and cross-box collaborative electricity theft mode, ensuring that parameter adjustment is based on reliable data, and enhancing the system's anti-evasion capability.

[0007] (II) Technical Solution This application provides a method for monitoring electricity theft prevention in electricity metering boxes, the technical solution of which is as follows: Acquire electricity consumption characteristic data from multiple electricity metering boxes, including basic characteristics and anti-spoofing characteristics; Acquire macroscopic characteristic data from the power grid side, including total regional load and line losses; Based on electricity consumption characteristic data and macroscopic characteristic data, a federated learning model is used to perform cross-box collaborative feature learning to generate collaborative anomaly features. Electricity consumption characteristic data and macroeconomic characteristic data are anchored to the blockchain, and smart contracts are used to perform two-dimensional data verification to verify the credibility of the data. Based on collaborative anomaly characteristics and data reliability, the dynamic parameters of the electricity metering box are adjusted; and Based on the adjusted dynamic parameters and collaborative anomaly characteristics, a tiered intervention mechanism is triggered.

[0008] Furthermore, this application also proposes a method for cross-box collaborative feature learning based on electricity consumption characteristic data and macroscopic characteristic data, using a federated learning model, including: Each electricity meter box is used as an edge training node in federated learning to extract local electricity consumption feature data. The model parameters of multiple edge training nodes are aggregated on the regional power grid edge server to train a cross-container collaborative feature model; and Based on the cross-enclosure collaborative feature model, a collaborative anomaly level is output, which is used to characterize the degree of anomaly in cross-enclosure collaborative electricity theft.

[0009] Furthermore, this application also proposes that the cross-container collaborative feature model learning includes at least one of spatiotemporal collaborative features, load transfer features, and parameter consistency features.

[0010] Furthermore, this application also proposes anchoring electricity consumption characteristic data and macroeconomic characteristic data to the blockchain, including: The electricity consumption characteristic data of each electricity metering box is used to generate micro data blocks, and the macro characteristic data of the power grid side is used to generate macro data blocks. Calculate the hash values ​​for micro-level and macro-level data blocks separately, and associate them with adjacent blocks in the blockchain; and Two-dimensional data verification is performed through smart contracts, including calculating the sum of individual bias rates and feature correlation bias.

[0011] Furthermore, this application also proposes adjusting the dynamic parameters of the electricity metering box based on collaborative anomaly characteristics and data reliability, including: Based on data credibility, reliable data is selected as the benchmark for parameter adjustment. Based on the collaborative anomaly level, local parameters are weighted and corrected; and The parameter tolerance is dynamically adjusted based on the total individual deviation rate.

[0012] Furthermore, this application proposes that the triggering mechanism for tiered intervention includes: When only electricity consumption data is abnormal and the data credibility verification is passed, a primary intervention is triggered. When electricity consumption characteristic data is abnormal, data credibility verification passes, and the collaborative anomaly level is below the threshold, intermediate intervention is triggered; and Advanced intervention is triggered when electricity consumption characteristic data is abnormal, data credibility verification fails, and the level of collaborative anomaly exceeds the threshold.

[0013] Furthermore, this application also proposes and includes lightweight optimizations to the federated learning model, including: A lightweight model is generated through model distillation and then deployed to an energy metering box; and The gradients of the model parameters are compressed and quantized to reduce the amount of data transmitted.

[0014] Furthermore, this application also proposes that obtaining electricity consumption characteristic data from multiple electricity metering boxes includes: At least one of the following is collected in real time from the power metering box: voltage ripple mutation frequency, current phase shift, and load curve time distribution entropy, as the basic feature and anti-spoofing feature.

[0015] Furthermore, this application also proposes, including: At least one processor; and The memory stores instructions that, when executed by at least one processor, cause the at least one processor to perform an anti-theft monitoring method for an electricity metering box that runs the above-described method.

[0016] Furthermore, this application also proposes a method for monitoring electricity theft in an electricity metering box that stores executable instructions, which, when executed, cause the machine to perform the above-described method.

[0017] (III) Beneficial Effects Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates federated learning to achieve cross-container collaborative feature learning to capture the spatiotemporal correlation of electricity theft. It combines blockchain data anchoring and smart contract verification to ensure data credibility, and dynamically adjusts parameters and triggers tiered interventions based on credible data. This effectively solves the technical bottlenecks of traditional monitoring methods in dealing with intelligent disguised electricity theft and distributed collaborative electricity theft. It has the advantages of improving the accuracy and reliability of anti-electricity theft monitoring, effectively identifying intelligent disguised electricity theft and cross-container collaborative electricity theft modes, ensuring that parameter adjustments are based on credible data, and enhancing the system's anti-evasion capabilities. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0019] Figure 1 This is a schematic diagram of the logic structure of the anti-theft monitoring method for electricity metering boxes. Detailed Implementation

[0020] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0022] like Figure 1 As shown in the embodiment of this application, a method for monitoring electricity theft in an electricity metering box is proposed, including: S100. Obtain electricity consumption characteristic data from multiple electricity metering boxes. The electricity consumption characteristic data includes basic characteristics and anti-spoofing characteristics. S200: Obtain macroscopic characteristic data from the power grid side, including total regional load and line losses; S300: Based on electricity consumption characteristic data and macroscopic characteristic data, a federated learning model is used to perform cross-box collaborative feature learning to generate collaborative anomaly features. S400 anchors electricity consumption characteristic data and macroeconomic characteristic data to the blockchain, and performs two-dimensional data verification through smart contracts to verify data credibility. S500 adjusts the dynamic parameters of the electricity metering box based on collaborative anomaly characteristics and data reliability; and triggers a graded intervention mechanism based on the adjusted dynamic parameters and collaborative anomaly characteristics.

[0023] In practical applications, electricity consumption characteristic data can be understood as multi-dimensional information related to electricity consumption behavior extracted from the electricity metering box. Basic features describe typical electricity consumption patterns, such as the amplitude variation patterns of voltage and current, or the overall trend of the load curve. Anti-spoofing features are specifically used to identify potential spoofing behavior, such as detecting abnormal fluctuations by analyzing the frequency of voltage ripple spikes or the time distribution entropy of the load curve. Furthermore, electricity consumption characteristic data can be acquired in real-time by sensors installed in the electricity metering box, or extracted after preprocessing historical data using edge computing devices.

[0024] Macro-level characteristic data refers to information obtained from the power grid side that reflects the overall electricity consumption status. Regional total load characterizes the total electricity consumption of all users within a specific region, while line losses reflect energy losses during transmission. Specifically, macro-level characteristic data can be obtained directly through the data interface of the power grid dispatch center, or by aggregating and calculating the local power grid status through a distributed sensor network.

[0025] Federated learning models can be understood as a distributed machine learning framework, whose core lies in achieving multi-node collaborative modeling while protecting data privacy. As a preferred implementation, the training process of a federated learning model can involve deploying a lightweight local model in each electricity metering box and uploading the model parameters to a regional server for aggregation, thereby generating a global model. Furthermore, the learning objective of the federated learning model can be set to identify cross-box collaborative electricity theft behavior, for example, by analyzing load transfer patterns or parameter consistency deviations between multiple boxes to generate collaborative anomaly features.

[0026] The process of anchoring data to the blockchain can be understood as recording electricity consumption characteristic data and macroeconomic characteristic data in the form of blocks in a distributed ledger to ensure the immutability of the data. Specifically, data anchoring can be achieved by generating independent micro-data blocks from the electricity consumption characteristic data of each electricity metering box and generating macro-data blocks from the macroeconomic characteristic data of the power grid side. Then, hash values ​​are calculated for each block and associated with adjacent blocks in the blockchain. Smart contracts are used to perform two-dimensional data verification, such as verifying the consistency and credibility of the data by comparing the individual total deviation rate or the deviation of characteristic correlations.

[0027] Dynamic parameter adjustment can be understood as the process of correcting the operating parameters of an electricity metering box based on the characteristics of coordinated anomalies and the reliability of the data. For example, when the data reliability is high, it can be selected as the benchmark for parameter adjustment; when coordinated anomalies indicate a risk of cross-box electricity theft, local parameters can be weighted and corrected. Furthermore, dynamic parameter adjustment can also incorporate other external constraints, such as ambient temperature or user electricity consumption habits, to improve the accuracy of the adjustment.

[0028] A tiered intervention mechanism can be understood as a process of adopting differentiated response strategies based on the type and severity of the anomaly. For example, when only a single electricity meter box is detected to have abnormal electricity consumption data and the data is highly reliable, primary intervention measures can be triggered; when cross-meter box coordinated electricity theft is detected and the level of coordination anomaly is high, higher-level intervention measures can be triggered.

[0029] The innovation of this application lies in the integration of multi-container micro-data and power grid macro-data to construct a cross-container collaborative analysis and reliable verification mechanism, effectively addressing the problems of intelligent spoofing and electricity theft evasion, as well as the blind spots of isolated single-point monitoring. Specifically, anti-spoofing features in electricity consumption characteristic data can identify AI-generated false load curves or data tampering behavior, overcoming the limitations of single-point monitoring in identifying intelligent spoofing electricity theft; macro-feature data reveals abnormal behaviors such as cross-container load transfer through global indicators such as regional total load and line loss, eliminating the blind spots of isolated single-point monitoring. The federated learning model realizes distributed aggregation and analysis of multi-container data, generating collaborative anomaly features to quantify global electricity theft risk; blockchain technology executes two-dimensional data verification through smart contracts, ensuring the immutability and consistency verification of data.

[0030] The dynamic parameter adjustment and tiered intervention mechanism further improve the system's response accuracy and targeting, forming a complete closed loop from data collection to precise intervention.

[0031] The working principle of this application embodiment is as follows: By acquiring electricity consumption characteristic data from multiple electricity metering boxes, basic features are used to capture regular electricity consumption patterns, while anti-spoofing features are specifically used to identify AI-generated false load curves or data tampering. Furthermore, by combining macro-level characteristic data from the power grid side, including global indicators such as regional total load and line losses, collaborative electricity theft behaviors such as cross-box load transfer and total load sharing are revealed. This forms a dual-dimensional data input, encompassing both micro and macro perspectives, providing comprehensive data support for subsequent analysis.

[0032] Based on the aforementioned electricity consumption characteristic data and macroscopic characteristic data, a federated learning model is used for cross-container collaborative feature learning. Specifically, the federated learning model achieves distributed aggregation and analysis of multi-container data without transmitting raw electricity consumption data, generating collaborative anomaly features to quantify the global electricity theft risk. This distributed modeling approach not only protects user privacy but also avoids the latency issues that may arise from centralized modeling, thus effectively addressing the core pain points of distributed collaborative electricity theft scenarios.

[0033] Furthermore, electricity consumption characteristic data and macroscopic characteristic data are anchored to the blockchain, and dual-dimensional data verification is performed through smart contracts. Specifically, the individual summation deviation rate is used to verify the consistency between the sum of loads across multiple containers and the total regional load, while the characteristic correlation deviation is used to verify the matching degree between the microscopic characteristics of a single container and the global collaborative characteristics. In this way, the immutability and consistency verification of the data are ensured, fundamentally addressing the problem of lost credibility caused by data tampering.

[0034] The dynamic parameters of the electricity metering box are adjusted based on collaborative anomaly characteristics and data credibility. Specifically, data credibility is used to screen reliable data sources and avoid misleading the system with locally distorted data; collaborative anomaly levels are used to quantify global risks, ensuring that parameter sensitivity matches the severity of electricity theft. This dual-dimensional parameter correction mechanism significantly improves the accuracy and relevance of parameter responses.

[0035] A tiered intervention mechanism is triggered based on adjusted dynamic parameters and collaborative anomaly characteristics. For example, a primary intervention is triggered when only electricity consumption characteristic data is abnormal and data credibility verification passes; a secondary intervention is triggered when electricity consumption characteristic data is abnormal, data credibility verification passes, and the collaborative anomaly level is below the threshold; and a high-level intervention is triggered when electricity consumption characteristic data is abnormal, data credibility verification fails, and the collaborative anomaly level is above the threshold. This forms a complete closed loop from data collection, feature extraction, collaborative analysis, credibility verification, parameter adjustment to precise intervention.

[0036] In summary, by working closely together and integrating microscopic data from multiple enclosures with macroscopic data from the power grid, a cross-enclosure collaborative analysis and reliable verification mechanism was constructed, effectively solving the problems of intelligent camouflage for electricity theft and isolated monitoring blind spots.

[0037] This application further proposes a method for cross-container collaborative feature learning based on electricity consumption characteristic data and macroscopic characteristic data, using a federated learning model. This method includes: using each electricity metering box as an edge training node for federated learning to extract local electricity consumption characteristic data; aggregating the model parameters of multiple edge training nodes on a regional power grid edge server to train a cross-container collaborative feature model; and outputting a collaborative anomaly level based on the cross-container collaborative feature model, whereby the collaborative anomaly level is used to characterize the degree of anomaly in cross-container collaborative electricity theft.

[0038] Specifically, edge training nodes refer to computing units that undertake local data processing tasks in a distributed federated learning architecture. They can be implemented using embedded processors, edge gateways, or dedicated AI chips, with the aim of ensuring the privacy and real-time nature of local electricity consumption characteristic data.

[0039] In practical applications, the cross-box collaborative feature model refers to a machine learning model that captures cross-regional coordinated electricity theft behavior generated by multi-node parameter aggregation. It can be constructed using algorithm frameworks such as deep neural networks, random forests, or support vector machines, with the aim of identifying electricity theft patterns that are normal locally but abnormal globally.

[0040] Among them, the collaborative anomaly level refers to an indicator system that quantifies the severity of collaborative electricity theft across enclosures. It can be expressed in the form of probability values, scoring systems, or graded labels, with the aim of providing a basis for decision-making for subsequent intervention measures.

[0041] In detail, this solution achieves cross-cabinet collaborative electricity theft monitoring by constructing a three-tier architecture of "local processing at edge nodes - parameter aggregation at regional servers - quantitative output of anomaly levels". First, each electricity metering box acts as an independent edge training node, completing the extraction and preliminary processing of electricity consumption feature data locally. This process not only avoids the centralized transmission of raw electricity consumption data, but also completes the preprocessing of anti-spoofing features at the edge, making the aggregated model parameters more reliable.

[0042] Building upon this foundation, the regional power grid edge server aggregates model parameters from multiple edge training nodes. Through aggregation logic specifically optimized for cross-container electricity theft characteristics, it enhances the ability to capture "locally normal, globally abnormal" patterns. Finally, based on the trained cross-container collaborative feature model, it outputs a collaborative anomaly level. This quantitative indicator directly correlates with the collaborative nature and severity of electricity theft behavior, providing a precise basis for subsequent graded intervention.

[0043] Considering the close coordination between the above scheme and the acquisition of electricity consumption characteristic data, macro-characteristic data, and triggering hierarchical intervention mechanisms, the scheme effectively solves the problems of data silos and anomaly quantification in cross-regional joint electricity theft monitoring by organically combining local feature extraction, parameter aggregation, and anomaly quantification, thus forming a complete technical chain from data processing to decision output.

[0044] This application further proposes a cross-container collaborative feature model learning that includes at least one of spatiotemporal collaborative features, load transfer features, and parameter consistency features.

[0045] Specifically, spatiotemporal coordination characteristics refer to the correlation between multiple electricity metering boxes in terms of time series and spatial distribution. This can be achieved by calculating the overlap of load fluctuation timestamps and the correlation of spatial locations, with the aim of capturing abnormal fluctuation patterns when electricity thieves coordinate to adjust electricity load across regions. Load transfer characteristics refer to the dynamic load transfer relationship between electricity metering boxes. This can be achieved by calculating the complementarity of load changes between boxes in real time and the time delay matching degree of load transfer, with the aim of identifying unnatural load redistribution relationships.

[0046] The parameter consistency feature is based on the physical laws of the power grid, requiring that the voltage, current and other parameters of the metering boxes in the same area maintain internal consistency. This can be achieved by comparing the similarity of parameters between the target box and other boxes in the same area, and the deviation of parameters from the regional power grid benchmark parameters. The purpose is to detect traces of human intervention.

[0047] Specifically, when the scheme performs cross-box collaborative feature learning through a federated learning model, it first extracts local electricity consumption feature data from each electricity metering box as edge training nodes, and then aggregates the model parameters of multiple edge training nodes on the regional power grid edge server to train the cross-box collaborative feature model.

[0048] In this process, spatiotemporal coordination features can identify synchronous abnormal fluctuations in the load of multiple containers within a specific time period, load transfer features can detect unnatural load redistribution relationships between containers, and parameter consistency features are used to verify the authenticity of the data. These features work together to construct a triple verification mechanism of behavior-total-data, thereby effectively solving the fundamental deficiency of isolated monitoring in dealing with cross-container coordinated attacks.

[0049] For example, in practical applications, when multiple adjacent merchants simultaneously reduce their load by 10% at 8 PM, although the load fluctuation of a single container does not exceed the normal threshold, the spatiotemporal coordination feature can detect anomalies by calculating the overlap of load fluctuation timestamps and the spatial correlation. Simultaneously, if the load of a container suddenly drops by 20kW while the loads of other containers simultaneously increase slightly, the load transfer feature will identify this anomaly by calculating the complementarity of load changes and the matching degree of time delay.

[0050] Furthermore, when the parameters of a certain enclosure deviate significantly from those of other enclosures in the same area, the parameter consistency feature can promptly detect traces of human intervention. This multi-feature combination monitoring method significantly improves the ability to identify distributed, coordinated electricity theft.

[0051] This application further proposes a specific implementation method for anchoring electricity consumption characteristic data and macroscopic characteristic data to the blockchain, including: generating micro data blocks from the electricity consumption characteristic data of each electricity metering box and generating macro data blocks from the macroscopic characteristic data of the power grid side; calculating hash values ​​for the micro data blocks and macro data blocks respectively and associating them with adjacent blocks of the blockchain; and performing two-dimensional data verification through smart contracts, including calculating the individual total deviation rate and feature correlation deviation.

[0052] In practical applications, a micro data block refers to a set of electricity consumption characteristic data extracted from a single electricity metering box. It can be implemented in the form of a distributed database or a structured file, with the aim of focusing on the refined electricity consumption behavior of a single box.

[0053] Macro-level data blocks refer to data sets reflecting the overall operational status of a regional power grid. These can be implemented through centralized data acquisition systems or cloud storage platforms, aiming to provide a reference benchmark for global electricity consumption patterns. Hash values ​​are fixed-length data digests generated based on specific algorithms, such as SHA-256 or MD5, designed to ensure data integrity and immutability. Smart contracts are automated execution scripts deployed on the blockchain. They can be written in Solidity and run on blockchain platforms that support smart contracts, such as Ethereum, aiming to automate and ensure the trustworthiness of two-dimensional data verification.

[0054] Specifically, this scheme achieves a structured separation of data sources and dimensions by generating micro data blocks from electricity consumption characteristic data and macro data blocks from macro characteristic data, providing an independent and comparable data foundation for subsequent deviation analysis.

[0055] Building upon this foundation, by calculating hash values ​​for both micro and macro data blocks and associating these hash values ​​with adjacent blocks in the blockchain, the chain structure of the blockchain ensures the consistency of data over time. This means that any tampering requires reconstructing the hash values ​​of all subsequent blocks, significantly increasing the cost of data tampering. Simultaneously, through smart contracts executing dual-dimensional data verification, combining the calculation of individual total deviation rate and feature correlation deviation, a verification closed loop integrating total consistency and feature logic is constructed. This can accurately detect load loss caused by electricity theft or identify abnormal patterns of AI masquerading as electricity theft.

[0056] The above solution not only addresses the lack of a specific path for anchoring micro and macro data, but also completely eliminates monitoring blind spots related to single-point data tampering and cross-dimensional data inconsistencies through a comprehensive mechanism of "structured data block generation - chained hash association - dual-dimensional smart contract verification." For example, when an electricity thief tampers with the micro data of a certain container, the chained hash association will expose hash breaks; when multiple containers collaboratively tamper with the micro data to make the total amount appear normal, feature correlation deviations will reveal logical contradictions between micro and macro features, thereby effectively improving the comprehensiveness and robustness of data credibility verification.

[0057] This application further proposes a method for adjusting the dynamic parameters of an energy metering box based on cooperative anomaly characteristics and data reliability, including: selecting reliable data as the parameter adjustment benchmark based on data reliability; performing weighted correction on local parameters based on cooperative anomaly levels; and dynamically adjusting parameter tolerance based on individual total deviation rates.

[0058] Specifically, data credibility refers to the data reliability assessment result verified by blockchain smart contracts. It can be achieved by methods such as hash value verification and micro and macro feature logical consistency analysis. The purpose is to ensure that the starting data for parameter adjustment has not been tampered with or forged.

[0059] The collaborative anomaly level can be understood as a quantitative indicator of the degree of anomaly in cross-container collaborative electricity theft. It can be characterized by anomaly scores or anomaly probability values ​​output by a federated learning model, with the aim of dynamically responding to multi-container coordinated electricity theft patterns. The individual total deviation rate refers to the ratio of the deviation between the sum of the loads of all containers in the area and the total load of the area on the grid side. It can be calculated by using the ratio of the real-time load sum to the total regional load, with the aim of reflecting the consistency of global data and constraining the parameter adjustment range.

[0060] In detail, this solution constructs a complete dynamic parameter adjustment mechanism by organically combining data credibility, collaborative anomaly level, and individual total deviation rate. First, based on data credibility, it filters out credible data that has passed the dual-dimensional verification of blockchain, such as data with unaltered hashes and consistent micro- and macro-level characteristics, to avoid parameter adjustment inaccuracies caused by AI tampering with local data.

[0061] Secondly, by using the level of coordinated anomaly as a weighting factor for local parameters, when the level of coordinated anomaly is high, the system automatically increases the sensitivity of local parameters to quickly detect subtle anomalies; when the level of coordinated anomaly is low, the parameter sensitivity returns to a normal level, thereby accurately matching the overall risk of electricity theft. Finally, the parameter tolerance is dynamically adjusted through the individual total deviation rate. When the deviation rate is large, the parameter tolerance is tightened, and priority is given to adjusting the box parameters in areas with high deviation rates, forming a closed loop of linkage between micro-parameters and macro-deviations.

[0062] Based on this, the aforementioned dynamic parameter adjustment mechanism is closely integrated with electricity consumption characteristic data collection, macro-level characteristic data analysis, and cross-container collaborative characteristic learning of the federated learning model. For example, when electricity consumption characteristic data is found to be untrustworthy after blockchain verification, the system automatically removes the data and uses trusted data from adjacent containers as the adjustment benchmark. At the same time, the collaborative anomaly level output by the federated learning model is directly used for weighted correction of local parameters, enabling parameter adjustments to dynamically adapt to changes in the risk of cross-container collaborative electricity theft.

[0063] The above technical solutions significantly improve the anti-evasion capabilities of the anti-electricity theft system and solve the core problems of data reliability being disconnected from parameter adjustment and lack of macro-constraints.

[0064] This application further proposes a tiered intervention mechanism to respond to abnormal electricity theft behavior, including: triggering primary intervention when only electricity consumption characteristic data is abnormal and the data credibility verification is passed; triggering intermediate intervention when electricity consumption characteristic data is abnormal, the data credibility verification is passed, and the level of collaborative abnormality is below a threshold; and triggering advanced intervention when electricity consumption characteristic data is abnormal, the data credibility verification is not passed, and the level of collaborative abnormality is above a threshold.

[0065] Specifically, electricity consumption characteristic data refers to basic and anti-spoofing characteristics such as voltage ripple mutation frequency, current phase shift, and load curve time distribution entropy collected in real time from the electricity metering box. These characteristics can be extracted and analyzed through smart sensors or edge computing devices.

[0066] Data credibility refers to an indicator generated based on the dual-dimensional verification results of blockchain. Specifically, it can be measured using the sum of individual deviation rates and feature correlation deviations, with the aim of ensuring the authenticity and reliability of the data. Collaborative anomaly level refers to the degree of anomaly in cross-measuring box collaborative electricity theft output by a federated learning model. It can be modeled using at least one of spatiotemporal collaborative features, load transfer features, and parameter consistency features, aiming to reflect the load pattern correlation between multiple metering boxes.

[0067] In detail, this solution constructs a three-dimensional logical judgment framework of "abnormal electricity consumption characteristics - data credibility - collaborative anomaly level" to achieve differentiated responses to different types of electricity theft. When only the electricity consumption characteristic data is abnormal and the data credibility verification passes, it indicates that there is a deviation in the local metering data but it has not been maliciously tampered with. This may be due to false alarms from equipment or minor electricity theft. At this time, a primary intervention is triggered, such as electromagnetic locking or anomaly marking, which can avoid excessive interference with normal electricity use and provide maintenance personnel with accurate clues for investigation. When the electricity consumption characteristic data is abnormal, the data credibility verification passes, and the collaborative anomaly level is below the threshold, it indicates that the anomaly is limited to a single enclosure and has no cross-regional linkage characteristics. It can be judged as a clear single-point electricity theft. At this time, a medium-level intervention is triggered, such as disconnecting the metering circuit, which can accurately block the current electricity theft without expanding the impact on the system.

[0068] When electricity consumption data is abnormal, data credibility verification fails, and the level of collaborative anomaly exceeds a threshold, it indicates data tampering or AI spoofing, accompanied by cross-enclosure collaborative behavior. At this point, advanced intervention is triggered, such as location alarms or area-linked protection, which can not only quickly stop the current electricity theft but also prevent its spread. This multi-dimensional dynamic grading mechanism, by using data credibility verification as a basic filtering layer and collaborative anomaly level as a risk amplifier, significantly improves the accuracy of intervention decisions and solves the misjudgment problems caused by traditional methods that rely on isolated local features or ignore macro-constraints.

[0069] In summary, the above technical solution effectively solves the problems of misjudgment of risk level and mismatch of intervention intensity in graded intervention by introducing a combined verification framework of data credibility and collaborative anomaly level, thereby realizing a refined graded response to electricity theft and significantly improving the adaptability and reliability of the anti-electricity theft system.

[0070] This application further proposes a lightweight optimization for federated learning models, including: generating a lightweight model through model distillation and deploying the lightweight model to an electricity metering box; and compressing and quantizing the gradients of model parameters to reduce the amount of data transmission.

[0071] Specifically, model distillation refers to a technique that uses the knowledge transfer capabilities of a large teacher model to train a small student model. Its purpose is to significantly reduce model complexity while preserving key feature extraction capabilities. In practical applications, model distillation can be achieved through knowledge distillation, network pruning, or parameter sharing, thus adapting to the limited computing power and storage space of the electricity metering box. Gradient compression and quantization refer to techniques for sparsifying model parameter gradients and reducing numerical precision. Their purpose is to reduce communication traffic between nodes and the server during federated learning. In practical applications, gradient compression can be achieved through Top-K sparsity, randomized sparsity, or numerical quantization, thereby alleviating network bandwidth pressure.

[0072] In detail, this solution generates a lightweight model through model distillation and deploys it to the electricity metering box. Based on a high-precision teacher model trained on a regional power grid edge server, the ability to identify electricity theft features is transferred to the student model. While reducing the number of model parameters by 60%-70% and the computational load by more than 50%, the accuracy of the student model in identifying key electricity theft features only decreases by 3%-5%. This design allows the lightweight model to perfectly adapt to the limited computing power and storage space of the embedded device in the electricity metering box, avoiding local inference latency caused by an excessively large model, and meeting the core requirement of real-time monitoring of electricity theft. Simultaneously, by compressing and quantizing the model parameter gradients, priority is given to retaining gradient parameters strongly correlated with electricity theft features. Using Top-K sparsity combined with 8-bit integer quantization, the gradient data volume is reduced by more than 80% while ensuring that the parameter aggregation stage can still accurately capture cross-box collaborative anomalies. This optimization not only alleviates the bandwidth pressure on the power communication network but also ensures the parameter aggregation efficiency of federated learning, adapting to the monitoring needs of distributed collaborative electricity theft. Building upon this foundation, model distillation reduces the computational burden on edge nodes at the model structure level, enabling edge devices to efficiently perform local feature extraction and gradient calculation. Gradient compression reduces communication between nodes and servers at the data transmission level, allowing parameter aggregation to be completed quickly. The combination of these two approaches achieves dual optimization of computation and communication, not only enabling federated learning to be deployed in low-computing-power electricity metering boxes but also ensuring the real-time performance and accuracy of collaborative feature learning, thus completely resolving the core bottleneck of deploying federated learning at the edge of electricity metering boxes.

[0073] This application further proposes to obtain electricity consumption characteristic data from multiple electricity metering boxes, including: real-time acquisition of at least one of voltage ripple mutation frequency, current phase shift, and load curve time distribution entropy from the electricity metering boxes as basic characteristics and anti-spoofing characteristics.

[0074] Specifically, voltage ripple surge frequency refers to the physical nature of instantaneous voltage fluctuations in the power grid. It can be achieved by monitoring the intervals, durations, and amplitude distribution of voltage surges during load start-up and shutdown, aiming to capture the unsimulable dynamic changes of real loads. Current phase shift can be understood as an electrical law directly related to load impedance characteristics. It can be detected by detecting current lag caused by motor loads or phase lead caused by capacitor compensation devices, aiming to reflect the strong correlation between load type and operating state. In practical applications, load curve time distribution entropy is a statistical indicator that measures the uncertainty of electricity consumption sequences. For example, entropy values ​​can be used to quantify the natural fluctuation patterns of residential electricity consumption during morning and evening peak hours, aiming to distinguish the pseudo-randomness of AI-generated data from the natural randomness of real electricity consumption.

[0075] In detail, this solution constructs a multi-dimensional anti-spoofing barrier by collecting the aforementioned three features in real time. First, the frequency of voltage ripple fluctuations depends on the real-time dynamics of the physical load, effectively identifying the correlation of fluctuations that are difficult to reproduce in AI-generated fake data. Second, current phase shifts are limited by the electrical characteristics of the load; by comparing the phase distribution of other enclosures in the same area, spoofed data can be quickly identified. Finally, the temporal distribution entropy of the load curve is rooted in the randomness of natural electricity consumption; by calculating the entropy value and comparing it with a benchmark range, the regularity defects of AI-generated data can be accurately identified. These three factors form a cross-validation mechanism from the three dimensions of physical process, electrical regularity, and statistical randomness, significantly increasing the difficulty of spoofing. In addition, these features not only meet the needs of routine monitoring but also provide a highly reliable input source for subsequent federated learning to mine cross-enclosure collaborative anomaly features and for blockchain verification, thus solving the core problem of unreliable features in intelligent spoofing electricity theft scenarios. Example

[0076] In another embodiment, this application also discloses a computing device, including: at least one processor; and a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the aforementioned method for monitoring electricity theft from an electricity metering box.

[0077] The core innovation of this embodiment lies in effectively solving the problems of intelligent camouflage and evasion of electricity theft in electricity metering box monitoring by combining at least one processor and memory in a collaborative manner, and introducing federated learning models and blockchain technology as key supports. Specifically, at least one processor is responsible for processing electricity consumption characteristic data and macroscopic characteristic data in real time, ensuring the automation and continuity of the monitoring process; the memory is used to store the complete instruction code of the monitoring method, ensuring the stability of logical calls. On this basis, the federated learning model realizes cross-box collaborative feature learning, generating collaborative abnormal features to quantify the global electricity theft risk, while blockchain technology executes two-dimensional data verification through smart contracts, ensuring the immutability and consistency verification of the data. The organic combination of these technologies significantly improves the system's responsiveness and anti-evasion capabilities.

[0078] This computing device provides the execution foundation for anti-electricity theft monitoring methods through its hardware architecture, ensuring the reliability and real-time performance of the monitoring process. This effectively addresses the challenges of intelligent camouflage and cross-cabinet collaborative evasion in electricity theft. At least one processor is responsible for executing the computational tasks of the monitoring method, processing real-time data streams, and generating monitoring results, ensuring the timeliness and accuracy of data processing. The memory stores the instruction code of the monitoring method, ensuring the integrity and callability of the method logic and preventing interruptions or distortions during data execution. The instruction execution logic automatically triggers the execution of the monitoring method when called by the processor, achieving automation and continuity of the monitoring process. These features work together, with the processor and memory cooperating to enable the monitoring method to be implemented stably and efficiently, capturing abnormal electricity consumption characteristics in real time, overcoming monitoring blind spots caused by isolated single-point data and local data tampering, and ultimately improving the overall response capability of the anti-electricity theft system.

[0079] This application proposes a method for monitoring electricity theft prevention in electricity metering boxes, comprising: acquiring electricity consumption characteristic data of multiple electricity metering boxes, including basic characteristics and anti-spoofing characteristics; acquiring macroscopic characteristic data from the power grid side, including regional total load and line loss; based on the electricity consumption characteristic data and macroscopic characteristic data, using a federated learning model to perform cross-box collaborative feature learning to generate collaborative anomaly features; anchoring the electricity consumption characteristic data and macroscopic characteristic data to a blockchain, and performing dual-dimensional data verification through smart contracts to verify data credibility; adjusting the dynamic parameters of the electricity metering boxes based on the collaborative anomaly features and data credibility; and triggering a tiered intervention mechanism based on the adjusted dynamic parameters and collaborative anomaly features.

[0080] In practical applications, electricity consumption characteristic data can be understood as multi-dimensional information related to electricity consumption behavior extracted from the electricity metering box. Basic features describe typical electricity consumption patterns, such as the amplitude variation patterns of voltage and current, or the overall trend of the load curve. Anti-spoofing features are specifically used to identify potential spoofing behavior, such as detecting abnormal fluctuations by analyzing the frequency of voltage ripple spikes or the time distribution entropy of the load curve. Furthermore, electricity consumption characteristic data can be acquired in real-time by sensors installed in the electricity metering box, or extracted after preprocessing historical data using edge computing devices.

[0081] Macro-level characteristic data refers to information obtained from the power grid side that reflects the overall electricity consumption status. Regional total load characterizes the total electricity consumption of all users within a specific region, while line losses reflect energy losses during transmission. Specifically, macro-level characteristic data can be obtained directly through the data interface of the power grid dispatch center, or by aggregating and calculating the local power grid status through a distributed sensor network.

[0082] Federated learning models can be understood as a distributed machine learning framework, whose core lies in achieving multi-node collaborative modeling while protecting data privacy. As a preferred implementation, the training process of a federated learning model can involve deploying a lightweight local model in each electricity metering box and uploading the model parameters to a regional server for aggregation, thereby generating a global model. Furthermore, the learning objective of the federated learning model can be set to identify cross-box collaborative electricity theft behavior, for example, by analyzing load transfer patterns or parameter consistency deviations between multiple boxes to generate collaborative anomaly features.

[0083] The process of anchoring data to the blockchain can be understood as recording electricity consumption characteristic data and macroeconomic characteristic data in the form of blocks in a distributed ledger to ensure the immutability of the data. Specifically, data anchoring can be achieved by generating independent micro-data blocks from the electricity consumption characteristic data of each electricity metering box and generating macro-data blocks from the macroeconomic characteristic data of the power grid side. Then, hash values ​​are calculated for each block and associated with adjacent blocks in the blockchain. Smart contracts are used to perform two-dimensional data verification, such as verifying the consistency and credibility of the data by comparing the individual total deviation rate or the deviation of characteristic correlations.

[0084] Dynamic parameter adjustment can be understood as the process of correcting the operating parameters of an electricity metering box based on the characteristics of coordinated anomalies and the reliability of the data. For example, when the data reliability is high, it can be selected as the benchmark for parameter adjustment; when coordinated anomalies indicate a risk of cross-box electricity theft, local parameters can be weighted and corrected. Furthermore, dynamic parameter adjustment can also incorporate other external constraints, such as ambient temperature or user electricity consumption habits, to improve the accuracy of the adjustment.

[0085] A tiered intervention mechanism can be understood as a process of adopting differentiated response strategies based on the type and severity of the anomaly. For example, when only a single electricity meter box is detected to have abnormal electricity consumption data and the data is highly reliable, primary intervention measures can be triggered; when cross-meter box coordinated electricity theft is detected and the level of coordination anomaly is high, higher-level intervention measures can be triggered.

[0086] The innovation of this application lies in the integration of multi-container micro-data and power grid macro-data to construct a cross-container collaborative analysis and reliable verification mechanism, effectively addressing the problems of intelligent camouflage-based electricity theft evasion and blind spots in isolated single-point monitoring. Specifically, anti-camouflage features in electricity consumption characteristic data can identify AI-generated false load curves or data tampering behavior, overcoming the limitations of single-point monitoring in identifying intelligent camouflage-based electricity theft; macro-feature data reveals abnormal behaviors such as cross-container load transfer through global indicators such as regional total load and line loss, eliminating blind spots in isolated single-point monitoring. The federated learning model realizes distributed aggregation and analysis of multi-container data, generating collaborative anomaly features to quantify global electricity theft risk; blockchain technology executes two-dimensional data verification through smart contracts, ensuring data immutability and consistency verification. Dynamic parameter adjustment and hierarchical intervention mechanisms further improve the system's response accuracy and targeting, forming a complete closed loop from data collection to precise intervention.

[0087] The working principle of this application embodiment is as follows: First, by acquiring electricity consumption characteristic data from multiple electricity metering boxes, basic features are used to capture regular electricity consumption patterns, while anti-spoofing features are specifically used to identify AI-generated false load curves or data tampering behavior. Furthermore, by combining macroscopic characteristic data from the power grid side, including global indicators such as regional total load and line losses, collaborative electricity theft behaviors such as cross-box load transfer and total load sharing are revealed. Thus, a dual-dimensional data input, encompassing both micro and macro dimensions, is formed, providing comprehensive data support for subsequent analysis.

[0088] Secondly, based on the aforementioned electricity consumption characteristic data and macroscopic characteristic data, a federated learning model is used for cross-container collaborative feature learning. Specifically, the federated learning model achieves distributed aggregation and analysis of multi-container data without transmitting raw electricity consumption data, generating collaborative anomaly features to quantify the global electricity theft risk. This distributed modeling approach not only protects user privacy but also avoids the latency issues that may arise from centralized modeling, thus effectively addressing the core pain points of distributed collaborative electricity theft scenarios.

[0089] Furthermore, electricity consumption characteristic data and macroscopic characteristic data are anchored to the blockchain, and dual-dimensional data verification is performed through smart contracts. Specifically, the individual summation deviation rate is used to verify the consistency between the sum of loads across multiple containers and the total regional load, while the characteristic correlation deviation is used to verify the matching degree between the microscopic characteristics of a single container and the global collaborative characteristics. In this way, the immutability and consistency verification of the data are ensured, fundamentally addressing the problem of lost credibility caused by data tampering.

[0090] Subsequently, the dynamic parameters of the electricity metering box were adjusted based on collaborative anomaly characteristics and data credibility. Specifically, data credibility was used to screen reliable data sources and avoid misleading the system with locally distorted data; collaborative anomaly levels were used to quantify global risks, ensuring that parameter sensitivity matched the severity of electricity theft. This dual-dimensional parameter correction mechanism significantly improved the accuracy and relevance of parameter responses.

[0091] Finally, a tiered intervention mechanism is triggered based on the adjusted dynamic parameters and collaborative anomaly characteristics. For example, a primary intervention is triggered when only electricity consumption characteristic data is abnormal and data credibility verification passes; a medium-level intervention is triggered when electricity consumption characteristic data is abnormal, data credibility verification passes, and the collaborative anomaly level is below the threshold; and a high-level intervention is triggered when electricity consumption characteristic data is abnormal, data credibility verification fails, and the collaborative anomaly level is above the threshold. This forms a complete closed loop from data collection, feature extraction, collaborative analysis, credibility verification, parameter adjustment to precise intervention.

[0092] In summary, by working closely together and integrating microscopic data from multiple enclosures with macroscopic data from the power grid, a cross-enclosure collaborative analysis and reliable verification mechanism was constructed, effectively solving the problems of intelligent camouflage for electricity theft and isolated monitoring blind spots. Example

[0093] In another embodiment, this application also discloses a non-transitory machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the above-described anti-theft monitoring method for the electricity metering box.

[0094] The core innovation of this application lies in combining electricity consumption characteristic data with macro-level characteristic data using a federated learning model and blockchain technology, and introducing a tiered intervention mechanism. This effectively solves the monitoring blind spots and evasion problems caused by data tampering and isolated monitoring. Specifically, the solution utilizes anti-spoofing features to identify intelligent spoofing electricity theft behavior, combines macro-level indicators such as regional total load and line loss to reveal cross-container collaborative electricity theft risks, and ensures data credibility through blockchain. Ultimately, it achieves a complete closed loop from data collection to precise intervention, significantly improving the reliability and accuracy of anti-electricity theft monitoring.

[0095] This application proposes a method for monitoring electricity theft prevention in electricity metering boxes, comprising: acquiring electricity consumption characteristic data of multiple electricity metering boxes, including basic characteristics and anti-spoofing characteristics; acquiring macroscopic characteristic data from the power grid side, including regional total load and line loss; based on the electricity consumption characteristic data and macroscopic characteristic data, using a federated learning model to perform cross-box collaborative feature learning to generate collaborative anomaly features; anchoring the electricity consumption characteristic data and macroscopic characteristic data to a blockchain, and performing dual-dimensional data verification through smart contracts to verify data credibility; adjusting the dynamic parameters of the electricity metering boxes based on the collaborative anomaly features and data credibility; and triggering a tiered intervention mechanism based on the adjusted dynamic parameters and collaborative anomaly features.

[0096] In practical applications, electricity consumption characteristic data can be understood as multi-dimensional information related to electricity consumption behavior extracted from the electricity metering box. Basic features describe typical electricity consumption patterns, such as the amplitude variation patterns of voltage and current, or the overall trend of the load curve. Anti-spoofing features are specifically used to identify potential spoofing behavior, such as detecting abnormal fluctuations by analyzing the frequency of voltage ripple spikes or the time distribution entropy of the load curve. Furthermore, electricity consumption characteristic data can be acquired in real-time by sensors installed in the electricity metering box, or extracted after preprocessing historical data using edge computing devices.

[0097] Macro-level characteristic data refers to information obtained from the power grid side that reflects the overall electricity consumption status. Regional total load characterizes the total electricity consumption of all users within a specific region, while line losses reflect energy losses during transmission. Specifically, macro-level characteristic data can be obtained directly through the data interface of the power grid dispatch center, or by aggregating and calculating the local power grid status through a distributed sensor network.

[0098] Federated learning models can be understood as a distributed machine learning framework, whose core lies in achieving multi-node collaborative modeling while protecting data privacy. As a preferred implementation, the training process of a federated learning model can involve deploying a lightweight local model in each electricity metering box and uploading the model parameters to a regional server for aggregation, thereby generating a global model. Furthermore, the learning objective of the federated learning model can be set to identify cross-box collaborative electricity theft behavior, for example, by analyzing load transfer patterns or parameter consistency deviations between multiple boxes to generate collaborative anomaly features.

[0099] The process of anchoring data to the blockchain can be understood as recording electricity consumption characteristic data and macroeconomic characteristic data in the form of blocks in a distributed ledger to ensure the immutability of the data. Specifically, data anchoring can be achieved by generating independent micro-data blocks from the electricity consumption characteristic data of each electricity metering box and generating macro-data blocks from the macroeconomic characteristic data of the power grid side. Then, hash values ​​are calculated for each block and associated with adjacent blocks in the blockchain. Smart contracts are used to perform two-dimensional data verification, such as verifying the consistency and credibility of the data by comparing the individual total deviation rate or the deviation of characteristic correlations.

[0100] Dynamic parameter adjustment can be understood as the process of correcting the operating parameters of an electricity metering box based on the characteristics of coordinated anomalies and the reliability of the data. For example, when the data reliability is high, it can be selected as the benchmark for parameter adjustment; when coordinated anomalies indicate a risk of cross-box electricity theft, local parameters can be weighted and corrected. Furthermore, dynamic parameter adjustment can also incorporate other external constraints, such as ambient temperature or user electricity consumption habits, to improve the accuracy of the adjustment.

[0101] A tiered intervention mechanism can be understood as a process of adopting differentiated response strategies based on the type and severity of the anomaly. For example, when only a single electricity meter box is detected to have abnormal electricity consumption data and the data is highly reliable, primary intervention measures can be triggered; when cross-meter box coordinated electricity theft is detected and the level of coordination anomaly is high, higher-level intervention measures can be triggered.

[0102] The working principle of this application embodiment is as follows: First, by acquiring electricity consumption characteristic data from multiple electricity metering boxes, basic features are used to capture regular electricity consumption patterns, while anti-spoofing features are specifically used to identify AI-generated false load curves or data tampering behavior. Furthermore, by combining macroscopic characteristic data from the power grid side, including global indicators such as regional total load and line losses, collaborative electricity theft behaviors such as cross-box load transfer and total load sharing are revealed. Thus, a dual-dimensional data input, encompassing both micro and macro dimensions, is formed, providing comprehensive data support for subsequent analysis.

[0103] Secondly, based on the aforementioned electricity consumption characteristic data and macroscopic characteristic data, a federated learning model is used for cross-container collaborative feature learning. Specifically, the federated learning model achieves distributed aggregation and analysis of multi-container data without transmitting raw electricity consumption data, generating collaborative anomaly features to quantify the global electricity theft risk. This distributed modeling approach not only protects user privacy but also avoids the latency issues that may arise from centralized modeling, thus effectively addressing the core pain points of distributed collaborative electricity theft scenarios.

[0104] Furthermore, electricity consumption characteristic data and macroscopic characteristic data are anchored to the blockchain, and dual-dimensional data verification is performed through smart contracts. Specifically, the individual summation deviation rate is used to verify the consistency between the sum of loads across multiple containers and the total regional load, while the characteristic correlation deviation is used to verify the matching degree between the microscopic characteristics of a single container and the global collaborative characteristics. In this way, the immutability and consistency verification of the data are ensured, fundamentally addressing the problem of lost credibility caused by data tampering.

[0105] Subsequently, the dynamic parameters of the electricity metering box were adjusted based on collaborative anomaly characteristics and data credibility. Specifically, data credibility was used to screen reliable data sources and avoid misleading the system with locally distorted data; collaborative anomaly levels were used to quantify global risks, ensuring that parameter sensitivity matched the severity of electricity theft. This dual-dimensional parameter correction mechanism significantly improved the accuracy and relevance of parameter responses.

[0106] Finally, a tiered intervention mechanism is triggered based on the adjusted dynamic parameters and collaborative anomaly characteristics. For example, a primary intervention is triggered when only electricity consumption characteristic data is abnormal and data credibility verification passes; a medium-level intervention is triggered when electricity consumption characteristic data is abnormal, data credibility verification passes, and the collaborative anomaly level is below the threshold; and a high-level intervention is triggered when electricity consumption characteristic data is abnormal, data credibility verification fails, and the collaborative anomaly level is above the threshold. This forms a complete closed loop from data collection, feature extraction, collaborative analysis, credibility verification, parameter adjustment to precise intervention.

[0107] In summary, by working closely together and integrating microscopic data from multiple enclosures with macroscopic data from the power grid, a cross-enclosure collaborative analysis and reliable verification mechanism was constructed, effectively solving the problems of intelligent camouflage for electricity theft and isolated monitoring blind spots.

[0108] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for monitoring electricity theft in an electricity metering box, characterized in that, include: Acquire electricity consumption characteristic data from multiple electricity metering boxes, the electricity consumption characteristic data including basic characteristics and anti-spoofing characteristics; Acquire macroscopic characteristic data from the power grid side, including regional total load and line losses; Based on the electricity consumption characteristic data and the macroscopic characteristic data, a federated learning model is used to perform cross-box collaborative feature learning in order to generate collaborative anomaly features. The electricity consumption characteristic data and the macroscopic characteristic data are anchored to the blockchain, and the data is verified through a two-dimensional data verification via smart contracts to verify the data credibility. Based on the aforementioned collaborative anomaly characteristics and the aforementioned data reliability, the dynamic parameters of the power metering box are adjusted; as well as Based on the adjusted dynamic parameters and the aforementioned collaborative anomaly characteristics, a tiered intervention mechanism is triggered.

2. The method for monitoring electricity theft in an electricity metering box according to claim 1, characterized in that, The cross-box collaborative feature learning based on the electricity consumption characteristic data and the macroscopic characteristic data using a federated learning model includes: Each electricity meter box is used as an edge training node in federated learning to extract local electricity consumption feature data. The model parameters of multiple edge training nodes are aggregated on the regional power grid edge server to train a cross-container collaborative feature model; and Based on the cross-enclosure collaborative feature model, a collaborative anomaly level is output, which is used to characterize the degree of anomaly in cross-enclosure collaborative electricity theft.

3. The method for monitoring electricity theft in an electricity metering box according to claim 2, characterized in that, The cross-container collaborative feature model learning includes at least one of spatiotemporal collaborative features, load transfer features, and parameter consistency features.

4. The method for monitoring electricity theft in an electricity metering box according to claim 1, characterized in that, The step of anchoring the electricity consumption characteristic data and the macroscopic characteristic data to the blockchain includes: The electricity consumption characteristic data of each electricity metering box is used to generate micro data blocks, and the macro characteristic data of the power grid side is used to generate macro data blocks. Calculate the hash values ​​of the micro-data blocks and macro-data blocks respectively, and associate them with adjacent blocks in the blockchain; and The smart contract performs two-dimensional data verification, including calculating the total deviation rate of individuals and the deviation of feature correlation.

5. The method for monitoring electricity theft in an electricity metering box according to claim 4, characterized in that, The adjustment of the dynamic parameters of the power metering box based on the cooperative anomaly characteristics and the data reliability includes: Based on the data credibility, reliable data is selected as the benchmark for parameter adjustment. Based on the aforementioned collaborative anomaly level, local parameters are weighted and corrected; and Based on the sum of individual deviation rates, the parameter tolerance is dynamically adjusted.

6. The method for monitoring electricity theft in an electricity metering box according to claim 1, characterized in that, The triggering tiered intervention mechanism includes: A primary intervention is triggered when only the electricity consumption characteristic data is abnormal and the data credibility verification is passed. When the electricity consumption characteristic data is abnormal, the data credibility verification passes, and the collaborative anomaly level is below the threshold, a medium-level intervention is triggered; and Advanced intervention is triggered when the electricity consumption characteristic data is abnormal, the data credibility verification fails, and the collaborative anomaly level is higher than the threshold.

7. The method for monitoring electricity theft in an electricity metering box according to claim 2, characterized in that, It also includes lightweight optimization of the federated learning model, including: A lightweight model is generated through model distillation, and the lightweight model is deployed to an energy metering box; and The gradients of the model parameters are compressed and quantized to reduce the amount of data transmitted.

8. The method for monitoring electricity theft in an electricity metering box according to claim 1, characterized in that, The acquisition of electricity consumption characteristic data from multiple electricity metering boxes includes: At least one of the following is collected in real time from the power metering box: voltage ripple mutation frequency, current phase shift, and load curve time distribution entropy, as the basic feature and anti-spoofing feature.

9. A computing device, characterized in that, include: At least one processor; as well as A memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the anti-theft monitoring method for the electricity metering box as described in any one of claims 1-8.

10. A non-transitory machine-readable storage medium, characterized in that, It stores executable instructions that, when executed, cause the machine to perform the anti-theft monitoring method for the electricity metering box as described in any one of claims 1-8.

Images