Electric power measurement data management system based on cloud-side collaboration and electric quantity distributed acquisition method

Abstract

The invention discloses an electric power measurement data management system based on cloud edge collaboration and an electric quantity distributed acquisition method, and relates to the technical field of electric power information acquisition and processing. The system comprises an intelligent electric meter terminal layer, an edge computing node layer, a regional collaborative gateway layer and a cloud data center. The intelligent terminal can dynamically adjust the sampling frequency according to the load change rate and the historical volatility, and identifies abnormity through an LSTM-ATT model; the edge nodes realize prediction caching based on Kalman filtering and RMSE judgment; the regional gateway executes dynamic time warping and consistency arbitration of topological weighting; and the cloud end completes digital twinborn modeling and exception evidence storage. According to the method, the real-time performance and accuracy of electric power data acquisition are improved, the communication bandwidth occupation is reduced, the system abnormity sensing capability and the operation stability are enhanced, and the method is suitable for an intelligent power distribution network and a distributed energy metering scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric power information collection and processing, and particularly relates to an electric power metering data management system based on a cloud-edge collaborative architecture and a power distribution type collection method, which is suitable for real-time acquisition of multi-node electric power data, edge analysis and cloud scheduling optimization in scenarios such as smart distribution networks, distributed energy management and industrial park energy consumption monitoring. BACKGROUND

[0002] In the current power distribution network and energy management system, the electric power metering system, as the infrastructure for obtaining power consumption data and supporting power scheduling, directly affects the economy, safety and intelligent level of the power grid in terms of operation efficiency and data accuracy. Traditional electric power metering systems mostly adopt a centralized collection structure, that is, the terminal smart meter periodically uploads power consumption data to the central server, and the central side uniformly completes data storage, processing and analysis. This architecture has certain universality and implementation cost advantage in the early stage, but under the background of current power grid scale expansion, multi-source load access and increasing real-time scheduling demand, it gradually exposes a series of key problems.

[0003] Firstly, the real-time performance of data collection is insufficient in the traditional centralized structure. Since the data of the terminal meter needs to be collected and transmitted to the central server through the communication gateway and relay node, limited by communication bandwidth, network congestion and processing bottleneck, the conventional sampling period is mostly 15 minutes, which cannot meet the needs of abnormal detection and demand control for second-level response. When a sudden event such as power surge or equipment failure occurs, the central system can only obtain the relevant information after several minutes, which has obvious monitoring time lag.

[0004] Secondly, the fixed sampling rate mechanism lacks dynamic response capability. Most existing meters use a fixed sampling frequency (such as 1 Hz or 15 minutes / second) to collect load information, which cannot be flexibly adjusted according to the user-side load fluctuation: if the sampling frequency is too low, key feature points may be missed, affecting load identification accuracy; if the sampling frequency is too high, a large amount of redundant data will be generated, increasing communication burden and storage pressure.

[0005] Thirdly, the lack of edge computing capability leads to heavy load of the center. Traditional meters and relay nodes lack local computing and intelligent analysis capability, and all data need to be uploaded to the center for processing, which leads to high computing pressure, high storage cost and algorithm response delay of the center system. At the same time, in the case of communication link interruption or node failure, the system is prone to data loss or abnormal detection, and the overall robustness is poor.

[0006] Fourth, it is difficult to guarantee the spatio-temporal consistency of multi-node data. In existing systems, due to network delay and imperfect synchronization mechanism, there are often problems such as time dislocation and inconsistent collection granularity among different electric meters, especially in multi-point correlation analysis (such as abnormal traceability and harmonic traceability) scenarios, which will seriously affect the accuracy of the analysis results. The common time deviation can reach hundreds of milliseconds, which is enough to cause misjudgment of fault identification.

[0007] Fifth, the intelligent mechanism for abnormal data processing is lacking. Current systems mostly rely on central rule libraries for threshold judgment and lack model-driven prediction and learning capabilities. In the face of complex volatile loads, distributed energy access and seasonal fluctuating load patterns, traditional algorithms are difficult to adapt to their dynamic evolution, with high false negative and false positive rates.

[0008] In addition, the weak data credibility and security guarantee mechanism is also one of the important problems of the current electric power metering system. Some scenarios such as user dispute electricity, malicious load manipulation detection, etc. urgently need data processing capabilities with traceability and auditability, but the existing system lacks data chain storage or trusted timestamp mechanism, which is difficult to support the compliance requirements of subsequent businesses such as power trading and load forecasting.

[0009] Therefore, there is an urgent need for a new electric power metering data management system that integrates dynamic sampling, edge intelligence, regional collaboration, spatio-temporal alignment and cloud analysis to solve the technical bottlenecks of poor real-time performance, high bandwidth occupation, high abnormality detection rate, and difficulty in synchronization in the current centralized structure, and to improve the real-time performance, accuracy, stability and intelligence level of the system as a whole, providing efficient and reliable data foundation support for the construction of a new power system. SUMMARY

[0010] To overcome the technical bottlenecks of poor data sampling real-time performance, high central load pressure, abnormal response lag, and weak spatio-temporal consistency of multi-source data in existing electric power metering systems, the present application proposes an electric power metering data management system architecture that integrates dynamic sampling, edge intelligence, regional collaboration and cloud analysis. Through hierarchical collaborative design at the system level and intelligent algorithm optimization at the method level, the present application can significantly improve the collection efficiency, response speed and system stability while ensuring metering accuracy. The following will describe the function modules and technical implementation schemes of each level of the present application in combination with the system composition and core methods of the present application.

[0011] In a possible implementation, a cloud-edge collaborative power metering data management system is provided, comprising: a smart meter terminal layer, including at least one smart meter integrated with an STM32U5 microcontroller, for collecting voltage, current and other waveform data at an adaptive frequency of 32-128 Hz, and performing 256-point fast Fourier transform (FFT) to extract frequency domain features; an edge computing node layer, deployed at a power distribution site or transformer room, using a Jetson XavierNX edge server, running an LSTM-ATT model built based on a PyTorch Lightning framework, for receiving compressed feature data uploaded by the meter, and identifying abnormalities according to a root mean square error (RMSE) threshold γ = 0.25; a regional collaborative gateway layer for time series alignment and consistency judgment of multi-node data, using a topologically weighted dynamic time warping (DTW) algorithm to process data with an error of no more than ±50 milliseconds between nodes; and a cloud data center for performing statistical analysis, digital twin modeling and abnormal data chain evidence, and regularly issuing optimized model parameters and detection rules through an MQTT protocol.

[0012] In a possible implementation, the smart meter terminal layer further comprises: a dynamic sampling controller for automatically adjusting the sampling frequency according to the current load change rate ΔP / Δt (unit: kW / min) and the historical data volatility σ 2 (unit: kW 2 ) of the meter. When ΔP / Δt > 0.1 or σ 2 > 0.05, the system automatically switches to a high-frequency sampling mode (128 Hz, sampling interval about 7.8 milliseconds); when ΔP / Δt < 0.05 and σ 2 < 0.02, it switches to a low-frequency sampling mode (32 Hz). The frequency switching interval is set to no less than 2 minutes to prevent system jitter caused by frequent fluctuations in the sampling rate. The layer also includes an abnormality identification module that marks the current data segment as abnormal and immediately uploads it if the local LSTM-ATT model output abnormality confidence is less than 0.6.

[0013] In a possible implementation, the edge computing node layer comprises:

[0014] a Kalman filter module, whose process noise Q k and observation noise R k are dynamically adjusted according to the following formulas:

[0015]

[0016] where ΔP / Δt is the load change rate per unit time (kW / min), and σ 2 is the load variance based on a 10-minute window (kW 2 );

[0017] The cache judgment module is used to compare the RMSE of the LSTM prediction results with the real-time sampled data. When the error exceeds 0.25, it is uploaded immediately.

[0018] The data persistence module uses a high-performance in-memory database such as Redis to retain data fragments of nearly 24 hours to support subsequent anomaly backtracking and AI model training.

[0019] In one possible implementation, the regional collaborative gateway layer includes:

[0020] The spatiotemporal alignment module employs a topology-weighted DTW algorithm based on a sliding window (±10 sampling points) for sequence synchronization, with the modified cost function as follows:

[0021] cost(i,j)=|x i -y j |(1+λ·d topo (i,j))

[0022] Where d topo (i,j) represents the power grid topology distance between nodes i and j, and λ is a correction coefficient with a preferred value of 0.05;

[0023] The data consistency arbitration module uses a voting mechanism of three or more nodes to arbitrate when there are conflicts in the data of multiple nodes. The node scoring weights include: reputation (0.6), data consistency (0.3), and topology connectivity (0.1). When the score is lower than 0.8, the blockchain evidence storage mechanism is triggered.

[0024] In one possible implementation, a distributed power acquisition method based on the aforementioned system is provided, comprising the following steps:

[0025] S1: The data acquisition terminal calculates the current ΔP / Δt and σ. 2 And the sampling frequency is set accordingly;

[0026] S2: Perform FFT feature extraction and LSTM-ATT anomaly detection, and compress and upload the data;

[0027] S3: An error was detected in the RMSE calculation of the edge node, triggering an upload.

[0028] S4: The regional collaborative gateway performs DTW alignment and consistency adjudication on multi-source data under topology constraints;

[0029] S5: Cloud-based execution load statistics, digital twin modeling, and model feedback training.

[0030] In one possible implementation, the LSTM-ATT model in S2 includes: an input dimension of 6 (voltage, current, frequency, harmonics, active power, reactive power), a 2-layer LSTM network with 64 hidden nodes in each layer, and a time step of 20; the attention mechanism consists of a 2-layer fully connected network with Tanh activation function, and the output is weighted by Softmax.

[0031] In one possible implementation, the cloud data center further includes: a digital twin module (integrating GIS and real-time load data to construct a virtual power grid), a blockchain notarization module implemented using the SM3 digest algorithm combined with Hyperledger Fabric, and a visualization chart component implemented based on ECharts or Plotly components.

[0032] In one possible implementation, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the method steps described above. The program modules include: an ADC control subroutine, an LSTM inference subroutine, an MQTT data upload subroutine, an edge prediction judgment subroutine, and a Flink streaming data analysis subroutine.

[0033] Based on the above technical solutions, the cloud-edge collaborative power metering data management system and distributed power acquisition method of the present invention constructs an efficient, intelligent, and low-latency distributed power metering data processing architecture for distribution network scenarios, and has the following significant technical effects:

[0034] 1. Significantly improves the real-time performance of data processing: This invention uses the local LSTM-ATT anomaly detection model of edge computing nodes to reduce the fault event detection and response time from minutes in traditional centralized processing to seconds (<3s), which greatly improves the system's agility and security early warning capabilities.

[0035] 2. Achieve efficient utilization of communication resources: through a dynamic sampling mechanism (based on ΔP / Δt and σ). 2 Based on the data upload logic driven by indicators and prediction deviations, this invention can reduce the amount of uplink data by 72% to 85% while ensuring data integrity, effectively alleviating the bandwidth pressure on the power distribution network communication link.

[0036] 3. Enhance multi-node data consistency and fusion capabilities: By introducing the topological constraint dynamic time warping (DTW) algorithm and weighted consistency voting mechanism, the system can control the time synchronization error of different acquisition terminals within ±50ms, improve the accuracy of edge collaborative processing, and solve the problems of "data silos" and "acquisition misalignment" in traditional solutions.

[0037] 4. Excellent hardware and software synergy and feasibility: The key modules of this system are based on mainstream embedded microcontrollers (such as STM32U5), edge AI platforms (such as JetsonNX), standard containerized deployment technologies (such as Docker), and mainstream deep learning frameworks (such as PyTorch), which ensures the feasibility, scalability, and maintainability of the solution, making it suitable for rapid deployment in various industrial power distribution scenarios.

[0038] 5. Support for digital twin and intelligent decision-making function expansion: The cloud system integrates GIS topology model and real-time collected data, which can build a digital twin view of the power distribution network and support multi-dimensional power consumption behavior analysis and equipment operation status modeling in combination with streaming engines such as Flink, further expanding the system's perception depth and proactive optimization capabilities.

[0039] 6. Enhance the credibility and auditing capabilities of abnormal data: This invention integrates blockchain technology to store important abnormal events or detection data in a chain, ensuring that subsequent scheduling decisions, accountability, and data sharing are verifiable and complete, meeting the energy industry's compliance requirements for key data to be "traceable, auditable, and non-repudiable".

[0040] 7. Possesses promising application and promotion prospects: This system can be widely applied to multiple typical scenarios such as urban smart distribution networks, industrial park power management, distributed photovoltaic grid-connected monitoring, and new energy load forecasting. It has strong adaptability and promotion potential, and provides key infrastructure support for building a collaborative perception system of "source-grid-load-storage" in a new power system.

[0041] In summary, this invention not only achieves three-level linkage between cloud, edge, and terminal in system architecture design, but also demonstrates originality, advancement, and industrial application value in multiple aspects such as algorithm model, data mechanism, and platform deployment. It can effectively solve the technical bottlenecks of existing power metering systems in terms of data acquisition efficiency, fault detection timeliness, and system stability, and has a clear foundation for patent licensing and commercial transformation prospects. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall system architecture of the present invention;

[0043] Figure 2 This is a schematic diagram of the smart meter terminal layer module of the present invention;

[0044] Figure 3 This is a schematic diagram of the edge computing node and data processing flow of the present invention;

[0045] Figure 4 This is a schematic diagram of the regional collaborative gateway and cloud functional modules of the present invention. Detailed Implementation

[0046] To make the technical solution of this invention clearer, more complete, and easier to understand, the following describes in detail the specific implementation of this invention in typical scenarios, combined with actual deployment architecture and system operation flow. It should be noted that those skilled in the art can make various modifications or substitutions to the described implementation without departing from the spirit and essence of this invention, and these should all be included within the protection scope of this invention. The implementation process of this invention revolves around the technical main line of "collection—analysis—judgment—upload—coordination—optimization," achieving key functions through engineering-level hardware and software collaboration. The following will describe each layer in detail according to the system composition and its operating logic.

[0047] like Figure 1 As shown in this embodiment, an industrial park is selected as the application scenario to build a power metering data management system based on a cloud-edge collaborative architecture. The system consists of four parts: intelligent terminal devices, edge computing nodes, regional collaborative gateways, and cloud data platform. It also constructs a complete data acquisition and analysis closed loop through multi-level communication and data processing logic.

[0048] like Figure 2 As shown, during the system initialization phase, 200 smart meters are first deployed at the main power distribution points within the park as terminal data acquisition devices. Each meter integrates an STM32U5 series microcontroller and a high-precision ADC module, possessing independent edge data processing capabilities. Upon startup, the meter's control software loads a set of preset dynamic sampling parameters, including a load change rate threshold α = 0.1 (unit: kW / min) and a load fluctuation threshold β = 0.05 (unit: kW). 2 The system offers three sampling frequency levels: low frequency (32Hz), medium frequency (64Hz), and high frequency (128Hz). The system defaults to medium frequency sampling and dynamically evaluates and reconfigures the sampling frequency every 2 minutes.

[0049] During the operation of the electricity meter, the control module calculates in real time the load change rate ΔP / Δt per unit time and the variance σ based on data from the past 5 minutes. 2 When ΔP / Δt exceeds 0.1 or σ 2 When the value is greater than 0.05, the system automatically switches to a high-frequency sampling mode to enhance its ability to capture high-dynamic load behavior; conversely, when...

[0050] ΔP / Δt is less than 0.05 and σ 2 When the value is less than 0.02, the frequency is reduced to a low-frequency mode to reduce redundant data generation. In each sampling period, the acquired waveform data such as voltage and current are processed by a 256-point Fast Fourier Transform (FFT) to extract frequency domain features and then fed into the locally loaded LSTM-ATT anomaly recognition model.

[0051] The anomaly detection model is trained and deployed using the PyTorch framework. Its input includes six feature parameters: RMS voltage, RMS current, fundamental frequency, third harmonic amplitude, active power, and reactive power. The model structure consists of a two-layer LSTM (64 hidden units per layer) and an attention mechanism for weighted scoring of key fluctuations in the time series. When the model output confidence score is below 0.6, or the RMSE between the predicted and actual sample values ​​exceeds 0.25, the current sample segment is marked as a "potential anomaly," then packaged, compressed, and sent to the upper-level edge computing node via LoRa or MQTT protocols.

[0052] like Figure 3 As shown, at the edge computing node layer, the system uses Jetson XavierNX as the hardware platform and runs containerized intelligent processing modules. Each edge node connects to several smart meters and is responsible for receiving, verifying, buffering, and analyzing data streams. Each edge node has a built-in set of adaptive Kalman filter modules for dynamically adjusting the noise of the uplink data. The process noise Q and observation noise R are calculated according to the following formulas:

[0053]

[0054] Where Q0 and R0 are the system-defined base values ​​(e.g., 1×10⁻⁶). -3 With 1×10 -2 ), ΔP / Δt and σ 2 Inherited from terminal sampling results. Through the adaptive characteristics of the above filters, the system can automatically adjust the signal-to-noise ratio weights according to load fluctuations, thereby improving the stability of predictions.

[0055] For each piece of terminal data, the edge node performs an RMSE calculation. If the prediction deviation is greater than 0.25, the data is uploaded to the regional collaborative gateway; otherwise, it is retained in the local Redis cache for 24 hours. The data in the cache is also used to support local model retraining, cloud anomaly backtracking, or manual auditing processes.

[0056] At the regional collaborative gateway layer, the system performs unified coordination and consistency judgment on data from multiple edge nodes. Considering potential time discrepancies due to communication delays between nodes, the regional gateway employs a topology-weighted Dynamic Time Warping (DTW) algorithm for timing alignment. This algorithm sets a maximum sliding window of ±10 sampling points (approximately ±78ms) and introduces a grid topology distance factor d. topo (i,j) and the adjusted weight λ=0.05 form the following cost function:

[0057] cost(i,j)=|x i -y j|(1+λ·d topo (i,j))

[0058] After time alignment is completed, the gateway system performs consistency scoring on the data uploaded by adjacent nodes and uses a hybrid voting mechanism to decide whether to store the data on the blockchain. The scoring mechanism combines node reputation (0.6%), data consistency (0.3%), and topology tightness (0.1%). When the total score is below 0.8, it is determined that the data set has potential credibility issues, triggering the blockchain notarization module and marking it as an anomaly.

[0059] To enhance the non-repudiation and auditability of data uploads and anomaly detection results, this embodiment deploys a blockchain evidence storage module on a regional collaborative gateway or cloud platform. This module is used to process data fragments that meet the consistency score threshold (e.g., below 0.8) for on-chain processing. The blockchain evidence storage module is built using a lightweight consortium blockchain architecture, supporting the SM3 hash algorithm and SM2 signature mechanism to ensure data digest integrity and signature authentication.

[0060] Each time evidence storage is triggered, the system automatically constructs a data on-chain transaction record, which includes the following fields:

[0061] Timestamp: Indicates the time when the abnormal data was identified, accurate to milliseconds;

[0062] Node ID (node_id): A unique identifier for the source edge node, which can be a MAC address or device serial number;

[0063] Data digest (data_hash): The abnormal power data content, RMSE value, predicted value and measured value uploaded this time are digested using SM3 to generate an irreversible digest value with a length of 256 bits;

[0064] Consistency score: Records the score of this data voting.

[0065] Signature verification: The regional gateway or cloud-based issuing node packages the above data and performs an SM2 signature to ensure the validity of the evidence.

[0066] On-chain record address (block_index / tx_id): identifies the block number and transaction number of the transaction in the blockchain, facilitating traceability.

[0067] The above records are triggered by the Fabric chaincode smart contract and simultaneously broadcast to other nodes in the consortium chain for evidence storage, realizing a verifiable and tamper-proof historical record of abnormal power behavior.

[0068] The system supports monthly archived summaries on the blockchain and allows authorized regulatory agencies or third parties to conduct security audits of data upload history through a blockchain explorer, effectively improving the credibility and compliance of the system in areas such as measurement anomalies and edge intelligence intervention records.

[0069] like Figure 4 As shown, in the cloud data center, the system adopts a fusion platform based on Flink stream processing architecture and Spark offline processing architecture to receive data from the upper-layer regional gateway and construct a digital twin power grid by combining it with a GIS power grid topology model. The cloud can dynamically analyze the current status of the power distribution system based on the data uploaded from the edge, identify hotspot nodes, load imbalances or potential trends, and generate visualized load trend maps, electricity consumption behavior distribution maps, and fault prediction models. At the same time, the cloud platform retrains the LSTM-ATT model uniformly every 7 days and decides whether to push a new version of the model to the terminal based on the model accuracy feedback uploaded by each edge node.

[0070] In a six-month pilot operation in the industrial park, the implementation of this invention reduced the average daily uplink data from 2GB to approximately 280-320MB across 200 metering points. The average system anomaly response time was shortened to within 2.6 seconds, the model recognition accuracy remained above 98.3%, and the communication bandwidth saving rate exceeded 75%, effectively improving the real-time performance, intelligence, and stability of the metering system.

[0071] Example 1: Multi-node power consumption monitoring system for commercial complexes

[0072] In this embodiment, the system of the present invention is deployed in a large commercial complex consisting of five commercial office buildings, with approximately 1,200 independent power metering circuits covering various loads such as lighting, air conditioning, elevators, and tenant-specific equipment. To meet the complex's needs for high-frequency energy consumption monitoring, dynamic load identification, and electricity consumption behavior analysis, the operator upgraded the system using the cloud-edge collaborative power metering data management system described in this invention.

[0073] Each building is equipped with approximately 240 smart meters, all of which feature an STM32U5 microcontroller and a local computing module. The dynamic sampling parameters are configured as follows: α = 0.12, β = 0.045, and a sampling frequency range of 32–128 Hz. The system uses a local LSTM-ATT model to learn and identify load changes in real time, developing adaptive recognition patterns for typical electricity consumption characteristics such as cold start-up, elevator start-up and shutdown, and peak lighting fluctuations.

[0074] Jetson NX edge node devices are deployed in the building's power distribution room, communicating with the corresponding meter groups via dual links of LoRa and Ethernet. Each edge node sends aggregated features to the area gateway every 5 minutes and dynamically reports raw anomaly data based on RMSE prediction deviations. The area gateway is deployed in the IoT control center, which uniformly performs time-series alignment and data consistency checks on all building edge nodes.

[0075] The cloud-based system generates daily energy consumption maps for the entire building and sub-areas, with an average response time of less than 3 seconds for abnormal load alarm data. Compared to the previous centralized upload architecture, the overall uplink data compression ratio of the system reaches 78%, and the comparison of energy consumption differences among buildings is clearer, providing an important basis for energy consumption allocation and operation and maintenance optimization for commercial tenants.

[0076] Example 2: Two-way energy metering and fluctuation control applied in distributed photovoltaic grid connection scenarios

[0077] In this embodiment, the system of the present invention is applied to a distributed photovoltaic access area in the urban-rural fringe of a prefecture-level city, comprising rooftop photovoltaic systems for 850 users with a total capacity of approximately 5MWp. Due to the drastic fluctuations in distributed photovoltaic power and uneven load distribution, which pose operational risks such as "peak backfeeding" and "voltage drop jumps" to the distribution network, the local power company plans to deploy a high-response power acquisition and anomaly prediction system to improve the accuracy of grid connection management.

[0078] According to the present invention, each household installs one smart meter, which has the ability to perform high-frequency sampling and local anomaly identification. Considering the coupling characteristics between the photovoltaic system output and the user-side load, the terminal LSTM model incorporates solar irradiance and external meteorological parameters as auxiliary inputs to enhance its ability to identify sudden changes in photovoltaic power.

[0079] Edge nodes are deployed in four village-level power distribution rooms to aggregate meter data from approximately 200 households. An RMSE dynamic caching mechanism is enabled at the edge, and reverse power surges are extracted with specific tags and uploaded to the regional gateway. The gateway uses a time window mechanism to perform multi-point calibration of instantaneous power changes and collaborates with the cloud platform to model the spatiotemporal characteristics of photovoltaic-load.

[0080] Within three months of deployment, the system identified approximately 160 reverse over-limit events caused by photovoltaic power surges, and issued alarms 15 to 60 seconds in advance, assisting the dispatch center in initiating active power mitigation control and significantly reducing transformer overload rate and tripping risk. This embodiment demonstrates that the system of the present invention possesses intelligent identification and efficient response capabilities in the management of renewable energy fluctuations.

[0081] In summary, this invention significantly improves the real-time performance, accuracy, and robustness of power metering data by constructing a cloud-edge-device collaborative power metering data management system and combining key technologies such as dynamic sampling, edge prediction, anomaly detection, spatiotemporal alignment, and blockchain-based evidence storage. The system and method described are not only applicable to energy consumption monitoring and load identification in traditional distribution networks, but also particularly suitable for emerging power system scenarios such as distributed energy access, multi-node collaborative management, and new energy fluctuation control, possessing broad engineering application prospects and commercial transformation potential.

[0082] It should be understood that the above description is only a preferred embodiment of the present invention, and all equivalent substitutions and modifications made within the spirit and essence of the present invention should be included within the protection scope of the present invention.

Claims

1. A cloud-edge collaborative power metering data management system, characterized in that, include: The smart meter terminal layer includes at least one smart meter with an integrated STM32U5 microcontroller, which is used to acquire voltage and current waveform data at an adaptive frequency of 32Hz to 128Hz and perform a 256-point Fast Fourier Transform (FFT) to extract frequency domain features. The edge computing node layer, deployed at power distribution sites or substations, includes Jetson XavierNX edge servers, equipped with an LSTM-ATT model (2-layer LSTM, 64-dimensional hidden layers, attention-weighted output) and running the PyTorch Lightning 1.9 framework. It is used to receive compressed feature data uploaded by electricity meters and identify anomalies based on the root mean square error (RMSE) threshold γ = 0.

25. The regional collaborative gateway layer supports alignment logic for processing data from multiple nodes with a time difference not exceeding ±50ms, and uses the topology-weighted improved DTW algorithm for alignment, consistency scoring, and voting arbitration. The cloud-based data center provides digital twin modeling, model retraining, and visual statistical analysis, and regularly distributes model updates via the MQTT protocol. The system has a four-level data processing architecture, supports multi-granularity data synchronization strategies, reports critical alarm information within 5 seconds, and uploads non-critical data in hourly batch processing.

2. The system according to claim 1, characterized in that, The smart meter terminal layer further includes: The dynamic sampling controller is based on the current load change rate ΔP / Δt and historical volatility σ. 2 The sampling frequency is dynamically set, and the high-frequency sampling mode (128Hz, corresponding to a sampling period of 7.8ms) is enabled when the following conditions are met: ΔP / Δt>0.1kW / min or σ 2 >0.05 (variance calculation based on historical data window); The anomaly detection module marks the corresponding data as an anomaly and uploads it in real time when the confidence score output by the LSTM-ATT model is less than 0.

6.

3. The system according to claim 1, characterized in that, The edge computing node layer also includes: Kalman filter module, its process noise Q k and observation noise R k Dynamically adjust according to the following formula: Where ΔP / Δt is the load change rate per unit time (kW / min), σ 2 Load variance (kW) based on a 10-minute window 2 ); The cache determination module calculates the RMSE between the predicted value and the measured value. If the RMSE > γ = 0.25, the upload action is triggered. The data persistence module uses an in-memory database (such as Redis) to store data fragments from the past 24 hours to support backtracking.

4. The system according to claim 1, characterized in that, The regional collaborative gateway layer includes: The timing alignment module performs DTW matching based on a sliding window of ±10 points (equivalent to ±78ms); The data consistency voting mechanism uses a weighted average consistency score when the number of nodes is greater than 3. The weights are: reputation 0.6, consistency score 0.3, and topology tightness 0.

1. When the score is lower than 0.8, the blockchain notarization mechanism is triggered.

5. A distributed power acquisition method based on the system described in any one of claims 1-4, characterized in that, include: S1: The data acquisition terminal calculates ΔP / Δt and σ. 2 And determine whether to enter high-frequency sampling mode; S2: Perform FFT feature extraction, LSTM-ATT model inference, and package and upload; S3: Edge nodes use a caching mechanism to determine whether there are any anomalies, and decide whether to send data based on the RMSE result; S4: The region layer performs DTW alignment and consensus voting decisions based on topology optimization; S5: The cloud performs system state modeling and predictive optimization based on the received data.

6. The method according to claim 5, characterized in that, The setting of the sampling frequency f in step S1 satisfies: When ΔP / Δt>0.1 or σ 2 When >0.05, f = 128Hz; When ΔP / Δt < 0.05 and σ 2 When <0.02, f = 32Hz; Otherwise, f = 64Hz; Furthermore, the frequency switching interval is no less than 2 minutes to prevent frequent fluctuations from causing resource jitter.

7. The method according to claim 5, characterized in that, The LTSTM-ATT model in step S2 adopts the following structure: The input feature dimensions are 6, including: RMS voltage value, RMS current value, frequency, harmonic amplitude, active power, and reactive power; The LSTM network consists of 2 layers, with a hidden layer size of 64 and a time step of 20. The attention mechanism uses a two-layer fully connected network. The first layer has an output dimension of 32 and uses the Tanh activation function. The second layer has an output dimension of 1. Finally, attention weights are obtained and weighted output using the Softmax method. The anomaly detection criteria for this model are: an output score below 0.6, or an RMSE with the predicted load curve exceeding γ = 0.

25.

8. The method according to claim 5, characterized in that, The spatiotemporal alignment in step S4 includes: The time series are aligned using the weighted DTW algorithm, with a topological distance correction coefficient λ added. The correction cost function is as follows: cost(i,j)=|x i -y j |(1+λ·d topo (i,j)) Where d topo (i,j) represents the power grid topology distance between nodes i and j, and λ is a correction coefficient with a preferred value of 0.05; The timestamp alignment error is controlled within ±50ms, and a 10-point sliding window is used to accelerate matching. When inconsistent data exists, a weighted arbitration mechanism with 3 or more nodes is used to decide whether to record it on the blockchain or trigger a repair.

9. The system according to claim 1, characterized in that, The cloud data center further includes: The digital twin modeling module is used to merge edge aggregation data with GIS topology data to build a real-time virtual power distribution network model; The blockchain evidence storage module uses the SM3 digest algorithm and the Hyperledger Fabric framework to achieve chained recording of key abnormal data. The visualization and analysis module, based on ECharts or Plotly components, supports multi-dimensional data presentation and filtering analysis by node, region, and load type. The model distribution and feedback module supports an online update mechanism for edge models. By default, a new version of the LSTM model is released every 7 days, and the distribution frequency can be dynamically adjusted based on the accuracy feedback from edge nodes.

10. A computer-readable storage medium storing a computer program, said program, when executed by a processor, implementing the steps of the method as described in any one of claims 5 to 8, characterized in that... The program module includes: The data acquisition subroutine is used to control the ADC sampling rate with configuration parameters and to calculate ΔP / Δt and σ. 2 ; The model inference subroutine loads the LSTM-ATT model weights and performs anomaly inference and confidence scoring. The upload control subroutine determines whether to send the data to the edge node via MQTT based on the inference results. The edge preprocessing subroutine performs prediction caching, anomaly RMSE judgment, time-series alignment, and consistency scoring. The cloud-based analytics subroutine performs full network load analysis and model retraining based on the Spark or Flink stream processing framework.

Images