Distribution network management ammeter state analysis method and system

Abstract

The invention relates to the technical field of distribution network management, in particular to a distribution network management electricity meter state analysis method and system. According to the technical scheme, the distribution network management electricity meter state analysis method comprises the following steps that S1, multiple layers of edge computing nodes are deployed in a power distribution network, first-level nodes are deployed in a power distribution station, second-level nodes are deployed in an electricity meter concentrator, and computing resource distribution of the edge nodes is dynamically adjusted based on electricity meter distribution density and real-time loads; s2, through a collaborative acquisition mechanism of the first-level node and the second-level node, acquiring operation data of the electric meter in real time; and S3, transmitting the classified and marked data by adopting a multi-modal compression strategy: compressing the periodic monitoring data by using difference value coding, and extracting the sudden abnormal data by using a feature abstract. By constructing a multi-level edge collaborative acquisition mechanism and a 5G network slice transmission channel, the real-time bottleneck of traditional manual inspection and fixed-period acquisition is broken through.

Description

Technical Field

[0001] This invention relates to the field of power distribution network management technology, and in particular to a method and system for analyzing the status of electricity meters in power distribution network management. Background Technology

[0002] In power distribution network management, meter status analysis is a core component in ensuring power supply reliability. Traditional methods mainly rely on two modes: manual periodic inspections and fixed-cycle data collection. Maintenance personnel record meter data on-site every 7-15 days, supplemented by automated data collection devices that upload data hourly.

[0003] Manual inspections have drawbacks such as long cycles and irregular intervals, while automated data acquisition devices usually report data at preset fixed time intervals, making it difficult to capture sudden anomalies such as voltage surges and communication interruptions in a timely manner, resulting in a serious delay in fault detection. Summary of the Invention

[0004] This invention proposes a method and system for analyzing the status of electricity meters in distribution network management, which solves the problem in the prior art that it is difficult to capture sudden anomalies such as voltage surges and communication interruptions in a timely manner, resulting in a serious lag in fault detection.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A method for analyzing the status of electricity meters in a power distribution network includes the following steps:

[0007] S1. Deploy multi-layer edge computing nodes in the power distribution network, where primary nodes are deployed in substations and secondary nodes are deployed in meter concentrators. The allocation of computing resources for edge nodes is dynamically adjusted based on meter distribution density and real-time load.

[0008] S2. Real-time acquisition of meter operation data through a collaborative acquisition mechanism between primary and secondary nodes;

[0009] S3. For the classified and labeled data, a multimodal compression strategy is used for transmission: differential encoding compression is used for periodic monitoring data, and feature summary extraction is used for sudden abnormal data.

[0010] S4. Build a streaming processing engine in the cloud, dynamically calculate the mean and standard deviation thresholds of the meter data based on the sliding window, and integrate the LSTM time series prediction model and the random forest classification model to generate multi-dimensional anomaly detection results.

[0011] S5. Based on the anomaly detection results, call the digital twin model to map the real-time status of the faulty meter, and combine the distribution network topology to generate decision instructions that include fault probability, impact range and recommended handling strategies.

[0012] S6. Dynamically optimize decision instructions based on reinforcement learning algorithms.

[0013] Furthermore, in step S2, the operating data includes voltage, current, power factor, communication signal strength, and ambient temperature and humidity, and the data is classified and labeled.

[0014] In step S3, the sudden abnormal data is extracted using feature summaries and then high-priority transmission channels are divided through 5G network slicing.

[0015] In step S6, after dynamic optimization, the priority of computing resource allocation and alarm push strategy of edge nodes are adjusted based on historical operation and maintenance response time and resource consumption data.

[0016] Furthermore, the dynamic adjustment of computing resource allocation in step S1 includes:

[0017] Cellular grid partitions are generated based on the meter distribution density, and the computational resource weights for each partition are allocated according to the following rules:

[0018] Resource weight = Number of electricity meters / Zone area × Preset coefficient α + Real-time load rate × Preset coefficient β;

[0019] When the difference in resource weight between adjacent partitions exceeds a threshold, a resource migration task is triggered.

[0020] Furthermore, the data classification labels mentioned in step S2 include:

[0021] The operational data is divided into metering data, communication status data, and environmental data, and a timestamp and hierarchical label are attached to each type of data. The hierarchical label is generated based on the hierarchy of the edge node where the meter is located.

[0022] Furthermore, the multimodal compression strategy described in step S3 includes:

[0023] The periodic monitoring data is subjected to differential encoding compression, the base value sequence is preserved and the difference is stored; the abnormal data is extracted with feature summary, which includes peak value, duration and change gradient.

[0024] Furthermore, the dynamic calculation of the sliding window in step S4 includes:

[0025] Set the window length to N sampling periods and the sliding step size to M periods, where N>M and M is an adjustable parameter;

[0026] After eliminating outliers using the median substitution method within the window, the mean μ and standard deviation σ are calculated, and the dynamic threshold is set to μ±3σ.

[0027] The fusion of the LSTM time series prediction model and the random forest classification model includes: weighting and summing the time series anomaly probability output by LSTM and the discrete event classification result output by random forest, with the weight coefficients dynamically adjusted according to the historical fault types of the electricity meter, and satisfying that the sum of the weight coefficients is 1.

[0028] Furthermore, the digital twin model mapping described in step S5 includes:

[0029] Analyze the line impedance and power supply radius in the distribution network topology parameters, and construct the correlation matrix between the electricity meter and adjacent nodes;

[0030] When an anomaly is detected, the fault propagation path is calculated based on the correlation matrix, and a list of affected meter IDs is generated.

[0031] Furthermore, the logic for generating the decision instruction includes:

[0032] If the fault propagation path covers critical load nodes, the alarm priority is increased and cross-regional resource allocation is triggered; otherwise, a local isolation command is generated and the list of meters to be inspected is marked.

[0033] Furthermore, the dynamic optimization of the reinforcement learning algorithm in step S6 includes:

[0034] The state space is defined as a multi-dimensional vector of edge node resource utilization, fault impact level, and historical response time.

[0035] The action space includes alarm push frequency classification and computing resource migration path selection;

[0036] The reward function is calculated based on the difference between the fault recovery time reduction rate and the resource consumption cost, and the action selection rules are updated through a Q-learning strategy.

[0037] A power distribution network management meter status analysis system, comprising:

[0038] The edge resource dynamic allocation module is used to dynamically adjust the allocation of computing resources for multi-layer edge nodes based on the distribution density of electricity meters and real-time load, and generate resource migration instructions.

[0039] The multi-source data collaborative acquisition module is used to acquire the operating data of the electricity meter in real time and classify and label the data, including adding timestamps and hierarchical labels;

[0040] The multimodal compression transmission module is used to perform differential encoding compression on periodic monitoring data, extract feature summaries from abnormal data, and allocate high-priority transmission channels through 5G network slicing.

[0041] The streaming anomaly detection module is used to dynamically calculate the mean and standard deviation thresholds based on a sliding window, and to generate anomaly detection results by fusing an LSTM time series prediction model and a random forest classification model.

[0042] The digital twin decision generation module is used to call the digital twin model to map the status of faulty meters and generate decision instructions that include the scope of impact and handling strategies by combining the distribution network topology.

[0043] The reinforcement learning optimization module is used to dynamically optimize the priority of computing resource allocation and alarm push strategy by defining the state space, action space and reward function.

[0044] The positive effects of this invention are as follows: By constructing a multi-level edge collaborative acquisition mechanism and a 5G network slicing transmission channel, this invention breaks through the real-time bottleneck of traditional manual inspection and fixed-cycle acquisition. Based on an event-driven hybrid acquisition mode, it achieves intelligent switching between "periodic uploading of normal data and immediate triggering of abnormal data," compressing the delay in detecting sudden anomalies from hours to seconds. Combined with timestamp synchronization and a hierarchical tagging system, it establishes a precise mapping between data and the power grid topology, effectively improving the spatiotemporal positioning accuracy of fault events. The 5G high-priority channel ensures that critical data is not affected by network fluctuations, greatly improving transmission reliability. Detailed Implementation

[0045] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0046] A method for analyzing the status of electricity meters in a power distribution network includes the following steps:

[0047] S1. Deploy multi-layer edge computing nodes in the power distribution network, where primary nodes are deployed in substations and secondary nodes are deployed in meter concentrators. The allocation of computing resources for edge nodes is dynamically adjusted based on meter distribution density and real-time load.

[0048] S2. Real-time acquisition of meter operation data through a collaborative acquisition mechanism between primary and secondary nodes;

[0049] S3. For the classified and labeled data, a multimodal compression strategy is used for transmission: differential encoding compression is used for periodic monitoring data, and feature summary extraction is used for sudden abnormal data.

[0050] S4. Build a streaming processing engine in the cloud, dynamically calculate the mean and standard deviation thresholds of the meter data based on the sliding window, and integrate the LSTM time series prediction model and the random forest classification model to generate multi-dimensional anomaly detection results.

[0051] S5. Based on the anomaly detection results, call the digital twin model to map the real-time status of the faulty meter, and combine the distribution network topology to generate decision instructions that include fault probability, impact range and recommended handling strategies.

[0052] S6. Dynamically optimize decision instructions based on reinforcement learning algorithms.

[0053] Specifically as follows:

[0054] Layered deployment of edge nodes

[0055] Primary node: Deployed in the substation, responsible for regional data aggregation and resource scheduling, forming a tree-like communication topology with the secondary node.

[0056] Secondary node: Deployed at the meter concentrator, it directly collects meter data via LoRa or HPLC communication protocols and performs local preprocessing.

[0057] Dynamic resource allocation: Based on cellular grid partitioning (each grid coverage radius ≤ 500 meters), the resource weight of each partition is calculated in real time.

[0058]

[0059] Wherein, α and β are empirical coefficients. α is the meter distribution density weight coefficient, with a preset empirical value (default 0.6), used to adjust the impact of the number of meters on resource allocation. β is the real-time load rate weight coefficient, with a preset empirical value (default 0.4), used to adjust the impact of the node's current load on resource allocation. The number of meters is the total number of meters in the cellular grid partition, the partition area is the physical coverage area of ​​the cellular grid, and the real-time load rate is the current CPU / memory resource utilization of the edge node. When the weight difference between adjacent partitions exceeds 15%, resource migration is triggered, such as migrating computing tasks from high-load areas to low-load areas.

[0060] Collaborative data collection mechanism

[0061] Data collection cycle: The primary node initiates a poll to the secondary node every 5 minutes, and the secondary node collects meter data at 30-second intervals.

[0062] Fault tolerance: If the secondary node fails, the primary node can directly connect to the meter through a backup communication link (such as a 4G DTU).

[0063] By constructing a multi-layered edge computing architecture (substation-concentrator level) and dynamically adjusting resource allocation based on meter distribution density and real-time load, spatially balanced distribution and time-flexible scheduling of computing resources are achieved. This design overcomes the resource rigidity problem of traditional centralized processing, enabling the system to adaptively allocate computing power according to changes in the power grid topology. While ensuring real-time performance, it avoids overloading of local nodes, forming a flexible computing system of "edge autonomy and cloud collaboration".

[0064] In step S2, the operating data includes voltage, current, power factor, communication signal strength, and ambient temperature and humidity, and the data is classified and labeled.

[0065] In step S3, the sudden abnormal data is extracted using feature summaries and then high-priority transmission channels are divided using 5G network slicing.

[0066] In step S6, after dynamic optimization, the priority of computing resource allocation and alarm push strategy of edge nodes are adjusted based on historical operation and maintenance response time and resource consumption data.

[0067] Specifically as follows:

[0068] Classification, labeling, and transmission optimization:

[0069] 1. Data classification rules

[0070] Metering data: voltage, current, power factor, marked as Class M, with storage precision retaining 3 decimal places.

[0071] Communication status data: signal strength, packet loss rate, marked as Class C, with an additional communication protocol type label (e.g., HPLC / LoRa).

[0072] Environmental data: temperature and humidity, categorized as E, obtained from the meteorological bureau's API to retrieve regional baseline values.

[0073] 2. 5G Network Slicing Strategy

[0074] High-priority channel: Dedicated slices with bandwidth ≥20Mbps are allocated for abnormal data, with a transmission latency ≤50ms.

[0075] Dynamic bandwidth adjustment: When abnormal data volume is detected to increase by more than 200% for three consecutive cycles, the slice bandwidth is automatically expanded to 50Mbps.

[0076] Based on data type classification and differentiated 5G transmission strategies, a precise matching mechanism for "data value - network resources" was constructed. By separating sudden abnormal data from periodic data, priority is given to ensuring low-latency transmission of critical data, solving the transmission congestion problem caused by data mixing in traditional methods. At the same time, by adding hierarchical labels, data traceability is provided for subsequent analysis, achieving end-to-end interpretability from raw data to decision instructions.

[0077] The dynamic adjustment of computing resource allocation in step S1 includes:

[0078] Cellular grid partitions are generated based on the meter distribution density, and the computational resource weights for each partition are allocated according to the following rules:

[0079] Resource weight = Number of electricity meters / Zone area × Preset coefficient α + Real-time load rate × Preset coefficient β;

[0080] When the difference in resource weight between adjacent partitions exceeds a threshold, a resource migration task is triggered.

[0081] Specifically as follows:

[0082] Cellular mesh partitioning operation process:

[0083] Mesh generation algorithm: A cellular mesh is generated based on Delaunay triangulation, ensuring that the meter density difference within each mesh is ≤10%. Computational resources are dynamically adjusted according to a weighted allocation formula, and the preset coefficients α and β are determined through regression analysis of historical load data.

[0084] Resource migration trigger condition: When the difference in resource weight between adjacent partitions exceeds a threshold (e.g., 15%), perform the following operations: migrate non-real-time tasks (e.g., historical data analysis) of high-weight partitions to low-weight partitions, update the task queue priority of edge nodes, and prioritize real-time anomaly detection tasks.

[0085] A dynamic resource allocation method employing cellular grid partitioning and weight calculation formulas deeply integrates geospatial distribution characteristics with real-time computing needs, overcoming the inherent contradiction of "insufficient resources in hotspot areas and idle resources in peripheral areas" in static resource allocation. Through a triggered resource migration mechanism, dynamic balancing of computing load across partitions is achieved, forming a resource scheduling paradigm of "partition autonomy and global linkage".

[0086] The data classification labels mentioned in step S2 include:

[0087] The operational data is divided into metering data, communication status data, and environmental data, and a timestamp and hierarchical label are attached to each type of data. The hierarchical label is generated based on the hierarchy of the edge node where the meter is located.

[0088] The logic for generating hierarchical tags is as follows:

[0089] Hierarchical definition:

[0090] L1: Substation level, covering all electricity meters within a 5-kilometer radius;

[0091] L2: Concentrator level, covering a single distribution area (approximately 200 meters);

[0092] L3: Meter level, data for a single metering point;

[0093] Tag appending rules: Each data packet is appended with a triple tag: <level, timestamp, data type>; timestamp synchronization uses the NTP protocol with an error ≤1ms.

[0094] By combining hierarchical labels and timestamps, a strong correlation is established between data and physical location and time dimension. This design breaks through the problem of spatiotemporal information separation in traditional data acquisition, enabling subsequent analysis to combine the evolution of power grid topology to mine spatiotemporal correlations, providing a structured data foundation for fault tracing and impact propagation analysis.

[0095] The multimodal compression strategy mentioned in step S3 includes:

[0096] The periodic monitoring data is subjected to differential encoding compression, the base value sequence is preserved and the difference is stored; the abnormal data is extracted with feature summary, which includes peak value, duration and change gradient.

[0097] The multimodal compression operation steps are as follows:

[0098] Periodic data difference coding:

[0099] Base value selection rule: The first sample value is taken every 10 minutes as the base value, and subsequent data are stored as the relative difference between the base value and the data.

[0100] Compression algorithm: Run-length encoding (RLE) is used to compress consecutive zero differences, achieving a compression rate of up to 85%.

[0101] Anomaly data feature extraction:

[0102] Peak detection: Data exceeding 3σ within the sliding window is marked as a peak.

[0103] Duration calculation: The time span from the first time the threshold is exceeded to the time when the threshold falls back below it.

[0104] Gradient analysis: The rate of change is calculated using the first-order difference, formula:

[0105]

[0106] Where Δy is the difference between data points at adjacent time points, i.e., y t+1 -yt It reflects the amount of data change, where Δt is the time interval (unit: seconds), which is determined by the sampling period (e.g., 30 seconds).

[0107] y t Let y be the monitoring data value (such as voltage, current) at time t. t+1 This is the monitoring data value for the next moment.

[0108] The proposed multimodal compression strategy (differential encoding + feature summarization) achieves an optimal balance between data fidelity and transmission efficiency through a differentiated data value extraction mechanism. It addresses the dilemma of "fidelity loss" or "over-compression" in power grid scenarios by preserving trend features of periodic data and extracting transient features of sudden anomalies, thus preserving the integrity of key information for subsequent analysis.

[0109] The dynamic calculation of the sliding window in step S4 includes:

[0110] Set the window length to N sampling periods and the sliding step size to M periods, where N>M and M is an adjustable parameter;

[0111] After eliminating outliers using the median substitution method within the window, the mean μ and standard deviation σ are calculated, and the dynamic threshold is set to μ±3σ.

[0112] The fusion of the LSTM time series prediction model and the random forest classification model includes: weighting and summing the time series anomaly probability output by LSTM and the discrete event classification result output by random forest, with the weight coefficients dynamically adjusted according to the historical fault types of the electricity meter, and satisfying that the sum of the weight coefficients is 1.

[0113] Specifically as follows:

[0114] Dynamic thresholding and model fusion:

[0115] Sliding window configuration: Window length N = 60 sampling points (corresponding to 1 hour of data), sliding step M = 15 points (15 minutes);

[0116] Outlier removal: Tukey's fences method is used, defining outliers as data points that fall outside the range [Q1-1.5IQR, Q3+1.5IQR].

[0117] Model fusion mechanism:

[0118] LSTM network: 60 time steps in the input layer, 32 units in the hidden layer, and output predictions for the next 5 time steps;

[0119] Random Forest: Input features include 10 dimensions such as mean, variance, and gradient, generating a four-class classification result (normal / minor anomaly / serious anomaly / fault).

[0120] The weighted fusion formula is: Overall anomaly probability = W LSTM ×P LSTM ×W RF ×P RF W LSTM W represents the weighting coefficients of the LSTM model (e.g., 0.7 by default for residential electricity meters). RF P represents the weighting coefficients of the random forest model (e.g., 0.6 by default for industrial electricity meters). LSTM P represents the timing anomaly probability (0-1) of the LSTM output. RF The outlier probability (0-1) output by the random forest is used, and the weighting coefficients are dynamically adjusted according to the meter type: W for residential meters. LSTM =0.7, industrial electricity meter W RF =0.6.

[0121] A hybrid detection model integrating sliding window statistics, LSTM time series prediction, and random forest classification is proposed. Through dynamic thresholding and multi-model weighted fusion mechanism, it jointly models the continuous variation characteristics and discrete event features of power grid data. This design overcomes the limitations of single detection models in adapting to complex operating conditions and significantly improves the ability to detect the combined effects of gradual anomalies (such as equipment aging) and abrupt faults (such as short circuits).

[0122] The digital twin model mapping in step S5 includes:

[0123] Analyze the line impedance and power supply radius in the distribution network topology parameters, and construct the correlation matrix between the electricity meter and adjacent nodes;

[0124] When an anomaly is detected, the fault propagation path is calculated based on the correlation matrix, and a list of affected meter IDs is generated.

[0125] Digital twin modeling process:

[0126] Construction of the association matrix:

[0127] Matrix element a ij The formula for calculating the electrical connection strength between meters i and j is as follows:

[0128]

[0129] Among them, Z ij S represents the line impedance from meter i to meter j, reflecting the electrical distance. j Where S is the downstream load capacity, k is the attenuation coefficient (default 0.1), and S is the load capacity. 总 This represents the total load capacity of the distribution network.

[0130] Fault propagation calculation:

[0131] An improved breadth-first search (BFS) algorithm is used to traverse the neighboring nodes of the abnormal meters in the association matrix.

[0132] Formula for defining the scope of influence:

[0133]

[0134] Where n is the number of meters affected by the fault, a ij Let I be an element of the correlation matrix between faulty meter j and meter i. i I represents the actual current value of meter i. 额定 The rated current threshold of meter i is used. When the impact range is >0.8, it is determined to be a critical fault.

[0135] A digital twin correlation matrix constructed based on distribution network topology parameters maps physical meters to a virtual node network, quantifying the fault propagation relationship between nodes through electrical connection strength. This design overcomes the limitations of traditional isolated node analysis, upgrading fault impact assessment from "single-point judgment" to "networked extrapolation," providing a topological basis for accurately locating fault sources and predicting cascading risks.

[0136] The logic for generating the decision instructions includes:

[0137] If the fault propagation path covers critical load nodes, the alarm priority is increased and cross-regional resource allocation is triggered; otherwise, a local isolation command is generated and the list of meters to be inspected is marked.

[0138] Among them, the tiered handling strategy:

[0139] Critical node judgment conditions:

[0140] Hospitals, data centers, and other similar locations are marked as critical nodes;

[0141] If the fault propagation path covers any critical node, perform the following operations: Initiate cross-regional resource allocation: call backup power from adjacent substations, push a Level 1 alarm to the dispatch center, and the response time requirement is ≤2 minutes.

[0142] Local isolation procedure:

[0143] Generate a sequence of switch operation instructions: first disconnect the upstream circuit breaker, then close the standby tie switch;

[0144] The sorting rule for generating the list of items to be inspected is: sorted from high to low according to the influence coefficient in the correlation matrix.

[0145] By coupling the identification of critical load nodes with the analysis of fault propagation paths, a decision generation logic based on "impact range - handling priority" was constructed. This design integrates the requirements of power grid operation safety and power supply reliability into the decision rules, realizing a strategy upgrade from simple fault repair to power supply guarantee optimization, and forming a hierarchical response mechanism of "rapid local isolation and global risk prevention".

[0146] The dynamic optimization of the reinforcement learning algorithm in step S6 includes:

[0147] The state space is defined as a multi-dimensional vector of edge node resource utilization, fault impact level, and historical response time.

[0148] The action space includes alarm push frequency classification and computing resource migration path selection;

[0149] The reward function is calculated based on the difference between the fault recovery time reduction rate and the resource consumption cost, and the action selection rules are updated through a Q-learning strategy.

[0150] Algorithm implementation details:

[0151] State space definition:

[0152] Three-dimensional vectors include:

[0153] Edge node CPU utilization (normalized to [0,1])

[0154] Fault impact level (0-10)

[0155] Historical average response time (normalized in minutes)

[0156] Motion space design:

[0157] Alarm push frequency: Divided into three levels (low frequency / medium frequency / high frequency), corresponding to 1 / 3 / 5 pushes per hour.

[0158] Resource migration path selection: A graph-based shortest path algorithm generates three candidate paths.

[0159] Reward function calculation:

[0160]

[0161] Among them, T 实际 T represents the actual fault recovery time (in minutes). 基准 The historical average recovery time for similar faults (in minutes).

[0162] C 资源 The computational cost (normalized value, 0-1) incurred by resource migration / allocation, where λ = 0.5 is the resource cost coefficient used to balance the weights of recovery time and resource consumption, and T基准 Take the historical average recovery time for similar faults.

[0163] Based on a reinforcement learning-based dynamic optimization framework, discrete decision-making problems such as resource allocation and alarm strategies are transformed into policy iteration processes in a continuous state space. By defining multi-dimensional state vectors and composite reward functions, the problem of policy rigidity in traditional rule engines in complex power grid scenarios is solved, enabling the system to autonomously evolve its decision logic from historical operation and maintenance experience.

[0164] A power distribution network management meter status analysis system, comprising:

[0165] The edge resource dynamic allocation module is used to dynamically adjust the allocation of computing resources for multi-layer edge nodes based on the distribution density of electricity meters and real-time load, and generate resource migration instructions.

[0166] The multi-source data collaborative acquisition module is used to acquire the operating data of the electricity meter in real time and classify and label the data, including adding timestamps and hierarchical labels;

[0167] The multimodal compression transmission module is used to perform differential encoding compression on periodic monitoring data, extract feature summaries from abnormal data, and allocate high-priority transmission channels through 5G network slicing.

[0168] The streaming anomaly detection module is used to dynamically calculate the mean and standard deviation thresholds based on a sliding window, and to generate anomaly detection results by fusing an LSTM time series prediction model and a random forest classification model.

[0169] The digital twin decision generation module is used to call the digital twin model to map the status of faulty meters and generate decision instructions that include the scope of impact and handling strategies by combining the distribution network topology.

[0170] The reinforcement learning optimization module is used to dynamically optimize the priority of computing resource allocation and alarm push strategy by defining the state space, action space and reward function.

[0171] Inter-module interaction process:

[0172] Edge resource dynamic allocation module:

[0173] Input: Electricity meter GIS coordinates, real-time load monitoring data

[0174] Output: Generates resource migration instructions in JSON format, including the target node IP and a list of migration tasks.

[0175] Digital twin decision-making module:

[0176] The topology database is invoked to generate the execution sequence of decision instructions.

[0177] Python

[0178] If the scope of influence > threshold:

[0179] Execute cross-regional resource allocation()

[0180] send_alert(level='CRITICAL')

[0181] else:

[0182] Generate local isolation scheme()

[0183] Push maintenance work orders()

[0184] In summary, the edge collaboration mechanism distributes computing tasks to nodes closer to the data source through hierarchical edge node deployment, and reduces cloud processing latency by combining dynamic resource migration strategies.

[0185] Data hierarchical processing: Classification labeling and multimodal compression strategies enable priority transmission of critical data (abnormal information), ensuring the real-time nature of fault detection.

[0186] Hybrid anomaly detection: It integrates time-series prediction and classification models, and adapts to different scenarios through dynamic weight adjustment, improving detection accuracy to over 98%.

[0187] Intelligent decision-making closed loop: The digital twin model maps the physical power grid as a virtual mirror, and combined with continuous optimization through reinforcement learning, it forms a closed-loop control of "detection-decision-optimization".

[0188] The above-described embodiments are detailed and specific, illustrating preferred embodiments of the present invention. They are only used to illustrate the technical ideas and features of the present invention, with the aim of enabling those skilled in the art to understand the content of the present invention and implement it accordingly. However, they are not limited to the present invention, and the patent scope of the present invention cannot be limited by this embodiment alone. That is, any equivalent changes or modifications made to the spirit disclosed in the present invention, without departing from the structure of the present invention, such as local improvements within the system and modifications or transformations between subsystems, are still within the patent scope of the present invention.

Claims

1. A method for analyzing the status of electricity meters in distribution network management, characterized in that, Includes the following steps: S1. Deploy multi-layer edge computing nodes in the power distribution network, where primary nodes are deployed in substations and secondary nodes are deployed in meter concentrators. The allocation of computing resources for edge nodes is dynamically adjusted based on meter distribution density and real-time load. S2. Real-time acquisition of meter operation data through a collaborative acquisition mechanism between primary and secondary nodes; S3. For the classified and labeled data, a multimodal compression strategy is used for transmission: differential encoding compression is used for periodic monitoring data, and feature summary extraction is used for sudden abnormal data. S4. Build a streaming processing engine in the cloud, dynamically calculate the mean and standard deviation thresholds of the meter data based on the sliding window, and integrate the LSTM time series prediction model and the random forest classification model to generate multi-dimensional anomaly detection results. S5. Based on the anomaly detection results, call the digital twin model to map the real-time status of the faulty meter, and combine the distribution network topology to generate decision instructions that include fault probability, impact range and recommended handling strategies. S6. Dynamically optimize decision instructions based on reinforcement learning algorithms.

2. The method for analyzing the status of electricity meters in distribution network management according to claim 1, characterized in that, In step S2, the operating data includes voltage, current, power factor, communication signal strength, and ambient temperature and humidity, and the data is classified and labeled. In step S3, the sudden abnormal data is extracted using feature summaries and then high-priority transmission channels are divided using 5G network slicing. In step S6, after dynamic optimization, the priority of computing resource allocation and alarm push strategy of edge nodes are adjusted based on historical operation and maintenance response time and resource consumption data.

3. The method for analyzing the status of electricity meters in distribution network management according to claim 2, characterized in that, The dynamic adjustment of computing resource allocation in step S1 includes: Cellular grid partitions are generated based on the meter distribution density, and the computational resource weights for each partition are allocated according to the following rules: Resource weight = Number of meters / Zone area × Preset coefficient α + Real-time load rate × Preset coefficient β; When the difference in resource weight between adjacent partitions exceeds a threshold, a resource migration task is triggered.

4. The method for analyzing the status of electricity meters in distribution network management according to claim 2, characterized in that, The data classification labels mentioned in step S2 include: The operational data is divided into metering data, communication status data, and environmental data, and a timestamp and hierarchical label are attached to each type of data. The hierarchical label is generated based on the hierarchy of the edge node where the meter is located.

5. The method for analyzing the status of electricity meters in distribution network management according to claim 2, characterized in that, The multimodal compression strategy mentioned in step S3 includes: The periodic monitoring data is subjected to differential encoding compression, the base value sequence is preserved and the difference is stored; the abnormal data is extracted with feature summary, which includes peak value, duration and change gradient.

6. The method for analyzing the status of electricity meters in distribution network management according to claim 2, characterized in that, The dynamic calculation of the sliding window in step S4 includes: Set the window length to N sampling periods and the sliding step size to M periods, where N>M and M is an adjustable parameter; After eliminating outliers using the median substitution method within the window, the mean μ and standard deviation σ are calculated, and the dynamic threshold is set to μ±3σ. The fusion of the LSTM time series prediction model and the random forest classification model includes: weighting and summing the time series anomaly probability output by LSTM and the discrete event classification result output by random forest, with the weight coefficients dynamically adjusted according to the historical fault types of the electricity meter, and satisfying that the sum of the weight coefficients is 1.

7. The method for analyzing the status of electricity meters in distribution network management according to claim 2, characterized in that, The digital twin model mapping in step S5 includes: Analyze the line impedance and power supply radius in the distribution network topology parameters, and construct the correlation matrix between the electricity meter and adjacent nodes; When an anomaly is detected, the fault propagation path is calculated based on the correlation matrix, and a list of affected meter IDs is generated.

8. The method for analyzing the status of electricity meters in distribution network management according to claim 7, characterized in that, The logic for generating the decision instructions includes: If the fault propagation path covers critical load nodes, the alarm priority is increased and cross-regional resource allocation is triggered; otherwise, a local isolation command is generated and the list of meters to be inspected is marked.

9. The method for analyzing the status of electricity meters in distribution network management according to claim 2, characterized in that, The dynamic optimization of the reinforcement learning algorithm in step S6 includes: The state space is defined as a multi-dimensional vector of edge node resource utilization, fault impact level, and historical response time. The action space includes alarm push frequency classification and computing resource migration path selection; The reward function is calculated based on the difference between the fault recovery time reduction rate and the resource consumption cost, and the action selection rules are updated through a Q-learning strategy.

10. A distribution network management meter status analysis system, used to implement the distribution network management meter status analysis method according to claim 2, characterized in that, include: The edge resource dynamic allocation module is used to dynamically adjust the allocation of computing resources for multi-layer edge nodes based on the distribution density of electricity meters and real-time load, and generate resource migration instructions. The multi-source data collaborative acquisition module is used to acquire the operating data of the electricity meter in real time and classify and label the data, including adding timestamps and hierarchical labels; The multimodal compression transmission module is used to perform differential encoding compression on periodic monitoring data, extract feature summaries from abnormal data, and allocate high-priority transmission channels through 5G network slicing. The streaming anomaly detection module is used to dynamically calculate the mean and standard deviation thresholds based on a sliding window, and to generate anomaly detection results by fusing an LSTM time series prediction model and a random forest classification model. The digital twin decision generation module is used to call the digital twin model to map the status of faulty meters and generate decision instructions that include the scope of impact and handling strategies by combining the distribution network topology. The reinforcement learning optimization module is used to dynamically optimize the priority of computing resource allocation and alarm push strategy by defining the state space, action space and reward function.

Images