Data collection modeling method for electricity consumption information collection

Abstract

This invention discloses a data acquisition modeling method for electricity consumption information collection. This method preprocesses historical meter collection records and grid load data to generate standardized meter response characteristic data. A six-dimensional networking behavior feature vector is constructed, and dynamic meter partitioning results are generated using a density clustering algorithm. A partition-level resource allocation strategy is generated by combining meter metadata with a data value assessment model. A two-layer time-wheel scheduling mechanism is constructed, with the outer layer managing inter-partition resource allocation and the inner layer controlling meter decrement time slice allocation, generating an optimized collection and scheduling execution plan. This method addresses the issues of traditional methods that ignore grid topology constraints and data value differences, significantly improving data acquisition success rates and resource utilization efficiency, and providing an intelligent technical solution for power information collection systems.

Description

Technical Field

[0001] The present invention relates to power electronics technology, and in particular to a data acquisition modeling method for electricity consumption information acquisition. Background Art

[0002] With the advancement of smart grid construction and the continuous improvement of power market reform, electricity consumption information collection systems have become core infrastructure for power companies to achieve refined management, improve service quality, and ensure the safe and stable operation of the power grid. By collecting real-time data from distributed smart meters, this system provides accurate and timely data support for power dispatch, load forecasting, electricity cost calculation, demand response, and other services. However, faced with the growing number of meters, complex and changing network environments, and diverse business needs, traditional data collection methods face severe challenges in efficiency, success rate, and resource utilization. Establishing an efficient and intelligent data collection modeling method has important theoretical and practical significance for improving the overall operational efficiency of the power system, reducing operating costs, and enhancing the intelligence level of the power grid.

[0003] Current electricity consumption information collection and modeling methods primarily employ static scheduling strategies based on fixed time windows, collecting data from meters in batches over preset time periods. Existing research focuses on building time series prediction models, leveraging historical response time data to develop ARIMA, neural network, or support vector machine models to predict meter response performance and formulate collection strategies based on the predicted results. Some research has introduced multi-period adaptive mechanisms to adjust collection parameters based on network load characteristics in different time periods. Regarding resource allocation, existing methods primarily statically group meters based on their geographic location or transformer ownership, employing polling or priority queues for scheduling management. Some advanced solutions have begun to consider the statistical characteristics of meter response times, using cluster analysis to group meters with similar response characteristics. However, these clustering features are relatively simple, primarily based on basic statistics such as the mean and variance of historical response times.

[0004] Existing technical solutions have significant deficiencies in dealing with the complexity and dynamics of meter networking behavior. For example, traditional clustering methods fail to fully consider the inherent relationship between meters and the grid topology, resulting in grouping meters that are electrically far apart but have similar behaviors. This violates the physical constraints of the power system, causing unnecessary network overhead and potential system instability. In addition, existing resource allocation strategies lack a deep understanding of the value differences in meter data, and fail to establish a comprehensive evaluation model for data timeliness, business importance, and data integrity. When resources are limited, it is impossible to achieve value-maximizing collection decisions, resulting in the possibility that key business data may be missing or delayed due to improper resource allocation. These two core technical issues directly affect the overall performance and reliability of the electricity information collection system. Summary of the Invention

[0005] The purpose of the invention is to provide a data collection modeling method for electricity consumption information collection, in order to solve at least one technical problem existing in the prior art.

[0006] Technical solution: A data collection modeling method for electricity consumption information collection, including:

[0007] Read historical meter collection records and grid load data, pre-process, and generate standardized meter response characteristic data

[0008] And perform cluster analysis to generate dynamic partition results of electric meters;

[0009] Obtain meter metadata from the business system, combine standardized meter response feature data with meter dynamic partitioning results, and generate a partition-level resource allocation strategy.

[0010] Based on the partition-level resource allocation strategy, a two-layer time wheel scheduling method is adopted. The outer time wheel manages the resource allocation between partitions and the inner time wheel controls the allocation of meter decreasing time slices within the partition to generate an optimized collection scheduling execution plan.

[0011] Furthermore, an optimized collection scheduling execution plan is generated, including:

[0012] Based on the partition-level resource allocation strategy, an outer time wheel data structure is constructed, the acquisition window is divided into multiple time slices, and the scheduling order is set according to the partition priority to generate an outer time wheel scheduling table;

[0013] Based on the pre-calculated meter-level priority index, an inner time wheel data structure is configured for each meter in the partition, and a meter scheduling sequence is generated within the partition. Each meter in the partition is assigned a decreasing time slice. Meters with higher acquisition success rates receive smaller time slices.

[0014] The outer time wheel scheduling table, the meter scheduling sequence within the partition, and the pre-configured retry strategy table are integrated to generate the collection scheduling execution plan.

[0015] Furthermore, a decreasing time slice is allocated to each meter in the meter scheduling sequence within the partition, including:

[0016] Read the meter scheduling sequence within the partition, extract the historical RTT sequence and success rate sequence,

[0017] Combined with the real-time RTT value, the RTT change rate, RTT variance mutation index, and success rate decline index are calculated. A network anomaly score is generated through a multi-dimensional network anomaly score fusion model. The network stability index is calculated and the anomaly detection threshold is adjusted accordingly.

[0018] Abnormal state meters are identified based on network anomaly scores and anomaly detection thresholds. Time slice quality adjustment factors are calculated for stable mode meters, and real-time quality adjustment factors are calculated for response mode meters based on real-time network quality. Time slice allocation values ​​are generated based on this and boundary constraints are processed.

[0019] Furthermore, standardized meter response characteristic data is generated, including:

[0020] Perform time range screening and outlier processing on historical meter collection records to generate cleaned historical data;

[0021] Perform correlation analysis on the cleaned historical data and grid load data, calculate the correlation coefficient between the response performance of each meter and the grid load, and generate meter-load correlation data;

[0022] The period performance statistics of the cleaned historical data and the meter-load correlation data are integrated to generate standardized meter response characteristic data.

[0023] Furthermore, correlation analysis is performed on the cleaned historical data and grid load data, including:

[0024] Obtain multi-level grid loads, including transformer-level loads, feeder-level loads, and regional total loads; perform time series decomposition on these loads along with meter response time series from cleaned historical data, calculate the delayed cross-correlation coefficients of each component, and generate component-level correlation features;

[0025] The load correlation of each level is weighted according to the number of electrical hops from the meter to the transformer to generate the topological weighted correlation;

[0026] Perform frequency domain transformation on the cleaned historical data and multi-level power grid loads, calculate the coherence spectrum, and generate frequency domain correlation features;

[0027] Calculate the correlation changes between meter response and grid load in different time windows, detect correlation mutation points, and generate correlation time-varying features;

[0028] The above outputs are weighted and integrated, including component-level correlation features, topology-weighted correlation, frequency-domain correlation features, and correlation time-varying features, to generate a meter-load correlation index.

[0029] Furthermore, the dynamic partitioning results of the electric meter are generated, including:

[0030] Based on the standardized meter response characteristic data, a networking behavior feature vector including average response time, response time standard deviation, short-term success rate, long-term success rate, stability index, and load correlation is constructed and normalized to generate a normalized feature vector.

[0031] The Mahalanobis distance between each pair of electricity meters is calculated based on the normalized eigenvectors to generate the electricity meter behavior similarity matrix;

[0032] An improved density clustering algorithm is used to analyze the similarity matrix of electric meter behavior and obtain clustering results.

[0033] Generate dynamic meter partitioning results based on clustering results, including partition ID, each partition meter set and partition center point features.

[0034] Furthermore, the improved density clustering algorithm is a topology-constrained hierarchical clustering algorithm:

[0035] Read the original grid topology data from the grid management system, construct the grid topology structure, calculate the electrical hop count matrix between meters, identify transformer nodes of different voltage levels as topology anchor points, and generate a set of topology constraint cluster centers;

[0036] Based on the meter behavior similarity matrix and the electrical distance matrix, a topology constraint similarity matrix is ​​generated through electrical distance weighted similarity reconstruction.

[0037] Based on the topologically constrained cluster center set and the topologically constrained similarity matrix, coarse-grained clustering is performed at the first-level anchor point level, and then fine-grained partitioning is performed within the cluster. Topological connectivity verification is implemented, and topological connectivity clustering results are generated and dynamic boundary optimization is performed to generate stabilized partitioning results and cluster output.

[0038] Furthermore, the electrical distance weighted similarity reconstruction includes:

[0039] Extracting the behavioral similarity and electrical distance of each pair of meters from the meter behavioral similarity matrix and electrical distance matrix, constructing a dual-constrained input data set, and multiplying the behavioral similarity by the exponential decay factor of the electrical distance to generate a topological weighted similarity matrix;

[0040] When the meters belong to different voltage levels, the level penalty factor is used to modify the meter behavior similarity matrix to generate a level-constrained similarity matrix.

[0041] The topologically weighted similarity matrix and the level-constrained similarity matrix are integrated to generate a topologically constrained similarity matrix.

[0042] Furthermore, the networking behavior characteristic vector includes: average response time, response time standard deviation, short-term success rate, long-term success rate, stability index, and correlation coefficient with grid load.

[0043] Furthermore, a partition-level resource allocation strategy is generated, including:

[0044] Obtain meter metadata from the business system and calculate the basic business value score based on user type and electricity consumption;

[0045] Calculate the data timeliness value based on the collection time interval in the standardized meter response characteristic data, and calculate the data gap value based on the collection success records;

[0046] Integrate the basic business value score, data timeliness value, and data gap value, and calculate the data value index of each meter through weighted summation;

[0047] Based on the dynamic partitioning results of electricity meters and the data value indicators of each meter, the total value of the partition and the historical success rate of the partition are calculated, and the partition resource allocation ratio is calculated accordingly. The partition-level resource allocation strategy is generated in combination with the total available collection time.

[0048] Furthermore, data value indicators are calculated, including:

[0049] Extract business value basic points, data timeliness value, and data gap value from historical data, normalize them, and generate three types of normalized value indicators; and adjust the indicator values ​​according to the current system status to generate dynamic weight coefficients;

[0050] The dynamic weight coefficient is used to weight the sum of the three normalized value indicators to obtain the basic value score, and the interaction enhancement items between the value indicators are added to obtain the interactive value score;

[0051] Set compensation coefficients, importance coefficients, and gain coefficients for meters that continuously fail to collect data, meters for key users, and meters in designated monitoring states, and calculate situational adjustment coefficients.

[0052] Integrate the basic value score, interaction value score and situational adjustment coefficient to calculate the data value index.

[0053] Furthermore, the partition resource allocation ratio is calculated, including:

[0054] Based on the total value of the partition, the historical success rate of the partition, the cohesion of the partition and the size of the partition, the partition importance index is comprehensively calculated; according to the global statistical characteristics of the system, the dynamic balance coefficient is generated;

[0055] Calculate the initial resource allocation ratio based on the partition importance index and dynamic balance coefficient; generate the minimum resource quota and maximum resource quota based on the resource upper and lower limit guarantee mechanism;

[0056] Integrate the preliminary resource allocation ratio, minimum resource quota and maximum resource quota to generate the optimized resource allocation ratio that meets the constraints and calculate the partition time quota based on it. Combined with the partition priority sorting, a partition-level resource allocation strategy is formed.

[0057] Beneficial effects: It solves the problem of traditional methods ignoring grid topology constraints and data value differences, significantly improves the data collection success rate and resource utilization efficiency, and provides an intelligent technical solution for the power information collection system. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart of the present invention.

[0059] Figure 2 It is a flow chart of the optimized acquisition scheduling execution plan generated by the present invention.

[0060] Figure 3 This is a flow chart of allocating decreasing time slices to each meter in the meter scheduling sequence within a partition according to the present invention.

[0061] Figure 4 The present invention is a flow chart of generating standardized electric meter response characteristic data. DETAILED DESCRIPTION

[0062] To make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless there is a conflict, the embodiments and features in the embodiments of this application can be combined with each other.

[0063] Embodiment 1: This embodiment provides a data collection modeling method for collecting electricity usage information.

[0064] like Figure 1 As shown, the method is an automated processing flow running on the background server of the power information acquisition system, including the following steps:

[0065] Step S101: read historical electricity meter collection records and grid load data, perform preprocessing, generate standardized electricity meter response feature data, perform cluster analysis on the data, and generate electricity meter dynamic partitioning results.

[0066] Historical meter collection records refer to the raw data obtained from the electricity consumption information collection system database, including meter ID, collection time, response time, success / failure status, etc. Grid load data refers to the total load power data of transformers, feeders, or areas related to the area where the meter is located, obtained from the grid dispatching system or SCADA system. Standardized meter response characteristic data refers to a set of structured data that can comprehensively and comparably describe the response characteristics of each meter after cleaning, statistics, and correlation analysis. Dynamic meter partitioning refers to the process of dividing meters with similar networking behavior characteristics and similar electrical topological locations into the same group (i.e., partition). The partitioning results will be dynamically updated based on changes in meter behavior.

[0067] In this embodiment, the database interface is used to first retrieve the last 30 days of meter data and the corresponding grid load data for the corresponding time period. Next, the response times in the data are processed for outliers, for example, using a modified Z-score method to identify outliers and replacing them with the median to generate cleaned data. Then, based on the cleaned data and load data, the correlation coefficient between each meter's response performance (e.g., average response time, success rate) and grid load is calculated. Combined with time period performance statistics, these data are finally integrated into standardized meter response feature data. Finally, based on this feature data, a density clustering algorithm (such as a topology-constrained hierarchical clustering algorithm) is applied to cluster all meters, resulting in dynamic partitioning results.

[0068] This step integrates historical and load data and conducts in-depth analysis to construct characteristic data that reflects the inherent operating patterns of the meter, laying the foundation for subsequent scientific partitioning. Dynamic partitioning achieved through cluster analysis overcomes the drawback of traditional static grouping, which cannot adapt to changes in the network environment.

[0069] Step S102: Obtaining meter metadata from the business system, combining the standardized meter response characteristic data and the meter dynamic partitioning result, and generating a partition-level resource allocation strategy.

[0070] Meter metadata refers to data that describes the service attributes of the meter, such as user type (industrial, residential), electricity consumption, and billing level. The partition-level resource allocation strategy is a decision-making scheme that assigns specific collection time quotas and priorities to each dynamic partition.

[0071] This step obtains meter metadata from the power marketing system or customer relationship management system (CRM). A comprehensive data value metric is then calculated for each meter, combining this metadata (to calculate business value), standardized meter response characteristic data (to calculate data timeliness and gap value), and a data value assessment model. Next, based on the dynamic meter partitioning results generated in step S101, the total value and historical success rate of each partition are summarized. Finally, using a value-success rate balance model, the resource allocation ratio for each partition is calculated. This is then combined with the total available collection time to form the final partition-level resource allocation strategy.

[0072] This step introduces a differentiated assessment of data value, allowing limited collection resources to be tilted towards high-value meters and partitions, solving the problem of traditional methods that average effort and fail to achieve value-maximizing collection.

[0073] Step S103: Based on the partition-level resource allocation strategy, a double-layer time wheel scheduling method is adopted to manage the resource allocation between partitions through the outer time wheel and control the decrement time slice allocation of the electricity meter within the partition through the inner time wheel, thereby generating an optimized collection scheduling execution plan.

[0074] In a two-tiered time wheel scheduling system, the slots of the outer time wheel correspond to different meter partitions, and the movement of its pointer determines which partition should be collected. The slots of the inner time wheel correspond to the individual meters within a partition, and the movement of its pointer determines the specific collection order and timing of the meters within the partition. Descending time slice allocation means that meters with higher collection success rates are assigned shorter collection timeouts (time slices), and vice versa.

[0075] For example, first, an outer time wheel is constructed based on the partition-level resource allocation strategy generated in step S102. For example, the total collection window (e.g., 2 hours) is divided into multiple time periods based on the resource allocation ratio of each partition, and these time periods are arranged on the time wheel in order of partition priority. Then, for each partition, an inner time wheel is constructed based on the priority index of its internal meters (which combines data value and historical response performance), and each meter is assigned a decreasing time slice. Finally, the outer time wheel's schedule, the inner time wheel scheduling sequence of all partitions, and the preset retry strategy (e.g., exponential backoff) are integrated to generate a detailed final execution plan that includes the specific collection time for each meter.

[0076] The two-tiered time wheel mechanism ensures fair resource allocation and priority across partitions at the macro level, while enabling efficient data collection scheduling within partitions at the micro level. The decreasing time slice allocation strategy saves time for high-quality meters, leaving more time resources for difficult meters, thereby improving overall data collection success rates and resource utilization efficiency.

[0077] Through the above steps, the present invention provides an intelligent data acquisition modeling method, which comprehensively considers the dynamic behavior of the electricity meter, the physical topology of the power grid, the business value of the data, and an efficient scheduling algorithm. It can significantly improve the success rate of data acquisition and resource utilization efficiency, and solve the technical problem that traditional methods ignore the power grid topology constraints and data value differences.

[0078] Embodiment 2 provides a method for generating an optimized acquisition scheduling execution plan and allocating time slices. This embodiment describes in detail how to generate an optimized acquisition scheduling execution plan, especially the core decreasing time slice allocation mechanism.

[0079] In this embodiment, the process of generating the optimized acquisition scheduling execution plan specifically includes:

[0080] Step S201: Based on the partition-level resource allocation strategy, an outer time wheel data structure is constructed, the acquisition window is divided into multiple time slices, and the scheduling order is set according to the partition priority to generate an outer time wheel scheduling table.

[0081] Assume that according to Example 1, a partition-level resource allocation policy has been obtained, which includes the priority ranking and time quotas [1238s, 1080s, 950s, 857s, 900s, 785s, 727s, 663s] for eight partitions (1B, 1A, 2A, 2B, 3, 4, 5, 6). The total acquisition window is 2 hours (7200 seconds). The outer time wheel is constructed as a scheduling list, where each item in the list defines the scheduling window for a partition.

[0082] For example, the first entry is: {zone_id:"1B",priority:18.5,time_start:0,time_end:1238}.

[0083] The second item is: {zone_id:"1A",priority:16.02,time_start:1238,time_end:2318}, and so on. This schedule defines the macro-collection process, ensuring that high-priority zones are collected first and that the resources (and time) they receive are consistent with the policy.

[0084] Step S202: Based on the pre-calculated meter-level priority index, an inner time wheel data structure is configured for each meter in the partition, a meter scheduling sequence in the partition is generated, and a decreasing time slice is allocated to each meter therein.

[0085] The meter-level priority index is calculated by combining the data value index and historical response performance of an individual meter and is used to determine its collection priority within a partition. The decreasing time slice refers to the collection timeout period assigned to a meter. The core concept is that meters with higher historical collection success rates generally have more stable network conditions, so a shorter timeout period can be set for them. Failures can be resolved through retries. Conversely, meters with lower success rates require longer timeout periods to account for potential network delays.

[0086] For example, consider zone 1A, which has a time quota of 1080 seconds and contains 20 meters. First, calculate the priority index P_meter for each meter in the zone. For example, the P_meter for meter M001 is 0.765, and for meter M003 it is 0.742. Then, sort all meters in the zone in descending order based on their P_meters to form a basic scheduling sequence. Next, assign a decreasing time slice to each meter in the sequence, with meters with higher data collection success rates receiving smaller time slices.

[0087] As a preferred embodiment, allocating a decreasing time slice to each meter in the meter scheduling sequence within the partition includes:

[0088] Step S202a: Read the historical RTT (round-trip time) and success rate series of the electricity meters within the zone. Combined with the real-time RTT values, the RTT change rate, RTT variance spike, and success rate drop indicators are calculated. A quantified network anomaly score is generated using a multi-dimensional network anomaly scoring fusion model, such as the weighted summation of Anomaly_score = w1 × RTT_change_rate + w2 × RTT_variance_spike + w3 × success_rate_drop. Simultaneously, a network stability index is calculated, and the Anomaly_threshold_adaptive threshold used to determine anomalies is dynamically adjusted based on this index.

[0089] Step S202b: Based on the network anomaly score and anomaly detection threshold, identify meters currently experiencing anomalies. The system maintains an operating mode identifier (stable mode or response mode) for each meter. For meters in stable mode, a damped smoothing mechanism is used to calculate a time-slice quality adjustment factor, damped_quality_factor, to prevent drastic time-slice changes caused by occasional network jitter. For meters in response mode (i.e., where a network anomaly has been detected), a real-time quality adjustment factor, real_time_quality_factor, is calculated based on real-time network quality (such as queue delay and processing jitter) to enable rapid adaptation to network changes.

[0090] Step S202c: Based on the meter's operating mode, the final time slice allocation is generated by combining the base time slice, success rate, priority, and corresponding adjustment factors. For example, for a stable mode meter, the final time slice allocation value T_final = T_base × (1-δ × SR) × damped_quality_factor. For a responsive mode meter, an emergency multiplier (emergency_multiplier) may also be applied. Finally, the calculated time slice is bounded to ensure that its value is within a preset reasonable range (e.g., no less than 15 seconds and no more than 120 seconds).

[0091] Step S203: Integrate the outer time wheel scheduling table, the meter scheduling sequence within the partition, and the pre-configured retry strategy table to generate a collection scheduling execution plan.

[0092] Integrate the outer scheduling table in step S201 and the inner scheduling sequence generated for each partition in step S202. At the same time, configure a general or hierarchical retry strategy. Preferably, an exponential backoff algorithm is used, that is, the waiting time for each retry increases exponentially, and random jitter is added to avoid conflicts. For example, Delay(retry_n)=T_base×2^(n-1)×(1+random(0,0.1)), where n is the number of retries. The final execution plan is a detailed instruction list for each meter, which clearly defines its initial collection time point, the allocated timeout (time slice), and the retry strategy after failure. For example:

[0093] {meter_id:"M001",start_time:1283.8,duration:32.46,retry_config:[51.4,105.8,201.7],zone_id:"1A"}.

[0094] Through the above steps, the present invention not only plans the macro-level collection sequence and resources, but also achieves refined, adaptive control of each meter's collection time at the micro level. In particular, the introduction of network anomaly detection and a dual-mode time-slice allocation mechanism enables the scheduling strategy to intelligently balance system stability with rapid response to network changes. This addresses the technical issue of traditional fixed or simply decreasing time-slices being unable to adapt to dynamic network environments. This ensures a high success rate while further improving the overall operational efficiency of the system.

[0095] Example 3: This example is used to describe how the basic data used as model input, namely, the standardized meter response characteristic data, is generated, especially the innovative meter-load correlation analysis process.

[0096] Specifically, the process of generating standardized meter response characteristic data includes:

[0097] Step S301: perform time range screening and outlier processing on historical electricity meter collection records to generate cleaned historical data.

[0098] The most recent period (e.g., 30 days) of data collected from the electricity consumption information collection system database is selected as the analysis sample. For the response time series, an adaptive threshold model is used for outlier detection. This model combines the median (RT_median) and interquartile range (RT_IQR) and dynamically adjusts based on the coefficient of variation (CV) of the data. The threshold formula can be set as: Upper_threshold = RT_median + α × RT_IQR × log(1 + CV). For identified outliers, local linear interpolation or replacement with the median of the recent period is used, depending on whether they are sporadic or continuous, to generate a clean and reliable historical dataset.

[0099] Step S302: performing correlation analysis on the cleaned historical data and the grid load data, calculating the correlation coefficient between the response performance of each meter and the grid load, and generating meter-load correlation data.

[0100] As a preferred embodiment, this step specifically includes:

[0101] Step S302a: Obtain multi-level grid loads and perform time series decomposition. Multi-level load data, including transformer-level, feeder-level, and regional total load, is obtained from the grid management system. Then, using time series decomposition techniques (such as STL decomposition), both the meter response time series and the load series at each level are decomposed into trend terms, periodic terms, and residual terms. This approach aims to separate patterns of change at different time scales, facilitating more refined correlation analysis.

[0102] Step S302b: Calculate component-level correlation features.

[0103] For each decomposed component, multi-scale cross-correlation analysis is used to calculate the delay cross-correlation coefficient between them. For example, the correlation between the meter response time trend term and the transformer load trend term within the delay window of -24 hours to +24 hours is calculated, and the maximum correlation and its corresponding time delay are found. This solves the problem that traditional correlation analysis ignores the effect of time delay.

[0104] Step S302c: Generate topologically weighted correlation. Based on the meter's location in the grid topology, calculate the number of electrical hops (Hop_distance) from the meter to each transformer level. Then, nonlinearly weight the correlation of loads at each level using the formula: WC(meter_i, load_j) = Corr(meter_i, load_j) × exp(-k × Hop_distance). k is the attenuation coefficient. This allows loads closer in electrical connection to contribute more to the correlation, reflecting physical constraints.

[0105] Step S302d: Generate frequency-domain correlation features and time-varying features. Perform a Fast Fourier Transform (FFT) on the response time series and the load series, and calculate the coherence spectrum in the frequency domain to identify strong correlation patterns at specific frequencies (e.g., daily or weekly cycles). Simultaneously, a sliding window method is used to calculate the change in correlation over different time windows and detect correlation mutation points to generate time-varying correlation features.

[0106] Step S302e: Weighted integration to generate a meter-load correlation index. Finally, the component-level correlation features, topology weighted correlation, frequency-domain correlation features, and time-varying correlation features generated in the above steps are integrated using a multi-feature fusion model (e.g., weighted summation) to generate a comprehensive meter-load correlation index that fully characterizes the relationship between the meter and the load.

[0107] Step S303: integrating the period performance statistics of the cleaned historical data and the meter-load correlation data to generate standardized meter response characteristic data.

[0108] The period performance statistics (e.g., average response time and success rate by time period) of the cleaned data generated in step S301 are combined with the meter-load correlation index generated in step S302. The resulting standardized meter response feature data is a structured table or dataset, with each row representing a meter and each column representing a feature (e.g., average response time from 0-4 o'clock, load correlation index, etc.). This dataset will serve as the basis for subsequently constructing the network behavior feature vector and performing dynamic partitioning.

[0109] Through the above steps, the present invention can generate a set of high-quality, high-dimensional meter response characteristic data. Compared with existing technologies, its innovation lies in the construction of a load correlation analysis model that considers time delay, topology, frequency domain characteristics, and dynamic changes. This allows the understanding of meter response behavior to be no longer limited to the meter response itself, but to be closely integrated with the macroscopic operating status of the power grid, providing solid data support for subsequent more accurate modeling and decision-making.

[0110] Embodiment 4: This embodiment describes the process of dynamically partitioning electricity meters based on standardized electricity meter response feature data, especially the topology-constrained hierarchical clustering algorithm used therein.

[0111] Step S401: Based on the standardized meter response characteristic data, a networking behavior characteristic vector including average response time, response time standard deviation, short-term success rate, long-term success rate, stability index and load correlation is constructed and normalized.

[0112] The networking behavior feature vector is a multi-dimensional vector, each dimension of which is an indicator extracted or calculated from the standardized meter response feature data and can characterize a certain aspect of the meter network behavior.

[0113] For each electricity meter, a six-dimensional networking behavior feature vector V = [RT_avg, RT_std, SR_short, SR_long, Stab_idx, Load_corr] is constructed. The calculation of each feature is optimized: RT_avg uses time-decay weighting to give more weight to recent data; RT_std integrates fluctuations at multiple time scales; SR_short and SR_long are also calculated using a time-decay function to achieve a smooth transition; Stab_idx evaluates stability from three dimensions: time, fluctuation, and period; and Load_corr is the comprehensive index calculated in Example 3. After constructing the vector, the feature vectors of all meters are normalized using an improved MinMax method to eliminate the influence of different feature dimensions.

[0114] Optionally, the interactive relationship between features can be mined to generate auxiliary combined features such as the response-success rate composite indicator RT_SR = RT_avg × (1-SR_short) to enhance the feature expression capability and add it to the feature vector.

[0115] Step S402: Calculate the Mahalanobis distance between each pair of electricity meters based on the normalized eigenvectors to generate an electricity meter behavior similarity matrix.

[0116] Unlike Euclidean distance, this embodiment uses Mahalanobis distance to calculate the similarity between electricity meters. Mahalanobis distance accounts for correlation between features and is more suitable for processing the high-dimensional correlated features constructed in this invention. Mahalanobis distances are calculated between all two meters and converted to similarities (for example, sim = 1 / (1 + dist_Mahal)), ultimately forming an N×N meter behavior similarity matrix, where N is the total number of meters.

[0117] Step S403: applying an improved density clustering algorithm to analyze the electricity meter behavior similarity matrix to obtain a clustering result.

[0118] As a preferred implementation, this step adopts a topology-constrained hierarchical clustering algorithm, specifically including:

[0119] Step S403a: Analyze the grid topology and pre-set anchor points. Topology data is read from the grid management system, a topology map is constructed, and the electrical hop count matrix between all meters is calculated. Transformer nodes of different voltage levels (e.g., 10kV transformers) are identified as topological anchor points, and the characteristic centroids of their downstream meter groups are calculated as anchor point features. This creates a topologically constrained cluster center set consistent with the physical structure of the grid.

[0120] Step S403b: Perform electrical distance weighted similarity reconstruction. This step is to modify the behavior similarity matrix generated in step S402 so as to incorporate the topology constraints.

[0121] First, the behavior similarity sim ij Multiply by an electrical distance ed ij The exponential decay factor exp(-λ×ed ij / ed_max) to generate a topological weighted similarity matrix. This means that the farther the electrical distance between two meters is, the more similar their behaviors will be.

[0122] Secondly, a level penalty factor is introduced. When two meters belong to different voltage levels, their behavioral similarity will be multiplied by a penalty factor less than 1 to generate a level-constrained similarity matrix.

[0123] Finally, the two modified similarity matrices are fused through adaptive weights to generate the final topology-constrained similarity matrix.

[0124] Step S403c: Perform hierarchical constrained clustering and connectivity verification. Based on the topologically constrained cluster center set and the topologically constrained similarity matrix, coarse-grained clustering is first performed at the first-level anchor point level. Within each coarse-grained cluster, refined partitioning is then performed using reference to the second-level anchor points. Importantly, topological connectivity verification is performed after partitioning to ensure that all meters within each partition are physically connected. Disconnected partitions are repartitioned.

[0125] Step S404: Generate a dynamic partitioning result of the electric meter based on the clustering result, including a partition ID, a set of electric meters in each partition, and a feature of a partition center point.

[0126] The topological connectivity clustering results generated in step S403c are also dynamically optimized. Fuzzy meters located at partition boundaries are identified and redistributed based on their topological connectivity (e.g., whether they are on the same feeder) and partition load balancing considerations. Furthermore, to prevent frequent and drastic changes in partition results, a partition stability enhancement mechanism is introduced to smooth the current partition results against historical results. Finally, a stabilized partition result is output, which includes each partition's ID, a list of meters, and updated partition center features and topological attributes.

[0127] For example, in one run, 200 electricity meters were divided into eight partitions. Partition 1A contained 20 meters, with topological characteristics of an average electrical hop count of 2.4, a connectivity of 0.132, and a partition quality index of 0.282.

[0128] The core advantage of the dynamic partitioning method constructed through the above steps lies in its deep integration of data-driven behavioral similarity with topological constraints driven by physical rules. This overcomes the fundamental problem in existing technologies where clustering results may violate the physical constraints of the power system (for example, classifying two meters that are far apart on the grid into the same category). This ensures the electrical rationality and practical operability of the partitioning scheme, laying a solid foundation for subsequent efficient resource allocation and scheduling.

[0129] Embodiment 5: This embodiment is used to describe how to formulate a reasonable resource allocation strategy for each dynamic partition.

[0130] Step S501: Obtain meter metadata from the business system, and calculate the data value index of each meter in combination with standardized meter response feature data.

[0131] The data value index is used to measure the comprehensive importance and urgency of collecting a certain meter data at the current moment.

[0132] Step S501a: Calculate three types of basic values.

[0133] Business Value Base (BV_base): This score is calculated through weighted calculation based on meter metadata, such as user type (industrial users have a higher weight than residential users) and historical electricity consumption (higher electricity consumption has a higher weight).

[0134] Data Timeliness Value (TV): Calculated based on the time interval since the last successful data collection, using an exponential decay function. The longer the time, the higher the value. The formula is TV = 1 / (1 + λ × (t_current - t_last_successful)). t_current and t_last_successful are the current time and the last successful data collection time, respectively.

[0135] Data Gap Value (GV): Based on the historical data collection success records, the data integrity assessment model is used to calculate the value. The more serious and critical the historical data is missing, the higher the value.

[0136] Step S501b: Perform nonlinear fusion and context adjustment.

[0137] First, the above three types of value indicators are adaptively normalized.

[0138] Then, based on the current business scenario (e.g., whether it is a critical point in the billing cycle), the dynamic weight coefficients (w1, w2, w3) for the three indicators are determined and weighted summed to obtain the base value score V_base. Optionally, an interaction enhancement term V_interact is added between value indicators to reflect the synergistic effect of value.

[0139] Next, a situational reward mechanism is introduced to calculate a situational adjustment factor (SAF) for meters in special situations. For example, a failure compensation factor is set for meters that continuously fail to collect data, a user importance factor is set for key user meters, and a monitoring gain factor is set for meters in special monitoring states.

[0140] Finally, the final comprehensive data value indicator V=(V_base+V_interact)×SAF is integrated and calculated.

[0141] For example, electricity meter M001 is an industrial user and recently experienced a data collection failure, resulting in a high timeliness value. Calculations show that its BV_norm = 0.528, TV_norm = 0.9999, and GV_norm = 0.001. During the billing cycle, the BV weight increases, ultimately resulting in a comprehensive data value indicator, V, of 0.800.

[0142] Step S502: Based on the dynamic partitioning result of the electric meter and the data value index of each electric meter, the total partition value and the historical partition success rate are calculated, and the partition resource allocation ratio is calculated accordingly.

[0143] The details are as follows:

[0144] Step S502a: Calculate the partition importance index. First, sum the data value index V of all meters within the partition to obtain the total partition value V_zone. Then, taking into account the total partition value, partition size (Size), and partition cohesion (Cohesion), calculate a partition importance index Imp_zone. For example, Imp_zone = V_zone × (1 + λ1 × log(Size / Size_avg)) × (1 + λ2 × (Cohesion - Cohesion_avg)). This approach aims to balance the impact of value, size, and partition quality.

[0145] Step S502b: Adaptively calculate resource allocation ratios. Dynamically determine the value-success rate balance coefficients α and β based on the system's global state (e.g., current acquisition load and data value variance). Then, use the formula R_raw(zone_i) = (α × Imp_zone_i + β × SR_zone_i × Size_i) / Σ(...) to calculate the initial resource allocation ratio for each zone.

[0146] Step S502c: Constrained optimization is performed. To prevent low-value partitions from being completely deprived of resources or high-value partitions from occupying excessive resources, a resource floor guarantee and ceiling control mechanism are designed to calculate the minimum resource quota R_min and maximum resource quota R_max for each partition. Finally, a constrained optimization algorithm (such as the Lagrange multiplier method) is applied to adjust the initial resource allocation ratio, subject to the upper and lower constraints and the sum of the constraints being 1, to obtain the final optimized resource allocation ratio R_opt.

[0147] Step S503: Generate a partition-level resource allocation strategy based on the total available acquisition time.

[0148] Multiply the total available collection time (e.g., 7200 seconds) by each partition's optimized resource allocation ratio, R_opt, to obtain the specific time quota, T_zone, for that partition. Furthermore, the partition priority ranking is calculated based on factors such as the partition importance index. Finally, the optimized resource allocation ratio, partition time quota table, and partition priority ranking are integrated to form a complete partition-level resource allocation strategy, which serves as the input for the two-tier time round scheduling.

[0149] For example, partition 1A has an importance index of 15.03 and an optimized resource allocation ratio of 15.0%. It receives a quota of 1080 seconds within a total time of 7200 seconds, and its priority ranks second among the eight partitions.

[0150] Through the above steps, this invention establishes a complete decision-making chain, from single-meter data value assessment to district-level resource macro-control. This approach ensures the purposefulness of resource allocation through a dynamic, multidimensional value assessment model. Through value-success rate balancing and constrained optimization, it achieves scientific and robust resource allocation. This addresses the technical issues of existing technologies, such as a single resource allocation strategy and a lack of value orientation. It enables limited communication and computing resources to be utilized where they are most needed, thereby maximizing the overall benefits of the system.

[0151] Example 6: Describing the preferred implementation of multi-dimensional meter-load correlation analysis

[0152] This sixth embodiment is a more detailed and preferred implementation of the correlation analysis of the cleaned historical data and the grid load data in step S302 in the third embodiment.

[0153] Step S601: Acquire multi-level grid loads and perform time series decomposition.

[0154] Multi-level grid load refers to load data obtained from different levels of the grid (such as transformers, feeders, and regions). It can reflect the local operating status of the grid more precisely than a single load value.

[0155] Specifically, a multidimensional load matrix, including transformer-level load (e.g., T1: 850kW), feeder-level load (e.g., F1: 280kW), and regional total load (2850kW), is obtained from the power grid management system. Then, using time series decomposition techniques (preferably STL decomposition), each meter's response time series and each level of load series are decomposed into trend, seasonal, and residual terms.

[0156] For example, the response time series of meter M001 is decomposed into three subsequences: Trend_RT, Seasonal_RT, and Residual_RT. The output of this step is multiple groups of decomposed time series components, which are used for subsequent refined correlation analysis.

[0157] Step S602: Calculate component-level correlation, topology-weighted correlation, and frequency-domain / time-varying features.

[0158] Calculate component-level correlations. For each component decomposed in step S601, perform multi-scale cross-correlation analysis to calculate the correlation coefficients at different time delays (e.g., -24 hours to +24 hours). Find the maximum correlation and the corresponding time delay. For example, the calculation found that the trend term for the response time of meter M001 and the trend term for the load of transformer T1 have the highest correlation at a delay of -2 hours, reaching 0.82.

[0159] Calculate the topology weighted correlation: Calculate the electrical hop count Hop_distance from the meter to the transformer based on the grid topology, and then perform exponential decay weighting on the original correlation coefficient:

[0160] WC(meter_i, load_j) = Corr(meter_i, load_j) × exp(-k × Hop_distance). For example, if the number of electrical hops from meter M001 to T1 is 3, its topological weighted correlation is corrected to 0.417. Corr represents the correlation coefficient.

[0161] Calculate frequency domain correlation features: Perform fast Fourier transform (FFT) on the original time series and calculate the coherence spectrum in the frequency domain to identify strong coupling relationships within a specific period (such as 24 hours).

[0162] Calculate time-varying correlation features: Use algorithms such as sliding windows and CUSUM to detect changes and mutation points in the correlation between meter response and load over time.

[0163] Step S603: generating a meter-load correlation index by weighted integration.

[0164] A multi-feature fusion model is constructed to fuse the component-level correlation features, topology weighted correlation, frequency domain correlation features, and correlation time-varying features generated in step S602 into a comprehensive meter-load correlation index through weighted integration.

[0165] For example, the final load dependency index of meter M001 is calculated as:

[0166] Load_corr_index = 0.3 × Max_corr + 0.25 × WC + 0.25 × Max_coherence + 0.2 × Corr_variance = 0.64. This index will serve as a key dimension of the networking behavior feature vector.

[0167] Through the above steps, this embodiment details an innovative method for analyzing meter-load correlation. This method fundamentally differs from the traditional, single, static correlation calculation method. By incorporating multi-dimensional perspectives such as time series decomposition, physical topology weighting, and frequency domain analysis, it can more deeply and accurately characterize the inherent connection between meter communication behavior and grid operation status, providing unprecedented high-quality data features for subsequent intelligent zoning and scheduling.

[0168] Example 7: Preferred Implementation of a Hierarchical Clustering Algorithm Describing Topological Constraints

[0169] This embodiment 7 is a more detailed and preferred implementation of the improved density clustering algorithm applied in step S403 of embodiment 4. It describes in detail how to deeply integrate the power grid topology into the clustering process.

[0170] Step S701: Analyze the power grid topology and preset topology anchor points.

[0171] Topological data is read from the power grid management system to construct a topological graph structure containing nodes such as transformers, feeders, and meters. Based on this graph, a multi-source shortest path algorithm (such as BFS) is used to calculate the electrical hop count matrix edij between all meters. Key nodes in the network, such as transformers of 10 kV and above, are then identified as primary topological anchor points. The weighted characteristic centroid of the downstream meter group is calculated as the feature vector of this anchor point, forming an initial set of cluster centers that are strongly correlated with the physical structure.

[0172] Step S702: Perform electrical distance weighted similarity reconstruction.

[0173] Similarity reconstruction refers to the modification of the purely data-driven behavioral similarity matrix so that it compulsorily considers physical topological constraints.

[0174] From the meter behavioral similarity matrix and electrical distance matrix, extract the behavioral similarity simij and electrical distance edij for each pair of meters. Apply an exponential decay function to correct them: Corrected_simij = simij × exp(-λ × edij / ed_max). This step reduces the final similarity of meters that are physically distant, even if they have similar behavioral characteristics. For example, the original behavioral similarity between M001 and M045 is 0.72, but due to an electrical distance of 5, the corrected similarity drops to 0.207.

[0175] When two meters belong to different voltage levels, a level penalty factor Level_penaltyij is applied to further correct their similarity. For example, the similarity between M001 (10kV) and M085 (400V) will be multiplied by a penalty factor of 0.9.

[0176] Finally, the above topologically weighted similarity and level-constrained similarity are fused through adaptive weights α and β to generate the final topologically constrained similarity matrix.

[0177] Step S703: Execute hierarchical constrained clustering and perform dynamic optimization.

[0178] Based on the topological anchor points of step S701 and the topological constraint similarity matrix of step S702, an improved hierarchical clustering algorithm is used to first perform coarse-grained clustering at the high-level anchor point level, and then perform refined partitioning within the clusters.

[0179] For all generated partitions, a topological connectivity check is performed to ensure that all meters within the partition physically form a connected subgraph, and disconnected partitions are re-segmented in a topologically aware manner.

[0180] Identify meters located at the partition boundary with ambiguous ownership, and make final allocations based on their upstream and downstream topological relationships (such as whether they belong to the same feeder) and partition load balancing considerations. A historical weight smoothing mechanism is introduced to enhance the stability of the partition results.

[0181] Through the above steps, this embodiment deeply and fully embeds the physical topology of the power grid into the meter clustering algorithm in a quantifiable and computable manner. This fundamentally addresses the technical problem of existing data-driven clustering methods that ignore the unique constraints of power systems, resulting in unrealistic partitioning results. The final partitioning results are not only cohesive in terms of data behavior, but also reasonable in terms of physical connectivity, providing a high-quality decision-making basis for subsequent resource allocation and network management.

[0182] Example 8: Describing the preferred implementation of comprehensive data value assessment and partition resource allocation strategy

[0183] This eighth embodiment is a more detailed and preferred implementation of steps S501 and S502 in the fifth embodiment, and describes in detail the entire process from the value assessment of a single electricity meter to the formulation of a macro resource strategy.

[0184] Step S801: Calculate the comprehensive data value index.

[0185] The comprehensive data value index V is a quantitative score that dynamically and multi-dimensionally measures the importance and urgency of collecting meter data at the current moment.

[0186] Calculate the basic value of three dimensions: calculate the business value base BV_base based on user type and electricity consumption; calculate the data timeliness value TV by applying the exponential decay function based on the time since the last successful data collection; and evaluate the data gap value GV based on historical missing information.

[0187] After adaptively normalizing the three value indicators, weight coefficients (w1, w2, w3) are dynamically determined based on the current business scenario (e.g., whether it is a billing cycle or whether the data integrity rate has reached an alarm), and a weighted sum is calculated to obtain the base value score V_base. Optionally, to enhance synergy, an interaction enhancement term (V_interact) between value indicators can be added.

[0188] To address special situations, a situational adjustment factor (SAF) is introduced. For example, a failure compensation factor (FC) is assigned to meters that have repeatedly failed to collect data; a user importance factor (UIC) is assigned to meters used by key users, such as hospitals and large factories; and a monitoring gain factor (MBC) is assigned to meters in special monitoring states, such as power protection or load control.

[0189] All components are integrated through the formula V=(V_base+V_interact)×SAF to calculate the final comprehensive data value index of each meter.

[0190] Step S802: Calculate the partition-level resource allocation ratio.

[0191] The value V of all meters within a zone is summed to obtain the total zone value V_zone. The zone's total value, size, and cohesion (derived from cluster quality assessment) are then considered to calculate the zone's importance index Imp_zone. This ensures that zones with high value, large size, and good quality receive higher importance scores.

[0192] The coefficients α and β used to balance value and success rate are dynamically determined based on the overall system operating status. For example, when network resources are tight, the weight of β is increased to prioritize the overall collection success rate; when the value of data varies greatly, the weight of α is increased to prioritize the collection of high-value data.

[0193] Based on the importance index and balance coefficient, a preliminary resource allocation ratio, R_raw, is calculated. Then, a resource floor guarantee mechanism (ensuring each partition has a minimum collection opportunity) and an upper limit control mechanism (preventing a single partition from occupying too many resources) are applied to generate the minimum R_min and maximum R_max quotas. Finally, a constrained optimization algorithm is used to generate the final optimized resource allocation ratio, R_opt, that satisfies all constraints and sums to 1.

[0194] Step S803: Generate a complete partition-level resource allocation strategy.

[0195] Multiply the total available collection time, T_total, by R_opt to obtain the specific time quota, T_zone, for each partition. This is then combined with the partition's importance index and urgency to calculate the partition priority ranking. The resulting strategy includes each partition's resource ratio, time quota, and scheduling order.

[0196] Through the above steps, this embodiment establishes a scientific and intelligent resource allocation decision-making system. Through refined data value assessment and a globally adaptive allocation algorithm, it addresses the blind and rigid resource allocation issues of traditional methods. It ensures that valuable collection resources are precisely allocated to where they can generate the most value, achieving a shift from completing collection tasks to maximizing collection value.

[0197] Example 9: Preferred implementation method for describing adaptive hybrid response time slice allocation and parameter self-learning

[0198] This ninth embodiment is a more detailed and preferred implementation of allocating decreasing time slices to each electricity meter in step S202 of the second embodiment, and incorporates the concept of parameter self-learning optimization.

[0199] Step S901: Construct a network anomaly detection trigger.

[0200] The network anomaly detection trigger is a real-time monitoring module used to determine whether the network communication quality of a single meter has abnormal fluctuations.

[0201] Monitor the RTT (round-trip time) and acquisition success rate of each meter in real time, and calculate three key abnormal indicators: RTT change rate (RTT_change_rate), RTT variance spike (RTT_variance_spike) (the ratio of the current variance to the historical benchmark), and success rate drop (success_rate_drop) (the decrease in short-term success rate relative to the long-term).

[0202] The three aforementioned metrics are combined into a comprehensive anomaly score using a multi-dimensional network anomaly scoring fusion model (e.g., weighted summation). Furthermore, a network stability index (network_stability_index) is calculated based on recent network quality statistics and used to dynamically adjust the anomaly threshold (anomaly_threshold_adaptive). This allows the threshold to be appropriately relaxed when the network is unstable, preventing false positives.

[0203] Step S902: Implement the dual-mode time slice allocation strategy.

[0204] Based on the anomaly_score and anomaly_threshold_adaptive values, the system determines whether the meter is currently in an abnormal state. The system maintains an operating mode flag for each meter. When anomalies are continuously detected, the meter switches to response mode. After the network stabilizes for a period of time, the system switches back to stable mode.

[0205] For meters in stable mode, the time slice adjustment factor is calculated using a damping smoothing mechanism to absorb occasional network jitter and maintain data collection stability. The formula is:

[0206] T_final=T_base×(1-δ×SR)×damped_quality_factor.

[0207] For meters in response mode, the time slice adjustment factor is calculated directly based on real-time network quality (such as queue delay determined by RTT differential analysis) and is multiplied by an emergency multiplier (emergency_multiplier) determined by the severity of the anomaly. This allows the time slice to be rapidly increased to address network deterioration and maximize data collection success. The formula is: T_final = T_base × (1-δ × SR) × real_time_quality_factor × emergency_multiplier.

[0208] Step S903: Final calculation of time slice and parameter self-learning optimization.

[0209] Boundary constraints are applied to T_final calculated in step S902 (maximum / minimum time slice checks), and time resource balancing is performed within the partition to ensure that the total time consumed by the partition does not exceed the quota.

[0210] After the scheduling plan is executed, the system will track the actual collection success rate actual_success_rate of each meter in the new time slice.

[0211] The actual success rate is compared with the model's predicted success rate to form a success rate loss function, success_loss. Then, using optimization algorithms such as gradient descent, key model parameters are automatically fine-tuned based on the loss function, such as the weights w1, w2, and w3 in the anomaly scoring model and the adjustment coefficients α, β, and γ in the time slice calculation formula. The update logic can be expressed as w_new = w_old - learning_rate × d(success_loss) / dw. learning_rate is the learning rate. success_loss is the success rate loss function, and learning_rate is the learning rate that controls the parameter update step size. w includes the weight coefficients w1, w2, and w3 and the adjustment coefficients α, β, and γ.

[0212] Through the above steps, this embodiment proposes an adaptive time-slice allocation mechanism. Through real-time anomaly detection and dual-mode switching, it addresses the fundamental problem that traditional static or simply decreasing time-slices are incapable of adapting to complex and dynamic network environments. More importantly, by introducing a closed loop of success rate feedback and parameter self-learning, the entire scheduling system is equipped with the ability to self-evolve and continuously optimize, ensuring optimal performance over the long term.

[0213] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. A data collection modeling method for electricity consumption information collection, characterized in that: include: Read historical meter collection records and grid load data, pre-process them, generate standardized meter response feature data, perform cluster analysis on them, and generate dynamic meter partitioning results; Obtain meter metadata from the business system, combine standardized meter response feature data with meter dynamic partitioning results, and generate a partition-level resource allocation strategy. Based on the partition-level resource allocation strategy, a two-layer time wheel scheduling method is adopted. The outer time wheel manages the resource allocation between partitions and the inner time wheel controls the allocation of meter decrement time slices within the partition to generate an optimized data collection scheduling execution plan. Generating a partition-level resource allocation strategy includes: Obtain meter metadata from the business system and calculate the basic business value score based on user type and electricity consumption; Calculate the data timeliness value based on the collection time interval in the standardized meter response characteristic data, and calculate the data gap value based on the collection success records; Integrate the basic business value score, data timeliness value, and data gap value, and calculate the data value index of each meter through weighted summation; Based on the dynamic partitioning results of electricity meters and the data value indicators of each meter, the total value of the partition and the historical success rate of the partition are calculated, and the partition resource allocation ratio is calculated accordingly. The partition-level resource allocation strategy is generated in combination with the total available collection time.

2. The method according to claim 1, characterized in that Generate an optimized collection scheduling execution plan, including: Based on the partition-level resource allocation strategy, an outer time wheel data structure is constructed, the acquisition window is divided into multiple time slices, and the scheduling order is set according to the partition priority to generate an outer time wheel scheduling table; Based on the pre-calculated meter-level priority index, an inner time wheel data structure is configured for each meter in the partition, and a meter scheduling sequence is generated within the partition. Each meter in the partition is assigned a decreasing time slice. Meters with higher acquisition success rates receive smaller time slices. The outer time wheel scheduling table, the meter scheduling sequence within the partition, and the pre-configured retry strategy table are integrated to generate the collection scheduling execution plan.

3. The method according to claim 2, characterized in that Assigning decreasing time slices to each meter in the meter scheduling sequence within the zone includes: Read the meter scheduling sequence within the partition, extract the historical RTT sequence and success rate sequence, Combined with the real-time RTT value, the RTT change rate, RTT variance mutation index, and success rate decline index are calculated. A network anomaly score is generated through a multi-dimensional network anomaly score fusion model. The network stability index is calculated and the anomaly detection threshold is adjusted accordingly. Abnormal state meters are identified based on network anomaly scores and anomaly detection thresholds. Time slice quality adjustment factors are calculated for stable mode meters, and real-time quality adjustment factors are calculated for response mode meters based on real-time network quality. Time slice allocation values ​​are generated based on this and boundary constraints are processed.

4. The method according to claim 1, wherein Generates standardized meter response characteristic data, including: Perform time range screening and outlier processing on historical meter collection records to generate cleaned historical data; Perform correlation analysis on the cleaned historical data and grid load data, calculate the correlation coefficient between the response performance of each meter and the grid load, and generate meter-load correlation data; The period performance statistics of the cleaned historical data and the meter-load correlation data are integrated to generate standardized meter response characteristic data.

5. The method according to claim 4, characterized in that Perform correlation analysis on cleaned historical data and grid load data, including: Obtain multi-level grid loads, including transformer-level loads, feeder-level loads, and regional total loads; perform time series decomposition on these loads along with meter response time series from cleaned historical data, calculate the delayed cross-correlation coefficients of each component, and generate component-level correlation features; The load correlation of each level is weighted according to the number of electrical hops from the meter to the transformer to generate the topological weighted correlation; Perform frequency domain transformation on the cleaned historical data and multi-level power grid loads, calculate the coherence spectrum, and generate frequency domain correlation features; Calculate the correlation changes between meter response and grid load in different time windows, detect correlation mutation points, and generate correlation time-varying features; The above outputs are weighted and integrated to generate the meter-load correlation index.

6. The method according to claim 1, wherein Generate dynamic partition results for electric meters, including: Based on the standardized meter response characteristic data, a networking behavior feature vector including average response time, response time standard deviation, short-term success rate, long-term success rate, stability index, and load correlation is constructed and normalized to generate a normalized feature vector. The Mahalanobis distance between each pair of electricity meters is calculated based on the normalized eigenvectors to generate the electricity meter behavior similarity matrix; Perform cluster analysis on the meter behavior similarity matrix to obtain clustering results; Generate dynamic meter partitioning results based on clustering results, including partition ID, each partition meter set and partition center point features.

7. The method according to claim 6, characterized in that The improved density clustering algorithm is a topology-constrained hierarchical clustering algorithm: Read the original grid topology data from the grid management system, construct the grid topology structure, calculate the electrical hop count matrix between meters, identify transformer nodes of different voltage levels as topology anchor points, and generate a set of topology constraint cluster centers; Based on the meter behavior similarity matrix and the electrical distance matrix, a topology constraint similarity matrix is ​​generated through electrical distance weighted similarity reconstruction. Based on the topologically constrained cluster center set and the topologically constrained similarity matrix, coarse-grained clustering is performed at the first-level anchor point level, and then fine-grained partitioning is performed within the cluster. Topological connectivity verification is implemented, and topological connectivity clustering results are generated and dynamic boundary optimization is performed to generate stabilized partitioning results and cluster output.

8. The method according to claim 7, characterized in that Electrical distance weighted similarity reconstruction, including: Extracting the behavioral similarity and electrical distance of each pair of meters from the meter behavioral similarity matrix and electrical distance matrix, constructing a dual-constrained input data set, and multiplying the behavioral similarity by the exponential decay factor of the electrical distance to generate a topological weighted similarity matrix; When the meters belong to different voltage levels, the level penalty factor is used to modify the meter behavior similarity matrix to generate a level-constrained similarity matrix. The topologically weighted similarity matrix and the level-constrained similarity matrix are integrated to generate a topologically constrained similarity matrix.

9. The method according to claim 6, characterized in that Networking behavior feature vector, including: average response time, response time standard deviation, short-term success rate, long-term success rate, stability index, and correlation coefficient with grid load.

10. The method according to claim 1, characterized in that Calculate data value indicators, including: Extract business value basic points, data timeliness value, and data gap value from historical data, normalize them, and generate three types of normalized value indicators; and adjust the indicator values ​​according to the current system status to generate dynamic weight coefficients; The dynamic weight coefficient is used to weight the sum of the three normalized value indicators to obtain the basic value score, and the interaction enhancement items between the value indicators are added to obtain the interactive value score; Set compensation coefficients, importance coefficients, and gain coefficients for meters that continuously fail to collect data, meters for key users, and meters in designated monitoring states, and calculate situational adjustment coefficients. Integrate the basic value score, interaction value score and situational adjustment coefficient to calculate the data value index.

11. The method according to claim 1, wherein Calculate the partition resource allocation ratio, including: Based on the total value of the partition, the historical success rate of the partition, the cohesion of the partition and the size of the partition, the partition importance index is comprehensively calculated; according to the global statistical characteristics of the system, the dynamic balance coefficient is generated; Calculate the initial resource allocation ratio based on the partition importance index and dynamic balance coefficient; generate the minimum resource quota and maximum resource quota based on the resource upper and lower limit guarantee mechanism; Integrate the preliminary resource allocation ratio, minimum resource quota and maximum resource quota to generate the optimized resource allocation ratio that meets the constraints and calculate the partition time quota based on it. Combined with the partition priority sorting, a partition-level resource allocation strategy is formed.

Images