Power distribution network data dynamic cleaning and completion method based on graph structure and timing association

Abstract

The application provides a power distribution network data dynamic cleaning and completing method based on a graph structure and timing association, relates to the technical field of intelligent power distribution network management, verifies the density reachable of data sample density, obtains an abnormal candidate point, determines the corresponding associated data structure of power distribution network operation data according to the data continuous characteristics of the abnormal candidate point and the power distribution network operation data; maps the abnormal candidate point to the associated data structure to obtain a rule constraint interval of the power distribution network operation data, extracts the detected outliers of the power distribution network operation data from the rule constraint interval; determines the completing repair characteristics according to the detected outliers and the completing candidate set, completes the power distribution network operation data according to the completing repair characteristics, and outputs the power distribution network completing data. The application can clean and reconstruct massive multi-source electrical monitoring data in the dynamic topology power distribution network with access to distributed power sources, so as to improve the restoration accuracy of missing data.

Description

Technical Field

[0001] This application relates to the field of smart distribution network management technology, and more specifically, to a method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation. Background Technology

[0002] Smart distribution network management is a digital power control technology supported by distribution network topology and comprehensive electrical timing monitoring data. Its mainstream applications cover scenarios such as urban 10kV medium-voltage feeders, county-level low-voltage distribution areas, and active power distribution systems in industrial parks including distributed photovoltaic and energy storage systems. The general implementation framework in this field relies on the synchronous collection of voltage, current, active and reactive power operation data from multiple terminals. Through a three-layer processing chain of basic data cleaning, missing data completion, and operating condition analysis, it outputs compliant data to support upper-level operations such as load forecasting, fault diagnosis, three-phase imbalance management, and distribution network coordinated control. Complete and accurate massive amounts of operational data are the core foundation for achieving refined and automated control of smart distribution networks. The operational efficiency and numerical restoration accuracy of the data processing system directly determine the effectiveness of digital dispatching and maintenance operations in the distribution network.

[0003] However, current dynamic cleaning and completion processes for distribution network data generally employ single-dimensional, independent computational logic, relying solely on local values ​​from single samples for anomaly identification and unidirectional time-series data filling. This prevents the data processing from simultaneously integrating the spatial electrical coupling correlations of distribution network equipment with long-term evolutionary patterns. Consequently, anomaly identification is prone to missed or false positives, and the restoration of missing values ​​does not accurately reflect real-world operating conditions. Therefore, the resulting data integrity and accuracy are insufficient to meet the standards for various advanced management and control services in active distribution networks. Thus, how to clean and reconstruct massive amounts of multi-source electrical monitoring data in dynamic topology distribution networks connected to distributed power sources to improve the accuracy of missing data restoration is a key challenge facing the industry. Summary of the Invention

[0004] This application provides a method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation. It can clean and reconstruct massive multi-source electrical monitoring data in dynamic topology distribution networks with distributed power sources to improve the accuracy of restoring missing data.

[0005] Firstly, this application provides a method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation, wherein the method: Collect power distribution network operation data and construct a time-series dataset of power distribution network operation; Determine the data sample density of the time series dataset within a set neighborhood radius, perform density reachability verification on the data sample density to obtain anomaly candidate points, lock the topology node constraint boundary with the anomaly candidate points, and divide the discrete feature interval by combining the data continuity characteristics of the distribution network operation data to determine the associated data structure corresponding to the distribution network operation data. The abnormal candidate points are mapped to the associated data structure to obtain the rule constraint interval of the distribution network operation data, and the detected outliers of the distribution network operation data are extracted from the rule constraint interval. Based on the associated data structure, a candidate set for the completion of distribution network operation data is constructed. The completion and repair features are determined according to the detected outliers and the candidate set. The missing locations, missing time intervals, and missing types in the distribution network operation data are identified. The electrical fields to be completed are locked. The electrical fields to be completed are synchronously repaired and completed according to the completion and repair features. The completed distribution network data is then output.

[0006] In this embodiment, constructing the time-series dataset for the operation of the distribution network specifically includes: Extract the original sampled data sequence from the power distribution network operation data; The original sampled data sequence is calibrated for deviation and sorted in ascending order by timestamp to obtain a time-series dataset.

[0007] In this embodiment, determining the data sample density of the time-series dataset within a set neighborhood radius specifically includes: Set the search step size interval for the neighborhood distance, and perform a grid search on the time series dataset with the silhouette coefficient as the target to obtain the neighborhood search parameters; The set of candidate samples within a set neighborhood radius is determined based on the neighborhood search parameters. Extract the data sample density of the time series dataset within a set neighborhood radius from the candidate sample set.

[0008] In this embodiment, the density reachability verification of the data sample density to obtain anomaly candidate points specifically includes: Construct a point connectivity chain within the data neighborhood based on the data sample density; The connectivity attributes of the power distribution network operation data are determined by tracing back hop by hop of the aforementioned point connectivity chain. The network operation data with inconsistent connectivity paths are extracted from the connectivity attributes as anomaly candidate points.

[0009] In this embodiment, the determination of the associated data structure corresponding to the distribution network operation data by locking the topology node constraint boundary using the abnormal candidate points and dividing the discrete feature intervals based on the continuous data characteristics of the distribution network operation data specifically includes: Extract continuous operational features corresponding to the power grid operation type from the power distribution network operation data, and determine the discrete interval of the power distribution network operation data; Determine the feature attribution matrix corresponding to the power distribution network operation data based on the discrete intervals; The associated data structure corresponding to the distribution network operation data is determined based on the feature attribution matrix.

[0010] In this embodiment, mapping the abnormal candidate points to the associated data structure to obtain the rule constraint intervals of the distribution network operation data specifically includes: Locate the corresponding node position of the abnormal candidate point in the topology graph, and extract the association rule attribute of the node position in the associated data structure; Based on the association rule attributes, the boundary values ​​of the value range corresponding to each electrical feature at the node location are determined, forming the rule constraint limits; The rule constraint range for distribution network operation data is determined based on the rule constraint limits.

[0011] In this embodiment, determining the boundary values ​​of the value intervals corresponding to each electrical feature at the node location based on the association rule attributes specifically includes: Read the discrete interval identifiers and pre-stored interval mapping tables corresponding to each electrical feature at the node location from the association rule attributes, and determine the upper bound and lower bound of each feature value; The boundary values ​​of the value intervals corresponding to each electrical feature at the node location are determined based on the upper bounds and lower bounds of all feature values.

[0012] In this embodiment, the extraction of outlier points from the distribution network operation data within the rule constraint interval specifically includes: The candidate outliers are obtained by comparing each data feature value of the data sample to be verified with the boundary of the rule constraint interval. The candidate outliers and the abnormal candidate points are traced back to their data sources. Duplicate samples are identified by joint matching based on three information: the sample's built-in node location identifier, the time sequence timestamp, and the device number. Duplicate records from the same source are removed to obtain a set of deduplicated outliers. The outliers detected in the power distribution network operation data are determined based on the set of deduplicated outliers.

[0013] In this embodiment, the candidate outliers and the abnormal candidate points are traced back to their data sources. Duplicate samples are identified by joint matching based on three pieces of information: the sample's built-in node location identifier, the time-series timestamp, and the device number. The process of removing duplicate records from the same source to obtain the deduplicated outlier set specifically includes: Extract the three types of data source identifiers carried by the abnormal candidate points and candidate outliers: node location identifier, time sequence timestamp, and device number; The node location identifier, the time sequence timestamp, and the device number are compared together. If all three pieces of information match, it is determined to be a duplicate sample. Based on the duplicate samples, the candidate outliers are redundantly removed to obtain a deduplicated outlier set.

[0014] In this embodiment, determining the completion and repair features based on the detected outliers and the completion candidate set specifically includes: Based on the associated data structure, the feature value range of the detected outliers is determined, and the rule box space is determined; All distribution network operation data falling within the rule box space are selected to form a repair reference sample subset. The Mahalanobis distance quantization similarity is calculated, and the Mahalanobis distance quantization result is determined as the completion and repair feature.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: Data collection is performed on power distribution network operation data, and a time-series dataset of the power distribution network operation is constructed. The data sample density of the time-series dataset within a set neighborhood radius is determined, and the data sample density is verified to obtain anomaly candidate points. The anomaly candidate points are used to lock the topological node constraint boundaries. The discrete feature intervals are divided based on the continuous data characteristics of the power distribution network operation data, and the associated data structure corresponding to the power distribution network operation data is determined. The anomaly candidate points are mapped to the associated data structure to obtain the rule constraint intervals of the power distribution network operation data. Outliers of the power distribution network operation data are extracted from the rule constraint intervals. A completion candidate set of the power distribution network operation data is constructed based on the associated data structure. Completion and repair features are determined based on the detected outliers and the completion candidate set. The missing locations, missing time intervals, and missing types in the power distribution network operation data are identified, and the electrical fields to be completed are locked. The electrical fields to be completed are synchronously repaired and completed according to the completion and repair features, and the power distribution network completed data is output.

[0016] Therefore, this application can identify the missing locations, time intervals, and types of missing data in distribution network operation data, pinpoint the electrical fields to be supplemented, and synchronously supplement the electrical fields to be supplemented based on the supplementation and repair features, outputting distribution network supplemented data. Specifically, determining the associated data structure provides a unified and standardized constraint carrier that integrates equipment topology constraints and time-series numerical patterns, thus overcoming the inherent defects of existing technologies that rely solely on isolated single-sample threshold judgments and cannot consider the electrical coupling relationships of the network structure. This carrier uniformly incorporates the legal electrical value ranges of all nodes and spatiotemporal linkage logic, eliminating the one-sidedness of rules caused by single-dimensional judgments. It can adapt to the dynamic changes in network topology caused by distributed power source switching, fully covering the legal value boundaries of data under multiple operating conditions, providing a unified and reliable judgment benchmark for the entire process of anomaly identification and sample repair, fundamentally reducing the probability of misjudgment and missed judgment of abnormal samples, and simultaneously improving the overall stability of batch processing of massive data. Identifying outliers allows for the completion of a set of true abnormal samples through a double-layer verification mechanism, thus overcoming the inherent defects of existing technologies that rely solely on single-layer density clustering for coarse screening and cannot distinguish between operating condition fluctuations and true data distortion. This set removes suspected noise with no abnormal density, filters invalid labeled samples caused by simple clustering bias, accurately distinguishes distorted data caused by equipment acquisition failure and transmission packet loss from normal load fluctuation data, greatly reduces the amount of invalid computation in subsequent data repair, and simultaneously ensures that the anomaly identification results have quantifiable evaluation criteria, significantly improving the overall accuracy of cleaning massive multi-source electrical monitoring data.

[0017] In summary, the technical solution adopted in this application can clean and reconstruct massive multi-source electrical monitoring data in a dynamic topology distribution network connected to distributed power sources, thereby improving the accuracy of restoring missing data. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an exemplary flowchart of a dynamic cleaning and completion method for distribution network data based on graph structure and time-series correlation provided in this application; Figure 2 This is a flowchart illustrating the process for determining outliers provided in this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] This application provides a method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation. The core of this method is to collect distribution network operation data and construct a time-series dataset of the distribution network during operation; determine the data sample density of the time-series dataset within a set neighborhood radius, perform density reachability verification on the data sample density to obtain anomaly candidate points, use the anomaly candidate points to lock the topological node constraint boundaries, and divide discrete feature intervals based on the continuous data characteristics of the distribution network operation data to determine the corresponding associated data structure; map the anomaly candidate points to the associated data structure to obtain the rule constraint intervals of the distribution network operation data, and extract outliers from the rule constraint intervals; construct a completion candidate set for the distribution network operation data based on the associated data structure; determine completion and repair features based on the detected outliers and the completion candidate set; identify the missing locations, missing time-series intervals, and missing types in the distribution network operation data; lock the electrical fields to be completed; and synchronously repair and complete the electrical fields to be completed according to the completion and repair features, outputting the completed distribution network data.

[0022] To better understand the above technical solutions, a detailed description of the technical solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. (Refer to...) Figure 1 As shown in the figure, this is an exemplary flowchart of a dynamic cleaning and completion method for distribution network data based on graph structure and time-series correlation according to an embodiment of this application. The method includes the following steps: In step S1, power distribution network operation data is collected, and a time-series dataset of power distribution network operation is constructed.

[0023] In practice, the collection of distribution network operation data can be achieved in the following way: synchronous acquisition terminals are deployed at each busbar, transformer, distributed photovoltaic, and load point of the 10 kV medium-voltage and 0.4 kV low-voltage distribution network, with a unified acquisition cycle set; the terminals synchronously collect four core electrical quantities: three-phase voltage, three-phase current, active power, and reactive power, and bind a unique and accurate timestamp to each set of acquired data; after acquisition, raw invalid data such as garbled characters, exceeding the rated extreme value of the equipment, and empty records are initially filtered out and uniformly converted into a standardized data storage format; the multi-source raw acquisition records uploaded by each terminal are collected, cross-equipment sampling clock calibration is completed, the time sequence offset caused by asynchronous sampling is eliminated, and the distribution network operation data is obtained.

[0024] It should be noted that, in this application, the distribution network operation data refers to the real-time electrical condition data of voltage, current, active power and reactive power collected by the distribution network terminal.

[0025] In this embodiment, the construction of a time-series dataset for the operation of the distribution network can be achieved through the following steps: Extract the original sampled data sequence from the power distribution network operation data; The original sampled data sequence is calibrated for deviation and sorted in ascending order by timestamp to obtain a time-series dataset.

[0026] In practice, the process begins by reading the complete distribution network operation data set after invalid data removal. This data is then divided into independent subsets based on the monitoring equipment number. All collected records within each subset are traversed, and each record is separated to include its equipment number, timestamp, three-phase voltage, three-phase current, active power, and reactive power values. All continuously collected records from the same location are grouped by equipment affiliation to form continuous numerical segments for each device. These segments are then spliced ​​together to generate a complete numerical sequence without time-series correction, thus obtaining the original sampled data sequence. Next, a global standard clock reference is established. Each record in the original sampled data sequence carries its own local timestamp. The global time offset between the terminal's local clock and the standard clock is calculated using the round-trip message delay formula. This offset is then used to correct the original timestamp value of each record. After correction, the data is uniformly converted to a standardized timestamp format. Using the standardized timestamp as the unique sorting key, all calibrated sampled records are rearranged in ascending order. Duplicate timestamp records appearing after sorting are removed, generating a time-series dataset.

[0027] It should be noted that in this application, the original sampling data sequence refers to the continuous electrical acquisition records of a single monitoring device without local time scale correction; the time series dataset refers to the set of continuous electrical quantities of the distribution network arranged in ascending order according to a unified standard time scale.

[0028] In step S2, the data sample density of the time series dataset within a set neighborhood radius is determined, density reachability verification is performed on the data sample density to obtain anomaly candidate points, the topology node constraint boundary is locked with the anomaly candidate points, and discrete feature intervals are divided in combination with the data continuity characteristics of the distribution network operation data to determine the associated data structure corresponding to the distribution network operation data.

[0029] In this embodiment, the data sample density of the time-series dataset within a set neighborhood radius can be determined in the following way: Set the search step size interval for the neighborhood distance, and perform a grid search on the time series dataset with the silhouette coefficient as the target to obtain the neighborhood search parameters; The set of candidate samples within a set neighborhood radius is determined based on the neighborhood search parameters. Extract the data sample density of the time series dataset within a set neighborhood radius from the candidate sample set.

[0030] In specific implementation, firstly, the range of neighborhood distance values ​​is defined as 0.01 to 0.5, with a fixed search step size of 0.01. Simultaneously, the range of minimum sample size is defined as 3 to 20, corresponding to a search step size of 1. All parameter combinations within these two ranges are traversed, and density clustering is performed for each parameter group. The profile coefficient is calculated for each sample, and discrete clusters deviating from the topology are simultaneously included in the distribution network connectivity constraints. The parameter combination with the maximum profile coefficient among the global parameter combinations is selected and used as the neighborhood search parameter. Then, the optimal neighborhood distance value within the neighborhood search parameter is extracted as the set neighborhood radius, and the optimal minimum sample size threshold is extracted. All data samples within the time-series dataset are traversed, and the feature space distance between the current sample and all other samples is calculated using the Euclidean distance formula. All other samples with distance values ​​less than or equal to the set neighborhood radius are selected, and all selected samples are aggregated with the current center sample to generate an initial sample group corresponding to that center sample. All initial sample groups are then integrated to form a candidate sample set. Finally, for each independent center sample in the time series dataset, the candidate sample set corresponding to that sample is retrieved, and the total number of all valid data samples contained in the set is counted. The total number of samples obtained is assigned as the data sample density corresponding to the current center sample, and the density value corresponding to each sample is recorded and bound to the unique identifier of the sample for storage. After completing the traversal and statistics of all samples in the time series dataset, the data sample density of the time series dataset within the set neighborhood radius is output.

[0031] It should be noted that, in this application, the set neighborhood radius refers to the distance threshold that defines the local feature space coverage of a single time series sample; the search step size interval refers to the fixed difference of a single parameter change when the grid search traverses the parameter interval; the silhouette coefficient refers to the evaluation index of the compactness within clusters and the separation between clusters of clustered samples; the grid search refers to the optimization mechanism that traverses all combinations of multiple parameters with a fixed step size; the neighborhood search parameter refers to the combination of the optimal neighborhood distance and the minimum number of samples output by the grid search; the candidate sample set refers to the collection of all related samples centered on a single time series sample and within a distance not exceeding the set neighborhood radius; and the data sample density refers to the total number of valid data samples contained in the candidate sample set corresponding to a single time series sample.

[0032] In this embodiment, the density reachability verification of the data sample density to obtain anomaly candidate points can be achieved through the following steps: Construct a point connectivity chain within the data neighborhood based on the data sample density; The connectivity attributes of the power distribution network operation data are determined by tracing back hop by hop of the aforementioned point connectivity chain. The network operation data with inconsistent connectivity paths are extracted from the connectivity attributes as anomaly candidate points.

[0033] In specific implementation, firstly, the density values ​​of all time-series samples that have been matched and the minimum sample number threshold within the neighborhood search parameters are retrieved. Samples with density values ​​greater than or equal to the threshold are marked as density core samples. The candidate sample set corresponding to each density core sample is traversed, and all samples in the set are established with the core sample through bidirectional connectivity. The remaining density core samples in the associated samples are retrieved sequentially, continuously expanding the sample connectivity relationship. All samples with bidirectional density connectivity are sequentially concatenated and sorted to generate multiple independent point connectivity chains within the data neighborhood. The point connectivity chain is a sample sequence formed by the orderly concatenation of time-series samples that mutually satisfy density connectivity constraints. Any adjacent samples in the chain satisfy both neighborhood distance and sample density constraints. Then, the end sample of a single point connectivity chain is selected as the backtracking starting node. The previous hop associated sample is retrieved in reverse order along the sample concatenation sequence stored in the link, and hop-by-hop backtracking verification is performed to check whether the distance and density constraints between adjacent samples continue to satisfy the connectivity judgment formula. The hop-by-hop backtracking refers to the judgment process of traversing adjacent associated samples in reverse order along the sample hierarchy stored in the point connectivity chain and verifying the connectivity constraints layer by layer. The entire process records the link number and link level position of each sample within the backtracking link to distinguish whether a sample belongs to a continuous and complete density-connected cluster. Samples with complete and continuous backtracking links that satisfy density connectivity constraints throughout are marked as connected samples. Samples whose distance to adjacent samples exceeds a set neighborhood radius, whose sample density is lower than the minimum sample number threshold, resulting in link breakage and inability to complete bidirectional density reachability retrieval, are marked as disconnected samples, forming the connectivity attributes of the distribution network operation data. Finally, the connectivity attribute labels corresponding to all samples in the time series dataset are traversed, and all distribution network operation data samples labeled as disconnected are filtered. The judgment criteria for inconsistent connectivity paths are defined: if any link is broken during the hop-by-hop backtracking verification process and no complete and continuous point connected chain can achieve bidirectional density reachability throughout the entire process, then the sample is determined to be an inconsistent connectivity path sample. All time series samples that meet the criteria for inconsistent connectivity paths are collected to obtain distribution network operation data with inconsistent connectivity paths as anomaly candidate points.

[0034] It should be noted that in this application, density reachability verification refers to a complete verification process that determines whether a sample can be classified into the same data cluster based on sample density and neighborhood constraints; a point connected chain refers to an ordered sample sequence carrier formed by sequentially linking time-series samples that satisfy the density connectivity conditions; hop-by-hop backtracking refers to a traversal verification method that searches for adjacent related samples in reverse order along the sample hierarchy stored in the point connected chain; connectivity attribute refers to a standardized judgment label for whether a single time-series sample belongs to a complete continuous density connected cluster; inconsistent connectivity path refers to the isolated state judgment result of a sample that cannot be connected to any complete point connected chain through density constraints; and abnormal candidate point refers to a time-series sample that has been judged by density reachability verification as having an inconsistent connectivity path and is subject to secondary verification by spatiotemporal rules.

[0035] In this embodiment, the determination of the associated data structure corresponding to the distribution network operation data by locking the topology node constraint boundaries using the anomaly candidate points and dividing the discrete feature intervals based on the continuous data characteristics of the distribution network operation data can be achieved through the following steps: Extract continuous operational features corresponding to the power grid operation type from the power distribution network operation data, and determine the discrete interval of the power distribution network operation data; Determine the feature attribution matrix corresponding to the power distribution network operation data based on the discrete intervals; The associated data structure corresponding to the distribution network operation data is determined based on the feature attribution matrix.

[0036] In practice, firstly, the power grid operation type is classified according to the power supply attributes of the equipment, distinguishing between three types of equipment: transformers, distributed photovoltaics, and conventional loads. Corresponding continuous operating features of voltage, current, active power, and reactive power are extracted for each type. Clustering optimization is used to divide intervals, traversing between clusters with values ​​ranging from two to fifteen. The optimal profile coefficient is used as the classification criterion. Voltage is divided into eight discrete intervals, current into five, active power into six, and reactive power into five. The boundaries of each interval are locked based on the extreme values ​​of the per-unit features, ensuring no numerical overlap between intervals. Then, a two-dimensional matrix is ​​constructed with rows equal to the total number of samples and columns equal to the total dimension of the features. Each continuous operating feature of each distribution network operation data sample is traversed, comparing the per-unit value of the feature to its corresponding discrete interval. If the feature value falls into the corresponding discrete interval, the corresponding row and column position in the matrix is ​​assigned a value of 1; if the value is outside all discrete intervals, it is assigned a value of 0. This process is repeated row by row to fill all samples, generating a feature attribution matrix. All elements in the matrix contain only two values: 0 and 1, with no intermediate values. The formula for determining the feature attribution matrix is ​​as follows: , It is the element in the i-th row and j-th column of the matrix; Let be the value of the j-th dimension feature of the i-th sample; This feature corresponds to any discrete interval. Finally, an electrical feature itemset is constructed based on the feature attribution matrix. The itemset is a combination set of multi-dimensional electrical feature value intervals for a single device in the same time series. The frequency of occurrence of the feature itemset is counted according to the binary elements in the feature attribution matrix, and the support and confidence of the association rule are calculated using the frequent itemset standard algorithm. Support and confidence thresholds are preset according to the rated operating conditions of the distribution network equipment and the fluctuation range of distributed power source switching. Rules with support and confidence values ​​higher than their respective preset thresholds are retained as strong spatial association rules in the time series. A topology carrier is built with the distribution network bus, load, and photovoltaic equipment as vertices and power feeders as edges. According to the acquisition device number and line topology attribution relationship corresponding to the rule, the selected strong association rules are bound to the corresponding topology vertex or topology edge unit. The discrete interval constraints of each vertex are entered simultaneously. The three types of information, namely topology connection relationship, discrete interval boundary, and spatiotemporal association rule, are integrated to generate an integrated association data structure.

[0037] It should be noted that, in this application, continuous data characteristics refer to the per-unit numerical sequence of electrical quantities such as voltage, current, and power that changes continuously over time; grid operation type refers to the equipment category to which different power distribution equipment such as transformers, distributed photovoltaics, and conventional loads belong; continuous operation characteristics refer to the non-overlapping numerical segment ranges with fixed upper and lower limits for continuous electrical quantities; discrete intervals refer to the non-overlapping numerical segment ranges with fixed upper and lower limits for continuous electrical quantities; feature attribution matrix refers to a binary two-dimensional data matrix indicating whether each dimension of the sample's features falls into the corresponding discrete interval; and associated data structure refers to a standardized constraint carrier that integrates distribution network topology, discrete intervals, and spatiotemporal association rules.

[0038] In step S3, the abnormal candidate points are mapped to the associated data structure to obtain the rule constraint interval of the distribution network operation data, and the detected outliers of the distribution network operation data are extracted from the rule constraint interval.

[0039] In this embodiment, mapping the abnormal candidate points to the associated data structure to obtain the rule constraint interval of the distribution network operation data can be achieved by the following steps: Locate the corresponding node position of the abnormal candidate point in the topology graph, and extract the association rule attribute of the node position in the associated data structure; Based on the association rule attributes, the boundary values ​​of the value range corresponding to each electrical feature at the node location are determined, forming the rule constraint limits; The rule constraint range for distribution network operation data is determined based on the rule constraint limits.

[0040] In specific implementation, firstly, the unique identifier of the built-in device of the anomaly candidate point is retrieved, and all vertex indices of the topology graph are traversed to complete identifier matching, locking the unique node position code of the anomaly candidate point. Using the node position code as the search keyword, the storage unit of the associated data structure is traversed to retrieve all temporal association rules and neighborhood spatial association rules bound to the node; the temporal conditions corresponding to the features in each rule and the coupling constraints of adjacent nodes are parsed one by one, and all the judgment logic of the node is completely collected and uniformly collected into the association rule attributes of the node position in the associated data structure. Then, the collected association rule attributes are parsed one by one, and the legal upper and lower limits corresponding to the four types of electrical features, namely voltage, current, active power, and reactive power, are separated. The lower limit value of a single feature is used as the left boundary of the interval, and the upper limit value is used as the right boundary of the interval. The coupling constraints of adjacent nodes are simultaneously incorporated to correct the boundary offset, eliminate the isolated judgment bias of a single device, and integrate the left and right boundaries of a single electrical feature pair into a single constraint. All single constraints of features are summarized to form the rule constraint boundary. Finally, all individual constraints within the rule constraint limit set are traversed, and each set of paired boundary values ​​is directly mapped to the complete value range of the corresponding electrical feature. The independent value ranges of the four types of electrical features are collected by node and merged to generate the overall rule constraint interval of the node. The interval contains all legal electrical quantity values ​​that conform to time-series evolution and neighborhood electrical coupling, while the interval outside is determined to violate the grid operation state, thus generating the rule constraint interval of the distribution network operation data.

[0041] It should be noted that, in this application, the topology graph refers to a network space mathematical model constructed with distribution network equipment as vertices and power lines as edges; the node position refers to the unique vertex storage index representing a single distribution network monitoring device in the topology graph; the association rule attribute refers to the set of temporal and spatial dual electrical operation constraint rules binding the topology nodes; the electrical characteristic refers to the voltage, current, and power numerical indicators of the operating status of distribution network equipment; the value interval boundary value refers to the minimum and maximum extreme values ​​that limit the legal operating values ​​of a single electrical characteristic; the rule constraint limit refers to the binary constraint combination formed by pairing the upper and lower boundary values ​​of a single electrical characteristic; and the rule constraint interval refers to the total range of legal values ​​obtained by integrating all electrical characteristic constraint limits of the nodes.

[0042] In addition, in this embodiment, determining the boundary values ​​of the value intervals corresponding to each electrical feature at the node location based on the association rule attributes can be achieved using the following steps: Read the discrete interval identifiers and pre-stored interval mapping tables corresponding to each electrical feature at the node location from the association rule attributes, and determine the upper bound and lower bound of each feature value; The boundary values ​​of the value intervals corresponding to each electrical feature at the node location are determined based on the upper bounds and lower bounds of all feature values.

[0043] In specific implementation, firstly, the complete association rule attribute dataset is retrieved using the current node position code as the retrieval primary key. The discrete interval identifier codes bound to voltage, current, active power, and reactive power are separated one by one. The system's globally pre-stored interval mapping table is retrieved, and each record in the mapping table is matched one-to-one with the discrete interval identifier and its corresponding extreme value. Using the discrete interval identifier as the matching index, the mapping table is traversed to retrieve the corresponding stored field, and the maximum per-unit value obtained from the matching identifier is extracted as the upper bound and the minimum per-unit value as the lower bound of the temporary value for a single feature. Then, multiple sets of temporary upper and lower bounds derived from all association rules for the same electrical feature are collected. The global maximum value of all temporary lower bounds for this feature is selected as the final lower bound, and the global minimum value of all temporary upper bounds for this feature is selected as the final upper bound. This completes multi-rule interval convergence. The converged final lower and upper bounds are paired and combined to form a set of value interval boundary values ​​specific to this electrical feature. The same convergence operation is performed sequentially on the four types of electrical features, outputting the value interval boundary values ​​corresponding to each electrical feature at the node position.

[0044] It should be noted that, in this application, the discrete interval identifier refers to the encoding symbol corresponding to a set of electrical quantity value segments; the interval mapping table refers to the static data table that stores the discrete interval identifier and the corresponding upper and lower limits of the values; the upper limit of the feature value refers to the maximum per-unit value that the electrical feature is allowed to appear within a single association rule; and the lower limit of the feature value refers to the minimum per-unit value that the electrical feature is allowed to appear within a single association rule.

[0045] Preferably, in this embodiment, outlier points in the distribution network operation data are extracted from the rule constraint interval, with reference to... Figure 2 As shown in the figure, this is a schematic flowchart of the process for determining outlier detection points in some embodiments of this application. In this embodiment, the determination of outlier detection points can be achieved by the following steps: In step S31, all data feature values ​​of the data sample to be verified are compared with the boundary of the rule constraint interval one by one to obtain candidate outliers; In step S32, the candidate outliers and the abnormal candidate points are traced back to their data sources. Duplicate samples are identified by joint matching based on the three information of the sample's built-in node location identifier, time sequence timestamp, and device number. Duplicate records from the same source are removed to obtain a set of deduplicated outliers. In step S33, outliers in the distribution network operation data are determined based on the deduplication outlier set.

[0046] In practice, firstly, all per-unit values ​​of voltage, current, active power, and reactive power contained in a single data sample to be verified are retrieved. These values ​​are then matched against the rule constraint intervals corresponding to the node to which the sample belongs. Each feature value is compared to the lower and upper bounds of the corresponding interval. If any feature value is less than the lower bound or greater than the upper bound, the sample is deemed to have exceeded the legal value range and is marked as a temporary outlier. After all dimensions are compared, all temporary outliers are aggregated to form candidate outliers. Next, a globally unique sampling identifier is extracted from both candidate outliers and candidate outliers. This identifier simultaneously records three types of data source information: the acquisition terminal, the acquisition time, and the device number. Using the sampling identifier as the matching primary key, two sets of sample sets are compared bidirectionally. Duplicate samples existing in both sets are filtered out, retaining only a single record. Samples existing only in a single set are retained in their entirety. Duplicate and redundant sample entries are removed, and all non-duplicate outliers are integrated to generate a deduplicated outlier set. Finally, after traversing all samples in the deduplicated outlier set obtained by jointly matching the node location identifier, time sequence timestamp, and device number to remove duplicate samples from the same source, the device-assigned node, feature out-of-bounds dimension, and out-of-bounds numerical deviation of each sample are reviewed. It is confirmed that all samples in the set simultaneously meet the dual judgment conditions of density connectivity path anomaly and rule interval out-of-bounds. All samples in the set are uniformly assigned a unique classification label for detected outliers, thus completing the solidification of sample attributes and outputting the detected outliers of the distribution network operation data.

[0047] It should be noted that in this application, the data sample to be verified refers to the original time-series data record of the distribution network that has completed density coarse screening and entered the regular interval verification; data source tracing refers to verifying the original source of the sample by relying on the unique collection identifier of the sample; the deduplicated outlier set refers to the summary of suspected abnormal samples that have eliminated duplicate records and simultaneously meet the two-level anomaly judgment conditions; and the detected outlier refers to the real abnormal distribution network sample that has been confirmed by both density verification and multi-dimensional regular interval verification.

[0048] In addition, in this embodiment, the data source tracing of the candidate outliers and the abnormal candidate points, and the identification of duplicate samples by jointly matching the three information of the sample's built-in node location identifier, time sequence timestamp, and device number, and the elimination of duplicate records from the same source to obtain the deduplicated outlier set can be achieved by the following steps: Extract the three types of data source identifiers carried by the abnormal candidate points and candidate outliers: node location identifier, time sequence timestamp, and device number; The node location identifier, the time sequence timestamp, and the device number are compared together. If all three pieces of information are consistent, it is determined to be a duplicate sample. Based on the duplicate samples, the candidate outliers are redundantly removed to obtain a deduplicated outlier set.

[0049] In specific implementation, firstly, the three types of data source identifiers carried by the candidate anomalies and candidate outliers are extracted: node location identifier, time sequence timestamp, and device number. All candidate anomaly samples are traversed, and the three types of identifiers bound to each candidate outlier sample are retrieved simultaneously to establish a dual-set multi-dimensional identifier comparison table. Then, the node location identifier, time sequence timestamp, and device number are jointly compared and verified. If all three types of information are consistent, the two samples are determined to be duplicate samples, and all duplicate samples are collected to generate a duplicate sample index list. Finally, based on the duplicate sample index list, a redundancy removal operation is performed on the entire set of candidate outliers, retaining only the single sample corresponding to the same source within the candidate anomalies as a unique record. All candidate outliers not included in the duplicate list are completely retained without deletion. The remaining non-duplicate anomaly samples are uniformly collected, reordered and reorganized according to chronological order to generate a deduplicated outlier set without duplicate records from the same source.

[0050] It should be noted that in this application, node location identifier refers to a unique coded symbol bound to a single distribution network device and used to distinguish topology vertices; duplicate samples refer to two data records produced by two screening processes that correspond to the same device at the same collection time; redundancy removal refers to the deletion operation performed on abnormal samples that are duplicates from the same source.

[0051] In step S4, a candidate set for the completion of distribution network operation data is constructed based on the associated data structure. The completion and repair features are determined according to the detected outliers and the candidate set. The missing locations, missing time intervals, and missing types in the distribution network operation data are identified. The electrical fields to be completed are locked. The electrical fields to be completed are synchronously repaired and completed according to the completion and repair features. The completed distribution network data is then output.

[0052] In specific implementation, the completion candidate set of distribution network operation data based on the aforementioned associated data structure can be constructed in the following way: First, locate the topological node location identifier corresponding to the outlier point, retrieve the complete rule constraint interval bound to the node in the associated data structure, the interval includes the legal upper and lower boundary values ​​of four electrical features: voltage, current, active power, and reactive power, then retrieve the basic normal dataset after time-series calibration, and perform full-dimensional boundary verification on each normal data sample in the basic dataset one by one, comparing whether all electrical feature values ​​of the sample fall within the rule constraint interval corresponding to the current node, retaining only compliant normal samples whose values ​​of all dimensions have not exceeded the upper and lower limits of the interval, and removing unqualified samples with any feature exceeding the boundary, and uniformly collecting and summarizing all compliant normal samples that have passed the verification, sorting and regularizing them in ascending order of collection timestamp, and synchronously binding the topological association rules and time-series change features of the corresponding node to generate the completion candidate set of distribution network operation data.

[0053] It should be noted that, in this application, the complete candidate set refers to a collection of normal power distribution network operation data that fits the spatiotemporal constraints of nodes and can be used for numerical repair of abnormal samples.

[0054] In this embodiment, determining the completion and repair features based on the detected outliers and the completion candidate set can be achieved through the following steps: Based on the associated data structure, the feature value range of the detected outliers is determined, and the rule box space is determined; All distribution network operation data falling within the rule box space are selected to form a repair reference sample subset. The Mahalanobis distance quantization similarity is calculated, and the Mahalanobis distance quantization result is determined as the completion and repair feature.

[0055] In practice, firstly, the topology node location identifiers bound to the detected outliers are read. Using the identifier as the retrieval key, the multi-dimensional rule constraint boundaries stored in the associated data structure are retrieved. The feature value ranges corresponding to voltage, current, active power, and reactive power are separated, and the upper and lower limits of each dimension are combined to form a multi-dimensional closed numerical space. This multi-dimensional closed numerical space is the rule box space. The space boundary completely matches the timing and electrical coupling constraints of the current node. The space contains all compliant electrical per-unit values, while the space outside is judged as abnormal values ​​that violate the network structure's operating state. Then, all normal data samples in the pre-constructed completion candidate set are traversed. Each sample's electrical feature values ​​are compared with the boundaries of each dimension of the rule box space. Only compliant samples whose values ​​in all dimensions are within the space are retained, and these compliant samples are collected to form a repair reference sample subset. The subset covariance matrix and covariance inverse matrix are solved, and Mahalanobis distance is used to quantify the numerical similarity between the detected outliers and each compliant sample. All Mahalanobis distance quantization results are used as completion and repair features.

[0056] It should be noted that in this application, the feature value range refers to the minimum to maximum legal per-unit value range of a single electrical feature constrained by the grid structure; the rule box space refers to the multi-dimensional legal value closed space formed by the joint of all electrical feature value ranges of the nodes; and the completion and repair feature is the Mahalanobis distance index for quantifying the similarity between detected outliers and compliant normal sample operating conditions.

[0057] In specific implementation, the missing locations, missing time intervals, and missing types in the distribution network operation data are identified, the electrical fields to be completed are located, and the electrical fields to be completed are synchronously repaired and completed according to the completion and repair features. The output of the distribution network completed data can be implemented in the following way: retrieve the Mahalanobis distance quantization results corresponding to all completion and repair features, traverse the repair reference sample subset, and select the sample with the smallest distance value as the repair benchmark sample group. If the benchmark sample group contains only a single compliant normal sample, directly reuse all electrical feature values ​​of that sample to replace the out-of-bounds abnormal values ​​of the detected outliers; if the benchmark sample contains multiple compliant samples, calculate the arithmetic mean of each dimension of voltage, current, active power, and reactive power, and use the mean result to cover the original abnormal records of the detected outliers. After all detected outliers have been replaced, for the missing intervals in the time segment where no data collection records are available, extract the spatiotemporal fusion features and adaptively fill the blank time series in two scenarios: single-value missing and continuous block missing. For single missing values, the numerical reconstruction output is completed directly using the benchmark sample with the highest matching degree. For missing values ​​in the middle continuous blocks, linear interpolation weighted correction is used. For missing values ​​at the beginning and end of the time series, only the matching benchmark sample is used to reconstruct the numerical values. After all anomaly repair and blank filling operations are completed, all data are regularized in ascending time series, duplicate and redundant records are removed, and finally the distribution network complete data is obtained.

[0058] It should be noted that, in this application, distribution network completion data refers to a set of distribution network time-series data that has completed the repair of abnormal samples and the filling of missing values, while satisfying the network topology and time-series operation constraints.

[0059] Therefore, this application can identify the missing locations, time intervals, and types of missing data in distribution network operation data, pinpoint the electrical fields to be supplemented, and synchronously supplement the electrical fields to be supplemented based on the supplementation and repair features, outputting distribution network supplemented data. Specifically, determining the associated data structure provides a unified and standardized constraint carrier that integrates equipment topology constraints and time-series numerical patterns, thus overcoming the inherent defects of existing technologies that rely solely on isolated single-sample threshold judgments and cannot consider the electrical coupling relationships of the network structure. This carrier uniformly incorporates the legal electrical value ranges of all nodes and spatiotemporal linkage logic, eliminating the one-sidedness of rules caused by single-dimensional judgments. It can adapt to the dynamic changes in network topology caused by distributed power source switching, fully covering the legal value boundaries of data under multiple operating conditions, providing a unified and reliable judgment benchmark for the entire process of anomaly identification and sample repair, fundamentally reducing the probability of misjudgment and missed judgment of abnormal samples, and simultaneously improving the overall stability of batch processing of massive data. Identifying outliers allows for the completion of a set of true abnormal samples through a double-layer verification mechanism, thus overcoming the inherent defects of existing technologies that rely solely on single-layer density clustering for coarse screening and cannot distinguish between operating condition fluctuations and true data distortion. This set removes suspected noise with no abnormal density, filters invalid labeled samples caused by simple clustering bias, accurately distinguishes distorted data caused by equipment acquisition failure and transmission packet loss from normal load fluctuation data, greatly reduces the amount of invalid computation in subsequent data repair, and simultaneously ensures that the anomaly identification results have quantifiable evaluation criteria, significantly improving the overall accuracy of cleaning massive multi-source electrical monitoring data.

[0060] In summary, the technical solution adopted in this application can clean and reconstruct massive multi-source electrical monitoring data in a dynamic topology distribution network connected to distributed power sources, thereby improving the accuracy of restoring missing data.

[0061] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0062] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0063] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation, characterized in that, The method includes: Collect power distribution network operation data and construct a time-series dataset of power distribution network operation; Determine the data sample density of the time series dataset within a set neighborhood radius, perform density reachability verification on the data sample density to obtain anomaly candidate points, lock the topology node constraint boundary with the anomaly candidate points, and divide the discrete feature interval by combining the data continuity characteristics of the distribution network operation data to determine the associated data structure corresponding to the distribution network operation data. The abnormal candidate points are mapped to the associated data structure to obtain the rule constraint interval of the distribution network operation data, and the detected outliers of the distribution network operation data are extracted from the rule constraint interval. Based on the associated data structure, a candidate set for the completion of distribution network operation data is constructed. The completion and repair features are determined according to the detected outliers and the candidate set. The missing locations, missing time intervals, and missing types in the distribution network operation data are identified. The electrical fields to be completed are locked. The electrical fields to be completed are synchronously repaired and completed according to the completion and repair features. The completed distribution network data is then output.

2. The method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation as described in claim 1, characterized in that, Constructing a time-series dataset for the operation of a power distribution network specifically includes: Extract the original sampled data sequence from the power distribution network operation data; The original sampled data sequence is calibrated for deviation and sorted in ascending order by timestamp to obtain a time-series dataset.

3. The method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation as described in claim 1, characterized in that, Determining the data sample density of the time series dataset within a set neighborhood radius specifically includes: Set the search step size interval for the neighborhood distance, and perform a grid search on the time series dataset with the silhouette coefficient as the target to obtain the neighborhood search parameters; The set of candidate samples within a set neighborhood radius is determined based on the neighborhood search parameters. Extract the data sample density of the time series dataset within a set neighborhood radius from the candidate sample set.

4. The method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation as described in claim 1, characterized in that, The density reachability verification of the data sample density yields the following anomaly candidate points: Construct a point connectivity chain within the data neighborhood based on the data sample density; The connectivity attributes of the power distribution network operation data are determined by tracing back hop by hop of the aforementioned point connectivity chain. The network operation data with inconsistent connectivity paths are extracted from the connectivity attributes as anomaly candidate points.

5. The method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation as described in claim 1, characterized in that, Using the aforementioned anomaly candidate points to lock the topology node constraint boundaries, and combining the continuous data characteristics of the distribution network operation data to divide discrete feature intervals, the specific determination of the associated data structure corresponding to the distribution network operation data includes: Extract continuous operational features corresponding to the power grid operation type from the power distribution network operation data, and determine the discrete interval of the power distribution network operation data; Determine the feature attribution matrix corresponding to the power distribution network operation data based on the discrete intervals; The associated data structure corresponding to the distribution network operation data is determined based on the feature attribution matrix.

6. The method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation as described in claim 1, characterized in that, Mapping the anomaly candidate points to the associated data structure to obtain the rule constraint intervals for distribution network operation data specifically includes: Locate the corresponding node position of the abnormal candidate point in the topology graph, and extract the association rule attribute of the node position in the associated data structure; Based on the association rule attributes, the boundary values ​​of the value range corresponding to each electrical feature at the node location are determined, forming the rule constraint limits; The rule constraint range for distribution network operation data is determined based on the rule constraint limits.

7. The method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation as described in claim 6, characterized in that, Determining the boundary values ​​of the value range corresponding to each electrical feature at the node location based on the association rule attributes specifically includes: Read the discrete interval identifiers and pre-stored interval mapping tables corresponding to each electrical feature at the node location from the association rule attributes, and determine the upper bound and lower bound of each feature value; The boundary values ​​of the value intervals corresponding to each electrical feature at the node location are determined based on the upper bounds and lower bounds of all feature values.

8. The method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation as described in claim 1, characterized in that, The specific outlier points detected from the rule-constrained interval for extracting distribution network operation data include: The data feature values ​​of the data sample to be verified are compared one by one with the boundary of the rule constraint interval to obtain candidate outliers; The candidate outliers and the abnormal candidate points are traced back to their data sources. Duplicate samples are identified by joint matching based on three information: the sample's built-in node location identifier, the time sequence timestamp, and the device number. Duplicate records from the same source are removed to obtain a set of deduplicated outliers. The outliers detected in the power distribution network operation data are determined based on the set of deduplicated outliers.

9. The method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation as described in claim 8, characterized in that, The candidate outliers and the abnormal candidate points are traced back to their data sources. Duplicate samples are identified by joint matching based on three pieces of information: the sample's built-in node location identifier, the time-series timestamp, and the device number. The deduplicated outlier set is obtained by removing duplicate records from the same source. Specifically, this includes: Extract the three types of data source identifiers carried by the abnormal candidate points and candidate outliers: node location identifier, time sequence timestamp, and device number; The node location identifier, the time sequence timestamp, and the device number are compared together. If all three pieces of information are consistent, it is determined to be a duplicate sample. Based on the duplicate samples, the candidate outliers are redundantly removed to obtain a deduplicated outlier set.

10. The method for dynamic cleaning and completion of distribution network data based on graph structure and time-series correlation as described in claim 1, characterized in that, Determining the completion and repair features based on the detected outliers and the completion candidate set specifically includes: Based on the associated data structure, the feature value range of the detected outliers is determined, and the rule box space is determined; All distribution network operation data falling within the rule box space are selected to form a repair reference sample subset. The Mahalanobis distance quantization similarity is calculated, and the Mahalanobis distance quantization result is determined as the completion and repair feature.

Images