Transformer area voltage data preprocessing method and device based on space-time double channels and medium

Abstract

The invention discloses a space-time dual-channel-based transformer area voltage data preprocessing method and device and a medium, obvious noise and missing values in original data are removed through unsupervised cleaning, input quality is ensured, space-time data processing is performed by constructing a space-time dual-channel model, a time sequence channel captures progressive anomaly of single-point equipment, and the time sequence channel captures the time sequence of the single-point equipment. The graph channel recognizes relevance faults based on power grid topology, can cover full scenes of independent anomalies and group anomalies, and can effectively eliminate noise and pseudo anomalies through dual quality control of state estimation physical verification and space-time continuity screening. The rationality of abnormal data captured by a space-time dual-channel model is verified through a state estimation correction loop, full-scene coverage of independent abnormity and group abnormity can be realized, original single-point, single-time and single-machine low-precision cleaning is upgraded to space-time dual-channel high-precision collaborative cleaning, the quality of electric energy quality data is improved, and the method is suitable for large-scale popularization and application. And the data preprocessing precision is improved.

Description

Technical Field

[0001] This invention relates to the field of big data acquisition and preprocessing technology, specifically to a method, device, and medium for preprocessing transformer area voltage data based on spatiotemporal dual channels. Background Technology

[0002] With the advancement of new power system construction, the number of distribution substations has exploded, leading to increasingly prominent power quality issues such as voltage, harmonics, and three-phase imbalance. To effectively manage power quality in these substations, the industry generally adopts a data-driven + model optimization approach: first, preprocessing massive amounts of measurement data; then calculating the voltage-harmonic-imbalance weight matrix; and finally, performing global voltage and reactive power optimization and substation-side rolling voltage regulation. However, existing substation voltage data preprocessing methods still follow the traditional single-channel approach, with main processes including outlier removal, missing value imputation, and standardization. These methods treat each substation as an isolated time series, failing to fully utilize the inherent spatial topology and long-range temporal coupling characteristics of the distribution network, resulting in the following shortcomings:

[0003] The lack of consideration for the electrical coupling relationship between adjacent transformer areas on the same feeder can easily lead to misjudging regional voltage events as global anomalies. Furthermore, the neglect of spatial correlation during missing value filling can result in significant estimation errors. Fixed-length sliding windows are difficult to adapt to slow disturbances spanning several hours, and the accuracy of anomaly detection decreases with seasonal and load pattern variations. The sampling intervals of smart meters, FTUs, and power quality monitoring devices are inconsistent, and traditional interpolation methods directly reduce or increase the frequency, introducing spurious fluctuations and losing details of harmonic transients.

[0004] Poor computational scalability: The complexity of loading the entire dataset into memory on a single machine is insufficient to handle massive data volumes, and a single data cleaning operation takes several hours, making it impossible to support daily rolling optimization needs.

[0005] Therefore, there is an urgent need to study a "spatiotemporal dual-channel" transformer voltage data preprocessing method that can simultaneously utilize spatial topological adjacency relationships and long-term temporal evolution characteristics, in order to improve data quality, computational efficiency, and the accuracy of subsequent optimization decisions. Summary of the Invention

[0006] The present invention addresses the problem that existing transformer substation voltage preprocessing uses single-point time series, which leads to large estimation errors and is prone to heterogeneous sampling distortion and false deletion of clustered anomalies. The aim is to provide a transformer substation voltage data preprocessing method, equipment, and medium based on a spatiotemporal dual-channel approach. This method removes obvious noise and missing values ​​from the original data through unsupervised cleaning, ensuring input quality. The spatiotemporal dual-channel model effectively removes noise and false anomalies. The time series channel captures progressive anomalies of single-point devices, while the graph channel identifies correlated faults based on the power grid topology, effectively removing noise and false anomalies. It provides full-scene coverage of both independent and group anomalies, upgrading the previous low-precision cleaning (single-point, single-time, single-machine) to high-precision collaborative cleaning using spatiotemporal dual channels, GPU parallelism, and joint verification of anomaly states. This improves the quality of power quality data and the accuracy of data preprocessing.

[0007] This invention is achieved through the following technical solution:

[0008] The first aspect of this invention provides a method for preprocessing transformer area voltage data based on spatiotemporal dual-channel technology, comprising the following specific steps:

[0009] Collect data from the target monitoring area and perform unsupervised data cleaning on the collected data;

[0010] Based on the cleaned data, the original voltage sequences and corresponding external covariate matrices of N transformer substations in the target monitoring area during time period T are extracted.

[0011] Spatiotemporal alignment is performed on the original voltage sequences of N transformer substations in the target monitoring area during time period T and the corresponding external covariate matrices to form a two-dimensional feature tensor covering all transformer substations and all time periods.

[0012] A spatiotemporal dual-channel anomaly detection model is built based on the two-dimensional feature tensor. The spatiotemporal dual-channel anomaly detection model is composed of a time-series channel and a graph channel connected in parallel, and is used to output the joint anomaly score of each station area at each time point.

[0013] Extract the joint anomaly score for each transformer area at each time point, and extract the anomaly candidate set that exceeds the dynamic threshold;

[0014] The anomaly candidate set is fed into the state estimation correction loop;

[0015] The data after state estimation correction is filtered again according to temporal continuity and graph node adjacency continuity, and a reliable joint anomaly label for each station area at each time point is output.

[0016] Furthermore, the unsupervised data cleaning of the collected data specifically includes:

[0017] An unsupervised data cleaning model is constructed, which consists of a local feature self-extraction network composed of a sparse autoencoder and principal component analysis, and is used to extract local features from the collected data and perform dimensionality reduction representation.

[0018] Clustering is performed based on the local features after dimensionality reduction, and outlier data represented by peripheral discrete points is removed based on the clustering results.

[0019] Furthermore, based on the cleaned data, the extraction of the original voltage sequences and corresponding external covariate matrices for N transformer substations in the target monitoring area during time period T specifically includes:

[0020] N public distribution transformer substations are selected within the target monitoring area. A continuous observation period T is set, and the original voltage operation sequence of each substation during the period is collected to obtain N one-dimensional time series of length T.

[0021] For the same time period T, external covariates related to voltage operation are collected synchronously, including meteorological elements at each time step, load levels of distribution areas, holiday markers, and adjacency vectors transformed from the distribution network topology. The four types of covariates are aligned along the time axis and then concatenated into a multidimensional external covariate matrix to obtain the voltage sequence-covariate matrix data pair corresponding to each distribution area.

[0022] Furthermore, the construction of the spatiotemporal dual-channel anomaly detection model based on the two-dimensional feature tensor specifically includes:

[0023] Construct time-series channels and graph channels;

[0024] Time series channel: The time series slices of each area are reconstructed using an autoencoder with an attention mechanism, the reconstruction error is calculated and converted into a time series anomaly score;

[0025] Graph Channel: A topology graph is constructed with transformer substations as graph nodes and electrical distance as edge weights. Graph embedding algorithm is used to extract the spatial vectors of nodes, and spatial anomaly scores are calculated by the cosine similarity between a node and its k-hop neighbors.

[0026] The temporal anomaly scores and spatial anomaly scores are merged using a weighted voting strategy to generate a joint anomaly score, and anomaly candidate sets above the dynamic threshold are extracted.

[0027] Furthermore, the anomaly candidate set is fed into the state estimation correction loop, specifically including:

[0028] Receive an anomaly candidate set generated from the initial state estimate, the anomaly candidate set containing one or more measurement data identified as potentially anomaly;

[0029] The anomaly candidate set is input into a state estimation correction loop, wherein the state estimation correction loop is configured to perform an iterative closed-loop processing procedure.

[0030] Furthermore, the state estimation correction loop performs an iterative closed-loop processing procedure including:

[0031] Observability analysis is performed on the power grid area associated with the anomaly candidate set, and the correctness of the network topology is verified based on the measurement residual distribution.

[0032] If a network topology error is identified, the network model is corrected, and the anomaly candidate set is updated based on the corrected network model.

[0033] After confirming that the network topology is correct, bad data is identified and marked from the set of abnormal candidates;

[0034] Remove or reduce the weight of the bad data from the measurement set for state estimation;

[0035] The state estimation calculation is re-executed using the updated measurement set to obtain new system state estimates;

[0036] Determine whether the new state estimate has converged and whether the corresponding measurement residual is lower than a preset threshold.

[0037] If the judgment result is yes, the final system state estimation result is output; if the judgment result is no, the steps are repeated to identify and mark bad data from the abnormal candidate set.

[0038] Furthermore, the data corrected by state estimation is further filtered based on temporal continuity and the adjacency continuity of graph nodes, specifically including:

[0039] Receive a power grid data sequence after state estimation correction, the data sequence containing measurement data of multiple transformer areas at consecutive time points;

[0040] The data is subjected to spatiotemporal continuity analysis, the analysis including:

[0041] For the same transformer area, detect whether its abnormal states occur consecutively over time.

[0042] For the same point in time, detect whether the abnormal distribution area and the adjacent distribution area in the power grid topology are abnormal at the same time;

[0043] Based on the results of the spatiotemporal continuity analysis, a reliable joint anomaly marker for each station area at each time point is output.

[0044] Furthermore, the selection criteria based on the spatiotemporal continuity analysis specifically include retaining only anomalous samples that meet one of the following conditions:

[0045] Anomalies occurring consecutively in time; or,

[0046] Anomalies that occur simultaneously with adjacent nodes in space.

[0047] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for preprocessing transformer voltage data based on a spatiotemporal dual-channel architecture.

[0048] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for preprocessing transformer area voltage data based on a spatiotemporal dual-channel architecture.

[0049] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0050] Unsupervised cleaning removes obvious noise and missing values ​​from the original data, ensuring input quality. A spatiotemporal dual-channel model is used, with the time-series channel capturing progressive anomalies of single-point devices and the graph channel identifying correlated faults based on the power grid topology. This provides full-scenario coverage of both independent and group anomalies. Dual quality control through physical verification of state estimation and spatiotemporal continuity screening effectively removes noise and pseudo-anomalies. The rationality of the abnormal data captured by the spatiotemporal dual-channel model is verified through a state estimation correction loop, effectively identifying and eliminating pseudo-anomalies that are physically impossible. This improves the accuracy and efficiency of data processing, enabling precise detection of fault initiation and accurate location of fault root causes, significantly shortening response time and improving power supply reliability. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0052] Figure 1 This is the flowchart of the transformer area voltage anomaly detection method in this embodiment of the invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0054] As one possible implementation method, such as Figure 1 As shown, this embodiment provides a method for preprocessing transformer area voltage data based on spatiotemporal dual-channel, including the following specific steps:

[0055] Data from the target monitoring area is collected, and unsupervised data cleaning is performed on the collected data. Based on the cleaned data, the original voltage sequences and corresponding external covariate matrices of N transformer substations in the target monitoring area at time period T are extracted. Spatiotemporal alignment is performed on the original voltage sequences and corresponding external covariate matrices of the N transformer substations in the target monitoring area at time period T to form a two-dimensional feature tensor covering all transformer substations and all time periods. A spatiotemporal dual-channel anomaly detection model is built based on the two-dimensional feature tensor. The spatiotemporal dual-channel anomaly detection model consists of a time-series channel and a graph channel connected in parallel, which is used to output the joint anomaly score of each transformer substation at each time point. The joint anomaly score of each transformer substation at each time point is extracted, and anomaly candidate sets above the dynamic threshold are extracted. The anomaly candidate sets are fed into the state estimation correction loop. The data after state estimation correction is further filtered according to temporal continuity and graph node adjacency continuity, and a reliable joint anomaly label of each transformer substation at each time point is output.

[0056] This embodiment achieves a significant leap in data preprocessing accuracy by industrial data processing, industrial data acquisition, and industrial data processing of transformer area voltage. By abandoning the traditional single-point-single-time isolated approach in the cleaning process, it constructs a four-level technical link with spatiotemporal dual channels and two-level verification. Through progressive means of signal enhancement, fusion, correction, and filtering, it improves the accuracy of anomaly detection in initial power quality data. In this embodiment, a stacked dilated convolutional architecture is adopted, and the receptive field covers multiple time steps at once, so that the slow perturbations of power data in the transformer area are fully incorporated into the feature map. External covariates such as illumination, temperature, and holidays are introduced in parallel to ensure the application of the method in multiple scenarios. An undirected graph is constructed using the feeder topology adjacency list, with the nodes being transformer areas. The edge weights are dynamically updated using the voltage between nodes, so that the spatial coupling information of voltage rise / fall in the same feeder is encoded into the anomaly score for the first time, which significantly reduces false alarms of isolated points. After the time-series channel and the graph channel output anomaly scores respectively, a joint anomaly score is generated through learnable weighting, which can accurately capture this spatially correlated concurrent anomaly pattern. When the joint score is higher than the initial threshold, it is not immediately marked as an anomaly, but is sent to the state estimation correction loop to ensure the reliability of the data. Through dual-rule filtering of time continuity and graph node adjacency continuity, residual false alarms caused by fragmented flash jumps are eliminated, improving the anomaly marking accuracy in the data preprocessing process.

[0057] The following are the specific implementation methods of this embodiment.

[0058] Step 1:

[0059] After data collection, the system first performs unsupervised data cleaning to remove obvious bad data and identify and filter out seemingly reasonable but actually inconsistent latent anomalies. An unsupervised data cleaning model, consisting of a local feature extraction network composed of a sparse autoencoder and principal component analysis, is built to extract local features from the collected data and perform dimensionality reduction. Clustering is then performed based on the dimensionality-reduced local features, and anomalous data represented by peripheral discrete points is removed according to the clustering results. No manual definition of cleaning rules or labeling of normal / abnormal data samples is required. The system can automatically learn the inherent distribution of the data itself and clean accordingly, exhibiting strong adaptability and applicability to distribution networks with different regions and operating characteristics. This provides a clean input for subsequent spatiotemporal anomaly detection models, allowing the models to focus more on uncovering real power grid operational anomalies rather than being interfered with by data quality issues, thereby improving the reliability of the final output results from the source.

[0060] Step Two:

[0061] Based on the cleaned data, the original voltage sequences and corresponding external covariate matrices of N distribution transformer substations in the target monitoring area during time period T are extracted. Specifically, this includes: selecting N public distribution transformer substations in the target monitoring area, setting a continuous observation period T, collecting the original voltage operation sequence of each substation during this period, and obtaining N one-dimensional time series of length T; for the same time period T, external covariates related to voltage operation are collected synchronously, including time-by-time meteorological elements, substation load levels, holiday markers, and adjacency relationship vectors transformed from the distribution network topology. The four types of covariates are aligned along the time axis and then concatenated into a multi-dimensional external covariate matrix to obtain the voltage sequence-covariate matrix data pair corresponding to each substation.

[0062] The output voltage sequence-covariate matrix data pairs are a highly standardized data structure. This greatly facilitates the design and training of subsequent models, enabling the spatiotemporal dual-channel model to process data from all transformer substations in a unified and efficient manner.

[0063] By setting a uniform observation period T for all transformer substations and ensuring precise temporal alignment of all data, analytical biases caused by time asynchrony are eliminated. Using adjacency vectors as covariates allows for the explicit and pre-emptive injection of spatial information implicitly learned in the next graph channel into the model input. This is equivalent to informing the model of the physical connections between transformer substations at the feature level, enabling the model to perceive spatial correlations earlier and more directly, thus improving the efficiency and effectiveness of spatial feature learning.

[0064] Step 3:

[0065] A spatiotemporal dual-channel anomaly detection model is built based on two-dimensional feature tensors. This step is the core intelligent engine of the entire embodiment. It uses a parallel dual-channel architecture to mine anomaly patterns from both temporal and spatial dimensions and perform information fusion. Specifically, it includes constructing a temporal channel and a graph channel, wherein:

[0066] Time Series Channel: For each voltage time series slice of a transformer area, an autoencoder with an attention mechanism is employed. The autoencoder learns a compressed representation of the data through a bottleneck layer and attempts to reconstruct the original sequence from it. The attention mechanism allows the model to assign different importance weights to different time points in the sequence during reconstruction. The degree of data anomaly is assessed by comparing the difference between the original sequence and the reconstructed sequence (i.e., the reconstruction error). The larger the reconstruction error, the more difficult it is for the model to learn the data pattern at that time point (i.e., the more anomalous it is), which is then converted into a time series anomaly score.

[0067] Graph Channel: Each distribution area is treated as a graph node, with electrical distances between nodes (such as electrical coupling strength, impedance, or physical connection relationships) used as edge weights to construct a distribution network topology. A graph embedding algorithm is used to map each node (distribution area) to a low-dimensional vector space. This spatial vector encodes the distribution area's location and connectivity within the power grid topology. For a distribution area node, the average cosine similarity between its spatial vector and the vectors of its k-hop neighbors (i.e., all neighboring nodes within a k-hop radius) is calculated. If a distribution area's vector differs significantly in direction from the vectors of most of its neighbors (low cosine similarity), it indicates an outlier in spatial pattern, and a spatial anomaly score is derived accordingly.

[0068] Step Four:

[0069] The temporal anomaly scores and spatial anomaly scores are weighted and merged according to preset weights to generate a joint anomaly score.

[0070] In the above, the temporal channel is used to detect anomalies in individual behavior (such as slow drifts or instantaneous spikes in voltage of a single transformer substation), while the graph channel is used to detect anomalies in group relationships (such as a transformer substation exhibiting voltage trends opposite to all its neighbors, i.e., local disconnection). The attention mechanism in the temporal channel allows the model to focus on truly important abrupt changes or anomalous pattern segments in the sequence, effectively avoiding overreaction to stationary noise and improving the detection accuracy of key anomalies. Graph embedding quantifies electrical distance into spatial vectors, enabling the model to learn complex, nonlinear spatial association patterns that go beyond simple physical connections. Cosine similarity calculations can sensitively capture any discordant nodes in the network. The weighted voting fusion strategy combines evidence from both temporal and spatial perspectives, improving the reliability of the final joint judgment.

[0071] The anomaly candidate set is fed into the state estimation correction loop, specifically including:

[0072] Receive an anomaly candidate set generated from the initial state estimation from the output of the previous spatiotemporal anomaly detection model. The anomaly candidate set contains one or more measurement data (such as voltage, power, etc.) that are identified as potentially anomalies.

[0073] The anomaly candidate set is input into the state estimation correction loop, which is configured to perform an iterative closed-loop process, including: performing an observability analysis on the power grid area associated with the anomaly candidate set, first confirming whether the current measurement configuration of the power grid area is sufficient to uniquely determine the state of all nodes. If unobservable, pseudo-measurements are added, and based on the initial state estimation results, the distribution of measurement residuals (the difference between actual measured values ​​and estimated values) is analyzed. The correctness of the network topology is verified based on the distribution of measurement residuals; if the residuals exhibit a specific pattern, the network topology model (switch states) may not match reality.

[0074] If a network topology error is identified, the network model is corrected, and the anomaly candidate set is updated based on the corrected network model.

[0075] After confirming that the network topology is correct, bad data is identified and marked from the abnormal candidate set;

[0076] Remove or reduce the weight of bad data from the measurement set of state estimation so that it has less influence in subsequent calculations;

[0077] The state estimation calculation is re-executed using the updated measurement set to obtain new system state estimates;

[0078] Determine whether the new state estimate has converged and whether the corresponding measurement residual is lower than a preset threshold.

[0079] If the judgment result is yes, the final system state estimation result is output. If the judgment result is no, the steps are repeated to identify and mark bad data from the anomaly candidate set, and a new round of identification and elimination begins until the new state estimation is determined to converge. This loop can not only correct data errors, but also identify and correct errors in the power grid topology model, and can eliminate spurious anomalies caused by model inaccuracies, communication noise, etc. to a great extent.

[0080] Step 5:

[0081] The data, after being corrected for state estimation, is further filtered based on temporal continuity and the adjacency continuity of graph nodes, specifically including:

[0082] Receive the power grid data sequence after state estimation correction. The data sequence contains measurement data from multiple transformer substations at consecutive time points. Perform spatiotemporal continuity analysis on the data, including:

[0083] For the same transformer substation, a sliding time window is defined on its abnormal state time series to detect whether the abnormal state continues to occur within this window. Based on the power grid topology, transformer substations are treated as nodes, and electrical connections (such as shared busbars and shared feeders) are treated as edges, constructing a transformer substation relationship graph. For a transformer substation node marked as abnormal at the same time point, all its directly connected or neighboring nodes within a certain electrical distance are checked. The number of these neighboring nodes that are simultaneously marked as abnormal is counted.

[0084] Based on the results of spatiotemporal continuity analysis, a credible joint anomaly label for each transformer area at each time point is output. The selection criteria are to retain only anomaly samples that meet one of the following conditions: anomalies that occur consecutively in time; or anomalies that occur simultaneously with adjacent nodes in space. By further filtering the spatiotemporal continuity of the data after state estimation correction, the output is no longer the original anomaly score, but is further accompanied by a spatiotemporal evidence chain, improving its reliability.

[0085] As one possible implementation, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a method for preprocessing transformer area voltage data based on spatiotemporal dual channels.

[0086] As one possible implementation, this embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for preprocessing transformer area voltage data based on a spatiotemporal dual-channel architecture.

[0087] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for preprocessing substation voltage data based on space-time dual channels, characterized in that, The method comprises the following specific steps: Collecting target monitoring area data, and performing unsupervised data cleaning on the collected data; Based on the cleaned data, extracting the original voltage sequence of N substation areas in the target monitoring area in T period and the corresponding external covariant matrix; Based on the original voltage sequence of N substation areas in the target monitoring area in T period and the corresponding external covariant matrix, spatiotemporal alignment is performed to form a two-dimensional feature tensor covering all substation areas and all time points; Based on the two-dimensional feature tensor, a spatiotemporal dual-channel anomaly detection model is built, which is composed of a time sequence channel and a graph channel in parallel, and is used to output a joint anomaly score of each substation area at each time point; Extracting the joint anomaly score of each substation area at each time point, and extracting an anomaly candidate set higher than a dynamic threshold; The anomaly candidate set is sent to a state estimation correction loop; Based on the data corrected by the state estimation, the time continuity and the graph node adjacency continuity are re-screened to output a reliable joint anomaly label of each substation area at each time point.

2. The method of claim 1, wherein the method is a method of preprocessing the data of the voltage of the area by using the space-time dual channel. The unsupervised data cleaning on the collected data specifically comprises: Building an unsupervised data cleaning model, which is composed of a local feature self-extraction network composed of a sparse autoencoder and a principal component analysis, and is used to extract local features from the collected data and perform dimension reduction expression; Based on the dimension-reduced local features, clustering is performed, and according to the clustering results, abnormal data represented by peripheral discrete points is removed.

3. The method of claim 1, wherein the method is a method of preprocessing the data of the voltage of the area by using the space-time dual channel. The extraction of the original voltage sequence of N substation areas in the target monitoring area in T period and the corresponding external covariant matrix based on the cleaned data specifically comprises: Selecting N public distribution transformer substation areas in the target monitoring area, setting a continuous observation period T, collecting the original voltage operation sequence of each substation area in the period, and obtaining N one-dimensional time sequences with a length of T; For the same T period, the external covariants related to voltage operation are synchronously collected, including time-by-time meteorological elements, substation load levels, holiday labels, and adjacency relationship vectors converted from the distribution network topology structure, and the four types of covariants are aligned along the time axis and spliced into a multi-dimensional external covariant matrix to obtain a voltage sequence-covariant matrix data pair corresponding to each substation area.

4. The method of claim 1, wherein the method is a method of preprocessing data of a district voltage based on a space-time dual channel. The building of the spatiotemporal dual-channel anomaly detection model based on the two-dimensional feature tensor specifically comprises: Constructing a time sequence channel and a graph channel; Time sequence channel: reconstructing each substation area time sequence slice using an autoencoder with attention mechanism, calculating the reconstruction error and converting it into a time sequence anomaly score; Graph channel: constructing a topology graph with substation areas as graph nodes and electrical distance as edge weight, extracting node spatial vectors using graph embedding algorithm, and calculating spatial anomaly score through cosine similarity of nodes and their k-hop neighbors; Combining the time sequence anomaly score and the spatial anomaly score according to the weighted voting strategy to generate a joint anomaly score, and extracting an anomaly candidate set higher than a dynamic threshold.

5. The method of claim 1, wherein the method is a method of preprocessing the data of the voltage of the area by using the space-time dual channel. The anomaly candidate set is sent to a state estimation correction loop, which specifically comprises: Receiving an anomaly candidate set generated by initial state estimation, the anomaly candidate set containing one or more measurement data identified as possible anomalies; inputting the abnormal candidate set into a state estimation correction loop, wherein the state estimation correction loop is configured to perform an iterative closed-loop processing procedure.

6. The method of claim 5, wherein the method is a method of preprocessing the data of the voltage of the area by using the space-time dual channel. The state estimation correction loop performs an iterative closed-loop processing procedure, which includes: performing observability analysis on the power grid area associated with the abnormal candidate set, and verifying the correctness of the network topology based on the measurement residual distribution; if a network topology error is identified, correcting the network model, and updating the abnormal candidate set based on the corrected network model; after confirming the correctness of the network topology, identifying and labeling bad data from the abnormal candidate set; eliminating or reducing the weight of the bad data from the measurement set of the state estimation; re-executing the state estimation calculation using the updated measurement set to obtain a new system state estimation value; judging whether the new state estimation converges and the corresponding measurement residual is lower than a preset threshold; if the judgment result is yes, outputting the final system state estimation result, and if the judgment result is no, re-executing the step of identifying and labeling bad data from the abnormal candidate set.

7. The method of claim 1, wherein the method is a method of preprocessing of substation voltage data based on space-time dual channels. The state estimation corrected data is screened again according to time continuity and graph node adjacency continuity, specifically including: receiving a power grid data sequence after state estimation correction, the data sequence containing measurement data of multiple areas at consecutive time points; performing spatiotemporal continuity analysis on the data, which includes: for the same area, detecting whether its abnormal state continuously appears in the time dimension; for the same time point, detecting whether the abnormal area and the adjacent area simultaneously appear abnormal in the power grid topology structure; based on the results of the spatiotemporal continuity analysis, screening and outputting the trusted joint abnormality label of each area at each time point.

8. The method according to claim 7, wherein, The conditions for screening based on the results of the spatiotemporal continuity analysis specifically include retaining only abnormal samples that meet one of the following conditions: abnormalities that continuously appear in time; or abnormalities that simultaneously appear with adjacent nodes in space.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the area voltage data preprocessing method based on the time-space dual channel according to any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the area voltage data preprocessing method based on the time-space dual channel according to any one of claims 1-8.

Images