A method, system, equipment, and medium for optimizing the processing of forced data transmission operations.

Abstract

This invention discloses a method, system, device, and medium for optimizing data processing in forced transmission operations. The method includes: constructing a multi-dimensional matrix framework model; filling the coordinate nodes corresponding to the first standardized data stream into the multi-dimensional matrix framework model to form a first matrix data block; performing a hierarchical parsing operation to extract time-layer feature tensors, spatial-layer feature tensors, electrical-layer feature tensors, and logical-layer feature tensors; performing a fusion operation on the extracted feature tensors using a preset set of fusion coefficients to generate first target feature data; converting the first target feature data into a tensor structure to be compressed and performing a decomposition operation; selecting a data decomposition strategy based on the operation results; decomposing the tensor structure to be compressed into multiple low-dimensional core factors according to the data decomposition strategy; and performing quantization and sparse coding operations on the low-dimensional core factors to obtain the final optimized data packet. This invention provides high-quality, low-latency data support for forced transmission operation decisions.

Description

Technical Field

[0001] This invention relates to the field of power system data processing and intelligent analysis technology, and in particular to a method, system, equipment and medium for optimizing the processing of forced transmission operation data. Background Technology

[0002] In power system fault handling, forced power restoration refers to the operation performed by operators to quickly restore power to tripped equipment or further determine the nature of the fault after a grid fault trips. The correctness and timeliness of this decision directly affect the safe and stable operation of the power grid and the reliability of power supply. Forced power restoration decisions require the simultaneous acquisition of multiple data sources, including real-time grid operation data, equipment status monitoring data, protection action information, fault waveform data, meteorological and environmental data, and historical case data. These data differ significantly in transmission protocols, data formats, sampling frequencies, and spatiotemporal reference systems. Traditional methods typically acquire data from each source sequentially, encountering efficiency bottlenecks at the data access level. Furthermore, each data source operates independently, with varying time synchronization accuracy and spatial positioning methods. This inconsistency in spatiotemporal reference systems necessitates complex alignment processing before data fusion and analysis, further increasing preprocessing overhead. In terms of feature extraction, existing data processing methods mostly adopt single-dimensional analysis or simple data splicing, tending to process each data source independently before summarizing the results. This severs the inherent connection between time evolution, spatial propagation, electrical logic and equipment status, and fails to fully explore the deep spatiotemporal correlation and coupling relationship between data, resulting in limited feature expression capabilities and difficulty in comprehensively depicting the complex physical process under the strong delivery operation scenario.

[0003] Furthermore, due to the large volume, high dimensionality, and serialized processing flow of traditional methods, the process from multi-source data acquisition to generating usable features is typically time-consuming, making it difficult to meet the actual demand for second-level response in emergency power system fault handling. Processing flows designed for specific scenarios are also difficult to adapt to the needs of strong transmission data analysis under different voltage levels, different fault types, and different power grid structures. When the scenario changes, it is often necessary to redesign the processing flow or manually adjust the parameters, lacking adaptive capabilities. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method, system, device and medium for optimizing the processing of forced delivery operation data, which solves the problems of low processing efficiency, weak feature representation ability and difficulty in adapting to different application scenarios caused by the spatiotemporal inconsistency of multi-source data and insufficient correlation analysis capability in existing forced delivery operation data processing methods.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for optimizing data processing during forced delivery operations, comprising: Acquire multi-source heterogeneous data and align it to a unified spatiotemporal reference frame to generate a first standardized data stream; Construct a multidimensional matrix framework model, and fill the coordinate nodes corresponding to the multidimensional matrix framework model with each data value in the first standardized data stream to form a first matrix data block; A hierarchical parsing operation is performed on the first matrix data block to extract time layer feature tensors, spatial layer feature tensors, electrical layer feature tensors and logic layer feature tensors. A fusion operation is performed on the extracted feature tensors using a preset set of fusion coefficients to generate the first target feature data. The first target feature data is converted into a tensor structure to be compressed and decomposition operation is performed. A data decomposition strategy is selected based on the operation result. The tensor structure to be compressed is decomposed into multiple low-dimensional core factors according to the data decomposition strategy. Quantization and sparse coding operations are performed on the low-dimensional core factors to obtain the final optimized data packet.

[0007] As a preferred embodiment of the forced delivery operation data optimization processing method described in this invention, the generation of the first standardized data stream includes: Obtain the original timestamps corresponding to the multi-source heterogeneous data, compare the original timestamps with a preset time reference protocol, determine the time deviation value, and perform time axis translation and alignment on the data records according to the time deviation value; The device number contained in the multi-source heterogeneous data is extracted, the physical geographical location corresponding to the device number is determined according to the power grid topology location information, the multi-source heterogeneous data is mapped to the spatial node corresponding to the physical geographical location, and a first standardized data stream with a unified spatiotemporal label is generated.

[0008] As a preferred embodiment of the forced data processing method of the present invention, the step of forming the first matrix data block includes: Based on the scope of the forced delivery task and the type of access data source, the set of coordinate axes for each dimension of the multidimensional matrix framework model is dynamically determined; Traverse each data record in the first standardized data stream and fill each data value in the first standardized data stream into the coordinate node corresponding to the multidimensional matrix framework model; When multiple values ​​appear at the same coordinate node, select the primary value according to the data source priority and record the deviation; Perform a coordinate-by-coordinate scan on the filled multidimensional matrix. For missing coordinates, use interpolation, nearest neighbor estimation, or marking as unavailable according to the missing ratio. Perform a dimensionless transformation on the values ​​of each coordinate node after missing data processing, and output the first matrix data block.

[0009] As a preferred embodiment of the forced delivery operation data optimization processing method described in this invention, the generation of the first target feature data includes: A sliding window is performed on the first matrix data block along the time axis to extract the numerical rate of change, spectral components and statistics within each time window, and the data are spliced ​​together to form a time layer feature tensor. The topological neighborhood of the first matrix data block is expanded layer by layer along the device dimension with the faulty device as the center, and the electrical quantity distribution and inter-layer attenuation at each level are calculated to form a spatial layer feature tensor. The first matrix data block is grouped according to physical attributes along the electrical quantity dimension, and the proportional relationship, phase difference and changes before and after the fault are analyzed within the group to form an electrical layer feature tensor; the first matrix data block is analyzed along the state dimension to parse the protection action sequence, equipment state transition path and setting margin to form a logic layer feature tensor. The time layer feature tensor, spatial layer feature tensor, electrical layer feature tensor, and logic layer feature tensor are cross-correlated and calculated along their respective dimensions to generate cross-layer coupling features; Using a preset set of fusion coefficients, the extracted feature tensor and the cross-layer coupling features are weighted and fused to obtain a fused feature vector. The correlation between each feature dimension in the fused feature vector and the preset strong result label is calculated. The feature dimensions with a correlation degree higher than a preset threshold are retained as valid features. Principal component transformation is performed on the retained valid feature set, and principal components with a cumulative contribution rate that meets the preset requirements are selected as the final feature set. The feature vectors of the standardized final feature set are packaged and output with the corresponding metadata information to generate the first target feature data.

[0010] As a preferred embodiment of the forced delivery operation data optimization processing method of the present invention, the step of converting the first target feature data into a tensor structure to be compressed and performing decomposition operation, selecting a data decomposition strategy according to the operation result, and decomposing the tensor structure to be compressed into multiple low-dimensional core factors according to the data decomposition strategy, including: The standardized feature tensors in the first target feature data are converted into standard tensor format and bound to metadata to obtain the tensor structure to be compressed. Analyze the dimensional distribution characteristics of the tensor structure to be compressed, and select one of the preset decomposition strategies as the data decomposition strategy for the current tensor to be compressed based on the combination of dimensional scale balance index and order interval. Depending on the type of data decomposition strategy, single-parameter traversal optimization, multi-dimensional independent rank determination, or dimension-by-dimensional chain truncation are performed respectively to determine the optimal rank parameter. Based on the data decomposition strategy and the optimal rank parameter, the tensor structure to be compressed is decomposed to obtain multiple low-dimensional core factors.

[0011] As a preferred embodiment of the forced data transmission optimization processing method described in this invention, the step of performing quantization and sparse coding operations on the low-dimensional core factors to obtain the final optimized data packet includes: Numerical distribution statistics are performed on the multiple low-dimensional core factors respectively, and the number of quantization bits and interval division method are determined according to the precision sensitivity of each type of factor; Based on the number of quantization bits, the continuous values ​​of each element in various low-dimensional core factors are mapped to discrete quantization intervals, and the original values ​​are replaced by the index values ​​of the quantization intervals to which each element belongs, thus completing the conversion from floating-point values ​​to integer indices. The quantized sequence is scanned for consecutive identical index values, which are encoded into a pair consisting of the index value and the number of repetitions. Frequency statistics are performed on the pairs, and a variable-length encoding table is constructed. Each pair is replaced with the corresponding variable-length codeword to generate the final encoded data stream. According to the preset encapsulation format, the strategy identifier, dimension information, quantization parameters and encoding table are written into the packet header, the encoded data stream is written into the packet body, and the final optimized data packet is packaged and output.

[0012] As a preferred embodiment of the forced data processing method described in this invention, the method further includes real-time incremental updates and feature library construction. By monitoring changes in each data source through event subscription and periodic polling, the time interval, device set, and data source type of the changed data are parsed to generate a change area descriptor. The change type is determined by comparing the time interval in the changed region descriptor with the time range covered by the current multidimensional matrix, and local feature recalculation is performed only on the affected matrix slices and adjacent windows to obtain the updated feature vector. The updated feature vector is compared with historical features in the feature library to obtain the maximum similarity value; In response to the maximum similarity value being lower than a preset similarity threshold, the currently updated feature vector is stored as a new feature and statistical information is initialized; in response to the maximum similarity value being not lower than the preset similarity threshold, the currently updated feature vector is marked as a non-new feature of the same type as the matched historical features. The newly stored feature vectors are incrementally inserted into the index structure to update the vector index. When the total number of features in the feature library increases by more than a preset ratio compared to the number during the last full index reconstruction, an asynchronous background task is triggered to perform a full index reconstruction on all feature vectors in the feature library. Regularly perform cluster analysis on all features in the database, extract cluster centers with sufficient samples as general patterns, and statistically analyze metadata to generate typical scenario labels.

[0013] Secondly, the present invention provides a forced delivery operation data optimization processing system, comprising: The data spatiotemporal alignment module is used to acquire multi-source heterogeneous data and align it to a unified spatiotemporal reference system to generate a first standardized data stream. A multidimensional matrix construction module is used to construct a multidimensional matrix framework model by filling each data value in the first standardized data stream into the coordinate nodes corresponding to the multidimensional matrix framework model to form a first matrix data block. The hierarchical feature parsing module is used to perform hierarchical parsing operations on the first matrix data block, extract time layer feature tensors, spatial layer feature tensors, electrical layer feature tensors and logic layer feature tensors, and perform fusion operations on the extracted feature tensors using a preset fusion coefficient set to generate the first target feature data; The compression encoding module is used to convert the first target feature data into a tensor structure to be compressed and perform decomposition operations. Based on the operation results, a data decomposition strategy is selected, and the tensor structure to be compressed is decomposed into multiple low-dimensional core factors according to the data decomposition strategy. Quantization and sparse encoding operations are performed on the low-dimensional core factors to obtain the final optimized data packet.

[0014] Thirdly, the present invention provides an electronic device, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor, when executing the computer-executable instructions, implements the steps of a forced data optimization processing method.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of a forced data optimization processing method.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention integrates multi-source heterogeneous data into the same spatiotemporal framework by constructing a unified multidimensional matrix model, effectively eliminating the fusion barriers caused by differences in data format, sampling frequency, and spatiotemporal reference system. On this basis, a hierarchical parallel parsing strategy is adopted to simultaneously extract features from four dimensions: time, space, electrical logic, and equipment status, and perform cross-layer correlation analysis. This fully explores the deep coupling relationship between data, overcomes the shortcomings of traditional single-dimensional analysis or simple splicing that result in insufficient feature expression capabilities, and significantly reduces data processing and storage overhead by combining tensor decomposition and incremental update mechanisms. Compared with traditional methods, it can significantly shorten the entire process time from data acquisition to feature output, and has the ability to adaptively process different voltage levels, fault types, and power grid structures, providing high-quality, low-latency data support for forced transmission operation decisions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the overall process logic of a forced delivery operation data optimization processing method provided in one embodiment of the present invention.

[0019] Figure 2 A schematic diagram of the structure of the first matrix data block of the forced delivery operation data optimization processing method provided in an embodiment of the present invention (showing four dimensions: T, S, C, and K).

[0020] Figure 3 The detailed process and data flow diagram of the intelligent hierarchical parsing of the forced delivery operation data optimization processing method provided in one embodiment of the present invention are shown below. Detailed Implementation

[0021] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0022] Example 1, referring to Figure 1 As one embodiment of the present invention, a method for optimizing data processing during forced transmission operations is provided, such as... Figure 1 The specific steps shown are as follows: S100: Acquire multi-source heterogeneous data and align it to a unified spatiotemporal reference frame to generate the first normalized data stream; S200: Construct a multidimensional matrix framework model, fill the data values ​​in the first standardized data stream into the coordinate nodes corresponding to the multidimensional matrix framework model, and form the first matrix data block; S300: Perform a hierarchical parsing operation on the first matrix data block to extract the time layer feature tensor, spatial layer feature tensor, electrical layer feature tensor and logic layer feature tensor. Use a preset set of fusion coefficients to perform a fusion operation on the extracted feature tensors to generate the first target feature data. S400: Convert the first target feature data into a tensor structure to be compressed and perform decomposition operation. Select a data decomposition strategy based on the operation result. Decompose the tensor structure to be compressed into multiple low-dimensional core factors according to the data decomposition strategy. Perform quantization and sparse coding operations on the low-dimensional core factors to obtain the final optimized data packet. It should be noted that this invention integrates multi-source heterogeneous data into the same spatiotemporal framework by constructing a unified multidimensional matrix model, effectively eliminating fusion barriers caused by differences in data format, sampling frequency, and spatiotemporal reference system. On this basis, a hierarchical parallel parsing strategy is adopted to simultaneously extract features from four dimensions: time, space, electrical logic, and equipment status, and perform cross-layer correlation analysis. This fully explores the deep coupling relationships between data, overcoming the shortcomings of traditional single-dimensional analysis or simple splicing that result in insufficient feature representation capabilities. At the same time, the combination of tensor decomposition and incremental update mechanisms significantly reduces data processing and storage overhead. Compared with traditional methods, it can significantly shorten the entire process time from data acquisition to feature output, and has the ability to adaptively process different voltage levels, fault types, and power grid structures, providing high-quality, low-latency data support for forced transmission operation decisions.

[0023] Example 2, refer to Figure 2 and Figure 3 Based on the previous embodiment, this embodiment provides a specific implementation method for optimizing the processing of forced delivery operation data.

[0024] In this embodiment of the invention, step S100, which involves acquiring multi-source heterogeneous data and aligning it to a unified spatiotemporal reference system to generate a first standardized data stream, includes the following sub-steps A1 and A2: In A1: Obtain the original timestamps corresponding to the multi-source heterogeneous data, compare the original timestamps with the preset time base protocol, determine the time deviation value, and perform time axis translation and alignment on the data records according to the time deviation value; Specifically, the system accesses data sources such as SCADA system, fault recording device, meteorological monitoring system and equipment ledger through multi-protocol adapter interfaces in parallel. It performs protocol parsing and format conversion on the raw data output from each data source and extracts the original timestamp information carried by each data record.

[0025] Specifically, the time synchronization unit compares the original timestamps of each data record with a preset time base protocol, calculating the time deviation of each record relative to the unified time base. For records with deviations within the allowable range, their original timestamps are directly corrected to the corresponding time on the unified timeline. For records with deviations exceeding the allowable range, software interpolation compensation is performed based on historical time synchronization accuracy records from the data source, ultimately synchronizing the timestamps of all data records to the same time base. Based on this, an integrity assessment is performed on each time-aligned data record to check for missing or abnormal data, and problematic data points are marked with quality indicators.

[0026] In A2: Extract the device number contained in the multi-source heterogeneous data, determine the physical geographical location corresponding to the device number based on the power grid topology location information, map the multi-source heterogeneous data to the spatial node corresponding to the physical geographical location, and generate the first standardized data stream with a unified spatiotemporal label.

[0027] Specifically, the spatial alignment engine extracts the device number or device name information carried in each data record, and determines the physical geographical location and electrical node corresponding to each device number based on the power grid topology connection relationship and device ledger information. For electrical quantity data, it directly maps it to the spatial coordinates of the corresponding device node in the power grid topology; for non-electrical quantity data such as meteorological data, it associates the geographical location of the meteorological monitoring point with the spatial distance of each electrical device node, mapping the meteorological data to the nearest electrical device node or distributing it to multiple surrounding nodes through spatial interpolation.

[0028] Specifically, after completing the spatial location mapping, the spatiotemporal coordinate information of each data record is bound to the numerical content to generate a unified spatiotemporal label for each data record, and finally outputs a first standardized data stream with a consistent time base and clear spatial attribution.

[0029] It should be noted that step S100 above breaks down the barriers of independent operation and incompatible formats of various data sources by parallel access to multi-source heterogeneous data and unified spatiotemporal alignment, and eliminates the fusion obstacles caused by inconsistent time bases and spatial positioning differences, so that the originally scattered and fragmented data can achieve collaborative representation under the same spatiotemporal framework.

[0030] In this embodiment of the invention, step S200 above constructs a multidimensional matrix framework model, filling each data value in the first standardized data stream into the coordinate nodes corresponding to the multidimensional matrix framework model to form a first matrix data block, and includes the following sub-steps B1-B5 in detail: In B1: Based on the scope of the forced delivery task and the type of access data source, dynamically determine the set of coordinate axes for each dimension of the multidimensional matrix framework model; Specifically, the value range of each dimension of data is dynamically determined based on the actual physical boundaries of the forced delivery operation task and the system configuration parameters. The time dimension T ranges from the fault initiation time to the completion of the forced transmission operation within a time window. The minimum time granularity is determined by the data source with the highest sampling frequency. The total length covers N cycles before the fault to M minutes after the fault. N and M are preset in the system configuration based on the voltage level and fault type. The equipment dimension S ranges from all circuit breakers, busbars, transformers, and adjacent feeders within the tripping range that are electrically at least L away from the fault point. L is automatically calculated based on the power grid topology. The electrical quantity dimension C ranges from, but is not limited to, three-phase voltage amplitude, three-phase current amplitude, active power, reactive power, frequency, zero-sequence voltage, zero-sequence current, and negative-sequence components. The specific number of channels is automatically enumerated from the data source access list. The environmental dimension E ranges from meteorological parameters such as temperature, humidity, wind speed, rainfall, and lightning intensity, which are derived from the real-time data interface of the meteorological monitoring system. The status dimension K ranges from the current operating status of each device and the activation / deactivation status of the protection pressure plate. The set of values ​​is defined by the real-time remote signaling point table of the SCADA system.

[0031] In B2: Traverse each data record in the first normalized data stream and fill each data value in the first normalized data stream into the coordinate nodes corresponding to the multidimensional matrix framework model; In B3: When multiple values ​​appear for the same coordinate node, the primary value is selected according to the data source priority and the deviation is recorded; In B4: Perform a coordinate-by-coordinate scan on the filled multidimensional matrix. For missing coordinates, use interpolation, nearest neighbor estimation, or marking as unavailable according to the missing ratio. In B5: Perform a dimensionless transformation on the values ​​of each coordinate node after missing data processing, and output the first matrix data block; In this embodiment of the invention, the specific implementation steps for abstracting the forced delivery operation scenario into a multi-dimensional matrix are as follows: Based on the above rules for determining the range of values ​​for each dimension, an ordered set of coordinate axes for each dimension is constructed. The time dimension T is arranged according to equally spaced sampling points, the equipment dimension S is arranged according to the grid topology adjacency relationship, the electrical quantity dimension C is grouped according to the physical quantity type, the environmental dimension E is arranged according to the data source access order, and the state dimension K is arranged according to the device ID index. Traverse each record in the spatiotemporally aligned standardized data stream, determine its coordinate position on the time axis based on the time stamp information carried by the record, determine its coordinate position on the device axis based on the device ID associated with the record, determine its coordinate position on the electrical quantity axis based on the physical quantity type of the record, determine its coordinate position on the environment axis or state axis based on the data source of the record, and fill the corresponding coordinates in the multidimensional matrix with the values ​​in the record. When multiple values ​​exist at the same coordinate position, they are merged according to the following multi-value processing rules to ensure that each coordinate position corresponds to a unique value. After traversing all data records, a normalization operation is performed on all elements in the matrix to eliminate dimensional differences between physical quantities, ultimately generating a standardized first matrix data block. ,like Figure 2 As shown.

[0032] Specifically, when multiple values ​​from different data sources or multiple acquisition periods exist at the same matrix coordinate position, the system performs fusion processing according to the following priority rules: If multiple values ​​come from the same type of sensor or the same source system, the arithmetic weighted average is used as the final value of the coordinate position, and the weighting coefficients are dynamically allocated according to the reliability score of each data source. If multiple values ​​include values ​​from the waveform data source, the value from the waveform data source shall be used first, and the values ​​from other sources shall be used as verification references and the deviation shall be recorded. If there are significant differences between multiple values ​​(the deviation exceeds the preset threshold TH, which defaults to 5% of the range and is configurable), a data consistency alarm will be triggered, the coordinate position will be marked as pending verification, and all original values ​​will be retained for manual review. At the same time, the median will be temporarily used as the fill value for this position to ensure the integrity of the matrix. The upper-level logic will decide whether to replace it when the subsequent module processes it.

[0033] Furthermore, the missing value processing unit works by scanning the original multidimensional matrix M element by element, marking the coordinates of all missing or NULL values, and calculating the missing rate in each dimension. If the missing rate For values ​​≤5%, linear interpolation is used to fill in missing values ​​along the time dimension; if 5% < ≤20%, the K-nearest neighbor algorithm (K=5) is used to fill in the missing position using a weighted average of the values ​​of the nearest points in the device and electrical quantity dimensions; if If the data loss rate is greater than 20%, it is determined that there is a systematic data gap in that dimension or data source. Instead of automatic filling, the coordinate position is marked as a "data unavailable" placeholder, and a missing data report is generated and fed back to the data acquisition layer, prompting a check of the data quality of the corresponding data source.

[0034] In an optional embodiment, the generation of the first matrix data block can also be based on a strategy of prioritizing data source type grouping. SCADA system data, fault recording data, meteorological data, etc., are grouped into the corresponding data source dimensions according to their sources. Within each group, sub-matrices are independently constructed using time as the index, and then the dimensions are spliced ​​together to form a complete five-dimensional matrix.

[0035] In an optional embodiment, the generation of the first matrix data block can also be carried out in a spatiotemporal grid pre-division manner. First, sampling nodes are defined at equal intervals on the time axis according to the range of the forced delivery task, and a device node grid is generated on the spatial axis according to the topological distance to form an empty matrix skeleton. Then, the spatiotemporal coordinates in each data record are used as search keys to locate and fill in the values ​​in the skeleton. If there is no corresponding data at a certain grid position, it is directly marked as missing for subsequent processing.

[0036] It should be noted that step S200 above solves the problem of losing interdimensional correlation information in traditional data splicing methods, enabling time evolution, equipment topology, electrical physical quantities and state variables to maintain an inherent coupling relationship under a unified mathematical structure. At the same time, the integrity and numerical consistency of the matrix are guaranteed through missing data processing and normalization operations, providing a data carrier with a clear structure and lossless information for deep feature mining.

[0037] In this embodiment of the invention, step S300 performs a hierarchical parsing operation on the first matrix data block to extract time-layer feature tensors, spatial-layer feature tensors, electrical-layer feature tensors, and logic-layer feature tensors. A preset set of fusion coefficients is then used to perform a fusion operation on the extracted feature tensors to generate the first target feature data. This includes the following sub-steps C1-C7: In C1: A sliding window is performed on the first matrix data block along the time axis. The numerical rate of change, spectral components and statistics are extracted in each time window and then spliced ​​together to form a time layer feature tensor. Specifically, the first matrix data block is sliced ​​using a multi-scale sliding window along the time dimension T, with three scale windows set: short-time window, long-time window, and long-time window. (Corresponding to 10 cycles, approximately 0.2 seconds, used to capture transient characteristics of faults), mid-time window (Corresponding to 300 sampling points, used to capture the transient evolution of the fault), long-term window (Corresponding to the entire time axis T, used to capture global situational changes before and after the fault).

[0038] Specifically, within each time window, the following calculations are performed on the time-series data of each device node and each electrical quantity channel: First-order and second-order differences are calculated to extract the starting point of voltage dips, the peak time of current surges, and the time of recovery to stability, forming the fault event time-series feature F_time; Fast Fourier Transform is used to extract the dominant frequency component and harmonic distortion rate of the spectrum; the mean μ and variance σ within the sliding window are calculated. 2 Peak-to-peak value and energy E form a statistical feature set; Specifically, the aforementioned temporal features are concatenated along the time axis to obtain the temporal layer feature tensor. ∈ ^(T'×S×C×E), where T' is the number of time points after downsampling via a sliding window.

[0039] In C2: The topological neighborhood of the first matrix data block is expanded layer by layer along the device dimension with the faulty device as the center. The electrical quantity distribution and inter-layer attenuation of each layer are calculated to form the spatial layer feature tensor. Specifically, spatial topology parsing is performed on the first matrix data block along the device dimension S. Taking the faulty device as the central node, the topology is expanded outward layer by layer according to the power grid connection relationship to construct a hierarchical spatial adjacency graph G=(V,E) containing the central node, first-order adjacent nodes (circuit breakers and buses directly connected to the faulty device), and second-order adjacent nodes (lines and transformers indirectly connected through first-order nodes).

[0040] Specifically, for each spatial level, the average electrical quantity of all device nodes within that level is calculated. ,variance and maximum value and minimum value ratio = / Extract the attenuation rate λ = ΔV / Δd (ΔV is the voltage amplitude change between adjacent levels, and Δd is the electrical distance difference) of the fault propagation along space; calculate the electrical quantity correlation coefficient matrix Corr between each equipment node, and extract the correlation coefficients that are higher than the threshold. ( Strongly correlated device pairs (default value 0.8, configurable) form a spatial correlation feature set. ; Specifically, the features from each level are concatenated according to their spatial adjacency relationships to obtain the spatial layer feature tensor. ∈ ^(T×S'×C×E), where S' is the number of spatial nodes after topological aggregation.

[0041] In C3: The first matrix data block is grouped according to physical attributes along the electrical quantity dimension, and the proportional relationship, phase difference and changes before and after the fault are analyzed within the group to form the electrical layer feature tensor; Specifically, electrical logic relationship mining is performed on the first matrix data block along the electrical quantity dimension C, and the electrical quantities are divided into voltage, current, power and derived quantity sets according to their physical attributes.

[0042] Specifically, the numerical ratios and phase differences between electrical quantities within each set are calculated, including: calculating the ratio of zero-sequence voltage U0 to zero-sequence current I0 to determine the nature of the ground fault; calculating the ratio of negative-sequence component to positive-sequence component to determine the fault asymmetry; and calculating the rate of change ΔP / of active power P and reactive power Q before and after the fault. With ΔQ / Extract power fluctuation characteristics; calculate Granger causality test values ​​between various electrical quantities to determine the strength of the drive-response relationship between electrical quantities; Specifically, the aforementioned electrical logic relationship features are integrated into an electrical layer feature tensor. ∈ ^(T×S×C'×E), where C' is the number of characteristic channels after electrical logic aggregation.

[0043] In C4: The protection action timing, equipment state transition path and setting margin are parsed along the state dimension of the first matrix data block to form a logic layer feature tensor; Specifically, along the state dimension K, the dependency rules for device state and protection logic are extracted from the first matrix data block. The time sequence of protection action sequences is parsed to extract the type of the first trip protection, whether the backup protection is activated, the timing and result of the reclosing action, and to construct the protection action logic chain A_seq={a_1,a_2,...,a_n}, where each action event a_i contains the action element name, action timestamp, and action setting value; the state transition path of each device before and after the fault is extracted, from "running" through "fault" to "tripping" and then to "pending recovery" complete state transition sequence, and the dwell time Δt_state of each state is calculated; the margin relationship between the protection setting value and the measured electrical quantity is analyzed to determine whether the sensitivity and selectivity of the protection action meet the requirements; Specifically, the results of the logic layer parsing are encoded into logic layer feature vectors. ∈ ^(T×S×C×K'), where K' is the feature dimension after the logical rule is encoded.

[0044] In C5: The time layer feature tensor, spatial layer feature tensor, electrical layer feature tensor, and logic layer feature tensor are cross-correlated and calculated along their respective dimensions to generate cross-layer coupling features; Specifically, the mutual information MI(F_T,F_S) between temporal and spatial features is calculated to extract strongly spatiotemporally coupled event-location correlation patterns; the time lag correlation coefficient between electrical and logical features is calculated to extract time delay correlation features between electrical quantity anomalies and protection actions. Calculate the joint distribution divergence of spatial layer features and electrical layer features within the time window before and after the fault, and extract the spatial electrical coupling strength index. .

[0045] Specifically, such as Figure 3 As shown, all cross-layer coupling features are aggregated into cross-layer coupling features. .

[0046] In C6: Using a preset set of fusion coefficients, the extracted feature tensors and cross-layer coupling features are weighted and fused to obtain a fused feature vector. The correlation between each feature dimension in the fused feature vector and the preset strong result label is calculated. Specifically, a weighted fusion is performed on the four independent features and cross-layer coupled features. The weight coefficient vector W = [ , , , , The initial value of ]^T is preset by the system (each weight is initially 0.2). During offline training using historical forced delivery cases, iterative optimization is performed based on the contribution of features to the prediction of forced delivery success rate. The fusion calculation formula is: in, This is a vector concatenation operation. Before concatenation, each feature vector is normalized to the same scale using the L2 norm.

[0047] In C7: retain feature dimensions with correlation higher than a preset threshold as effective features, perform principal component transformation on the retained effective feature set and select principal components with cumulative contribution rates that meet preset requirements as the final feature set, package and output the feature vectors of the standardized final feature set with the corresponding metadata information to generate the first target feature data; Specifically, the Pearson correlation coefficient r between each feature dimension and the label of the forced transmission result is calculated, and feature dimensions with |r|>0.3 are retained as effective features; principal component analysis is performed on the retained features, and the top m principal components with a cumulative variance contribution rate of 95% are selected as the final feature set; a standardization transformation Z=(x-μ) / σ is performed on each feature in the final feature set to make it follow a standard normal distribution; the standardized feature vector is packaged with the corresponding spatiotemporal labels (fault equipment ID, fault timestamp, fault type code) and metadata (voltage level, weather condition code, protection action sequence code) to output the first target feature data. .

[0048] It should be noted that step S300 overcomes the shortcomings of traditional single-dimensional analysis or simple summarization after independent processing of each source, which leads to insufficient feature expression capabilities. It enables a comprehensive characterization of the transient evolution process of the fault, the spatial propagation law, the electrical quantity coupling relationship and the protection action logic. The fused feature data not only achieves dimensionality reduction but also retains key decision information, significantly improving the representativeness and distinguishability of the features.

[0049] In this embodiment of the invention, step S400 converts the first target feature data into a tensor structure to be compressed and performs decomposition operations. Based on the operation results, a data decomposition strategy is selected. The tensor structure to be compressed is decomposed into multiple low-dimensional core factors according to the data decomposition strategy. Quantization and sparse coding operations are performed on the low-dimensional core factors to obtain the final optimized data packet. The detailed steps include the following sub-steps D1-D8: In D1: The normalized feature tensor in the first target feature data is converted into a standard tensor format and bound to the metadata to obtain the tensor structure to be compressed; Specifically, receiving the first target feature data This includes the standardized feature tensor. ∈ ^( × × ) and the corresponding metadata (faulty device ID, fault timestamp, fault type code, voltage level, weather condition code, protection action sequence code).

[0050] Specifically, the tensor converter will standardize the feature tensor. Unified conversion to standard tensor format X∈ ^(I1×I2×...× ), where N is the tensor order, I1 to These represent the magnitudes of each dimension. If the input feature is a third-order tensor (time × space × feature channels), then N = 3; if an environment dimension or state dimension is subsequently introduced, then N increases accordingly.

[0051] Specifically, during the tensor transformation, the metadata is associated and bound to the tensor in the form of attribute tensors to obtain the tensor structure to be compressed.

[0052] In D2: Analyze the dimensional distribution characteristics of the tensor structure to be compressed, and select one of the preset decomposition strategies as the data decomposition strategy for the current tensor to be compressed based on the combination of dimensional scale balance index and order interval. Specifically, the decomposition algorithm library includes three built-in algorithms: CP decomposition, Tucker decomposition, and Tensor-Train decomposition. The system automatically analyzes the characteristics of the current input tensor 𝒳, calculates the following indicators, and then selects the optimal decomposition strategy: Calculate the ratio of the maximum to the minimum value of each dimension of the tensor. = max(I1,I2,..., ) / min(I1,I2,..., ).like If the value is >10, it is determined that the differences in scale across dimensions are significant, and Tucker decomposition should be preferred. Calculate the ratio of the tensor's mean multilinear rank to its maximum rank. If the ranks of each dimension are nearly balanced and the ratio of the tensor's mean multilinear rank to its maximum rank is >0.7, the tensor is considered to have a uniform information distribution across its dimensions, and CP decomposition is preferred. Calculate the tensor order N. If N≥4, it is determined to be a high-order tensor, and Tensor-Train decomposition is preferred.

[0053] If any of the above three criteria conflict, the decision will be made based on the principle of prioritizing computational complexity. The computational complexity of TensorTrain decomposition is O(N·I·R). 3 The time complexity of Tucker decomposition is O(∏). ·R + ∑ The complexity of CP decomposition is O(N·I·R), and after comprehensive comparison, the decomposition strategy with the lowest complexity is selected as the final solution.

[0054] In D3: Depending on the type of data decomposition strategy, single-parameter traversal optimization, multi-dimensional independent rank determination, or dimension-by-dimensional chain truncation are performed to determine the optimal rank parameter. Specifically, if the selected decomposition strategy is CP decomposition, a single rank R needs to be determined. The rank selector uses alternating least squares to fit the tensor to different candidate values ​​of rank R, with candidate values ​​R ranging from 1 to R_max. The relative fitting error RE(R) is calculated for each R value. Simultaneously, the tensor compression ratio CR(R) is calculated. The weighted composite score of the fitting error RE and the compression ratio CR is used as the optimization objective, and the rank R at which the weighted composite score reaches its maximum value is selected as the optimal rank.

[0055] Specifically, if the selected decomposition strategy is Tucker decomposition, then the independent ranks R1, R2, ..., R_N of each dimension need to be determined. The rank selector uses higher-order singular value decomposition to expand the tensor 𝒳 in each dimension, obtaining the expanded matrices X_(1), X_(2), ..., X_(N) for each dimension. Singular value decomposition is then performed on each expanded matrix to obtain a sequence of singular values. For each dimension n, the cumulative contribution rate of the first k singular values ​​is calculated, and the smallest k value that makes the cumulative contribution rate of the first k singular values ​​≥ 0.95 is selected as the optimal rank for that dimension. If there is a significant jump between adjacent singular values, the position before the jump is used as the cutoff point to determine the rank, prioritizing information fidelity.

[0056] Specifically, if the selected decomposition strategy is Tensor-Train decomposition, the rank of each tensor chain core needs to be determined. The rank selector uses a dimension-wise SVD decomposition and truncation method to determine each rank: tensor chain decomposition is performed sequentially from the 1st dimension to the Nth dimension. At each step, singular value decomposition is performed on the current second-order ladder matrix. The minimum number of singular values ​​that makes the cumulative singular value contribution rate reach 95% is taken as the current truncation rank. This process is continued dimension by dimension until the rank of all dimensions is determined.

[0057] In D4: Based on the data decomposition strategy and the optimal rank parameter, a decomposition operation is performed on the tensor structure to be compressed to obtain multiple low-dimensional core factors; Specifically, if CP decomposition is used, X is decomposed into the sum of R rank tensors. The decomposition results are stored as factor matrices and weight vectors.

[0058] Specifically, if Tucker decomposition is used, X is decomposed into the product of the core tensor G and the factor matrices of each dimension. The decomposition result is stored as the core tensor G and the factor matrices.

[0059] Specifically, if Tensor-Train decomposition is used, X is decomposed into a chain product of N third-order core tensors.

[0060] Furthermore, the decomposed and stored results are each core tensor. .

[0061] In D5: Numerical distribution statistics are performed on multiple low-dimensional core factors, and the number of quantization bits and interval division method are determined according to the precision sensitivity of each factor. Specifically, the distribution of the data to be quantized is statistically analyzed, and the quantization interval boundaries are determined based on the quantization ...

[0062] In D6: Based on the number of quantization bits, the continuous values ​​of each element in various low-dimensional core factors are mapped to discrete quantization intervals, and the index value of the quantization interval to which each element belongs replaces the original value, thus completing the conversion from floating-point values ​​to integer indices. In D7: Scan consecutive identical index values ​​in the quantized sequence and encode them as a pair of index value and repetition count; perform frequency statistics on the pairs and construct a variable-length encoding table; replace each pair with the corresponding variable-length codeword to generate the final encoded data stream. Specifically, sparse coding employs a hybrid coding scheme combining run-length encoding and Huffman coding. The specific execution method is as follows: the quantized factor matrix and core tensor are scanned row by row, and consecutively occurring identical quantization index values ​​are encoded into (index value, number of consecutive occurrences) tuples; the tuple sequence is further processed using Huffman coding, constructing a Huffman tree based on the frequency of each tuple, assigning shorter codewords to high-frequency tuples and longer codewords to low-frequency tuples, thus achieving secondary compression.

[0063] In D8: According to the preset encapsulation format, the strategy identifier, dimension information, quantization parameters and encoding table are written into the packet header, the encoded data stream is written into the packet body, and the final optimized data packet is packaged and output. The decomposition and encoding results are packaged and output according to a unified data format. The output data structure consists of two parts: a file header and a data body. The file header contains the following fields: decomposition strategy type identifier (CP / Tucker / Tensor Train correspond to identifier codes 0x01 / 0x02 / 0x03 respectively), size of each dimension of the original tensor, dimension information of each factor matrix or core tensor after decomposition, number of quantization bits B and quantization interval boundary table, Huffman coding table, timestamp and associated metadata fingerprint (generated by concatenating fault device ID, fault timestamp, and fault type code and then hashing it using MD5). The total length of the file header is fixed at 512 bytes. The data body contains all factor matrix data, core tensor data, and weight vector data after quantization and sparse encoding. They are arranged in the order of storage of factor matrix / core tensor, and each data block is preceded by a 4-byte length identifier for location during decoding.

[0064] Furthermore, this embodiment also includes approximate reconstruction and quality verification steps: When subsequent modules need to read compressed data, the system performs inverse operations to reconstruct it according to the output format. The file header is parsed to obtain the decomposition strategy and parameters for each dimension; Huffman decoding and run-length decoding are performed on the data volume to recover the quantization index values; then, the quantization index values ​​are restored to approximate floating-point values ​​through dequantization mapping; finally, based on the decomposition strategy identifier in the file header, each factor matrix and the core tensor are multiplied and summed according to the corresponding tensor decomposition inverse operations to reconstruct the approximate tensor X. The system automatically calculates the relative error between the reconstructed tensor and the original input tensor. If the relative error between the reconstructed tensor and the original input tensor is >5%, a reconstruction quality alarm is triggered, and the compressed data block is marked as "needs to be re-decomposed," and the rank selection strategy is readjusted; if the relative error between the reconstructed tensor and the original input tensor is ≤5%, the compression is considered valid, and the reconstructed tensor is output to subsequent modules for use.

[0065] It should be noted that step S400 above decomposes high-dimensional feature data into low-dimensional core factors through tensor decomposition and compresses and stores them with quantization and sparse coding. While maintaining the integrity of the main information components, it significantly reduces the amount of data and storage overhead, solving the problems of heavy data storage burden and low transmission efficiency of high-dimensional feature data. This makes the optimized data package easy to retrieve and call quickly and conducive to long-term accumulation and reuse, providing an efficient storage solution for real-time incremental updates and feature knowledge base construction.

[0066] In this embodiment of the invention, the method further includes a real-time incremental update and feature library construction step, which specifically includes the following sub-steps E1-E6: In E1: By subscribing to events and periodically polling, changes to each data source are monitored, the time interval, device set, and data source type of the changed data are parsed, and a change area descriptor is generated. Specifically, the incremental listener registers as a subscriber to data change events for each data source interface (SCADA system, fault recording device, meteorological monitoring system, and equipment online monitoring system). When any data source generates a new data record, the data source interface proactively pushes a change notification to the incremental listener (the notification format includes the data source identifier, change type (add / update / delete), change timestamp, time range of affected data, and device range). Upon receiving the notification, the incremental listener immediately triggers the incremental update process. Specifically, for traditional data source interfaces that do not support event pushing, the incremental listener uses a fixed period of T_poll (default) The polling query (within 30 seconds, configurable) retrieves new or changed records since the last polling timestamp. The polling query uses an incremental query condition (WHERE timestamp > last_poll_time) to retrieve only changed data, avoiding a full data fetch.

[0067] Specifically, the incremental listener parses the received change notifications or polling results, extracting the spatiotemporal coordinate range corresponding to the changed data, specifically including: time intervals [ , (Start and end times of the changed data coverage), device set (List of device nodes where data changes have occurred), Data source type identifier Package the above information into a change region descriptor ΔR={ , , , }, and pass it to the incremental update engine.

[0068] In E2: The change type is determined by comparing the time interval in the changed region descriptor with the time range covered by the current multidimensional matrix, and local feature recalculation is performed only on the affected matrix slices and adjacent windows to obtain the updated feature vector; Specifically, if the changed data only affects new time slices in the time dimension (i.e., the timestamp of the new data is later than the latest time covered by the current matrix), it is determined to be a "time appending" change, and only the data corresponding to the new time slice needs to be processed. If the timestamp of the changed data is within the time range covered by the current matrix (i.e., historical data has been corrected or supplemented), it is determined to be a "history correction" change, and the matrix slice position corresponding to the timestamp needs to be located, and the affected features need to be locally recalculated. If the changed data involves new device nodes or new data source types (i.e., ... The device is not in the existing device dimension set, or If the data source dimension is not in the existing data source dimension set, it is determined to be a "structure extension" change, and the dimensions of the multidimensional matrix need to be extended before processing can be performed.

[0069] Specifically, for "time-appending" changes: the newly added time slice data is directly appended to the corresponding device nodes and electrical quantity coordinates of the multidimensional matrix M; a feature extraction process is performed on the newly added matrix slices (only performed on the newly added time slices, the features of the original time slices directly reuse historical results) to generate new feature vectors. ; Specifically, for "historical correction" changes: locate the affected historical time slice in matrix M, erase the original values ​​on the slice and fill in the corrected values; perform the feature extraction process only on the time slice and its adjacent time windows (one window length before and after to ensure the continuity of the sliding window boundary features) to generate updated feature vectors. ;Will The feature is compared with the original feature of the corresponding timestamp in the historical feature database, and only the differences are updated.

[0070] Specifically, for "structural expansion" changes: insert new coordinate axes into the corresponding dimensions of matrix M (device dimension S or data source dimension K) and expand the matrix size; fill the values ​​at the newly added coordinate positions according to the rules; perform feature extraction only on the newly expanded dimension slices and their associated intersections, and retain and reuse all existing features.

[0071] In E3: The updated feature vector is compared with the historical features in the feature library to obtain the maximum similarity value; Specifically, the feature vector f to be determined is compared with all historical feature vectors in the feature library. Calculate the cosine similarity sim(f, respectively) )=(f· ) / (||f||·|| ||), take the maximum similarity value. = (sim(f, )).

[0072] In E4: When the maximum similarity value is lower than the preset similarity threshold, the current updated feature vector is stored as a new feature and the statistical information is initialized; when the maximum similarity value is not lower than the preset similarity threshold, the current updated feature vector is marked as a non-new feature of the same type as the matched historical features. Specifically, in this embodiment, the preset similarity threshold is set to 0.85, which is configurable.

[0073] Specifically, when a feature to be updated is determined to be similar to a historical feature (the historical feature corresponding to the highest similarity), a weighted average method is used to update the corresponding record in the feature library. This design ensures that as the frequency of the same feature increases, the disturbance of new data to existing features gradually decreases, maintaining the stability of the feature library. Simultaneously, the statistical field of this feature, namely its frequency of occurrence, is updated. Add 1, update the most recent occurrence time to the current time, and append the current fault event ID to the related case list; Specifically, in response to the maximum similarity value being lower than the preset similarity threshold, the currently updated feature vector is stored as a new feature in the feature library, a unique feature ID is assigned to it, and the statistical fields are initialized (the frequency of occurrence is set to 1, the first occurrence time and the most recent occurrence time are both set to the current time, and the associated case list records the current fault event ID).

[0074] In E5: Incremental insertion of the index structure is performed on newly stored feature vectors to update the vector index. When the total number of features in the feature library increases by more than a preset ratio compared to the number during the last full index reconstruction, an asynchronous background task is triggered to perform a full index reconstruction on all feature vectors in the feature library. Specifically, the feature indexer constructs an efficient vector index for all feature vectors in the feature library, supporting fast similar feature retrieval in subsequent modules. The index uses the HNSW (Hierarchical Navigable Small World Graph) algorithm, configured as follows: number of layers L = log2(N) (N is the total number of features in the feature library, rounded up), maximum number of connections per layer 16, dynamic list size = 200, and cosine similarity as the distance metric. The index construction and maintenance method is as follows: Initial index construction: When deploying the system for the first time, the full structure of the HNSW index is constructed using a static method for all existing feature vectors in the feature library; Index update during incremental feature addition: When a new feature is added to the feature library, the index is updated by incremental insertion. That is, the new feature vector is directly inserted into the existing HNSW graph structure, and adjacency search and connection update are performed from the top to the bottom. Only the local graph structure is adjusted. The complexity of a single insertion is O(log2(N)), and there is no need to reconstruct the entire index. Periodic Index Reconstruction: When the total number of features in the feature library increases by more than a certain percentage (β) compared to the last full reconstruction (default β=50%, configurable), the system triggers an asynchronous background reconstruction task. This task utilizes system idle time to perform a full HNSW index reconstruction on all feature vectors, optimizing query performance. During the reconstruction, the existing index continues to provide query services, and a seamless switch to the new index occurs upon completion.

[0075] In E6: Periodically perform cluster analysis on all features in the database, extract cluster centers with sufficient samples as general patterns, and statistically analyze metadata to generate typical scenario labels; Specifically, cluster analysis is performed on the feature library periodically to extract general strong feature patterns. The specific implementation method is as follows: The DBSCAN density clustering algorithm is used to cluster all feature vectors in the feature library, with a neighborhood radius ε=0.3 and a minimum number of samples MinPts=5. For clusters with more than a threshold (default 10) in the clustering results, the cluster center vector is extracted as the "general feature pattern" of the cluster and stored in the pattern library. The frequency of metadata such as fault type, voltage level, and weather conditions corresponding to the features in each cluster is statistically analyzed to form typical scene labels, which are provided to the upper-level intelligent judgment system for rapid scene matching and decision reference.

[0076] Example 3: This example provides a forced delivery operation data optimization processing system, including: The data spatiotemporal alignment module is used to acquire multi-source heterogeneous data and align it to a unified spatiotemporal reference system to generate a first standardized data stream. The multidimensional matrix construction module is used to construct a multidimensional matrix framework model by filling the data values ​​in the first standardized data stream into the coordinate nodes corresponding to the multidimensional matrix framework model to form the first matrix data block. The hierarchical feature parsing module is used to perform hierarchical parsing operations on the first matrix data block, extract the time layer feature tensor, spatial layer feature tensor, electrical layer feature tensor and logic layer feature tensor, and perform fusion operation on the extracted feature tensors using a preset fusion coefficient set to generate the first target feature data; The compression coding module is used to convert the first target feature data into a tensor structure to be compressed and perform decomposition operations. Based on the operation results, a data decomposition strategy is selected. The tensor structure to be compressed is decomposed into multiple low-dimensional core factors according to the data decomposition strategy. Quantization and sparse coding operations are performed on the low-dimensional core factors to obtain the final optimized data packet.

[0077] It should be noted that the technical solution of the forced delivery operation data optimization processing system and the technical solution of the forced delivery operation data optimization processing method described above belong to the same concept. For details not described in detail in the technical solution of the forced delivery operation data optimization processing system in this embodiment, please refer to the description of the technical solution of the forced delivery operation data optimization processing method described above.

[0078] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0079] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through WiFi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a forced data delivery optimization processing method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0080] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method proposed in the above embodiments.

[0081] The storage medium proposed in this embodiment belongs to the same inventive concept as the method proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0082] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory, random access memory, flash memory, hard disk, or optical disk, and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute the method of the embodiments of the present invention.

[0083] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the present invention.

Claims

1. A method for optimizing data processing during forced data transmission operations, characterized in that, include: Acquire multi-source heterogeneous data and align it to a unified spatiotemporal reference frame to generate a first standardized data stream; Construct a multidimensional matrix framework model, and fill the coordinate nodes corresponding to the multidimensional matrix framework model with each data value in the first standardized data stream to form a first matrix data block; A hierarchical parsing operation is performed on the first matrix data block to extract time layer feature tensors, spatial layer feature tensors, electrical layer feature tensors and logic layer feature tensors. A fusion operation is performed on the extracted feature tensors using a preset set of fusion coefficients to generate the first target feature data. The first target feature data is converted into a tensor structure to be compressed and decomposition operation is performed. A data decomposition strategy is selected based on the operation result. The tensor structure to be compressed is decomposed into multiple low-dimensional core factors according to the data decomposition strategy. Quantization and sparse coding operations are performed on the low-dimensional core factors to obtain the final optimized data packet.

2. The data optimization processing method for forced delivery operation as described in claim 1, characterized in that, The generation of the first standardized data stream includes: Obtain the original timestamps corresponding to the multi-source heterogeneous data, compare the original timestamps with a preset time reference protocol, determine the time deviation value, and perform time axis translation and alignment on the data records according to the time deviation value; The device number contained in the multi-source heterogeneous data is extracted, the physical geographical location corresponding to the device number is determined according to the power grid topology location information, the multi-source heterogeneous data is mapped to the spatial node corresponding to the physical geographical location, and a first standardized data stream with a unified spatiotemporal label is generated.

3. The data optimization processing method for forced delivery operation as described in claim 2, characterized in that, The formation of the first matrix data block includes: Based on the scope of the forced delivery task and the type of access data source, the set of coordinate axes for each dimension of the multidimensional matrix framework model is dynamically determined; Traverse each data record in the first standardized data stream and fill each data value in the first standardized data stream into the coordinate node corresponding to the multidimensional matrix framework model; When multiple values ​​appear at the same coordinate node, select the primary value according to the data source priority and record the deviation; Perform a coordinate-by-coordinate scan on the filled multidimensional matrix. For missing coordinates, use interpolation, nearest neighbor estimation, or marking as unavailable according to the missing ratio. Perform a dimensionless transformation on the values ​​of each coordinate node after missing data processing, and output the first matrix data block.

4. The data optimization processing method for forced delivery operation as described in claim 3, characterized in that, The generation of the first target feature data includes: A sliding window is performed on the first matrix data block along the time axis to extract the numerical rate of change, spectral components and statistics within each time window, and the data are spliced ​​together to form a time layer feature tensor. The topological neighborhood of the first matrix data block is expanded layer by layer along the device dimension with the faulty device as the center, and the electrical quantity distribution and inter-layer attenuation at each level are calculated to form a spatial layer feature tensor. The first matrix data block is grouped according to physical attributes along the electrical quantity dimension, and the proportional relationship, phase difference and changes before and after the fault are analyzed within the group to form an electrical layer feature tensor; the first matrix data block is analyzed along the state dimension to parse the protection action sequence, equipment state transition path and setting margin to form a logic layer feature tensor. The time layer feature tensor, spatial layer feature tensor, electrical layer feature tensor, and logic layer feature tensor are cross-correlated and calculated along their respective dimensions to generate cross-layer coupling features; Using a preset set of fusion coefficients, the extracted feature tensor and the cross-layer coupling features are weighted and fused to obtain a fused feature vector. The correlation between each feature dimension in the fused feature vector and the preset strong result label is calculated. The feature dimensions with a correlation degree higher than a preset threshold are retained as valid features. Principal component transformation is performed on the retained valid feature set, and principal components with a cumulative contribution rate that meets the preset requirements are selected as the final feature set. The feature vectors of the standardized final feature set are packaged and output with the corresponding metadata information to generate the first target feature data.

5. The data optimization processing method for forced delivery operation as described in claim 4, characterized in that, The process involves converting the first target feature data into a tensor structure to be compressed and performing a decomposition operation. Based on the operation result, a data decomposition strategy is selected, and the tensor structure to be compressed is decomposed into multiple low-dimensional core factors according to the data decomposition strategy, including: The standardized feature tensors in the first target feature data are converted into standard tensor format and bound to metadata to obtain the tensor structure to be compressed. Analyze the dimensional distribution characteristics of the tensor structure to be compressed, and select one of the preset decomposition strategies as the data decomposition strategy for the current tensor to be compressed based on the combination of dimensional scale balance index and order interval. Depending on the type of data decomposition strategy, single-parameter traversal optimization, multi-dimensional independent rank determination, or dimension-by-dimensional chain truncation are performed respectively to determine the optimal rank parameter. Based on the data decomposition strategy and the optimal rank parameter, the tensor structure to be compressed is decomposed to obtain multiple low-dimensional core factors.

6. The data optimization processing method for forced delivery operation as described in claim 5, characterized in that, The step of performing quantization and sparse coding operations on the low-dimensional core factors to obtain the final optimized data packet includes: Numerical distribution statistics are performed on the multiple low-dimensional core factors respectively, and the number of quantization bits and interval division method are determined according to the precision sensitivity of each type of factor; Based on the number of quantization bits, the continuous values ​​of each element in various low-dimensional core factors are mapped to discrete quantization intervals, and the original values ​​are replaced by the index values ​​of the quantization intervals to which each element belongs, thus completing the conversion from floating-point values ​​to integer indices. The quantized sequence is scanned for consecutive identical index values ​​and encoded into a pair consisting of the index value and the number of repetitions. Frequency statistics are performed on the pairs and a variable-length encoding table is constructed. Each pair is replaced with the corresponding variable-length codeword to generate the final encoded data stream. According to the preset encapsulation format, the strategy identifier, dimension information, quantization parameters and encoding table are written into the packet header, the encoded data stream is written into the packet body, and the final optimized data packet is packaged and output.

7. The data optimization processing method for forced delivery operation as described in claim 6, characterized in that, The method also includes real-time incremental updates and feature library construction: By monitoring changes in each data source through event subscription and periodic polling, the time interval, device set, and data source type of the changed data are parsed to generate a change area descriptor. The change type is determined by comparing the time interval in the changed region descriptor with the time range covered by the current multidimensional matrix, and local feature recalculation is performed only on the affected matrix slices and adjacent windows to obtain the updated feature vector. The updated feature vector is compared with historical features in the feature library to obtain the maximum similarity value; In response to the maximum similarity value being lower than a preset similarity threshold, the currently updated feature vector is stored as a new feature and statistical information is initialized; in response to the maximum similarity value being not lower than the preset similarity threshold, the currently updated feature vector is marked as a non-new feature of the same type as the matched historical features. The newly stored feature vectors are incrementally inserted into the index structure to update the vector index. When the total number of features in the feature library increases by more than a preset ratio compared to the number during the last full index reconstruction, an asynchronous background task is triggered to perform a full index reconstruction on all feature vectors in the feature library. Regularly perform cluster analysis on all features in the database, extract cluster centers with sufficient samples as general patterns, and statistically analyze metadata to generate typical scenario labels.

8. A forced delivery operation data optimization processing system, employing the forced delivery operation data optimization processing method as described in any one of claims 1 to 7, characterized in that, include: The data spatiotemporal alignment module is used to acquire multi-source heterogeneous data and align it to a unified spatiotemporal reference system to generate a first standardized data stream. A multidimensional matrix construction module is used to construct a multidimensional matrix framework model by filling each data value in the first standardized data stream into the coordinate nodes corresponding to the multidimensional matrix framework model to form a first matrix data block. The hierarchical feature parsing module is used to perform hierarchical parsing operations on the first matrix data block, extract time layer feature tensors, spatial layer feature tensors, electrical layer feature tensors and logic layer feature tensors, and perform fusion operations on the extracted feature tensors using a preset fusion coefficient set to generate the first target feature data; The compression encoding module is used to convert the first target feature data into a tensor structure to be compressed and perform decomposition operations. Based on the operation results, a data decomposition strategy is selected, and the tensor structure to be compressed is decomposed into multiple low-dimensional core factors according to the data decomposition strategy. Quantization and sparse encoding operations are performed on the low-dimensional core factors to obtain the final optimized data packet.

9. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store computer-executable instructions, and when the processor executes the computer-executable instructions, it implements the steps of the forced data delivery optimization processing method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer-executable instructions stored thereon, characterized in that: When the computer-executable instructions are executed by the processor, they implement the steps of the forced data delivery optimization processing method according to any one of claims 1 to 7.

Images