Space-time data fusion-based power grid transient fault intelligent early warning method and system

Abstract

The application relates to the technical field of power grid fault diagnosis, and discloses a power grid transient fault intelligent early warning method and system based on space-time data fusion, which comprises the following steps: receiving a power grid monitoring node real-time data block and extracting a waveform fingerprint vector, constructing a credible space-time island through dynamic magnetic puzzle space adsorption alignment; constructing a local dynamic causal graph through traveling wave causal verification based on the credible space-time island, identifying a disturbance source and upgrading the disturbance source into a high-risk space-time island; extracting an island-level space-time feature vector, calculating an early warning deviation index to trigger a hierarchical asynchronous advanced early warning; when data asynchronous transmission disorder reaches, the application considers the early warning real-time performance and accuracy, and realizes robust asynchronous advanced early warning of the power grid transient fault.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid fault diagnosis, and more particularly to a power grid transient fault intelligent early warning method and system based on spatiotemporal data fusion. BACKGROUND

[0002] With the deepening of the construction of smart grid, real-time and accurate early warning of power grid transient faults has become a key technical link to ensure the safety of power grid. At present, power grid monitoring systems generally rely on PMU (Phasor Measurement Unit) devices and other devices to collect multi-source heterogeneous data to provide a data basis for fault warning. However, the non-synchronous nature of data transmission poses a serious challenge to fault warning.

[0003] A power grid transient voltage stability evaluation method based on spatiotemporal information synchronous learning is disclosed in Chinese Patent No. CN115935264B, which extracts the time sequence response trajectory of each monitoring node through time domain simulation, constructs a spatiotemporal adjacency matrix, and uses a graph convolutional neural network for synchronous learning. A power distribution network fault warning method and system based on transient waveform signals is disclosed in Chinese Patent Application No. CN116540015A, which collects transient waveform signal data, performs preprocessing and feature extraction, and establishes a fault diagnosis model to realize fault warning. The above-mentioned patents all use timestamp alignment to process multi-source data, relying on the synchronous arrival characteristics of data.

[0004] However, the existing technology has a fundamental defect in dealing with the non-synchronous transmission of power grid multi-source heterogeneous data. Although the system relies on PMU devices for synchronous sampling, the time of different substation data arriving at the cloud is inconsistent due to network congestion or device processing delay in actual transmission. Although the data has been stamped with GPS timestamps, in the streaming computing process, the model processes data according to "arrival time" rather than "event occurrence time". Specifically, when the upstream node's fault precursor data (such as voltage mutation and current distortion) arrives later than the downstream node's secondary wave data (such as fault propagation waveform) due to transmission delay, the model will mistakenly regard the downstream data as the fault source, resulting in a complete reversal of the cause-and-effect relationship. This cause-and-effect inversion not only makes the early warning system unable to accurately identify the fault propagation path, but also triggers false dispatch interventions. Forcing timestamp alignment requires waiting for the slowest data packet, resulting in a delay of hundreds of milliseconds in early warning, which cannot meet the real-time requirements of power grid transient faults. How to achieve accurate spatiotemporal alignment without waiting for the data to arrive in disorder has become a core technical bottleneck for intelligent early warning of power grid transient faults. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the application provides a power grid transient fault intelligent early warning method and system based on spatiotemporal data fusion, which innovatively uses "content alignment" instead of traditional "time alignment", constructs a credible spatiotemporal island through waveform fingerprint extraction and dynamic magnetic puzzle space adsorption alignment, and accurately identifies the disturbance source by combining with the traveling wave causality check, thereby solving the problems of spatiotemporal misplacement and causality inversion of non-synchronous transmission data. The application takes into account the real-time and accuracy of early warning without waiting for slow data, and provides robust and efficient technical support for power grid transient fault prevention and control.

[0006] In order to achieve the above-mentioned purpose, the application provides the following technical scheme:

[0007] The power grid transient fault intelligent early warning method based on spatiotemporal data fusion comprises the following steps:

[0008] Receiving real-time data blocks of each monitoring node of the power grid in real time, and extracting a waveform fingerprint vector; mapping the waveform fingerprint vector to a dynamic magnetic puzzle space for magnetic adsorption alignment, and constructing a credible spatiotemporal island;

[0009] Based on the credible spatiotemporal island, a local dynamic causal graph is constructed by using a traveling wave causality check method, potential disturbance sources are identified in the local dynamic causal graph, and the credible spatiotemporal island containing the potential disturbance sources is upgraded to a high-risk spatiotemporal island;

[0010] Extracting an island-level spatiotemporal feature vector of the high-risk spatiotemporal island, calculating an asynchronous early warning deviation index based on the island-level spatiotemporal feature vector, and triggering a hierarchical asynchronous early warning according to the asynchronous early warning deviation index.

[0011] The method for extracting the waveform fingerprint vector comprises the following steps:

[0012] Calculating a first-order difference sequence of the real-time data block, performing wavelet packet decomposition on the first-order difference sequence, calculating the energy proportion of each high-frequency subband obtained by the wavelet packet decomposition, screening a target high-frequency subband, combining the energy proportions of the target high-frequency subbands in a set order to form a waveform fingerprint vector representing the texture features of the real-time data block, and recording the relative time coordinates of the waveform fingerprint vector.

[0013] The method for performing magnetic adsorption alignment comprises the following steps:

[0014] Inputting the waveform fingerprint vector into a pre-trained Siamese network to map it into a high-dimensional embedding vector, and inserting the high-dimensional embedding vector into a dynamic magnetic puzzle space maintained in the cloud in real time, searching for k1 existing embedding vectors closest to the high-dimensional embedding vector in the dynamic magnetic puzzle space as the nearest neighbors of the high-dimensional embedding vector; the high-dimensional embedding vector inherits the relative time coordinates of the waveform fingerprint vector.

[0015] Topological check is performed on k1 nearest neighbors of the high-dimensional embedding vector of the current real-time data block, and if the topological check passes, magnetic adsorption alignment is performed.

[0016] The method for performing topological check comprises:

[0017] It is judged whether the monitoring nodes corresponding to the source of each nearest neighbor and the source of the current real-time data block belong to the same physical line or substation bus, and the number m of nearest neighbors satisfying the same physical line or substation bus is recorded, and if m is greater than the preset minimum number of topologically consistent neighbors m min , it is determined that the topological check passes.

[0018] The method for performing magnetic adsorption alignment comprises:

[0019] The relative time coordinates of the m nearest neighbors passing the topological check are extracted, the arithmetic mean value thereof is calculated as the average time coordinate, the relative time coordinate of the current real-time data block is aligned to the average time coordinate, the magnetic adsorption alignment is completed, and a spatiotemporal local is constructed.

[0020] The method for constructing a credible spatiotemporal island comprises:

[0021] The completeness score of the spatiotemporal local is calculated, and when the completeness score first exceeds the preset completeness threshold, the spatiotemporal local is marked as a credible spatiotemporal island.

[0022] The method for constructing a local dynamic causal graph comprises:

[0023] All real-time data blocks that have been magnetically adsorbed and aligned in the credible spatiotemporal island are taken as nodes, and the actual electrical connection between the nodes is taken as a candidate edge, to construct a local dynamic candidate graph;

[0024] In the local dynamic candidate graph, each candidate edge is subjected to traveling wave causal check using a waveform fingerprint vector, if the traveling wave causal check passes, the candidate edge is confirmed as an effective causal edge, otherwise the candidate edge is deleted, to obtain a local dynamic causal graph.

[0025] The method for traveling wave causal check comprises:

[0026] For each candidate edge, the time shift cosine similarity of the waveform fingerprint vectors of the nodes at both ends of the candidate edge is calculated, to obtain a plurality of similarity values, and if the maximum similarity value is greater than a preset similarity threshold, it is determined that the traveling wave causal check passes.

[0027] The method for identifying a potential disturbance source comprises:

[0028] The statistical causal in-degree and causal out-degree of each node in the local dynamic causal graph are counted, and waveform fingerprint abnormality degrees of each node are calculated; and potential disturbance sources in the local dynamic causal graph are identified according to the causal in-degree, the causal out-degree and the waveform fingerprint abnormality degrees of each node.

[0029] The method for identifying the potential disturbance sources in the local dynamic causal graph according to the causal in-degree, the causal out-degree and the waveform fingerprint abnormality degrees of each node comprises the following steps of:

[0030] If the waveform fingerprint abnormality degree of a node is greater than a preset abnormality degree threshold, and the causal in-degree is equal to 0, and the causal out-degree is greater than a preset minimum out-degree threshold, it is determined that the node is a potential disturbance source.

[0031] The method for calculating the asynchronous early warning deviation index comprises the following steps of:

[0032] The island-level spatiotemporal feature vector of the high-risk spatiotemporal island is input into an LSTM predictor, the internal evolution state of the high-risk spatiotemporal island is predicted, and a predicted feature vector is output;

[0033] A reference feature vector is acquired, the Euclidean distance between the predicted feature vector and the reference feature vector is calculated as an asynchronous early warning deviation index.

[0034] The power grid transient fault intelligent early warning system for spatiotemporal data fusion is used for realizing the power grid transient fault intelligent early warning method for spatiotemporal data fusion, and comprises the following modules:

[0035] The spatiotemporal island construction module is used for receiving real-time data blocks of each monitoring node of the power grid in real time, extracting waveform fingerprint vectors, mapping the waveform fingerprint vectors to a dynamic magnetic puzzle space for magnetic adsorption alignment, and constructing a credible spatiotemporal island.

[0036] The disturbance positioning module is used for constructing a local dynamic causal graph based on the credible spatiotemporal island by using a traveling wave causal verification method, identifying potential disturbance sources in the local dynamic causal graph, and upgrading the credible spatiotemporal island containing the potential disturbance sources to a high-risk spatiotemporal island.

[0037] The early warning module is used for extracting an island-level spatiotemporal feature vector of the high-risk spatiotemporal island, calculating an asynchronous early warning deviation index based on the island-level spatiotemporal feature vector, and triggering a hierarchical asynchronous early warning according to the asynchronous early warning deviation index.

[0038] Compared with the prior art, the power grid transient fault intelligent early warning system for spatiotemporal data fusion has the following beneficial effects:

[0039] The application constructs a 'content alignment' mechanism without GPS timestamp by taking the texture characteristics of power grid data as the alignment basis, effectively solving the problems of space-time dislocation and cause inversion caused by non-synchronous transmission of multi-source heterogeneous data. The mechanism realizes the synchronization of the relative time coordinates of data blocks through the magnetic adsorption alignment of waveform fingerprint vectors in the dynamic magnetic puzzle space, avoiding the loss of real-time caused by waiting for slow data packets; at the same time, by constructing a trusted space-time island and using a traveling wave cause and effect checking method, the disturbance source is accurately identified and the disturbance propagation characteristics are quantified, so that the system can still accurately determine the fault propagation path and source even in the case of data out-of-order arrival. The scheme not only guarantees the real-time of the early warning, but also ensures the accuracy of the early warning, so that the power grid fault prevention can be started at an early stage of disturbance evolution, leaving sufficient time for dispatch intervention, and significantly improving the safety and reliability of power grid operation. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, a brief introduction will be given to the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0041] Figure 1 The principle flow chart of the power grid transient fault intelligent early warning method of space-time data fusion provided by the embodiment of the present application;

[0042] Figure 2 The principle flow chart of constructing a trusted space-time island provided by the embodiment of the present application;

[0043] Figure 3 The principle flow chart of the island-level space-time feature vector extraction method provided by the embodiment of the present application;

[0044] Figure 4 The flow chart of the hierarchical asynchronous early warning method provided by the embodiment of the present application;

[0045] Figure 5 The functional module diagram of the power grid transient fault intelligent early warning system of space-time data fusion provided by the embodiment of the present application. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0047] Embodiment 1

[0048] Please refer to Figure 1 The embodiment provides a power grid transient fault intelligent early warning method based on spatio-temporal data fusion, and the method comprises the following steps:

[0049] In step S10, real-time data blocks of each monitoring node of the power grid are received, and a waveform fingerprint vector is extracted; the waveform fingerprint vector is mapped to a dynamic magnetic puzzle space for magnetic adsorption alignment, and a trusted spatio-temporal island is constructed.

[0050] In step S10, the inherent texture features of power grid data are used as a link to realize spatio-temporal alignment of non-synchronous transmission data, to construct a trusted data unit that does not depend on a GPS timestamp, to resolve the spatio-temporal dislocation problem caused by the out-of-order arrival of multi-source heterogeneous data from the source, and to provide basic data support for subsequent causal relationship analysis. This step breaks the inherent idea of traditional "timestamp alignment" and innovatively uses "content alignment" instead of "time alignment", which not only avoids the loss of real-time caused by waiting for slow data, but also ensures the spatio-temporal consistency of the data through topological association and geometric adsorption, and solves the core contradiction between real-time and accuracy in power grid transient fault early warning.

[0051] Further, referring to Figure 2 , step S10 comprises:

[0052] In step S11, a first-order difference sequence of the real-time data block is calculated, the first-order difference sequence is decomposed by a wavelet packet, the energy proportion of each high-frequency subband obtained by the wavelet packet decomposition is calculated, a target high-frequency subband is screened, the energy proportion of the target high-frequency subband is combined in a set order, a waveform fingerprint vector representing the texture features of the real-time data block is formed, and the relative time coordinates of the waveform fingerprint vector are recorded.

[0053] The real-time data block refers to a continuous sampling sequence collected by a synchronous sampling device within a fixed time window at each monitoring node of the power grid. Its essence is a set of values of a specific physical quantity (voltage or current) within a set period, which is used to capture the electrical operating state of the monitoring node during the corresponding period. The first-order difference sequence is obtained by calculating the difference between adjacent sampling points in the real-time data block. Its core function is to amplify the mutation characteristics in the data and suppress the interference of smooth components. Wavelet packet decomposition is a time-frequency analysis technique that can finely divide the full frequency band of a signal. Unlike traditional wavelet decomposition, which only subdivides the low frequency band, it can simultaneously decompose the high and low frequency bands, thereby accurately capturing the high frequency band concentrated in the transient fault traveling wave. The high frequency sub-band refers to the signal segment in the high frequency range after wavelet packet decomposition. Its energy ratio is the ratio of the energy of the sub-band to the total energy of all sub-bands after decomposition, which can represent the energy distribution characteristics of the signal in the corresponding high frequency band. The waveform fingerprint vector is a feature vector formed by the energy ratio of the high frequency sub-band, which can uniquely represent the high frequency texture characteristics of the real-time data block and serve as an identity identifier in subsequent spatial mapping and geometric adsorption.

[0054] In the specific implementation process, first, the real-time data blocks arriving in real time from each monitoring node of the power grid are received. The data block length is set to a fixed value N. The determination of N is based on the typical duration of the transient fault of the power grid, which needs to cover the key period from fault occurrence to propagation, ensuring that the complete transient waveform characteristics can be captured. For example, N can be set to 1024 sampling points, which is suitable for a transient process of 2ms under a sampling rate of 500kHz. Each real-time data block uniquely corresponds to a monitoring node, which can be a bus monitoring node of a substation, a line monitoring node, etc. The data block carries the identification information of the source monitoring node. For each received real-time data block, its first-order difference sequence is calculated. The calculation method is to subtract the value of the previous sampling point from the value of the current sampling point, forming a sequence with a length of N-1. When a transient fault occurs, the voltage or current signal will have a significant mutation. The first-order difference can amplify this mutation, making it easier to extract fault features, while the first-order difference of normal smooth signals tends to zero, achieving preliminary separation of fault features and background noise. Then, wavelet packet decomposition is performed on the first-order difference sequence. Before decomposition, a wavelet basis function needs to be selected. The selection is based on the ability to capture the mutation signal. For example, the db series wavelet basis can be selected because it can accurately depict the steep front of the transient traveling wave. The decomposition level is set to L decomp , L decompThe determination basis is the frequency resolution requirement, and it is necessary to ensure that the high-frequency sub-band after decomposition can cover the main frequency band of the transient fault traveling wave. After decomposition, the energy proportion of each high-frequency sub-band is calculated, and the calculation method is the energy of a single high-frequency sub-band divided by the total energy of all sub-bands, and the energy calculation adopts the coefficient modulus square sum after wavelet packet decomposition. Then, the target high-frequency sub-band is selected, and the selection standard is that the frequency band corresponding to the sub-band matches the concentrated frequency band of the transient fault traveling wave, so as to ensure that the energy proportion of the selected sub-band can effectively represent the texture feature related to the fault; the energy proportions of the target high-frequency sub-bands are combined in a set order to form a waveform fingerprint vector F. While extracting the waveform fingerprint vector, the system records the time stamp t arrival of the arrival of each real-time data block to the cloud, which is based on the cloud system clock and is used for subsequent construction of a relative time coordinate system. Taking the arrival time t0 of the first data block after the system starts as a reference, the relative time coordinate of the current data block is initialized as t relative =t arrival -t0, t relative is the initialized relative time coordinate, and the unit is millisecond. The relative time coordinate will be transmitted to the subsequent processing step together with the waveform fingerprint vector F.

[0055] For example, the db4 wavelet base in the db series is selected, which has orthogonal and compactly supported properties and can accurately depict the mutation details of the traveling wave and avoid information redundancy or loss in the signal decomposition process. The decomposition layer number L decomp is set to 4, and 4-layer decomposition can subdivide the full frequency band of the signal into 2 4 =16 sub-bands. The sub-band index is represented in the format of "(decomposition layer number, sub-band number)", that is, (4, 0), (4, 1), …, (4, 15), which breaks through the limitation of traditional wavelet decomposition that only subdivides the low-frequency band, and realizes synchronous fine division of high and low frequency bands. The frequency band attributes of the 16 sub-bands are matched and verified, for example, if the frequency band ranges of sub-bands (4, 3), (4, 5), (4, 7), and (4, 11) completely coincide with the 500 kHz-2 MHz frequency band in the transient fault traveling wave set, which meets the standard of "selecting high-frequency sub-bands related to the fault", then the waveform fingerprint vector F=[e 43 ,e 45 ,e 47 ,e 411 ] is constructed, wherein e 43 is the energy proportion of sub-band (4, 3), e 45 is the energy proportion of sub-band (4, 5), e 47 is the energy proportion of sub-band (4, 7), and e 411 is the energy proportion of sub-band (4, 11).

[0056] The first-order difference sequence is selected for subsequent processing, because when the traditional feature extraction is directly performed on the original data, the smooth component of the normal operation of the power grid will cover the fault mutation characteristics, and the first-order difference can effectively highlight the mutation information, improve the recognition degree of the fault characteristics, can realize the preliminary separation of the fault characteristics and the background noise, reduce the noise interference in the subsequent feature extraction process, and make the extracted texture features more accurately reflect the essential properties of the data. Instead of traditional wavelet decomposition, wavelet packet decomposition is used, because the transient fault traveling wave is concentrated in the high frequency band, the traditional wavelet decomposition is insufficient for subdivision of the high frequency band, and the energy distribution difference in the high frequency range cannot be accurately captured, while the full-band subdivision capability of the wavelet packet decomposition can realize the fine division of the high frequency band, accurately split the high frequency band signal, obtain the subdivided high frequency subband, and capture the subtle energy distribution difference of the transient fault traveling wave in the high frequency band, providing support for subsequent construction of unique waveform fingerprint vectors. In the prior art, data alignment depends on time stamp or single amplitude feature, the former needs to wait for slow data when the data arrives in disorder, resulting in loss of real-time performance, and the latter cannot accurately capture the subtle differences of transient faults, resulting in low matching accuracy of associated node data. This step constructs a waveform fingerprint vector by combining multi-dimensional high frequency features, so that the "texture" of the data block has uniqueness, effectively solving the problem of insufficient discrimination of single feature, enabling the data of topologically adjacent nodes to be accurately associated, providing core feature support for subsequent content alignment without time stamp, and ensuring that even if the data arrives in disorder, it can be associated and matched through feature similarity. Provides standardized and high-identification input features for subsequent embedding and mapping of S12. If this step is missing, the subsequent spatial mapping will not be able to accurately find the associated data block due to the lack of discrimination of the input features, resulting in a loss of basis for time and space alignment.

[0057] Step S12, inputting the waveform fingerprint vector into a pre-trained Siamese network to map it into a high-dimensional embedding vector, and inserting the high-dimensional embedding vector into a dynamic magnetic puzzle space maintained in the cloud in real time, searching for k1 existing embedding vectors closest to the high-dimensional embedding vector in the dynamic magnetic puzzle space as the nearest neighbors of the high-dimensional embedding vector; the high-dimensional embedding vector inherits the relative time coordinates of the waveform fingerprint vector;

[0058] The Siamese network is a deep learning network with a double-branch structure, which can learn the similarity mapping rule of input features through contrastive learning training, so that similar features are closer in the embedding space and dissimilar features are farther apart. The dynamic magnetic puzzle space is a high-dimensional embedding space maintained in the cloud, which has the ability to insert and delete embedding vectors in real time, and is used to store the embedding vectors of each real-time data block after mapping, simulating the spatial environment of puzzle assembly. The high-dimensional embedding vector is a high-dimensional feature vector obtained by mapping the waveform fingerprint vector through the Siamese network, denoted as E, which has the same dimension as the dynamic magnetic puzzle space and can preserve the core correlation features of the original waveform fingerprint vector. In the specific implementation process, the Siamese network is first pre-trained, and the training data is the real-time data block and its corresponding topological correlation label in the historical operation of the power grid. The topological correlation label is used to identify whether two data blocks belong to topologically adjacent nodes or belong to the same bus of a substation; the training target is to make the Euclidean distance between the high-dimensional embedding vectors of two data blocks with a topological correlation label of "yes" less than a set value, and the Euclidean distance between the embedding vectors of two data blocks with a label of "no" greater than a set value, so that the network learns the feature mapping rule of topologically correlated data blocks through this training process. The waveform fingerprint vector F generated in S11 is input into the pre-trained Siamese network, which calculates the output high-dimensional embedding vector E through forward propagation. This mapping process converts the low-dimensional waveform fingerprint vector into a feature vector in a high-dimensional space, preserving the similarity correlation between vectors. Then the high-dimensional embedding vector E is inserted into the dynamic magnetic puzzle space maintained in the cloud in real time, and the topological information of the real-time data block corresponding to the vector is recorded synchronously during the insertion process, including the physical line, substation bus, etc. At the same time, the relative time coordinate t relative is recorded as the time identifier of the high-dimensional embedding vector. Each embedding vector in the dynamic magnetic puzzle space is associated with three types of information: high-dimensional feature vector E, topological information, and relative time coordinate t relative , where the relative time coordinate represents the relative position of the data block on the time axis, providing a reference benchmark for the time dimension in subsequent magnetic adsorption alignment. In the dynamic magnetic puzzle space, the Euclidean distance between the high-dimensional embedding vector E and all existing embedding vectors in the space is calculated, and the first k1 embedding vectors are selected as the nearest neighbors of the current vector in order from small to large. The determination of k1 is based on the density of the power grid topology to ensure that there are enough samples for subsequent topology verification, for example, set to 6.

[0059] The Siamese network is used for embedding mapping because when directly performing neighborhood search on the waveform fingerprint vector, the low-dimensional space can not fully reflect the complex correlation between features, resulting in inaccurate similarity judgment. The Siamese network can mine deep correlations between features through deep learning, map them to a high-dimensional space that is more conducive to distinguishing similarities, strengthen the feature similarity of topologically correlated data blocks, amplify the feature difference of non-correlated data blocks, improve the accuracy of subsequent neighborhood search, and ensure that most of the found nearest neighbors are topologically correlated data blocks. The dynamic magnetic puzzle space is constructed and the embedding vector is inserted in real time because the power grid data is streamed and arrives in real time. A static space cannot meet the dynamic increase and decrease requirements of data. A dynamic space can accept embedding vectors of new data blocks in real time, while deleting invalid vectors that have not been attracted for a long time. It provides a storage and retrieval environment that adapts to streaming data, ensuring that neighborhood search can be based on all valid data blocks at the latest, improving the timeliness and comprehensiveness of the search results.

[0060] In step S13, the k1 nearest neighbors of the high-dimensional embedding vector of the current real-time data block are topologically checked, and if the topological check is passed, magnetic attraction alignment is performed.

[0061] It is judged whether the monitoring nodes of the sources of the real-time data blocks corresponding to each nearest neighbor and the monitoring node of the source of the current real-time data block belong to the same physical line or the same bus of a transformer substation, and the number m of nearest neighbors that meet the condition is recorded. If m is greater than the preset minimum number of topologically consistent neighbors m min , it is determined that the topological check is passed. The relative time coordinates of the m nearest neighbors that pass the topological check are extracted, and the arithmetic mean thereof is calculated as the average time coordinate. The relative time coordinate of the current real-time data block is forcibly aligned to the average time coordinate, the magnetic attraction alignment is completed, and the spatiotemporal local is constructed.

[0062] The topological check refers to a checking process of querying a power grid topology library to verify whether the monitoring nodes corresponding to two real-time data blocks belong to the same physical line or the same bus of a transformer substation, and is used to confirm the physical correlation between data blocks. The relative time coordinate refers to a relative time identifier of each data block in the spatiotemporal local set with reference to the arrival time of a reference data block, and does not need to rely on a GPS time stamp, and is only used to represent the time sequence of data blocks in a local range. The magnetic attraction alignment simulates the characteristics of mutual attraction of magnetic substances, and realizes time synchronization of data blocks. The spatiotemporal local refers to a local data set composed of multiple real-time data blocks that have been attracted and aligned and have a physical correlation, and is a basic unit for constructing a credible spatiotemporal island.

[0063] In the implementation process, first, the topology information of the real-time data blocks corresponding to the k1 nearest neighbors searched in S12 is acquired, which includes the physical line identifier and the substation bus identifier to which each data block belongs, and is stored in a power grid topology database. The power grid topology database contains the physical connection relationship of all monitoring nodes in the power grid, and provides a basis for topology verification. Topology verification is performed on each nearest neighbor: the power grid topology database is queried to determine whether the monitoring node corresponding to the neighbor and the monitoring node corresponding to the current real-time data block belong to the same physical line or the same substation bus, and the number m of nearest neighbors satisfying the same physical line or substation bus is recorded. The minimum number of topologically consistent neighbors m min is set min The determination basis is the proportional range of k1, which ensures that the selected neighbors have sufficient topological consistency and avoids alignment deviation caused by a small number of associated neighbors. The value needs to balance the topological consistency and the number of available neighbors, and is usually selected between one-third and one-half of k1. If m is greater than or equal to m min , it is determined that the current real-time data block belongs to the spatiotemporal local part of the m neighbors, the relative time coordinates of the m neighbors recorded in step S12 are extracted, the arithmetic mean of the relative time coordinates of the m neighbors is calculated, the relative time coordinates of the current real-time data block are forcibly set to the arithmetic mean, and the magnetic adsorption alignment is completed; the average value of the relative time coordinates is calculated by dividing the sum of the relative time coordinates of all effective neighbors by the number of effective neighbors. This method can keep the time coordinates of the current data block consistent with the time coordinates of the associated data block, and realize relative alignment without GPS dependence. If m is less than m min , that is, it does not pass the topology verification, it is determined that the current real-time data block has no sufficient topologically associated neighbors, and it is temporarily suspended at the edge of the dynamic magnetic jigsaw puzzle space and does not participate in the construction of the current spatiotemporal local part. After a new data block arrives and becomes an effective neighbor, the topology verification and adsorption alignment process of this step are re-executed.

[0064] The reason for performing topology verification is that the nearest neighbors filtered by feature similarity in step S12 may have the case of "feature similarity but no physical correlation", and if time alignment is directly based on these neighbors, it will lead to irrelevant data blocks in the spatiotemporal local area, destroying the spatiotemporal consistency. Topology verification can exclude such invalid neighbors, ensure that the neighbors aligned by adsorption have real physical correlation, and make the constructed spatiotemporal local area comply with the physical topology of the power grid, providing a reliable spatiotemporal basis for subsequent causal analysis. The reason for using the average relative time coordinate for alignment is that in the GPS-independent scenario, it is impossible to obtain absolute timestamps for alignment, while the relative time coordinates of the associated data blocks have inherent consistency, and their average value can represent the reference time of the spatiotemporal local area. This method realizes the time synchronization of data blocks, solves the problem of time misplacement caused by non-synchronous transmission, and makes the time coordinates of all data blocks in the spatiotemporal local area relatively consistent. Temporarily suspending data blocks that do not meet the conditions is to avoid mixing invalid data blocks into the spatiotemporal local area, leading to a decrease in local data quality, filtering invalid data that cannot be aligned temporarily, ensuring the purity of the spatiotemporal local area, and avoiding data pollution affecting subsequent completeness evaluation and credible island generation. In the prior art, if the GPS signal is lost or delayed when aligning data relying on GPS timestamps, alignment will fail, and waiting for all data block timestamps to align will lose real-time performance; without time alignment, direct analysis will lead to incorrect causal relationship judgment due to spatiotemporal misplacement. This step solves this contradiction by filtering valid neighbors through topology verification and aligning based on the average value of the relative time coordinates, realizing data time alignment without GPS dependence and waiting for slow data, balancing real-time performance and spatiotemporal consistency, and enabling data blocks to complete alignment quickly after arrival while ensuring that the alignment result complies with the physical topology.

[0065] In step S14, the completeness score of the spatiotemporal local area is calculated, and when the completeness score first exceeds the preset completeness threshold, the spatiotemporal local area is marked as a credible spatiotemporal island.

[0066] The completeness score is a proportional index for measuring the number of adsorbed data blocks in the spatiotemporal local area relative to the number of data blocks that the local area should theoretically have, denoted as C local , which can quantify the data completeness of the spatiotemporal local area. In the specific implementation process, as soon as a new real-time data block is successfully adsorbed to a certain spatiotemporal local area through step S13, the completeness score calculation of the spatiotemporal local area is immediately started. The calculation method of the completeness score is: the number of adsorbed and aligned real-time data blocks in the spatiotemporal local area divided by the number of theoretical real-time data blocks of the spatiotemporal local area. The number of theoretical real-time data blocks refers to the number of all monitoring nodes included in the physical line or bus of the substation corresponding to the spatiotemporal local area according to the power grid topology library, i.e., the total number of all monitoring nodes that should collect data in the region, and each monitoring node corresponds to a theoretical data block. The completeness threshold C threshThe determination method is: through a large number of simulation tests, the accuracy rate and early warning delay time of subsequent causal analysis under different data completeness are simulated, and the completeness value meeting the requirement of accuracy rate and having the minimum early warning delay is selected as C thresh , for example, 0.75; if the completeness threshold C thresh is too high, although the data integrity can be ensured, more data blocks need to be waited for, which leads to the decrease of real-time performance; if the completeness threshold C thresh is too low, although the trusted island can be quickly generated, the data may be incomplete, which affects the accuracy of subsequent analysis. When the completeness score C local of a certain spatiotemporal local part exceeds the completeness threshold C thresh for the first time, the spatiotemporal local part is marked as a trusted spatiotemporal island, which is used as an input object for subsequent local dynamic causal graph construction; if the completeness score C local does not exceed the completeness threshold C thresh , the spatiotemporal local part is kept and the completeness is evaluated again after new data blocks are adsorbed.

[0067] The data completeness of the spatiotemporal local part is quantified by the completeness score, because the spatiotemporal local part cannot contain enough data blocks only by adsorption alignment, if the subsequent causal analysis is based on the local part with insufficient data, the disturbance source may be misjudged or missed due to the lack of key node data, the completeness score can directly reflect the data completeness, quantify the completeness of the data set, and provide a data quality screening standard for subsequent analysis to avoid analysis based on incomplete data. The trusted spatiotemporal island is marked when the completeness score exceeds the completeness threshold for the first time, because exceeding the completeness threshold for the first time can not only ensure that the data quantity meets the analysis requirement, but also minimize the early warning delay, if higher completeness is waited for, the real-time performance will be lost, which cannot meet the requirement of advanced warning, so as to balance the data integrity and real-time performance, and ensure that the trusted island can support accurate causal analysis and meet the real-time requirement of the early warning system. In the prior art, when the causal analysis is based on incomplete data, the causal relationship is easy to be misjudged, and waiting for complete data will lose the real-time performance, which cannot realize advanced warning. This step solves the problem by completeness evaluation and threshold judgment, selects the data set with sufficient data quantity and real-time performance, provides a double-qualified data unit of “spatiotemporal consistency + data completeness” for subsequent causal graph construction, so that the causal analysis can ensure accuracy while considering real-time performance. If this step is missed, S20 will directly analyze the spatiotemporal local part without verified completeness, which may cause disturbance source locking error, and thus the whole early warning scheme is invalid.

[0068] Step S20: Based on the trusted spatiotemporal islands, a local dynamic causal graph is constructed using the traveling wave causal verification method. Potential disturbance sources are identified in the local dynamic causal graphs, and trusted spatiotemporal islands containing potential disturbance sources are upgraded to high-risk spatiotemporal islands.

[0069] Step S20 uses the reliable spatiotemporal island output from S10 as the core input. Based on the physical topology characteristics of the power grid and the traveling wave propagation law of waveform fingerprints, it constructs a local dynamic causal graph that accurately reflects the disturbance propagation relationship. This solves the problem of causal inversion caused by out-of-order data arrival, and simultaneously achieves accurate location of the disturbance source and determination of high-risk states, providing data support for the asynchronous early warning in S30. This step breaks through the limitations of traditional methods that rely solely on topological structure or a single timestamp to determine causal relationships. It innovatively combines topological correlation, traveling wave propagation characteristics, and waveform texture similarity, ensuring that the causal relationship judgment conforms to physical laws while avoiding interference from out-of-order data on causal logic. This resolves the key technical contradictions of "misjudgment of causal relationships" and "inaccurate location of disturbance sources" in power grid transient fault analysis.

[0070] Further, step S20 includes:

[0071] Step S21: Using all the real-time data blocks that have been magnetically adsorbed and aligned within the trusted spatiotemporal island as nodes, and the actual electrical connections between nodes as candidate edges, construct a local dynamic candidate graph.

[0072] Local dynamic candidate graph refers to the initial graph structure G constructed using aligned data blocks within trusted spatiotemporal islands as nodes and actual electrical connections of the power grid as candidate edges. local denoted as G local =(V local E candid ), where V local Let E be a set of nodes. candid This is a set of candidate edges, used to provide the basic graph structure for subsequent causal direction verification. A node refers to a single real-time data block within a trusted spatiotemporal island that has completed magnetic adsorption alignment. Each node uniquely corresponds to a monitoring node in the power grid, such as a substation busbar or line monitoring point. Its core attributes include the corresponding waveform fingerprint vector, relative time coordinates, and topology identification information. A candidate edge refers to a potential causal relationship edge established between two nodes based on the actual electrical connections existing in the power grid topology library, used to characterize the possible disturbance propagation paths between the two nodes. In the specific implementation, the newly emerging trusted spatiotemporal island output by S14 is first received, and all real-time data blocks within the trusted spatiotemporal island that have completed magnetic adsorption alignment are extracted. Each data block is then included as an independent node in the node set V. local Simultaneously, each node is associated with its corresponding power grid topology identifier, such as its line number and bus number, ensuring a one-to-one correspondence between nodes and power grid physical monitoring nodes. Then, the power grid topology database is queried for V.local For any two nodes i and j in the topology database, check if there is a direct electrical connection, including adjacent monitoring points on the same line, different nodes connected to the same busbar, etc. If the topology database records that the two nodes have an actual electrical connection, then establish an undirected candidate edge between nodes i and j and add it to the candidate edge set E. candid If no actual electrical connection exists, no candidate edge is established. Through the above process, the local dynamic candidate graph G is completed. local =(V local E candid The construction of ), where the node set V local This ensures that all units participating in the causal analysis possess spatiotemporal consistency, and the candidate edge set E candid This ensures that all potential causal paths conform to the physical topology of the power grid.

[0073] Choosing aligned data blocks within a trusted spatiotemporal island as nodes is crucial because original data blocks may exhibit spatiotemporal misalignment. Directly using these blocks as nodes would render the basic unit for causal relationship analysis unreliable. Data blocks within a trusted spatiotemporal island, however, are spatiotemporally aligned and consistent. This method provides a unified spatiotemporally basic unit for causal analysis, avoiding causal logic confusion caused by node spatiotemporal misalignment and ensuring the reliability of the premises for subsequent causal relationship judgments. Using actual electrical connections between nodes as candidate edges is essential because causal relationships involving disturbance propagation are impossible between nodes without physical connections. Constructing edges based on virtual connections introduces numerous invalid candidate edges, increasing the computational complexity of subsequent verification and reducing accuracy. This method filters out potential causal paths that conform to physical laws, reducing the number of invalid candidate edges and improving the efficiency and accuracy of subsequent causal verification, ensuring that the candidate edge set only contains paths with physical propagation potential. In existing technologies, some causal graph construction methods directly construct edges based on data arrival order or virtual connections, resulting in candidate edges containing numerous invalid paths or distorted causal relationship analysis due to node spatiotemporal misalignment. This step solves the problem by selecting spatiotemporally consistent nodes and physically connected candidate edges, constructing an initial graph structure with clear physical meaning and reliable node quality. This provides a high-quality basic framework for subsequent causal direction verification and avoids interference from invalid paths and misaligned nodes on the analysis results.

[0074] Step S22: In the local dynamic candidate graph, perform a traveling wave causality check on each candidate edge using the waveform fingerprint vector. If the traveling wave causality check passes, the candidate edge is confirmed as a valid causal edge; otherwise, the candidate edge is deleted to obtain the local dynamic causal graph.

[0075] The method of the traveling wave causality check is: for each candidate edge, the time shift cosine similarity of the waveform fingerprint vectors of the nodes at both ends of the candidate edge is calculated to obtain a plurality of similarity values, and if the maximum similarity value is greater than a preset similarity threshold, it is determined that the traveling wave causality check is passed.

[0076] The traveling wave causality check refers to a check process of judging whether there is a real traveling wave propagation and a propagation direction by verifying the similarity of the waveform fingerprints of the nodes at both ends of the candidate edge after time shift. The core basis is that the waveform texture features remain consistent when the traveling wave propagates in the power grid, and there is a fixed delay in the propagation. The time shift cosine similarity refers to the cosine similarity calculated by adjusting the waveform fingerprint vector of one node by a set time shift amount and the waveform fingerprint vector of another node, which is used to quantify the matching degree of the two waveforms after time shift compensation, denoted as S i→j (δ), where δ is a positive time shift amount, indicating that the waveform of the previous node is delayed by δ sampling points, and δ is a negative time shift amount, indicating that the waveform of the previous node is advanced by δ sampling points.

[0077] In the specific implementation process, for each candidate edge (i→j) in the local dynamic candidate graph constructed in step S21, the traveling wave causality check process is started. First, the waveform fingerprint vectors F i and F j generated by nodes i and j are extracted to ensure that the texture features of the vectors are consistent with the high-frequency characteristics of the corresponding data blocks. The time shift range is set as δ∈[-δ max ,+δ max ], δ max is the maximum time shift range, and the determination of δ max is based on the ratio of the length of the longest line in the power grid to the traveling wave propagation speed. The calculation method is: first, obtain the lengths of all lines in the power grid, determine the length L max of the longest line, query the traveling wave propagation speed v in the line of this type, and the commonly used traveling wave propagation speed in the power grid is a fixed range value, which can be calibrated through historical test data. The propagation time t delay of the traveling wave on the longest line is calculated as L max / v, and then the propagation time is converted into the number of sampling points according to the sampling rate fs, i.e. δ max =t delay ×fs, to ensure that all possible traveling wave propagation delays are included in the time shift range. For each δ value in the time shift range, the waveform fingerprint vector F i of node i is time-shifted to obtain the time-shifted vector F i,shifted (δ): when δ is positive, the subsequent vector segment of F i starting from the δth sampling point is intercepted, and the length is consistent with F j ; when δ is negative, F i is time-shifted by |δ| sampling points forward, i.e. F iFrom the first sampling point to the (N)th sampling point sample A vector segment with +δ) sampling points, the length of which is F j Consistent, where N sample N represents the number of sampling points for the real-time data block; note that δ is negative. sample +δ is actually a subtraction operation; when δ=0, the original F is used directly. i Then calculate F. i,shifted (δ) and F j cosine similarity S i→j (δ), the calculation of cosine similarity follows a well-known formula. The core logic is to quantify the consistency of vector directions using the cosine of the angle between the vectors; the closer the value is to 1, the more similar the texture features of the two vectors. After iterating through all δ values, the maximum similarity Smax is recorded. i→j and its corresponding time shift δ opt .

[0078] Set a similarity threshold S thresh The method for determining this is as follows: collect effective traveling wave propagation data (data of node pairs where actual traveling wave propagation is known) and noise data (data of node pairs where only random noise exists and there is no traveling wave propagation) from the historical operation of the power grid. Calculate the maximum time-shifted cosine similarity of the two types of data respectively, and calculate the minimum similarity S of the effective traveling wave data. valid,min Maximum similarity S with noisy data noise,max The similarity threshold S thresh Set as between S valid,min and S noise,max The values ​​between these ranges ensure that the similarity of all valid traveling wave data is greater than the similarity threshold S. thresh The similarity of the noisy data is all less than the similarity threshold S. thresh For example, the similarity threshold S can be determined by statistical analysis of 1000 sets of historical data. thresh The specific numerical value. If the maximum similarity Smax of the current candidate edges... i→j >S thresh If the candidate edge is found to have a real physical traveling wave propagation, the propagation direction is i→j, and the propagation delay is δ, then it is determined that the candidate edge has a real physical traveling wave propagation, the propagation direction is i→j, and the propagation delay is δ. opt If Smax, retain the edge as a valid causal edge; i→j ≤S thresh If the similarity of the candidate edge is determined to be due to noise interference and no real traveling wave propagation, the edge will be deleted from the local dynamic candidate graph.

[0079] The time-shifted cosine similarity is used for verification because traveling waves have an inherent propagation delay in the power grid; directly calculating F would be insufficient. i With F jThe similarity of the real causal relationship will be low due to the delay, and the time shift processing can compensate for the propagation delay, so that the two waveform fingerprint vectors originating from the same traveling wave are accurately matched, the propagation delay of the traveling wave is compensated, the similarity recognition accuracy of the real causal relationship is improved, the missed judgment of the causal relationship due to the propagation delay is avoided, and the candidate edge of the effective traveling wave propagation is accurately retained. Set a reasonable δ max , so as to avoid missing the real propagation delay due to too small time shift range, or increasing the calculation complexity due to too large time shift range, control the calculation cost while ensuring the comprehensive verification, and ensure the real-time requirement. A similarity threshold S thresh is introduced for screening, because there is random noise in the power grid data, which may cause occasional high similarity of the node pair without traveling wave propagation, the threshold screening can exclude such invalid interference, distinguish the effective traveling wave and the noise interference, avoid the invalid edge being misjudged as the effective causal edge, and ensure the accuracy of the causal graph. In the prior art, only the topological structure is used to determine the causal direction, which cannot consider the traveling wave propagation delay, and is easy to cause the causal inversion due to the out-of-order data arrival, for example, the downstream node data arrives first, but the actual upstream node is the disturbance source. When the similarity is not considered in the waveform similarity, the similarity will be low due to the propagation delay, and the real causal relationship cannot be recognized. This step solves the problem by combining the time shift compensation and the similarity verification, accurately identifies the effective causal edge and the real propagation direction under the consideration of the propagation delay, completely solves the problem of the causal inversion caused by the out-of-order data arrival, makes the causal relationship judgment comply with the physical law of the traveling wave propagation, and ensures that the upstream disturbance source node is accurately identified as the starting point of the causal relationship. If this step is omitted, the causal graph will contain a large number of invalid edges without real propagation relationship, and the causal direction is chaotic, which causes the subsequent disturbance source locking to be completely invalid.

[0080] In step S23, the causal in-degree and the causal out-degree of each node in the local dynamic causal graph are counted, and the waveform fingerprint abnormality degree of each node is calculated; and the potential disturbance source head in the local dynamic causal graph is identified according to the causal in-degree, the causal out-degree and the waveform fingerprint abnormality degree of each node.

[0081] The identification method of the potential disturbance source head is as follows: if the waveform fingerprint abnormality degree of a node is greater than a preset abnormality degree threshold, the causal in-degree is equal to 0, and the causal out-degree is greater than a preset minimum out-degree threshold, the node is determined as the potential disturbance source head.

[0082] The local dynamic causal graph refers to a graph structure G cause composed of effective causal edges and original node sets after the traveling wave causal verification in S22, and is denoted as G cause =(V local ,E cause ), wherein E causeThis is the set of valid causal edges confirmed in step S22, used to visually represent the propagation paths and causal relationships of disturbances within the credible spatiotemporal island. Causal in-degree D in D refers to the number of valid causal edges pointing to a node in a local dynamic causal graph, reflecting the number of upstream disturbances received by that node. in =0 indicates that the node has no upstream disturbance input and is a potential source of disturbance. Causal out-degree D out D refers to the number of valid causal edges originating from a given node, reflecting the range of perturbations propagating downstream from that node. out A larger value indicates a wider propagation of the disturbance. Waveform fingerprint anomaly degree A k It refers to the degree of difference between the current waveform fingerprint vector of node k and the waveform fingerprint vector under normal operating conditions, and is used to quantify whether there is abnormal disturbance in the node.

[0083] In the specific implementation process, firstly, all valid causal edges confirmed in step S22 are collected to form a set of valid edges E. cause The local dynamic candidate graph G constructed by S21 local The candidate edge set E in candid Replace with E cause The local dynamic causal graph G is obtained. cause =(V local E cause Then, the causal in-degree D of each node is calculated. in and causal exit D out For each node k, traverse E cause Given all edges in D, if the endpoint of an edge is k, then D in (k) Increment the count by 1; if the starting point of the edge is k, then D out (k) Increment the count by 1. The statistical process ensures that each valid causal edge is counted only once, avoiding duplicate counting. D in (k) represents the causal in-degree of node k, D out (k) represents the causal out-degree of node k. Calculate the waveform fingerprint anomaly degree A for each node k. k Formula A k =||F k -F k,normal ||2, where,|| ||2 is the Euclidean distance, F k F is the waveform fingerprint vector of node k; k,normal The waveform fingerprint vector of node k under normal operating conditions is obtained as follows: M real-time data blocks from the historical normal operating period of node k are collected. M must ensure coverage of different load conditions to guarantee the representativeness of the normal state. For each normal data block, the waveform fingerprint vector is extracted using the same process as in step S11, resulting in F. k1 F k2 FkM F kM For the M-th waveform fingerprint vector of the k-th node under its historical normal operating conditions, calculate the arithmetic mean of these vectors to obtain F. k,normal , ensure F k,normal It can represent the typical texture features of node k during normal operation; || ||2 represents the Euclidean distance, used to quantify the distance between two vectors in a high-dimensional space. A larger distance indicates a more significant difference between the current vector and the normal vector. During normal operation, the waveform fingerprint vector of a node is stable and consistent with F. k,normal The Euclidean distance is small; when a disturbance occurs, the high-frequency texture characteristics of the waveform change abruptly, F k With F k,normal The Euclidean distance is increased, and abnormal states can be identified through distance quantization. A k With F k With F k,normal The difference increases monotonically as it increases, which is completely consistent with the waveform distortion law when the disturbance occurs, ensuring that the anomaly degree can accurately reflect the existence and intensity of the disturbance.

[0084] Set an anomaly threshold A thresh and minimum out-degree threshold D out,thresh A thresh The method for determining this is to statistically analyze the historical normal operation data of node k. k The value is used to obtain the normal anomaly distribution, and the upper limit of the confidence interval of this normal anomaly distribution is taken as A. thresh To ensure A operates normally k All are less than A thresh When the disturbance occurs, A k All are greater than A thresh ;D out,thresh The method for determining this is to statistically analyze historical isolated noise data, that is, random anomalies without actual disturbance propagation. out The maximum value plus the safety margin is taken as D. out,thresh Ensure the D of isolated noise out Less than D out,thresh The D of the real disturbance source out Greater than D out,thresh For example, D out,thresh =2. For each node k, the disturbance source determination must simultaneously satisfy three core conditions: First, A k >A thresh This indicates that the waveform characteristics of node k show significant anomalies, which may be the point where the disturbance occurs; secondly, D in (k)=0 indicates that there are no valid causal edges pointing to node k in the local dynamic causal graph. This node has no upstream disturbance source and is the starting point of the causal relationship; third, D out (k)≥D out,thresh, indicating that the disturbance of node k has propagated to at least D out,thresh nodes downstream, excluding the interference of isolated noise or end burr. If multiple nodes simultaneously satisfy the above three conditions, they are arranged in descending order of D out (k), and the node with the largest D out (k) is the primary disturbance source node, and the rest are secondary disturbance source nodes. The sorting logic is that the larger the D out (k), the wider the range of disturbance propagation, and the more significant the impact on the power grid. Prior locking of the primary source can concentrate resources for intervention. After identifying at least one potential disturbance source, the current credible spatiotemporal island is marked as a high-risk spatiotemporal island, and its primary disturbance source node information, including node identification, waveform fingerprint abnormality, and causal out-degree, is recorded as input objects for step S30 asynchronous early warning.

[0085] The abnormality is calculated using the Euclidean distance because the waveform fingerprint vector is a high-dimensional feature. The Euclidean distance can effectively quantify the difference between vectors in high-dimensional space. Compared with other distance measures, such as Manhattan distance, the Euclidean distance is more sensitive to subtle differences in features, can accurately capture waveform distortions in transient faults, accurately quantify the abnormality of waveform fingerprints, accurately distinguish between normal operation and disturbance state, and avoid false negatives or false positives. The three core conditions for determining the disturbance source are set because a single condition cannot ensure the accuracy of the determination (e.g., A k may be isolated noise, and D in =0 may be data missing). Multiple condition combinations can achieve threefold verification of "abnormality exists + no upstream source + downstream propagation", exclude invalid interference, accurately lock the real disturbance source, avoid misjudging noise or secondary nodes as disturbance sources, and ensure that intervention measures can be targeted at the core source. In the prior art, disturbance source positioning relies mainly on a single feature (such as amplitude mutation) or only on topological structure, which is easily disturbed by noise, leading to inaccurate positioning, and cannot distinguish between primary and secondary sources and evaluate the impact. This step solves this problem by combining multiple conditions, sorting primary and secondary sources, and quantifying the impact, accurately locking the real disturbance source, distinguishing between primary and secondary sources, and quantifying the impact, providing a clear target and priority for subsequent early warning and intervention, ensuring efficient and accurate intervention measures, and avoiding resource waste.

[0086] Step S30: Extract the island-level spatiotemporal feature vector of the high-risk spatiotemporal island, calculate the asynchronous early warning deviation index based on the island-level spatiotemporal feature vector, and trigger the hierarchical asynchronous early warning according to the asynchronous early warning deviation index.

[0087] Step S30 inputs the high-risk spatiotemporal island output by S20 as the core, realizes accurate asynchronous early warning in the case of incomplete network data through spatiotemporal feature fusion, local state prediction and multi-dimensional deviation discrimination, solves the problem of delay in traditional warning caused by the dependence of the traditional warning on complete network data, and provides targeted intervention basis for power grid dispatching through a hierarchical triggering mechanism to adapt to different disturbance hazards. This step breaks the traditional warning mode of "full data + single threshold", and innovatively combines local spatiotemporal features, evolution prediction and multi-dimensional risk indicators, ensuring the advancement of the warning and controlling the false alarm rate through joint discrimination, thus solving the core technical contradictions of "incompatibility between advancement and reliability" and "incompatibility between warning level and hazard degree" in power grid transient fault warning.

[0088] Further, step S30 includes:

[0089] Step S31 combines all real-time data blocks in the high-risk spatiotemporal island that have completed magnetic adsorption alignment into a spatiotemporal data matrix, and extracts an island-level spatiotemporal feature vector of the high-risk spatiotemporal island.

[0090] Spatiotemporal data matrix X island is a matrix formed by arranging all real-time data blocks in the high-risk spatiotemporal island that have completed magnetic adsorption alignment according to the set dimensions, which is used to integrate the spatiotemporal correlation information of all nodes in the island, and has dimensions of (number of nodes × data block length × feature dimension). It not only retains the time sequence features of a single data block, but also contains the spatial topological correlation between nodes. In the specific implementation process, first, the high-risk spatiotemporal island pushed by step S23 is received, and all real-time data blocks in the island that have completed magnetic adsorption alignment are extracted. The spatiotemporal data matrix X island is constructed in the order of "node dimension-time dimension-feature dimension". The node dimension corresponds to the number of all nodes in the island, and each node occupies a row group in the matrix; the time dimension corresponds to the number of sampling points of the real-time data block, and each sampling point is a coordinate on the time axis; and the feature dimension corresponds to the core features of the real-time data block, which are the four-dimensional waveform fingerprint vectors extracted by S11, ensuring that the matrix contains three types of information, i.e., spatial node distribution, time sequence change and feature attribute. For example, if the high-risk spatiotemporal island contains K nodes, each data block has a length of N sample , and the feature dimension is 4 (four-dimensional waveform fingerprint vector), then the dimension of X island is (K × N sample × 4). The element X island [k, t, f] in the matrix represents the value of the f-th feature of the k-th node at the t-th sampling point.

[0091] X island is input into the preset spatiotemporal feature fusion module, which is described in detail in Figure 3, the internal space-time feature fusion module realizes global feature condensation through three-level feature extraction and fusion: the first level is 1D-CNN (one-dimensional convolutional neural network) to extract time sequence features, the convolution kernel of 1D-CNN slides along the time dimension, and the local features of each node time sequence data are extracted to capture the mutation trend, periodic law and other time sequence information in the data block, and the determination basis of the convolution kernel size and number is the data block length and the time sequence feature complexity, which ensures that the typical time sequence mode of transient fault can be covered; the second level is GCN (graph convolutional neural network) to extract spatial features, GCN takes the physical connection relationship in the power grid topology library as the adjacency matrix, and takes the time sequence features of each node extracted by 1D-CNN as the node feature input, and aggregates the feature information of adjacent nodes through graph convolution operation to capture the spatial propagation correlation of the disturbance between nodes, and the adjacency matrix is directly called from the power grid topology library to ensure consistency with the actual physical connection; the third level is the transformer attention mechanism to realize the fusion of space-time features, the attention mechanism calculates the correlation weight of different node and time step features, and the time sequence features output by 1D-CNN and the spatial features output by GCN are weighted and fused, which automatically focuses on the nodes and time segments that play a key role in the evolution of the disturbance and suppresses the interference of secondary information. After the three-level processing is completed, the high-dimensional features are mapped to a fixed-dimensional low-dimensional dense vector through global average pooling, that is, the island-level space-time feature vector H island , and the determination basis of the vector dimension is the feature expression requirement and the input adaptability of the subsequent predictor, which needs to be verified through experiments to ensure that the core information can be completely retained and the calculation complexity can be controlled at this dimension.

[0092] The reason for using 1D-CNN to extract time sequence features is that traditional time sequence feature extraction methods (such as sliding window mean) cannot capture the local mutation characteristics of transient faults, while the local receptive field characteristics of 1D-CNN can accurately capture subtle mutations and trend changes in time sequence data, accurately extract time sequence evolution features of single node data blocks, and retain the propagation law of the disturbance in the time dimension, providing time sequence basis for subsequent state prediction. The reason for using GCN to extract spatial features is that traditional spatial feature extraction methods ignore the topological association between nodes and only process single node features in isolation, while GCN can aggregate adjacent node information using the adjacency matrix to reflect the spatial propagation correlation of the disturbance, integrate the spatial topological features between nodes, and retain the propagation path information of the disturbance in the spatial dimension, so that the feature vector can reflect the spatial correlation state of the island as a whole. The reason for introducing the transformer attention mechanism for fusion is that simply stacking time sequence features and spatial features cannot highlight key information, the attention mechanism can automatically assign weights, focus on core nodes and key time segments, strengthen key information and suppress redundant interference, so that the extracted H islandMore accurate characterization of the core evolution state of the island, and improve the accuracy of subsequent prediction. In the prior art, feature extraction is mostly single time series or single space extraction method, which cannot consider the spatio-temporal correlation, resulting in incomplete feature expression and affecting the subsequent prediction accuracy. This step solves this problem through a multi-mechanism fusion feature extraction method, integrates spatio-temporal correlation information, generates comprehensive and accurate global feature vectors, provides high-quality input for subsequent local state prediction, ensures that the prediction result can reflect the overall evolution trend of the island, and avoids prediction deviation caused by one-sided features. If this step is missing, step S32 will directly predict based on the original data or single features, resulting in low prediction efficiency and poor accuracy due to high data dimension and much redundant information, and thus unable to achieve effective early warning.

[0093] Step S32, input the island-level spatio-temporal feature vector of the high-risk spatio-temporal island into the LSTM predictor, predict the internal evolution state of the high-risk spatio-temporal island, and output a predicted feature vector;

[0094] The LSTM predictor is a time series prediction model based on long short-term memory network, which is good at capturing long-term dependencies of time series data and can extrapolate future states based on historical time series features. Its network structure includes input gate, forget gate and output gate, which can effectively avoid the gradient disappearance problem in time series prediction. In the specific implementation process, first, the LSTM predictor is pre-trained: the training data is the spatio-temporal data matrix corresponding to the high-risk spatio-temporal island in the historical operation of the power grid and the subsequent actual evolution state feature vector, and the training target is to minimize the Euclidean distance between the predicted feature vector and the actual feature vector; The determination of network hyperparameters (such as the number of hidden layers, the number of hidden units, and the learning rate) is based on the balance between prediction error and training efficiency, and the combination of hyperparameters that meet the requirements of prediction accuracy and fast training convergence speed is selected through multiple experiments. The island-level spatio-temporal feature vector H island The pre-trained LSTM predictor is input, and the predictor receives H island through the input gate, the forget gate filters irrelevant historical information, and the output gate extrapolates the future state based on the effective features, and finally outputs the predicted feature vector H pred after P pred sampling points. The determination method of P pred is: according to the typical evolution period of power grid transient fault and the response time requirement of dispatching intervention, combined with the sampling rate to calculate the required early warning time, and then convert the early warning time into the number of sampling points; if the early warning time requirement is T pred , the sampling rate is fs, then P pred =T predXfs, ensuring that the prediction duration can both reserve sufficient time for scheduling intervention and not cause accuracy to decrease due to too large prediction span. The LSTM predictor is selected for state extrapolation because traditional time series prediction models (such as ARIMA) cannot handle nonlinear and long-dependent transient evolution data, while the gating mechanism of LSTM can effectively capture the nonlinear laws and long-term dependence of disturbance evolution, accurately extrapolate the future evolution trend of the local state, provide a quantitative basis for early warning of the future state, and avoid the lag of early warning caused by the inability to predict the evolution trend. Limiting the prediction range to the interior of the high-risk spatio-temporal island is because the nodes outside the island have no direct correlation with the current disturbance, and including them in the prediction will introduce irrelevant variables, increase the computational complexity and reduce the prediction accuracy. The present method focuses on the core prediction object, simplifies the calculation model, improves the prediction efficiency and accuracy, and meets the time requirements of real-time warning. In the prior art, early warning relies on global prediction of full data, which not only has large computational complexity and poor real-time performance, but also is prone to prediction deviation due to irrelevant node interference, and cannot start when the global data is not complete. This step solves this problem by combining local state extrapolation with an LSTM predictor, achieving accurate future state prediction under the condition of complete local data and incomplete global data, ensuring the advancement of early warning, while controlling the computational complexity, so that early warning can start early in the evolution of the disturbance, leaving sufficient time for scheduling intervention. Without this step, S33 will have no prediction reference, and it will not be able to quantify the difference between the future state and the reference state, and thus cannot trigger effective early warning.

[0095] In step S33, a reference feature vector is obtained, and the Euclidean distance between the predicted feature vector and the reference feature vector is calculated as an asynchronous early warning deviation index.

[0096] The reference feature vector H refer is the future state feature vector obtained by linear extrapolation based on the current data of the high-risk spatio-temporal island, which is used as a reference for judging whether the predicted state deviates abnormally, and its dimension is the same as H island and H pred . In the specific implementation process, first, based on the spatio-temporal data matrix X island constructed in step S31, the recent evolution trend of the current high-risk spatio-temporal island is extracted: the feature data of the last continuous multiple sampling points of each node data block is selected, the change slope of each node feature is fitted by linear regression, and the time series evolution trend parameter (slope, intercept) of each node is obtained. Based on this trend parameter, the features of each node are linearly extrapolated to obtain the node feature data after P pred sampling points in the future, ensuring that the extrapolation process follows the current evolution law and does not introduce additional disturbance assumptions. The extrapolated feature data of all nodes is constructed into X islandthe same dimension sequence, combined into an extrapolated space-time data matrix, input into the space-time feature fusion module of step S31, and output to obtain the reference feature vector H refer , ensuring that the extraction logic of the reference feature and the prediction feature is consistent, and the deviation calculation is comparable.

[0097] Calculate the asynchronous early warning deviation index Δ island Use the Euclidean distance formula: Δ island =||H pred -H refer ||2, where H pred is the future prediction feature vector output by step S32, reflecting the future state containing potential malignant evolution trend; H refer is the reference feature vector obtained by linear extrapolation, reflecting the future state of natural continuation of the current evolution trend; and ||2 is the Euclidean distance, used to quantify the spatial distance of two high-dimensional vectors. If the disturbance evolves naturally according to the current trend (without malignant development), the difference between H pred and H refer is small, and the value of Δ island is small. If the disturbance evolves in a malignant direction, such as fault expansion and rapid spread, H pred will deviate significantly from the natural evolution trajectory, and the value of Δ island will be larger. Δ island monotonically increases with the difference between H pred and H refer , which is completely consistent with the law of malignant evolution of the disturbance, ensuring that the deviation index can accurately quantify the evolution risk.

[0098] The reason for using linear extrapolation to construct the reference feature vector is that linear extrapolation can preserve the existing evolution trend to the greatest extent, and is simple to calculate and has strong real-time performance, making it suitable as a reference benchmark for rapid warning. Compared with nonlinear extrapolation, linear extrapolation can better reflect the natural evolution state without additional disturbance intervention, and can truly reflect the degree of abnormal deviation of the predicted state, avoiding deviation misjudgment caused by distorted reference benchmarks. The reason for using Euclidean distance to calculate the deviation is that H pred and H referAll are high-dimensional feature vectors, and the Euclidean distance can comprehensively reflect the deviation of the overall evolution state, compared with the single-dimensional deviation, which can avoid misjudgment of the overall evolution trend due to single-dimensional deviation and ensure the reliability of the deviation index. In the prior art, the early warning lacks a clear reference benchmark, and only the difference between the current state and the normal state is used for early warning, which cannot distinguish between natural attenuation and malignant evolution of the disturbance, and is prone to false alarm or missed alarm. This step solves this problem by constructing a reference feature vector and calculating the deviation, quantifies the difference between the predicted state and the natural evolution state, accurately identifies the malignant evolution trend, provides an objective quantitative basis for early warning triggering, and avoids false alarms or missed alarms caused by subjective judgment.

[0099] Step S34, four-element joint discrimination is performed on the asynchronous early warning deviation index, the integrity score of the high-risk spatiotemporal island, the waveform fingerprint abnormality of the potential disturbance source, and the causal out-degree to trigger the hierarchical asynchronous early warning.

[0100] The four-element joint discrimination refers to comprehensive judgment based on the asynchronous early warning deviation index Δ island , the integrity score C local of the high-risk spatiotemporal island, the waveform fingerprint abnormality A k of the main disturbance source node, and the causal out-degree D out (k). Specifically, in the implementation process, first, collect Δ island , the integrity score C local of the high-risk spatiotemporal island output by step S14, A k and D out (k) output by step S23. Referring to Figure 4 , the early warning triggering judgment is performed according to the four-element joint discrimination logic: the first discrimination is the basic triggering condition. If Δ island > Δ thresh and C local > C alert , it indicates that the disturbance evolution trend deviates from the natural trajectory, and the current island data integrity meets the reliability requirement of early warning, and the early warning is triggered. If any condition is not met, the early warning is not triggered, and the island state evolution is continued to be monitored.

[0101] Δ thresh is the early warning deviation threshold, and the determination method is as follows: collect the “benign disturbance” data, i.e., the disturbance that does not cause fault expansion after evolution, and the “malignant disturbance” data, i.e., the disturbance that causes fault expansion after evolution, calculate the Δ island values of the two types of data respectively, and count the maximum deviation value Δ benign,max of the benign disturbance and the minimum deviation value Δ malign,min of the malignant disturbance., Δ thresh is set to a value between Δ benign,max and Δ malign,min , ensuring that Δ island of benign disturbances are all less than Δ thresh , Δ island of malignant disturbances are all greater than Δ thresh , and the specific value of Δ thresh needs to be calibrated in combination with the system noise level to avoid false deviation triggered false alarms caused by noise.C alert For early warning trigger completeness, the false alarm rate of early warning under different C local is simulated through experiments, and the C local with false alarm rate lower than the set threshold and minimum early warning delay is selected as C alert , ensuring that the data completeness when the early warning is triggered can support reliable judgment.

[0102] After the basic trigger condition is met, the following rules are followed for grading, as shown in Table 1.

[0103] Grade I early warning, i.e., the highest level of early warning: A k > A high , C local ≥ C high and D out (k) ≥ D out,high must be met; in particular, if D out (k) ≥ D out,extreme , even if other conditions are not fully met, such as C local is in the interval [C alert , C high ) or A k is slightly lower than the high abnormality threshold A high but still higher than the abnormality threshold A thresh , it is also forced to upgrade to Grade I early warning, because the extreme propagation range indicates that the disturbance has shown a malignant diffusion trend and needs to be immediately intervened. This level corresponds to the scene of high abnormal intensity of the disturbance, high data reliability, wide propagation range or extreme propagation range, and the response measure is immediate network-wide broadcast + emergency intervention to ensure rapid containment of the disturbance diffusion.

[0104] Grade II early warning, i.e., the next highest level of early warning: A k > A high , C local ≥ C high but D out (k) ∈ [D out,thresh , D out,high ). This level corresponds to the scene of high abnormal intensity of the disturbance, high data reliability but limited propagation range, and the response measure is to push to the dispatcher + prepare for intervention, keeping close monitoring of the evolution of the disturbance.

[0105] Level III alert, also known as a concern alert: All scenarios that meet the basic triggering conditions but do not reach the Level I or Level II alert standards, such as D. out (k) is in [D out,thresh D out,high But A k Slightly lower, or C local Located in [C alert C high This level corresponds to scenarios where the disturbance poses a potential risk but the current harm is limited. The response measure is to issue only an in-station alarm and continue monitoring to avoid excessive consumption of scheduling resources.

[0106] Table 1. Tiered Early Warning Rules

[0107]

[0108] The methods for setting various thresholds are as follows: C high To determine the high confidence completeness threshold, based on historical data statistics, the smallest C corresponding to a causal analysis accuracy of 95% (95% confidence interval) was selected. local As C high This ensures that the indicator data at this level of completeness is highly reliable. high To establish a high anomaly threshold, statistically analyze historical malicious perturbations A. k The distribution is chosen, and the median of this distribution is selected as A. high Ensure A k Perturbations exceeding this value exhibit significant anomalous strength. D out,high To determine the high propagation threshold, and considering the power grid topology, the minimum D value is selected that ensures the disturbance's impact covers critical load areas. out As D out,high This ensures that disturbances at this propagation level pose a threat to the core area of ​​the power grid. out,extreme As the threshold for the maximum propagation degree, the D values ​​of historical extreme failures are statistically analyzed. out The distribution is chosen, and its lower bound is selected as D. out,extreme This indicates that the disturbance at this level of propagation has shown a malignant spread trend and requires urgent intervention.

[0109] The quaternary joint discrimination method is used instead of a single indicator because a single indicator cannot comprehensively assess the severity of the disturbance, such as Δ... island It's most likely caused by noise, only A kThe anomaly is likely an isolated one. Combining multiple indicators enables quadruple verification of "evolutionary risk + data reliability + anomaly intensity + propagation range," comprehensively screening for real high-risk scenarios, significantly reducing false alarm and missed alarm rates, and ensuring the accuracy of early warnings. A tiered early warning mechanism is established because disturbances of different severity require different response resources. A single warning level would lead to resource waste or insufficient intervention. This method matches response measures according to severity, improving intervention efficiency and focusing core resources on high-severity disturbances while avoiding excessive resource consumption by low-severity disturbances. A special escalation clause is included because disturbances with extreme propagation ranges can occur even if data integrity does not reach level C. high Its probability of malignant evolution is extremely high, and delayed intervention can lead to serious consequences. Prioritizing the handling of extreme risks can avoid missing major faults due to mechanically following integrity thresholds. In existing technologies, early warning systems mostly use single threshold triggers and lack a hierarchical mechanism. This not only results in a high false alarm rate but also fails to provide clear intervention priorities for dispatching, leading to poor targeting of intervention measures. This step solves this problem by combining quaternary joint discrimination with hierarchical triggering. It accurately identifies high-risk scenarios and matches them with corresponding intervention levels, ensuring that early warnings are both accurate and instructive. This enables dispatchers to quickly formulate intervention strategies based on the warning level, improving the efficiency of power grid fault prevention and control.

[0110] Example 2

[0111] This embodiment, based on Embodiment 1, provides an intelligent early warning system for power grid transient faults using spatiotemporal data fusion, such as... Figure 5 As shown, it includes:

[0112] Spatiotemporal island construction module: used to receive real-time data blocks arriving from each monitoring node of the power grid and extract waveform fingerprint vectors; map the waveform fingerprint vectors to the dynamic magnetic mosaic space for magnetic adsorption alignment to construct a reliable spatiotemporal island;

[0113] Disturbance localization module: Based on trusted spatiotemporal islands, a local dynamic causal graph is constructed using the traveling wave causal verification method. Potential disturbance sources are identified in the local dynamic causal graph, and trusted spatiotemporal islands containing potential disturbance sources are upgraded to high-risk spatiotemporal islands.

[0114] Advanced early warning module: used to extract island-level spatiotemporal feature vectors of high-risk spatiotemporal islands, calculate asynchronous advanced early warning deviation index based on island-level spatiotemporal feature vectors, and trigger graded asynchronous advanced early warning based on asynchronous advanced early warning deviation index.

[0115] Furthermore, in the spatiotemporal island construction module, the method for mapping the waveform fingerprint vector to the dynamic magnetic mosaic space for magnetic adsorption alignment includes:

[0116] inputting the waveform fingerprint vector into a pre-trained Siamese network to map the waveform fingerprint vector into a high-dimensional embedding vector, and inserting the high-dimensional embedding vector into a dynamic magnetic puzzle space maintained in the cloud in real time, searching for k1 existing embedding vectors closest to the high-dimensional embedding vector in the dynamic magnetic puzzle space as the nearest neighbors of the high-dimensional embedding vector;

[0117] topology checking is passed, magnetic adsorption alignment is performed.

[0118] The method of topology checking is to determine whether the monitoring nodes from which the real-time data blocks corresponding to each nearest neighbor are sourced and the monitoring node from which the current real-time data block is sourced belong to the same physical line or substation bus, and record the number m of nearest neighbors that meet the condition, if m is greater than the preset minimum number of topologically consistent neighbors m min , it is determined that the topology checking is passed.

[0119] Further, in the disturbance positioning module, the method of constructing a local dynamic causal graph comprises:

[0120] all real-time data blocks that have been magnetically adsorbed and aligned in the trusted space-time island are taken as nodes, and actual electrical connections between the nodes are taken as candidate edges, to construct a local dynamic candidate graph;

[0121] In the local dynamic candidate graph, a traveling wave causal check is performed on each candidate edge using a waveform fingerprint vector, if the traveling wave causal check is passed, the candidate edge is confirmed as an effective causal edge, otherwise the candidate edge is deleted, to obtain a local dynamic causal graph.

[0122] The method and system of the present application can be implemented in many ways. For example, the method and system of the present application can be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is only for illustration, and the steps of the method of the present application are not limited to the above specifically described order, unless otherwise specifically stated.

[0123] In addition, parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of corresponding technical solutions in the prior art are not described in detail to avoid excessive repetition.

[0124] The specific embodiments described above further illustrate the objects, technical solutions, and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A power grid transient fault intelligent early warning method of spatiotemporal data fusion, characterized in that, The method comprises: Receiving real-time data blocks of each monitoring node of the power grid in real time, and extracting waveform fingerprint vectors; mapping the waveform fingerprint vectors to a dynamic magnetic puzzle space for magnetic adsorption alignment, and constructing a trusted space-time island; the method for extracting the waveform fingerprint vectors comprises: calculating a first-order difference sequence of the real-time data blocks, performing wavelet packet decomposition on the first-order difference sequence, calculating the energy proportion of each high-frequency subband obtained by the wavelet packet decomposition, screening target high-frequency subbands, combining the energy proportions of the target high-frequency subbands in a set order to form a waveform fingerprint vector representing the texture features of the real-time data blocks, and recording the relative time coordinates of the waveform fingerprint vector; The method for performing magnetic adsorption alignment comprises: inputting the waveform fingerprint vector into a pre-trained Siamese network to map it into a high-dimensional embedding vector, and inserting the high-dimensional embedding vector into a dynamic magnetic puzzle space maintained in the cloud in real time, searching for k1 existing embedding vectors closest to the high-dimensional embedding vector in the dynamic magnetic puzzle space as the nearest neighbors of the high-dimensional embedding vector; the high-dimensional embedding vector inherits the relative time coordinates of the waveform fingerprint vector; performing topological verification on the k1 nearest neighbors of the high-dimensional embedding vector of the current real-time data block, and if the topological verification is passed, performing magnetic adsorption alignment; Based on the trusted space-time island, a local dynamic causal graph is constructed by using a traveling wave causal verification method, potential disturbance sources are identified in the local dynamic causal graph, and the trusted space-time island containing the potential disturbance sources is upgraded to a high-risk space-time island; An island-level space-time feature vector of the high-risk space-time island is extracted, an asynchronous early warning deviation index is calculated based on the island-level space-time feature vector, and a hierarchical asynchronous early warning is triggered according to the asynchronous early warning deviation index.

2. The method for intelligent early warning of power grid transient fault according to spatio-temporal data fusion of claim 1, characterized in that, The method for performing topological verification comprises: determining whether the monitoring nodes corresponding to the real-time data block sources of each nearest neighbor and the monitoring node of the current real-time data block source belong to the same physical line or substation bus, and recording the number m of nearest neighbors satisfying the same physical line or substation bus, if m is greater than the preset minimum number of topologically consistent neighbors m min topology verification is passed.

3. The method of claim 2, wherein, The method for performing magnetic adsorption alignment comprises: Extracting the relative time coordinates of the m nearest neighbors that pass the topological verification, calculating the arithmetic mean value thereof as an average time coordinate, aligning the relative time coordinates of the current real-time data block to the average time coordinate, completing the magnetic adsorption alignment, and constructing a space-time local.

4. The method of claim 3, wherein, The method for constructing a trusted space-time island comprises: When the completeness score exceeds a preset completeness threshold for the first time, the space-time local is marked as a trusted space-time island.

5. The method for intelligent early warning of power grid transient fault according to spatio-temporal data fusion of claim 4, characterized in that, The method for constructing a local dynamic causal graph comprises: Taking all real-time data blocks that have been magnetically adsorbed and aligned in the trusted space-time island as nodes, and taking the actual electrical connection between the nodes as candidate edges, a local dynamic candidate graph is constructed; In the local dynamic candidate graph, the waveform fingerprint vector is used to perform traveling wave causal verification on each candidate edge, if the traveling wave causal verification is passed, the candidate edge is confirmed as an effective causal edge, otherwise the candidate edge is deleted, and a local dynamic causal graph is obtained.

6. The method of claim 5, wherein, The method for performing traveling wave causal verification comprises: For each candidate edge, the time shift cosine similarity of the waveform fingerprint vectors of the nodes at both ends of the candidate edge is calculated to obtain a plurality of similarity values, and if the maximum similarity value is greater than a preset similarity threshold, it is determined that the traveling wave causal verification is passed.

7. The method of claim 6, wherein, The method for identifying potential disturbance sources comprises: The statistical causality in-degree and causality out-degree of each node in the local dynamic causal graph are counted, and waveform fingerprint anomaly degrees of each node are calculated; potential disturbance sources in the local dynamic causal graph are identified according to the causality in-degree, causality out-degree and waveform fingerprint anomaly degree of each node.

8. The method of claim 7, wherein, The method for identifying the potential disturbance source in the local dynamic causal graph according to the causality in-degree, causality out-degree and waveform fingerprint anomaly degree of each node comprises: If the waveform fingerprint anomaly degree of a node is greater than a preset anomaly degree threshold, the causality in-degree is equal to 0, and the causality out-degree is greater than a preset minimum out-degree threshold, it is determined that the node is a potential disturbance source.

9. The method of claim 8, wherein, The method for calculating the asynchronous early warning deviation index comprises: The island-level spatiotemporal feature vector of the high-risk spatiotemporal island is input into an LSTM predictor, the internal evolution state of the high-risk spatiotemporal island is predicted, and a predicted feature vector is output; A reference feature vector is obtained, and the Euclidean distance between the predicted feature vector and the reference feature vector is calculated as the asynchronous early warning deviation index.

10. The intelligent early warning system for power grid transient fault based on spatiotemporal data fusion, for implementing the intelligent early warning method for power grid transient fault based on spatiotemporal data fusion according to any one of claims 1-9, characterized in that, The system comprises: A spatiotemporal island construction module is configured to receive real-time data blocks of each monitoring node of the power grid in real time, extract waveform fingerprint vectors, map the waveform fingerprint vectors to a dynamic magnetic puzzle space for magnetic adsorption alignment, and construct a trusted spatiotemporal island; A disturbance positioning module is configured to construct a local dynamic causal graph based on the trusted spatiotemporal island by using a traveling wave causality checking method, identify potential disturbance sources in the local dynamic causal graph, and upgrade the trusted spatiotemporal island containing the potential disturbance sources to a high-risk spatiotemporal island; An early warning module is configured to extract an island-level spatiotemporal feature vector of the high-risk spatiotemporal island, calculate an asynchronous early warning deviation index based on the island-level spatiotemporal feature vector, and trigger a hierarchical asynchronous early warning according to the asynchronous early warning deviation index.

Images