Space-time associated transformer area voltage risk event identification method, equipment and medium

Abstract

The invention discloses a space-time associated transformer area voltage risk event identification method and device and a medium, and the method comprises the steps: constructing a voltage risk space-time cube covering a future preset time period for each dynamic cluster, and enabling the space-time cube to form a three-dimensional data structure through a two-dimensional space and a one-dimensional time; the voltage distribution of each node in the cluster in a future time period is obtained through time sequence load flow calculation, a continuous risk scoring function is adopted to quantify a voltage value into a risk index, a visual risk space-time distribution picture is formed, a 3D convolutional neural network is adopted to carry out deep processing on a space-time cube, a space-time spot pattern of a low-voltage risk is extracted, and the risk of the low-voltage risk is obtained. According to the method, propagation and evolution characteristics of risks in a space-time dimension are effectively captured, a space-time pattern is segmented into independent risk events through multi-dimensional connected domain analysis, feature parameters such as space-time boundaries and risk strength are labeled for each event, data authenticity in a risk identification process is ensured, risk capture capability is improved, and risk identification efficiency is improved. And the overall risk identification efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of transformer area voltage pattern recognition, specifically to a method, device, and medium for identifying transformer area voltage risk events with spatiotemporal correlation. Background Technology

[0002] With the accelerated evolution of new power systems, distribution networks are undergoing deep penetration and high-frequency interaction between power sources, grids, loads, and energy storage. Traditional voltage quality management methods, centered on single-point, static, and low-dimensional ledgers, are no longer adequate to adapt to the strong uncertainty and rapid spatiotemporal evolution brought about by a high proportion of distributed energy resources. Against this backdrop, traditional distribution network operation and control models have revealed numerous limitations: the data processing and risk identification paradigm of existing spatiotemporally correlated transformer area voltage risk identification methods are essentially still linear pipelines of static slicing and threshold comparison. Their technical approach can be summarized as data compression: reducing the dimensionality of the 24-hour voltage curve to mean, range, and number of exceedances, losing a large amount of temporal morphological information; constructing a one-time adjacency matrix using electrical or geographical straight-line distances between transformer areas, creating a static graph, and then clustering by modularity to obtain a lifetime transformer area group; setting a fixed threshold for the compressed statistics, generating a single-point alarm once the threshold is exceeded, and then simply merging scattered alarms occurring within the same time slice, calling them risk events. This dimensionality reduction-piecing-threshold paradigm exposes four types of technical defects in new power systems:

[0003] The failure to fully consider the temporal characteristics of voltage patterns in distribution areas, the spatial correlation between distribution areas, and the dynamic response characteristics to external environmental factors makes it difficult for the cluster division results to truly reflect the actual operating characteristics of the distribution network, which can easily lead to high-risk distribution areas being misjudged as normal.

[0004] Static adjacency matrices cannot reflect real-time electrical couplings such as inverter reactive power sharing and energy storage cluster discharge, resulting in discrepancies between cluster boundaries and actual power flow.

[0005] Existing systems mainly rely on threshold-based passive alarm mechanisms, which only issue alarms after voltage exceedances occur, lacking the ability to accurately predict and visualize the spatiotemporal distribution characteristics of voltage risks in future periods.

[0006] The same voltage instability is cut into isolated alarms in adjacent transformer areas and continuous time periods, making it impossible to reconstruct the risk evolution trajectory and thus impossible to accurately locate and quantify the risk. Summary of the Invention

[0007] The technical problem this invention aims to solve is that existing risk identification methods suffer from feature distortion, spatial staticity, temporal lag, and event fragmentation. The goal is to provide a spatiotemporally correlated method, device, and medium for identifying voltage risk events in transformer substations. This involves constructing a voltage risk spatiotemporal cube covering a preset future time period for each dynamic cluster. This cube uses a two-dimensional space and a one-dimensional time to form a three-dimensional data structure. Voltage distribution of each node within the cluster in the future time period is obtained through time-series power flow calculation. A continuous risk scoring function quantifies voltage values ​​into risk indicators, forming a visualized spatiotemporal risk distribution map. A 3D convolutional neural network is used for deep processing of the spatiotemporal cube to extract spatiotemporal patches of low-voltage risk, effectively capturing the propagation and evolution characteristics of risk in the spatiotemporal dimension. Multidimensional connected component analysis segments the spatiotemporal patches into independent risk events, labeling each event with spatiotemporal boundaries and risk intensity, ensuring the data authenticity of the risk identification process. By capturing the spatiotemporal evolution of risk, the ability to capture and identify risk is improved, enhancing the overall efficiency and accuracy of risk identification.

[0008] This invention is achieved through the following technical solution:

[0009] The first aspect of this invention provides a method for identifying voltage risk events in transformer substations based on spatiotemporal correlation, comprising the following specific steps:

[0010] Obtain high-reliability transformer area voltage data, and construct an enhanced multidimensional feature vector for each transformer area based on the high-reliability transformer area voltage data, which integrates voltage characteristics, spatial correlation, and dynamic external correlation.

[0011] The optimal number of clusters for the enhanced multidimensional feature vector is determined based on a multi-objective optimization algorithm, resulting in a dynamic cluster that divides the transformer area into cohesive clusters at the feature, geographical, and electrical levels.

[0012] Construct a spatiotemporal cube for voltage risk covering a preset future time period for each dynamic cluster;

[0013] A 3D convolutional neural network is used to process the spatiotemporal cube to extract spatiotemporal patterns of low voltage risk.

[0014] The spatiotemporal pattern is segmented into independent risk events through multidimensional connected component analysis;

[0015] The list of risk events and their corresponding preset control strategies are automatically pushed to the scheduling and control system via the message bus, triggering the control equipment to take action.

[0016] Furthermore, based on high-reliability transformer area voltage data, an enhanced multidimensional feature vector is constructed for each transformer area, fusing voltage characteristics, spatial correlation, and dynamic external correlation. Specifically, this includes:

[0017] Based on high-reliability transformer area voltage data, extract the phase-time matrix for time synchronization;

[0018] Variational mode decomposition is performed on each phase voltage sequence to remove the highest frequency and DC intrinsic modes, resulting in a detrended and denoised voltage sequence.

[0019] Construct a sliding window and extract multidimensional morphological indicators based on the voltage sequence;

[0020] Construct a transformer area adjacency graph with transformer areas as nodes and co-feed lines or low-voltage connections as initial edges;

[0021] The node impedance matrix is ​​calculated based on the power distribution GIS and conductor parameters, and the edges are weighted according to the electrical distance to obtain the weighted adjacency matrix;

[0022] By using graph convolutional networks, the voltage morphology features of nodes themselves and their neighborhood features are aggregated to generate spatially enhanced features;

[0023] Simultaneously collect external variables such as meteorological data, load data, and event data to form an external time series vector;

[0024] The external temporal vector is encoded using a temporal coding network to obtain external dynamic features;

[0025] Using the spatial enhancement features as queries and the external dynamic features as key-value pairs, a multi-head cross-attention mechanism is employed to calculate and output fused interactive features.

[0026] Calculate the correlation coefficient between spatial augmentation features and external dynamic features, and adaptively adjust the fusion weights based on this coefficient;

[0027] The voltage morphology features, spatial enhancement features, external dynamic features, and fusion interaction features are concatenated to obtain an enhanced multidimensional feature vector.

[0028] Furthermore, the determination of the optimal number of clusters for the enhanced multidimensional feature vectors based on the multi-objective optimization algorithm specifically includes:

[0029] Based on one or more key features of the enhanced multidimensional feature vector, the first clustering algorithm is used to perform initial coarse clustering on all transformer areas, forming M coarse clusters;

[0030] Define an optimization problem that includes at least three objective functions, which are used to simultaneously evaluate the quality of a clustering scheme in three dimensions: feature cohesion, geographic proximity, and electrical coupling strength.

[0031] For each of the M coarse clusters, a multi-objective optimization algorithm is used to perform fine clustering within the coarse cluster, with the at least three objective functions as the optimization objectives, to obtain a dynamic cluster that divides the transformer area into cohesive clusters at the feature, geographical and electrical levels.

[0032] Furthermore, the at least three objective functions include:

[0033] The first objective function is configured to minimize the sum of intra-cluster distances based on the feature space of the transformer area in order to maximize feature cohesion.

[0034] The second objective function is configured to minimize the sum of intra-cluster distances based on the geospatial area of ​​the station, in order to maximize geographic proximity;

[0035] The third objective function is configured to maximize the intra-cluster electrical connection strength based on the grid topology, or minimize the sum of intra-cluster electrical distances calculated based on electrical parameters, in order to maximize the electrical coupling strength.

[0036] Furthermore, the construction of a voltage risk spatiotemporal cube covering a preset future time period for each dynamic cluster specifically includes:

[0037] Power grid modeling is performed on the target dynamic cluster, and time-series power flow calculation is performed to obtain the voltage time-series data of each node in the cluster within a preset future time period;

[0038] Based on the voltage time series data, a predefined risk function is constructed to quantify the voltage risk value of each node at each time section;

[0039] Based on the voltage risk value, a voltage risk spatiotemporal cube is constructed; wherein, the two dimensions of the spatiotemporal cube represent the spatial information of the cluster, and the third dimension represents the temporal information, and the value of the cube element corresponds to the voltage risk value of the spatial location at the corresponding time section.

[0040] Furthermore, the timing power flow calculation is performed, specifically including:

[0041] The preset future time period is discretized into multiple continuous time segments;

[0042] For each time segment, the predicted load value and new energy power generation value for that time segment are injected into the power grid model;

[0043] Perform power flow calculations to solve for and obtain the voltage amplitude of each node in the cluster at that time section.

[0044] Furthermore, the step of processing the spatiotemporal cube using a 3D convolutional neural network to extract spatiotemporal patterns of low voltage risk specifically includes:

[0045] The voltage risk spatiotemporal cube is input into a convolutional neural network model. The encoder path is downsampled through 3D convolutional layers and 3D pooling layers to extract multi-scale spatiotemporal features. The decoder path is upsampled through 3D upsampling layers and jump-connected with the feature maps of the corresponding scale in the encoder path to restore spatial details and output a spatiotemporal patch map with the same size as the input. The spatiotemporal patch map is used to represent the probability that each spatiotemporal point in the spatiotemporal cube belongs to the core risk patch map.

[0046] Furthermore, the spatiotemporal pattern is segmented into independent risk events through multidimensional connected component analysis;

[0047] A probability threshold is set, and voxels with probability values ​​greater than the threshold in the spatiotemporal patch diagram are set as first values, representing the prospect of risk events, while voxels with probability values ​​less than or equal to the threshold are set as second values, representing the background.

[0048] Generate a 3D risk mask based on binary data of risk event prospects and background;

[0049] The connectivity criterion is used to determine the connectivity between voxels in the three-dimensional risk mask, and to identify and mark the isolated connected regions. Each marked connected region is defined as an independent low-voltage risk event.

[0050] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a spatiotemporally correlated method for identifying transformer voltage risk events.

[0051] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a spatiotemporally correlated method for identifying transformer area voltage risk events.

[0052] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0053] By acquiring highly reliable transformer area voltage data and constructing an enhanced multidimensional feature vector for each transformer area, the deep integration of spatial and external features is achieved through voltage morphology feature extraction, spatial correlation feature construction, and external correlation feature extraction, thereby improving the comprehensiveness of risk event monitoring.

[0054] The optimal number of clusters is automatically determined by a multi-objective optimization algorithm. Taking into account multiple optimization objectives such as feature similarity, geographical proximity and electrical coupling strength, the transformer area is divided into highly cohesive dynamic clusters through a two-stage clustering method. This ensures that the cluster division results conform to the electrical characteristics and meet the spatial constraints of actual operation and maintenance management.

[0055] By constructing a voltage risk spatiotemporal cube covering a preset future time period for each dynamic cluster, this spatiotemporal cube forms a three-dimensional data structure with two-dimensional space and one-dimensional time. The voltage distribution of each node within the cluster in the future time period is obtained through time-series power flow calculation, and the voltage values ​​are quantified into risk indicators using a continuous risk scoring function, forming a visualized spatiotemporal risk distribution map. A 3D convolutional neural network is used for deep processing of the spatiotemporal cube to extract spatiotemporal patches of low voltage risk. Through the encoder-decoder network design, the propagation and evolution characteristics of risk in the spatiotemporal dimension can be effectively captured.

[0056] By segmenting the spatiotemporal patch into independent risk events through multidimensional connected component analysis, and labeling each event with characteristic parameters such as spatiotemporal boundaries and risk intensity, the system automatically pushes the list of risk events and their corresponding preset control strategies to the scheduling and control system via a message bus. This enables intelligent matching and rapid triggering of control strategies, ensuring that effective control measures are taken before or in the early stages of a risk event. Attached Figure Description

[0057] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0058] Figure 1 This is the risk identification process in an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0060] As one possible implementation method, such as Figure 1 As shown, this embodiment provides a method for identifying voltage risk events in transformer substations based on spatiotemporal correlation, specifically including the following steps:

[0061] S1. Obtain high-reliability transformer area voltage data, and construct an enhanced multi-dimensional feature vector for each transformer area based on the high-reliability transformer area voltage data, which integrates voltage characteristics, spatial correlation and dynamic external correlation.

[0062] S2. Based on the multi-objective optimization algorithm, determine the optimal number of clusters for the enhanced multidimensional feature vectors, and obtain a dynamic cluster that divides the transformer area into features, geography and electrical levels.

[0063] S3. Construct a spatiotemporal cube of voltage risk covering a preset future time period for each dynamic cluster;

[0064] S4. The spatiotemporal cube is processed using a 3D convolutional neural network to extract spatiotemporal patterns of low voltage risk.

[0065] S5. The spatiotemporal pattern is segmented into independent risk events through multidimensional connected component analysis;

[0066] S6. The risk event list and its corresponding preset control strategies are automatically pushed to the scheduling and control system via the message bus to trigger the control equipment to take action.

[0067] This embodiment collects high-reliability transformer voltage data and performs pattern recognition on the voltage patterns of the transformer areas based on this data to determine whether the current system is in normal operating mode. If not, it is identified as a risk mode. A voltage risk spatiotemporal cube covering a preset future time period is constructed for each dynamic cluster. This spatiotemporal cube uses a two-dimensional space and a one-dimensional time to form a three-dimensional data structure. The voltage distribution of each node in the cluster in the future time period is obtained through time-series power flow calculation, and a continuous risk scoring function is used to quantify the voltage values ​​into risk indicators, forming a visualized spatiotemporal risk distribution map. A 3D convolutional neural network is used to perform deep processing on the spatiotemporal cube to extract spatiotemporal patches of low voltage risk, effectively capturing the propagation and evolution characteristics of risk in the spatiotemporal dimension. Through multidimensional connected component analysis, the spatiotemporal patches are segmented into independent risk events, and each event is labeled with spatiotemporal boundaries and risk intensity, ensuring the data authenticity of the risk identification process. By capturing the evolution of risk in the spatiotemporal dimension, the ability to capture and identify risks is improved, enhancing the overall efficiency and accuracy of risk identification.

[0068] The specific implementation steps are as follows:

[0069] S11. Obtain highly reliable three-phase voltage time series data from the measurement system, align the timestamps by phase, and form a structured phase-time two-dimensional matrix; for the voltage time series of each phase, apply the variational mode decomposition algorithm to decompose it into K eigenmode functions with different center frequencies, remove the highest frequency IMF component representing noise and the DC IMF component representing long-term trend, and reconstruct the remaining intermediate frequency IMF components to obtain a detrended and denoised pure voltage series.

[0070] S12. Construct a sliding window. In this embodiment, a 24-hour time window is set, with a step size of 1 hour, sliding over the denoised voltage sequence. Within each window, multi-dimensional morphological indicators are calculated, including statistical indicators such as mean, standard deviation, skewness, kurtosis, and root mean square, and physical indicators such as dynamic time warping (DTW) distance (compared with a typical curve), volatility, and load factor. The calculation results of all windows are integrated, and the daily average is taken to form the final voltage morphological feature vector.

[0071] S13. Taking each transformer substation as a node, according to the distribution network wiring diagram, if two substations are directly powered by the same feeder or connected through a low-voltage tie line, an edge is established between them to form an initial substation adjacency graph. Based on the distribution GIS system and conductor type parameters, a node admittance matrix is ​​established, and the node impedance matrix is ​​calculated using its inverse matrix or related algorithms. The mutual impedance values ​​between nodes are extracted as electrical distances, and the adjacent edges in the substation adjacency graph are weighted to obtain a weighted adjacency matrix. This matrix quantifies the electrical coupling strength between substations.

[0072] S14. Using the voltage morphology features obtained above as the initial features of each node, a two-layer GraphSAGE graph convolutional network is used. Each layer aggregates the features of the node itself and its first-order neighbors. After propagation through the network, each node obtains new features that fuse its own and its neighborhood information, i.e., spatially enhanced features.

[0073] S15. Synchronously collect time-series data concurrent with voltage data, including meteorological data (temperature, humidity, light intensity), regional total load data, and event data. Align all external variables to the same time granularity to form an external time-series vector. Use a bidirectional long short-term memory network as the time-series encoding network to encode the external time-series vector. This embodiment uses a Bi-LSTM encoding network, which can capture the dependencies between time series. The hidden state of the last time step is output as an external dynamic feature, which encodes the dynamic pattern of external environmental changes. Spatial augmented features are used as queries, and external dynamic features are used as keys and values. A multi-head cross-attention mechanism is used for computation, allowing the model to learn the complex relationship between spatial features and external features, and output a fused interactive feature.

[0074] S16. Calculate the Pearson correlation coefficient matrix between spatial enhancement features and external dynamic features. Adaptively adjust their weights in the final fusion based on the absolute value of the correlation coefficients. For example, increase the weight of feature pairs with high correlation and decrease the weight of those with low correlation. Concatenate the voltage morphology features (2n-dimensional), spatial enhancement features (2n-dimensional), external dynamic features (n-dimensional), and fusion interaction features (3n-dimensional) in sequence to obtain a comprehensive 8n-dimensional enhanced multidimensional feature vector describing the characteristics of the transformer area.

[0075] S21. From the enhanced multidimensional feature vector, select one or more core features with high discriminative power, including principal components in voltage morphology features to represent the core electricity consumption behavior patterns of the transformer substations, and the first few dimensions in spatial enhancement features to capture a rough geographical and electrical community structure. Then, use hierarchical clustering to perform initial coarse clustering on all transformer substations, forming M coarse clusters. Define an optimization problem containing at least three objective functions, which are used to simultaneously evaluate the quality of the clustering scheme in three dimensions: feature cohesion, geographical proximity, and electrical coupling strength. For each of the M coarse clusters, use a multi-objective optimization algorithm with the at least three objective functions as the optimization objectives. Within each coarse cluster, for each coarse cluster, determine the optimal number of clusters and the cluster label assignment for each transformer substation sample. Merge the final fine division results of all coarse clusters to obtain the final dynamic cluster division of all transformer substations in the entire region. Each final cluster exhibits high cohesion at the characteristics, geographical, and electrical levels, and can be business-tagged based on its typical characteristics to guide differentiated power grid planning and operation strategies.

[0076] S211. In the above content, at least three objective functions include: a first objective function, which is configured to minimize the sum of intra-cluster distances based on the characteristic space of the transformer area to maximize feature cohesion; a second objective function, which is configured to minimize the sum of intra-cluster distances based on the geographic space of the transformer area to maximize geographic proximity; and a third objective function, which is configured to maximize the intra-cluster electrical connection strength based on the power grid topology, or minimize the sum of intra-cluster electrical distances calculated based on electrical parameters to maximize electrical coupling strength.

[0077] S31. Determine the preset future time period and resolution, perform power grid modeling on the target dynamic cluster, input the complete power grid model data of the dynamic cluster, including: equipment parameters of transformers, lines, and switches, network topology, geographical location information of all load nodes and distributed power generation nodes, and perform time-series power flow calculation. Discretize the preset future time period into multiple continuous time segments, and obtain the voltage time-series data of each node in the cluster within the preset future time period; inject the load value and generation value of the current time segment into the corresponding node of the power grid model, perform a power flow calculation, solve for the steady-state operating point of the entire network, and extract the voltage amplitude of each node in the cluster at the current time segment from the power flow solution results. After the loop is completed, a three-dimensional data structure can be obtained, that is, the voltage time-series data of each node in the entire preset time period.

[0078] S32. Based on the voltage time series data, construct a predefined risk function to quantify the voltage risk value of each node at each time segment, that is, calculate the risk value of each node at each time segment.

[0079] S33. Based on the voltage risk value, construct a voltage risk spatiotemporal cube; wherein, the two dimensions of the spatiotemporal cube represent the spatial information of the cluster, the third dimension represents the temporal information, and the value of the cube element corresponds to the voltage risk value of the spatial location at the corresponding time section.

[0080] S34. Map all nodes in the cluster to a two-dimensional plane according to their latitude and longitude coordinates. This can divide the area into a regular geographic grid, with each grid cell representing a spatial location. Alternatively, take the geographic location of each node as a spatial cell to form an irregular spatial scatter layer. Arrange all time segments of the preset future period in sequence to form a time axis. Use nodes as spatial cells, traverse each spatial cell (or node) and each time segment, and fill the calculated risk value into the corresponding cell of the cube. Finally, generate a complete voltage risk spatiotemporal cube.

[0081] S41. Construct a pre-trained convolutional neural network model. Input the voltage risk spatiotemporal cube into the convolutional neural network model, normalize the cube data, and scale the risk values ​​to the [0,1] interval to accelerate model convergence and improve stability. In the convolutional neural network model, the encoder path is downsampled through 3D convolutional layers and 3D pooling layers to extract multi-scale spatiotemporal features. The decoder path is upsampled through 3D upsampling layers and jump-connected with the feature maps of the corresponding scale in the encoder path to restore spatial details and output a spatiotemporal patch map with the same size as the input. The spatiotemporal patch map is used to determine the probability that each spatiotemporal point in the spatiotemporal cube belongs to the core risk patch map, providing an accurate and high-quality data foundation for subsequent risk event segmentation. The spatiotemporal patch map filters out noise in the original data and highlights the coherent low-voltage risk core region with significant spatiotemporal structure.

[0082] S51. To transform continuous probability outputs into explicit yes / no judgments, a probability threshold is set. If a voxel in the spatiotemporal patch has a probability value greater than the threshold, it is set to a first value at the corresponding position in the three-dimensional risk mask, representing the foreground of a risk event. If a voxel has a probability value less than or equal to the threshold, it is set to a second value at the corresponding position in the three-dimensional risk mask, representing the background. The output is a binary matrix with the same size as the spatiotemporal cube, where the connected regions with the first value are potential risk events.

[0083] S52. The connectivity criteria are used to determine the connectivity between voxels in the 3D risk mask. Each voxel in the 3D risk mask is traversed sequentially. For a voxel with the first value, its visited neighbors are checked. Based on the neighbors' labels, a disjoint-set data structure is applied to manage equivalent label relationships. Then, a second scan is performed: the disjoint-set is parsed, equivalent labels are merged, and the temporary labels in each connected region are replaced with the final, unique label values. A new 3D label array is output. All voxels with a value of 0 remain 0, and each isolated connected region consisting of the first value is assigned a unique integer label. Each unique label represents an independent low-voltage risk event.

[0084] As one possible implementation, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a spatiotemporally correlated method for identifying transformer area voltage risk events.

[0085] As one possible implementation, this embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a spatiotemporally correlated method for identifying transformer area voltage risk events.

[0086] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying a time-space correlated substation voltage risk event, characterized in that, The method comprises the following specific steps: Obtaining high-trust substation area voltage data, and constructing an enhanced multi-dimensional feature vector for each substation area based on the high-trust substation area voltage data, the enhanced multi-dimensional feature vector fusing voltage patterns, spatial correlations and dynamic external correlations; Determining an optimal cluster number of the enhanced multi-dimensional feature vector based on a multi-objective optimization algorithm, and obtaining dynamic clusters that divide the substation areas into clusters that have internal cohesion in the aspects of features, geography and electricity; Constructing a voltage risk space-time cube covering a preset time period in the future for each dynamic cluster; Processing the space-time cube by using a 3D convolutional neural network, and extracting a low-voltage risk space-time pattern; Segmenting the space-time pattern into independent risk events by using multi-dimensional connected domain analysis; Pushing a risk event list and a corresponding preset control strategy to a dispatching control system, and triggering a control device to act.

2. The spatio-temporally correlated zonal voltage risk event identification method of claim 1, wherein, The method of constructing an enhanced multi-dimensional feature vector for each substation area based on high-trust substation area voltage data comprises the following steps: Based on the high-trust substation area voltage data, a phase-time matrix synchronized in time is extracted; Each phase voltage sequence is subjected to variational mode decomposition, and after the highest frequency and direct current intrinsic mode are removed, a voltage sequence after detrending and denoising is obtained; A sliding window is constructed, and multi-dimensional pattern indexes are extracted based on the voltage sequence; A substation area adjacency graph is constructed, taking the substation areas as nodes and taking the same feeder or low-voltage connection as an initial edge; A node impedance matrix is calculated based on distribution GIS and conductor parameters, and the edges are weighted according to electrical distances, to obtain a weighted adjacency matrix; The voltage pattern features of the nodes themselves and the features of the neighborhoods of the nodes are aggregated by using a graph convolution network, to generate spatial enhanced features; Meteorological, load and event type external variables are synchronously collected, to form external time series vectors; The external time series vectors are encoded by using a time series encoding network, to obtain external dynamic features; The spatial enhanced features are taken as queries, the external dynamic features are taken as keys and values, a multi-head cross attention mechanism is used for calculation, and fused interaction features are outputted; Correlation coefficients of the spatial enhanced features and the external dynamic features are calculated, and the fused weights are adaptively adjusted according to the correlation coefficients; The voltage pattern features, the spatial enhanced features, the external dynamic features and the fused interaction features are spliced, to obtain an enhanced multi-dimensional feature vector.

3. The spatio-temporally correlated zonal voltage risk event identification method of claim 1, wherein, The method of determining an optimal cluster number of the enhanced multi-dimensional feature vector based on a multi-objective optimization algorithm comprises the following steps: Based on one or more key features of the enhanced multi-dimensional feature vector, a first clustering algorithm is used to perform initial coarse clustering on all substation areas, to form M coarse clustering clusters; An optimization problem including at least three objective functions is defined, the objective functions being used to evaluate the quality of a clustering scheme in three dimensions of feature cohesion, geographical proximity and electrical coupling strength; For each of the M coarse clustering clusters, a multi-objective optimization algorithm is used to perform fine clustering in the coarse clustering cluster, taking the at least three objective functions as optimization objectives, to obtain dynamic clusters that divide the substation areas into clusters that have internal cohesion in the aspects of features, geography and electricity.

4. The spatio-temporally correlated zonal voltage risk event identification method of claim 3, wherein, The at least three objective functions include: A first objective function configured to minimize the sum of intra-cluster distances based on a substation feature space, to maximize feature cohesion; a second objective function configured to minimize the sum of intra-cluster distances based on geographical space of the feeder area to maximize geographical proximity; a third objective function configured to maximize intra-cluster electrical connection strength based on power grid topology, or minimize the sum of intra-cluster electrical distances calculated based on electrical parameters, to maximize electrical coupling strength.

5. The spatio-temporally correlated zonal voltage risk event identification method of claim 1, wherein, constructing a voltage risk spatio-temporal cube covering a preset future period for each dynamic cluster, specifically comprising: performing power grid modeling on the target dynamic cluster and executing time-series power flow calculation to obtain voltage time-series data of each node in the cluster within the preset future period; constructing a predefined risk function to quantify voltage risk values of each node at each time section based on the voltage time-series data; constructing a voltage risk spatio-temporal cube based on the voltage risk values; wherein two dimensions of the spatio-temporal cube represent spatial information of the cluster, and the third dimension represents time information, and the value of a cube unit corresponds to the voltage risk value of the spatial position at the corresponding time section.

6. The spatio-temporally correlated zonal voltage risk event identification method of claim 5, wherein, performing time-series power flow calculation, specifically comprising: discretizing the preset future period into multiple continuous time sections; for each time section, injecting the predicted load value and new energy generation value of the time section into the power grid model; performing power flow calculation to solve and obtain the voltage amplitude of each node in the cluster at the time section.

7. The spatio-temporally correlated zonal voltage risk event identification method of claim 1, wherein, processing the spatio-temporal cube using a 3D convolutional neural network to extract a low-voltage risk spatio-temporal pattern, specifically comprising: inputting the voltage risk spatio-temporal cube into a convolutional neural network model, and an encoder path performs down-sampling through 3D convolutional layers and 3D pooling layers to extract multi-scale spatio-temporal features, and a decoder path performs up-sampling through 3D up-sampling layers and performs skip connection with feature maps of corresponding scales in the encoder path to restore spatial details and output a spatio-temporal pattern with the same input size; the spatio-temporal pattern is used for the probability that each spatio-temporal point in the spatio-temporal cube belongs to a core risk pattern.

8. The spatio-temporally correlated zonal voltage risk event identification method of claim 1, wherein, segmenting the spatio-temporal pattern into independent risk events through multi-dimensional connected component analysis; setting a probability threshold, setting voxels with probability values greater than the threshold in the spatio-temporal pattern to a first value representing risk event foreground, and setting voxels with probability values less than or equal to the threshold to a second value representing background; generating a three-dimensional risk mask based on the binary data of the risk event foreground and background; judging the connectivity between voxels in the three-dimensional risk mask using a connectivity criterion to identify and label mutually isolated connected regions therein, wherein each labeled connected region is defined as an independent low-voltage risk event.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the spatio-temporally correlated feeder voltage risk event identification method of any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the spatio-temporally correlated feeder voltage risk event identification method of any one of claims 1-8.

Images